5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
88 #include "dquote_static.c"
95 # if defined(BUGGY_MSC6)
96 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
97 # pragma optimize("a",off)
98 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
99 # pragma optimize("w",on )
100 # endif /* BUGGY_MSC6 */
104 #define STATIC static
107 typedef struct RExC_state_t {
108 U32 flags; /* are we folding, multilining? */
109 char *precomp; /* uncompiled string. */
110 REGEXP *rx_sv; /* The SV that is the regexp. */
111 regexp *rx; /* perl core regexp structure */
112 regexp_internal *rxi; /* internal data for regexp object pprivate field */
113 char *start; /* Start of input for compile */
114 char *end; /* End of input for compile */
115 char *parse; /* Input-scan pointer. */
116 I32 whilem_seen; /* number of WHILEM in this expr */
117 regnode *emit_start; /* Start of emitted-code area */
118 regnode *emit_bound; /* First regnode outside of the allocated space */
119 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
120 I32 naughty; /* How bad is this pattern? */
121 I32 sawback; /* Did we see \1, ...? */
123 I32 size; /* Code size. */
124 I32 npar; /* Capture buffer count, (OPEN). */
125 I32 cpar; /* Capture buffer count, (CLOSE). */
126 I32 nestroot; /* root parens we are in - used by accept */
130 regnode **open_parens; /* pointers to open parens */
131 regnode **close_parens; /* pointers to close parens */
132 regnode *opend; /* END node in program */
133 I32 utf8; /* whether the pattern is utf8 or not */
134 I32 orig_utf8; /* whether the pattern was originally in utf8 */
135 /* XXX use this for future optimisation of case
136 * where pattern must be upgraded to utf8. */
137 I32 uni_semantics; /* If a d charset modifier should use unicode
138 rules, even if the pattern is not in
140 HV *paren_names; /* Paren names */
142 regnode **recurse; /* Recurse regops */
143 I32 recurse_count; /* Number of recurse regops */
146 I32 override_recoding;
148 char *starttry; /* -Dr: where regtry was called. */
149 #define RExC_starttry (pRExC_state->starttry)
152 const char *lastparse;
154 AV *paren_name_list; /* idx -> name */
155 #define RExC_lastparse (pRExC_state->lastparse)
156 #define RExC_lastnum (pRExC_state->lastnum)
157 #define RExC_paren_name_list (pRExC_state->paren_name_list)
161 #define RExC_flags (pRExC_state->flags)
162 #define RExC_precomp (pRExC_state->precomp)
163 #define RExC_rx_sv (pRExC_state->rx_sv)
164 #define RExC_rx (pRExC_state->rx)
165 #define RExC_rxi (pRExC_state->rxi)
166 #define RExC_start (pRExC_state->start)
167 #define RExC_end (pRExC_state->end)
168 #define RExC_parse (pRExC_state->parse)
169 #define RExC_whilem_seen (pRExC_state->whilem_seen)
170 #ifdef RE_TRACK_PATTERN_OFFSETS
171 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
173 #define RExC_emit (pRExC_state->emit)
174 #define RExC_emit_start (pRExC_state->emit_start)
175 #define RExC_emit_bound (pRExC_state->emit_bound)
176 #define RExC_naughty (pRExC_state->naughty)
177 #define RExC_sawback (pRExC_state->sawback)
178 #define RExC_seen (pRExC_state->seen)
179 #define RExC_size (pRExC_state->size)
180 #define RExC_npar (pRExC_state->npar)
181 #define RExC_nestroot (pRExC_state->nestroot)
182 #define RExC_extralen (pRExC_state->extralen)
183 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
184 #define RExC_seen_evals (pRExC_state->seen_evals)
185 #define RExC_utf8 (pRExC_state->utf8)
186 #define RExC_uni_semantics (pRExC_state->uni_semantics)
187 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
188 #define RExC_open_parens (pRExC_state->open_parens)
189 #define RExC_close_parens (pRExC_state->close_parens)
190 #define RExC_opend (pRExC_state->opend)
191 #define RExC_paren_names (pRExC_state->paren_names)
192 #define RExC_recurse (pRExC_state->recurse)
193 #define RExC_recurse_count (pRExC_state->recurse_count)
194 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
195 #define RExC_contains_locale (pRExC_state->contains_locale)
196 #define RExC_override_recoding (pRExC_state->override_recoding)
199 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
200 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
201 ((*s) == '{' && regcurly(s)))
204 #undef SPSTART /* dratted cpp namespace... */
207 * Flags to be passed up and down.
209 #define WORST 0 /* Worst case. */
210 #define HASWIDTH 0x01 /* Known to match non-null strings. */
212 /* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
213 * character, and if utf8, must be invariant. Note that this is not the same thing as REGNODE_SIMPLE */
215 #define SPSTART 0x04 /* Starts with * or +. */
216 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
217 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
219 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
221 /* whether trie related optimizations are enabled */
222 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
223 #define TRIE_STUDY_OPT
224 #define FULL_TRIE_STUDY
230 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
231 #define PBITVAL(paren) (1 << ((paren) & 7))
232 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
233 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
234 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
236 /* If not already in utf8, do a longjmp back to the beginning */
237 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
238 #define REQUIRE_UTF8 STMT_START { \
239 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
242 /* About scan_data_t.
244 During optimisation we recurse through the regexp program performing
245 various inplace (keyhole style) optimisations. In addition study_chunk
246 and scan_commit populate this data structure with information about
247 what strings MUST appear in the pattern. We look for the longest
248 string that must appear at a fixed location, and we look for the
249 longest string that may appear at a floating location. So for instance
254 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
255 strings (because they follow a .* construct). study_chunk will identify
256 both FOO and BAR as being the longest fixed and floating strings respectively.
258 The strings can be composites, for instance
262 will result in a composite fixed substring 'foo'.
264 For each string some basic information is maintained:
266 - offset or min_offset
267 This is the position the string must appear at, or not before.
268 It also implicitly (when combined with minlenp) tells us how many
269 characters must match before the string we are searching for.
270 Likewise when combined with minlenp and the length of the string it
271 tells us how many characters must appear after the string we have
275 Only used for floating strings. This is the rightmost point that
276 the string can appear at. If set to I32 max it indicates that the
277 string can occur infinitely far to the right.
280 A pointer to the minimum length of the pattern that the string
281 was found inside. This is important as in the case of positive
282 lookahead or positive lookbehind we can have multiple patterns
287 The minimum length of the pattern overall is 3, the minimum length
288 of the lookahead part is 3, but the minimum length of the part that
289 will actually match is 1. So 'FOO's minimum length is 3, but the
290 minimum length for the F is 1. This is important as the minimum length
291 is used to determine offsets in front of and behind the string being
292 looked for. Since strings can be composites this is the length of the
293 pattern at the time it was committed with a scan_commit. Note that
294 the length is calculated by study_chunk, so that the minimum lengths
295 are not known until the full pattern has been compiled, thus the
296 pointer to the value.
300 In the case of lookbehind the string being searched for can be
301 offset past the start point of the final matching string.
302 If this value was just blithely removed from the min_offset it would
303 invalidate some of the calculations for how many chars must match
304 before or after (as they are derived from min_offset and minlen and
305 the length of the string being searched for).
306 When the final pattern is compiled and the data is moved from the
307 scan_data_t structure into the regexp structure the information
308 about lookbehind is factored in, with the information that would
309 have been lost precalculated in the end_shift field for the
312 The fields pos_min and pos_delta are used to store the minimum offset
313 and the delta to the maximum offset at the current point in the pattern.
317 typedef struct scan_data_t {
318 /*I32 len_min; unused */
319 /*I32 len_delta; unused */
323 I32 last_end; /* min value, <0 unless valid. */
326 SV **longest; /* Either &l_fixed, or &l_float. */
327 SV *longest_fixed; /* longest fixed string found in pattern */
328 I32 offset_fixed; /* offset where it starts */
329 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
330 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
331 SV *longest_float; /* longest floating string found in pattern */
332 I32 offset_float_min; /* earliest point in string it can appear */
333 I32 offset_float_max; /* latest point in string it can appear */
334 I32 *minlen_float; /* pointer to the minlen relevant to the string */
335 I32 lookbehind_float; /* is the position of the string modified by LB */
339 struct regnode_charclass_class *start_class;
343 * Forward declarations for pregcomp()'s friends.
346 static const scan_data_t zero_scan_data =
347 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
349 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
350 #define SF_BEFORE_SEOL 0x0001
351 #define SF_BEFORE_MEOL 0x0002
352 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
353 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
356 # define SF_FIX_SHIFT_EOL (0+2)
357 # define SF_FL_SHIFT_EOL (0+4)
359 # define SF_FIX_SHIFT_EOL (+2)
360 # define SF_FL_SHIFT_EOL (+4)
363 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
364 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
366 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
367 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
368 #define SF_IS_INF 0x0040
369 #define SF_HAS_PAR 0x0080
370 #define SF_IN_PAR 0x0100
371 #define SF_HAS_EVAL 0x0200
372 #define SCF_DO_SUBSTR 0x0400
373 #define SCF_DO_STCLASS_AND 0x0800
374 #define SCF_DO_STCLASS_OR 0x1000
375 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
376 #define SCF_WHILEM_VISITED_POS 0x2000
378 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
379 #define SCF_SEEN_ACCEPT 0x8000
381 #define UTF cBOOL(RExC_utf8)
382 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
383 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
384 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
385 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
386 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
387 #define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
388 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
390 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
392 #define OOB_UNICODE 12345678
393 #define OOB_NAMEDCLASS -1
395 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
396 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
399 /* length of regex to show in messages that don't mark a position within */
400 #define RegexLengthToShowInErrorMessages 127
403 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
404 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
405 * op/pragma/warn/regcomp.
407 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
408 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
410 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
413 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
414 * arg. Show regex, up to a maximum length. If it's too long, chop and add
417 #define _FAIL(code) STMT_START { \
418 const char *ellipses = ""; \
419 IV len = RExC_end - RExC_precomp; \
422 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
423 if (len > RegexLengthToShowInErrorMessages) { \
424 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
425 len = RegexLengthToShowInErrorMessages - 10; \
431 #define FAIL(msg) _FAIL( \
432 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
433 msg, (int)len, RExC_precomp, ellipses))
435 #define FAIL2(msg,arg) _FAIL( \
436 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
437 arg, (int)len, RExC_precomp, ellipses))
440 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
442 #define Simple_vFAIL(m) STMT_START { \
443 const IV offset = RExC_parse - RExC_precomp; \
444 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
445 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
449 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
451 #define vFAIL(m) STMT_START { \
453 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
458 * Like Simple_vFAIL(), but accepts two arguments.
460 #define Simple_vFAIL2(m,a1) STMT_START { \
461 const IV offset = RExC_parse - RExC_precomp; \
462 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
463 (int)offset, RExC_precomp, RExC_precomp + offset); \
467 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
469 #define vFAIL2(m,a1) STMT_START { \
471 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
472 Simple_vFAIL2(m, a1); \
477 * Like Simple_vFAIL(), but accepts three arguments.
479 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
480 const IV offset = RExC_parse - RExC_precomp; \
481 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
482 (int)offset, RExC_precomp, RExC_precomp + offset); \
486 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
488 #define vFAIL3(m,a1,a2) STMT_START { \
490 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
491 Simple_vFAIL3(m, a1, a2); \
495 * Like Simple_vFAIL(), but accepts four arguments.
497 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
498 const IV offset = RExC_parse - RExC_precomp; \
499 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
500 (int)offset, RExC_precomp, RExC_precomp + offset); \
503 #define ckWARNreg(loc,m) STMT_START { \
504 const IV offset = loc - RExC_precomp; \
505 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
506 (int)offset, RExC_precomp, RExC_precomp + offset); \
509 #define ckWARNregdep(loc,m) STMT_START { \
510 const IV offset = loc - RExC_precomp; \
511 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
513 (int)offset, RExC_precomp, RExC_precomp + offset); \
516 #define ckWARN2regdep(loc,m, a1) STMT_START { \
517 const IV offset = loc - RExC_precomp; \
518 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
520 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
523 #define ckWARN2reg(loc, m, a1) STMT_START { \
524 const IV offset = loc - RExC_precomp; \
525 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
526 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
529 #define vWARN3(loc, m, a1, a2) STMT_START { \
530 const IV offset = loc - RExC_precomp; \
531 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
532 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
535 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
536 const IV offset = loc - RExC_precomp; \
537 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
538 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
541 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
542 const IV offset = loc - RExC_precomp; \
543 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
544 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
547 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
548 const IV offset = loc - RExC_precomp; \
549 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
550 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
553 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
554 const IV offset = loc - RExC_precomp; \
555 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
556 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
560 /* Allow for side effects in s */
561 #define REGC(c,s) STMT_START { \
562 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
565 /* Macros for recording node offsets. 20001227 mjd@plover.com
566 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
567 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
568 * Element 0 holds the number n.
569 * Position is 1 indexed.
571 #ifndef RE_TRACK_PATTERN_OFFSETS
572 #define Set_Node_Offset_To_R(node,byte)
573 #define Set_Node_Offset(node,byte)
574 #define Set_Cur_Node_Offset
575 #define Set_Node_Length_To_R(node,len)
576 #define Set_Node_Length(node,len)
577 #define Set_Node_Cur_Length(node)
578 #define Node_Offset(n)
579 #define Node_Length(n)
580 #define Set_Node_Offset_Length(node,offset,len)
581 #define ProgLen(ri) ri->u.proglen
582 #define SetProgLen(ri,x) ri->u.proglen = x
584 #define ProgLen(ri) ri->u.offsets[0]
585 #define SetProgLen(ri,x) ri->u.offsets[0] = x
586 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
588 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
589 __LINE__, (int)(node), (int)(byte))); \
591 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
593 RExC_offsets[2*(node)-1] = (byte); \
598 #define Set_Node_Offset(node,byte) \
599 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
600 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
602 #define Set_Node_Length_To_R(node,len) STMT_START { \
604 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
605 __LINE__, (int)(node), (int)(len))); \
607 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
609 RExC_offsets[2*(node)] = (len); \
614 #define Set_Node_Length(node,len) \
615 Set_Node_Length_To_R((node)-RExC_emit_start, len)
616 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
617 #define Set_Node_Cur_Length(node) \
618 Set_Node_Length(node, RExC_parse - parse_start)
620 /* Get offsets and lengths */
621 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
622 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
624 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
625 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
626 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
630 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
631 #define EXPERIMENTAL_INPLACESCAN
632 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
634 #define DEBUG_STUDYDATA(str,data,depth) \
635 DEBUG_OPTIMISE_MORE_r(if(data){ \
636 PerlIO_printf(Perl_debug_log, \
637 "%*s" str "Pos:%"IVdf"/%"IVdf \
638 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
639 (int)(depth)*2, "", \
640 (IV)((data)->pos_min), \
641 (IV)((data)->pos_delta), \
642 (UV)((data)->flags), \
643 (IV)((data)->whilem_c), \
644 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
645 is_inf ? "INF " : "" \
647 if ((data)->last_found) \
648 PerlIO_printf(Perl_debug_log, \
649 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
650 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
651 SvPVX_const((data)->last_found), \
652 (IV)((data)->last_end), \
653 (IV)((data)->last_start_min), \
654 (IV)((data)->last_start_max), \
655 ((data)->longest && \
656 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
657 SvPVX_const((data)->longest_fixed), \
658 (IV)((data)->offset_fixed), \
659 ((data)->longest && \
660 (data)->longest==&((data)->longest_float)) ? "*" : "", \
661 SvPVX_const((data)->longest_float), \
662 (IV)((data)->offset_float_min), \
663 (IV)((data)->offset_float_max) \
665 PerlIO_printf(Perl_debug_log,"\n"); \
668 static void clear_re(pTHX_ void *r);
670 /* Mark that we cannot extend a found fixed substring at this point.
671 Update the longest found anchored substring and the longest found
672 floating substrings if needed. */
675 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
677 const STRLEN l = CHR_SVLEN(data->last_found);
678 const STRLEN old_l = CHR_SVLEN(*data->longest);
679 GET_RE_DEBUG_FLAGS_DECL;
681 PERL_ARGS_ASSERT_SCAN_COMMIT;
683 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
684 SvSetMagicSV(*data->longest, data->last_found);
685 if (*data->longest == data->longest_fixed) {
686 data->offset_fixed = l ? data->last_start_min : data->pos_min;
687 if (data->flags & SF_BEFORE_EOL)
689 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
691 data->flags &= ~SF_FIX_BEFORE_EOL;
692 data->minlen_fixed=minlenp;
693 data->lookbehind_fixed=0;
695 else { /* *data->longest == data->longest_float */
696 data->offset_float_min = l ? data->last_start_min : data->pos_min;
697 data->offset_float_max = (l
698 ? data->last_start_max
699 : data->pos_min + data->pos_delta);
700 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
701 data->offset_float_max = I32_MAX;
702 if (data->flags & SF_BEFORE_EOL)
704 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
706 data->flags &= ~SF_FL_BEFORE_EOL;
707 data->minlen_float=minlenp;
708 data->lookbehind_float=0;
711 SvCUR_set(data->last_found, 0);
713 SV * const sv = data->last_found;
714 if (SvUTF8(sv) && SvMAGICAL(sv)) {
715 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
721 data->flags &= ~SF_BEFORE_EOL;
722 DEBUG_STUDYDATA("commit: ",data,0);
725 /* Can match anything (initialization) */
727 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
729 PERL_ARGS_ASSERT_CL_ANYTHING;
731 ANYOF_BITMAP_SETALL(cl);
732 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
733 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
735 /* If any portion of the regex is to operate under locale rules,
736 * initialization includes it. The reason this isn't done for all regexes
737 * is that the optimizer was written under the assumption that locale was
738 * all-or-nothing. Given the complexity and lack of documentation in the
739 * optimizer, and that there are inadequate test cases for locale, so many
740 * parts of it may not work properly, it is safest to avoid locale unless
742 if (RExC_contains_locale) {
743 ANYOF_CLASS_SETALL(cl); /* /l uses class */
744 cl->flags |= ANYOF_LOCALE;
747 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
751 /* Can match anything (initialization) */
753 S_cl_is_anything(const struct regnode_charclass_class *cl)
757 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
759 for (value = 0; value <= ANYOF_MAX; value += 2)
760 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
762 if (!(cl->flags & ANYOF_UNICODE_ALL))
764 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
769 /* Can match anything (initialization) */
771 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
773 PERL_ARGS_ASSERT_CL_INIT;
775 Zero(cl, 1, struct regnode_charclass_class);
777 cl_anything(pRExC_state, cl);
778 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
781 /* These two functions currently do the exact same thing */
782 #define cl_init_zero S_cl_init
784 /* 'AND' a given class with another one. Can create false positives. 'cl'
785 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
786 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
788 S_cl_and(struct regnode_charclass_class *cl,
789 const struct regnode_charclass_class *and_with)
791 PERL_ARGS_ASSERT_CL_AND;
793 assert(and_with->type == ANYOF);
795 /* I (khw) am not sure all these restrictions are necessary XXX */
796 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
797 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
798 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
799 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
800 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
803 if (and_with->flags & ANYOF_INVERT)
804 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
805 cl->bitmap[i] &= ~and_with->bitmap[i];
807 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
808 cl->bitmap[i] &= and_with->bitmap[i];
809 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
811 if (and_with->flags & ANYOF_INVERT) {
813 /* Here, the and'ed node is inverted. Get the AND of the flags that
814 * aren't affected by the inversion. Those that are affected are
815 * handled individually below */
816 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
817 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
818 cl->flags |= affected_flags;
820 /* We currently don't know how to deal with things that aren't in the
821 * bitmap, but we know that the intersection is no greater than what
822 * is already in cl, so let there be false positives that get sorted
823 * out after the synthetic start class succeeds, and the node is
824 * matched for real. */
826 /* The inversion of these two flags indicate that the resulting
827 * intersection doesn't have them */
828 if (and_with->flags & ANYOF_UNICODE_ALL) {
829 cl->flags &= ~ANYOF_UNICODE_ALL;
831 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
832 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
835 else { /* and'd node is not inverted */
836 U8 outside_bitmap_but_not_utf8; /* Temp variable */
838 if (! ANYOF_NONBITMAP(and_with)) {
840 /* Here 'and_with' doesn't match anything outside the bitmap
841 * (except possibly ANYOF_UNICODE_ALL), which means the
842 * intersection can't either, except for ANYOF_UNICODE_ALL, in
843 * which case we don't know what the intersection is, but it's no
844 * greater than what cl already has, so can just leave it alone,
845 * with possible false positives */
846 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
847 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
848 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
851 else if (! ANYOF_NONBITMAP(cl)) {
853 /* Here, 'and_with' does match something outside the bitmap, and cl
854 * doesn't have a list of things to match outside the bitmap. If
855 * cl can match all code points above 255, the intersection will
856 * be those above-255 code points that 'and_with' matches. If cl
857 * can't match all Unicode code points, it means that it can't
858 * match anything outside the bitmap (since the 'if' that got us
859 * into this block tested for that), so we leave the bitmap empty.
861 if (cl->flags & ANYOF_UNICODE_ALL) {
862 ARG_SET(cl, ARG(and_with));
864 /* and_with's ARG may match things that don't require UTF8.
865 * And now cl's will too, in spite of this being an 'and'. See
866 * the comments below about the kludge */
867 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
871 /* Here, both 'and_with' and cl match something outside the
872 * bitmap. Currently we do not do the intersection, so just match
873 * whatever cl had at the beginning. */
877 /* Take the intersection of the two sets of flags. However, the
878 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
879 * kludge around the fact that this flag is not treated like the others
880 * which are initialized in cl_anything(). The way the optimizer works
881 * is that the synthetic start class (SSC) is initialized to match
882 * anything, and then the first time a real node is encountered, its
883 * values are AND'd with the SSC's with the result being the values of
884 * the real node. However, there are paths through the optimizer where
885 * the AND never gets called, so those initialized bits are set
886 * inappropriately, which is not usually a big deal, as they just cause
887 * false positives in the SSC, which will just mean a probably
888 * imperceptible slow down in execution. However this bit has a
889 * higher false positive consequence in that it can cause utf8.pm,
890 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
891 * bigger slowdown and also causes significant extra memory to be used.
892 * In order to prevent this, the code now takes a different tack. The
893 * bit isn't set unless some part of the regular expression needs it,
894 * but once set it won't get cleared. This means that these extra
895 * modules won't get loaded unless there was some path through the
896 * pattern that would have required them anyway, and so any false
897 * positives that occur by not ANDing them out when they could be
898 * aren't as severe as they would be if we treated this bit like all
900 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
901 & ANYOF_NONBITMAP_NON_UTF8;
902 cl->flags &= and_with->flags;
903 cl->flags |= outside_bitmap_but_not_utf8;
907 /* 'OR' a given class with another one. Can create false positives. 'cl'
908 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
909 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
911 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
913 PERL_ARGS_ASSERT_CL_OR;
915 if (or_with->flags & ANYOF_INVERT) {
917 /* Here, the or'd node is to be inverted. This means we take the
918 * complement of everything not in the bitmap, but currently we don't
919 * know what that is, so give up and match anything */
920 if (ANYOF_NONBITMAP(or_with)) {
921 cl_anything(pRExC_state, cl);
924 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
925 * <= (B1 | !B2) | (CL1 | !CL2)
926 * which is wasteful if CL2 is small, but we ignore CL2:
927 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
928 * XXXX Can we handle case-fold? Unclear:
929 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
930 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
932 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
933 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
934 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
937 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
938 cl->bitmap[i] |= ~or_with->bitmap[i];
939 } /* XXXX: logic is complicated otherwise */
941 cl_anything(pRExC_state, cl);
944 /* And, we can just take the union of the flags that aren't affected
945 * by the inversion */
946 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
948 /* For the remaining flags:
949 ANYOF_UNICODE_ALL and inverted means to not match anything above
950 255, which means that the union with cl should just be
951 what cl has in it, so can ignore this flag
952 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
953 is 127-255 to match them, but then invert that, so the
954 union with cl should just be what cl has in it, so can
957 } else { /* 'or_with' is not inverted */
958 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
959 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
960 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
961 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
964 /* OR char bitmap and class bitmap separately */
965 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
966 cl->bitmap[i] |= or_with->bitmap[i];
967 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
968 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
969 cl->classflags[i] |= or_with->classflags[i];
970 cl->flags |= ANYOF_CLASS;
973 else { /* XXXX: logic is complicated, leave it along for a moment. */
974 cl_anything(pRExC_state, cl);
977 if (ANYOF_NONBITMAP(or_with)) {
979 /* Use the added node's outside-the-bit-map match if there isn't a
980 * conflict. If there is a conflict (both nodes match something
981 * outside the bitmap, but what they match outside is not the same
982 * pointer, and hence not easily compared until XXX we extend
983 * inversion lists this far), give up and allow the start class to
984 * match everything outside the bitmap. If that stuff is all above
985 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
986 if (! ANYOF_NONBITMAP(cl)) {
987 ARG_SET(cl, ARG(or_with));
989 else if (ARG(cl) != ARG(or_with)) {
991 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
992 cl_anything(pRExC_state, cl);
995 cl->flags |= ANYOF_UNICODE_ALL;
1000 /* Take the union */
1001 cl->flags |= or_with->flags;
1005 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1006 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1007 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1008 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1013 dump_trie(trie,widecharmap,revcharmap)
1014 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1015 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1017 These routines dump out a trie in a somewhat readable format.
1018 The _interim_ variants are used for debugging the interim
1019 tables that are used to generate the final compressed
1020 representation which is what dump_trie expects.
1022 Part of the reason for their existence is to provide a form
1023 of documentation as to how the different representations function.
1028 Dumps the final compressed table form of the trie to Perl_debug_log.
1029 Used for debugging make_trie().
1033 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1034 AV *revcharmap, U32 depth)
1037 SV *sv=sv_newmortal();
1038 int colwidth= widecharmap ? 6 : 4;
1040 GET_RE_DEBUG_FLAGS_DECL;
1042 PERL_ARGS_ASSERT_DUMP_TRIE;
1044 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1045 (int)depth * 2 + 2,"",
1046 "Match","Base","Ofs" );
1048 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1049 SV ** const tmp = av_fetch( revcharmap, state, 0);
1051 PerlIO_printf( Perl_debug_log, "%*s",
1053 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1054 PL_colors[0], PL_colors[1],
1055 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1056 PERL_PV_ESCAPE_FIRSTCHAR
1061 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1062 (int)depth * 2 + 2,"");
1064 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1065 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1066 PerlIO_printf( Perl_debug_log, "\n");
1068 for( state = 1 ; state < trie->statecount ; state++ ) {
1069 const U32 base = trie->states[ state ].trans.base;
1071 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1073 if ( trie->states[ state ].wordnum ) {
1074 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1076 PerlIO_printf( Perl_debug_log, "%6s", "" );
1079 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1084 while( ( base + ofs < trie->uniquecharcount ) ||
1085 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1086 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1089 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1091 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1092 if ( ( base + ofs >= trie->uniquecharcount ) &&
1093 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1094 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1096 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1098 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1100 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1104 PerlIO_printf( Perl_debug_log, "]");
1107 PerlIO_printf( Perl_debug_log, "\n" );
1109 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1110 for (word=1; word <= trie->wordcount; word++) {
1111 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1112 (int)word, (int)(trie->wordinfo[word].prev),
1113 (int)(trie->wordinfo[word].len));
1115 PerlIO_printf(Perl_debug_log, "\n" );
1118 Dumps a fully constructed but uncompressed trie in list form.
1119 List tries normally only are used for construction when the number of
1120 possible chars (trie->uniquecharcount) is very high.
1121 Used for debugging make_trie().
1124 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1125 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1129 SV *sv=sv_newmortal();
1130 int colwidth= widecharmap ? 6 : 4;
1131 GET_RE_DEBUG_FLAGS_DECL;
1133 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1135 /* print out the table precompression. */
1136 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1137 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1138 "------:-----+-----------------\n" );
1140 for( state=1 ; state < next_alloc ; state ++ ) {
1143 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1144 (int)depth * 2 + 2,"", (UV)state );
1145 if ( ! trie->states[ state ].wordnum ) {
1146 PerlIO_printf( Perl_debug_log, "%5s| ","");
1148 PerlIO_printf( Perl_debug_log, "W%4x| ",
1149 trie->states[ state ].wordnum
1152 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1153 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1155 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1157 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1158 PL_colors[0], PL_colors[1],
1159 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1160 PERL_PV_ESCAPE_FIRSTCHAR
1162 TRIE_LIST_ITEM(state,charid).forid,
1163 (UV)TRIE_LIST_ITEM(state,charid).newstate
1166 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1167 (int)((depth * 2) + 14), "");
1170 PerlIO_printf( Perl_debug_log, "\n");
1175 Dumps a fully constructed but uncompressed trie in table form.
1176 This is the normal DFA style state transition table, with a few
1177 twists to facilitate compression later.
1178 Used for debugging make_trie().
1181 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1182 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1187 SV *sv=sv_newmortal();
1188 int colwidth= widecharmap ? 6 : 4;
1189 GET_RE_DEBUG_FLAGS_DECL;
1191 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1194 print out the table precompression so that we can do a visual check
1195 that they are identical.
1198 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1200 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1201 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1203 PerlIO_printf( Perl_debug_log, "%*s",
1205 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1206 PL_colors[0], PL_colors[1],
1207 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1208 PERL_PV_ESCAPE_FIRSTCHAR
1214 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1216 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1217 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1220 PerlIO_printf( Perl_debug_log, "\n" );
1222 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1224 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1225 (int)depth * 2 + 2,"",
1226 (UV)TRIE_NODENUM( state ) );
1228 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1229 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1231 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1233 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1235 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1236 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1238 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1239 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1247 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1248 startbranch: the first branch in the whole branch sequence
1249 first : start branch of sequence of branch-exact nodes.
1250 May be the same as startbranch
1251 last : Thing following the last branch.
1252 May be the same as tail.
1253 tail : item following the branch sequence
1254 count : words in the sequence
1255 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1256 depth : indent depth
1258 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1260 A trie is an N'ary tree where the branches are determined by digital
1261 decomposition of the key. IE, at the root node you look up the 1st character and
1262 follow that branch repeat until you find the end of the branches. Nodes can be
1263 marked as "accepting" meaning they represent a complete word. Eg:
1267 would convert into the following structure. Numbers represent states, letters
1268 following numbers represent valid transitions on the letter from that state, if
1269 the number is in square brackets it represents an accepting state, otherwise it
1270 will be in parenthesis.
1272 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1276 (1) +-i->(6)-+-s->[7]
1278 +-s->(3)-+-h->(4)-+-e->[5]
1280 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1282 This shows that when matching against the string 'hers' we will begin at state 1
1283 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1284 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1285 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1286 single traverse. We store a mapping from accepting to state to which word was
1287 matched, and then when we have multiple possibilities we try to complete the
1288 rest of the regex in the order in which they occured in the alternation.
1290 The only prior NFA like behaviour that would be changed by the TRIE support is
1291 the silent ignoring of duplicate alternations which are of the form:
1293 / (DUPE|DUPE) X? (?{ ... }) Y /x
1295 Thus EVAL blocks following a trie may be called a different number of times with
1296 and without the optimisation. With the optimisations dupes will be silently
1297 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1298 the following demonstrates:
1300 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1302 which prints out 'word' three times, but
1304 'words'=~/(word|word|word)(?{ print $1 })S/
1306 which doesnt print it out at all. This is due to other optimisations kicking in.
1308 Example of what happens on a structural level:
1310 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1312 1: CURLYM[1] {1,32767}(18)
1323 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1324 and should turn into:
1326 1: CURLYM[1] {1,32767}(18)
1328 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1336 Cases where tail != last would be like /(?foo|bar)baz/:
1346 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1347 and would end up looking like:
1350 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1357 d = uvuni_to_utf8_flags(d, uv, 0);
1359 is the recommended Unicode-aware way of saying
1364 #define TRIE_STORE_REVCHAR \
1367 SV *zlopp = newSV(2); \
1368 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1369 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
1370 SvCUR_set(zlopp, kapow - flrbbbbb); \
1373 av_push(revcharmap, zlopp); \
1375 char ooooff = (char)uvc; \
1376 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1380 #define TRIE_READ_CHAR STMT_START { \
1384 if ( foldlen > 0 ) { \
1385 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1390 len = UTF8SKIP(uc);\
1391 uvc = to_utf8_fold( uc, foldbuf, &foldlen); \
1392 foldlen -= UNISKIP( uvc ); \
1393 scan = foldbuf + UNISKIP( uvc ); \
1396 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1406 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1407 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1408 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1409 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1411 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1412 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1413 TRIE_LIST_CUR( state )++; \
1416 #define TRIE_LIST_NEW(state) STMT_START { \
1417 Newxz( trie->states[ state ].trans.list, \
1418 4, reg_trie_trans_le ); \
1419 TRIE_LIST_CUR( state ) = 1; \
1420 TRIE_LIST_LEN( state ) = 4; \
1423 #define TRIE_HANDLE_WORD(state) STMT_START { \
1424 U16 dupe= trie->states[ state ].wordnum; \
1425 regnode * const noper_next = regnext( noper ); \
1428 /* store the word for dumping */ \
1430 if (OP(noper) != NOTHING) \
1431 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1433 tmp = newSVpvn_utf8( "", 0, UTF ); \
1434 av_push( trie_words, tmp ); \
1438 trie->wordinfo[curword].prev = 0; \
1439 trie->wordinfo[curword].len = wordlen; \
1440 trie->wordinfo[curword].accept = state; \
1442 if ( noper_next < tail ) { \
1444 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1445 trie->jump[curword] = (U16)(noper_next - convert); \
1447 jumper = noper_next; \
1449 nextbranch= regnext(cur); \
1453 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1454 /* chain, so that when the bits of chain are later */\
1455 /* linked together, the dups appear in the chain */\
1456 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1457 trie->wordinfo[dupe].prev = curword; \
1459 /* we haven't inserted this word yet. */ \
1460 trie->states[ state ].wordnum = curword; \
1465 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1466 ( ( base + charid >= ucharcount \
1467 && base + charid < ubound \
1468 && state == trie->trans[ base - ucharcount + charid ].check \
1469 && trie->trans[ base - ucharcount + charid ].next ) \
1470 ? trie->trans[ base - ucharcount + charid ].next \
1471 : ( state==1 ? special : 0 ) \
1475 #define MADE_JUMP_TRIE 2
1476 #define MADE_EXACT_TRIE 4
1479 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1482 /* first pass, loop through and scan words */
1483 reg_trie_data *trie;
1484 HV *widecharmap = NULL;
1485 AV *revcharmap = newAV();
1487 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1492 regnode *jumper = NULL;
1493 regnode *nextbranch = NULL;
1494 regnode *convert = NULL;
1495 U32 *prev_states; /* temp array mapping each state to previous one */
1496 /* we just use folder as a flag in utf8 */
1497 const U8 * folder = NULL;
1500 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1501 AV *trie_words = NULL;
1502 /* along with revcharmap, this only used during construction but both are
1503 * useful during debugging so we store them in the struct when debugging.
1506 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1507 STRLEN trie_charcount=0;
1509 SV *re_trie_maxbuff;
1510 GET_RE_DEBUG_FLAGS_DECL;
1512 PERL_ARGS_ASSERT_MAKE_TRIE;
1514 PERL_UNUSED_ARG(depth);
1519 case EXACTFU: folder = PL_fold_latin1; break;
1520 case EXACTF: folder = PL_fold; break;
1521 case EXACTFL: folder = PL_fold_locale; break;
1524 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1526 trie->startstate = 1;
1527 trie->wordcount = word_count;
1528 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1529 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1530 if (!(UTF && folder))
1531 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1532 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1533 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1536 trie_words = newAV();
1539 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1540 if (!SvIOK(re_trie_maxbuff)) {
1541 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1544 PerlIO_printf( Perl_debug_log,
1545 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1546 (int)depth * 2 + 2, "",
1547 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1548 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1552 /* Find the node we are going to overwrite */
1553 if ( first == startbranch && OP( last ) != BRANCH ) {
1554 /* whole branch chain */
1557 /* branch sub-chain */
1558 convert = NEXTOPER( first );
1561 /* -- First loop and Setup --
1563 We first traverse the branches and scan each word to determine if it
1564 contains widechars, and how many unique chars there are, this is
1565 important as we have to build a table with at least as many columns as we
1568 We use an array of integers to represent the character codes 0..255
1569 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1570 native representation of the character value as the key and IV's for the
1573 *TODO* If we keep track of how many times each character is used we can
1574 remap the columns so that the table compression later on is more
1575 efficient in terms of memory by ensuring the most common value is in the
1576 middle and the least common are on the outside. IMO this would be better
1577 than a most to least common mapping as theres a decent chance the most
1578 common letter will share a node with the least common, meaning the node
1579 will not be compressible. With a middle is most common approach the worst
1580 case is when we have the least common nodes twice.
1584 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1585 regnode * const noper = NEXTOPER( cur );
1586 const U8 *uc = (U8*)STRING( noper );
1587 const U8 * const e = uc + STR_LEN( noper );
1589 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1590 const U8 *scan = (U8*)NULL;
1591 U32 wordlen = 0; /* required init */
1593 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1595 if (OP(noper) == NOTHING) {
1599 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1600 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1601 regardless of encoding */
1603 for ( ; uc < e ; uc += len ) {
1604 TRIE_CHARCOUNT(trie)++;
1608 if ( !trie->charmap[ uvc ] ) {
1609 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1611 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
1615 /* store the codepoint in the bitmap, and its folded
1617 TRIE_BITMAP_SET(trie,uvc);
1619 /* store the folded codepoint */
1620 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1623 /* store first byte of utf8 representation of
1624 variant codepoints */
1625 if (! UNI_IS_INVARIANT(uvc)) {
1626 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1629 set_bit = 0; /* We've done our bit :-) */
1634 widecharmap = newHV();
1636 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1639 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1641 if ( !SvTRUE( *svpp ) ) {
1642 sv_setiv( *svpp, ++trie->uniquecharcount );
1647 if( cur == first ) {
1650 } else if (chars < trie->minlen) {
1652 } else if (chars > trie->maxlen) {
1656 } /* end first pass */
1657 DEBUG_TRIE_COMPILE_r(
1658 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1659 (int)depth * 2 + 2,"",
1660 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1661 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1662 (int)trie->minlen, (int)trie->maxlen )
1666 We now know what we are dealing with in terms of unique chars and
1667 string sizes so we can calculate how much memory a naive
1668 representation using a flat table will take. If it's over a reasonable
1669 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1670 conservative but potentially much slower representation using an array
1673 At the end we convert both representations into the same compressed
1674 form that will be used in regexec.c for matching with. The latter
1675 is a form that cannot be used to construct with but has memory
1676 properties similar to the list form and access properties similar
1677 to the table form making it both suitable for fast searches and
1678 small enough that its feasable to store for the duration of a program.
1680 See the comment in the code where the compressed table is produced
1681 inplace from the flat tabe representation for an explanation of how
1682 the compression works.
1687 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1690 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1692 Second Pass -- Array Of Lists Representation
1694 Each state will be represented by a list of charid:state records
1695 (reg_trie_trans_le) the first such element holds the CUR and LEN
1696 points of the allocated array. (See defines above).
1698 We build the initial structure using the lists, and then convert
1699 it into the compressed table form which allows faster lookups
1700 (but cant be modified once converted).
1703 STRLEN transcount = 1;
1705 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1706 "%*sCompiling trie using list compiler\n",
1707 (int)depth * 2 + 2, ""));
1709 trie->states = (reg_trie_state *)
1710 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1711 sizeof(reg_trie_state) );
1715 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1717 regnode * const noper = NEXTOPER( cur );
1718 U8 *uc = (U8*)STRING( noper );
1719 const U8 * const e = uc + STR_LEN( noper );
1720 U32 state = 1; /* required init */
1721 U16 charid = 0; /* sanity init */
1722 U8 *scan = (U8*)NULL; /* sanity init */
1723 STRLEN foldlen = 0; /* required init */
1724 U32 wordlen = 0; /* required init */
1725 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1727 if (OP(noper) != NOTHING) {
1728 for ( ; uc < e ; uc += len ) {
1733 charid = trie->charmap[ uvc ];
1735 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1739 charid=(U16)SvIV( *svpp );
1742 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1749 if ( !trie->states[ state ].trans.list ) {
1750 TRIE_LIST_NEW( state );
1752 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1753 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1754 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1759 newstate = next_alloc++;
1760 prev_states[newstate] = state;
1761 TRIE_LIST_PUSH( state, charid, newstate );
1766 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1770 TRIE_HANDLE_WORD(state);
1772 } /* end second pass */
1774 /* next alloc is the NEXT state to be allocated */
1775 trie->statecount = next_alloc;
1776 trie->states = (reg_trie_state *)
1777 PerlMemShared_realloc( trie->states,
1779 * sizeof(reg_trie_state) );
1781 /* and now dump it out before we compress it */
1782 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1783 revcharmap, next_alloc,
1787 trie->trans = (reg_trie_trans *)
1788 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1795 for( state=1 ; state < next_alloc ; state ++ ) {
1799 DEBUG_TRIE_COMPILE_MORE_r(
1800 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1804 if (trie->states[state].trans.list) {
1805 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1809 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1810 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1811 if ( forid < minid ) {
1813 } else if ( forid > maxid ) {
1817 if ( transcount < tp + maxid - minid + 1) {
1819 trie->trans = (reg_trie_trans *)
1820 PerlMemShared_realloc( trie->trans,
1822 * sizeof(reg_trie_trans) );
1823 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1825 base = trie->uniquecharcount + tp - minid;
1826 if ( maxid == minid ) {
1828 for ( ; zp < tp ; zp++ ) {
1829 if ( ! trie->trans[ zp ].next ) {
1830 base = trie->uniquecharcount + zp - minid;
1831 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1832 trie->trans[ zp ].check = state;
1838 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1839 trie->trans[ tp ].check = state;
1844 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1845 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1846 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1847 trie->trans[ tid ].check = state;
1849 tp += ( maxid - minid + 1 );
1851 Safefree(trie->states[ state ].trans.list);
1854 DEBUG_TRIE_COMPILE_MORE_r(
1855 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1858 trie->states[ state ].trans.base=base;
1860 trie->lasttrans = tp + 1;
1864 Second Pass -- Flat Table Representation.
1866 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1867 We know that we will need Charcount+1 trans at most to store the data
1868 (one row per char at worst case) So we preallocate both structures
1869 assuming worst case.
1871 We then construct the trie using only the .next slots of the entry
1874 We use the .check field of the first entry of the node temporarily to
1875 make compression both faster and easier by keeping track of how many non
1876 zero fields are in the node.
1878 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1881 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1882 number representing the first entry of the node, and state as a
1883 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1884 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1885 are 2 entrys per node. eg:
1893 The table is internally in the right hand, idx form. However as we also
1894 have to deal with the states array which is indexed by nodenum we have to
1895 use TRIE_NODENUM() to convert.
1898 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1899 "%*sCompiling trie using table compiler\n",
1900 (int)depth * 2 + 2, ""));
1902 trie->trans = (reg_trie_trans *)
1903 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1904 * trie->uniquecharcount + 1,
1905 sizeof(reg_trie_trans) );
1906 trie->states = (reg_trie_state *)
1907 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1908 sizeof(reg_trie_state) );
1909 next_alloc = trie->uniquecharcount + 1;
1912 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1914 regnode * const noper = NEXTOPER( cur );
1915 const U8 *uc = (U8*)STRING( noper );
1916 const U8 * const e = uc + STR_LEN( noper );
1918 U32 state = 1; /* required init */
1920 U16 charid = 0; /* sanity init */
1921 U32 accept_state = 0; /* sanity init */
1922 U8 *scan = (U8*)NULL; /* sanity init */
1924 STRLEN foldlen = 0; /* required init */
1925 U32 wordlen = 0; /* required init */
1926 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1928 if ( OP(noper) != NOTHING ) {
1929 for ( ; uc < e ; uc += len ) {
1934 charid = trie->charmap[ uvc ];
1936 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1937 charid = svpp ? (U16)SvIV(*svpp) : 0;
1941 if ( !trie->trans[ state + charid ].next ) {
1942 trie->trans[ state + charid ].next = next_alloc;
1943 trie->trans[ state ].check++;
1944 prev_states[TRIE_NODENUM(next_alloc)]
1945 = TRIE_NODENUM(state);
1946 next_alloc += trie->uniquecharcount;
1948 state = trie->trans[ state + charid ].next;
1950 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1952 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1955 accept_state = TRIE_NODENUM( state );
1956 TRIE_HANDLE_WORD(accept_state);
1958 } /* end second pass */
1960 /* and now dump it out before we compress it */
1961 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1963 next_alloc, depth+1));
1967 * Inplace compress the table.*
1969 For sparse data sets the table constructed by the trie algorithm will
1970 be mostly 0/FAIL transitions or to put it another way mostly empty.
1971 (Note that leaf nodes will not contain any transitions.)
1973 This algorithm compresses the tables by eliminating most such
1974 transitions, at the cost of a modest bit of extra work during lookup:
1976 - Each states[] entry contains a .base field which indicates the
1977 index in the state[] array wheres its transition data is stored.
1979 - If .base is 0 there are no valid transitions from that node.
1981 - If .base is nonzero then charid is added to it to find an entry in
1984 -If trans[states[state].base+charid].check!=state then the
1985 transition is taken to be a 0/Fail transition. Thus if there are fail
1986 transitions at the front of the node then the .base offset will point
1987 somewhere inside the previous nodes data (or maybe even into a node
1988 even earlier), but the .check field determines if the transition is
1992 The following process inplace converts the table to the compressed
1993 table: We first do not compress the root node 1,and mark all its
1994 .check pointers as 1 and set its .base pointer as 1 as well. This
1995 allows us to do a DFA construction from the compressed table later,
1996 and ensures that any .base pointers we calculate later are greater
1999 - We set 'pos' to indicate the first entry of the second node.
2001 - We then iterate over the columns of the node, finding the first and
2002 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2003 and set the .check pointers accordingly, and advance pos
2004 appropriately and repreat for the next node. Note that when we copy
2005 the next pointers we have to convert them from the original
2006 NODEIDX form to NODENUM form as the former is not valid post
2009 - If a node has no transitions used we mark its base as 0 and do not
2010 advance the pos pointer.
2012 - If a node only has one transition we use a second pointer into the
2013 structure to fill in allocated fail transitions from other states.
2014 This pointer is independent of the main pointer and scans forward
2015 looking for null transitions that are allocated to a state. When it
2016 finds one it writes the single transition into the "hole". If the
2017 pointer doesnt find one the single transition is appended as normal.
2019 - Once compressed we can Renew/realloc the structures to release the
2022 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2023 specifically Fig 3.47 and the associated pseudocode.
2027 const U32 laststate = TRIE_NODENUM( next_alloc );
2030 trie->statecount = laststate;
2032 for ( state = 1 ; state < laststate ; state++ ) {
2034 const U32 stateidx = TRIE_NODEIDX( state );
2035 const U32 o_used = trie->trans[ stateidx ].check;
2036 U32 used = trie->trans[ stateidx ].check;
2037 trie->trans[ stateidx ].check = 0;
2039 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2040 if ( flag || trie->trans[ stateidx + charid ].next ) {
2041 if ( trie->trans[ stateidx + charid ].next ) {
2043 for ( ; zp < pos ; zp++ ) {
2044 if ( ! trie->trans[ zp ].next ) {
2048 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2049 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2050 trie->trans[ zp ].check = state;
2051 if ( ++zp > pos ) pos = zp;
2058 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2060 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2061 trie->trans[ pos ].check = state;
2066 trie->lasttrans = pos + 1;
2067 trie->states = (reg_trie_state *)
2068 PerlMemShared_realloc( trie->states, laststate
2069 * sizeof(reg_trie_state) );
2070 DEBUG_TRIE_COMPILE_MORE_r(
2071 PerlIO_printf( Perl_debug_log,
2072 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2073 (int)depth * 2 + 2,"",
2074 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2077 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2080 } /* end table compress */
2082 DEBUG_TRIE_COMPILE_MORE_r(
2083 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2084 (int)depth * 2 + 2, "",
2085 (UV)trie->statecount,
2086 (UV)trie->lasttrans)
2088 /* resize the trans array to remove unused space */
2089 trie->trans = (reg_trie_trans *)
2090 PerlMemShared_realloc( trie->trans, trie->lasttrans
2091 * sizeof(reg_trie_trans) );
2093 { /* Modify the program and insert the new TRIE node */
2094 U8 nodetype =(U8)(flags & 0xFF);
2098 regnode *optimize = NULL;
2099 #ifdef RE_TRACK_PATTERN_OFFSETS
2102 U32 mjd_nodelen = 0;
2103 #endif /* RE_TRACK_PATTERN_OFFSETS */
2104 #endif /* DEBUGGING */
2106 This means we convert either the first branch or the first Exact,
2107 depending on whether the thing following (in 'last') is a branch
2108 or not and whther first is the startbranch (ie is it a sub part of
2109 the alternation or is it the whole thing.)
2110 Assuming its a sub part we convert the EXACT otherwise we convert
2111 the whole branch sequence, including the first.
2113 /* Find the node we are going to overwrite */
2114 if ( first != startbranch || OP( last ) == BRANCH ) {
2115 /* branch sub-chain */
2116 NEXT_OFF( first ) = (U16)(last - first);
2117 #ifdef RE_TRACK_PATTERN_OFFSETS
2119 mjd_offset= Node_Offset((convert));
2120 mjd_nodelen= Node_Length((convert));
2123 /* whole branch chain */
2125 #ifdef RE_TRACK_PATTERN_OFFSETS
2128 const regnode *nop = NEXTOPER( convert );
2129 mjd_offset= Node_Offset((nop));
2130 mjd_nodelen= Node_Length((nop));
2134 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2135 (int)depth * 2 + 2, "",
2136 (UV)mjd_offset, (UV)mjd_nodelen)
2139 /* But first we check to see if there is a common prefix we can
2140 split out as an EXACT and put in front of the TRIE node. */
2141 trie->startstate= 1;
2142 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2144 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2148 const U32 base = trie->states[ state ].trans.base;
2150 if ( trie->states[state].wordnum )
2153 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2154 if ( ( base + ofs >= trie->uniquecharcount ) &&
2155 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2156 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2158 if ( ++count > 1 ) {
2159 SV **tmp = av_fetch( revcharmap, ofs, 0);
2160 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2161 if ( state == 1 ) break;
2163 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2165 PerlIO_printf(Perl_debug_log,
2166 "%*sNew Start State=%"UVuf" Class: [",
2167 (int)depth * 2 + 2, "",
2170 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2171 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2173 TRIE_BITMAP_SET(trie,*ch);
2175 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2177 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2181 TRIE_BITMAP_SET(trie,*ch);
2183 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2184 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2190 SV **tmp = av_fetch( revcharmap, idx, 0);
2192 char *ch = SvPV( *tmp, len );
2194 SV *sv=sv_newmortal();
2195 PerlIO_printf( Perl_debug_log,
2196 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2197 (int)depth * 2 + 2, "",
2199 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2200 PL_colors[0], PL_colors[1],
2201 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2202 PERL_PV_ESCAPE_FIRSTCHAR
2207 OP( convert ) = nodetype;
2208 str=STRING(convert);
2211 STR_LEN(convert) += len;
2217 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2222 trie->prefixlen = (state-1);
2224 regnode *n = convert+NODE_SZ_STR(convert);
2225 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2226 trie->startstate = state;
2227 trie->minlen -= (state - 1);
2228 trie->maxlen -= (state - 1);
2230 /* At least the UNICOS C compiler choked on this
2231 * being argument to DEBUG_r(), so let's just have
2234 #ifdef PERL_EXT_RE_BUILD
2240 regnode *fix = convert;
2241 U32 word = trie->wordcount;
2243 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2244 while( ++fix < n ) {
2245 Set_Node_Offset_Length(fix, 0, 0);
2248 SV ** const tmp = av_fetch( trie_words, word, 0 );
2250 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2251 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2253 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2261 NEXT_OFF(convert) = (U16)(tail - convert);
2262 DEBUG_r(optimize= n);
2268 if ( trie->maxlen ) {
2269 NEXT_OFF( convert ) = (U16)(tail - convert);
2270 ARG_SET( convert, data_slot );
2271 /* Store the offset to the first unabsorbed branch in
2272 jump[0], which is otherwise unused by the jump logic.
2273 We use this when dumping a trie and during optimisation. */
2275 trie->jump[0] = (U16)(nextbranch - convert);
2277 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2278 * and there is a bitmap
2279 * and the first "jump target" node we found leaves enough room
2280 * then convert the TRIE node into a TRIEC node, with the bitmap
2281 * embedded inline in the opcode - this is hypothetically faster.
2283 if ( !trie->states[trie->startstate].wordnum
2285 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2287 OP( convert ) = TRIEC;
2288 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2289 PerlMemShared_free(trie->bitmap);
2292 OP( convert ) = TRIE;
2294 /* store the type in the flags */
2295 convert->flags = nodetype;
2299 + regarglen[ OP( convert ) ];
2301 /* XXX We really should free up the resource in trie now,
2302 as we won't use them - (which resources?) dmq */
2304 /* needed for dumping*/
2305 DEBUG_r(if (optimize) {
2306 regnode *opt = convert;
2308 while ( ++opt < optimize) {
2309 Set_Node_Offset_Length(opt,0,0);
2312 Try to clean up some of the debris left after the
2315 while( optimize < jumper ) {
2316 mjd_nodelen += Node_Length((optimize));
2317 OP( optimize ) = OPTIMIZED;
2318 Set_Node_Offset_Length(optimize,0,0);
2321 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2323 } /* end node insert */
2324 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, convert);
2326 /* Finish populating the prev field of the wordinfo array. Walk back
2327 * from each accept state until we find another accept state, and if
2328 * so, point the first word's .prev field at the second word. If the
2329 * second already has a .prev field set, stop now. This will be the
2330 * case either if we've already processed that word's accept state,
2331 * or that state had multiple words, and the overspill words were
2332 * already linked up earlier.
2339 for (word=1; word <= trie->wordcount; word++) {
2341 if (trie->wordinfo[word].prev)
2343 state = trie->wordinfo[word].accept;
2345 state = prev_states[state];
2348 prev = trie->states[state].wordnum;
2352 trie->wordinfo[word].prev = prev;
2354 Safefree(prev_states);
2358 /* and now dump out the compressed format */
2359 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2361 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2363 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2364 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2366 SvREFCNT_dec(revcharmap);
2370 : trie->startstate>1
2376 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2378 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2380 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2381 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2384 We find the fail state for each state in the trie, this state is the longest proper
2385 suffix of the current state's 'word' that is also a proper prefix of another word in our
2386 trie. State 1 represents the word '' and is thus the default fail state. This allows
2387 the DFA not to have to restart after its tried and failed a word at a given point, it
2388 simply continues as though it had been matching the other word in the first place.
2390 'abcdgu'=~/abcdefg|cdgu/
2391 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2392 fail, which would bring us to the state representing 'd' in the second word where we would
2393 try 'g' and succeed, proceeding to match 'cdgu'.
2395 /* add a fail transition */
2396 const U32 trie_offset = ARG(source);
2397 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2399 const U32 ucharcount = trie->uniquecharcount;
2400 const U32 numstates = trie->statecount;
2401 const U32 ubound = trie->lasttrans + ucharcount;
2405 U32 base = trie->states[ 1 ].trans.base;
2408 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2409 GET_RE_DEBUG_FLAGS_DECL;
2411 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2413 PERL_UNUSED_ARG(depth);
2417 ARG_SET( stclass, data_slot );
2418 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2419 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2420 aho->trie=trie_offset;
2421 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2422 Copy( trie->states, aho->states, numstates, reg_trie_state );
2423 Newxz( q, numstates, U32);
2424 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2427 /* initialize fail[0..1] to be 1 so that we always have
2428 a valid final fail state */
2429 fail[ 0 ] = fail[ 1 ] = 1;
2431 for ( charid = 0; charid < ucharcount ; charid++ ) {
2432 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2434 q[ q_write ] = newstate;
2435 /* set to point at the root */
2436 fail[ q[ q_write++ ] ]=1;
2439 while ( q_read < q_write) {
2440 const U32 cur = q[ q_read++ % numstates ];
2441 base = trie->states[ cur ].trans.base;
2443 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2444 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2446 U32 fail_state = cur;
2449 fail_state = fail[ fail_state ];
2450 fail_base = aho->states[ fail_state ].trans.base;
2451 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2453 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2454 fail[ ch_state ] = fail_state;
2455 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2457 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2459 q[ q_write++ % numstates] = ch_state;
2463 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2464 when we fail in state 1, this allows us to use the
2465 charclass scan to find a valid start char. This is based on the principle
2466 that theres a good chance the string being searched contains lots of stuff
2467 that cant be a start char.
2469 fail[ 0 ] = fail[ 1 ] = 0;
2470 DEBUG_TRIE_COMPILE_r({
2471 PerlIO_printf(Perl_debug_log,
2472 "%*sStclass Failtable (%"UVuf" states): 0",
2473 (int)(depth * 2), "", (UV)numstates
2475 for( q_read=1; q_read<numstates; q_read++ ) {
2476 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2478 PerlIO_printf(Perl_debug_log, "\n");
2481 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2486 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2487 * These need to be revisited when a newer toolchain becomes available.
2489 #if defined(__sparc64__) && defined(__GNUC__)
2490 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2491 # undef SPARC64_GCC_WORKAROUND
2492 # define SPARC64_GCC_WORKAROUND 1
2496 #define DEBUG_PEEP(str,scan,depth) \
2497 DEBUG_OPTIMISE_r({if (scan){ \
2498 SV * const mysv=sv_newmortal(); \
2499 regnode *Next = regnext(scan); \
2500 regprop(RExC_rx, mysv, scan); \
2501 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2502 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2503 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2510 #define JOIN_EXACT(scan,min,flags) \
2511 if (PL_regkind[OP(scan)] == EXACT) \
2512 join_exact(pRExC_state,(scan),(min),(flags),NULL,depth+1)
2515 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, I32 *min, U32 flags,regnode *val, U32 depth) {
2516 /* Merge several consecutive EXACTish nodes into one. */
2517 regnode *n = regnext(scan);
2519 regnode *next = scan + NODE_SZ_STR(scan);
2523 regnode *stop = scan;
2524 GET_RE_DEBUG_FLAGS_DECL;
2526 PERL_UNUSED_ARG(depth);
2529 PERL_ARGS_ASSERT_JOIN_EXACT;
2530 #ifndef EXPERIMENTAL_INPLACESCAN
2531 PERL_UNUSED_ARG(flags);
2532 PERL_UNUSED_ARG(val);
2534 DEBUG_PEEP("join",scan,depth);
2536 /* Skip NOTHING, merge EXACT*. */
2538 ( PL_regkind[OP(n)] == NOTHING ||
2539 (stringok && (OP(n) == OP(scan))))
2541 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX) {
2543 if (OP(n) == TAIL || n > next)
2545 if (PL_regkind[OP(n)] == NOTHING) {
2546 DEBUG_PEEP("skip:",n,depth);
2547 NEXT_OFF(scan) += NEXT_OFF(n);
2548 next = n + NODE_STEP_REGNODE;
2555 else if (stringok) {
2556 const unsigned int oldl = STR_LEN(scan);
2557 regnode * const nnext = regnext(n);
2559 DEBUG_PEEP("merg",n,depth);
2562 if (oldl + STR_LEN(n) > U8_MAX)
2564 NEXT_OFF(scan) += NEXT_OFF(n);
2565 STR_LEN(scan) += STR_LEN(n);
2566 next = n + NODE_SZ_STR(n);
2567 /* Now we can overwrite *n : */
2568 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2576 #ifdef EXPERIMENTAL_INPLACESCAN
2577 if (flags && !NEXT_OFF(n)) {
2578 DEBUG_PEEP("atch", val, depth);
2579 if (reg_off_by_arg[OP(n)]) {
2580 ARG_SET(n, val - n);
2583 NEXT_OFF(n) = val - n;
2589 #define GREEK_SMALL_LETTER_IOTA_WITH_DIALYTIKA_AND_TONOS 0x0390
2590 #define IOTA_D_T GREEK_SMALL_LETTER_IOTA_WITH_DIALYTIKA_AND_TONOS
2591 #define GREEK_SMALL_LETTER_UPSILON_WITH_DIALYTIKA_AND_TONOS 0x03B0
2592 #define UPSILON_D_T GREEK_SMALL_LETTER_UPSILON_WITH_DIALYTIKA_AND_TONOS
2595 && ( OP(scan) == EXACTF || OP(scan) == EXACTFU || OP(scan) == EXACTFA)
2596 && ( STR_LEN(scan) >= 6 ) )
2599 Two problematic code points in Unicode casefolding of EXACT nodes:
2601 U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2602 U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2608 U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2609 U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2611 This means that in case-insensitive matching (or "loose matching",
2612 as Unicode calls it), an EXACTF of length six (the UTF-8 encoded byte
2613 length of the above casefolded versions) can match a target string
2614 of length two (the byte length of UTF-8 encoded U+0390 or U+03B0).
2615 This would rather mess up the minimum length computation.
2617 What we'll do is to look for the tail four bytes, and then peek
2618 at the preceding two bytes to see whether we need to decrease
2619 the minimum length by four (six minus two).
2621 Thanks to the design of UTF-8, there cannot be false matches:
2622 A sequence of valid UTF-8 bytes cannot be a subsequence of
2623 another valid sequence of UTF-8 bytes.
2626 char * const s0 = STRING(scan), *s, *t;
2627 char * const s1 = s0 + STR_LEN(scan) - 1;
2628 char * const s2 = s1 - 4;
2629 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2630 const char t0[] = "\xaf\x49\xaf\x42";
2632 const char t0[] = "\xcc\x88\xcc\x81";
2634 const char * const t1 = t0 + 3;
2637 s < s2 && (t = ninstr(s, s1, t0, t1));
2640 if (((U8)t[-1] == 0x68 && (U8)t[-2] == 0xB4) ||
2641 ((U8)t[-1] == 0x46 && (U8)t[-2] == 0xB5))
2643 if (((U8)t[-1] == 0xB9 && (U8)t[-2] == 0xCE) ||
2644 ((U8)t[-1] == 0x85 && (U8)t[-2] == 0xCF))
2651 /* Allow dumping but overwriting the collection of skipped
2652 * ops and/or strings with fake optimized ops */
2653 n = scan + NODE_SZ_STR(scan);
2661 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2665 /* REx optimizer. Converts nodes into quicker variants "in place".
2666 Finds fixed substrings. */
2668 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2669 to the position after last scanned or to NULL. */
2671 #define INIT_AND_WITHP \
2672 assert(!and_withp); \
2673 Newx(and_withp,1,struct regnode_charclass_class); \
2674 SAVEFREEPV(and_withp)
2676 /* this is a chain of data about sub patterns we are processing that
2677 need to be handled separately/specially in study_chunk. Its so
2678 we can simulate recursion without losing state. */
2680 typedef struct scan_frame {
2681 regnode *last; /* last node to process in this frame */
2682 regnode *next; /* next node to process when last is reached */
2683 struct scan_frame *prev; /*previous frame*/
2684 I32 stop; /* what stopparen do we use */
2688 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2690 #define CASE_SYNST_FNC(nAmE) \
2692 if (flags & SCF_DO_STCLASS_AND) { \
2693 for (value = 0; value < 256; value++) \
2694 if (!is_ ## nAmE ## _cp(value)) \
2695 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2698 for (value = 0; value < 256; value++) \
2699 if (is_ ## nAmE ## _cp(value)) \
2700 ANYOF_BITMAP_SET(data->start_class, value); \
2704 if (flags & SCF_DO_STCLASS_AND) { \
2705 for (value = 0; value < 256; value++) \
2706 if (is_ ## nAmE ## _cp(value)) \
2707 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2710 for (value = 0; value < 256; value++) \
2711 if (!is_ ## nAmE ## _cp(value)) \
2712 ANYOF_BITMAP_SET(data->start_class, value); \
2719 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2720 I32 *minlenp, I32 *deltap,
2725 struct regnode_charclass_class *and_withp,
2726 U32 flags, U32 depth)
2727 /* scanp: Start here (read-write). */
2728 /* deltap: Write maxlen-minlen here. */
2729 /* last: Stop before this one. */
2730 /* data: string data about the pattern */
2731 /* stopparen: treat close N as END */
2732 /* recursed: which subroutines have we recursed into */
2733 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2736 I32 min = 0, pars = 0, code;
2737 regnode *scan = *scanp, *next;
2739 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2740 int is_inf_internal = 0; /* The studied chunk is infinite */
2741 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2742 scan_data_t data_fake;
2743 SV *re_trie_maxbuff = NULL;
2744 regnode *first_non_open = scan;
2745 I32 stopmin = I32_MAX;
2746 scan_frame *frame = NULL;
2747 GET_RE_DEBUG_FLAGS_DECL;
2749 PERL_ARGS_ASSERT_STUDY_CHUNK;
2752 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
2756 while (first_non_open && OP(first_non_open) == OPEN)
2757 first_non_open=regnext(first_non_open);
2762 while ( scan && OP(scan) != END && scan < last ){
2763 /* Peephole optimizer: */
2764 DEBUG_STUDYDATA("Peep:", data,depth);
2765 DEBUG_PEEP("Peep",scan,depth);
2766 JOIN_EXACT(scan,&min,0);
2768 /* Follow the next-chain of the current node and optimize
2769 away all the NOTHINGs from it. */
2770 if (OP(scan) != CURLYX) {
2771 const int max = (reg_off_by_arg[OP(scan)]
2773 /* I32 may be smaller than U16 on CRAYs! */
2774 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
2775 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
2779 /* Skip NOTHING and LONGJMP. */
2780 while ((n = regnext(n))
2781 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
2782 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
2783 && off + noff < max)
2785 if (reg_off_by_arg[OP(scan)])
2788 NEXT_OFF(scan) = off;
2793 /* The principal pseudo-switch. Cannot be a switch, since we
2794 look into several different things. */
2795 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
2796 || OP(scan) == IFTHEN) {
2797 next = regnext(scan);
2799 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
2801 if (OP(next) == code || code == IFTHEN) {
2802 /* NOTE - There is similar code to this block below for handling
2803 TRIE nodes on a re-study. If you change stuff here check there
2805 I32 max1 = 0, min1 = I32_MAX, num = 0;
2806 struct regnode_charclass_class accum;
2807 regnode * const startbranch=scan;
2809 if (flags & SCF_DO_SUBSTR)
2810 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
2811 if (flags & SCF_DO_STCLASS)
2812 cl_init_zero(pRExC_state, &accum);
2814 while (OP(scan) == code) {
2815 I32 deltanext, minnext, f = 0, fake;
2816 struct regnode_charclass_class this_class;
2819 data_fake.flags = 0;
2821 data_fake.whilem_c = data->whilem_c;
2822 data_fake.last_closep = data->last_closep;
2825 data_fake.last_closep = &fake;
2827 data_fake.pos_delta = delta;
2828 next = regnext(scan);
2829 scan = NEXTOPER(scan);
2831 scan = NEXTOPER(scan);
2832 if (flags & SCF_DO_STCLASS) {
2833 cl_init(pRExC_state, &this_class);
2834 data_fake.start_class = &this_class;
2835 f = SCF_DO_STCLASS_AND;
2837 if (flags & SCF_WHILEM_VISITED_POS)
2838 f |= SCF_WHILEM_VISITED_POS;
2840 /* we suppose the run is continuous, last=next...*/
2841 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
2843 stopparen, recursed, NULL, f,depth+1);
2846 if (max1 < minnext + deltanext)
2847 max1 = minnext + deltanext;
2848 if (deltanext == I32_MAX)
2849 is_inf = is_inf_internal = 1;
2851 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
2853 if (data_fake.flags & SCF_SEEN_ACCEPT) {
2854 if ( stopmin > minnext)
2855 stopmin = min + min1;
2856 flags &= ~SCF_DO_SUBSTR;
2858 data->flags |= SCF_SEEN_ACCEPT;
2861 if (data_fake.flags & SF_HAS_EVAL)
2862 data->flags |= SF_HAS_EVAL;
2863 data->whilem_c = data_fake.whilem_c;
2865 if (flags & SCF_DO_STCLASS)
2866 cl_or(pRExC_state, &accum, &this_class);
2868 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
2870 if (flags & SCF_DO_SUBSTR) {
2871 data->pos_min += min1;
2872 data->pos_delta += max1 - min1;
2873 if (max1 != min1 || is_inf)
2874 data->longest = &(data->longest_float);
2877 delta += max1 - min1;
2878 if (flags & SCF_DO_STCLASS_OR) {
2879 cl_or(pRExC_state, data->start_class, &accum);
2881 cl_and(data->start_class, and_withp);
2882 flags &= ~SCF_DO_STCLASS;
2885 else if (flags & SCF_DO_STCLASS_AND) {
2887 cl_and(data->start_class, &accum);
2888 flags &= ~SCF_DO_STCLASS;
2891 /* Switch to OR mode: cache the old value of
2892 * data->start_class */
2894 StructCopy(data->start_class, and_withp,
2895 struct regnode_charclass_class);
2896 flags &= ~SCF_DO_STCLASS_AND;
2897 StructCopy(&accum, data->start_class,
2898 struct regnode_charclass_class);
2899 flags |= SCF_DO_STCLASS_OR;
2900 data->start_class->flags |= ANYOF_EOS;
2904 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
2907 Assuming this was/is a branch we are dealing with: 'scan' now
2908 points at the item that follows the branch sequence, whatever
2909 it is. We now start at the beginning of the sequence and look
2916 which would be constructed from a pattern like /A|LIST|OF|WORDS/
2918 If we can find such a subsequence we need to turn the first
2919 element into a trie and then add the subsequent branch exact
2920 strings to the trie.
2924 1. patterns where the whole set of branches can be converted.
2926 2. patterns where only a subset can be converted.
2928 In case 1 we can replace the whole set with a single regop
2929 for the trie. In case 2 we need to keep the start and end
2932 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
2933 becomes BRANCH TRIE; BRANCH X;
2935 There is an additional case, that being where there is a
2936 common prefix, which gets split out into an EXACT like node
2937 preceding the TRIE node.
2939 If x(1..n)==tail then we can do a simple trie, if not we make
2940 a "jump" trie, such that when we match the appropriate word
2941 we "jump" to the appropriate tail node. Essentially we turn
2942 a nested if into a case structure of sorts.
2947 if (!re_trie_maxbuff) {
2948 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2949 if (!SvIOK(re_trie_maxbuff))
2950 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2952 if ( SvIV(re_trie_maxbuff)>=0 ) {
2954 regnode *first = (regnode *)NULL;
2955 regnode *last = (regnode *)NULL;
2956 regnode *tail = scan;
2961 SV * const mysv = sv_newmortal(); /* for dumping */
2963 /* var tail is used because there may be a TAIL
2964 regop in the way. Ie, the exacts will point to the
2965 thing following the TAIL, but the last branch will
2966 point at the TAIL. So we advance tail. If we
2967 have nested (?:) we may have to move through several
2971 while ( OP( tail ) == TAIL ) {
2972 /* this is the TAIL generated by (?:) */
2973 tail = regnext( tail );
2978 regprop(RExC_rx, mysv, tail );
2979 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
2980 (int)depth * 2 + 2, "",
2981 "Looking for TRIE'able sequences. Tail node is: ",
2982 SvPV_nolen_const( mysv )
2988 step through the branches, cur represents each
2989 branch, noper is the first thing to be matched
2990 as part of that branch and noper_next is the
2991 regnext() of that node. if noper is an EXACT
2992 and noper_next is the same as scan (our current
2993 position in the regex) then the EXACT branch is
2994 a possible optimization target. Once we have
2995 two or more consecutive such branches we can
2996 create a trie of the EXACT's contents and stich
2997 it in place. If the sequence represents all of
2998 the branches we eliminate the whole thing and
2999 replace it with a single TRIE. If it is a
3000 subsequence then we need to stitch it in. This
3001 means the first branch has to remain, and needs
3002 to be repointed at the item on the branch chain
3003 following the last branch optimized. This could
3004 be either a BRANCH, in which case the
3005 subsequence is internal, or it could be the
3006 item following the branch sequence in which
3007 case the subsequence is at the end.
3011 /* dont use tail as the end marker for this traverse */
3012 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3013 regnode * const noper = NEXTOPER( cur );
3014 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3015 regnode * const noper_next = regnext( noper );
3019 regprop(RExC_rx, mysv, cur);
3020 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3021 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3023 regprop(RExC_rx, mysv, noper);
3024 PerlIO_printf( Perl_debug_log, " -> %s",
3025 SvPV_nolen_const(mysv));
3028 regprop(RExC_rx, mysv, noper_next );
3029 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3030 SvPV_nolen_const(mysv));
3032 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
3033 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
3035 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
3036 : PL_regkind[ OP( noper ) ] == EXACT )
3037 || OP(noper) == NOTHING )
3039 && noper_next == tail
3044 if ( !first || optype == NOTHING ) {
3045 if (!first) first = cur;
3046 optype = OP( noper );
3052 Currently the trie logic handles case insensitive matching properly only
3053 when the pattern is UTF-8 and the node is EXACTFU (thus forcing unicode
3056 If/when this is fixed the following define can be swapped
3057 in below to fully enable trie logic.
3059 #define TRIE_TYPE_IS_SAFE 1
3062 #define TRIE_TYPE_IS_SAFE ((UTF && optype == EXACTFU) || optype==EXACT)
3064 if ( last && TRIE_TYPE_IS_SAFE ) {
3065 make_trie( pRExC_state,
3066 startbranch, first, cur, tail, count,
3069 if ( PL_regkind[ OP( noper ) ] == EXACT
3071 && noper_next == tail
3076 optype = OP( noper );
3086 regprop(RExC_rx, mysv, cur);
3087 PerlIO_printf( Perl_debug_log,
3088 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3089 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3093 if ( last && TRIE_TYPE_IS_SAFE ) {
3094 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
3095 #ifdef TRIE_STUDY_OPT
3096 if ( ((made == MADE_EXACT_TRIE &&
3097 startbranch == first)
3098 || ( first_non_open == first )) &&
3100 flags |= SCF_TRIE_RESTUDY;
3101 if ( startbranch == first
3104 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3114 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3115 scan = NEXTOPER(NEXTOPER(scan));
3116 } else /* single branch is optimized. */
3117 scan = NEXTOPER(scan);
3119 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3120 scan_frame *newframe = NULL;
3125 if (OP(scan) != SUSPEND) {
3126 /* set the pointer */
3127 if (OP(scan) == GOSUB) {
3129 RExC_recurse[ARG2L(scan)] = scan;
3130 start = RExC_open_parens[paren-1];
3131 end = RExC_close_parens[paren-1];
3134 start = RExC_rxi->program + 1;
3138 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3139 SAVEFREEPV(recursed);
3141 if (!PAREN_TEST(recursed,paren+1)) {
3142 PAREN_SET(recursed,paren+1);
3143 Newx(newframe,1,scan_frame);
3145 if (flags & SCF_DO_SUBSTR) {
3146 SCAN_COMMIT(pRExC_state,data,minlenp);
3147 data->longest = &(data->longest_float);
3149 is_inf = is_inf_internal = 1;
3150 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3151 cl_anything(pRExC_state, data->start_class);
3152 flags &= ~SCF_DO_STCLASS;
3155 Newx(newframe,1,scan_frame);
3158 end = regnext(scan);
3163 SAVEFREEPV(newframe);
3164 newframe->next = regnext(scan);
3165 newframe->last = last;
3166 newframe->stop = stopparen;
3167 newframe->prev = frame;
3177 else if (OP(scan) == EXACT) {
3178 I32 l = STR_LEN(scan);
3181 const U8 * const s = (U8*)STRING(scan);
3182 l = utf8_length(s, s + l);
3183 uc = utf8_to_uvchr(s, NULL);
3185 uc = *((U8*)STRING(scan));
3188 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3189 /* The code below prefers earlier match for fixed
3190 offset, later match for variable offset. */
3191 if (data->last_end == -1) { /* Update the start info. */
3192 data->last_start_min = data->pos_min;
3193 data->last_start_max = is_inf
3194 ? I32_MAX : data->pos_min + data->pos_delta;
3196 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3198 SvUTF8_on(data->last_found);
3200 SV * const sv = data->last_found;
3201 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3202 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3203 if (mg && mg->mg_len >= 0)
3204 mg->mg_len += utf8_length((U8*)STRING(scan),
3205 (U8*)STRING(scan)+STR_LEN(scan));
3207 data->last_end = data->pos_min + l;
3208 data->pos_min += l; /* As in the first entry. */
3209 data->flags &= ~SF_BEFORE_EOL;
3211 if (flags & SCF_DO_STCLASS_AND) {
3212 /* Check whether it is compatible with what we know already! */
3216 /* If compatible, we or it in below. It is compatible if is
3217 * in the bitmp and either 1) its bit or its fold is set, or 2)
3218 * it's for a locale. Even if there isn't unicode semantics
3219 * here, at runtime there may be because of matching against a
3220 * utf8 string, so accept a possible false positive for
3221 * latin1-range folds */
3223 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3224 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3225 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
3226 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3231 ANYOF_CLASS_ZERO(data->start_class);
3232 ANYOF_BITMAP_ZERO(data->start_class);
3234 ANYOF_BITMAP_SET(data->start_class, uc);
3235 else if (uc >= 0x100) {
3238 /* Some Unicode code points fold to the Latin1 range; as
3239 * XXX temporary code, instead of figuring out if this is
3240 * one, just assume it is and set all the start class bits
3241 * that could be some such above 255 code point's fold
3242 * which will generate fals positives. As the code
3243 * elsewhere that does compute the fold settles down, it
3244 * can be extracted out and re-used here */
3245 for (i = 0; i < 256; i++){
3246 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3247 ANYOF_BITMAP_SET(data->start_class, i);
3251 data->start_class->flags &= ~ANYOF_EOS;
3253 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3255 else if (flags & SCF_DO_STCLASS_OR) {
3256 /* false positive possible if the class is case-folded */
3258 ANYOF_BITMAP_SET(data->start_class, uc);
3260 data->start_class->flags |= ANYOF_UNICODE_ALL;
3261 data->start_class->flags &= ~ANYOF_EOS;
3262 cl_and(data->start_class, and_withp);
3264 flags &= ~SCF_DO_STCLASS;
3266 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3267 I32 l = STR_LEN(scan);
3268 UV uc = *((U8*)STRING(scan));
3270 /* Search for fixed substrings supports EXACT only. */
3271 if (flags & SCF_DO_SUBSTR) {
3273 SCAN_COMMIT(pRExC_state, data, minlenp);
3276 const U8 * const s = (U8 *)STRING(scan);
3277 l = utf8_length(s, s + l);
3278 uc = utf8_to_uvchr(s, NULL);
3281 if (flags & SCF_DO_SUBSTR)
3283 if (flags & SCF_DO_STCLASS_AND) {
3284 /* Check whether it is compatible with what we know already! */
3287 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3288 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3289 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3293 ANYOF_CLASS_ZERO(data->start_class);
3294 ANYOF_BITMAP_ZERO(data->start_class);
3296 ANYOF_BITMAP_SET(data->start_class, uc);
3297 data->start_class->flags &= ~ANYOF_EOS;
3298 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
3299 if (OP(scan) == EXACTFL) {
3300 /* XXX This set is probably no longer necessary, and
3301 * probably wrong as LOCALE now is on in the initial
3303 data->start_class->flags |= ANYOF_LOCALE;
3307 /* Also set the other member of the fold pair. In case
3308 * that unicode semantics is called for at runtime, use
3309 * the full latin1 fold. (Can't do this for locale,
3310 * because not known until runtime */
3311 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3314 else if (uc >= 0x100) {
3316 for (i = 0; i < 256; i++){
3317 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3318 ANYOF_BITMAP_SET(data->start_class, i);
3323 else if (flags & SCF_DO_STCLASS_OR) {
3324 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
3325 /* false positive possible if the class is case-folded.
3326 Assume that the locale settings are the same... */
3328 ANYOF_BITMAP_SET(data->start_class, uc);
3329 if (OP(scan) != EXACTFL) {
3331 /* And set the other member of the fold pair, but
3332 * can't do that in locale because not known until
3334 ANYOF_BITMAP_SET(data->start_class,
3335 PL_fold_latin1[uc]);
3338 data->start_class->flags &= ~ANYOF_EOS;
3340 cl_and(data->start_class, and_withp);
3342 flags &= ~SCF_DO_STCLASS;
3344 else if (REGNODE_VARIES(OP(scan))) {
3345 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3346 I32 f = flags, pos_before = 0;
3347 regnode * const oscan = scan;
3348 struct regnode_charclass_class this_class;
3349 struct regnode_charclass_class *oclass = NULL;
3350 I32 next_is_eval = 0;
3352 switch (PL_regkind[OP(scan)]) {
3353 case WHILEM: /* End of (?:...)* . */
3354 scan = NEXTOPER(scan);
3357 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3358 next = NEXTOPER(scan);
3359 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3361 maxcount = REG_INFTY;
3362 next = regnext(scan);
3363 scan = NEXTOPER(scan);
3367 if (flags & SCF_DO_SUBSTR)
3372 if (flags & SCF_DO_STCLASS) {
3374 maxcount = REG_INFTY;
3375 next = regnext(scan);
3376 scan = NEXTOPER(scan);
3379 is_inf = is_inf_internal = 1;
3380 scan = regnext(scan);
3381 if (flags & SCF_DO_SUBSTR) {
3382 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3383 data->longest = &(data->longest_float);
3385 goto optimize_curly_tail;
3387 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3388 && (scan->flags == stopparen))
3393 mincount = ARG1(scan);
3394 maxcount = ARG2(scan);
3396 next = regnext(scan);
3397 if (OP(scan) == CURLYX) {
3398 I32 lp = (data ? *(data->last_closep) : 0);
3399 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3401 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3402 next_is_eval = (OP(scan) == EVAL);
3404 if (flags & SCF_DO_SUBSTR) {
3405 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3406 pos_before = data->pos_min;
3410 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3412 data->flags |= SF_IS_INF;
3414 if (flags & SCF_DO_STCLASS) {
3415 cl_init(pRExC_state, &this_class);
3416 oclass = data->start_class;
3417 data->start_class = &this_class;
3418 f |= SCF_DO_STCLASS_AND;
3419 f &= ~SCF_DO_STCLASS_OR;
3421 /* Exclude from super-linear cache processing any {n,m}
3422 regops for which the combination of input pos and regex
3423 pos is not enough information to determine if a match
3426 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3427 regex pos at the \s*, the prospects for a match depend not
3428 only on the input position but also on how many (bar\s*)
3429 repeats into the {4,8} we are. */
3430 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3431 f &= ~SCF_WHILEM_VISITED_POS;
3433 /* This will finish on WHILEM, setting scan, or on NULL: */
3434 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3435 last, data, stopparen, recursed, NULL,
3437 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3439 if (flags & SCF_DO_STCLASS)
3440 data->start_class = oclass;
3441 if (mincount == 0 || minnext == 0) {
3442 if (flags & SCF_DO_STCLASS_OR) {
3443 cl_or(pRExC_state, data->start_class, &this_class);
3445 else if (flags & SCF_DO_STCLASS_AND) {
3446 /* Switch to OR mode: cache the old value of
3447 * data->start_class */
3449 StructCopy(data->start_class, and_withp,
3450 struct regnode_charclass_class);
3451 flags &= ~SCF_DO_STCLASS_AND;
3452 StructCopy(&this_class, data->start_class,
3453 struct regnode_charclass_class);
3454 flags |= SCF_DO_STCLASS_OR;
3455 data->start_class->flags |= ANYOF_EOS;
3457 } else { /* Non-zero len */
3458 if (flags & SCF_DO_STCLASS_OR) {
3459 cl_or(pRExC_state, data->start_class, &this_class);
3460 cl_and(data->start_class, and_withp);
3462 else if (flags & SCF_DO_STCLASS_AND)
3463 cl_and(data->start_class, &this_class);
3464 flags &= ~SCF_DO_STCLASS;
3466 if (!scan) /* It was not CURLYX, but CURLY. */
3468 if ( /* ? quantifier ok, except for (?{ ... }) */
3469 (next_is_eval || !(mincount == 0 && maxcount == 1))
3470 && (minnext == 0) && (deltanext == 0)
3471 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3472 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3474 ckWARNreg(RExC_parse,
3475 "Quantifier unexpected on zero-length expression");
3478 min += minnext * mincount;
3479 is_inf_internal |= ((maxcount == REG_INFTY
3480 && (minnext + deltanext) > 0)
3481 || deltanext == I32_MAX);
3482 is_inf |= is_inf_internal;
3483 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3485 /* Try powerful optimization CURLYX => CURLYN. */
3486 if ( OP(oscan) == CURLYX && data
3487 && data->flags & SF_IN_PAR
3488 && !(data->flags & SF_HAS_EVAL)
3489 && !deltanext && minnext == 1 ) {
3490 /* Try to optimize to CURLYN. */
3491 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3492 regnode * const nxt1 = nxt;
3499 if (!REGNODE_SIMPLE(OP(nxt))
3500 && !(PL_regkind[OP(nxt)] == EXACT
3501 && STR_LEN(nxt) == 1))
3507 if (OP(nxt) != CLOSE)
3509 if (RExC_open_parens) {
3510 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3511 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3513 /* Now we know that nxt2 is the only contents: */
3514 oscan->flags = (U8)ARG(nxt);
3516 OP(nxt1) = NOTHING; /* was OPEN. */
3519 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3520 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3521 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3522 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3523 OP(nxt + 1) = OPTIMIZED; /* was count. */
3524 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3529 /* Try optimization CURLYX => CURLYM. */
3530 if ( OP(oscan) == CURLYX && data
3531 && !(data->flags & SF_HAS_PAR)
3532 && !(data->flags & SF_HAS_EVAL)
3533 && !deltanext /* atom is fixed width */
3534 && minnext != 0 /* CURLYM can't handle zero width */
3536 /* XXXX How to optimize if data == 0? */
3537 /* Optimize to a simpler form. */
3538 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3542 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3543 && (OP(nxt2) != WHILEM))
3545 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3546 /* Need to optimize away parenths. */
3547 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
3548 /* Set the parenth number. */
3549 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3551 oscan->flags = (U8)ARG(nxt);
3552 if (RExC_open_parens) {
3553 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3554 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3556 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3557 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3560 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3561 OP(nxt + 1) = OPTIMIZED; /* was count. */
3562 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
3563 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
3566 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3567 regnode *nnxt = regnext(nxt1);
3569 if (reg_off_by_arg[OP(nxt1)])
3570 ARG_SET(nxt1, nxt2 - nxt1);
3571 else if (nxt2 - nxt1 < U16_MAX)
3572 NEXT_OFF(nxt1) = nxt2 - nxt1;
3574 OP(nxt) = NOTHING; /* Cannot beautify */
3579 /* Optimize again: */
3580 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
3581 NULL, stopparen, recursed, NULL, 0,depth+1);
3586 else if ((OP(oscan) == CURLYX)
3587 && (flags & SCF_WHILEM_VISITED_POS)
3588 /* See the comment on a similar expression above.
3589 However, this time it's not a subexpression
3590 we care about, but the expression itself. */
3591 && (maxcount == REG_INFTY)
3592 && data && ++data->whilem_c < 16) {
3593 /* This stays as CURLYX, we can put the count/of pair. */
3594 /* Find WHILEM (as in regexec.c) */
3595 regnode *nxt = oscan + NEXT_OFF(oscan);
3597 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
3599 PREVOPER(nxt)->flags = (U8)(data->whilem_c
3600 | (RExC_whilem_seen << 4)); /* On WHILEM */
3602 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
3604 if (flags & SCF_DO_SUBSTR) {
3605 SV *last_str = NULL;
3606 int counted = mincount != 0;
3608 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
3609 #if defined(SPARC64_GCC_WORKAROUND)
3612 const char *s = NULL;
3615 if (pos_before >= data->last_start_min)
3618 b = data->last_start_min;
3621 s = SvPV_const(data->last_found, l);
3622 old = b - data->last_start_min;
3625 I32 b = pos_before >= data->last_start_min
3626 ? pos_before : data->last_start_min;
3628 const char * const s = SvPV_const(data->last_found, l);
3629 I32 old = b - data->last_start_min;
3633 old = utf8_hop((U8*)s, old) - (U8*)s;
3635 /* Get the added string: */
3636 last_str = newSVpvn_utf8(s + old, l, UTF);
3637 if (deltanext == 0 && pos_before == b) {
3638 /* What was added is a constant string */
3640 SvGROW(last_str, (mincount * l) + 1);
3641 repeatcpy(SvPVX(last_str) + l,
3642 SvPVX_const(last_str), l, mincount - 1);
3643 SvCUR_set(last_str, SvCUR(last_str) * mincount);
3644 /* Add additional parts. */
3645 SvCUR_set(data->last_found,
3646 SvCUR(data->last_found) - l);
3647 sv_catsv(data->last_found, last_str);
3649 SV * sv = data->last_found;
3651 SvUTF8(sv) && SvMAGICAL(sv) ?
3652 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3653 if (mg && mg->mg_len >= 0)
3654 mg->mg_len += CHR_SVLEN(last_str) - l;
3656 data->last_end += l * (mincount - 1);
3659 /* start offset must point into the last copy */
3660 data->last_start_min += minnext * (mincount - 1);
3661 data->last_start_max += is_inf ? I32_MAX
3662 : (maxcount - 1) * (minnext + data->pos_delta);
3665 /* It is counted once already... */
3666 data->pos_min += minnext * (mincount - counted);
3667 data->pos_delta += - counted * deltanext +
3668 (minnext + deltanext) * maxcount - minnext * mincount;
3669 if (mincount != maxcount) {
3670 /* Cannot extend fixed substrings found inside
3672 SCAN_COMMIT(pRExC_state,data,minlenp);
3673 if (mincount && last_str) {
3674 SV * const sv = data->last_found;
3675 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3676 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3680 sv_setsv(sv, last_str);
3681 data->last_end = data->pos_min;
3682 data->last_start_min =
3683 data->pos_min - CHR_SVLEN(last_str);
3684 data->last_start_max = is_inf
3686 : data->pos_min + data->pos_delta
3687 - CHR_SVLEN(last_str);
3689 data->longest = &(data->longest_float);
3691 SvREFCNT_dec(last_str);
3693 if (data && (fl & SF_HAS_EVAL))
3694 data->flags |= SF_HAS_EVAL;
3695 optimize_curly_tail:
3696 if (OP(oscan) != CURLYX) {
3697 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
3699 NEXT_OFF(oscan) += NEXT_OFF(next);
3702 default: /* REF, ANYOFV, and CLUMP only? */
3703 if (flags & SCF_DO_SUBSTR) {
3704 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3705 data->longest = &(data->longest_float);
3707 is_inf = is_inf_internal = 1;
3708 if (flags & SCF_DO_STCLASS_OR)
3709 cl_anything(pRExC_state, data->start_class);
3710 flags &= ~SCF_DO_STCLASS;
3714 else if (OP(scan) == LNBREAK) {
3715 if (flags & SCF_DO_STCLASS) {
3717 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3718 if (flags & SCF_DO_STCLASS_AND) {
3719 for (value = 0; value < 256; value++)
3720 if (!is_VERTWS_cp(value))
3721 ANYOF_BITMAP_CLEAR(data->start_class, value);
3724 for (value = 0; value < 256; value++)
3725 if (is_VERTWS_cp(value))
3726 ANYOF_BITMAP_SET(data->start_class, value);
3728 if (flags & SCF_DO_STCLASS_OR)
3729 cl_and(data->start_class, and_withp);
3730 flags &= ~SCF_DO_STCLASS;
3734 if (flags & SCF_DO_SUBSTR) {
3735 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3737 data->pos_delta += 1;
3738 data->longest = &(data->longest_float);
3741 else if (OP(scan) == FOLDCHAR) {
3742 int d = ARG(scan) == LATIN_SMALL_LETTER_SHARP_S ? 1 : 2;
3743 flags &= ~SCF_DO_STCLASS;
3746 if (flags & SCF_DO_SUBSTR) {
3747 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3749 data->pos_delta += d;
3750 data->longest = &(data->longest_float);
3753 else if (REGNODE_SIMPLE(OP(scan))) {
3756 if (flags & SCF_DO_SUBSTR) {
3757 SCAN_COMMIT(pRExC_state,data,minlenp);
3761 if (flags & SCF_DO_STCLASS) {
3762 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3764 /* Some of the logic below assumes that switching
3765 locale on will only add false positives. */
3766 switch (PL_regkind[OP(scan)]) {
3770 /* Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan)); */
3771 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3772 cl_anything(pRExC_state, data->start_class);
3775 if (OP(scan) == SANY)
3777 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
3778 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
3779 || ANYOF_CLASS_TEST_ANY_SET(data->start_class));
3780 cl_anything(pRExC_state, data->start_class);
3782 if (flags & SCF_DO_STCLASS_AND || !value)
3783 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
3786 if (flags & SCF_DO_STCLASS_AND)
3787 cl_and(data->start_class,
3788 (struct regnode_charclass_class*)scan);
3790 cl_or(pRExC_state, data->start_class,
3791 (struct regnode_charclass_class*)scan);
3794 if (flags & SCF_DO_STCLASS_AND) {
3795 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3796 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3797 if (OP(scan) == ALNUMU) {
3798 for (value = 0; value < 256; value++) {
3799 if (!isWORDCHAR_L1(value)) {
3800 ANYOF_BITMAP_CLEAR(data->start_class, value);
3804 for (value = 0; value < 256; value++) {
3805 if (!isALNUM(value)) {
3806 ANYOF_BITMAP_CLEAR(data->start_class, value);
3813 if (data->start_class->flags & ANYOF_LOCALE)
3814 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3816 /* Even if under locale, set the bits for non-locale
3817 * in case it isn't a true locale-node. This will
3818 * create false positives if it truly is locale */
3819 if (OP(scan) == ALNUMU) {
3820 for (value = 0; value < 256; value++) {
3821 if (isWORDCHAR_L1(value)) {
3822 ANYOF_BITMAP_SET(data->start_class, value);
3826 for (value = 0; value < 256; value++) {
3827 if (isALNUM(value)) {
3828 ANYOF_BITMAP_SET(data->start_class, value);
3835 if (flags & SCF_DO_STCLASS_AND) {
3836 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3837 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3838 if (OP(scan) == NALNUMU) {
3839 for (value = 0; value < 256; value++) {
3840 if (isWORDCHAR_L1(value)) {
3841 ANYOF_BITMAP_CLEAR(data->start_class, value);
3845 for (value = 0; value < 256; value++) {
3846 if (isALNUM(value)) {
3847 ANYOF_BITMAP_CLEAR(data->start_class, value);
3854 if (data->start_class->flags & ANYOF_LOCALE)
3855 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3857 /* Even if under locale, set the bits for non-locale in
3858 * case it isn't a true locale-node. This will create
3859 * false positives if it truly is locale */
3860 if (OP(scan) == NALNUMU) {
3861 for (value = 0; value < 256; value++) {
3862 if (! isWORDCHAR_L1(value)) {
3863 ANYOF_BITMAP_SET(data->start_class, value);
3867 for (value = 0; value < 256; value++) {
3868 if (! isALNUM(value)) {
3869 ANYOF_BITMAP_SET(data->start_class, value);
3876 if (flags & SCF_DO_STCLASS_AND) {
3877 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3878 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3879 if (OP(scan) == SPACEU) {
3880 for (value = 0; value < 256; value++) {
3881 if (!isSPACE_L1(value)) {
3882 ANYOF_BITMAP_CLEAR(data->start_class, value);
3886 for (value = 0; value < 256; value++) {
3887 if (!isSPACE(value)) {
3888 ANYOF_BITMAP_CLEAR(data->start_class, value);
3895 if (data->start_class->flags & ANYOF_LOCALE) {
3896 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3898 if (OP(scan) == SPACEU) {
3899 for (value = 0; value < 256; value++) {
3900 if (isSPACE_L1(value)) {
3901 ANYOF_BITMAP_SET(data->start_class, value);
3905 for (value = 0; value < 256; value++) {
3906 if (isSPACE(value)) {
3907 ANYOF_BITMAP_SET(data->start_class, value);
3914 if (flags & SCF_DO_STCLASS_AND) {
3915 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3916 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3917 if (OP(scan) == NSPACEU) {
3918 for (value = 0; value < 256; value++) {
3919 if (isSPACE_L1(value)) {
3920 ANYOF_BITMAP_CLEAR(data->start_class, value);
3924 for (value = 0; value < 256; value++) {
3925 if (isSPACE(value)) {
3926 ANYOF_BITMAP_CLEAR(data->start_class, value);
3933 if (data->start_class->flags & ANYOF_LOCALE)
3934 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3935 if (OP(scan) == NSPACEU) {
3936 for (value = 0; value < 256; value++) {
3937 if (!isSPACE_L1(value)) {
3938 ANYOF_BITMAP_SET(data->start_class, value);
3943 for (value = 0; value < 256; value++) {
3944 if (!isSPACE(value)) {
3945 ANYOF_BITMAP_SET(data->start_class, value);
3952 if (flags & SCF_DO_STCLASS_AND) {
3953 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3954 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NDIGIT);
3955 for (value = 0; value < 256; value++)
3956 if (!isDIGIT(value))
3957 ANYOF_BITMAP_CLEAR(data->start_class, value);
3961 if (data->start_class->flags & ANYOF_LOCALE)
3962 ANYOF_CLASS_SET(data->start_class,ANYOF_DIGIT);
3963 for (value = 0; value < 256; value++)
3965 ANYOF_BITMAP_SET(data->start_class, value);
3969 if (flags & SCF_DO_STCLASS_AND) {
3970 if (!(data->start_class->flags & ANYOF_LOCALE))
3971 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_DIGIT);
3972 for (value = 0; value < 256; value++)
3974 ANYOF_BITMAP_CLEAR(data->start_class, value);
3977 if (data->start_class->flags & ANYOF_LOCALE)
3978 ANYOF_CLASS_SET(data->start_class,ANYOF_NDIGIT);
3979 for (value = 0; value < 256; value++)
3980 if (!isDIGIT(value))
3981 ANYOF_BITMAP_SET(data->start_class, value);
3984 CASE_SYNST_FNC(VERTWS);
3985 CASE_SYNST_FNC(HORIZWS);
3988 if (flags & SCF_DO_STCLASS_OR)
3989 cl_and(data->start_class, and_withp);
3990 flags &= ~SCF_DO_STCLASS;
3993 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
3994 data->flags |= (OP(scan) == MEOL
3998 else if ( PL_regkind[OP(scan)] == BRANCHJ
3999 /* Lookbehind, or need to calculate parens/evals/stclass: */
4000 && (scan->flags || data || (flags & SCF_DO_STCLASS))
4001 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
4002 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4003 || OP(scan) == UNLESSM )
4005 /* Negative Lookahead/lookbehind
4006 In this case we can't do fixed string optimisation.
4009 I32 deltanext, minnext, fake = 0;
4011 struct regnode_charclass_class intrnl;
4014 data_fake.flags = 0;
4016 data_fake.whilem_c = data->whilem_c;
4017 data_fake.last_closep = data->last_closep;
4020 data_fake.last_closep = &fake;
4021 data_fake.pos_delta = delta;
4022 if ( flags & SCF_DO_STCLASS && !scan->flags
4023 && OP(scan) == IFMATCH ) { /* Lookahead */
4024 cl_init(pRExC_state, &intrnl);
4025 data_fake.start_class = &intrnl;
4026 f |= SCF_DO_STCLASS_AND;
4028 if (flags & SCF_WHILEM_VISITED_POS)
4029 f |= SCF_WHILEM_VISITED_POS;
4030 next = regnext(scan);
4031 nscan = NEXTOPER(NEXTOPER(scan));
4032 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
4033 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
4036 FAIL("Variable length lookbehind not implemented");
4038 else if (minnext > (I32)U8_MAX) {
4039 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4041 scan->flags = (U8)minnext;
4044 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4046 if (data_fake.flags & SF_HAS_EVAL)
4047 data->flags |= SF_HAS_EVAL;
4048 data->whilem_c = data_fake.whilem_c;
4050 if (f & SCF_DO_STCLASS_AND) {
4051 if (flags & SCF_DO_STCLASS_OR) {
4052 /* OR before, AND after: ideally we would recurse with
4053 * data_fake to get the AND applied by study of the
4054 * remainder of the pattern, and then derecurse;
4055 * *** HACK *** for now just treat as "no information".
4056 * See [perl #56690].
4058 cl_init(pRExC_state, data->start_class);
4060 /* AND before and after: combine and continue */
4061 const int was = (data->start_class->flags & ANYOF_EOS);
4063 cl_and(data->start_class, &intrnl);
4065 data->start_class->flags |= ANYOF_EOS;
4069 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4071 /* Positive Lookahead/lookbehind
4072 In this case we can do fixed string optimisation,
4073 but we must be careful about it. Note in the case of
4074 lookbehind the positions will be offset by the minimum
4075 length of the pattern, something we won't know about
4076 until after the recurse.
4078 I32 deltanext, fake = 0;
4080 struct regnode_charclass_class intrnl;
4082 /* We use SAVEFREEPV so that when the full compile
4083 is finished perl will clean up the allocated
4084 minlens when it's all done. This way we don't
4085 have to worry about freeing them when we know
4086 they wont be used, which would be a pain.
4089 Newx( minnextp, 1, I32 );
4090 SAVEFREEPV(minnextp);
4093 StructCopy(data, &data_fake, scan_data_t);
4094 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
4097 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
4098 data_fake.last_found=newSVsv(data->last_found);
4102 data_fake.last_closep = &fake;
4103 data_fake.flags = 0;
4104 data_fake.pos_delta = delta;
4106 data_fake.flags |= SF_IS_INF;
4107 if ( flags & SCF_DO_STCLASS && !scan->flags
4108 && OP(scan) == IFMATCH ) { /* Lookahead */
4109 cl_init(pRExC_state, &intrnl);
4110 data_fake.start_class = &intrnl;
4111 f |= SCF_DO_STCLASS_AND;
4113 if (flags & SCF_WHILEM_VISITED_POS)
4114 f |= SCF_WHILEM_VISITED_POS;
4115 next = regnext(scan);
4116 nscan = NEXTOPER(NEXTOPER(scan));
4118 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
4119 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
4122 FAIL("Variable length lookbehind not implemented");
4124 else if (*minnextp > (I32)U8_MAX) {
4125 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4127 scan->flags = (U8)*minnextp;
4132 if (f & SCF_DO_STCLASS_AND) {
4133 const int was = (data->start_class->flags & ANYOF_EOS);
4135 cl_and(data->start_class, &intrnl);
4137 data->start_class->flags |= ANYOF_EOS;
4140 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4142 if (data_fake.flags & SF_HAS_EVAL)
4143 data->flags |= SF_HAS_EVAL;
4144 data->whilem_c = data_fake.whilem_c;
4145 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
4146 if (RExC_rx->minlen<*minnextp)
4147 RExC_rx->minlen=*minnextp;
4148 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
4149 SvREFCNT_dec(data_fake.last_found);
4151 if ( data_fake.minlen_fixed != minlenp )
4153 data->offset_fixed= data_fake.offset_fixed;
4154 data->minlen_fixed= data_fake.minlen_fixed;
4155 data->lookbehind_fixed+= scan->flags;
4157 if ( data_fake.minlen_float != minlenp )
4159 data->minlen_float= data_fake.minlen_float;
4160 data->offset_float_min=data_fake.offset_float_min;
4161 data->offset_float_max=data_fake.offset_float_max;
4162 data->lookbehind_float+= scan->flags;
4171 else if (OP(scan) == OPEN) {
4172 if (stopparen != (I32)ARG(scan))
4175 else if (OP(scan) == CLOSE) {
4176 if (stopparen == (I32)ARG(scan)) {
4179 if ((I32)ARG(scan) == is_par) {
4180 next = regnext(scan);
4182 if ( next && (OP(next) != WHILEM) && next < last)
4183 is_par = 0; /* Disable optimization */
4186 *(data->last_closep) = ARG(scan);
4188 else if (OP(scan) == EVAL) {
4190 data->flags |= SF_HAS_EVAL;
4192 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
4193 if (flags & SCF_DO_SUBSTR) {
4194 SCAN_COMMIT(pRExC_state,data,minlenp);
4195 flags &= ~SCF_DO_SUBSTR;
4197 if (data && OP(scan)==ACCEPT) {
4198 data->flags |= SCF_SEEN_ACCEPT;
4203 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
4205 if (flags & SCF_DO_SUBSTR) {
4206 SCAN_COMMIT(pRExC_state,data,minlenp);
4207 data->longest = &(data->longest_float);
4209 is_inf = is_inf_internal = 1;
4210 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4211 cl_anything(pRExC_state, data->start_class);
4212 flags &= ~SCF_DO_STCLASS;
4214 else if (OP(scan) == GPOS) {
4215 if (!(RExC_rx->extflags & RXf_GPOS_FLOAT) &&
4216 !(delta || is_inf || (data && data->pos_delta)))
4218 if (!(RExC_rx->extflags & RXf_ANCH) && (flags & SCF_DO_SUBSTR))
4219 RExC_rx->extflags |= RXf_ANCH_GPOS;
4220 if (RExC_rx->gofs < (U32)min)
4221 RExC_rx->gofs = min;
4223 RExC_rx->extflags |= RXf_GPOS_FLOAT;
4227 #ifdef TRIE_STUDY_OPT
4228 #ifdef FULL_TRIE_STUDY
4229 else if (PL_regkind[OP(scan)] == TRIE) {
4230 /* NOTE - There is similar code to this block above for handling
4231 BRANCH nodes on the initial study. If you change stuff here
4233 regnode *trie_node= scan;
4234 regnode *tail= regnext(scan);
4235 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4236 I32 max1 = 0, min1 = I32_MAX;
4237 struct regnode_charclass_class accum;
4239 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
4240 SCAN_COMMIT(pRExC_state, data,minlenp); /* Cannot merge strings after this. */
4241 if (flags & SCF_DO_STCLASS)
4242 cl_init_zero(pRExC_state, &accum);
4248 const regnode *nextbranch= NULL;
4251 for ( word=1 ; word <= trie->wordcount ; word++)
4253 I32 deltanext=0, minnext=0, f = 0, fake;
4254 struct regnode_charclass_class this_class;
4256 data_fake.flags = 0;
4258 data_fake.whilem_c = data->whilem_c;
4259 data_fake.last_closep = data->last_closep;
4262 data_fake.last_closep = &fake;
4263 data_fake.pos_delta = delta;
4264 if (flags & SCF_DO_STCLASS) {
4265 cl_init(pRExC_state, &this_class);
4266 data_fake.start_class = &this_class;
4267 f = SCF_DO_STCLASS_AND;
4269 if (flags & SCF_WHILEM_VISITED_POS)
4270 f |= SCF_WHILEM_VISITED_POS;
4272 if (trie->jump[word]) {
4274 nextbranch = trie_node + trie->jump[0];
4275 scan= trie_node + trie->jump[word];
4276 /* We go from the jump point to the branch that follows
4277 it. Note this means we need the vestigal unused branches
4278 even though they arent otherwise used.
4280 minnext = study_chunk(pRExC_state, &scan, minlenp,
4281 &deltanext, (regnode *)nextbranch, &data_fake,
4282 stopparen, recursed, NULL, f,depth+1);
4284 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
4285 nextbranch= regnext((regnode*)nextbranch);
4287 if (min1 > (I32)(minnext + trie->minlen))
4288 min1 = minnext + trie->minlen;
4289 if (max1 < (I32)(minnext + deltanext + trie->maxlen))
4290 max1 = minnext + deltanext + trie->maxlen;
4291 if (deltanext == I32_MAX)
4292 is_inf = is_inf_internal = 1;
4294 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4296 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4297 if ( stopmin > min + min1)
4298 stopmin = min + min1;
4299 flags &= ~SCF_DO_SUBSTR;
4301 data->flags |= SCF_SEEN_ACCEPT;
4304 if (data_fake.flags & SF_HAS_EVAL)
4305 data->flags |= SF_HAS_EVAL;
4306 data->whilem_c = data_fake.whilem_c;
4308 if (flags & SCF_DO_STCLASS)
4309 cl_or(pRExC_state, &accum, &this_class);
4312 if (flags & SCF_DO_SUBSTR) {
4313 data->pos_min += min1;
4314 data->pos_delta += max1 - min1;
4315 if (max1 != min1 || is_inf)
4316 data->longest = &(data->longest_float);
4319 delta += max1 - min1;
4320 if (flags & SCF_DO_STCLASS_OR) {
4321 cl_or(pRExC_state, data->start_class, &accum);
4323 cl_and(data->start_class, and_withp);
4324 flags &= ~SCF_DO_STCLASS;
4327 else if (flags & SCF_DO_STCLASS_AND) {
4329 cl_and(data->start_class, &accum);
4330 flags &= ~SCF_DO_STCLASS;
4333 /* Switch to OR mode: cache the old value of
4334 * data->start_class */
4336 StructCopy(data->start_class, and_withp,
4337 struct regnode_charclass_class);
4338 flags &= ~SCF_DO_STCLASS_AND;
4339 StructCopy(&accum, data->start_class,
4340 struct regnode_charclass_class);
4341 flags |= SCF_DO_STCLASS_OR;
4342 data->start_class->flags |= ANYOF_EOS;
4349 else if (PL_regkind[OP(scan)] == TRIE) {
4350 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4353 min += trie->minlen;
4354 delta += (trie->maxlen - trie->minlen);
4355 flags &= ~SCF_DO_STCLASS; /* xxx */
4356 if (flags & SCF_DO_SUBSTR) {
4357 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4358 data->pos_min += trie->minlen;
4359 data->pos_delta += (trie->maxlen - trie->minlen);
4360 if (trie->maxlen != trie->minlen)
4361 data->longest = &(data->longest_float);
4363 if (trie->jump) /* no more substrings -- for now /grr*/
4364 flags &= ~SCF_DO_SUBSTR;
4366 #endif /* old or new */
4367 #endif /* TRIE_STUDY_OPT */
4369 /* Else: zero-length, ignore. */
4370 scan = regnext(scan);
4375 stopparen = frame->stop;
4376 frame = frame->prev;
4377 goto fake_study_recurse;
4382 DEBUG_STUDYDATA("pre-fin:",data,depth);
4385 *deltap = is_inf_internal ? I32_MAX : delta;
4386 if (flags & SCF_DO_SUBSTR && is_inf)
4387 data->pos_delta = I32_MAX - data->pos_min;
4388 if (is_par > (I32)U8_MAX)
4390 if (is_par && pars==1 && data) {
4391 data->flags |= SF_IN_PAR;
4392 data->flags &= ~SF_HAS_PAR;
4394 else if (pars && data) {
4395 data->flags |= SF_HAS_PAR;
4396 data->flags &= ~SF_IN_PAR;
4398 if (flags & SCF_DO_STCLASS_OR)
4399 cl_and(data->start_class, and_withp);
4400 if (flags & SCF_TRIE_RESTUDY)
4401 data->flags |= SCF_TRIE_RESTUDY;
4403 DEBUG_STUDYDATA("post-fin:",data,depth);
4405 return min < stopmin ? min : stopmin;
4409 S_add_data(RExC_state_t *pRExC_state, U32 n, const char *s)
4411 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
4413 PERL_ARGS_ASSERT_ADD_DATA;
4415 Renewc(RExC_rxi->data,
4416 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
4417 char, struct reg_data);
4419 Renew(RExC_rxi->data->what, count + n, U8);
4421 Newx(RExC_rxi->data->what, n, U8);
4422 RExC_rxi->data->count = count + n;
4423 Copy(s, RExC_rxi->data->what + count, n, U8);
4427 /*XXX: todo make this not included in a non debugging perl */
4428 #ifndef PERL_IN_XSUB_RE
4430 Perl_reginitcolors(pTHX)
4433 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
4435 char *t = savepv(s);
4439 t = strchr(t, '\t');
4445 PL_colors[i] = t = (char *)"";
4450 PL_colors[i++] = (char *)"";
4457 #ifdef TRIE_STUDY_OPT
4458 #define CHECK_RESTUDY_GOTO \
4460 (data.flags & SCF_TRIE_RESTUDY) \
4464 #define CHECK_RESTUDY_GOTO
4468 - pregcomp - compile a regular expression into internal code
4470 * We can't allocate space until we know how big the compiled form will be,
4471 * but we can't compile it (and thus know how big it is) until we've got a
4472 * place to put the code. So we cheat: we compile it twice, once with code
4473 * generation turned off and size counting turned on, and once "for real".
4474 * This also means that we don't allocate space until we are sure that the
4475 * thing really will compile successfully, and we never have to move the
4476 * code and thus invalidate pointers into it. (Note that it has to be in
4477 * one piece because free() must be able to free it all.) [NB: not true in perl]
4479 * Beware that the optimization-preparation code in here knows about some
4480 * of the structure of the compiled regexp. [I'll say.]
4485 #ifndef PERL_IN_XSUB_RE
4486 #define RE_ENGINE_PTR &PL_core_reg_engine
4488 extern const struct regexp_engine my_reg_engine;
4489 #define RE_ENGINE_PTR &my_reg_engine
4492 #ifndef PERL_IN_XSUB_RE
4494 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
4497 HV * const table = GvHV(PL_hintgv);
4499 PERL_ARGS_ASSERT_PREGCOMP;
4501 /* Dispatch a request to compile a regexp to correct
4504 SV **ptr= hv_fetchs(table, "regcomp", FALSE);
4505 GET_RE_DEBUG_FLAGS_DECL;
4506 if (ptr && SvIOK(*ptr) && SvIV(*ptr)) {
4507 const regexp_engine *eng=INT2PTR(regexp_engine*,SvIV(*ptr));
4509 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
4512 return CALLREGCOMP_ENG(eng, pattern, flags);
4515 return Perl_re_compile(aTHX_ pattern, flags);
4520 Perl_re_compile(pTHX_ SV * const pattern, U32 orig_pm_flags)
4525 register regexp_internal *ri;
4534 /* these are all flags - maybe they should be turned
4535 * into a single int with different bit masks */
4536 I32 sawlookahead = 0;
4539 bool used_setjump = FALSE;
4540 regex_charset initial_charset = get_regex_charset(orig_pm_flags);
4545 RExC_state_t RExC_state;
4546 RExC_state_t * const pRExC_state = &RExC_state;
4547 #ifdef TRIE_STUDY_OPT
4549 RExC_state_t copyRExC_state;
4551 GET_RE_DEBUG_FLAGS_DECL;
4553 PERL_ARGS_ASSERT_RE_COMPILE;
4555 DEBUG_r(if (!PL_colorset) reginitcolors());
4557 exp = SvPV(pattern, plen);
4559 if (plen == 0) { /* ignore the utf8ness if the pattern is 0 length */
4560 RExC_utf8 = RExC_orig_utf8 = 0;
4563 RExC_utf8 = RExC_orig_utf8 = SvUTF8(pattern);
4565 RExC_uni_semantics = 0;
4566 RExC_contains_locale = 0;
4568 /****************** LONG JUMP TARGET HERE***********************/
4569 /* Longjmp back to here if have to switch in midstream to utf8 */
4570 if (! RExC_orig_utf8) {
4571 JMPENV_PUSH(jump_ret);
4572 used_setjump = TRUE;
4575 if (jump_ret == 0) { /* First time through */
4579 SV *dsv= sv_newmortal();
4580 RE_PV_QUOTED_DECL(s, RExC_utf8,
4581 dsv, exp, plen, 60);
4582 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
4583 PL_colors[4],PL_colors[5],s);
4586 else { /* longjumped back */
4589 /* If the cause for the longjmp was other than changing to utf8, pop
4590 * our own setjmp, and longjmp to the correct handler */
4591 if (jump_ret != UTF8_LONGJMP) {
4593 JMPENV_JUMP(jump_ret);
4598 /* It's possible to write a regexp in ascii that represents Unicode
4599 codepoints outside of the byte range, such as via \x{100}. If we
4600 detect such a sequence we have to convert the entire pattern to utf8
4601 and then recompile, as our sizing calculation will have been based
4602 on 1 byte == 1 character, but we will need to use utf8 to encode
4603 at least some part of the pattern, and therefore must convert the whole
4606 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
4607 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
4608 exp = (char*)Perl_bytes_to_utf8(aTHX_
4609 (U8*)SvPV_nomg(pattern, plen),
4612 RExC_orig_utf8 = RExC_utf8 = 1;
4616 #ifdef TRIE_STUDY_OPT
4620 pm_flags = orig_pm_flags;
4622 if (initial_charset == REGEX_LOCALE_CHARSET) {
4623 RExC_contains_locale = 1;
4625 else if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
4627 /* Set to use unicode semantics if the pattern is in utf8 and has the
4628 * 'depends' charset specified, as it means unicode when utf8 */
4629 set_regex_charset(&pm_flags, REGEX_UNICODE_CHARSET);
4633 RExC_flags = pm_flags;
4637 RExC_in_lookbehind = 0;
4638 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
4639 RExC_seen_evals = 0;
4641 RExC_override_recoding = 0;
4643 /* First pass: determine size, legality. */
4651 RExC_emit = &PL_regdummy;
4652 RExC_whilem_seen = 0;
4653 RExC_open_parens = NULL;
4654 RExC_close_parens = NULL;
4656 RExC_paren_names = NULL;
4658 RExC_paren_name_list = NULL;
4660 RExC_recurse = NULL;
4661 RExC_recurse_count = 0;
4663 #if 0 /* REGC() is (currently) a NOP at the first pass.
4664 * Clever compilers notice this and complain. --jhi */
4665 REGC((U8)REG_MAGIC, (char*)RExC_emit);
4667 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n"));
4668 if (reg(pRExC_state, 0, &flags,1) == NULL) {
4669 RExC_precomp = NULL;
4673 /* Here, finished first pass. Get rid of any added setjmp */
4679 PerlIO_printf(Perl_debug_log,
4680 "Required size %"IVdf" nodes\n"
4681 "Starting second pass (creation)\n",
4684 RExC_lastparse=NULL;
4687 /* The first pass could have found things that force Unicode semantics */
4688 if ((RExC_utf8 || RExC_uni_semantics)
4689 && get_regex_charset(pm_flags) == REGEX_DEPENDS_CHARSET)
4691 set_regex_charset(&pm_flags, REGEX_UNICODE_CHARSET);
4694 /* Small enough for pointer-storage convention?
4695 If extralen==0, this means that we will not need long jumps. */
4696 if (RExC_size >= 0x10000L && RExC_extralen)
4697 RExC_size += RExC_extralen;
4700 if (RExC_whilem_seen > 15)
4701 RExC_whilem_seen = 15;
4703 /* Allocate space and zero-initialize. Note, the two step process
4704 of zeroing when in debug mode, thus anything assigned has to
4705 happen after that */
4706 rx = (REGEXP*) newSV_type(SVt_REGEXP);
4707 r = (struct regexp*)SvANY(rx);
4708 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
4709 char, regexp_internal);
4710 if ( r == NULL || ri == NULL )
4711 FAIL("Regexp out of space");
4713 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
4714 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode), char);
4716 /* bulk initialize base fields with 0. */
4717 Zero(ri, sizeof(regexp_internal), char);
4720 /* non-zero initialization begins here */
4722 r->engine= RE_ENGINE_PTR;
4723 r->extflags = pm_flags;
4725 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
4726 bool has_charset = (get_regex_charset(r->extflags) != REGEX_DEPENDS_CHARSET);
4728 /* The caret is output if there are any defaults: if not all the STD
4729 * flags are set, or if no character set specifier is needed */
4731 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
4733 bool has_runon = ((RExC_seen & REG_SEEN_RUN_ON_COMMENT)==REG_SEEN_RUN_ON_COMMENT);
4734 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
4735 >> RXf_PMf_STD_PMMOD_SHIFT);
4736 const char *fptr = STD_PAT_MODS; /*"msix"*/
4738 /* Allocate for the worst case, which is all the std flags are turned
4739 * on. If more precision is desired, we could do a population count of
4740 * the flags set. This could be done with a small lookup table, or by
4741 * shifting, masking and adding, or even, when available, assembly
4742 * language for a machine-language population count.
4743 * We never output a minus, as all those are defaults, so are
4744 * covered by the caret */
4745 const STRLEN wraplen = plen + has_p + has_runon
4746 + has_default /* If needs a caret */
4748 /* If needs a character set specifier */
4749 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
4750 + (sizeof(STD_PAT_MODS) - 1)
4751 + (sizeof("(?:)") - 1);
4753 p = sv_grow(MUTABLE_SV(rx), wraplen + 1); /* +1 for the ending NUL */
4755 SvFLAGS(rx) |= SvUTF8(pattern);
4758 /* If a default, cover it using the caret */
4760 *p++= DEFAULT_PAT_MOD;
4764 const char* const name = get_regex_charset_name(r->extflags, &len);
4765 Copy(name, p, len, char);
4769 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
4772 while((ch = *fptr++)) {
4780 Copy(RExC_precomp, p, plen, char);
4781 assert ((RX_WRAPPED(rx) - p) < 16);
4782 r->pre_prefix = p - RX_WRAPPED(rx);
4788 SvCUR_set(rx, p - SvPVX_const(rx));
4792 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
4794 if (RExC_seen & REG_SEEN_RECURSE) {
4795 Newxz(RExC_open_parens, RExC_npar,regnode *);
4796 SAVEFREEPV(RExC_open_parens);
4797 Newxz(RExC_close_parens,RExC_npar,regnode *);
4798 SAVEFREEPV(RExC_close_parens);
4801 /* Useful during FAIL. */
4802 #ifdef RE_TRACK_PATTERN_OFFSETS
4803 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
4804 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
4805 "%s %"UVuf" bytes for offset annotations.\n",
4806 ri->u.offsets ? "Got" : "Couldn't get",
4807 (UV)((2*RExC_size+1) * sizeof(U32))));
4809 SetProgLen(ri,RExC_size);
4813 REH_CALL_COMP_BEGIN_HOOK(pRExC_state->rx);
4815 /* Second pass: emit code. */
4816 RExC_flags = pm_flags; /* don't let top level (?i) bleed */
4821 RExC_emit_start = ri->program;
4822 RExC_emit = ri->program;
4823 RExC_emit_bound = ri->program + RExC_size + 1;
4825 /* Store the count of eval-groups for security checks: */
4826 RExC_rx->seen_evals = RExC_seen_evals;
4827 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
4828 if (reg(pRExC_state, 0, &flags,1) == NULL) {
4832 /* XXXX To minimize changes to RE engine we always allocate
4833 3-units-long substrs field. */
4834 Newx(r->substrs, 1, struct reg_substr_data);
4835 if (RExC_recurse_count) {
4836 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
4837 SAVEFREEPV(RExC_recurse);
4841 r->minlen = minlen = sawlookahead = sawplus = sawopen = 0;
4842 Zero(r->substrs, 1, struct reg_substr_data);
4844 #ifdef TRIE_STUDY_OPT
4846 StructCopy(&zero_scan_data, &data, scan_data_t);
4847 copyRExC_state = RExC_state;
4850 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
4852 RExC_state = copyRExC_state;
4853 if (seen & REG_TOP_LEVEL_BRANCHES)
4854 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
4856 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES;
4857 if (data.last_found) {
4858 SvREFCNT_dec(data.longest_fixed);
4859 SvREFCNT_dec(data.longest_float);
4860 SvREFCNT_dec(data.last_found);
4862 StructCopy(&zero_scan_data, &data, scan_data_t);
4865 StructCopy(&zero_scan_data, &data, scan_data_t);
4868 /* Dig out information for optimizations. */
4869 r->extflags = RExC_flags; /* was pm_op */
4870 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
4873 SvUTF8_on(rx); /* Unicode in it? */
4874 ri->regstclass = NULL;
4875 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
4876 r->intflags |= PREGf_NAUGHTY;
4877 scan = ri->program + 1; /* First BRANCH. */
4879 /* testing for BRANCH here tells us whether there is "must appear"
4880 data in the pattern. If there is then we can use it for optimisations */
4881 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES)) { /* Only one top-level choice. */
4883 STRLEN longest_float_length, longest_fixed_length;
4884 struct regnode_charclass_class ch_class; /* pointed to by data */
4886 I32 last_close = 0; /* pointed to by data */
4887 regnode *first= scan;
4888 regnode *first_next= regnext(first);
4890 * Skip introductions and multiplicators >= 1
4891 * so that we can extract the 'meat' of the pattern that must
4892 * match in the large if() sequence following.
4893 * NOTE that EXACT is NOT covered here, as it is normally
4894 * picked up by the optimiser separately.
4896 * This is unfortunate as the optimiser isnt handling lookahead
4897 * properly currently.
4900 while ((OP(first) == OPEN && (sawopen = 1)) ||
4901 /* An OR of *one* alternative - should not happen now. */
4902 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
4903 /* for now we can't handle lookbehind IFMATCH*/
4904 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
4905 (OP(first) == PLUS) ||
4906 (OP(first) == MINMOD) ||
4907 /* An {n,m} with n>0 */
4908 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
4909 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
4912 * the only op that could be a regnode is PLUS, all the rest
4913 * will be regnode_1 or regnode_2.
4916 if (OP(first) == PLUS)
4919 first += regarglen[OP(first)];
4921 first = NEXTOPER(first);
4922 first_next= regnext(first);
4925 /* Starting-point info. */
4927 DEBUG_PEEP("first:",first,0);
4928 /* Ignore EXACT as we deal with it later. */
4929 if (PL_regkind[OP(first)] == EXACT) {
4930 if (OP(first) == EXACT)
4931 NOOP; /* Empty, get anchored substr later. */
4933 ri->regstclass = first;
4936 else if (PL_regkind[OP(first)] == TRIE &&
4937 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
4940 /* this can happen only on restudy */
4941 if ( OP(first) == TRIE ) {
4942 struct regnode_1 *trieop = (struct regnode_1 *)
4943 PerlMemShared_calloc(1, sizeof(struct regnode_1));
4944 StructCopy(first,trieop,struct regnode_1);
4945 trie_op=(regnode *)trieop;
4947 struct regnode_charclass *trieop = (struct regnode_charclass *)
4948 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
4949 StructCopy(first,trieop,struct regnode_charclass);
4950 trie_op=(regnode *)trieop;
4953 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
4954 ri->regstclass = trie_op;
4957 else if (REGNODE_SIMPLE(OP(first)))
4958 ri->regstclass = first;
4959 else if (PL_regkind[OP(first)] == BOUND ||
4960 PL_regkind[OP(first)] == NBOUND)
4961 ri->regstclass = first;
4962 else if (PL_regkind[OP(first)] == BOL) {
4963 r->extflags |= (OP(first) == MBOL
4965 : (OP(first) == SBOL
4968 first = NEXTOPER(first);
4971 else if (OP(first) == GPOS) {
4972 r->extflags |= RXf_ANCH_GPOS;
4973 first = NEXTOPER(first);
4976 else if ((!sawopen || !RExC_sawback) &&
4977 (OP(first) == STAR &&
4978 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
4979 !(r->extflags & RXf_ANCH) && !(RExC_seen & REG_SEEN_EVAL))
4981 /* turn .* into ^.* with an implied $*=1 */
4983 (OP(NEXTOPER(first)) == REG_ANY)
4986 r->extflags |= type;
4987 r->intflags |= PREGf_IMPLICIT;
4988 first = NEXTOPER(first);
4991 if (sawplus && !sawlookahead && (!sawopen || !RExC_sawback)
4992 && !(RExC_seen & REG_SEEN_EVAL)) /* May examine pos and $& */
4993 /* x+ must match at the 1st pos of run of x's */
4994 r->intflags |= PREGf_SKIP;
4996 /* Scan is after the zeroth branch, first is atomic matcher. */
4997 #ifdef TRIE_STUDY_OPT
5000 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
5001 (IV)(first - scan + 1))
5005 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
5006 (IV)(first - scan + 1))
5012 * If there's something expensive in the r.e., find the
5013 * longest literal string that must appear and make it the
5014 * regmust. Resolve ties in favor of later strings, since
5015 * the regstart check works with the beginning of the r.e.
5016 * and avoiding duplication strengthens checking. Not a
5017 * strong reason, but sufficient in the absence of others.
5018 * [Now we resolve ties in favor of the earlier string if
5019 * it happens that c_offset_min has been invalidated, since the
5020 * earlier string may buy us something the later one won't.]
5023 data.longest_fixed = newSVpvs("");
5024 data.longest_float = newSVpvs("");
5025 data.last_found = newSVpvs("");
5026 data.longest = &(data.longest_fixed);
5028 if (!ri->regstclass) {
5029 cl_init(pRExC_state, &ch_class);
5030 data.start_class = &ch_class;
5031 stclass_flag = SCF_DO_STCLASS_AND;
5032 } else /* XXXX Check for BOUND? */
5034 data.last_closep = &last_close;
5036 minlen = study_chunk(pRExC_state, &first, &minlen, &fake, scan + RExC_size, /* Up to end */
5037 &data, -1, NULL, NULL,
5038 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag,0);
5044 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
5045 && data.last_start_min == 0 && data.last_end > 0
5046 && !RExC_seen_zerolen
5047 && !(RExC_seen & REG_SEEN_VERBARG)
5048 && (!(RExC_seen & REG_SEEN_GPOS) || (r->extflags & RXf_ANCH_GPOS)))
5049 r->extflags |= RXf_CHECK_ALL;
5050 scan_commit(pRExC_state, &data,&minlen,0);
5051 SvREFCNT_dec(data.last_found);
5053 /* Note that code very similar to this but for anchored string
5054 follows immediately below, changes may need to be made to both.
5057 longest_float_length = CHR_SVLEN(data.longest_float);
5058 if (longest_float_length
5059 || (data.flags & SF_FL_BEFORE_EOL
5060 && (!(data.flags & SF_FL_BEFORE_MEOL)
5061 || (RExC_flags & RXf_PMf_MULTILINE))))
5065 if (SvCUR(data.longest_fixed) /* ok to leave SvCUR */
5066 && data.offset_fixed == data.offset_float_min
5067 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float))
5068 goto remove_float; /* As in (a)+. */
5070 /* copy the information about the longest float from the reg_scan_data
5071 over to the program. */
5072 if (SvUTF8(data.longest_float)) {
5073 r->float_utf8 = data.longest_float;
5074 r->float_substr = NULL;
5076 r->float_substr = data.longest_float;
5077 r->float_utf8 = NULL;
5079 /* float_end_shift is how many chars that must be matched that
5080 follow this item. We calculate it ahead of time as once the
5081 lookbehind offset is added in we lose the ability to correctly
5083 ml = data.minlen_float ? *(data.minlen_float)
5084 : (I32)longest_float_length;
5085 r->float_end_shift = ml - data.offset_float_min
5086 - longest_float_length + (SvTAIL(data.longest_float) != 0)
5087 + data.lookbehind_float;
5088 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
5089 r->float_max_offset = data.offset_float_max;
5090 if (data.offset_float_max < I32_MAX) /* Don't offset infinity */
5091 r->float_max_offset -= data.lookbehind_float;
5093 t = (data.flags & SF_FL_BEFORE_EOL /* Can't have SEOL and MULTI */
5094 && (!(data.flags & SF_FL_BEFORE_MEOL)
5095 || (RExC_flags & RXf_PMf_MULTILINE)));
5096 fbm_compile(data.longest_float, t ? FBMcf_TAIL : 0);
5100 r->float_substr = r->float_utf8 = NULL;
5101 SvREFCNT_dec(data.longest_float);
5102 longest_float_length = 0;
5105 /* Note that code very similar to this but for floating string
5106 is immediately above, changes may need to be made to both.
5109 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
5110 if (longest_fixed_length
5111 || (data.flags & SF_FIX_BEFORE_EOL /* Cannot have SEOL and MULTI */
5112 && (!(data.flags & SF_FIX_BEFORE_MEOL)
5113 || (RExC_flags & RXf_PMf_MULTILINE))))
5117 /* copy the information about the longest fixed
5118 from the reg_scan_data over to the program. */
5119 if (SvUTF8(data.longest_fixed)) {
5120 r->anchored_utf8 = data.longest_fixed;
5121 r->anchored_substr = NULL;
5123 r->anchored_substr = data.longest_fixed;
5124 r->anchored_utf8 = NULL;
5126 /* fixed_end_shift is how many chars that must be matched that
5127 follow this item. We calculate it ahead of time as once the
5128 lookbehind offset is added in we lose the ability to correctly
5130 ml = data.minlen_fixed ? *(data.minlen_fixed)
5131 : (I32)longest_fixed_length;
5132 r->anchored_end_shift = ml - data.offset_fixed
5133 - longest_fixed_length + (SvTAIL(data.longest_fixed) != 0)
5134 + data.lookbehind_fixed;
5135 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
5137 t = (data.flags & SF_FIX_BEFORE_EOL /* Can't have SEOL and MULTI */
5138 && (!(data.flags & SF_FIX_BEFORE_MEOL)
5139 || (RExC_flags & RXf_PMf_MULTILINE)));
5140 fbm_compile(data.longest_fixed, t ? FBMcf_TAIL : 0);
5143 r->anchored_substr = r->anchored_utf8 = NULL;
5144 SvREFCNT_dec(data.longest_fixed);
5145 longest_fixed_length = 0;
5148 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
5149 ri->regstclass = NULL;
5151 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
5153 && !(data.start_class->flags & ANYOF_EOS)
5154 && !cl_is_anything(data.start_class))
5156 const U32 n = add_data(pRExC_state, 1, "f");
5157 data.start_class->flags |= ANYOF_IS_SYNTHETIC;
5159 Newx(RExC_rxi->data->data[n], 1,
5160 struct regnode_charclass_class);
5161 StructCopy(data.start_class,
5162 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
5163 struct regnode_charclass_class);
5164 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
5165 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
5166 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
5167 regprop(r, sv, (regnode*)data.start_class);
5168 PerlIO_printf(Perl_debug_log,
5169 "synthetic stclass \"%s\".\n",
5170 SvPVX_const(sv));});
5173 /* A temporary algorithm prefers floated substr to fixed one to dig more info. */
5174 if (longest_fixed_length > longest_float_length) {
5175 r->check_end_shift = r->anchored_end_shift;
5176 r->check_substr = r->anchored_substr;
5177 r->check_utf8 = r->anchored_utf8;
5178 r->check_offset_min = r->check_offset_max = r->anchored_offset;
5179 if (r->extflags & RXf_ANCH_SINGLE)
5180 r->extflags |= RXf_NOSCAN;
5183 r->check_end_shift = r->float_end_shift;
5184 r->check_substr = r->float_substr;
5185 r->check_utf8 = r->float_utf8;
5186 r->check_offset_min = r->float_min_offset;
5187 r->check_offset_max = r->float_max_offset;
5189 /* XXXX Currently intuiting is not compatible with ANCH_GPOS.
5190 This should be changed ASAP! */
5191 if ((r->check_substr || r->check_utf8) && !(r->extflags & RXf_ANCH_GPOS)) {
5192 r->extflags |= RXf_USE_INTUIT;
5193 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
5194 r->extflags |= RXf_INTUIT_TAIL;
5196 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
5197 if ( (STRLEN)minlen < longest_float_length )
5198 minlen= longest_float_length;
5199 if ( (STRLEN)minlen < longest_fixed_length )
5200 minlen= longest_fixed_length;
5204 /* Several toplevels. Best we can is to set minlen. */
5206 struct regnode_charclass_class ch_class;
5209 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
5211 scan = ri->program + 1;
5212 cl_init(pRExC_state, &ch_class);
5213 data.start_class = &ch_class;
5214 data.last_closep = &last_close;
5217 minlen = study_chunk(pRExC_state, &scan, &minlen, &fake, scan + RExC_size,
5218 &data, -1, NULL, NULL, SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS,0);
5222 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
5223 = r->float_substr = r->float_utf8 = NULL;
5225 if (!(data.start_class->flags & ANYOF_EOS)
5226 && !cl_is_anything(data.start_class))
5228 const U32 n = add_data(pRExC_state, 1, "f");
5229 data.start_class->flags |= ANYOF_IS_SYNTHETIC;
5231 Newx(RExC_rxi->data->data[n], 1,
5232 struct regnode_charclass_class);
5233 StructCopy(data.start_class,
5234 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
5235 struct regnode_charclass_class);
5236 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
5237 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
5238 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
5239 regprop(r, sv, (regnode*)data.start_class);
5240 PerlIO_printf(Perl_debug_log,
5241 "synthetic stclass \"%s\".\n",
5242 SvPVX_const(sv));});
5246 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
5247 the "real" pattern. */
5249 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf"\n",
5250 (IV)minlen, (IV)r->minlen);
5252 r->minlenret = minlen;
5253 if (r->minlen < minlen)
5256 if (RExC_seen & REG_SEEN_GPOS)
5257 r->extflags |= RXf_GPOS_SEEN;
5258 if (RExC_seen & REG_SEEN_LOOKBEHIND)
5259 r->extflags |= RXf_LOOKBEHIND_SEEN;
5260 if (RExC_seen & REG_SEEN_EVAL)
5261 r->extflags |= RXf_EVAL_SEEN;
5262 if (RExC_seen & REG_SEEN_CANY)
5263 r->extflags |= RXf_CANY_SEEN;
5264 if (RExC_seen & REG_SEEN_VERBARG)
5265 r->intflags |= PREGf_VERBARG_SEEN;
5266 if (RExC_seen & REG_SEEN_CUTGROUP)
5267 r->intflags |= PREGf_CUTGROUP_SEEN;
5268 if (RExC_paren_names)
5269 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
5271 RXp_PAREN_NAMES(r) = NULL;
5273 #ifdef STUPID_PATTERN_CHECKS
5274 if (RX_PRELEN(rx) == 0)
5275 r->extflags |= RXf_NULL;
5276 if (r->extflags & RXf_SPLIT && RX_PRELEN(rx) == 1 && RX_PRECOMP(rx)[0] == ' ')
5277 /* XXX: this should happen BEFORE we compile */
5278 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
5279 else if (RX_PRELEN(rx) == 3 && memEQ("\\s+", RX_PRECOMP(rx), 3))
5280 r->extflags |= RXf_WHITE;
5281 else if (RX_PRELEN(rx) == 1 && RXp_PRECOMP(rx)[0] == '^')
5282 r->extflags |= RXf_START_ONLY;
5284 if (r->extflags & RXf_SPLIT && RX_PRELEN(rx) == 1 && RX_PRECOMP(rx)[0] == ' ')
5285 /* XXX: this should happen BEFORE we compile */
5286 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
5288 regnode *first = ri->program + 1;
5291 if (PL_regkind[fop] == NOTHING && OP(NEXTOPER(first)) == END)
5292 r->extflags |= RXf_NULL;
5293 else if (PL_regkind[fop] == BOL && OP(NEXTOPER(first)) == END)
5294 r->extflags |= RXf_START_ONLY;
5295 else if (fop == PLUS && OP(NEXTOPER(first)) == SPACE
5296 && OP(regnext(first)) == END)
5297 r->extflags |= RXf_WHITE;
5301 if (RExC_paren_names) {
5302 ri->name_list_idx = add_data( pRExC_state, 1, "a" );
5303 ri->data->data[ri->name_list_idx] = (void*)SvREFCNT_inc(RExC_paren_name_list);
5306 ri->name_list_idx = 0;
5308 if (RExC_recurse_count) {
5309 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
5310 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
5311 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
5314 Newxz(r->offs, RExC_npar, regexp_paren_pair);
5315 /* assume we don't need to swap parens around before we match */
5318 PerlIO_printf(Perl_debug_log,"Final program:\n");
5321 #ifdef RE_TRACK_PATTERN_OFFSETS
5322 DEBUG_OFFSETS_r(if (ri->u.offsets) {
5323 const U32 len = ri->u.offsets[0];
5325 GET_RE_DEBUG_FLAGS_DECL;
5326 PerlIO_printf(Perl_debug_log, "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
5327 for (i = 1; i <= len; i++) {
5328 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
5329 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
5330 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
5332 PerlIO_printf(Perl_debug_log, "\n");
5338 #undef RE_ENGINE_PTR
5342 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
5345 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
5347 PERL_UNUSED_ARG(value);
5349 if (flags & RXapif_FETCH) {
5350 return reg_named_buff_fetch(rx, key, flags);
5351 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
5352 Perl_croak_no_modify(aTHX);
5354 } else if (flags & RXapif_EXISTS) {
5355 return reg_named_buff_exists(rx, key, flags)
5358 } else if (flags & RXapif_REGNAMES) {
5359 return reg_named_buff_all(rx, flags);
5360 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
5361 return reg_named_buff_scalar(rx, flags);
5363 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
5369 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
5372 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
5373 PERL_UNUSED_ARG(lastkey);
5375 if (flags & RXapif_FIRSTKEY)
5376 return reg_named_buff_firstkey(rx, flags);
5377 else if (flags & RXapif_NEXTKEY)
5378 return reg_named_buff_nextkey(rx, flags);
5380 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter", (int)flags);
5386 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
5389 AV *retarray = NULL;
5391 struct regexp *const rx = (struct regexp *)SvANY(r);
5393 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
5395 if (flags & RXapif_ALL)
5398 if (rx && RXp_PAREN_NAMES(rx)) {
5399 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
5402 SV* sv_dat=HeVAL(he_str);
5403 I32 *nums=(I32*)SvPVX(sv_dat);
5404 for ( i=0; i<SvIVX(sv_dat); i++ ) {
5405 if ((I32)(rx->nparens) >= nums[i]
5406 && rx->offs[nums[i]].start != -1
5407 && rx->offs[nums[i]].end != -1)
5410 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
5415 ret = newSVsv(&PL_sv_undef);
5418 av_push(retarray, ret);
5421 return newRV_noinc(MUTABLE_SV(retarray));
5428 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
5431 struct regexp *const rx = (struct regexp *)SvANY(r);
5433 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
5435 if (rx && RXp_PAREN_NAMES(rx)) {
5436 if (flags & RXapif_ALL) {
5437 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
5439 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
5453 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
5455 struct regexp *const rx = (struct regexp *)SvANY(r);
5457 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
5459 if ( rx && RXp_PAREN_NAMES(rx) ) {
5460 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
5462 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
5469 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
5471 struct regexp *const rx = (struct regexp *)SvANY(r);
5472 GET_RE_DEBUG_FLAGS_DECL;
5474 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
5476 if (rx && RXp_PAREN_NAMES(rx)) {
5477 HV *hv = RXp_PAREN_NAMES(rx);
5479 while ( (temphe = hv_iternext_flags(hv,0)) ) {
5482 SV* sv_dat = HeVAL(temphe);
5483 I32 *nums = (I32*)SvPVX(sv_dat);
5484 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
5485 if ((I32)(rx->lastparen) >= nums[i] &&
5486 rx->offs[nums[i]].start != -1 &&
5487 rx->offs[nums[i]].end != -1)
5493 if (parno || flags & RXapif_ALL) {
5494 return newSVhek(HeKEY_hek(temphe));
5502 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
5507 struct regexp *const rx = (struct regexp *)SvANY(r);
5509 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
5511 if (rx && RXp_PAREN_NAMES(rx)) {
5512 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
5513 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
5514 } else if (flags & RXapif_ONE) {
5515 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
5516 av = MUTABLE_AV(SvRV(ret));
5517 length = av_len(av);
5519 return newSViv(length + 1);
5521 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar", (int)flags);
5525 return &PL_sv_undef;
5529 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
5531 struct regexp *const rx = (struct regexp *)SvANY(r);
5534 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
5536 if (rx && RXp_PAREN_NAMES(rx)) {
5537 HV *hv= RXp_PAREN_NAMES(rx);
5539 (void)hv_iterinit(hv);
5540 while ( (temphe = hv_iternext_flags(hv,0)) ) {
5543 SV* sv_dat = HeVAL(temphe);
5544 I32 *nums = (I32*)SvPVX(sv_dat);
5545 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
5546 if ((I32)(rx->lastparen) >= nums[i] &&
5547 rx->offs[nums[i]].start != -1 &&
5548 rx->offs[nums[i]].end != -1)
5554 if (parno || flags & RXapif_ALL) {
5555 av_push(av, newSVhek(HeKEY_hek(temphe)));
5560 return newRV_noinc(MUTABLE_SV(av));
5564 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
5567 struct regexp *const rx = (struct regexp *)SvANY(r);
5572 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
5575 sv_setsv(sv,&PL_sv_undef);
5579 if (paren == RX_BUFF_IDX_PREMATCH && rx->offs[0].start != -1) {
5581 i = rx->offs[0].start;
5585 if (paren == RX_BUFF_IDX_POSTMATCH && rx->offs[0].end != -1) {
5587 s = rx->subbeg + rx->offs[0].end;
5588 i = rx->sublen - rx->offs[0].end;
5591 if ( 0 <= paren && paren <= (I32)rx->nparens &&
5592 (s1 = rx->offs[paren].start) != -1 &&
5593 (t1 = rx->offs[paren].end) != -1)
5597 s = rx->subbeg + s1;
5599 sv_setsv(sv,&PL_sv_undef);
5602 assert(rx->sublen >= (s - rx->subbeg) + i );
5604 const int oldtainted = PL_tainted;
5606 sv_setpvn(sv, s, i);
5607 PL_tainted = oldtainted;
5608 if ( (rx->extflags & RXf_CANY_SEEN)
5609 ? (RXp_MATCH_UTF8(rx)
5610 && (!i || is_utf8_string((U8*)s, i)))
5611 : (RXp_MATCH_UTF8(rx)) )
5618 if (RXp_MATCH_TAINTED(rx)) {
5619 if (SvTYPE(sv) >= SVt_PVMG) {
5620 MAGIC* const mg = SvMAGIC(sv);
5623 SvMAGIC_set(sv, mg->mg_moremagic);
5625 if ((mgt = SvMAGIC(sv))) {
5626 mg->mg_moremagic = mgt;
5627 SvMAGIC_set(sv, mg);
5637 sv_setsv(sv,&PL_sv_undef);
5643 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
5644 SV const * const value)
5646 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
5648 PERL_UNUSED_ARG(rx);
5649 PERL_UNUSED_ARG(paren);
5650 PERL_UNUSED_ARG(value);
5653 Perl_croak_no_modify(aTHX);
5657 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
5660 struct regexp *const rx = (struct regexp *)SvANY(r);
5664 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
5666 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
5668 /* $` / ${^PREMATCH} */
5669 case RX_BUFF_IDX_PREMATCH:
5670 if (rx->offs[0].start != -1) {
5671 i = rx->offs[0].start;
5679 /* $' / ${^POSTMATCH} */
5680 case RX_BUFF_IDX_POSTMATCH:
5681 if (rx->offs[0].end != -1) {
5682 i = rx->sublen - rx->offs[0].end;
5684 s1 = rx->offs[0].end;
5690 /* $& / ${^MATCH}, $1, $2, ... */
5692 if (paren <= (I32)rx->nparens &&
5693 (s1 = rx->offs[paren].start) != -1 &&
5694 (t1 = rx->offs[paren].end) != -1)
5699 if (ckWARN(WARN_UNINITIALIZED))
5700 report_uninit((const SV *)sv);
5705 if (i > 0 && RXp_MATCH_UTF8(rx)) {
5706 const char * const s = rx->subbeg + s1;
5711 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
5718 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
5720 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
5721 PERL_UNUSED_ARG(rx);
5725 return newSVpvs("Regexp");
5728 /* Scans the name of a named buffer from the pattern.
5729 * If flags is REG_RSN_RETURN_NULL returns null.
5730 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
5731 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
5732 * to the parsed name as looked up in the RExC_paren_names hash.
5733 * If there is an error throws a vFAIL().. type exception.
5736 #define REG_RSN_RETURN_NULL 0
5737 #define REG_RSN_RETURN_NAME 1
5738 #define REG_RSN_RETURN_DATA 2
5741 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
5743 char *name_start = RExC_parse;
5745 PERL_ARGS_ASSERT_REG_SCAN_NAME;
5747 if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
5748 /* skip IDFIRST by using do...while */
5751 RExC_parse += UTF8SKIP(RExC_parse);
5752 } while (isALNUM_utf8((U8*)RExC_parse));
5756 } while (isALNUM(*RExC_parse));
5761 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
5762 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
5763 if ( flags == REG_RSN_RETURN_NAME)
5765 else if (flags==REG_RSN_RETURN_DATA) {
5768 if ( ! sv_name ) /* should not happen*/
5769 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
5770 if (RExC_paren_names)
5771 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
5773 sv_dat = HeVAL(he_str);
5775 vFAIL("Reference to nonexistent named group");
5779 Perl_croak(aTHX_ "panic: bad flag in reg_scan_name");
5786 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
5787 int rem=(int)(RExC_end - RExC_parse); \
5796 if (RExC_lastparse!=RExC_parse) \
5797 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
5800 iscut ? "..." : "<" \
5803 PerlIO_printf(Perl_debug_log,"%16s",""); \
5806 num = RExC_size + 1; \
5808 num=REG_NODE_NUM(RExC_emit); \
5809 if (RExC_lastnum!=num) \
5810 PerlIO_printf(Perl_debug_log,"|%4d",num); \
5812 PerlIO_printf(Perl_debug_log,"|%4s",""); \
5813 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
5814 (int)((depth*2)), "", \
5818 RExC_lastparse=RExC_parse; \
5823 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
5824 DEBUG_PARSE_MSG((funcname)); \
5825 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
5827 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
5828 DEBUG_PARSE_MSG((funcname)); \
5829 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
5832 /* This section of code defines the inversion list object and its methods. The
5833 * interfaces are highly subject to change, so as much as possible is static to
5834 * this file. An inversion list is here implemented as a malloc'd C UV array
5835 * with some added info that is placed as UVs at the beginning in a header
5836 * portion. An inversion list for Unicode is an array of code points, sorted
5837 * by ordinal number. The zeroth element is the first code point in the list.
5838 * The 1th element is the first element beyond that not in the list. In other
5839 * words, the first range is
5840 * invlist[0]..(invlist[1]-1)
5841 * The other ranges follow. Thus every element that is divisible by two marks
5842 * the beginning of a range that is in the list, and every element not
5843 * divisible by two marks the beginning of a range not in the list. A single
5844 * element inversion list that contains the single code point N generally
5845 * consists of two elements
5848 * (The exception is when N is the highest representable value on the
5849 * machine, in which case the list containing just it would be a single
5850 * element, itself. By extension, if the last range in the list extends to
5851 * infinity, then the first element of that range will be in the inversion list
5852 * at a position that is divisible by two, and is the final element in the
5854 * Taking the complement (inverting) an inversion list is quite simple, if the
5855 * first element is 0, remove it; otherwise add a 0 element at the beginning.
5856 * This implementation reserves an element at the beginning of each inversion list
5857 * to contain 0 when the list contains 0, and contains 1 otherwise. The actual
5858 * beginning of the list is either that element if 0, or the next one if 1.
5860 * More about inversion lists can be found in "Unicode Demystified"
5861 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
5862 * More will be coming when functionality is added later.
5864 * The inversion list data structure is currently implemented as an SV pointing
5865 * to an array of UVs that the SV thinks are bytes. This allows us to have an
5866 * array of UV whose memory management is automatically handled by the existing
5867 * facilities for SV's.
5869 * Some of the methods should always be private to the implementation, and some
5870 * should eventually be made public */
5872 #define INVLIST_LEN_OFFSET 0 /* Number of elements in the inversion list */
5873 #define INVLIST_ITER_OFFSET 1 /* Current iteration position */
5875 #define INVLIST_ZERO_OFFSET 2 /* 0 or 1; must be last element in header */
5876 /* The UV at position ZERO contains either 0 or 1. If 0, the inversion list
5877 * contains the code point U+00000, and begins here. If 1, the inversion list
5878 * doesn't contain U+0000, and it begins at the next UV in the array.
5879 * Inverting an inversion list consists of adding or removing the 0 at the
5880 * beginning of it. By reserving a space for that 0, inversion can be made
5883 #define HEADER_LENGTH (INVLIST_ZERO_OFFSET + 1)
5885 /* Internally things are UVs */
5886 #define TO_INTERNAL_SIZE(x) ((x + HEADER_LENGTH) * sizeof(UV))
5887 #define FROM_INTERNAL_SIZE(x) ((x / sizeof(UV)) - HEADER_LENGTH)
5889 #define INVLIST_INITIAL_LEN 10
5891 PERL_STATIC_INLINE UV*
5892 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
5894 /* Returns a pointer to the first element in the inversion list's array.
5895 * This is called upon initialization of an inversion list. Where the
5896 * array begins depends on whether the list has the code point U+0000
5897 * in it or not. The other parameter tells it whether the code that
5898 * follows this call is about to put a 0 in the inversion list or not.
5899 * The first element is either the element with 0, if 0, or the next one,
5902 UV* zero = get_invlist_zero_addr(invlist);
5904 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
5907 assert(! *get_invlist_len_addr(invlist));
5909 /* 1^1 = 0; 1^0 = 1 */
5910 *zero = 1 ^ will_have_0;
5911 return zero + *zero;
5914 PERL_STATIC_INLINE UV*
5915 S_invlist_array(pTHX_ SV* const invlist)
5917 /* Returns the pointer to the inversion list's array. Every time the
5918 * length changes, this needs to be called in case malloc or realloc moved
5921 PERL_ARGS_ASSERT_INVLIST_ARRAY;
5923 /* Must not be empty */
5924 assert(*get_invlist_len_addr(invlist));
5925 assert(*get_invlist_zero_addr(invlist) == 0
5926 || *get_invlist_zero_addr(invlist) == 1);
5928 /* The array begins either at the element reserved for zero if the
5929 * list contains 0 (that element will be set to 0), or otherwise the next
5930 * element (in which case the reserved element will be set to 1). */
5931 return (UV *) (get_invlist_zero_addr(invlist)
5932 + *get_invlist_zero_addr(invlist));
5935 PERL_STATIC_INLINE UV*
5936 S_get_invlist_len_addr(pTHX_ SV* invlist)
5938 /* Return the address of the UV that contains the current number
5939 * of used elements in the inversion list */
5941 PERL_ARGS_ASSERT_GET_INVLIST_LEN_ADDR;
5943 return (UV *) (SvPVX(invlist) + (INVLIST_LEN_OFFSET * sizeof (UV)));
5946 PERL_STATIC_INLINE UV
5947 S_invlist_len(pTHX_ SV* const invlist)
5949 /* Returns the current number of elements in the inversion list's array */
5951 PERL_ARGS_ASSERT_INVLIST_LEN;
5953 return *get_invlist_len_addr(invlist);
5956 PERL_STATIC_INLINE void
5957 S_invlist_set_len(pTHX_ SV* const invlist, const UV len)
5959 /* Sets the current number of elements stored in the inversion list */
5961 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
5963 *get_invlist_len_addr(invlist) = len;
5965 assert(len <= SvLEN(invlist));
5967 SvCUR_set(invlist, TO_INTERNAL_SIZE(len));
5968 /* If the list contains U+0000, that element is part of the header,
5969 * and should not be counted as part of the array. It will contain
5970 * 0 in that case, and 1 otherwise. So we could flop 0=>1, 1=>0 and
5972 * SvCUR_set(invlist,
5973 * TO_INTERNAL_SIZE(len
5974 * - (*get_invlist_zero_addr(inv_list) ^ 1)));
5975 * But, this is only valid if len is not 0. The consequences of not doing
5976 * this is that the memory allocation code may think that 1 more UV is
5977 * being used than actually is, and so might do an unnecessary grow. That
5978 * seems worth not bothering to make this the precise amount.
5980 * Note that when inverting, SvCUR shouldn't change */
5983 PERL_STATIC_INLINE UV
5984 S_invlist_max(pTHX_ SV* const invlist)
5986 /* Returns the maximum number of elements storable in the inversion list's
5987 * array, without having to realloc() */
5989 PERL_ARGS_ASSERT_INVLIST_MAX;
5991 return FROM_INTERNAL_SIZE(SvLEN(invlist));
5994 PERL_STATIC_INLINE UV*
5995 S_get_invlist_zero_addr(pTHX_ SV* invlist)
5997 /* Return the address of the UV that is reserved to hold 0 if the inversion
5998 * list contains 0. This has to be the last element of the heading, as the
5999 * list proper starts with either it if 0, or the next element if not.
6000 * (But we force it to contain either 0 or 1) */
6002 PERL_ARGS_ASSERT_GET_INVLIST_ZERO_ADDR;
6004 return (UV *) (SvPVX(invlist) + (INVLIST_ZERO_OFFSET * sizeof (UV)));
6007 #ifndef PERL_IN_XSUB_RE
6009 Perl__new_invlist(pTHX_ IV initial_size)
6012 /* Return a pointer to a newly constructed inversion list, with enough
6013 * space to store 'initial_size' elements. If that number is negative, a
6014 * system default is used instead */
6018 if (initial_size < 0) {
6019 initial_size = INVLIST_INITIAL_LEN;
6022 /* Allocate the initial space */
6023 new_list = newSV(TO_INTERNAL_SIZE(initial_size));
6024 invlist_set_len(new_list, 0);
6026 /* Force iterinit() to be used to get iteration to work */
6027 *get_invlist_iter_addr(new_list) = UV_MAX;
6029 /* This should force a segfault if a method doesn't initialize this
6031 *get_invlist_zero_addr(new_list) = UV_MAX;
6038 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
6040 /* Grow the maximum size of an inversion list */
6042 PERL_ARGS_ASSERT_INVLIST_EXTEND;
6044 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max));
6047 PERL_STATIC_INLINE void
6048 S_invlist_trim(pTHX_ SV* const invlist)
6050 PERL_ARGS_ASSERT_INVLIST_TRIM;
6052 /* Change the length of the inversion list to how many entries it currently
6055 SvPV_shrink_to_cur((SV *) invlist);
6058 /* An element is in an inversion list iff its index is even numbered: 0, 2, 4,
6061 #define ELEMENT_IN_INVLIST_SET(i) (! ((i) & 1))
6062 #define PREV_ELEMENT_IN_INVLIST_SET(i) (! ELEMENT_IN_INVLIST_SET(i))
6064 #ifndef PERL_IN_XSUB_RE
6066 Perl__append_range_to_invlist(pTHX_ SV* const invlist, const UV start, const UV end)
6068 /* Subject to change or removal. Append the range from 'start' to 'end' at
6069 * the end of the inversion list. The range must be above any existing
6073 UV max = invlist_max(invlist);
6074 UV len = invlist_len(invlist);
6076 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
6078 if (len == 0) { /* Empty lists must be initialized */
6079 array = _invlist_array_init(invlist, start == 0);
6082 /* Here, the existing list is non-empty. The current max entry in the
6083 * list is generally the first value not in the set, except when the
6084 * set extends to the end of permissible values, in which case it is
6085 * the first entry in that final set, and so this call is an attempt to
6086 * append out-of-order */
6088 UV final_element = len - 1;
6089 array = invlist_array(invlist);
6090 if (array[final_element] > start
6091 || ELEMENT_IN_INVLIST_SET(final_element))
6093 Perl_croak(aTHX_ "panic: attempting to append to an inversion list, but wasn't at the end of the list");
6096 /* Here, it is a legal append. If the new range begins with the first
6097 * value not in the set, it is extending the set, so the new first
6098 * value not in the set is one greater than the newly extended range.
6100 if (array[final_element] == start) {
6101 if (end != UV_MAX) {
6102 array[final_element] = end + 1;
6105 /* But if the end is the maximum representable on the machine,
6106 * just let the range that this would extend have no end */
6107 invlist_set_len(invlist, len - 1);
6113 /* Here the new range doesn't extend any existing set. Add it */
6115 len += 2; /* Includes an element each for the start and end of range */
6117 /* If overflows the existing space, extend, which may cause the array to be
6120 invlist_extend(invlist, len);
6121 invlist_set_len(invlist, len); /* Have to set len here to avoid assert
6122 failure in invlist_array() */
6123 array = invlist_array(invlist);
6126 invlist_set_len(invlist, len);
6129 /* The next item on the list starts the range, the one after that is
6130 * one past the new range. */
6131 array[len - 2] = start;
6132 if (end != UV_MAX) {
6133 array[len - 1] = end + 1;
6136 /* But if the end is the maximum representable on the machine, just let
6137 * the range have no end */
6138 invlist_set_len(invlist, len - 1);
6143 Perl__invlist_union(pTHX_ SV* const a, SV* const b, SV** output)
6145 /* Take the union of two inversion lists and point 'result' to it. If
6146 * 'result' on input points to one of the two lists, the reference count to
6147 * that list will be decremented.
6148 * The basis for this comes from "Unicode Demystified" Chapter 13 by
6149 * Richard Gillam, published by Addison-Wesley, and explained at some
6150 * length there. The preface says to incorporate its examples into your
6151 * code at your own risk.
6153 * The algorithm is like a merge sort.
6155 * XXX A potential performance improvement is to keep track as we go along
6156 * if only one of the inputs contributes to the result, meaning the other
6157 * is a subset of that one. In that case, we can skip the final copy and
6158 * return the larger of the input lists, but then outside code might need
6159 * to keep track of whether to free the input list or not */
6161 UV* array_a; /* a's array */
6163 UV len_a; /* length of a's array */
6166 SV* u; /* the resulting union */
6170 UV i_a = 0; /* current index into a's array */
6174 /* running count, as explained in the algorithm source book; items are
6175 * stopped accumulating and are output when the count changes to/from 0.
6176 * The count is incremented when we start a range that's in the set, and
6177 * decremented when we start a range that's not in the set. So its range
6178 * is 0 to 2. Only when the count is zero is something not in the set.
6182 PERL_ARGS_ASSERT__INVLIST_UNION;
6184 /* If either one is empty, the union is the other one */
6185 len_a = invlist_len(a);
6190 else if (output != &b) {
6191 *output = invlist_clone(b);
6193 /* else *output already = b; */
6196 else if ((len_b = invlist_len(b)) == 0) {
6200 else if (output != &a) {
6201 *output = invlist_clone(a);
6203 /* else *output already = a; */
6207 /* Here both lists exist and are non-empty */
6208 array_a = invlist_array(a);
6209 array_b = invlist_array(b);
6211 /* Size the union for the worst case: that the sets are completely
6213 u = _new_invlist(len_a + len_b);
6215 /* Will contain U+0000 if either component does */
6216 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
6217 || (len_b > 0 && array_b[0] == 0));
6219 /* Go through each list item by item, stopping when exhausted one of
6221 while (i_a < len_a && i_b < len_b) {
6222 UV cp; /* The element to potentially add to the union's array */
6223 bool cp_in_set; /* is it in the the input list's set or not */
6225 /* We need to take one or the other of the two inputs for the union.
6226 * Since we are merging two sorted lists, we take the smaller of the
6227 * next items. In case of a tie, we take the one that is in its set
6228 * first. If we took one not in the set first, it would decrement the
6229 * count, possibly to 0 which would cause it to be output as ending the
6230 * range, and the next time through we would take the same number, and
6231 * output it again as beginning the next range. By doing it the
6232 * opposite way, there is no possibility that the count will be
6233 * momentarily decremented to 0, and thus the two adjoining ranges will
6234 * be seamlessly merged. (In a tie and both are in the set or both not
6235 * in the set, it doesn't matter which we take first.) */
6236 if (array_a[i_a] < array_b[i_b]
6237 || (array_a[i_a] == array_b[i_b] && ELEMENT_IN_INVLIST_SET(i_a)))
6239 cp_in_set = ELEMENT_IN_INVLIST_SET(i_a);
6243 cp_in_set = ELEMENT_IN_INVLIST_SET(i_b);
6247 /* Here, have chosen which of the two inputs to look at. Only output
6248 * if the running count changes to/from 0, which marks the
6249 * beginning/end of a range in that's in the set */
6252 array_u[i_u++] = cp;
6259 array_u[i_u++] = cp;
6264 /* Here, we are finished going through at least one of the lists, which
6265 * means there is something remaining in at most one. We check if the list
6266 * that hasn't been exhausted is positioned such that we are in the middle
6267 * of a range in its set or not. (i_a and i_b point to the element beyond
6268 * the one we care about.) If in the set, we decrement 'count'; if 0, there
6269 * is potentially more to output.
6270 * There are four cases:
6271 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
6272 * in the union is entirely from the non-exhausted set.
6273 * 2) Both were in their sets, count is 2. Nothing further should
6274 * be output, as everything that remains will be in the exhausted
6275 * list's set, hence in the union; decrementing to 1 but not 0 insures
6277 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
6278 * Nothing further should be output because the union includes
6279 * everything from the exhausted set. Not decrementing ensures that.
6280 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
6281 * decrementing to 0 insures that we look at the remainder of the
6282 * non-exhausted set */
6283 if ((i_a != len_a && PREV_ELEMENT_IN_INVLIST_SET(i_a))
6284 || (i_b != len_b && PREV_ELEMENT_IN_INVLIST_SET(i_b)))
6289 /* The final length is what we've output so far, plus what else is about to
6290 * be output. (If 'count' is non-zero, then the input list we exhausted
6291 * has everything remaining up to the machine's limit in its set, and hence
6292 * in the union, so there will be no further output. */
6295 /* At most one of the subexpressions will be non-zero */
6296 len_u += (len_a - i_a) + (len_b - i_b);
6299 /* Set result to final length, which can change the pointer to array_u, so
6301 if (len_u != invlist_len(u)) {
6302 invlist_set_len(u, len_u);
6304 array_u = invlist_array(u);
6307 /* When 'count' is 0, the list that was exhausted (if one was shorter than
6308 * the other) ended with everything above it not in its set. That means
6309 * that the remaining part of the union is precisely the same as the
6310 * non-exhausted list, so can just copy it unchanged. (If both list were
6311 * exhausted at the same time, then the operations below will be both 0.)
6314 IV copy_count; /* At most one will have a non-zero copy count */
6315 if ((copy_count = len_a - i_a) > 0) {
6316 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
6318 else if ((copy_count = len_b - i_b) > 0) {
6319 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
6323 /* We may be removing a reference to one of the inputs */
6324 if (&a == output || &b == output) {
6325 SvREFCNT_dec(*output);
6333 Perl__invlist_intersection(pTHX_ SV* const a, SV* const b, SV** i)
6335 /* Take the intersection of two inversion lists and point 'i' to it. If
6336 * 'i' on input points to one of the two lists, the reference count to that
6337 * list will be decremented.
6338 * The basis for this comes from "Unicode Demystified" Chapter 13 by
6339 * Richard Gillam, published by Addison-Wesley, and explained at some
6340 * length there. The preface says to incorporate its examples into your
6341 * code at your own risk. In fact, it had bugs
6343 * The algorithm is like a merge sort, and is essentially the same as the
6347 UV* array_a; /* a's array */
6349 UV len_a; /* length of a's array */
6352 SV* r; /* the resulting intersection */
6356 UV i_a = 0; /* current index into a's array */
6360 /* running count, as explained in the algorithm source book; items are
6361 * stopped accumulating and are output when the count changes to/from 2.
6362 * The count is incremented when we start a range that's in the set, and
6363 * decremented when we start a range that's not in the set. So its range
6364 * is 0 to 2. Only when the count is 2 is something in the intersection.
6368 PERL_ARGS_ASSERT__INVLIST_INTERSECTION;
6370 /* If either one is empty, the intersection is null */
6371 len_a = invlist_len(a);
6372 if ((len_a == 0) || ((len_b = invlist_len(b)) == 0)) {
6373 *i = _new_invlist(0);
6375 /* If the result is the same as one of the inputs, the input is being
6386 /* Here both lists exist and are non-empty */
6387 array_a = invlist_array(a);
6388 array_b = invlist_array(b);
6390 /* Size the intersection for the worst case: that the intersection ends up
6391 * fragmenting everything to be completely disjoint */
6392 r= _new_invlist(len_a + len_b);
6394 /* Will contain U+0000 iff both components do */
6395 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
6396 && len_b > 0 && array_b[0] == 0);
6398 /* Go through each list item by item, stopping when exhausted one of
6400 while (i_a < len_a && i_b < len_b) {
6401 UV cp; /* The element to potentially add to the intersection's
6403 bool cp_in_set; /* Is it in the input list's set or not */
6405 /* We need to take one or the other of the two inputs for the
6406 * intersection. Since we are merging two sorted lists, we take the
6407 * smaller of the next items. In case of a tie, we take the one that
6408 * is not in its set first (a difference from the union algorithm). If
6409 * we took one in the set first, it would increment the count, possibly
6410 * to 2 which would cause it to be output as starting a range in the
6411 * intersection, and the next time through we would take that same
6412 * number, and output it again as ending the set. By doing it the
6413 * opposite of this, there is no possibility that the count will be
6414 * momentarily incremented to 2. (In a tie and both are in the set or
6415 * both not in the set, it doesn't matter which we take first.) */
6416 if (array_a[i_a] < array_b[i_b]
6417 || (array_a[i_a] == array_b[i_b] && ! ELEMENT_IN_INVLIST_SET(i_a)))
6419 cp_in_set = ELEMENT_IN_INVLIST_SET(i_a);
6423 cp_in_set = ELEMENT_IN_INVLIST_SET(i_b);
6427 /* Here, have chosen which of the two inputs to look at. Only output
6428 * if the running count changes to/from 2, which marks the
6429 * beginning/end of a range that's in the intersection */
6433 array_r[i_r++] = cp;
6438 array_r[i_r++] = cp;
6444 /* Here, we are finished going through at least one of the lists, which
6445 * means there is something remaining in at most one. We check if the list
6446 * that has been exhausted is positioned such that we are in the middle
6447 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
6448 * the ones we care about.) There are four cases:
6449 * 1) Both weren't in their sets, count is 0, and remains 0. There's
6450 * nothing left in the intersection.
6451 * 2) Both were in their sets, count is 2 and perhaps is incremented to
6452 * above 2. What should be output is exactly that which is in the
6453 * non-exhausted set, as everything it has is also in the intersection
6454 * set, and everything it doesn't have can't be in the intersection
6455 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
6456 * gets incremented to 2. Like the previous case, the intersection is
6457 * everything that remains in the non-exhausted set.
6458 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
6459 * remains 1. And the intersection has nothing more. */
6460 if ((i_a == len_a && PREV_ELEMENT_IN_INVLIST_SET(i_a))
6461 || (i_b == len_b && PREV_ELEMENT_IN_INVLIST_SET(i_b)))
6466 /* The final length is what we've output so far plus what else is in the
6467 * intersection. At most one of the subexpressions below will be non-zero */
6470 len_r += (len_a - i_a) + (len_b - i_b);
6473 /* Set result to final length, which can change the pointer to array_r, so
6475 if (len_r != invlist_len(r)) {
6476 invlist_set_len(r, len_r);
6478 array_r = invlist_array(r);
6481 /* Finish outputting any remaining */
6482 if (count >= 2) { /* At most one will have a non-zero copy count */
6484 if ((copy_count = len_a - i_a) > 0) {
6485 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
6487 else if ((copy_count = len_b - i_b) > 0) {
6488 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
6492 /* We may be removing a reference to one of the inputs */
6493 if (&a == i || &b == i) {
6504 S_add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
6506 /* Add the range from 'start' to 'end' inclusive to the inversion list's
6507 * set. A pointer to the inversion list is returned. This may actually be
6508 * a new list, in which case the passed in one has been destroyed. The
6509 * passed in inversion list can be NULL, in which case a new one is created
6510 * with just the one range in it */
6515 if (invlist == NULL) {
6516 invlist = _new_invlist(2);
6520 len = invlist_len(invlist);
6523 /* If comes after the final entry, can just append it to the end */
6525 || start >= invlist_array(invlist)
6526 [invlist_len(invlist) - 1])
6528 _append_range_to_invlist(invlist, start, end);
6532 /* Here, can't just append things, create and return a new inversion list
6533 * which is the union of this range and the existing inversion list */
6534 range_invlist = _new_invlist(2);
6535 _append_range_to_invlist(range_invlist, start, end);
6537 _invlist_union(invlist, range_invlist, &invlist);
6539 /* The temporary can be freed */
6540 SvREFCNT_dec(range_invlist);
6545 PERL_STATIC_INLINE SV*
6546 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
6547 return add_range_to_invlist(invlist, cp, cp);
6550 #ifndef PERL_IN_XSUB_RE
6552 Perl__invlist_invert(pTHX_ SV* const invlist)
6554 /* Complement the input inversion list. This adds a 0 if the list didn't
6555 * have a zero; removes it otherwise. As described above, the data
6556 * structure is set up so that this is very efficient */
6558 UV* len_pos = get_invlist_len_addr(invlist);
6560 PERL_ARGS_ASSERT__INVLIST_INVERT;
6562 /* The inverse of matching nothing is matching everything */
6563 if (*len_pos == 0) {
6564 _append_range_to_invlist(invlist, 0, UV_MAX);
6568 /* The exclusive or complents 0 to 1; and 1 to 0. If the result is 1, the
6569 * zero element was a 0, so it is being removed, so the length decrements
6570 * by 1; and vice-versa. SvCUR is unaffected */
6571 if (*get_invlist_zero_addr(invlist) ^= 1) {
6580 Perl__invlist_invert_prop(pTHX_ SV* const invlist)
6582 /* Complement the input inversion list (which must be a Unicode property,
6583 * all of which don't match above the Unicode maximum code point.) And
6584 * Perl has chosen to not have the inversion match above that either. This
6585 * adds a 0x110000 if the list didn't end with it, and removes it if it did
6591 PERL_ARGS_ASSERT__INVLIST_INVERT_PROP;
6593 _invlist_invert(invlist);
6595 len = invlist_len(invlist);
6597 if (len != 0) { /* If empty do nothing */
6598 array = invlist_array(invlist);
6599 if (array[len - 1] != PERL_UNICODE_MAX + 1) {
6600 /* Add 0x110000. First, grow if necessary */
6602 if (invlist_max(invlist) < len) {
6603 invlist_extend(invlist, len);
6604 array = invlist_array(invlist);
6606 invlist_set_len(invlist, len);
6607 array[len - 1] = PERL_UNICODE_MAX + 1;
6609 else { /* Remove the 0x110000 */
6610 invlist_set_len(invlist, len - 1);
6618 PERL_STATIC_INLINE SV*
6619 S_invlist_clone(pTHX_ SV* const invlist)
6622 /* Return a new inversion list that is a copy of the input one, which is
6625 SV* new_invlist = _new_invlist(SvCUR(invlist));
6627 PERL_ARGS_ASSERT_INVLIST_CLONE;
6629 Copy(SvPVX(invlist), SvPVX(new_invlist), SvCUR(invlist), char);
6633 #ifndef PERL_IN_XSUB_RE
6635 Perl__invlist_subtract(pTHX_ SV* const a, SV* const b, SV** result)
6637 /* Point result to an inversion list which consists of all elements in 'a'
6638 * that aren't also in 'b' */
6640 PERL_ARGS_ASSERT__INVLIST_SUBTRACT;
6642 /* Subtracting nothing retains the original */
6643 if (invlist_len(b) == 0) {
6645 /* If the result is not to be the same variable as the original, create
6648 *result = invlist_clone(a);
6651 SV *b_copy = invlist_clone(b);
6652 _invlist_invert(b_copy); /* Everything not in 'b' */
6653 _invlist_intersection(a, b_copy, result); /* Everything in 'a' not in
6655 SvREFCNT_dec(b_copy);
6666 PERL_STATIC_INLINE UV*
6667 S_get_invlist_iter_addr(pTHX_ SV* invlist)
6669 /* Return the address of the UV that contains the current iteration
6672 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
6674 return (UV *) (SvPVX(invlist) + (INVLIST_ITER_OFFSET * sizeof (UV)));
6677 PERL_STATIC_INLINE void
6678 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
6680 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
6682 *get_invlist_iter_addr(invlist) = 0;
6686 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
6688 UV* pos = get_invlist_iter_addr(invlist);
6689 UV len = invlist_len(invlist);
6692 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
6695 *pos = UV_MAX; /* Force iternit() to be required next time */
6699 array = invlist_array(invlist);
6701 *start = array[(*pos)++];
6707 *end = array[(*pos)++] - 1;
6715 S_invlist_dump(pTHX_ SV* const invlist, const char * const header)
6717 /* Dumps out the ranges in an inversion list. The string 'header'
6718 * if present is output on a line before the first range */
6722 if (header && strlen(header)) {
6723 PerlIO_printf(Perl_debug_log, "%s\n", header);
6725 invlist_iterinit(invlist);
6726 while (invlist_iternext(invlist, &start, &end)) {
6727 if (end == UV_MAX) {
6728 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. INFINITY\n", start);
6731 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. 0x%04"UVXf"\n", start, end);
6737 #undef HEADER_LENGTH
6738 #undef INVLIST_INITIAL_LENGTH
6739 #undef TO_INTERNAL_SIZE
6740 #undef FROM_INTERNAL_SIZE
6741 #undef INVLIST_LEN_OFFSET
6742 #undef INVLIST_ZERO_OFFSET
6743 #undef INVLIST_ITER_OFFSET
6745 /* End of inversion list object */
6748 - reg - regular expression, i.e. main body or parenthesized thing
6750 * Caller must absorb opening parenthesis.
6752 * Combining parenthesis handling with the base level of regular expression
6753 * is a trifle forced, but the need to tie the tails of the branches to what
6754 * follows makes it hard to avoid.
6756 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
6758 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
6760 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
6764 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
6765 /* paren: Parenthesized? 0=top, 1=(, inside: changed to letter. */
6768 register regnode *ret; /* Will be the head of the group. */
6769 register regnode *br;
6770 register regnode *lastbr;
6771 register regnode *ender = NULL;
6772 register I32 parno = 0;
6774 U32 oregflags = RExC_flags;
6775 bool have_branch = 0;
6777 I32 freeze_paren = 0;
6778 I32 after_freeze = 0;
6780 /* for (?g), (?gc), and (?o) warnings; warning
6781 about (?c) will warn about (?g) -- japhy */
6783 #define WASTED_O 0x01
6784 #define WASTED_G 0x02
6785 #define WASTED_C 0x04
6786 #define WASTED_GC (0x02|0x04)
6787 I32 wastedflags = 0x00;
6789 char * parse_start = RExC_parse; /* MJD */
6790 char * const oregcomp_parse = RExC_parse;
6792 GET_RE_DEBUG_FLAGS_DECL;
6794 PERL_ARGS_ASSERT_REG;
6795 DEBUG_PARSE("reg ");
6797 *flagp = 0; /* Tentatively. */
6800 /* Make an OPEN node, if parenthesized. */
6802 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
6803 char *start_verb = RExC_parse;
6804 STRLEN verb_len = 0;
6805 char *start_arg = NULL;
6806 unsigned char op = 0;
6808 int internal_argval = 0; /* internal_argval is only useful if !argok */
6809 while ( *RExC_parse && *RExC_parse != ')' ) {
6810 if ( *RExC_parse == ':' ) {
6811 start_arg = RExC_parse + 1;
6817 verb_len = RExC_parse - start_verb;
6820 while ( *RExC_parse && *RExC_parse != ')' )
6822 if ( *RExC_parse != ')' )
6823 vFAIL("Unterminated verb pattern argument");
6824 if ( RExC_parse == start_arg )
6827 if ( *RExC_parse != ')' )
6828 vFAIL("Unterminated verb pattern");
6831 switch ( *start_verb ) {
6832 case 'A': /* (*ACCEPT) */
6833 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
6835 internal_argval = RExC_nestroot;
6838 case 'C': /* (*COMMIT) */
6839 if ( memEQs(start_verb,verb_len,"COMMIT") )
6842 case 'F': /* (*FAIL) */
6843 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
6848 case ':': /* (*:NAME) */
6849 case 'M': /* (*MARK:NAME) */
6850 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
6855 case 'P': /* (*PRUNE) */
6856 if ( memEQs(start_verb,verb_len,"PRUNE") )
6859 case 'S': /* (*SKIP) */
6860 if ( memEQs(start_verb,verb_len,"SKIP") )
6863 case 'T': /* (*THEN) */
6864 /* [19:06] <TimToady> :: is then */
6865 if ( memEQs(start_verb,verb_len,"THEN") ) {
6867 RExC_seen |= REG_SEEN_CUTGROUP;
6873 vFAIL3("Unknown verb pattern '%.*s'",
6874 verb_len, start_verb);
6877 if ( start_arg && internal_argval ) {
6878 vFAIL3("Verb pattern '%.*s' may not have an argument",
6879 verb_len, start_verb);
6880 } else if ( argok < 0 && !start_arg ) {
6881 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
6882 verb_len, start_verb);
6884 ret = reganode(pRExC_state, op, internal_argval);
6885 if ( ! internal_argval && ! SIZE_ONLY ) {
6887 SV *sv = newSVpvn( start_arg, RExC_parse - start_arg);
6888 ARG(ret) = add_data( pRExC_state, 1, "S" );
6889 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
6896 if (!internal_argval)
6897 RExC_seen |= REG_SEEN_VERBARG;
6898 } else if ( start_arg ) {
6899 vFAIL3("Verb pattern '%.*s' may not have an argument",
6900 verb_len, start_verb);
6902 ret = reg_node(pRExC_state, op);
6904 nextchar(pRExC_state);
6907 if (*RExC_parse == '?') { /* (?...) */
6908 bool is_logical = 0;
6909 const char * const seqstart = RExC_parse;
6910 bool has_use_defaults = FALSE;
6913 paren = *RExC_parse++;
6914 ret = NULL; /* For look-ahead/behind. */
6917 case 'P': /* (?P...) variants for those used to PCRE/Python */
6918 paren = *RExC_parse++;
6919 if ( paren == '<') /* (?P<...>) named capture */
6921 else if (paren == '>') { /* (?P>name) named recursion */
6922 goto named_recursion;
6924 else if (paren == '=') { /* (?P=...) named backref */
6925 /* this pretty much dupes the code for \k<NAME> in regatom(), if
6926 you change this make sure you change that */
6927 char* name_start = RExC_parse;
6929 SV *sv_dat = reg_scan_name(pRExC_state,
6930 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
6931 if (RExC_parse == name_start || *RExC_parse != ')')
6932 vFAIL2("Sequence %.3s... not terminated",parse_start);
6935 num = add_data( pRExC_state, 1, "S" );
6936 RExC_rxi->data->data[num]=(void*)sv_dat;
6937 SvREFCNT_inc_simple_void(sv_dat);
6940 ret = reganode(pRExC_state,
6943 : (MORE_ASCII_RESTRICTED)
6945 : (AT_LEAST_UNI_SEMANTICS)
6953 Set_Node_Offset(ret, parse_start+1);
6954 Set_Node_Cur_Length(ret); /* MJD */
6956 nextchar(pRExC_state);
6960 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
6962 case '<': /* (?<...) */
6963 if (*RExC_parse == '!')
6965 else if (*RExC_parse != '=')
6971 case '\'': /* (?'...') */
6972 name_start= RExC_parse;
6973 svname = reg_scan_name(pRExC_state,
6974 SIZE_ONLY ? /* reverse test from the others */
6975 REG_RSN_RETURN_NAME :
6976 REG_RSN_RETURN_NULL);
6977 if (RExC_parse == name_start) {
6979 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
6982 if (*RExC_parse != paren)
6983 vFAIL2("Sequence (?%c... not terminated",
6984 paren=='>' ? '<' : paren);
6988 if (!svname) /* shouldn't happen */
6990 "panic: reg_scan_name returned NULL");
6991 if (!RExC_paren_names) {
6992 RExC_paren_names= newHV();
6993 sv_2mortal(MUTABLE_SV(RExC_paren_names));
6995 RExC_paren_name_list= newAV();
6996 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
6999 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
7001 sv_dat = HeVAL(he_str);
7003 /* croak baby croak */
7005 "panic: paren_name hash element allocation failed");
7006 } else if ( SvPOK(sv_dat) ) {
7007 /* (?|...) can mean we have dupes so scan to check
7008 its already been stored. Maybe a flag indicating
7009 we are inside such a construct would be useful,
7010 but the arrays are likely to be quite small, so
7011 for now we punt -- dmq */
7012 IV count = SvIV(sv_dat);
7013 I32 *pv = (I32*)SvPVX(sv_dat);
7015 for ( i = 0 ; i < count ; i++ ) {
7016 if ( pv[i] == RExC_npar ) {
7022 pv = (I32*)SvGROW(sv_dat, SvCUR(sv_dat) + sizeof(I32)+1);
7023 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
7024 pv[count] = RExC_npar;
7025 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
7028 (void)SvUPGRADE(sv_dat,SVt_PVNV);
7029 sv_setpvn(sv_dat, (char *)&(RExC_npar), sizeof(I32));
7031 SvIV_set(sv_dat, 1);
7034 /* Yes this does cause a memory leak in debugging Perls */
7035 if (!av_store(RExC_paren_name_list, RExC_npar, SvREFCNT_inc(svname)))
7036 SvREFCNT_dec(svname);
7039 /*sv_dump(sv_dat);*/
7041 nextchar(pRExC_state);
7043 goto capturing_parens;
7045 RExC_seen |= REG_SEEN_LOOKBEHIND;
7046 RExC_in_lookbehind++;
7048 case '=': /* (?=...) */
7049 RExC_seen_zerolen++;
7051 case '!': /* (?!...) */
7052 RExC_seen_zerolen++;
7053 if (*RExC_parse == ')') {
7054 ret=reg_node(pRExC_state, OPFAIL);
7055 nextchar(pRExC_state);
7059 case '|': /* (?|...) */
7060 /* branch reset, behave like a (?:...) except that
7061 buffers in alternations share the same numbers */
7063 after_freeze = freeze_paren = RExC_npar;
7065 case ':': /* (?:...) */
7066 case '>': /* (?>...) */
7068 case '$': /* (?$...) */
7069 case '@': /* (?@...) */
7070 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
7072 case '#': /* (?#...) */
7073 while (*RExC_parse && *RExC_parse != ')')
7075 if (*RExC_parse != ')')
7076 FAIL("Sequence (?#... not terminated");
7077 nextchar(pRExC_state);
7080 case '0' : /* (?0) */
7081 case 'R' : /* (?R) */
7082 if (*RExC_parse != ')')
7083 FAIL("Sequence (?R) not terminated");
7084 ret = reg_node(pRExC_state, GOSTART);
7085 *flagp |= POSTPONED;
7086 nextchar(pRExC_state);
7089 { /* named and numeric backreferences */
7091 case '&': /* (?&NAME) */
7092 parse_start = RExC_parse - 1;
7095 SV *sv_dat = reg_scan_name(pRExC_state,
7096 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
7097 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
7099 goto gen_recurse_regop;
7102 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
7104 vFAIL("Illegal pattern");
7106 goto parse_recursion;
7108 case '-': /* (?-1) */
7109 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
7110 RExC_parse--; /* rewind to let it be handled later */
7114 case '1': case '2': case '3': case '4': /* (?1) */
7115 case '5': case '6': case '7': case '8': case '9':
7118 num = atoi(RExC_parse);
7119 parse_start = RExC_parse - 1; /* MJD */
7120 if (*RExC_parse == '-')
7122 while (isDIGIT(*RExC_parse))
7124 if (*RExC_parse!=')')
7125 vFAIL("Expecting close bracket");
7128 if ( paren == '-' ) {
7130 Diagram of capture buffer numbering.
7131 Top line is the normal capture buffer numbers
7132 Bottom line is the negative indexing as from
7136 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
7140 num = RExC_npar + num;
7143 vFAIL("Reference to nonexistent group");
7145 } else if ( paren == '+' ) {
7146 num = RExC_npar + num - 1;
7149 ret = reganode(pRExC_state, GOSUB, num);
7151 if (num > (I32)RExC_rx->nparens) {
7153 vFAIL("Reference to nonexistent group");
7155 ARG2L_SET( ret, RExC_recurse_count++);
7157 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
7158 "Recurse #%"UVuf" to %"IVdf"\n", (UV)ARG(ret), (IV)ARG2L(ret)));
7162 RExC_seen |= REG_SEEN_RECURSE;
7163 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
7164 Set_Node_Offset(ret, parse_start); /* MJD */
7166 *flagp |= POSTPONED;
7167 nextchar(pRExC_state);
7169 } /* named and numeric backreferences */
7172 case '?': /* (??...) */
7174 if (*RExC_parse != '{') {
7176 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
7179 *flagp |= POSTPONED;
7180 paren = *RExC_parse++;
7182 case '{': /* (?{...}) */
7187 char *s = RExC_parse;
7189 RExC_seen_zerolen++;
7190 RExC_seen |= REG_SEEN_EVAL;
7191 while (count && (c = *RExC_parse)) {
7202 if (*RExC_parse != ')') {
7204 vFAIL("Sequence (?{...}) not terminated or not {}-balanced");
7208 OP_4tree *sop, *rop;
7209 SV * const sv = newSVpvn(s, RExC_parse - 1 - s);
7212 Perl_save_re_context(aTHX);
7213 rop = Perl_sv_compile_2op_is_broken(aTHX_ sv, &sop, "re", &pad);
7214 sop->op_private |= OPpREFCOUNTED;
7215 /* re_dup will OpREFCNT_inc */
7216 OpREFCNT_set(sop, 1);
7219 n = add_data(pRExC_state, 3, "nop");
7220 RExC_rxi->data->data[n] = (void*)rop;
7221 RExC_rxi->data->data[n+1] = (void*)sop;
7222 RExC_rxi->data->data[n+2] = (void*)pad;
7225 else { /* First pass */
7226 if (PL_reginterp_cnt < ++RExC_seen_evals
7228 /* No compiled RE interpolated, has runtime
7229 components ===> unsafe. */
7230 FAIL("Eval-group not allowed at runtime, use re 'eval'");
7231 if (PL_tainting && PL_tainted)
7232 FAIL("Eval-group in insecure regular expression");
7233 #if PERL_VERSION > 8
7234 if (IN_PERL_COMPILETIME)
7239 nextchar(pRExC_state);
7241 ret = reg_node(pRExC_state, LOGICAL);
7244 REGTAIL(pRExC_state, ret, reganode(pRExC_state, EVAL, n));
7245 /* deal with the length of this later - MJD */
7248 ret = reganode(pRExC_state, EVAL, n);
7249 Set_Node_Length(ret, RExC_parse - parse_start + 1);
7250 Set_Node_Offset(ret, parse_start);
7253 case '(': /* (?(?{...})...) and (?(?=...)...) */
7256 if (RExC_parse[0] == '?') { /* (?(?...)) */
7257 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
7258 || RExC_parse[1] == '<'
7259 || RExC_parse[1] == '{') { /* Lookahead or eval. */
7262 ret = reg_node(pRExC_state, LOGICAL);
7265 REGTAIL(pRExC_state, ret, reg(pRExC_state, 1, &flag,depth+1));
7269 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
7270 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
7272 char ch = RExC_parse[0] == '<' ? '>' : '\'';
7273 char *name_start= RExC_parse++;
7275 SV *sv_dat=reg_scan_name(pRExC_state,
7276 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
7277 if (RExC_parse == name_start || *RExC_parse != ch)
7278 vFAIL2("Sequence (?(%c... not terminated",
7279 (ch == '>' ? '<' : ch));
7282 num = add_data( pRExC_state, 1, "S" );
7283 RExC_rxi->data->data[num]=(void*)sv_dat;
7284 SvREFCNT_inc_simple_void(sv_dat);
7286 ret = reganode(pRExC_state,NGROUPP,num);
7287 goto insert_if_check_paren;
7289 else if (RExC_parse[0] == 'D' &&
7290 RExC_parse[1] == 'E' &&
7291 RExC_parse[2] == 'F' &&
7292 RExC_parse[3] == 'I' &&
7293 RExC_parse[4] == 'N' &&
7294 RExC_parse[5] == 'E')
7296 ret = reganode(pRExC_state,DEFINEP,0);
7299 goto insert_if_check_paren;
7301 else if (RExC_parse[0] == 'R') {
7304 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
7305 parno = atoi(RExC_parse++);
7306 while (isDIGIT(*RExC_parse))
7308 } else if (RExC_parse[0] == '&') {
7311 sv_dat = reg_scan_name(pRExC_state,
7312 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
7313 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
7315 ret = reganode(pRExC_state,INSUBP,parno);
7316 goto insert_if_check_paren;
7318 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
7321 parno = atoi(RExC_parse++);
7323 while (isDIGIT(*RExC_parse))
7325 ret = reganode(pRExC_state, GROUPP, parno);
7327 insert_if_check_paren:
7328 if ((c = *nextchar(pRExC_state)) != ')')
7329 vFAIL("Switch condition not recognized");
7331 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
7332 br = regbranch(pRExC_state, &flags, 1,depth+1);
7334 br = reganode(pRExC_state, LONGJMP, 0);
7336 REGTAIL(pRExC_state, br, reganode(pRExC_state, LONGJMP, 0));
7337 c = *nextchar(pRExC_state);
7342 vFAIL("(?(DEFINE)....) does not allow branches");
7343 lastbr = reganode(pRExC_state, IFTHEN, 0); /* Fake one for optimizer. */
7344 regbranch(pRExC_state, &flags, 1,depth+1);
7345 REGTAIL(pRExC_state, ret, lastbr);
7348 c = *nextchar(pRExC_state);
7353 vFAIL("Switch (?(condition)... contains too many branches");
7354 ender = reg_node(pRExC_state, TAIL);
7355 REGTAIL(pRExC_state, br, ender);
7357 REGTAIL(pRExC_state, lastbr, ender);
7358 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
7361 REGTAIL(pRExC_state, ret, ender);
7362 RExC_size++; /* XXX WHY do we need this?!!
7363 For large programs it seems to be required
7364 but I can't figure out why. -- dmq*/
7368 vFAIL2("Unknown switch condition (?(%.2s", RExC_parse);
7372 RExC_parse--; /* for vFAIL to print correctly */
7373 vFAIL("Sequence (? incomplete");
7375 case DEFAULT_PAT_MOD: /* Use default flags with the exceptions
7377 has_use_defaults = TRUE;
7378 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
7379 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
7380 ? REGEX_UNICODE_CHARSET
7381 : REGEX_DEPENDS_CHARSET);
7385 parse_flags: /* (?i) */
7387 U32 posflags = 0, negflags = 0;
7388 U32 *flagsp = &posflags;
7389 char has_charset_modifier = '\0';
7390 regex_charset cs = (RExC_utf8 || RExC_uni_semantics)
7391 ? REGEX_UNICODE_CHARSET
7392 : REGEX_DEPENDS_CHARSET;
7394 while (*RExC_parse) {
7395 /* && strchr("iogcmsx", *RExC_parse) */
7396 /* (?g), (?gc) and (?o) are useless here
7397 and must be globally applied -- japhy */
7398 switch (*RExC_parse) {
7399 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
7400 case LOCALE_PAT_MOD:
7401 if (has_charset_modifier) {
7402 goto excess_modifier;
7404 else if (flagsp == &negflags) {
7407 cs = REGEX_LOCALE_CHARSET;
7408 has_charset_modifier = LOCALE_PAT_MOD;
7409 RExC_contains_locale = 1;
7411 case UNICODE_PAT_MOD:
7412 if (has_charset_modifier) {
7413 goto excess_modifier;
7415 else if (flagsp == &negflags) {
7418 cs = REGEX_UNICODE_CHARSET;
7419 has_charset_modifier = UNICODE_PAT_MOD;
7421 case ASCII_RESTRICT_PAT_MOD:
7422 if (flagsp == &negflags) {
7425 if (has_charset_modifier) {
7426 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
7427 goto excess_modifier;
7429 /* Doubled modifier implies more restricted */
7430 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
7433 cs = REGEX_ASCII_RESTRICTED_CHARSET;
7435 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
7437 case DEPENDS_PAT_MOD:
7438 if (has_use_defaults) {
7439 goto fail_modifiers;
7441 else if (flagsp == &negflags) {
7444 else if (has_charset_modifier) {
7445 goto excess_modifier;
7448 /* The dual charset means unicode semantics if the
7449 * pattern (or target, not known until runtime) are
7450 * utf8, or something in the pattern indicates unicode
7452 cs = (RExC_utf8 || RExC_uni_semantics)
7453 ? REGEX_UNICODE_CHARSET
7454 : REGEX_DEPENDS_CHARSET;
7455 has_charset_modifier = DEPENDS_PAT_MOD;
7459 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
7460 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
7462 else if (has_charset_modifier == *(RExC_parse - 1)) {
7463 vFAIL2("Regexp modifier \"%c\" may not appear twice", *(RExC_parse - 1));
7466 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
7471 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"", *(RExC_parse - 1));
7473 case ONCE_PAT_MOD: /* 'o' */
7474 case GLOBAL_PAT_MOD: /* 'g' */
7475 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
7476 const I32 wflagbit = *RExC_parse == 'o' ? WASTED_O : WASTED_G;
7477 if (! (wastedflags & wflagbit) ) {
7478 wastedflags |= wflagbit;
7481 "Useless (%s%c) - %suse /%c modifier",
7482 flagsp == &negflags ? "?-" : "?",
7484 flagsp == &negflags ? "don't " : "",
7491 case CONTINUE_PAT_MOD: /* 'c' */
7492 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
7493 if (! (wastedflags & WASTED_C) ) {
7494 wastedflags |= WASTED_GC;
7497 "Useless (%sc) - %suse /gc modifier",
7498 flagsp == &negflags ? "?-" : "?",
7499 flagsp == &negflags ? "don't " : ""
7504 case KEEPCOPY_PAT_MOD: /* 'p' */
7505 if (flagsp == &negflags) {
7507 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
7509 *flagsp |= RXf_PMf_KEEPCOPY;
7513 /* A flag is a default iff it is following a minus, so
7514 * if there is a minus, it means will be trying to
7515 * re-specify a default which is an error */
7516 if (has_use_defaults || flagsp == &negflags) {
7519 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
7523 wastedflags = 0; /* reset so (?g-c) warns twice */
7529 RExC_flags |= posflags;
7530 RExC_flags &= ~negflags;
7531 set_regex_charset(&RExC_flags, cs);
7533 oregflags |= posflags;
7534 oregflags &= ~negflags;
7535 set_regex_charset(&oregflags, cs);
7537 nextchar(pRExC_state);
7548 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
7553 }} /* one for the default block, one for the switch */
7560 ret = reganode(pRExC_state, OPEN, parno);
7563 RExC_nestroot = parno;
7564 if (RExC_seen & REG_SEEN_RECURSE
7565 && !RExC_open_parens[parno-1])
7567 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
7568 "Setting open paren #%"IVdf" to %d\n",
7569 (IV)parno, REG_NODE_NUM(ret)));
7570 RExC_open_parens[parno-1]= ret;
7573 Set_Node_Length(ret, 1); /* MJD */
7574 Set_Node_Offset(ret, RExC_parse); /* MJD */
7582 /* Pick up the branches, linking them together. */
7583 parse_start = RExC_parse; /* MJD */
7584 br = regbranch(pRExC_state, &flags, 1,depth+1);
7586 /* branch_len = (paren != 0); */
7590 if (*RExC_parse == '|') {
7591 if (!SIZE_ONLY && RExC_extralen) {
7592 reginsert(pRExC_state, BRANCHJ, br, depth+1);
7595 reginsert(pRExC_state, BRANCH, br, depth+1);
7596 Set_Node_Length(br, paren != 0);
7597 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
7601 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
7603 else if (paren == ':') {
7604 *flagp |= flags&SIMPLE;
7606 if (is_open) { /* Starts with OPEN. */
7607 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
7609 else if (paren != '?') /* Not Conditional */
7611 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
7613 while (*RExC_parse == '|') {
7614 if (!SIZE_ONLY && RExC_extralen) {
7615 ender = reganode(pRExC_state, LONGJMP,0);
7616 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender); /* Append to the previous. */
7619 RExC_extralen += 2; /* Account for LONGJMP. */
7620 nextchar(pRExC_state);
7622 if (RExC_npar > after_freeze)
7623 after_freeze = RExC_npar;
7624 RExC_npar = freeze_paren;
7626 br = regbranch(pRExC_state, &flags, 0, depth+1);
7630 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
7632 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
7635 if (have_branch || paren != ':') {
7636 /* Make a closing node, and hook it on the end. */
7639 ender = reg_node(pRExC_state, TAIL);
7642 ender = reganode(pRExC_state, CLOSE, parno);
7643 if (!SIZE_ONLY && RExC_seen & REG_SEEN_RECURSE) {
7644 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
7645 "Setting close paren #%"IVdf" to %d\n",
7646 (IV)parno, REG_NODE_NUM(ender)));
7647 RExC_close_parens[parno-1]= ender;
7648 if (RExC_nestroot == parno)
7651 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
7652 Set_Node_Length(ender,1); /* MJD */
7658 *flagp &= ~HASWIDTH;
7661 ender = reg_node(pRExC_state, SUCCEED);
7664 ender = reg_node(pRExC_state, END);
7666 assert(!RExC_opend); /* there can only be one! */
7671 REGTAIL(pRExC_state, lastbr, ender);
7673 if (have_branch && !SIZE_ONLY) {
7675 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
7677 /* Hook the tails of the branches to the closing node. */
7678 for (br = ret; br; br = regnext(br)) {
7679 const U8 op = PL_regkind[OP(br)];
7681 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
7683 else if (op == BRANCHJ) {
7684 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
7692 static const char parens[] = "=!<,>";
7694 if (paren && (p = strchr(parens, paren))) {
7695 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
7696 int flag = (p - parens) > 1;
7699 node = SUSPEND, flag = 0;
7700 reginsert(pRExC_state, node,ret, depth+1);
7701 Set_Node_Cur_Length(ret);
7702 Set_Node_Offset(ret, parse_start + 1);
7704 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
7708 /* Check for proper termination. */
7710 RExC_flags = oregflags;
7711 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
7712 RExC_parse = oregcomp_parse;
7713 vFAIL("Unmatched (");
7716 else if (!paren && RExC_parse < RExC_end) {
7717 if (*RExC_parse == ')') {
7719 vFAIL("Unmatched )");
7722 FAIL("Junk on end of regexp"); /* "Can't happen". */
7726 if (RExC_in_lookbehind) {
7727 RExC_in_lookbehind--;
7729 if (after_freeze > RExC_npar)
7730 RExC_npar = after_freeze;
7735 - regbranch - one alternative of an | operator
7737 * Implements the concatenation operator.
7740 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
7743 register regnode *ret;
7744 register regnode *chain = NULL;
7745 register regnode *latest;
7746 I32 flags = 0, c = 0;
7747 GET_RE_DEBUG_FLAGS_DECL;
7749 PERL_ARGS_ASSERT_REGBRANCH;
7751 DEBUG_PARSE("brnc");
7756 if (!SIZE_ONLY && RExC_extralen)
7757 ret = reganode(pRExC_state, BRANCHJ,0);
7759 ret = reg_node(pRExC_state, BRANCH);
7760 Set_Node_Length(ret, 1);
7764 if (!first && SIZE_ONLY)
7765 RExC_extralen += 1; /* BRANCHJ */
7767 *flagp = WORST; /* Tentatively. */
7770 nextchar(pRExC_state);
7771 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
7773 latest = regpiece(pRExC_state, &flags,depth+1);
7774 if (latest == NULL) {
7775 if (flags & TRYAGAIN)
7779 else if (ret == NULL)
7781 *flagp |= flags&(HASWIDTH|POSTPONED);
7782 if (chain == NULL) /* First piece. */
7783 *flagp |= flags&SPSTART;
7786 REGTAIL(pRExC_state, chain, latest);
7791 if (chain == NULL) { /* Loop ran zero times. */
7792 chain = reg_node(pRExC_state, NOTHING);
7797 *flagp |= flags&SIMPLE;
7804 - regpiece - something followed by possible [*+?]
7806 * Note that the branching code sequences used for ? and the general cases
7807 * of * and + are somewhat optimized: they use the same NOTHING node as
7808 * both the endmarker for their branch list and the body of the last branch.
7809 * It might seem that this node could be dispensed with entirely, but the
7810 * endmarker role is not redundant.
7813 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
7816 register regnode *ret;
7818 register char *next;
7820 const char * const origparse = RExC_parse;
7822 I32 max = REG_INFTY;
7823 #ifdef RE_TRACK_PATTERN_OFFSETS
7826 const char *maxpos = NULL;
7827 GET_RE_DEBUG_FLAGS_DECL;
7829 PERL_ARGS_ASSERT_REGPIECE;
7831 DEBUG_PARSE("piec");
7833 ret = regatom(pRExC_state, &flags,depth+1);
7835 if (flags & TRYAGAIN)
7842 if (op == '{' && regcurly(RExC_parse)) {
7844 #ifdef RE_TRACK_PATTERN_OFFSETS
7845 parse_start = RExC_parse; /* MJD */
7847 next = RExC_parse + 1;
7848 while (isDIGIT(*next) || *next == ',') {
7857 if (*next == '}') { /* got one */
7861 min = atoi(RExC_parse);
7865 maxpos = RExC_parse;
7867 if (!max && *maxpos != '0')
7868 max = REG_INFTY; /* meaning "infinity" */
7869 else if (max >= REG_INFTY)
7870 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
7872 nextchar(pRExC_state);
7875 if ((flags&SIMPLE)) {
7876 RExC_naughty += 2 + RExC_naughty / 2;
7877 reginsert(pRExC_state, CURLY, ret, depth+1);
7878 Set_Node_Offset(ret, parse_start+1); /* MJD */
7879 Set_Node_Cur_Length(ret);
7882 regnode * const w = reg_node(pRExC_state, WHILEM);
7885 REGTAIL(pRExC_state, ret, w);
7886 if (!SIZE_ONLY && RExC_extralen) {
7887 reginsert(pRExC_state, LONGJMP,ret, depth+1);
7888 reginsert(pRExC_state, NOTHING,ret, depth+1);
7889 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
7891 reginsert(pRExC_state, CURLYX,ret, depth+1);
7893 Set_Node_Offset(ret, parse_start+1);
7894 Set_Node_Length(ret,
7895 op == '{' ? (RExC_parse - parse_start) : 1);
7897 if (!SIZE_ONLY && RExC_extralen)
7898 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
7899 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
7901 RExC_whilem_seen++, RExC_extralen += 3;
7902 RExC_naughty += 4 + RExC_naughty; /* compound interest */
7911 vFAIL("Can't do {n,m} with n > m");
7913 ARG1_SET(ret, (U16)min);
7914 ARG2_SET(ret, (U16)max);
7926 #if 0 /* Now runtime fix should be reliable. */
7928 /* if this is reinstated, don't forget to put this back into perldiag:
7930 =item Regexp *+ operand could be empty at {#} in regex m/%s/
7932 (F) The part of the regexp subject to either the * or + quantifier
7933 could match an empty string. The {#} shows in the regular
7934 expression about where the problem was discovered.
7938 if (!(flags&HASWIDTH) && op != '?')
7939 vFAIL("Regexp *+ operand could be empty");
7942 #ifdef RE_TRACK_PATTERN_OFFSETS
7943 parse_start = RExC_parse;
7945 nextchar(pRExC_state);
7947 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
7949 if (op == '*' && (flags&SIMPLE)) {
7950 reginsert(pRExC_state, STAR, ret, depth+1);
7954 else if (op == '*') {
7958 else if (op == '+' && (flags&SIMPLE)) {
7959 reginsert(pRExC_state, PLUS, ret, depth+1);
7963 else if (op == '+') {
7967 else if (op == '?') {
7972 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
7973 ckWARN3reg(RExC_parse,
7974 "%.*s matches null string many times",
7975 (int)(RExC_parse >= origparse ? RExC_parse - origparse : 0),
7979 if (RExC_parse < RExC_end && *RExC_parse == '?') {
7980 nextchar(pRExC_state);
7981 reginsert(pRExC_state, MINMOD, ret, depth+1);
7982 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
7984 #ifndef REG_ALLOW_MINMOD_SUSPEND
7987 if (RExC_parse < RExC_end && *RExC_parse == '+') {
7989 nextchar(pRExC_state);
7990 ender = reg_node(pRExC_state, SUCCEED);
7991 REGTAIL(pRExC_state, ret, ender);
7992 reginsert(pRExC_state, SUSPEND, ret, depth+1);
7994 ender = reg_node(pRExC_state, TAIL);
7995 REGTAIL(pRExC_state, ret, ender);
7999 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
8001 vFAIL("Nested quantifiers");
8008 /* reg_namedseq(pRExC_state,UVp, UV depth)
8010 This is expected to be called by a parser routine that has
8011 recognized '\N' and needs to handle the rest. RExC_parse is
8012 expected to point at the first char following the N at the time
8015 The \N may be inside (indicated by valuep not being NULL) or outside a
8018 \N may begin either a named sequence, or if outside a character class, mean
8019 to match a non-newline. For non single-quoted regexes, the tokenizer has
8020 attempted to decide which, and in the case of a named sequence converted it
8021 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
8022 where c1... are the characters in the sequence. For single-quoted regexes,
8023 the tokenizer passes the \N sequence through unchanged; this code will not
8024 attempt to determine this nor expand those. The net effect is that if the
8025 beginning of the passed-in pattern isn't '{U+' or there is no '}', it
8026 signals that this \N occurrence means to match a non-newline.
8028 Only the \N{U+...} form should occur in a character class, for the same
8029 reason that '.' inside a character class means to just match a period: it
8030 just doesn't make sense.
8032 If valuep is non-null then it is assumed that we are parsing inside
8033 of a charclass definition and the first codepoint in the resolved
8034 string is returned via *valuep and the routine will return NULL.
8035 In this mode if a multichar string is returned from the charnames
8036 handler, a warning will be issued, and only the first char in the
8037 sequence will be examined. If the string returned is zero length
8038 then the value of *valuep is undefined and NON-NULL will
8039 be returned to indicate failure. (This will NOT be a valid pointer
8042 If valuep is null then it is assumed that we are parsing normal text and a
8043 new EXACT node is inserted into the program containing the resolved string,
8044 and a pointer to the new node is returned. But if the string is zero length
8045 a NOTHING node is emitted instead.
8047 On success RExC_parse is set to the char following the endbrace.
8048 Parsing failures will generate a fatal error via vFAIL(...)
8051 S_reg_namedseq(pTHX_ RExC_state_t *pRExC_state, UV *valuep, I32 *flagp, U32 depth)
8053 char * endbrace; /* '}' following the name */
8054 regnode *ret = NULL;
8057 GET_RE_DEBUG_FLAGS_DECL;
8059 PERL_ARGS_ASSERT_REG_NAMEDSEQ;
8063 /* The [^\n] meaning of \N ignores spaces and comments under the /x
8064 * modifier. The other meaning does not */
8065 p = (RExC_flags & RXf_PMf_EXTENDED)
8066 ? regwhite( pRExC_state, RExC_parse )
8069 /* Disambiguate between \N meaning a named character versus \N meaning
8070 * [^\n]. The former is assumed when it can't be the latter. */
8071 if (*p != '{' || regcurly(p)) {
8074 /* no bare \N in a charclass */
8075 vFAIL("\\N in a character class must be a named character: \\N{...}");
8077 nextchar(pRExC_state);
8078 ret = reg_node(pRExC_state, REG_ANY);
8079 *flagp |= HASWIDTH|SIMPLE;
8082 Set_Node_Length(ret, 1); /* MJD */
8086 /* Here, we have decided it should be a named sequence */
8088 /* The test above made sure that the next real character is a '{', but
8089 * under the /x modifier, it could be separated by space (or a comment and
8090 * \n) and this is not allowed (for consistency with \x{...} and the
8091 * tokenizer handling of \N{NAME}). */
8092 if (*RExC_parse != '{') {
8093 vFAIL("Missing braces on \\N{}");
8096 RExC_parse++; /* Skip past the '{' */
8098 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
8099 || ! (endbrace == RExC_parse /* nothing between the {} */
8100 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below */
8101 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg) */
8103 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
8104 vFAIL("\\N{NAME} must be resolved by the lexer");
8107 if (endbrace == RExC_parse) { /* empty: \N{} */
8109 RExC_parse = endbrace + 1;
8110 return reg_node(pRExC_state,NOTHING);
8114 ckWARNreg(RExC_parse,
8115 "Ignoring zero length \\N{} in character class"
8117 RExC_parse = endbrace + 1;
8120 return (regnode *) &RExC_parse; /* Invalid regnode pointer */
8123 REQUIRE_UTF8; /* named sequences imply Unicode semantics */
8124 RExC_parse += 2; /* Skip past the 'U+' */
8126 if (valuep) { /* In a bracketed char class */
8127 /* We only pay attention to the first char of
8128 multichar strings being returned. I kinda wonder
8129 if this makes sense as it does change the behaviour
8130 from earlier versions, OTOH that behaviour was broken
8131 as well. XXX Solution is to recharacterize as
8132 [rest-of-class]|multi1|multi2... */
8134 STRLEN length_of_hex;
8135 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES
8136 | PERL_SCAN_DISALLOW_PREFIX
8137 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
8139 char * endchar = RExC_parse + strcspn(RExC_parse, ".}");
8140 if (endchar < endbrace) {
8141 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
8144 length_of_hex = (STRLEN)(endchar - RExC_parse);
8145 *valuep = grok_hex(RExC_parse, &length_of_hex, &flags, NULL);
8147 /* The tokenizer should have guaranteed validity, but it's possible to
8148 * bypass it by using single quoting, so check */
8149 if (length_of_hex == 0
8150 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
8152 RExC_parse += length_of_hex; /* Includes all the valid */
8153 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
8154 ? UTF8SKIP(RExC_parse)
8156 /* Guard against malformed utf8 */
8157 if (RExC_parse >= endchar) RExC_parse = endchar;
8158 vFAIL("Invalid hexadecimal number in \\N{U+...}");
8161 RExC_parse = endbrace + 1;
8162 if (endchar == endbrace) return NULL;
8164 ret = (regnode *) &RExC_parse; /* Invalid regnode pointer */
8166 else { /* Not a char class */
8168 /* What is done here is to convert this to a sub-pattern of the form
8169 * (?:\x{char1}\x{char2}...)
8170 * and then call reg recursively. That way, it retains its atomicness,
8171 * while not having to worry about special handling that some code
8172 * points may have. toke.c has converted the original Unicode values
8173 * to native, so that we can just pass on the hex values unchanged. We
8174 * do have to set a flag to keep recoding from happening in the
8177 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
8179 char *endchar; /* Points to '.' or '}' ending cur char in the input
8181 char *orig_end = RExC_end;
8183 while (RExC_parse < endbrace) {
8185 /* Code points are separated by dots. If none, there is only one
8186 * code point, and is terminated by the brace */
8187 endchar = RExC_parse + strcspn(RExC_parse, ".}");
8189 /* Convert to notation the rest of the code understands */
8190 sv_catpv(substitute_parse, "\\x{");
8191 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
8192 sv_catpv(substitute_parse, "}");
8194 /* Point to the beginning of the next character in the sequence. */
8195 RExC_parse = endchar + 1;
8197 sv_catpv(substitute_parse, ")");
8199 RExC_parse = SvPV(substitute_parse, len);
8201 /* Don't allow empty number */
8203 vFAIL("Invalid hexadecimal number in \\N{U+...}");
8205 RExC_end = RExC_parse + len;
8207 /* The values are Unicode, and therefore not subject to recoding */
8208 RExC_override_recoding = 1;
8210 ret = reg(pRExC_state, 1, flagp, depth+1);
8212 RExC_parse = endbrace;
8213 RExC_end = orig_end;
8214 RExC_override_recoding = 0;
8216 nextchar(pRExC_state);
8226 * It returns the code point in utf8 for the value in *encp.
8227 * value: a code value in the source encoding
8228 * encp: a pointer to an Encode object
8230 * If the result from Encode is not a single character,
8231 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
8234 S_reg_recode(pTHX_ const char value, SV **encp)
8237 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
8238 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
8239 const STRLEN newlen = SvCUR(sv);
8240 UV uv = UNICODE_REPLACEMENT;
8242 PERL_ARGS_ASSERT_REG_RECODE;
8246 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
8249 if (!newlen || numlen != newlen) {
8250 uv = UNICODE_REPLACEMENT;
8258 - regatom - the lowest level
8260 Try to identify anything special at the start of the pattern. If there
8261 is, then handle it as required. This may involve generating a single regop,
8262 such as for an assertion; or it may involve recursing, such as to
8263 handle a () structure.
8265 If the string doesn't start with something special then we gobble up
8266 as much literal text as we can.
8268 Once we have been able to handle whatever type of thing started the
8269 sequence, we return.
8271 Note: we have to be careful with escapes, as they can be both literal
8272 and special, and in the case of \10 and friends can either, depending
8273 on context. Specifically there are two separate switches for handling
8274 escape sequences, with the one for handling literal escapes requiring
8275 a dummy entry for all of the special escapes that are actually handled
8280 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
8283 register regnode *ret = NULL;
8285 char *parse_start = RExC_parse;
8287 GET_RE_DEBUG_FLAGS_DECL;
8288 DEBUG_PARSE("atom");
8289 *flagp = WORST; /* Tentatively. */
8291 PERL_ARGS_ASSERT_REGATOM;
8294 switch ((U8)*RExC_parse) {
8296 RExC_seen_zerolen++;
8297 nextchar(pRExC_state);
8298 if (RExC_flags & RXf_PMf_MULTILINE)
8299 ret = reg_node(pRExC_state, MBOL);
8300 else if (RExC_flags & RXf_PMf_SINGLELINE)
8301 ret = reg_node(pRExC_state, SBOL);
8303 ret = reg_node(pRExC_state, BOL);
8304 Set_Node_Length(ret, 1); /* MJD */
8307 nextchar(pRExC_state);
8309 RExC_seen_zerolen++;
8310 if (RExC_flags & RXf_PMf_MULTILINE)
8311 ret = reg_node(pRExC_state, MEOL);
8312 else if (RExC_flags & RXf_PMf_SINGLELINE)
8313 ret = reg_node(pRExC_state, SEOL);
8315 ret = reg_node(pRExC_state, EOL);
8316 Set_Node_Length(ret, 1); /* MJD */
8319 nextchar(pRExC_state);
8320 if (RExC_flags & RXf_PMf_SINGLELINE)
8321 ret = reg_node(pRExC_state, SANY);
8323 ret = reg_node(pRExC_state, REG_ANY);
8324 *flagp |= HASWIDTH|SIMPLE;
8326 Set_Node_Length(ret, 1); /* MJD */
8330 char * const oregcomp_parse = ++RExC_parse;
8331 ret = regclass(pRExC_state,depth+1);
8332 if (*RExC_parse != ']') {
8333 RExC_parse = oregcomp_parse;
8334 vFAIL("Unmatched [");
8336 nextchar(pRExC_state);
8337 *flagp |= HASWIDTH|SIMPLE;
8338 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
8342 nextchar(pRExC_state);
8343 ret = reg(pRExC_state, 1, &flags,depth+1);
8345 if (flags & TRYAGAIN) {
8346 if (RExC_parse == RExC_end) {
8347 /* Make parent create an empty node if needed. */
8355 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
8359 if (flags & TRYAGAIN) {
8363 vFAIL("Internal urp");
8364 /* Supposed to be caught earlier. */
8367 if (!regcurly(RExC_parse)) {
8376 vFAIL("Quantifier follows nothing");
8381 This switch handles escape sequences that resolve to some kind
8382 of special regop and not to literal text. Escape sequnces that
8383 resolve to literal text are handled below in the switch marked
8386 Every entry in this switch *must* have a corresponding entry
8387 in the literal escape switch. However, the opposite is not
8388 required, as the default for this switch is to jump to the
8389 literal text handling code.
8391 switch ((U8)*++RExC_parse) {
8392 /* Special Escapes */
8394 RExC_seen_zerolen++;
8395 ret = reg_node(pRExC_state, SBOL);
8397 goto finish_meta_pat;
8399 ret = reg_node(pRExC_state, GPOS);
8400 RExC_seen |= REG_SEEN_GPOS;
8402 goto finish_meta_pat;
8404 RExC_seen_zerolen++;
8405 ret = reg_node(pRExC_state, KEEPS);
8407 /* XXX:dmq : disabling in-place substitution seems to
8408 * be necessary here to avoid cases of memory corruption, as
8409 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
8411 RExC_seen |= REG_SEEN_LOOKBEHIND;
8412 goto finish_meta_pat;
8414 ret = reg_node(pRExC_state, SEOL);
8416 RExC_seen_zerolen++; /* Do not optimize RE away */
8417 goto finish_meta_pat;
8419 ret = reg_node(pRExC_state, EOS);
8421 RExC_seen_zerolen++; /* Do not optimize RE away */
8422 goto finish_meta_pat;
8424 ret = reg_node(pRExC_state, CANY);
8425 RExC_seen |= REG_SEEN_CANY;
8426 *flagp |= HASWIDTH|SIMPLE;
8427 goto finish_meta_pat;
8429 ret = reg_node(pRExC_state, CLUMP);
8431 goto finish_meta_pat;
8433 switch (get_regex_charset(RExC_flags)) {
8434 case REGEX_LOCALE_CHARSET:
8437 case REGEX_UNICODE_CHARSET:
8440 case REGEX_ASCII_RESTRICTED_CHARSET:
8441 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
8444 case REGEX_DEPENDS_CHARSET:
8450 ret = reg_node(pRExC_state, op);
8451 *flagp |= HASWIDTH|SIMPLE;
8452 goto finish_meta_pat;
8454 switch (get_regex_charset(RExC_flags)) {
8455 case REGEX_LOCALE_CHARSET:
8458 case REGEX_UNICODE_CHARSET:
8461 case REGEX_ASCII_RESTRICTED_CHARSET:
8462 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
8465 case REGEX_DEPENDS_CHARSET:
8471 ret = reg_node(pRExC_state, op);
8472 *flagp |= HASWIDTH|SIMPLE;
8473 goto finish_meta_pat;
8475 RExC_seen_zerolen++;
8476 RExC_seen |= REG_SEEN_LOOKBEHIND;
8477 switch (get_regex_charset(RExC_flags)) {
8478 case REGEX_LOCALE_CHARSET:
8481 case REGEX_UNICODE_CHARSET:
8484 case REGEX_ASCII_RESTRICTED_CHARSET:
8485 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
8488 case REGEX_DEPENDS_CHARSET:
8494 ret = reg_node(pRExC_state, op);
8495 FLAGS(ret) = get_regex_charset(RExC_flags);
8497 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
8498 ckWARNregdep(RExC_parse, "\"\\b{\" is deprecated; use \"\\b\\{\" instead");
8500 goto finish_meta_pat;
8502 RExC_seen_zerolen++;
8503 RExC_seen |= REG_SEEN_LOOKBEHIND;
8504 switch (get_regex_charset(RExC_flags)) {
8505 case REGEX_LOCALE_CHARSET:
8508 case REGEX_UNICODE_CHARSET:
8511 case REGEX_ASCII_RESTRICTED_CHARSET:
8512 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
8515 case REGEX_DEPENDS_CHARSET:
8521 ret = reg_node(pRExC_state, op);
8522 FLAGS(ret) = get_regex_charset(RExC_flags);
8524 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
8525 ckWARNregdep(RExC_parse, "\"\\B{\" is deprecated; use \"\\B\\{\" instead");
8527 goto finish_meta_pat;
8529 switch (get_regex_charset(RExC_flags)) {
8530 case REGEX_LOCALE_CHARSET:
8533 case REGEX_UNICODE_CHARSET:
8536 case REGEX_ASCII_RESTRICTED_CHARSET:
8537 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
8540 case REGEX_DEPENDS_CHARSET:
8546 ret = reg_node(pRExC_state, op);
8547 *flagp |= HASWIDTH|SIMPLE;
8548 goto finish_meta_pat;
8550 switch (get_regex_charset(RExC_flags)) {
8551 case REGEX_LOCALE_CHARSET:
8554 case REGEX_UNICODE_CHARSET:
8557 case REGEX_ASCII_RESTRICTED_CHARSET:
8558 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
8561 case REGEX_DEPENDS_CHARSET:
8567 ret = reg_node(pRExC_state, op);
8568 *flagp |= HASWIDTH|SIMPLE;
8569 goto finish_meta_pat;
8571 switch (get_regex_charset(RExC_flags)) {
8572 case REGEX_LOCALE_CHARSET:
8575 case REGEX_ASCII_RESTRICTED_CHARSET:
8576 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
8579 case REGEX_DEPENDS_CHARSET: /* No difference between these */
8580 case REGEX_UNICODE_CHARSET:
8586 ret = reg_node(pRExC_state, op);
8587 *flagp |= HASWIDTH|SIMPLE;
8588 goto finish_meta_pat;
8590 switch (get_regex_charset(RExC_flags)) {
8591 case REGEX_LOCALE_CHARSET:
8594 case REGEX_ASCII_RESTRICTED_CHARSET:
8595 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
8598 case REGEX_DEPENDS_CHARSET: /* No difference between these */
8599 case REGEX_UNICODE_CHARSET:
8605 ret = reg_node(pRExC_state, op);
8606 *flagp |= HASWIDTH|SIMPLE;
8607 goto finish_meta_pat;
8609 ret = reg_node(pRExC_state, LNBREAK);
8610 *flagp |= HASWIDTH|SIMPLE;
8611 goto finish_meta_pat;
8613 ret = reg_node(pRExC_state, HORIZWS);
8614 *flagp |= HASWIDTH|SIMPLE;
8615 goto finish_meta_pat;
8617 ret = reg_node(pRExC_state, NHORIZWS);
8618 *flagp |= HASWIDTH|SIMPLE;
8619 goto finish_meta_pat;
8621 ret = reg_node(pRExC_state, VERTWS);
8622 *flagp |= HASWIDTH|SIMPLE;
8623 goto finish_meta_pat;
8625 ret = reg_node(pRExC_state, NVERTWS);
8626 *flagp |= HASWIDTH|SIMPLE;
8628 nextchar(pRExC_state);
8629 Set_Node_Length(ret, 2); /* MJD */
8634 char* const oldregxend = RExC_end;
8636 char* parse_start = RExC_parse - 2;
8639 if (RExC_parse[1] == '{') {
8640 /* a lovely hack--pretend we saw [\pX] instead */
8641 RExC_end = strchr(RExC_parse, '}');
8643 const U8 c = (U8)*RExC_parse;
8645 RExC_end = oldregxend;
8646 vFAIL2("Missing right brace on \\%c{}", c);
8651 RExC_end = RExC_parse + 2;
8652 if (RExC_end > oldregxend)
8653 RExC_end = oldregxend;
8657 ret = regclass(pRExC_state,depth+1);
8659 RExC_end = oldregxend;
8662 Set_Node_Offset(ret, parse_start + 2);
8663 Set_Node_Cur_Length(ret);
8664 nextchar(pRExC_state);
8665 *flagp |= HASWIDTH|SIMPLE;
8669 /* Handle \N and \N{NAME} here and not below because it can be
8670 multicharacter. join_exact() will join them up later on.
8671 Also this makes sure that things like /\N{BLAH}+/ and
8672 \N{BLAH} being multi char Just Happen. dmq*/
8674 ret= reg_namedseq(pRExC_state, NULL, flagp, depth);
8676 case 'k': /* Handle \k<NAME> and \k'NAME' */
8679 char ch= RExC_parse[1];
8680 if (ch != '<' && ch != '\'' && ch != '{') {
8682 vFAIL2("Sequence %.2s... not terminated",parse_start);
8684 /* this pretty much dupes the code for (?P=...) in reg(), if
8685 you change this make sure you change that */
8686 char* name_start = (RExC_parse += 2);
8688 SV *sv_dat = reg_scan_name(pRExC_state,
8689 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8690 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
8691 if (RExC_parse == name_start || *RExC_parse != ch)
8692 vFAIL2("Sequence %.3s... not terminated",parse_start);
8695 num = add_data( pRExC_state, 1, "S" );
8696 RExC_rxi->data->data[num]=(void*)sv_dat;
8697 SvREFCNT_inc_simple_void(sv_dat);
8701 ret = reganode(pRExC_state,
8704 : (MORE_ASCII_RESTRICTED)
8706 : (AT_LEAST_UNI_SEMANTICS)
8714 /* override incorrect value set in reganode MJD */
8715 Set_Node_Offset(ret, parse_start+1);
8716 Set_Node_Cur_Length(ret); /* MJD */
8717 nextchar(pRExC_state);
8723 case '1': case '2': case '3': case '4':
8724 case '5': case '6': case '7': case '8': case '9':
8727 bool isg = *RExC_parse == 'g';
8732 if (*RExC_parse == '{') {
8736 if (*RExC_parse == '-') {
8740 if (hasbrace && !isDIGIT(*RExC_parse)) {
8741 if (isrel) RExC_parse--;
8743 goto parse_named_seq;
8745 num = atoi(RExC_parse);
8746 if (isg && num == 0)
8747 vFAIL("Reference to invalid group 0");
8749 num = RExC_npar - num;
8751 vFAIL("Reference to nonexistent or unclosed group");
8753 if (!isg && num > 9 && num >= RExC_npar)
8756 char * const parse_start = RExC_parse - 1; /* MJD */
8757 while (isDIGIT(*RExC_parse))
8759 if (parse_start == RExC_parse - 1)
8760 vFAIL("Unterminated \\g... pattern");
8762 if (*RExC_parse != '}')
8763 vFAIL("Unterminated \\g{...} pattern");
8767 if (num > (I32)RExC_rx->nparens)
8768 vFAIL("Reference to nonexistent group");
8771 ret = reganode(pRExC_state,
8774 : (MORE_ASCII_RESTRICTED)
8776 : (AT_LEAST_UNI_SEMANTICS)
8784 /* override incorrect value set in reganode MJD */
8785 Set_Node_Offset(ret, parse_start+1);
8786 Set_Node_Cur_Length(ret); /* MJD */
8788 nextchar(pRExC_state);
8793 if (RExC_parse >= RExC_end)
8794 FAIL("Trailing \\");
8797 /* Do not generate "unrecognized" warnings here, we fall
8798 back into the quick-grab loop below */
8805 if (RExC_flags & RXf_PMf_EXTENDED) {
8806 if ( reg_skipcomment( pRExC_state ) )
8813 parse_start = RExC_parse - 1;
8826 char_state latest_char_state = generic_char;
8827 register STRLEN len;
8832 U8 tmpbuf[UTF8_MAXBYTES_CASE+1], *foldbuf;
8833 regnode * orig_emit;
8836 orig_emit = RExC_emit; /* Save the original output node position in
8837 case we need to output a different node
8839 ret = reg_node(pRExC_state,
8840 (U8) ((! FOLD) ? EXACT
8843 : (MORE_ASCII_RESTRICTED)
8845 : (AT_LEAST_UNI_SEMANTICS)
8850 for (len = 0, p = RExC_parse - 1;
8851 len < 127 && p < RExC_end;
8854 char * const oldp = p;
8856 if (RExC_flags & RXf_PMf_EXTENDED)
8857 p = regwhite( pRExC_state, p );
8868 /* Literal Escapes Switch
8870 This switch is meant to handle escape sequences that
8871 resolve to a literal character.
8873 Every escape sequence that represents something
8874 else, like an assertion or a char class, is handled
8875 in the switch marked 'Special Escapes' above in this
8876 routine, but also has an entry here as anything that
8877 isn't explicitly mentioned here will be treated as
8878 an unescaped equivalent literal.
8882 /* These are all the special escapes. */
8883 case 'A': /* Start assertion */
8884 case 'b': case 'B': /* Word-boundary assertion*/
8885 case 'C': /* Single char !DANGEROUS! */
8886 case 'd': case 'D': /* digit class */
8887 case 'g': case 'G': /* generic-backref, pos assertion */
8888 case 'h': case 'H': /* HORIZWS */
8889 case 'k': case 'K': /* named backref, keep marker */
8890 case 'N': /* named char sequence */
8891 case 'p': case 'P': /* Unicode property */
8892 case 'R': /* LNBREAK */
8893 case 's': case 'S': /* space class */
8894 case 'v': case 'V': /* VERTWS */
8895 case 'w': case 'W': /* word class */
8896 case 'X': /* eXtended Unicode "combining character sequence" */
8897 case 'z': case 'Z': /* End of line/string assertion */
8901 /* Anything after here is an escape that resolves to a
8902 literal. (Except digits, which may or may not)
8921 ender = ASCII_TO_NATIVE('\033');
8925 ender = ASCII_TO_NATIVE('\007');
8930 STRLEN brace_len = len;
8932 const char* error_msg;
8934 bool valid = grok_bslash_o(p,
8941 RExC_parse = p; /* going to die anyway; point
8942 to exact spot of failure */
8949 if (PL_encoding && ender < 0x100) {
8950 goto recode_encoding;
8959 char* const e = strchr(p, '}');
8963 vFAIL("Missing right brace on \\x{}");
8966 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES
8967 | PERL_SCAN_DISALLOW_PREFIX;
8968 STRLEN numlen = e - p - 1;
8969 ender = grok_hex(p + 1, &numlen, &flags, NULL);
8976 I32 flags = PERL_SCAN_DISALLOW_PREFIX;
8978 ender = grok_hex(p, &numlen, &flags, NULL);
8981 if (PL_encoding && ender < 0x100)
8982 goto recode_encoding;
8986 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
8988 case '0': case '1': case '2': case '3':case '4':
8989 case '5': case '6': case '7': case '8':case '9':
8991 (isDIGIT(p[1]) && atoi(p) >= RExC_npar))
8993 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
8995 ender = grok_oct(p, &numlen, &flags, NULL);
9005 if (PL_encoding && ender < 0x100)
9006 goto recode_encoding;
9009 if (! RExC_override_recoding) {
9010 SV* enc = PL_encoding;
9011 ender = reg_recode((const char)(U8)ender, &enc);
9012 if (!enc && SIZE_ONLY)
9013 ckWARNreg(p, "Invalid escape in the specified encoding");
9019 FAIL("Trailing \\");
9022 if (!SIZE_ONLY&& isALPHA(*p)) {
9023 /* Include any { following the alpha to emphasize
9024 * that it could be part of an escape at some point
9026 int len = (*(p + 1) == '{') ? 2 : 1;
9027 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
9029 goto normal_default;
9034 if (UTF8_IS_START(*p) && UTF) {
9036 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
9037 &numlen, UTF8_ALLOW_DEFAULT);
9043 } /* End of switch on the literal */
9045 /* Certain characters are problematic because their folded
9046 * length is so different from their original length that it
9047 * isn't handleable by the optimizer. They are therefore not
9048 * placed in an EXACTish node; and are here handled specially.
9049 * (Even if the optimizer handled LATIN_SMALL_LETTER_SHARP_S,
9050 * putting it in a special node keeps regexec from having to
9051 * deal with a non-utf8 multi-char fold */
9053 && (ender > 255 || (! MORE_ASCII_RESTRICTED && ! LOC)))
9055 /* We look for either side of the fold. For example \xDF
9056 * folds to 'ss'. We look for both the single character
9057 * \xDF and the sequence 'ss'. When we find something that
9058 * could be one of those, we stop and flush whatever we
9059 * have output so far into the EXACTish node that was being
9060 * built. Then restore the input pointer to what it was.
9061 * regatom will return that EXACT node, and will be called
9062 * again, positioned so the first character is the one in
9063 * question, which we return in a different node type.
9064 * The multi-char folds are a sequence, so the occurrence
9065 * of the first character in that sequence doesn't
9066 * necessarily mean that what follows is the rest of the
9067 * sequence. We keep track of that with a state machine,
9068 * with the state being set to the latest character
9069 * processed before the current one. Most characters will
9070 * set the state to 0, but if one occurs that is part of a
9071 * potential tricky fold sequence, the state is set to that
9072 * character, and the next loop iteration sees if the state
9073 * should progress towards the final folded-from character,
9074 * or if it was a false alarm. If it turns out to be a
9075 * false alarm, the character(s) will be output in a new
9076 * EXACTish node, and join_exact() will later combine them.
9077 * In the case of the 'ss' sequence, which is more common
9078 * and more easily checked, some look-ahead is done to
9079 * save time by ruling-out some false alarms */
9082 latest_char_state = generic_char;
9086 case 0x17F: /* LATIN SMALL LETTER LONG S */
9087 if (AT_LEAST_UNI_SEMANTICS) {
9088 if (latest_char_state == char_s) { /* 'ss' */
9089 ender = LATIN_SMALL_LETTER_SHARP_S;
9092 else if (p < RExC_end) {
9094 /* Look-ahead at the next character. If it
9095 * is also an s, we handle as a sharp s
9096 * tricky regnode. */
9097 if (*p == 's' || *p == 'S') {
9099 /* But first flush anything in the
9100 * EXACTish buffer */
9105 p++; /* Account for swallowing this
9107 ender = LATIN_SMALL_LETTER_SHARP_S;
9110 /* Here, the next character is not a
9111 * literal 's', but still could
9112 * evaluate to one if part of a \o{},
9113 * \x or \OCTAL-DIGIT. The minimum
9114 * length required for that is 4, eg
9118 && (isDIGIT(*(p + 1))
9120 || *(p + 1) == 'o' ))
9123 /* Here, it could be an 's', too much
9124 * bother to figure it out here. Flush
9125 * the buffer if any; when come back
9126 * here, set the state so know that the
9127 * previous char was an 's' */
9129 latest_char_state = generic_char;
9133 latest_char_state = char_s;
9139 /* Here, can't be an 'ss' sequence, or at least not
9140 * one that could fold to/from the sharp ss */
9141 latest_char_state = generic_char;
9143 case 0x03C5: /* First char in upsilon series */
9144 case 0x03A5: /* Also capital UPSILON, which folds to
9145 03C5, and hence exhibits the same
9147 if (p < RExC_end - 4) { /* Need >= 4 bytes left */
9148 latest_char_state = upsilon_1;
9155 latest_char_state = generic_char;
9158 case 0x03B9: /* First char in iota series */
9159 case 0x0399: /* Also capital IOTA */
9160 case 0x1FBE: /* GREEK PROSGEGRAMMENI folds to 3B9 */
9161 case 0x0345: /* COMBINING GREEK YPOGEGRAMMENI folds
9163 if (p < RExC_end - 4) {
9164 latest_char_state = iota_1;
9171 latest_char_state = generic_char;
9175 if (latest_char_state == upsilon_1) {
9176 latest_char_state = upsilon_2;
9178 else if (latest_char_state == iota_1) {
9179 latest_char_state = iota_2;
9182 latest_char_state = generic_char;
9186 if (latest_char_state == upsilon_2) {
9187 ender = GREEK_SMALL_LETTER_UPSILON_WITH_DIALYTIKA_AND_TONOS;
9190 else if (latest_char_state == iota_2) {
9191 ender = GREEK_SMALL_LETTER_IOTA_WITH_DIALYTIKA_AND_TONOS;
9194 latest_char_state = generic_char;
9197 /* These are the tricky fold characters. Flush any
9198 * buffer first. (When adding to this list, also should
9199 * add them to fold_grind.t to make sure get tested) */
9200 case GREEK_SMALL_LETTER_UPSILON_WITH_DIALYTIKA_AND_TONOS:
9201 case GREEK_SMALL_LETTER_IOTA_WITH_DIALYTIKA_AND_TONOS:
9202 case LATIN_SMALL_LETTER_SHARP_S:
9203 case LATIN_CAPITAL_LETTER_SHARP_S:
9204 case 0x1FD3: /* GREEK SMALL LETTER IOTA WITH DIALYTIKA AND OXIA */
9205 case 0x1FE3: /* GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND OXIA */
9212 char* const oldregxend = RExC_end;
9213 U8 tmpbuf[UTF8_MAXBYTES+1];
9215 /* Here, we know we need to generate a special
9216 * regnode, and 'ender' contains the tricky
9217 * character. What's done is to pretend it's in a
9218 * [bracketed] class, and let the code that deals
9219 * with those handle it, as that code has all the
9220 * intelligence necessary. First save the current
9221 * parse state, get rid of the already allocated
9222 * but empty EXACT node that the ANYOFV node will
9223 * replace, and point the parse to a buffer which
9224 * we fill with the character we want the regclass
9225 * code to think is being parsed */
9226 RExC_emit = orig_emit;
9227 RExC_parse = (char *) tmpbuf;
9229 U8 *d = uvchr_to_utf8(tmpbuf, ender);
9231 RExC_end = (char *) d;
9233 else { /* ender above 255 already excluded */
9234 tmpbuf[0] = (U8) ender;
9236 RExC_end = RExC_parse + 1;
9239 ret = regclass(pRExC_state,depth+1);
9241 /* Here, have parsed the buffer. Reset the parse to
9242 * the actual input, and return */
9243 RExC_end = oldregxend;
9246 Set_Node_Offset(ret, RExC_parse);
9247 Set_Node_Cur_Length(ret);
9248 nextchar(pRExC_state);
9249 *flagp |= HASWIDTH|SIMPLE;
9255 if ( RExC_flags & RXf_PMf_EXTENDED)
9256 p = regwhite( pRExC_state, p );
9258 /* Prime the casefolded buffer. Locale rules, which apply
9259 * only to code points < 256, aren't known until execution,
9260 * so for them, just output the original character using
9262 if (LOC && ender < 256) {
9263 if (UNI_IS_INVARIANT(ender)) {
9264 *tmpbuf = (U8) ender;
9267 *tmpbuf = UTF8_TWO_BYTE_HI(ender);
9268 *(tmpbuf + 1) = UTF8_TWO_BYTE_LO(ender);
9272 else if (isASCII(ender)) { /* Note: Here can't also be LOC
9274 ender = toLOWER(ender);
9275 *tmpbuf = (U8) ender;
9278 else if (! MORE_ASCII_RESTRICTED && ! LOC) {
9280 /* Locale and /aa require more selectivity about the
9281 * fold, so are handled below. Otherwise, here, just
9283 ender = toFOLD_uni(ender, tmpbuf, &foldlen);
9286 /* Under locale rules or /aa we are not to mix,
9287 * respectively, ords < 256 or ASCII with non-. So
9288 * reject folds that mix them, using only the
9289 * non-folded code point. So do the fold to a
9290 * temporary, and inspect each character in it. */
9291 U8 trialbuf[UTF8_MAXBYTES_CASE+1];
9293 UV tmpender = toFOLD_uni(ender, trialbuf, &foldlen);
9294 U8* e = s + foldlen;
9295 bool fold_ok = TRUE;
9299 || (LOC && (UTF8_IS_INVARIANT(*s)
9300 || UTF8_IS_DOWNGRADEABLE_START(*s))))
9308 Copy(trialbuf, tmpbuf, foldlen, U8);
9312 uvuni_to_utf8(tmpbuf, ender);
9313 foldlen = UNISKIP(ender);
9317 if (p < RExC_end && ISMULT2(p)) { /* Back off on ?+*. */
9322 /* Emit all the Unicode characters. */
9324 for (foldbuf = tmpbuf;
9326 foldlen -= numlen) {
9327 ender = utf8_to_uvchr(foldbuf, &numlen);
9329 const STRLEN unilen = reguni(pRExC_state, ender, s);
9332 /* In EBCDIC the numlen
9333 * and unilen can differ. */
9335 if (numlen >= foldlen)
9339 break; /* "Can't happen." */
9343 const STRLEN unilen = reguni(pRExC_state, ender, s);
9352 REGC((char)ender, s++);
9358 /* Emit all the Unicode characters. */
9360 for (foldbuf = tmpbuf;
9362 foldlen -= numlen) {
9363 ender = utf8_to_uvchr(foldbuf, &numlen);
9365 const STRLEN unilen = reguni(pRExC_state, ender, s);
9368 /* In EBCDIC the numlen
9369 * and unilen can differ. */
9371 if (numlen >= foldlen)
9379 const STRLEN unilen = reguni(pRExC_state, ender, s);
9388 REGC((char)ender, s++);
9391 loopdone: /* Jumped to when encounters something that shouldn't be in
9394 Set_Node_Cur_Length(ret); /* MJD */
9395 nextchar(pRExC_state);
9397 /* len is STRLEN which is unsigned, need to copy to signed */
9400 vFAIL("Internal disaster");
9404 if (len == 1 && UNI_IS_INVARIANT(ender))
9408 RExC_size += STR_SZ(len);
9411 RExC_emit += STR_SZ(len);
9419 /* Jumped to when an unrecognized character set is encountered */
9421 Perl_croak(aTHX_ "panic: Unknown regex character set encoding: %u", get_regex_charset(RExC_flags));
9426 S_regwhite( RExC_state_t *pRExC_state, char *p )
9428 const char *e = RExC_end;
9430 PERL_ARGS_ASSERT_REGWHITE;
9435 else if (*p == '#') {
9444 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
9452 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
9453 Character classes ([:foo:]) can also be negated ([:^foo:]).
9454 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
9455 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
9456 but trigger failures because they are currently unimplemented. */
9458 #define POSIXCC_DONE(c) ((c) == ':')
9459 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
9460 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
9463 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value)
9466 I32 namedclass = OOB_NAMEDCLASS;
9468 PERL_ARGS_ASSERT_REGPPOSIXCC;
9470 if (value == '[' && RExC_parse + 1 < RExC_end &&
9471 /* I smell either [: or [= or [. -- POSIX has been here, right? */
9472 POSIXCC(UCHARAT(RExC_parse))) {
9473 const char c = UCHARAT(RExC_parse);
9474 char* const s = RExC_parse++;
9476 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
9478 if (RExC_parse == RExC_end)
9479 /* Grandfather lone [:, [=, [. */
9482 const char* const t = RExC_parse++; /* skip over the c */
9485 if (UCHARAT(RExC_parse) == ']') {
9486 const char *posixcc = s + 1;
9487 RExC_parse++; /* skip over the ending ] */
9490 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
9491 const I32 skip = t - posixcc;
9493 /* Initially switch on the length of the name. */
9496 if (memEQ(posixcc, "word", 4)) /* this is not POSIX, this is the Perl \w */
9497 namedclass = complement ? ANYOF_NALNUM : ANYOF_ALNUM;
9500 /* Names all of length 5. */
9501 /* alnum alpha ascii blank cntrl digit graph lower
9502 print punct space upper */
9503 /* Offset 4 gives the best switch position. */
9504 switch (posixcc[4]) {
9506 if (memEQ(posixcc, "alph", 4)) /* alpha */
9507 namedclass = complement ? ANYOF_NALPHA : ANYOF_ALPHA;
9510 if (memEQ(posixcc, "spac", 4)) /* space */
9511 namedclass = complement ? ANYOF_NPSXSPC : ANYOF_PSXSPC;
9514 if (memEQ(posixcc, "grap", 4)) /* graph */
9515 namedclass = complement ? ANYOF_NGRAPH : ANYOF_GRAPH;
9518 if (memEQ(posixcc, "asci", 4)) /* ascii */
9519 namedclass = complement ? ANYOF_NASCII : ANYOF_ASCII;
9522 if (memEQ(posixcc, "blan", 4)) /* blank */
9523 namedclass = complement ? ANYOF_NBLANK : ANYOF_BLANK;
9526 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
9527 namedclass = complement ? ANYOF_NCNTRL : ANYOF_CNTRL;
9530 if (memEQ(posixcc, "alnu", 4)) /* alnum */
9531 namedclass = complement ? ANYOF_NALNUMC : ANYOF_ALNUMC;
9534 if (memEQ(posixcc, "lowe", 4)) /* lower */
9535 namedclass = complement ? ANYOF_NLOWER : ANYOF_LOWER;
9536 else if (memEQ(posixcc, "uppe", 4)) /* upper */
9537 namedclass = complement ? ANYOF_NUPPER : ANYOF_UPPER;
9540 if (memEQ(posixcc, "digi", 4)) /* digit */
9541 namedclass = complement ? ANYOF_NDIGIT : ANYOF_DIGIT;
9542 else if (memEQ(posixcc, "prin", 4)) /* print */
9543 namedclass = complement ? ANYOF_NPRINT : ANYOF_PRINT;
9544 else if (memEQ(posixcc, "punc", 4)) /* punct */
9545 namedclass = complement ? ANYOF_NPUNCT : ANYOF_PUNCT;
9550 if (memEQ(posixcc, "xdigit", 6))
9551 namedclass = complement ? ANYOF_NXDIGIT : ANYOF_XDIGIT;
9555 if (namedclass == OOB_NAMEDCLASS)
9556 Simple_vFAIL3("POSIX class [:%.*s:] unknown",
9558 assert (posixcc[skip] == ':');
9559 assert (posixcc[skip+1] == ']');
9560 } else if (!SIZE_ONLY) {
9561 /* [[=foo=]] and [[.foo.]] are still future. */
9563 /* adjust RExC_parse so the warning shows after
9565 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
9567 Simple_vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
9570 /* Maternal grandfather:
9571 * "[:" ending in ":" but not in ":]" */
9581 S_checkposixcc(pTHX_ RExC_state_t *pRExC_state)
9585 PERL_ARGS_ASSERT_CHECKPOSIXCC;
9587 if (POSIXCC(UCHARAT(RExC_parse))) {
9588 const char *s = RExC_parse;
9589 const char c = *s++;
9593 if (*s && c == *s && s[1] == ']') {
9595 "POSIX syntax [%c %c] belongs inside character classes",
9598 /* [[=foo=]] and [[.foo.]] are still future. */
9599 if (POSIXCC_NOTYET(c)) {
9600 /* adjust RExC_parse so the error shows after
9602 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse++) != ']')
9604 Simple_vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
9610 /* No locale test, and always Unicode semantics, no ignore-case differences */
9611 #define _C_C_T_NOLOC_(NAME,TEST,WORD) \
9613 for (value = 0; value < 256; value++) \
9615 stored += set_regclass_bit(pRExC_state, ret, (U8) value, &l1_fold_invlist, &unicode_alternate); \
9619 case ANYOF_N##NAME: \
9620 for (value = 0; value < 256; value++) \
9622 stored += set_regclass_bit(pRExC_state, ret, (U8) value, &l1_fold_invlist, &unicode_alternate); \
9627 /* Like the above, but there are differences if we are in uni-8-bit or not, so
9628 * there are two tests passed in, to use depending on that. There aren't any
9629 * cases where the label is different from the name, so no need for that
9631 * Sets 'what' to WORD which is the property name for non-bitmap code points;
9632 * But, uses FOLD_WORD instead if /i has been selected, to allow a different
9634 #define _C_C_T_(NAME, TEST_8, TEST_7, WORD, FOLD_WORD) \
9636 if (LOC) ANYOF_CLASS_SET(ret, ANYOF_##NAME); \
9637 else if (UNI_SEMANTICS) { \
9638 for (value = 0; value < 256; value++) { \
9639 if (TEST_8(value)) stored += \
9640 set_regclass_bit(pRExC_state, ret, (U8) value, &l1_fold_invlist, &unicode_alternate); \
9644 for (value = 0; value < 128; value++) { \
9645 if (TEST_7(UNI_TO_NATIVE(value))) stored += \
9646 set_regclass_bit(pRExC_state, ret, \
9647 (U8) UNI_TO_NATIVE(value), &l1_fold_invlist, &unicode_alternate); \
9658 case ANYOF_N##NAME: \
9659 if (LOC) ANYOF_CLASS_SET(ret, ANYOF_N##NAME); \
9660 else if (UNI_SEMANTICS) { \
9661 for (value = 0; value < 256; value++) { \
9662 if (! TEST_8(value)) stored += \
9663 set_regclass_bit(pRExC_state, ret, (U8) value, &l1_fold_invlist, &unicode_alternate); \
9667 for (value = 0; value < 128; value++) { \
9668 if (! TEST_7(UNI_TO_NATIVE(value))) stored += set_regclass_bit( \
9669 pRExC_state, ret, (U8) UNI_TO_NATIVE(value), &l1_fold_invlist, &unicode_alternate); \
9671 if (AT_LEAST_ASCII_RESTRICTED) { \
9672 for (value = 128; value < 256; value++) { \
9673 stored += set_regclass_bit( \
9674 pRExC_state, ret, (U8) UNI_TO_NATIVE(value), &l1_fold_invlist, &unicode_alternate); \
9676 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL; \
9679 /* For a non-ut8 target string with DEPENDS semantics, all above \
9680 * ASCII Latin1 code points match the complement of any of the \
9681 * classes. But in utf8, they have their Unicode semantics, so \
9682 * can't just set them in the bitmap, or else regexec.c will think \
9683 * they matched when they shouldn't. */ \
9684 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL; \
9697 S_set_regclass_bit_fold(pTHX_ RExC_state_t *pRExC_state, regnode* node, const U8 value, SV** invlist_ptr, AV** alternate_ptr)
9700 /* Handle the setting of folds in the bitmap for non-locale ANYOF nodes.
9701 * Locale folding is done at run-time, so this function should not be
9702 * called for nodes that are for locales.
9704 * This function sets the bit corresponding to the fold of the input
9705 * 'value', if not already set. The fold of 'f' is 'F', and the fold of
9708 * It also knows about the characters that are in the bitmap that have
9709 * folds that are matchable only outside it, and sets the appropriate lists
9712 * It returns the number of bits that actually changed from 0 to 1 */
9717 PERL_ARGS_ASSERT_SET_REGCLASS_BIT_FOLD;
9719 fold = (AT_LEAST_UNI_SEMANTICS) ? PL_fold_latin1[value]
9722 /* It assumes the bit for 'value' has already been set */
9723 if (fold != value && ! ANYOF_BITMAP_TEST(node, fold)) {
9724 ANYOF_BITMAP_SET(node, fold);
9727 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(value) && (! isASCII(value) || ! MORE_ASCII_RESTRICTED)) {
9728 /* Certain Latin1 characters have matches outside the bitmap. To get
9729 * here, 'value' is one of those characters. None of these matches is
9730 * valid for ASCII characters under /aa, which have been excluded by
9731 * the 'if' above. The matches fall into three categories:
9732 * 1) They are singly folded-to or -from an above 255 character, as
9733 * LATIN SMALL LETTER Y WITH DIAERESIS and LATIN CAPITAL LETTER Y
9735 * 2) They are part of a multi-char fold with another character in the
9736 * bitmap, only LATIN SMALL LETTER SHARP S => "ss" fits that bill;
9737 * 3) They are part of a multi-char fold with a character not in the
9738 * bitmap, such as various ligatures.
9739 * We aren't dealing fully with multi-char folds, except we do deal
9740 * with the pattern containing a character that has a multi-char fold
9741 * (not so much the inverse).
9742 * For types 1) and 3), the matches only happen when the target string
9743 * is utf8; that's not true for 2), and we set a flag for it.
9745 * The code below adds to the passed in inversion list the single fold
9746 * closures for 'value'. The values are hard-coded here so that an
9747 * innocent-looking character class, like /[ks]/i won't have to go out
9748 * to disk to find the possible matches. XXX It would be better to
9749 * generate these via regen, in case a new version of the Unicode
9750 * standard adds new mappings, though that is not really likely. */
9755 *invlist_ptr = add_cp_to_invlist(*invlist_ptr, 0x212A);
9759 /* LATIN SMALL LETTER LONG S */
9760 *invlist_ptr = add_cp_to_invlist(*invlist_ptr, 0x017F);
9763 *invlist_ptr = add_cp_to_invlist(*invlist_ptr,
9764 GREEK_SMALL_LETTER_MU);
9765 *invlist_ptr = add_cp_to_invlist(*invlist_ptr,
9766 GREEK_CAPITAL_LETTER_MU);
9768 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
9769 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
9771 *invlist_ptr = add_cp_to_invlist(*invlist_ptr, 0x212B);
9772 if (DEPENDS_SEMANTICS) { /* See DEPENDS comment below */
9773 *invlist_ptr = add_cp_to_invlist(*invlist_ptr,
9774 PL_fold_latin1[value]);
9777 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
9778 *invlist_ptr = add_cp_to_invlist(*invlist_ptr,
9779 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
9781 case LATIN_SMALL_LETTER_SHARP_S:
9782 *invlist_ptr = add_cp_to_invlist(*invlist_ptr,
9783 LATIN_CAPITAL_LETTER_SHARP_S);
9785 /* Under /a, /d, and /u, this can match the two chars "ss" */
9786 if (! MORE_ASCII_RESTRICTED) {
9787 add_alternate(alternate_ptr, (U8 *) "ss", 2);
9789 /* And under /u or /a, it can match even if the target is
9791 if (AT_LEAST_UNI_SEMANTICS) {
9792 ANYOF_FLAGS(node) |= ANYOF_NONBITMAP_NON_UTF8;
9806 /* These all are targets of multi-character folds from code
9807 * points that require UTF8 to express, so they can't match
9808 * unless the target string is in UTF-8, so no action here is
9809 * necessary, as regexec.c properly handles the general case
9810 * for UTF-8 matching */
9813 /* Use deprecated warning to increase the chances of this
9815 ckWARN2regdep(RExC_parse, "Perl folding rules are not up-to-date for 0x%x; please use the perlbug utility to report;", value);
9819 else if (DEPENDS_SEMANTICS
9821 && PL_fold_latin1[value] != value)
9823 /* Under DEPENDS rules, non-ASCII Latin1 characters match their
9824 * folds only when the target string is in UTF-8. We add the fold
9825 * here to the list of things to match outside the bitmap, which
9826 * won't be looked at unless it is UTF8 (or else if something else
9827 * says to look even if not utf8, but those things better not happen
9828 * under DEPENDS semantics. */
9829 *invlist_ptr = add_cp_to_invlist(*invlist_ptr, PL_fold_latin1[value]);
9836 PERL_STATIC_INLINE U8
9837 S_set_regclass_bit(pTHX_ RExC_state_t *pRExC_state, regnode* node, const U8 value, SV** invlist_ptr, AV** alternate_ptr)
9839 /* This inline function sets a bit in the bitmap if not already set, and if
9840 * appropriate, its fold, returning the number of bits that actually
9841 * changed from 0 to 1 */
9845 PERL_ARGS_ASSERT_SET_REGCLASS_BIT;
9847 if (ANYOF_BITMAP_TEST(node, value)) { /* Already set */
9851 ANYOF_BITMAP_SET(node, value);
9854 if (FOLD && ! LOC) { /* Locale folds aren't known until runtime */
9855 stored += set_regclass_bit_fold(pRExC_state, node, value, invlist_ptr, alternate_ptr);
9862 S_add_alternate(pTHX_ AV** alternate_ptr, U8* string, STRLEN len)
9864 /* Adds input 'string' with length 'len' to the ANYOF node's unicode
9865 * alternate list, pointed to by 'alternate_ptr'. This is an array of
9866 * the multi-character folds of characters in the node */
9869 PERL_ARGS_ASSERT_ADD_ALTERNATE;
9871 if (! *alternate_ptr) {
9872 *alternate_ptr = newAV();
9874 sv = newSVpvn_utf8((char*)string, len, TRUE);
9875 av_push(*alternate_ptr, sv);
9880 parse a class specification and produce either an ANYOF node that
9881 matches the pattern or perhaps will be optimized into an EXACTish node
9882 instead. The node contains a bit map for the first 256 characters, with the
9883 corresponding bit set if that character is in the list. For characters
9884 above 255, a range list is used */
9887 S_regclass(pTHX_ RExC_state_t *pRExC_state, U32 depth)
9890 register UV nextvalue;
9891 register IV prevvalue = OOB_UNICODE;
9892 register IV range = 0;
9893 UV value = 0; /* XXX:dmq: needs to be referenceable (unfortunately) */
9894 register regnode *ret;
9897 char *rangebegin = NULL;
9898 bool need_class = 0;
9899 bool allow_full_fold = TRUE; /* Assume wants multi-char folding */
9901 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
9902 than just initialized. */
9905 /* code points this node matches that can't be stored in the bitmap */
9906 SV* nonbitmap = NULL;
9908 /* The items that are to match that aren't stored in the bitmap, but are a
9909 * result of things that are stored there. This is the fold closure of
9910 * such a character, either because it has DEPENDS semantics and shouldn't
9911 * be matched unless the target string is utf8, or is a code point that is
9912 * too large for the bit map, as for example, the fold of the MICRO SIGN is
9913 * above 255. This all is solely for performance reasons. By having this
9914 * code know the outside-the-bitmap folds that the bitmapped characters are
9915 * involved with, we don't have to go out to disk to find the list of
9916 * matches, unless the character class includes code points that aren't
9917 * storable in the bit map. That means that a character class with an 's'
9918 * in it, for example, doesn't need to go out to disk to find everything
9919 * that matches. A 2nd list is used so that the 'nonbitmap' list is kept
9920 * empty unless there is something whose fold we don't know about, and will
9921 * have to go out to the disk to find. */
9922 SV* l1_fold_invlist = NULL;
9924 /* List of multi-character folds that are matched by this node */
9925 AV* unicode_alternate = NULL;
9927 UV literal_endpoint = 0;
9929 UV stored = 0; /* how many chars stored in the bitmap */
9931 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
9932 case we need to change the emitted regop to an EXACT. */
9933 const char * orig_parse = RExC_parse;
9934 GET_RE_DEBUG_FLAGS_DECL;
9936 PERL_ARGS_ASSERT_REGCLASS;
9938 PERL_UNUSED_ARG(depth);
9941 DEBUG_PARSE("clas");
9943 /* Assume we are going to generate an ANYOF node. */
9944 ret = reganode(pRExC_state, ANYOF, 0);
9948 ANYOF_FLAGS(ret) = 0;
9951 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
9955 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
9957 /* We have decided to not allow multi-char folds in inverted character
9958 * classes, due to the confusion that can happen, especially with
9959 * classes that are designed for a non-Unicode world: You have the
9960 * peculiar case that:
9961 "s s" =~ /^[^\xDF]+$/i => Y
9962 "ss" =~ /^[^\xDF]+$/i => N
9964 * See [perl #89750] */
9965 allow_full_fold = FALSE;
9969 RExC_size += ANYOF_SKIP;
9970 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
9973 RExC_emit += ANYOF_SKIP;
9975 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
9977 ANYOF_BITMAP_ZERO(ret);
9978 listsv = newSVpvs("# comment\n");
9979 initial_listsv_len = SvCUR(listsv);
9982 nextvalue = RExC_parse < RExC_end ? UCHARAT(RExC_parse) : 0;
9984 if (!SIZE_ONLY && POSIXCC(nextvalue))
9985 checkposixcc(pRExC_state);
9987 /* allow 1st char to be ] (allowing it to be - is dealt with later) */
9988 if (UCHARAT(RExC_parse) == ']')
9992 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
9996 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
9999 rangebegin = RExC_parse;
10001 value = utf8n_to_uvchr((U8*)RExC_parse,
10002 RExC_end - RExC_parse,
10003 &numlen, UTF8_ALLOW_DEFAULT);
10004 RExC_parse += numlen;
10007 value = UCHARAT(RExC_parse++);
10009 nextvalue = RExC_parse < RExC_end ? UCHARAT(RExC_parse) : 0;
10010 if (value == '[' && POSIXCC(nextvalue))
10011 namedclass = regpposixcc(pRExC_state, value);
10012 else if (value == '\\') {
10014 value = utf8n_to_uvchr((U8*)RExC_parse,
10015 RExC_end - RExC_parse,
10016 &numlen, UTF8_ALLOW_DEFAULT);
10017 RExC_parse += numlen;
10020 value = UCHARAT(RExC_parse++);
10021 /* Some compilers cannot handle switching on 64-bit integer
10022 * values, therefore value cannot be an UV. Yes, this will
10023 * be a problem later if we want switch on Unicode.
10024 * A similar issue a little bit later when switching on
10025 * namedclass. --jhi */
10026 switch ((I32)value) {
10027 case 'w': namedclass = ANYOF_ALNUM; break;
10028 case 'W': namedclass = ANYOF_NALNUM; break;
10029 case 's': namedclass = ANYOF_SPACE; break;
10030 case 'S': namedclass = ANYOF_NSPACE; break;
10031 case 'd': namedclass = ANYOF_DIGIT; break;
10032 case 'D': namedclass = ANYOF_NDIGIT; break;
10033 case 'v': namedclass = ANYOF_VERTWS; break;
10034 case 'V': namedclass = ANYOF_NVERTWS; break;
10035 case 'h': namedclass = ANYOF_HORIZWS; break;
10036 case 'H': namedclass = ANYOF_NHORIZWS; break;
10037 case 'N': /* Handle \N{NAME} in class */
10039 /* We only pay attention to the first char of
10040 multichar strings being returned. I kinda wonder
10041 if this makes sense as it does change the behaviour
10042 from earlier versions, OTOH that behaviour was broken
10044 UV v; /* value is register so we cant & it /grrr */
10045 if (reg_namedseq(pRExC_state, &v, NULL, depth)) {
10055 if (RExC_parse >= RExC_end)
10056 vFAIL2("Empty \\%c{}", (U8)value);
10057 if (*RExC_parse == '{') {
10058 const U8 c = (U8)value;
10059 e = strchr(RExC_parse++, '}');
10061 vFAIL2("Missing right brace on \\%c{}", c);
10062 while (isSPACE(UCHARAT(RExC_parse)))
10064 if (e == RExC_parse)
10065 vFAIL2("Empty \\%c{}", c);
10066 n = e - RExC_parse;
10067 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
10075 if (UCHARAT(RExC_parse) == '^') {
10078 value = value == 'p' ? 'P' : 'p'; /* toggle */
10079 while (isSPACE(UCHARAT(RExC_parse))) {
10085 /* Add the property name to the list. If /i matching, give
10086 * a different name which consists of the normal name
10087 * sandwiched between two underscores and '_i'. The design
10088 * is discussed in the commit message for this. */
10089 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s%.*s%s\n",
10090 (value=='p' ? '+' : '!'),
10091 (FOLD) ? "__" : "",
10097 RExC_parse = e + 1;
10099 /* The \p could match something in the Latin1 range, hence
10100 * something that isn't utf8 */
10101 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
10102 namedclass = ANYOF_MAX; /* no official name, but it's named */
10104 /* \p means they want Unicode semantics */
10105 RExC_uni_semantics = 1;
10108 case 'n': value = '\n'; break;
10109 case 'r': value = '\r'; break;
10110 case 't': value = '\t'; break;
10111 case 'f': value = '\f'; break;
10112 case 'b': value = '\b'; break;
10113 case 'e': value = ASCII_TO_NATIVE('\033');break;
10114 case 'a': value = ASCII_TO_NATIVE('\007');break;
10116 RExC_parse--; /* function expects to be pointed at the 'o' */
10118 const char* error_msg;
10119 bool valid = grok_bslash_o(RExC_parse,
10124 RExC_parse += numlen;
10129 if (PL_encoding && value < 0x100) {
10130 goto recode_encoding;
10134 if (*RExC_parse == '{') {
10135 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES
10136 | PERL_SCAN_DISALLOW_PREFIX;
10137 char * const e = strchr(RExC_parse++, '}');
10139 vFAIL("Missing right brace on \\x{}");
10141 numlen = e - RExC_parse;
10142 value = grok_hex(RExC_parse, &numlen, &flags, NULL);
10143 RExC_parse = e + 1;
10146 I32 flags = PERL_SCAN_DISALLOW_PREFIX;
10148 value = grok_hex(RExC_parse, &numlen, &flags, NULL);
10149 RExC_parse += numlen;
10151 if (PL_encoding && value < 0x100)
10152 goto recode_encoding;
10155 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
10157 case '0': case '1': case '2': case '3': case '4':
10158 case '5': case '6': case '7':
10160 /* Take 1-3 octal digits */
10161 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
10163 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
10164 RExC_parse += numlen;
10165 if (PL_encoding && value < 0x100)
10166 goto recode_encoding;
10170 if (! RExC_override_recoding) {
10171 SV* enc = PL_encoding;
10172 value = reg_recode((const char)(U8)value, &enc);
10173 if (!enc && SIZE_ONLY)
10174 ckWARNreg(RExC_parse,
10175 "Invalid escape in the specified encoding");
10179 /* Allow \_ to not give an error */
10180 if (!SIZE_ONLY && isALNUM(value) && value != '_') {
10181 ckWARN2reg(RExC_parse,
10182 "Unrecognized escape \\%c in character class passed through",
10187 } /* end of \blah */
10190 literal_endpoint++;
10193 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
10195 /* What matches in a locale is not known until runtime, so need to
10196 * (one time per class) allocate extra space to pass to regexec.
10197 * The space will contain a bit for each named class that is to be
10198 * matched against. This isn't needed for \p{} and pseudo-classes,
10199 * as they are not affected by locale, and hence are dealt with
10201 if (LOC && namedclass < ANYOF_MAX && ! need_class) {
10204 RExC_size += ANYOF_CLASS_SKIP - ANYOF_SKIP;
10207 RExC_emit += ANYOF_CLASS_SKIP - ANYOF_SKIP;
10208 ANYOF_CLASS_ZERO(ret);
10210 ANYOF_FLAGS(ret) |= ANYOF_CLASS;
10213 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
10214 * literal, as is the character that began the false range, i.e.
10215 * the 'a' in the examples */
10219 RExC_parse >= rangebegin ?
10220 RExC_parse - rangebegin : 0;
10221 ckWARN4reg(RExC_parse,
10222 "False [] range \"%*.*s\"",
10226 set_regclass_bit(pRExC_state, ret, '-', &l1_fold_invlist, &unicode_alternate);
10227 if (prevvalue < 256) {
10229 set_regclass_bit(pRExC_state, ret, (U8) prevvalue, &l1_fold_invlist, &unicode_alternate);
10232 nonbitmap = add_cp_to_invlist(nonbitmap, prevvalue);
10236 range = 0; /* this was not a true range */
10242 const char *what = NULL;
10245 /* Possible truncation here but in some 64-bit environments
10246 * the compiler gets heartburn about switch on 64-bit values.
10247 * A similar issue a little earlier when switching on value.
10249 switch ((I32)namedclass) {
10251 case _C_C_T_(ALNUMC, isALNUMC_L1, isALNUMC, "XPosixAlnum", "XPosixAlnum");
10252 case _C_C_T_(ALPHA, isALPHA_L1, isALPHA, "XPosixAlpha", "XPosixAlpha");
10253 case _C_C_T_(BLANK, isBLANK_L1, isBLANK, "XPosixBlank", "XPosixBlank");
10254 case _C_C_T_(CNTRL, isCNTRL_L1, isCNTRL, "XPosixCntrl", "XPosixCntrl");
10255 case _C_C_T_(GRAPH, isGRAPH_L1, isGRAPH, "XPosixGraph", "XPosixGraph");
10256 case _C_C_T_(LOWER, isLOWER_L1, isLOWER, "XPosixLower", "__XPosixLower_i");
10257 case _C_C_T_(PRINT, isPRINT_L1, isPRINT, "XPosixPrint", "XPosixPrint");
10258 case _C_C_T_(PSXSPC, isPSXSPC_L1, isPSXSPC, "XPosixSpace", "XPosixSpace");
10259 case _C_C_T_(PUNCT, isPUNCT_L1, isPUNCT, "XPosixPunct", "XPosixPunct");
10260 case _C_C_T_(UPPER, isUPPER_L1, isUPPER, "XPosixUpper", "__XPosixUpper_i");
10261 /* \s, \w match all unicode if utf8. */
10262 case _C_C_T_(SPACE, isSPACE_L1, isSPACE, "SpacePerl", "SpacePerl");
10263 case _C_C_T_(ALNUM, isWORDCHAR_L1, isALNUM, "Word", "Word");
10264 case _C_C_T_(XDIGIT, isXDIGIT_L1, isXDIGIT, "XPosixXDigit", "XPosixXDigit");
10265 case _C_C_T_NOLOC_(VERTWS, is_VERTWS_latin1(&value), "VertSpace");
10266 case _C_C_T_NOLOC_(HORIZWS, is_HORIZWS_latin1(&value), "HorizSpace");
10269 ANYOF_CLASS_SET(ret, ANYOF_ASCII);
10271 for (value = 0; value < 128; value++)
10273 set_regclass_bit(pRExC_state, ret, (U8) ASCII_TO_NATIVE(value), &l1_fold_invlist, &unicode_alternate);
10276 what = NULL; /* Doesn't match outside ascii, so
10277 don't want to add +utf8:: */
10281 ANYOF_CLASS_SET(ret, ANYOF_NASCII);
10283 for (value = 128; value < 256; value++)
10285 set_regclass_bit(pRExC_state, ret, (U8) ASCII_TO_NATIVE(value), &l1_fold_invlist, &unicode_alternate);
10287 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
10293 ANYOF_CLASS_SET(ret, ANYOF_DIGIT);
10295 /* consecutive digits assumed */
10296 for (value = '0'; value <= '9'; value++)
10298 set_regclass_bit(pRExC_state, ret, (U8) value, &l1_fold_invlist, &unicode_alternate);
10305 ANYOF_CLASS_SET(ret, ANYOF_NDIGIT);
10307 /* consecutive digits assumed */
10308 for (value = 0; value < '0'; value++)
10310 set_regclass_bit(pRExC_state, ret, (U8) value, &l1_fold_invlist, &unicode_alternate);
10311 for (value = '9' + 1; value < 256; value++)
10313 set_regclass_bit(pRExC_state, ret, (U8) value, &l1_fold_invlist, &unicode_alternate);
10317 if (AT_LEAST_ASCII_RESTRICTED ) {
10318 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
10322 /* this is to handle \p and \P */
10325 vFAIL("Invalid [::] class");
10328 if (what && ! (AT_LEAST_ASCII_RESTRICTED)) {
10329 /* Strings such as "+utf8::isWord\n" */
10330 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s\n", yesno, what);
10335 } /* end of namedclass \blah */
10338 if (prevvalue > (IV)value) /* b-a */ {
10339 const int w = RExC_parse - rangebegin;
10340 Simple_vFAIL4("Invalid [] range \"%*.*s\"", w, w, rangebegin);
10341 range = 0; /* not a valid range */
10345 prevvalue = value; /* save the beginning of the range */
10346 if (RExC_parse+1 < RExC_end
10347 && *RExC_parse == '-'
10348 && RExC_parse[1] != ']')
10352 /* a bad range like \w-, [:word:]- ? */
10353 if (namedclass > OOB_NAMEDCLASS) {
10354 if (ckWARN(WARN_REGEXP)) {
10356 RExC_parse >= rangebegin ?
10357 RExC_parse - rangebegin : 0;
10359 "False [] range \"%*.*s\"",
10364 set_regclass_bit(pRExC_state, ret, '-', &l1_fold_invlist, &unicode_alternate);
10366 range = 1; /* yeah, it's a range! */
10367 continue; /* but do it the next time */
10371 /* non-Latin1 code point implies unicode semantics. Must be set in
10372 * pass1 so is there for the whole of pass 2 */
10374 RExC_uni_semantics = 1;
10377 /* now is the next time */
10379 if (prevvalue < 256) {
10380 const IV ceilvalue = value < 256 ? value : 255;
10383 /* In EBCDIC [\x89-\x91] should include
10384 * the \x8e but [i-j] should not. */
10385 if (literal_endpoint == 2 &&
10386 ((isLOWER(prevvalue) && isLOWER(ceilvalue)) ||
10387 (isUPPER(prevvalue) && isUPPER(ceilvalue))))
10389 if (isLOWER(prevvalue)) {
10390 for (i = prevvalue; i <= ceilvalue; i++)
10391 if (isLOWER(i) && !ANYOF_BITMAP_TEST(ret,i)) {
10393 set_regclass_bit(pRExC_state, ret, (U8) i, &l1_fold_invlist, &unicode_alternate);
10396 for (i = prevvalue; i <= ceilvalue; i++)
10397 if (isUPPER(i) && !ANYOF_BITMAP_TEST(ret,i)) {
10399 set_regclass_bit(pRExC_state, ret, (U8) i, &l1_fold_invlist, &unicode_alternate);
10405 for (i = prevvalue; i <= ceilvalue; i++) {
10406 stored += set_regclass_bit(pRExC_state, ret, (U8) i, &l1_fold_invlist, &unicode_alternate);
10410 const UV prevnatvalue = NATIVE_TO_UNI(prevvalue);
10411 const UV natvalue = NATIVE_TO_UNI(value);
10412 nonbitmap = add_range_to_invlist(nonbitmap, prevnatvalue, natvalue);
10415 literal_endpoint = 0;
10419 range = 0; /* this range (if it was one) is done now */
10426 /****** !SIZE_ONLY AFTER HERE *********/
10428 /* If folding and there are code points above 255, we calculate all
10429 * characters that could fold to or from the ones already on the list */
10430 if (FOLD && nonbitmap) {
10431 UV start, end; /* End points of code point ranges */
10433 SV* fold_intersection;
10435 /* This is a list of all the characters that participate in folds
10436 * (except marks, etc in multi-char folds */
10437 if (! PL_utf8_foldable) {
10438 SV* swash = swash_init("utf8", "Cased", &PL_sv_undef, 1, 0);
10439 PL_utf8_foldable = _swash_to_invlist(swash);
10442 /* This is a hash that for a particular fold gives all characters
10443 * that are involved in it */
10444 if (! PL_utf8_foldclosures) {
10446 /* If we were unable to find any folds, then we likely won't be
10447 * able to find the closures. So just create an empty list.
10448 * Folding will effectively be restricted to the non-Unicode rules
10449 * hard-coded into Perl. (This case happens legitimately during
10450 * compilation of Perl itself before the Unicode tables are
10452 if (invlist_len(PL_utf8_foldable) == 0) {
10453 PL_utf8_foldclosures = newHV();
10455 /* If the folds haven't been read in, call a fold function
10457 if (! PL_utf8_tofold) {
10458 U8 dummy[UTF8_MAXBYTES+1];
10461 /* This particular string is above \xff in both UTF-8 and
10463 to_utf8_fold((U8*) "\xC8\x80", dummy, &dummy_len);
10464 assert(PL_utf8_tofold); /* Verify that worked */
10466 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
10470 /* Only the characters in this class that participate in folds need
10471 * be checked. Get the intersection of this class and all the
10472 * possible characters that are foldable. This can quickly narrow
10473 * down a large class */
10474 _invlist_intersection(PL_utf8_foldable, nonbitmap, &fold_intersection);
10476 /* Now look at the foldable characters in this class individually */
10477 invlist_iterinit(fold_intersection);
10478 while (invlist_iternext(fold_intersection, &start, &end)) {
10481 /* Look at every character in the range */
10482 for (j = start; j <= end; j++) {
10485 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
10488 _to_uni_fold_flags(j, foldbuf, &foldlen, allow_full_fold);
10490 if (foldlen > (STRLEN)UNISKIP(f)) {
10492 /* Any multicharacter foldings (disallowed in
10493 * lookbehind patterns) require the following
10494 * transform: [ABCDEF] -> (?:[ABCabcDEFd]|pq|rst) where
10495 * E folds into "pq" and F folds into "rst", all other
10496 * characters fold to single characters. We save away
10497 * these multicharacter foldings, to be later saved as
10498 * part of the additional "s" data. */
10499 if (! RExC_in_lookbehind) {
10501 U8* e = foldbuf + foldlen;
10503 /* If any of the folded characters of this are in
10504 * the Latin1 range, tell the regex engine that
10505 * this can match a non-utf8 target string. The
10506 * only multi-byte fold whose source is in the
10507 * Latin1 range (U+00DF) applies only when the
10508 * target string is utf8, or under unicode rules */
10509 if (j > 255 || AT_LEAST_UNI_SEMANTICS) {
10512 /* Can't mix ascii with non- under /aa */
10513 if (MORE_ASCII_RESTRICTED
10514 && (isASCII(*loc) != isASCII(j)))
10516 goto end_multi_fold;
10518 if (UTF8_IS_INVARIANT(*loc)
10519 || UTF8_IS_DOWNGRADEABLE_START(*loc))
10521 /* Can't mix above and below 256 under
10524 goto end_multi_fold;
10527 |= ANYOF_NONBITMAP_NON_UTF8;
10530 loc += UTF8SKIP(loc);
10534 add_alternate(&unicode_alternate, foldbuf, foldlen);
10538 /* This is special-cased, as it is the only letter which
10539 * has both a multi-fold and single-fold in Latin1. All
10540 * the other chars that have single and multi-folds are
10541 * always in utf8, and the utf8 folding algorithm catches
10543 if (! LOC && j == LATIN_CAPITAL_LETTER_SHARP_S) {
10544 stored += set_regclass_bit(pRExC_state,
10546 LATIN_SMALL_LETTER_SHARP_S,
10547 &l1_fold_invlist, &unicode_alternate);
10551 /* Single character fold. Add everything in its fold
10552 * closure to the list that this node should match */
10555 /* The fold closures data structure is a hash with the
10556 * keys being every character that is folded to, like
10557 * 'k', and the values each an array of everything that
10558 * folds to its key. e.g. [ 'k', 'K', KELVIN_SIGN ] */
10559 if ((listp = hv_fetch(PL_utf8_foldclosures,
10560 (char *) foldbuf, foldlen, FALSE)))
10562 AV* list = (AV*) *listp;
10564 for (k = 0; k <= av_len(list); k++) {
10565 SV** c_p = av_fetch(list, k, FALSE);
10568 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
10572 /* /aa doesn't allow folds between ASCII and
10573 * non-; /l doesn't allow them between above
10575 if ((MORE_ASCII_RESTRICTED
10576 && (isASCII(c) != isASCII(j)))
10577 || (LOC && ((c < 256) != (j < 256))))
10582 if (c < 256 && AT_LEAST_UNI_SEMANTICS) {
10583 stored += set_regclass_bit(pRExC_state,
10586 &l1_fold_invlist, &unicode_alternate);
10588 /* It may be that the code point is already
10589 * in this range or already in the bitmap,
10590 * in which case we need do nothing */
10591 else if ((c < start || c > end)
10593 || ! ANYOF_BITMAP_TEST(ret, c)))
10595 nonbitmap = add_cp_to_invlist(nonbitmap, c);
10602 SvREFCNT_dec(fold_intersection);
10605 /* Combine the two lists into one. */
10606 if (l1_fold_invlist) {
10608 _invlist_union(nonbitmap, l1_fold_invlist, &nonbitmap);
10609 SvREFCNT_dec(l1_fold_invlist);
10612 nonbitmap = l1_fold_invlist;
10616 /* Here, we have calculated what code points should be in the character
10617 * class. Now we can see about various optimizations. Fold calculation
10618 * needs to take place before inversion. Otherwise /[^k]/i would invert to
10619 * include K, which under /i would match k. */
10621 /* Optimize inverted simple patterns (e.g. [^a-z]). Note that we haven't
10622 * set the FOLD flag yet, so this this does optimize those. It doesn't
10623 * optimize locale. Doing so perhaps could be done as long as there is
10624 * nothing like \w in it; some thought also would have to be given to the
10625 * interaction with above 0x100 chars */
10627 && (ANYOF_FLAGS(ret) & ANYOF_INVERT)
10628 && ! unicode_alternate
10629 /* In case of /d, there are some things that should match only when in
10630 * not in the bitmap, i.e., they require UTF8 to match. These are
10631 * listed in nonbitmap. */
10633 || ! DEPENDS_SEMANTICS
10634 || (ANYOF_FLAGS(ret) & ANYOF_NONBITMAP_NON_UTF8))
10635 && SvCUR(listsv) == initial_listsv_len)
10638 for (value = 0; value < ANYOF_BITMAP_SIZE; ++value)
10639 ANYOF_BITMAP(ret)[value] ^= 0xFF;
10640 /* The inversion means that everything above 255 is matched */
10641 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
10644 /* Here, also has things outside the bitmap. Go through each bit
10645 * individually and add it to the list to get rid of from those
10646 * things not in the bitmap */
10647 SV *remove_list = _new_invlist(2);
10648 _invlist_invert(nonbitmap);
10649 for (value = 0; value < 256; ++value) {
10650 if (ANYOF_BITMAP_TEST(ret, value)) {
10651 ANYOF_BITMAP_CLEAR(ret, value);
10652 remove_list = add_cp_to_invlist(remove_list, value);
10655 ANYOF_BITMAP_SET(ret, value);
10658 _invlist_subtract(nonbitmap, remove_list, &nonbitmap);
10659 SvREFCNT_dec(remove_list);
10662 stored = 256 - stored;
10664 /* Clear the invert flag since have just done it here */
10665 ANYOF_FLAGS(ret) &= ~ANYOF_INVERT;
10668 /* Folding in the bitmap is taken care of above, but not for locale (for
10669 * which we have to wait to see what folding is in effect at runtime), and
10670 * for things not in the bitmap. Set run-time fold flag for these */
10671 if (FOLD && (LOC || nonbitmap || unicode_alternate)) {
10672 ANYOF_FLAGS(ret) |= ANYOF_LOC_NONBITMAP_FOLD;
10675 /* A single character class can be "optimized" into an EXACTish node.
10676 * Note that since we don't currently count how many characters there are
10677 * outside the bitmap, we are XXX missing optimization possibilities for
10678 * them. This optimization can't happen unless this is a truly single
10679 * character class, which means that it can't be an inversion into a
10680 * many-character class, and there must be no possibility of there being
10681 * things outside the bitmap. 'stored' (only) for locales doesn't include
10682 * \w, etc, so have to make a special test that they aren't present
10684 * Similarly A 2-character class of the very special form like [bB] can be
10685 * optimized into an EXACTFish node, but only for non-locales, and for
10686 * characters which only have the two folds; so things like 'fF' and 'Ii'
10687 * wouldn't work because they are part of the fold of 'LATIN SMALL LIGATURE
10690 && ! unicode_alternate
10691 && SvCUR(listsv) == initial_listsv_len
10692 && ! (ANYOF_FLAGS(ret) & (ANYOF_INVERT|ANYOF_UNICODE_ALL))
10693 && (((stored == 1 && ((! (ANYOF_FLAGS(ret) & ANYOF_LOCALE))
10694 || (! ANYOF_CLASS_TEST_ANY_SET(ret)))))
10695 || (stored == 2 && ((! (ANYOF_FLAGS(ret) & ANYOF_LOCALE))
10696 && (! _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(value))
10697 /* If the latest code point has a fold whose
10698 * bit is set, it must be the only other one */
10699 && ((prevvalue = PL_fold_latin1[value]) != (IV)value)
10700 && ANYOF_BITMAP_TEST(ret, prevvalue)))))
10702 /* Note that the information needed to decide to do this optimization
10703 * is not currently available until the 2nd pass, and that the actually
10704 * used EXACTish node takes less space than the calculated ANYOF node,
10705 * and hence the amount of space calculated in the first pass is larger
10706 * than actually used, so this optimization doesn't gain us any space.
10707 * But an EXACT node is faster than an ANYOF node, and can be combined
10708 * with any adjacent EXACT nodes later by the optimizer for further
10709 * gains. The speed of executing an EXACTF is similar to an ANYOF
10710 * node, so the optimization advantage comes from the ability to join
10711 * it to adjacent EXACT nodes */
10713 const char * cur_parse= RExC_parse;
10715 RExC_emit = (regnode *)orig_emit;
10716 RExC_parse = (char *)orig_parse;
10720 /* A locale node with one point can be folded; all the other cases
10721 * with folding will have two points, since we calculate them above
10723 if (ANYOF_FLAGS(ret) & ANYOF_LOC_NONBITMAP_FOLD) {
10730 else { /* else 2 chars in the bit map: the folds of each other */
10732 /* Use the folded value, which for the cases where we get here,
10733 * is just the lower case of the current one (which may resolve to
10734 * itself, or to the other one */
10735 value = toLOWER_LATIN1(value);
10736 if (AT_LEAST_UNI_SEMANTICS || !isASCII(value)) {
10738 /* To join adjacent nodes, they must be the exact EXACTish
10739 * type. Try to use the most likely type, by using EXACTFU if
10740 * the regex calls for them, or is required because the
10741 * character is non-ASCII */
10744 else { /* Otherwise, more likely to be EXACTF type */
10749 ret = reg_node(pRExC_state, op);
10750 RExC_parse = (char *)cur_parse;
10751 if (UTF && ! NATIVE_IS_INVARIANT(value)) {
10752 *STRING(ret)= UTF8_EIGHT_BIT_HI((U8) value);
10753 *(STRING(ret) + 1)= UTF8_EIGHT_BIT_LO((U8) value);
10755 RExC_emit += STR_SZ(2);
10758 *STRING(ret)= (char)value;
10760 RExC_emit += STR_SZ(1);
10762 SvREFCNT_dec(listsv);
10768 invlist_iterinit(nonbitmap);
10769 while (invlist_iternext(nonbitmap, &start, &end)) {
10770 if (start == end) {
10771 Perl_sv_catpvf(aTHX_ listsv, "%04"UVxf"\n", start);
10774 /* The \t sets the whole range */
10775 Perl_sv_catpvf(aTHX_ listsv, "%04"UVxf"\t%04"UVxf"\n",
10780 SvREFCNT_dec(nonbitmap);
10783 if (SvCUR(listsv) == initial_listsv_len && ! unicode_alternate) {
10784 ARG_SET(ret, ANYOF_NONBITMAP_EMPTY);
10785 SvREFCNT_dec(listsv);
10786 SvREFCNT_dec(unicode_alternate);
10790 AV * const av = newAV();
10792 /* The 0th element stores the character class description
10793 * in its textual form: used later (regexec.c:Perl_regclass_swash())
10794 * to initialize the appropriate swash (which gets stored in
10795 * the 1st element), and also useful for dumping the regnode.
10796 * The 2nd element stores the multicharacter foldings,
10797 * used later (regexec.c:S_reginclass()). */
10798 av_store(av, 0, listsv);
10799 av_store(av, 1, NULL);
10801 /* Store any computed multi-char folds only if we are allowing
10803 if (allow_full_fold) {
10804 av_store(av, 2, MUTABLE_SV(unicode_alternate));
10805 if (unicode_alternate) { /* This node is variable length */
10810 av_store(av, 2, NULL);
10812 rv = newRV_noinc(MUTABLE_SV(av));
10813 n = add_data(pRExC_state, 1, "s");
10814 RExC_rxi->data->data[n] = (void*)rv;
10822 /* reg_skipcomment()
10824 Absorbs an /x style # comments from the input stream.
10825 Returns true if there is more text remaining in the stream.
10826 Will set the REG_SEEN_RUN_ON_COMMENT flag if the comment
10827 terminates the pattern without including a newline.
10829 Note its the callers responsibility to ensure that we are
10830 actually in /x mode
10835 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
10839 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
10841 while (RExC_parse < RExC_end)
10842 if (*RExC_parse++ == '\n') {
10847 /* we ran off the end of the pattern without ending
10848 the comment, so we have to add an \n when wrapping */
10849 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
10857 Advances the parse position, and optionally absorbs
10858 "whitespace" from the inputstream.
10860 Without /x "whitespace" means (?#...) style comments only,
10861 with /x this means (?#...) and # comments and whitespace proper.
10863 Returns the RExC_parse point from BEFORE the scan occurs.
10865 This is the /x friendly way of saying RExC_parse++.
10869 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
10871 char* const retval = RExC_parse++;
10873 PERL_ARGS_ASSERT_NEXTCHAR;
10876 if (*RExC_parse == '(' && RExC_parse[1] == '?' &&
10877 RExC_parse[2] == '#') {
10878 while (*RExC_parse != ')') {
10879 if (RExC_parse == RExC_end)
10880 FAIL("Sequence (?#... not terminated");
10886 if (RExC_flags & RXf_PMf_EXTENDED) {
10887 if (isSPACE(*RExC_parse)) {
10891 else if (*RExC_parse == '#') {
10892 if ( reg_skipcomment( pRExC_state ) )
10901 - reg_node - emit a node
10903 STATIC regnode * /* Location. */
10904 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
10907 register regnode *ptr;
10908 regnode * const ret = RExC_emit;
10909 GET_RE_DEBUG_FLAGS_DECL;
10911 PERL_ARGS_ASSERT_REG_NODE;
10914 SIZE_ALIGN(RExC_size);
10918 if (RExC_emit >= RExC_emit_bound)
10919 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d", op);
10921 NODE_ALIGN_FILL(ret);
10923 FILL_ADVANCE_NODE(ptr, op);
10924 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (ptr) - 1);
10925 #ifdef RE_TRACK_PATTERN_OFFSETS
10926 if (RExC_offsets) { /* MJD */
10927 MJD_OFFSET_DEBUG(("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
10928 "reg_node", __LINE__,
10930 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
10931 ? "Overwriting end of array!\n" : "OK",
10932 (UV)(RExC_emit - RExC_emit_start),
10933 (UV)(RExC_parse - RExC_start),
10934 (UV)RExC_offsets[0]));
10935 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
10943 - reganode - emit a node with an argument
10945 STATIC regnode * /* Location. */
10946 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
10949 register regnode *ptr;
10950 regnode * const ret = RExC_emit;
10951 GET_RE_DEBUG_FLAGS_DECL;
10953 PERL_ARGS_ASSERT_REGANODE;
10956 SIZE_ALIGN(RExC_size);
10961 assert(2==regarglen[op]+1);
10963 Anything larger than this has to allocate the extra amount.
10964 If we changed this to be:
10966 RExC_size += (1 + regarglen[op]);
10968 then it wouldn't matter. Its not clear what side effect
10969 might come from that so its not done so far.
10974 if (RExC_emit >= RExC_emit_bound)
10975 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d", op);
10977 NODE_ALIGN_FILL(ret);
10979 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
10980 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (ptr) - 2);
10981 #ifdef RE_TRACK_PATTERN_OFFSETS
10982 if (RExC_offsets) { /* MJD */
10983 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
10987 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
10988 "Overwriting end of array!\n" : "OK",
10989 (UV)(RExC_emit - RExC_emit_start),
10990 (UV)(RExC_parse - RExC_start),
10991 (UV)RExC_offsets[0]));
10992 Set_Cur_Node_Offset;
11000 - reguni - emit (if appropriate) a Unicode character
11003 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
11007 PERL_ARGS_ASSERT_REGUNI;
11009 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
11013 - reginsert - insert an operator in front of already-emitted operand
11015 * Means relocating the operand.
11018 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
11021 register regnode *src;
11022 register regnode *dst;
11023 register regnode *place;
11024 const int offset = regarglen[(U8)op];
11025 const int size = NODE_STEP_REGNODE + offset;
11026 GET_RE_DEBUG_FLAGS_DECL;
11028 PERL_ARGS_ASSERT_REGINSERT;
11029 PERL_UNUSED_ARG(depth);
11030 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
11031 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
11040 if (RExC_open_parens) {
11042 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
11043 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
11044 if ( RExC_open_parens[paren] >= opnd ) {
11045 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
11046 RExC_open_parens[paren] += size;
11048 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
11050 if ( RExC_close_parens[paren] >= opnd ) {
11051 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
11052 RExC_close_parens[paren] += size;
11054 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
11059 while (src > opnd) {
11060 StructCopy(--src, --dst, regnode);
11061 #ifdef RE_TRACK_PATTERN_OFFSETS
11062 if (RExC_offsets) { /* MJD 20010112 */
11063 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
11067 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
11068 ? "Overwriting end of array!\n" : "OK",
11069 (UV)(src - RExC_emit_start),
11070 (UV)(dst - RExC_emit_start),
11071 (UV)RExC_offsets[0]));
11072 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
11073 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
11079 place = opnd; /* Op node, where operand used to be. */
11080 #ifdef RE_TRACK_PATTERN_OFFSETS
11081 if (RExC_offsets) { /* MJD */
11082 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
11086 (UV)(place - RExC_emit_start) > RExC_offsets[0]
11087 ? "Overwriting end of array!\n" : "OK",
11088 (UV)(place - RExC_emit_start),
11089 (UV)(RExC_parse - RExC_start),
11090 (UV)RExC_offsets[0]));
11091 Set_Node_Offset(place, RExC_parse);
11092 Set_Node_Length(place, 1);
11095 src = NEXTOPER(place);
11096 FILL_ADVANCE_NODE(place, op);
11097 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (place) - 1);
11098 Zero(src, offset, regnode);
11102 - regtail - set the next-pointer at the end of a node chain of p to val.
11103 - SEE ALSO: regtail_study
11105 /* TODO: All three parms should be const */
11107 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
11110 register regnode *scan;
11111 GET_RE_DEBUG_FLAGS_DECL;
11113 PERL_ARGS_ASSERT_REGTAIL;
11115 PERL_UNUSED_ARG(depth);
11121 /* Find last node. */
11124 regnode * const temp = regnext(scan);
11126 SV * const mysv=sv_newmortal();
11127 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
11128 regprop(RExC_rx, mysv, scan);
11129 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
11130 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
11131 (temp == NULL ? "->" : ""),
11132 (temp == NULL ? PL_reg_name[OP(val)] : "")
11140 if (reg_off_by_arg[OP(scan)]) {
11141 ARG_SET(scan, val - scan);
11144 NEXT_OFF(scan) = val - scan;
11150 - regtail_study - set the next-pointer at the end of a node chain of p to val.
11151 - Look for optimizable sequences at the same time.
11152 - currently only looks for EXACT chains.
11154 This is experimental code. The idea is to use this routine to perform
11155 in place optimizations on branches and groups as they are constructed,
11156 with the long term intention of removing optimization from study_chunk so
11157 that it is purely analytical.
11159 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
11160 to control which is which.
11163 /* TODO: All four parms should be const */
11166 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
11169 register regnode *scan;
11171 #ifdef EXPERIMENTAL_INPLACESCAN
11174 GET_RE_DEBUG_FLAGS_DECL;
11176 PERL_ARGS_ASSERT_REGTAIL_STUDY;
11182 /* Find last node. */
11186 regnode * const temp = regnext(scan);
11187 #ifdef EXPERIMENTAL_INPLACESCAN
11188 if (PL_regkind[OP(scan)] == EXACT)
11189 if (join_exact(pRExC_state,scan,&min,1,val,depth+1))
11193 switch (OP(scan)) {
11199 if( exact == PSEUDO )
11201 else if ( exact != OP(scan) )
11210 SV * const mysv=sv_newmortal();
11211 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
11212 regprop(RExC_rx, mysv, scan);
11213 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
11214 SvPV_nolen_const(mysv),
11215 REG_NODE_NUM(scan),
11216 PL_reg_name[exact]);
11223 SV * const mysv_val=sv_newmortal();
11224 DEBUG_PARSE_MSG("");
11225 regprop(RExC_rx, mysv_val, val);
11226 PerlIO_printf(Perl_debug_log, "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
11227 SvPV_nolen_const(mysv_val),
11228 (IV)REG_NODE_NUM(val),
11232 if (reg_off_by_arg[OP(scan)]) {
11233 ARG_SET(scan, val - scan);
11236 NEXT_OFF(scan) = val - scan;
11244 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
11248 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
11254 for (bit=0; bit<32; bit++) {
11255 if (flags & (1<<bit)) {
11256 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
11259 if (!set++ && lead)
11260 PerlIO_printf(Perl_debug_log, "%s",lead);
11261 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
11264 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
11265 if (!set++ && lead) {
11266 PerlIO_printf(Perl_debug_log, "%s",lead);
11269 case REGEX_UNICODE_CHARSET:
11270 PerlIO_printf(Perl_debug_log, "UNICODE");
11272 case REGEX_LOCALE_CHARSET:
11273 PerlIO_printf(Perl_debug_log, "LOCALE");
11275 case REGEX_ASCII_RESTRICTED_CHARSET:
11276 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
11278 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
11279 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
11282 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
11288 PerlIO_printf(Perl_debug_log, "\n");
11290 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
11296 Perl_regdump(pTHX_ const regexp *r)
11300 SV * const sv = sv_newmortal();
11301 SV *dsv= sv_newmortal();
11302 RXi_GET_DECL(r,ri);
11303 GET_RE_DEBUG_FLAGS_DECL;
11305 PERL_ARGS_ASSERT_REGDUMP;
11307 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
11309 /* Header fields of interest. */
11310 if (r->anchored_substr) {
11311 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
11312 RE_SV_DUMPLEN(r->anchored_substr), 30);
11313 PerlIO_printf(Perl_debug_log,
11314 "anchored %s%s at %"IVdf" ",
11315 s, RE_SV_TAIL(r->anchored_substr),
11316 (IV)r->anchored_offset);
11317 } else if (r->anchored_utf8) {
11318 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
11319 RE_SV_DUMPLEN(r->anchored_utf8), 30);
11320 PerlIO_printf(Perl_debug_log,
11321 "anchored utf8 %s%s at %"IVdf" ",
11322 s, RE_SV_TAIL(r->anchored_utf8),
11323 (IV)r->anchored_offset);
11325 if (r->float_substr) {
11326 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
11327 RE_SV_DUMPLEN(r->float_substr), 30);
11328 PerlIO_printf(Perl_debug_log,
11329 "floating %s%s at %"IVdf"..%"UVuf" ",
11330 s, RE_SV_TAIL(r->float_substr),
11331 (IV)r->float_min_offset, (UV)r->float_max_offset);
11332 } else if (r->float_utf8) {
11333 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
11334 RE_SV_DUMPLEN(r->float_utf8), 30);
11335 PerlIO_printf(Perl_debug_log,
11336 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
11337 s, RE_SV_TAIL(r->float_utf8),
11338 (IV)r->float_min_offset, (UV)r->float_max_offset);
11340 if (r->check_substr || r->check_utf8)
11341 PerlIO_printf(Perl_debug_log,
11343 (r->check_substr == r->float_substr
11344 && r->check_utf8 == r->float_utf8
11345 ? "(checking floating" : "(checking anchored"));
11346 if (r->extflags & RXf_NOSCAN)
11347 PerlIO_printf(Perl_debug_log, " noscan");
11348 if (r->extflags & RXf_CHECK_ALL)
11349 PerlIO_printf(Perl_debug_log, " isall");
11350 if (r->check_substr || r->check_utf8)
11351 PerlIO_printf(Perl_debug_log, ") ");
11353 if (ri->regstclass) {
11354 regprop(r, sv, ri->regstclass);
11355 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
11357 if (r->extflags & RXf_ANCH) {
11358 PerlIO_printf(Perl_debug_log, "anchored");
11359 if (r->extflags & RXf_ANCH_BOL)
11360 PerlIO_printf(Perl_debug_log, "(BOL)");
11361 if (r->extflags & RXf_ANCH_MBOL)
11362 PerlIO_printf(Perl_debug_log, "(MBOL)");
11363 if (r->extflags & RXf_ANCH_SBOL)
11364 PerlIO_printf(Perl_debug_log, "(SBOL)");
11365 if (r->extflags & RXf_ANCH_GPOS)
11366 PerlIO_printf(Perl_debug_log, "(GPOS)");
11367 PerlIO_putc(Perl_debug_log, ' ');
11369 if (r->extflags & RXf_GPOS_SEEN)
11370 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
11371 if (r->intflags & PREGf_SKIP)
11372 PerlIO_printf(Perl_debug_log, "plus ");
11373 if (r->intflags & PREGf_IMPLICIT)
11374 PerlIO_printf(Perl_debug_log, "implicit ");
11375 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
11376 if (r->extflags & RXf_EVAL_SEEN)
11377 PerlIO_printf(Perl_debug_log, "with eval ");
11378 PerlIO_printf(Perl_debug_log, "\n");
11379 DEBUG_FLAGS_r(regdump_extflags("r->extflags: ",r->extflags));
11381 PERL_ARGS_ASSERT_REGDUMP;
11382 PERL_UNUSED_CONTEXT;
11383 PERL_UNUSED_ARG(r);
11384 #endif /* DEBUGGING */
11388 - regprop - printable representation of opcode
11390 #define EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags) \
11393 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]); \
11394 if (flags & ANYOF_INVERT) \
11395 /*make sure the invert info is in each */ \
11396 sv_catpvs(sv, "^"); \
11402 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
11407 RXi_GET_DECL(prog,progi);
11408 GET_RE_DEBUG_FLAGS_DECL;
11410 PERL_ARGS_ASSERT_REGPROP;
11414 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
11415 /* It would be nice to FAIL() here, but this may be called from
11416 regexec.c, and it would be hard to supply pRExC_state. */
11417 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(o), (int)REGNODE_MAX);
11418 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
11420 k = PL_regkind[OP(o)];
11423 sv_catpvs(sv, " ");
11424 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
11425 * is a crude hack but it may be the best for now since
11426 * we have no flag "this EXACTish node was UTF-8"
11428 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
11429 PERL_PV_ESCAPE_UNI_DETECT |
11430 PERL_PV_ESCAPE_NONASCII |
11431 PERL_PV_PRETTY_ELLIPSES |
11432 PERL_PV_PRETTY_LTGT |
11433 PERL_PV_PRETTY_NOCLEAR
11435 } else if (k == TRIE) {
11436 /* print the details of the trie in dumpuntil instead, as
11437 * progi->data isn't available here */
11438 const char op = OP(o);
11439 const U32 n = ARG(o);
11440 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
11441 (reg_ac_data *)progi->data->data[n] :
11443 const reg_trie_data * const trie
11444 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
11446 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
11447 DEBUG_TRIE_COMPILE_r(
11448 Perl_sv_catpvf(aTHX_ sv,
11449 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
11450 (UV)trie->startstate,
11451 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
11452 (UV)trie->wordcount,
11455 (UV)TRIE_CHARCOUNT(trie),
11456 (UV)trie->uniquecharcount
11459 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
11461 int rangestart = -1;
11462 U8* bitmap = IS_ANYOF_TRIE(op) ? (U8*)ANYOF_BITMAP(o) : (U8*)TRIE_BITMAP(trie);
11463 sv_catpvs(sv, "[");
11464 for (i = 0; i <= 256; i++) {
11465 if (i < 256 && BITMAP_TEST(bitmap,i)) {
11466 if (rangestart == -1)
11468 } else if (rangestart != -1) {
11469 if (i <= rangestart + 3)
11470 for (; rangestart < i; rangestart++)
11471 put_byte(sv, rangestart);
11473 put_byte(sv, rangestart);
11474 sv_catpvs(sv, "-");
11475 put_byte(sv, i - 1);
11480 sv_catpvs(sv, "]");
11483 } else if (k == CURLY) {
11484 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
11485 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
11486 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
11488 else if (k == WHILEM && o->flags) /* Ordinal/of */
11489 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
11490 else if (k == REF || k == OPEN || k == CLOSE || k == GROUPP || OP(o)==ACCEPT) {
11491 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
11492 if ( RXp_PAREN_NAMES(prog) ) {
11493 if ( k != REF || (OP(o) < NREF)) {
11494 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
11495 SV **name= av_fetch(list, ARG(o), 0 );
11497 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
11500 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
11501 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
11502 I32 *nums=(I32*)SvPVX(sv_dat);
11503 SV **name= av_fetch(list, nums[0], 0 );
11506 for ( n=0; n<SvIVX(sv_dat); n++ ) {
11507 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
11508 (n ? "," : ""), (IV)nums[n]);
11510 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
11514 } else if (k == GOSUB)
11515 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o)); /* Paren and offset */
11516 else if (k == VERB) {
11518 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
11519 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
11520 } else if (k == LOGICAL)
11521 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* 2: embedded, otherwise 1 */
11522 else if (k == FOLDCHAR)
11523 Perl_sv_catpvf(aTHX_ sv, "[0x%"UVXf"]", PTR2UV(ARG(o)) );
11524 else if (k == ANYOF) {
11525 int i, rangestart = -1;
11526 const U8 flags = ANYOF_FLAGS(o);
11529 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
11530 static const char * const anyofs[] = {
11563 if (flags & ANYOF_LOCALE)
11564 sv_catpvs(sv, "{loc}");
11565 if (flags & ANYOF_LOC_NONBITMAP_FOLD)
11566 sv_catpvs(sv, "{i}");
11567 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
11568 if (flags & ANYOF_INVERT)
11569 sv_catpvs(sv, "^");
11571 /* output what the standard cp 0-255 bitmap matches */
11572 for (i = 0; i <= 256; i++) {
11573 if (i < 256 && ANYOF_BITMAP_TEST(o,i)) {
11574 if (rangestart == -1)
11576 } else if (rangestart != -1) {
11577 if (i <= rangestart + 3)
11578 for (; rangestart < i; rangestart++)
11579 put_byte(sv, rangestart);
11581 put_byte(sv, rangestart);
11582 sv_catpvs(sv, "-");
11583 put_byte(sv, i - 1);
11590 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
11591 /* output any special charclass tests (used entirely under use locale) */
11592 if (ANYOF_CLASS_TEST_ANY_SET(o))
11593 for (i = 0; i < (int)(sizeof(anyofs)/sizeof(char*)); i++)
11594 if (ANYOF_CLASS_TEST(o,i)) {
11595 sv_catpv(sv, anyofs[i]);
11599 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
11601 if (flags & ANYOF_NON_UTF8_LATIN1_ALL) {
11602 sv_catpvs(sv, "{non-utf8-latin1-all}");
11605 /* output information about the unicode matching */
11606 if (flags & ANYOF_UNICODE_ALL)
11607 sv_catpvs(sv, "{unicode_all}");
11608 else if (ANYOF_NONBITMAP(o))
11609 sv_catpvs(sv, "{unicode}");
11610 if (flags & ANYOF_NONBITMAP_NON_UTF8)
11611 sv_catpvs(sv, "{outside bitmap}");
11613 if (ANYOF_NONBITMAP(o)) {
11615 SV * const sw = regclass_swash(prog, o, FALSE, &lv, 0);
11619 U8 s[UTF8_MAXBYTES_CASE+1];
11621 for (i = 0; i <= 256; i++) { /* just the first 256 */
11622 uvchr_to_utf8(s, i);
11624 if (i < 256 && swash_fetch(sw, s, TRUE)) {
11625 if (rangestart == -1)
11627 } else if (rangestart != -1) {
11628 if (i <= rangestart + 3)
11629 for (; rangestart < i; rangestart++) {
11630 const U8 * const e = uvchr_to_utf8(s,rangestart);
11632 for(p = s; p < e; p++)
11636 const U8 *e = uvchr_to_utf8(s,rangestart);
11638 for (p = s; p < e; p++)
11640 sv_catpvs(sv, "-");
11641 e = uvchr_to_utf8(s, i-1);
11642 for (p = s; p < e; p++)
11649 sv_catpvs(sv, "..."); /* et cetera */
11653 char *s = savesvpv(lv);
11654 char * const origs = s;
11656 while (*s && *s != '\n')
11660 const char * const t = ++s;
11678 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
11680 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
11681 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
11683 PERL_UNUSED_CONTEXT;
11684 PERL_UNUSED_ARG(sv);
11685 PERL_UNUSED_ARG(o);
11686 PERL_UNUSED_ARG(prog);
11687 #endif /* DEBUGGING */
11691 Perl_re_intuit_string(pTHX_ REGEXP * const r)
11692 { /* Assume that RE_INTUIT is set */
11694 struct regexp *const prog = (struct regexp *)SvANY(r);
11695 GET_RE_DEBUG_FLAGS_DECL;
11697 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
11698 PERL_UNUSED_CONTEXT;
11702 const char * const s = SvPV_nolen_const(prog->check_substr
11703 ? prog->check_substr : prog->check_utf8);
11705 if (!PL_colorset) reginitcolors();
11706 PerlIO_printf(Perl_debug_log,
11707 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
11709 prog->check_substr ? "" : "utf8 ",
11710 PL_colors[5],PL_colors[0],
11713 (strlen(s) > 60 ? "..." : ""));
11716 return prog->check_substr ? prog->check_substr : prog->check_utf8;
11722 handles refcounting and freeing the perl core regexp structure. When
11723 it is necessary to actually free the structure the first thing it
11724 does is call the 'free' method of the regexp_engine associated to
11725 the regexp, allowing the handling of the void *pprivate; member
11726 first. (This routine is not overridable by extensions, which is why
11727 the extensions free is called first.)
11729 See regdupe and regdupe_internal if you change anything here.
11731 #ifndef PERL_IN_XSUB_RE
11733 Perl_pregfree(pTHX_ REGEXP *r)
11739 Perl_pregfree2(pTHX_ REGEXP *rx)
11742 struct regexp *const r = (struct regexp *)SvANY(rx);
11743 GET_RE_DEBUG_FLAGS_DECL;
11745 PERL_ARGS_ASSERT_PREGFREE2;
11747 if (r->mother_re) {
11748 ReREFCNT_dec(r->mother_re);
11750 CALLREGFREE_PVT(rx); /* free the private data */
11751 SvREFCNT_dec(RXp_PAREN_NAMES(r));
11754 SvREFCNT_dec(r->anchored_substr);
11755 SvREFCNT_dec(r->anchored_utf8);
11756 SvREFCNT_dec(r->float_substr);
11757 SvREFCNT_dec(r->float_utf8);
11758 Safefree(r->substrs);
11760 RX_MATCH_COPY_FREE(rx);
11761 #ifdef PERL_OLD_COPY_ON_WRITE
11762 SvREFCNT_dec(r->saved_copy);
11769 This is a hacky workaround to the structural issue of match results
11770 being stored in the regexp structure which is in turn stored in
11771 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
11772 could be PL_curpm in multiple contexts, and could require multiple
11773 result sets being associated with the pattern simultaneously, such
11774 as when doing a recursive match with (??{$qr})
11776 The solution is to make a lightweight copy of the regexp structure
11777 when a qr// is returned from the code executed by (??{$qr}) this
11778 lightweight copy doesn't actually own any of its data except for
11779 the starp/end and the actual regexp structure itself.
11785 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
11787 struct regexp *ret;
11788 struct regexp *const r = (struct regexp *)SvANY(rx);
11789 register const I32 npar = r->nparens+1;
11791 PERL_ARGS_ASSERT_REG_TEMP_COPY;
11794 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
11795 ret = (struct regexp *)SvANY(ret_x);
11797 (void)ReREFCNT_inc(rx);
11798 /* We can take advantage of the existing "copied buffer" mechanism in SVs
11799 by pointing directly at the buffer, but flagging that the allocated
11800 space in the copy is zero. As we've just done a struct copy, it's now
11801 a case of zero-ing that, rather than copying the current length. */
11802 SvPV_set(ret_x, RX_WRAPPED(rx));
11803 SvFLAGS(ret_x) |= SvFLAGS(rx) & (SVf_POK|SVp_POK|SVf_UTF8);
11804 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
11805 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
11806 SvLEN_set(ret_x, 0);
11807 SvSTASH_set(ret_x, NULL);
11808 SvMAGIC_set(ret_x, NULL);
11809 Newx(ret->offs, npar, regexp_paren_pair);
11810 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
11812 Newx(ret->substrs, 1, struct reg_substr_data);
11813 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
11815 SvREFCNT_inc_void(ret->anchored_substr);
11816 SvREFCNT_inc_void(ret->anchored_utf8);
11817 SvREFCNT_inc_void(ret->float_substr);
11818 SvREFCNT_inc_void(ret->float_utf8);
11820 /* check_substr and check_utf8, if non-NULL, point to either their
11821 anchored or float namesakes, and don't hold a second reference. */
11823 RX_MATCH_COPIED_off(ret_x);
11824 #ifdef PERL_OLD_COPY_ON_WRITE
11825 ret->saved_copy = NULL;
11827 ret->mother_re = rx;
11833 /* regfree_internal()
11835 Free the private data in a regexp. This is overloadable by
11836 extensions. Perl takes care of the regexp structure in pregfree(),
11837 this covers the *pprivate pointer which technically perl doesn't
11838 know about, however of course we have to handle the
11839 regexp_internal structure when no extension is in use.
11841 Note this is called before freeing anything in the regexp
11846 Perl_regfree_internal(pTHX_ REGEXP * const rx)
11849 struct regexp *const r = (struct regexp *)SvANY(rx);
11850 RXi_GET_DECL(r,ri);
11851 GET_RE_DEBUG_FLAGS_DECL;
11853 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
11859 SV *dsv= sv_newmortal();
11860 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
11861 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
11862 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
11863 PL_colors[4],PL_colors[5],s);
11866 #ifdef RE_TRACK_PATTERN_OFFSETS
11868 Safefree(ri->u.offsets); /* 20010421 MJD */
11871 int n = ri->data->count;
11872 PAD* new_comppad = NULL;
11877 /* If you add a ->what type here, update the comment in regcomp.h */
11878 switch (ri->data->what[n]) {
11883 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
11886 Safefree(ri->data->data[n]);
11889 new_comppad = MUTABLE_AV(ri->data->data[n]);
11892 if (new_comppad == NULL)
11893 Perl_croak(aTHX_ "panic: pregfree comppad");
11894 PAD_SAVE_LOCAL(old_comppad,
11895 /* Watch out for global destruction's random ordering. */
11896 (SvTYPE(new_comppad) == SVt_PVAV) ? new_comppad : NULL
11899 refcnt = OpREFCNT_dec((OP_4tree*)ri->data->data[n]);
11902 op_free((OP_4tree*)ri->data->data[n]);
11904 PAD_RESTORE_LOCAL(old_comppad);
11905 SvREFCNT_dec(MUTABLE_SV(new_comppad));
11906 new_comppad = NULL;
11911 { /* Aho Corasick add-on structure for a trie node.
11912 Used in stclass optimization only */
11914 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
11916 refcount = --aho->refcount;
11919 PerlMemShared_free(aho->states);
11920 PerlMemShared_free(aho->fail);
11921 /* do this last!!!! */
11922 PerlMemShared_free(ri->data->data[n]);
11923 PerlMemShared_free(ri->regstclass);
11929 /* trie structure. */
11931 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
11933 refcount = --trie->refcount;
11936 PerlMemShared_free(trie->charmap);
11937 PerlMemShared_free(trie->states);
11938 PerlMemShared_free(trie->trans);
11940 PerlMemShared_free(trie->bitmap);
11942 PerlMemShared_free(trie->jump);
11943 PerlMemShared_free(trie->wordinfo);
11944 /* do this last!!!! */
11945 PerlMemShared_free(ri->data->data[n]);
11950 Perl_croak(aTHX_ "panic: regfree data code '%c'", ri->data->what[n]);
11953 Safefree(ri->data->what);
11954 Safefree(ri->data);
11960 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
11961 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
11962 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
11965 re_dup - duplicate a regexp.
11967 This routine is expected to clone a given regexp structure. It is only
11968 compiled under USE_ITHREADS.
11970 After all of the core data stored in struct regexp is duplicated
11971 the regexp_engine.dupe method is used to copy any private data
11972 stored in the *pprivate pointer. This allows extensions to handle
11973 any duplication it needs to do.
11975 See pregfree() and regfree_internal() if you change anything here.
11977 #if defined(USE_ITHREADS)
11978 #ifndef PERL_IN_XSUB_RE
11980 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
11984 const struct regexp *r = (const struct regexp *)SvANY(sstr);
11985 struct regexp *ret = (struct regexp *)SvANY(dstr);
11987 PERL_ARGS_ASSERT_RE_DUP_GUTS;
11989 npar = r->nparens+1;
11990 Newx(ret->offs, npar, regexp_paren_pair);
11991 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
11993 /* no need to copy these */
11994 Newx(ret->swap, npar, regexp_paren_pair);
11997 if (ret->substrs) {
11998 /* Do it this way to avoid reading from *r after the StructCopy().
11999 That way, if any of the sv_dup_inc()s dislodge *r from the L1
12000 cache, it doesn't matter. */
12001 const bool anchored = r->check_substr
12002 ? r->check_substr == r->anchored_substr
12003 : r->check_utf8 == r->anchored_utf8;
12004 Newx(ret->substrs, 1, struct reg_substr_data);
12005 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
12007 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
12008 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
12009 ret->float_substr = sv_dup_inc(ret->float_substr, param);
12010 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
12012 /* check_substr and check_utf8, if non-NULL, point to either their
12013 anchored or float namesakes, and don't hold a second reference. */
12015 if (ret->check_substr) {
12017 assert(r->check_utf8 == r->anchored_utf8);
12018 ret->check_substr = ret->anchored_substr;
12019 ret->check_utf8 = ret->anchored_utf8;
12021 assert(r->check_substr == r->float_substr);
12022 assert(r->check_utf8 == r->float_utf8);
12023 ret->check_substr = ret->float_substr;
12024 ret->check_utf8 = ret->float_utf8;
12026 } else if (ret->check_utf8) {
12028 ret->check_utf8 = ret->anchored_utf8;
12030 ret->check_utf8 = ret->float_utf8;
12035 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
12038 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
12040 if (RX_MATCH_COPIED(dstr))
12041 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
12043 ret->subbeg = NULL;
12044 #ifdef PERL_OLD_COPY_ON_WRITE
12045 ret->saved_copy = NULL;
12048 if (ret->mother_re) {
12049 if (SvPVX_const(dstr) == SvPVX_const(ret->mother_re)) {
12050 /* Our storage points directly to our mother regexp, but that's
12051 1: a buffer in a different thread
12052 2: something we no longer hold a reference on
12053 so we need to copy it locally. */
12054 /* Note we need to use SvCUR(), rather than
12055 SvLEN(), on our mother_re, because it, in
12056 turn, may well be pointing to its own mother_re. */
12057 SvPV_set(dstr, SAVEPVN(SvPVX_const(ret->mother_re),
12058 SvCUR(ret->mother_re)+1));
12059 SvLEN_set(dstr, SvCUR(ret->mother_re)+1);
12061 ret->mother_re = NULL;
12065 #endif /* PERL_IN_XSUB_RE */
12070 This is the internal complement to regdupe() which is used to copy
12071 the structure pointed to by the *pprivate pointer in the regexp.
12072 This is the core version of the extension overridable cloning hook.
12073 The regexp structure being duplicated will be copied by perl prior
12074 to this and will be provided as the regexp *r argument, however
12075 with the /old/ structures pprivate pointer value. Thus this routine
12076 may override any copying normally done by perl.
12078 It returns a pointer to the new regexp_internal structure.
12082 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
12085 struct regexp *const r = (struct regexp *)SvANY(rx);
12086 regexp_internal *reti;
12088 RXi_GET_DECL(r,ri);
12090 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
12094 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode), char, regexp_internal);
12095 Copy(ri->program, reti->program, len+1, regnode);
12098 reti->regstclass = NULL;
12101 struct reg_data *d;
12102 const int count = ri->data->count;
12105 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
12106 char, struct reg_data);
12107 Newx(d->what, count, U8);
12110 for (i = 0; i < count; i++) {
12111 d->what[i] = ri->data->what[i];
12112 switch (d->what[i]) {
12113 /* legal options are one of: sSfpontTua
12114 see also regcomp.h and pregfree() */
12115 case 'a': /* actually an AV, but the dup function is identical. */
12118 case 'p': /* actually an AV, but the dup function is identical. */
12119 case 'u': /* actually an HV, but the dup function is identical. */
12120 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
12123 /* This is cheating. */
12124 Newx(d->data[i], 1, struct regnode_charclass_class);
12125 StructCopy(ri->data->data[i], d->data[i],
12126 struct regnode_charclass_class);
12127 reti->regstclass = (regnode*)d->data[i];
12130 /* Compiled op trees are readonly and in shared memory,
12131 and can thus be shared without duplication. */
12133 d->data[i] = (void*)OpREFCNT_inc((OP*)ri->data->data[i]);
12137 /* Trie stclasses are readonly and can thus be shared
12138 * without duplication. We free the stclass in pregfree
12139 * when the corresponding reg_ac_data struct is freed.
12141 reti->regstclass= ri->regstclass;
12145 ((reg_trie_data*)ri->data->data[i])->refcount++;
12149 d->data[i] = ri->data->data[i];
12152 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'", ri->data->what[i]);
12161 reti->name_list_idx = ri->name_list_idx;
12163 #ifdef RE_TRACK_PATTERN_OFFSETS
12164 if (ri->u.offsets) {
12165 Newx(reti->u.offsets, 2*len+1, U32);
12166 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
12169 SetProgLen(reti,len);
12172 return (void*)reti;
12175 #endif /* USE_ITHREADS */
12177 #ifndef PERL_IN_XSUB_RE
12180 - regnext - dig the "next" pointer out of a node
12183 Perl_regnext(pTHX_ register regnode *p)
12186 register I32 offset;
12191 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
12192 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(p), (int)REGNODE_MAX);
12195 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
12204 S_re_croak2(pTHX_ const char* pat1,const char* pat2,...)
12207 STRLEN l1 = strlen(pat1);
12208 STRLEN l2 = strlen(pat2);
12211 const char *message;
12213 PERL_ARGS_ASSERT_RE_CROAK2;
12219 Copy(pat1, buf, l1 , char);
12220 Copy(pat2, buf + l1, l2 , char);
12221 buf[l1 + l2] = '\n';
12222 buf[l1 + l2 + 1] = '\0';
12224 /* ANSI variant takes additional second argument */
12225 va_start(args, pat2);
12229 msv = vmess(buf, &args);
12231 message = SvPV_const(msv,l1);
12234 Copy(message, buf, l1 , char);
12235 buf[l1-1] = '\0'; /* Overwrite \n */
12236 Perl_croak(aTHX_ "%s", buf);
12239 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
12241 #ifndef PERL_IN_XSUB_RE
12243 Perl_save_re_context(pTHX)
12247 struct re_save_state *state;
12249 SAVEVPTR(PL_curcop);
12250 SSGROW(SAVESTACK_ALLOC_FOR_RE_SAVE_STATE + 1);
12252 state = (struct re_save_state *)(PL_savestack + PL_savestack_ix);
12253 PL_savestack_ix += SAVESTACK_ALLOC_FOR_RE_SAVE_STATE;
12254 SSPUSHUV(SAVEt_RE_STATE);
12256 Copy(&PL_reg_state, state, 1, struct re_save_state);
12258 PL_reg_start_tmp = 0;
12259 PL_reg_start_tmpl = 0;
12260 PL_reg_oldsaved = NULL;
12261 PL_reg_oldsavedlen = 0;
12262 PL_reg_maxiter = 0;
12263 PL_reg_leftiter = 0;
12264 PL_reg_poscache = NULL;
12265 PL_reg_poscache_size = 0;
12266 #ifdef PERL_OLD_COPY_ON_WRITE
12270 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
12272 const REGEXP * const rx = PM_GETRE(PL_curpm);
12275 for (i = 1; i <= RX_NPARENS(rx); i++) {
12276 char digits[TYPE_CHARS(long)];
12277 const STRLEN len = my_snprintf(digits, sizeof(digits), "%lu", (long)i);
12278 GV *const *const gvp
12279 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
12282 GV * const gv = *gvp;
12283 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
12293 clear_re(pTHX_ void *r)
12296 ReREFCNT_dec((REGEXP *)r);
12302 S_put_byte(pTHX_ SV *sv, int c)
12304 PERL_ARGS_ASSERT_PUT_BYTE;
12306 /* Our definition of isPRINT() ignores locales, so only bytes that are
12307 not part of UTF-8 are considered printable. I assume that the same
12308 holds for UTF-EBCDIC.
12309 Also, code point 255 is not printable in either (it's E0 in EBCDIC,
12310 which Wikipedia says:
12312 EO, or Eight Ones, is an 8-bit EBCDIC character code represented as all
12313 ones (binary 1111 1111, hexadecimal FF). It is similar, but not
12314 identical, to the ASCII delete (DEL) or rubout control character.
12315 ) So the old condition can be simplified to !isPRINT(c) */
12318 Perl_sv_catpvf(aTHX_ sv, "\\x%02x", c);
12321 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
12325 const char string = c;
12326 if (c == '-' || c == ']' || c == '\\' || c == '^')
12327 sv_catpvs(sv, "\\");
12328 sv_catpvn(sv, &string, 1);
12333 #define CLEAR_OPTSTART \
12334 if (optstart) STMT_START { \
12335 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
12339 #define DUMPUNTIL(b,e) CLEAR_OPTSTART; node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
12341 STATIC const regnode *
12342 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
12343 const regnode *last, const regnode *plast,
12344 SV* sv, I32 indent, U32 depth)
12347 register U8 op = PSEUDO; /* Arbitrary non-END op. */
12348 register const regnode *next;
12349 const regnode *optstart= NULL;
12351 RXi_GET_DECL(r,ri);
12352 GET_RE_DEBUG_FLAGS_DECL;
12354 PERL_ARGS_ASSERT_DUMPUNTIL;
12356 #ifdef DEBUG_DUMPUNTIL
12357 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
12358 last ? last-start : 0,plast ? plast-start : 0);
12361 if (plast && plast < last)
12364 while (PL_regkind[op] != END && (!last || node < last)) {
12365 /* While that wasn't END last time... */
12368 if (op == CLOSE || op == WHILEM)
12370 next = regnext((regnode *)node);
12373 if (OP(node) == OPTIMIZED) {
12374 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
12381 regprop(r, sv, node);
12382 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
12383 (int)(2*indent + 1), "", SvPVX_const(sv));
12385 if (OP(node) != OPTIMIZED) {
12386 if (next == NULL) /* Next ptr. */
12387 PerlIO_printf(Perl_debug_log, " (0)");
12388 else if (PL_regkind[(U8)op] == BRANCH && PL_regkind[OP(next)] != BRANCH )
12389 PerlIO_printf(Perl_debug_log, " (FAIL)");
12391 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
12392 (void)PerlIO_putc(Perl_debug_log, '\n');
12396 if (PL_regkind[(U8)op] == BRANCHJ) {
12399 register const regnode *nnode = (OP(next) == LONGJMP
12400 ? regnext((regnode *)next)
12402 if (last && nnode > last)
12404 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
12407 else if (PL_regkind[(U8)op] == BRANCH) {
12409 DUMPUNTIL(NEXTOPER(node), next);
12411 else if ( PL_regkind[(U8)op] == TRIE ) {
12412 const regnode *this_trie = node;
12413 const char op = OP(node);
12414 const U32 n = ARG(node);
12415 const reg_ac_data * const ac = op>=AHOCORASICK ?
12416 (reg_ac_data *)ri->data->data[n] :
12418 const reg_trie_data * const trie =
12419 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
12421 AV *const trie_words = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
12423 const regnode *nextbranch= NULL;
12426 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
12427 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
12429 PerlIO_printf(Perl_debug_log, "%*s%s ",
12430 (int)(2*(indent+3)), "",
12431 elem_ptr ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr), SvCUR(*elem_ptr), 60,
12432 PL_colors[0], PL_colors[1],
12433 (SvUTF8(*elem_ptr) ? PERL_PV_ESCAPE_UNI : 0) |
12434 PERL_PV_PRETTY_ELLIPSES |
12435 PERL_PV_PRETTY_LTGT
12440 U16 dist= trie->jump[word_idx+1];
12441 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
12442 (UV)((dist ? this_trie + dist : next) - start));
12445 nextbranch= this_trie + trie->jump[0];
12446 DUMPUNTIL(this_trie + dist, nextbranch);
12448 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
12449 nextbranch= regnext((regnode *)nextbranch);
12451 PerlIO_printf(Perl_debug_log, "\n");
12454 if (last && next > last)
12459 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
12460 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
12461 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
12463 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
12465 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
12467 else if ( op == PLUS || op == STAR) {
12468 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
12470 else if (PL_regkind[(U8)op] == ANYOF) {
12471 /* arglen 1 + class block */
12472 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_CLASS)
12473 ? ANYOF_CLASS_SKIP : ANYOF_SKIP);
12474 node = NEXTOPER(node);
12476 else if (PL_regkind[(U8)op] == EXACT) {
12477 /* Literal string, where present. */
12478 node += NODE_SZ_STR(node) - 1;
12479 node = NEXTOPER(node);
12482 node = NEXTOPER(node);
12483 node += regarglen[(U8)op];
12485 if (op == CURLYX || op == OPEN)
12489 #ifdef DEBUG_DUMPUNTIL
12490 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
12495 #endif /* DEBUGGING */
12499 * c-indentation-style: bsd
12500 * c-basic-offset: 4
12501 * indent-tabs-mode: t
12504 * ex: set ts=8 sts=4 sw=4 noet: