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
84 extern const struct regexp_engine my_reg_engine;
89 #include "dquote_static.c"
90 #include "charclass_invlists.h"
91 #include "inline_invlist.c"
92 #include "unicode_constants.h"
94 #define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
96 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
103 # if defined(BUGGY_MSC6)
104 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
105 # pragma optimize("a",off)
106 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
107 # pragma optimize("w",on )
108 # endif /* BUGGY_MSC6 */
112 #define STATIC static
116 typedef struct RExC_state_t {
117 U32 flags; /* RXf_* are we folding, multilining? */
118 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
119 char *precomp; /* uncompiled string. */
120 REGEXP *rx_sv; /* The SV that is the regexp. */
121 regexp *rx; /* perl core regexp structure */
122 regexp_internal *rxi; /* internal data for regexp object pprivate field */
123 char *start; /* Start of input for compile */
124 char *end; /* End of input for compile */
125 char *parse; /* Input-scan pointer. */
126 I32 whilem_seen; /* number of WHILEM in this expr */
127 regnode *emit_start; /* Start of emitted-code area */
128 regnode *emit_bound; /* First regnode outside of the allocated space */
129 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
130 I32 naughty; /* How bad is this pattern? */
131 I32 sawback; /* Did we see \1, ...? */
133 I32 size; /* Code size. */
134 I32 npar; /* Capture buffer count, (OPEN). */
135 I32 cpar; /* Capture buffer count, (CLOSE). */
136 I32 nestroot; /* root parens we are in - used by accept */
139 regnode **open_parens; /* pointers to open parens */
140 regnode **close_parens; /* pointers to close parens */
141 regnode *opend; /* END node in program */
142 I32 utf8; /* whether the pattern is utf8 or not */
143 I32 orig_utf8; /* whether the pattern was originally in utf8 */
144 /* XXX use this for future optimisation of case
145 * where pattern must be upgraded to utf8. */
146 I32 uni_semantics; /* If a d charset modifier should use unicode
147 rules, even if the pattern is not in
149 HV *paren_names; /* Paren names */
151 regnode **recurse; /* Recurse regops */
152 I32 recurse_count; /* Number of recurse regops */
155 I32 override_recoding;
156 I32 in_multi_char_class;
157 struct reg_code_block *code_blocks; /* positions of literal (?{})
159 int num_code_blocks; /* size of code_blocks[] */
160 int code_index; /* next code_blocks[] slot */
162 char *starttry; /* -Dr: where regtry was called. */
163 #define RExC_starttry (pRExC_state->starttry)
165 SV *runtime_code_qr; /* qr with the runtime code blocks */
167 const char *lastparse;
169 AV *paren_name_list; /* idx -> name */
170 #define RExC_lastparse (pRExC_state->lastparse)
171 #define RExC_lastnum (pRExC_state->lastnum)
172 #define RExC_paren_name_list (pRExC_state->paren_name_list)
176 #define RExC_flags (pRExC_state->flags)
177 #define RExC_pm_flags (pRExC_state->pm_flags)
178 #define RExC_precomp (pRExC_state->precomp)
179 #define RExC_rx_sv (pRExC_state->rx_sv)
180 #define RExC_rx (pRExC_state->rx)
181 #define RExC_rxi (pRExC_state->rxi)
182 #define RExC_start (pRExC_state->start)
183 #define RExC_end (pRExC_state->end)
184 #define RExC_parse (pRExC_state->parse)
185 #define RExC_whilem_seen (pRExC_state->whilem_seen)
186 #ifdef RE_TRACK_PATTERN_OFFSETS
187 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
189 #define RExC_emit (pRExC_state->emit)
190 #define RExC_emit_start (pRExC_state->emit_start)
191 #define RExC_emit_bound (pRExC_state->emit_bound)
192 #define RExC_naughty (pRExC_state->naughty)
193 #define RExC_sawback (pRExC_state->sawback)
194 #define RExC_seen (pRExC_state->seen)
195 #define RExC_size (pRExC_state->size)
196 #define RExC_npar (pRExC_state->npar)
197 #define RExC_nestroot (pRExC_state->nestroot)
198 #define RExC_extralen (pRExC_state->extralen)
199 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
200 #define RExC_utf8 (pRExC_state->utf8)
201 #define RExC_uni_semantics (pRExC_state->uni_semantics)
202 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
203 #define RExC_open_parens (pRExC_state->open_parens)
204 #define RExC_close_parens (pRExC_state->close_parens)
205 #define RExC_opend (pRExC_state->opend)
206 #define RExC_paren_names (pRExC_state->paren_names)
207 #define RExC_recurse (pRExC_state->recurse)
208 #define RExC_recurse_count (pRExC_state->recurse_count)
209 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
210 #define RExC_contains_locale (pRExC_state->contains_locale)
211 #define RExC_override_recoding (pRExC_state->override_recoding)
212 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
215 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
216 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
217 ((*s) == '{' && regcurly(s, FALSE)))
220 #undef SPSTART /* dratted cpp namespace... */
223 * Flags to be passed up and down.
225 #define WORST 0 /* Worst case. */
226 #define HASWIDTH 0x01 /* Known to match non-null strings. */
228 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
229 * character. (There needs to be a case: in the switch statement in regexec.c
230 * for any node marked SIMPLE.) Note that this is not the same thing as
233 #define SPSTART 0x04 /* Starts with * or + */
234 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
235 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
236 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
238 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
240 /* whether trie related optimizations are enabled */
241 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
242 #define TRIE_STUDY_OPT
243 #define FULL_TRIE_STUDY
249 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
250 #define PBITVAL(paren) (1 << ((paren) & 7))
251 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
252 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
253 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
255 #define REQUIRE_UTF8 STMT_START { \
257 *flagp = RESTART_UTF8; \
262 /* This converts the named class defined in regcomp.h to its equivalent class
263 * number defined in handy.h. */
264 #define namedclass_to_classnum(class) ((int) ((class) / 2))
265 #define classnum_to_namedclass(classnum) ((classnum) * 2)
267 /* About scan_data_t.
269 During optimisation we recurse through the regexp program performing
270 various inplace (keyhole style) optimisations. In addition study_chunk
271 and scan_commit populate this data structure with information about
272 what strings MUST appear in the pattern. We look for the longest
273 string that must appear at a fixed location, and we look for the
274 longest string that may appear at a floating location. So for instance
279 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
280 strings (because they follow a .* construct). study_chunk will identify
281 both FOO and BAR as being the longest fixed and floating strings respectively.
283 The strings can be composites, for instance
287 will result in a composite fixed substring 'foo'.
289 For each string some basic information is maintained:
291 - offset or min_offset
292 This is the position the string must appear at, or not before.
293 It also implicitly (when combined with minlenp) tells us how many
294 characters must match before the string we are searching for.
295 Likewise when combined with minlenp and the length of the string it
296 tells us how many characters must appear after the string we have
300 Only used for floating strings. This is the rightmost point that
301 the string can appear at. If set to I32 max it indicates that the
302 string can occur infinitely far to the right.
305 A pointer to the minimum number of characters of the pattern that the
306 string was found inside. This is important as in the case of positive
307 lookahead or positive lookbehind we can have multiple patterns
312 The minimum length of the pattern overall is 3, the minimum length
313 of the lookahead part is 3, but the minimum length of the part that
314 will actually match is 1. So 'FOO's minimum length is 3, but the
315 minimum length for the F is 1. This is important as the minimum length
316 is used to determine offsets in front of and behind the string being
317 looked for. Since strings can be composites this is the length of the
318 pattern at the time it was committed with a scan_commit. Note that
319 the length is calculated by study_chunk, so that the minimum lengths
320 are not known until the full pattern has been compiled, thus the
321 pointer to the value.
325 In the case of lookbehind the string being searched for can be
326 offset past the start point of the final matching string.
327 If this value was just blithely removed from the min_offset it would
328 invalidate some of the calculations for how many chars must match
329 before or after (as they are derived from min_offset and minlen and
330 the length of the string being searched for).
331 When the final pattern is compiled and the data is moved from the
332 scan_data_t structure into the regexp structure the information
333 about lookbehind is factored in, with the information that would
334 have been lost precalculated in the end_shift field for the
337 The fields pos_min and pos_delta are used to store the minimum offset
338 and the delta to the maximum offset at the current point in the pattern.
342 typedef struct scan_data_t {
343 /*I32 len_min; unused */
344 /*I32 len_delta; unused */
348 I32 last_end; /* min value, <0 unless valid. */
351 SV **longest; /* Either &l_fixed, or &l_float. */
352 SV *longest_fixed; /* longest fixed string found in pattern */
353 I32 offset_fixed; /* offset where it starts */
354 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
355 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
356 SV *longest_float; /* longest floating string found in pattern */
357 I32 offset_float_min; /* earliest point in string it can appear */
358 I32 offset_float_max; /* latest point in string it can appear */
359 I32 *minlen_float; /* pointer to the minlen relevant to the string */
360 I32 lookbehind_float; /* is the position of the string modified by LB */
364 struct regnode_charclass_class *start_class;
368 * Forward declarations for pregcomp()'s friends.
371 static const scan_data_t zero_scan_data =
372 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
374 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
375 #define SF_BEFORE_SEOL 0x0001
376 #define SF_BEFORE_MEOL 0x0002
377 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
378 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
381 # define SF_FIX_SHIFT_EOL (0+2)
382 # define SF_FL_SHIFT_EOL (0+4)
384 # define SF_FIX_SHIFT_EOL (+2)
385 # define SF_FL_SHIFT_EOL (+4)
388 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
389 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
391 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
392 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
393 #define SF_IS_INF 0x0040
394 #define SF_HAS_PAR 0x0080
395 #define SF_IN_PAR 0x0100
396 #define SF_HAS_EVAL 0x0200
397 #define SCF_DO_SUBSTR 0x0400
398 #define SCF_DO_STCLASS_AND 0x0800
399 #define SCF_DO_STCLASS_OR 0x1000
400 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
401 #define SCF_WHILEM_VISITED_POS 0x2000
403 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
404 #define SCF_SEEN_ACCEPT 0x8000
406 #define UTF cBOOL(RExC_utf8)
408 /* The enums for all these are ordered so things work out correctly */
409 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
410 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
411 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
412 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
413 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
414 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
415 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
417 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
419 #define OOB_NAMEDCLASS -1
421 /* There is no code point that is out-of-bounds, so this is problematic. But
422 * its only current use is to initialize a variable that is always set before
424 #define OOB_UNICODE 0xDEADBEEF
426 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
427 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
430 /* length of regex to show in messages that don't mark a position within */
431 #define RegexLengthToShowInErrorMessages 127
434 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
435 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
436 * op/pragma/warn/regcomp.
438 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
439 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
441 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
444 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
445 * arg. Show regex, up to a maximum length. If it's too long, chop and add
448 #define _FAIL(code) STMT_START { \
449 const char *ellipses = ""; \
450 IV len = RExC_end - RExC_precomp; \
453 SAVEFREESV(RExC_rx_sv); \
454 if (len > RegexLengthToShowInErrorMessages) { \
455 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
456 len = RegexLengthToShowInErrorMessages - 10; \
462 #define FAIL(msg) _FAIL( \
463 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
464 msg, (int)len, RExC_precomp, ellipses))
466 #define FAIL2(msg,arg) _FAIL( \
467 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
468 arg, (int)len, RExC_precomp, ellipses))
471 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
473 #define Simple_vFAIL(m) STMT_START { \
474 const IV offset = RExC_parse - RExC_precomp; \
475 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
476 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
480 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
482 #define vFAIL(m) STMT_START { \
484 SAVEFREESV(RExC_rx_sv); \
489 * Like Simple_vFAIL(), but accepts two arguments.
491 #define Simple_vFAIL2(m,a1) STMT_START { \
492 const IV offset = RExC_parse - RExC_precomp; \
493 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
494 (int)offset, RExC_precomp, RExC_precomp + offset); \
498 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
500 #define vFAIL2(m,a1) STMT_START { \
502 SAVEFREESV(RExC_rx_sv); \
503 Simple_vFAIL2(m, a1); \
508 * Like Simple_vFAIL(), but accepts three arguments.
510 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
511 const IV offset = RExC_parse - RExC_precomp; \
512 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
513 (int)offset, RExC_precomp, RExC_precomp + offset); \
517 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
519 #define vFAIL3(m,a1,a2) STMT_START { \
521 SAVEFREESV(RExC_rx_sv); \
522 Simple_vFAIL3(m, a1, a2); \
526 * Like Simple_vFAIL(), but accepts four arguments.
528 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
529 const IV offset = RExC_parse - RExC_precomp; \
530 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
531 (int)offset, RExC_precomp, RExC_precomp + offset); \
534 #define vFAIL4(m,a1,a2,a3) STMT_START { \
536 SAVEFREESV(RExC_rx_sv); \
537 Simple_vFAIL4(m, a1, a2, a3); \
540 /* m is not necessarily a "literal string", in this macro */
541 #define reg_warn_non_literal_string(loc, m) STMT_START { \
542 const IV offset = loc - RExC_precomp; \
543 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
544 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
547 #define ckWARNreg(loc,m) STMT_START { \
548 const IV offset = loc - RExC_precomp; \
549 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
550 (int)offset, RExC_precomp, RExC_precomp + offset); \
553 #define vWARN_dep(loc, m) STMT_START { \
554 const IV offset = loc - RExC_precomp; \
555 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
556 (int)offset, RExC_precomp, RExC_precomp + offset); \
559 #define ckWARNdep(loc,m) STMT_START { \
560 const IV offset = loc - RExC_precomp; \
561 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
563 (int)offset, RExC_precomp, RExC_precomp + offset); \
566 #define ckWARNregdep(loc,m) STMT_START { \
567 const IV offset = loc - RExC_precomp; \
568 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
570 (int)offset, RExC_precomp, RExC_precomp + offset); \
573 #define ckWARN2regdep(loc,m, a1) STMT_START { \
574 const IV offset = loc - RExC_precomp; \
575 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
577 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
580 #define ckWARN2reg(loc, m, a1) STMT_START { \
581 const IV offset = loc - RExC_precomp; \
582 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
583 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
586 #define vWARN3(loc, m, a1, a2) STMT_START { \
587 const IV offset = loc - RExC_precomp; \
588 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
589 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
592 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
593 const IV offset = loc - RExC_precomp; \
594 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
595 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
598 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
599 const IV offset = loc - RExC_precomp; \
600 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
601 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
604 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
605 const IV offset = loc - RExC_precomp; \
606 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
607 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
610 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
611 const IV offset = loc - RExC_precomp; \
612 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
613 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
617 /* Allow for side effects in s */
618 #define REGC(c,s) STMT_START { \
619 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
622 /* Macros for recording node offsets. 20001227 mjd@plover.com
623 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
624 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
625 * Element 0 holds the number n.
626 * Position is 1 indexed.
628 #ifndef RE_TRACK_PATTERN_OFFSETS
629 #define Set_Node_Offset_To_R(node,byte)
630 #define Set_Node_Offset(node,byte)
631 #define Set_Cur_Node_Offset
632 #define Set_Node_Length_To_R(node,len)
633 #define Set_Node_Length(node,len)
634 #define Set_Node_Cur_Length(node)
635 #define Node_Offset(n)
636 #define Node_Length(n)
637 #define Set_Node_Offset_Length(node,offset,len)
638 #define ProgLen(ri) ri->u.proglen
639 #define SetProgLen(ri,x) ri->u.proglen = x
641 #define ProgLen(ri) ri->u.offsets[0]
642 #define SetProgLen(ri,x) ri->u.offsets[0] = x
643 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
645 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
646 __LINE__, (int)(node), (int)(byte))); \
648 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
650 RExC_offsets[2*(node)-1] = (byte); \
655 #define Set_Node_Offset(node,byte) \
656 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
657 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
659 #define Set_Node_Length_To_R(node,len) STMT_START { \
661 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
662 __LINE__, (int)(node), (int)(len))); \
664 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
666 RExC_offsets[2*(node)] = (len); \
671 #define Set_Node_Length(node,len) \
672 Set_Node_Length_To_R((node)-RExC_emit_start, len)
673 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
674 #define Set_Node_Cur_Length(node) \
675 Set_Node_Length(node, RExC_parse - parse_start)
677 /* Get offsets and lengths */
678 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
679 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
681 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
682 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
683 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
687 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
688 #define EXPERIMENTAL_INPLACESCAN
689 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
691 #define DEBUG_STUDYDATA(str,data,depth) \
692 DEBUG_OPTIMISE_MORE_r(if(data){ \
693 PerlIO_printf(Perl_debug_log, \
694 "%*s" str "Pos:%"IVdf"/%"IVdf \
695 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
696 (int)(depth)*2, "", \
697 (IV)((data)->pos_min), \
698 (IV)((data)->pos_delta), \
699 (UV)((data)->flags), \
700 (IV)((data)->whilem_c), \
701 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
702 is_inf ? "INF " : "" \
704 if ((data)->last_found) \
705 PerlIO_printf(Perl_debug_log, \
706 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
707 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
708 SvPVX_const((data)->last_found), \
709 (IV)((data)->last_end), \
710 (IV)((data)->last_start_min), \
711 (IV)((data)->last_start_max), \
712 ((data)->longest && \
713 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
714 SvPVX_const((data)->longest_fixed), \
715 (IV)((data)->offset_fixed), \
716 ((data)->longest && \
717 (data)->longest==&((data)->longest_float)) ? "*" : "", \
718 SvPVX_const((data)->longest_float), \
719 (IV)((data)->offset_float_min), \
720 (IV)((data)->offset_float_max) \
722 PerlIO_printf(Perl_debug_log,"\n"); \
725 /* Mark that we cannot extend a found fixed substring at this point.
726 Update the longest found anchored substring and the longest found
727 floating substrings if needed. */
730 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
732 const STRLEN l = CHR_SVLEN(data->last_found);
733 const STRLEN old_l = CHR_SVLEN(*data->longest);
734 GET_RE_DEBUG_FLAGS_DECL;
736 PERL_ARGS_ASSERT_SCAN_COMMIT;
738 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
739 SvSetMagicSV(*data->longest, data->last_found);
740 if (*data->longest == data->longest_fixed) {
741 data->offset_fixed = l ? data->last_start_min : data->pos_min;
742 if (data->flags & SF_BEFORE_EOL)
744 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
746 data->flags &= ~SF_FIX_BEFORE_EOL;
747 data->minlen_fixed=minlenp;
748 data->lookbehind_fixed=0;
750 else { /* *data->longest == data->longest_float */
751 data->offset_float_min = l ? data->last_start_min : data->pos_min;
752 data->offset_float_max = (l
753 ? data->last_start_max
754 : (data->pos_delta == I32_MAX ? I32_MAX : data->pos_min + data->pos_delta));
755 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
756 data->offset_float_max = I32_MAX;
757 if (data->flags & SF_BEFORE_EOL)
759 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
761 data->flags &= ~SF_FL_BEFORE_EOL;
762 data->minlen_float=minlenp;
763 data->lookbehind_float=0;
766 SvCUR_set(data->last_found, 0);
768 SV * const sv = data->last_found;
769 if (SvUTF8(sv) && SvMAGICAL(sv)) {
770 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
776 data->flags &= ~SF_BEFORE_EOL;
777 DEBUG_STUDYDATA("commit: ",data,0);
780 /* These macros set, clear and test whether the synthetic start class ('ssc',
781 * given by the parameter) matches an empty string (EOS). This uses the
782 * 'next_off' field in the node, to save a bit in the flags field. The ssc
783 * stands alone, so there is never a next_off, so this field is otherwise
784 * unused. The EOS information is used only for compilation, but theoretically
785 * it could be passed on to the execution code. This could be used to store
786 * more than one bit of information, but only this one is currently used. */
787 #define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
788 #define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
789 #define TEST_SSC_EOS(node) cBOOL((node)->next_off)
791 /* Can match anything (initialization) */
793 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
795 PERL_ARGS_ASSERT_CL_ANYTHING;
797 ANYOF_BITMAP_SETALL(cl);
798 cl->flags = ANYOF_UNICODE_ALL;
801 /* If any portion of the regex is to operate under locale rules,
802 * initialization includes it. The reason this isn't done for all regexes
803 * is that the optimizer was written under the assumption that locale was
804 * all-or-nothing. Given the complexity and lack of documentation in the
805 * optimizer, and that there are inadequate test cases for locale, so many
806 * parts of it may not work properly, it is safest to avoid locale unless
808 if (RExC_contains_locale) {
809 ANYOF_CLASS_SETALL(cl); /* /l uses class */
810 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
813 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
817 /* Can match anything (initialization) */
819 S_cl_is_anything(const struct regnode_charclass_class *cl)
823 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
825 for (value = 0; value < ANYOF_MAX; value += 2)
826 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
828 if (!(cl->flags & ANYOF_UNICODE_ALL))
830 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
835 /* Can match anything (initialization) */
837 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
839 PERL_ARGS_ASSERT_CL_INIT;
841 Zero(cl, 1, struct regnode_charclass_class);
843 cl_anything(pRExC_state, cl);
844 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
847 /* These two functions currently do the exact same thing */
848 #define cl_init_zero S_cl_init
850 /* 'AND' a given class with another one. Can create false positives. 'cl'
851 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
852 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
854 S_cl_and(struct regnode_charclass_class *cl,
855 const struct regnode_charclass_class *and_with)
857 PERL_ARGS_ASSERT_CL_AND;
859 assert(PL_regkind[and_with->type] == ANYOF);
861 /* I (khw) am not sure all these restrictions are necessary XXX */
862 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
863 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
864 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
865 && !(and_with->flags & ANYOF_LOC_FOLD)
866 && !(cl->flags & ANYOF_LOC_FOLD)) {
869 if (and_with->flags & ANYOF_INVERT)
870 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
871 cl->bitmap[i] &= ~and_with->bitmap[i];
873 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
874 cl->bitmap[i] &= and_with->bitmap[i];
875 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
877 if (and_with->flags & ANYOF_INVERT) {
879 /* Here, the and'ed node is inverted. Get the AND of the flags that
880 * aren't affected by the inversion. Those that are affected are
881 * handled individually below */
882 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
883 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
884 cl->flags |= affected_flags;
886 /* We currently don't know how to deal with things that aren't in the
887 * bitmap, but we know that the intersection is no greater than what
888 * is already in cl, so let there be false positives that get sorted
889 * out after the synthetic start class succeeds, and the node is
890 * matched for real. */
892 /* The inversion of these two flags indicate that the resulting
893 * intersection doesn't have them */
894 if (and_with->flags & ANYOF_UNICODE_ALL) {
895 cl->flags &= ~ANYOF_UNICODE_ALL;
897 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
898 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
901 else { /* and'd node is not inverted */
902 U8 outside_bitmap_but_not_utf8; /* Temp variable */
904 if (! ANYOF_NONBITMAP(and_with)) {
906 /* Here 'and_with' doesn't match anything outside the bitmap
907 * (except possibly ANYOF_UNICODE_ALL), which means the
908 * intersection can't either, except for ANYOF_UNICODE_ALL, in
909 * which case we don't know what the intersection is, but it's no
910 * greater than what cl already has, so can just leave it alone,
911 * with possible false positives */
912 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
913 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
914 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
917 else if (! ANYOF_NONBITMAP(cl)) {
919 /* Here, 'and_with' does match something outside the bitmap, and cl
920 * doesn't have a list of things to match outside the bitmap. If
921 * cl can match all code points above 255, the intersection will
922 * be those above-255 code points that 'and_with' matches. If cl
923 * can't match all Unicode code points, it means that it can't
924 * match anything outside the bitmap (since the 'if' that got us
925 * into this block tested for that), so we leave the bitmap empty.
927 if (cl->flags & ANYOF_UNICODE_ALL) {
928 ARG_SET(cl, ARG(and_with));
930 /* and_with's ARG may match things that don't require UTF8.
931 * And now cl's will too, in spite of this being an 'and'. See
932 * the comments below about the kludge */
933 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
937 /* Here, both 'and_with' and cl match something outside the
938 * bitmap. Currently we do not do the intersection, so just match
939 * whatever cl had at the beginning. */
943 /* Take the intersection of the two sets of flags. However, the
944 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
945 * kludge around the fact that this flag is not treated like the others
946 * which are initialized in cl_anything(). The way the optimizer works
947 * is that the synthetic start class (SSC) is initialized to match
948 * anything, and then the first time a real node is encountered, its
949 * values are AND'd with the SSC's with the result being the values of
950 * the real node. However, there are paths through the optimizer where
951 * the AND never gets called, so those initialized bits are set
952 * inappropriately, which is not usually a big deal, as they just cause
953 * false positives in the SSC, which will just mean a probably
954 * imperceptible slow down in execution. However this bit has a
955 * higher false positive consequence in that it can cause utf8.pm,
956 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
957 * bigger slowdown and also causes significant extra memory to be used.
958 * In order to prevent this, the code now takes a different tack. The
959 * bit isn't set unless some part of the regular expression needs it,
960 * but once set it won't get cleared. This means that these extra
961 * modules won't get loaded unless there was some path through the
962 * pattern that would have required them anyway, and so any false
963 * positives that occur by not ANDing them out when they could be
964 * aren't as severe as they would be if we treated this bit like all
966 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
967 & ANYOF_NONBITMAP_NON_UTF8;
968 cl->flags &= and_with->flags;
969 cl->flags |= outside_bitmap_but_not_utf8;
973 /* 'OR' a given class with another one. Can create false positives. 'cl'
974 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
975 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
977 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
979 PERL_ARGS_ASSERT_CL_OR;
981 if (or_with->flags & ANYOF_INVERT) {
983 /* Here, the or'd node is to be inverted. This means we take the
984 * complement of everything not in the bitmap, but currently we don't
985 * know what that is, so give up and match anything */
986 if (ANYOF_NONBITMAP(or_with)) {
987 cl_anything(pRExC_state, cl);
990 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
991 * <= (B1 | !B2) | (CL1 | !CL2)
992 * which is wasteful if CL2 is small, but we ignore CL2:
993 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
994 * XXXX Can we handle case-fold? Unclear:
995 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
996 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
998 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
999 && !(or_with->flags & ANYOF_LOC_FOLD)
1000 && !(cl->flags & ANYOF_LOC_FOLD) ) {
1003 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1004 cl->bitmap[i] |= ~or_with->bitmap[i];
1005 } /* XXXX: logic is complicated otherwise */
1007 cl_anything(pRExC_state, cl);
1010 /* And, we can just take the union of the flags that aren't affected
1011 * by the inversion */
1012 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1014 /* For the remaining flags:
1015 ANYOF_UNICODE_ALL and inverted means to not match anything above
1016 255, which means that the union with cl should just be
1017 what cl has in it, so can ignore this flag
1018 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1019 is 127-255 to match them, but then invert that, so the
1020 union with cl should just be what cl has in it, so can
1023 } else { /* 'or_with' is not inverted */
1024 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1025 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1026 && (!(or_with->flags & ANYOF_LOC_FOLD)
1027 || (cl->flags & ANYOF_LOC_FOLD)) ) {
1030 /* OR char bitmap and class bitmap separately */
1031 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1032 cl->bitmap[i] |= or_with->bitmap[i];
1033 if (or_with->flags & ANYOF_CLASS) {
1034 ANYOF_CLASS_OR(or_with, cl);
1037 else { /* XXXX: logic is complicated, leave it along for a moment. */
1038 cl_anything(pRExC_state, cl);
1041 if (ANYOF_NONBITMAP(or_with)) {
1043 /* Use the added node's outside-the-bit-map match if there isn't a
1044 * conflict. If there is a conflict (both nodes match something
1045 * outside the bitmap, but what they match outside is not the same
1046 * pointer, and hence not easily compared until XXX we extend
1047 * inversion lists this far), give up and allow the start class to
1048 * match everything outside the bitmap. If that stuff is all above
1049 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1050 if (! ANYOF_NONBITMAP(cl)) {
1051 ARG_SET(cl, ARG(or_with));
1053 else if (ARG(cl) != ARG(or_with)) {
1055 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1056 cl_anything(pRExC_state, cl);
1059 cl->flags |= ANYOF_UNICODE_ALL;
1064 /* Take the union */
1065 cl->flags |= or_with->flags;
1069 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1070 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1071 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1072 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1077 dump_trie(trie,widecharmap,revcharmap)
1078 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1079 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1081 These routines dump out a trie in a somewhat readable format.
1082 The _interim_ variants are used for debugging the interim
1083 tables that are used to generate the final compressed
1084 representation which is what dump_trie expects.
1086 Part of the reason for their existence is to provide a form
1087 of documentation as to how the different representations function.
1092 Dumps the final compressed table form of the trie to Perl_debug_log.
1093 Used for debugging make_trie().
1097 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1098 AV *revcharmap, U32 depth)
1101 SV *sv=sv_newmortal();
1102 int colwidth= widecharmap ? 6 : 4;
1104 GET_RE_DEBUG_FLAGS_DECL;
1106 PERL_ARGS_ASSERT_DUMP_TRIE;
1108 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1109 (int)depth * 2 + 2,"",
1110 "Match","Base","Ofs" );
1112 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1113 SV ** const tmp = av_fetch( revcharmap, state, 0);
1115 PerlIO_printf( Perl_debug_log, "%*s",
1117 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1118 PL_colors[0], PL_colors[1],
1119 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1120 PERL_PV_ESCAPE_FIRSTCHAR
1125 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1126 (int)depth * 2 + 2,"");
1128 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1129 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1130 PerlIO_printf( Perl_debug_log, "\n");
1132 for( state = 1 ; state < trie->statecount ; state++ ) {
1133 const U32 base = trie->states[ state ].trans.base;
1135 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1137 if ( trie->states[ state ].wordnum ) {
1138 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1140 PerlIO_printf( Perl_debug_log, "%6s", "" );
1143 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1148 while( ( base + ofs < trie->uniquecharcount ) ||
1149 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1150 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1153 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1155 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1156 if ( ( base + ofs >= trie->uniquecharcount ) &&
1157 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1158 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1160 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1162 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1164 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1168 PerlIO_printf( Perl_debug_log, "]");
1171 PerlIO_printf( Perl_debug_log, "\n" );
1173 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1174 for (word=1; word <= trie->wordcount; word++) {
1175 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1176 (int)word, (int)(trie->wordinfo[word].prev),
1177 (int)(trie->wordinfo[word].len));
1179 PerlIO_printf(Perl_debug_log, "\n" );
1182 Dumps a fully constructed but uncompressed trie in list form.
1183 List tries normally only are used for construction when the number of
1184 possible chars (trie->uniquecharcount) is very high.
1185 Used for debugging make_trie().
1188 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1189 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1193 SV *sv=sv_newmortal();
1194 int colwidth= widecharmap ? 6 : 4;
1195 GET_RE_DEBUG_FLAGS_DECL;
1197 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1199 /* print out the table precompression. */
1200 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1201 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1202 "------:-----+-----------------\n" );
1204 for( state=1 ; state < next_alloc ; state ++ ) {
1207 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1208 (int)depth * 2 + 2,"", (UV)state );
1209 if ( ! trie->states[ state ].wordnum ) {
1210 PerlIO_printf( Perl_debug_log, "%5s| ","");
1212 PerlIO_printf( Perl_debug_log, "W%4x| ",
1213 trie->states[ state ].wordnum
1216 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1217 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1219 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1221 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1222 PL_colors[0], PL_colors[1],
1223 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1224 PERL_PV_ESCAPE_FIRSTCHAR
1226 TRIE_LIST_ITEM(state,charid).forid,
1227 (UV)TRIE_LIST_ITEM(state,charid).newstate
1230 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1231 (int)((depth * 2) + 14), "");
1234 PerlIO_printf( Perl_debug_log, "\n");
1239 Dumps a fully constructed but uncompressed trie in table form.
1240 This is the normal DFA style state transition table, with a few
1241 twists to facilitate compression later.
1242 Used for debugging make_trie().
1245 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1246 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1251 SV *sv=sv_newmortal();
1252 int colwidth= widecharmap ? 6 : 4;
1253 GET_RE_DEBUG_FLAGS_DECL;
1255 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1258 print out the table precompression so that we can do a visual check
1259 that they are identical.
1262 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1264 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1265 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1267 PerlIO_printf( Perl_debug_log, "%*s",
1269 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1270 PL_colors[0], PL_colors[1],
1271 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1272 PERL_PV_ESCAPE_FIRSTCHAR
1278 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1280 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1281 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1284 PerlIO_printf( Perl_debug_log, "\n" );
1286 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1288 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1289 (int)depth * 2 + 2,"",
1290 (UV)TRIE_NODENUM( state ) );
1292 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1293 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1295 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1297 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1299 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1300 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1302 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1303 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1311 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1312 startbranch: the first branch in the whole branch sequence
1313 first : start branch of sequence of branch-exact nodes.
1314 May be the same as startbranch
1315 last : Thing following the last branch.
1316 May be the same as tail.
1317 tail : item following the branch sequence
1318 count : words in the sequence
1319 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1320 depth : indent depth
1322 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1324 A trie is an N'ary tree where the branches are determined by digital
1325 decomposition of the key. IE, at the root node you look up the 1st character and
1326 follow that branch repeat until you find the end of the branches. Nodes can be
1327 marked as "accepting" meaning they represent a complete word. Eg:
1331 would convert into the following structure. Numbers represent states, letters
1332 following numbers represent valid transitions on the letter from that state, if
1333 the number is in square brackets it represents an accepting state, otherwise it
1334 will be in parenthesis.
1336 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1340 (1) +-i->(6)-+-s->[7]
1342 +-s->(3)-+-h->(4)-+-e->[5]
1344 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1346 This shows that when matching against the string 'hers' we will begin at state 1
1347 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1348 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1349 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1350 single traverse. We store a mapping from accepting to state to which word was
1351 matched, and then when we have multiple possibilities we try to complete the
1352 rest of the regex in the order in which they occured in the alternation.
1354 The only prior NFA like behaviour that would be changed by the TRIE support is
1355 the silent ignoring of duplicate alternations which are of the form:
1357 / (DUPE|DUPE) X? (?{ ... }) Y /x
1359 Thus EVAL blocks following a trie may be called a different number of times with
1360 and without the optimisation. With the optimisations dupes will be silently
1361 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1362 the following demonstrates:
1364 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1366 which prints out 'word' three times, but
1368 'words'=~/(word|word|word)(?{ print $1 })S/
1370 which doesnt print it out at all. This is due to other optimisations kicking in.
1372 Example of what happens on a structural level:
1374 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1376 1: CURLYM[1] {1,32767}(18)
1387 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1388 and should turn into:
1390 1: CURLYM[1] {1,32767}(18)
1392 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1400 Cases where tail != last would be like /(?foo|bar)baz/:
1410 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1411 and would end up looking like:
1414 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1421 d = uvuni_to_utf8_flags(d, uv, 0);
1423 is the recommended Unicode-aware way of saying
1428 #define TRIE_STORE_REVCHAR(val) \
1431 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1432 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1433 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1434 SvCUR_set(zlopp, kapow - flrbbbbb); \
1437 av_push(revcharmap, zlopp); \
1439 char ooooff = (char)val; \
1440 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1444 #define TRIE_READ_CHAR STMT_START { \
1447 /* if it is UTF then it is either already folded, or does not need folding */ \
1448 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1450 else if (folder == PL_fold_latin1) { \
1451 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1452 if ( foldlen > 0 ) { \
1453 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1459 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1460 skiplen = UNISKIP(uvc); \
1461 foldlen -= skiplen; \
1462 scan = foldbuf + skiplen; \
1465 /* raw data, will be folded later if needed */ \
1473 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1474 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1475 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1476 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1478 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1479 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1480 TRIE_LIST_CUR( state )++; \
1483 #define TRIE_LIST_NEW(state) STMT_START { \
1484 Newxz( trie->states[ state ].trans.list, \
1485 4, reg_trie_trans_le ); \
1486 TRIE_LIST_CUR( state ) = 1; \
1487 TRIE_LIST_LEN( state ) = 4; \
1490 #define TRIE_HANDLE_WORD(state) STMT_START { \
1491 U16 dupe= trie->states[ state ].wordnum; \
1492 regnode * const noper_next = regnext( noper ); \
1495 /* store the word for dumping */ \
1497 if (OP(noper) != NOTHING) \
1498 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1500 tmp = newSVpvn_utf8( "", 0, UTF ); \
1501 av_push( trie_words, tmp ); \
1505 trie->wordinfo[curword].prev = 0; \
1506 trie->wordinfo[curword].len = wordlen; \
1507 trie->wordinfo[curword].accept = state; \
1509 if ( noper_next < tail ) { \
1511 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1512 trie->jump[curword] = (U16)(noper_next - convert); \
1514 jumper = noper_next; \
1516 nextbranch= regnext(cur); \
1520 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1521 /* chain, so that when the bits of chain are later */\
1522 /* linked together, the dups appear in the chain */\
1523 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1524 trie->wordinfo[dupe].prev = curword; \
1526 /* we haven't inserted this word yet. */ \
1527 trie->states[ state ].wordnum = curword; \
1532 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1533 ( ( base + charid >= ucharcount \
1534 && base + charid < ubound \
1535 && state == trie->trans[ base - ucharcount + charid ].check \
1536 && trie->trans[ base - ucharcount + charid ].next ) \
1537 ? trie->trans[ base - ucharcount + charid ].next \
1538 : ( state==1 ? special : 0 ) \
1542 #define MADE_JUMP_TRIE 2
1543 #define MADE_EXACT_TRIE 4
1546 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1549 /* first pass, loop through and scan words */
1550 reg_trie_data *trie;
1551 HV *widecharmap = NULL;
1552 AV *revcharmap = newAV();
1554 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1559 regnode *jumper = NULL;
1560 regnode *nextbranch = NULL;
1561 regnode *convert = NULL;
1562 U32 *prev_states; /* temp array mapping each state to previous one */
1563 /* we just use folder as a flag in utf8 */
1564 const U8 * folder = NULL;
1567 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1568 AV *trie_words = NULL;
1569 /* along with revcharmap, this only used during construction but both are
1570 * useful during debugging so we store them in the struct when debugging.
1573 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1574 STRLEN trie_charcount=0;
1576 SV *re_trie_maxbuff;
1577 GET_RE_DEBUG_FLAGS_DECL;
1579 PERL_ARGS_ASSERT_MAKE_TRIE;
1581 PERL_UNUSED_ARG(depth);
1588 case EXACTFU_TRICKYFOLD:
1589 case EXACTFU: folder = PL_fold_latin1; break;
1590 case EXACTF: folder = PL_fold; break;
1591 case EXACTFL: folder = PL_fold_locale; break;
1592 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1595 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1597 trie->startstate = 1;
1598 trie->wordcount = word_count;
1599 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1600 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1602 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1603 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1604 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1607 trie_words = newAV();
1610 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1611 if (!SvIOK(re_trie_maxbuff)) {
1612 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1614 DEBUG_TRIE_COMPILE_r({
1615 PerlIO_printf( Perl_debug_log,
1616 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1617 (int)depth * 2 + 2, "",
1618 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1619 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1623 /* Find the node we are going to overwrite */
1624 if ( first == startbranch && OP( last ) != BRANCH ) {
1625 /* whole branch chain */
1628 /* branch sub-chain */
1629 convert = NEXTOPER( first );
1632 /* -- First loop and Setup --
1634 We first traverse the branches and scan each word to determine if it
1635 contains widechars, and how many unique chars there are, this is
1636 important as we have to build a table with at least as many columns as we
1639 We use an array of integers to represent the character codes 0..255
1640 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1641 native representation of the character value as the key and IV's for the
1644 *TODO* If we keep track of how many times each character is used we can
1645 remap the columns so that the table compression later on is more
1646 efficient in terms of memory by ensuring the most common value is in the
1647 middle and the least common are on the outside. IMO this would be better
1648 than a most to least common mapping as theres a decent chance the most
1649 common letter will share a node with the least common, meaning the node
1650 will not be compressible. With a middle is most common approach the worst
1651 case is when we have the least common nodes twice.
1655 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1656 regnode *noper = NEXTOPER( cur );
1657 const U8 *uc = (U8*)STRING( noper );
1658 const U8 *e = uc + STR_LEN( noper );
1660 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1662 const U8 *scan = (U8*)NULL;
1663 U32 wordlen = 0; /* required init */
1665 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1667 if (OP(noper) == NOTHING) {
1668 regnode *noper_next= regnext(noper);
1669 if (noper_next != tail && OP(noper_next) == flags) {
1671 uc= (U8*)STRING(noper);
1672 e= uc + STR_LEN(noper);
1673 trie->minlen= STR_LEN(noper);
1680 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1681 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1682 regardless of encoding */
1683 if (OP( noper ) == EXACTFU_SS) {
1684 /* false positives are ok, so just set this */
1685 TRIE_BITMAP_SET(trie,0xDF);
1688 for ( ; uc < e ; uc += len ) {
1689 TRIE_CHARCOUNT(trie)++;
1694 U8 folded= folder[ (U8) uvc ];
1695 if ( !trie->charmap[ folded ] ) {
1696 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1697 TRIE_STORE_REVCHAR( folded );
1700 if ( !trie->charmap[ uvc ] ) {
1701 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1702 TRIE_STORE_REVCHAR( uvc );
1705 /* store the codepoint in the bitmap, and its folded
1707 TRIE_BITMAP_SET(trie, uvc);
1709 /* store the folded codepoint */
1710 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1713 /* store first byte of utf8 representation of
1714 variant codepoints */
1715 if (! UNI_IS_INVARIANT(uvc)) {
1716 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1719 set_bit = 0; /* We've done our bit :-) */
1724 widecharmap = newHV();
1726 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1729 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1731 if ( !SvTRUE( *svpp ) ) {
1732 sv_setiv( *svpp, ++trie->uniquecharcount );
1733 TRIE_STORE_REVCHAR(uvc);
1737 if( cur == first ) {
1738 trie->minlen = chars;
1739 trie->maxlen = chars;
1740 } else if (chars < trie->minlen) {
1741 trie->minlen = chars;
1742 } else if (chars > trie->maxlen) {
1743 trie->maxlen = chars;
1745 if (OP( noper ) == EXACTFU_SS) {
1746 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1747 if (trie->minlen > 1)
1750 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1751 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1752 * - We assume that any such sequence might match a 2 byte string */
1753 if (trie->minlen > 2 )
1757 } /* end first pass */
1758 DEBUG_TRIE_COMPILE_r(
1759 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1760 (int)depth * 2 + 2,"",
1761 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1762 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1763 (int)trie->minlen, (int)trie->maxlen )
1767 We now know what we are dealing with in terms of unique chars and
1768 string sizes so we can calculate how much memory a naive
1769 representation using a flat table will take. If it's over a reasonable
1770 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1771 conservative but potentially much slower representation using an array
1774 At the end we convert both representations into the same compressed
1775 form that will be used in regexec.c for matching with. The latter
1776 is a form that cannot be used to construct with but has memory
1777 properties similar to the list form and access properties similar
1778 to the table form making it both suitable for fast searches and
1779 small enough that its feasable to store for the duration of a program.
1781 See the comment in the code where the compressed table is produced
1782 inplace from the flat tabe representation for an explanation of how
1783 the compression works.
1788 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1791 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1793 Second Pass -- Array Of Lists Representation
1795 Each state will be represented by a list of charid:state records
1796 (reg_trie_trans_le) the first such element holds the CUR and LEN
1797 points of the allocated array. (See defines above).
1799 We build the initial structure using the lists, and then convert
1800 it into the compressed table form which allows faster lookups
1801 (but cant be modified once converted).
1804 STRLEN transcount = 1;
1806 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1807 "%*sCompiling trie using list compiler\n",
1808 (int)depth * 2 + 2, ""));
1810 trie->states = (reg_trie_state *)
1811 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1812 sizeof(reg_trie_state) );
1816 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1818 regnode *noper = NEXTOPER( cur );
1819 U8 *uc = (U8*)STRING( noper );
1820 const U8 *e = uc + STR_LEN( noper );
1821 U32 state = 1; /* required init */
1822 U16 charid = 0; /* sanity init */
1823 U8 *scan = (U8*)NULL; /* sanity init */
1824 STRLEN foldlen = 0; /* required init */
1825 U32 wordlen = 0; /* required init */
1826 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1829 if (OP(noper) == NOTHING) {
1830 regnode *noper_next= regnext(noper);
1831 if (noper_next != tail && OP(noper_next) == flags) {
1833 uc= (U8*)STRING(noper);
1834 e= uc + STR_LEN(noper);
1838 if (OP(noper) != NOTHING) {
1839 for ( ; uc < e ; uc += len ) {
1844 charid = trie->charmap[ uvc ];
1846 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1850 charid=(U16)SvIV( *svpp );
1853 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1860 if ( !trie->states[ state ].trans.list ) {
1861 TRIE_LIST_NEW( state );
1863 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1864 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1865 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1870 newstate = next_alloc++;
1871 prev_states[newstate] = state;
1872 TRIE_LIST_PUSH( state, charid, newstate );
1877 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1881 TRIE_HANDLE_WORD(state);
1883 } /* end second pass */
1885 /* next alloc is the NEXT state to be allocated */
1886 trie->statecount = next_alloc;
1887 trie->states = (reg_trie_state *)
1888 PerlMemShared_realloc( trie->states,
1890 * sizeof(reg_trie_state) );
1892 /* and now dump it out before we compress it */
1893 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1894 revcharmap, next_alloc,
1898 trie->trans = (reg_trie_trans *)
1899 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1906 for( state=1 ; state < next_alloc ; state ++ ) {
1910 DEBUG_TRIE_COMPILE_MORE_r(
1911 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1915 if (trie->states[state].trans.list) {
1916 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1920 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1921 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1922 if ( forid < minid ) {
1924 } else if ( forid > maxid ) {
1928 if ( transcount < tp + maxid - minid + 1) {
1930 trie->trans = (reg_trie_trans *)
1931 PerlMemShared_realloc( trie->trans,
1933 * sizeof(reg_trie_trans) );
1934 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1936 base = trie->uniquecharcount + tp - minid;
1937 if ( maxid == minid ) {
1939 for ( ; zp < tp ; zp++ ) {
1940 if ( ! trie->trans[ zp ].next ) {
1941 base = trie->uniquecharcount + zp - minid;
1942 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1943 trie->trans[ zp ].check = state;
1949 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1950 trie->trans[ tp ].check = state;
1955 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1956 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1957 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1958 trie->trans[ tid ].check = state;
1960 tp += ( maxid - minid + 1 );
1962 Safefree(trie->states[ state ].trans.list);
1965 DEBUG_TRIE_COMPILE_MORE_r(
1966 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1969 trie->states[ state ].trans.base=base;
1971 trie->lasttrans = tp + 1;
1975 Second Pass -- Flat Table Representation.
1977 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1978 We know that we will need Charcount+1 trans at most to store the data
1979 (one row per char at worst case) So we preallocate both structures
1980 assuming worst case.
1982 We then construct the trie using only the .next slots of the entry
1985 We use the .check field of the first entry of the node temporarily to
1986 make compression both faster and easier by keeping track of how many non
1987 zero fields are in the node.
1989 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1992 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1993 number representing the first entry of the node, and state as a
1994 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1995 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1996 are 2 entrys per node. eg:
2004 The table is internally in the right hand, idx form. However as we also
2005 have to deal with the states array which is indexed by nodenum we have to
2006 use TRIE_NODENUM() to convert.
2009 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2010 "%*sCompiling trie using table compiler\n",
2011 (int)depth * 2 + 2, ""));
2013 trie->trans = (reg_trie_trans *)
2014 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2015 * trie->uniquecharcount + 1,
2016 sizeof(reg_trie_trans) );
2017 trie->states = (reg_trie_state *)
2018 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2019 sizeof(reg_trie_state) );
2020 next_alloc = trie->uniquecharcount + 1;
2023 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2025 regnode *noper = NEXTOPER( cur );
2026 const U8 *uc = (U8*)STRING( noper );
2027 const U8 *e = uc + STR_LEN( noper );
2029 U32 state = 1; /* required init */
2031 U16 charid = 0; /* sanity init */
2032 U32 accept_state = 0; /* sanity init */
2033 U8 *scan = (U8*)NULL; /* sanity init */
2035 STRLEN foldlen = 0; /* required init */
2036 U32 wordlen = 0; /* required init */
2038 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2040 if (OP(noper) == NOTHING) {
2041 regnode *noper_next= regnext(noper);
2042 if (noper_next != tail && OP(noper_next) == flags) {
2044 uc= (U8*)STRING(noper);
2045 e= uc + STR_LEN(noper);
2049 if ( OP(noper) != NOTHING ) {
2050 for ( ; uc < e ; uc += len ) {
2055 charid = trie->charmap[ uvc ];
2057 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2058 charid = svpp ? (U16)SvIV(*svpp) : 0;
2062 if ( !trie->trans[ state + charid ].next ) {
2063 trie->trans[ state + charid ].next = next_alloc;
2064 trie->trans[ state ].check++;
2065 prev_states[TRIE_NODENUM(next_alloc)]
2066 = TRIE_NODENUM(state);
2067 next_alloc += trie->uniquecharcount;
2069 state = trie->trans[ state + charid ].next;
2071 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2073 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2076 accept_state = TRIE_NODENUM( state );
2077 TRIE_HANDLE_WORD(accept_state);
2079 } /* end second pass */
2081 /* and now dump it out before we compress it */
2082 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2084 next_alloc, depth+1));
2088 * Inplace compress the table.*
2090 For sparse data sets the table constructed by the trie algorithm will
2091 be mostly 0/FAIL transitions or to put it another way mostly empty.
2092 (Note that leaf nodes will not contain any transitions.)
2094 This algorithm compresses the tables by eliminating most such
2095 transitions, at the cost of a modest bit of extra work during lookup:
2097 - Each states[] entry contains a .base field which indicates the
2098 index in the state[] array wheres its transition data is stored.
2100 - If .base is 0 there are no valid transitions from that node.
2102 - If .base is nonzero then charid is added to it to find an entry in
2105 -If trans[states[state].base+charid].check!=state then the
2106 transition is taken to be a 0/Fail transition. Thus if there are fail
2107 transitions at the front of the node then the .base offset will point
2108 somewhere inside the previous nodes data (or maybe even into a node
2109 even earlier), but the .check field determines if the transition is
2113 The following process inplace converts the table to the compressed
2114 table: We first do not compress the root node 1,and mark all its
2115 .check pointers as 1 and set its .base pointer as 1 as well. This
2116 allows us to do a DFA construction from the compressed table later,
2117 and ensures that any .base pointers we calculate later are greater
2120 - We set 'pos' to indicate the first entry of the second node.
2122 - We then iterate over the columns of the node, finding the first and
2123 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2124 and set the .check pointers accordingly, and advance pos
2125 appropriately and repreat for the next node. Note that when we copy
2126 the next pointers we have to convert them from the original
2127 NODEIDX form to NODENUM form as the former is not valid post
2130 - If a node has no transitions used we mark its base as 0 and do not
2131 advance the pos pointer.
2133 - If a node only has one transition we use a second pointer into the
2134 structure to fill in allocated fail transitions from other states.
2135 This pointer is independent of the main pointer and scans forward
2136 looking for null transitions that are allocated to a state. When it
2137 finds one it writes the single transition into the "hole". If the
2138 pointer doesnt find one the single transition is appended as normal.
2140 - Once compressed we can Renew/realloc the structures to release the
2143 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2144 specifically Fig 3.47 and the associated pseudocode.
2148 const U32 laststate = TRIE_NODENUM( next_alloc );
2151 trie->statecount = laststate;
2153 for ( state = 1 ; state < laststate ; state++ ) {
2155 const U32 stateidx = TRIE_NODEIDX( state );
2156 const U32 o_used = trie->trans[ stateidx ].check;
2157 U32 used = trie->trans[ stateidx ].check;
2158 trie->trans[ stateidx ].check = 0;
2160 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2161 if ( flag || trie->trans[ stateidx + charid ].next ) {
2162 if ( trie->trans[ stateidx + charid ].next ) {
2164 for ( ; zp < pos ; zp++ ) {
2165 if ( ! trie->trans[ zp ].next ) {
2169 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2170 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2171 trie->trans[ zp ].check = state;
2172 if ( ++zp > pos ) pos = zp;
2179 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2181 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2182 trie->trans[ pos ].check = state;
2187 trie->lasttrans = pos + 1;
2188 trie->states = (reg_trie_state *)
2189 PerlMemShared_realloc( trie->states, laststate
2190 * sizeof(reg_trie_state) );
2191 DEBUG_TRIE_COMPILE_MORE_r(
2192 PerlIO_printf( Perl_debug_log,
2193 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2194 (int)depth * 2 + 2,"",
2195 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2198 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2201 } /* end table compress */
2203 DEBUG_TRIE_COMPILE_MORE_r(
2204 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2205 (int)depth * 2 + 2, "",
2206 (UV)trie->statecount,
2207 (UV)trie->lasttrans)
2209 /* resize the trans array to remove unused space */
2210 trie->trans = (reg_trie_trans *)
2211 PerlMemShared_realloc( trie->trans, trie->lasttrans
2212 * sizeof(reg_trie_trans) );
2214 { /* Modify the program and insert the new TRIE node */
2215 U8 nodetype =(U8)(flags & 0xFF);
2219 regnode *optimize = NULL;
2220 #ifdef RE_TRACK_PATTERN_OFFSETS
2223 U32 mjd_nodelen = 0;
2224 #endif /* RE_TRACK_PATTERN_OFFSETS */
2225 #endif /* DEBUGGING */
2227 This means we convert either the first branch or the first Exact,
2228 depending on whether the thing following (in 'last') is a branch
2229 or not and whther first is the startbranch (ie is it a sub part of
2230 the alternation or is it the whole thing.)
2231 Assuming its a sub part we convert the EXACT otherwise we convert
2232 the whole branch sequence, including the first.
2234 /* Find the node we are going to overwrite */
2235 if ( first != startbranch || OP( last ) == BRANCH ) {
2236 /* branch sub-chain */
2237 NEXT_OFF( first ) = (U16)(last - first);
2238 #ifdef RE_TRACK_PATTERN_OFFSETS
2240 mjd_offset= Node_Offset((convert));
2241 mjd_nodelen= Node_Length((convert));
2244 /* whole branch chain */
2246 #ifdef RE_TRACK_PATTERN_OFFSETS
2249 const regnode *nop = NEXTOPER( convert );
2250 mjd_offset= Node_Offset((nop));
2251 mjd_nodelen= Node_Length((nop));
2255 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2256 (int)depth * 2 + 2, "",
2257 (UV)mjd_offset, (UV)mjd_nodelen)
2260 /* But first we check to see if there is a common prefix we can
2261 split out as an EXACT and put in front of the TRIE node. */
2262 trie->startstate= 1;
2263 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2265 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2269 const U32 base = trie->states[ state ].trans.base;
2271 if ( trie->states[state].wordnum )
2274 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2275 if ( ( base + ofs >= trie->uniquecharcount ) &&
2276 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2277 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2279 if ( ++count > 1 ) {
2280 SV **tmp = av_fetch( revcharmap, ofs, 0);
2281 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2282 if ( state == 1 ) break;
2284 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2286 PerlIO_printf(Perl_debug_log,
2287 "%*sNew Start State=%"UVuf" Class: [",
2288 (int)depth * 2 + 2, "",
2291 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2292 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2294 TRIE_BITMAP_SET(trie,*ch);
2296 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2298 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2302 TRIE_BITMAP_SET(trie,*ch);
2304 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2305 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2311 SV **tmp = av_fetch( revcharmap, idx, 0);
2313 char *ch = SvPV( *tmp, len );
2315 SV *sv=sv_newmortal();
2316 PerlIO_printf( Perl_debug_log,
2317 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2318 (int)depth * 2 + 2, "",
2320 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2321 PL_colors[0], PL_colors[1],
2322 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2323 PERL_PV_ESCAPE_FIRSTCHAR
2328 OP( convert ) = nodetype;
2329 str=STRING(convert);
2332 STR_LEN(convert) += len;
2338 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2343 trie->prefixlen = (state-1);
2345 regnode *n = convert+NODE_SZ_STR(convert);
2346 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2347 trie->startstate = state;
2348 trie->minlen -= (state - 1);
2349 trie->maxlen -= (state - 1);
2351 /* At least the UNICOS C compiler choked on this
2352 * being argument to DEBUG_r(), so let's just have
2355 #ifdef PERL_EXT_RE_BUILD
2361 regnode *fix = convert;
2362 U32 word = trie->wordcount;
2364 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2365 while( ++fix < n ) {
2366 Set_Node_Offset_Length(fix, 0, 0);
2369 SV ** const tmp = av_fetch( trie_words, word, 0 );
2371 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2372 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2374 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2382 NEXT_OFF(convert) = (U16)(tail - convert);
2383 DEBUG_r(optimize= n);
2389 if ( trie->maxlen ) {
2390 NEXT_OFF( convert ) = (U16)(tail - convert);
2391 ARG_SET( convert, data_slot );
2392 /* Store the offset to the first unabsorbed branch in
2393 jump[0], which is otherwise unused by the jump logic.
2394 We use this when dumping a trie and during optimisation. */
2396 trie->jump[0] = (U16)(nextbranch - convert);
2398 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2399 * and there is a bitmap
2400 * and the first "jump target" node we found leaves enough room
2401 * then convert the TRIE node into a TRIEC node, with the bitmap
2402 * embedded inline in the opcode - this is hypothetically faster.
2404 if ( !trie->states[trie->startstate].wordnum
2406 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2408 OP( convert ) = TRIEC;
2409 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2410 PerlMemShared_free(trie->bitmap);
2413 OP( convert ) = TRIE;
2415 /* store the type in the flags */
2416 convert->flags = nodetype;
2420 + regarglen[ OP( convert ) ];
2422 /* XXX We really should free up the resource in trie now,
2423 as we won't use them - (which resources?) dmq */
2425 /* needed for dumping*/
2426 DEBUG_r(if (optimize) {
2427 regnode *opt = convert;
2429 while ( ++opt < optimize) {
2430 Set_Node_Offset_Length(opt,0,0);
2433 Try to clean up some of the debris left after the
2436 while( optimize < jumper ) {
2437 mjd_nodelen += Node_Length((optimize));
2438 OP( optimize ) = OPTIMIZED;
2439 Set_Node_Offset_Length(optimize,0,0);
2442 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2444 } /* end node insert */
2445 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, convert);
2447 /* Finish populating the prev field of the wordinfo array. Walk back
2448 * from each accept state until we find another accept state, and if
2449 * so, point the first word's .prev field at the second word. If the
2450 * second already has a .prev field set, stop now. This will be the
2451 * case either if we've already processed that word's accept state,
2452 * or that state had multiple words, and the overspill words were
2453 * already linked up earlier.
2460 for (word=1; word <= trie->wordcount; word++) {
2462 if (trie->wordinfo[word].prev)
2464 state = trie->wordinfo[word].accept;
2466 state = prev_states[state];
2469 prev = trie->states[state].wordnum;
2473 trie->wordinfo[word].prev = prev;
2475 Safefree(prev_states);
2479 /* and now dump out the compressed format */
2480 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2482 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2484 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2485 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2487 SvREFCNT_dec_NN(revcharmap);
2491 : trie->startstate>1
2497 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2499 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2501 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2502 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2505 We find the fail state for each state in the trie, this state is the longest proper
2506 suffix of the current state's 'word' that is also a proper prefix of another word in our
2507 trie. State 1 represents the word '' and is thus the default fail state. This allows
2508 the DFA not to have to restart after its tried and failed a word at a given point, it
2509 simply continues as though it had been matching the other word in the first place.
2511 'abcdgu'=~/abcdefg|cdgu/
2512 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2513 fail, which would bring us to the state representing 'd' in the second word where we would
2514 try 'g' and succeed, proceeding to match 'cdgu'.
2516 /* add a fail transition */
2517 const U32 trie_offset = ARG(source);
2518 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2520 const U32 ucharcount = trie->uniquecharcount;
2521 const U32 numstates = trie->statecount;
2522 const U32 ubound = trie->lasttrans + ucharcount;
2526 U32 base = trie->states[ 1 ].trans.base;
2529 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2530 GET_RE_DEBUG_FLAGS_DECL;
2532 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2534 PERL_UNUSED_ARG(depth);
2538 ARG_SET( stclass, data_slot );
2539 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2540 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2541 aho->trie=trie_offset;
2542 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2543 Copy( trie->states, aho->states, numstates, reg_trie_state );
2544 Newxz( q, numstates, U32);
2545 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2548 /* initialize fail[0..1] to be 1 so that we always have
2549 a valid final fail state */
2550 fail[ 0 ] = fail[ 1 ] = 1;
2552 for ( charid = 0; charid < ucharcount ; charid++ ) {
2553 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2555 q[ q_write ] = newstate;
2556 /* set to point at the root */
2557 fail[ q[ q_write++ ] ]=1;
2560 while ( q_read < q_write) {
2561 const U32 cur = q[ q_read++ % numstates ];
2562 base = trie->states[ cur ].trans.base;
2564 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2565 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2567 U32 fail_state = cur;
2570 fail_state = fail[ fail_state ];
2571 fail_base = aho->states[ fail_state ].trans.base;
2572 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2574 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2575 fail[ ch_state ] = fail_state;
2576 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2578 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2580 q[ q_write++ % numstates] = ch_state;
2584 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2585 when we fail in state 1, this allows us to use the
2586 charclass scan to find a valid start char. This is based on the principle
2587 that theres a good chance the string being searched contains lots of stuff
2588 that cant be a start char.
2590 fail[ 0 ] = fail[ 1 ] = 0;
2591 DEBUG_TRIE_COMPILE_r({
2592 PerlIO_printf(Perl_debug_log,
2593 "%*sStclass Failtable (%"UVuf" states): 0",
2594 (int)(depth * 2), "", (UV)numstates
2596 for( q_read=1; q_read<numstates; q_read++ ) {
2597 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2599 PerlIO_printf(Perl_debug_log, "\n");
2602 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2607 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2608 * These need to be revisited when a newer toolchain becomes available.
2610 #if defined(__sparc64__) && defined(__GNUC__)
2611 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2612 # undef SPARC64_GCC_WORKAROUND
2613 # define SPARC64_GCC_WORKAROUND 1
2617 #define DEBUG_PEEP(str,scan,depth) \
2618 DEBUG_OPTIMISE_r({if (scan){ \
2619 SV * const mysv=sv_newmortal(); \
2620 regnode *Next = regnext(scan); \
2621 regprop(RExC_rx, mysv, scan); \
2622 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2623 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2624 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2628 /* The below joins as many adjacent EXACTish nodes as possible into a single
2629 * one. The regop may be changed if the node(s) contain certain sequences that
2630 * require special handling. The joining is only done if:
2631 * 1) there is room in the current conglomerated node to entirely contain the
2633 * 2) they are the exact same node type
2635 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2636 * these get optimized out
2638 * If a node is to match under /i (folded), the number of characters it matches
2639 * can be different than its character length if it contains a multi-character
2640 * fold. *min_subtract is set to the total delta of the input nodes.
2642 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2643 * and contains LATIN SMALL LETTER SHARP S
2645 * This is as good a place as any to discuss the design of handling these
2646 * multi-character fold sequences. It's been wrong in Perl for a very long
2647 * time. There are three code points in Unicode whose multi-character folds
2648 * were long ago discovered to mess things up. The previous designs for
2649 * dealing with these involved assigning a special node for them. This
2650 * approach doesn't work, as evidenced by this example:
2651 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2652 * Both these fold to "sss", but if the pattern is parsed to create a node that
2653 * would match just the \xDF, it won't be able to handle the case where a
2654 * successful match would have to cross the node's boundary. The new approach
2655 * that hopefully generally solves the problem generates an EXACTFU_SS node
2658 * It turns out that there are problems with all multi-character folds, and not
2659 * just these three. Now the code is general, for all such cases, but the
2660 * three still have some special handling. The approach taken is:
2661 * 1) This routine examines each EXACTFish node that could contain multi-
2662 * character fold sequences. It returns in *min_subtract how much to
2663 * subtract from the the actual length of the string to get a real minimum
2664 * match length; it is 0 if there are no multi-char folds. This delta is
2665 * used by the caller to adjust the min length of the match, and the delta
2666 * between min and max, so that the optimizer doesn't reject these
2667 * possibilities based on size constraints.
2668 * 2) Certain of these sequences require special handling by the trie code,
2669 * so, if found, this code changes the joined node type to special ops:
2670 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2671 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2672 * is used for an EXACTFU node that contains at least one "ss" sequence in
2673 * it. For non-UTF-8 patterns and strings, this is the only case where
2674 * there is a possible fold length change. That means that a regular
2675 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2676 * with length changes, and so can be processed faster. regexec.c takes
2677 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2678 * pre-folded by regcomp.c. This saves effort in regex matching.
2679 * However, the pre-folding isn't done for non-UTF8 patterns because the
2680 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2681 * down by forcing the pattern into UTF8 unless necessary. Also what
2682 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2683 * possibilities for the non-UTF8 patterns are quite simple, except for
2684 * the sharp s. All the ones that don't involve a UTF-8 target string are
2685 * members of a fold-pair, and arrays are set up for all of them so that
2686 * the other member of the pair can be found quickly. Code elsewhere in
2687 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2688 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2689 * described in the next item.
2690 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2691 * 'ss' or not is not knowable at compile time. It will match iff the
2692 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2693 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2694 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
2695 * described in item 3). An assumption that the optimizer part of
2696 * regexec.c (probably unwittingly) makes is that a character in the
2697 * pattern corresponds to at most a single character in the target string.
2698 * (And I do mean character, and not byte here, unlike other parts of the
2699 * documentation that have never been updated to account for multibyte
2700 * Unicode.) This assumption is wrong only in this case, as all other
2701 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2702 * virtue of having this file pre-fold UTF-8 patterns. I'm
2703 * reluctant to try to change this assumption, so instead the code punts.
2704 * This routine examines EXACTF nodes for the sharp s, and returns a
2705 * boolean indicating whether or not the node is an EXACTF node that
2706 * contains a sharp s. When it is true, the caller sets a flag that later
2707 * causes the optimizer in this file to not set values for the floating
2708 * and fixed string lengths, and thus avoids the optimizer code in
2709 * regexec.c that makes the invalid assumption. Thus, there is no
2710 * optimization based on string lengths for EXACTF nodes that contain the
2711 * sharp s. This only happens for /id rules (which means the pattern
2715 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2716 if (PL_regkind[OP(scan)] == EXACT) \
2717 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2720 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, UV *min_subtract, bool *has_exactf_sharp_s, U32 flags,regnode *val, U32 depth) {
2721 /* Merge several consecutive EXACTish nodes into one. */
2722 regnode *n = regnext(scan);
2724 regnode *next = scan + NODE_SZ_STR(scan);
2728 regnode *stop = scan;
2729 GET_RE_DEBUG_FLAGS_DECL;
2731 PERL_UNUSED_ARG(depth);
2734 PERL_ARGS_ASSERT_JOIN_EXACT;
2735 #ifndef EXPERIMENTAL_INPLACESCAN
2736 PERL_UNUSED_ARG(flags);
2737 PERL_UNUSED_ARG(val);
2739 DEBUG_PEEP("join",scan,depth);
2741 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2742 * EXACT ones that are mergeable to the current one. */
2744 && (PL_regkind[OP(n)] == NOTHING
2745 || (stringok && OP(n) == OP(scan)))
2747 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2750 if (OP(n) == TAIL || n > next)
2752 if (PL_regkind[OP(n)] == NOTHING) {
2753 DEBUG_PEEP("skip:",n,depth);
2754 NEXT_OFF(scan) += NEXT_OFF(n);
2755 next = n + NODE_STEP_REGNODE;
2762 else if (stringok) {
2763 const unsigned int oldl = STR_LEN(scan);
2764 regnode * const nnext = regnext(n);
2766 /* XXX I (khw) kind of doubt that this works on platforms where
2767 * U8_MAX is above 255 because of lots of other assumptions */
2768 /* Don't join if the sum can't fit into a single node */
2769 if (oldl + STR_LEN(n) > U8_MAX)
2772 DEBUG_PEEP("merg",n,depth);
2775 NEXT_OFF(scan) += NEXT_OFF(n);
2776 STR_LEN(scan) += STR_LEN(n);
2777 next = n + NODE_SZ_STR(n);
2778 /* Now we can overwrite *n : */
2779 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2787 #ifdef EXPERIMENTAL_INPLACESCAN
2788 if (flags && !NEXT_OFF(n)) {
2789 DEBUG_PEEP("atch", val, depth);
2790 if (reg_off_by_arg[OP(n)]) {
2791 ARG_SET(n, val - n);
2794 NEXT_OFF(n) = val - n;
2802 *has_exactf_sharp_s = FALSE;
2804 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2805 * can now analyze for sequences of problematic code points. (Prior to
2806 * this final joining, sequences could have been split over boundaries, and
2807 * hence missed). The sequences only happen in folding, hence for any
2808 * non-EXACT EXACTish node */
2809 if (OP(scan) != EXACT) {
2810 const U8 * const s0 = (U8*) STRING(scan);
2812 const U8 * const s_end = s0 + STR_LEN(scan);
2814 /* One pass is made over the node's string looking for all the
2815 * possibilities. to avoid some tests in the loop, there are two main
2816 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2820 /* Examine the string for a multi-character fold sequence. UTF-8
2821 * patterns have all characters pre-folded by the time this code is
2823 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2824 length sequence we are looking for is 2 */
2827 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2828 if (! len) { /* Not a multi-char fold: get next char */
2833 /* Nodes with 'ss' require special handling, except for EXACTFL
2834 * and EXACTFA for which there is no multi-char fold to this */
2835 if (len == 2 && *s == 's' && *(s+1) == 's'
2836 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2839 OP(scan) = EXACTFU_SS;
2842 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2843 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2844 COMBINING_DIAERESIS_UTF8
2845 COMBINING_ACUTE_ACCENT_UTF8,
2847 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2848 COMBINING_DIAERESIS_UTF8
2849 COMBINING_ACUTE_ACCENT_UTF8,
2854 /* These two folds require special handling by trie's, so
2855 * change the node type to indicate this. If EXACTFA and
2856 * EXACTFL were ever to be handled by trie's, this would
2857 * have to be changed. If this node has already been
2858 * changed to EXACTFU_SS in this loop, leave it as is. (I
2859 * (khw) think it doesn't matter in regexec.c for UTF
2860 * patterns, but no need to change it */
2861 if (OP(scan) == EXACTFU) {
2862 OP(scan) = EXACTFU_TRICKYFOLD;
2866 else { /* Here is a generic multi-char fold. */
2867 const U8* multi_end = s + len;
2869 /* Count how many characters in it. In the case of /l and
2870 * /aa, no folds which contain ASCII code points are
2871 * allowed, so check for those, and skip if found. (In
2872 * EXACTFL, no folds are allowed to any Latin1 code point,
2873 * not just ASCII. But there aren't any of these
2874 * currently, nor ever likely, so don't take the time to
2875 * test for them. The code that generates the
2876 * is_MULTI_foo() macros croaks should one actually get put
2877 * into Unicode .) */
2878 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2879 count = utf8_length(s, multi_end);
2883 while (s < multi_end) {
2886 goto next_iteration;
2896 /* The delta is how long the sequence is minus 1 (1 is how long
2897 * the character that folds to the sequence is) */
2898 *min_subtract += count - 1;
2902 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2904 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2905 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2906 * nodes can't have multi-char folds to this range (and there are
2907 * no existing ones in the upper latin1 range). In the EXACTF
2908 * case we look also for the sharp s, which can be in the final
2909 * position. Otherwise we can stop looking 1 byte earlier because
2910 * have to find at least two characters for a multi-fold */
2911 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2913 /* The below is perhaps overboard, but this allows us to save a
2914 * test each time through the loop at the expense of a mask. This
2915 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2916 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2917 * are 64. This uses an exclusive 'or' to find that bit and then
2918 * inverts it to form a mask, with just a single 0, in the bit
2919 * position where 'S' and 's' differ. */
2920 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2921 const U8 s_masked = 's' & S_or_s_mask;
2924 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2925 if (! len) { /* Not a multi-char fold. */
2926 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2928 *has_exactf_sharp_s = TRUE;
2935 && ((*s & S_or_s_mask) == s_masked)
2936 && ((*(s+1) & S_or_s_mask) == s_masked))
2939 /* EXACTF nodes need to know that the minimum length
2940 * changed so that a sharp s in the string can match this
2941 * ss in the pattern, but they remain EXACTF nodes, as they
2942 * won't match this unless the target string is is UTF-8,
2943 * which we don't know until runtime */
2944 if (OP(scan) != EXACTF) {
2945 OP(scan) = EXACTFU_SS;
2949 *min_subtract += len - 1;
2956 /* Allow dumping but overwriting the collection of skipped
2957 * ops and/or strings with fake optimized ops */
2958 n = scan + NODE_SZ_STR(scan);
2966 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2970 /* REx optimizer. Converts nodes into quicker variants "in place".
2971 Finds fixed substrings. */
2973 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2974 to the position after last scanned or to NULL. */
2976 #define INIT_AND_WITHP \
2977 assert(!and_withp); \
2978 Newx(and_withp,1,struct regnode_charclass_class); \
2979 SAVEFREEPV(and_withp)
2981 /* this is a chain of data about sub patterns we are processing that
2982 need to be handled separately/specially in study_chunk. Its so
2983 we can simulate recursion without losing state. */
2985 typedef struct scan_frame {
2986 regnode *last; /* last node to process in this frame */
2987 regnode *next; /* next node to process when last is reached */
2988 struct scan_frame *prev; /*previous frame*/
2989 I32 stop; /* what stopparen do we use */
2993 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2996 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2997 I32 *minlenp, I32 *deltap,
3002 struct regnode_charclass_class *and_withp,
3003 U32 flags, U32 depth)
3004 /* scanp: Start here (read-write). */
3005 /* deltap: Write maxlen-minlen here. */
3006 /* last: Stop before this one. */
3007 /* data: string data about the pattern */
3008 /* stopparen: treat close N as END */
3009 /* recursed: which subroutines have we recursed into */
3010 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3013 I32 min = 0; /* There must be at least this number of characters to match */
3015 regnode *scan = *scanp, *next;
3017 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3018 int is_inf_internal = 0; /* The studied chunk is infinite */
3019 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3020 scan_data_t data_fake;
3021 SV *re_trie_maxbuff = NULL;
3022 regnode *first_non_open = scan;
3023 I32 stopmin = I32_MAX;
3024 scan_frame *frame = NULL;
3025 GET_RE_DEBUG_FLAGS_DECL;
3027 PERL_ARGS_ASSERT_STUDY_CHUNK;
3030 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3034 while (first_non_open && OP(first_non_open) == OPEN)
3035 first_non_open=regnext(first_non_open);
3040 while ( scan && OP(scan) != END && scan < last ){
3041 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3042 node length to get a real minimum (because
3043 the folded version may be shorter) */
3044 bool has_exactf_sharp_s = FALSE;
3045 /* Peephole optimizer: */
3046 DEBUG_STUDYDATA("Peep:", data,depth);
3047 DEBUG_PEEP("Peep",scan,depth);
3049 /* Its not clear to khw or hv why this is done here, and not in the
3050 * clauses that deal with EXACT nodes. khw's guess is that it's
3051 * because of a previous design */
3052 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3054 /* Follow the next-chain of the current node and optimize
3055 away all the NOTHINGs from it. */
3056 if (OP(scan) != CURLYX) {
3057 const int max = (reg_off_by_arg[OP(scan)]
3059 /* I32 may be smaller than U16 on CRAYs! */
3060 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3061 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3065 /* Skip NOTHING and LONGJMP. */
3066 while ((n = regnext(n))
3067 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3068 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3069 && off + noff < max)
3071 if (reg_off_by_arg[OP(scan)])
3074 NEXT_OFF(scan) = off;
3079 /* The principal pseudo-switch. Cannot be a switch, since we
3080 look into several different things. */
3081 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3082 || OP(scan) == IFTHEN) {
3083 next = regnext(scan);
3085 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3087 if (OP(next) == code || code == IFTHEN) {
3088 /* NOTE - There is similar code to this block below for handling
3089 TRIE nodes on a re-study. If you change stuff here check there
3091 I32 max1 = 0, min1 = I32_MAX, num = 0;
3092 struct regnode_charclass_class accum;
3093 regnode * const startbranch=scan;
3095 if (flags & SCF_DO_SUBSTR)
3096 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3097 if (flags & SCF_DO_STCLASS)
3098 cl_init_zero(pRExC_state, &accum);
3100 while (OP(scan) == code) {
3101 I32 deltanext, minnext, f = 0, fake;
3102 struct regnode_charclass_class this_class;
3105 data_fake.flags = 0;
3107 data_fake.whilem_c = data->whilem_c;
3108 data_fake.last_closep = data->last_closep;
3111 data_fake.last_closep = &fake;
3113 data_fake.pos_delta = delta;
3114 next = regnext(scan);
3115 scan = NEXTOPER(scan);
3117 scan = NEXTOPER(scan);
3118 if (flags & SCF_DO_STCLASS) {
3119 cl_init(pRExC_state, &this_class);
3120 data_fake.start_class = &this_class;
3121 f = SCF_DO_STCLASS_AND;
3123 if (flags & SCF_WHILEM_VISITED_POS)
3124 f |= SCF_WHILEM_VISITED_POS;
3126 /* we suppose the run is continuous, last=next...*/
3127 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3129 stopparen, recursed, NULL, f,depth+1);
3132 if (deltanext == I32_MAX) {
3133 is_inf = is_inf_internal = 1;
3135 } else if (max1 < minnext + deltanext)
3136 max1 = minnext + deltanext;
3138 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3140 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3141 if ( stopmin > minnext)
3142 stopmin = min + min1;
3143 flags &= ~SCF_DO_SUBSTR;
3145 data->flags |= SCF_SEEN_ACCEPT;
3148 if (data_fake.flags & SF_HAS_EVAL)
3149 data->flags |= SF_HAS_EVAL;
3150 data->whilem_c = data_fake.whilem_c;
3152 if (flags & SCF_DO_STCLASS)
3153 cl_or(pRExC_state, &accum, &this_class);
3155 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3157 if (flags & SCF_DO_SUBSTR) {
3158 data->pos_min += min1;
3159 if (data->pos_delta >= I32_MAX - (max1 - min1))
3160 data->pos_delta = I32_MAX;
3162 data->pos_delta += max1 - min1;
3163 if (max1 != min1 || is_inf)
3164 data->longest = &(data->longest_float);
3167 if (delta == I32_MAX || I32_MAX - delta - (max1 - min1) < 0)
3170 delta += max1 - min1;
3171 if (flags & SCF_DO_STCLASS_OR) {
3172 cl_or(pRExC_state, data->start_class, &accum);
3174 cl_and(data->start_class, and_withp);
3175 flags &= ~SCF_DO_STCLASS;
3178 else if (flags & SCF_DO_STCLASS_AND) {
3180 cl_and(data->start_class, &accum);
3181 flags &= ~SCF_DO_STCLASS;
3184 /* Switch to OR mode: cache the old value of
3185 * data->start_class */
3187 StructCopy(data->start_class, and_withp,
3188 struct regnode_charclass_class);
3189 flags &= ~SCF_DO_STCLASS_AND;
3190 StructCopy(&accum, data->start_class,
3191 struct regnode_charclass_class);
3192 flags |= SCF_DO_STCLASS_OR;
3193 SET_SSC_EOS(data->start_class);
3197 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3200 Assuming this was/is a branch we are dealing with: 'scan' now
3201 points at the item that follows the branch sequence, whatever
3202 it is. We now start at the beginning of the sequence and look
3209 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3211 If we can find such a subsequence we need to turn the first
3212 element into a trie and then add the subsequent branch exact
3213 strings to the trie.
3217 1. patterns where the whole set of branches can be converted.
3219 2. patterns where only a subset can be converted.
3221 In case 1 we can replace the whole set with a single regop
3222 for the trie. In case 2 we need to keep the start and end
3225 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3226 becomes BRANCH TRIE; BRANCH X;
3228 There is an additional case, that being where there is a
3229 common prefix, which gets split out into an EXACT like node
3230 preceding the TRIE node.
3232 If x(1..n)==tail then we can do a simple trie, if not we make
3233 a "jump" trie, such that when we match the appropriate word
3234 we "jump" to the appropriate tail node. Essentially we turn
3235 a nested if into a case structure of sorts.
3240 if (!re_trie_maxbuff) {
3241 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3242 if (!SvIOK(re_trie_maxbuff))
3243 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3245 if ( SvIV(re_trie_maxbuff)>=0 ) {
3247 regnode *first = (regnode *)NULL;
3248 regnode *last = (regnode *)NULL;
3249 regnode *tail = scan;
3254 SV * const mysv = sv_newmortal(); /* for dumping */
3256 /* var tail is used because there may be a TAIL
3257 regop in the way. Ie, the exacts will point to the
3258 thing following the TAIL, but the last branch will
3259 point at the TAIL. So we advance tail. If we
3260 have nested (?:) we may have to move through several
3264 while ( OP( tail ) == TAIL ) {
3265 /* this is the TAIL generated by (?:) */
3266 tail = regnext( tail );
3270 DEBUG_TRIE_COMPILE_r({
3271 regprop(RExC_rx, mysv, tail );
3272 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3273 (int)depth * 2 + 2, "",
3274 "Looking for TRIE'able sequences. Tail node is: ",
3275 SvPV_nolen_const( mysv )
3281 Step through the branches
3282 cur represents each branch,
3283 noper is the first thing to be matched as part of that branch
3284 noper_next is the regnext() of that node.
3286 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3287 via a "jump trie" but we also support building with NOJUMPTRIE,
3288 which restricts the trie logic to structures like /FOO|BAR/.
3290 If noper is a trieable nodetype then the branch is a possible optimization
3291 target. If we are building under NOJUMPTRIE then we require that noper_next
3292 is the same as scan (our current position in the regex program).
3294 Once we have two or more consecutive such branches we can create a
3295 trie of the EXACT's contents and stitch it in place into the program.
3297 If the sequence represents all of the branches in the alternation we
3298 replace the entire thing with a single TRIE node.
3300 Otherwise when it is a subsequence we need to stitch it in place and
3301 replace only the relevant branches. This means the first branch has
3302 to remain as it is used by the alternation logic, and its next pointer,
3303 and needs to be repointed at the item on the branch chain following
3304 the last branch we have optimized away.
3306 This could be either a BRANCH, in which case the subsequence is internal,
3307 or it could be the item following the branch sequence in which case the
3308 subsequence is at the end (which does not necessarily mean the first node
3309 is the start of the alternation).
3311 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3314 ----------------+-----------
3318 EXACTFU_SS | EXACTFU
3319 EXACTFU_TRICKYFOLD | EXACTFU
3324 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3325 ( EXACT == (X) ) ? EXACT : \
3326 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3329 /* dont use tail as the end marker for this traverse */
3330 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3331 regnode * const noper = NEXTOPER( cur );
3332 U8 noper_type = OP( noper );
3333 U8 noper_trietype = TRIE_TYPE( noper_type );
3334 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3335 regnode * const noper_next = regnext( noper );
3336 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3337 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3340 DEBUG_TRIE_COMPILE_r({
3341 regprop(RExC_rx, mysv, cur);
3342 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3343 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3345 regprop(RExC_rx, mysv, noper);
3346 PerlIO_printf( Perl_debug_log, " -> %s",
3347 SvPV_nolen_const(mysv));
3350 regprop(RExC_rx, mysv, noper_next );
3351 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3352 SvPV_nolen_const(mysv));
3354 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3355 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3356 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3360 /* Is noper a trieable nodetype that can be merged with the
3361 * current trie (if there is one)? */
3365 ( noper_trietype == NOTHING)
3366 || ( trietype == NOTHING )
3367 || ( trietype == noper_trietype )
3370 && noper_next == tail
3374 /* Handle mergable triable node
3375 * Either we are the first node in a new trieable sequence,
3376 * in which case we do some bookkeeping, otherwise we update
3377 * the end pointer. */
3380 if ( noper_trietype == NOTHING ) {
3381 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3382 regnode * const noper_next = regnext( noper );
3383 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3384 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3387 if ( noper_next_trietype ) {
3388 trietype = noper_next_trietype;
3389 } else if (noper_next_type) {
3390 /* a NOTHING regop is 1 regop wide. We need at least two
3391 * for a trie so we can't merge this in */
3395 trietype = noper_trietype;
3398 if ( trietype == NOTHING )
3399 trietype = noper_trietype;
3404 } /* end handle mergable triable node */
3406 /* handle unmergable node -
3407 * noper may either be a triable node which can not be tried
3408 * together with the current trie, or a non triable node */
3410 /* If last is set and trietype is not NOTHING then we have found
3411 * at least two triable branch sequences in a row of a similar
3412 * trietype so we can turn them into a trie. If/when we
3413 * allow NOTHING to start a trie sequence this condition will be
3414 * required, and it isn't expensive so we leave it in for now. */
3415 if ( trietype && trietype != NOTHING )
3416 make_trie( pRExC_state,
3417 startbranch, first, cur, tail, count,
3418 trietype, depth+1 );
3419 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3423 && noper_next == tail
3426 /* noper is triable, so we can start a new trie sequence */
3429 trietype = noper_trietype;
3431 /* if we already saw a first but the current node is not triable then we have
3432 * to reset the first information. */
3437 } /* end handle unmergable node */
3438 } /* loop over branches */
3439 DEBUG_TRIE_COMPILE_r({
3440 regprop(RExC_rx, mysv, cur);
3441 PerlIO_printf( Perl_debug_log,
3442 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3443 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3446 if ( last && trietype ) {
3447 if ( trietype != NOTHING ) {
3448 /* the last branch of the sequence was part of a trie,
3449 * so we have to construct it here outside of the loop
3451 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3452 #ifdef TRIE_STUDY_OPT
3453 if ( ((made == MADE_EXACT_TRIE &&
3454 startbranch == first)
3455 || ( first_non_open == first )) &&
3457 flags |= SCF_TRIE_RESTUDY;
3458 if ( startbranch == first
3461 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3466 /* at this point we know whatever we have is a NOTHING sequence/branch
3467 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3469 if ( startbranch == first ) {
3471 /* the entire thing is a NOTHING sequence, something like this:
3472 * (?:|) So we can turn it into a plain NOTHING op. */
3473 DEBUG_TRIE_COMPILE_r({
3474 regprop(RExC_rx, mysv, cur);
3475 PerlIO_printf( Perl_debug_log,
3476 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3477 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3480 OP(startbranch)= NOTHING;
3481 NEXT_OFF(startbranch)= tail - startbranch;
3482 for ( opt= startbranch + 1; opt < tail ; opt++ )
3486 } /* end if ( last) */
3487 } /* TRIE_MAXBUF is non zero */
3492 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3493 scan = NEXTOPER(NEXTOPER(scan));
3494 } else /* single branch is optimized. */
3495 scan = NEXTOPER(scan);
3497 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3498 scan_frame *newframe = NULL;
3503 if (OP(scan) != SUSPEND) {
3504 /* set the pointer */
3505 if (OP(scan) == GOSUB) {
3507 RExC_recurse[ARG2L(scan)] = scan;
3508 start = RExC_open_parens[paren-1];
3509 end = RExC_close_parens[paren-1];
3512 start = RExC_rxi->program + 1;
3516 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3517 SAVEFREEPV(recursed);
3519 if (!PAREN_TEST(recursed,paren+1)) {
3520 PAREN_SET(recursed,paren+1);
3521 Newx(newframe,1,scan_frame);
3523 if (flags & SCF_DO_SUBSTR) {
3524 SCAN_COMMIT(pRExC_state,data,minlenp);
3525 data->longest = &(data->longest_float);
3527 is_inf = is_inf_internal = 1;
3528 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3529 cl_anything(pRExC_state, data->start_class);
3530 flags &= ~SCF_DO_STCLASS;
3533 Newx(newframe,1,scan_frame);
3536 end = regnext(scan);
3541 SAVEFREEPV(newframe);
3542 newframe->next = regnext(scan);
3543 newframe->last = last;
3544 newframe->stop = stopparen;
3545 newframe->prev = frame;
3555 else if (OP(scan) == EXACT) {
3556 I32 l = STR_LEN(scan);
3559 const U8 * const s = (U8*)STRING(scan);
3560 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3561 l = utf8_length(s, s + l);
3563 uc = *((U8*)STRING(scan));
3566 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3567 /* The code below prefers earlier match for fixed
3568 offset, later match for variable offset. */
3569 if (data->last_end == -1) { /* Update the start info. */
3570 data->last_start_min = data->pos_min;
3571 data->last_start_max = is_inf
3572 ? I32_MAX : data->pos_min + data->pos_delta;
3574 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3576 SvUTF8_on(data->last_found);
3578 SV * const sv = data->last_found;
3579 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3580 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3581 if (mg && mg->mg_len >= 0)
3582 mg->mg_len += utf8_length((U8*)STRING(scan),
3583 (U8*)STRING(scan)+STR_LEN(scan));
3585 data->last_end = data->pos_min + l;
3586 data->pos_min += l; /* As in the first entry. */
3587 data->flags &= ~SF_BEFORE_EOL;
3589 if (flags & SCF_DO_STCLASS_AND) {
3590 /* Check whether it is compatible with what we know already! */
3594 /* If compatible, we or it in below. It is compatible if is
3595 * in the bitmp and either 1) its bit or its fold is set, or 2)
3596 * it's for a locale. Even if there isn't unicode semantics
3597 * here, at runtime there may be because of matching against a
3598 * utf8 string, so accept a possible false positive for
3599 * latin1-range folds */
3601 (!(data->start_class->flags & ANYOF_LOCALE)
3602 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3603 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3604 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3609 ANYOF_CLASS_ZERO(data->start_class);
3610 ANYOF_BITMAP_ZERO(data->start_class);
3612 ANYOF_BITMAP_SET(data->start_class, uc);
3613 else if (uc >= 0x100) {
3616 /* Some Unicode code points fold to the Latin1 range; as
3617 * XXX temporary code, instead of figuring out if this is
3618 * one, just assume it is and set all the start class bits
3619 * that could be some such above 255 code point's fold
3620 * which will generate fals positives. As the code
3621 * elsewhere that does compute the fold settles down, it
3622 * can be extracted out and re-used here */
3623 for (i = 0; i < 256; i++){
3624 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3625 ANYOF_BITMAP_SET(data->start_class, i);
3629 CLEAR_SSC_EOS(data->start_class);
3631 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3633 else if (flags & SCF_DO_STCLASS_OR) {
3634 /* false positive possible if the class is case-folded */
3636 ANYOF_BITMAP_SET(data->start_class, uc);
3638 data->start_class->flags |= ANYOF_UNICODE_ALL;
3639 CLEAR_SSC_EOS(data->start_class);
3640 cl_and(data->start_class, and_withp);
3642 flags &= ~SCF_DO_STCLASS;
3644 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3645 I32 l = STR_LEN(scan);
3646 UV uc = *((U8*)STRING(scan));
3648 /* Search for fixed substrings supports EXACT only. */
3649 if (flags & SCF_DO_SUBSTR) {
3651 SCAN_COMMIT(pRExC_state, data, minlenp);
3654 const U8 * const s = (U8 *)STRING(scan);
3655 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3656 l = utf8_length(s, s + l);
3658 if (has_exactf_sharp_s) {
3659 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3661 min += l - min_subtract;
3663 delta += min_subtract;
3664 if (flags & SCF_DO_SUBSTR) {
3665 data->pos_min += l - min_subtract;
3666 if (data->pos_min < 0) {
3669 data->pos_delta += min_subtract;
3671 data->longest = &(data->longest_float);
3674 if (flags & SCF_DO_STCLASS_AND) {
3675 /* Check whether it is compatible with what we know already! */
3678 (!(data->start_class->flags & ANYOF_LOCALE)
3679 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3680 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3684 ANYOF_CLASS_ZERO(data->start_class);
3685 ANYOF_BITMAP_ZERO(data->start_class);
3687 ANYOF_BITMAP_SET(data->start_class, uc);
3688 CLEAR_SSC_EOS(data->start_class);
3689 if (OP(scan) == EXACTFL) {
3690 /* XXX This set is probably no longer necessary, and
3691 * probably wrong as LOCALE now is on in the initial
3693 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3697 /* Also set the other member of the fold pair. In case
3698 * that unicode semantics is called for at runtime, use
3699 * the full latin1 fold. (Can't do this for locale,
3700 * because not known until runtime) */
3701 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3703 /* All other (EXACTFL handled above) folds except under
3704 * /iaa that include s, S, and sharp_s also may include
3706 if (OP(scan) != EXACTFA) {
3707 if (uc == 's' || uc == 'S') {
3708 ANYOF_BITMAP_SET(data->start_class,
3709 LATIN_SMALL_LETTER_SHARP_S);
3711 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3712 ANYOF_BITMAP_SET(data->start_class, 's');
3713 ANYOF_BITMAP_SET(data->start_class, 'S');
3718 else if (uc >= 0x100) {
3720 for (i = 0; i < 256; i++){
3721 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3722 ANYOF_BITMAP_SET(data->start_class, i);
3727 else if (flags & SCF_DO_STCLASS_OR) {
3728 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3729 /* false positive possible if the class is case-folded.
3730 Assume that the locale settings are the same... */
3732 ANYOF_BITMAP_SET(data->start_class, uc);
3733 if (OP(scan) != EXACTFL) {
3735 /* And set the other member of the fold pair, but
3736 * can't do that in locale because not known until
3738 ANYOF_BITMAP_SET(data->start_class,
3739 PL_fold_latin1[uc]);
3741 /* All folds except under /iaa that include s, S,
3742 * and sharp_s also may include the others */
3743 if (OP(scan) != EXACTFA) {
3744 if (uc == 's' || uc == 'S') {
3745 ANYOF_BITMAP_SET(data->start_class,
3746 LATIN_SMALL_LETTER_SHARP_S);
3748 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3749 ANYOF_BITMAP_SET(data->start_class, 's');
3750 ANYOF_BITMAP_SET(data->start_class, 'S');
3755 CLEAR_SSC_EOS(data->start_class);
3757 cl_and(data->start_class, and_withp);
3759 flags &= ~SCF_DO_STCLASS;
3761 else if (REGNODE_VARIES(OP(scan))) {
3762 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3763 I32 f = flags, pos_before = 0;
3764 regnode * const oscan = scan;
3765 struct regnode_charclass_class this_class;
3766 struct regnode_charclass_class *oclass = NULL;
3767 I32 next_is_eval = 0;
3769 switch (PL_regkind[OP(scan)]) {
3770 case WHILEM: /* End of (?:...)* . */
3771 scan = NEXTOPER(scan);
3774 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3775 next = NEXTOPER(scan);
3776 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3778 maxcount = REG_INFTY;
3779 next = regnext(scan);
3780 scan = NEXTOPER(scan);
3784 if (flags & SCF_DO_SUBSTR)
3789 if (flags & SCF_DO_STCLASS) {
3791 maxcount = REG_INFTY;
3792 next = regnext(scan);
3793 scan = NEXTOPER(scan);
3796 is_inf = is_inf_internal = 1;
3797 scan = regnext(scan);
3798 if (flags & SCF_DO_SUBSTR) {
3799 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3800 data->longest = &(data->longest_float);
3802 goto optimize_curly_tail;
3804 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3805 && (scan->flags == stopparen))
3810 mincount = ARG1(scan);
3811 maxcount = ARG2(scan);
3813 next = regnext(scan);
3814 if (OP(scan) == CURLYX) {
3815 I32 lp = (data ? *(data->last_closep) : 0);
3816 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3818 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3819 next_is_eval = (OP(scan) == EVAL);
3821 if (flags & SCF_DO_SUBSTR) {
3822 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3823 pos_before = data->pos_min;
3827 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3829 data->flags |= SF_IS_INF;
3831 if (flags & SCF_DO_STCLASS) {
3832 cl_init(pRExC_state, &this_class);
3833 oclass = data->start_class;
3834 data->start_class = &this_class;
3835 f |= SCF_DO_STCLASS_AND;
3836 f &= ~SCF_DO_STCLASS_OR;
3838 /* Exclude from super-linear cache processing any {n,m}
3839 regops for which the combination of input pos and regex
3840 pos is not enough information to determine if a match
3843 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3844 regex pos at the \s*, the prospects for a match depend not
3845 only on the input position but also on how many (bar\s*)
3846 repeats into the {4,8} we are. */
3847 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3848 f &= ~SCF_WHILEM_VISITED_POS;
3850 /* This will finish on WHILEM, setting scan, or on NULL: */
3851 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3852 last, data, stopparen, recursed, NULL,
3854 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3856 if (flags & SCF_DO_STCLASS)
3857 data->start_class = oclass;
3858 if (mincount == 0 || minnext == 0) {
3859 if (flags & SCF_DO_STCLASS_OR) {
3860 cl_or(pRExC_state, data->start_class, &this_class);
3862 else if (flags & SCF_DO_STCLASS_AND) {
3863 /* Switch to OR mode: cache the old value of
3864 * data->start_class */
3866 StructCopy(data->start_class, and_withp,
3867 struct regnode_charclass_class);
3868 flags &= ~SCF_DO_STCLASS_AND;
3869 StructCopy(&this_class, data->start_class,
3870 struct regnode_charclass_class);
3871 flags |= SCF_DO_STCLASS_OR;
3872 SET_SSC_EOS(data->start_class);
3874 } else { /* Non-zero len */
3875 if (flags & SCF_DO_STCLASS_OR) {
3876 cl_or(pRExC_state, data->start_class, &this_class);
3877 cl_and(data->start_class, and_withp);
3879 else if (flags & SCF_DO_STCLASS_AND)
3880 cl_and(data->start_class, &this_class);
3881 flags &= ~SCF_DO_STCLASS;
3883 if (!scan) /* It was not CURLYX, but CURLY. */
3885 if ( /* ? quantifier ok, except for (?{ ... }) */
3886 (next_is_eval || !(mincount == 0 && maxcount == 1))
3887 && (minnext == 0) && (deltanext == 0)
3888 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3889 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3891 /* Fatal warnings may leak the regexp without this: */
3892 SAVEFREESV(RExC_rx_sv);
3893 ckWARNreg(RExC_parse,
3894 "Quantifier unexpected on zero-length expression");
3895 (void)ReREFCNT_inc(RExC_rx_sv);
3898 min += minnext * mincount;
3899 is_inf_internal |= deltanext == I32_MAX
3900 || (maxcount == REG_INFTY && minnext + deltanext > 0);
3901 is_inf |= is_inf_internal;
3905 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3907 /* Try powerful optimization CURLYX => CURLYN. */
3908 if ( OP(oscan) == CURLYX && data
3909 && data->flags & SF_IN_PAR
3910 && !(data->flags & SF_HAS_EVAL)
3911 && !deltanext && minnext == 1 ) {
3912 /* Try to optimize to CURLYN. */
3913 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3914 regnode * const nxt1 = nxt;
3921 if (!REGNODE_SIMPLE(OP(nxt))
3922 && !(PL_regkind[OP(nxt)] == EXACT
3923 && STR_LEN(nxt) == 1))
3929 if (OP(nxt) != CLOSE)
3931 if (RExC_open_parens) {
3932 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3933 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3935 /* Now we know that nxt2 is the only contents: */
3936 oscan->flags = (U8)ARG(nxt);
3938 OP(nxt1) = NOTHING; /* was OPEN. */
3941 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3942 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3943 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3944 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3945 OP(nxt + 1) = OPTIMIZED; /* was count. */
3946 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3951 /* Try optimization CURLYX => CURLYM. */
3952 if ( OP(oscan) == CURLYX && data
3953 && !(data->flags & SF_HAS_PAR)
3954 && !(data->flags & SF_HAS_EVAL)
3955 && !deltanext /* atom is fixed width */
3956 && minnext != 0 /* CURLYM can't handle zero width */
3957 && ! (RExC_seen & REG_SEEN_EXACTF_SHARP_S) /* Nor \xDF */
3959 /* XXXX How to optimize if data == 0? */
3960 /* Optimize to a simpler form. */
3961 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3965 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3966 && (OP(nxt2) != WHILEM))
3968 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3969 /* Need to optimize away parenths. */
3970 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
3971 /* Set the parenth number. */
3972 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3974 oscan->flags = (U8)ARG(nxt);
3975 if (RExC_open_parens) {
3976 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3977 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3979 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3980 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3983 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3984 OP(nxt + 1) = OPTIMIZED; /* was count. */
3985 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
3986 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
3989 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3990 regnode *nnxt = regnext(nxt1);
3992 if (reg_off_by_arg[OP(nxt1)])
3993 ARG_SET(nxt1, nxt2 - nxt1);
3994 else if (nxt2 - nxt1 < U16_MAX)
3995 NEXT_OFF(nxt1) = nxt2 - nxt1;
3997 OP(nxt) = NOTHING; /* Cannot beautify */
4002 /* Optimize again: */
4003 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4004 NULL, stopparen, recursed, NULL, 0,depth+1);
4009 else if ((OP(oscan) == CURLYX)
4010 && (flags & SCF_WHILEM_VISITED_POS)
4011 /* See the comment on a similar expression above.
4012 However, this time it's not a subexpression
4013 we care about, but the expression itself. */
4014 && (maxcount == REG_INFTY)
4015 && data && ++data->whilem_c < 16) {
4016 /* This stays as CURLYX, we can put the count/of pair. */
4017 /* Find WHILEM (as in regexec.c) */
4018 regnode *nxt = oscan + NEXT_OFF(oscan);
4020 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4022 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4023 | (RExC_whilem_seen << 4)); /* On WHILEM */
4025 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4027 if (flags & SCF_DO_SUBSTR) {
4028 SV *last_str = NULL;
4029 int counted = mincount != 0;
4031 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
4032 #if defined(SPARC64_GCC_WORKAROUND)
4035 const char *s = NULL;
4038 if (pos_before >= data->last_start_min)
4041 b = data->last_start_min;
4044 s = SvPV_const(data->last_found, l);
4045 old = b - data->last_start_min;
4048 I32 b = pos_before >= data->last_start_min
4049 ? pos_before : data->last_start_min;
4051 const char * const s = SvPV_const(data->last_found, l);
4052 I32 old = b - data->last_start_min;
4056 old = utf8_hop((U8*)s, old) - (U8*)s;
4058 /* Get the added string: */
4059 last_str = newSVpvn_utf8(s + old, l, UTF);
4060 if (deltanext == 0 && pos_before == b) {
4061 /* What was added is a constant string */
4063 SvGROW(last_str, (mincount * l) + 1);
4064 repeatcpy(SvPVX(last_str) + l,
4065 SvPVX_const(last_str), l, mincount - 1);
4066 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4067 /* Add additional parts. */
4068 SvCUR_set(data->last_found,
4069 SvCUR(data->last_found) - l);
4070 sv_catsv(data->last_found, last_str);
4072 SV * sv = data->last_found;
4074 SvUTF8(sv) && SvMAGICAL(sv) ?
4075 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4076 if (mg && mg->mg_len >= 0)
4077 mg->mg_len += CHR_SVLEN(last_str) - l;
4079 data->last_end += l * (mincount - 1);
4082 /* start offset must point into the last copy */
4083 data->last_start_min += minnext * (mincount - 1);
4084 data->last_start_max += is_inf ? I32_MAX
4085 : (maxcount - 1) * (minnext + data->pos_delta);
4088 /* It is counted once already... */
4089 data->pos_min += minnext * (mincount - counted);
4091 PerlIO_printf(Perl_debug_log, "counted=%d deltanext=%d I32_MAX=%d minnext=%d maxcount=%d mincount=%d\n",
4092 counted, deltanext, I32_MAX, minnext, maxcount, mincount);
4093 if (deltanext != I32_MAX)
4094 PerlIO_printf(Perl_debug_log, "LHS=%d RHS=%d\n", -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount, I32_MAX - data->pos_delta);
4096 if (deltanext == I32_MAX || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= I32_MAX - data->pos_delta)
4097 data->pos_delta = I32_MAX;
4099 data->pos_delta += - counted * deltanext +
4100 (minnext + deltanext) * maxcount - minnext * mincount;
4101 if (mincount != maxcount) {
4102 /* Cannot extend fixed substrings found inside
4104 SCAN_COMMIT(pRExC_state,data,minlenp);
4105 if (mincount && last_str) {
4106 SV * const sv = data->last_found;
4107 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4108 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4112 sv_setsv(sv, last_str);
4113 data->last_end = data->pos_min;
4114 data->last_start_min =
4115 data->pos_min - CHR_SVLEN(last_str);
4116 data->last_start_max = is_inf
4118 : data->pos_min + data->pos_delta
4119 - CHR_SVLEN(last_str);
4121 data->longest = &(data->longest_float);
4123 SvREFCNT_dec(last_str);
4125 if (data && (fl & SF_HAS_EVAL))
4126 data->flags |= SF_HAS_EVAL;
4127 optimize_curly_tail:
4128 if (OP(oscan) != CURLYX) {
4129 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4131 NEXT_OFF(oscan) += NEXT_OFF(next);
4134 default: /* REF, and CLUMP only? */
4135 if (flags & SCF_DO_SUBSTR) {
4136 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4137 data->longest = &(data->longest_float);
4139 is_inf = is_inf_internal = 1;
4140 if (flags & SCF_DO_STCLASS_OR)
4141 cl_anything(pRExC_state, data->start_class);
4142 flags &= ~SCF_DO_STCLASS;
4146 else if (OP(scan) == LNBREAK) {
4147 if (flags & SCF_DO_STCLASS) {
4149 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4150 if (flags & SCF_DO_STCLASS_AND) {
4151 for (value = 0; value < 256; value++)
4152 if (!is_VERTWS_cp(value))
4153 ANYOF_BITMAP_CLEAR(data->start_class, value);
4156 for (value = 0; value < 256; value++)
4157 if (is_VERTWS_cp(value))
4158 ANYOF_BITMAP_SET(data->start_class, value);
4160 if (flags & SCF_DO_STCLASS_OR)
4161 cl_and(data->start_class, and_withp);
4162 flags &= ~SCF_DO_STCLASS;
4165 delta++; /* Because of the 2 char string cr-lf */
4166 if (flags & SCF_DO_SUBSTR) {
4167 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4169 data->pos_delta += 1;
4170 data->longest = &(data->longest_float);
4173 else if (REGNODE_SIMPLE(OP(scan))) {
4176 if (flags & SCF_DO_SUBSTR) {
4177 SCAN_COMMIT(pRExC_state,data,minlenp);
4181 if (flags & SCF_DO_STCLASS) {
4183 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4185 /* Some of the logic below assumes that switching
4186 locale on will only add false positives. */
4187 switch (PL_regkind[OP(scan)]) {
4193 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan));
4196 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4197 cl_anything(pRExC_state, data->start_class);
4200 if (OP(scan) == SANY)
4202 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
4203 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
4204 || ANYOF_CLASS_TEST_ANY_SET(data->start_class));
4205 cl_anything(pRExC_state, data->start_class);
4207 if (flags & SCF_DO_STCLASS_AND || !value)
4208 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
4211 if (flags & SCF_DO_STCLASS_AND)
4212 cl_and(data->start_class,
4213 (struct regnode_charclass_class*)scan);
4215 cl_or(pRExC_state, data->start_class,
4216 (struct regnode_charclass_class*)scan);
4224 classnum = FLAGS(scan);
4225 if (flags & SCF_DO_STCLASS_AND) {
4226 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4227 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum) + 1);
4228 for (value = 0; value < loop_max; value++) {
4229 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4230 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4236 if (data->start_class->flags & ANYOF_LOCALE) {
4237 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum));
4241 /* Even if under locale, set the bits for non-locale
4242 * in case it isn't a true locale-node. This will
4243 * create false positives if it truly is locale */
4244 for (value = 0; value < loop_max; value++) {
4245 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4246 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4258 classnum = FLAGS(scan);
4259 if (flags & SCF_DO_STCLASS_AND) {
4260 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4261 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum));
4262 for (value = 0; value < loop_max; value++) {
4263 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4264 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4270 if (data->start_class->flags & ANYOF_LOCALE) {
4271 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum) + 1);
4275 /* Even if under locale, set the bits for non-locale in
4276 * case it isn't a true locale-node. This will create
4277 * false positives if it truly is locale */
4278 for (value = 0; value < loop_max; value++) {
4279 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4280 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4283 if (PL_regkind[OP(scan)] == NPOSIXD) {
4284 data->start_class->flags |= ANYOF_NON_UTF8_LATIN1_ALL;
4290 if (flags & SCF_DO_STCLASS_OR)
4291 cl_and(data->start_class, and_withp);
4292 flags &= ~SCF_DO_STCLASS;
4295 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
4296 data->flags |= (OP(scan) == MEOL
4299 SCAN_COMMIT(pRExC_state, data, minlenp);
4302 else if ( PL_regkind[OP(scan)] == BRANCHJ
4303 /* Lookbehind, or need to calculate parens/evals/stclass: */
4304 && (scan->flags || data || (flags & SCF_DO_STCLASS))
4305 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
4306 if ( OP(scan) == UNLESSM &&
4308 OP(NEXTOPER(NEXTOPER(scan))) == NOTHING &&
4309 OP(regnext(NEXTOPER(NEXTOPER(scan)))) == SUCCEED
4312 regnode *upto= regnext(scan);
4314 SV * const mysv_val=sv_newmortal();
4315 DEBUG_STUDYDATA("OPFAIL",data,depth);
4317 /*DEBUG_PARSE_MSG("opfail");*/
4318 regprop(RExC_rx, mysv_val, upto);
4319 PerlIO_printf(Perl_debug_log, "~ replace with OPFAIL pointed at %s (%"IVdf") offset %"IVdf"\n",
4320 SvPV_nolen_const(mysv_val),
4321 (IV)REG_NODE_NUM(upto),
4326 NEXT_OFF(scan) = upto - scan;
4327 for (opt= scan + 1; opt < upto ; opt++)
4328 OP(opt) = OPTIMIZED;
4332 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4333 || OP(scan) == UNLESSM )
4335 /* Negative Lookahead/lookbehind
4336 In this case we can't do fixed string optimisation.
4339 I32 deltanext, minnext, fake = 0;
4341 struct regnode_charclass_class intrnl;
4344 data_fake.flags = 0;
4346 data_fake.whilem_c = data->whilem_c;
4347 data_fake.last_closep = data->last_closep;
4350 data_fake.last_closep = &fake;
4351 data_fake.pos_delta = delta;
4352 if ( flags & SCF_DO_STCLASS && !scan->flags
4353 && OP(scan) == IFMATCH ) { /* Lookahead */
4354 cl_init(pRExC_state, &intrnl);
4355 data_fake.start_class = &intrnl;
4356 f |= SCF_DO_STCLASS_AND;
4358 if (flags & SCF_WHILEM_VISITED_POS)
4359 f |= SCF_WHILEM_VISITED_POS;
4360 next = regnext(scan);
4361 nscan = NEXTOPER(NEXTOPER(scan));
4362 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
4363 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
4366 FAIL("Variable length lookbehind not implemented");
4368 else if (minnext > (I32)U8_MAX) {
4369 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4371 scan->flags = (U8)minnext;
4374 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4376 if (data_fake.flags & SF_HAS_EVAL)
4377 data->flags |= SF_HAS_EVAL;
4378 data->whilem_c = data_fake.whilem_c;
4380 if (f & SCF_DO_STCLASS_AND) {
4381 if (flags & SCF_DO_STCLASS_OR) {
4382 /* OR before, AND after: ideally we would recurse with
4383 * data_fake to get the AND applied by study of the
4384 * remainder of the pattern, and then derecurse;
4385 * *** HACK *** for now just treat as "no information".
4386 * See [perl #56690].
4388 cl_init(pRExC_state, data->start_class);
4390 /* AND before and after: combine and continue */
4391 const int was = TEST_SSC_EOS(data->start_class);
4393 cl_and(data->start_class, &intrnl);
4395 SET_SSC_EOS(data->start_class);
4399 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4401 /* Positive Lookahead/lookbehind
4402 In this case we can do fixed string optimisation,
4403 but we must be careful about it. Note in the case of
4404 lookbehind the positions will be offset by the minimum
4405 length of the pattern, something we won't know about
4406 until after the recurse.
4408 I32 deltanext, fake = 0;
4410 struct regnode_charclass_class intrnl;
4412 /* We use SAVEFREEPV so that when the full compile
4413 is finished perl will clean up the allocated
4414 minlens when it's all done. This way we don't
4415 have to worry about freeing them when we know
4416 they wont be used, which would be a pain.
4419 Newx( minnextp, 1, I32 );
4420 SAVEFREEPV(minnextp);
4423 StructCopy(data, &data_fake, scan_data_t);
4424 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
4427 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
4428 data_fake.last_found=newSVsv(data->last_found);
4432 data_fake.last_closep = &fake;
4433 data_fake.flags = 0;
4434 data_fake.pos_delta = delta;
4436 data_fake.flags |= SF_IS_INF;
4437 if ( flags & SCF_DO_STCLASS && !scan->flags
4438 && OP(scan) == IFMATCH ) { /* Lookahead */
4439 cl_init(pRExC_state, &intrnl);
4440 data_fake.start_class = &intrnl;
4441 f |= SCF_DO_STCLASS_AND;
4443 if (flags & SCF_WHILEM_VISITED_POS)
4444 f |= SCF_WHILEM_VISITED_POS;
4445 next = regnext(scan);
4446 nscan = NEXTOPER(NEXTOPER(scan));
4448 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
4449 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
4452 FAIL("Variable length lookbehind not implemented");
4454 else if (*minnextp > (I32)U8_MAX) {
4455 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4457 scan->flags = (U8)*minnextp;
4462 if (f & SCF_DO_STCLASS_AND) {
4463 const int was = TEST_SSC_EOS(data.start_class);
4465 cl_and(data->start_class, &intrnl);
4467 SET_SSC_EOS(data->start_class);
4470 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4472 if (data_fake.flags & SF_HAS_EVAL)
4473 data->flags |= SF_HAS_EVAL;
4474 data->whilem_c = data_fake.whilem_c;
4475 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
4476 if (RExC_rx->minlen<*minnextp)
4477 RExC_rx->minlen=*minnextp;
4478 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
4479 SvREFCNT_dec_NN(data_fake.last_found);
4481 if ( data_fake.minlen_fixed != minlenp )
4483 data->offset_fixed= data_fake.offset_fixed;
4484 data->minlen_fixed= data_fake.minlen_fixed;
4485 data->lookbehind_fixed+= scan->flags;
4487 if ( data_fake.minlen_float != minlenp )
4489 data->minlen_float= data_fake.minlen_float;
4490 data->offset_float_min=data_fake.offset_float_min;
4491 data->offset_float_max=data_fake.offset_float_max;
4492 data->lookbehind_float+= scan->flags;
4499 else if (OP(scan) == OPEN) {
4500 if (stopparen != (I32)ARG(scan))
4503 else if (OP(scan) == CLOSE) {
4504 if (stopparen == (I32)ARG(scan)) {
4507 if ((I32)ARG(scan) == is_par) {
4508 next = regnext(scan);
4510 if ( next && (OP(next) != WHILEM) && next < last)
4511 is_par = 0; /* Disable optimization */
4514 *(data->last_closep) = ARG(scan);
4516 else if (OP(scan) == EVAL) {
4518 data->flags |= SF_HAS_EVAL;
4520 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
4521 if (flags & SCF_DO_SUBSTR) {
4522 SCAN_COMMIT(pRExC_state,data,minlenp);
4523 flags &= ~SCF_DO_SUBSTR;
4525 if (data && OP(scan)==ACCEPT) {
4526 data->flags |= SCF_SEEN_ACCEPT;
4531 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
4533 if (flags & SCF_DO_SUBSTR) {
4534 SCAN_COMMIT(pRExC_state,data,minlenp);
4535 data->longest = &(data->longest_float);
4537 is_inf = is_inf_internal = 1;
4538 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4539 cl_anything(pRExC_state, data->start_class);
4540 flags &= ~SCF_DO_STCLASS;
4542 else if (OP(scan) == GPOS) {
4543 if (!(RExC_rx->extflags & RXf_GPOS_FLOAT) &&
4544 !(delta || is_inf || (data && data->pos_delta)))
4546 if (!(RExC_rx->extflags & RXf_ANCH) && (flags & SCF_DO_SUBSTR))
4547 RExC_rx->extflags |= RXf_ANCH_GPOS;
4548 if (RExC_rx->gofs < (U32)min)
4549 RExC_rx->gofs = min;
4551 RExC_rx->extflags |= RXf_GPOS_FLOAT;
4555 #ifdef TRIE_STUDY_OPT
4556 #ifdef FULL_TRIE_STUDY
4557 else if (PL_regkind[OP(scan)] == TRIE) {
4558 /* NOTE - There is similar code to this block above for handling
4559 BRANCH nodes on the initial study. If you change stuff here
4561 regnode *trie_node= scan;
4562 regnode *tail= regnext(scan);
4563 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4564 I32 max1 = 0, min1 = I32_MAX;
4565 struct regnode_charclass_class accum;
4567 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
4568 SCAN_COMMIT(pRExC_state, data,minlenp); /* Cannot merge strings after this. */
4569 if (flags & SCF_DO_STCLASS)
4570 cl_init_zero(pRExC_state, &accum);
4576 const regnode *nextbranch= NULL;
4579 for ( word=1 ; word <= trie->wordcount ; word++)
4581 I32 deltanext=0, minnext=0, f = 0, fake;
4582 struct regnode_charclass_class this_class;
4584 data_fake.flags = 0;
4586 data_fake.whilem_c = data->whilem_c;
4587 data_fake.last_closep = data->last_closep;
4590 data_fake.last_closep = &fake;
4591 data_fake.pos_delta = delta;
4592 if (flags & SCF_DO_STCLASS) {
4593 cl_init(pRExC_state, &this_class);
4594 data_fake.start_class = &this_class;
4595 f = SCF_DO_STCLASS_AND;
4597 if (flags & SCF_WHILEM_VISITED_POS)
4598 f |= SCF_WHILEM_VISITED_POS;
4600 if (trie->jump[word]) {
4602 nextbranch = trie_node + trie->jump[0];
4603 scan= trie_node + trie->jump[word];
4604 /* We go from the jump point to the branch that follows
4605 it. Note this means we need the vestigal unused branches
4606 even though they arent otherwise used.
4608 minnext = study_chunk(pRExC_state, &scan, minlenp,
4609 &deltanext, (regnode *)nextbranch, &data_fake,
4610 stopparen, recursed, NULL, f,depth+1);
4612 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
4613 nextbranch= regnext((regnode*)nextbranch);
4615 if (min1 > (I32)(minnext + trie->minlen))
4616 min1 = minnext + trie->minlen;
4617 if (deltanext == I32_MAX) {
4618 is_inf = is_inf_internal = 1;
4620 } else if (max1 < (I32)(minnext + deltanext + trie->maxlen))
4621 max1 = minnext + deltanext + trie->maxlen;
4623 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4625 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4626 if ( stopmin > min + min1)
4627 stopmin = min + min1;
4628 flags &= ~SCF_DO_SUBSTR;
4630 data->flags |= SCF_SEEN_ACCEPT;
4633 if (data_fake.flags & SF_HAS_EVAL)
4634 data->flags |= SF_HAS_EVAL;
4635 data->whilem_c = data_fake.whilem_c;
4637 if (flags & SCF_DO_STCLASS)
4638 cl_or(pRExC_state, &accum, &this_class);
4641 if (flags & SCF_DO_SUBSTR) {
4642 data->pos_min += min1;
4643 data->pos_delta += max1 - min1;
4644 if (max1 != min1 || is_inf)
4645 data->longest = &(data->longest_float);
4648 delta += max1 - min1;
4649 if (flags & SCF_DO_STCLASS_OR) {
4650 cl_or(pRExC_state, data->start_class, &accum);
4652 cl_and(data->start_class, and_withp);
4653 flags &= ~SCF_DO_STCLASS;
4656 else if (flags & SCF_DO_STCLASS_AND) {
4658 cl_and(data->start_class, &accum);
4659 flags &= ~SCF_DO_STCLASS;
4662 /* Switch to OR mode: cache the old value of
4663 * data->start_class */
4665 StructCopy(data->start_class, and_withp,
4666 struct regnode_charclass_class);
4667 flags &= ~SCF_DO_STCLASS_AND;
4668 StructCopy(&accum, data->start_class,
4669 struct regnode_charclass_class);
4670 flags |= SCF_DO_STCLASS_OR;
4671 SET_SSC_EOS(data->start_class);
4678 else if (PL_regkind[OP(scan)] == TRIE) {
4679 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4682 min += trie->minlen;
4683 delta += (trie->maxlen - trie->minlen);
4684 flags &= ~SCF_DO_STCLASS; /* xxx */
4685 if (flags & SCF_DO_SUBSTR) {
4686 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4687 data->pos_min += trie->minlen;
4688 data->pos_delta += (trie->maxlen - trie->minlen);
4689 if (trie->maxlen != trie->minlen)
4690 data->longest = &(data->longest_float);
4692 if (trie->jump) /* no more substrings -- for now /grr*/
4693 flags &= ~SCF_DO_SUBSTR;
4695 #endif /* old or new */
4696 #endif /* TRIE_STUDY_OPT */
4698 /* Else: zero-length, ignore. */
4699 scan = regnext(scan);
4704 stopparen = frame->stop;
4705 frame = frame->prev;
4706 goto fake_study_recurse;
4711 DEBUG_STUDYDATA("pre-fin:",data,depth);
4714 *deltap = is_inf_internal ? I32_MAX : delta;
4715 if (flags & SCF_DO_SUBSTR && is_inf)
4716 data->pos_delta = I32_MAX - data->pos_min;
4717 if (is_par > (I32)U8_MAX)
4719 if (is_par && pars==1 && data) {
4720 data->flags |= SF_IN_PAR;
4721 data->flags &= ~SF_HAS_PAR;
4723 else if (pars && data) {
4724 data->flags |= SF_HAS_PAR;
4725 data->flags &= ~SF_IN_PAR;
4727 if (flags & SCF_DO_STCLASS_OR)
4728 cl_and(data->start_class, and_withp);
4729 if (flags & SCF_TRIE_RESTUDY)
4730 data->flags |= SCF_TRIE_RESTUDY;
4732 DEBUG_STUDYDATA("post-fin:",data,depth);
4734 return min < stopmin ? min : stopmin;
4738 S_add_data(RExC_state_t *pRExC_state, U32 n, const char *s)
4740 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
4742 PERL_ARGS_ASSERT_ADD_DATA;
4744 Renewc(RExC_rxi->data,
4745 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
4746 char, struct reg_data);
4748 Renew(RExC_rxi->data->what, count + n, U8);
4750 Newx(RExC_rxi->data->what, n, U8);
4751 RExC_rxi->data->count = count + n;
4752 Copy(s, RExC_rxi->data->what + count, n, U8);
4756 /*XXX: todo make this not included in a non debugging perl */
4757 #ifndef PERL_IN_XSUB_RE
4759 Perl_reginitcolors(pTHX)
4762 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
4764 char *t = savepv(s);
4768 t = strchr(t, '\t');
4774 PL_colors[i] = t = (char *)"";
4779 PL_colors[i++] = (char *)"";
4786 #ifdef TRIE_STUDY_OPT
4787 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
4790 (data.flags & SCF_TRIE_RESTUDY) \
4798 #define CHECK_RESTUDY_GOTO_butfirst
4802 * pregcomp - compile a regular expression into internal code
4804 * Decides which engine's compiler to call based on the hint currently in
4808 #ifndef PERL_IN_XSUB_RE
4810 /* return the currently in-scope regex engine (or the default if none) */
4812 regexp_engine const *
4813 Perl_current_re_engine(pTHX)
4817 if (IN_PERL_COMPILETIME) {
4818 HV * const table = GvHV(PL_hintgv);
4822 return &reh_regexp_engine;
4823 ptr = hv_fetchs(table, "regcomp", FALSE);
4824 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
4825 return &reh_regexp_engine;
4826 return INT2PTR(regexp_engine*,SvIV(*ptr));
4830 if (!PL_curcop->cop_hints_hash)
4831 return &reh_regexp_engine;
4832 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
4833 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
4834 return &reh_regexp_engine;
4835 return INT2PTR(regexp_engine*,SvIV(ptr));
4841 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
4844 regexp_engine const *eng = current_re_engine();
4845 GET_RE_DEBUG_FLAGS_DECL;
4847 PERL_ARGS_ASSERT_PREGCOMP;
4849 /* Dispatch a request to compile a regexp to correct regexp engine. */
4851 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
4854 return CALLREGCOMP_ENG(eng, pattern, flags);
4858 /* public(ish) entry point for the perl core's own regex compiling code.
4859 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
4860 * pattern rather than a list of OPs, and uses the internal engine rather
4861 * than the current one */
4864 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
4866 SV *pat = pattern; /* defeat constness! */
4867 PERL_ARGS_ASSERT_RE_COMPILE;
4868 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
4869 #ifdef PERL_IN_XSUB_RE
4874 NULL, NULL, rx_flags, 0);
4877 /* see if there are any run-time code blocks in the pattern.
4878 * False positives are allowed */
4881 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state, OP *expr,
4882 U32 pm_flags, char *pat, STRLEN plen)
4887 /* avoid infinitely recursing when we recompile the pattern parcelled up
4888 * as qr'...'. A single constant qr// string can't have have any
4889 * run-time component in it, and thus, no runtime code. (A non-qr
4890 * string, however, can, e.g. $x =~ '(?{})') */
4891 if ((pm_flags & PMf_IS_QR) && expr && expr->op_type == OP_CONST)
4894 for (s = 0; s < plen; s++) {
4895 if (n < pRExC_state->num_code_blocks
4896 && s == pRExC_state->code_blocks[n].start)
4898 s = pRExC_state->code_blocks[n].end;
4902 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
4904 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
4906 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
4913 /* Handle run-time code blocks. We will already have compiled any direct
4914 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
4915 * copy of it, but with any literal code blocks blanked out and
4916 * appropriate chars escaped; then feed it into
4918 * eval "qr'modified_pattern'"
4922 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
4926 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
4928 * After eval_sv()-ing that, grab any new code blocks from the returned qr
4929 * and merge them with any code blocks of the original regexp.
4931 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
4932 * instead, just save the qr and return FALSE; this tells our caller that
4933 * the original pattern needs upgrading to utf8.
4937 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
4938 char *pat, STRLEN plen)
4942 GET_RE_DEBUG_FLAGS_DECL;
4944 if (pRExC_state->runtime_code_qr) {
4945 /* this is the second time we've been called; this should
4946 * only happen if the main pattern got upgraded to utf8
4947 * during compilation; re-use the qr we compiled first time
4948 * round (which should be utf8 too)
4950 qr = pRExC_state->runtime_code_qr;
4951 pRExC_state->runtime_code_qr = NULL;
4952 assert(RExC_utf8 && SvUTF8(qr));
4958 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
4962 /* determine how many extra chars we need for ' and \ escaping */
4963 for (s = 0; s < plen; s++) {
4964 if (pat[s] == '\'' || pat[s] == '\\')
4968 Newx(newpat, newlen, char);
4970 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
4972 for (s = 0; s < plen; s++) {
4973 if (n < pRExC_state->num_code_blocks
4974 && s == pRExC_state->code_blocks[n].start)
4976 /* blank out literal code block */
4977 assert(pat[s] == '(');
4978 while (s <= pRExC_state->code_blocks[n].end) {
4986 if (pat[s] == '\'' || pat[s] == '\\')
4991 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
4995 PerlIO_printf(Perl_debug_log,
4996 "%sre-parsing pattern for runtime code:%s %s\n",
4997 PL_colors[4],PL_colors[5],newpat);
5000 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5006 PUSHSTACKi(PERLSI_REQUIRE);
5007 /* this causes the toker to collapse \\ into \ when parsing
5008 * qr''; normally only q'' does this. It also alters hints
5010 PL_reg_state.re_reparsing = TRUE;
5011 eval_sv(sv, G_SCALAR);
5012 SvREFCNT_dec_NN(sv);
5017 SV * const errsv = ERRSV;
5018 if (SvTRUE_NN(errsv))
5020 Safefree(pRExC_state->code_blocks);
5021 /* use croak_sv ? */
5022 Perl_croak_nocontext("%s", SvPV_nolen_const(errsv));
5025 assert(SvROK(qr_ref));
5027 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5028 /* the leaving below frees the tmp qr_ref.
5029 * Give qr a life of its own */
5037 if (!RExC_utf8 && SvUTF8(qr)) {
5038 /* first time through; the pattern got upgraded; save the
5039 * qr for the next time through */
5040 assert(!pRExC_state->runtime_code_qr);
5041 pRExC_state->runtime_code_qr = qr;
5046 /* extract any code blocks within the returned qr// */
5049 /* merge the main (r1) and run-time (r2) code blocks into one */
5051 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
5052 struct reg_code_block *new_block, *dst;
5053 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
5056 if (!r2->num_code_blocks) /* we guessed wrong */
5058 SvREFCNT_dec_NN(qr);
5063 r1->num_code_blocks + r2->num_code_blocks,
5064 struct reg_code_block);
5067 while ( i1 < r1->num_code_blocks
5068 || i2 < r2->num_code_blocks)
5070 struct reg_code_block *src;
5073 if (i1 == r1->num_code_blocks) {
5074 src = &r2->code_blocks[i2++];
5077 else if (i2 == r2->num_code_blocks)
5078 src = &r1->code_blocks[i1++];
5079 else if ( r1->code_blocks[i1].start
5080 < r2->code_blocks[i2].start)
5082 src = &r1->code_blocks[i1++];
5083 assert(src->end < r2->code_blocks[i2].start);
5086 assert( r1->code_blocks[i1].start
5087 > r2->code_blocks[i2].start);
5088 src = &r2->code_blocks[i2++];
5090 assert(src->end < r1->code_blocks[i1].start);
5093 assert(pat[src->start] == '(');
5094 assert(pat[src->end] == ')');
5095 dst->start = src->start;
5096 dst->end = src->end;
5097 dst->block = src->block;
5098 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
5102 r1->num_code_blocks += r2->num_code_blocks;
5103 Safefree(r1->code_blocks);
5104 r1->code_blocks = new_block;
5107 SvREFCNT_dec_NN(qr);
5113 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest, SV** rx_utf8, SV** rx_substr, I32* rx_end_shift, I32 lookbehind, I32 offset, I32 *minlen, STRLEN longest_length, bool eol, bool meol)
5115 /* This is the common code for setting up the floating and fixed length
5116 * string data extracted from Perlre_op_compile() below. Returns a boolean
5117 * as to whether succeeded or not */
5121 if (! (longest_length
5122 || (eol /* Can't have SEOL and MULTI */
5123 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
5125 /* See comments for join_exact for why REG_SEEN_EXACTF_SHARP_S */
5126 || (RExC_seen & REG_SEEN_EXACTF_SHARP_S))
5131 /* copy the information about the longest from the reg_scan_data
5132 over to the program. */
5133 if (SvUTF8(sv_longest)) {
5134 *rx_utf8 = sv_longest;
5137 *rx_substr = sv_longest;
5140 /* end_shift is how many chars that must be matched that
5141 follow this item. We calculate it ahead of time as once the
5142 lookbehind offset is added in we lose the ability to correctly
5144 ml = minlen ? *(minlen) : (I32)longest_length;
5145 *rx_end_shift = ml - offset
5146 - longest_length + (SvTAIL(sv_longest) != 0)
5149 t = (eol/* Can't have SEOL and MULTI */
5150 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
5151 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
5157 * Perl_re_op_compile - the perl internal RE engine's function to compile a
5158 * regular expression into internal code.
5159 * The pattern may be passed either as:
5160 * a list of SVs (patternp plus pat_count)
5161 * a list of OPs (expr)
5162 * If both are passed, the SV list is used, but the OP list indicates
5163 * which SVs are actually pre-compiled code blocks
5165 * The SVs in the list have magic and qr overloading applied to them (and
5166 * the list may be modified in-place with replacement SVs in the latter
5169 * If the pattern hasn't changed from old_re, then old_re will be
5172 * eng is the current engine. If that engine has an op_comp method, then
5173 * handle directly (i.e. we assume that op_comp was us); otherwise, just
5174 * do the initial concatenation of arguments and pass on to the external
5177 * If is_bare_re is not null, set it to a boolean indicating whether the
5178 * arg list reduced (after overloading) to a single bare regex which has
5179 * been returned (i.e. /$qr/).
5181 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
5183 * pm_flags contains the PMf_* flags, typically based on those from the
5184 * pm_flags field of the related PMOP. Currently we're only interested in
5185 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
5187 * We can't allocate space until we know how big the compiled form will be,
5188 * but we can't compile it (and thus know how big it is) until we've got a
5189 * place to put the code. So we cheat: we compile it twice, once with code
5190 * generation turned off and size counting turned on, and once "for real".
5191 * This also means that we don't allocate space until we are sure that the
5192 * thing really will compile successfully, and we never have to move the
5193 * code and thus invalidate pointers into it. (Note that it has to be in
5194 * one piece because free() must be able to free it all.) [NB: not true in perl]
5196 * Beware that the optimization-preparation code in here knows about some
5197 * of the structure of the compiled regexp. [I'll say.]
5201 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
5202 OP *expr, const regexp_engine* eng, REGEXP *old_re,
5203 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
5208 regexp_internal *ri;
5217 SV *code_blocksv = NULL;
5219 /* these are all flags - maybe they should be turned
5220 * into a single int with different bit masks */
5221 I32 sawlookahead = 0;
5224 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
5225 bool code_is_utf8 = 0;
5227 bool runtime_code = 0;
5229 RExC_state_t RExC_state;
5230 RExC_state_t * const pRExC_state = &RExC_state;
5231 #ifdef TRIE_STUDY_OPT
5233 RExC_state_t copyRExC_state;
5235 GET_RE_DEBUG_FLAGS_DECL;
5237 PERL_ARGS_ASSERT_RE_OP_COMPILE;
5239 DEBUG_r(if (!PL_colorset) reginitcolors());
5241 #ifndef PERL_IN_XSUB_RE
5242 /* Initialize these here instead of as-needed, as is quick and avoids
5243 * having to test them each time otherwise */
5244 if (! PL_AboveLatin1) {
5245 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
5246 PL_ASCII = _new_invlist_C_array(ASCII_invlist);
5247 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
5249 PL_L1Posix_ptrs[_CC_ALPHANUMERIC]
5250 = _new_invlist_C_array(L1PosixAlnum_invlist);
5251 PL_Posix_ptrs[_CC_ALPHANUMERIC]
5252 = _new_invlist_C_array(PosixAlnum_invlist);
5254 PL_L1Posix_ptrs[_CC_ALPHA]
5255 = _new_invlist_C_array(L1PosixAlpha_invlist);
5256 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(PosixAlpha_invlist);
5258 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(PosixBlank_invlist);
5259 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(XPosixBlank_invlist);
5261 /* Cased is the same as Alpha in the ASCII range */
5262 PL_L1Posix_ptrs[_CC_CASED] = _new_invlist_C_array(L1Cased_invlist);
5263 PL_Posix_ptrs[_CC_CASED] = _new_invlist_C_array(PosixAlpha_invlist);
5265 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(PosixCntrl_invlist);
5266 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(XPosixCntrl_invlist);
5268 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5269 PL_L1Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5271 PL_L1Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(L1PosixGraph_invlist);
5272 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(PosixGraph_invlist);
5274 PL_L1Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(L1PosixLower_invlist);
5275 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(PosixLower_invlist);
5277 PL_L1Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(L1PosixPrint_invlist);
5278 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(PosixPrint_invlist);
5280 PL_L1Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(L1PosixPunct_invlist);
5281 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(PosixPunct_invlist);
5283 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(PerlSpace_invlist);
5284 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(XPerlSpace_invlist);
5285 PL_Posix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(PosixSpace_invlist);
5286 PL_XPosix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(XPosixSpace_invlist);
5288 PL_L1Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(L1PosixUpper_invlist);
5289 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(PosixUpper_invlist);
5291 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(VertSpace_invlist);
5293 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(PosixWord_invlist);
5294 PL_L1Posix_ptrs[_CC_WORDCHAR]
5295 = _new_invlist_C_array(L1PosixWord_invlist);
5297 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(PosixXDigit_invlist);
5298 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(XPosixXDigit_invlist);
5300 PL_HasMultiCharFold = _new_invlist_C_array(_Perl_Multi_Char_Folds_invlist);
5304 pRExC_state->code_blocks = NULL;
5305 pRExC_state->num_code_blocks = 0;
5308 *is_bare_re = FALSE;
5310 if (expr && (expr->op_type == OP_LIST ||
5311 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
5313 /* is the source UTF8, and how many code blocks are there? */
5317 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5318 if (o->op_type == OP_CONST) {
5319 /* skip if we have SVs as well as OPs. In this case,
5320 * a) we decide utf8 based on SVs not OPs;
5321 * b) the current pad may not match that which the ops
5322 * were compiled in, so, so on threaded builds,
5323 * cSVOPo_sv would look in the wrong pad */
5324 if (!pat_count && SvUTF8(cSVOPo_sv))
5327 else if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
5328 /* count of DO blocks */
5332 pRExC_state->num_code_blocks = ncode;
5333 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
5338 /* handle a list of SVs */
5342 /* apply magic and RE overloading to each arg */
5343 for (svp = patternp; svp < patternp + pat_count; svp++) {
5346 if (SvROK(rx) && SvAMAGIC(rx)) {
5347 SV *sv = AMG_CALLunary(rx, regexp_amg);
5351 if (SvTYPE(sv) != SVt_REGEXP)
5352 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5358 if (pat_count > 1) {
5359 /* concat multiple args and find any code block indexes */
5364 STRLEN orig_patlen = 0;
5366 if (pRExC_state->num_code_blocks) {
5367 o = cLISTOPx(expr)->op_first;
5368 assert( o->op_type == OP_PUSHMARK
5369 || (o->op_type == OP_NULL && o->op_targ == OP_PUSHMARK)
5370 || o->op_type == OP_PADRANGE);
5374 pat = newSVpvn("", 0);
5377 /* determine if the pattern is going to be utf8 (needed
5378 * in advance to align code block indices correctly).
5379 * XXX This could fail to be detected for an arg with
5380 * overloading but not concat overloading; but the main effect
5381 * in this obscure case is to need a 'use re eval' for a
5382 * literal code block */
5383 for (svp = patternp; svp < patternp + pat_count; svp++) {
5390 for (svp = patternp; svp < patternp + pat_count; svp++) {
5391 SV *sv, *msv = *svp;
5394 /* we make the assumption here that each op in the list of
5395 * op_siblings maps to one SV pushed onto the stack,
5396 * except for code blocks, with have both an OP_NULL and
5398 * This allows us to match up the list of SVs against the
5399 * list of OPs to find the next code block.
5401 * Note that PUSHMARK PADSV PADSV ..
5403 * PADRANGE NULL NULL ..
5404 * so the alignment still works. */
5406 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL)) {
5407 assert(n < pRExC_state->num_code_blocks);
5408 pRExC_state->code_blocks[n].start = SvCUR(pat);
5409 pRExC_state->code_blocks[n].block = o;
5410 pRExC_state->code_blocks[n].src_regex = NULL;
5413 o = o->op_sibling; /* skip CONST */
5419 if ((SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5420 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5423 /* overloading involved: all bets are off over literal
5424 * code. Pretend we haven't seen it */
5425 pRExC_state->num_code_blocks -= n;
5431 while (SvAMAGIC(msv)
5432 && (sv = AMG_CALLunary(msv, string_amg))
5436 && SvRV(msv) == SvRV(sv))
5441 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5443 orig_patlen = SvCUR(pat);
5444 sv_catsv_nomg(pat, msv);
5447 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5450 /* extract any code blocks within any embedded qr//'s */
5451 if (rx && SvTYPE(rx) == SVt_REGEXP
5452 && RX_ENGINE((REGEXP*)rx)->op_comp)
5455 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5456 if (ri->num_code_blocks) {
5458 /* the presence of an embedded qr// with code means
5459 * we should always recompile: the text of the
5460 * qr// may not have changed, but it may be a
5461 * different closure than last time */
5463 Renew(pRExC_state->code_blocks,
5464 pRExC_state->num_code_blocks + ri->num_code_blocks,
5465 struct reg_code_block);
5466 pRExC_state->num_code_blocks += ri->num_code_blocks;
5467 for (i=0; i < ri->num_code_blocks; i++) {
5468 struct reg_code_block *src, *dst;
5469 STRLEN offset = orig_patlen
5470 + ReANY((REGEXP *)rx)->pre_prefix;
5471 assert(n < pRExC_state->num_code_blocks);
5472 src = &ri->code_blocks[i];
5473 dst = &pRExC_state->code_blocks[n];
5474 dst->start = src->start + offset;
5475 dst->end = src->end + offset;
5476 dst->block = src->block;
5477 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5491 while (SvAMAGIC(pat)
5492 && (sv = AMG_CALLunary(pat, string_amg))
5500 /* handle bare regex: foo =~ $re */
5505 if (SvTYPE(re) == SVt_REGEXP) {
5509 Safefree(pRExC_state->code_blocks);
5515 /* not a list of SVs, so must be a list of OPs */
5517 if (expr->op_type == OP_LIST) {
5522 pat = newSVpvn("", 0);
5527 /* given a list of CONSTs and DO blocks in expr, append all
5528 * the CONSTs to pat, and record the start and end of each
5529 * code block in code_blocks[] (each DO{} op is followed by an
5530 * OP_CONST containing the corresponding literal '(?{...})
5533 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5534 if (o->op_type == OP_CONST) {
5535 sv_catsv(pat, cSVOPo_sv);
5537 pRExC_state->code_blocks[i].end = SvCUR(pat)-1;
5541 else if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL)) {
5542 assert(i+1 < pRExC_state->num_code_blocks);
5543 pRExC_state->code_blocks[++i].start = SvCUR(pat);
5544 pRExC_state->code_blocks[i].block = o;
5545 pRExC_state->code_blocks[i].src_regex = NULL;
5551 assert(expr->op_type == OP_CONST);
5552 pat = cSVOPx_sv(expr);
5556 exp = SvPV_nomg(pat, plen);
5559 if (!eng->op_comp) {
5560 if ((SvUTF8(pat) && IN_BYTES)
5561 || SvGMAGICAL(pat) || SvAMAGIC(pat))
5563 /* make a temporary copy; either to convert to bytes,
5564 * or to avoid repeating get-magic / overloaded stringify */
5565 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
5566 (IN_BYTES ? 0 : SvUTF8(pat)));
5568 Safefree(pRExC_state->code_blocks);
5569 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
5572 /* ignore the utf8ness if the pattern is 0 length */
5573 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
5574 RExC_uni_semantics = 0;
5575 RExC_contains_locale = 0;
5576 pRExC_state->runtime_code_qr = NULL;
5579 SV *dsv= sv_newmortal();
5580 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
5581 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
5582 PL_colors[4],PL_colors[5],s);
5588 U8 *const src = (U8*)exp;
5591 STRLEN s = 0, d = 0;
5594 /* It's possible to write a regexp in ascii that represents Unicode
5595 codepoints outside of the byte range, such as via \x{100}. If we
5596 detect such a sequence we have to convert the entire pattern to utf8
5597 and then recompile, as our sizing calculation will have been based
5598 on 1 byte == 1 character, but we will need to use utf8 to encode
5599 at least some part of the pattern, and therefore must convert the whole
5602 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5603 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
5605 /* upgrade pattern to UTF8, and if there are code blocks,
5606 * recalculate the indices.
5607 * This is essentially an unrolled Perl_bytes_to_utf8() */
5609 Newx(dst, plen * 2 + 1, U8);
5612 const UV uv = NATIVE_TO_ASCII(src[s]);
5613 if (UNI_IS_INVARIANT(uv))
5614 dst[d] = (U8)UTF_TO_NATIVE(uv);
5616 dst[d++] = (U8)UTF8_EIGHT_BIT_HI(uv);
5617 dst[d] = (U8)UTF8_EIGHT_BIT_LO(uv);
5619 if (n < pRExC_state->num_code_blocks) {
5620 if (!do_end && pRExC_state->code_blocks[n].start == s) {
5621 pRExC_state->code_blocks[n].start = d;
5622 assert(dst[d] == '(');
5625 else if (do_end && pRExC_state->code_blocks[n].end == s) {
5626 pRExC_state->code_blocks[n].end = d;
5627 assert(dst[d] == ')');
5640 RExC_orig_utf8 = RExC_utf8 = 1;
5644 /* return old regex if pattern hasn't changed */
5648 && !!RX_UTF8(old_re) == !!RExC_utf8
5649 && RX_PRECOMP(old_re)
5650 && RX_PRELEN(old_re) == plen
5651 && memEQ(RX_PRECOMP(old_re), exp, plen))
5653 /* with runtime code, always recompile */
5654 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, expr, pm_flags,
5656 if (!runtime_code) {
5657 Safefree(pRExC_state->code_blocks);
5661 else if ((pm_flags & PMf_USE_RE_EVAL)
5662 /* this second condition covers the non-regex literal case,
5663 * i.e. $foo =~ '(?{})'. */
5664 || ( !PL_reg_state.re_reparsing && IN_PERL_COMPILETIME
5665 && (PL_hints & HINT_RE_EVAL))
5667 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, expr, pm_flags,
5670 rx_flags = orig_rx_flags;
5672 if (initial_charset == REGEX_LOCALE_CHARSET) {
5673 RExC_contains_locale = 1;
5675 else if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
5677 /* Set to use unicode semantics if the pattern is in utf8 and has the
5678 * 'depends' charset specified, as it means unicode when utf8 */
5679 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5683 RExC_flags = rx_flags;
5684 RExC_pm_flags = pm_flags;
5687 if (TAINTING_get && TAINT_get)
5688 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
5690 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
5691 /* whoops, we have a non-utf8 pattern, whilst run-time code
5692 * got compiled as utf8. Try again with a utf8 pattern */
5693 goto redo_first_pass;
5696 assert(!pRExC_state->runtime_code_qr);
5701 RExC_in_lookbehind = 0;
5702 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
5704 RExC_override_recoding = 0;
5705 RExC_in_multi_char_class = 0;
5707 /* First pass: determine size, legality. */
5715 RExC_emit = &PL_regdummy;
5716 RExC_whilem_seen = 0;
5717 RExC_open_parens = NULL;
5718 RExC_close_parens = NULL;
5720 RExC_paren_names = NULL;
5722 RExC_paren_name_list = NULL;
5724 RExC_recurse = NULL;
5725 RExC_recurse_count = 0;
5726 pRExC_state->code_index = 0;
5728 #if 0 /* REGC() is (currently) a NOP at the first pass.
5729 * Clever compilers notice this and complain. --jhi */
5730 REGC((U8)REG_MAGIC, (char*)RExC_emit);
5733 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
5735 RExC_lastparse=NULL;
5737 /* reg may croak on us, not giving us a chance to free
5738 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
5739 need it to survive as long as the regexp (qr/(?{})/).
5740 We must check that code_blocksv is not already set, because we may
5741 have jumped back to restart the sizing pass. */
5742 if (pRExC_state->code_blocks && !code_blocksv) {
5743 code_blocksv = newSV_type(SVt_PV);
5744 SAVEFREESV(code_blocksv);
5745 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
5746 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
5748 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5749 if (flags & RESTART_UTF8) {
5750 goto redo_first_pass;
5752 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#X", flags);
5755 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
5758 PerlIO_printf(Perl_debug_log,
5759 "Required size %"IVdf" nodes\n"
5760 "Starting second pass (creation)\n",
5763 RExC_lastparse=NULL;
5766 /* The first pass could have found things that force Unicode semantics */
5767 if ((RExC_utf8 || RExC_uni_semantics)
5768 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
5770 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5773 /* Small enough for pointer-storage convention?
5774 If extralen==0, this means that we will not need long jumps. */
5775 if (RExC_size >= 0x10000L && RExC_extralen)
5776 RExC_size += RExC_extralen;
5779 if (RExC_whilem_seen > 15)
5780 RExC_whilem_seen = 15;
5782 /* Allocate space and zero-initialize. Note, the two step process
5783 of zeroing when in debug mode, thus anything assigned has to
5784 happen after that */
5785 rx = (REGEXP*) newSV_type(SVt_REGEXP);
5787 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
5788 char, regexp_internal);
5789 if ( r == NULL || ri == NULL )
5790 FAIL("Regexp out of space");
5792 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
5793 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode), char);
5795 /* bulk initialize base fields with 0. */
5796 Zero(ri, sizeof(regexp_internal), char);
5799 /* non-zero initialization begins here */
5802 r->extflags = rx_flags;
5803 if (pm_flags & PMf_IS_QR) {
5804 ri->code_blocks = pRExC_state->code_blocks;
5805 ri->num_code_blocks = pRExC_state->num_code_blocks;
5810 for (n = 0; n < pRExC_state->num_code_blocks; n++)
5811 if (pRExC_state->code_blocks[n].src_regex)
5812 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
5813 SAVEFREEPV(pRExC_state->code_blocks);
5817 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
5818 bool has_charset = (get_regex_charset(r->extflags) != REGEX_DEPENDS_CHARSET);
5820 /* The caret is output if there are any defaults: if not all the STD
5821 * flags are set, or if no character set specifier is needed */
5823 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
5825 bool has_runon = ((RExC_seen & REG_SEEN_RUN_ON_COMMENT)==REG_SEEN_RUN_ON_COMMENT);
5826 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
5827 >> RXf_PMf_STD_PMMOD_SHIFT);
5828 const char *fptr = STD_PAT_MODS; /*"msix"*/
5830 /* Allocate for the worst case, which is all the std flags are turned
5831 * on. If more precision is desired, we could do a population count of
5832 * the flags set. This could be done with a small lookup table, or by
5833 * shifting, masking and adding, or even, when available, assembly
5834 * language for a machine-language population count.
5835 * We never output a minus, as all those are defaults, so are
5836 * covered by the caret */
5837 const STRLEN wraplen = plen + has_p + has_runon
5838 + has_default /* If needs a caret */
5840 /* If needs a character set specifier */
5841 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
5842 + (sizeof(STD_PAT_MODS) - 1)
5843 + (sizeof("(?:)") - 1);
5845 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
5846 r->xpv_len_u.xpvlenu_pv = p;
5848 SvFLAGS(rx) |= SVf_UTF8;
5851 /* If a default, cover it using the caret */
5853 *p++= DEFAULT_PAT_MOD;
5857 const char* const name = get_regex_charset_name(r->extflags, &len);
5858 Copy(name, p, len, char);
5862 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
5865 while((ch = *fptr++)) {
5873 Copy(RExC_precomp, p, plen, char);
5874 assert ((RX_WRAPPED(rx) - p) < 16);
5875 r->pre_prefix = p - RX_WRAPPED(rx);
5881 SvCUR_set(rx, p - RX_WRAPPED(rx));
5885 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
5887 if (RExC_seen & REG_SEEN_RECURSE) {
5888 Newxz(RExC_open_parens, RExC_npar,regnode *);
5889 SAVEFREEPV(RExC_open_parens);
5890 Newxz(RExC_close_parens,RExC_npar,regnode *);
5891 SAVEFREEPV(RExC_close_parens);
5894 /* Useful during FAIL. */
5895 #ifdef RE_TRACK_PATTERN_OFFSETS
5896 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
5897 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
5898 "%s %"UVuf" bytes for offset annotations.\n",
5899 ri->u.offsets ? "Got" : "Couldn't get",
5900 (UV)((2*RExC_size+1) * sizeof(U32))));
5902 SetProgLen(ri,RExC_size);
5906 REH_CALL_COMP_BEGIN_HOOK(pRExC_state->rx);
5908 /* Second pass: emit code. */
5909 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
5910 RExC_pm_flags = pm_flags;
5915 RExC_emit_start = ri->program;
5916 RExC_emit = ri->program;
5917 RExC_emit_bound = ri->program + RExC_size + 1;
5918 pRExC_state->code_index = 0;
5920 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
5921 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5923 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#X", flags);
5925 /* XXXX To minimize changes to RE engine we always allocate
5926 3-units-long substrs field. */
5927 Newx(r->substrs, 1, struct reg_substr_data);
5928 if (RExC_recurse_count) {
5929 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
5930 SAVEFREEPV(RExC_recurse);
5934 r->minlen = minlen = sawlookahead = sawplus = sawopen = 0;
5935 Zero(r->substrs, 1, struct reg_substr_data);
5937 #ifdef TRIE_STUDY_OPT
5939 StructCopy(&zero_scan_data, &data, scan_data_t);
5940 copyRExC_state = RExC_state;
5943 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
5945 RExC_state = copyRExC_state;
5946 if (seen & REG_TOP_LEVEL_BRANCHES)
5947 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
5949 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES;
5950 StructCopy(&zero_scan_data, &data, scan_data_t);
5953 StructCopy(&zero_scan_data, &data, scan_data_t);
5956 /* Dig out information for optimizations. */
5957 r->extflags = RExC_flags; /* was pm_op */
5958 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
5961 SvUTF8_on(rx); /* Unicode in it? */
5962 ri->regstclass = NULL;
5963 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
5964 r->intflags |= PREGf_NAUGHTY;
5965 scan = ri->program + 1; /* First BRANCH. */
5967 /* testing for BRANCH here tells us whether there is "must appear"
5968 data in the pattern. If there is then we can use it for optimisations */
5969 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES)) { /* Only one top-level choice. */
5971 STRLEN longest_float_length, longest_fixed_length;
5972 struct regnode_charclass_class ch_class; /* pointed to by data */
5974 I32 last_close = 0; /* pointed to by data */
5975 regnode *first= scan;
5976 regnode *first_next= regnext(first);
5978 * Skip introductions and multiplicators >= 1
5979 * so that we can extract the 'meat' of the pattern that must
5980 * match in the large if() sequence following.
5981 * NOTE that EXACT is NOT covered here, as it is normally
5982 * picked up by the optimiser separately.
5984 * This is unfortunate as the optimiser isnt handling lookahead
5985 * properly currently.
5988 while ((OP(first) == OPEN && (sawopen = 1)) ||
5989 /* An OR of *one* alternative - should not happen now. */
5990 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
5991 /* for now we can't handle lookbehind IFMATCH*/
5992 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
5993 (OP(first) == PLUS) ||
5994 (OP(first) == MINMOD) ||
5995 /* An {n,m} with n>0 */
5996 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
5997 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6000 * the only op that could be a regnode is PLUS, all the rest
6001 * will be regnode_1 or regnode_2.
6004 if (OP(first) == PLUS)
6007 first += regarglen[OP(first)];
6009 first = NEXTOPER(first);
6010 first_next= regnext(first);
6013 /* Starting-point info. */
6015 DEBUG_PEEP("first:",first,0);
6016 /* Ignore EXACT as we deal with it later. */
6017 if (PL_regkind[OP(first)] == EXACT) {
6018 if (OP(first) == EXACT)
6019 NOOP; /* Empty, get anchored substr later. */
6021 ri->regstclass = first;
6024 else if (PL_regkind[OP(first)] == TRIE &&
6025 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6028 /* this can happen only on restudy */
6029 if ( OP(first) == TRIE ) {
6030 struct regnode_1 *trieop = (struct regnode_1 *)
6031 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6032 StructCopy(first,trieop,struct regnode_1);
6033 trie_op=(regnode *)trieop;
6035 struct regnode_charclass *trieop = (struct regnode_charclass *)
6036 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6037 StructCopy(first,trieop,struct regnode_charclass);
6038 trie_op=(regnode *)trieop;
6041 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6042 ri->regstclass = trie_op;
6045 else if (REGNODE_SIMPLE(OP(first)))
6046 ri->regstclass = first;
6047 else if (PL_regkind[OP(first)] == BOUND ||
6048 PL_regkind[OP(first)] == NBOUND)
6049 ri->regstclass = first;
6050 else if (PL_regkind[OP(first)] == BOL) {
6051 r->extflags |= (OP(first) == MBOL
6053 : (OP(first) == SBOL
6056 first = NEXTOPER(first);
6059 else if (OP(first) == GPOS) {
6060 r->extflags |= RXf_ANCH_GPOS;
6061 first = NEXTOPER(first);
6064 else if ((!sawopen || !RExC_sawback) &&
6065 (OP(first) == STAR &&
6066 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6067 !(r->extflags & RXf_ANCH) && !pRExC_state->num_code_blocks)
6069 /* turn .* into ^.* with an implied $*=1 */
6071 (OP(NEXTOPER(first)) == REG_ANY)
6074 r->extflags |= type;
6075 r->intflags |= PREGf_IMPLICIT;
6076 first = NEXTOPER(first);
6079 if (sawplus && !sawlookahead && (!sawopen || !RExC_sawback)
6080 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6081 /* x+ must match at the 1st pos of run of x's */
6082 r->intflags |= PREGf_SKIP;
6084 /* Scan is after the zeroth branch, first is atomic matcher. */
6085 #ifdef TRIE_STUDY_OPT
6088 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6089 (IV)(first - scan + 1))
6093 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6094 (IV)(first - scan + 1))
6100 * If there's something expensive in the r.e., find the
6101 * longest literal string that must appear and make it the
6102 * regmust. Resolve ties in favor of later strings, since
6103 * the regstart check works with the beginning of the r.e.
6104 * and avoiding duplication strengthens checking. Not a
6105 * strong reason, but sufficient in the absence of others.
6106 * [Now we resolve ties in favor of the earlier string if
6107 * it happens that c_offset_min has been invalidated, since the
6108 * earlier string may buy us something the later one won't.]
6111 data.longest_fixed = newSVpvs("");
6112 data.longest_float = newSVpvs("");
6113 data.last_found = newSVpvs("");
6114 data.longest = &(data.longest_fixed);
6115 ENTER_with_name("study_chunk");
6116 SAVEFREESV(data.longest_fixed);
6117 SAVEFREESV(data.longest_float);
6118 SAVEFREESV(data.last_found);
6120 if (!ri->regstclass) {
6121 cl_init(pRExC_state, &ch_class);
6122 data.start_class = &ch_class;
6123 stclass_flag = SCF_DO_STCLASS_AND;
6124 } else /* XXXX Check for BOUND? */
6126 data.last_closep = &last_close;
6128 minlen = study_chunk(pRExC_state, &first, &minlen, &fake, scan + RExC_size, /* Up to end */
6129 &data, -1, NULL, NULL,
6130 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag,0);
6133 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6136 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6137 && data.last_start_min == 0 && data.last_end > 0
6138 && !RExC_seen_zerolen
6139 && !(RExC_seen & REG_SEEN_VERBARG)
6140 && (!(RExC_seen & REG_SEEN_GPOS) || (r->extflags & RXf_ANCH_GPOS)))
6141 r->extflags |= RXf_CHECK_ALL;
6142 scan_commit(pRExC_state, &data,&minlen,0);
6144 longest_float_length = CHR_SVLEN(data.longest_float);
6146 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6147 && data.offset_fixed == data.offset_float_min
6148 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6149 && S_setup_longest (aTHX_ pRExC_state,
6153 &(r->float_end_shift),
6154 data.lookbehind_float,
6155 data.offset_float_min,
6157 longest_float_length,
6158 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6159 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6161 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6162 r->float_max_offset = data.offset_float_max;
6163 if (data.offset_float_max < I32_MAX) /* Don't offset infinity */
6164 r->float_max_offset -= data.lookbehind_float;
6165 SvREFCNT_inc_simple_void_NN(data.longest_float);
6168 r->float_substr = r->float_utf8 = NULL;
6169 longest_float_length = 0;
6172 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6174 if (S_setup_longest (aTHX_ pRExC_state,
6176 &(r->anchored_utf8),
6177 &(r->anchored_substr),
6178 &(r->anchored_end_shift),
6179 data.lookbehind_fixed,
6182 longest_fixed_length,
6183 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6184 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6186 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6187 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6190 r->anchored_substr = r->anchored_utf8 = NULL;
6191 longest_fixed_length = 0;
6193 LEAVE_with_name("study_chunk");
6196 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6197 ri->regstclass = NULL;
6199 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6201 && ! TEST_SSC_EOS(data.start_class)
6202 && !cl_is_anything(data.start_class))
6204 const U32 n = add_data(pRExC_state, 1, "f");
6205 OP(data.start_class) = ANYOF_SYNTHETIC;
6207 Newx(RExC_rxi->data->data[n], 1,
6208 struct regnode_charclass_class);
6209 StructCopy(data.start_class,
6210 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6211 struct regnode_charclass_class);
6212 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6213 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6214 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6215 regprop(r, sv, (regnode*)data.start_class);
6216 PerlIO_printf(Perl_debug_log,
6217 "synthetic stclass \"%s\".\n",
6218 SvPVX_const(sv));});
6221 /* A temporary algorithm prefers floated substr to fixed one to dig more info. */
6222 if (longest_fixed_length > longest_float_length) {
6223 r->check_end_shift = r->anchored_end_shift;
6224 r->check_substr = r->anchored_substr;
6225 r->check_utf8 = r->anchored_utf8;
6226 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6227 if (r->extflags & RXf_ANCH_SINGLE)
6228 r->extflags |= RXf_NOSCAN;
6231 r->check_end_shift = r->float_end_shift;
6232 r->check_substr = r->float_substr;
6233 r->check_utf8 = r->float_utf8;
6234 r->check_offset_min = r->float_min_offset;
6235 r->check_offset_max = r->float_max_offset;
6237 /* XXXX Currently intuiting is not compatible with ANCH_GPOS.
6238 This should be changed ASAP! */
6239 if ((r->check_substr || r->check_utf8) && !(r->extflags & RXf_ANCH_GPOS)) {
6240 r->extflags |= RXf_USE_INTUIT;
6241 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6242 r->extflags |= RXf_INTUIT_TAIL;
6244 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6245 if ( (STRLEN)minlen < longest_float_length )
6246 minlen= longest_float_length;
6247 if ( (STRLEN)minlen < longest_fixed_length )
6248 minlen= longest_fixed_length;
6252 /* Several toplevels. Best we can is to set minlen. */
6254 struct regnode_charclass_class ch_class;
6257 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6259 scan = ri->program + 1;
6260 cl_init(pRExC_state, &ch_class);
6261 data.start_class = &ch_class;
6262 data.last_closep = &last_close;
6265 minlen = study_chunk(pRExC_state, &scan, &minlen, &fake, scan + RExC_size,
6266 &data, -1, NULL, NULL, SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS,0);
6268 CHECK_RESTUDY_GOTO_butfirst(NOOP);
6270 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
6271 = r->float_substr = r->float_utf8 = NULL;
6273 if (! TEST_SSC_EOS(data.start_class)
6274 && !cl_is_anything(data.start_class))
6276 const U32 n = add_data(pRExC_state, 1, "f");
6277 OP(data.start_class) = ANYOF_SYNTHETIC;
6279 Newx(RExC_rxi->data->data[n], 1,
6280 struct regnode_charclass_class);
6281 StructCopy(data.start_class,
6282 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6283 struct regnode_charclass_class);
6284 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6285 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6286 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
6287 regprop(r, sv, (regnode*)data.start_class);
6288 PerlIO_printf(Perl_debug_log,
6289 "synthetic stclass \"%s\".\n",
6290 SvPVX_const(sv));});
6294 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
6295 the "real" pattern. */
6297 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf"\n",
6298 (IV)minlen, (IV)r->minlen);
6300 r->minlenret = minlen;
6301 if (r->minlen < minlen)
6304 if (RExC_seen & REG_SEEN_GPOS)
6305 r->extflags |= RXf_GPOS_SEEN;
6306 if (RExC_seen & REG_SEEN_LOOKBEHIND)
6307 r->extflags |= RXf_LOOKBEHIND_SEEN;
6308 if (pRExC_state->num_code_blocks)
6309 r->extflags |= RXf_EVAL_SEEN;
6310 if (RExC_seen & REG_SEEN_CANY)
6311 r->extflags |= RXf_CANY_SEEN;
6312 if (RExC_seen & REG_SEEN_VERBARG)
6314 r->intflags |= PREGf_VERBARG_SEEN;
6315 r->extflags |= RXf_MODIFIES_VARS;
6317 if (RExC_seen & REG_SEEN_CUTGROUP)
6318 r->intflags |= PREGf_CUTGROUP_SEEN;
6319 if (pm_flags & PMf_USE_RE_EVAL)
6320 r->intflags |= PREGf_USE_RE_EVAL;
6321 if (RExC_paren_names)
6322 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
6324 RXp_PAREN_NAMES(r) = NULL;
6326 #ifdef STUPID_PATTERN_CHECKS
6327 if (RX_PRELEN(rx) == 0)
6328 r->extflags |= RXf_NULL;
6329 if (RX_PRELEN(rx) == 3 && memEQ("\\s+", RX_PRECOMP(rx), 3))
6330 r->extflags |= RXf_WHITE;
6331 else if (RX_PRELEN(rx) == 1 && RXp_PRECOMP(rx)[0] == '^')
6332 r->extflags |= RXf_START_ONLY;
6335 regnode *first = ri->program + 1;
6338 if (PL_regkind[fop] == NOTHING && OP(NEXTOPER(first)) == END)
6339 r->extflags |= RXf_NULL;
6340 else if (PL_regkind[fop] == BOL && OP(NEXTOPER(first)) == END)
6341 r->extflags |= RXf_START_ONLY;
6342 else if (fop == PLUS && PL_regkind[OP(NEXTOPER(first))] == POSIXD && FLAGS(NEXTOPER(first)) == _CC_SPACE
6343 && OP(regnext(first)) == END)
6344 r->extflags |= RXf_WHITE;
6348 if (RExC_paren_names) {
6349 ri->name_list_idx = add_data( pRExC_state, 1, "a" );
6350 ri->data->data[ri->name_list_idx] = (void*)SvREFCNT_inc(RExC_paren_name_list);
6353 ri->name_list_idx = 0;
6355 if (RExC_recurse_count) {
6356 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
6357 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
6358 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
6361 Newxz(r->offs, RExC_npar, regexp_paren_pair);
6362 /* assume we don't need to swap parens around before we match */
6365 PerlIO_printf(Perl_debug_log,"Final program:\n");
6368 #ifdef RE_TRACK_PATTERN_OFFSETS
6369 DEBUG_OFFSETS_r(if (ri->u.offsets) {
6370 const U32 len = ri->u.offsets[0];
6372 GET_RE_DEBUG_FLAGS_DECL;
6373 PerlIO_printf(Perl_debug_log, "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
6374 for (i = 1; i <= len; i++) {
6375 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
6376 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
6377 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
6379 PerlIO_printf(Perl_debug_log, "\n");
6384 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
6385 * by setting the regexp SV to readonly-only instead. If the
6386 * pattern's been recompiled, the USEDness should remain. */
6387 if (old_re && SvREADONLY(old_re))
6395 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
6398 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
6400 PERL_UNUSED_ARG(value);
6402 if (flags & RXapif_FETCH) {
6403 return reg_named_buff_fetch(rx, key, flags);
6404 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
6405 Perl_croak_no_modify();
6407 } else if (flags & RXapif_EXISTS) {
6408 return reg_named_buff_exists(rx, key, flags)
6411 } else if (flags & RXapif_REGNAMES) {
6412 return reg_named_buff_all(rx, flags);
6413 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
6414 return reg_named_buff_scalar(rx, flags);
6416 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
6422 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
6425 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
6426 PERL_UNUSED_ARG(lastkey);
6428 if (flags & RXapif_FIRSTKEY)
6429 return reg_named_buff_firstkey(rx, flags);
6430 else if (flags & RXapif_NEXTKEY)
6431 return reg_named_buff_nextkey(rx, flags);
6433 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter", (int)flags);
6439 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
6442 AV *retarray = NULL;
6444 struct regexp *const rx = ReANY(r);
6446 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
6448 if (flags & RXapif_ALL)
6451 if (rx && RXp_PAREN_NAMES(rx)) {
6452 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
6455 SV* sv_dat=HeVAL(he_str);
6456 I32 *nums=(I32*)SvPVX(sv_dat);
6457 for ( i=0; i<SvIVX(sv_dat); i++ ) {
6458 if ((I32)(rx->nparens) >= nums[i]
6459 && rx->offs[nums[i]].start != -1
6460 && rx->offs[nums[i]].end != -1)
6463 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
6468 ret = newSVsv(&PL_sv_undef);
6471 av_push(retarray, ret);
6474 return newRV_noinc(MUTABLE_SV(retarray));
6481 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
6484 struct regexp *const rx = ReANY(r);
6486 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
6488 if (rx && RXp_PAREN_NAMES(rx)) {
6489 if (flags & RXapif_ALL) {
6490 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
6492 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
6494 SvREFCNT_dec_NN(sv);
6506 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
6508 struct regexp *const rx = ReANY(r);
6510 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
6512 if ( rx && RXp_PAREN_NAMES(rx) ) {
6513 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
6515 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
6522 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
6524 struct regexp *const rx = ReANY(r);
6525 GET_RE_DEBUG_FLAGS_DECL;
6527 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
6529 if (rx && RXp_PAREN_NAMES(rx)) {
6530 HV *hv = RXp_PAREN_NAMES(rx);
6532 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6535 SV* sv_dat = HeVAL(temphe);
6536 I32 *nums = (I32*)SvPVX(sv_dat);
6537 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6538 if ((I32)(rx->lastparen) >= nums[i] &&
6539 rx->offs[nums[i]].start != -1 &&
6540 rx->offs[nums[i]].end != -1)
6546 if (parno || flags & RXapif_ALL) {
6547 return newSVhek(HeKEY_hek(temphe));
6555 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
6560 struct regexp *const rx = ReANY(r);
6562 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
6564 if (rx && RXp_PAREN_NAMES(rx)) {
6565 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
6566 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
6567 } else if (flags & RXapif_ONE) {
6568 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
6569 av = MUTABLE_AV(SvRV(ret));
6570 length = av_len(av);
6571 SvREFCNT_dec_NN(ret);
6572 return newSViv(length + 1);
6574 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar", (int)flags);
6578 return &PL_sv_undef;
6582 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
6584 struct regexp *const rx = ReANY(r);
6587 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
6589 if (rx && RXp_PAREN_NAMES(rx)) {
6590 HV *hv= RXp_PAREN_NAMES(rx);
6592 (void)hv_iterinit(hv);
6593 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6596 SV* sv_dat = HeVAL(temphe);
6597 I32 *nums = (I32*)SvPVX(sv_dat);
6598 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6599 if ((I32)(rx->lastparen) >= nums[i] &&
6600 rx->offs[nums[i]].start != -1 &&
6601 rx->offs[nums[i]].end != -1)
6607 if (parno || flags & RXapif_ALL) {
6608 av_push(av, newSVhek(HeKEY_hek(temphe)));
6613 return newRV_noinc(MUTABLE_SV(av));
6617 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
6620 struct regexp *const rx = ReANY(r);
6626 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
6628 if ( ( n == RX_BUFF_IDX_CARET_PREMATCH
6629 || n == RX_BUFF_IDX_CARET_FULLMATCH
6630 || n == RX_BUFF_IDX_CARET_POSTMATCH
6632 && !(rx->extflags & RXf_PMf_KEEPCOPY)
6639 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
6640 /* no need to distinguish between them any more */
6641 n = RX_BUFF_IDX_FULLMATCH;
6643 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
6644 && rx->offs[0].start != -1)
6646 /* $`, ${^PREMATCH} */
6647 i = rx->offs[0].start;
6651 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
6652 && rx->offs[0].end != -1)
6654 /* $', ${^POSTMATCH} */
6655 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
6656 i = rx->sublen + rx->suboffset - rx->offs[0].end;
6659 if ( 0 <= n && n <= (I32)rx->nparens &&
6660 (s1 = rx->offs[n].start) != -1 &&
6661 (t1 = rx->offs[n].end) != -1)
6663 /* $&, ${^MATCH}, $1 ... */
6665 s = rx->subbeg + s1 - rx->suboffset;
6670 assert(s >= rx->subbeg);
6671 assert(rx->sublen >= (s - rx->subbeg) + i );
6673 #if NO_TAINT_SUPPORT
6674 sv_setpvn(sv, s, i);
6676 const int oldtainted = TAINT_get;
6678 sv_setpvn(sv, s, i);
6679 TAINT_set(oldtainted);
6681 if ( (rx->extflags & RXf_CANY_SEEN)
6682 ? (RXp_MATCH_UTF8(rx)
6683 && (!i || is_utf8_string((U8*)s, i)))
6684 : (RXp_MATCH_UTF8(rx)) )
6691 if (RXp_MATCH_TAINTED(rx)) {
6692 if (SvTYPE(sv) >= SVt_PVMG) {
6693 MAGIC* const mg = SvMAGIC(sv);
6696 SvMAGIC_set(sv, mg->mg_moremagic);
6698 if ((mgt = SvMAGIC(sv))) {
6699 mg->mg_moremagic = mgt;
6700 SvMAGIC_set(sv, mg);
6711 sv_setsv(sv,&PL_sv_undef);
6717 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
6718 SV const * const value)
6720 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
6722 PERL_UNUSED_ARG(rx);
6723 PERL_UNUSED_ARG(paren);
6724 PERL_UNUSED_ARG(value);
6727 Perl_croak_no_modify();
6731 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
6734 struct regexp *const rx = ReANY(r);
6738 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
6740 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
6742 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
6743 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6747 case RX_BUFF_IDX_PREMATCH: /* $` */
6748 if (rx->offs[0].start != -1) {
6749 i = rx->offs[0].start;
6758 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
6759 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6761 case RX_BUFF_IDX_POSTMATCH: /* $' */
6762 if (rx->offs[0].end != -1) {
6763 i = rx->sublen - rx->offs[0].end;
6765 s1 = rx->offs[0].end;
6772 case RX_BUFF_IDX_CARET_FULLMATCH: /* ${^MATCH} */
6773 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6777 /* $& / ${^MATCH}, $1, $2, ... */
6779 if (paren <= (I32)rx->nparens &&
6780 (s1 = rx->offs[paren].start) != -1 &&
6781 (t1 = rx->offs[paren].end) != -1)
6787 if (ckWARN(WARN_UNINITIALIZED))
6788 report_uninit((const SV *)sv);
6793 if (i > 0 && RXp_MATCH_UTF8(rx)) {
6794 const char * const s = rx->subbeg - rx->suboffset + s1;
6799 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
6806 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
6808 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
6809 PERL_UNUSED_ARG(rx);
6813 return newSVpvs("Regexp");
6816 /* Scans the name of a named buffer from the pattern.
6817 * If flags is REG_RSN_RETURN_NULL returns null.
6818 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
6819 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
6820 * to the parsed name as looked up in the RExC_paren_names hash.
6821 * If there is an error throws a vFAIL().. type exception.
6824 #define REG_RSN_RETURN_NULL 0
6825 #define REG_RSN_RETURN_NAME 1
6826 #define REG_RSN_RETURN_DATA 2
6829 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
6831 char *name_start = RExC_parse;
6833 PERL_ARGS_ASSERT_REG_SCAN_NAME;
6835 if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
6836 /* skip IDFIRST by using do...while */
6839 RExC_parse += UTF8SKIP(RExC_parse);
6840 } while (isWORDCHAR_utf8((U8*)RExC_parse));
6844 } while (isWORDCHAR(*RExC_parse));
6846 RExC_parse++; /* so the <- from the vFAIL is after the offending character */
6847 vFAIL("Group name must start with a non-digit word character");
6851 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
6852 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
6853 if ( flags == REG_RSN_RETURN_NAME)
6855 else if (flags==REG_RSN_RETURN_DATA) {
6858 if ( ! sv_name ) /* should not happen*/
6859 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
6860 if (RExC_paren_names)
6861 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
6863 sv_dat = HeVAL(he_str);
6865 vFAIL("Reference to nonexistent named group");
6869 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
6870 (unsigned long) flags);
6872 assert(0); /* NOT REACHED */
6877 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
6878 int rem=(int)(RExC_end - RExC_parse); \
6887 if (RExC_lastparse!=RExC_parse) \
6888 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
6891 iscut ? "..." : "<" \
6894 PerlIO_printf(Perl_debug_log,"%16s",""); \
6897 num = RExC_size + 1; \
6899 num=REG_NODE_NUM(RExC_emit); \
6900 if (RExC_lastnum!=num) \
6901 PerlIO_printf(Perl_debug_log,"|%4d",num); \
6903 PerlIO_printf(Perl_debug_log,"|%4s",""); \
6904 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
6905 (int)((depth*2)), "", \
6909 RExC_lastparse=RExC_parse; \
6914 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
6915 DEBUG_PARSE_MSG((funcname)); \
6916 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
6918 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
6919 DEBUG_PARSE_MSG((funcname)); \
6920 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
6923 /* This section of code defines the inversion list object and its methods. The
6924 * interfaces are highly subject to change, so as much as possible is static to
6925 * this file. An inversion list is here implemented as a malloc'd C UV array
6926 * with some added info that is placed as UVs at the beginning in a header
6927 * portion. An inversion list for Unicode is an array of code points, sorted
6928 * by ordinal number. The zeroth element is the first code point in the list.
6929 * The 1th element is the first element beyond that not in the list. In other
6930 * words, the first range is
6931 * invlist[0]..(invlist[1]-1)
6932 * The other ranges follow. Thus every element whose index is divisible by two
6933 * marks the beginning of a range that is in the list, and every element not
6934 * divisible by two marks the beginning of a range not in the list. A single
6935 * element inversion list that contains the single code point N generally
6936 * consists of two elements
6939 * (The exception is when N is the highest representable value on the
6940 * machine, in which case the list containing just it would be a single
6941 * element, itself. By extension, if the last range in the list extends to
6942 * infinity, then the first element of that range will be in the inversion list
6943 * at a position that is divisible by two, and is the final element in the
6945 * Taking the complement (inverting) an inversion list is quite simple, if the
6946 * first element is 0, remove it; otherwise add a 0 element at the beginning.
6947 * This implementation reserves an element at the beginning of each inversion
6948 * list to contain 0 when the list contains 0, and contains 1 otherwise. The
6949 * actual beginning of the list is either that element if 0, or the next one if
6952 * More about inversion lists can be found in "Unicode Demystified"
6953 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
6954 * More will be coming when functionality is added later.
6956 * The inversion list data structure is currently implemented as an SV pointing
6957 * to an array of UVs that the SV thinks are bytes. This allows us to have an
6958 * array of UV whose memory management is automatically handled by the existing
6959 * facilities for SV's.
6961 * Some of the methods should always be private to the implementation, and some
6962 * should eventually be made public */
6964 /* The header definitions are in F<inline_invlist.c> */
6965 #define TO_INTERNAL_SIZE(x) (((x) + HEADER_LENGTH) * sizeof(UV))
6966 #define FROM_INTERNAL_SIZE(x) (((x)/ sizeof(UV)) - HEADER_LENGTH)
6968 #define INVLIST_INITIAL_LEN 10
6970 PERL_STATIC_INLINE UV*
6971 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
6973 /* Returns a pointer to the first element in the inversion list's array.
6974 * This is called upon initialization of an inversion list. Where the
6975 * array begins depends on whether the list has the code point U+0000
6976 * in it or not. The other parameter tells it whether the code that
6977 * follows this call is about to put a 0 in the inversion list or not.
6978 * The first element is either the element with 0, if 0, or the next one,
6981 UV* zero = get_invlist_zero_addr(invlist);
6983 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
6986 assert(! *_get_invlist_len_addr(invlist));
6988 /* 1^1 = 0; 1^0 = 1 */
6989 *zero = 1 ^ will_have_0;
6990 return zero + *zero;
6993 PERL_STATIC_INLINE UV*
6994 S_invlist_array(pTHX_ SV* const invlist)
6996 /* Returns the pointer to the inversion list's array. Every time the
6997 * length changes, this needs to be called in case malloc or realloc moved
7000 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7002 /* Must not be empty. If these fail, you probably didn't check for <len>
7003 * being non-zero before trying to get the array */
7004 assert(*_get_invlist_len_addr(invlist));
7005 assert(*get_invlist_zero_addr(invlist) == 0
7006 || *get_invlist_zero_addr(invlist) == 1);
7008 /* The array begins either at the element reserved for zero if the
7009 * list contains 0 (that element will be set to 0), or otherwise the next
7010 * element (in which case the reserved element will be set to 1). */
7011 return (UV *) (get_invlist_zero_addr(invlist)
7012 + *get_invlist_zero_addr(invlist));
7015 PERL_STATIC_INLINE void
7016 S_invlist_set_len(pTHX_ SV* const invlist, const UV len)
7018 /* Sets the current number of elements stored in the inversion list */
7020 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7022 *_get_invlist_len_addr(invlist) = len;
7024 assert(len <= SvLEN(invlist));
7026 SvCUR_set(invlist, TO_INTERNAL_SIZE(len));
7027 /* If the list contains U+0000, that element is part of the header,
7028 * and should not be counted as part of the array. It will contain
7029 * 0 in that case, and 1 otherwise. So we could flop 0=>1, 1=>0 and
7031 * SvCUR_set(invlist,
7032 * TO_INTERNAL_SIZE(len
7033 * - (*get_invlist_zero_addr(inv_list) ^ 1)));
7034 * But, this is only valid if len is not 0. The consequences of not doing
7035 * this is that the memory allocation code may think that 1 more UV is
7036 * being used than actually is, and so might do an unnecessary grow. That
7037 * seems worth not bothering to make this the precise amount.
7039 * Note that when inverting, SvCUR shouldn't change */
7042 PERL_STATIC_INLINE IV*
7043 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7045 /* Return the address of the UV that is reserved to hold the cached index
7048 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7050 return (IV *) (SvPVX(invlist) + (INVLIST_PREVIOUS_INDEX_OFFSET * sizeof (UV)));
7053 PERL_STATIC_INLINE IV
7054 S_invlist_previous_index(pTHX_ SV* const invlist)
7056 /* Returns cached index of previous search */
7058 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7060 return *get_invlist_previous_index_addr(invlist);
7063 PERL_STATIC_INLINE void
7064 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7066 /* Caches <index> for later retrieval */
7068 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7070 assert(index == 0 || index < (int) _invlist_len(invlist));
7072 *get_invlist_previous_index_addr(invlist) = index;
7075 PERL_STATIC_INLINE UV
7076 S_invlist_max(pTHX_ SV* const invlist)
7078 /* Returns the maximum number of elements storable in the inversion list's
7079 * array, without having to realloc() */
7081 PERL_ARGS_ASSERT_INVLIST_MAX;
7083 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7084 ? _invlist_len(invlist)
7085 : FROM_INTERNAL_SIZE(SvLEN(invlist));
7088 PERL_STATIC_INLINE UV*
7089 S_get_invlist_zero_addr(pTHX_ SV* invlist)
7091 /* Return the address of the UV that is reserved to hold 0 if the inversion
7092 * list contains 0. This has to be the last element of the heading, as the
7093 * list proper starts with either it if 0, or the next element if not.
7094 * (But we force it to contain either 0 or 1) */
7096 PERL_ARGS_ASSERT_GET_INVLIST_ZERO_ADDR;
7098 return (UV *) (SvPVX(invlist) + (INVLIST_ZERO_OFFSET * sizeof (UV)));
7101 #ifndef PERL_IN_XSUB_RE
7103 Perl__new_invlist(pTHX_ IV initial_size)
7106 /* Return a pointer to a newly constructed inversion list, with enough
7107 * space to store 'initial_size' elements. If that number is negative, a
7108 * system default is used instead */
7112 if (initial_size < 0) {
7113 initial_size = INVLIST_INITIAL_LEN;
7116 /* Allocate the initial space */
7117 new_list = newSV(TO_INTERNAL_SIZE(initial_size));
7118 invlist_set_len(new_list, 0);
7120 /* Force iterinit() to be used to get iteration to work */
7121 *get_invlist_iter_addr(new_list) = UV_MAX;
7123 /* This should force a segfault if a method doesn't initialize this
7125 *get_invlist_zero_addr(new_list) = UV_MAX;
7127 *get_invlist_previous_index_addr(new_list) = 0;
7128 *get_invlist_version_id_addr(new_list) = INVLIST_VERSION_ID;
7129 #if HEADER_LENGTH != 5
7130 # error Need to regenerate INVLIST_VERSION_ID by running perl -E 'say int(rand 2**31-1)', and then changing the #if to the new length
7138 S__new_invlist_C_array(pTHX_ UV* list)
7140 /* Return a pointer to a newly constructed inversion list, initialized to
7141 * point to <list>, which has to be in the exact correct inversion list
7142 * form, including internal fields. Thus this is a dangerous routine that
7143 * should not be used in the wrong hands */
7145 SV* invlist = newSV_type(SVt_PV);
7147 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7149 SvPV_set(invlist, (char *) list);
7150 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7151 shouldn't touch it */
7152 SvCUR_set(invlist, TO_INTERNAL_SIZE(_invlist_len(invlist)));
7154 if (*get_invlist_version_id_addr(invlist) != INVLIST_VERSION_ID) {
7155 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7158 /* Initialize the iteration pointer.
7159 * XXX This could be done at compile time in charclass_invlists.h, but I
7160 * (khw) am not confident that the suffixes for specifying the C constant
7161 * UV_MAX are portable, e.g. 'ull' on a 32 bit machine that is configured
7162 * to use 64 bits; might need a Configure probe */
7163 invlist_iterfinish(invlist);
7169 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7171 /* Grow the maximum size of an inversion list */
7173 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7175 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max));
7178 PERL_STATIC_INLINE void
7179 S_invlist_trim(pTHX_ SV* const invlist)
7181 PERL_ARGS_ASSERT_INVLIST_TRIM;
7183 /* Change the length of the inversion list to how many entries it currently
7186 SvPV_shrink_to_cur((SV *) invlist);
7189 #define _invlist_union_complement_2nd(a, b, output) _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
7192 S__append_range_to_invlist(pTHX_ SV* const invlist, const UV start, const UV end)
7194 /* Subject to change or removal. Append the range from 'start' to 'end' at
7195 * the end of the inversion list. The range must be above any existing
7199 UV max = invlist_max(invlist);
7200 UV len = _invlist_len(invlist);
7202 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7204 if (len == 0) { /* Empty lists must be initialized */
7205 array = _invlist_array_init(invlist, start == 0);
7208 /* Here, the existing list is non-empty. The current max entry in the
7209 * list is generally the first value not in the set, except when the
7210 * set extends to the end of permissible values, in which case it is
7211 * the first entry in that final set, and so this call is an attempt to
7212 * append out-of-order */
7214 UV final_element = len - 1;
7215 array = invlist_array(invlist);
7216 if (array[final_element] > start
7217 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
7219 Perl_croak(aTHX_ "panic: attempting to append to an inversion list, but wasn't at the end of the list, final=%"UVuf", start=%"UVuf", match=%c",
7220 array[final_element], start,
7221 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
7224 /* Here, it is a legal append. If the new range begins with the first
7225 * value not in the set, it is extending the set, so the new first
7226 * value not in the set is one greater than the newly extended range.
7228 if (array[final_element] == start) {
7229 if (end != UV_MAX) {
7230 array[final_element] = end + 1;
7233 /* But if the end is the maximum representable on the machine,
7234 * just let the range that this would extend to have no end */
7235 invlist_set_len(invlist, len - 1);
7241 /* Here the new range doesn't extend any existing set. Add it */
7243 len += 2; /* Includes an element each for the start and end of range */
7245 /* If overflows the existing space, extend, which may cause the array to be
7248 invlist_extend(invlist, len);
7249 invlist_set_len(invlist, len); /* Have to set len here to avoid assert
7250 failure in invlist_array() */
7251 array = invlist_array(invlist);
7254 invlist_set_len(invlist, len);
7257 /* The next item on the list starts the range, the one after that is
7258 * one past the new range. */
7259 array[len - 2] = start;
7260 if (end != UV_MAX) {
7261 array[len - 1] = end + 1;
7264 /* But if the end is the maximum representable on the machine, just let
7265 * the range have no end */
7266 invlist_set_len(invlist, len - 1);
7270 #ifndef PERL_IN_XSUB_RE
7273 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
7275 /* Searches the inversion list for the entry that contains the input code
7276 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
7277 * return value is the index into the list's array of the range that
7282 IV high = _invlist_len(invlist);
7283 const IV highest_element = high - 1;
7286 PERL_ARGS_ASSERT__INVLIST_SEARCH;
7288 /* If list is empty, return failure. */
7293 /* (We can't get the array unless we know the list is non-empty) */
7294 array = invlist_array(invlist);
7296 mid = invlist_previous_index(invlist);
7297 assert(mid >=0 && mid <= highest_element);
7299 /* <mid> contains the cache of the result of the previous call to this
7300 * function (0 the first time). See if this call is for the same result,
7301 * or if it is for mid-1. This is under the theory that calls to this
7302 * function will often be for related code points that are near each other.
7303 * And benchmarks show that caching gives better results. We also test
7304 * here if the code point is within the bounds of the list. These tests
7305 * replace others that would have had to be made anyway to make sure that
7306 * the array bounds were not exceeded, and these give us extra information
7307 * at the same time */
7308 if (cp >= array[mid]) {
7309 if (cp >= array[highest_element]) {
7310 return highest_element;
7313 /* Here, array[mid] <= cp < array[highest_element]. This means that
7314 * the final element is not the answer, so can exclude it; it also
7315 * means that <mid> is not the final element, so can refer to 'mid + 1'
7317 if (cp < array[mid + 1]) {
7323 else { /* cp < aray[mid] */
7324 if (cp < array[0]) { /* Fail if outside the array */
7328 if (cp >= array[mid - 1]) {
7333 /* Binary search. What we are looking for is <i> such that
7334 * array[i] <= cp < array[i+1]
7335 * The loop below converges on the i+1. Note that there may not be an
7336 * (i+1)th element in the array, and things work nonetheless */
7337 while (low < high) {
7338 mid = (low + high) / 2;
7339 assert(mid <= highest_element);
7340 if (array[mid] <= cp) { /* cp >= array[mid] */
7343 /* We could do this extra test to exit the loop early.
7344 if (cp < array[low]) {
7349 else { /* cp < array[mid] */
7356 invlist_set_previous_index(invlist, high);
7361 Perl__invlist_populate_swatch(pTHX_ SV* const invlist, const UV start, const UV end, U8* swatch)
7363 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
7364 * but is used when the swash has an inversion list. This makes this much
7365 * faster, as it uses a binary search instead of a linear one. This is
7366 * intimately tied to that function, and perhaps should be in utf8.c,
7367 * except it is intimately tied to inversion lists as well. It assumes
7368 * that <swatch> is all 0's on input */
7371 const IV len = _invlist_len(invlist);
7375 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
7377 if (len == 0) { /* Empty inversion list */
7381 array = invlist_array(invlist);
7383 /* Find which element it is */
7384 i = _invlist_search(invlist, start);
7386 /* We populate from <start> to <end> */
7387 while (current < end) {
7390 /* The inversion list gives the results for every possible code point
7391 * after the first one in the list. Only those ranges whose index is
7392 * even are ones that the inversion list matches. For the odd ones,
7393 * and if the initial code point is not in the list, we have to skip
7394 * forward to the next element */
7395 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
7397 if (i >= len) { /* Finished if beyond the end of the array */
7401 if (current >= end) { /* Finished if beyond the end of what we
7403 if (LIKELY(end < UV_MAX)) {
7407 /* We get here when the upper bound is the maximum
7408 * representable on the machine, and we are looking for just
7409 * that code point. Have to special case it */
7411 goto join_end_of_list;
7414 assert(current >= start);
7416 /* The current range ends one below the next one, except don't go past
7419 upper = (i < len && array[i] < end) ? array[i] : end;
7421 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
7422 * for each code point in it */
7423 for (; current < upper; current++) {
7424 const STRLEN offset = (STRLEN)(current - start);
7425 swatch[offset >> 3] |= 1 << (offset & 7);
7430 /* Quit if at the end of the list */
7433 /* But first, have to deal with the highest possible code point on
7434 * the platform. The previous code assumes that <end> is one
7435 * beyond where we want to populate, but that is impossible at the
7436 * platform's infinity, so have to handle it specially */
7437 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
7439 const STRLEN offset = (STRLEN)(end - start);
7440 swatch[offset >> 3] |= 1 << (offset & 7);
7445 /* Advance to the next range, which will be for code points not in the
7454 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** output)
7456 /* Take the union of two inversion lists and point <output> to it. *output
7457 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7458 * the reference count to that list will be decremented. The first list,
7459 * <a>, may be NULL, in which case a copy of the second list is returned.
7460 * If <complement_b> is TRUE, the union is taken of the complement
7461 * (inversion) of <b> instead of b itself.
7463 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7464 * Richard Gillam, published by Addison-Wesley, and explained at some
7465 * length there. The preface says to incorporate its examples into your
7466 * code at your own risk.
7468 * The algorithm is like a merge sort.
7470 * XXX A potential performance improvement is to keep track as we go along
7471 * if only one of the inputs contributes to the result, meaning the other
7472 * is a subset of that one. In that case, we can skip the final copy and
7473 * return the larger of the input lists, but then outside code might need
7474 * to keep track of whether to free the input list or not */
7476 UV* array_a; /* a's array */
7478 UV len_a; /* length of a's array */
7481 SV* u; /* the resulting union */
7485 UV i_a = 0; /* current index into a's array */
7489 /* running count, as explained in the algorithm source book; items are
7490 * stopped accumulating and are output when the count changes to/from 0.
7491 * The count is incremented when we start a range that's in the set, and
7492 * decremented when we start a range that's not in the set. So its range
7493 * is 0 to 2. Only when the count is zero is something not in the set.
7497 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
7500 /* If either one is empty, the union is the other one */
7501 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
7508 *output = invlist_clone(b);
7510 _invlist_invert(*output);
7512 } /* else *output already = b; */
7515 else if ((len_b = _invlist_len(b)) == 0) {
7520 /* The complement of an empty list is a list that has everything in it,
7521 * so the union with <a> includes everything too */
7526 *output = _new_invlist(1);
7527 _append_range_to_invlist(*output, 0, UV_MAX);
7529 else if (*output != a) {
7530 *output = invlist_clone(a);
7532 /* else *output already = a; */
7536 /* Here both lists exist and are non-empty */
7537 array_a = invlist_array(a);
7538 array_b = invlist_array(b);
7540 /* If are to take the union of 'a' with the complement of b, set it
7541 * up so are looking at b's complement. */
7544 /* To complement, we invert: if the first element is 0, remove it. To
7545 * do this, we just pretend the array starts one later, and clear the
7546 * flag as we don't have to do anything else later */
7547 if (array_b[0] == 0) {
7550 complement_b = FALSE;
7554 /* But if the first element is not zero, we unshift a 0 before the
7555 * array. The data structure reserves a space for that 0 (which
7556 * should be a '1' right now), so physical shifting is unneeded,
7557 * but temporarily change that element to 0. Before exiting the
7558 * routine, we must restore the element to '1' */
7565 /* Size the union for the worst case: that the sets are completely
7567 u = _new_invlist(len_a + len_b);
7569 /* Will contain U+0000 if either component does */
7570 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
7571 || (len_b > 0 && array_b[0] == 0));
7573 /* Go through each list item by item, stopping when exhausted one of
7575 while (i_a < len_a && i_b < len_b) {
7576 UV cp; /* The element to potentially add to the union's array */
7577 bool cp_in_set; /* is it in the the input list's set or not */
7579 /* We need to take one or the other of the two inputs for the union.
7580 * Since we are merging two sorted lists, we take the smaller of the
7581 * next items. In case of a tie, we take the one that is in its set
7582 * first. If we took one not in the set first, it would decrement the
7583 * count, possibly to 0 which would cause it to be output as ending the
7584 * range, and the next time through we would take the same number, and
7585 * output it again as beginning the next range. By doing it the
7586 * opposite way, there is no possibility that the count will be
7587 * momentarily decremented to 0, and thus the two adjoining ranges will
7588 * be seamlessly merged. (In a tie and both are in the set or both not
7589 * in the set, it doesn't matter which we take first.) */
7590 if (array_a[i_a] < array_b[i_b]
7591 || (array_a[i_a] == array_b[i_b]
7592 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7594 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7598 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7599 cp = array_b[i_b++];
7602 /* Here, have chosen which of the two inputs to look at. Only output
7603 * if the running count changes to/from 0, which marks the
7604 * beginning/end of a range in that's in the set */
7607 array_u[i_u++] = cp;
7614 array_u[i_u++] = cp;
7619 /* Here, we are finished going through at least one of the lists, which
7620 * means there is something remaining in at most one. We check if the list
7621 * that hasn't been exhausted is positioned such that we are in the middle
7622 * of a range in its set or not. (i_a and i_b point to the element beyond
7623 * the one we care about.) If in the set, we decrement 'count'; if 0, there
7624 * is potentially more to output.
7625 * There are four cases:
7626 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
7627 * in the union is entirely from the non-exhausted set.
7628 * 2) Both were in their sets, count is 2. Nothing further should
7629 * be output, as everything that remains will be in the exhausted
7630 * list's set, hence in the union; decrementing to 1 but not 0 insures
7632 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
7633 * Nothing further should be output because the union includes
7634 * everything from the exhausted set. Not decrementing ensures that.
7635 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
7636 * decrementing to 0 insures that we look at the remainder of the
7637 * non-exhausted set */
7638 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7639 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7644 /* The final length is what we've output so far, plus what else is about to
7645 * be output. (If 'count' is non-zero, then the input list we exhausted
7646 * has everything remaining up to the machine's limit in its set, and hence
7647 * in the union, so there will be no further output. */
7650 /* At most one of the subexpressions will be non-zero */
7651 len_u += (len_a - i_a) + (len_b - i_b);
7654 /* Set result to final length, which can change the pointer to array_u, so
7656 if (len_u != _invlist_len(u)) {
7657 invlist_set_len(u, len_u);
7659 array_u = invlist_array(u);
7662 /* When 'count' is 0, the list that was exhausted (if one was shorter than
7663 * the other) ended with everything above it not in its set. That means
7664 * that the remaining part of the union is precisely the same as the
7665 * non-exhausted list, so can just copy it unchanged. (If both list were
7666 * exhausted at the same time, then the operations below will be both 0.)
7669 IV copy_count; /* At most one will have a non-zero copy count */
7670 if ((copy_count = len_a - i_a) > 0) {
7671 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
7673 else if ((copy_count = len_b - i_b) > 0) {
7674 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
7678 /* If we've changed b, restore it */
7683 /* We may be removing a reference to one of the inputs */
7684 if (a == *output || b == *output) {
7685 assert(! invlist_is_iterating(*output));
7686 SvREFCNT_dec_NN(*output);
7694 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** i)
7696 /* Take the intersection of two inversion lists and point <i> to it. *i
7697 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7698 * the reference count to that list will be decremented.
7699 * If <complement_b> is TRUE, the result will be the intersection of <a>
7700 * and the complement (or inversion) of <b> instead of <b> directly.
7702 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7703 * Richard Gillam, published by Addison-Wesley, and explained at some
7704 * length there. The preface says to incorporate its examples into your
7705 * code at your own risk. In fact, it had bugs
7707 * The algorithm is like a merge sort, and is essentially the same as the
7711 UV* array_a; /* a's array */
7713 UV len_a; /* length of a's array */
7716 SV* r; /* the resulting intersection */
7720 UV i_a = 0; /* current index into a's array */
7724 /* running count, as explained in the algorithm source book; items are
7725 * stopped accumulating and are output when the count changes to/from 2.
7726 * The count is incremented when we start a range that's in the set, and
7727 * decremented when we start a range that's not in the set. So its range
7728 * is 0 to 2. Only when the count is 2 is something in the intersection.
7732 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
7735 /* Special case if either one is empty */
7736 len_a = _invlist_len(a);
7737 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
7739 if (len_a != 0 && complement_b) {
7741 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
7742 * be empty. Here, also we are using 'b's complement, which hence
7743 * must be every possible code point. Thus the intersection is
7746 *i = invlist_clone(a);
7752 /* else *i is already 'a' */
7756 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
7757 * intersection must be empty */
7764 *i = _new_invlist(0);
7768 /* Here both lists exist and are non-empty */
7769 array_a = invlist_array(a);
7770 array_b = invlist_array(b);
7772 /* If are to take the intersection of 'a' with the complement of b, set it
7773 * up so are looking at b's complement. */
7776 /* To complement, we invert: if the first element is 0, remove it. To
7777 * do this, we just pretend the array starts one later, and clear the
7778 * flag as we don't have to do anything else later */
7779 if (array_b[0] == 0) {
7782 complement_b = FALSE;
7786 /* But if the first element is not zero, we unshift a 0 before the
7787 * array. The data structure reserves a space for that 0 (which
7788 * should be a '1' right now), so physical shifting is unneeded,
7789 * but temporarily change that element to 0. Before exiting the
7790 * routine, we must restore the element to '1' */
7797 /* Size the intersection for the worst case: that the intersection ends up
7798 * fragmenting everything to be completely disjoint */
7799 r= _new_invlist(len_a + len_b);
7801 /* Will contain U+0000 iff both components do */
7802 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
7803 && len_b > 0 && array_b[0] == 0);
7805 /* Go through each list item by item, stopping when exhausted one of
7807 while (i_a < len_a && i_b < len_b) {
7808 UV cp; /* The element to potentially add to the intersection's
7810 bool cp_in_set; /* Is it in the input list's set or not */
7812 /* We need to take one or the other of the two inputs for the
7813 * intersection. Since we are merging two sorted lists, we take the
7814 * smaller of the next items. In case of a tie, we take the one that
7815 * is not in its set first (a difference from the union algorithm). If
7816 * we took one in the set first, it would increment the count, possibly
7817 * to 2 which would cause it to be output as starting a range in the
7818 * intersection, and the next time through we would take that same
7819 * number, and output it again as ending the set. By doing it the
7820 * opposite of this, there is no possibility that the count will be
7821 * momentarily incremented to 2. (In a tie and both are in the set or
7822 * both not in the set, it doesn't matter which we take first.) */
7823 if (array_a[i_a] < array_b[i_b]
7824 || (array_a[i_a] == array_b[i_b]
7825 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7827 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7831 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7835 /* Here, have chosen which of the two inputs to look at. Only output
7836 * if the running count changes to/from 2, which marks the
7837 * beginning/end of a range that's in the intersection */
7841 array_r[i_r++] = cp;
7846 array_r[i_r++] = cp;
7852 /* Here, we are finished going through at least one of the lists, which
7853 * means there is something remaining in at most one. We check if the list
7854 * that has been exhausted is positioned such that we are in the middle
7855 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
7856 * the ones we care about.) There are four cases:
7857 * 1) Both weren't in their sets, count is 0, and remains 0. There's
7858 * nothing left in the intersection.
7859 * 2) Both were in their sets, count is 2 and perhaps is incremented to
7860 * above 2. What should be output is exactly that which is in the
7861 * non-exhausted set, as everything it has is also in the intersection
7862 * set, and everything it doesn't have can't be in the intersection
7863 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
7864 * gets incremented to 2. Like the previous case, the intersection is
7865 * everything that remains in the non-exhausted set.
7866 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
7867 * remains 1. And the intersection has nothing more. */
7868 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7869 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7874 /* The final length is what we've output so far plus what else is in the
7875 * intersection. At most one of the subexpressions below will be non-zero */
7878 len_r += (len_a - i_a) + (len_b - i_b);
7881 /* Set result to final length, which can change the pointer to array_r, so
7883 if (len_r != _invlist_len(r)) {
7884 invlist_set_len(r, len_r);
7886 array_r = invlist_array(r);
7889 /* Finish outputting any remaining */
7890 if (count >= 2) { /* At most one will have a non-zero copy count */
7892 if ((copy_count = len_a - i_a) > 0) {
7893 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
7895 else if ((copy_count = len_b - i_b) > 0) {
7896 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
7900 /* If we've changed b, restore it */
7905 /* We may be removing a reference to one of the inputs */
7906 if (a == *i || b == *i) {
7907 assert(! invlist_is_iterating(*i));
7908 SvREFCNT_dec_NN(*i);
7916 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
7918 /* Add the range from 'start' to 'end' inclusive to the inversion list's
7919 * set. A pointer to the inversion list is returned. This may actually be
7920 * a new list, in which case the passed in one has been destroyed. The
7921 * passed in inversion list can be NULL, in which case a new one is created
7922 * with just the one range in it */
7927 if (invlist == NULL) {
7928 invlist = _new_invlist(2);
7932 len = _invlist_len(invlist);
7935 /* If comes after the final entry actually in the list, can just append it
7938 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
7939 && start >= invlist_array(invlist)[len - 1]))
7941 _append_range_to_invlist(invlist, start, end);
7945 /* Here, can't just append things, create and return a new inversion list
7946 * which is the union of this range and the existing inversion list */
7947 range_invlist = _new_invlist(2);
7948 _append_range_to_invlist(range_invlist, start, end);
7950 _invlist_union(invlist, range_invlist, &invlist);
7952 /* The temporary can be freed */
7953 SvREFCNT_dec_NN(range_invlist);
7960 PERL_STATIC_INLINE SV*
7961 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
7962 return _add_range_to_invlist(invlist, cp, cp);
7965 #ifndef PERL_IN_XSUB_RE
7967 Perl__invlist_invert(pTHX_ SV* const invlist)
7969 /* Complement the input inversion list. This adds a 0 if the list didn't
7970 * have a zero; removes it otherwise. As described above, the data
7971 * structure is set up so that this is very efficient */
7973 UV* len_pos = _get_invlist_len_addr(invlist);
7975 PERL_ARGS_ASSERT__INVLIST_INVERT;
7977 assert(! invlist_is_iterating(invlist));
7979 /* The inverse of matching nothing is matching everything */
7980 if (*len_pos == 0) {
7981 _append_range_to_invlist(invlist, 0, UV_MAX);
7985 /* The exclusive or complents 0 to 1; and 1 to 0. If the result is 1, the
7986 * zero element was a 0, so it is being removed, so the length decrements
7987 * by 1; and vice-versa. SvCUR is unaffected */
7988 if (*get_invlist_zero_addr(invlist) ^= 1) {
7997 Perl__invlist_invert_prop(pTHX_ SV* const invlist)
7999 /* Complement the input inversion list (which must be a Unicode property,
8000 * all of which don't match above the Unicode maximum code point.) And
8001 * Perl has chosen to not have the inversion match above that either. This
8002 * adds a 0x110000 if the list didn't end with it, and removes it if it did
8008 PERL_ARGS_ASSERT__INVLIST_INVERT_PROP;
8010 _invlist_invert(invlist);
8012 len = _invlist_len(invlist);
8014 if (len != 0) { /* If empty do nothing */
8015 array = invlist_array(invlist);
8016 if (array[len - 1] != PERL_UNICODE_MAX + 1) {
8017 /* Add 0x110000. First, grow if necessary */
8019 if (invlist_max(invlist) < len) {
8020 invlist_extend(invlist, len);
8021 array = invlist_array(invlist);
8023 invlist_set_len(invlist, len);
8024 array[len - 1] = PERL_UNICODE_MAX + 1;
8026 else { /* Remove the 0x110000 */
8027 invlist_set_len(invlist, len - 1);
8035 PERL_STATIC_INLINE SV*
8036 S_invlist_clone(pTHX_ SV* const invlist)
8039 /* Return a new inversion list that is a copy of the input one, which is
8042 /* Need to allocate extra space to accommodate Perl's addition of a
8043 * trailing NUL to SvPV's, since it thinks they are always strings */
8044 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8045 STRLEN length = SvCUR(invlist);
8047 PERL_ARGS_ASSERT_INVLIST_CLONE;
8049 SvCUR_set(new_invlist, length); /* This isn't done automatically */
8050 Copy(SvPVX(invlist), SvPVX(new_invlist), length, char);
8055 PERL_STATIC_INLINE UV*
8056 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8058 /* Return the address of the UV that contains the current iteration
8061 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8063 return (UV *) (SvPVX(invlist) + (INVLIST_ITER_OFFSET * sizeof (UV)));
8066 PERL_STATIC_INLINE UV*
8067 S_get_invlist_version_id_addr(pTHX_ SV* invlist)
8069 /* Return the address of the UV that contains the version id. */
8071 PERL_ARGS_ASSERT_GET_INVLIST_VERSION_ID_ADDR;
8073 return (UV *) (SvPVX(invlist) + (INVLIST_VERSION_ID_OFFSET * sizeof (UV)));
8076 PERL_STATIC_INLINE void
8077 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8079 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8081 *get_invlist_iter_addr(invlist) = 0;
8084 PERL_STATIC_INLINE void
8085 S_invlist_iterfinish(pTHX_ SV* invlist)
8087 /* Terminate iterator for invlist. This is to catch development errors.
8088 * Any iteration that is interrupted before completed should call this
8089 * function. Functions that add code points anywhere else but to the end
8090 * of an inversion list assert that they are not in the middle of an
8091 * iteration. If they were, the addition would make the iteration
8092 * problematical: if the iteration hadn't reached the place where things
8093 * were being added, it would be ok */
8095 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8097 *get_invlist_iter_addr(invlist) = UV_MAX;
8101 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8103 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8104 * This call sets in <*start> and <*end>, the next range in <invlist>.
8105 * Returns <TRUE> if successful and the next call will return the next
8106 * range; <FALSE> if was already at the end of the list. If the latter,
8107 * <*start> and <*end> are unchanged, and the next call to this function
8108 * will start over at the beginning of the list */
8110 UV* pos = get_invlist_iter_addr(invlist);
8111 UV len = _invlist_len(invlist);
8114 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8117 *pos = UV_MAX; /* Force iterinit() to be required next time */
8121 array = invlist_array(invlist);
8123 *start = array[(*pos)++];
8129 *end = array[(*pos)++] - 1;
8135 PERL_STATIC_INLINE bool
8136 S_invlist_is_iterating(pTHX_ SV* const invlist)
8138 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8140 return *(get_invlist_iter_addr(invlist)) < UV_MAX;
8143 PERL_STATIC_INLINE UV
8144 S_invlist_highest(pTHX_ SV* const invlist)
8146 /* Returns the highest code point that matches an inversion list. This API
8147 * has an ambiguity, as it returns 0 under either the highest is actually
8148 * 0, or if the list is empty. If this distinction matters to you, check
8149 * for emptiness before calling this function */
8151 UV len = _invlist_len(invlist);
8154 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8160 array = invlist_array(invlist);
8162 /* The last element in the array in the inversion list always starts a
8163 * range that goes to infinity. That range may be for code points that are
8164 * matched in the inversion list, or it may be for ones that aren't
8165 * matched. In the latter case, the highest code point in the set is one
8166 * less than the beginning of this range; otherwise it is the final element
8167 * of this range: infinity */
8168 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8170 : array[len - 1] - 1;
8173 #ifndef PERL_IN_XSUB_RE
8175 Perl__invlist_contents(pTHX_ SV* const invlist)
8177 /* Get the contents of an inversion list into a string SV so that they can
8178 * be printed out. It uses the format traditionally done for debug tracing
8182 SV* output = newSVpvs("\n");
8184 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8186 assert(! invlist_is_iterating(invlist));
8188 invlist_iterinit(invlist);
8189 while (invlist_iternext(invlist, &start, &end)) {
8190 if (end == UV_MAX) {
8191 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8193 else if (end != start) {
8194 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8198 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
8206 #ifdef PERL_ARGS_ASSERT__INVLIST_DUMP
8208 Perl__invlist_dump(pTHX_ SV* const invlist, const char * const header)
8210 /* Dumps out the ranges in an inversion list. The string 'header'
8211 * if present is output on a line before the first range */
8215 PERL_ARGS_ASSERT__INVLIST_DUMP;
8217 if (header && strlen(header)) {
8218 PerlIO_printf(Perl_debug_log, "%s\n", header);
8220 if (invlist_is_iterating(invlist)) {
8221 PerlIO_printf(Perl_debug_log, "Can't dump because is in middle of iterating\n");
8225 invlist_iterinit(invlist);
8226 while (invlist_iternext(invlist, &start, &end)) {
8227 if (end == UV_MAX) {
8228 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. INFINITY\n", start);
8230 else if (end != start) {
8231 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. 0x%04"UVXf"\n",
8235 PerlIO_printf(Perl_debug_log, "0x%04"UVXf"\n", start);
8243 S__invlistEQ(pTHX_ SV* const a, SV* const b, bool complement_b)
8245 /* Return a boolean as to if the two passed in inversion lists are
8246 * identical. The final argument, if TRUE, says to take the complement of
8247 * the second inversion list before doing the comparison */
8249 UV* array_a = invlist_array(a);
8250 UV* array_b = invlist_array(b);
8251 UV len_a = _invlist_len(a);
8252 UV len_b = _invlist_len(b);
8254 UV i = 0; /* current index into the arrays */
8255 bool retval = TRUE; /* Assume are identical until proven otherwise */
8257 PERL_ARGS_ASSERT__INVLISTEQ;
8259 /* If are to compare 'a' with the complement of b, set it
8260 * up so are looking at b's complement. */
8263 /* The complement of nothing is everything, so <a> would have to have
8264 * just one element, starting at zero (ending at infinity) */
8266 return (len_a == 1 && array_a[0] == 0);
8268 else if (array_b[0] == 0) {
8270 /* Otherwise, to complement, we invert. Here, the first element is
8271 * 0, just remove it. To do this, we just pretend the array starts
8272 * one later, and clear the flag as we don't have to do anything
8277 complement_b = FALSE;
8281 /* But if the first element is not zero, we unshift a 0 before the
8282 * array. The data structure reserves a space for that 0 (which
8283 * should be a '1' right now), so physical shifting is unneeded,
8284 * but temporarily change that element to 0. Before exiting the
8285 * routine, we must restore the element to '1' */
8292 /* Make sure that the lengths are the same, as well as the final element
8293 * before looping through the remainder. (Thus we test the length, final,
8294 * and first elements right off the bat) */
8295 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
8298 else for (i = 0; i < len_a - 1; i++) {
8299 if (array_a[i] != array_b[i]) {
8312 #undef HEADER_LENGTH
8313 #undef INVLIST_INITIAL_LENGTH
8314 #undef TO_INTERNAL_SIZE
8315 #undef FROM_INTERNAL_SIZE
8316 #undef INVLIST_LEN_OFFSET
8317 #undef INVLIST_ZERO_OFFSET
8318 #undef INVLIST_ITER_OFFSET
8319 #undef INVLIST_VERSION_ID
8320 #undef INVLIST_PREVIOUS_INDEX_OFFSET
8322 /* End of inversion list object */
8325 S_parse_lparen_question_flags(pTHX_ struct RExC_state_t *pRExC_state)
8327 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
8328 * constructs, and updates RExC_flags with them. On input, RExC_parse
8329 * should point to the first flag; it is updated on output to point to the
8330 * final ')' or ':'. There needs to be at least one flag, or this will
8333 /* for (?g), (?gc), and (?o) warnings; warning
8334 about (?c) will warn about (?g) -- japhy */
8336 #define WASTED_O 0x01
8337 #define WASTED_G 0x02
8338 #define WASTED_C 0x04
8339 #define WASTED_GC (0x02|0x04)
8340 I32 wastedflags = 0x00;
8341 U32 posflags = 0, negflags = 0;
8342 U32 *flagsp = &posflags;
8343 char has_charset_modifier = '\0';
8345 bool has_use_defaults = FALSE;
8346 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
8348 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
8350 /* '^' as an initial flag sets certain defaults */
8351 if (UCHARAT(RExC_parse) == '^') {
8353 has_use_defaults = TRUE;
8354 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
8355 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
8356 ? REGEX_UNICODE_CHARSET
8357 : REGEX_DEPENDS_CHARSET);
8360 cs = get_regex_charset(RExC_flags);
8361 if (cs == REGEX_DEPENDS_CHARSET
8362 && (RExC_utf8 || RExC_uni_semantics))
8364 cs = REGEX_UNICODE_CHARSET;
8367 while (*RExC_parse) {
8368 /* && strchr("iogcmsx", *RExC_parse) */
8369 /* (?g), (?gc) and (?o) are useless here
8370 and must be globally applied -- japhy */
8371 switch (*RExC_parse) {
8373 /* Code for the imsx flags */
8374 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
8376 case LOCALE_PAT_MOD:
8377 if (has_charset_modifier) {
8378 goto excess_modifier;
8380 else if (flagsp == &negflags) {
8383 cs = REGEX_LOCALE_CHARSET;
8384 has_charset_modifier = LOCALE_PAT_MOD;
8385 RExC_contains_locale = 1;
8387 case UNICODE_PAT_MOD:
8388 if (has_charset_modifier) {
8389 goto excess_modifier;
8391 else if (flagsp == &negflags) {
8394 cs = REGEX_UNICODE_CHARSET;
8395 has_charset_modifier = UNICODE_PAT_MOD;
8397 case ASCII_RESTRICT_PAT_MOD:
8398 if (flagsp == &negflags) {
8401 if (has_charset_modifier) {
8402 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
8403 goto excess_modifier;
8405 /* Doubled modifier implies more restricted */
8406 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
8409 cs = REGEX_ASCII_RESTRICTED_CHARSET;
8411 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
8413 case DEPENDS_PAT_MOD:
8414 if (has_use_defaults) {
8415 goto fail_modifiers;
8417 else if (flagsp == &negflags) {
8420 else if (has_charset_modifier) {
8421 goto excess_modifier;
8424 /* The dual charset means unicode semantics if the
8425 * pattern (or target, not known until runtime) are
8426 * utf8, or something in the pattern indicates unicode
8428 cs = (RExC_utf8 || RExC_uni_semantics)
8429 ? REGEX_UNICODE_CHARSET
8430 : REGEX_DEPENDS_CHARSET;
8431 has_charset_modifier = DEPENDS_PAT_MOD;
8435 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
8436 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
8438 else if (has_charset_modifier == *(RExC_parse - 1)) {
8439 vFAIL2("Regexp modifier \"%c\" may not appear twice", *(RExC_parse - 1));
8442 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
8447 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"", *(RExC_parse - 1));
8449 case ONCE_PAT_MOD: /* 'o' */
8450 case GLOBAL_PAT_MOD: /* 'g' */
8451 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8452 const I32 wflagbit = *RExC_parse == 'o' ? WASTED_O : WASTED_G;
8453 if (! (wastedflags & wflagbit) ) {
8454 wastedflags |= wflagbit;
8457 "Useless (%s%c) - %suse /%c modifier",
8458 flagsp == &negflags ? "?-" : "?",
8460 flagsp == &negflags ? "don't " : "",
8467 case CONTINUE_PAT_MOD: /* 'c' */
8468 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8469 if (! (wastedflags & WASTED_C) ) {
8470 wastedflags |= WASTED_GC;
8473 "Useless (%sc) - %suse /gc modifier",
8474 flagsp == &negflags ? "?-" : "?",
8475 flagsp == &negflags ? "don't " : ""
8480 case KEEPCOPY_PAT_MOD: /* 'p' */
8481 if (flagsp == &negflags) {
8483 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
8485 *flagsp |= RXf_PMf_KEEPCOPY;
8489 /* A flag is a default iff it is following a minus, so
8490 * if there is a minus, it means will be trying to
8491 * re-specify a default which is an error */
8492 if (has_use_defaults || flagsp == &negflags) {
8493 goto fail_modifiers;
8496 wastedflags = 0; /* reset so (?g-c) warns twice */
8500 RExC_flags |= posflags;
8501 RExC_flags &= ~negflags;
8502 set_regex_charset(&RExC_flags, cs);
8508 vFAIL3("Sequence (%.*s...) not recognized",
8509 RExC_parse-seqstart, seqstart);
8518 - reg - regular expression, i.e. main body or parenthesized thing
8520 * Caller must absorb opening parenthesis.
8522 * Combining parenthesis handling with the base level of regular expression
8523 * is a trifle forced, but the need to tie the tails of the branches to what
8524 * follows makes it hard to avoid.
8526 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
8528 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
8530 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
8533 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
8534 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
8535 needs to be restarted.
8536 Otherwise would only return NULL if regbranch() returns NULL, which
8539 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
8540 /* paren: Parenthesized? 0=top, 1=(, inside: changed to letter. */
8543 regnode *ret; /* Will be the head of the group. */
8546 regnode *ender = NULL;
8549 U32 oregflags = RExC_flags;
8550 bool have_branch = 0;
8552 I32 freeze_paren = 0;
8553 I32 after_freeze = 0;
8555 char * parse_start = RExC_parse; /* MJD */
8556 char * const oregcomp_parse = RExC_parse;
8558 GET_RE_DEBUG_FLAGS_DECL;
8560 PERL_ARGS_ASSERT_REG;
8561 DEBUG_PARSE("reg ");
8563 *flagp = 0; /* Tentatively. */
8566 /* Make an OPEN node, if parenthesized. */
8568 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
8569 char *start_verb = RExC_parse;
8570 STRLEN verb_len = 0;
8571 char *start_arg = NULL;
8572 unsigned char op = 0;
8574 int internal_argval = 0; /* internal_argval is only useful if !argok */
8575 while ( *RExC_parse && *RExC_parse != ')' ) {
8576 if ( *RExC_parse == ':' ) {
8577 start_arg = RExC_parse + 1;
8583 verb_len = RExC_parse - start_verb;
8586 while ( *RExC_parse && *RExC_parse != ')' )
8588 if ( *RExC_parse != ')' )
8589 vFAIL("Unterminated verb pattern argument");
8590 if ( RExC_parse == start_arg )
8593 if ( *RExC_parse != ')' )
8594 vFAIL("Unterminated verb pattern");
8597 switch ( *start_verb ) {
8598 case 'A': /* (*ACCEPT) */
8599 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
8601 internal_argval = RExC_nestroot;
8604 case 'C': /* (*COMMIT) */
8605 if ( memEQs(start_verb,verb_len,"COMMIT") )
8608 case 'F': /* (*FAIL) */
8609 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
8614 case ':': /* (*:NAME) */
8615 case 'M': /* (*MARK:NAME) */
8616 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
8621 case 'P': /* (*PRUNE) */
8622 if ( memEQs(start_verb,verb_len,"PRUNE") )
8625 case 'S': /* (*SKIP) */
8626 if ( memEQs(start_verb,verb_len,"SKIP") )
8629 case 'T': /* (*THEN) */
8630 /* [19:06] <TimToady> :: is then */
8631 if ( memEQs(start_verb,verb_len,"THEN") ) {
8633 RExC_seen |= REG_SEEN_CUTGROUP;
8639 vFAIL3("Unknown verb pattern '%.*s'",
8640 verb_len, start_verb);
8643 if ( start_arg && internal_argval ) {
8644 vFAIL3("Verb pattern '%.*s' may not have an argument",
8645 verb_len, start_verb);
8646 } else if ( argok < 0 && !start_arg ) {
8647 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
8648 verb_len, start_verb);
8650 ret = reganode(pRExC_state, op, internal_argval);
8651 if ( ! internal_argval && ! SIZE_ONLY ) {
8653 SV *sv = newSVpvn( start_arg, RExC_parse - start_arg);
8654 ARG(ret) = add_data( pRExC_state, 1, "S" );
8655 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
8662 if (!internal_argval)
8663 RExC_seen |= REG_SEEN_VERBARG;
8664 } else if ( start_arg ) {
8665 vFAIL3("Verb pattern '%.*s' may not have an argument",
8666 verb_len, start_verb);
8668 ret = reg_node(pRExC_state, op);
8670 nextchar(pRExC_state);
8673 if (*RExC_parse == '?') { /* (?...) */
8674 bool is_logical = 0;
8675 const char * const seqstart = RExC_parse;
8678 paren = *RExC_parse++;
8679 ret = NULL; /* For look-ahead/behind. */
8682 case 'P': /* (?P...) variants for those used to PCRE/Python */
8683 paren = *RExC_parse++;
8684 if ( paren == '<') /* (?P<...>) named capture */
8686 else if (paren == '>') { /* (?P>name) named recursion */
8687 goto named_recursion;
8689 else if (paren == '=') { /* (?P=...) named backref */
8690 /* this pretty much dupes the code for \k<NAME> in regatom(), if
8691 you change this make sure you change that */
8692 char* name_start = RExC_parse;
8694 SV *sv_dat = reg_scan_name(pRExC_state,
8695 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8696 if (RExC_parse == name_start || *RExC_parse != ')')
8697 vFAIL2("Sequence %.3s... not terminated",parse_start);
8700 num = add_data( pRExC_state, 1, "S" );
8701 RExC_rxi->data->data[num]=(void*)sv_dat;
8702 SvREFCNT_inc_simple_void(sv_dat);
8705 ret = reganode(pRExC_state,
8708 : (ASCII_FOLD_RESTRICTED)
8710 : (AT_LEAST_UNI_SEMANTICS)
8718 Set_Node_Offset(ret, parse_start+1);
8719 Set_Node_Cur_Length(ret); /* MJD */
8721 nextchar(pRExC_state);
8725 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8727 case '<': /* (?<...) */
8728 if (*RExC_parse == '!')
8730 else if (*RExC_parse != '=')
8736 case '\'': /* (?'...') */
8737 name_start= RExC_parse;
8738 svname = reg_scan_name(pRExC_state,
8739 SIZE_ONLY ? /* reverse test from the others */
8740 REG_RSN_RETURN_NAME :
8741 REG_RSN_RETURN_NULL);
8742 if (RExC_parse == name_start) {
8744 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8747 if (*RExC_parse != paren)
8748 vFAIL2("Sequence (?%c... not terminated",
8749 paren=='>' ? '<' : paren);
8753 if (!svname) /* shouldn't happen */
8755 "panic: reg_scan_name returned NULL");
8756 if (!RExC_paren_names) {
8757 RExC_paren_names= newHV();
8758 sv_2mortal(MUTABLE_SV(RExC_paren_names));
8760 RExC_paren_name_list= newAV();
8761 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
8764 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
8766 sv_dat = HeVAL(he_str);
8768 /* croak baby croak */
8770 "panic: paren_name hash element allocation failed");
8771 } else if ( SvPOK(sv_dat) ) {
8772 /* (?|...) can mean we have dupes so scan to check
8773 its already been stored. Maybe a flag indicating
8774 we are inside such a construct would be useful,
8775 but the arrays are likely to be quite small, so
8776 for now we punt -- dmq */
8777 IV count = SvIV(sv_dat);
8778 I32 *pv = (I32*)SvPVX(sv_dat);
8780 for ( i = 0 ; i < count ; i++ ) {
8781 if ( pv[i] == RExC_npar ) {
8787 pv = (I32*)SvGROW(sv_dat, SvCUR(sv_dat) + sizeof(I32)+1);
8788 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
8789 pv[count] = RExC_npar;
8790 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
8793 (void)SvUPGRADE(sv_dat,SVt_PVNV);
8794 sv_setpvn(sv_dat, (char *)&(RExC_npar), sizeof(I32));
8796 SvIV_set(sv_dat, 1);
8799 /* Yes this does cause a memory leak in debugging Perls */
8800 if (!av_store(RExC_paren_name_list, RExC_npar, SvREFCNT_inc(svname)))
8801 SvREFCNT_dec_NN(svname);
8804 /*sv_dump(sv_dat);*/
8806 nextchar(pRExC_state);
8808 goto capturing_parens;
8810 RExC_seen |= REG_SEEN_LOOKBEHIND;
8811 RExC_in_lookbehind++;
8813 case '=': /* (?=...) */
8814 RExC_seen_zerolen++;
8816 case '!': /* (?!...) */
8817 RExC_seen_zerolen++;
8818 if (*RExC_parse == ')') {
8819 ret=reg_node(pRExC_state, OPFAIL);
8820 nextchar(pRExC_state);
8824 case '|': /* (?|...) */
8825 /* branch reset, behave like a (?:...) except that
8826 buffers in alternations share the same numbers */
8828 after_freeze = freeze_paren = RExC_npar;
8830 case ':': /* (?:...) */
8831 case '>': /* (?>...) */
8833 case '$': /* (?$...) */
8834 case '@': /* (?@...) */
8835 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
8837 case '0' : /* (?0) */
8838 case 'R' : /* (?R) */
8839 if (*RExC_parse != ')')
8840 FAIL("Sequence (?R) not terminated");
8841 ret = reg_node(pRExC_state, GOSTART);
8842 *flagp |= POSTPONED;
8843 nextchar(pRExC_state);
8846 { /* named and numeric backreferences */
8848 case '&': /* (?&NAME) */
8849 parse_start = RExC_parse - 1;
8852 SV *sv_dat = reg_scan_name(pRExC_state,
8853 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8854 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8856 goto gen_recurse_regop;
8857 assert(0); /* NOT REACHED */
8859 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8861 vFAIL("Illegal pattern");
8863 goto parse_recursion;
8865 case '-': /* (?-1) */
8866 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8867 RExC_parse--; /* rewind to let it be handled later */
8871 case '1': case '2': case '3': case '4': /* (?1) */
8872 case '5': case '6': case '7': case '8': case '9':
8875 num = atoi(RExC_parse);
8876 parse_start = RExC_parse - 1; /* MJD */
8877 if (*RExC_parse == '-')
8879 while (isDIGIT(*RExC_parse))
8881 if (*RExC_parse!=')')
8882 vFAIL("Expecting close bracket");
8885 if ( paren == '-' ) {
8887 Diagram of capture buffer numbering.
8888 Top line is the normal capture buffer numbers
8889 Bottom line is the negative indexing as from
8893 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
8897 num = RExC_npar + num;
8900 vFAIL("Reference to nonexistent group");
8902 } else if ( paren == '+' ) {
8903 num = RExC_npar + num - 1;
8906 ret = reganode(pRExC_state, GOSUB, num);
8908 if (num > (I32)RExC_rx->nparens) {
8910 vFAIL("Reference to nonexistent group");
8912 ARG2L_SET( ret, RExC_recurse_count++);
8914 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
8915 "Recurse #%"UVuf" to %"IVdf"\n", (UV)ARG(ret), (IV)ARG2L(ret)));
8919 RExC_seen |= REG_SEEN_RECURSE;
8920 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
8921 Set_Node_Offset(ret, parse_start); /* MJD */
8923 *flagp |= POSTPONED;
8924 nextchar(pRExC_state);
8926 } /* named and numeric backreferences */
8927 assert(0); /* NOT REACHED */
8929 case '?': /* (??...) */
8931 if (*RExC_parse != '{') {
8933 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8936 *flagp |= POSTPONED;
8937 paren = *RExC_parse++;
8939 case '{': /* (?{...}) */
8942 struct reg_code_block *cb;
8944 RExC_seen_zerolen++;
8946 if ( !pRExC_state->num_code_blocks
8947 || pRExC_state->code_index >= pRExC_state->num_code_blocks
8948 || pRExC_state->code_blocks[pRExC_state->code_index].start
8949 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
8952 if (RExC_pm_flags & PMf_USE_RE_EVAL)
8953 FAIL("panic: Sequence (?{...}): no code block found\n");
8954 FAIL("Eval-group not allowed at runtime, use re 'eval'");
8956 /* this is a pre-compiled code block (?{...}) */
8957 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
8958 RExC_parse = RExC_start + cb->end;
8961 if (cb->src_regex) {
8962 n = add_data(pRExC_state, 2, "rl");
8963 RExC_rxi->data->data[n] =
8964 (void*)SvREFCNT_inc((SV*)cb->src_regex);
8965 RExC_rxi->data->data[n+1] = (void*)o;
8968 n = add_data(pRExC_state, 1,
8969 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l");
8970 RExC_rxi->data->data[n] = (void*)o;
8973 pRExC_state->code_index++;
8974 nextchar(pRExC_state);
8978 ret = reg_node(pRExC_state, LOGICAL);
8979 eval = reganode(pRExC_state, EVAL, n);
8982 /* for later propagation into (??{}) return value */
8983 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
8985 REGTAIL(pRExC_state, ret, eval);
8986 /* deal with the length of this later - MJD */
8989 ret = reganode(pRExC_state, EVAL, n);
8990 Set_Node_Length(ret, RExC_parse - parse_start + 1);
8991 Set_Node_Offset(ret, parse_start);
8994 case '(': /* (?(?{...})...) and (?(?=...)...) */
8997 if (RExC_parse[0] == '?') { /* (?(?...)) */
8998 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
8999 || RExC_parse[1] == '<'
9000 || RExC_parse[1] == '{') { /* Lookahead or eval. */
9004 ret = reg_node(pRExC_state, LOGICAL);
9008 tail = reg(pRExC_state, 1, &flag, depth+1);
9009 if (flag & RESTART_UTF8) {
9010 *flagp = RESTART_UTF8;
9013 REGTAIL(pRExC_state, ret, tail);
9017 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9018 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9020 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9021 char *name_start= RExC_parse++;
9023 SV *sv_dat=reg_scan_name(pRExC_state,
9024 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9025 if (RExC_parse == name_start || *RExC_parse != ch)
9026 vFAIL2("Sequence (?(%c... not terminated",
9027 (ch == '>' ? '<' : ch));
9030 num = add_data( pRExC_state, 1, "S" );
9031 RExC_rxi->data->data[num]=(void*)sv_dat;
9032 SvREFCNT_inc_simple_void(sv_dat);
9034 ret = reganode(pRExC_state,NGROUPP,num);
9035 goto insert_if_check_paren;
9037 else if (RExC_parse[0] == 'D' &&
9038 RExC_parse[1] == 'E' &&
9039 RExC_parse[2] == 'F' &&
9040 RExC_parse[3] == 'I' &&
9041 RExC_parse[4] == 'N' &&
9042 RExC_parse[5] == 'E')
9044 ret = reganode(pRExC_state,DEFINEP,0);
9047 goto insert_if_check_paren;
9049 else if (RExC_parse[0] == 'R') {
9052 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9053 parno = atoi(RExC_parse++);
9054 while (isDIGIT(*RExC_parse))
9056 } else if (RExC_parse[0] == '&') {
9059 sv_dat = reg_scan_name(pRExC_state,
9060 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9061 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9063 ret = reganode(pRExC_state,INSUBP,parno);
9064 goto insert_if_check_paren;
9066 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9069 parno = atoi(RExC_parse++);
9071 while (isDIGIT(*RExC_parse))
9073 ret = reganode(pRExC_state, GROUPP, parno);
9075 insert_if_check_paren:
9076 if ((c = *nextchar(pRExC_state)) != ')')
9077 vFAIL("Switch condition not recognized");
9079 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9080 br = regbranch(pRExC_state, &flags, 1,depth+1);
9082 if (flags & RESTART_UTF8) {
9083 *flagp = RESTART_UTF8;
9086 FAIL2("panic: regbranch returned NULL, flags=%#X",
9089 REGTAIL(pRExC_state, br, reganode(pRExC_state, LONGJMP, 0));
9090 c = *nextchar(pRExC_state);
9095 vFAIL("(?(DEFINE)....) does not allow branches");
9096 lastbr = reganode(pRExC_state, IFTHEN, 0); /* Fake one for optimizer. */
9097 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
9098 if (flags & RESTART_UTF8) {
9099 *flagp = RESTART_UTF8;
9102 FAIL2("panic: regbranch returned NULL, flags=%#X",
9105 REGTAIL(pRExC_state, ret, lastbr);
9108 c = *nextchar(pRExC_state);
9113 vFAIL("Switch (?(condition)... contains too many branches");
9114 ender = reg_node(pRExC_state, TAIL);
9115 REGTAIL(pRExC_state, br, ender);
9117 REGTAIL(pRExC_state, lastbr, ender);
9118 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
9121 REGTAIL(pRExC_state, ret, ender);
9122 RExC_size++; /* XXX WHY do we need this?!!
9123 For large programs it seems to be required
9124 but I can't figure out why. -- dmq*/
9128 vFAIL2("Unknown switch condition (?(%.2s", RExC_parse);
9131 case '[': /* (?[ ... ]) */
9132 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
9135 RExC_parse--; /* for vFAIL to print correctly */
9136 vFAIL("Sequence (? incomplete");
9138 default: /* e.g., (?i) */
9141 parse_lparen_question_flags(pRExC_state);
9142 if (UCHARAT(RExC_parse) != ':') {
9143 nextchar(pRExC_state);
9148 nextchar(pRExC_state);
9158 ret = reganode(pRExC_state, OPEN, parno);
9161 RExC_nestroot = parno;
9162 if (RExC_seen & REG_SEEN_RECURSE
9163 && !RExC_open_parens[parno-1])
9165 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9166 "Setting open paren #%"IVdf" to %d\n",
9167 (IV)parno, REG_NODE_NUM(ret)));
9168 RExC_open_parens[parno-1]= ret;
9171 Set_Node_Length(ret, 1); /* MJD */
9172 Set_Node_Offset(ret, RExC_parse); /* MJD */
9180 /* Pick up the branches, linking them together. */
9181 parse_start = RExC_parse; /* MJD */
9182 br = regbranch(pRExC_state, &flags, 1,depth+1);
9184 /* branch_len = (paren != 0); */
9187 if (flags & RESTART_UTF8) {
9188 *flagp = RESTART_UTF8;
9191 FAIL2("panic: regbranch returned NULL, flags=%#X", flags);
9193 if (*RExC_parse == '|') {
9194 if (!SIZE_ONLY && RExC_extralen) {
9195 reginsert(pRExC_state, BRANCHJ, br, depth+1);
9198 reginsert(pRExC_state, BRANCH, br, depth+1);
9199 Set_Node_Length(br, paren != 0);
9200 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
9204 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
9206 else if (paren == ':') {
9207 *flagp |= flags&SIMPLE;
9209 if (is_open) { /* Starts with OPEN. */
9210 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
9212 else if (paren != '?') /* Not Conditional */
9214 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9216 while (*RExC_parse == '|') {
9217 if (!SIZE_ONLY && RExC_extralen) {
9218 ender = reganode(pRExC_state, LONGJMP,0);
9219 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender); /* Append to the previous. */
9222 RExC_extralen += 2; /* Account for LONGJMP. */
9223 nextchar(pRExC_state);
9225 if (RExC_npar > after_freeze)
9226 after_freeze = RExC_npar;
9227 RExC_npar = freeze_paren;
9229 br = regbranch(pRExC_state, &flags, 0, depth+1);
9232 if (flags & RESTART_UTF8) {
9233 *flagp = RESTART_UTF8;
9236 FAIL2("panic: regbranch returned NULL, flags=%#X", flags);
9238 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
9240 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9243 if (have_branch || paren != ':') {
9244 /* Make a closing node, and hook it on the end. */
9247 ender = reg_node(pRExC_state, TAIL);
9250 ender = reganode(pRExC_state, CLOSE, parno);
9251 if (!SIZE_ONLY && RExC_seen & REG_SEEN_RECURSE) {
9252 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9253 "Setting close paren #%"IVdf" to %d\n",
9254 (IV)parno, REG_NODE_NUM(ender)));
9255 RExC_close_parens[parno-1]= ender;
9256 if (RExC_nestroot == parno)
9259 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
9260 Set_Node_Length(ender,1); /* MJD */
9266 *flagp &= ~HASWIDTH;
9269 ender = reg_node(pRExC_state, SUCCEED);
9272 ender = reg_node(pRExC_state, END);
9274 assert(!RExC_opend); /* there can only be one! */
9279 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9280 SV * const mysv_val1=sv_newmortal();
9281 SV * const mysv_val2=sv_newmortal();
9282 DEBUG_PARSE_MSG("lsbr");
9283 regprop(RExC_rx, mysv_val1, lastbr);
9284 regprop(RExC_rx, mysv_val2, ender);
9285 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9286 SvPV_nolen_const(mysv_val1),
9287 (IV)REG_NODE_NUM(lastbr),
9288 SvPV_nolen_const(mysv_val2),
9289 (IV)REG_NODE_NUM(ender),
9290 (IV)(ender - lastbr)
9293 REGTAIL(pRExC_state, lastbr, ender);
9295 if (have_branch && !SIZE_ONLY) {
9298 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
9300 /* Hook the tails of the branches to the closing node. */
9301 for (br = ret; br; br = regnext(br)) {
9302 const U8 op = PL_regkind[OP(br)];
9304 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
9305 if (OP(NEXTOPER(br)) != NOTHING || regnext(NEXTOPER(br)) != ender)
9308 else if (op == BRANCHJ) {
9309 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
9310 /* for now we always disable this optimisation * /
9311 if (OP(NEXTOPER(NEXTOPER(br))) != NOTHING || regnext(NEXTOPER(NEXTOPER(br))) != ender)
9317 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
9318 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9319 SV * const mysv_val1=sv_newmortal();
9320 SV * const mysv_val2=sv_newmortal();
9321 DEBUG_PARSE_MSG("NADA");
9322 regprop(RExC_rx, mysv_val1, ret);
9323 regprop(RExC_rx, mysv_val2, ender);
9324 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9325 SvPV_nolen_const(mysv_val1),
9326 (IV)REG_NODE_NUM(ret),
9327 SvPV_nolen_const(mysv_val2),
9328 (IV)REG_NODE_NUM(ender),
9333 if (OP(ender) == TAIL) {
9338 for ( opt= br + 1; opt < ender ; opt++ )
9340 NEXT_OFF(br)= ender - br;
9348 static const char parens[] = "=!<,>";
9350 if (paren && (p = strchr(parens, paren))) {
9351 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
9352 int flag = (p - parens) > 1;
9355 node = SUSPEND, flag = 0;
9356 reginsert(pRExC_state, node,ret, depth+1);
9357 Set_Node_Cur_Length(ret);
9358 Set_Node_Offset(ret, parse_start + 1);
9360 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
9364 /* Check for proper termination. */
9366 RExC_flags = oregflags;
9367 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
9368 RExC_parse = oregcomp_parse;
9369 vFAIL("Unmatched (");
9372 else if (!paren && RExC_parse < RExC_end) {
9373 if (*RExC_parse == ')') {
9375 vFAIL("Unmatched )");
9378 FAIL("Junk on end of regexp"); /* "Can't happen". */
9379 assert(0); /* NOTREACHED */
9382 if (RExC_in_lookbehind) {
9383 RExC_in_lookbehind--;
9385 if (after_freeze > RExC_npar)
9386 RExC_npar = after_freeze;
9391 - regbranch - one alternative of an | operator
9393 * Implements the concatenation operator.
9395 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9399 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
9403 regnode *chain = NULL;
9405 I32 flags = 0, c = 0;
9406 GET_RE_DEBUG_FLAGS_DECL;
9408 PERL_ARGS_ASSERT_REGBRANCH;
9410 DEBUG_PARSE("brnc");
9415 if (!SIZE_ONLY && RExC_extralen)
9416 ret = reganode(pRExC_state, BRANCHJ,0);
9418 ret = reg_node(pRExC_state, BRANCH);
9419 Set_Node_Length(ret, 1);
9423 if (!first && SIZE_ONLY)
9424 RExC_extralen += 1; /* BRANCHJ */
9426 *flagp = WORST; /* Tentatively. */
9429 nextchar(pRExC_state);
9430 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
9432 latest = regpiece(pRExC_state, &flags,depth+1);
9433 if (latest == NULL) {
9434 if (flags & TRYAGAIN)
9436 if (flags & RESTART_UTF8) {
9437 *flagp = RESTART_UTF8;
9440 FAIL2("panic: regpiece returned NULL, flags=%#X", flags);
9442 else if (ret == NULL)
9444 *flagp |= flags&(HASWIDTH|POSTPONED);
9445 if (chain == NULL) /* First piece. */
9446 *flagp |= flags&SPSTART;
9449 REGTAIL(pRExC_state, chain, latest);
9454 if (chain == NULL) { /* Loop ran zero times. */
9455 chain = reg_node(pRExC_state, NOTHING);
9460 *flagp |= flags&SIMPLE;
9467 - regpiece - something followed by possible [*+?]
9469 * Note that the branching code sequences used for ? and the general cases
9470 * of * and + are somewhat optimized: they use the same NOTHING node as
9471 * both the endmarker for their branch list and the body of the last branch.
9472 * It might seem that this node could be dispensed with entirely, but the
9473 * endmarker role is not redundant.
9475 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
9477 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9481 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
9488 const char * const origparse = RExC_parse;
9490 I32 max = REG_INFTY;
9491 #ifdef RE_TRACK_PATTERN_OFFSETS
9494 const char *maxpos = NULL;
9496 /* Save the original in case we change the emitted regop to a FAIL. */
9497 regnode * const orig_emit = RExC_emit;
9499 GET_RE_DEBUG_FLAGS_DECL;
9501 PERL_ARGS_ASSERT_REGPIECE;
9503 DEBUG_PARSE("piec");
9505 ret = regatom(pRExC_state, &flags,depth+1);
9507 if (flags & (TRYAGAIN|RESTART_UTF8))
9508 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
9510 FAIL2("panic: regatom returned NULL, flags=%#X", flags);
9516 if (op == '{' && regcurly(RExC_parse, FALSE)) {
9518 #ifdef RE_TRACK_PATTERN_OFFSETS
9519 parse_start = RExC_parse; /* MJD */
9521 next = RExC_parse + 1;
9522 while (isDIGIT(*next) || *next == ',') {
9531 if (*next == '}') { /* got one */
9535 min = atoi(RExC_parse);
9539 maxpos = RExC_parse;
9541 if (!max && *maxpos != '0')
9542 max = REG_INFTY; /* meaning "infinity" */
9543 else if (max >= REG_INFTY)
9544 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
9546 nextchar(pRExC_state);
9547 if (max < min) { /* If can't match, warn and optimize to fail
9550 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
9552 /* We can't back off the size because we have to reserve
9553 * enough space for all the things we are about to throw
9554 * away, but we can shrink it by the ammount we are about
9556 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
9559 RExC_emit = orig_emit;
9561 ret = reg_node(pRExC_state, OPFAIL);
9564 else if (max == 0) { /* replace {0} with a nothing node */
9566 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)NOTHING];
9569 RExC_emit = orig_emit;
9571 ret = reg_node(pRExC_state, NOTHING);
9576 if ((flags&SIMPLE)) {
9577 RExC_naughty += 2 + RExC_naughty / 2;
9578 reginsert(pRExC_state, CURLY, ret, depth+1);
9579 Set_Node_Offset(ret, parse_start+1); /* MJD */
9580 Set_Node_Cur_Length(ret);
9583 regnode * const w = reg_node(pRExC_state, WHILEM);
9586 REGTAIL(pRExC_state, ret, w);
9587 if (!SIZE_ONLY && RExC_extralen) {
9588 reginsert(pRExC_state, LONGJMP,ret, depth+1);
9589 reginsert(pRExC_state, NOTHING,ret, depth+1);
9590 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
9592 reginsert(pRExC_state, CURLYX,ret, depth+1);
9594 Set_Node_Offset(ret, parse_start+1);
9595 Set_Node_Length(ret,
9596 op == '{' ? (RExC_parse - parse_start) : 1);
9598 if (!SIZE_ONLY && RExC_extralen)
9599 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
9600 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
9602 RExC_whilem_seen++, RExC_extralen += 3;
9603 RExC_naughty += 4 + RExC_naughty; /* compound interest */
9612 ARG1_SET(ret, (U16)min);
9613 ARG2_SET(ret, (U16)max);
9625 #if 0 /* Now runtime fix should be reliable. */
9627 /* if this is reinstated, don't forget to put this back into perldiag:
9629 =item Regexp *+ operand could be empty at {#} in regex m/%s/
9631 (F) The part of the regexp subject to either the * or + quantifier
9632 could match an empty string. The {#} shows in the regular
9633 expression about where the problem was discovered.
9637 if (!(flags&HASWIDTH) && op != '?')
9638 vFAIL("Regexp *+ operand could be empty");
9641 #ifdef RE_TRACK_PATTERN_OFFSETS
9642 parse_start = RExC_parse;
9644 nextchar(pRExC_state);
9646 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
9648 if (op == '*' && (flags&SIMPLE)) {
9649 reginsert(pRExC_state, STAR, ret, depth+1);
9653 else if (op == '*') {
9657 else if (op == '+' && (flags&SIMPLE)) {
9658 reginsert(pRExC_state, PLUS, ret, depth+1);
9662 else if (op == '+') {
9666 else if (op == '?') {
9671 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
9672 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
9673 ckWARN3reg(RExC_parse,
9674 "%.*s matches null string many times",
9675 (int)(RExC_parse >= origparse ? RExC_parse - origparse : 0),
9677 (void)ReREFCNT_inc(RExC_rx_sv);
9680 if (RExC_parse < RExC_end && *RExC_parse == '?') {
9681 nextchar(pRExC_state);
9682 reginsert(pRExC_state, MINMOD, ret, depth+1);
9683 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
9685 #ifndef REG_ALLOW_MINMOD_SUSPEND
9688 if (RExC_parse < RExC_end && *RExC_parse == '+') {
9690 nextchar(pRExC_state);
9691 ender = reg_node(pRExC_state, SUCCEED);
9692 REGTAIL(pRExC_state, ret, ender);
9693 reginsert(pRExC_state, SUSPEND, ret, depth+1);
9695 ender = reg_node(pRExC_state, TAIL);
9696 REGTAIL(pRExC_state, ret, ender);
9700 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
9702 vFAIL("Nested quantifiers");
9709 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p, UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
9710 const bool strict /* Apply stricter parsing rules? */
9714 /* This is expected to be called by a parser routine that has recognized '\N'
9715 and needs to handle the rest. RExC_parse is expected to point at the first
9716 char following the N at the time of the call. On successful return,
9717 RExC_parse has been updated to point to just after the sequence identified
9718 by this routine, and <*flagp> has been updated.
9720 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
9723 \N may begin either a named sequence, or if outside a character class, mean
9724 to match a non-newline. For non single-quoted regexes, the tokenizer has
9725 attempted to decide which, and in the case of a named sequence, converted it
9726 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
9727 where c1... are the characters in the sequence. For single-quoted regexes,
9728 the tokenizer passes the \N sequence through unchanged; this code will not
9729 attempt to determine this nor expand those, instead raising a syntax error.
9730 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
9731 or there is no '}', it signals that this \N occurrence means to match a
9734 Only the \N{U+...} form should occur in a character class, for the same
9735 reason that '.' inside a character class means to just match a period: it
9736 just doesn't make sense.
9738 The function raises an error (via vFAIL), and doesn't return for various
9739 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
9740 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
9741 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
9742 only possible if node_p is non-NULL.
9745 If <valuep> is non-null, it means the caller can accept an input sequence
9746 consisting of a just a single code point; <*valuep> is set to that value
9747 if the input is such.
9749 If <node_p> is non-null it signifies that the caller can accept any other
9750 legal sequence (i.e., one that isn't just a single code point). <*node_p>
9752 1) \N means not-a-NL: points to a newly created REG_ANY node;
9753 2) \N{}: points to a new NOTHING node;
9754 3) otherwise: points to a new EXACT node containing the resolved
9756 Note that FALSE is returned for single code point sequences if <valuep> is
9760 char * endbrace; /* '}' following the name */
9762 char *endchar; /* Points to '.' or '}' ending cur char in the input
9764 bool has_multiple_chars; /* true if the input stream contains a sequence of
9765 more than one character */
9767 GET_RE_DEBUG_FLAGS_DECL;
9769 PERL_ARGS_ASSERT_GROK_BSLASH_N;
9773 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
9775 /* The [^\n] meaning of \N ignores spaces and comments under the /x
9776 * modifier. The other meaning does not */
9777 p = (RExC_flags & RXf_PMf_EXTENDED)
9778 ? regwhite( pRExC_state, RExC_parse )
9781 /* Disambiguate between \N meaning a named character versus \N meaning
9782 * [^\n]. The former is assumed when it can't be the latter. */
9783 if (*p != '{' || regcurly(p, FALSE)) {
9786 /* no bare \N in a charclass */
9787 if (in_char_class) {
9788 vFAIL("\\N in a character class must be a named character: \\N{...}");
9792 nextchar(pRExC_state);
9793 *node_p = reg_node(pRExC_state, REG_ANY);
9794 *flagp |= HASWIDTH|SIMPLE;
9797 Set_Node_Length(*node_p, 1); /* MJD */
9801 /* Here, we have decided it should be a named character or sequence */
9803 /* The test above made sure that the next real character is a '{', but
9804 * under the /x modifier, it could be separated by space (or a comment and
9805 * \n) and this is not allowed (for consistency with \x{...} and the
9806 * tokenizer handling of \N{NAME}). */
9807 if (*RExC_parse != '{') {
9808 vFAIL("Missing braces on \\N{}");
9811 RExC_parse++; /* Skip past the '{' */
9813 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
9814 || ! (endbrace == RExC_parse /* nothing between the {} */
9815 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below */
9816 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg) */
9818 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
9819 vFAIL("\\N{NAME} must be resolved by the lexer");
9822 if (endbrace == RExC_parse) { /* empty: \N{} */
9825 *node_p = reg_node(pRExC_state,NOTHING);
9827 else if (in_char_class) {
9828 if (SIZE_ONLY && in_char_class) {
9830 RExC_parse++; /* Position after the "}" */
9831 vFAIL("Zero length \\N{}");
9834 ckWARNreg(RExC_parse,
9835 "Ignoring zero length \\N{} in character class");
9843 nextchar(pRExC_state);
9847 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
9848 RExC_parse += 2; /* Skip past the 'U+' */
9850 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9852 /* Code points are separated by dots. If none, there is only one code
9853 * point, and is terminated by the brace */
9854 has_multiple_chars = (endchar < endbrace);
9856 if (valuep && (! has_multiple_chars || in_char_class)) {
9857 /* We only pay attention to the first char of
9858 multichar strings being returned in char classes. I kinda wonder
9859 if this makes sense as it does change the behaviour
9860 from earlier versions, OTOH that behaviour was broken
9861 as well. XXX Solution is to recharacterize as
9862 [rest-of-class]|multi1|multi2... */
9864 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
9865 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
9866 | PERL_SCAN_DISALLOW_PREFIX
9867 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
9869 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
9871 /* The tokenizer should have guaranteed validity, but it's possible to
9872 * bypass it by using single quoting, so check */
9873 if (length_of_hex == 0
9874 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
9876 RExC_parse += length_of_hex; /* Includes all the valid */
9877 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
9878 ? UTF8SKIP(RExC_parse)
9880 /* Guard against malformed utf8 */
9881 if (RExC_parse >= endchar) {
9882 RExC_parse = endchar;
9884 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9887 if (in_char_class && has_multiple_chars) {
9889 RExC_parse = endbrace;
9890 vFAIL("\\N{} in character class restricted to one character");
9893 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
9897 RExC_parse = endbrace + 1;
9899 else if (! node_p || ! has_multiple_chars) {
9901 /* Here, the input is legal, but not according to the caller's
9902 * options. We fail without advancing the parse, so that the
9903 * caller can try again */
9909 /* What is done here is to convert this to a sub-pattern of the form
9910 * (?:\x{char1}\x{char2}...)
9911 * and then call reg recursively. That way, it retains its atomicness,
9912 * while not having to worry about special handling that some code
9913 * points may have. toke.c has converted the original Unicode values
9914 * to native, so that we can just pass on the hex values unchanged. We
9915 * do have to set a flag to keep recoding from happening in the
9918 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
9920 char *orig_end = RExC_end;
9923 while (RExC_parse < endbrace) {
9925 /* Convert to notation the rest of the code understands */
9926 sv_catpv(substitute_parse, "\\x{");
9927 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
9928 sv_catpv(substitute_parse, "}");
9930 /* Point to the beginning of the next character in the sequence. */
9931 RExC_parse = endchar + 1;
9932 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9934 sv_catpv(substitute_parse, ")");
9936 RExC_parse = SvPV(substitute_parse, len);
9938 /* Don't allow empty number */
9940 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9942 RExC_end = RExC_parse + len;
9944 /* The values are Unicode, and therefore not subject to recoding */
9945 RExC_override_recoding = 1;
9947 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
9948 if (flags & RESTART_UTF8) {
9949 *flagp = RESTART_UTF8;
9952 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#X",
9955 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
9957 RExC_parse = endbrace;
9958 RExC_end = orig_end;
9959 RExC_override_recoding = 0;
9961 nextchar(pRExC_state);
9971 * It returns the code point in utf8 for the value in *encp.
9972 * value: a code value in the source encoding
9973 * encp: a pointer to an Encode object
9975 * If the result from Encode is not a single character,
9976 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
9979 S_reg_recode(pTHX_ const char value, SV **encp)
9982 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
9983 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
9984 const STRLEN newlen = SvCUR(sv);
9985 UV uv = UNICODE_REPLACEMENT;
9987 PERL_ARGS_ASSERT_REG_RECODE;
9991 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
9994 if (!newlen || numlen != newlen) {
9995 uv = UNICODE_REPLACEMENT;
10001 PERL_STATIC_INLINE U8
10002 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
10006 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
10012 op = get_regex_charset(RExC_flags);
10013 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
10014 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
10015 been, so there is no hole */
10018 return op + EXACTF;
10021 PERL_STATIC_INLINE void
10022 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state, regnode *node, I32* flagp, STRLEN len, UV code_point)
10024 /* This knows the details about sizing an EXACTish node, setting flags for
10025 * it (by setting <*flagp>, and potentially populating it with a single
10028 * If <len> (the length in bytes) is non-zero, this function assumes that
10029 * the node has already been populated, and just does the sizing. In this
10030 * case <code_point> should be the final code point that has already been
10031 * placed into the node. This value will be ignored except that under some
10032 * circumstances <*flagp> is set based on it.
10034 * If <len> is zero, the function assumes that the node is to contain only
10035 * the single character given by <code_point> and calculates what <len>
10036 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10037 * additionally will populate the node's STRING with <code_point>, if <len>
10038 * is 0. In both cases <*flagp> is appropriately set
10040 * It knows that under FOLD, UTF characters and the Latin Sharp S must be
10041 * folded (the latter only when the rules indicate it can match 'ss') */
10043 bool len_passed_in = cBOOL(len != 0);
10044 U8 character[UTF8_MAXBYTES_CASE+1];
10046 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10048 if (! len_passed_in) {
10051 to_uni_fold(NATIVE_TO_UNI(code_point), character, &len);
10054 uvchr_to_utf8( character, code_point);
10055 len = UTF8SKIP(character);
10059 || code_point != LATIN_SMALL_LETTER_SHARP_S
10060 || ASCII_FOLD_RESTRICTED
10061 || ! AT_LEAST_UNI_SEMANTICS)
10063 *character = (U8) code_point;
10068 *(character + 1) = 's';
10074 RExC_size += STR_SZ(len);
10077 RExC_emit += STR_SZ(len);
10078 STR_LEN(node) = len;
10079 if (! len_passed_in) {
10080 Copy((char *) character, STRING(node), len, char);
10084 *flagp |= HASWIDTH;
10086 /* A single character node is SIMPLE, except for the special-cased SHARP S
10088 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
10089 && (code_point != LATIN_SMALL_LETTER_SHARP_S
10090 || ! FOLD || ! DEPENDS_SEMANTICS))
10097 - regatom - the lowest level
10099 Try to identify anything special at the start of the pattern. If there
10100 is, then handle it as required. This may involve generating a single regop,
10101 such as for an assertion; or it may involve recursing, such as to
10102 handle a () structure.
10104 If the string doesn't start with something special then we gobble up
10105 as much literal text as we can.
10107 Once we have been able to handle whatever type of thing started the
10108 sequence, we return.
10110 Note: we have to be careful with escapes, as they can be both literal
10111 and special, and in the case of \10 and friends, context determines which.
10113 A summary of the code structure is:
10115 switch (first_byte) {
10116 cases for each special:
10117 handle this special;
10120 switch (2nd byte) {
10121 cases for each unambiguous special:
10122 handle this special;
10124 cases for each ambigous special/literal:
10126 if (special) handle here
10128 default: // unambiguously literal:
10131 default: // is a literal char
10134 create EXACTish node for literal;
10135 while (more input and node isn't full) {
10136 switch (input_byte) {
10137 cases for each special;
10138 make sure parse pointer is set so that the next call to
10139 regatom will see this special first
10140 goto loopdone; // EXACTish node terminated by prev. char
10142 append char to EXACTISH node;
10144 get next input byte;
10148 return the generated node;
10150 Specifically there are two separate switches for handling
10151 escape sequences, with the one for handling literal escapes requiring
10152 a dummy entry for all of the special escapes that are actually handled
10155 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
10157 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10159 Otherwise does not return NULL.
10163 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10166 regnode *ret = NULL;
10168 char *parse_start = RExC_parse;
10172 GET_RE_DEBUG_FLAGS_DECL;
10174 *flagp = WORST; /* Tentatively. */
10176 DEBUG_PARSE("atom");
10178 PERL_ARGS_ASSERT_REGATOM;
10181 switch ((U8)*RExC_parse) {
10183 RExC_seen_zerolen++;
10184 nextchar(pRExC_state);
10185 if (RExC_flags & RXf_PMf_MULTILINE)
10186 ret = reg_node(pRExC_state, MBOL);
10187 else if (RExC_flags & RXf_PMf_SINGLELINE)
10188 ret = reg_node(pRExC_state, SBOL);
10190 ret = reg_node(pRExC_state, BOL);
10191 Set_Node_Length(ret, 1); /* MJD */
10194 nextchar(pRExC_state);
10196 RExC_seen_zerolen++;
10197 if (RExC_flags & RXf_PMf_MULTILINE)
10198 ret = reg_node(pRExC_state, MEOL);
10199 else if (RExC_flags & RXf_PMf_SINGLELINE)
10200 ret = reg_node(pRExC_state, SEOL);
10202 ret = reg_node(pRExC_state, EOL);
10203 Set_Node_Length(ret, 1); /* MJD */
10206 nextchar(pRExC_state);
10207 if (RExC_flags & RXf_PMf_SINGLELINE)
10208 ret = reg_node(pRExC_state, SANY);
10210 ret = reg_node(pRExC_state, REG_ANY);
10211 *flagp |= HASWIDTH|SIMPLE;
10213 Set_Node_Length(ret, 1); /* MJD */
10217 char * const oregcomp_parse = ++RExC_parse;
10218 ret = regclass(pRExC_state, flagp,depth+1,
10219 FALSE, /* means parse the whole char class */
10220 TRUE, /* allow multi-char folds */
10221 FALSE, /* don't silence non-portable warnings. */
10223 if (*RExC_parse != ']') {
10224 RExC_parse = oregcomp_parse;
10225 vFAIL("Unmatched [");
10228 if (*flagp & RESTART_UTF8)
10230 FAIL2("panic: regclass returned NULL to regatom, flags=%#X",
10233 nextchar(pRExC_state);
10234 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
10238 nextchar(pRExC_state);
10239 ret = reg(pRExC_state, 1, &flags,depth+1);
10241 if (flags & TRYAGAIN) {
10242 if (RExC_parse == RExC_end) {
10243 /* Make parent create an empty node if needed. */
10244 *flagp |= TRYAGAIN;
10249 if (flags & RESTART_UTF8) {
10250 *flagp = RESTART_UTF8;
10253 FAIL2("panic: reg returned NULL to regatom, flags=%#X", flags);
10255 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10259 if (flags & TRYAGAIN) {
10260 *flagp |= TRYAGAIN;
10263 vFAIL("Internal urp");
10264 /* Supposed to be caught earlier. */
10267 if (!regcurly(RExC_parse, FALSE)) {
10276 vFAIL("Quantifier follows nothing");
10281 This switch handles escape sequences that resolve to some kind
10282 of special regop and not to literal text. Escape sequnces that
10283 resolve to literal text are handled below in the switch marked
10286 Every entry in this switch *must* have a corresponding entry
10287 in the literal escape switch. However, the opposite is not
10288 required, as the default for this switch is to jump to the
10289 literal text handling code.
10291 switch ((U8)*++RExC_parse) {
10293 /* Special Escapes */
10295 RExC_seen_zerolen++;
10296 ret = reg_node(pRExC_state, SBOL);
10298 goto finish_meta_pat;
10300 ret = reg_node(pRExC_state, GPOS);
10301 RExC_seen |= REG_SEEN_GPOS;
10303 goto finish_meta_pat;
10305 RExC_seen_zerolen++;
10306 ret = reg_node(pRExC_state, KEEPS);
10308 /* XXX:dmq : disabling in-place substitution seems to
10309 * be necessary here to avoid cases of memory corruption, as
10310 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
10312 RExC_seen |= REG_SEEN_LOOKBEHIND;
10313 goto finish_meta_pat;
10315 ret = reg_node(pRExC_state, SEOL);
10317 RExC_seen_zerolen++; /* Do not optimize RE away */
10318 goto finish_meta_pat;
10320 ret = reg_node(pRExC_state, EOS);
10322 RExC_seen_zerolen++; /* Do not optimize RE away */
10323 goto finish_meta_pat;
10325 ret = reg_node(pRExC_state, CANY);
10326 RExC_seen |= REG_SEEN_CANY;
10327 *flagp |= HASWIDTH|SIMPLE;
10328 goto finish_meta_pat;
10330 ret = reg_node(pRExC_state, CLUMP);
10331 *flagp |= HASWIDTH;
10332 goto finish_meta_pat;
10338 arg = ANYOF_WORDCHAR;
10342 RExC_seen_zerolen++;
10343 RExC_seen |= REG_SEEN_LOOKBEHIND;
10344 op = BOUND + get_regex_charset(RExC_flags);
10345 if (op > BOUNDA) { /* /aa is same as /a */
10348 ret = reg_node(pRExC_state, op);
10349 FLAGS(ret) = get_regex_charset(RExC_flags);
10351 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10352 ckWARNdep(RExC_parse, "\"\\b{\" is deprecated; use \"\\b\\{\" or \"\\b[{]\" instead");
10354 goto finish_meta_pat;
10356 RExC_seen_zerolen++;
10357 RExC_seen |= REG_SEEN_LOOKBEHIND;
10358 op = NBOUND + get_regex_charset(RExC_flags);
10359 if (op > NBOUNDA) { /* /aa is same as /a */
10362 ret = reg_node(pRExC_state, op);
10363 FLAGS(ret) = get_regex_charset(RExC_flags);
10365 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10366 ckWARNdep(RExC_parse, "\"\\B{\" is deprecated; use \"\\B\\{\" or \"\\B[{]\" instead");
10368 goto finish_meta_pat;
10378 ret = reg_node(pRExC_state, LNBREAK);
10379 *flagp |= HASWIDTH|SIMPLE;
10380 goto finish_meta_pat;
10388 goto join_posix_op_known;
10394 arg = ANYOF_VERTWS;
10396 goto join_posix_op_known;
10406 op = POSIXD + get_regex_charset(RExC_flags);
10407 if (op > POSIXA) { /* /aa is same as /a */
10411 join_posix_op_known:
10414 op += NPOSIXD - POSIXD;
10417 ret = reg_node(pRExC_state, op);
10419 FLAGS(ret) = namedclass_to_classnum(arg);
10422 *flagp |= HASWIDTH|SIMPLE;
10426 nextchar(pRExC_state);
10427 Set_Node_Length(ret, 2); /* MJD */
10433 char* parse_start = RExC_parse - 2;
10438 ret = regclass(pRExC_state, flagp,depth+1,
10439 TRUE, /* means just parse this element */
10440 FALSE, /* don't allow multi-char folds */
10441 FALSE, /* don't silence non-portable warnings.
10442 It would be a bug if these returned
10445 /* regclass() can only return RESTART_UTF8 if multi-char folds
10448 FAIL2("panic: regclass returned NULL to regatom, flags=%#X",
10453 Set_Node_Offset(ret, parse_start + 2);
10454 Set_Node_Cur_Length(ret);
10455 nextchar(pRExC_state);
10459 /* Handle \N and \N{NAME} with multiple code points here and not
10460 * below because it can be multicharacter. join_exact() will join
10461 * them up later on. Also this makes sure that things like
10462 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
10463 * The options to the grok function call causes it to fail if the
10464 * sequence is just a single code point. We then go treat it as
10465 * just another character in the current EXACT node, and hence it
10466 * gets uniform treatment with all the other characters. The
10467 * special treatment for quantifiers is not needed for such single
10468 * character sequences */
10470 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
10471 FALSE /* not strict */ )) {
10472 if (*flagp & RESTART_UTF8)
10478 case 'k': /* Handle \k<NAME> and \k'NAME' */
10481 char ch= RExC_parse[1];
10482 if (ch != '<' && ch != '\'' && ch != '{') {
10484 vFAIL2("Sequence %.2s... not terminated",parse_start);
10486 /* this pretty much dupes the code for (?P=...) in reg(), if
10487 you change this make sure you change that */
10488 char* name_start = (RExC_parse += 2);
10490 SV *sv_dat = reg_scan_name(pRExC_state,
10491 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10492 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
10493 if (RExC_parse == name_start || *RExC_parse != ch)
10494 vFAIL2("Sequence %.3s... not terminated",parse_start);
10497 num = add_data( pRExC_state, 1, "S" );
10498 RExC_rxi->data->data[num]=(void*)sv_dat;
10499 SvREFCNT_inc_simple_void(sv_dat);
10503 ret = reganode(pRExC_state,
10506 : (ASCII_FOLD_RESTRICTED)
10508 : (AT_LEAST_UNI_SEMANTICS)
10514 *flagp |= HASWIDTH;
10516 /* override incorrect value set in reganode MJD */
10517 Set_Node_Offset(ret, parse_start+1);
10518 Set_Node_Cur_Length(ret); /* MJD */
10519 nextchar(pRExC_state);
10525 case '1': case '2': case '3': case '4':
10526 case '5': case '6': case '7': case '8': case '9':
10529 bool isg = *RExC_parse == 'g';
10534 if (*RExC_parse == '{') {
10538 if (*RExC_parse == '-') {
10542 if (hasbrace && !isDIGIT(*RExC_parse)) {
10543 if (isrel) RExC_parse--;
10545 goto parse_named_seq;
10547 num = atoi(RExC_parse);
10548 if (isg && num == 0)
10549 vFAIL("Reference to invalid group 0");
10551 num = RExC_npar - num;
10553 vFAIL("Reference to nonexistent or unclosed group");
10555 if (!isg && num > 9 && num >= RExC_npar)
10556 /* Probably a character specified in octal, e.g. \35 */
10559 char * const parse_start = RExC_parse - 1; /* MJD */
10560 while (isDIGIT(*RExC_parse))
10562 if (parse_start == RExC_parse - 1)
10563 vFAIL("Unterminated \\g... pattern");
10565 if (*RExC_parse != '}')
10566 vFAIL("Unterminated \\g{...} pattern");
10570 if (num > (I32)RExC_rx->nparens)
10571 vFAIL("Reference to nonexistent group");
10574 ret = reganode(pRExC_state,
10577 : (ASCII_FOLD_RESTRICTED)
10579 : (AT_LEAST_UNI_SEMANTICS)
10585 *flagp |= HASWIDTH;
10587 /* override incorrect value set in reganode MJD */
10588 Set_Node_Offset(ret, parse_start+1);
10589 Set_Node_Cur_Length(ret); /* MJD */
10591 nextchar(pRExC_state);
10596 if (RExC_parse >= RExC_end)
10597 FAIL("Trailing \\");
10600 /* Do not generate "unrecognized" warnings here, we fall
10601 back into the quick-grab loop below */
10608 if (RExC_flags & RXf_PMf_EXTENDED) {
10609 if ( reg_skipcomment( pRExC_state ) )
10616 parse_start = RExC_parse - 1;
10625 #define MAX_NODE_STRING_SIZE 127
10626 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
10628 U8 upper_parse = MAX_NODE_STRING_SIZE;
10631 bool next_is_quantifier;
10632 char * oldp = NULL;
10634 /* If a folding node contains only code points that don't
10635 * participate in folds, it can be changed into an EXACT node,
10636 * which allows the optimizer more things to look for */
10640 node_type = compute_EXACTish(pRExC_state);
10641 ret = reg_node(pRExC_state, node_type);
10643 /* In pass1, folded, we use a temporary buffer instead of the
10644 * actual node, as the node doesn't exist yet */
10645 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
10651 /* We do the EXACTFish to EXACT node only if folding, and not if in
10652 * locale, as whether a character folds or not isn't known until
10654 maybe_exact = FOLD && ! LOC;
10656 /* XXX The node can hold up to 255 bytes, yet this only goes to
10657 * 127. I (khw) do not know why. Keeping it somewhat less than
10658 * 255 allows us to not have to worry about overflow due to
10659 * converting to utf8 and fold expansion, but that value is
10660 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
10661 * split up by this limit into a single one using the real max of
10662 * 255. Even at 127, this breaks under rare circumstances. If
10663 * folding, we do not want to split a node at a character that is a
10664 * non-final in a multi-char fold, as an input string could just
10665 * happen to want to match across the node boundary. The join
10666 * would solve that problem if the join actually happens. But a
10667 * series of more than two nodes in a row each of 127 would cause
10668 * the first join to succeed to get to 254, but then there wouldn't
10669 * be room for the next one, which could at be one of those split
10670 * multi-char folds. I don't know of any fool-proof solution. One
10671 * could back off to end with only a code point that isn't such a
10672 * non-final, but it is possible for there not to be any in the
10674 for (p = RExC_parse - 1;
10675 len < upper_parse && p < RExC_end;
10680 if (RExC_flags & RXf_PMf_EXTENDED)
10681 p = regwhite( pRExC_state, p );
10692 /* Literal Escapes Switch
10694 This switch is meant to handle escape sequences that
10695 resolve to a literal character.
10697 Every escape sequence that represents something
10698 else, like an assertion or a char class, is handled
10699 in the switch marked 'Special Escapes' above in this
10700 routine, but also has an entry here as anything that
10701 isn't explicitly mentioned here will be treated as
10702 an unescaped equivalent literal.
10705 switch ((U8)*++p) {
10706 /* These are all the special escapes. */
10707 case 'A': /* Start assertion */
10708 case 'b': case 'B': /* Word-boundary assertion*/
10709 case 'C': /* Single char !DANGEROUS! */
10710 case 'd': case 'D': /* digit class */
10711 case 'g': case 'G': /* generic-backref, pos assertion */
10712 case 'h': case 'H': /* HORIZWS */
10713 case 'k': case 'K': /* named backref, keep marker */
10714 case 'p': case 'P': /* Unicode property */
10715 case 'R': /* LNBREAK */
10716 case 's': case 'S': /* space class */
10717 case 'v': case 'V': /* VERTWS */
10718 case 'w': case 'W': /* word class */
10719 case 'X': /* eXtended Unicode "combining character sequence" */
10720 case 'z': case 'Z': /* End of line/string assertion */
10724 /* Anything after here is an escape that resolves to a
10725 literal. (Except digits, which may or may not)
10731 case 'N': /* Handle a single-code point named character. */
10732 /* The options cause it to fail if a multiple code
10733 * point sequence. Handle those in the switch() above
10735 RExC_parse = p + 1;
10736 if (! grok_bslash_N(pRExC_state, NULL, &ender,
10737 flagp, depth, FALSE,
10738 FALSE /* not strict */ ))
10740 if (*flagp & RESTART_UTF8)
10741 FAIL("panic: grok_bslash_N set RESTART_UTF8");
10742 RExC_parse = p = oldp;
10746 if (ender > 0xff) {
10763 ender = ASCII_TO_NATIVE('\033');
10767 ender = ASCII_TO_NATIVE('\007');
10773 const char* error_msg;
10775 bool valid = grok_bslash_o(&p,
10778 TRUE, /* out warnings */
10779 FALSE, /* not strict */
10780 TRUE, /* Output warnings
10785 RExC_parse = p; /* going to die anyway; point
10786 to exact spot of failure */
10790 if (PL_encoding && ender < 0x100) {
10791 goto recode_encoding;
10793 if (ender > 0xff) {
10800 UV result = UV_MAX; /* initialize to erroneous
10802 const char* error_msg;
10804 bool valid = grok_bslash_x(&p,
10807 TRUE, /* out warnings */
10808 FALSE, /* not strict */
10809 TRUE, /* Output warnings
10814 RExC_parse = p; /* going to die anyway; point
10815 to exact spot of failure */
10820 if (PL_encoding && ender < 0x100) {
10821 goto recode_encoding;
10823 if (ender > 0xff) {
10830 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
10832 case '0': case '1': case '2': case '3':case '4':
10833 case '5': case '6': case '7':
10835 (isDIGIT(p[1]) && atoi(p) >= RExC_npar))
10837 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
10839 ender = grok_oct(p, &numlen, &flags, NULL);
10840 if (ender > 0xff) {
10844 if (SIZE_ONLY /* like \08, \178 */
10847 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
10849 reg_warn_non_literal_string(
10851 form_short_octal_warning(p, numlen));
10854 else { /* Not to be treated as an octal constant, go
10859 if (PL_encoding && ender < 0x100)
10860 goto recode_encoding;
10863 if (! RExC_override_recoding) {
10864 SV* enc = PL_encoding;
10865 ender = reg_recode((const char)(U8)ender, &enc);
10866 if (!enc && SIZE_ONLY)
10867 ckWARNreg(p, "Invalid escape in the specified encoding");
10873 FAIL("Trailing \\");
10876 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
10877 /* Include any { following the alpha to emphasize
10878 * that it could be part of an escape at some point
10880 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
10881 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
10883 goto normal_default;
10884 } /* End of switch on '\' */
10886 default: /* A literal character */
10889 && RExC_flags & RXf_PMf_EXTENDED
10890 && ckWARN(WARN_DEPRECATED)
10891 && is_PATWS_non_low(p, UTF))
10893 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
10894 "Escape literal pattern white space under /x");
10898 if (UTF8_IS_START(*p) && UTF) {
10900 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
10901 &numlen, UTF8_ALLOW_DEFAULT);
10907 } /* End of switch on the literal */
10909 /* Here, have looked at the literal character and <ender>
10910 * contains its ordinal, <p> points to the character after it
10913 if ( RExC_flags & RXf_PMf_EXTENDED)
10914 p = regwhite( pRExC_state, p );
10916 /* If the next thing is a quantifier, it applies to this
10917 * character only, which means that this character has to be in
10918 * its own node and can't just be appended to the string in an
10919 * existing node, so if there are already other characters in
10920 * the node, close the node with just them, and set up to do
10921 * this character again next time through, when it will be the
10922 * only thing in its new node */
10923 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
10931 /* See comments for join_exact() as to why we fold
10932 * this non-UTF at compile time */
10933 || (node_type == EXACTFU
10934 && ender == LATIN_SMALL_LETTER_SHARP_S))
10938 /* Prime the casefolded buffer. Locale rules, which
10939 * apply only to code points < 256, aren't known until
10940 * execution, so for them, just output the original
10941 * character using utf8. If we start to fold non-UTF
10942 * patterns, be sure to update join_exact() */
10943 if (LOC && ender < 256) {
10944 if (UNI_IS_INVARIANT(ender)) {
10948 *s = UTF8_TWO_BYTE_HI(ender);
10949 *(s + 1) = UTF8_TWO_BYTE_LO(ender);
10954 UV folded = _to_uni_fold_flags(
10959 | ((LOC) ? FOLD_FLAGS_LOCALE
10960 : (ASCII_FOLD_RESTRICTED)
10961 ? FOLD_FLAGS_NOMIX_ASCII
10965 /* If this node only contains non-folding code
10966 * points so far, see if this new one is also
10969 if (folded != ender) {
10970 maybe_exact = FALSE;
10973 /* Here the fold is the original; we have
10974 * to check further to see if anything
10976 if (! PL_utf8_foldable) {
10977 SV* swash = swash_init("utf8",
10979 &PL_sv_undef, 1, 0);
10981 _get_swash_invlist(swash);
10982 SvREFCNT_dec_NN(swash);
10984 if (_invlist_contains_cp(PL_utf8_foldable,
10987 maybe_exact = FALSE;
10995 /* The loop increments <len> each time, as all but this
10996 * path (and the one just below for UTF) through it add
10997 * a single byte to the EXACTish node. But this one
10998 * has changed len to be the correct final value, so
10999 * subtract one to cancel out the increment that
11001 len += foldlen - 1;
11004 *(s++) = (char) ender;
11005 maybe_exact &= ! IS_IN_SOME_FOLD_L1(ender);
11009 const STRLEN unilen = reguni(pRExC_state, ender, s);
11015 /* See comment just above for - 1 */
11019 REGC((char)ender, s++);
11022 if (next_is_quantifier) {
11024 /* Here, the next input is a quantifier, and to get here,
11025 * the current character is the only one in the node.
11026 * Also, here <len> doesn't include the final byte for this
11032 } /* End of loop through literal characters */
11034 /* Here we have either exhausted the input or ran out of room in
11035 * the node. (If we encountered a character that can't be in the
11036 * node, transfer is made directly to <loopdone>, and so we
11037 * wouldn't have fallen off the end of the loop.) In the latter
11038 * case, we artificially have to split the node into two, because
11039 * we just don't have enough space to hold everything. This
11040 * creates a problem if the final character participates in a
11041 * multi-character fold in the non-final position, as a match that
11042 * should have occurred won't, due to the way nodes are matched,
11043 * and our artificial boundary. So back off until we find a non-
11044 * problematic character -- one that isn't at the beginning or
11045 * middle of such a fold. (Either it doesn't participate in any
11046 * folds, or appears only in the final position of all the folds it
11047 * does participate in.) A better solution with far fewer false
11048 * positives, and that would fill the nodes more completely, would
11049 * be to actually have available all the multi-character folds to
11050 * test against, and to back-off only far enough to be sure that
11051 * this node isn't ending with a partial one. <upper_parse> is set
11052 * further below (if we need to reparse the node) to include just
11053 * up through that final non-problematic character that this code
11054 * identifies, so when it is set to less than the full node, we can
11055 * skip the rest of this */
11056 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
11058 const STRLEN full_len = len;
11060 assert(len >= MAX_NODE_STRING_SIZE);
11062 /* Here, <s> points to the final byte of the final character.
11063 * Look backwards through the string until find a non-
11064 * problematic character */
11068 /* These two have no multi-char folds to non-UTF characters
11070 if (ASCII_FOLD_RESTRICTED || LOC) {
11074 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
11078 if (! PL_NonL1NonFinalFold) {
11079 PL_NonL1NonFinalFold = _new_invlist_C_array(
11080 NonL1_Perl_Non_Final_Folds_invlist);
11083 /* Point to the first byte of the final character */
11084 s = (char *) utf8_hop((U8 *) s, -1);
11086 while (s >= s0) { /* Search backwards until find
11087 non-problematic char */
11088 if (UTF8_IS_INVARIANT(*s)) {
11090 /* There are no ascii characters that participate
11091 * in multi-char folds under /aa. In EBCDIC, the
11092 * non-ascii invariants are all control characters,
11093 * so don't ever participate in any folds. */
11094 if (ASCII_FOLD_RESTRICTED
11095 || ! IS_NON_FINAL_FOLD(*s))
11100 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
11102 /* No Latin1 characters participate in multi-char
11103 * folds under /l */
11105 || ! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_UNI(
11111 else if (! _invlist_contains_cp(
11112 PL_NonL1NonFinalFold,
11113 valid_utf8_to_uvchr((U8 *) s, NULL)))
11118 /* Here, the current character is problematic in that
11119 * it does occur in the non-final position of some
11120 * fold, so try the character before it, but have to
11121 * special case the very first byte in the string, so
11122 * we don't read outside the string */
11123 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
11124 } /* End of loop backwards through the string */
11126 /* If there were only problematic characters in the string,
11127 * <s> will point to before s0, in which case the length
11128 * should be 0, otherwise include the length of the
11129 * non-problematic character just found */
11130 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
11133 /* Here, have found the final character, if any, that is
11134 * non-problematic as far as ending the node without splitting
11135 * it across a potential multi-char fold. <len> contains the
11136 * number of bytes in the node up-to and including that
11137 * character, or is 0 if there is no such character, meaning
11138 * the whole node contains only problematic characters. In
11139 * this case, give up and just take the node as-is. We can't
11145 /* Here, the node does contain some characters that aren't
11146 * problematic. If one such is the final character in the
11147 * node, we are done */
11148 if (len == full_len) {
11151 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
11153 /* If the final character is problematic, but the
11154 * penultimate is not, back-off that last character to
11155 * later start a new node with it */
11160 /* Here, the final non-problematic character is earlier
11161 * in the input than the penultimate character. What we do
11162 * is reparse from the beginning, going up only as far as
11163 * this final ok one, thus guaranteeing that the node ends
11164 * in an acceptable character. The reason we reparse is
11165 * that we know how far in the character is, but we don't
11166 * know how to correlate its position with the input parse.
11167 * An alternate implementation would be to build that
11168 * correlation as we go along during the original parse,
11169 * but that would entail extra work for every node, whereas
11170 * this code gets executed only when the string is too
11171 * large for the node, and the final two characters are
11172 * problematic, an infrequent occurrence. Yet another
11173 * possible strategy would be to save the tail of the
11174 * string, and the next time regatom is called, initialize
11175 * with that. The problem with this is that unless you
11176 * back off one more character, you won't be guaranteed
11177 * regatom will get called again, unless regbranch,
11178 * regpiece ... are also changed. If you do back off that
11179 * extra character, so that there is input guaranteed to
11180 * force calling regatom, you can't handle the case where
11181 * just the first character in the node is acceptable. I
11182 * (khw) decided to try this method which doesn't have that
11183 * pitfall; if performance issues are found, we can do a
11184 * combination of the current approach plus that one */
11190 } /* End of verifying node ends with an appropriate char */
11192 loopdone: /* Jumped to when encounters something that shouldn't be in
11195 /* If 'maybe_exact' is still set here, means there are no
11196 * code points in the node that participate in folds */
11197 if (FOLD && maybe_exact) {
11201 /* I (khw) don't know if you can get here with zero length, but the
11202 * old code handled this situation by creating a zero-length EXACT
11203 * node. Might as well be NOTHING instead */
11208 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
11211 RExC_parse = p - 1;
11212 Set_Node_Cur_Length(ret); /* MJD */
11213 nextchar(pRExC_state);
11215 /* len is STRLEN which is unsigned, need to copy to signed */
11218 vFAIL("Internal disaster");
11221 } /* End of label 'defchar:' */
11223 } /* End of giant switch on input character */
11229 S_regwhite( RExC_state_t *pRExC_state, char *p )
11231 const char *e = RExC_end;
11233 PERL_ARGS_ASSERT_REGWHITE;
11238 else if (*p == '#') {
11241 if (*p++ == '\n') {
11247 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11256 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
11258 /* Returns the next non-pattern-white space, non-comment character (the
11259 * latter only if 'recognize_comment is true) in the string p, which is
11260 * ended by RExC_end. If there is no line break ending a comment,
11261 * RExC_seen has added the REG_SEEN_RUN_ON_COMMENT flag; */
11262 const char *e = RExC_end;
11264 PERL_ARGS_ASSERT_REGPATWS;
11268 if ((len = is_PATWS_safe(p, e, UTF))) {
11271 else if (recognize_comment && *p == '#') {
11275 if (is_LNBREAK_safe(p, e, UTF)) {
11281 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11289 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
11290 Character classes ([:foo:]) can also be negated ([:^foo:]).
11291 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
11292 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
11293 but trigger failures because they are currently unimplemented. */
11295 #define POSIXCC_DONE(c) ((c) == ':')
11296 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
11297 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
11299 PERL_STATIC_INLINE I32
11300 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
11303 I32 namedclass = OOB_NAMEDCLASS;
11305 PERL_ARGS_ASSERT_REGPPOSIXCC;
11307 if (value == '[' && RExC_parse + 1 < RExC_end &&
11308 /* I smell either [: or [= or [. -- POSIX has been here, right? */
11309 POSIXCC(UCHARAT(RExC_parse)))
11311 const char c = UCHARAT(RExC_parse);
11312 char* const s = RExC_parse++;
11314 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
11316 if (RExC_parse == RExC_end) {
11319 /* Try to give a better location for the error (than the end of
11320 * the string) by looking for the matching ']' */
11322 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
11325 vFAIL2("Unmatched '%c' in POSIX class", c);
11327 /* Grandfather lone [:, [=, [. */
11331 const char* const t = RExC_parse++; /* skip over the c */
11334 if (UCHARAT(RExC_parse) == ']') {
11335 const char *posixcc = s + 1;
11336 RExC_parse++; /* skip over the ending ] */
11339 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
11340 const I32 skip = t - posixcc;
11342 /* Initially switch on the length of the name. */
11345 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
11346 this is the Perl \w
11348 namedclass = ANYOF_WORDCHAR;
11351 /* Names all of length 5. */
11352 /* alnum alpha ascii blank cntrl digit graph lower
11353 print punct space upper */
11354 /* Offset 4 gives the best switch position. */
11355 switch (posixcc[4]) {
11357 if (memEQ(posixcc, "alph", 4)) /* alpha */
11358 namedclass = ANYOF_ALPHA;
11361 if (memEQ(posixcc, "spac", 4)) /* space */
11362 namedclass = ANYOF_PSXSPC;
11365 if (memEQ(posixcc, "grap", 4)) /* graph */
11366 namedclass = ANYOF_GRAPH;
11369 if (memEQ(posixcc, "asci", 4)) /* ascii */
11370 namedclass = ANYOF_ASCII;
11373 if (memEQ(posixcc, "blan", 4)) /* blank */
11374 namedclass = ANYOF_BLANK;
11377 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
11378 namedclass = ANYOF_CNTRL;
11381 if (memEQ(posixcc, "alnu", 4)) /* alnum */
11382 namedclass = ANYOF_ALPHANUMERIC;
11385 if (memEQ(posixcc, "lowe", 4)) /* lower */
11386 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
11387 else if (memEQ(posixcc, "uppe", 4)) /* upper */
11388 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
11391 if (memEQ(posixcc, "digi", 4)) /* digit */
11392 namedclass = ANYOF_DIGIT;
11393 else if (memEQ(posixcc, "prin", 4)) /* print */
11394 namedclass = ANYOF_PRINT;
11395 else if (memEQ(posixcc, "punc", 4)) /* punct */
11396 namedclass = ANYOF_PUNCT;
11401 if (memEQ(posixcc, "xdigit", 6))
11402 namedclass = ANYOF_XDIGIT;
11406 if (namedclass == OOB_NAMEDCLASS)
11407 Simple_vFAIL3("POSIX class [:%.*s:] unknown",
11410 /* The #defines are structured so each complement is +1 to
11411 * the normal one */
11415 assert (posixcc[skip] == ':');
11416 assert (posixcc[skip+1] == ']');
11417 } else if (!SIZE_ONLY) {
11418 /* [[=foo=]] and [[.foo.]] are still future. */
11420 /* adjust RExC_parse so the warning shows after
11421 the class closes */
11422 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
11424 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11427 /* Maternal grandfather:
11428 * "[:" ending in ":" but not in ":]" */
11430 vFAIL("Unmatched '[' in POSIX class");
11433 /* Grandfather lone [:, [=, [. */
11443 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
11445 /* This applies some heuristics at the current parse position (which should
11446 * be at a '[') to see if what follows might be intended to be a [:posix:]
11447 * class. It returns true if it really is a posix class, of course, but it
11448 * also can return true if it thinks that what was intended was a posix
11449 * class that didn't quite make it.
11451 * It will return true for
11453 * [:alphanumerics] (as long as the ] isn't followed immediately by a
11454 * ')' indicating the end of the (?[
11455 * [:any garbage including %^&$ punctuation:]
11457 * This is designed to be called only from S_handle_regex_sets; it could be
11458 * easily adapted to be called from the spot at the beginning of regclass()
11459 * that checks to see in a normal bracketed class if the surrounding []
11460 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
11461 * change long-standing behavior, so I (khw) didn't do that */
11462 char* p = RExC_parse + 1;
11463 char first_char = *p;
11465 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
11467 assert(*(p - 1) == '[');
11469 if (! POSIXCC(first_char)) {
11474 while (p < RExC_end && isWORDCHAR(*p)) p++;
11476 if (p >= RExC_end) {
11480 if (p - RExC_parse > 2 /* Got at least 1 word character */
11481 && (*p == first_char
11482 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
11487 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
11490 && p - RExC_parse > 2 /* [:] evaluates to colon;
11491 [::] is a bad posix class. */
11492 && first_char == *(p - 1));
11496 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist, I32 *flagp, U32 depth,
11497 char * const oregcomp_parse)
11499 /* Handle the (?[...]) construct to do set operations */
11502 UV start, end; /* End points of code point ranges */
11504 char *save_end, *save_parse;
11509 const bool save_fold = FOLD;
11511 GET_RE_DEBUG_FLAGS_DECL;
11513 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
11516 vFAIL("(?[...]) not valid in locale");
11518 RExC_uni_semantics = 1;
11520 /* This will return only an ANYOF regnode, or (unlikely) something smaller
11521 * (such as EXACT). Thus we can skip most everything if just sizing. We
11522 * call regclass to handle '[]' so as to not have to reinvent its parsing
11523 * rules here (throwing away the size it computes each time). And, we exit
11524 * upon an unescaped ']' that isn't one ending a regclass. To do both
11525 * these things, we need to realize that something preceded by a backslash
11526 * is escaped, so we have to keep track of backslashes */
11529 Perl_ck_warner_d(aTHX_
11530 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
11531 "The regex_sets feature is experimental" REPORT_LOCATION,
11532 (int) (RExC_parse - RExC_precomp) , RExC_precomp, RExC_parse);
11534 while (RExC_parse < RExC_end) {
11535 SV* current = NULL;
11536 RExC_parse = regpatws(pRExC_state, RExC_parse,
11537 TRUE); /* means recognize comments */
11538 switch (*RExC_parse) {
11542 /* Skip the next byte (which could cause us to end up in
11543 * the middle of a UTF-8 character, but since none of those
11544 * are confusable with anything we currently handle in this
11545 * switch (invariants all), it's safe. We'll just hit the
11546 * default: case next time and keep on incrementing until
11547 * we find one of the invariants we do handle. */
11552 /* If this looks like it is a [:posix:] class, leave the
11553 * parse pointer at the '[' to fool regclass() into
11554 * thinking it is part of a '[[:posix:]]'. That function
11555 * will use strict checking to force a syntax error if it
11556 * doesn't work out to a legitimate class */
11557 bool is_posix_class
11558 = could_it_be_a_POSIX_class(pRExC_state);
11559 if (! is_posix_class) {
11563 /* regclass() can only return RESTART_UTF8 if multi-char
11564 folds are allowed. */
11565 if (!regclass(pRExC_state, flagp,depth+1,
11566 is_posix_class, /* parse the whole char
11567 class only if not a
11569 FALSE, /* don't allow multi-char folds */
11570 TRUE, /* silence non-portable warnings. */
11572 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X",
11575 /* function call leaves parse pointing to the ']', except
11576 * if we faked it */
11577 if (is_posix_class) {
11581 SvREFCNT_dec(current); /* In case it returned something */
11587 if (RExC_parse < RExC_end
11588 && *RExC_parse == ')')
11590 node = reganode(pRExC_state, ANYOF, 0);
11591 RExC_size += ANYOF_SKIP;
11592 nextchar(pRExC_state);
11593 Set_Node_Length(node,
11594 RExC_parse - oregcomp_parse + 1); /* MJD */
11603 FAIL("Syntax error in (?[...])");
11606 /* Pass 2 only after this. Everything in this construct is a
11607 * metacharacter. Operands begin with either a '\' (for an escape
11608 * sequence), or a '[' for a bracketed character class. Any other
11609 * character should be an operator, or parenthesis for grouping. Both
11610 * types of operands are handled by calling regclass() to parse them. It
11611 * is called with a parameter to indicate to return the computed inversion
11612 * list. The parsing here is implemented via a stack. Each entry on the
11613 * stack is a single character representing one of the operators, or the
11614 * '('; or else a pointer to an operand inversion list. */
11616 #define IS_OPERAND(a) (! SvIOK(a))
11618 /* The stack starts empty. It is a syntax error if the first thing parsed
11619 * is a binary operator; everything else is pushed on the stack. When an
11620 * operand is parsed, the top of the stack is examined. If it is a binary
11621 * operator, the item before it should be an operand, and both are replaced
11622 * by the result of doing that operation on the new operand and the one on
11623 * the stack. Thus a sequence of binary operands is reduced to a single
11624 * one before the next one is parsed.
11626 * A unary operator may immediately follow a binary in the input, for
11629 * When an operand is parsed and the top of the stack is a unary operator,
11630 * the operation is performed, and then the stack is rechecked to see if
11631 * this new operand is part of a binary operation; if so, it is handled as
11634 * A '(' is simply pushed on the stack; it is valid only if the stack is
11635 * empty, or the top element of the stack is an operator or another '('
11636 * (for which the parenthesized expression will become an operand). By the
11637 * time the corresponding ')' is parsed everything in between should have
11638 * been parsed and evaluated to a single operand (or else is a syntax
11639 * error), and is handled as a regular operand */
11643 while (RExC_parse < RExC_end) {
11644 I32 top_index = av_tindex(stack);
11646 SV* current = NULL;
11648 /* Skip white space */
11649 RExC_parse = regpatws(pRExC_state, RExC_parse,
11650 TRUE); /* means recognize comments */
11651 if (RExC_parse >= RExC_end) {
11652 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
11654 if ((curchar = UCHARAT(RExC_parse)) == ']') {
11661 if (av_tindex(stack) >= 0 /* This makes sure that we can
11662 safely subtract 1 from
11663 RExC_parse in the next clause.
11664 If we have something on the
11665 stack, we have parsed something
11667 && UCHARAT(RExC_parse - 1) == '('
11668 && RExC_parse < RExC_end)
11670 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
11671 * This happens when we have some thing like
11673 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
11675 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
11677 * Here we would be handling the interpolated
11678 * '$thai_or_lao'. We handle this by a recursive call to
11679 * ourselves which returns the inversion list the
11680 * interpolated expression evaluates to. We use the flags
11681 * from the interpolated pattern. */
11682 U32 save_flags = RExC_flags;
11683 const char * const save_parse = ++RExC_parse;
11685 parse_lparen_question_flags(pRExC_state);
11687 if (RExC_parse == save_parse /* Makes sure there was at
11688 least one flag (or this
11689 embedding wasn't compiled)
11691 || RExC_parse >= RExC_end - 4
11692 || UCHARAT(RExC_parse) != ':'
11693 || UCHARAT(++RExC_parse) != '('
11694 || UCHARAT(++RExC_parse) != '?'
11695 || UCHARAT(++RExC_parse) != '[')
11698 /* In combination with the above, this moves the
11699 * pointer to the point just after the first erroneous
11700 * character (or if there are no flags, to where they
11701 * should have been) */
11702 if (RExC_parse >= RExC_end - 4) {
11703 RExC_parse = RExC_end;
11705 else if (RExC_parse != save_parse) {
11706 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11708 vFAIL("Expecting '(?flags:(?[...'");
11711 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
11712 depth+1, oregcomp_parse);
11714 /* Here, 'current' contains the embedded expression's
11715 * inversion list, and RExC_parse points to the trailing
11716 * ']'; the next character should be the ')' which will be
11717 * paired with the '(' that has been put on the stack, so
11718 * the whole embedded expression reduces to '(operand)' */
11721 RExC_flags = save_flags;
11722 goto handle_operand;
11727 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11728 vFAIL("Unexpected character");
11731 /* regclass() can only return RESTART_UTF8 if multi-char
11732 folds are allowed. */
11733 if (!regclass(pRExC_state, flagp,depth+1,
11734 TRUE, /* means parse just the next thing */
11735 FALSE, /* don't allow multi-char folds */
11736 FALSE, /* don't silence non-portable warnings. */
11738 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X",
11740 /* regclass() will return with parsing just the \ sequence,
11741 * leaving the parse pointer at the next thing to parse */
11743 goto handle_operand;
11745 case '[': /* Is a bracketed character class */
11747 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
11749 if (! is_posix_class) {
11753 /* regclass() can only return RESTART_UTF8 if multi-char
11754 folds are allowed. */
11755 if(!regclass(pRExC_state, flagp,depth+1,
11756 is_posix_class, /* parse the whole char class
11757 only if not a posix class */
11758 FALSE, /* don't allow multi-char folds */
11759 FALSE, /* don't silence non-portable warnings. */
11761 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X",
11763 /* function call leaves parse pointing to the ']', except if we
11765 if (is_posix_class) {
11769 goto handle_operand;
11778 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
11779 || ! IS_OPERAND(*top_ptr))
11782 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
11784 av_push(stack, newSVuv(curchar));
11788 av_push(stack, newSVuv(curchar));
11792 if (top_index >= 0) {
11793 top_ptr = av_fetch(stack, top_index, FALSE);
11795 if (IS_OPERAND(*top_ptr)) {
11797 vFAIL("Unexpected '(' with no preceding operator");
11800 av_push(stack, newSVuv(curchar));
11807 || ! (current = av_pop(stack))
11808 || ! IS_OPERAND(current)
11809 || ! (lparen = av_pop(stack))
11810 || IS_OPERAND(lparen)
11811 || SvUV(lparen) != '(')
11814 vFAIL("Unexpected ')'");
11817 SvREFCNT_dec_NN(lparen);
11824 /* Here, we have an operand to process, in 'current' */
11826 if (top_index < 0) { /* Just push if stack is empty */
11827 av_push(stack, current);
11830 SV* top = av_pop(stack);
11831 char current_operator;
11833 if (IS_OPERAND(top)) {
11834 vFAIL("Operand with no preceding operator");
11836 current_operator = (char) SvUV(top);
11837 switch (current_operator) {
11838 case '(': /* Push the '(' back on followed by the new
11840 av_push(stack, top);
11841 av_push(stack, current);
11842 SvREFCNT_inc(top); /* Counters the '_dec' done
11843 just after the 'break', so
11844 it doesn't get wrongly freed
11849 _invlist_invert(current);
11851 /* Unlike binary operators, the top of the stack,
11852 * now that this unary one has been popped off, may
11853 * legally be an operator, and we now have operand
11856 SvREFCNT_dec_NN(top);
11857 goto handle_operand;
11860 _invlist_intersection(av_pop(stack),
11863 av_push(stack, current);
11868 _invlist_union(av_pop(stack), current, ¤t);
11869 av_push(stack, current);
11873 _invlist_subtract(av_pop(stack), current, ¤t);
11874 av_push(stack, current);
11877 case '^': /* The union minus the intersection */
11883 element = av_pop(stack);
11884 _invlist_union(element, current, &u);
11885 _invlist_intersection(element, current, &i);
11886 _invlist_subtract(u, i, ¤t);
11887 av_push(stack, current);
11888 SvREFCNT_dec_NN(i);
11889 SvREFCNT_dec_NN(u);
11890 SvREFCNT_dec_NN(element);
11895 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
11897 SvREFCNT_dec_NN(top);
11901 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11904 if (av_tindex(stack) < 0 /* Was empty */
11905 || ((final = av_pop(stack)) == NULL)
11906 || ! IS_OPERAND(final)
11907 || av_tindex(stack) >= 0) /* More left on stack */
11909 vFAIL("Incomplete expression within '(?[ ])'");
11912 /* Here, 'final' is the resultant inversion list from evaluating the
11913 * expression. Return it if so requested */
11914 if (return_invlist) {
11915 *return_invlist = final;
11919 /* Otherwise generate a resultant node, based on 'final'. regclass() is
11920 * expecting a string of ranges and individual code points */
11921 invlist_iterinit(final);
11922 result_string = newSVpvs("");
11923 while (invlist_iternext(final, &start, &end)) {
11924 if (start == end) {
11925 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
11928 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
11933 save_parse = RExC_parse;
11934 RExC_parse = SvPV(result_string, len);
11935 save_end = RExC_end;
11936 RExC_end = RExC_parse + len;
11938 /* We turn off folding around the call, as the class we have constructed
11939 * already has all folding taken into consideration, and we don't want
11940 * regclass() to add to that */
11941 RExC_flags &= ~RXf_PMf_FOLD;
11942 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
11944 node = regclass(pRExC_state, flagp,depth+1,
11945 FALSE, /* means parse the whole char class */
11946 FALSE, /* don't allow multi-char folds */
11947 TRUE, /* silence non-portable warnings. The above may very
11948 well have generated non-portable code points, but
11949 they're valid on this machine */
11952 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X", flagp);
11954 RExC_flags |= RXf_PMf_FOLD;
11956 RExC_parse = save_parse + 1;
11957 RExC_end = save_end;
11958 SvREFCNT_dec_NN(final);
11959 SvREFCNT_dec_NN(result_string);
11960 SvREFCNT_dec_NN(stack);
11962 nextchar(pRExC_state);
11963 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
11968 /* The names of properties whose definitions are not known at compile time are
11969 * stored in this SV, after a constant heading. So if the length has been
11970 * changed since initialization, then there is a run-time definition. */
11971 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION (SvCUR(listsv) != initial_listsv_len)
11974 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
11975 const bool stop_at_1, /* Just parse the next thing, don't
11976 look for a full character class */
11977 bool allow_multi_folds,
11978 const bool silence_non_portable, /* Don't output warnings
11981 SV** ret_invlist) /* Return an inversion list, not a node */
11983 /* parse a bracketed class specification. Most of these will produce an
11984 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
11985 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
11986 * under /i with multi-character folds: it will be rewritten following the
11987 * paradigm of this example, where the <multi-fold>s are characters which
11988 * fold to multiple character sequences:
11989 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
11990 * gets effectively rewritten as:
11991 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
11992 * reg() gets called (recursively) on the rewritten version, and this
11993 * function will return what it constructs. (Actually the <multi-fold>s
11994 * aren't physically removed from the [abcdefghi], it's just that they are
11995 * ignored in the recursion by means of a flag:
11996 * <RExC_in_multi_char_class>.)
11998 * ANYOF nodes contain a bit map for the first 256 characters, with the
11999 * corresponding bit set if that character is in the list. For characters
12000 * above 255, a range list or swash is used. There are extra bits for \w,
12001 * etc. in locale ANYOFs, as what these match is not determinable at
12004 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
12005 * to be restarted. This can only happen if ret_invlist is non-NULL.
12009 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
12011 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
12014 IV namedclass = OOB_NAMEDCLASS;
12015 char *rangebegin = NULL;
12016 bool need_class = 0;
12018 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
12019 than just initialized. */
12020 SV* properties = NULL; /* Code points that match \p{} \P{} */
12021 SV* posixes = NULL; /* Code points that match classes like, [:word:],
12022 extended beyond the Latin1 range */
12023 UV element_count = 0; /* Number of distinct elements in the class.
12024 Optimizations may be possible if this is tiny */
12025 AV * multi_char_matches = NULL; /* Code points that fold to more than one
12026 character; used under /i */
12028 char * stop_ptr = RExC_end; /* where to stop parsing */
12029 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
12031 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
12033 /* Unicode properties are stored in a swash; this holds the current one
12034 * being parsed. If this swash is the only above-latin1 component of the
12035 * character class, an optimization is to pass it directly on to the
12036 * execution engine. Otherwise, it is set to NULL to indicate that there
12037 * are other things in the class that have to be dealt with at execution
12039 SV* swash = NULL; /* Code points that match \p{} \P{} */
12041 /* Set if a component of this character class is user-defined; just passed
12042 * on to the engine */
12043 bool has_user_defined_property = FALSE;
12045 /* inversion list of code points this node matches only when the target
12046 * string is in UTF-8. (Because is under /d) */
12047 SV* depends_list = NULL;
12049 /* inversion list of code points this node matches. For much of the
12050 * function, it includes only those that match regardless of the utf8ness
12051 * of the target string */
12052 SV* cp_list = NULL;
12055 /* In a range, counts how many 0-2 of the ends of it came from literals,
12056 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
12057 UV literal_endpoint = 0;
12059 bool invert = FALSE; /* Is this class to be complemented */
12061 /* Is there any thing like \W or [:^digit:] that matches above the legal
12062 * Unicode range? */
12063 bool runtime_posix_matches_above_Unicode = FALSE;
12065 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
12066 case we need to change the emitted regop to an EXACT. */
12067 const char * orig_parse = RExC_parse;
12068 const I32 orig_size = RExC_size;
12069 GET_RE_DEBUG_FLAGS_DECL;
12071 PERL_ARGS_ASSERT_REGCLASS;
12073 PERL_UNUSED_ARG(depth);
12076 DEBUG_PARSE("clas");
12078 /* Assume we are going to generate an ANYOF node. */
12079 ret = reganode(pRExC_state, ANYOF, 0);
12082 RExC_size += ANYOF_SKIP;
12083 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
12086 ANYOF_FLAGS(ret) = 0;
12088 RExC_emit += ANYOF_SKIP;
12090 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
12092 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
12093 initial_listsv_len = SvCUR(listsv);
12094 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
12098 RExC_parse = regpatws(pRExC_state, RExC_parse,
12099 FALSE /* means don't recognize comments */);
12102 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
12105 allow_multi_folds = FALSE;
12108 RExC_parse = regpatws(pRExC_state, RExC_parse,
12109 FALSE /* means don't recognize comments */);
12113 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
12114 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
12115 const char *s = RExC_parse;
12116 const char c = *s++;
12118 while (isWORDCHAR(*s))
12120 if (*s && c == *s && s[1] == ']') {
12121 SAVEFREESV(RExC_rx_sv);
12123 "POSIX syntax [%c %c] belongs inside character classes",
12125 (void)ReREFCNT_inc(RExC_rx_sv);
12129 /* If the caller wants us to just parse a single element, accomplish this
12130 * by faking the loop ending condition */
12131 if (stop_at_1 && RExC_end > RExC_parse) {
12132 stop_ptr = RExC_parse + 1;
12135 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
12136 if (UCHARAT(RExC_parse) == ']')
12137 goto charclassloop;
12141 if (RExC_parse >= stop_ptr) {
12146 RExC_parse = regpatws(pRExC_state, RExC_parse,
12147 FALSE /* means don't recognize comments */);
12150 if (UCHARAT(RExC_parse) == ']') {
12156 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
12157 save_value = value;
12158 save_prevvalue = prevvalue;
12161 rangebegin = RExC_parse;
12165 value = utf8n_to_uvchr((U8*)RExC_parse,
12166 RExC_end - RExC_parse,
12167 &numlen, UTF8_ALLOW_DEFAULT);
12168 RExC_parse += numlen;
12171 value = UCHARAT(RExC_parse++);
12174 && RExC_parse < RExC_end
12175 && POSIXCC(UCHARAT(RExC_parse)))
12177 namedclass = regpposixcc(pRExC_state, value, strict);
12179 else if (value == '\\') {
12181 value = utf8n_to_uvchr((U8*)RExC_parse,
12182 RExC_end - RExC_parse,
12183 &numlen, UTF8_ALLOW_DEFAULT);
12184 RExC_parse += numlen;
12187 value = UCHARAT(RExC_parse++);
12189 /* Some compilers cannot handle switching on 64-bit integer
12190 * values, therefore value cannot be an UV. Yes, this will
12191 * be a problem later if we want switch on Unicode.
12192 * A similar issue a little bit later when switching on
12193 * namedclass. --jhi */
12195 /* If the \ is escaping white space when white space is being
12196 * skipped, it means that that white space is wanted literally, and
12197 * is already in 'value'. Otherwise, need to translate the escape
12198 * into what it signifies. */
12199 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
12201 case 'w': namedclass = ANYOF_WORDCHAR; break;
12202 case 'W': namedclass = ANYOF_NWORDCHAR; break;
12203 case 's': namedclass = ANYOF_SPACE; break;
12204 case 'S': namedclass = ANYOF_NSPACE; break;
12205 case 'd': namedclass = ANYOF_DIGIT; break;
12206 case 'D': namedclass = ANYOF_NDIGIT; break;
12207 case 'v': namedclass = ANYOF_VERTWS; break;
12208 case 'V': namedclass = ANYOF_NVERTWS; break;
12209 case 'h': namedclass = ANYOF_HORIZWS; break;
12210 case 'H': namedclass = ANYOF_NHORIZWS; break;
12211 case 'N': /* Handle \N{NAME} in class */
12213 /* We only pay attention to the first char of
12214 multichar strings being returned. I kinda wonder
12215 if this makes sense as it does change the behaviour
12216 from earlier versions, OTOH that behaviour was broken
12218 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
12219 TRUE, /* => charclass */
12222 if (*flagp & RESTART_UTF8)
12223 FAIL("panic: grok_bslash_N set RESTART_UTF8");
12233 /* We will handle any undefined properties ourselves */
12234 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF;
12236 if (RExC_parse >= RExC_end)
12237 vFAIL2("Empty \\%c{}", (U8)value);
12238 if (*RExC_parse == '{') {
12239 const U8 c = (U8)value;
12240 e = strchr(RExC_parse++, '}');
12242 vFAIL2("Missing right brace on \\%c{}", c);
12243 while (isSPACE(UCHARAT(RExC_parse)))
12245 if (e == RExC_parse)
12246 vFAIL2("Empty \\%c{}", c);
12247 n = e - RExC_parse;
12248 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
12259 if (UCHARAT(RExC_parse) == '^') {
12262 /* toggle. (The rhs xor gets the single bit that
12263 * differs between P and p; the other xor inverts just
12265 value ^= 'P' ^ 'p';
12267 while (isSPACE(UCHARAT(RExC_parse))) {
12272 /* Try to get the definition of the property into
12273 * <invlist>. If /i is in effect, the effective property
12274 * will have its name be <__NAME_i>. The design is
12275 * discussed in commit
12276 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
12277 Newx(name, n + sizeof("_i__\n"), char);
12279 sprintf(name, "%s%.*s%s\n",
12280 (FOLD) ? "__" : "",
12286 /* Look up the property name, and get its swash and
12287 * inversion list, if the property is found */
12289 SvREFCNT_dec_NN(swash);
12291 swash = _core_swash_init("utf8", name, &PL_sv_undef,
12294 NULL, /* No inversion list */
12297 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
12299 SvREFCNT_dec_NN(swash);
12303 /* Here didn't find it. It could be a user-defined
12304 * property that will be available at run-time. If we
12305 * accept only compile-time properties, is an error;
12306 * otherwise add it to the list for run-time look up */
12308 RExC_parse = e + 1;
12309 vFAIL3("Property '%.*s' is unknown", (int) n, name);
12311 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s\n",
12312 (value == 'p' ? '+' : '!'),
12314 has_user_defined_property = TRUE;
12316 /* We don't know yet, so have to assume that the
12317 * property could match something in the Latin1 range,
12318 * hence something that isn't utf8. Note that this
12319 * would cause things in <depends_list> to match
12320 * inappropriately, except that any \p{}, including
12321 * this one forces Unicode semantics, which means there
12322 * is <no depends_list> */
12323 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
12327 /* Here, did get the swash and its inversion list. If
12328 * the swash is from a user-defined property, then this
12329 * whole character class should be regarded as such */
12330 has_user_defined_property =
12332 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY);
12334 /* Invert if asking for the complement */
12335 if (value == 'P') {
12336 _invlist_union_complement_2nd(properties,
12340 /* The swash can't be used as-is, because we've
12341 * inverted things; delay removing it to here after
12342 * have copied its invlist above */
12343 SvREFCNT_dec_NN(swash);
12347 _invlist_union(properties, invlist, &properties);
12352 RExC_parse = e + 1;
12353 namedclass = ANYOF_UNIPROP; /* no official name, but it's
12356 /* \p means they want Unicode semantics */
12357 RExC_uni_semantics = 1;
12360 case 'n': value = '\n'; break;
12361 case 'r': value = '\r'; break;
12362 case 't': value = '\t'; break;
12363 case 'f': value = '\f'; break;
12364 case 'b': value = '\b'; break;
12365 case 'e': value = ASCII_TO_NATIVE('\033');break;
12366 case 'a': value = ASCII_TO_NATIVE('\007');break;
12368 RExC_parse--; /* function expects to be pointed at the 'o' */
12370 const char* error_msg;
12371 bool valid = grok_bslash_o(&RExC_parse,
12374 SIZE_ONLY, /* warnings in pass
12377 silence_non_portable,
12383 if (PL_encoding && value < 0x100) {
12384 goto recode_encoding;
12388 RExC_parse--; /* function expects to be pointed at the 'x' */
12390 const char* error_msg;
12391 bool valid = grok_bslash_x(&RExC_parse,
12394 TRUE, /* Output warnings */
12396 silence_non_portable,
12402 if (PL_encoding && value < 0x100)
12403 goto recode_encoding;
12406 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
12408 case '0': case '1': case '2': case '3': case '4':
12409 case '5': case '6': case '7':
12411 /* Take 1-3 octal digits */
12412 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12413 numlen = (strict) ? 4 : 3;
12414 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
12415 RExC_parse += numlen;
12418 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12419 vFAIL("Need exactly 3 octal digits");
12421 else if (! SIZE_ONLY /* like \08, \178 */
12423 && RExC_parse < RExC_end
12424 && isDIGIT(*RExC_parse)
12425 && ckWARN(WARN_REGEXP))
12427 SAVEFREESV(RExC_rx_sv);
12428 reg_warn_non_literal_string(
12430 form_short_octal_warning(RExC_parse, numlen));
12431 (void)ReREFCNT_inc(RExC_rx_sv);
12434 if (PL_encoding && value < 0x100)
12435 goto recode_encoding;
12439 if (! RExC_override_recoding) {
12440 SV* enc = PL_encoding;
12441 value = reg_recode((const char)(U8)value, &enc);
12444 vFAIL("Invalid escape in the specified encoding");
12446 else if (SIZE_ONLY) {
12447 ckWARNreg(RExC_parse,
12448 "Invalid escape in the specified encoding");
12454 /* Allow \_ to not give an error */
12455 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
12457 vFAIL2("Unrecognized escape \\%c in character class",
12461 SAVEFREESV(RExC_rx_sv);
12462 ckWARN2reg(RExC_parse,
12463 "Unrecognized escape \\%c in character class passed through",
12465 (void)ReREFCNT_inc(RExC_rx_sv);
12469 } /* End of switch on char following backslash */
12470 } /* end of handling backslash escape sequences */
12473 literal_endpoint++;
12476 /* Here, we have the current token in 'value' */
12478 /* What matches in a locale is not known until runtime. This includes
12479 * what the Posix classes (like \w, [:space:]) match. Room must be
12480 * reserved (one time per class) to store such classes, either if Perl
12481 * is compiled so that locale nodes always should have this space, or
12482 * if there is such class info to be stored. The space will contain a
12483 * bit for each named class that is to be matched against. This isn't
12484 * needed for \p{} and pseudo-classes, as they are not affected by
12485 * locale, and hence are dealt with separately */
12488 && (ANYOF_LOCALE == ANYOF_CLASS
12489 || (namedclass > OOB_NAMEDCLASS && namedclass < ANYOF_MAX)))
12493 RExC_size += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12496 RExC_emit += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12497 ANYOF_CLASS_ZERO(ret);
12499 ANYOF_FLAGS(ret) |= ANYOF_CLASS;
12502 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
12504 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
12505 * literal, as is the character that began the false range, i.e.
12506 * the 'a' in the examples */
12509 const int w = (RExC_parse >= rangebegin)
12510 ? RExC_parse - rangebegin
12513 vFAIL4("False [] range \"%*.*s\"", w, w, rangebegin);
12516 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
12517 ckWARN4reg(RExC_parse,
12518 "False [] range \"%*.*s\"",
12520 (void)ReREFCNT_inc(RExC_rx_sv);
12521 cp_list = add_cp_to_invlist(cp_list, '-');
12522 cp_list = add_cp_to_invlist(cp_list, prevvalue);
12526 range = 0; /* this was not a true range */
12527 element_count += 2; /* So counts for three values */
12531 U8 classnum = namedclass_to_classnum(namedclass);
12532 if (namedclass >= ANYOF_MAX) { /* If a special class */
12533 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
12535 /* Here, should be \h, \H, \v, or \V. Neither /d nor
12536 * /l make a difference in what these match. There
12537 * would be problems if these characters had folds
12538 * other than themselves, as cp_list is subject to
12540 if (classnum != _CC_VERTSPACE) {
12541 assert( namedclass == ANYOF_HORIZWS
12542 || namedclass == ANYOF_NHORIZWS);
12544 /* It turns out that \h is just a synonym for
12546 classnum = _CC_BLANK;
12549 _invlist_union_maybe_complement_2nd(
12551 PL_XPosix_ptrs[classnum],
12552 cBOOL(namedclass % 2), /* Complement if odd
12553 (NHORIZWS, NVERTWS)
12558 else if (classnum == _CC_ASCII) {
12561 ANYOF_CLASS_SET(ret, namedclass);
12564 #endif /* Not isascii(); just use the hard-coded definition for it */
12565 _invlist_union_maybe_complement_2nd(
12568 cBOOL(namedclass % 2), /* Complement if odd
12572 else { /* Garden variety class */
12574 /* The ascii range inversion list */
12575 SV* ascii_source = PL_Posix_ptrs[classnum];
12577 /* The full Latin1 range inversion list */
12578 SV* l1_source = PL_L1Posix_ptrs[classnum];
12580 /* This code is structured into two major clauses. The
12581 * first is for classes whose complete definitions may not
12582 * already be known. It not, the Latin1 definition
12583 * (guaranteed to already known) is used plus code is
12584 * generated to load the rest at run-time (only if needed).
12585 * If the complete definition is known, it drops down to
12586 * the second clause, where the complete definition is
12589 if (classnum < _FIRST_NON_SWASH_CC) {
12591 /* Here, the class has a swash, which may or not
12592 * already be loaded */
12594 /* The name of the property to use to match the full
12595 * eXtended Unicode range swash for this character
12597 const char *Xname = swash_property_names[classnum];
12599 /* If returning the inversion list, we can't defer
12600 * getting this until runtime */
12601 if (ret_invlist && ! PL_utf8_swash_ptrs[classnum]) {
12602 PL_utf8_swash_ptrs[classnum] =
12603 _core_swash_init("utf8", Xname, &PL_sv_undef,
12606 NULL, /* No inversion list */
12607 NULL /* No flags */
12609 assert(PL_utf8_swash_ptrs[classnum]);
12611 if ( ! PL_utf8_swash_ptrs[classnum]) {
12612 if (namedclass % 2 == 0) { /* A non-complemented
12614 /* If not /a matching, there are code points we
12615 * don't know at compile time. Arrange for the
12616 * unknown matches to be loaded at run-time, if
12618 if (! AT_LEAST_ASCII_RESTRICTED) {
12619 Perl_sv_catpvf(aTHX_ listsv, "+utf8::%s\n",
12622 if (LOC) { /* Under locale, set run-time
12624 ANYOF_CLASS_SET(ret, namedclass);
12627 /* Add the current class's code points to
12628 * the running total */
12629 _invlist_union(posixes,
12630 (AT_LEAST_ASCII_RESTRICTED)
12636 else { /* A complemented class */
12637 if (AT_LEAST_ASCII_RESTRICTED) {
12638 /* Under /a should match everything above
12639 * ASCII, plus the complement of the set's
12641 _invlist_union_complement_2nd(posixes,
12646 /* Arrange for the unknown matches to be
12647 * loaded at run-time, if needed */
12648 Perl_sv_catpvf(aTHX_ listsv, "!utf8::%s\n",
12650 runtime_posix_matches_above_Unicode = TRUE;
12652 ANYOF_CLASS_SET(ret, namedclass);
12656 /* We want to match everything in
12657 * Latin1, except those things that
12658 * l1_source matches */
12659 SV* scratch_list = NULL;
12660 _invlist_subtract(PL_Latin1, l1_source,
12663 /* Add the list from this class to the
12666 posixes = scratch_list;
12669 _invlist_union(posixes,
12672 SvREFCNT_dec_NN(scratch_list);
12674 if (DEPENDS_SEMANTICS) {
12676 |= ANYOF_NON_UTF8_LATIN1_ALL;
12681 goto namedclass_done;
12684 /* Here, there is a swash loaded for the class. If no
12685 * inversion list for it yet, get it */
12686 if (! PL_XPosix_ptrs[classnum]) {
12687 PL_XPosix_ptrs[classnum]
12688 = _swash_to_invlist(PL_utf8_swash_ptrs[classnum]);
12692 /* Here there is an inversion list already loaded for the
12695 if (namedclass % 2 == 0) { /* A non-complemented class,
12696 like ANYOF_PUNCT */
12698 /* For non-locale, just add it to any existing list
12700 _invlist_union(posixes,
12701 (AT_LEAST_ASCII_RESTRICTED)
12703 : PL_XPosix_ptrs[classnum],
12706 else { /* Locale */
12707 SV* scratch_list = NULL;
12709 /* For above Latin1 code points, we use the full
12711 _invlist_intersection(PL_AboveLatin1,
12712 PL_XPosix_ptrs[classnum],
12714 /* And set the output to it, adding instead if
12715 * there already is an output. Checking if
12716 * 'posixes' is NULL first saves an extra clone.
12717 * Its reference count will be decremented at the
12718 * next union, etc, or if this is the only
12719 * instance, at the end of the routine */
12721 posixes = scratch_list;
12724 _invlist_union(posixes, scratch_list, &posixes);
12725 SvREFCNT_dec_NN(scratch_list);
12728 #ifndef HAS_ISBLANK
12729 if (namedclass != ANYOF_BLANK) {
12731 /* Set this class in the node for runtime
12733 ANYOF_CLASS_SET(ret, namedclass);
12734 #ifndef HAS_ISBLANK
12737 /* No isblank(), use the hard-coded ASCII-range
12738 * blanks, adding them to the running total. */
12740 _invlist_union(posixes, ascii_source, &posixes);
12745 else { /* A complemented class, like ANYOF_NPUNCT */
12747 _invlist_union_complement_2nd(
12749 (AT_LEAST_ASCII_RESTRICTED)
12751 : PL_XPosix_ptrs[classnum],
12753 /* Under /d, everything in the upper half of the
12754 * Latin1 range matches this complement */
12755 if (DEPENDS_SEMANTICS) {
12756 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL;
12759 else { /* Locale */
12760 SV* scratch_list = NULL;
12761 _invlist_subtract(PL_AboveLatin1,
12762 PL_XPosix_ptrs[classnum],
12765 posixes = scratch_list;
12768 _invlist_union(posixes, scratch_list, &posixes);
12769 SvREFCNT_dec_NN(scratch_list);
12771 #ifndef HAS_ISBLANK
12772 if (namedclass != ANYOF_NBLANK) {
12774 ANYOF_CLASS_SET(ret, namedclass);
12775 #ifndef HAS_ISBLANK
12778 /* Get the list of all code points in Latin1
12779 * that are not ASCII blanks, and add them to
12780 * the running total */
12781 _invlist_subtract(PL_Latin1, ascii_source,
12783 _invlist_union(posixes, scratch_list, &posixes);
12784 SvREFCNT_dec_NN(scratch_list);
12791 continue; /* Go get next character */
12793 } /* end of namedclass \blah */
12795 /* Here, we have a single value. If 'range' is set, it is the ending
12796 * of a range--check its validity. Later, we will handle each
12797 * individual code point in the range. If 'range' isn't set, this
12798 * could be the beginning of a range, so check for that by looking
12799 * ahead to see if the next real character to be processed is the range
12800 * indicator--the minus sign */
12803 RExC_parse = regpatws(pRExC_state, RExC_parse,
12804 FALSE /* means don't recognize comments */);
12808 if (prevvalue > value) /* b-a */ {
12809 const int w = RExC_parse - rangebegin;
12810 Simple_vFAIL4("Invalid [] range \"%*.*s\"", w, w, rangebegin);
12811 range = 0; /* not a valid range */
12815 prevvalue = value; /* save the beginning of the potential range */
12816 if (! stop_at_1 /* Can't be a range if parsing just one thing */
12817 && *RExC_parse == '-')
12819 char* next_char_ptr = RExC_parse + 1;
12820 if (skip_white) { /* Get the next real char after the '-' */
12821 next_char_ptr = regpatws(pRExC_state,
12823 FALSE); /* means don't recognize
12827 /* If the '-' is at the end of the class (just before the ']',
12828 * it is a literal minus; otherwise it is a range */
12829 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
12830 RExC_parse = next_char_ptr;
12832 /* a bad range like \w-, [:word:]- ? */
12833 if (namedclass > OOB_NAMEDCLASS) {
12834 if (strict || ckWARN(WARN_REGEXP)) {
12836 RExC_parse >= rangebegin ?
12837 RExC_parse - rangebegin : 0;
12839 vFAIL4("False [] range \"%*.*s\"",
12844 "False [] range \"%*.*s\"",
12849 cp_list = add_cp_to_invlist(cp_list, '-');
12853 range = 1; /* yeah, it's a range! */
12854 continue; /* but do it the next time */
12859 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
12862 /* non-Latin1 code point implies unicode semantics. Must be set in
12863 * pass1 so is there for the whole of pass 2 */
12865 RExC_uni_semantics = 1;
12868 /* Ready to process either the single value, or the completed range.
12869 * For single-valued non-inverted ranges, we consider the possibility
12870 * of multi-char folds. (We made a conscious decision to not do this
12871 * for the other cases because it can often lead to non-intuitive
12872 * results. For example, you have the peculiar case that:
12873 * "s s" =~ /^[^\xDF]+$/i => Y
12874 * "ss" =~ /^[^\xDF]+$/i => N
12876 * See [perl #89750] */
12877 if (FOLD && allow_multi_folds && value == prevvalue) {
12878 if (value == LATIN_SMALL_LETTER_SHARP_S
12879 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
12882 /* Here <value> is indeed a multi-char fold. Get what it is */
12884 U8 foldbuf[UTF8_MAXBYTES_CASE];
12887 UV folded = _to_uni_fold_flags(
12892 | ((LOC) ? FOLD_FLAGS_LOCALE
12893 : (ASCII_FOLD_RESTRICTED)
12894 ? FOLD_FLAGS_NOMIX_ASCII
12898 /* Here, <folded> should be the first character of the
12899 * multi-char fold of <value>, with <foldbuf> containing the
12900 * whole thing. But, if this fold is not allowed (because of
12901 * the flags), <fold> will be the same as <value>, and should
12902 * be processed like any other character, so skip the special
12904 if (folded != value) {
12906 /* Skip if we are recursed, currently parsing the class
12907 * again. Otherwise add this character to the list of
12908 * multi-char folds. */
12909 if (! RExC_in_multi_char_class) {
12910 AV** this_array_ptr;
12912 STRLEN cp_count = utf8_length(foldbuf,
12913 foldbuf + foldlen);
12914 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
12916 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
12919 if (! multi_char_matches) {
12920 multi_char_matches = newAV();
12923 /* <multi_char_matches> is actually an array of arrays.
12924 * There will be one or two top-level elements: [2],
12925 * and/or [3]. The [2] element is an array, each
12926 * element thereof is a character which folds to two
12927 * characters; likewise for [3]. (Unicode guarantees a
12928 * maximum of 3 characters in any fold.) When we
12929 * rewrite the character class below, we will do so
12930 * such that the longest folds are written first, so
12931 * that it prefers the longest matching strings first.
12932 * This is done even if it turns out that any
12933 * quantifier is non-greedy, out of programmer
12934 * laziness. Tom Christiansen has agreed that this is
12935 * ok. This makes the test for the ligature 'ffi' come
12936 * before the test for 'ff' */
12937 if (av_exists(multi_char_matches, cp_count)) {
12938 this_array_ptr = (AV**) av_fetch(multi_char_matches,
12940 this_array = *this_array_ptr;
12943 this_array = newAV();
12944 av_store(multi_char_matches, cp_count,
12947 av_push(this_array, multi_fold);
12950 /* This element should not be processed further in this
12953 value = save_value;
12954 prevvalue = save_prevvalue;
12960 /* Deal with this element of the class */
12963 cp_list = _add_range_to_invlist(cp_list, prevvalue, value);
12965 SV* this_range = _new_invlist(1);
12966 _append_range_to_invlist(this_range, prevvalue, value);
12968 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
12969 * If this range was specified using something like 'i-j', we want
12970 * to include only the 'i' and the 'j', and not anything in
12971 * between, so exclude non-ASCII, non-alphabetics from it.
12972 * However, if the range was specified with something like
12973 * [\x89-\x91] or [\x89-j], all code points within it should be
12974 * included. literal_endpoint==2 means both ends of the range used
12975 * a literal character, not \x{foo} */
12976 if (literal_endpoint == 2
12977 && (prevvalue >= 'a' && value <= 'z')
12978 || (prevvalue >= 'A' && value <= 'Z'))
12980 _invlist_intersection(this_range, PL_Posix_ptrs[_CC_ALPHA],
12983 _invlist_union(cp_list, this_range, &cp_list);
12984 literal_endpoint = 0;
12988 range = 0; /* this range (if it was one) is done now */
12989 } /* End of loop through all the text within the brackets */
12991 /* If anything in the class expands to more than one character, we have to
12992 * deal with them by building up a substitute parse string, and recursively
12993 * calling reg() on it, instead of proceeding */
12994 if (multi_char_matches) {
12995 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
12998 char *save_end = RExC_end;
12999 char *save_parse = RExC_parse;
13000 bool first_time = TRUE; /* First multi-char occurrence doesn't get
13005 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
13006 because too confusing */
13008 sv_catpv(substitute_parse, "(?:");
13012 /* Look at the longest folds first */
13013 for (cp_count = av_len(multi_char_matches); cp_count > 0; cp_count--) {
13015 if (av_exists(multi_char_matches, cp_count)) {
13016 AV** this_array_ptr;
13019 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13021 while ((this_sequence = av_pop(*this_array_ptr)) !=
13024 if (! first_time) {
13025 sv_catpv(substitute_parse, "|");
13027 first_time = FALSE;
13029 sv_catpv(substitute_parse, SvPVX(this_sequence));
13034 /* If the character class contains anything else besides these
13035 * multi-character folds, have to include it in recursive parsing */
13036 if (element_count) {
13037 sv_catpv(substitute_parse, "|[");
13038 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
13039 sv_catpv(substitute_parse, "]");
13042 sv_catpv(substitute_parse, ")");
13045 /* This is a way to get the parse to skip forward a whole named
13046 * sequence instead of matching the 2nd character when it fails the
13048 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
13052 RExC_parse = SvPV(substitute_parse, len);
13053 RExC_end = RExC_parse + len;
13054 RExC_in_multi_char_class = 1;
13055 RExC_emit = (regnode *)orig_emit;
13057 ret = reg(pRExC_state, 1, ®_flags, depth+1);
13059 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
13061 RExC_parse = save_parse;
13062 RExC_end = save_end;
13063 RExC_in_multi_char_class = 0;
13064 SvREFCNT_dec_NN(multi_char_matches);
13068 /* If the character class contains only a single element, it may be
13069 * optimizable into another node type which is smaller and runs faster.
13070 * Check if this is the case for this class */
13071 if (element_count == 1 && ! ret_invlist) {
13075 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like \w or
13076 [:digit:] or \p{foo} */
13078 /* All named classes are mapped into POSIXish nodes, with its FLAG
13079 * argument giving which class it is */
13080 switch ((I32)namedclass) {
13081 case ANYOF_UNIPROP:
13084 /* These don't depend on the charset modifiers. They always
13085 * match under /u rules */
13086 case ANYOF_NHORIZWS:
13087 case ANYOF_HORIZWS:
13088 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
13091 case ANYOF_NVERTWS:
13096 /* The actual POSIXish node for all the rest depends on the
13097 * charset modifier. The ones in the first set depend only on
13098 * ASCII or, if available on this platform, locale */
13102 op = (LOC) ? POSIXL : POSIXA;
13113 /* under /a could be alpha */
13115 if (ASCII_RESTRICTED) {
13116 namedclass = ANYOF_ALPHA + (namedclass % 2);
13124 /* The rest have more possibilities depending on the charset.
13125 * We take advantage of the enum ordering of the charset
13126 * modifiers to get the exact node type, */
13128 op = POSIXD + get_regex_charset(RExC_flags);
13129 if (op > POSIXA) { /* /aa is same as /a */
13132 #ifndef HAS_ISBLANK
13134 && (namedclass == ANYOF_BLANK
13135 || namedclass == ANYOF_NBLANK))
13142 /* The odd numbered ones are the complements of the
13143 * next-lower even number one */
13144 if (namedclass % 2 == 1) {
13148 arg = namedclass_to_classnum(namedclass);
13152 else if (value == prevvalue) {
13154 /* Here, the class consists of just a single code point */
13157 if (! LOC && value == '\n') {
13158 op = REG_ANY; /* Optimize [^\n] */
13159 *flagp |= HASWIDTH|SIMPLE;
13163 else if (value < 256 || UTF) {
13165 /* Optimize a single value into an EXACTish node, but not if it
13166 * would require converting the pattern to UTF-8. */
13167 op = compute_EXACTish(pRExC_state);
13169 } /* Otherwise is a range */
13170 else if (! LOC) { /* locale could vary these */
13171 if (prevvalue == '0') {
13172 if (value == '9') {
13179 /* Here, we have changed <op> away from its initial value iff we found
13180 * an optimization */
13183 /* Throw away this ANYOF regnode, and emit the calculated one,
13184 * which should correspond to the beginning, not current, state of
13186 const char * cur_parse = RExC_parse;
13187 RExC_parse = (char *)orig_parse;
13191 /* To get locale nodes to not use the full ANYOF size would
13192 * require moving the code above that writes the portions
13193 * of it that aren't in other nodes to after this point.
13194 * e.g. ANYOF_CLASS_SET */
13195 RExC_size = orig_size;
13199 RExC_emit = (regnode *)orig_emit;
13200 if (PL_regkind[op] == POSIXD) {
13202 op += NPOSIXD - POSIXD;
13207 ret = reg_node(pRExC_state, op);
13209 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
13213 *flagp |= HASWIDTH|SIMPLE;
13215 else if (PL_regkind[op] == EXACT) {
13216 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13219 RExC_parse = (char *) cur_parse;
13221 SvREFCNT_dec(posixes);
13222 SvREFCNT_dec(cp_list);
13229 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
13231 /* If folding, we calculate all characters that could fold to or from the
13232 * ones already on the list */
13233 if (FOLD && cp_list) {
13234 UV start, end; /* End points of code point ranges */
13236 SV* fold_intersection = NULL;
13238 /* If the highest code point is within Latin1, we can use the
13239 * compiled-in Alphas list, and not have to go out to disk. This
13240 * yields two false positives, the masculine and feminine ordinal
13241 * indicators, which are weeded out below using the
13242 * IS_IN_SOME_FOLD_L1() macro */
13243 if (invlist_highest(cp_list) < 256) {
13244 _invlist_intersection(PL_L1Posix_ptrs[_CC_ALPHA], cp_list,
13245 &fold_intersection);
13249 /* Here, there are non-Latin1 code points, so we will have to go
13250 * fetch the list of all the characters that participate in folds
13252 if (! PL_utf8_foldable) {
13253 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13254 &PL_sv_undef, 1, 0);
13255 PL_utf8_foldable = _get_swash_invlist(swash);
13256 SvREFCNT_dec_NN(swash);
13259 /* This is a hash that for a particular fold gives all characters
13260 * that are involved in it */
13261 if (! PL_utf8_foldclosures) {
13263 /* If we were unable to find any folds, then we likely won't be
13264 * able to find the closures. So just create an empty list.
13265 * Folding will effectively be restricted to the non-Unicode
13266 * rules hard-coded into Perl. (This case happens legitimately
13267 * during compilation of Perl itself before the Unicode tables
13268 * are generated) */
13269 if (_invlist_len(PL_utf8_foldable) == 0) {
13270 PL_utf8_foldclosures = newHV();
13273 /* If the folds haven't been read in, call a fold function
13275 if (! PL_utf8_tofold) {
13276 U8 dummy[UTF8_MAXBYTES+1];
13278 /* This string is just a short named one above \xff */
13279 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
13280 assert(PL_utf8_tofold); /* Verify that worked */
13282 PL_utf8_foldclosures =
13283 _swash_inversion_hash(PL_utf8_tofold);
13287 /* Only the characters in this class that participate in folds need
13288 * be checked. Get the intersection of this class and all the
13289 * possible characters that are foldable. This can quickly narrow
13290 * down a large class */
13291 _invlist_intersection(PL_utf8_foldable, cp_list,
13292 &fold_intersection);
13295 /* Now look at the foldable characters in this class individually */
13296 invlist_iterinit(fold_intersection);
13297 while (invlist_iternext(fold_intersection, &start, &end)) {
13300 /* Locale folding for Latin1 characters is deferred until runtime */
13301 if (LOC && start < 256) {
13305 /* Look at every character in the range */
13306 for (j = start; j <= end; j++) {
13308 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
13314 /* We have the latin1 folding rules hard-coded here so that
13315 * an innocent-looking character class, like /[ks]/i won't
13316 * have to go out to disk to find the possible matches.
13317 * XXX It would be better to generate these via regen, in
13318 * case a new version of the Unicode standard adds new
13319 * mappings, though that is not really likely, and may be
13320 * caught by the default: case of the switch below. */
13322 if (IS_IN_SOME_FOLD_L1(j)) {
13324 /* ASCII is always matched; non-ASCII is matched only
13325 * under Unicode rules */
13326 if (isASCII(j) || AT_LEAST_UNI_SEMANTICS) {
13328 add_cp_to_invlist(cp_list, PL_fold_latin1[j]);
13332 add_cp_to_invlist(depends_list, PL_fold_latin1[j]);
13336 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
13337 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
13339 /* Certain Latin1 characters have matches outside
13340 * Latin1. To get here, <j> is one of those
13341 * characters. None of these matches is valid for
13342 * ASCII characters under /aa, which is why the 'if'
13343 * just above excludes those. These matches only
13344 * happen when the target string is utf8. The code
13345 * below adds the single fold closures for <j> to the
13346 * inversion list. */
13351 add_cp_to_invlist(cp_list, KELVIN_SIGN);
13355 cp_list = add_cp_to_invlist(cp_list,
13356 LATIN_SMALL_LETTER_LONG_S);
13359 cp_list = add_cp_to_invlist(cp_list,
13360 GREEK_CAPITAL_LETTER_MU);
13361 cp_list = add_cp_to_invlist(cp_list,
13362 GREEK_SMALL_LETTER_MU);
13364 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
13365 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
13367 add_cp_to_invlist(cp_list, ANGSTROM_SIGN);
13369 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
13370 cp_list = add_cp_to_invlist(cp_list,
13371 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
13373 case LATIN_SMALL_LETTER_SHARP_S:
13374 cp_list = add_cp_to_invlist(cp_list,
13375 LATIN_CAPITAL_LETTER_SHARP_S);
13377 case 'F': case 'f':
13378 case 'I': case 'i':
13379 case 'L': case 'l':
13380 case 'T': case 't':
13381 case 'A': case 'a':
13382 case 'H': case 'h':
13383 case 'J': case 'j':
13384 case 'N': case 'n':
13385 case 'W': case 'w':
13386 case 'Y': case 'y':
13387 /* These all are targets of multi-character
13388 * folds from code points that require UTF8 to
13389 * express, so they can't match unless the
13390 * target string is in UTF-8, so no action here
13391 * is necessary, as regexec.c properly handles
13392 * the general case for UTF-8 matching and
13393 * multi-char folds */
13396 /* Use deprecated warning to increase the
13397 * chances of this being output */
13398 ckWARN2regdep(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
13405 /* Here is an above Latin1 character. We don't have the rules
13406 * hard-coded for it. First, get its fold. This is the simple
13407 * fold, as the multi-character folds have been handled earlier
13408 * and separated out */
13409 _to_uni_fold_flags(j, foldbuf, &foldlen,
13411 ? FOLD_FLAGS_LOCALE
13412 : (ASCII_FOLD_RESTRICTED)
13413 ? FOLD_FLAGS_NOMIX_ASCII
13416 /* Single character fold of above Latin1. Add everything in
13417 * its fold closure to the list that this node should match.
13418 * The fold closures data structure is a hash with the keys
13419 * being the UTF-8 of every character that is folded to, like
13420 * 'k', and the values each an array of all code points that
13421 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
13422 * Multi-character folds are not included */
13423 if ((listp = hv_fetch(PL_utf8_foldclosures,
13424 (char *) foldbuf, foldlen, FALSE)))
13426 AV* list = (AV*) *listp;
13428 for (k = 0; k <= av_len(list); k++) {
13429 SV** c_p = av_fetch(list, k, FALSE);
13432 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
13436 /* /aa doesn't allow folds between ASCII and non-; /l
13437 * doesn't allow them between above and below 256 */
13438 if ((ASCII_FOLD_RESTRICTED
13439 && (isASCII(c) != isASCII(j)))
13440 || (LOC && ((c < 256) != (j < 256))))
13445 /* Folds involving non-ascii Latin1 characters
13446 * under /d are added to a separate list */
13447 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
13449 cp_list = add_cp_to_invlist(cp_list, c);
13452 depends_list = add_cp_to_invlist(depends_list, c);
13458 SvREFCNT_dec_NN(fold_intersection);
13461 /* And combine the result (if any) with any inversion list from posix
13462 * classes. The lists are kept separate up to now because we don't want to
13463 * fold the classes (folding of those is automatically handled by the swash
13464 * fetching code) */
13466 if (! DEPENDS_SEMANTICS) {
13468 _invlist_union(cp_list, posixes, &cp_list);
13469 SvREFCNT_dec_NN(posixes);
13476 /* Under /d, we put into a separate list the Latin1 things that
13477 * match only when the target string is utf8 */
13478 SV* nonascii_but_latin1_properties = NULL;
13479 _invlist_intersection(posixes, PL_Latin1,
13480 &nonascii_but_latin1_properties);
13481 _invlist_subtract(nonascii_but_latin1_properties, PL_ASCII,
13482 &nonascii_but_latin1_properties);
13483 _invlist_subtract(posixes, nonascii_but_latin1_properties,
13486 _invlist_union(cp_list, posixes, &cp_list);
13487 SvREFCNT_dec_NN(posixes);
13493 if (depends_list) {
13494 _invlist_union(depends_list, nonascii_but_latin1_properties,
13496 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
13499 depends_list = nonascii_but_latin1_properties;
13504 /* And combine the result (if any) with any inversion list from properties.
13505 * The lists are kept separate up to now so that we can distinguish the two
13506 * in regards to matching above-Unicode. A run-time warning is generated
13507 * if a Unicode property is matched against a non-Unicode code point. But,
13508 * we allow user-defined properties to match anything, without any warning,
13509 * and we also suppress the warning if there is a portion of the character
13510 * class that isn't a Unicode property, and which matches above Unicode, \W
13511 * or [\x{110000}] for example.
13512 * (Note that in this case, unlike the Posix one above, there is no
13513 * <depends_list>, because having a Unicode property forces Unicode
13516 bool warn_super = ! has_user_defined_property;
13519 /* If it matters to the final outcome, see if a non-property
13520 * component of the class matches above Unicode. If so, the
13521 * warning gets suppressed. This is true even if just a single
13522 * such code point is specified, as though not strictly correct if
13523 * another such code point is matched against, the fact that they
13524 * are using above-Unicode code points indicates they should know
13525 * the issues involved */
13527 bool non_prop_matches_above_Unicode =
13528 runtime_posix_matches_above_Unicode
13529 | (invlist_highest(cp_list) > PERL_UNICODE_MAX);
13531 non_prop_matches_above_Unicode =
13532 ! non_prop_matches_above_Unicode;
13534 warn_super = ! non_prop_matches_above_Unicode;
13537 _invlist_union(properties, cp_list, &cp_list);
13538 SvREFCNT_dec_NN(properties);
13541 cp_list = properties;
13545 OP(ret) = ANYOF_WARN_SUPER;
13549 /* Here, we have calculated what code points should be in the character
13552 * Now we can see about various optimizations. Fold calculation (which we
13553 * did above) needs to take place before inversion. Otherwise /[^k]/i
13554 * would invert to include K, which under /i would match k, which it
13555 * shouldn't. Therefore we can't invert folded locale now, as it won't be
13556 * folded until runtime */
13558 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
13559 * at compile time. Besides not inverting folded locale now, we can't
13560 * invert if there are things such as \w, which aren't known until runtime
13563 && ! (LOC && (FOLD || (ANYOF_FLAGS(ret) & ANYOF_CLASS)))
13565 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13567 _invlist_invert(cp_list);
13569 /* Any swash can't be used as-is, because we've inverted things */
13571 SvREFCNT_dec_NN(swash);
13575 /* Clear the invert flag since have just done it here */
13580 *ret_invlist = cp_list;
13582 /* Discard the generated node */
13584 RExC_size = orig_size;
13587 RExC_emit = orig_emit;
13592 /* If we didn't do folding, it's because some information isn't available
13593 * until runtime; set the run-time fold flag for these. (We don't have to
13594 * worry about properties folding, as that is taken care of by the swash
13598 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
13601 /* Some character classes are equivalent to other nodes. Such nodes take
13602 * up less room and generally fewer operations to execute than ANYOF nodes.
13603 * Above, we checked for and optimized into some such equivalents for
13604 * certain common classes that are easy to test. Getting to this point in
13605 * the code means that the class didn't get optimized there. Since this
13606 * code is only executed in Pass 2, it is too late to save space--it has
13607 * been allocated in Pass 1, and currently isn't given back. But turning
13608 * things into an EXACTish node can allow the optimizer to join it to any
13609 * adjacent such nodes. And if the class is equivalent to things like /./,
13610 * expensive run-time swashes can be avoided. Now that we have more
13611 * complete information, we can find things necessarily missed by the
13612 * earlier code. I (khw) am not sure how much to look for here. It would
13613 * be easy, but perhaps too slow, to check any candidates against all the
13614 * node types they could possibly match using _invlistEQ(). */
13619 && ! (ANYOF_FLAGS(ret) & ANYOF_CLASS)
13620 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13623 U8 op = END; /* The optimzation node-type */
13624 const char * cur_parse= RExC_parse;
13626 invlist_iterinit(cp_list);
13627 if (! invlist_iternext(cp_list, &start, &end)) {
13629 /* Here, the list is empty. This happens, for example, when a
13630 * Unicode property is the only thing in the character class, and
13631 * it doesn't match anything. (perluniprops.pod notes such
13634 *flagp |= HASWIDTH|SIMPLE;
13636 else if (start == end) { /* The range is a single code point */
13637 if (! invlist_iternext(cp_list, &start, &end)
13639 /* Don't do this optimization if it would require changing
13640 * the pattern to UTF-8 */
13641 && (start < 256 || UTF))
13643 /* Here, the list contains a single code point. Can optimize
13644 * into an EXACT node */
13653 /* A locale node under folding with one code point can be
13654 * an EXACTFL, as its fold won't be calculated until
13660 /* Here, we are generally folding, but there is only one
13661 * code point to match. If we have to, we use an EXACT
13662 * node, but it would be better for joining with adjacent
13663 * nodes in the optimization pass if we used the same
13664 * EXACTFish node that any such are likely to be. We can
13665 * do this iff the code point doesn't participate in any
13666 * folds. For example, an EXACTF of a colon is the same as
13667 * an EXACT one, since nothing folds to or from a colon. */
13669 if (IS_IN_SOME_FOLD_L1(value)) {
13674 if (! PL_utf8_foldable) {
13675 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13676 &PL_sv_undef, 1, 0);
13677 PL_utf8_foldable = _get_swash_invlist(swash);
13678 SvREFCNT_dec_NN(swash);
13680 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
13685 /* If we haven't found the node type, above, it means we
13686 * can use the prevailing one */
13688 op = compute_EXACTish(pRExC_state);
13693 else if (start == 0) {
13694 if (end == UV_MAX) {
13696 *flagp |= HASWIDTH|SIMPLE;
13699 else if (end == '\n' - 1
13700 && invlist_iternext(cp_list, &start, &end)
13701 && start == '\n' + 1 && end == UV_MAX)
13704 *flagp |= HASWIDTH|SIMPLE;
13708 invlist_iterfinish(cp_list);
13711 RExC_parse = (char *)orig_parse;
13712 RExC_emit = (regnode *)orig_emit;
13714 ret = reg_node(pRExC_state, op);
13716 RExC_parse = (char *)cur_parse;
13718 if (PL_regkind[op] == EXACT) {
13719 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13722 SvREFCNT_dec_NN(cp_list);
13727 /* Here, <cp_list> contains all the code points we can determine at
13728 * compile time that match under all conditions. Go through it, and
13729 * for things that belong in the bitmap, put them there, and delete from
13730 * <cp_list>. While we are at it, see if everything above 255 is in the
13731 * list, and if so, set a flag to speed up execution */
13732 ANYOF_BITMAP_ZERO(ret);
13735 /* This gets set if we actually need to modify things */
13736 bool change_invlist = FALSE;
13740 /* Start looking through <cp_list> */
13741 invlist_iterinit(cp_list);
13742 while (invlist_iternext(cp_list, &start, &end)) {
13746 if (end == UV_MAX && start <= 256) {
13747 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
13750 /* Quit if are above what we should change */
13755 change_invlist = TRUE;
13757 /* Set all the bits in the range, up to the max that we are doing */
13758 high = (end < 255) ? end : 255;
13759 for (i = start; i <= (int) high; i++) {
13760 if (! ANYOF_BITMAP_TEST(ret, i)) {
13761 ANYOF_BITMAP_SET(ret, i);
13767 invlist_iterfinish(cp_list);
13769 /* Done with loop; remove any code points that are in the bitmap from
13771 if (change_invlist) {
13772 _invlist_subtract(cp_list, PL_Latin1, &cp_list);
13775 /* If have completely emptied it, remove it completely */
13776 if (_invlist_len(cp_list) == 0) {
13777 SvREFCNT_dec_NN(cp_list);
13783 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
13786 /* Here, the bitmap has been populated with all the Latin1 code points that
13787 * always match. Can now add to the overall list those that match only
13788 * when the target string is UTF-8 (<depends_list>). */
13789 if (depends_list) {
13791 _invlist_union(cp_list, depends_list, &cp_list);
13792 SvREFCNT_dec_NN(depends_list);
13795 cp_list = depends_list;
13799 /* If there is a swash and more than one element, we can't use the swash in
13800 * the optimization below. */
13801 if (swash && element_count > 1) {
13802 SvREFCNT_dec_NN(swash);
13807 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13809 ARG_SET(ret, ANYOF_NONBITMAP_EMPTY);
13812 /* av[0] stores the character class description in its textual form:
13813 * used later (regexec.c:Perl_regclass_swash()) to initialize the
13814 * appropriate swash, and is also useful for dumping the regnode.
13815 * av[1] if NULL, is a placeholder to later contain the swash computed
13816 * from av[0]. But if no further computation need be done, the
13817 * swash is stored there now.
13818 * av[2] stores the cp_list inversion list for use in addition or
13819 * instead of av[0]; used only if av[1] is NULL
13820 * av[3] is set if any component of the class is from a user-defined
13821 * property; used only if av[1] is NULL */
13822 AV * const av = newAV();
13825 av_store(av, 0, (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13826 ? SvREFCNT_inc(listsv) : &PL_sv_undef);
13828 av_store(av, 1, swash);
13829 SvREFCNT_dec_NN(cp_list);
13832 av_store(av, 1, NULL);
13834 av_store(av, 2, cp_list);
13835 av_store(av, 3, newSVuv(has_user_defined_property));
13839 rv = newRV_noinc(MUTABLE_SV(av));
13840 n = add_data(pRExC_state, 1, "s");
13841 RExC_rxi->data->data[n] = (void*)rv;
13845 *flagp |= HASWIDTH|SIMPLE;
13848 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
13851 /* reg_skipcomment()
13853 Absorbs an /x style # comments from the input stream.
13854 Returns true if there is more text remaining in the stream.
13855 Will set the REG_SEEN_RUN_ON_COMMENT flag if the comment
13856 terminates the pattern without including a newline.
13858 Note its the callers responsibility to ensure that we are
13859 actually in /x mode
13864 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
13868 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
13870 while (RExC_parse < RExC_end)
13871 if (*RExC_parse++ == '\n') {
13876 /* we ran off the end of the pattern without ending
13877 the comment, so we have to add an \n when wrapping */
13878 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
13886 Advances the parse position, and optionally absorbs
13887 "whitespace" from the inputstream.
13889 Without /x "whitespace" means (?#...) style comments only,
13890 with /x this means (?#...) and # comments and whitespace proper.
13892 Returns the RExC_parse point from BEFORE the scan occurs.
13894 This is the /x friendly way of saying RExC_parse++.
13898 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
13900 char* const retval = RExC_parse++;
13902 PERL_ARGS_ASSERT_NEXTCHAR;
13905 if (RExC_end - RExC_parse >= 3
13906 && *RExC_parse == '('
13907 && RExC_parse[1] == '?'
13908 && RExC_parse[2] == '#')
13910 while (*RExC_parse != ')') {
13911 if (RExC_parse == RExC_end)
13912 FAIL("Sequence (?#... not terminated");
13918 if (RExC_flags & RXf_PMf_EXTENDED) {
13919 if (isSPACE(*RExC_parse)) {
13923 else if (*RExC_parse == '#') {
13924 if ( reg_skipcomment( pRExC_state ) )
13933 - reg_node - emit a node
13935 STATIC regnode * /* Location. */
13936 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
13940 regnode * const ret = RExC_emit;
13941 GET_RE_DEBUG_FLAGS_DECL;
13943 PERL_ARGS_ASSERT_REG_NODE;
13946 SIZE_ALIGN(RExC_size);
13950 if (RExC_emit >= RExC_emit_bound)
13951 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
13952 op, RExC_emit, RExC_emit_bound);
13954 NODE_ALIGN_FILL(ret);
13956 FILL_ADVANCE_NODE(ptr, op);
13957 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (ptr) - 1);
13958 #ifdef RE_TRACK_PATTERN_OFFSETS
13959 if (RExC_offsets) { /* MJD */
13960 MJD_OFFSET_DEBUG(("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
13961 "reg_node", __LINE__,
13963 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
13964 ? "Overwriting end of array!\n" : "OK",
13965 (UV)(RExC_emit - RExC_emit_start),
13966 (UV)(RExC_parse - RExC_start),
13967 (UV)RExC_offsets[0]));
13968 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
13976 - reganode - emit a node with an argument
13978 STATIC regnode * /* Location. */
13979 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
13983 regnode * const ret = RExC_emit;
13984 GET_RE_DEBUG_FLAGS_DECL;
13986 PERL_ARGS_ASSERT_REGANODE;
13989 SIZE_ALIGN(RExC_size);
13994 assert(2==regarglen[op]+1);
13996 Anything larger than this has to allocate the extra amount.
13997 If we changed this to be:
13999 RExC_size += (1 + regarglen[op]);
14001 then it wouldn't matter. Its not clear what side effect
14002 might come from that so its not done so far.
14007 if (RExC_emit >= RExC_emit_bound)
14008 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14009 op, RExC_emit, RExC_emit_bound);
14011 NODE_ALIGN_FILL(ret);
14013 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
14014 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (ptr) - 2);
14015 #ifdef RE_TRACK_PATTERN_OFFSETS
14016 if (RExC_offsets) { /* MJD */
14017 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14021 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
14022 "Overwriting end of array!\n" : "OK",
14023 (UV)(RExC_emit - RExC_emit_start),
14024 (UV)(RExC_parse - RExC_start),
14025 (UV)RExC_offsets[0]));
14026 Set_Cur_Node_Offset;
14034 - reguni - emit (if appropriate) a Unicode character
14037 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
14041 PERL_ARGS_ASSERT_REGUNI;
14043 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
14047 - reginsert - insert an operator in front of already-emitted operand
14049 * Means relocating the operand.
14052 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
14058 const int offset = regarglen[(U8)op];
14059 const int size = NODE_STEP_REGNODE + offset;
14060 GET_RE_DEBUG_FLAGS_DECL;
14062 PERL_ARGS_ASSERT_REGINSERT;
14063 PERL_UNUSED_ARG(depth);
14064 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
14065 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
14074 if (RExC_open_parens) {
14076 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
14077 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
14078 if ( RExC_open_parens[paren] >= opnd ) {
14079 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
14080 RExC_open_parens[paren] += size;
14082 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
14084 if ( RExC_close_parens[paren] >= opnd ) {
14085 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
14086 RExC_close_parens[paren] += size;
14088 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
14093 while (src > opnd) {
14094 StructCopy(--src, --dst, regnode);
14095 #ifdef RE_TRACK_PATTERN_OFFSETS
14096 if (RExC_offsets) { /* MJD 20010112 */
14097 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
14101 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
14102 ? "Overwriting end of array!\n" : "OK",
14103 (UV)(src - RExC_emit_start),
14104 (UV)(dst - RExC_emit_start),
14105 (UV)RExC_offsets[0]));
14106 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
14107 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
14113 place = opnd; /* Op node, where operand used to be. */
14114 #ifdef RE_TRACK_PATTERN_OFFSETS
14115 if (RExC_offsets) { /* MJD */
14116 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14120 (UV)(place - RExC_emit_start) > RExC_offsets[0]
14121 ? "Overwriting end of array!\n" : "OK",
14122 (UV)(place - RExC_emit_start),
14123 (UV)(RExC_parse - RExC_start),
14124 (UV)RExC_offsets[0]));
14125 Set_Node_Offset(place, RExC_parse);
14126 Set_Node_Length(place, 1);
14129 src = NEXTOPER(place);
14130 FILL_ADVANCE_NODE(place, op);
14131 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (place) - 1);
14132 Zero(src, offset, regnode);
14136 - regtail - set the next-pointer at the end of a node chain of p to val.
14137 - SEE ALSO: regtail_study
14139 /* TODO: All three parms should be const */
14141 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14145 GET_RE_DEBUG_FLAGS_DECL;
14147 PERL_ARGS_ASSERT_REGTAIL;
14149 PERL_UNUSED_ARG(depth);
14155 /* Find last node. */
14158 regnode * const temp = regnext(scan);
14160 SV * const mysv=sv_newmortal();
14161 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
14162 regprop(RExC_rx, mysv, scan);
14163 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
14164 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
14165 (temp == NULL ? "->" : ""),
14166 (temp == NULL ? PL_reg_name[OP(val)] : "")
14174 if (reg_off_by_arg[OP(scan)]) {
14175 ARG_SET(scan, val - scan);
14178 NEXT_OFF(scan) = val - scan;
14184 - regtail_study - set the next-pointer at the end of a node chain of p to val.
14185 - Look for optimizable sequences at the same time.
14186 - currently only looks for EXACT chains.
14188 This is experimental code. The idea is to use this routine to perform
14189 in place optimizations on branches and groups as they are constructed,
14190 with the long term intention of removing optimization from study_chunk so
14191 that it is purely analytical.
14193 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
14194 to control which is which.
14197 /* TODO: All four parms should be const */
14200 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14205 #ifdef EXPERIMENTAL_INPLACESCAN
14208 GET_RE_DEBUG_FLAGS_DECL;
14210 PERL_ARGS_ASSERT_REGTAIL_STUDY;
14216 /* Find last node. */
14220 regnode * const temp = regnext(scan);
14221 #ifdef EXPERIMENTAL_INPLACESCAN
14222 if (PL_regkind[OP(scan)] == EXACT) {
14223 bool has_exactf_sharp_s; /* Unexamined in this routine */
14224 if (join_exact(pRExC_state,scan,&min, &has_exactf_sharp_s, 1,val,depth+1))
14229 switch (OP(scan)) {
14235 case EXACTFU_TRICKYFOLD:
14237 if( exact == PSEUDO )
14239 else if ( exact != OP(scan) )
14248 SV * const mysv=sv_newmortal();
14249 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
14250 regprop(RExC_rx, mysv, scan);
14251 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
14252 SvPV_nolen_const(mysv),
14253 REG_NODE_NUM(scan),
14254 PL_reg_name[exact]);
14261 SV * const mysv_val=sv_newmortal();
14262 DEBUG_PARSE_MSG("");
14263 regprop(RExC_rx, mysv_val, val);
14264 PerlIO_printf(Perl_debug_log, "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
14265 SvPV_nolen_const(mysv_val),
14266 (IV)REG_NODE_NUM(val),
14270 if (reg_off_by_arg[OP(scan)]) {
14271 ARG_SET(scan, val - scan);
14274 NEXT_OFF(scan) = val - scan;
14282 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
14286 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
14292 for (bit=0; bit<32; bit++) {
14293 if (flags & (1<<bit)) {
14294 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
14297 if (!set++ && lead)
14298 PerlIO_printf(Perl_debug_log, "%s",lead);
14299 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
14302 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
14303 if (!set++ && lead) {
14304 PerlIO_printf(Perl_debug_log, "%s",lead);
14307 case REGEX_UNICODE_CHARSET:
14308 PerlIO_printf(Perl_debug_log, "UNICODE");
14310 case REGEX_LOCALE_CHARSET:
14311 PerlIO_printf(Perl_debug_log, "LOCALE");
14313 case REGEX_ASCII_RESTRICTED_CHARSET:
14314 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
14316 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
14317 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
14320 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
14326 PerlIO_printf(Perl_debug_log, "\n");
14328 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14334 Perl_regdump(pTHX_ const regexp *r)
14338 SV * const sv = sv_newmortal();
14339 SV *dsv= sv_newmortal();
14340 RXi_GET_DECL(r,ri);
14341 GET_RE_DEBUG_FLAGS_DECL;
14343 PERL_ARGS_ASSERT_REGDUMP;
14345 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
14347 /* Header fields of interest. */
14348 if (r->anchored_substr) {
14349 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
14350 RE_SV_DUMPLEN(r->anchored_substr), 30);
14351 PerlIO_printf(Perl_debug_log,
14352 "anchored %s%s at %"IVdf" ",
14353 s, RE_SV_TAIL(r->anchored_substr),
14354 (IV)r->anchored_offset);
14355 } else if (r->anchored_utf8) {
14356 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
14357 RE_SV_DUMPLEN(r->anchored_utf8), 30);
14358 PerlIO_printf(Perl_debug_log,
14359 "anchored utf8 %s%s at %"IVdf" ",
14360 s, RE_SV_TAIL(r->anchored_utf8),
14361 (IV)r->anchored_offset);
14363 if (r->float_substr) {
14364 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
14365 RE_SV_DUMPLEN(r->float_substr), 30);
14366 PerlIO_printf(Perl_debug_log,
14367 "floating %s%s at %"IVdf"..%"UVuf" ",
14368 s, RE_SV_TAIL(r->float_substr),
14369 (IV)r->float_min_offset, (UV)r->float_max_offset);
14370 } else if (r->float_utf8) {
14371 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
14372 RE_SV_DUMPLEN(r->float_utf8), 30);
14373 PerlIO_printf(Perl_debug_log,
14374 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
14375 s, RE_SV_TAIL(r->float_utf8),
14376 (IV)r->float_min_offset, (UV)r->float_max_offset);
14378 if (r->check_substr || r->check_utf8)
14379 PerlIO_printf(Perl_debug_log,
14381 (r->check_substr == r->float_substr
14382 && r->check_utf8 == r->float_utf8
14383 ? "(checking floating" : "(checking anchored"));
14384 if (r->extflags & RXf_NOSCAN)
14385 PerlIO_printf(Perl_debug_log, " noscan");
14386 if (r->extflags & RXf_CHECK_ALL)
14387 PerlIO_printf(Perl_debug_log, " isall");
14388 if (r->check_substr || r->check_utf8)
14389 PerlIO_printf(Perl_debug_log, ") ");
14391 if (ri->regstclass) {
14392 regprop(r, sv, ri->regstclass);
14393 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
14395 if (r->extflags & RXf_ANCH) {
14396 PerlIO_printf(Perl_debug_log, "anchored");
14397 if (r->extflags & RXf_ANCH_BOL)
14398 PerlIO_printf(Perl_debug_log, "(BOL)");
14399 if (r->extflags & RXf_ANCH_MBOL)
14400 PerlIO_printf(Perl_debug_log, "(MBOL)");
14401 if (r->extflags & RXf_ANCH_SBOL)
14402 PerlIO_printf(Perl_debug_log, "(SBOL)");
14403 if (r->extflags & RXf_ANCH_GPOS)
14404 PerlIO_printf(Perl_debug_log, "(GPOS)");
14405 PerlIO_putc(Perl_debug_log, ' ');
14407 if (r->extflags & RXf_GPOS_SEEN)
14408 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
14409 if (r->intflags & PREGf_SKIP)
14410 PerlIO_printf(Perl_debug_log, "plus ");
14411 if (r->intflags & PREGf_IMPLICIT)
14412 PerlIO_printf(Perl_debug_log, "implicit ");
14413 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
14414 if (r->extflags & RXf_EVAL_SEEN)
14415 PerlIO_printf(Perl_debug_log, "with eval ");
14416 PerlIO_printf(Perl_debug_log, "\n");
14417 DEBUG_FLAGS_r(regdump_extflags("r->extflags: ",r->extflags));
14419 PERL_ARGS_ASSERT_REGDUMP;
14420 PERL_UNUSED_CONTEXT;
14421 PERL_UNUSED_ARG(r);
14422 #endif /* DEBUGGING */
14426 - regprop - printable representation of opcode
14428 #define EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags) \
14431 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]); \
14432 if (flags & ANYOF_INVERT) \
14433 /*make sure the invert info is in each */ \
14434 sv_catpvs(sv, "^"); \
14440 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
14446 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
14447 static const char * const anyofs[] = {
14448 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
14449 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
14450 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
14451 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
14452 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
14453 || _CC_VERTSPACE != 16
14454 #error Need to adjust order of anyofs[]
14491 RXi_GET_DECL(prog,progi);
14492 GET_RE_DEBUG_FLAGS_DECL;
14494 PERL_ARGS_ASSERT_REGPROP;
14498 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
14499 /* It would be nice to FAIL() here, but this may be called from
14500 regexec.c, and it would be hard to supply pRExC_state. */
14501 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(o), (int)REGNODE_MAX);
14502 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
14504 k = PL_regkind[OP(o)];
14507 sv_catpvs(sv, " ");
14508 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
14509 * is a crude hack but it may be the best for now since
14510 * we have no flag "this EXACTish node was UTF-8"
14512 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
14513 PERL_PV_ESCAPE_UNI_DETECT |
14514 PERL_PV_ESCAPE_NONASCII |
14515 PERL_PV_PRETTY_ELLIPSES |
14516 PERL_PV_PRETTY_LTGT |
14517 PERL_PV_PRETTY_NOCLEAR
14519 } else if (k == TRIE) {
14520 /* print the details of the trie in dumpuntil instead, as
14521 * progi->data isn't available here */
14522 const char op = OP(o);
14523 const U32 n = ARG(o);
14524 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
14525 (reg_ac_data *)progi->data->data[n] :
14527 const reg_trie_data * const trie
14528 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
14530 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
14531 DEBUG_TRIE_COMPILE_r(
14532 Perl_sv_catpvf(aTHX_ sv,
14533 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
14534 (UV)trie->startstate,
14535 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
14536 (UV)trie->wordcount,
14539 (UV)TRIE_CHARCOUNT(trie),
14540 (UV)trie->uniquecharcount
14543 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
14545 int rangestart = -1;
14546 U8* bitmap = IS_ANYOF_TRIE(op) ? (U8*)ANYOF_BITMAP(o) : (U8*)TRIE_BITMAP(trie);
14547 sv_catpvs(sv, "[");
14548 for (i = 0; i <= 256; i++) {
14549 if (i < 256 && BITMAP_TEST(bitmap,i)) {
14550 if (rangestart == -1)
14552 } else if (rangestart != -1) {
14553 if (i <= rangestart + 3)
14554 for (; rangestart < i; rangestart++)
14555 put_byte(sv, rangestart);
14557 put_byte(sv, rangestart);
14558 sv_catpvs(sv, "-");
14559 put_byte(sv, i - 1);
14564 sv_catpvs(sv, "]");
14567 } else if (k == CURLY) {
14568 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
14569 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
14570 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
14572 else if (k == WHILEM && o->flags) /* Ordinal/of */
14573 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
14574 else if (k == REF || k == OPEN || k == CLOSE || k == GROUPP || OP(o)==ACCEPT) {
14575 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
14576 if ( RXp_PAREN_NAMES(prog) ) {
14577 if ( k != REF || (OP(o) < NREF)) {
14578 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
14579 SV **name= av_fetch(list, ARG(o), 0 );
14581 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14584 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
14585 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
14586 I32 *nums=(I32*)SvPVX(sv_dat);
14587 SV **name= av_fetch(list, nums[0], 0 );
14590 for ( n=0; n<SvIVX(sv_dat); n++ ) {
14591 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
14592 (n ? "," : ""), (IV)nums[n]);
14594 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14598 } else if (k == GOSUB)
14599 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o)); /* Paren and offset */
14600 else if (k == VERB) {
14602 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
14603 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
14604 } else if (k == LOGICAL)
14605 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* 2: embedded, otherwise 1 */
14606 else if (k == ANYOF) {
14607 int i, rangestart = -1;
14608 const U8 flags = ANYOF_FLAGS(o);
14612 if (flags & ANYOF_LOCALE)
14613 sv_catpvs(sv, "{loc}");
14614 if (flags & ANYOF_LOC_FOLD)
14615 sv_catpvs(sv, "{i}");
14616 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
14617 if (flags & ANYOF_INVERT)
14618 sv_catpvs(sv, "^");
14620 /* output what the standard cp 0-255 bitmap matches */
14621 for (i = 0; i <= 256; i++) {
14622 if (i < 256 && ANYOF_BITMAP_TEST(o,i)) {
14623 if (rangestart == -1)
14625 } else if (rangestart != -1) {
14626 if (i <= rangestart + 3)
14627 for (; rangestart < i; rangestart++)
14628 put_byte(sv, rangestart);
14630 put_byte(sv, rangestart);
14631 sv_catpvs(sv, "-");
14632 put_byte(sv, i - 1);
14639 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14640 /* output any special charclass tests (used entirely under use locale) */
14641 if (ANYOF_CLASS_TEST_ANY_SET(o))
14642 for (i = 0; i < (int)(sizeof(anyofs)/sizeof(char*)); i++)
14643 if (ANYOF_CLASS_TEST(o,i)) {
14644 sv_catpv(sv, anyofs[i]);
14648 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14650 if (flags & ANYOF_NON_UTF8_LATIN1_ALL) {
14651 sv_catpvs(sv, "{non-utf8-latin1-all}");
14654 /* output information about the unicode matching */
14655 if (flags & ANYOF_UNICODE_ALL)
14656 sv_catpvs(sv, "{unicode_all}");
14657 else if (ANYOF_NONBITMAP(o))
14658 sv_catpvs(sv, "{unicode}");
14659 if (flags & ANYOF_NONBITMAP_NON_UTF8)
14660 sv_catpvs(sv, "{outside bitmap}");
14662 if (ANYOF_NONBITMAP(o)) {
14663 SV *lv; /* Set if there is something outside the bit map */
14664 SV * const sw = regclass_swash(prog, o, FALSE, &lv, NULL);
14665 bool byte_output = FALSE; /* If something in the bitmap has been
14668 if (lv && lv != &PL_sv_undef) {
14670 U8 s[UTF8_MAXBYTES_CASE+1];
14672 for (i = 0; i <= 256; i++) { /* Look at chars in bitmap */
14673 uvchr_to_utf8(s, i);
14676 && ! ANYOF_BITMAP_TEST(o, i) /* Don't duplicate
14680 && swash_fetch(sw, s, TRUE))
14682 if (rangestart == -1)
14684 } else if (rangestart != -1) {
14685 byte_output = TRUE;
14686 if (i <= rangestart + 3)
14687 for (; rangestart < i; rangestart++) {
14688 put_byte(sv, rangestart);
14691 put_byte(sv, rangestart);
14692 sv_catpvs(sv, "-");
14701 char *s = savesvpv(lv);
14702 char * const origs = s;
14704 while (*s && *s != '\n')
14708 const char * const t = ++s;
14711 sv_catpvs(sv, " ");
14717 /* Truncate very long output */
14718 if (s - origs > 256) {
14719 Perl_sv_catpvf(aTHX_ sv,
14721 (int) (s - origs - 1),
14727 else if (*s == '\t') {
14742 SvREFCNT_dec_NN(lv);
14746 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
14748 else if (k == POSIXD || k == NPOSIXD) {
14749 U8 index = FLAGS(o) * 2;
14750 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
14751 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
14754 sv_catpv(sv, anyofs[index]);
14757 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
14758 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
14760 PERL_UNUSED_CONTEXT;
14761 PERL_UNUSED_ARG(sv);
14762 PERL_UNUSED_ARG(o);
14763 PERL_UNUSED_ARG(prog);
14764 #endif /* DEBUGGING */
14768 Perl_re_intuit_string(pTHX_ REGEXP * const r)
14769 { /* Assume that RE_INTUIT is set */
14771 struct regexp *const prog = ReANY(r);
14772 GET_RE_DEBUG_FLAGS_DECL;
14774 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
14775 PERL_UNUSED_CONTEXT;
14779 const char * const s = SvPV_nolen_const(prog->check_substr
14780 ? prog->check_substr : prog->check_utf8);
14782 if (!PL_colorset) reginitcolors();
14783 PerlIO_printf(Perl_debug_log,
14784 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
14786 prog->check_substr ? "" : "utf8 ",
14787 PL_colors[5],PL_colors[0],
14790 (strlen(s) > 60 ? "..." : ""));
14793 return prog->check_substr ? prog->check_substr : prog->check_utf8;
14799 handles refcounting and freeing the perl core regexp structure. When
14800 it is necessary to actually free the structure the first thing it
14801 does is call the 'free' method of the regexp_engine associated to
14802 the regexp, allowing the handling of the void *pprivate; member
14803 first. (This routine is not overridable by extensions, which is why
14804 the extensions free is called first.)
14806 See regdupe and regdupe_internal if you change anything here.
14808 #ifndef PERL_IN_XSUB_RE
14810 Perl_pregfree(pTHX_ REGEXP *r)
14816 Perl_pregfree2(pTHX_ REGEXP *rx)
14819 struct regexp *const r = ReANY(rx);
14820 GET_RE_DEBUG_FLAGS_DECL;
14822 PERL_ARGS_ASSERT_PREGFREE2;
14824 if (r->mother_re) {
14825 ReREFCNT_dec(r->mother_re);
14827 CALLREGFREE_PVT(rx); /* free the private data */
14828 SvREFCNT_dec(RXp_PAREN_NAMES(r));
14829 Safefree(r->xpv_len_u.xpvlenu_pv);
14832 SvREFCNT_dec(r->anchored_substr);
14833 SvREFCNT_dec(r->anchored_utf8);
14834 SvREFCNT_dec(r->float_substr);
14835 SvREFCNT_dec(r->float_utf8);
14836 Safefree(r->substrs);
14838 RX_MATCH_COPY_FREE(rx);
14839 #ifdef PERL_ANY_COW
14840 SvREFCNT_dec(r->saved_copy);
14843 SvREFCNT_dec(r->qr_anoncv);
14844 rx->sv_u.svu_rx = 0;
14849 This is a hacky workaround to the structural issue of match results
14850 being stored in the regexp structure which is in turn stored in
14851 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
14852 could be PL_curpm in multiple contexts, and could require multiple
14853 result sets being associated with the pattern simultaneously, such
14854 as when doing a recursive match with (??{$qr})
14856 The solution is to make a lightweight copy of the regexp structure
14857 when a qr// is returned from the code executed by (??{$qr}) this
14858 lightweight copy doesn't actually own any of its data except for
14859 the starp/end and the actual regexp structure itself.
14865 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
14867 struct regexp *ret;
14868 struct regexp *const r = ReANY(rx);
14869 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
14871 PERL_ARGS_ASSERT_REG_TEMP_COPY;
14874 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
14876 SvOK_off((SV *)ret_x);
14878 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
14879 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
14880 made both spots point to the same regexp body.) */
14881 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
14882 assert(!SvPVX(ret_x));
14883 ret_x->sv_u.svu_rx = temp->sv_any;
14884 temp->sv_any = NULL;
14885 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
14886 SvREFCNT_dec_NN(temp);
14887 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
14888 ing below will not set it. */
14889 SvCUR_set(ret_x, SvCUR(rx));
14892 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
14893 sv_force_normal(sv) is called. */
14895 ret = ReANY(ret_x);
14897 SvFLAGS(ret_x) |= SvUTF8(rx);
14898 /* We share the same string buffer as the original regexp, on which we
14899 hold a reference count, incremented when mother_re is set below.
14900 The string pointer is copied here, being part of the regexp struct.
14902 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
14903 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
14905 const I32 npar = r->nparens+1;
14906 Newx(ret->offs, npar, regexp_paren_pair);
14907 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
14910 Newx(ret->substrs, 1, struct reg_substr_data);
14911 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
14913 SvREFCNT_inc_void(ret->anchored_substr);
14914 SvREFCNT_inc_void(ret->anchored_utf8);
14915 SvREFCNT_inc_void(ret->float_substr);
14916 SvREFCNT_inc_void(ret->float_utf8);
14918 /* check_substr and check_utf8, if non-NULL, point to either their
14919 anchored or float namesakes, and don't hold a second reference. */
14921 RX_MATCH_COPIED_off(ret_x);
14922 #ifdef PERL_ANY_COW
14923 ret->saved_copy = NULL;
14925 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
14926 SvREFCNT_inc_void(ret->qr_anoncv);
14932 /* regfree_internal()
14934 Free the private data in a regexp. This is overloadable by
14935 extensions. Perl takes care of the regexp structure in pregfree(),
14936 this covers the *pprivate pointer which technically perl doesn't
14937 know about, however of course we have to handle the
14938 regexp_internal structure when no extension is in use.
14940 Note this is called before freeing anything in the regexp
14945 Perl_regfree_internal(pTHX_ REGEXP * const rx)
14948 struct regexp *const r = ReANY(rx);
14949 RXi_GET_DECL(r,ri);
14950 GET_RE_DEBUG_FLAGS_DECL;
14952 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
14958 SV *dsv= sv_newmortal();
14959 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
14960 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
14961 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
14962 PL_colors[4],PL_colors[5],s);
14965 #ifdef RE_TRACK_PATTERN_OFFSETS
14967 Safefree(ri->u.offsets); /* 20010421 MJD */
14969 if (ri->code_blocks) {
14971 for (n = 0; n < ri->num_code_blocks; n++)
14972 SvREFCNT_dec(ri->code_blocks[n].src_regex);
14973 Safefree(ri->code_blocks);
14977 int n = ri->data->count;
14980 /* If you add a ->what type here, update the comment in regcomp.h */
14981 switch (ri->data->what[n]) {
14987 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
14990 Safefree(ri->data->data[n]);
14996 { /* Aho Corasick add-on structure for a trie node.
14997 Used in stclass optimization only */
14999 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
15001 refcount = --aho->refcount;
15004 PerlMemShared_free(aho->states);
15005 PerlMemShared_free(aho->fail);
15006 /* do this last!!!! */
15007 PerlMemShared_free(ri->data->data[n]);
15008 PerlMemShared_free(ri->regstclass);
15014 /* trie structure. */
15016 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
15018 refcount = --trie->refcount;
15021 PerlMemShared_free(trie->charmap);
15022 PerlMemShared_free(trie->states);
15023 PerlMemShared_free(trie->trans);
15025 PerlMemShared_free(trie->bitmap);
15027 PerlMemShared_free(trie->jump);
15028 PerlMemShared_free(trie->wordinfo);
15029 /* do this last!!!! */
15030 PerlMemShared_free(ri->data->data[n]);
15035 Perl_croak(aTHX_ "panic: regfree data code '%c'", ri->data->what[n]);
15038 Safefree(ri->data->what);
15039 Safefree(ri->data);
15045 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
15046 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
15047 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
15050 re_dup - duplicate a regexp.
15052 This routine is expected to clone a given regexp structure. It is only
15053 compiled under USE_ITHREADS.
15055 After all of the core data stored in struct regexp is duplicated
15056 the regexp_engine.dupe method is used to copy any private data
15057 stored in the *pprivate pointer. This allows extensions to handle
15058 any duplication it needs to do.
15060 See pregfree() and regfree_internal() if you change anything here.
15062 #if defined(USE_ITHREADS)
15063 #ifndef PERL_IN_XSUB_RE
15065 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
15069 const struct regexp *r = ReANY(sstr);
15070 struct regexp *ret = ReANY(dstr);
15072 PERL_ARGS_ASSERT_RE_DUP_GUTS;
15074 npar = r->nparens+1;
15075 Newx(ret->offs, npar, regexp_paren_pair);
15076 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15078 if (ret->substrs) {
15079 /* Do it this way to avoid reading from *r after the StructCopy().
15080 That way, if any of the sv_dup_inc()s dislodge *r from the L1
15081 cache, it doesn't matter. */
15082 const bool anchored = r->check_substr
15083 ? r->check_substr == r->anchored_substr
15084 : r->check_utf8 == r->anchored_utf8;
15085 Newx(ret->substrs, 1, struct reg_substr_data);
15086 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15088 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
15089 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
15090 ret->float_substr = sv_dup_inc(ret->float_substr, param);
15091 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
15093 /* check_substr and check_utf8, if non-NULL, point to either their
15094 anchored or float namesakes, and don't hold a second reference. */
15096 if (ret->check_substr) {
15098 assert(r->check_utf8 == r->anchored_utf8);
15099 ret->check_substr = ret->anchored_substr;
15100 ret->check_utf8 = ret->anchored_utf8;
15102 assert(r->check_substr == r->float_substr);
15103 assert(r->check_utf8 == r->float_utf8);
15104 ret->check_substr = ret->float_substr;
15105 ret->check_utf8 = ret->float_utf8;
15107 } else if (ret->check_utf8) {
15109 ret->check_utf8 = ret->anchored_utf8;
15111 ret->check_utf8 = ret->float_utf8;
15116 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
15117 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
15120 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
15122 if (RX_MATCH_COPIED(dstr))
15123 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
15125 ret->subbeg = NULL;
15126 #ifdef PERL_ANY_COW
15127 ret->saved_copy = NULL;
15130 /* Whether mother_re be set or no, we need to copy the string. We
15131 cannot refrain from copying it when the storage points directly to
15132 our mother regexp, because that's
15133 1: a buffer in a different thread
15134 2: something we no longer hold a reference on
15135 so we need to copy it locally. */
15136 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
15137 ret->mother_re = NULL;
15140 #endif /* PERL_IN_XSUB_RE */
15145 This is the internal complement to regdupe() which is used to copy
15146 the structure pointed to by the *pprivate pointer in the regexp.
15147 This is the core version of the extension overridable cloning hook.
15148 The regexp structure being duplicated will be copied by perl prior
15149 to this and will be provided as the regexp *r argument, however
15150 with the /old/ structures pprivate pointer value. Thus this routine
15151 may override any copying normally done by perl.
15153 It returns a pointer to the new regexp_internal structure.
15157 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
15160 struct regexp *const r = ReANY(rx);
15161 regexp_internal *reti;
15163 RXi_GET_DECL(r,ri);
15165 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
15169 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode), char, regexp_internal);
15170 Copy(ri->program, reti->program, len+1, regnode);
15172 reti->num_code_blocks = ri->num_code_blocks;
15173 if (ri->code_blocks) {
15175 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
15176 struct reg_code_block);
15177 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
15178 struct reg_code_block);
15179 for (n = 0; n < ri->num_code_blocks; n++)
15180 reti->code_blocks[n].src_regex = (REGEXP*)
15181 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
15184 reti->code_blocks = NULL;
15186 reti->regstclass = NULL;
15189 struct reg_data *d;
15190 const int count = ri->data->count;
15193 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
15194 char, struct reg_data);
15195 Newx(d->what, count, U8);
15198 for (i = 0; i < count; i++) {
15199 d->what[i] = ri->data->what[i];
15200 switch (d->what[i]) {
15201 /* see also regcomp.h and regfree_internal() */
15202 case 'a': /* actually an AV, but the dup function is identical. */
15206 case 'u': /* actually an HV, but the dup function is identical. */
15207 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
15210 /* This is cheating. */
15211 Newx(d->data[i], 1, struct regnode_charclass_class);
15212 StructCopy(ri->data->data[i], d->data[i],
15213 struct regnode_charclass_class);
15214 reti->regstclass = (regnode*)d->data[i];
15217 /* Trie stclasses are readonly and can thus be shared
15218 * without duplication. We free the stclass in pregfree
15219 * when the corresponding reg_ac_data struct is freed.
15221 reti->regstclass= ri->regstclass;
15225 ((reg_trie_data*)ri->data->data[i])->refcount++;
15230 d->data[i] = ri->data->data[i];
15233 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'", ri->data->what[i]);
15242 reti->name_list_idx = ri->name_list_idx;
15244 #ifdef RE_TRACK_PATTERN_OFFSETS
15245 if (ri->u.offsets) {
15246 Newx(reti->u.offsets, 2*len+1, U32);
15247 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
15250 SetProgLen(reti,len);
15253 return (void*)reti;
15256 #endif /* USE_ITHREADS */
15258 #ifndef PERL_IN_XSUB_RE
15261 - regnext - dig the "next" pointer out of a node
15264 Perl_regnext(pTHX_ regnode *p)
15272 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
15273 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(p), (int)REGNODE_MAX);
15276 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
15285 S_re_croak2(pTHX_ const char* pat1,const char* pat2,...)
15288 STRLEN l1 = strlen(pat1);
15289 STRLEN l2 = strlen(pat2);
15292 const char *message;
15294 PERL_ARGS_ASSERT_RE_CROAK2;
15300 Copy(pat1, buf, l1 , char);
15301 Copy(pat2, buf + l1, l2 , char);
15302 buf[l1 + l2] = '\n';
15303 buf[l1 + l2 + 1] = '\0';
15305 /* ANSI variant takes additional second argument */
15306 va_start(args, pat2);
15310 msv = vmess(buf, &args);
15312 message = SvPV_const(msv,l1);
15315 Copy(message, buf, l1 , char);
15316 buf[l1-1] = '\0'; /* Overwrite \n */
15317 Perl_croak(aTHX_ "%s", buf);
15320 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
15322 #ifndef PERL_IN_XSUB_RE
15324 Perl_save_re_context(pTHX)
15328 struct re_save_state *state;
15330 SAVEVPTR(PL_curcop);
15331 SSGROW(SAVESTACK_ALLOC_FOR_RE_SAVE_STATE + 1);
15333 state = (struct re_save_state *)(PL_savestack + PL_savestack_ix);
15334 PL_savestack_ix += SAVESTACK_ALLOC_FOR_RE_SAVE_STATE;
15335 SSPUSHUV(SAVEt_RE_STATE);
15337 Copy(&PL_reg_state, state, 1, struct re_save_state);
15339 PL_reg_oldsaved = NULL;
15340 PL_reg_oldsavedlen = 0;
15341 PL_reg_oldsavedoffset = 0;
15342 PL_reg_oldsavedcoffset = 0;
15343 PL_reg_maxiter = 0;
15344 PL_reg_leftiter = 0;
15345 PL_reg_poscache = NULL;
15346 PL_reg_poscache_size = 0;
15347 #ifdef PERL_ANY_COW
15351 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
15353 const REGEXP * const rx = PM_GETRE(PL_curpm);
15356 for (i = 1; i <= RX_NPARENS(rx); i++) {
15357 char digits[TYPE_CHARS(long)];
15358 const STRLEN len = my_snprintf(digits, sizeof(digits), "%lu", (long)i);
15359 GV *const *const gvp
15360 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
15363 GV * const gv = *gvp;
15364 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
15376 S_put_byte(pTHX_ SV *sv, int c)
15378 PERL_ARGS_ASSERT_PUT_BYTE;
15380 /* Our definition of isPRINT() ignores locales, so only bytes that are
15381 not part of UTF-8 are considered printable. I assume that the same
15382 holds for UTF-EBCDIC.
15383 Also, code point 255 is not printable in either (it's E0 in EBCDIC,
15384 which Wikipedia says:
15386 EO, or Eight Ones, is an 8-bit EBCDIC character code represented as all
15387 ones (binary 1111 1111, hexadecimal FF). It is similar, but not
15388 identical, to the ASCII delete (DEL) or rubout control character. ...
15389 it is typically mapped to hexadecimal code 9F, in order to provide a
15390 unique character mapping in both directions)
15392 So the old condition can be simplified to !isPRINT(c) */
15395 Perl_sv_catpvf(aTHX_ sv, "\\x%02x", c);
15398 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
15402 const char string = c;
15403 if (c == '-' || c == ']' || c == '\\' || c == '^')
15404 sv_catpvs(sv, "\\");
15405 sv_catpvn(sv, &string, 1);
15410 #define CLEAR_OPTSTART \
15411 if (optstart) STMT_START { \
15412 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
15416 #define DUMPUNTIL(b,e) CLEAR_OPTSTART; node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
15418 STATIC const regnode *
15419 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
15420 const regnode *last, const regnode *plast,
15421 SV* sv, I32 indent, U32 depth)
15424 U8 op = PSEUDO; /* Arbitrary non-END op. */
15425 const regnode *next;
15426 const regnode *optstart= NULL;
15428 RXi_GET_DECL(r,ri);
15429 GET_RE_DEBUG_FLAGS_DECL;
15431 PERL_ARGS_ASSERT_DUMPUNTIL;
15433 #ifdef DEBUG_DUMPUNTIL
15434 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
15435 last ? last-start : 0,plast ? plast-start : 0);
15438 if (plast && plast < last)
15441 while (PL_regkind[op] != END && (!last || node < last)) {
15442 /* While that wasn't END last time... */
15445 if (op == CLOSE || op == WHILEM)
15447 next = regnext((regnode *)node);
15450 if (OP(node) == OPTIMIZED) {
15451 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
15458 regprop(r, sv, node);
15459 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
15460 (int)(2*indent + 1), "", SvPVX_const(sv));
15462 if (OP(node) != OPTIMIZED) {
15463 if (next == NULL) /* Next ptr. */
15464 PerlIO_printf(Perl_debug_log, " (0)");
15465 else if (PL_regkind[(U8)op] == BRANCH && PL_regkind[OP(next)] != BRANCH )
15466 PerlIO_printf(Perl_debug_log, " (FAIL)");
15468 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
15469 (void)PerlIO_putc(Perl_debug_log, '\n');
15473 if (PL_regkind[(U8)op] == BRANCHJ) {
15476 const regnode *nnode = (OP(next) == LONGJMP
15477 ? regnext((regnode *)next)
15479 if (last && nnode > last)
15481 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
15484 else if (PL_regkind[(U8)op] == BRANCH) {
15486 DUMPUNTIL(NEXTOPER(node), next);
15488 else if ( PL_regkind[(U8)op] == TRIE ) {
15489 const regnode *this_trie = node;
15490 const char op = OP(node);
15491 const U32 n = ARG(node);
15492 const reg_ac_data * const ac = op>=AHOCORASICK ?
15493 (reg_ac_data *)ri->data->data[n] :
15495 const reg_trie_data * const trie =
15496 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
15498 AV *const trie_words = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
15500 const regnode *nextbranch= NULL;
15503 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
15504 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
15506 PerlIO_printf(Perl_debug_log, "%*s%s ",
15507 (int)(2*(indent+3)), "",
15508 elem_ptr ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr), SvCUR(*elem_ptr), 60,
15509 PL_colors[0], PL_colors[1],
15510 (SvUTF8(*elem_ptr) ? PERL_PV_ESCAPE_UNI : 0) |
15511 PERL_PV_PRETTY_ELLIPSES |
15512 PERL_PV_PRETTY_LTGT
15517 U16 dist= trie->jump[word_idx+1];
15518 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
15519 (UV)((dist ? this_trie + dist : next) - start));
15522 nextbranch= this_trie + trie->jump[0];
15523 DUMPUNTIL(this_trie + dist, nextbranch);
15525 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
15526 nextbranch= regnext((regnode *)nextbranch);
15528 PerlIO_printf(Perl_debug_log, "\n");
15531 if (last && next > last)
15536 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
15537 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
15538 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
15540 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
15542 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
15544 else if ( op == PLUS || op == STAR) {
15545 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
15547 else if (PL_regkind[(U8)op] == ANYOF) {
15548 /* arglen 1 + class block */
15549 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_CLASS)
15550 ? ANYOF_CLASS_SKIP : ANYOF_SKIP);
15551 node = NEXTOPER(node);
15553 else if (PL_regkind[(U8)op] == EXACT) {
15554 /* Literal string, where present. */
15555 node += NODE_SZ_STR(node) - 1;
15556 node = NEXTOPER(node);
15559 node = NEXTOPER(node);
15560 node += regarglen[(U8)op];
15562 if (op == CURLYX || op == OPEN)
15566 #ifdef DEBUG_DUMPUNTIL
15567 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
15572 #endif /* DEBUGGING */
15576 * c-indentation-style: bsd
15577 * c-basic-offset: 4
15578 * indent-tabs-mode: nil
15581 * ex: set ts=8 sts=4 sw=4 et: