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 */
2446 /* Finish populating the prev field of the wordinfo array. Walk back
2447 * from each accept state until we find another accept state, and if
2448 * so, point the first word's .prev field at the second word. If the
2449 * second already has a .prev field set, stop now. This will be the
2450 * case either if we've already processed that word's accept state,
2451 * or that state had multiple words, and the overspill words were
2452 * already linked up earlier.
2459 for (word=1; word <= trie->wordcount; word++) {
2461 if (trie->wordinfo[word].prev)
2463 state = trie->wordinfo[word].accept;
2465 state = prev_states[state];
2468 prev = trie->states[state].wordnum;
2472 trie->wordinfo[word].prev = prev;
2474 Safefree(prev_states);
2478 /* and now dump out the compressed format */
2479 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2481 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2483 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2484 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2486 SvREFCNT_dec_NN(revcharmap);
2490 : trie->startstate>1
2496 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2498 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2500 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2501 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2504 We find the fail state for each state in the trie, this state is the longest proper
2505 suffix of the current state's 'word' that is also a proper prefix of another word in our
2506 trie. State 1 represents the word '' and is thus the default fail state. This allows
2507 the DFA not to have to restart after its tried and failed a word at a given point, it
2508 simply continues as though it had been matching the other word in the first place.
2510 'abcdgu'=~/abcdefg|cdgu/
2511 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2512 fail, which would bring us to the state representing 'd' in the second word where we would
2513 try 'g' and succeed, proceeding to match 'cdgu'.
2515 /* add a fail transition */
2516 const U32 trie_offset = ARG(source);
2517 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2519 const U32 ucharcount = trie->uniquecharcount;
2520 const U32 numstates = trie->statecount;
2521 const U32 ubound = trie->lasttrans + ucharcount;
2525 U32 base = trie->states[ 1 ].trans.base;
2528 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2529 GET_RE_DEBUG_FLAGS_DECL;
2531 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2533 PERL_UNUSED_ARG(depth);
2537 ARG_SET( stclass, data_slot );
2538 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2539 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2540 aho->trie=trie_offset;
2541 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2542 Copy( trie->states, aho->states, numstates, reg_trie_state );
2543 Newxz( q, numstates, U32);
2544 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2547 /* initialize fail[0..1] to be 1 so that we always have
2548 a valid final fail state */
2549 fail[ 0 ] = fail[ 1 ] = 1;
2551 for ( charid = 0; charid < ucharcount ; charid++ ) {
2552 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2554 q[ q_write ] = newstate;
2555 /* set to point at the root */
2556 fail[ q[ q_write++ ] ]=1;
2559 while ( q_read < q_write) {
2560 const U32 cur = q[ q_read++ % numstates ];
2561 base = trie->states[ cur ].trans.base;
2563 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2564 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2566 U32 fail_state = cur;
2569 fail_state = fail[ fail_state ];
2570 fail_base = aho->states[ fail_state ].trans.base;
2571 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2573 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2574 fail[ ch_state ] = fail_state;
2575 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2577 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2579 q[ q_write++ % numstates] = ch_state;
2583 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2584 when we fail in state 1, this allows us to use the
2585 charclass scan to find a valid start char. This is based on the principle
2586 that theres a good chance the string being searched contains lots of stuff
2587 that cant be a start char.
2589 fail[ 0 ] = fail[ 1 ] = 0;
2590 DEBUG_TRIE_COMPILE_r({
2591 PerlIO_printf(Perl_debug_log,
2592 "%*sStclass Failtable (%"UVuf" states): 0",
2593 (int)(depth * 2), "", (UV)numstates
2595 for( q_read=1; q_read<numstates; q_read++ ) {
2596 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2598 PerlIO_printf(Perl_debug_log, "\n");
2601 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2606 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2607 * These need to be revisited when a newer toolchain becomes available.
2609 #if defined(__sparc64__) && defined(__GNUC__)
2610 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2611 # undef SPARC64_GCC_WORKAROUND
2612 # define SPARC64_GCC_WORKAROUND 1
2616 #define DEBUG_PEEP(str,scan,depth) \
2617 DEBUG_OPTIMISE_r({if (scan){ \
2618 SV * const mysv=sv_newmortal(); \
2619 regnode *Next = regnext(scan); \
2620 regprop(RExC_rx, mysv, scan); \
2621 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2622 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2623 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2627 /* The below joins as many adjacent EXACTish nodes as possible into a single
2628 * one. The regop may be changed if the node(s) contain certain sequences that
2629 * require special handling. The joining is only done if:
2630 * 1) there is room in the current conglomerated node to entirely contain the
2632 * 2) they are the exact same node type
2634 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2635 * these get optimized out
2637 * If a node is to match under /i (folded), the number of characters it matches
2638 * can be different than its character length if it contains a multi-character
2639 * fold. *min_subtract is set to the total delta of the input nodes.
2641 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2642 * and contains LATIN SMALL LETTER SHARP S
2644 * This is as good a place as any to discuss the design of handling these
2645 * multi-character fold sequences. It's been wrong in Perl for a very long
2646 * time. There are three code points in Unicode whose multi-character folds
2647 * were long ago discovered to mess things up. The previous designs for
2648 * dealing with these involved assigning a special node for them. This
2649 * approach doesn't work, as evidenced by this example:
2650 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2651 * Both these fold to "sss", but if the pattern is parsed to create a node that
2652 * would match just the \xDF, it won't be able to handle the case where a
2653 * successful match would have to cross the node's boundary. The new approach
2654 * that hopefully generally solves the problem generates an EXACTFU_SS node
2657 * It turns out that there are problems with all multi-character folds, and not
2658 * just these three. Now the code is general, for all such cases, but the
2659 * three still have some special handling. The approach taken is:
2660 * 1) This routine examines each EXACTFish node that could contain multi-
2661 * character fold sequences. It returns in *min_subtract how much to
2662 * subtract from the the actual length of the string to get a real minimum
2663 * match length; it is 0 if there are no multi-char folds. This delta is
2664 * used by the caller to adjust the min length of the match, and the delta
2665 * between min and max, so that the optimizer doesn't reject these
2666 * possibilities based on size constraints.
2667 * 2) Certain of these sequences require special handling by the trie code,
2668 * so, if found, this code changes the joined node type to special ops:
2669 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2670 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2671 * is used for an EXACTFU node that contains at least one "ss" sequence in
2672 * it. For non-UTF-8 patterns and strings, this is the only case where
2673 * there is a possible fold length change. That means that a regular
2674 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2675 * with length changes, and so can be processed faster. regexec.c takes
2676 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2677 * pre-folded by regcomp.c. This saves effort in regex matching.
2678 * However, the pre-folding isn't done for non-UTF8 patterns because the
2679 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2680 * down by forcing the pattern into UTF8 unless necessary. Also what
2681 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2682 * possibilities for the non-UTF8 patterns are quite simple, except for
2683 * the sharp s. All the ones that don't involve a UTF-8 target string are
2684 * members of a fold-pair, and arrays are set up for all of them so that
2685 * the other member of the pair can be found quickly. Code elsewhere in
2686 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2687 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2688 * described in the next item.
2689 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2690 * 'ss' or not is not knowable at compile time. It will match iff the
2691 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2692 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2693 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
2694 * described in item 3). An assumption that the optimizer part of
2695 * regexec.c (probably unwittingly) makes is that a character in the
2696 * pattern corresponds to at most a single character in the target string.
2697 * (And I do mean character, and not byte here, unlike other parts of the
2698 * documentation that have never been updated to account for multibyte
2699 * Unicode.) This assumption is wrong only in this case, as all other
2700 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2701 * virtue of having this file pre-fold UTF-8 patterns. I'm
2702 * reluctant to try to change this assumption, so instead the code punts.
2703 * This routine examines EXACTF nodes for the sharp s, and returns a
2704 * boolean indicating whether or not the node is an EXACTF node that
2705 * contains a sharp s. When it is true, the caller sets a flag that later
2706 * causes the optimizer in this file to not set values for the floating
2707 * and fixed string lengths, and thus avoids the optimizer code in
2708 * regexec.c that makes the invalid assumption. Thus, there is no
2709 * optimization based on string lengths for EXACTF nodes that contain the
2710 * sharp s. This only happens for /id rules (which means the pattern
2714 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2715 if (PL_regkind[OP(scan)] == EXACT) \
2716 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2719 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) {
2720 /* Merge several consecutive EXACTish nodes into one. */
2721 regnode *n = regnext(scan);
2723 regnode *next = scan + NODE_SZ_STR(scan);
2727 regnode *stop = scan;
2728 GET_RE_DEBUG_FLAGS_DECL;
2730 PERL_UNUSED_ARG(depth);
2733 PERL_ARGS_ASSERT_JOIN_EXACT;
2734 #ifndef EXPERIMENTAL_INPLACESCAN
2735 PERL_UNUSED_ARG(flags);
2736 PERL_UNUSED_ARG(val);
2738 DEBUG_PEEP("join",scan,depth);
2740 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2741 * EXACT ones that are mergeable to the current one. */
2743 && (PL_regkind[OP(n)] == NOTHING
2744 || (stringok && OP(n) == OP(scan)))
2746 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2749 if (OP(n) == TAIL || n > next)
2751 if (PL_regkind[OP(n)] == NOTHING) {
2752 DEBUG_PEEP("skip:",n,depth);
2753 NEXT_OFF(scan) += NEXT_OFF(n);
2754 next = n + NODE_STEP_REGNODE;
2761 else if (stringok) {
2762 const unsigned int oldl = STR_LEN(scan);
2763 regnode * const nnext = regnext(n);
2765 /* XXX I (khw) kind of doubt that this works on platforms where
2766 * U8_MAX is above 255 because of lots of other assumptions */
2767 /* Don't join if the sum can't fit into a single node */
2768 if (oldl + STR_LEN(n) > U8_MAX)
2771 DEBUG_PEEP("merg",n,depth);
2774 NEXT_OFF(scan) += NEXT_OFF(n);
2775 STR_LEN(scan) += STR_LEN(n);
2776 next = n + NODE_SZ_STR(n);
2777 /* Now we can overwrite *n : */
2778 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2786 #ifdef EXPERIMENTAL_INPLACESCAN
2787 if (flags && !NEXT_OFF(n)) {
2788 DEBUG_PEEP("atch", val, depth);
2789 if (reg_off_by_arg[OP(n)]) {
2790 ARG_SET(n, val - n);
2793 NEXT_OFF(n) = val - n;
2801 *has_exactf_sharp_s = FALSE;
2803 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2804 * can now analyze for sequences of problematic code points. (Prior to
2805 * this final joining, sequences could have been split over boundaries, and
2806 * hence missed). The sequences only happen in folding, hence for any
2807 * non-EXACT EXACTish node */
2808 if (OP(scan) != EXACT) {
2809 const U8 * const s0 = (U8*) STRING(scan);
2811 const U8 * const s_end = s0 + STR_LEN(scan);
2813 /* One pass is made over the node's string looking for all the
2814 * possibilities. to avoid some tests in the loop, there are two main
2815 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2819 /* Examine the string for a multi-character fold sequence. UTF-8
2820 * patterns have all characters pre-folded by the time this code is
2822 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2823 length sequence we are looking for is 2 */
2826 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2827 if (! len) { /* Not a multi-char fold: get next char */
2832 /* Nodes with 'ss' require special handling, except for EXACTFL
2833 * and EXACTFA for which there is no multi-char fold to this */
2834 if (len == 2 && *s == 's' && *(s+1) == 's'
2835 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2838 OP(scan) = EXACTFU_SS;
2841 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2842 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2843 COMBINING_DIAERESIS_UTF8
2844 COMBINING_ACUTE_ACCENT_UTF8,
2846 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2847 COMBINING_DIAERESIS_UTF8
2848 COMBINING_ACUTE_ACCENT_UTF8,
2853 /* These two folds require special handling by trie's, so
2854 * change the node type to indicate this. If EXACTFA and
2855 * EXACTFL were ever to be handled by trie's, this would
2856 * have to be changed. If this node has already been
2857 * changed to EXACTFU_SS in this loop, leave it as is. (I
2858 * (khw) think it doesn't matter in regexec.c for UTF
2859 * patterns, but no need to change it */
2860 if (OP(scan) == EXACTFU) {
2861 OP(scan) = EXACTFU_TRICKYFOLD;
2865 else { /* Here is a generic multi-char fold. */
2866 const U8* multi_end = s + len;
2868 /* Count how many characters in it. In the case of /l and
2869 * /aa, no folds which contain ASCII code points are
2870 * allowed, so check for those, and skip if found. (In
2871 * EXACTFL, no folds are allowed to any Latin1 code point,
2872 * not just ASCII. But there aren't any of these
2873 * currently, nor ever likely, so don't take the time to
2874 * test for them. The code that generates the
2875 * is_MULTI_foo() macros croaks should one actually get put
2876 * into Unicode .) */
2877 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2878 count = utf8_length(s, multi_end);
2882 while (s < multi_end) {
2885 goto next_iteration;
2895 /* The delta is how long the sequence is minus 1 (1 is how long
2896 * the character that folds to the sequence is) */
2897 *min_subtract += count - 1;
2901 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2903 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2904 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2905 * nodes can't have multi-char folds to this range (and there are
2906 * no existing ones in the upper latin1 range). In the EXACTF
2907 * case we look also for the sharp s, which can be in the final
2908 * position. Otherwise we can stop looking 1 byte earlier because
2909 * have to find at least two characters for a multi-fold */
2910 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2912 /* The below is perhaps overboard, but this allows us to save a
2913 * test each time through the loop at the expense of a mask. This
2914 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2915 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2916 * are 64. This uses an exclusive 'or' to find that bit and then
2917 * inverts it to form a mask, with just a single 0, in the bit
2918 * position where 'S' and 's' differ. */
2919 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2920 const U8 s_masked = 's' & S_or_s_mask;
2923 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2924 if (! len) { /* Not a multi-char fold. */
2925 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2927 *has_exactf_sharp_s = TRUE;
2934 && ((*s & S_or_s_mask) == s_masked)
2935 && ((*(s+1) & S_or_s_mask) == s_masked))
2938 /* EXACTF nodes need to know that the minimum length
2939 * changed so that a sharp s in the string can match this
2940 * ss in the pattern, but they remain EXACTF nodes, as they
2941 * won't match this unless the target string is is UTF-8,
2942 * which we don't know until runtime */
2943 if (OP(scan) != EXACTF) {
2944 OP(scan) = EXACTFU_SS;
2948 *min_subtract += len - 1;
2955 /* Allow dumping but overwriting the collection of skipped
2956 * ops and/or strings with fake optimized ops */
2957 n = scan + NODE_SZ_STR(scan);
2965 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2969 /* REx optimizer. Converts nodes into quicker variants "in place".
2970 Finds fixed substrings. */
2972 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2973 to the position after last scanned or to NULL. */
2975 #define INIT_AND_WITHP \
2976 assert(!and_withp); \
2977 Newx(and_withp,1,struct regnode_charclass_class); \
2978 SAVEFREEPV(and_withp)
2980 /* this is a chain of data about sub patterns we are processing that
2981 need to be handled separately/specially in study_chunk. Its so
2982 we can simulate recursion without losing state. */
2984 typedef struct scan_frame {
2985 regnode *last; /* last node to process in this frame */
2986 regnode *next; /* next node to process when last is reached */
2987 struct scan_frame *prev; /*previous frame*/
2988 I32 stop; /* what stopparen do we use */
2992 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2995 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2996 I32 *minlenp, I32 *deltap,
3001 struct regnode_charclass_class *and_withp,
3002 U32 flags, U32 depth)
3003 /* scanp: Start here (read-write). */
3004 /* deltap: Write maxlen-minlen here. */
3005 /* last: Stop before this one. */
3006 /* data: string data about the pattern */
3007 /* stopparen: treat close N as END */
3008 /* recursed: which subroutines have we recursed into */
3009 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3012 I32 min = 0; /* There must be at least this number of characters to match */
3014 regnode *scan = *scanp, *next;
3016 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3017 int is_inf_internal = 0; /* The studied chunk is infinite */
3018 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3019 scan_data_t data_fake;
3020 SV *re_trie_maxbuff = NULL;
3021 regnode *first_non_open = scan;
3022 I32 stopmin = I32_MAX;
3023 scan_frame *frame = NULL;
3024 GET_RE_DEBUG_FLAGS_DECL;
3026 PERL_ARGS_ASSERT_STUDY_CHUNK;
3029 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3033 while (first_non_open && OP(first_non_open) == OPEN)
3034 first_non_open=regnext(first_non_open);
3039 while ( scan && OP(scan) != END && scan < last ){
3040 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3041 node length to get a real minimum (because
3042 the folded version may be shorter) */
3043 bool has_exactf_sharp_s = FALSE;
3044 /* Peephole optimizer: */
3045 DEBUG_STUDYDATA("Peep:", data,depth);
3046 DEBUG_PEEP("Peep",scan,depth);
3048 /* Its not clear to khw or hv why this is done here, and not in the
3049 * clauses that deal with EXACT nodes. khw's guess is that it's
3050 * because of a previous design */
3051 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3053 /* Follow the next-chain of the current node and optimize
3054 away all the NOTHINGs from it. */
3055 if (OP(scan) != CURLYX) {
3056 const int max = (reg_off_by_arg[OP(scan)]
3058 /* I32 may be smaller than U16 on CRAYs! */
3059 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3060 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3064 /* Skip NOTHING and LONGJMP. */
3065 while ((n = regnext(n))
3066 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3067 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3068 && off + noff < max)
3070 if (reg_off_by_arg[OP(scan)])
3073 NEXT_OFF(scan) = off;
3078 /* The principal pseudo-switch. Cannot be a switch, since we
3079 look into several different things. */
3080 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3081 || OP(scan) == IFTHEN) {
3082 next = regnext(scan);
3084 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3086 if (OP(next) == code || code == IFTHEN) {
3087 /* NOTE - There is similar code to this block below for handling
3088 TRIE nodes on a re-study. If you change stuff here check there
3090 I32 max1 = 0, min1 = I32_MAX, num = 0;
3091 struct regnode_charclass_class accum;
3092 regnode * const startbranch=scan;
3094 if (flags & SCF_DO_SUBSTR)
3095 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3096 if (flags & SCF_DO_STCLASS)
3097 cl_init_zero(pRExC_state, &accum);
3099 while (OP(scan) == code) {
3100 I32 deltanext, minnext, f = 0, fake;
3101 struct regnode_charclass_class this_class;
3104 data_fake.flags = 0;
3106 data_fake.whilem_c = data->whilem_c;
3107 data_fake.last_closep = data->last_closep;
3110 data_fake.last_closep = &fake;
3112 data_fake.pos_delta = delta;
3113 next = regnext(scan);
3114 scan = NEXTOPER(scan);
3116 scan = NEXTOPER(scan);
3117 if (flags & SCF_DO_STCLASS) {
3118 cl_init(pRExC_state, &this_class);
3119 data_fake.start_class = &this_class;
3120 f = SCF_DO_STCLASS_AND;
3122 if (flags & SCF_WHILEM_VISITED_POS)
3123 f |= SCF_WHILEM_VISITED_POS;
3125 /* we suppose the run is continuous, last=next...*/
3126 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3128 stopparen, recursed, NULL, f,depth+1);
3131 if (deltanext == I32_MAX) {
3132 is_inf = is_inf_internal = 1;
3134 } else if (max1 < minnext + deltanext)
3135 max1 = minnext + deltanext;
3137 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3139 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3140 if ( stopmin > minnext)
3141 stopmin = min + min1;
3142 flags &= ~SCF_DO_SUBSTR;
3144 data->flags |= SCF_SEEN_ACCEPT;
3147 if (data_fake.flags & SF_HAS_EVAL)
3148 data->flags |= SF_HAS_EVAL;
3149 data->whilem_c = data_fake.whilem_c;
3151 if (flags & SCF_DO_STCLASS)
3152 cl_or(pRExC_state, &accum, &this_class);
3154 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3156 if (flags & SCF_DO_SUBSTR) {
3157 data->pos_min += min1;
3158 if (data->pos_delta >= I32_MAX - (max1 - min1))
3159 data->pos_delta = I32_MAX;
3161 data->pos_delta += max1 - min1;
3162 if (max1 != min1 || is_inf)
3163 data->longest = &(data->longest_float);
3166 if (delta == I32_MAX || I32_MAX - delta - (max1 - min1) < 0)
3169 delta += max1 - min1;
3170 if (flags & SCF_DO_STCLASS_OR) {
3171 cl_or(pRExC_state, data->start_class, &accum);
3173 cl_and(data->start_class, and_withp);
3174 flags &= ~SCF_DO_STCLASS;
3177 else if (flags & SCF_DO_STCLASS_AND) {
3179 cl_and(data->start_class, &accum);
3180 flags &= ~SCF_DO_STCLASS;
3183 /* Switch to OR mode: cache the old value of
3184 * data->start_class */
3186 StructCopy(data->start_class, and_withp,
3187 struct regnode_charclass_class);
3188 flags &= ~SCF_DO_STCLASS_AND;
3189 StructCopy(&accum, data->start_class,
3190 struct regnode_charclass_class);
3191 flags |= SCF_DO_STCLASS_OR;
3192 SET_SSC_EOS(data->start_class);
3196 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3199 Assuming this was/is a branch we are dealing with: 'scan' now
3200 points at the item that follows the branch sequence, whatever
3201 it is. We now start at the beginning of the sequence and look
3208 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3210 If we can find such a subsequence we need to turn the first
3211 element into a trie and then add the subsequent branch exact
3212 strings to the trie.
3216 1. patterns where the whole set of branches can be converted.
3218 2. patterns where only a subset can be converted.
3220 In case 1 we can replace the whole set with a single regop
3221 for the trie. In case 2 we need to keep the start and end
3224 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3225 becomes BRANCH TRIE; BRANCH X;
3227 There is an additional case, that being where there is a
3228 common prefix, which gets split out into an EXACT like node
3229 preceding the TRIE node.
3231 If x(1..n)==tail then we can do a simple trie, if not we make
3232 a "jump" trie, such that when we match the appropriate word
3233 we "jump" to the appropriate tail node. Essentially we turn
3234 a nested if into a case structure of sorts.
3239 if (!re_trie_maxbuff) {
3240 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3241 if (!SvIOK(re_trie_maxbuff))
3242 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3244 if ( SvIV(re_trie_maxbuff)>=0 ) {
3246 regnode *first = (regnode *)NULL;
3247 regnode *last = (regnode *)NULL;
3248 regnode *tail = scan;
3253 SV * const mysv = sv_newmortal(); /* for dumping */
3255 /* var tail is used because there may be a TAIL
3256 regop in the way. Ie, the exacts will point to the
3257 thing following the TAIL, but the last branch will
3258 point at the TAIL. So we advance tail. If we
3259 have nested (?:) we may have to move through several
3263 while ( OP( tail ) == TAIL ) {
3264 /* this is the TAIL generated by (?:) */
3265 tail = regnext( tail );
3269 DEBUG_TRIE_COMPILE_r({
3270 regprop(RExC_rx, mysv, tail );
3271 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3272 (int)depth * 2 + 2, "",
3273 "Looking for TRIE'able sequences. Tail node is: ",
3274 SvPV_nolen_const( mysv )
3280 Step through the branches
3281 cur represents each branch,
3282 noper is the first thing to be matched as part of that branch
3283 noper_next is the regnext() of that node.
3285 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3286 via a "jump trie" but we also support building with NOJUMPTRIE,
3287 which restricts the trie logic to structures like /FOO|BAR/.
3289 If noper is a trieable nodetype then the branch is a possible optimization
3290 target. If we are building under NOJUMPTRIE then we require that noper_next
3291 is the same as scan (our current position in the regex program).
3293 Once we have two or more consecutive such branches we can create a
3294 trie of the EXACT's contents and stitch it in place into the program.
3296 If the sequence represents all of the branches in the alternation we
3297 replace the entire thing with a single TRIE node.
3299 Otherwise when it is a subsequence we need to stitch it in place and
3300 replace only the relevant branches. This means the first branch has
3301 to remain as it is used by the alternation logic, and its next pointer,
3302 and needs to be repointed at the item on the branch chain following
3303 the last branch we have optimized away.
3305 This could be either a BRANCH, in which case the subsequence is internal,
3306 or it could be the item following the branch sequence in which case the
3307 subsequence is at the end (which does not necessarily mean the first node
3308 is the start of the alternation).
3310 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3313 ----------------+-----------
3317 EXACTFU_SS | EXACTFU
3318 EXACTFU_TRICKYFOLD | EXACTFU
3323 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3324 ( EXACT == (X) ) ? EXACT : \
3325 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3328 /* dont use tail as the end marker for this traverse */
3329 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3330 regnode * const noper = NEXTOPER( cur );
3331 U8 noper_type = OP( noper );
3332 U8 noper_trietype = TRIE_TYPE( noper_type );
3333 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3334 regnode * const noper_next = regnext( noper );
3335 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3336 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3339 DEBUG_TRIE_COMPILE_r({
3340 regprop(RExC_rx, mysv, cur);
3341 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3342 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3344 regprop(RExC_rx, mysv, noper);
3345 PerlIO_printf( Perl_debug_log, " -> %s",
3346 SvPV_nolen_const(mysv));
3349 regprop(RExC_rx, mysv, noper_next );
3350 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3351 SvPV_nolen_const(mysv));
3353 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3354 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3355 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3359 /* Is noper a trieable nodetype that can be merged with the
3360 * current trie (if there is one)? */
3364 ( noper_trietype == NOTHING)
3365 || ( trietype == NOTHING )
3366 || ( trietype == noper_trietype )
3369 && noper_next == tail
3373 /* Handle mergable triable node
3374 * Either we are the first node in a new trieable sequence,
3375 * in which case we do some bookkeeping, otherwise we update
3376 * the end pointer. */
3379 if ( noper_trietype == NOTHING ) {
3380 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3381 regnode * const noper_next = regnext( noper );
3382 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3383 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3386 if ( noper_next_trietype ) {
3387 trietype = noper_next_trietype;
3388 } else if (noper_next_type) {
3389 /* a NOTHING regop is 1 regop wide. We need at least two
3390 * for a trie so we can't merge this in */
3394 trietype = noper_trietype;
3397 if ( trietype == NOTHING )
3398 trietype = noper_trietype;
3403 } /* end handle mergable triable node */
3405 /* handle unmergable node -
3406 * noper may either be a triable node which can not be tried
3407 * together with the current trie, or a non triable node */
3409 /* If last is set and trietype is not NOTHING then we have found
3410 * at least two triable branch sequences in a row of a similar
3411 * trietype so we can turn them into a trie. If/when we
3412 * allow NOTHING to start a trie sequence this condition will be
3413 * required, and it isn't expensive so we leave it in for now. */
3414 if ( trietype && trietype != NOTHING )
3415 make_trie( pRExC_state,
3416 startbranch, first, cur, tail, count,
3417 trietype, depth+1 );
3418 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3422 && noper_next == tail
3425 /* noper is triable, so we can start a new trie sequence */
3428 trietype = noper_trietype;
3430 /* if we already saw a first but the current node is not triable then we have
3431 * to reset the first information. */
3436 } /* end handle unmergable node */
3437 } /* loop over branches */
3438 DEBUG_TRIE_COMPILE_r({
3439 regprop(RExC_rx, mysv, cur);
3440 PerlIO_printf( Perl_debug_log,
3441 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3442 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3445 if ( last && trietype ) {
3446 if ( trietype != NOTHING ) {
3447 /* the last branch of the sequence was part of a trie,
3448 * so we have to construct it here outside of the loop
3450 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3451 #ifdef TRIE_STUDY_OPT
3452 if ( ((made == MADE_EXACT_TRIE &&
3453 startbranch == first)
3454 || ( first_non_open == first )) &&
3456 flags |= SCF_TRIE_RESTUDY;
3457 if ( startbranch == first
3460 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3465 /* at this point we know whatever we have is a NOTHING sequence/branch
3466 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3468 if ( startbranch == first ) {
3470 /* the entire thing is a NOTHING sequence, something like this:
3471 * (?:|) So we can turn it into a plain NOTHING op. */
3472 DEBUG_TRIE_COMPILE_r({
3473 regprop(RExC_rx, mysv, cur);
3474 PerlIO_printf( Perl_debug_log,
3475 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3476 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3479 OP(startbranch)= NOTHING;
3480 NEXT_OFF(startbranch)= tail - startbranch;
3481 for ( opt= startbranch + 1; opt < tail ; opt++ )
3485 } /* end if ( last) */
3486 } /* TRIE_MAXBUF is non zero */
3491 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3492 scan = NEXTOPER(NEXTOPER(scan));
3493 } else /* single branch is optimized. */
3494 scan = NEXTOPER(scan);
3496 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3497 scan_frame *newframe = NULL;
3502 if (OP(scan) != SUSPEND) {
3503 /* set the pointer */
3504 if (OP(scan) == GOSUB) {
3506 RExC_recurse[ARG2L(scan)] = scan;
3507 start = RExC_open_parens[paren-1];
3508 end = RExC_close_parens[paren-1];
3511 start = RExC_rxi->program + 1;
3515 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3516 SAVEFREEPV(recursed);
3518 if (!PAREN_TEST(recursed,paren+1)) {
3519 PAREN_SET(recursed,paren+1);
3520 Newx(newframe,1,scan_frame);
3522 if (flags & SCF_DO_SUBSTR) {
3523 SCAN_COMMIT(pRExC_state,data,minlenp);
3524 data->longest = &(data->longest_float);
3526 is_inf = is_inf_internal = 1;
3527 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3528 cl_anything(pRExC_state, data->start_class);
3529 flags &= ~SCF_DO_STCLASS;
3532 Newx(newframe,1,scan_frame);
3535 end = regnext(scan);
3540 SAVEFREEPV(newframe);
3541 newframe->next = regnext(scan);
3542 newframe->last = last;
3543 newframe->stop = stopparen;
3544 newframe->prev = frame;
3554 else if (OP(scan) == EXACT) {
3555 I32 l = STR_LEN(scan);
3558 const U8 * const s = (U8*)STRING(scan);
3559 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3560 l = utf8_length(s, s + l);
3562 uc = *((U8*)STRING(scan));
3565 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3566 /* The code below prefers earlier match for fixed
3567 offset, later match for variable offset. */
3568 if (data->last_end == -1) { /* Update the start info. */
3569 data->last_start_min = data->pos_min;
3570 data->last_start_max = is_inf
3571 ? I32_MAX : data->pos_min + data->pos_delta;
3573 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3575 SvUTF8_on(data->last_found);
3577 SV * const sv = data->last_found;
3578 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3579 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3580 if (mg && mg->mg_len >= 0)
3581 mg->mg_len += utf8_length((U8*)STRING(scan),
3582 (U8*)STRING(scan)+STR_LEN(scan));
3584 data->last_end = data->pos_min + l;
3585 data->pos_min += l; /* As in the first entry. */
3586 data->flags &= ~SF_BEFORE_EOL;
3588 if (flags & SCF_DO_STCLASS_AND) {
3589 /* Check whether it is compatible with what we know already! */
3593 /* If compatible, we or it in below. It is compatible if is
3594 * in the bitmp and either 1) its bit or its fold is set, or 2)
3595 * it's for a locale. Even if there isn't unicode semantics
3596 * here, at runtime there may be because of matching against a
3597 * utf8 string, so accept a possible false positive for
3598 * latin1-range folds */
3600 (!(data->start_class->flags & ANYOF_LOCALE)
3601 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3602 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3603 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3608 ANYOF_CLASS_ZERO(data->start_class);
3609 ANYOF_BITMAP_ZERO(data->start_class);
3611 ANYOF_BITMAP_SET(data->start_class, uc);
3612 else if (uc >= 0x100) {
3615 /* Some Unicode code points fold to the Latin1 range; as
3616 * XXX temporary code, instead of figuring out if this is
3617 * one, just assume it is and set all the start class bits
3618 * that could be some such above 255 code point's fold
3619 * which will generate fals positives. As the code
3620 * elsewhere that does compute the fold settles down, it
3621 * can be extracted out and re-used here */
3622 for (i = 0; i < 256; i++){
3623 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3624 ANYOF_BITMAP_SET(data->start_class, i);
3628 CLEAR_SSC_EOS(data->start_class);
3630 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3632 else if (flags & SCF_DO_STCLASS_OR) {
3633 /* false positive possible if the class is case-folded */
3635 ANYOF_BITMAP_SET(data->start_class, uc);
3637 data->start_class->flags |= ANYOF_UNICODE_ALL;
3638 CLEAR_SSC_EOS(data->start_class);
3639 cl_and(data->start_class, and_withp);
3641 flags &= ~SCF_DO_STCLASS;
3643 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3644 I32 l = STR_LEN(scan);
3645 UV uc = *((U8*)STRING(scan));
3647 /* Search for fixed substrings supports EXACT only. */
3648 if (flags & SCF_DO_SUBSTR) {
3650 SCAN_COMMIT(pRExC_state, data, minlenp);
3653 const U8 * const s = (U8 *)STRING(scan);
3654 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3655 l = utf8_length(s, s + l);
3657 if (has_exactf_sharp_s) {
3658 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3660 min += l - min_subtract;
3662 delta += min_subtract;
3663 if (flags & SCF_DO_SUBSTR) {
3664 data->pos_min += l - min_subtract;
3665 if (data->pos_min < 0) {
3668 data->pos_delta += min_subtract;
3670 data->longest = &(data->longest_float);
3673 if (flags & SCF_DO_STCLASS_AND) {
3674 /* Check whether it is compatible with what we know already! */
3677 (!(data->start_class->flags & ANYOF_LOCALE)
3678 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3679 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3683 ANYOF_CLASS_ZERO(data->start_class);
3684 ANYOF_BITMAP_ZERO(data->start_class);
3686 ANYOF_BITMAP_SET(data->start_class, uc);
3687 CLEAR_SSC_EOS(data->start_class);
3688 if (OP(scan) == EXACTFL) {
3689 /* XXX This set is probably no longer necessary, and
3690 * probably wrong as LOCALE now is on in the initial
3692 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3696 /* Also set the other member of the fold pair. In case
3697 * that unicode semantics is called for at runtime, use
3698 * the full latin1 fold. (Can't do this for locale,
3699 * because not known until runtime) */
3700 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3702 /* All other (EXACTFL handled above) folds except under
3703 * /iaa that include s, S, and sharp_s also may include
3705 if (OP(scan) != EXACTFA) {
3706 if (uc == 's' || uc == 'S') {
3707 ANYOF_BITMAP_SET(data->start_class,
3708 LATIN_SMALL_LETTER_SHARP_S);
3710 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3711 ANYOF_BITMAP_SET(data->start_class, 's');
3712 ANYOF_BITMAP_SET(data->start_class, 'S');
3717 else if (uc >= 0x100) {
3719 for (i = 0; i < 256; i++){
3720 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3721 ANYOF_BITMAP_SET(data->start_class, i);
3726 else if (flags & SCF_DO_STCLASS_OR) {
3727 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3728 /* false positive possible if the class is case-folded.
3729 Assume that the locale settings are the same... */
3731 ANYOF_BITMAP_SET(data->start_class, uc);
3732 if (OP(scan) != EXACTFL) {
3734 /* And set the other member of the fold pair, but
3735 * can't do that in locale because not known until
3737 ANYOF_BITMAP_SET(data->start_class,
3738 PL_fold_latin1[uc]);
3740 /* All folds except under /iaa that include s, S,
3741 * and sharp_s also may include the others */
3742 if (OP(scan) != EXACTFA) {
3743 if (uc == 's' || uc == 'S') {
3744 ANYOF_BITMAP_SET(data->start_class,
3745 LATIN_SMALL_LETTER_SHARP_S);
3747 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3748 ANYOF_BITMAP_SET(data->start_class, 's');
3749 ANYOF_BITMAP_SET(data->start_class, 'S');
3754 CLEAR_SSC_EOS(data->start_class);
3756 cl_and(data->start_class, and_withp);
3758 flags &= ~SCF_DO_STCLASS;
3760 else if (REGNODE_VARIES(OP(scan))) {
3761 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3762 I32 f = flags, pos_before = 0;
3763 regnode * const oscan = scan;
3764 struct regnode_charclass_class this_class;
3765 struct regnode_charclass_class *oclass = NULL;
3766 I32 next_is_eval = 0;
3768 switch (PL_regkind[OP(scan)]) {
3769 case WHILEM: /* End of (?:...)* . */
3770 scan = NEXTOPER(scan);
3773 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3774 next = NEXTOPER(scan);
3775 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3777 maxcount = REG_INFTY;
3778 next = regnext(scan);
3779 scan = NEXTOPER(scan);
3783 if (flags & SCF_DO_SUBSTR)
3788 if (flags & SCF_DO_STCLASS) {
3790 maxcount = REG_INFTY;
3791 next = regnext(scan);
3792 scan = NEXTOPER(scan);
3795 is_inf = is_inf_internal = 1;
3796 scan = regnext(scan);
3797 if (flags & SCF_DO_SUBSTR) {
3798 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3799 data->longest = &(data->longest_float);
3801 goto optimize_curly_tail;
3803 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3804 && (scan->flags == stopparen))
3809 mincount = ARG1(scan);
3810 maxcount = ARG2(scan);
3812 next = regnext(scan);
3813 if (OP(scan) == CURLYX) {
3814 I32 lp = (data ? *(data->last_closep) : 0);
3815 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3817 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3818 next_is_eval = (OP(scan) == EVAL);
3820 if (flags & SCF_DO_SUBSTR) {
3821 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3822 pos_before = data->pos_min;
3826 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3828 data->flags |= SF_IS_INF;
3830 if (flags & SCF_DO_STCLASS) {
3831 cl_init(pRExC_state, &this_class);
3832 oclass = data->start_class;
3833 data->start_class = &this_class;
3834 f |= SCF_DO_STCLASS_AND;
3835 f &= ~SCF_DO_STCLASS_OR;
3837 /* Exclude from super-linear cache processing any {n,m}
3838 regops for which the combination of input pos and regex
3839 pos is not enough information to determine if a match
3842 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3843 regex pos at the \s*, the prospects for a match depend not
3844 only on the input position but also on how many (bar\s*)
3845 repeats into the {4,8} we are. */
3846 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3847 f &= ~SCF_WHILEM_VISITED_POS;
3849 /* This will finish on WHILEM, setting scan, or on NULL: */
3850 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3851 last, data, stopparen, recursed, NULL,
3853 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3855 if (flags & SCF_DO_STCLASS)
3856 data->start_class = oclass;
3857 if (mincount == 0 || minnext == 0) {
3858 if (flags & SCF_DO_STCLASS_OR) {
3859 cl_or(pRExC_state, data->start_class, &this_class);
3861 else if (flags & SCF_DO_STCLASS_AND) {
3862 /* Switch to OR mode: cache the old value of
3863 * data->start_class */
3865 StructCopy(data->start_class, and_withp,
3866 struct regnode_charclass_class);
3867 flags &= ~SCF_DO_STCLASS_AND;
3868 StructCopy(&this_class, data->start_class,
3869 struct regnode_charclass_class);
3870 flags |= SCF_DO_STCLASS_OR;
3871 SET_SSC_EOS(data->start_class);
3873 } else { /* Non-zero len */
3874 if (flags & SCF_DO_STCLASS_OR) {
3875 cl_or(pRExC_state, data->start_class, &this_class);
3876 cl_and(data->start_class, and_withp);
3878 else if (flags & SCF_DO_STCLASS_AND)
3879 cl_and(data->start_class, &this_class);
3880 flags &= ~SCF_DO_STCLASS;
3882 if (!scan) /* It was not CURLYX, but CURLY. */
3884 if ( /* ? quantifier ok, except for (?{ ... }) */
3885 (next_is_eval || !(mincount == 0 && maxcount == 1))
3886 && (minnext == 0) && (deltanext == 0)
3887 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3888 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3890 /* Fatal warnings may leak the regexp without this: */
3891 SAVEFREESV(RExC_rx_sv);
3892 ckWARNreg(RExC_parse,
3893 "Quantifier unexpected on zero-length expression");
3894 (void)ReREFCNT_inc(RExC_rx_sv);
3897 min += minnext * mincount;
3898 is_inf_internal |= deltanext == I32_MAX
3899 || (maxcount == REG_INFTY && minnext + deltanext > 0);
3900 is_inf |= is_inf_internal;
3904 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3906 /* Try powerful optimization CURLYX => CURLYN. */
3907 if ( OP(oscan) == CURLYX && data
3908 && data->flags & SF_IN_PAR
3909 && !(data->flags & SF_HAS_EVAL)
3910 && !deltanext && minnext == 1 ) {
3911 /* Try to optimize to CURLYN. */
3912 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3913 regnode * const nxt1 = nxt;
3920 if (!REGNODE_SIMPLE(OP(nxt))
3921 && !(PL_regkind[OP(nxt)] == EXACT
3922 && STR_LEN(nxt) == 1))
3928 if (OP(nxt) != CLOSE)
3930 if (RExC_open_parens) {
3931 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3932 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3934 /* Now we know that nxt2 is the only contents: */
3935 oscan->flags = (U8)ARG(nxt);
3937 OP(nxt1) = NOTHING; /* was OPEN. */
3940 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3941 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3942 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3943 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3944 OP(nxt + 1) = OPTIMIZED; /* was count. */
3945 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3950 /* Try optimization CURLYX => CURLYM. */
3951 if ( OP(oscan) == CURLYX && data
3952 && !(data->flags & SF_HAS_PAR)
3953 && !(data->flags & SF_HAS_EVAL)
3954 && !deltanext /* atom is fixed width */
3955 && minnext != 0 /* CURLYM can't handle zero width */
3956 && ! (RExC_seen & REG_SEEN_EXACTF_SHARP_S) /* Nor \xDF */
3958 /* XXXX How to optimize if data == 0? */
3959 /* Optimize to a simpler form. */
3960 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3964 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3965 && (OP(nxt2) != WHILEM))
3967 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3968 /* Need to optimize away parenths. */
3969 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
3970 /* Set the parenth number. */
3971 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3973 oscan->flags = (U8)ARG(nxt);
3974 if (RExC_open_parens) {
3975 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3976 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3978 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3979 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3982 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3983 OP(nxt + 1) = OPTIMIZED; /* was count. */
3984 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
3985 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
3988 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3989 regnode *nnxt = regnext(nxt1);
3991 if (reg_off_by_arg[OP(nxt1)])
3992 ARG_SET(nxt1, nxt2 - nxt1);
3993 else if (nxt2 - nxt1 < U16_MAX)
3994 NEXT_OFF(nxt1) = nxt2 - nxt1;
3996 OP(nxt) = NOTHING; /* Cannot beautify */
4001 /* Optimize again: */
4002 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4003 NULL, stopparen, recursed, NULL, 0,depth+1);
4008 else if ((OP(oscan) == CURLYX)
4009 && (flags & SCF_WHILEM_VISITED_POS)
4010 /* See the comment on a similar expression above.
4011 However, this time it's not a subexpression
4012 we care about, but the expression itself. */
4013 && (maxcount == REG_INFTY)
4014 && data && ++data->whilem_c < 16) {
4015 /* This stays as CURLYX, we can put the count/of pair. */
4016 /* Find WHILEM (as in regexec.c) */
4017 regnode *nxt = oscan + NEXT_OFF(oscan);
4019 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4021 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4022 | (RExC_whilem_seen << 4)); /* On WHILEM */
4024 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4026 if (flags & SCF_DO_SUBSTR) {
4027 SV *last_str = NULL;
4028 int counted = mincount != 0;
4030 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
4031 #if defined(SPARC64_GCC_WORKAROUND)
4034 const char *s = NULL;
4037 if (pos_before >= data->last_start_min)
4040 b = data->last_start_min;
4043 s = SvPV_const(data->last_found, l);
4044 old = b - data->last_start_min;
4047 I32 b = pos_before >= data->last_start_min
4048 ? pos_before : data->last_start_min;
4050 const char * const s = SvPV_const(data->last_found, l);
4051 I32 old = b - data->last_start_min;
4055 old = utf8_hop((U8*)s, old) - (U8*)s;
4057 /* Get the added string: */
4058 last_str = newSVpvn_utf8(s + old, l, UTF);
4059 if (deltanext == 0 && pos_before == b) {
4060 /* What was added is a constant string */
4062 SvGROW(last_str, (mincount * l) + 1);
4063 repeatcpy(SvPVX(last_str) + l,
4064 SvPVX_const(last_str), l, mincount - 1);
4065 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4066 /* Add additional parts. */
4067 SvCUR_set(data->last_found,
4068 SvCUR(data->last_found) - l);
4069 sv_catsv(data->last_found, last_str);
4071 SV * sv = data->last_found;
4073 SvUTF8(sv) && SvMAGICAL(sv) ?
4074 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4075 if (mg && mg->mg_len >= 0)
4076 mg->mg_len += CHR_SVLEN(last_str) - l;
4078 data->last_end += l * (mincount - 1);
4081 /* start offset must point into the last copy */
4082 data->last_start_min += minnext * (mincount - 1);
4083 data->last_start_max += is_inf ? I32_MAX
4084 : (maxcount - 1) * (minnext + data->pos_delta);
4087 /* It is counted once already... */
4088 data->pos_min += minnext * (mincount - counted);
4090 PerlIO_printf(Perl_debug_log, "counted=%d deltanext=%d I32_MAX=%d minnext=%d maxcount=%d mincount=%d\n",
4091 counted, deltanext, I32_MAX, minnext, maxcount, mincount);
4092 if (deltanext != I32_MAX)
4093 PerlIO_printf(Perl_debug_log, "LHS=%d RHS=%d\n", -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount, I32_MAX - data->pos_delta);
4095 if (deltanext == I32_MAX || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= I32_MAX - data->pos_delta)
4096 data->pos_delta = I32_MAX;
4098 data->pos_delta += - counted * deltanext +
4099 (minnext + deltanext) * maxcount - minnext * mincount;
4100 if (mincount != maxcount) {
4101 /* Cannot extend fixed substrings found inside
4103 SCAN_COMMIT(pRExC_state,data,minlenp);
4104 if (mincount && last_str) {
4105 SV * const sv = data->last_found;
4106 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4107 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4111 sv_setsv(sv, last_str);
4112 data->last_end = data->pos_min;
4113 data->last_start_min =
4114 data->pos_min - CHR_SVLEN(last_str);
4115 data->last_start_max = is_inf
4117 : data->pos_min + data->pos_delta
4118 - CHR_SVLEN(last_str);
4120 data->longest = &(data->longest_float);
4122 SvREFCNT_dec(last_str);
4124 if (data && (fl & SF_HAS_EVAL))
4125 data->flags |= SF_HAS_EVAL;
4126 optimize_curly_tail:
4127 if (OP(oscan) != CURLYX) {
4128 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4130 NEXT_OFF(oscan) += NEXT_OFF(next);
4133 default: /* REF, and CLUMP only? */
4134 if (flags & SCF_DO_SUBSTR) {
4135 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4136 data->longest = &(data->longest_float);
4138 is_inf = is_inf_internal = 1;
4139 if (flags & SCF_DO_STCLASS_OR)
4140 cl_anything(pRExC_state, data->start_class);
4141 flags &= ~SCF_DO_STCLASS;
4145 else if (OP(scan) == LNBREAK) {
4146 if (flags & SCF_DO_STCLASS) {
4148 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4149 if (flags & SCF_DO_STCLASS_AND) {
4150 for (value = 0; value < 256; value++)
4151 if (!is_VERTWS_cp(value))
4152 ANYOF_BITMAP_CLEAR(data->start_class, value);
4155 for (value = 0; value < 256; value++)
4156 if (is_VERTWS_cp(value))
4157 ANYOF_BITMAP_SET(data->start_class, value);
4159 if (flags & SCF_DO_STCLASS_OR)
4160 cl_and(data->start_class, and_withp);
4161 flags &= ~SCF_DO_STCLASS;
4164 delta++; /* Because of the 2 char string cr-lf */
4165 if (flags & SCF_DO_SUBSTR) {
4166 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4168 data->pos_delta += 1;
4169 data->longest = &(data->longest_float);
4172 else if (REGNODE_SIMPLE(OP(scan))) {
4175 if (flags & SCF_DO_SUBSTR) {
4176 SCAN_COMMIT(pRExC_state,data,minlenp);
4180 if (flags & SCF_DO_STCLASS) {
4182 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4184 /* Some of the logic below assumes that switching
4185 locale on will only add false positives. */
4186 switch (PL_regkind[OP(scan)]) {
4192 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan));
4195 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4196 cl_anything(pRExC_state, data->start_class);
4199 if (OP(scan) == SANY)
4201 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
4202 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
4203 || ANYOF_CLASS_TEST_ANY_SET(data->start_class));
4204 cl_anything(pRExC_state, data->start_class);
4206 if (flags & SCF_DO_STCLASS_AND || !value)
4207 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
4210 if (flags & SCF_DO_STCLASS_AND)
4211 cl_and(data->start_class,
4212 (struct regnode_charclass_class*)scan);
4214 cl_or(pRExC_state, data->start_class,
4215 (struct regnode_charclass_class*)scan);
4223 classnum = FLAGS(scan);
4224 if (flags & SCF_DO_STCLASS_AND) {
4225 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4226 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum) + 1);
4227 for (value = 0; value < loop_max; value++) {
4228 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4229 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4235 if (data->start_class->flags & ANYOF_LOCALE) {
4236 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum));
4240 /* Even if under locale, set the bits for non-locale
4241 * in case it isn't a true locale-node. This will
4242 * create false positives if it truly is locale */
4243 for (value = 0; value < loop_max; value++) {
4244 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4245 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4257 classnum = FLAGS(scan);
4258 if (flags & SCF_DO_STCLASS_AND) {
4259 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4260 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum));
4261 for (value = 0; value < loop_max; value++) {
4262 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4263 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4269 if (data->start_class->flags & ANYOF_LOCALE) {
4270 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum) + 1);
4274 /* Even if under locale, set the bits for non-locale in
4275 * case it isn't a true locale-node. This will create
4276 * false positives if it truly is locale */
4277 for (value = 0; value < loop_max; value++) {
4278 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4279 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4282 if (PL_regkind[OP(scan)] == NPOSIXD) {
4283 data->start_class->flags |= ANYOF_NON_UTF8_LATIN1_ALL;
4289 if (flags & SCF_DO_STCLASS_OR)
4290 cl_and(data->start_class, and_withp);
4291 flags &= ~SCF_DO_STCLASS;
4294 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
4295 data->flags |= (OP(scan) == MEOL
4298 SCAN_COMMIT(pRExC_state, data, minlenp);
4301 else if ( PL_regkind[OP(scan)] == BRANCHJ
4302 /* Lookbehind, or need to calculate parens/evals/stclass: */
4303 && (scan->flags || data || (flags & SCF_DO_STCLASS))
4304 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
4305 if ( OP(scan) == UNLESSM &&
4307 OP(NEXTOPER(NEXTOPER(scan))) == NOTHING &&
4308 OP(regnext(NEXTOPER(NEXTOPER(scan)))) == SUCCEED
4311 regnode *upto= regnext(scan);
4313 SV * const mysv_val=sv_newmortal();
4314 DEBUG_STUDYDATA("OPFAIL",data,depth);
4316 /*DEBUG_PARSE_MSG("opfail");*/
4317 regprop(RExC_rx, mysv_val, upto);
4318 PerlIO_printf(Perl_debug_log, "~ replace with OPFAIL pointed at %s (%"IVdf") offset %"IVdf"\n",
4319 SvPV_nolen_const(mysv_val),
4320 (IV)REG_NODE_NUM(upto),
4325 NEXT_OFF(scan) = upto - scan;
4326 for (opt= scan + 1; opt < upto ; opt++)
4327 OP(opt) = OPTIMIZED;
4331 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4332 || OP(scan) == UNLESSM )
4334 /* Negative Lookahead/lookbehind
4335 In this case we can't do fixed string optimisation.
4338 I32 deltanext, minnext, fake = 0;
4340 struct regnode_charclass_class intrnl;
4343 data_fake.flags = 0;
4345 data_fake.whilem_c = data->whilem_c;
4346 data_fake.last_closep = data->last_closep;
4349 data_fake.last_closep = &fake;
4350 data_fake.pos_delta = delta;
4351 if ( flags & SCF_DO_STCLASS && !scan->flags
4352 && OP(scan) == IFMATCH ) { /* Lookahead */
4353 cl_init(pRExC_state, &intrnl);
4354 data_fake.start_class = &intrnl;
4355 f |= SCF_DO_STCLASS_AND;
4357 if (flags & SCF_WHILEM_VISITED_POS)
4358 f |= SCF_WHILEM_VISITED_POS;
4359 next = regnext(scan);
4360 nscan = NEXTOPER(NEXTOPER(scan));
4361 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
4362 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
4365 FAIL("Variable length lookbehind not implemented");
4367 else if (minnext > (I32)U8_MAX) {
4368 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4370 scan->flags = (U8)minnext;
4373 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4375 if (data_fake.flags & SF_HAS_EVAL)
4376 data->flags |= SF_HAS_EVAL;
4377 data->whilem_c = data_fake.whilem_c;
4379 if (f & SCF_DO_STCLASS_AND) {
4380 if (flags & SCF_DO_STCLASS_OR) {
4381 /* OR before, AND after: ideally we would recurse with
4382 * data_fake to get the AND applied by study of the
4383 * remainder of the pattern, and then derecurse;
4384 * *** HACK *** for now just treat as "no information".
4385 * See [perl #56690].
4387 cl_init(pRExC_state, data->start_class);
4389 /* AND before and after: combine and continue */
4390 const int was = TEST_SSC_EOS(data->start_class);
4392 cl_and(data->start_class, &intrnl);
4394 SET_SSC_EOS(data->start_class);
4398 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4400 /* Positive Lookahead/lookbehind
4401 In this case we can do fixed string optimisation,
4402 but we must be careful about it. Note in the case of
4403 lookbehind the positions will be offset by the minimum
4404 length of the pattern, something we won't know about
4405 until after the recurse.
4407 I32 deltanext, fake = 0;
4409 struct regnode_charclass_class intrnl;
4411 /* We use SAVEFREEPV so that when the full compile
4412 is finished perl will clean up the allocated
4413 minlens when it's all done. This way we don't
4414 have to worry about freeing them when we know
4415 they wont be used, which would be a pain.
4418 Newx( minnextp, 1, I32 );
4419 SAVEFREEPV(minnextp);
4422 StructCopy(data, &data_fake, scan_data_t);
4423 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
4426 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
4427 data_fake.last_found=newSVsv(data->last_found);
4431 data_fake.last_closep = &fake;
4432 data_fake.flags = 0;
4433 data_fake.pos_delta = delta;
4435 data_fake.flags |= SF_IS_INF;
4436 if ( flags & SCF_DO_STCLASS && !scan->flags
4437 && OP(scan) == IFMATCH ) { /* Lookahead */
4438 cl_init(pRExC_state, &intrnl);
4439 data_fake.start_class = &intrnl;
4440 f |= SCF_DO_STCLASS_AND;
4442 if (flags & SCF_WHILEM_VISITED_POS)
4443 f |= SCF_WHILEM_VISITED_POS;
4444 next = regnext(scan);
4445 nscan = NEXTOPER(NEXTOPER(scan));
4447 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
4448 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
4451 FAIL("Variable length lookbehind not implemented");
4453 else if (*minnextp > (I32)U8_MAX) {
4454 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4456 scan->flags = (U8)*minnextp;
4461 if (f & SCF_DO_STCLASS_AND) {
4462 const int was = TEST_SSC_EOS(data.start_class);
4464 cl_and(data->start_class, &intrnl);
4466 SET_SSC_EOS(data->start_class);
4469 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4471 if (data_fake.flags & SF_HAS_EVAL)
4472 data->flags |= SF_HAS_EVAL;
4473 data->whilem_c = data_fake.whilem_c;
4474 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
4475 if (RExC_rx->minlen<*minnextp)
4476 RExC_rx->minlen=*minnextp;
4477 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
4478 SvREFCNT_dec_NN(data_fake.last_found);
4480 if ( data_fake.minlen_fixed != minlenp )
4482 data->offset_fixed= data_fake.offset_fixed;
4483 data->minlen_fixed= data_fake.minlen_fixed;
4484 data->lookbehind_fixed+= scan->flags;
4486 if ( data_fake.minlen_float != minlenp )
4488 data->minlen_float= data_fake.minlen_float;
4489 data->offset_float_min=data_fake.offset_float_min;
4490 data->offset_float_max=data_fake.offset_float_max;
4491 data->lookbehind_float+= scan->flags;
4498 else if (OP(scan) == OPEN) {
4499 if (stopparen != (I32)ARG(scan))
4502 else if (OP(scan) == CLOSE) {
4503 if (stopparen == (I32)ARG(scan)) {
4506 if ((I32)ARG(scan) == is_par) {
4507 next = regnext(scan);
4509 if ( next && (OP(next) != WHILEM) && next < last)
4510 is_par = 0; /* Disable optimization */
4513 *(data->last_closep) = ARG(scan);
4515 else if (OP(scan) == EVAL) {
4517 data->flags |= SF_HAS_EVAL;
4519 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
4520 if (flags & SCF_DO_SUBSTR) {
4521 SCAN_COMMIT(pRExC_state,data,minlenp);
4522 flags &= ~SCF_DO_SUBSTR;
4524 if (data && OP(scan)==ACCEPT) {
4525 data->flags |= SCF_SEEN_ACCEPT;
4530 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
4532 if (flags & SCF_DO_SUBSTR) {
4533 SCAN_COMMIT(pRExC_state,data,minlenp);
4534 data->longest = &(data->longest_float);
4536 is_inf = is_inf_internal = 1;
4537 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4538 cl_anything(pRExC_state, data->start_class);
4539 flags &= ~SCF_DO_STCLASS;
4541 else if (OP(scan) == GPOS) {
4542 if (!(RExC_rx->extflags & RXf_GPOS_FLOAT) &&
4543 !(delta || is_inf || (data && data->pos_delta)))
4545 if (!(RExC_rx->extflags & RXf_ANCH) && (flags & SCF_DO_SUBSTR))
4546 RExC_rx->extflags |= RXf_ANCH_GPOS;
4547 if (RExC_rx->gofs < (U32)min)
4548 RExC_rx->gofs = min;
4550 RExC_rx->extflags |= RXf_GPOS_FLOAT;
4554 #ifdef TRIE_STUDY_OPT
4555 #ifdef FULL_TRIE_STUDY
4556 else if (PL_regkind[OP(scan)] == TRIE) {
4557 /* NOTE - There is similar code to this block above for handling
4558 BRANCH nodes on the initial study. If you change stuff here
4560 regnode *trie_node= scan;
4561 regnode *tail= regnext(scan);
4562 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4563 I32 max1 = 0, min1 = I32_MAX;
4564 struct regnode_charclass_class accum;
4566 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
4567 SCAN_COMMIT(pRExC_state, data,minlenp); /* Cannot merge strings after this. */
4568 if (flags & SCF_DO_STCLASS)
4569 cl_init_zero(pRExC_state, &accum);
4575 const regnode *nextbranch= NULL;
4578 for ( word=1 ; word <= trie->wordcount ; word++)
4580 I32 deltanext=0, minnext=0, f = 0, fake;
4581 struct regnode_charclass_class this_class;
4583 data_fake.flags = 0;
4585 data_fake.whilem_c = data->whilem_c;
4586 data_fake.last_closep = data->last_closep;
4589 data_fake.last_closep = &fake;
4590 data_fake.pos_delta = delta;
4591 if (flags & SCF_DO_STCLASS) {
4592 cl_init(pRExC_state, &this_class);
4593 data_fake.start_class = &this_class;
4594 f = SCF_DO_STCLASS_AND;
4596 if (flags & SCF_WHILEM_VISITED_POS)
4597 f |= SCF_WHILEM_VISITED_POS;
4599 if (trie->jump[word]) {
4601 nextbranch = trie_node + trie->jump[0];
4602 scan= trie_node + trie->jump[word];
4603 /* We go from the jump point to the branch that follows
4604 it. Note this means we need the vestigal unused branches
4605 even though they arent otherwise used.
4607 minnext = study_chunk(pRExC_state, &scan, minlenp,
4608 &deltanext, (regnode *)nextbranch, &data_fake,
4609 stopparen, recursed, NULL, f,depth+1);
4611 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
4612 nextbranch= regnext((regnode*)nextbranch);
4614 if (min1 > (I32)(minnext + trie->minlen))
4615 min1 = minnext + trie->minlen;
4616 if (deltanext == I32_MAX) {
4617 is_inf = is_inf_internal = 1;
4619 } else if (max1 < (I32)(minnext + deltanext + trie->maxlen))
4620 max1 = minnext + deltanext + trie->maxlen;
4622 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4624 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4625 if ( stopmin > min + min1)
4626 stopmin = min + min1;
4627 flags &= ~SCF_DO_SUBSTR;
4629 data->flags |= SCF_SEEN_ACCEPT;
4632 if (data_fake.flags & SF_HAS_EVAL)
4633 data->flags |= SF_HAS_EVAL;
4634 data->whilem_c = data_fake.whilem_c;
4636 if (flags & SCF_DO_STCLASS)
4637 cl_or(pRExC_state, &accum, &this_class);
4640 if (flags & SCF_DO_SUBSTR) {
4641 data->pos_min += min1;
4642 data->pos_delta += max1 - min1;
4643 if (max1 != min1 || is_inf)
4644 data->longest = &(data->longest_float);
4647 delta += max1 - min1;
4648 if (flags & SCF_DO_STCLASS_OR) {
4649 cl_or(pRExC_state, data->start_class, &accum);
4651 cl_and(data->start_class, and_withp);
4652 flags &= ~SCF_DO_STCLASS;
4655 else if (flags & SCF_DO_STCLASS_AND) {
4657 cl_and(data->start_class, &accum);
4658 flags &= ~SCF_DO_STCLASS;
4661 /* Switch to OR mode: cache the old value of
4662 * data->start_class */
4664 StructCopy(data->start_class, and_withp,
4665 struct regnode_charclass_class);
4666 flags &= ~SCF_DO_STCLASS_AND;
4667 StructCopy(&accum, data->start_class,
4668 struct regnode_charclass_class);
4669 flags |= SCF_DO_STCLASS_OR;
4670 SET_SSC_EOS(data->start_class);
4677 else if (PL_regkind[OP(scan)] == TRIE) {
4678 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4681 min += trie->minlen;
4682 delta += (trie->maxlen - trie->minlen);
4683 flags &= ~SCF_DO_STCLASS; /* xxx */
4684 if (flags & SCF_DO_SUBSTR) {
4685 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4686 data->pos_min += trie->minlen;
4687 data->pos_delta += (trie->maxlen - trie->minlen);
4688 if (trie->maxlen != trie->minlen)
4689 data->longest = &(data->longest_float);
4691 if (trie->jump) /* no more substrings -- for now /grr*/
4692 flags &= ~SCF_DO_SUBSTR;
4694 #endif /* old or new */
4695 #endif /* TRIE_STUDY_OPT */
4697 /* Else: zero-length, ignore. */
4698 scan = regnext(scan);
4703 stopparen = frame->stop;
4704 frame = frame->prev;
4705 goto fake_study_recurse;
4710 DEBUG_STUDYDATA("pre-fin:",data,depth);
4713 *deltap = is_inf_internal ? I32_MAX : delta;
4714 if (flags & SCF_DO_SUBSTR && is_inf)
4715 data->pos_delta = I32_MAX - data->pos_min;
4716 if (is_par > (I32)U8_MAX)
4718 if (is_par && pars==1 && data) {
4719 data->flags |= SF_IN_PAR;
4720 data->flags &= ~SF_HAS_PAR;
4722 else if (pars && data) {
4723 data->flags |= SF_HAS_PAR;
4724 data->flags &= ~SF_IN_PAR;
4726 if (flags & SCF_DO_STCLASS_OR)
4727 cl_and(data->start_class, and_withp);
4728 if (flags & SCF_TRIE_RESTUDY)
4729 data->flags |= SCF_TRIE_RESTUDY;
4731 DEBUG_STUDYDATA("post-fin:",data,depth);
4733 return min < stopmin ? min : stopmin;
4737 S_add_data(RExC_state_t *pRExC_state, U32 n, const char *s)
4739 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
4741 PERL_ARGS_ASSERT_ADD_DATA;
4743 Renewc(RExC_rxi->data,
4744 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
4745 char, struct reg_data);
4747 Renew(RExC_rxi->data->what, count + n, U8);
4749 Newx(RExC_rxi->data->what, n, U8);
4750 RExC_rxi->data->count = count + n;
4751 Copy(s, RExC_rxi->data->what + count, n, U8);
4755 /*XXX: todo make this not included in a non debugging perl */
4756 #ifndef PERL_IN_XSUB_RE
4758 Perl_reginitcolors(pTHX)
4761 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
4763 char *t = savepv(s);
4767 t = strchr(t, '\t');
4773 PL_colors[i] = t = (char *)"";
4778 PL_colors[i++] = (char *)"";
4785 #ifdef TRIE_STUDY_OPT
4786 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
4789 (data.flags & SCF_TRIE_RESTUDY) \
4797 #define CHECK_RESTUDY_GOTO_butfirst
4801 * pregcomp - compile a regular expression into internal code
4803 * Decides which engine's compiler to call based on the hint currently in
4807 #ifndef PERL_IN_XSUB_RE
4809 /* return the currently in-scope regex engine (or the default if none) */
4811 regexp_engine const *
4812 Perl_current_re_engine(pTHX)
4816 if (IN_PERL_COMPILETIME) {
4817 HV * const table = GvHV(PL_hintgv);
4821 return &PL_core_reg_engine;
4822 ptr = hv_fetchs(table, "regcomp", FALSE);
4823 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
4824 return &PL_core_reg_engine;
4825 return INT2PTR(regexp_engine*,SvIV(*ptr));
4829 if (!PL_curcop->cop_hints_hash)
4830 return &PL_core_reg_engine;
4831 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
4832 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
4833 return &PL_core_reg_engine;
4834 return INT2PTR(regexp_engine*,SvIV(ptr));
4840 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
4843 regexp_engine const *eng = current_re_engine();
4844 GET_RE_DEBUG_FLAGS_DECL;
4846 PERL_ARGS_ASSERT_PREGCOMP;
4848 /* Dispatch a request to compile a regexp to correct regexp engine. */
4850 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
4853 return CALLREGCOMP_ENG(eng, pattern, flags);
4857 /* public(ish) entry point for the perl core's own regex compiling code.
4858 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
4859 * pattern rather than a list of OPs, and uses the internal engine rather
4860 * than the current one */
4863 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
4865 SV *pat = pattern; /* defeat constness! */
4866 PERL_ARGS_ASSERT_RE_COMPILE;
4867 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
4868 #ifdef PERL_IN_XSUB_RE
4871 &PL_core_reg_engine,
4873 NULL, NULL, rx_flags, 0);
4877 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
4878 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
4879 * point to the realloced string and length.
4881 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
4885 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
4886 char **pat_p, STRLEN *plen_p)
4888 U8 *const src = (U8*)*pat_p;
4891 STRLEN s = 0, d = 0;
4893 GET_RE_DEBUG_FLAGS_DECL;
4895 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
4896 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
4898 Newx(dst, *plen_p * 2 + 1, U8);
4900 while (s < *plen_p) {
4901 const UV uv = NATIVE_TO_ASCII(src[s]);
4902 if (UNI_IS_INVARIANT(uv))
4903 dst[d] = (U8)UTF_TO_NATIVE(uv);
4905 dst[d++] = (U8)UTF8_EIGHT_BIT_HI(uv);
4906 dst[d] = (U8)UTF8_EIGHT_BIT_LO(uv);
4908 if (n < pRExC_state->num_code_blocks) {
4909 if (!do_end && pRExC_state->code_blocks[n].start == s) {
4910 pRExC_state->code_blocks[n].start = d;
4911 assert(dst[d] == '(');
4914 else if (do_end && pRExC_state->code_blocks[n].end == s) {
4915 pRExC_state->code_blocks[n].end = d;
4916 assert(dst[d] == ')');
4926 *pat_p = (char*) dst;
4928 RExC_orig_utf8 = RExC_utf8 = 1;
4932 /* see if there are any run-time code blocks in the pattern.
4933 * False positives are allowed */
4936 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
4937 char *pat, STRLEN plen)
4942 for (s = 0; s < plen; s++) {
4943 if (n < pRExC_state->num_code_blocks
4944 && s == pRExC_state->code_blocks[n].start)
4946 s = pRExC_state->code_blocks[n].end;
4950 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
4952 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
4954 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
4961 /* Handle run-time code blocks. We will already have compiled any direct
4962 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
4963 * copy of it, but with any literal code blocks blanked out and
4964 * appropriate chars escaped; then feed it into
4966 * eval "qr'modified_pattern'"
4970 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
4974 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
4976 * After eval_sv()-ing that, grab any new code blocks from the returned qr
4977 * and merge them with any code blocks of the original regexp.
4979 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
4980 * instead, just save the qr and return FALSE; this tells our caller that
4981 * the original pattern needs upgrading to utf8.
4985 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
4986 char *pat, STRLEN plen)
4990 GET_RE_DEBUG_FLAGS_DECL;
4992 if (pRExC_state->runtime_code_qr) {
4993 /* this is the second time we've been called; this should
4994 * only happen if the main pattern got upgraded to utf8
4995 * during compilation; re-use the qr we compiled first time
4996 * round (which should be utf8 too)
4998 qr = pRExC_state->runtime_code_qr;
4999 pRExC_state->runtime_code_qr = NULL;
5000 assert(RExC_utf8 && SvUTF8(qr));
5006 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
5010 /* determine how many extra chars we need for ' and \ escaping */
5011 for (s = 0; s < plen; s++) {
5012 if (pat[s] == '\'' || pat[s] == '\\')
5016 Newx(newpat, newlen, char);
5018 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
5020 for (s = 0; s < plen; s++) {
5021 if (n < pRExC_state->num_code_blocks
5022 && s == pRExC_state->code_blocks[n].start)
5024 /* blank out literal code block */
5025 assert(pat[s] == '(');
5026 while (s <= pRExC_state->code_blocks[n].end) {
5034 if (pat[s] == '\'' || pat[s] == '\\')
5039 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
5043 PerlIO_printf(Perl_debug_log,
5044 "%sre-parsing pattern for runtime code:%s %s\n",
5045 PL_colors[4],PL_colors[5],newpat);
5048 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5054 PUSHSTACKi(PERLSI_REQUIRE);
5055 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
5056 * parsing qr''; normally only q'' does this. It also alters
5058 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
5059 SvREFCNT_dec_NN(sv);
5064 SV * const errsv = ERRSV;
5065 if (SvTRUE_NN(errsv))
5067 Safefree(pRExC_state->code_blocks);
5068 /* use croak_sv ? */
5069 Perl_croak_nocontext("%s", SvPV_nolen_const(errsv));
5072 assert(SvROK(qr_ref));
5074 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5075 /* the leaving below frees the tmp qr_ref.
5076 * Give qr a life of its own */
5084 if (!RExC_utf8 && SvUTF8(qr)) {
5085 /* first time through; the pattern got upgraded; save the
5086 * qr for the next time through */
5087 assert(!pRExC_state->runtime_code_qr);
5088 pRExC_state->runtime_code_qr = qr;
5093 /* extract any code blocks within the returned qr// */
5096 /* merge the main (r1) and run-time (r2) code blocks into one */
5098 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
5099 struct reg_code_block *new_block, *dst;
5100 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
5103 if (!r2->num_code_blocks) /* we guessed wrong */
5105 SvREFCNT_dec_NN(qr);
5110 r1->num_code_blocks + r2->num_code_blocks,
5111 struct reg_code_block);
5114 while ( i1 < r1->num_code_blocks
5115 || i2 < r2->num_code_blocks)
5117 struct reg_code_block *src;
5120 if (i1 == r1->num_code_blocks) {
5121 src = &r2->code_blocks[i2++];
5124 else if (i2 == r2->num_code_blocks)
5125 src = &r1->code_blocks[i1++];
5126 else if ( r1->code_blocks[i1].start
5127 < r2->code_blocks[i2].start)
5129 src = &r1->code_blocks[i1++];
5130 assert(src->end < r2->code_blocks[i2].start);
5133 assert( r1->code_blocks[i1].start
5134 > r2->code_blocks[i2].start);
5135 src = &r2->code_blocks[i2++];
5137 assert(src->end < r1->code_blocks[i1].start);
5140 assert(pat[src->start] == '(');
5141 assert(pat[src->end] == ')');
5142 dst->start = src->start;
5143 dst->end = src->end;
5144 dst->block = src->block;
5145 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
5149 r1->num_code_blocks += r2->num_code_blocks;
5150 Safefree(r1->code_blocks);
5151 r1->code_blocks = new_block;
5154 SvREFCNT_dec_NN(qr);
5160 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)
5162 /* This is the common code for setting up the floating and fixed length
5163 * string data extracted from Perl_re_op_compile() below. Returns a boolean
5164 * as to whether succeeded or not */
5168 if (! (longest_length
5169 || (eol /* Can't have SEOL and MULTI */
5170 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
5172 /* See comments for join_exact for why REG_SEEN_EXACTF_SHARP_S */
5173 || (RExC_seen & REG_SEEN_EXACTF_SHARP_S))
5178 /* copy the information about the longest from the reg_scan_data
5179 over to the program. */
5180 if (SvUTF8(sv_longest)) {
5181 *rx_utf8 = sv_longest;
5184 *rx_substr = sv_longest;
5187 /* end_shift is how many chars that must be matched that
5188 follow this item. We calculate it ahead of time as once the
5189 lookbehind offset is added in we lose the ability to correctly
5191 ml = minlen ? *(minlen) : (I32)longest_length;
5192 *rx_end_shift = ml - offset
5193 - longest_length + (SvTAIL(sv_longest) != 0)
5196 t = (eol/* Can't have SEOL and MULTI */
5197 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
5198 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
5204 * Perl_re_op_compile - the perl internal RE engine's function to compile a
5205 * regular expression into internal code.
5206 * The pattern may be passed either as:
5207 * a list of SVs (patternp plus pat_count)
5208 * a list of OPs (expr)
5209 * If both are passed, the SV list is used, but the OP list indicates
5210 * which SVs are actually pre-compiled code blocks
5212 * The SVs in the list have magic and qr overloading applied to them (and
5213 * the list may be modified in-place with replacement SVs in the latter
5216 * If the pattern hasn't changed from old_re, then old_re will be
5219 * eng is the current engine. If that engine has an op_comp method, then
5220 * handle directly (i.e. we assume that op_comp was us); otherwise, just
5221 * do the initial concatenation of arguments and pass on to the external
5224 * If is_bare_re is not null, set it to a boolean indicating whether the
5225 * arg list reduced (after overloading) to a single bare regex which has
5226 * been returned (i.e. /$qr/).
5228 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
5230 * pm_flags contains the PMf_* flags, typically based on those from the
5231 * pm_flags field of the related PMOP. Currently we're only interested in
5232 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
5234 * We can't allocate space until we know how big the compiled form will be,
5235 * but we can't compile it (and thus know how big it is) until we've got a
5236 * place to put the code. So we cheat: we compile it twice, once with code
5237 * generation turned off and size counting turned on, and once "for real".
5238 * This also means that we don't allocate space until we are sure that the
5239 * thing really will compile successfully, and we never have to move the
5240 * code and thus invalidate pointers into it. (Note that it has to be in
5241 * one piece because free() must be able to free it all.) [NB: not true in perl]
5243 * Beware that the optimization-preparation code in here knows about some
5244 * of the structure of the compiled regexp. [I'll say.]
5248 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
5249 OP *expr, const regexp_engine* eng, REGEXP *old_re,
5250 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
5255 regexp_internal *ri;
5263 SV *code_blocksv = NULL;
5264 SV** new_patternp = patternp;
5266 /* these are all flags - maybe they should be turned
5267 * into a single int with different bit masks */
5268 I32 sawlookahead = 0;
5271 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
5273 bool runtime_code = 0;
5275 RExC_state_t RExC_state;
5276 RExC_state_t * const pRExC_state = &RExC_state;
5277 #ifdef TRIE_STUDY_OPT
5279 RExC_state_t copyRExC_state;
5281 GET_RE_DEBUG_FLAGS_DECL;
5283 PERL_ARGS_ASSERT_RE_OP_COMPILE;
5285 DEBUG_r(if (!PL_colorset) reginitcolors());
5287 #ifndef PERL_IN_XSUB_RE
5288 /* Initialize these here instead of as-needed, as is quick and avoids
5289 * having to test them each time otherwise */
5290 if (! PL_AboveLatin1) {
5291 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
5292 PL_ASCII = _new_invlist_C_array(ASCII_invlist);
5293 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
5295 PL_L1Posix_ptrs[_CC_ALPHANUMERIC]
5296 = _new_invlist_C_array(L1PosixAlnum_invlist);
5297 PL_Posix_ptrs[_CC_ALPHANUMERIC]
5298 = _new_invlist_C_array(PosixAlnum_invlist);
5300 PL_L1Posix_ptrs[_CC_ALPHA]
5301 = _new_invlist_C_array(L1PosixAlpha_invlist);
5302 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(PosixAlpha_invlist);
5304 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(PosixBlank_invlist);
5305 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(XPosixBlank_invlist);
5307 /* Cased is the same as Alpha in the ASCII range */
5308 PL_L1Posix_ptrs[_CC_CASED] = _new_invlist_C_array(L1Cased_invlist);
5309 PL_Posix_ptrs[_CC_CASED] = _new_invlist_C_array(PosixAlpha_invlist);
5311 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(PosixCntrl_invlist);
5312 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(XPosixCntrl_invlist);
5314 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5315 PL_L1Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5317 PL_L1Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(L1PosixGraph_invlist);
5318 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(PosixGraph_invlist);
5320 PL_L1Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(L1PosixLower_invlist);
5321 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(PosixLower_invlist);
5323 PL_L1Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(L1PosixPrint_invlist);
5324 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(PosixPrint_invlist);
5326 PL_L1Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(L1PosixPunct_invlist);
5327 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(PosixPunct_invlist);
5329 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(PerlSpace_invlist);
5330 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(XPerlSpace_invlist);
5331 PL_Posix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(PosixSpace_invlist);
5332 PL_XPosix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(XPosixSpace_invlist);
5334 PL_L1Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(L1PosixUpper_invlist);
5335 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(PosixUpper_invlist);
5337 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(VertSpace_invlist);
5339 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(PosixWord_invlist);
5340 PL_L1Posix_ptrs[_CC_WORDCHAR]
5341 = _new_invlist_C_array(L1PosixWord_invlist);
5343 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(PosixXDigit_invlist);
5344 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(XPosixXDigit_invlist);
5346 PL_HasMultiCharFold = _new_invlist_C_array(_Perl_Multi_Char_Folds_invlist);
5350 pRExC_state->code_blocks = NULL;
5351 pRExC_state->num_code_blocks = 0;
5354 *is_bare_re = FALSE;
5356 if (expr && (expr->op_type == OP_LIST ||
5357 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
5358 /* allocate code_blocks if needed */
5362 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling)
5363 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
5364 ncode++; /* count of DO blocks */
5366 pRExC_state->num_code_blocks = ncode;
5367 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
5372 /* compile-time pattern with just OP_CONSTs and DO blocks */
5377 /* find how many CONSTs there are */
5380 if (expr->op_type == OP_CONST)
5383 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5384 if (o->op_type == OP_CONST)
5388 /* fake up an SV array */
5390 assert(!new_patternp);
5391 Newx(new_patternp, n, SV*);
5392 SAVEFREEPV(new_patternp);
5396 if (expr->op_type == OP_CONST)
5397 new_patternp[n] = cSVOPx_sv(expr);
5399 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5400 if (o->op_type == OP_CONST)
5401 new_patternp[n++] = cSVOPo_sv;
5407 /* concat args, handling magic, overloading etc */
5412 STRLEN orig_patlen = 0;
5414 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5415 "Assembling pattern from %d elements%s\n", pat_count,
5416 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5418 /* apply magic and RE overloading to each arg */
5419 for (svp = new_patternp; svp < new_patternp + pat_count; svp++) {
5422 if (SvROK(rx) && SvAMAGIC(rx)) {
5423 SV *sv = AMG_CALLunary(rx, regexp_amg);
5427 if (SvTYPE(sv) != SVt_REGEXP)
5428 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5434 if (pRExC_state->num_code_blocks) {
5435 if (expr->op_type == OP_CONST)
5438 o = cLISTOPx(expr)->op_first;
5439 assert( o->op_type == OP_PUSHMARK
5440 || (o->op_type == OP_NULL && o->op_targ == OP_PUSHMARK)
5441 || o->op_type == OP_PADRANGE);
5446 if (pat_count > 1) {
5448 pat = newSVpvn("", 0);
5451 /* determine if the pattern is going to be utf8 (needed
5452 * in advance to align code block indices correctly).
5453 * XXX This could fail to be detected for an arg with
5454 * overloading but not concat overloading; but the main effect
5455 * in this obscure case is to need a 'use re eval' for a
5456 * literal code block */
5457 for (svp = new_patternp; svp < new_patternp + pat_count; svp++) {
5463 /* process args, concat them if there are multiple ones,
5464 * and find any code block indexes */
5467 for (svp = new_patternp; svp < new_patternp + pat_count; svp++) {
5468 SV *sv, *msv = *svp;
5471 /* we make the assumption here that each op in the list of
5472 * op_siblings maps to one SV pushed onto the stack,
5473 * except for code blocks, with have both an OP_NULL and
5475 * This allows us to match up the list of SVs against the
5476 * list of OPs to find the next code block.
5478 * Note that PUSHMARK PADSV PADSV ..
5480 * PADRANGE NULL NULL ..
5481 * so the alignment still works. */
5483 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL)) {
5484 assert(n < pRExC_state->num_code_blocks);
5485 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5486 pRExC_state->code_blocks[n].block = o;
5487 pRExC_state->code_blocks[n].src_regex = NULL;
5490 o = o->op_sibling; /* skip CONST */
5496 /* try concatenation overload ... */
5497 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5498 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5501 /* overloading involved: all bets are off over literal
5502 * code. Pretend we haven't seen it */
5503 pRExC_state->num_code_blocks -= n;
5507 /* ... or failing that, try "" overload */
5508 while (SvAMAGIC(msv)
5509 && (sv = AMG_CALLunary(msv, string_amg))
5513 && SvRV(msv) == SvRV(sv))
5518 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5521 orig_patlen = SvCUR(pat);
5522 sv_catsv_nomg(pat, msv);
5528 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5531 /* extract any code blocks within any embedded qr//'s */
5532 if (rx && SvTYPE(rx) == SVt_REGEXP
5533 && RX_ENGINE((REGEXP*)rx)->op_comp)
5536 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5537 if (ri->num_code_blocks) {
5539 /* the presence of an embedded qr// with code means
5540 * we should always recompile: the text of the
5541 * qr// may not have changed, but it may be a
5542 * different closure than last time */
5544 Renew(pRExC_state->code_blocks,
5545 pRExC_state->num_code_blocks + ri->num_code_blocks,
5546 struct reg_code_block);
5547 pRExC_state->num_code_blocks += ri->num_code_blocks;
5548 for (i=0; i < ri->num_code_blocks; i++) {
5549 struct reg_code_block *src, *dst;
5550 STRLEN offset = orig_patlen
5551 + ReANY((REGEXP *)rx)->pre_prefix;
5552 assert(n < pRExC_state->num_code_blocks);
5553 src = &ri->code_blocks[i];
5554 dst = &pRExC_state->code_blocks[n];
5555 dst->start = src->start + offset;
5556 dst->end = src->end + offset;
5557 dst->block = src->block;
5558 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5570 /* handle bare (possibly after overloading) regex: foo =~ $re */
5575 if (SvTYPE(re) == SVt_REGEXP) {
5579 Safefree(pRExC_state->code_blocks);
5580 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5581 "Precompiled pattern%s\n",
5582 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5589 exp = SvPV_nomg(pat, plen);
5591 if (!eng->op_comp) {
5592 if ((SvUTF8(pat) && IN_BYTES)
5593 || SvGMAGICAL(pat) || SvAMAGIC(pat))
5595 /* make a temporary copy; either to convert to bytes,
5596 * or to avoid repeating get-magic / overloaded stringify */
5597 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
5598 (IN_BYTES ? 0 : SvUTF8(pat)));
5600 Safefree(pRExC_state->code_blocks);
5601 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
5604 /* ignore the utf8ness if the pattern is 0 length */
5605 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
5606 RExC_uni_semantics = 0;
5607 RExC_contains_locale = 0;
5608 pRExC_state->runtime_code_qr = NULL;
5611 SV *dsv= sv_newmortal();
5612 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
5613 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
5614 PL_colors[4],PL_colors[5],s);
5619 /* It's possible to write a regexp in ascii that represents Unicode
5620 codepoints outside of the byte range, such as via \x{100}. If we
5621 detect such a sequence we have to convert the entire pattern to utf8
5622 and then recompile, as our sizing calculation will have been based
5623 on 1 byte == 1 character, but we will need to use utf8 to encode
5624 at least some part of the pattern, and therefore must convert the whole
5628 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen);
5631 if ((pm_flags & PMf_USE_RE_EVAL)
5632 /* this second condition covers the non-regex literal case,
5633 * i.e. $foo =~ '(?{})'. */
5634 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
5636 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
5638 /* return old regex if pattern hasn't changed */
5639 /* XXX: note in the below we have to check the flags as well as the pattern.
5641 * Things get a touch tricky as we have to compare the utf8 flag independently
5642 * from the compile flags.
5647 && !!RX_UTF8(old_re) == !!RExC_utf8
5648 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
5649 && RX_PRECOMP(old_re)
5650 && RX_PRELEN(old_re) == plen
5651 && memEQ(RX_PRECOMP(old_re), exp, plen)
5652 && !runtime_code /* with runtime code, always recompile */ )
5654 Safefree(pRExC_state->code_blocks);
5658 rx_flags = orig_rx_flags;
5660 if (initial_charset == REGEX_LOCALE_CHARSET) {
5661 RExC_contains_locale = 1;
5663 else if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
5665 /* Set to use unicode semantics if the pattern is in utf8 and has the
5666 * 'depends' charset specified, as it means unicode when utf8 */
5667 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5671 RExC_flags = rx_flags;
5672 RExC_pm_flags = pm_flags;
5675 if (TAINTING_get && TAINT_get)
5676 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
5678 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
5679 /* whoops, we have a non-utf8 pattern, whilst run-time code
5680 * got compiled as utf8. Try again with a utf8 pattern */
5681 goto redo_first_pass;
5684 assert(!pRExC_state->runtime_code_qr);
5689 RExC_in_lookbehind = 0;
5690 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
5692 RExC_override_recoding = 0;
5693 RExC_in_multi_char_class = 0;
5695 /* First pass: determine size, legality. */
5698 RExC_end = exp + plen;
5703 RExC_emit = &PL_regdummy;
5704 RExC_whilem_seen = 0;
5705 RExC_open_parens = NULL;
5706 RExC_close_parens = NULL;
5708 RExC_paren_names = NULL;
5710 RExC_paren_name_list = NULL;
5712 RExC_recurse = NULL;
5713 RExC_recurse_count = 0;
5714 pRExC_state->code_index = 0;
5716 #if 0 /* REGC() is (currently) a NOP at the first pass.
5717 * Clever compilers notice this and complain. --jhi */
5718 REGC((U8)REG_MAGIC, (char*)RExC_emit);
5721 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
5723 RExC_lastparse=NULL;
5725 /* reg may croak on us, not giving us a chance to free
5726 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
5727 need it to survive as long as the regexp (qr/(?{})/).
5728 We must check that code_blocksv is not already set, because we may
5729 have jumped back to restart the sizing pass. */
5730 if (pRExC_state->code_blocks && !code_blocksv) {
5731 code_blocksv = newSV_type(SVt_PV);
5732 SAVEFREESV(code_blocksv);
5733 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
5734 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
5736 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5737 if (flags & RESTART_UTF8) {
5738 goto redo_first_pass;
5740 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#X", flags);
5743 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
5746 PerlIO_printf(Perl_debug_log,
5747 "Required size %"IVdf" nodes\n"
5748 "Starting second pass (creation)\n",
5751 RExC_lastparse=NULL;
5754 /* The first pass could have found things that force Unicode semantics */
5755 if ((RExC_utf8 || RExC_uni_semantics)
5756 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
5758 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5761 /* Small enough for pointer-storage convention?
5762 If extralen==0, this means that we will not need long jumps. */
5763 if (RExC_size >= 0x10000L && RExC_extralen)
5764 RExC_size += RExC_extralen;
5767 if (RExC_whilem_seen > 15)
5768 RExC_whilem_seen = 15;
5770 /* Allocate space and zero-initialize. Note, the two step process
5771 of zeroing when in debug mode, thus anything assigned has to
5772 happen after that */
5773 rx = (REGEXP*) newSV_type(SVt_REGEXP);
5775 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
5776 char, regexp_internal);
5777 if ( r == NULL || ri == NULL )
5778 FAIL("Regexp out of space");
5780 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
5781 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode), char);
5783 /* bulk initialize base fields with 0. */
5784 Zero(ri, sizeof(regexp_internal), char);
5787 /* non-zero initialization begins here */
5790 r->extflags = rx_flags;
5791 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
5793 if (pm_flags & PMf_IS_QR) {
5794 ri->code_blocks = pRExC_state->code_blocks;
5795 ri->num_code_blocks = pRExC_state->num_code_blocks;
5800 for (n = 0; n < pRExC_state->num_code_blocks; n++)
5801 if (pRExC_state->code_blocks[n].src_regex)
5802 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
5803 SAVEFREEPV(pRExC_state->code_blocks);
5807 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
5808 bool has_charset = (get_regex_charset(r->extflags) != REGEX_DEPENDS_CHARSET);
5810 /* The caret is output if there are any defaults: if not all the STD
5811 * flags are set, or if no character set specifier is needed */
5813 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
5815 bool has_runon = ((RExC_seen & REG_SEEN_RUN_ON_COMMENT)==REG_SEEN_RUN_ON_COMMENT);
5816 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
5817 >> RXf_PMf_STD_PMMOD_SHIFT);
5818 const char *fptr = STD_PAT_MODS; /*"msix"*/
5820 /* Allocate for the worst case, which is all the std flags are turned
5821 * on. If more precision is desired, we could do a population count of
5822 * the flags set. This could be done with a small lookup table, or by
5823 * shifting, masking and adding, or even, when available, assembly
5824 * language for a machine-language population count.
5825 * We never output a minus, as all those are defaults, so are
5826 * covered by the caret */
5827 const STRLEN wraplen = plen + has_p + has_runon
5828 + has_default /* If needs a caret */
5830 /* If needs a character set specifier */
5831 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
5832 + (sizeof(STD_PAT_MODS) - 1)
5833 + (sizeof("(?:)") - 1);
5835 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
5836 r->xpv_len_u.xpvlenu_pv = p;
5838 SvFLAGS(rx) |= SVf_UTF8;
5841 /* If a default, cover it using the caret */
5843 *p++= DEFAULT_PAT_MOD;
5847 const char* const name = get_regex_charset_name(r->extflags, &len);
5848 Copy(name, p, len, char);
5852 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
5855 while((ch = *fptr++)) {
5863 Copy(RExC_precomp, p, plen, char);
5864 assert ((RX_WRAPPED(rx) - p) < 16);
5865 r->pre_prefix = p - RX_WRAPPED(rx);
5871 SvCUR_set(rx, p - RX_WRAPPED(rx));
5875 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
5877 if (RExC_seen & REG_SEEN_RECURSE) {
5878 Newxz(RExC_open_parens, RExC_npar,regnode *);
5879 SAVEFREEPV(RExC_open_parens);
5880 Newxz(RExC_close_parens,RExC_npar,regnode *);
5881 SAVEFREEPV(RExC_close_parens);
5884 /* Useful during FAIL. */
5885 #ifdef RE_TRACK_PATTERN_OFFSETS
5886 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
5887 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
5888 "%s %"UVuf" bytes for offset annotations.\n",
5889 ri->u.offsets ? "Got" : "Couldn't get",
5890 (UV)((2*RExC_size+1) * sizeof(U32))));
5892 SetProgLen(ri,RExC_size);
5897 /* Second pass: emit code. */
5898 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
5899 RExC_pm_flags = pm_flags;
5901 RExC_end = exp + plen;
5904 RExC_emit_start = ri->program;
5905 RExC_emit = ri->program;
5906 RExC_emit_bound = ri->program + RExC_size + 1;
5907 pRExC_state->code_index = 0;
5909 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
5910 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5912 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#X", flags);
5914 /* XXXX To minimize changes to RE engine we always allocate
5915 3-units-long substrs field. */
5916 Newx(r->substrs, 1, struct reg_substr_data);
5917 if (RExC_recurse_count) {
5918 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
5919 SAVEFREEPV(RExC_recurse);
5923 r->minlen = minlen = sawlookahead = sawplus = sawopen = 0;
5924 Zero(r->substrs, 1, struct reg_substr_data);
5926 #ifdef TRIE_STUDY_OPT
5928 StructCopy(&zero_scan_data, &data, scan_data_t);
5929 copyRExC_state = RExC_state;
5932 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
5934 RExC_state = copyRExC_state;
5935 if (seen & REG_TOP_LEVEL_BRANCHES)
5936 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
5938 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES;
5939 StructCopy(&zero_scan_data, &data, scan_data_t);
5942 StructCopy(&zero_scan_data, &data, scan_data_t);
5945 /* Dig out information for optimizations. */
5946 r->extflags = RExC_flags; /* was pm_op */
5947 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
5950 SvUTF8_on(rx); /* Unicode in it? */
5951 ri->regstclass = NULL;
5952 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
5953 r->intflags |= PREGf_NAUGHTY;
5954 scan = ri->program + 1; /* First BRANCH. */
5956 /* testing for BRANCH here tells us whether there is "must appear"
5957 data in the pattern. If there is then we can use it for optimisations */
5958 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES)) { /* Only one top-level choice. */
5960 STRLEN longest_float_length, longest_fixed_length;
5961 struct regnode_charclass_class ch_class; /* pointed to by data */
5963 I32 last_close = 0; /* pointed to by data */
5964 regnode *first= scan;
5965 regnode *first_next= regnext(first);
5967 * Skip introductions and multiplicators >= 1
5968 * so that we can extract the 'meat' of the pattern that must
5969 * match in the large if() sequence following.
5970 * NOTE that EXACT is NOT covered here, as it is normally
5971 * picked up by the optimiser separately.
5973 * This is unfortunate as the optimiser isnt handling lookahead
5974 * properly currently.
5977 while ((OP(first) == OPEN && (sawopen = 1)) ||
5978 /* An OR of *one* alternative - should not happen now. */
5979 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
5980 /* for now we can't handle lookbehind IFMATCH*/
5981 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
5982 (OP(first) == PLUS) ||
5983 (OP(first) == MINMOD) ||
5984 /* An {n,m} with n>0 */
5985 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
5986 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
5989 * the only op that could be a regnode is PLUS, all the rest
5990 * will be regnode_1 or regnode_2.
5993 if (OP(first) == PLUS)
5996 first += regarglen[OP(first)];
5998 first = NEXTOPER(first);
5999 first_next= regnext(first);
6002 /* Starting-point info. */
6004 DEBUG_PEEP("first:",first,0);
6005 /* Ignore EXACT as we deal with it later. */
6006 if (PL_regkind[OP(first)] == EXACT) {
6007 if (OP(first) == EXACT)
6008 NOOP; /* Empty, get anchored substr later. */
6010 ri->regstclass = first;
6013 else if (PL_regkind[OP(first)] == TRIE &&
6014 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6017 /* this can happen only on restudy */
6018 if ( OP(first) == TRIE ) {
6019 struct regnode_1 *trieop = (struct regnode_1 *)
6020 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6021 StructCopy(first,trieop,struct regnode_1);
6022 trie_op=(regnode *)trieop;
6024 struct regnode_charclass *trieop = (struct regnode_charclass *)
6025 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6026 StructCopy(first,trieop,struct regnode_charclass);
6027 trie_op=(regnode *)trieop;
6030 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6031 ri->regstclass = trie_op;
6034 else if (REGNODE_SIMPLE(OP(first)))
6035 ri->regstclass = first;
6036 else if (PL_regkind[OP(first)] == BOUND ||
6037 PL_regkind[OP(first)] == NBOUND)
6038 ri->regstclass = first;
6039 else if (PL_regkind[OP(first)] == BOL) {
6040 r->extflags |= (OP(first) == MBOL
6042 : (OP(first) == SBOL
6045 first = NEXTOPER(first);
6048 else if (OP(first) == GPOS) {
6049 r->extflags |= RXf_ANCH_GPOS;
6050 first = NEXTOPER(first);
6053 else if ((!sawopen || !RExC_sawback) &&
6054 (OP(first) == STAR &&
6055 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6056 !(r->extflags & RXf_ANCH) && !pRExC_state->num_code_blocks)
6058 /* turn .* into ^.* with an implied $*=1 */
6060 (OP(NEXTOPER(first)) == REG_ANY)
6063 r->extflags |= type;
6064 r->intflags |= PREGf_IMPLICIT;
6065 first = NEXTOPER(first);
6068 if (sawplus && !sawlookahead && (!sawopen || !RExC_sawback)
6069 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6070 /* x+ must match at the 1st pos of run of x's */
6071 r->intflags |= PREGf_SKIP;
6073 /* Scan is after the zeroth branch, first is atomic matcher. */
6074 #ifdef TRIE_STUDY_OPT
6077 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6078 (IV)(first - scan + 1))
6082 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6083 (IV)(first - scan + 1))
6089 * If there's something expensive in the r.e., find the
6090 * longest literal string that must appear and make it the
6091 * regmust. Resolve ties in favor of later strings, since
6092 * the regstart check works with the beginning of the r.e.
6093 * and avoiding duplication strengthens checking. Not a
6094 * strong reason, but sufficient in the absence of others.
6095 * [Now we resolve ties in favor of the earlier string if
6096 * it happens that c_offset_min has been invalidated, since the
6097 * earlier string may buy us something the later one won't.]
6100 data.longest_fixed = newSVpvs("");
6101 data.longest_float = newSVpvs("");
6102 data.last_found = newSVpvs("");
6103 data.longest = &(data.longest_fixed);
6104 ENTER_with_name("study_chunk");
6105 SAVEFREESV(data.longest_fixed);
6106 SAVEFREESV(data.longest_float);
6107 SAVEFREESV(data.last_found);
6109 if (!ri->regstclass) {
6110 cl_init(pRExC_state, &ch_class);
6111 data.start_class = &ch_class;
6112 stclass_flag = SCF_DO_STCLASS_AND;
6113 } else /* XXXX Check for BOUND? */
6115 data.last_closep = &last_close;
6117 minlen = study_chunk(pRExC_state, &first, &minlen, &fake, scan + RExC_size, /* Up to end */
6118 &data, -1, NULL, NULL,
6119 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag,0);
6122 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6125 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6126 && data.last_start_min == 0 && data.last_end > 0
6127 && !RExC_seen_zerolen
6128 && !(RExC_seen & REG_SEEN_VERBARG)
6129 && (!(RExC_seen & REG_SEEN_GPOS) || (r->extflags & RXf_ANCH_GPOS)))
6130 r->extflags |= RXf_CHECK_ALL;
6131 scan_commit(pRExC_state, &data,&minlen,0);
6133 longest_float_length = CHR_SVLEN(data.longest_float);
6135 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6136 && data.offset_fixed == data.offset_float_min
6137 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6138 && S_setup_longest (aTHX_ pRExC_state,
6142 &(r->float_end_shift),
6143 data.lookbehind_float,
6144 data.offset_float_min,
6146 longest_float_length,
6147 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6148 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6150 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6151 r->float_max_offset = data.offset_float_max;
6152 if (data.offset_float_max < I32_MAX) /* Don't offset infinity */
6153 r->float_max_offset -= data.lookbehind_float;
6154 SvREFCNT_inc_simple_void_NN(data.longest_float);
6157 r->float_substr = r->float_utf8 = NULL;
6158 longest_float_length = 0;
6161 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6163 if (S_setup_longest (aTHX_ pRExC_state,
6165 &(r->anchored_utf8),
6166 &(r->anchored_substr),
6167 &(r->anchored_end_shift),
6168 data.lookbehind_fixed,
6171 longest_fixed_length,
6172 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6173 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6175 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6176 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6179 r->anchored_substr = r->anchored_utf8 = NULL;
6180 longest_fixed_length = 0;
6182 LEAVE_with_name("study_chunk");
6185 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6186 ri->regstclass = NULL;
6188 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6190 && ! TEST_SSC_EOS(data.start_class)
6191 && !cl_is_anything(data.start_class))
6193 const U32 n = add_data(pRExC_state, 1, "f");
6194 OP(data.start_class) = ANYOF_SYNTHETIC;
6196 Newx(RExC_rxi->data->data[n], 1,
6197 struct regnode_charclass_class);
6198 StructCopy(data.start_class,
6199 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6200 struct regnode_charclass_class);
6201 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6202 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6203 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6204 regprop(r, sv, (regnode*)data.start_class);
6205 PerlIO_printf(Perl_debug_log,
6206 "synthetic stclass \"%s\".\n",
6207 SvPVX_const(sv));});
6210 /* A temporary algorithm prefers floated substr to fixed one to dig more info. */
6211 if (longest_fixed_length > longest_float_length) {
6212 r->check_end_shift = r->anchored_end_shift;
6213 r->check_substr = r->anchored_substr;
6214 r->check_utf8 = r->anchored_utf8;
6215 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6216 if (r->extflags & RXf_ANCH_SINGLE)
6217 r->extflags |= RXf_NOSCAN;
6220 r->check_end_shift = r->float_end_shift;
6221 r->check_substr = r->float_substr;
6222 r->check_utf8 = r->float_utf8;
6223 r->check_offset_min = r->float_min_offset;
6224 r->check_offset_max = r->float_max_offset;
6226 /* XXXX Currently intuiting is not compatible with ANCH_GPOS.
6227 This should be changed ASAP! */
6228 if ((r->check_substr || r->check_utf8) && !(r->extflags & RXf_ANCH_GPOS)) {
6229 r->extflags |= RXf_USE_INTUIT;
6230 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6231 r->extflags |= RXf_INTUIT_TAIL;
6233 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6234 if ( (STRLEN)minlen < longest_float_length )
6235 minlen= longest_float_length;
6236 if ( (STRLEN)minlen < longest_fixed_length )
6237 minlen= longest_fixed_length;
6241 /* Several toplevels. Best we can is to set minlen. */
6243 struct regnode_charclass_class ch_class;
6246 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6248 scan = ri->program + 1;
6249 cl_init(pRExC_state, &ch_class);
6250 data.start_class = &ch_class;
6251 data.last_closep = &last_close;
6254 minlen = study_chunk(pRExC_state, &scan, &minlen, &fake, scan + RExC_size,
6255 &data, -1, NULL, NULL, SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS,0);
6257 CHECK_RESTUDY_GOTO_butfirst(NOOP);
6259 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
6260 = r->float_substr = r->float_utf8 = NULL;
6262 if (! TEST_SSC_EOS(data.start_class)
6263 && !cl_is_anything(data.start_class))
6265 const U32 n = add_data(pRExC_state, 1, "f");
6266 OP(data.start_class) = ANYOF_SYNTHETIC;
6268 Newx(RExC_rxi->data->data[n], 1,
6269 struct regnode_charclass_class);
6270 StructCopy(data.start_class,
6271 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6272 struct regnode_charclass_class);
6273 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6274 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6275 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
6276 regprop(r, sv, (regnode*)data.start_class);
6277 PerlIO_printf(Perl_debug_log,
6278 "synthetic stclass \"%s\".\n",
6279 SvPVX_const(sv));});
6283 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
6284 the "real" pattern. */
6286 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf"\n",
6287 (IV)minlen, (IV)r->minlen);
6289 r->minlenret = minlen;
6290 if (r->minlen < minlen)
6293 if (RExC_seen & REG_SEEN_GPOS)
6294 r->extflags |= RXf_GPOS_SEEN;
6295 if (RExC_seen & REG_SEEN_LOOKBEHIND)
6296 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the lookbehind */
6297 if (pRExC_state->num_code_blocks)
6298 r->extflags |= RXf_EVAL_SEEN;
6299 if (RExC_seen & REG_SEEN_CANY)
6300 r->extflags |= RXf_CANY_SEEN;
6301 if (RExC_seen & REG_SEEN_VERBARG)
6303 r->intflags |= PREGf_VERBARG_SEEN;
6304 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
6306 if (RExC_seen & REG_SEEN_CUTGROUP)
6307 r->intflags |= PREGf_CUTGROUP_SEEN;
6308 if (pm_flags & PMf_USE_RE_EVAL)
6309 r->intflags |= PREGf_USE_RE_EVAL;
6310 if (RExC_paren_names)
6311 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
6313 RXp_PAREN_NAMES(r) = NULL;
6316 regnode *first = ri->program + 1;
6318 regnode *next = NEXTOPER(first);
6321 if (PL_regkind[fop] == NOTHING && nop == END)
6322 r->extflags |= RXf_NULL;
6323 else if (PL_regkind[fop] == BOL && nop == END)
6324 r->extflags |= RXf_START_ONLY;
6325 else if (fop == PLUS && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE && OP(regnext(first)) == END)
6326 r->extflags |= RXf_WHITE;
6327 else if ( r->extflags & RXf_SPLIT && fop == EXACT && STR_LEN(first) == 1 && *(STRING(first)) == ' ' && OP(regnext(first)) == END )
6328 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
6332 if (RExC_paren_names) {
6333 ri->name_list_idx = add_data( pRExC_state, 1, "a" );
6334 ri->data->data[ri->name_list_idx] = (void*)SvREFCNT_inc(RExC_paren_name_list);
6337 ri->name_list_idx = 0;
6339 if (RExC_recurse_count) {
6340 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
6341 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
6342 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
6345 Newxz(r->offs, RExC_npar, regexp_paren_pair);
6346 /* assume we don't need to swap parens around before we match */
6349 PerlIO_printf(Perl_debug_log,"Final program:\n");
6352 #ifdef RE_TRACK_PATTERN_OFFSETS
6353 DEBUG_OFFSETS_r(if (ri->u.offsets) {
6354 const U32 len = ri->u.offsets[0];
6356 GET_RE_DEBUG_FLAGS_DECL;
6357 PerlIO_printf(Perl_debug_log, "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
6358 for (i = 1; i <= len; i++) {
6359 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
6360 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
6361 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
6363 PerlIO_printf(Perl_debug_log, "\n");
6368 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
6369 * by setting the regexp SV to readonly-only instead. If the
6370 * pattern's been recompiled, the USEDness should remain. */
6371 if (old_re && SvREADONLY(old_re))
6379 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
6382 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
6384 PERL_UNUSED_ARG(value);
6386 if (flags & RXapif_FETCH) {
6387 return reg_named_buff_fetch(rx, key, flags);
6388 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
6389 Perl_croak_no_modify();
6391 } else if (flags & RXapif_EXISTS) {
6392 return reg_named_buff_exists(rx, key, flags)
6395 } else if (flags & RXapif_REGNAMES) {
6396 return reg_named_buff_all(rx, flags);
6397 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
6398 return reg_named_buff_scalar(rx, flags);
6400 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
6406 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
6409 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
6410 PERL_UNUSED_ARG(lastkey);
6412 if (flags & RXapif_FIRSTKEY)
6413 return reg_named_buff_firstkey(rx, flags);
6414 else if (flags & RXapif_NEXTKEY)
6415 return reg_named_buff_nextkey(rx, flags);
6417 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter", (int)flags);
6423 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
6426 AV *retarray = NULL;
6428 struct regexp *const rx = ReANY(r);
6430 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
6432 if (flags & RXapif_ALL)
6435 if (rx && RXp_PAREN_NAMES(rx)) {
6436 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
6439 SV* sv_dat=HeVAL(he_str);
6440 I32 *nums=(I32*)SvPVX(sv_dat);
6441 for ( i=0; i<SvIVX(sv_dat); i++ ) {
6442 if ((I32)(rx->nparens) >= nums[i]
6443 && rx->offs[nums[i]].start != -1
6444 && rx->offs[nums[i]].end != -1)
6447 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
6452 ret = newSVsv(&PL_sv_undef);
6455 av_push(retarray, ret);
6458 return newRV_noinc(MUTABLE_SV(retarray));
6465 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
6468 struct regexp *const rx = ReANY(r);
6470 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
6472 if (rx && RXp_PAREN_NAMES(rx)) {
6473 if (flags & RXapif_ALL) {
6474 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
6476 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
6478 SvREFCNT_dec_NN(sv);
6490 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
6492 struct regexp *const rx = ReANY(r);
6494 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
6496 if ( rx && RXp_PAREN_NAMES(rx) ) {
6497 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
6499 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
6506 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
6508 struct regexp *const rx = ReANY(r);
6509 GET_RE_DEBUG_FLAGS_DECL;
6511 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
6513 if (rx && RXp_PAREN_NAMES(rx)) {
6514 HV *hv = RXp_PAREN_NAMES(rx);
6516 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6519 SV* sv_dat = HeVAL(temphe);
6520 I32 *nums = (I32*)SvPVX(sv_dat);
6521 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6522 if ((I32)(rx->lastparen) >= nums[i] &&
6523 rx->offs[nums[i]].start != -1 &&
6524 rx->offs[nums[i]].end != -1)
6530 if (parno || flags & RXapif_ALL) {
6531 return newSVhek(HeKEY_hek(temphe));
6539 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
6544 struct regexp *const rx = ReANY(r);
6546 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
6548 if (rx && RXp_PAREN_NAMES(rx)) {
6549 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
6550 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
6551 } else if (flags & RXapif_ONE) {
6552 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
6553 av = MUTABLE_AV(SvRV(ret));
6554 length = av_len(av);
6555 SvREFCNT_dec_NN(ret);
6556 return newSViv(length + 1);
6558 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar", (int)flags);
6562 return &PL_sv_undef;
6566 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
6568 struct regexp *const rx = ReANY(r);
6571 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
6573 if (rx && RXp_PAREN_NAMES(rx)) {
6574 HV *hv= RXp_PAREN_NAMES(rx);
6576 (void)hv_iterinit(hv);
6577 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6580 SV* sv_dat = HeVAL(temphe);
6581 I32 *nums = (I32*)SvPVX(sv_dat);
6582 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6583 if ((I32)(rx->lastparen) >= nums[i] &&
6584 rx->offs[nums[i]].start != -1 &&
6585 rx->offs[nums[i]].end != -1)
6591 if (parno || flags & RXapif_ALL) {
6592 av_push(av, newSVhek(HeKEY_hek(temphe)));
6597 return newRV_noinc(MUTABLE_SV(av));
6601 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
6604 struct regexp *const rx = ReANY(r);
6610 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
6612 if ( ( n == RX_BUFF_IDX_CARET_PREMATCH
6613 || n == RX_BUFF_IDX_CARET_FULLMATCH
6614 || n == RX_BUFF_IDX_CARET_POSTMATCH
6616 && !(rx->extflags & RXf_PMf_KEEPCOPY)
6623 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
6624 /* no need to distinguish between them any more */
6625 n = RX_BUFF_IDX_FULLMATCH;
6627 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
6628 && rx->offs[0].start != -1)
6630 /* $`, ${^PREMATCH} */
6631 i = rx->offs[0].start;
6635 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
6636 && rx->offs[0].end != -1)
6638 /* $', ${^POSTMATCH} */
6639 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
6640 i = rx->sublen + rx->suboffset - rx->offs[0].end;
6643 if ( 0 <= n && n <= (I32)rx->nparens &&
6644 (s1 = rx->offs[n].start) != -1 &&
6645 (t1 = rx->offs[n].end) != -1)
6647 /* $&, ${^MATCH}, $1 ... */
6649 s = rx->subbeg + s1 - rx->suboffset;
6654 assert(s >= rx->subbeg);
6655 assert(rx->sublen >= (s - rx->subbeg) + i );
6657 #if NO_TAINT_SUPPORT
6658 sv_setpvn(sv, s, i);
6660 const int oldtainted = TAINT_get;
6662 sv_setpvn(sv, s, i);
6663 TAINT_set(oldtainted);
6665 if ( (rx->extflags & RXf_CANY_SEEN)
6666 ? (RXp_MATCH_UTF8(rx)
6667 && (!i || is_utf8_string((U8*)s, i)))
6668 : (RXp_MATCH_UTF8(rx)) )
6675 if (RXp_MATCH_TAINTED(rx)) {
6676 if (SvTYPE(sv) >= SVt_PVMG) {
6677 MAGIC* const mg = SvMAGIC(sv);
6680 SvMAGIC_set(sv, mg->mg_moremagic);
6682 if ((mgt = SvMAGIC(sv))) {
6683 mg->mg_moremagic = mgt;
6684 SvMAGIC_set(sv, mg);
6695 sv_setsv(sv,&PL_sv_undef);
6701 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
6702 SV const * const value)
6704 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
6706 PERL_UNUSED_ARG(rx);
6707 PERL_UNUSED_ARG(paren);
6708 PERL_UNUSED_ARG(value);
6711 Perl_croak_no_modify();
6715 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
6718 struct regexp *const rx = ReANY(r);
6722 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
6724 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
6726 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
6727 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6731 case RX_BUFF_IDX_PREMATCH: /* $` */
6732 if (rx->offs[0].start != -1) {
6733 i = rx->offs[0].start;
6742 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
6743 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6745 case RX_BUFF_IDX_POSTMATCH: /* $' */
6746 if (rx->offs[0].end != -1) {
6747 i = rx->sublen - rx->offs[0].end;
6749 s1 = rx->offs[0].end;
6756 case RX_BUFF_IDX_CARET_FULLMATCH: /* ${^MATCH} */
6757 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6761 /* $& / ${^MATCH}, $1, $2, ... */
6763 if (paren <= (I32)rx->nparens &&
6764 (s1 = rx->offs[paren].start) != -1 &&
6765 (t1 = rx->offs[paren].end) != -1)
6771 if (ckWARN(WARN_UNINITIALIZED))
6772 report_uninit((const SV *)sv);
6777 if (i > 0 && RXp_MATCH_UTF8(rx)) {
6778 const char * const s = rx->subbeg - rx->suboffset + s1;
6783 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
6790 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
6792 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
6793 PERL_UNUSED_ARG(rx);
6797 return newSVpvs("Regexp");
6800 /* Scans the name of a named buffer from the pattern.
6801 * If flags is REG_RSN_RETURN_NULL returns null.
6802 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
6803 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
6804 * to the parsed name as looked up in the RExC_paren_names hash.
6805 * If there is an error throws a vFAIL().. type exception.
6808 #define REG_RSN_RETURN_NULL 0
6809 #define REG_RSN_RETURN_NAME 1
6810 #define REG_RSN_RETURN_DATA 2
6813 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
6815 char *name_start = RExC_parse;
6817 PERL_ARGS_ASSERT_REG_SCAN_NAME;
6819 if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
6820 /* skip IDFIRST by using do...while */
6823 RExC_parse += UTF8SKIP(RExC_parse);
6824 } while (isWORDCHAR_utf8((U8*)RExC_parse));
6828 } while (isWORDCHAR(*RExC_parse));
6830 RExC_parse++; /* so the <- from the vFAIL is after the offending character */
6831 vFAIL("Group name must start with a non-digit word character");
6835 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
6836 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
6837 if ( flags == REG_RSN_RETURN_NAME)
6839 else if (flags==REG_RSN_RETURN_DATA) {
6842 if ( ! sv_name ) /* should not happen*/
6843 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
6844 if (RExC_paren_names)
6845 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
6847 sv_dat = HeVAL(he_str);
6849 vFAIL("Reference to nonexistent named group");
6853 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
6854 (unsigned long) flags);
6856 assert(0); /* NOT REACHED */
6861 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
6862 int rem=(int)(RExC_end - RExC_parse); \
6871 if (RExC_lastparse!=RExC_parse) \
6872 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
6875 iscut ? "..." : "<" \
6878 PerlIO_printf(Perl_debug_log,"%16s",""); \
6881 num = RExC_size + 1; \
6883 num=REG_NODE_NUM(RExC_emit); \
6884 if (RExC_lastnum!=num) \
6885 PerlIO_printf(Perl_debug_log,"|%4d",num); \
6887 PerlIO_printf(Perl_debug_log,"|%4s",""); \
6888 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
6889 (int)((depth*2)), "", \
6893 RExC_lastparse=RExC_parse; \
6898 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
6899 DEBUG_PARSE_MSG((funcname)); \
6900 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
6902 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
6903 DEBUG_PARSE_MSG((funcname)); \
6904 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
6907 /* This section of code defines the inversion list object and its methods. The
6908 * interfaces are highly subject to change, so as much as possible is static to
6909 * this file. An inversion list is here implemented as a malloc'd C UV array
6910 * with some added info that is placed as UVs at the beginning in a header
6911 * portion. An inversion list for Unicode is an array of code points, sorted
6912 * by ordinal number. The zeroth element is the first code point in the list.
6913 * The 1th element is the first element beyond that not in the list. In other
6914 * words, the first range is
6915 * invlist[0]..(invlist[1]-1)
6916 * The other ranges follow. Thus every element whose index is divisible by two
6917 * marks the beginning of a range that is in the list, and every element not
6918 * divisible by two marks the beginning of a range not in the list. A single
6919 * element inversion list that contains the single code point N generally
6920 * consists of two elements
6923 * (The exception is when N is the highest representable value on the
6924 * machine, in which case the list containing just it would be a single
6925 * element, itself. By extension, if the last range in the list extends to
6926 * infinity, then the first element of that range will be in the inversion list
6927 * at a position that is divisible by two, and is the final element in the
6929 * Taking the complement (inverting) an inversion list is quite simple, if the
6930 * first element is 0, remove it; otherwise add a 0 element at the beginning.
6931 * This implementation reserves an element at the beginning of each inversion
6932 * list to contain 0 when the list contains 0, and contains 1 otherwise. The
6933 * actual beginning of the list is either that element if 0, or the next one if
6936 * More about inversion lists can be found in "Unicode Demystified"
6937 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
6938 * More will be coming when functionality is added later.
6940 * The inversion list data structure is currently implemented as an SV pointing
6941 * to an array of UVs that the SV thinks are bytes. This allows us to have an
6942 * array of UV whose memory management is automatically handled by the existing
6943 * facilities for SV's.
6945 * Some of the methods should always be private to the implementation, and some
6946 * should eventually be made public */
6948 /* The header definitions are in F<inline_invlist.c> */
6949 #define TO_INTERNAL_SIZE(x) (((x) + HEADER_LENGTH) * sizeof(UV))
6950 #define FROM_INTERNAL_SIZE(x) (((x)/ sizeof(UV)) - HEADER_LENGTH)
6952 #define INVLIST_INITIAL_LEN 10
6954 PERL_STATIC_INLINE UV*
6955 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
6957 /* Returns a pointer to the first element in the inversion list's array.
6958 * This is called upon initialization of an inversion list. Where the
6959 * array begins depends on whether the list has the code point U+0000
6960 * in it or not. The other parameter tells it whether the code that
6961 * follows this call is about to put a 0 in the inversion list or not.
6962 * The first element is either the element with 0, if 0, or the next one,
6965 UV* zero = get_invlist_zero_addr(invlist);
6967 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
6970 assert(! *_get_invlist_len_addr(invlist));
6972 /* 1^1 = 0; 1^0 = 1 */
6973 *zero = 1 ^ will_have_0;
6974 return zero + *zero;
6977 PERL_STATIC_INLINE UV*
6978 S_invlist_array(pTHX_ SV* const invlist)
6980 /* Returns the pointer to the inversion list's array. Every time the
6981 * length changes, this needs to be called in case malloc or realloc moved
6984 PERL_ARGS_ASSERT_INVLIST_ARRAY;
6986 /* Must not be empty. If these fail, you probably didn't check for <len>
6987 * being non-zero before trying to get the array */
6988 assert(*_get_invlist_len_addr(invlist));
6989 assert(*get_invlist_zero_addr(invlist) == 0
6990 || *get_invlist_zero_addr(invlist) == 1);
6992 /* The array begins either at the element reserved for zero if the
6993 * list contains 0 (that element will be set to 0), or otherwise the next
6994 * element (in which case the reserved element will be set to 1). */
6995 return (UV *) (get_invlist_zero_addr(invlist)
6996 + *get_invlist_zero_addr(invlist));
6999 PERL_STATIC_INLINE void
7000 S_invlist_set_len(pTHX_ SV* const invlist, const UV len)
7002 /* Sets the current number of elements stored in the inversion list */
7004 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7006 *_get_invlist_len_addr(invlist) = len;
7008 assert(len <= SvLEN(invlist));
7010 SvCUR_set(invlist, TO_INTERNAL_SIZE(len));
7011 /* If the list contains U+0000, that element is part of the header,
7012 * and should not be counted as part of the array. It will contain
7013 * 0 in that case, and 1 otherwise. So we could flop 0=>1, 1=>0 and
7015 * SvCUR_set(invlist,
7016 * TO_INTERNAL_SIZE(len
7017 * - (*get_invlist_zero_addr(inv_list) ^ 1)));
7018 * But, this is only valid if len is not 0. The consequences of not doing
7019 * this is that the memory allocation code may think that 1 more UV is
7020 * being used than actually is, and so might do an unnecessary grow. That
7021 * seems worth not bothering to make this the precise amount.
7023 * Note that when inverting, SvCUR shouldn't change */
7026 PERL_STATIC_INLINE IV*
7027 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7029 /* Return the address of the UV that is reserved to hold the cached index
7032 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7034 return (IV *) (SvPVX(invlist) + (INVLIST_PREVIOUS_INDEX_OFFSET * sizeof (UV)));
7037 PERL_STATIC_INLINE IV
7038 S_invlist_previous_index(pTHX_ SV* const invlist)
7040 /* Returns cached index of previous search */
7042 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7044 return *get_invlist_previous_index_addr(invlist);
7047 PERL_STATIC_INLINE void
7048 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7050 /* Caches <index> for later retrieval */
7052 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7054 assert(index == 0 || index < (int) _invlist_len(invlist));
7056 *get_invlist_previous_index_addr(invlist) = index;
7059 PERL_STATIC_INLINE UV
7060 S_invlist_max(pTHX_ SV* const invlist)
7062 /* Returns the maximum number of elements storable in the inversion list's
7063 * array, without having to realloc() */
7065 PERL_ARGS_ASSERT_INVLIST_MAX;
7067 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7068 ? _invlist_len(invlist)
7069 : FROM_INTERNAL_SIZE(SvLEN(invlist));
7072 PERL_STATIC_INLINE UV*
7073 S_get_invlist_zero_addr(pTHX_ SV* invlist)
7075 /* Return the address of the UV that is reserved to hold 0 if the inversion
7076 * list contains 0. This has to be the last element of the heading, as the
7077 * list proper starts with either it if 0, or the next element if not.
7078 * (But we force it to contain either 0 or 1) */
7080 PERL_ARGS_ASSERT_GET_INVLIST_ZERO_ADDR;
7082 return (UV *) (SvPVX(invlist) + (INVLIST_ZERO_OFFSET * sizeof (UV)));
7085 #ifndef PERL_IN_XSUB_RE
7087 Perl__new_invlist(pTHX_ IV initial_size)
7090 /* Return a pointer to a newly constructed inversion list, with enough
7091 * space to store 'initial_size' elements. If that number is negative, a
7092 * system default is used instead */
7096 if (initial_size < 0) {
7097 initial_size = INVLIST_INITIAL_LEN;
7100 /* Allocate the initial space */
7101 new_list = newSV(TO_INTERNAL_SIZE(initial_size));
7102 invlist_set_len(new_list, 0);
7104 /* Force iterinit() to be used to get iteration to work */
7105 *get_invlist_iter_addr(new_list) = UV_MAX;
7107 /* This should force a segfault if a method doesn't initialize this
7109 *get_invlist_zero_addr(new_list) = UV_MAX;
7111 *get_invlist_previous_index_addr(new_list) = 0;
7112 *get_invlist_version_id_addr(new_list) = INVLIST_VERSION_ID;
7113 #if HEADER_LENGTH != 5
7114 # 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
7122 S__new_invlist_C_array(pTHX_ UV* list)
7124 /* Return a pointer to a newly constructed inversion list, initialized to
7125 * point to <list>, which has to be in the exact correct inversion list
7126 * form, including internal fields. Thus this is a dangerous routine that
7127 * should not be used in the wrong hands */
7129 SV* invlist = newSV_type(SVt_PV);
7131 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7133 SvPV_set(invlist, (char *) list);
7134 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7135 shouldn't touch it */
7136 SvCUR_set(invlist, TO_INTERNAL_SIZE(_invlist_len(invlist)));
7138 if (*get_invlist_version_id_addr(invlist) != INVLIST_VERSION_ID) {
7139 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7142 /* Initialize the iteration pointer.
7143 * XXX This could be done at compile time in charclass_invlists.h, but I
7144 * (khw) am not confident that the suffixes for specifying the C constant
7145 * UV_MAX are portable, e.g. 'ull' on a 32 bit machine that is configured
7146 * to use 64 bits; might need a Configure probe */
7147 invlist_iterfinish(invlist);
7153 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7155 /* Grow the maximum size of an inversion list */
7157 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7159 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max));
7162 PERL_STATIC_INLINE void
7163 S_invlist_trim(pTHX_ SV* const invlist)
7165 PERL_ARGS_ASSERT_INVLIST_TRIM;
7167 /* Change the length of the inversion list to how many entries it currently
7170 SvPV_shrink_to_cur((SV *) invlist);
7173 #define _invlist_union_complement_2nd(a, b, output) _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
7176 S__append_range_to_invlist(pTHX_ SV* const invlist, const UV start, const UV end)
7178 /* Subject to change or removal. Append the range from 'start' to 'end' at
7179 * the end of the inversion list. The range must be above any existing
7183 UV max = invlist_max(invlist);
7184 UV len = _invlist_len(invlist);
7186 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7188 if (len == 0) { /* Empty lists must be initialized */
7189 array = _invlist_array_init(invlist, start == 0);
7192 /* Here, the existing list is non-empty. The current max entry in the
7193 * list is generally the first value not in the set, except when the
7194 * set extends to the end of permissible values, in which case it is
7195 * the first entry in that final set, and so this call is an attempt to
7196 * append out-of-order */
7198 UV final_element = len - 1;
7199 array = invlist_array(invlist);
7200 if (array[final_element] > start
7201 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
7203 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",
7204 array[final_element], start,
7205 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
7208 /* Here, it is a legal append. If the new range begins with the first
7209 * value not in the set, it is extending the set, so the new first
7210 * value not in the set is one greater than the newly extended range.
7212 if (array[final_element] == start) {
7213 if (end != UV_MAX) {
7214 array[final_element] = end + 1;
7217 /* But if the end is the maximum representable on the machine,
7218 * just let the range that this would extend to have no end */
7219 invlist_set_len(invlist, len - 1);
7225 /* Here the new range doesn't extend any existing set. Add it */
7227 len += 2; /* Includes an element each for the start and end of range */
7229 /* If overflows the existing space, extend, which may cause the array to be
7232 invlist_extend(invlist, len);
7233 invlist_set_len(invlist, len); /* Have to set len here to avoid assert
7234 failure in invlist_array() */
7235 array = invlist_array(invlist);
7238 invlist_set_len(invlist, len);
7241 /* The next item on the list starts the range, the one after that is
7242 * one past the new range. */
7243 array[len - 2] = start;
7244 if (end != UV_MAX) {
7245 array[len - 1] = end + 1;
7248 /* But if the end is the maximum representable on the machine, just let
7249 * the range have no end */
7250 invlist_set_len(invlist, len - 1);
7254 #ifndef PERL_IN_XSUB_RE
7257 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
7259 /* Searches the inversion list for the entry that contains the input code
7260 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
7261 * return value is the index into the list's array of the range that
7266 IV high = _invlist_len(invlist);
7267 const IV highest_element = high - 1;
7270 PERL_ARGS_ASSERT__INVLIST_SEARCH;
7272 /* If list is empty, return failure. */
7277 /* (We can't get the array unless we know the list is non-empty) */
7278 array = invlist_array(invlist);
7280 mid = invlist_previous_index(invlist);
7281 assert(mid >=0 && mid <= highest_element);
7283 /* <mid> contains the cache of the result of the previous call to this
7284 * function (0 the first time). See if this call is for the same result,
7285 * or if it is for mid-1. This is under the theory that calls to this
7286 * function will often be for related code points that are near each other.
7287 * And benchmarks show that caching gives better results. We also test
7288 * here if the code point is within the bounds of the list. These tests
7289 * replace others that would have had to be made anyway to make sure that
7290 * the array bounds were not exceeded, and these give us extra information
7291 * at the same time */
7292 if (cp >= array[mid]) {
7293 if (cp >= array[highest_element]) {
7294 return highest_element;
7297 /* Here, array[mid] <= cp < array[highest_element]. This means that
7298 * the final element is not the answer, so can exclude it; it also
7299 * means that <mid> is not the final element, so can refer to 'mid + 1'
7301 if (cp < array[mid + 1]) {
7307 else { /* cp < aray[mid] */
7308 if (cp < array[0]) { /* Fail if outside the array */
7312 if (cp >= array[mid - 1]) {
7317 /* Binary search. What we are looking for is <i> such that
7318 * array[i] <= cp < array[i+1]
7319 * The loop below converges on the i+1. Note that there may not be an
7320 * (i+1)th element in the array, and things work nonetheless */
7321 while (low < high) {
7322 mid = (low + high) / 2;
7323 assert(mid <= highest_element);
7324 if (array[mid] <= cp) { /* cp >= array[mid] */
7327 /* We could do this extra test to exit the loop early.
7328 if (cp < array[low]) {
7333 else { /* cp < array[mid] */
7340 invlist_set_previous_index(invlist, high);
7345 Perl__invlist_populate_swatch(pTHX_ SV* const invlist, const UV start, const UV end, U8* swatch)
7347 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
7348 * but is used when the swash has an inversion list. This makes this much
7349 * faster, as it uses a binary search instead of a linear one. This is
7350 * intimately tied to that function, and perhaps should be in utf8.c,
7351 * except it is intimately tied to inversion lists as well. It assumes
7352 * that <swatch> is all 0's on input */
7355 const IV len = _invlist_len(invlist);
7359 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
7361 if (len == 0) { /* Empty inversion list */
7365 array = invlist_array(invlist);
7367 /* Find which element it is */
7368 i = _invlist_search(invlist, start);
7370 /* We populate from <start> to <end> */
7371 while (current < end) {
7374 /* The inversion list gives the results for every possible code point
7375 * after the first one in the list. Only those ranges whose index is
7376 * even are ones that the inversion list matches. For the odd ones,
7377 * and if the initial code point is not in the list, we have to skip
7378 * forward to the next element */
7379 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
7381 if (i >= len) { /* Finished if beyond the end of the array */
7385 if (current >= end) { /* Finished if beyond the end of what we
7387 if (LIKELY(end < UV_MAX)) {
7391 /* We get here when the upper bound is the maximum
7392 * representable on the machine, and we are looking for just
7393 * that code point. Have to special case it */
7395 goto join_end_of_list;
7398 assert(current >= start);
7400 /* The current range ends one below the next one, except don't go past
7403 upper = (i < len && array[i] < end) ? array[i] : end;
7405 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
7406 * for each code point in it */
7407 for (; current < upper; current++) {
7408 const STRLEN offset = (STRLEN)(current - start);
7409 swatch[offset >> 3] |= 1 << (offset & 7);
7414 /* Quit if at the end of the list */
7417 /* But first, have to deal with the highest possible code point on
7418 * the platform. The previous code assumes that <end> is one
7419 * beyond where we want to populate, but that is impossible at the
7420 * platform's infinity, so have to handle it specially */
7421 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
7423 const STRLEN offset = (STRLEN)(end - start);
7424 swatch[offset >> 3] |= 1 << (offset & 7);
7429 /* Advance to the next range, which will be for code points not in the
7438 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** output)
7440 /* Take the union of two inversion lists and point <output> to it. *output
7441 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7442 * the reference count to that list will be decremented. The first list,
7443 * <a>, may be NULL, in which case a copy of the second list is returned.
7444 * If <complement_b> is TRUE, the union is taken of the complement
7445 * (inversion) of <b> instead of b itself.
7447 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7448 * Richard Gillam, published by Addison-Wesley, and explained at some
7449 * length there. The preface says to incorporate its examples into your
7450 * code at your own risk.
7452 * The algorithm is like a merge sort.
7454 * XXX A potential performance improvement is to keep track as we go along
7455 * if only one of the inputs contributes to the result, meaning the other
7456 * is a subset of that one. In that case, we can skip the final copy and
7457 * return the larger of the input lists, but then outside code might need
7458 * to keep track of whether to free the input list or not */
7460 UV* array_a; /* a's array */
7462 UV len_a; /* length of a's array */
7465 SV* u; /* the resulting union */
7469 UV i_a = 0; /* current index into a's array */
7473 /* running count, as explained in the algorithm source book; items are
7474 * stopped accumulating and are output when the count changes to/from 0.
7475 * The count is incremented when we start a range that's in the set, and
7476 * decremented when we start a range that's not in the set. So its range
7477 * is 0 to 2. Only when the count is zero is something not in the set.
7481 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
7484 /* If either one is empty, the union is the other one */
7485 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
7492 *output = invlist_clone(b);
7494 _invlist_invert(*output);
7496 } /* else *output already = b; */
7499 else if ((len_b = _invlist_len(b)) == 0) {
7504 /* The complement of an empty list is a list that has everything in it,
7505 * so the union with <a> includes everything too */
7510 *output = _new_invlist(1);
7511 _append_range_to_invlist(*output, 0, UV_MAX);
7513 else if (*output != a) {
7514 *output = invlist_clone(a);
7516 /* else *output already = a; */
7520 /* Here both lists exist and are non-empty */
7521 array_a = invlist_array(a);
7522 array_b = invlist_array(b);
7524 /* If are to take the union of 'a' with the complement of b, set it
7525 * up so are looking at b's complement. */
7528 /* To complement, we invert: if the first element is 0, remove it. To
7529 * do this, we just pretend the array starts one later, and clear the
7530 * flag as we don't have to do anything else later */
7531 if (array_b[0] == 0) {
7534 complement_b = FALSE;
7538 /* But if the first element is not zero, we unshift a 0 before the
7539 * array. The data structure reserves a space for that 0 (which
7540 * should be a '1' right now), so physical shifting is unneeded,
7541 * but temporarily change that element to 0. Before exiting the
7542 * routine, we must restore the element to '1' */
7549 /* Size the union for the worst case: that the sets are completely
7551 u = _new_invlist(len_a + len_b);
7553 /* Will contain U+0000 if either component does */
7554 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
7555 || (len_b > 0 && array_b[0] == 0));
7557 /* Go through each list item by item, stopping when exhausted one of
7559 while (i_a < len_a && i_b < len_b) {
7560 UV cp; /* The element to potentially add to the union's array */
7561 bool cp_in_set; /* is it in the the input list's set or not */
7563 /* We need to take one or the other of the two inputs for the union.
7564 * Since we are merging two sorted lists, we take the smaller of the
7565 * next items. In case of a tie, we take the one that is in its set
7566 * first. If we took one not in the set first, it would decrement the
7567 * count, possibly to 0 which would cause it to be output as ending the
7568 * range, and the next time through we would take the same number, and
7569 * output it again as beginning the next range. By doing it the
7570 * opposite way, there is no possibility that the count will be
7571 * momentarily decremented to 0, and thus the two adjoining ranges will
7572 * be seamlessly merged. (In a tie and both are in the set or both not
7573 * in the set, it doesn't matter which we take first.) */
7574 if (array_a[i_a] < array_b[i_b]
7575 || (array_a[i_a] == array_b[i_b]
7576 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7578 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7582 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7583 cp = array_b[i_b++];
7586 /* Here, have chosen which of the two inputs to look at. Only output
7587 * if the running count changes to/from 0, which marks the
7588 * beginning/end of a range in that's in the set */
7591 array_u[i_u++] = cp;
7598 array_u[i_u++] = cp;
7603 /* Here, we are finished going through at least one of the lists, which
7604 * means there is something remaining in at most one. We check if the list
7605 * that hasn't been exhausted is positioned such that we are in the middle
7606 * of a range in its set or not. (i_a and i_b point to the element beyond
7607 * the one we care about.) If in the set, we decrement 'count'; if 0, there
7608 * is potentially more to output.
7609 * There are four cases:
7610 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
7611 * in the union is entirely from the non-exhausted set.
7612 * 2) Both were in their sets, count is 2. Nothing further should
7613 * be output, as everything that remains will be in the exhausted
7614 * list's set, hence in the union; decrementing to 1 but not 0 insures
7616 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
7617 * Nothing further should be output because the union includes
7618 * everything from the exhausted set. Not decrementing ensures that.
7619 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
7620 * decrementing to 0 insures that we look at the remainder of the
7621 * non-exhausted set */
7622 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7623 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7628 /* The final length is what we've output so far, plus what else is about to
7629 * be output. (If 'count' is non-zero, then the input list we exhausted
7630 * has everything remaining up to the machine's limit in its set, and hence
7631 * in the union, so there will be no further output. */
7634 /* At most one of the subexpressions will be non-zero */
7635 len_u += (len_a - i_a) + (len_b - i_b);
7638 /* Set result to final length, which can change the pointer to array_u, so
7640 if (len_u != _invlist_len(u)) {
7641 invlist_set_len(u, len_u);
7643 array_u = invlist_array(u);
7646 /* When 'count' is 0, the list that was exhausted (if one was shorter than
7647 * the other) ended with everything above it not in its set. That means
7648 * that the remaining part of the union is precisely the same as the
7649 * non-exhausted list, so can just copy it unchanged. (If both list were
7650 * exhausted at the same time, then the operations below will be both 0.)
7653 IV copy_count; /* At most one will have a non-zero copy count */
7654 if ((copy_count = len_a - i_a) > 0) {
7655 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
7657 else if ((copy_count = len_b - i_b) > 0) {
7658 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
7662 /* If we've changed b, restore it */
7667 /* We may be removing a reference to one of the inputs */
7668 if (a == *output || b == *output) {
7669 assert(! invlist_is_iterating(*output));
7670 SvREFCNT_dec_NN(*output);
7678 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** i)
7680 /* Take the intersection of two inversion lists and point <i> to it. *i
7681 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7682 * the reference count to that list will be decremented.
7683 * If <complement_b> is TRUE, the result will be the intersection of <a>
7684 * and the complement (or inversion) of <b> instead of <b> directly.
7686 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7687 * Richard Gillam, published by Addison-Wesley, and explained at some
7688 * length there. The preface says to incorporate its examples into your
7689 * code at your own risk. In fact, it had bugs
7691 * The algorithm is like a merge sort, and is essentially the same as the
7695 UV* array_a; /* a's array */
7697 UV len_a; /* length of a's array */
7700 SV* r; /* the resulting intersection */
7704 UV i_a = 0; /* current index into a's array */
7708 /* running count, as explained in the algorithm source book; items are
7709 * stopped accumulating and are output when the count changes to/from 2.
7710 * The count is incremented when we start a range that's in the set, and
7711 * decremented when we start a range that's not in the set. So its range
7712 * is 0 to 2. Only when the count is 2 is something in the intersection.
7716 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
7719 /* Special case if either one is empty */
7720 len_a = _invlist_len(a);
7721 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
7723 if (len_a != 0 && complement_b) {
7725 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
7726 * be empty. Here, also we are using 'b's complement, which hence
7727 * must be every possible code point. Thus the intersection is
7730 *i = invlist_clone(a);
7736 /* else *i is already 'a' */
7740 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
7741 * intersection must be empty */
7748 *i = _new_invlist(0);
7752 /* Here both lists exist and are non-empty */
7753 array_a = invlist_array(a);
7754 array_b = invlist_array(b);
7756 /* If are to take the intersection of 'a' with the complement of b, set it
7757 * up so are looking at b's complement. */
7760 /* To complement, we invert: if the first element is 0, remove it. To
7761 * do this, we just pretend the array starts one later, and clear the
7762 * flag as we don't have to do anything else later */
7763 if (array_b[0] == 0) {
7766 complement_b = FALSE;
7770 /* But if the first element is not zero, we unshift a 0 before the
7771 * array. The data structure reserves a space for that 0 (which
7772 * should be a '1' right now), so physical shifting is unneeded,
7773 * but temporarily change that element to 0. Before exiting the
7774 * routine, we must restore the element to '1' */
7781 /* Size the intersection for the worst case: that the intersection ends up
7782 * fragmenting everything to be completely disjoint */
7783 r= _new_invlist(len_a + len_b);
7785 /* Will contain U+0000 iff both components do */
7786 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
7787 && len_b > 0 && array_b[0] == 0);
7789 /* Go through each list item by item, stopping when exhausted one of
7791 while (i_a < len_a && i_b < len_b) {
7792 UV cp; /* The element to potentially add to the intersection's
7794 bool cp_in_set; /* Is it in the input list's set or not */
7796 /* We need to take one or the other of the two inputs for the
7797 * intersection. Since we are merging two sorted lists, we take the
7798 * smaller of the next items. In case of a tie, we take the one that
7799 * is not in its set first (a difference from the union algorithm). If
7800 * we took one in the set first, it would increment the count, possibly
7801 * to 2 which would cause it to be output as starting a range in the
7802 * intersection, and the next time through we would take that same
7803 * number, and output it again as ending the set. By doing it the
7804 * opposite of this, there is no possibility that the count will be
7805 * momentarily incremented to 2. (In a tie and both are in the set or
7806 * both not in the set, it doesn't matter which we take first.) */
7807 if (array_a[i_a] < array_b[i_b]
7808 || (array_a[i_a] == array_b[i_b]
7809 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7811 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7815 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7819 /* Here, have chosen which of the two inputs to look at. Only output
7820 * if the running count changes to/from 2, which marks the
7821 * beginning/end of a range that's in the intersection */
7825 array_r[i_r++] = cp;
7830 array_r[i_r++] = cp;
7836 /* Here, we are finished going through at least one of the lists, which
7837 * means there is something remaining in at most one. We check if the list
7838 * that has been exhausted is positioned such that we are in the middle
7839 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
7840 * the ones we care about.) There are four cases:
7841 * 1) Both weren't in their sets, count is 0, and remains 0. There's
7842 * nothing left in the intersection.
7843 * 2) Both were in their sets, count is 2 and perhaps is incremented to
7844 * above 2. What should be output is exactly that which is in the
7845 * non-exhausted set, as everything it has is also in the intersection
7846 * set, and everything it doesn't have can't be in the intersection
7847 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
7848 * gets incremented to 2. Like the previous case, the intersection is
7849 * everything that remains in the non-exhausted set.
7850 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
7851 * remains 1. And the intersection has nothing more. */
7852 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7853 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7858 /* The final length is what we've output so far plus what else is in the
7859 * intersection. At most one of the subexpressions below will be non-zero */
7862 len_r += (len_a - i_a) + (len_b - i_b);
7865 /* Set result to final length, which can change the pointer to array_r, so
7867 if (len_r != _invlist_len(r)) {
7868 invlist_set_len(r, len_r);
7870 array_r = invlist_array(r);
7873 /* Finish outputting any remaining */
7874 if (count >= 2) { /* At most one will have a non-zero copy count */
7876 if ((copy_count = len_a - i_a) > 0) {
7877 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
7879 else if ((copy_count = len_b - i_b) > 0) {
7880 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
7884 /* If we've changed b, restore it */
7889 /* We may be removing a reference to one of the inputs */
7890 if (a == *i || b == *i) {
7891 assert(! invlist_is_iterating(*i));
7892 SvREFCNT_dec_NN(*i);
7900 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
7902 /* Add the range from 'start' to 'end' inclusive to the inversion list's
7903 * set. A pointer to the inversion list is returned. This may actually be
7904 * a new list, in which case the passed in one has been destroyed. The
7905 * passed in inversion list can be NULL, in which case a new one is created
7906 * with just the one range in it */
7911 if (invlist == NULL) {
7912 invlist = _new_invlist(2);
7916 len = _invlist_len(invlist);
7919 /* If comes after the final entry actually in the list, can just append it
7922 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
7923 && start >= invlist_array(invlist)[len - 1]))
7925 _append_range_to_invlist(invlist, start, end);
7929 /* Here, can't just append things, create and return a new inversion list
7930 * which is the union of this range and the existing inversion list */
7931 range_invlist = _new_invlist(2);
7932 _append_range_to_invlist(range_invlist, start, end);
7934 _invlist_union(invlist, range_invlist, &invlist);
7936 /* The temporary can be freed */
7937 SvREFCNT_dec_NN(range_invlist);
7944 PERL_STATIC_INLINE SV*
7945 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
7946 return _add_range_to_invlist(invlist, cp, cp);
7949 #ifndef PERL_IN_XSUB_RE
7951 Perl__invlist_invert(pTHX_ SV* const invlist)
7953 /* Complement the input inversion list. This adds a 0 if the list didn't
7954 * have a zero; removes it otherwise. As described above, the data
7955 * structure is set up so that this is very efficient */
7957 UV* len_pos = _get_invlist_len_addr(invlist);
7959 PERL_ARGS_ASSERT__INVLIST_INVERT;
7961 assert(! invlist_is_iterating(invlist));
7963 /* The inverse of matching nothing is matching everything */
7964 if (*len_pos == 0) {
7965 _append_range_to_invlist(invlist, 0, UV_MAX);
7969 /* The exclusive or complents 0 to 1; and 1 to 0. If the result is 1, the
7970 * zero element was a 0, so it is being removed, so the length decrements
7971 * by 1; and vice-versa. SvCUR is unaffected */
7972 if (*get_invlist_zero_addr(invlist) ^= 1) {
7981 Perl__invlist_invert_prop(pTHX_ SV* const invlist)
7983 /* Complement the input inversion list (which must be a Unicode property,
7984 * all of which don't match above the Unicode maximum code point.) And
7985 * Perl has chosen to not have the inversion match above that either. This
7986 * adds a 0x110000 if the list didn't end with it, and removes it if it did
7992 PERL_ARGS_ASSERT__INVLIST_INVERT_PROP;
7994 _invlist_invert(invlist);
7996 len = _invlist_len(invlist);
7998 if (len != 0) { /* If empty do nothing */
7999 array = invlist_array(invlist);
8000 if (array[len - 1] != PERL_UNICODE_MAX + 1) {
8001 /* Add 0x110000. First, grow if necessary */
8003 if (invlist_max(invlist) < len) {
8004 invlist_extend(invlist, len);
8005 array = invlist_array(invlist);
8007 invlist_set_len(invlist, len);
8008 array[len - 1] = PERL_UNICODE_MAX + 1;
8010 else { /* Remove the 0x110000 */
8011 invlist_set_len(invlist, len - 1);
8019 PERL_STATIC_INLINE SV*
8020 S_invlist_clone(pTHX_ SV* const invlist)
8023 /* Return a new inversion list that is a copy of the input one, which is
8026 /* Need to allocate extra space to accommodate Perl's addition of a
8027 * trailing NUL to SvPV's, since it thinks they are always strings */
8028 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8029 STRLEN length = SvCUR(invlist);
8031 PERL_ARGS_ASSERT_INVLIST_CLONE;
8033 SvCUR_set(new_invlist, length); /* This isn't done automatically */
8034 Copy(SvPVX(invlist), SvPVX(new_invlist), length, char);
8039 PERL_STATIC_INLINE UV*
8040 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8042 /* Return the address of the UV that contains the current iteration
8045 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8047 return (UV *) (SvPVX(invlist) + (INVLIST_ITER_OFFSET * sizeof (UV)));
8050 PERL_STATIC_INLINE UV*
8051 S_get_invlist_version_id_addr(pTHX_ SV* invlist)
8053 /* Return the address of the UV that contains the version id. */
8055 PERL_ARGS_ASSERT_GET_INVLIST_VERSION_ID_ADDR;
8057 return (UV *) (SvPVX(invlist) + (INVLIST_VERSION_ID_OFFSET * sizeof (UV)));
8060 PERL_STATIC_INLINE void
8061 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8063 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8065 *get_invlist_iter_addr(invlist) = 0;
8068 PERL_STATIC_INLINE void
8069 S_invlist_iterfinish(pTHX_ SV* invlist)
8071 /* Terminate iterator for invlist. This is to catch development errors.
8072 * Any iteration that is interrupted before completed should call this
8073 * function. Functions that add code points anywhere else but to the end
8074 * of an inversion list assert that they are not in the middle of an
8075 * iteration. If they were, the addition would make the iteration
8076 * problematical: if the iteration hadn't reached the place where things
8077 * were being added, it would be ok */
8079 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8081 *get_invlist_iter_addr(invlist) = UV_MAX;
8085 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8087 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8088 * This call sets in <*start> and <*end>, the next range in <invlist>.
8089 * Returns <TRUE> if successful and the next call will return the next
8090 * range; <FALSE> if was already at the end of the list. If the latter,
8091 * <*start> and <*end> are unchanged, and the next call to this function
8092 * will start over at the beginning of the list */
8094 UV* pos = get_invlist_iter_addr(invlist);
8095 UV len = _invlist_len(invlist);
8098 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8101 *pos = UV_MAX; /* Force iterinit() to be required next time */
8105 array = invlist_array(invlist);
8107 *start = array[(*pos)++];
8113 *end = array[(*pos)++] - 1;
8119 PERL_STATIC_INLINE bool
8120 S_invlist_is_iterating(pTHX_ SV* const invlist)
8122 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8124 return *(get_invlist_iter_addr(invlist)) < UV_MAX;
8127 PERL_STATIC_INLINE UV
8128 S_invlist_highest(pTHX_ SV* const invlist)
8130 /* Returns the highest code point that matches an inversion list. This API
8131 * has an ambiguity, as it returns 0 under either the highest is actually
8132 * 0, or if the list is empty. If this distinction matters to you, check
8133 * for emptiness before calling this function */
8135 UV len = _invlist_len(invlist);
8138 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8144 array = invlist_array(invlist);
8146 /* The last element in the array in the inversion list always starts a
8147 * range that goes to infinity. That range may be for code points that are
8148 * matched in the inversion list, or it may be for ones that aren't
8149 * matched. In the latter case, the highest code point in the set is one
8150 * less than the beginning of this range; otherwise it is the final element
8151 * of this range: infinity */
8152 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8154 : array[len - 1] - 1;
8157 #ifndef PERL_IN_XSUB_RE
8159 Perl__invlist_contents(pTHX_ SV* const invlist)
8161 /* Get the contents of an inversion list into a string SV so that they can
8162 * be printed out. It uses the format traditionally done for debug tracing
8166 SV* output = newSVpvs("\n");
8168 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8170 assert(! invlist_is_iterating(invlist));
8172 invlist_iterinit(invlist);
8173 while (invlist_iternext(invlist, &start, &end)) {
8174 if (end == UV_MAX) {
8175 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8177 else if (end != start) {
8178 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8182 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
8190 #ifdef PERL_ARGS_ASSERT__INVLIST_DUMP
8192 Perl__invlist_dump(pTHX_ SV* const invlist, const char * const header)
8194 /* Dumps out the ranges in an inversion list. The string 'header'
8195 * if present is output on a line before the first range */
8199 PERL_ARGS_ASSERT__INVLIST_DUMP;
8201 if (header && strlen(header)) {
8202 PerlIO_printf(Perl_debug_log, "%s\n", header);
8204 if (invlist_is_iterating(invlist)) {
8205 PerlIO_printf(Perl_debug_log, "Can't dump because is in middle of iterating\n");
8209 invlist_iterinit(invlist);
8210 while (invlist_iternext(invlist, &start, &end)) {
8211 if (end == UV_MAX) {
8212 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. INFINITY\n", start);
8214 else if (end != start) {
8215 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. 0x%04"UVXf"\n",
8219 PerlIO_printf(Perl_debug_log, "0x%04"UVXf"\n", start);
8227 S__invlistEQ(pTHX_ SV* const a, SV* const b, bool complement_b)
8229 /* Return a boolean as to if the two passed in inversion lists are
8230 * identical. The final argument, if TRUE, says to take the complement of
8231 * the second inversion list before doing the comparison */
8233 UV* array_a = invlist_array(a);
8234 UV* array_b = invlist_array(b);
8235 UV len_a = _invlist_len(a);
8236 UV len_b = _invlist_len(b);
8238 UV i = 0; /* current index into the arrays */
8239 bool retval = TRUE; /* Assume are identical until proven otherwise */
8241 PERL_ARGS_ASSERT__INVLISTEQ;
8243 /* If are to compare 'a' with the complement of b, set it
8244 * up so are looking at b's complement. */
8247 /* The complement of nothing is everything, so <a> would have to have
8248 * just one element, starting at zero (ending at infinity) */
8250 return (len_a == 1 && array_a[0] == 0);
8252 else if (array_b[0] == 0) {
8254 /* Otherwise, to complement, we invert. Here, the first element is
8255 * 0, just remove it. To do this, we just pretend the array starts
8256 * one later, and clear the flag as we don't have to do anything
8261 complement_b = FALSE;
8265 /* But if the first element is not zero, we unshift a 0 before the
8266 * array. The data structure reserves a space for that 0 (which
8267 * should be a '1' right now), so physical shifting is unneeded,
8268 * but temporarily change that element to 0. Before exiting the
8269 * routine, we must restore the element to '1' */
8276 /* Make sure that the lengths are the same, as well as the final element
8277 * before looping through the remainder. (Thus we test the length, final,
8278 * and first elements right off the bat) */
8279 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
8282 else for (i = 0; i < len_a - 1; i++) {
8283 if (array_a[i] != array_b[i]) {
8296 #undef HEADER_LENGTH
8297 #undef INVLIST_INITIAL_LENGTH
8298 #undef TO_INTERNAL_SIZE
8299 #undef FROM_INTERNAL_SIZE
8300 #undef INVLIST_LEN_OFFSET
8301 #undef INVLIST_ZERO_OFFSET
8302 #undef INVLIST_ITER_OFFSET
8303 #undef INVLIST_VERSION_ID
8304 #undef INVLIST_PREVIOUS_INDEX_OFFSET
8306 /* End of inversion list object */
8309 S_parse_lparen_question_flags(pTHX_ struct RExC_state_t *pRExC_state)
8311 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
8312 * constructs, and updates RExC_flags with them. On input, RExC_parse
8313 * should point to the first flag; it is updated on output to point to the
8314 * final ')' or ':'. There needs to be at least one flag, or this will
8317 /* for (?g), (?gc), and (?o) warnings; warning
8318 about (?c) will warn about (?g) -- japhy */
8320 #define WASTED_O 0x01
8321 #define WASTED_G 0x02
8322 #define WASTED_C 0x04
8323 #define WASTED_GC (0x02|0x04)
8324 I32 wastedflags = 0x00;
8325 U32 posflags = 0, negflags = 0;
8326 U32 *flagsp = &posflags;
8327 char has_charset_modifier = '\0';
8329 bool has_use_defaults = FALSE;
8330 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
8332 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
8334 /* '^' as an initial flag sets certain defaults */
8335 if (UCHARAT(RExC_parse) == '^') {
8337 has_use_defaults = TRUE;
8338 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
8339 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
8340 ? REGEX_UNICODE_CHARSET
8341 : REGEX_DEPENDS_CHARSET);
8344 cs = get_regex_charset(RExC_flags);
8345 if (cs == REGEX_DEPENDS_CHARSET
8346 && (RExC_utf8 || RExC_uni_semantics))
8348 cs = REGEX_UNICODE_CHARSET;
8351 while (*RExC_parse) {
8352 /* && strchr("iogcmsx", *RExC_parse) */
8353 /* (?g), (?gc) and (?o) are useless here
8354 and must be globally applied -- japhy */
8355 switch (*RExC_parse) {
8357 /* Code for the imsx flags */
8358 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
8360 case LOCALE_PAT_MOD:
8361 if (has_charset_modifier) {
8362 goto excess_modifier;
8364 else if (flagsp == &negflags) {
8367 cs = REGEX_LOCALE_CHARSET;
8368 has_charset_modifier = LOCALE_PAT_MOD;
8369 RExC_contains_locale = 1;
8371 case UNICODE_PAT_MOD:
8372 if (has_charset_modifier) {
8373 goto excess_modifier;
8375 else if (flagsp == &negflags) {
8378 cs = REGEX_UNICODE_CHARSET;
8379 has_charset_modifier = UNICODE_PAT_MOD;
8381 case ASCII_RESTRICT_PAT_MOD:
8382 if (flagsp == &negflags) {
8385 if (has_charset_modifier) {
8386 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
8387 goto excess_modifier;
8389 /* Doubled modifier implies more restricted */
8390 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
8393 cs = REGEX_ASCII_RESTRICTED_CHARSET;
8395 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
8397 case DEPENDS_PAT_MOD:
8398 if (has_use_defaults) {
8399 goto fail_modifiers;
8401 else if (flagsp == &negflags) {
8404 else if (has_charset_modifier) {
8405 goto excess_modifier;
8408 /* The dual charset means unicode semantics if the
8409 * pattern (or target, not known until runtime) are
8410 * utf8, or something in the pattern indicates unicode
8412 cs = (RExC_utf8 || RExC_uni_semantics)
8413 ? REGEX_UNICODE_CHARSET
8414 : REGEX_DEPENDS_CHARSET;
8415 has_charset_modifier = DEPENDS_PAT_MOD;
8419 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
8420 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
8422 else if (has_charset_modifier == *(RExC_parse - 1)) {
8423 vFAIL2("Regexp modifier \"%c\" may not appear twice", *(RExC_parse - 1));
8426 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
8431 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"", *(RExC_parse - 1));
8433 case ONCE_PAT_MOD: /* 'o' */
8434 case GLOBAL_PAT_MOD: /* 'g' */
8435 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8436 const I32 wflagbit = *RExC_parse == 'o' ? WASTED_O : WASTED_G;
8437 if (! (wastedflags & wflagbit) ) {
8438 wastedflags |= wflagbit;
8441 "Useless (%s%c) - %suse /%c modifier",
8442 flagsp == &negflags ? "?-" : "?",
8444 flagsp == &negflags ? "don't " : "",
8451 case CONTINUE_PAT_MOD: /* 'c' */
8452 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8453 if (! (wastedflags & WASTED_C) ) {
8454 wastedflags |= WASTED_GC;
8457 "Useless (%sc) - %suse /gc modifier",
8458 flagsp == &negflags ? "?-" : "?",
8459 flagsp == &negflags ? "don't " : ""
8464 case KEEPCOPY_PAT_MOD: /* 'p' */
8465 if (flagsp == &negflags) {
8467 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
8469 *flagsp |= RXf_PMf_KEEPCOPY;
8473 /* A flag is a default iff it is following a minus, so
8474 * if there is a minus, it means will be trying to
8475 * re-specify a default which is an error */
8476 if (has_use_defaults || flagsp == &negflags) {
8477 goto fail_modifiers;
8480 wastedflags = 0; /* reset so (?g-c) warns twice */
8484 RExC_flags |= posflags;
8485 RExC_flags &= ~negflags;
8486 set_regex_charset(&RExC_flags, cs);
8492 vFAIL3("Sequence (%.*s...) not recognized",
8493 RExC_parse-seqstart, seqstart);
8502 - reg - regular expression, i.e. main body or parenthesized thing
8504 * Caller must absorb opening parenthesis.
8506 * Combining parenthesis handling with the base level of regular expression
8507 * is a trifle forced, but the need to tie the tails of the branches to what
8508 * follows makes it hard to avoid.
8510 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
8512 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
8514 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
8517 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
8518 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
8519 needs to be restarted.
8520 Otherwise would only return NULL if regbranch() returns NULL, which
8523 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
8524 /* paren: Parenthesized? 0=top, 1=(, inside: changed to letter. */
8527 regnode *ret; /* Will be the head of the group. */
8530 regnode *ender = NULL;
8533 U32 oregflags = RExC_flags;
8534 bool have_branch = 0;
8536 I32 freeze_paren = 0;
8537 I32 after_freeze = 0;
8539 char * parse_start = RExC_parse; /* MJD */
8540 char * const oregcomp_parse = RExC_parse;
8542 GET_RE_DEBUG_FLAGS_DECL;
8544 PERL_ARGS_ASSERT_REG;
8545 DEBUG_PARSE("reg ");
8547 *flagp = 0; /* Tentatively. */
8550 /* Make an OPEN node, if parenthesized. */
8552 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
8553 char *start_verb = RExC_parse;
8554 STRLEN verb_len = 0;
8555 char *start_arg = NULL;
8556 unsigned char op = 0;
8558 int internal_argval = 0; /* internal_argval is only useful if !argok */
8559 while ( *RExC_parse && *RExC_parse != ')' ) {
8560 if ( *RExC_parse == ':' ) {
8561 start_arg = RExC_parse + 1;
8567 verb_len = RExC_parse - start_verb;
8570 while ( *RExC_parse && *RExC_parse != ')' )
8572 if ( *RExC_parse != ')' )
8573 vFAIL("Unterminated verb pattern argument");
8574 if ( RExC_parse == start_arg )
8577 if ( *RExC_parse != ')' )
8578 vFAIL("Unterminated verb pattern");
8581 switch ( *start_verb ) {
8582 case 'A': /* (*ACCEPT) */
8583 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
8585 internal_argval = RExC_nestroot;
8588 case 'C': /* (*COMMIT) */
8589 if ( memEQs(start_verb,verb_len,"COMMIT") )
8592 case 'F': /* (*FAIL) */
8593 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
8598 case ':': /* (*:NAME) */
8599 case 'M': /* (*MARK:NAME) */
8600 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
8605 case 'P': /* (*PRUNE) */
8606 if ( memEQs(start_verb,verb_len,"PRUNE") )
8609 case 'S': /* (*SKIP) */
8610 if ( memEQs(start_verb,verb_len,"SKIP") )
8613 case 'T': /* (*THEN) */
8614 /* [19:06] <TimToady> :: is then */
8615 if ( memEQs(start_verb,verb_len,"THEN") ) {
8617 RExC_seen |= REG_SEEN_CUTGROUP;
8623 vFAIL3("Unknown verb pattern '%.*s'",
8624 verb_len, start_verb);
8627 if ( start_arg && internal_argval ) {
8628 vFAIL3("Verb pattern '%.*s' may not have an argument",
8629 verb_len, start_verb);
8630 } else if ( argok < 0 && !start_arg ) {
8631 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
8632 verb_len, start_verb);
8634 ret = reganode(pRExC_state, op, internal_argval);
8635 if ( ! internal_argval && ! SIZE_ONLY ) {
8637 SV *sv = newSVpvn( start_arg, RExC_parse - start_arg);
8638 ARG(ret) = add_data( pRExC_state, 1, "S" );
8639 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
8646 if (!internal_argval)
8647 RExC_seen |= REG_SEEN_VERBARG;
8648 } else if ( start_arg ) {
8649 vFAIL3("Verb pattern '%.*s' may not have an argument",
8650 verb_len, start_verb);
8652 ret = reg_node(pRExC_state, op);
8654 nextchar(pRExC_state);
8657 if (*RExC_parse == '?') { /* (?...) */
8658 bool is_logical = 0;
8659 const char * const seqstart = RExC_parse;
8662 paren = *RExC_parse++;
8663 ret = NULL; /* For look-ahead/behind. */
8666 case 'P': /* (?P...) variants for those used to PCRE/Python */
8667 paren = *RExC_parse++;
8668 if ( paren == '<') /* (?P<...>) named capture */
8670 else if (paren == '>') { /* (?P>name) named recursion */
8671 goto named_recursion;
8673 else if (paren == '=') { /* (?P=...) named backref */
8674 /* this pretty much dupes the code for \k<NAME> in regatom(), if
8675 you change this make sure you change that */
8676 char* name_start = RExC_parse;
8678 SV *sv_dat = reg_scan_name(pRExC_state,
8679 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8680 if (RExC_parse == name_start || *RExC_parse != ')')
8681 vFAIL2("Sequence %.3s... not terminated",parse_start);
8684 num = add_data( pRExC_state, 1, "S" );
8685 RExC_rxi->data->data[num]=(void*)sv_dat;
8686 SvREFCNT_inc_simple_void(sv_dat);
8689 ret = reganode(pRExC_state,
8692 : (ASCII_FOLD_RESTRICTED)
8694 : (AT_LEAST_UNI_SEMANTICS)
8702 Set_Node_Offset(ret, parse_start+1);
8703 Set_Node_Cur_Length(ret); /* MJD */
8705 nextchar(pRExC_state);
8709 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8711 case '<': /* (?<...) */
8712 if (*RExC_parse == '!')
8714 else if (*RExC_parse != '=')
8720 case '\'': /* (?'...') */
8721 name_start= RExC_parse;
8722 svname = reg_scan_name(pRExC_state,
8723 SIZE_ONLY ? /* reverse test from the others */
8724 REG_RSN_RETURN_NAME :
8725 REG_RSN_RETURN_NULL);
8726 if (RExC_parse == name_start) {
8728 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8731 if (*RExC_parse != paren)
8732 vFAIL2("Sequence (?%c... not terminated",
8733 paren=='>' ? '<' : paren);
8737 if (!svname) /* shouldn't happen */
8739 "panic: reg_scan_name returned NULL");
8740 if (!RExC_paren_names) {
8741 RExC_paren_names= newHV();
8742 sv_2mortal(MUTABLE_SV(RExC_paren_names));
8744 RExC_paren_name_list= newAV();
8745 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
8748 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
8750 sv_dat = HeVAL(he_str);
8752 /* croak baby croak */
8754 "panic: paren_name hash element allocation failed");
8755 } else if ( SvPOK(sv_dat) ) {
8756 /* (?|...) can mean we have dupes so scan to check
8757 its already been stored. Maybe a flag indicating
8758 we are inside such a construct would be useful,
8759 but the arrays are likely to be quite small, so
8760 for now we punt -- dmq */
8761 IV count = SvIV(sv_dat);
8762 I32 *pv = (I32*)SvPVX(sv_dat);
8764 for ( i = 0 ; i < count ; i++ ) {
8765 if ( pv[i] == RExC_npar ) {
8771 pv = (I32*)SvGROW(sv_dat, SvCUR(sv_dat) + sizeof(I32)+1);
8772 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
8773 pv[count] = RExC_npar;
8774 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
8777 (void)SvUPGRADE(sv_dat,SVt_PVNV);
8778 sv_setpvn(sv_dat, (char *)&(RExC_npar), sizeof(I32));
8780 SvIV_set(sv_dat, 1);
8783 /* Yes this does cause a memory leak in debugging Perls */
8784 if (!av_store(RExC_paren_name_list, RExC_npar, SvREFCNT_inc(svname)))
8785 SvREFCNT_dec_NN(svname);
8788 /*sv_dump(sv_dat);*/
8790 nextchar(pRExC_state);
8792 goto capturing_parens;
8794 RExC_seen |= REG_SEEN_LOOKBEHIND;
8795 RExC_in_lookbehind++;
8797 case '=': /* (?=...) */
8798 RExC_seen_zerolen++;
8800 case '!': /* (?!...) */
8801 RExC_seen_zerolen++;
8802 if (*RExC_parse == ')') {
8803 ret=reg_node(pRExC_state, OPFAIL);
8804 nextchar(pRExC_state);
8808 case '|': /* (?|...) */
8809 /* branch reset, behave like a (?:...) except that
8810 buffers in alternations share the same numbers */
8812 after_freeze = freeze_paren = RExC_npar;
8814 case ':': /* (?:...) */
8815 case '>': /* (?>...) */
8817 case '$': /* (?$...) */
8818 case '@': /* (?@...) */
8819 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
8821 case '0' : /* (?0) */
8822 case 'R' : /* (?R) */
8823 if (*RExC_parse != ')')
8824 FAIL("Sequence (?R) not terminated");
8825 ret = reg_node(pRExC_state, GOSTART);
8826 *flagp |= POSTPONED;
8827 nextchar(pRExC_state);
8830 { /* named and numeric backreferences */
8832 case '&': /* (?&NAME) */
8833 parse_start = RExC_parse - 1;
8836 SV *sv_dat = reg_scan_name(pRExC_state,
8837 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8838 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8840 goto gen_recurse_regop;
8841 assert(0); /* NOT REACHED */
8843 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8845 vFAIL("Illegal pattern");
8847 goto parse_recursion;
8849 case '-': /* (?-1) */
8850 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8851 RExC_parse--; /* rewind to let it be handled later */
8855 case '1': case '2': case '3': case '4': /* (?1) */
8856 case '5': case '6': case '7': case '8': case '9':
8859 num = atoi(RExC_parse);
8860 parse_start = RExC_parse - 1; /* MJD */
8861 if (*RExC_parse == '-')
8863 while (isDIGIT(*RExC_parse))
8865 if (*RExC_parse!=')')
8866 vFAIL("Expecting close bracket");
8869 if ( paren == '-' ) {
8871 Diagram of capture buffer numbering.
8872 Top line is the normal capture buffer numbers
8873 Bottom line is the negative indexing as from
8877 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
8881 num = RExC_npar + num;
8884 vFAIL("Reference to nonexistent group");
8886 } else if ( paren == '+' ) {
8887 num = RExC_npar + num - 1;
8890 ret = reganode(pRExC_state, GOSUB, num);
8892 if (num > (I32)RExC_rx->nparens) {
8894 vFAIL("Reference to nonexistent group");
8896 ARG2L_SET( ret, RExC_recurse_count++);
8898 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
8899 "Recurse #%"UVuf" to %"IVdf"\n", (UV)ARG(ret), (IV)ARG2L(ret)));
8903 RExC_seen |= REG_SEEN_RECURSE;
8904 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
8905 Set_Node_Offset(ret, parse_start); /* MJD */
8907 *flagp |= POSTPONED;
8908 nextchar(pRExC_state);
8910 } /* named and numeric backreferences */
8911 assert(0); /* NOT REACHED */
8913 case '?': /* (??...) */
8915 if (*RExC_parse != '{') {
8917 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8920 *flagp |= POSTPONED;
8921 paren = *RExC_parse++;
8923 case '{': /* (?{...}) */
8926 struct reg_code_block *cb;
8928 RExC_seen_zerolen++;
8930 if ( !pRExC_state->num_code_blocks
8931 || pRExC_state->code_index >= pRExC_state->num_code_blocks
8932 || pRExC_state->code_blocks[pRExC_state->code_index].start
8933 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
8936 if (RExC_pm_flags & PMf_USE_RE_EVAL)
8937 FAIL("panic: Sequence (?{...}): no code block found\n");
8938 FAIL("Eval-group not allowed at runtime, use re 'eval'");
8940 /* this is a pre-compiled code block (?{...}) */
8941 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
8942 RExC_parse = RExC_start + cb->end;
8945 if (cb->src_regex) {
8946 n = add_data(pRExC_state, 2, "rl");
8947 RExC_rxi->data->data[n] =
8948 (void*)SvREFCNT_inc((SV*)cb->src_regex);
8949 RExC_rxi->data->data[n+1] = (void*)o;
8952 n = add_data(pRExC_state, 1,
8953 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l");
8954 RExC_rxi->data->data[n] = (void*)o;
8957 pRExC_state->code_index++;
8958 nextchar(pRExC_state);
8962 ret = reg_node(pRExC_state, LOGICAL);
8963 eval = reganode(pRExC_state, EVAL, n);
8966 /* for later propagation into (??{}) return value */
8967 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
8969 REGTAIL(pRExC_state, ret, eval);
8970 /* deal with the length of this later - MJD */
8973 ret = reganode(pRExC_state, EVAL, n);
8974 Set_Node_Length(ret, RExC_parse - parse_start + 1);
8975 Set_Node_Offset(ret, parse_start);
8978 case '(': /* (?(?{...})...) and (?(?=...)...) */
8981 if (RExC_parse[0] == '?') { /* (?(?...)) */
8982 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
8983 || RExC_parse[1] == '<'
8984 || RExC_parse[1] == '{') { /* Lookahead or eval. */
8988 ret = reg_node(pRExC_state, LOGICAL);
8992 tail = reg(pRExC_state, 1, &flag, depth+1);
8993 if (flag & RESTART_UTF8) {
8994 *flagp = RESTART_UTF8;
8997 REGTAIL(pRExC_state, ret, tail);
9001 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9002 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9004 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9005 char *name_start= RExC_parse++;
9007 SV *sv_dat=reg_scan_name(pRExC_state,
9008 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9009 if (RExC_parse == name_start || *RExC_parse != ch)
9010 vFAIL2("Sequence (?(%c... not terminated",
9011 (ch == '>' ? '<' : ch));
9014 num = add_data( pRExC_state, 1, "S" );
9015 RExC_rxi->data->data[num]=(void*)sv_dat;
9016 SvREFCNT_inc_simple_void(sv_dat);
9018 ret = reganode(pRExC_state,NGROUPP,num);
9019 goto insert_if_check_paren;
9021 else if (RExC_parse[0] == 'D' &&
9022 RExC_parse[1] == 'E' &&
9023 RExC_parse[2] == 'F' &&
9024 RExC_parse[3] == 'I' &&
9025 RExC_parse[4] == 'N' &&
9026 RExC_parse[5] == 'E')
9028 ret = reganode(pRExC_state,DEFINEP,0);
9031 goto insert_if_check_paren;
9033 else if (RExC_parse[0] == 'R') {
9036 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9037 parno = atoi(RExC_parse++);
9038 while (isDIGIT(*RExC_parse))
9040 } else if (RExC_parse[0] == '&') {
9043 sv_dat = reg_scan_name(pRExC_state,
9044 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9045 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9047 ret = reganode(pRExC_state,INSUBP,parno);
9048 goto insert_if_check_paren;
9050 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9053 parno = atoi(RExC_parse++);
9055 while (isDIGIT(*RExC_parse))
9057 ret = reganode(pRExC_state, GROUPP, parno);
9059 insert_if_check_paren:
9060 if ((c = *nextchar(pRExC_state)) != ')')
9061 vFAIL("Switch condition not recognized");
9063 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9064 br = regbranch(pRExC_state, &flags, 1,depth+1);
9066 if (flags & RESTART_UTF8) {
9067 *flagp = RESTART_UTF8;
9070 FAIL2("panic: regbranch returned NULL, flags=%#X",
9073 REGTAIL(pRExC_state, br, reganode(pRExC_state, LONGJMP, 0));
9074 c = *nextchar(pRExC_state);
9079 vFAIL("(?(DEFINE)....) does not allow branches");
9080 lastbr = reganode(pRExC_state, IFTHEN, 0); /* Fake one for optimizer. */
9081 if (!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, ret, lastbr);
9092 c = *nextchar(pRExC_state);
9097 vFAIL("Switch (?(condition)... contains too many branches");
9098 ender = reg_node(pRExC_state, TAIL);
9099 REGTAIL(pRExC_state, br, ender);
9101 REGTAIL(pRExC_state, lastbr, ender);
9102 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
9105 REGTAIL(pRExC_state, ret, ender);
9106 RExC_size++; /* XXX WHY do we need this?!!
9107 For large programs it seems to be required
9108 but I can't figure out why. -- dmq*/
9112 vFAIL2("Unknown switch condition (?(%.2s", RExC_parse);
9115 case '[': /* (?[ ... ]) */
9116 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
9119 RExC_parse--; /* for vFAIL to print correctly */
9120 vFAIL("Sequence (? incomplete");
9122 default: /* e.g., (?i) */
9125 parse_lparen_question_flags(pRExC_state);
9126 if (UCHARAT(RExC_parse) != ':') {
9127 nextchar(pRExC_state);
9132 nextchar(pRExC_state);
9142 ret = reganode(pRExC_state, OPEN, parno);
9145 RExC_nestroot = parno;
9146 if (RExC_seen & REG_SEEN_RECURSE
9147 && !RExC_open_parens[parno-1])
9149 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9150 "Setting open paren #%"IVdf" to %d\n",
9151 (IV)parno, REG_NODE_NUM(ret)));
9152 RExC_open_parens[parno-1]= ret;
9155 Set_Node_Length(ret, 1); /* MJD */
9156 Set_Node_Offset(ret, RExC_parse); /* MJD */
9164 /* Pick up the branches, linking them together. */
9165 parse_start = RExC_parse; /* MJD */
9166 br = regbranch(pRExC_state, &flags, 1,depth+1);
9168 /* branch_len = (paren != 0); */
9171 if (flags & RESTART_UTF8) {
9172 *flagp = RESTART_UTF8;
9175 FAIL2("panic: regbranch returned NULL, flags=%#X", flags);
9177 if (*RExC_parse == '|') {
9178 if (!SIZE_ONLY && RExC_extralen) {
9179 reginsert(pRExC_state, BRANCHJ, br, depth+1);
9182 reginsert(pRExC_state, BRANCH, br, depth+1);
9183 Set_Node_Length(br, paren != 0);
9184 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
9188 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
9190 else if (paren == ':') {
9191 *flagp |= flags&SIMPLE;
9193 if (is_open) { /* Starts with OPEN. */
9194 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
9196 else if (paren != '?') /* Not Conditional */
9198 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9200 while (*RExC_parse == '|') {
9201 if (!SIZE_ONLY && RExC_extralen) {
9202 ender = reganode(pRExC_state, LONGJMP,0);
9203 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender); /* Append to the previous. */
9206 RExC_extralen += 2; /* Account for LONGJMP. */
9207 nextchar(pRExC_state);
9209 if (RExC_npar > after_freeze)
9210 after_freeze = RExC_npar;
9211 RExC_npar = freeze_paren;
9213 br = regbranch(pRExC_state, &flags, 0, depth+1);
9216 if (flags & RESTART_UTF8) {
9217 *flagp = RESTART_UTF8;
9220 FAIL2("panic: regbranch returned NULL, flags=%#X", flags);
9222 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
9224 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9227 if (have_branch || paren != ':') {
9228 /* Make a closing node, and hook it on the end. */
9231 ender = reg_node(pRExC_state, TAIL);
9234 ender = reganode(pRExC_state, CLOSE, parno);
9235 if (!SIZE_ONLY && RExC_seen & REG_SEEN_RECURSE) {
9236 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9237 "Setting close paren #%"IVdf" to %d\n",
9238 (IV)parno, REG_NODE_NUM(ender)));
9239 RExC_close_parens[parno-1]= ender;
9240 if (RExC_nestroot == parno)
9243 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
9244 Set_Node_Length(ender,1); /* MJD */
9250 *flagp &= ~HASWIDTH;
9253 ender = reg_node(pRExC_state, SUCCEED);
9256 ender = reg_node(pRExC_state, END);
9258 assert(!RExC_opend); /* there can only be one! */
9263 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9264 SV * const mysv_val1=sv_newmortal();
9265 SV * const mysv_val2=sv_newmortal();
9266 DEBUG_PARSE_MSG("lsbr");
9267 regprop(RExC_rx, mysv_val1, lastbr);
9268 regprop(RExC_rx, mysv_val2, ender);
9269 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9270 SvPV_nolen_const(mysv_val1),
9271 (IV)REG_NODE_NUM(lastbr),
9272 SvPV_nolen_const(mysv_val2),
9273 (IV)REG_NODE_NUM(ender),
9274 (IV)(ender - lastbr)
9277 REGTAIL(pRExC_state, lastbr, ender);
9279 if (have_branch && !SIZE_ONLY) {
9282 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
9284 /* Hook the tails of the branches to the closing node. */
9285 for (br = ret; br; br = regnext(br)) {
9286 const U8 op = PL_regkind[OP(br)];
9288 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
9289 if (OP(NEXTOPER(br)) != NOTHING || regnext(NEXTOPER(br)) != ender)
9292 else if (op == BRANCHJ) {
9293 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
9294 /* for now we always disable this optimisation * /
9295 if (OP(NEXTOPER(NEXTOPER(br))) != NOTHING || regnext(NEXTOPER(NEXTOPER(br))) != ender)
9301 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
9302 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9303 SV * const mysv_val1=sv_newmortal();
9304 SV * const mysv_val2=sv_newmortal();
9305 DEBUG_PARSE_MSG("NADA");
9306 regprop(RExC_rx, mysv_val1, ret);
9307 regprop(RExC_rx, mysv_val2, ender);
9308 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9309 SvPV_nolen_const(mysv_val1),
9310 (IV)REG_NODE_NUM(ret),
9311 SvPV_nolen_const(mysv_val2),
9312 (IV)REG_NODE_NUM(ender),
9317 if (OP(ender) == TAIL) {
9322 for ( opt= br + 1; opt < ender ; opt++ )
9324 NEXT_OFF(br)= ender - br;
9332 static const char parens[] = "=!<,>";
9334 if (paren && (p = strchr(parens, paren))) {
9335 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
9336 int flag = (p - parens) > 1;
9339 node = SUSPEND, flag = 0;
9340 reginsert(pRExC_state, node,ret, depth+1);
9341 Set_Node_Cur_Length(ret);
9342 Set_Node_Offset(ret, parse_start + 1);
9344 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
9348 /* Check for proper termination. */
9350 RExC_flags = oregflags;
9351 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
9352 RExC_parse = oregcomp_parse;
9353 vFAIL("Unmatched (");
9356 else if (!paren && RExC_parse < RExC_end) {
9357 if (*RExC_parse == ')') {
9359 vFAIL("Unmatched )");
9362 FAIL("Junk on end of regexp"); /* "Can't happen". */
9363 assert(0); /* NOTREACHED */
9366 if (RExC_in_lookbehind) {
9367 RExC_in_lookbehind--;
9369 if (after_freeze > RExC_npar)
9370 RExC_npar = after_freeze;
9375 - regbranch - one alternative of an | operator
9377 * Implements the concatenation operator.
9379 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9383 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
9387 regnode *chain = NULL;
9389 I32 flags = 0, c = 0;
9390 GET_RE_DEBUG_FLAGS_DECL;
9392 PERL_ARGS_ASSERT_REGBRANCH;
9394 DEBUG_PARSE("brnc");
9399 if (!SIZE_ONLY && RExC_extralen)
9400 ret = reganode(pRExC_state, BRANCHJ,0);
9402 ret = reg_node(pRExC_state, BRANCH);
9403 Set_Node_Length(ret, 1);
9407 if (!first && SIZE_ONLY)
9408 RExC_extralen += 1; /* BRANCHJ */
9410 *flagp = WORST; /* Tentatively. */
9413 nextchar(pRExC_state);
9414 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
9416 latest = regpiece(pRExC_state, &flags,depth+1);
9417 if (latest == NULL) {
9418 if (flags & TRYAGAIN)
9420 if (flags & RESTART_UTF8) {
9421 *flagp = RESTART_UTF8;
9424 FAIL2("panic: regpiece returned NULL, flags=%#X", flags);
9426 else if (ret == NULL)
9428 *flagp |= flags&(HASWIDTH|POSTPONED);
9429 if (chain == NULL) /* First piece. */
9430 *flagp |= flags&SPSTART;
9433 REGTAIL(pRExC_state, chain, latest);
9438 if (chain == NULL) { /* Loop ran zero times. */
9439 chain = reg_node(pRExC_state, NOTHING);
9444 *flagp |= flags&SIMPLE;
9451 - regpiece - something followed by possible [*+?]
9453 * Note that the branching code sequences used for ? and the general cases
9454 * of * and + are somewhat optimized: they use the same NOTHING node as
9455 * both the endmarker for their branch list and the body of the last branch.
9456 * It might seem that this node could be dispensed with entirely, but the
9457 * endmarker role is not redundant.
9459 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
9461 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9465 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
9472 const char * const origparse = RExC_parse;
9474 I32 max = REG_INFTY;
9475 #ifdef RE_TRACK_PATTERN_OFFSETS
9478 const char *maxpos = NULL;
9480 /* Save the original in case we change the emitted regop to a FAIL. */
9481 regnode * const orig_emit = RExC_emit;
9483 GET_RE_DEBUG_FLAGS_DECL;
9485 PERL_ARGS_ASSERT_REGPIECE;
9487 DEBUG_PARSE("piec");
9489 ret = regatom(pRExC_state, &flags,depth+1);
9491 if (flags & (TRYAGAIN|RESTART_UTF8))
9492 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
9494 FAIL2("panic: regatom returned NULL, flags=%#X", flags);
9500 if (op == '{' && regcurly(RExC_parse, FALSE)) {
9502 #ifdef RE_TRACK_PATTERN_OFFSETS
9503 parse_start = RExC_parse; /* MJD */
9505 next = RExC_parse + 1;
9506 while (isDIGIT(*next) || *next == ',') {
9515 if (*next == '}') { /* got one */
9519 min = atoi(RExC_parse);
9523 maxpos = RExC_parse;
9525 if (!max && *maxpos != '0')
9526 max = REG_INFTY; /* meaning "infinity" */
9527 else if (max >= REG_INFTY)
9528 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
9530 nextchar(pRExC_state);
9531 if (max < min) { /* If can't match, warn and optimize to fail
9534 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
9536 /* We can't back off the size because we have to reserve
9537 * enough space for all the things we are about to throw
9538 * away, but we can shrink it by the ammount we are about
9540 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
9543 RExC_emit = orig_emit;
9545 ret = reg_node(pRExC_state, OPFAIL);
9548 else if (max == 0) { /* replace {0} with a nothing node */
9550 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)NOTHING];
9553 RExC_emit = orig_emit;
9555 ret = reg_node(pRExC_state, NOTHING);
9560 if ((flags&SIMPLE)) {
9561 RExC_naughty += 2 + RExC_naughty / 2;
9562 reginsert(pRExC_state, CURLY, ret, depth+1);
9563 Set_Node_Offset(ret, parse_start+1); /* MJD */
9564 Set_Node_Cur_Length(ret);
9567 regnode * const w = reg_node(pRExC_state, WHILEM);
9570 REGTAIL(pRExC_state, ret, w);
9571 if (!SIZE_ONLY && RExC_extralen) {
9572 reginsert(pRExC_state, LONGJMP,ret, depth+1);
9573 reginsert(pRExC_state, NOTHING,ret, depth+1);
9574 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
9576 reginsert(pRExC_state, CURLYX,ret, depth+1);
9578 Set_Node_Offset(ret, parse_start+1);
9579 Set_Node_Length(ret,
9580 op == '{' ? (RExC_parse - parse_start) : 1);
9582 if (!SIZE_ONLY && RExC_extralen)
9583 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
9584 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
9586 RExC_whilem_seen++, RExC_extralen += 3;
9587 RExC_naughty += 4 + RExC_naughty; /* compound interest */
9596 ARG1_SET(ret, (U16)min);
9597 ARG2_SET(ret, (U16)max);
9609 #if 0 /* Now runtime fix should be reliable. */
9611 /* if this is reinstated, don't forget to put this back into perldiag:
9613 =item Regexp *+ operand could be empty at {#} in regex m/%s/
9615 (F) The part of the regexp subject to either the * or + quantifier
9616 could match an empty string. The {#} shows in the regular
9617 expression about where the problem was discovered.
9621 if (!(flags&HASWIDTH) && op != '?')
9622 vFAIL("Regexp *+ operand could be empty");
9625 #ifdef RE_TRACK_PATTERN_OFFSETS
9626 parse_start = RExC_parse;
9628 nextchar(pRExC_state);
9630 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
9632 if (op == '*' && (flags&SIMPLE)) {
9633 reginsert(pRExC_state, STAR, ret, depth+1);
9637 else if (op == '*') {
9641 else if (op == '+' && (flags&SIMPLE)) {
9642 reginsert(pRExC_state, PLUS, ret, depth+1);
9646 else if (op == '+') {
9650 else if (op == '?') {
9655 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
9656 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
9657 ckWARN3reg(RExC_parse,
9658 "%.*s matches null string many times",
9659 (int)(RExC_parse >= origparse ? RExC_parse - origparse : 0),
9661 (void)ReREFCNT_inc(RExC_rx_sv);
9664 if (RExC_parse < RExC_end && *RExC_parse == '?') {
9665 nextchar(pRExC_state);
9666 reginsert(pRExC_state, MINMOD, ret, depth+1);
9667 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
9669 #ifndef REG_ALLOW_MINMOD_SUSPEND
9672 if (RExC_parse < RExC_end && *RExC_parse == '+') {
9674 nextchar(pRExC_state);
9675 ender = reg_node(pRExC_state, SUCCEED);
9676 REGTAIL(pRExC_state, ret, ender);
9677 reginsert(pRExC_state, SUSPEND, ret, depth+1);
9679 ender = reg_node(pRExC_state, TAIL);
9680 REGTAIL(pRExC_state, ret, ender);
9684 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
9686 vFAIL("Nested quantifiers");
9693 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p, UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
9694 const bool strict /* Apply stricter parsing rules? */
9698 /* This is expected to be called by a parser routine that has recognized '\N'
9699 and needs to handle the rest. RExC_parse is expected to point at the first
9700 char following the N at the time of the call. On successful return,
9701 RExC_parse has been updated to point to just after the sequence identified
9702 by this routine, and <*flagp> has been updated.
9704 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
9707 \N may begin either a named sequence, or if outside a character class, mean
9708 to match a non-newline. For non single-quoted regexes, the tokenizer has
9709 attempted to decide which, and in the case of a named sequence, converted it
9710 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
9711 where c1... are the characters in the sequence. For single-quoted regexes,
9712 the tokenizer passes the \N sequence through unchanged; this code will not
9713 attempt to determine this nor expand those, instead raising a syntax error.
9714 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
9715 or there is no '}', it signals that this \N occurrence means to match a
9718 Only the \N{U+...} form should occur in a character class, for the same
9719 reason that '.' inside a character class means to just match a period: it
9720 just doesn't make sense.
9722 The function raises an error (via vFAIL), and doesn't return for various
9723 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
9724 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
9725 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
9726 only possible if node_p is non-NULL.
9729 If <valuep> is non-null, it means the caller can accept an input sequence
9730 consisting of a just a single code point; <*valuep> is set to that value
9731 if the input is such.
9733 If <node_p> is non-null it signifies that the caller can accept any other
9734 legal sequence (i.e., one that isn't just a single code point). <*node_p>
9736 1) \N means not-a-NL: points to a newly created REG_ANY node;
9737 2) \N{}: points to a new NOTHING node;
9738 3) otherwise: points to a new EXACT node containing the resolved
9740 Note that FALSE is returned for single code point sequences if <valuep> is
9744 char * endbrace; /* '}' following the name */
9746 char *endchar; /* Points to '.' or '}' ending cur char in the input
9748 bool has_multiple_chars; /* true if the input stream contains a sequence of
9749 more than one character */
9751 GET_RE_DEBUG_FLAGS_DECL;
9753 PERL_ARGS_ASSERT_GROK_BSLASH_N;
9757 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
9759 /* The [^\n] meaning of \N ignores spaces and comments under the /x
9760 * modifier. The other meaning does not */
9761 p = (RExC_flags & RXf_PMf_EXTENDED)
9762 ? regwhite( pRExC_state, RExC_parse )
9765 /* Disambiguate between \N meaning a named character versus \N meaning
9766 * [^\n]. The former is assumed when it can't be the latter. */
9767 if (*p != '{' || regcurly(p, FALSE)) {
9770 /* no bare \N in a charclass */
9771 if (in_char_class) {
9772 vFAIL("\\N in a character class must be a named character: \\N{...}");
9776 nextchar(pRExC_state);
9777 *node_p = reg_node(pRExC_state, REG_ANY);
9778 *flagp |= HASWIDTH|SIMPLE;
9781 Set_Node_Length(*node_p, 1); /* MJD */
9785 /* Here, we have decided it should be a named character or sequence */
9787 /* The test above made sure that the next real character is a '{', but
9788 * under the /x modifier, it could be separated by space (or a comment and
9789 * \n) and this is not allowed (for consistency with \x{...} and the
9790 * tokenizer handling of \N{NAME}). */
9791 if (*RExC_parse != '{') {
9792 vFAIL("Missing braces on \\N{}");
9795 RExC_parse++; /* Skip past the '{' */
9797 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
9798 || ! (endbrace == RExC_parse /* nothing between the {} */
9799 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below */
9800 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg) */
9802 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
9803 vFAIL("\\N{NAME} must be resolved by the lexer");
9806 if (endbrace == RExC_parse) { /* empty: \N{} */
9809 *node_p = reg_node(pRExC_state,NOTHING);
9811 else if (in_char_class) {
9812 if (SIZE_ONLY && in_char_class) {
9814 RExC_parse++; /* Position after the "}" */
9815 vFAIL("Zero length \\N{}");
9818 ckWARNreg(RExC_parse,
9819 "Ignoring zero length \\N{} in character class");
9827 nextchar(pRExC_state);
9831 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
9832 RExC_parse += 2; /* Skip past the 'U+' */
9834 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9836 /* Code points are separated by dots. If none, there is only one code
9837 * point, and is terminated by the brace */
9838 has_multiple_chars = (endchar < endbrace);
9840 if (valuep && (! has_multiple_chars || in_char_class)) {
9841 /* We only pay attention to the first char of
9842 multichar strings being returned in char classes. I kinda wonder
9843 if this makes sense as it does change the behaviour
9844 from earlier versions, OTOH that behaviour was broken
9845 as well. XXX Solution is to recharacterize as
9846 [rest-of-class]|multi1|multi2... */
9848 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
9849 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
9850 | PERL_SCAN_DISALLOW_PREFIX
9851 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
9853 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
9855 /* The tokenizer should have guaranteed validity, but it's possible to
9856 * bypass it by using single quoting, so check */
9857 if (length_of_hex == 0
9858 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
9860 RExC_parse += length_of_hex; /* Includes all the valid */
9861 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
9862 ? UTF8SKIP(RExC_parse)
9864 /* Guard against malformed utf8 */
9865 if (RExC_parse >= endchar) {
9866 RExC_parse = endchar;
9868 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9871 if (in_char_class && has_multiple_chars) {
9873 RExC_parse = endbrace;
9874 vFAIL("\\N{} in character class restricted to one character");
9877 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
9881 RExC_parse = endbrace + 1;
9883 else if (! node_p || ! has_multiple_chars) {
9885 /* Here, the input is legal, but not according to the caller's
9886 * options. We fail without advancing the parse, so that the
9887 * caller can try again */
9893 /* What is done here is to convert this to a sub-pattern of the form
9894 * (?:\x{char1}\x{char2}...)
9895 * and then call reg recursively. That way, it retains its atomicness,
9896 * while not having to worry about special handling that some code
9897 * points may have. toke.c has converted the original Unicode values
9898 * to native, so that we can just pass on the hex values unchanged. We
9899 * do have to set a flag to keep recoding from happening in the
9902 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
9904 char *orig_end = RExC_end;
9907 while (RExC_parse < endbrace) {
9909 /* Convert to notation the rest of the code understands */
9910 sv_catpv(substitute_parse, "\\x{");
9911 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
9912 sv_catpv(substitute_parse, "}");
9914 /* Point to the beginning of the next character in the sequence. */
9915 RExC_parse = endchar + 1;
9916 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9918 sv_catpv(substitute_parse, ")");
9920 RExC_parse = SvPV(substitute_parse, len);
9922 /* Don't allow empty number */
9924 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9926 RExC_end = RExC_parse + len;
9928 /* The values are Unicode, and therefore not subject to recoding */
9929 RExC_override_recoding = 1;
9931 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
9932 if (flags & RESTART_UTF8) {
9933 *flagp = RESTART_UTF8;
9936 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#X",
9939 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
9941 RExC_parse = endbrace;
9942 RExC_end = orig_end;
9943 RExC_override_recoding = 0;
9945 nextchar(pRExC_state);
9955 * It returns the code point in utf8 for the value in *encp.
9956 * value: a code value in the source encoding
9957 * encp: a pointer to an Encode object
9959 * If the result from Encode is not a single character,
9960 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
9963 S_reg_recode(pTHX_ const char value, SV **encp)
9966 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
9967 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
9968 const STRLEN newlen = SvCUR(sv);
9969 UV uv = UNICODE_REPLACEMENT;
9971 PERL_ARGS_ASSERT_REG_RECODE;
9975 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
9978 if (!newlen || numlen != newlen) {
9979 uv = UNICODE_REPLACEMENT;
9985 PERL_STATIC_INLINE U8
9986 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
9990 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
9996 op = get_regex_charset(RExC_flags);
9997 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
9998 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
9999 been, so there is no hole */
10002 return op + EXACTF;
10005 PERL_STATIC_INLINE void
10006 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state, regnode *node, I32* flagp, STRLEN len, UV code_point)
10008 /* This knows the details about sizing an EXACTish node, setting flags for
10009 * it (by setting <*flagp>, and potentially populating it with a single
10012 * If <len> (the length in bytes) is non-zero, this function assumes that
10013 * the node has already been populated, and just does the sizing. In this
10014 * case <code_point> should be the final code point that has already been
10015 * placed into the node. This value will be ignored except that under some
10016 * circumstances <*flagp> is set based on it.
10018 * If <len> is zero, the function assumes that the node is to contain only
10019 * the single character given by <code_point> and calculates what <len>
10020 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10021 * additionally will populate the node's STRING with <code_point>, if <len>
10022 * is 0. In both cases <*flagp> is appropriately set
10024 * It knows that under FOLD, UTF characters and the Latin Sharp S must be
10025 * folded (the latter only when the rules indicate it can match 'ss') */
10027 bool len_passed_in = cBOOL(len != 0);
10028 U8 character[UTF8_MAXBYTES_CASE+1];
10030 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10032 if (! len_passed_in) {
10035 to_uni_fold(NATIVE_TO_UNI(code_point), character, &len);
10038 uvchr_to_utf8( character, code_point);
10039 len = UTF8SKIP(character);
10043 || code_point != LATIN_SMALL_LETTER_SHARP_S
10044 || ASCII_FOLD_RESTRICTED
10045 || ! AT_LEAST_UNI_SEMANTICS)
10047 *character = (U8) code_point;
10052 *(character + 1) = 's';
10058 RExC_size += STR_SZ(len);
10061 RExC_emit += STR_SZ(len);
10062 STR_LEN(node) = len;
10063 if (! len_passed_in) {
10064 Copy((char *) character, STRING(node), len, char);
10068 *flagp |= HASWIDTH;
10070 /* A single character node is SIMPLE, except for the special-cased SHARP S
10072 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
10073 && (code_point != LATIN_SMALL_LETTER_SHARP_S
10074 || ! FOLD || ! DEPENDS_SEMANTICS))
10081 - regatom - the lowest level
10083 Try to identify anything special at the start of the pattern. If there
10084 is, then handle it as required. This may involve generating a single regop,
10085 such as for an assertion; or it may involve recursing, such as to
10086 handle a () structure.
10088 If the string doesn't start with something special then we gobble up
10089 as much literal text as we can.
10091 Once we have been able to handle whatever type of thing started the
10092 sequence, we return.
10094 Note: we have to be careful with escapes, as they can be both literal
10095 and special, and in the case of \10 and friends, context determines which.
10097 A summary of the code structure is:
10099 switch (first_byte) {
10100 cases for each special:
10101 handle this special;
10104 switch (2nd byte) {
10105 cases for each unambiguous special:
10106 handle this special;
10108 cases for each ambigous special/literal:
10110 if (special) handle here
10112 default: // unambiguously literal:
10115 default: // is a literal char
10118 create EXACTish node for literal;
10119 while (more input and node isn't full) {
10120 switch (input_byte) {
10121 cases for each special;
10122 make sure parse pointer is set so that the next call to
10123 regatom will see this special first
10124 goto loopdone; // EXACTish node terminated by prev. char
10126 append char to EXACTISH node;
10128 get next input byte;
10132 return the generated node;
10134 Specifically there are two separate switches for handling
10135 escape sequences, with the one for handling literal escapes requiring
10136 a dummy entry for all of the special escapes that are actually handled
10139 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
10141 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10143 Otherwise does not return NULL.
10147 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10150 regnode *ret = NULL;
10152 char *parse_start = RExC_parse;
10156 GET_RE_DEBUG_FLAGS_DECL;
10158 *flagp = WORST; /* Tentatively. */
10160 DEBUG_PARSE("atom");
10162 PERL_ARGS_ASSERT_REGATOM;
10165 switch ((U8)*RExC_parse) {
10167 RExC_seen_zerolen++;
10168 nextchar(pRExC_state);
10169 if (RExC_flags & RXf_PMf_MULTILINE)
10170 ret = reg_node(pRExC_state, MBOL);
10171 else if (RExC_flags & RXf_PMf_SINGLELINE)
10172 ret = reg_node(pRExC_state, SBOL);
10174 ret = reg_node(pRExC_state, BOL);
10175 Set_Node_Length(ret, 1); /* MJD */
10178 nextchar(pRExC_state);
10180 RExC_seen_zerolen++;
10181 if (RExC_flags & RXf_PMf_MULTILINE)
10182 ret = reg_node(pRExC_state, MEOL);
10183 else if (RExC_flags & RXf_PMf_SINGLELINE)
10184 ret = reg_node(pRExC_state, SEOL);
10186 ret = reg_node(pRExC_state, EOL);
10187 Set_Node_Length(ret, 1); /* MJD */
10190 nextchar(pRExC_state);
10191 if (RExC_flags & RXf_PMf_SINGLELINE)
10192 ret = reg_node(pRExC_state, SANY);
10194 ret = reg_node(pRExC_state, REG_ANY);
10195 *flagp |= HASWIDTH|SIMPLE;
10197 Set_Node_Length(ret, 1); /* MJD */
10201 char * const oregcomp_parse = ++RExC_parse;
10202 ret = regclass(pRExC_state, flagp,depth+1,
10203 FALSE, /* means parse the whole char class */
10204 TRUE, /* allow multi-char folds */
10205 FALSE, /* don't silence non-portable warnings. */
10207 if (*RExC_parse != ']') {
10208 RExC_parse = oregcomp_parse;
10209 vFAIL("Unmatched [");
10212 if (*flagp & RESTART_UTF8)
10214 FAIL2("panic: regclass returned NULL to regatom, flags=%#X",
10217 nextchar(pRExC_state);
10218 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
10222 nextchar(pRExC_state);
10223 ret = reg(pRExC_state, 1, &flags,depth+1);
10225 if (flags & TRYAGAIN) {
10226 if (RExC_parse == RExC_end) {
10227 /* Make parent create an empty node if needed. */
10228 *flagp |= TRYAGAIN;
10233 if (flags & RESTART_UTF8) {
10234 *flagp = RESTART_UTF8;
10237 FAIL2("panic: reg returned NULL to regatom, flags=%#X", flags);
10239 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10243 if (flags & TRYAGAIN) {
10244 *flagp |= TRYAGAIN;
10247 vFAIL("Internal urp");
10248 /* Supposed to be caught earlier. */
10251 if (!regcurly(RExC_parse, FALSE)) {
10260 vFAIL("Quantifier follows nothing");
10265 This switch handles escape sequences that resolve to some kind
10266 of special regop and not to literal text. Escape sequnces that
10267 resolve to literal text are handled below in the switch marked
10270 Every entry in this switch *must* have a corresponding entry
10271 in the literal escape switch. However, the opposite is not
10272 required, as the default for this switch is to jump to the
10273 literal text handling code.
10275 switch ((U8)*++RExC_parse) {
10277 /* Special Escapes */
10279 RExC_seen_zerolen++;
10280 ret = reg_node(pRExC_state, SBOL);
10282 goto finish_meta_pat;
10284 ret = reg_node(pRExC_state, GPOS);
10285 RExC_seen |= REG_SEEN_GPOS;
10287 goto finish_meta_pat;
10289 RExC_seen_zerolen++;
10290 ret = reg_node(pRExC_state, KEEPS);
10292 /* XXX:dmq : disabling in-place substitution seems to
10293 * be necessary here to avoid cases of memory corruption, as
10294 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
10296 RExC_seen |= REG_SEEN_LOOKBEHIND;
10297 goto finish_meta_pat;
10299 ret = reg_node(pRExC_state, SEOL);
10301 RExC_seen_zerolen++; /* Do not optimize RE away */
10302 goto finish_meta_pat;
10304 ret = reg_node(pRExC_state, EOS);
10306 RExC_seen_zerolen++; /* Do not optimize RE away */
10307 goto finish_meta_pat;
10309 ret = reg_node(pRExC_state, CANY);
10310 RExC_seen |= REG_SEEN_CANY;
10311 *flagp |= HASWIDTH|SIMPLE;
10312 goto finish_meta_pat;
10314 ret = reg_node(pRExC_state, CLUMP);
10315 *flagp |= HASWIDTH;
10316 goto finish_meta_pat;
10322 arg = ANYOF_WORDCHAR;
10326 RExC_seen_zerolen++;
10327 RExC_seen |= REG_SEEN_LOOKBEHIND;
10328 op = BOUND + get_regex_charset(RExC_flags);
10329 if (op > BOUNDA) { /* /aa is same as /a */
10332 ret = reg_node(pRExC_state, op);
10333 FLAGS(ret) = get_regex_charset(RExC_flags);
10335 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10336 ckWARNdep(RExC_parse, "\"\\b{\" is deprecated; use \"\\b\\{\" or \"\\b[{]\" instead");
10338 goto finish_meta_pat;
10340 RExC_seen_zerolen++;
10341 RExC_seen |= REG_SEEN_LOOKBEHIND;
10342 op = NBOUND + get_regex_charset(RExC_flags);
10343 if (op > NBOUNDA) { /* /aa is same as /a */
10346 ret = reg_node(pRExC_state, op);
10347 FLAGS(ret) = get_regex_charset(RExC_flags);
10349 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10350 ckWARNdep(RExC_parse, "\"\\B{\" is deprecated; use \"\\B\\{\" or \"\\B[{]\" instead");
10352 goto finish_meta_pat;
10362 ret = reg_node(pRExC_state, LNBREAK);
10363 *flagp |= HASWIDTH|SIMPLE;
10364 goto finish_meta_pat;
10372 goto join_posix_op_known;
10378 arg = ANYOF_VERTWS;
10380 goto join_posix_op_known;
10390 op = POSIXD + get_regex_charset(RExC_flags);
10391 if (op > POSIXA) { /* /aa is same as /a */
10395 join_posix_op_known:
10398 op += NPOSIXD - POSIXD;
10401 ret = reg_node(pRExC_state, op);
10403 FLAGS(ret) = namedclass_to_classnum(arg);
10406 *flagp |= HASWIDTH|SIMPLE;
10410 nextchar(pRExC_state);
10411 Set_Node_Length(ret, 2); /* MJD */
10417 char* parse_start = RExC_parse - 2;
10422 ret = regclass(pRExC_state, flagp,depth+1,
10423 TRUE, /* means just parse this element */
10424 FALSE, /* don't allow multi-char folds */
10425 FALSE, /* don't silence non-portable warnings.
10426 It would be a bug if these returned
10429 /* regclass() can only return RESTART_UTF8 if multi-char folds
10432 FAIL2("panic: regclass returned NULL to regatom, flags=%#X",
10437 Set_Node_Offset(ret, parse_start + 2);
10438 Set_Node_Cur_Length(ret);
10439 nextchar(pRExC_state);
10443 /* Handle \N and \N{NAME} with multiple code points here and not
10444 * below because it can be multicharacter. join_exact() will join
10445 * them up later on. Also this makes sure that things like
10446 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
10447 * The options to the grok function call causes it to fail if the
10448 * sequence is just a single code point. We then go treat it as
10449 * just another character in the current EXACT node, and hence it
10450 * gets uniform treatment with all the other characters. The
10451 * special treatment for quantifiers is not needed for such single
10452 * character sequences */
10454 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
10455 FALSE /* not strict */ )) {
10456 if (*flagp & RESTART_UTF8)
10462 case 'k': /* Handle \k<NAME> and \k'NAME' */
10465 char ch= RExC_parse[1];
10466 if (ch != '<' && ch != '\'' && ch != '{') {
10468 vFAIL2("Sequence %.2s... not terminated",parse_start);
10470 /* this pretty much dupes the code for (?P=...) in reg(), if
10471 you change this make sure you change that */
10472 char* name_start = (RExC_parse += 2);
10474 SV *sv_dat = reg_scan_name(pRExC_state,
10475 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10476 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
10477 if (RExC_parse == name_start || *RExC_parse != ch)
10478 vFAIL2("Sequence %.3s... not terminated",parse_start);
10481 num = add_data( pRExC_state, 1, "S" );
10482 RExC_rxi->data->data[num]=(void*)sv_dat;
10483 SvREFCNT_inc_simple_void(sv_dat);
10487 ret = reganode(pRExC_state,
10490 : (ASCII_FOLD_RESTRICTED)
10492 : (AT_LEAST_UNI_SEMANTICS)
10498 *flagp |= HASWIDTH;
10500 /* override incorrect value set in reganode MJD */
10501 Set_Node_Offset(ret, parse_start+1);
10502 Set_Node_Cur_Length(ret); /* MJD */
10503 nextchar(pRExC_state);
10509 case '1': case '2': case '3': case '4':
10510 case '5': case '6': case '7': case '8': case '9':
10513 bool isg = *RExC_parse == 'g';
10518 if (*RExC_parse == '{') {
10522 if (*RExC_parse == '-') {
10526 if (hasbrace && !isDIGIT(*RExC_parse)) {
10527 if (isrel) RExC_parse--;
10529 goto parse_named_seq;
10531 num = atoi(RExC_parse);
10532 if (isg && num == 0)
10533 vFAIL("Reference to invalid group 0");
10535 num = RExC_npar - num;
10537 vFAIL("Reference to nonexistent or unclosed group");
10539 if (!isg && num > 9 && num >= RExC_npar)
10540 /* Probably a character specified in octal, e.g. \35 */
10543 char * const parse_start = RExC_parse - 1; /* MJD */
10544 while (isDIGIT(*RExC_parse))
10546 if (parse_start == RExC_parse - 1)
10547 vFAIL("Unterminated \\g... pattern");
10549 if (*RExC_parse != '}')
10550 vFAIL("Unterminated \\g{...} pattern");
10554 if (num > (I32)RExC_rx->nparens)
10555 vFAIL("Reference to nonexistent group");
10558 ret = reganode(pRExC_state,
10561 : (ASCII_FOLD_RESTRICTED)
10563 : (AT_LEAST_UNI_SEMANTICS)
10569 *flagp |= HASWIDTH;
10571 /* override incorrect value set in reganode MJD */
10572 Set_Node_Offset(ret, parse_start+1);
10573 Set_Node_Cur_Length(ret); /* MJD */
10575 nextchar(pRExC_state);
10580 if (RExC_parse >= RExC_end)
10581 FAIL("Trailing \\");
10584 /* Do not generate "unrecognized" warnings here, we fall
10585 back into the quick-grab loop below */
10592 if (RExC_flags & RXf_PMf_EXTENDED) {
10593 if ( reg_skipcomment( pRExC_state ) )
10600 parse_start = RExC_parse - 1;
10609 #define MAX_NODE_STRING_SIZE 127
10610 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
10612 U8 upper_parse = MAX_NODE_STRING_SIZE;
10615 bool next_is_quantifier;
10616 char * oldp = NULL;
10618 /* If a folding node contains only code points that don't
10619 * participate in folds, it can be changed into an EXACT node,
10620 * which allows the optimizer more things to look for */
10624 node_type = compute_EXACTish(pRExC_state);
10625 ret = reg_node(pRExC_state, node_type);
10627 /* In pass1, folded, we use a temporary buffer instead of the
10628 * actual node, as the node doesn't exist yet */
10629 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
10635 /* We do the EXACTFish to EXACT node only if folding, and not if in
10636 * locale, as whether a character folds or not isn't known until
10638 maybe_exact = FOLD && ! LOC;
10640 /* XXX The node can hold up to 255 bytes, yet this only goes to
10641 * 127. I (khw) do not know why. Keeping it somewhat less than
10642 * 255 allows us to not have to worry about overflow due to
10643 * converting to utf8 and fold expansion, but that value is
10644 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
10645 * split up by this limit into a single one using the real max of
10646 * 255. Even at 127, this breaks under rare circumstances. If
10647 * folding, we do not want to split a node at a character that is a
10648 * non-final in a multi-char fold, as an input string could just
10649 * happen to want to match across the node boundary. The join
10650 * would solve that problem if the join actually happens. But a
10651 * series of more than two nodes in a row each of 127 would cause
10652 * the first join to succeed to get to 254, but then there wouldn't
10653 * be room for the next one, which could at be one of those split
10654 * multi-char folds. I don't know of any fool-proof solution. One
10655 * could back off to end with only a code point that isn't such a
10656 * non-final, but it is possible for there not to be any in the
10658 for (p = RExC_parse - 1;
10659 len < upper_parse && p < RExC_end;
10664 if (RExC_flags & RXf_PMf_EXTENDED)
10665 p = regwhite( pRExC_state, p );
10676 /* Literal Escapes Switch
10678 This switch is meant to handle escape sequences that
10679 resolve to a literal character.
10681 Every escape sequence that represents something
10682 else, like an assertion or a char class, is handled
10683 in the switch marked 'Special Escapes' above in this
10684 routine, but also has an entry here as anything that
10685 isn't explicitly mentioned here will be treated as
10686 an unescaped equivalent literal.
10689 switch ((U8)*++p) {
10690 /* These are all the special escapes. */
10691 case 'A': /* Start assertion */
10692 case 'b': case 'B': /* Word-boundary assertion*/
10693 case 'C': /* Single char !DANGEROUS! */
10694 case 'd': case 'D': /* digit class */
10695 case 'g': case 'G': /* generic-backref, pos assertion */
10696 case 'h': case 'H': /* HORIZWS */
10697 case 'k': case 'K': /* named backref, keep marker */
10698 case 'p': case 'P': /* Unicode property */
10699 case 'R': /* LNBREAK */
10700 case 's': case 'S': /* space class */
10701 case 'v': case 'V': /* VERTWS */
10702 case 'w': case 'W': /* word class */
10703 case 'X': /* eXtended Unicode "combining character sequence" */
10704 case 'z': case 'Z': /* End of line/string assertion */
10708 /* Anything after here is an escape that resolves to a
10709 literal. (Except digits, which may or may not)
10715 case 'N': /* Handle a single-code point named character. */
10716 /* The options cause it to fail if a multiple code
10717 * point sequence. Handle those in the switch() above
10719 RExC_parse = p + 1;
10720 if (! grok_bslash_N(pRExC_state, NULL, &ender,
10721 flagp, depth, FALSE,
10722 FALSE /* not strict */ ))
10724 if (*flagp & RESTART_UTF8)
10725 FAIL("panic: grok_bslash_N set RESTART_UTF8");
10726 RExC_parse = p = oldp;
10730 if (ender > 0xff) {
10747 ender = ASCII_TO_NATIVE('\033');
10751 ender = ASCII_TO_NATIVE('\007');
10757 const char* error_msg;
10759 bool valid = grok_bslash_o(&p,
10762 TRUE, /* out warnings */
10763 FALSE, /* not strict */
10764 TRUE, /* Output warnings
10769 RExC_parse = p; /* going to die anyway; point
10770 to exact spot of failure */
10774 if (PL_encoding && ender < 0x100) {
10775 goto recode_encoding;
10777 if (ender > 0xff) {
10784 UV result = UV_MAX; /* initialize to erroneous
10786 const char* error_msg;
10788 bool valid = grok_bslash_x(&p,
10791 TRUE, /* out warnings */
10792 FALSE, /* not strict */
10793 TRUE, /* Output warnings
10798 RExC_parse = p; /* going to die anyway; point
10799 to exact spot of failure */
10804 if (PL_encoding && ender < 0x100) {
10805 goto recode_encoding;
10807 if (ender > 0xff) {
10814 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
10816 case '0': case '1': case '2': case '3':case '4':
10817 case '5': case '6': case '7':
10819 (isDIGIT(p[1]) && atoi(p) >= RExC_npar))
10821 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
10823 ender = grok_oct(p, &numlen, &flags, NULL);
10824 if (ender > 0xff) {
10828 if (SIZE_ONLY /* like \08, \178 */
10831 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
10833 reg_warn_non_literal_string(
10835 form_short_octal_warning(p, numlen));
10838 else { /* Not to be treated as an octal constant, go
10843 if (PL_encoding && ender < 0x100)
10844 goto recode_encoding;
10847 if (! RExC_override_recoding) {
10848 SV* enc = PL_encoding;
10849 ender = reg_recode((const char)(U8)ender, &enc);
10850 if (!enc && SIZE_ONLY)
10851 ckWARNreg(p, "Invalid escape in the specified encoding");
10857 FAIL("Trailing \\");
10860 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
10861 /* Include any { following the alpha to emphasize
10862 * that it could be part of an escape at some point
10864 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
10865 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
10867 goto normal_default;
10868 } /* End of switch on '\' */
10870 default: /* A literal character */
10873 && RExC_flags & RXf_PMf_EXTENDED
10874 && ckWARN(WARN_DEPRECATED)
10875 && is_PATWS_non_low(p, UTF))
10877 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
10878 "Escape literal pattern white space under /x");
10882 if (UTF8_IS_START(*p) && UTF) {
10884 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
10885 &numlen, UTF8_ALLOW_DEFAULT);
10891 } /* End of switch on the literal */
10893 /* Here, have looked at the literal character and <ender>
10894 * contains its ordinal, <p> points to the character after it
10897 if ( RExC_flags & RXf_PMf_EXTENDED)
10898 p = regwhite( pRExC_state, p );
10900 /* If the next thing is a quantifier, it applies to this
10901 * character only, which means that this character has to be in
10902 * its own node and can't just be appended to the string in an
10903 * existing node, so if there are already other characters in
10904 * the node, close the node with just them, and set up to do
10905 * this character again next time through, when it will be the
10906 * only thing in its new node */
10907 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
10915 /* See comments for join_exact() as to why we fold
10916 * this non-UTF at compile time */
10917 || (node_type == EXACTFU
10918 && ender == LATIN_SMALL_LETTER_SHARP_S))
10922 /* Prime the casefolded buffer. Locale rules, which
10923 * apply only to code points < 256, aren't known until
10924 * execution, so for them, just output the original
10925 * character using utf8. If we start to fold non-UTF
10926 * patterns, be sure to update join_exact() */
10927 if (LOC && ender < 256) {
10928 if (UNI_IS_INVARIANT(ender)) {
10932 *s = UTF8_TWO_BYTE_HI(ender);
10933 *(s + 1) = UTF8_TWO_BYTE_LO(ender);
10938 UV folded = _to_uni_fold_flags(
10943 | ((LOC) ? FOLD_FLAGS_LOCALE
10944 : (ASCII_FOLD_RESTRICTED)
10945 ? FOLD_FLAGS_NOMIX_ASCII
10949 /* If this node only contains non-folding code
10950 * points so far, see if this new one is also
10953 if (folded != ender) {
10954 maybe_exact = FALSE;
10957 /* Here the fold is the original; we have
10958 * to check further to see if anything
10960 if (! PL_utf8_foldable) {
10961 SV* swash = swash_init("utf8",
10963 &PL_sv_undef, 1, 0);
10965 _get_swash_invlist(swash);
10966 SvREFCNT_dec_NN(swash);
10968 if (_invlist_contains_cp(PL_utf8_foldable,
10971 maybe_exact = FALSE;
10979 /* The loop increments <len> each time, as all but this
10980 * path (and the one just below for UTF) through it add
10981 * a single byte to the EXACTish node. But this one
10982 * has changed len to be the correct final value, so
10983 * subtract one to cancel out the increment that
10985 len += foldlen - 1;
10988 *(s++) = (char) ender;
10989 maybe_exact &= ! IS_IN_SOME_FOLD_L1(ender);
10993 const STRLEN unilen = reguni(pRExC_state, ender, s);
10999 /* See comment just above for - 1 */
11003 REGC((char)ender, s++);
11006 if (next_is_quantifier) {
11008 /* Here, the next input is a quantifier, and to get here,
11009 * the current character is the only one in the node.
11010 * Also, here <len> doesn't include the final byte for this
11016 } /* End of loop through literal characters */
11018 /* Here we have either exhausted the input or ran out of room in
11019 * the node. (If we encountered a character that can't be in the
11020 * node, transfer is made directly to <loopdone>, and so we
11021 * wouldn't have fallen off the end of the loop.) In the latter
11022 * case, we artificially have to split the node into two, because
11023 * we just don't have enough space to hold everything. This
11024 * creates a problem if the final character participates in a
11025 * multi-character fold in the non-final position, as a match that
11026 * should have occurred won't, due to the way nodes are matched,
11027 * and our artificial boundary. So back off until we find a non-
11028 * problematic character -- one that isn't at the beginning or
11029 * middle of such a fold. (Either it doesn't participate in any
11030 * folds, or appears only in the final position of all the folds it
11031 * does participate in.) A better solution with far fewer false
11032 * positives, and that would fill the nodes more completely, would
11033 * be to actually have available all the multi-character folds to
11034 * test against, and to back-off only far enough to be sure that
11035 * this node isn't ending with a partial one. <upper_parse> is set
11036 * further below (if we need to reparse the node) to include just
11037 * up through that final non-problematic character that this code
11038 * identifies, so when it is set to less than the full node, we can
11039 * skip the rest of this */
11040 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
11042 const STRLEN full_len = len;
11044 assert(len >= MAX_NODE_STRING_SIZE);
11046 /* Here, <s> points to the final byte of the final character.
11047 * Look backwards through the string until find a non-
11048 * problematic character */
11052 /* These two have no multi-char folds to non-UTF characters
11054 if (ASCII_FOLD_RESTRICTED || LOC) {
11058 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
11062 if (! PL_NonL1NonFinalFold) {
11063 PL_NonL1NonFinalFold = _new_invlist_C_array(
11064 NonL1_Perl_Non_Final_Folds_invlist);
11067 /* Point to the first byte of the final character */
11068 s = (char *) utf8_hop((U8 *) s, -1);
11070 while (s >= s0) { /* Search backwards until find
11071 non-problematic char */
11072 if (UTF8_IS_INVARIANT(*s)) {
11074 /* There are no ascii characters that participate
11075 * in multi-char folds under /aa. In EBCDIC, the
11076 * non-ascii invariants are all control characters,
11077 * so don't ever participate in any folds. */
11078 if (ASCII_FOLD_RESTRICTED
11079 || ! IS_NON_FINAL_FOLD(*s))
11084 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
11086 /* No Latin1 characters participate in multi-char
11087 * folds under /l */
11089 || ! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_UNI(
11095 else if (! _invlist_contains_cp(
11096 PL_NonL1NonFinalFold,
11097 valid_utf8_to_uvchr((U8 *) s, NULL)))
11102 /* Here, the current character is problematic in that
11103 * it does occur in the non-final position of some
11104 * fold, so try the character before it, but have to
11105 * special case the very first byte in the string, so
11106 * we don't read outside the string */
11107 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
11108 } /* End of loop backwards through the string */
11110 /* If there were only problematic characters in the string,
11111 * <s> will point to before s0, in which case the length
11112 * should be 0, otherwise include the length of the
11113 * non-problematic character just found */
11114 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
11117 /* Here, have found the final character, if any, that is
11118 * non-problematic as far as ending the node without splitting
11119 * it across a potential multi-char fold. <len> contains the
11120 * number of bytes in the node up-to and including that
11121 * character, or is 0 if there is no such character, meaning
11122 * the whole node contains only problematic characters. In
11123 * this case, give up and just take the node as-is. We can't
11129 /* Here, the node does contain some characters that aren't
11130 * problematic. If one such is the final character in the
11131 * node, we are done */
11132 if (len == full_len) {
11135 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
11137 /* If the final character is problematic, but the
11138 * penultimate is not, back-off that last character to
11139 * later start a new node with it */
11144 /* Here, the final non-problematic character is earlier
11145 * in the input than the penultimate character. What we do
11146 * is reparse from the beginning, going up only as far as
11147 * this final ok one, thus guaranteeing that the node ends
11148 * in an acceptable character. The reason we reparse is
11149 * that we know how far in the character is, but we don't
11150 * know how to correlate its position with the input parse.
11151 * An alternate implementation would be to build that
11152 * correlation as we go along during the original parse,
11153 * but that would entail extra work for every node, whereas
11154 * this code gets executed only when the string is too
11155 * large for the node, and the final two characters are
11156 * problematic, an infrequent occurrence. Yet another
11157 * possible strategy would be to save the tail of the
11158 * string, and the next time regatom is called, initialize
11159 * with that. The problem with this is that unless you
11160 * back off one more character, you won't be guaranteed
11161 * regatom will get called again, unless regbranch,
11162 * regpiece ... are also changed. If you do back off that
11163 * extra character, so that there is input guaranteed to
11164 * force calling regatom, you can't handle the case where
11165 * just the first character in the node is acceptable. I
11166 * (khw) decided to try this method which doesn't have that
11167 * pitfall; if performance issues are found, we can do a
11168 * combination of the current approach plus that one */
11174 } /* End of verifying node ends with an appropriate char */
11176 loopdone: /* Jumped to when encounters something that shouldn't be in
11179 /* If 'maybe_exact' is still set here, means there are no
11180 * code points in the node that participate in folds */
11181 if (FOLD && maybe_exact) {
11185 /* I (khw) don't know if you can get here with zero length, but the
11186 * old code handled this situation by creating a zero-length EXACT
11187 * node. Might as well be NOTHING instead */
11192 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
11195 RExC_parse = p - 1;
11196 Set_Node_Cur_Length(ret); /* MJD */
11197 nextchar(pRExC_state);
11199 /* len is STRLEN which is unsigned, need to copy to signed */
11202 vFAIL("Internal disaster");
11205 } /* End of label 'defchar:' */
11207 } /* End of giant switch on input character */
11213 S_regwhite( RExC_state_t *pRExC_state, char *p )
11215 const char *e = RExC_end;
11217 PERL_ARGS_ASSERT_REGWHITE;
11222 else if (*p == '#') {
11225 if (*p++ == '\n') {
11231 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11240 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
11242 /* Returns the next non-pattern-white space, non-comment character (the
11243 * latter only if 'recognize_comment is true) in the string p, which is
11244 * ended by RExC_end. If there is no line break ending a comment,
11245 * RExC_seen has added the REG_SEEN_RUN_ON_COMMENT flag; */
11246 const char *e = RExC_end;
11248 PERL_ARGS_ASSERT_REGPATWS;
11252 if ((len = is_PATWS_safe(p, e, UTF))) {
11255 else if (recognize_comment && *p == '#') {
11259 if (is_LNBREAK_safe(p, e, UTF)) {
11265 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11273 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
11274 Character classes ([:foo:]) can also be negated ([:^foo:]).
11275 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
11276 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
11277 but trigger failures because they are currently unimplemented. */
11279 #define POSIXCC_DONE(c) ((c) == ':')
11280 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
11281 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
11283 PERL_STATIC_INLINE I32
11284 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
11287 I32 namedclass = OOB_NAMEDCLASS;
11289 PERL_ARGS_ASSERT_REGPPOSIXCC;
11291 if (value == '[' && RExC_parse + 1 < RExC_end &&
11292 /* I smell either [: or [= or [. -- POSIX has been here, right? */
11293 POSIXCC(UCHARAT(RExC_parse)))
11295 const char c = UCHARAT(RExC_parse);
11296 char* const s = RExC_parse++;
11298 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
11300 if (RExC_parse == RExC_end) {
11303 /* Try to give a better location for the error (than the end of
11304 * the string) by looking for the matching ']' */
11306 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
11309 vFAIL2("Unmatched '%c' in POSIX class", c);
11311 /* Grandfather lone [:, [=, [. */
11315 const char* const t = RExC_parse++; /* skip over the c */
11318 if (UCHARAT(RExC_parse) == ']') {
11319 const char *posixcc = s + 1;
11320 RExC_parse++; /* skip over the ending ] */
11323 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
11324 const I32 skip = t - posixcc;
11326 /* Initially switch on the length of the name. */
11329 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
11330 this is the Perl \w
11332 namedclass = ANYOF_WORDCHAR;
11335 /* Names all of length 5. */
11336 /* alnum alpha ascii blank cntrl digit graph lower
11337 print punct space upper */
11338 /* Offset 4 gives the best switch position. */
11339 switch (posixcc[4]) {
11341 if (memEQ(posixcc, "alph", 4)) /* alpha */
11342 namedclass = ANYOF_ALPHA;
11345 if (memEQ(posixcc, "spac", 4)) /* space */
11346 namedclass = ANYOF_PSXSPC;
11349 if (memEQ(posixcc, "grap", 4)) /* graph */
11350 namedclass = ANYOF_GRAPH;
11353 if (memEQ(posixcc, "asci", 4)) /* ascii */
11354 namedclass = ANYOF_ASCII;
11357 if (memEQ(posixcc, "blan", 4)) /* blank */
11358 namedclass = ANYOF_BLANK;
11361 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
11362 namedclass = ANYOF_CNTRL;
11365 if (memEQ(posixcc, "alnu", 4)) /* alnum */
11366 namedclass = ANYOF_ALPHANUMERIC;
11369 if (memEQ(posixcc, "lowe", 4)) /* lower */
11370 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
11371 else if (memEQ(posixcc, "uppe", 4)) /* upper */
11372 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
11375 if (memEQ(posixcc, "digi", 4)) /* digit */
11376 namedclass = ANYOF_DIGIT;
11377 else if (memEQ(posixcc, "prin", 4)) /* print */
11378 namedclass = ANYOF_PRINT;
11379 else if (memEQ(posixcc, "punc", 4)) /* punct */
11380 namedclass = ANYOF_PUNCT;
11385 if (memEQ(posixcc, "xdigit", 6))
11386 namedclass = ANYOF_XDIGIT;
11390 if (namedclass == OOB_NAMEDCLASS)
11391 Simple_vFAIL3("POSIX class [:%.*s:] unknown",
11394 /* The #defines are structured so each complement is +1 to
11395 * the normal one */
11399 assert (posixcc[skip] == ':');
11400 assert (posixcc[skip+1] == ']');
11401 } else if (!SIZE_ONLY) {
11402 /* [[=foo=]] and [[.foo.]] are still future. */
11404 /* adjust RExC_parse so the warning shows after
11405 the class closes */
11406 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
11408 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11411 /* Maternal grandfather:
11412 * "[:" ending in ":" but not in ":]" */
11414 vFAIL("Unmatched '[' in POSIX class");
11417 /* Grandfather lone [:, [=, [. */
11427 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
11429 /* This applies some heuristics at the current parse position (which should
11430 * be at a '[') to see if what follows might be intended to be a [:posix:]
11431 * class. It returns true if it really is a posix class, of course, but it
11432 * also can return true if it thinks that what was intended was a posix
11433 * class that didn't quite make it.
11435 * It will return true for
11437 * [:alphanumerics] (as long as the ] isn't followed immediately by a
11438 * ')' indicating the end of the (?[
11439 * [:any garbage including %^&$ punctuation:]
11441 * This is designed to be called only from S_handle_regex_sets; it could be
11442 * easily adapted to be called from the spot at the beginning of regclass()
11443 * that checks to see in a normal bracketed class if the surrounding []
11444 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
11445 * change long-standing behavior, so I (khw) didn't do that */
11446 char* p = RExC_parse + 1;
11447 char first_char = *p;
11449 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
11451 assert(*(p - 1) == '[');
11453 if (! POSIXCC(first_char)) {
11458 while (p < RExC_end && isWORDCHAR(*p)) p++;
11460 if (p >= RExC_end) {
11464 if (p - RExC_parse > 2 /* Got at least 1 word character */
11465 && (*p == first_char
11466 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
11471 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
11474 && p - RExC_parse > 2 /* [:] evaluates to colon;
11475 [::] is a bad posix class. */
11476 && first_char == *(p - 1));
11480 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist, I32 *flagp, U32 depth,
11481 char * const oregcomp_parse)
11483 /* Handle the (?[...]) construct to do set operations */
11486 UV start, end; /* End points of code point ranges */
11488 char *save_end, *save_parse;
11493 const bool save_fold = FOLD;
11495 GET_RE_DEBUG_FLAGS_DECL;
11497 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
11500 vFAIL("(?[...]) not valid in locale");
11502 RExC_uni_semantics = 1;
11504 /* This will return only an ANYOF regnode, or (unlikely) something smaller
11505 * (such as EXACT). Thus we can skip most everything if just sizing. We
11506 * call regclass to handle '[]' so as to not have to reinvent its parsing
11507 * rules here (throwing away the size it computes each time). And, we exit
11508 * upon an unescaped ']' that isn't one ending a regclass. To do both
11509 * these things, we need to realize that something preceded by a backslash
11510 * is escaped, so we have to keep track of backslashes */
11513 Perl_ck_warner_d(aTHX_
11514 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
11515 "The regex_sets feature is experimental" REPORT_LOCATION,
11516 (int) (RExC_parse - RExC_precomp) , RExC_precomp, RExC_parse);
11518 while (RExC_parse < RExC_end) {
11519 SV* current = NULL;
11520 RExC_parse = regpatws(pRExC_state, RExC_parse,
11521 TRUE); /* means recognize comments */
11522 switch (*RExC_parse) {
11526 /* Skip the next byte (which could cause us to end up in
11527 * the middle of a UTF-8 character, but since none of those
11528 * are confusable with anything we currently handle in this
11529 * switch (invariants all), it's safe. We'll just hit the
11530 * default: case next time and keep on incrementing until
11531 * we find one of the invariants we do handle. */
11536 /* If this looks like it is a [:posix:] class, leave the
11537 * parse pointer at the '[' to fool regclass() into
11538 * thinking it is part of a '[[:posix:]]'. That function
11539 * will use strict checking to force a syntax error if it
11540 * doesn't work out to a legitimate class */
11541 bool is_posix_class
11542 = could_it_be_a_POSIX_class(pRExC_state);
11543 if (! is_posix_class) {
11547 /* regclass() can only return RESTART_UTF8 if multi-char
11548 folds are allowed. */
11549 if (!regclass(pRExC_state, flagp,depth+1,
11550 is_posix_class, /* parse the whole char
11551 class only if not a
11553 FALSE, /* don't allow multi-char folds */
11554 TRUE, /* silence non-portable warnings. */
11556 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X",
11559 /* function call leaves parse pointing to the ']', except
11560 * if we faked it */
11561 if (is_posix_class) {
11565 SvREFCNT_dec(current); /* In case it returned something */
11571 if (RExC_parse < RExC_end
11572 && *RExC_parse == ')')
11574 node = reganode(pRExC_state, ANYOF, 0);
11575 RExC_size += ANYOF_SKIP;
11576 nextchar(pRExC_state);
11577 Set_Node_Length(node,
11578 RExC_parse - oregcomp_parse + 1); /* MJD */
11587 FAIL("Syntax error in (?[...])");
11590 /* Pass 2 only after this. Everything in this construct is a
11591 * metacharacter. Operands begin with either a '\' (for an escape
11592 * sequence), or a '[' for a bracketed character class. Any other
11593 * character should be an operator, or parenthesis for grouping. Both
11594 * types of operands are handled by calling regclass() to parse them. It
11595 * is called with a parameter to indicate to return the computed inversion
11596 * list. The parsing here is implemented via a stack. Each entry on the
11597 * stack is a single character representing one of the operators, or the
11598 * '('; or else a pointer to an operand inversion list. */
11600 #define IS_OPERAND(a) (! SvIOK(a))
11602 /* The stack starts empty. It is a syntax error if the first thing parsed
11603 * is a binary operator; everything else is pushed on the stack. When an
11604 * operand is parsed, the top of the stack is examined. If it is a binary
11605 * operator, the item before it should be an operand, and both are replaced
11606 * by the result of doing that operation on the new operand and the one on
11607 * the stack. Thus a sequence of binary operands is reduced to a single
11608 * one before the next one is parsed.
11610 * A unary operator may immediately follow a binary in the input, for
11613 * When an operand is parsed and the top of the stack is a unary operator,
11614 * the operation is performed, and then the stack is rechecked to see if
11615 * this new operand is part of a binary operation; if so, it is handled as
11618 * A '(' is simply pushed on the stack; it is valid only if the stack is
11619 * empty, or the top element of the stack is an operator or another '('
11620 * (for which the parenthesized expression will become an operand). By the
11621 * time the corresponding ')' is parsed everything in between should have
11622 * been parsed and evaluated to a single operand (or else is a syntax
11623 * error), and is handled as a regular operand */
11627 while (RExC_parse < RExC_end) {
11628 I32 top_index = av_tindex(stack);
11630 SV* current = NULL;
11632 /* Skip white space */
11633 RExC_parse = regpatws(pRExC_state, RExC_parse,
11634 TRUE); /* means recognize comments */
11635 if (RExC_parse >= RExC_end) {
11636 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
11638 if ((curchar = UCHARAT(RExC_parse)) == ']') {
11645 if (av_tindex(stack) >= 0 /* This makes sure that we can
11646 safely subtract 1 from
11647 RExC_parse in the next clause.
11648 If we have something on the
11649 stack, we have parsed something
11651 && UCHARAT(RExC_parse - 1) == '('
11652 && RExC_parse < RExC_end)
11654 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
11655 * This happens when we have some thing like
11657 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
11659 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
11661 * Here we would be handling the interpolated
11662 * '$thai_or_lao'. We handle this by a recursive call to
11663 * ourselves which returns the inversion list the
11664 * interpolated expression evaluates to. We use the flags
11665 * from the interpolated pattern. */
11666 U32 save_flags = RExC_flags;
11667 const char * const save_parse = ++RExC_parse;
11669 parse_lparen_question_flags(pRExC_state);
11671 if (RExC_parse == save_parse /* Makes sure there was at
11672 least one flag (or this
11673 embedding wasn't compiled)
11675 || RExC_parse >= RExC_end - 4
11676 || UCHARAT(RExC_parse) != ':'
11677 || UCHARAT(++RExC_parse) != '('
11678 || UCHARAT(++RExC_parse) != '?'
11679 || UCHARAT(++RExC_parse) != '[')
11682 /* In combination with the above, this moves the
11683 * pointer to the point just after the first erroneous
11684 * character (or if there are no flags, to where they
11685 * should have been) */
11686 if (RExC_parse >= RExC_end - 4) {
11687 RExC_parse = RExC_end;
11689 else if (RExC_parse != save_parse) {
11690 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11692 vFAIL("Expecting '(?flags:(?[...'");
11695 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
11696 depth+1, oregcomp_parse);
11698 /* Here, 'current' contains the embedded expression's
11699 * inversion list, and RExC_parse points to the trailing
11700 * ']'; the next character should be the ')' which will be
11701 * paired with the '(' that has been put on the stack, so
11702 * the whole embedded expression reduces to '(operand)' */
11705 RExC_flags = save_flags;
11706 goto handle_operand;
11711 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11712 vFAIL("Unexpected character");
11715 /* regclass() can only return RESTART_UTF8 if multi-char
11716 folds are allowed. */
11717 if (!regclass(pRExC_state, flagp,depth+1,
11718 TRUE, /* means parse just the next thing */
11719 FALSE, /* don't allow multi-char folds */
11720 FALSE, /* don't silence non-portable warnings. */
11722 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X",
11724 /* regclass() will return with parsing just the \ sequence,
11725 * leaving the parse pointer at the next thing to parse */
11727 goto handle_operand;
11729 case '[': /* Is a bracketed character class */
11731 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
11733 if (! is_posix_class) {
11737 /* regclass() can only return RESTART_UTF8 if multi-char
11738 folds are allowed. */
11739 if(!regclass(pRExC_state, flagp,depth+1,
11740 is_posix_class, /* parse the whole char class
11741 only if not a posix class */
11742 FALSE, /* don't allow multi-char folds */
11743 FALSE, /* don't silence non-portable warnings. */
11745 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X",
11747 /* function call leaves parse pointing to the ']', except if we
11749 if (is_posix_class) {
11753 goto handle_operand;
11762 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
11763 || ! IS_OPERAND(*top_ptr))
11766 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
11768 av_push(stack, newSVuv(curchar));
11772 av_push(stack, newSVuv(curchar));
11776 if (top_index >= 0) {
11777 top_ptr = av_fetch(stack, top_index, FALSE);
11779 if (IS_OPERAND(*top_ptr)) {
11781 vFAIL("Unexpected '(' with no preceding operator");
11784 av_push(stack, newSVuv(curchar));
11791 || ! (current = av_pop(stack))
11792 || ! IS_OPERAND(current)
11793 || ! (lparen = av_pop(stack))
11794 || IS_OPERAND(lparen)
11795 || SvUV(lparen) != '(')
11798 vFAIL("Unexpected ')'");
11801 SvREFCNT_dec_NN(lparen);
11808 /* Here, we have an operand to process, in 'current' */
11810 if (top_index < 0) { /* Just push if stack is empty */
11811 av_push(stack, current);
11814 SV* top = av_pop(stack);
11815 char current_operator;
11817 if (IS_OPERAND(top)) {
11818 vFAIL("Operand with no preceding operator");
11820 current_operator = (char) SvUV(top);
11821 switch (current_operator) {
11822 case '(': /* Push the '(' back on followed by the new
11824 av_push(stack, top);
11825 av_push(stack, current);
11826 SvREFCNT_inc(top); /* Counters the '_dec' done
11827 just after the 'break', so
11828 it doesn't get wrongly freed
11833 _invlist_invert(current);
11835 /* Unlike binary operators, the top of the stack,
11836 * now that this unary one has been popped off, may
11837 * legally be an operator, and we now have operand
11840 SvREFCNT_dec_NN(top);
11841 goto handle_operand;
11844 _invlist_intersection(av_pop(stack),
11847 av_push(stack, current);
11852 _invlist_union(av_pop(stack), current, ¤t);
11853 av_push(stack, current);
11857 _invlist_subtract(av_pop(stack), current, ¤t);
11858 av_push(stack, current);
11861 case '^': /* The union minus the intersection */
11867 element = av_pop(stack);
11868 _invlist_union(element, current, &u);
11869 _invlist_intersection(element, current, &i);
11870 _invlist_subtract(u, i, ¤t);
11871 av_push(stack, current);
11872 SvREFCNT_dec_NN(i);
11873 SvREFCNT_dec_NN(u);
11874 SvREFCNT_dec_NN(element);
11879 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
11881 SvREFCNT_dec_NN(top);
11885 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11888 if (av_tindex(stack) < 0 /* Was empty */
11889 || ((final = av_pop(stack)) == NULL)
11890 || ! IS_OPERAND(final)
11891 || av_tindex(stack) >= 0) /* More left on stack */
11893 vFAIL("Incomplete expression within '(?[ ])'");
11896 /* Here, 'final' is the resultant inversion list from evaluating the
11897 * expression. Return it if so requested */
11898 if (return_invlist) {
11899 *return_invlist = final;
11903 /* Otherwise generate a resultant node, based on 'final'. regclass() is
11904 * expecting a string of ranges and individual code points */
11905 invlist_iterinit(final);
11906 result_string = newSVpvs("");
11907 while (invlist_iternext(final, &start, &end)) {
11908 if (start == end) {
11909 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
11912 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
11917 save_parse = RExC_parse;
11918 RExC_parse = SvPV(result_string, len);
11919 save_end = RExC_end;
11920 RExC_end = RExC_parse + len;
11922 /* We turn off folding around the call, as the class we have constructed
11923 * already has all folding taken into consideration, and we don't want
11924 * regclass() to add to that */
11925 RExC_flags &= ~RXf_PMf_FOLD;
11926 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
11928 node = regclass(pRExC_state, flagp,depth+1,
11929 FALSE, /* means parse the whole char class */
11930 FALSE, /* don't allow multi-char folds */
11931 TRUE, /* silence non-portable warnings. The above may very
11932 well have generated non-portable code points, but
11933 they're valid on this machine */
11936 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
11939 RExC_flags |= RXf_PMf_FOLD;
11941 RExC_parse = save_parse + 1;
11942 RExC_end = save_end;
11943 SvREFCNT_dec_NN(final);
11944 SvREFCNT_dec_NN(result_string);
11945 SvREFCNT_dec_NN(stack);
11947 nextchar(pRExC_state);
11948 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
11953 /* The names of properties whose definitions are not known at compile time are
11954 * stored in this SV, after a constant heading. So if the length has been
11955 * changed since initialization, then there is a run-time definition. */
11956 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION (SvCUR(listsv) != initial_listsv_len)
11959 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
11960 const bool stop_at_1, /* Just parse the next thing, don't
11961 look for a full character class */
11962 bool allow_multi_folds,
11963 const bool silence_non_portable, /* Don't output warnings
11966 SV** ret_invlist) /* Return an inversion list, not a node */
11968 /* parse a bracketed class specification. Most of these will produce an
11969 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
11970 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
11971 * under /i with multi-character folds: it will be rewritten following the
11972 * paradigm of this example, where the <multi-fold>s are characters which
11973 * fold to multiple character sequences:
11974 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
11975 * gets effectively rewritten as:
11976 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
11977 * reg() gets called (recursively) on the rewritten version, and this
11978 * function will return what it constructs. (Actually the <multi-fold>s
11979 * aren't physically removed from the [abcdefghi], it's just that they are
11980 * ignored in the recursion by means of a flag:
11981 * <RExC_in_multi_char_class>.)
11983 * ANYOF nodes contain a bit map for the first 256 characters, with the
11984 * corresponding bit set if that character is in the list. For characters
11985 * above 255, a range list or swash is used. There are extra bits for \w,
11986 * etc. in locale ANYOFs, as what these match is not determinable at
11989 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
11990 * to be restarted. This can only happen if ret_invlist is non-NULL.
11994 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
11996 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
11999 IV namedclass = OOB_NAMEDCLASS;
12000 char *rangebegin = NULL;
12001 bool need_class = 0;
12003 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
12004 than just initialized. */
12005 SV* properties = NULL; /* Code points that match \p{} \P{} */
12006 SV* posixes = NULL; /* Code points that match classes like, [:word:],
12007 extended beyond the Latin1 range */
12008 UV element_count = 0; /* Number of distinct elements in the class.
12009 Optimizations may be possible if this is tiny */
12010 AV * multi_char_matches = NULL; /* Code points that fold to more than one
12011 character; used under /i */
12013 char * stop_ptr = RExC_end; /* where to stop parsing */
12014 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
12016 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
12018 /* Unicode properties are stored in a swash; this holds the current one
12019 * being parsed. If this swash is the only above-latin1 component of the
12020 * character class, an optimization is to pass it directly on to the
12021 * execution engine. Otherwise, it is set to NULL to indicate that there
12022 * are other things in the class that have to be dealt with at execution
12024 SV* swash = NULL; /* Code points that match \p{} \P{} */
12026 /* Set if a component of this character class is user-defined; just passed
12027 * on to the engine */
12028 bool has_user_defined_property = FALSE;
12030 /* inversion list of code points this node matches only when the target
12031 * string is in UTF-8. (Because is under /d) */
12032 SV* depends_list = NULL;
12034 /* inversion list of code points this node matches. For much of the
12035 * function, it includes only those that match regardless of the utf8ness
12036 * of the target string */
12037 SV* cp_list = NULL;
12040 /* In a range, counts how many 0-2 of the ends of it came from literals,
12041 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
12042 UV literal_endpoint = 0;
12044 bool invert = FALSE; /* Is this class to be complemented */
12046 /* Is there any thing like \W or [:^digit:] that matches above the legal
12047 * Unicode range? */
12048 bool runtime_posix_matches_above_Unicode = FALSE;
12050 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
12051 case we need to change the emitted regop to an EXACT. */
12052 const char * orig_parse = RExC_parse;
12053 const I32 orig_size = RExC_size;
12054 GET_RE_DEBUG_FLAGS_DECL;
12056 PERL_ARGS_ASSERT_REGCLASS;
12058 PERL_UNUSED_ARG(depth);
12061 DEBUG_PARSE("clas");
12063 /* Assume we are going to generate an ANYOF node. */
12064 ret = reganode(pRExC_state, ANYOF, 0);
12067 RExC_size += ANYOF_SKIP;
12068 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
12071 ANYOF_FLAGS(ret) = 0;
12073 RExC_emit += ANYOF_SKIP;
12075 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
12077 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
12078 initial_listsv_len = SvCUR(listsv);
12079 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
12083 RExC_parse = regpatws(pRExC_state, RExC_parse,
12084 FALSE /* means don't recognize comments */);
12087 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
12090 allow_multi_folds = FALSE;
12093 RExC_parse = regpatws(pRExC_state, RExC_parse,
12094 FALSE /* means don't recognize comments */);
12098 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
12099 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
12100 const char *s = RExC_parse;
12101 const char c = *s++;
12103 while (isWORDCHAR(*s))
12105 if (*s && c == *s && s[1] == ']') {
12106 SAVEFREESV(RExC_rx_sv);
12108 "POSIX syntax [%c %c] belongs inside character classes",
12110 (void)ReREFCNT_inc(RExC_rx_sv);
12114 /* If the caller wants us to just parse a single element, accomplish this
12115 * by faking the loop ending condition */
12116 if (stop_at_1 && RExC_end > RExC_parse) {
12117 stop_ptr = RExC_parse + 1;
12120 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
12121 if (UCHARAT(RExC_parse) == ']')
12122 goto charclassloop;
12126 if (RExC_parse >= stop_ptr) {
12131 RExC_parse = regpatws(pRExC_state, RExC_parse,
12132 FALSE /* means don't recognize comments */);
12135 if (UCHARAT(RExC_parse) == ']') {
12141 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
12142 save_value = value;
12143 save_prevvalue = prevvalue;
12146 rangebegin = RExC_parse;
12150 value = utf8n_to_uvchr((U8*)RExC_parse,
12151 RExC_end - RExC_parse,
12152 &numlen, UTF8_ALLOW_DEFAULT);
12153 RExC_parse += numlen;
12156 value = UCHARAT(RExC_parse++);
12159 && RExC_parse < RExC_end
12160 && POSIXCC(UCHARAT(RExC_parse)))
12162 namedclass = regpposixcc(pRExC_state, value, strict);
12164 else if (value == '\\') {
12166 value = utf8n_to_uvchr((U8*)RExC_parse,
12167 RExC_end - RExC_parse,
12168 &numlen, UTF8_ALLOW_DEFAULT);
12169 RExC_parse += numlen;
12172 value = UCHARAT(RExC_parse++);
12174 /* Some compilers cannot handle switching on 64-bit integer
12175 * values, therefore value cannot be an UV. Yes, this will
12176 * be a problem later if we want switch on Unicode.
12177 * A similar issue a little bit later when switching on
12178 * namedclass. --jhi */
12180 /* If the \ is escaping white space when white space is being
12181 * skipped, it means that that white space is wanted literally, and
12182 * is already in 'value'. Otherwise, need to translate the escape
12183 * into what it signifies. */
12184 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
12186 case 'w': namedclass = ANYOF_WORDCHAR; break;
12187 case 'W': namedclass = ANYOF_NWORDCHAR; break;
12188 case 's': namedclass = ANYOF_SPACE; break;
12189 case 'S': namedclass = ANYOF_NSPACE; break;
12190 case 'd': namedclass = ANYOF_DIGIT; break;
12191 case 'D': namedclass = ANYOF_NDIGIT; break;
12192 case 'v': namedclass = ANYOF_VERTWS; break;
12193 case 'V': namedclass = ANYOF_NVERTWS; break;
12194 case 'h': namedclass = ANYOF_HORIZWS; break;
12195 case 'H': namedclass = ANYOF_NHORIZWS; break;
12196 case 'N': /* Handle \N{NAME} in class */
12198 /* We only pay attention to the first char of
12199 multichar strings being returned. I kinda wonder
12200 if this makes sense as it does change the behaviour
12201 from earlier versions, OTOH that behaviour was broken
12203 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
12204 TRUE, /* => charclass */
12207 if (*flagp & RESTART_UTF8)
12208 FAIL("panic: grok_bslash_N set RESTART_UTF8");
12218 /* We will handle any undefined properties ourselves */
12219 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF;
12221 if (RExC_parse >= RExC_end)
12222 vFAIL2("Empty \\%c{}", (U8)value);
12223 if (*RExC_parse == '{') {
12224 const U8 c = (U8)value;
12225 e = strchr(RExC_parse++, '}');
12227 vFAIL2("Missing right brace on \\%c{}", c);
12228 while (isSPACE(UCHARAT(RExC_parse)))
12230 if (e == RExC_parse)
12231 vFAIL2("Empty \\%c{}", c);
12232 n = e - RExC_parse;
12233 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
12244 if (UCHARAT(RExC_parse) == '^') {
12247 /* toggle. (The rhs xor gets the single bit that
12248 * differs between P and p; the other xor inverts just
12250 value ^= 'P' ^ 'p';
12252 while (isSPACE(UCHARAT(RExC_parse))) {
12257 /* Try to get the definition of the property into
12258 * <invlist>. If /i is in effect, the effective property
12259 * will have its name be <__NAME_i>. The design is
12260 * discussed in commit
12261 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
12262 Newx(name, n + sizeof("_i__\n"), char);
12264 sprintf(name, "%s%.*s%s\n",
12265 (FOLD) ? "__" : "",
12271 /* Look up the property name, and get its swash and
12272 * inversion list, if the property is found */
12274 SvREFCNT_dec_NN(swash);
12276 swash = _core_swash_init("utf8", name, &PL_sv_undef,
12279 NULL, /* No inversion list */
12282 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
12284 SvREFCNT_dec_NN(swash);
12288 /* Here didn't find it. It could be a user-defined
12289 * property that will be available at run-time. If we
12290 * accept only compile-time properties, is an error;
12291 * otherwise add it to the list for run-time look up */
12293 RExC_parse = e + 1;
12294 vFAIL3("Property '%.*s' is unknown", (int) n, name);
12296 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s\n",
12297 (value == 'p' ? '+' : '!'),
12299 has_user_defined_property = TRUE;
12301 /* We don't know yet, so have to assume that the
12302 * property could match something in the Latin1 range,
12303 * hence something that isn't utf8. Note that this
12304 * would cause things in <depends_list> to match
12305 * inappropriately, except that any \p{}, including
12306 * this one forces Unicode semantics, which means there
12307 * is <no depends_list> */
12308 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
12312 /* Here, did get the swash and its inversion list. If
12313 * the swash is from a user-defined property, then this
12314 * whole character class should be regarded as such */
12315 has_user_defined_property =
12317 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY);
12319 /* Invert if asking for the complement */
12320 if (value == 'P') {
12321 _invlist_union_complement_2nd(properties,
12325 /* The swash can't be used as-is, because we've
12326 * inverted things; delay removing it to here after
12327 * have copied its invlist above */
12328 SvREFCNT_dec_NN(swash);
12332 _invlist_union(properties, invlist, &properties);
12337 RExC_parse = e + 1;
12338 namedclass = ANYOF_UNIPROP; /* no official name, but it's
12341 /* \p means they want Unicode semantics */
12342 RExC_uni_semantics = 1;
12345 case 'n': value = '\n'; break;
12346 case 'r': value = '\r'; break;
12347 case 't': value = '\t'; break;
12348 case 'f': value = '\f'; break;
12349 case 'b': value = '\b'; break;
12350 case 'e': value = ASCII_TO_NATIVE('\033');break;
12351 case 'a': value = ASCII_TO_NATIVE('\007');break;
12353 RExC_parse--; /* function expects to be pointed at the 'o' */
12355 const char* error_msg;
12356 bool valid = grok_bslash_o(&RExC_parse,
12359 SIZE_ONLY, /* warnings in pass
12362 silence_non_portable,
12368 if (PL_encoding && value < 0x100) {
12369 goto recode_encoding;
12373 RExC_parse--; /* function expects to be pointed at the 'x' */
12375 const char* error_msg;
12376 bool valid = grok_bslash_x(&RExC_parse,
12379 TRUE, /* Output warnings */
12381 silence_non_portable,
12387 if (PL_encoding && value < 0x100)
12388 goto recode_encoding;
12391 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
12393 case '0': case '1': case '2': case '3': case '4':
12394 case '5': case '6': case '7':
12396 /* Take 1-3 octal digits */
12397 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12398 numlen = (strict) ? 4 : 3;
12399 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
12400 RExC_parse += numlen;
12403 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12404 vFAIL("Need exactly 3 octal digits");
12406 else if (! SIZE_ONLY /* like \08, \178 */
12408 && RExC_parse < RExC_end
12409 && isDIGIT(*RExC_parse)
12410 && ckWARN(WARN_REGEXP))
12412 SAVEFREESV(RExC_rx_sv);
12413 reg_warn_non_literal_string(
12415 form_short_octal_warning(RExC_parse, numlen));
12416 (void)ReREFCNT_inc(RExC_rx_sv);
12419 if (PL_encoding && value < 0x100)
12420 goto recode_encoding;
12424 if (! RExC_override_recoding) {
12425 SV* enc = PL_encoding;
12426 value = reg_recode((const char)(U8)value, &enc);
12429 vFAIL("Invalid escape in the specified encoding");
12431 else if (SIZE_ONLY) {
12432 ckWARNreg(RExC_parse,
12433 "Invalid escape in the specified encoding");
12439 /* Allow \_ to not give an error */
12440 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
12442 vFAIL2("Unrecognized escape \\%c in character class",
12446 SAVEFREESV(RExC_rx_sv);
12447 ckWARN2reg(RExC_parse,
12448 "Unrecognized escape \\%c in character class passed through",
12450 (void)ReREFCNT_inc(RExC_rx_sv);
12454 } /* End of switch on char following backslash */
12455 } /* end of handling backslash escape sequences */
12458 literal_endpoint++;
12461 /* Here, we have the current token in 'value' */
12463 /* What matches in a locale is not known until runtime. This includes
12464 * what the Posix classes (like \w, [:space:]) match. Room must be
12465 * reserved (one time per class) to store such classes, either if Perl
12466 * is compiled so that locale nodes always should have this space, or
12467 * if there is such class info to be stored. The space will contain a
12468 * bit for each named class that is to be matched against. This isn't
12469 * needed for \p{} and pseudo-classes, as they are not affected by
12470 * locale, and hence are dealt with separately */
12473 && (ANYOF_LOCALE == ANYOF_CLASS
12474 || (namedclass > OOB_NAMEDCLASS && namedclass < ANYOF_MAX)))
12478 RExC_size += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12481 RExC_emit += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12482 ANYOF_CLASS_ZERO(ret);
12484 ANYOF_FLAGS(ret) |= ANYOF_CLASS;
12487 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
12489 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
12490 * literal, as is the character that began the false range, i.e.
12491 * the 'a' in the examples */
12494 const int w = (RExC_parse >= rangebegin)
12495 ? RExC_parse - rangebegin
12498 vFAIL4("False [] range \"%*.*s\"", w, w, rangebegin);
12501 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
12502 ckWARN4reg(RExC_parse,
12503 "False [] range \"%*.*s\"",
12505 (void)ReREFCNT_inc(RExC_rx_sv);
12506 cp_list = add_cp_to_invlist(cp_list, '-');
12507 cp_list = add_cp_to_invlist(cp_list, prevvalue);
12511 range = 0; /* this was not a true range */
12512 element_count += 2; /* So counts for three values */
12516 U8 classnum = namedclass_to_classnum(namedclass);
12517 if (namedclass >= ANYOF_MAX) { /* If a special class */
12518 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
12520 /* Here, should be \h, \H, \v, or \V. Neither /d nor
12521 * /l make a difference in what these match. There
12522 * would be problems if these characters had folds
12523 * other than themselves, as cp_list is subject to
12525 if (classnum != _CC_VERTSPACE) {
12526 assert( namedclass == ANYOF_HORIZWS
12527 || namedclass == ANYOF_NHORIZWS);
12529 /* It turns out that \h is just a synonym for
12531 classnum = _CC_BLANK;
12534 _invlist_union_maybe_complement_2nd(
12536 PL_XPosix_ptrs[classnum],
12537 cBOOL(namedclass % 2), /* Complement if odd
12538 (NHORIZWS, NVERTWS)
12543 else if (classnum == _CC_ASCII) {
12546 ANYOF_CLASS_SET(ret, namedclass);
12549 #endif /* Not isascii(); just use the hard-coded definition for it */
12550 _invlist_union_maybe_complement_2nd(
12553 cBOOL(namedclass % 2), /* Complement if odd
12557 else { /* Garden variety class */
12559 /* The ascii range inversion list */
12560 SV* ascii_source = PL_Posix_ptrs[classnum];
12562 /* The full Latin1 range inversion list */
12563 SV* l1_source = PL_L1Posix_ptrs[classnum];
12565 /* This code is structured into two major clauses. The
12566 * first is for classes whose complete definitions may not
12567 * already be known. It not, the Latin1 definition
12568 * (guaranteed to already known) is used plus code is
12569 * generated to load the rest at run-time (only if needed).
12570 * If the complete definition is known, it drops down to
12571 * the second clause, where the complete definition is
12574 if (classnum < _FIRST_NON_SWASH_CC) {
12576 /* Here, the class has a swash, which may or not
12577 * already be loaded */
12579 /* The name of the property to use to match the full
12580 * eXtended Unicode range swash for this character
12582 const char *Xname = swash_property_names[classnum];
12584 /* If returning the inversion list, we can't defer
12585 * getting this until runtime */
12586 if (ret_invlist && ! PL_utf8_swash_ptrs[classnum]) {
12587 PL_utf8_swash_ptrs[classnum] =
12588 _core_swash_init("utf8", Xname, &PL_sv_undef,
12591 NULL, /* No inversion list */
12592 NULL /* No flags */
12594 assert(PL_utf8_swash_ptrs[classnum]);
12596 if ( ! PL_utf8_swash_ptrs[classnum]) {
12597 if (namedclass % 2 == 0) { /* A non-complemented
12599 /* If not /a matching, there are code points we
12600 * don't know at compile time. Arrange for the
12601 * unknown matches to be loaded at run-time, if
12603 if (! AT_LEAST_ASCII_RESTRICTED) {
12604 Perl_sv_catpvf(aTHX_ listsv, "+utf8::%s\n",
12607 if (LOC) { /* Under locale, set run-time
12609 ANYOF_CLASS_SET(ret, namedclass);
12612 /* Add the current class's code points to
12613 * the running total */
12614 _invlist_union(posixes,
12615 (AT_LEAST_ASCII_RESTRICTED)
12621 else { /* A complemented class */
12622 if (AT_LEAST_ASCII_RESTRICTED) {
12623 /* Under /a should match everything above
12624 * ASCII, plus the complement of the set's
12626 _invlist_union_complement_2nd(posixes,
12631 /* Arrange for the unknown matches to be
12632 * loaded at run-time, if needed */
12633 Perl_sv_catpvf(aTHX_ listsv, "!utf8::%s\n",
12635 runtime_posix_matches_above_Unicode = TRUE;
12637 ANYOF_CLASS_SET(ret, namedclass);
12641 /* We want to match everything in
12642 * Latin1, except those things that
12643 * l1_source matches */
12644 SV* scratch_list = NULL;
12645 _invlist_subtract(PL_Latin1, l1_source,
12648 /* Add the list from this class to the
12651 posixes = scratch_list;
12654 _invlist_union(posixes,
12657 SvREFCNT_dec_NN(scratch_list);
12659 if (DEPENDS_SEMANTICS) {
12661 |= ANYOF_NON_UTF8_LATIN1_ALL;
12666 goto namedclass_done;
12669 /* Here, there is a swash loaded for the class. If no
12670 * inversion list for it yet, get it */
12671 if (! PL_XPosix_ptrs[classnum]) {
12672 PL_XPosix_ptrs[classnum]
12673 = _swash_to_invlist(PL_utf8_swash_ptrs[classnum]);
12677 /* Here there is an inversion list already loaded for the
12680 if (namedclass % 2 == 0) { /* A non-complemented class,
12681 like ANYOF_PUNCT */
12683 /* For non-locale, just add it to any existing list
12685 _invlist_union(posixes,
12686 (AT_LEAST_ASCII_RESTRICTED)
12688 : PL_XPosix_ptrs[classnum],
12691 else { /* Locale */
12692 SV* scratch_list = NULL;
12694 /* For above Latin1 code points, we use the full
12696 _invlist_intersection(PL_AboveLatin1,
12697 PL_XPosix_ptrs[classnum],
12699 /* And set the output to it, adding instead if
12700 * there already is an output. Checking if
12701 * 'posixes' is NULL first saves an extra clone.
12702 * Its reference count will be decremented at the
12703 * next union, etc, or if this is the only
12704 * instance, at the end of the routine */
12706 posixes = scratch_list;
12709 _invlist_union(posixes, scratch_list, &posixes);
12710 SvREFCNT_dec_NN(scratch_list);
12713 #ifndef HAS_ISBLANK
12714 if (namedclass != ANYOF_BLANK) {
12716 /* Set this class in the node for runtime
12718 ANYOF_CLASS_SET(ret, namedclass);
12719 #ifndef HAS_ISBLANK
12722 /* No isblank(), use the hard-coded ASCII-range
12723 * blanks, adding them to the running total. */
12725 _invlist_union(posixes, ascii_source, &posixes);
12730 else { /* A complemented class, like ANYOF_NPUNCT */
12732 _invlist_union_complement_2nd(
12734 (AT_LEAST_ASCII_RESTRICTED)
12736 : PL_XPosix_ptrs[classnum],
12738 /* Under /d, everything in the upper half of the
12739 * Latin1 range matches this complement */
12740 if (DEPENDS_SEMANTICS) {
12741 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL;
12744 else { /* Locale */
12745 SV* scratch_list = NULL;
12746 _invlist_subtract(PL_AboveLatin1,
12747 PL_XPosix_ptrs[classnum],
12750 posixes = scratch_list;
12753 _invlist_union(posixes, scratch_list, &posixes);
12754 SvREFCNT_dec_NN(scratch_list);
12756 #ifndef HAS_ISBLANK
12757 if (namedclass != ANYOF_NBLANK) {
12759 ANYOF_CLASS_SET(ret, namedclass);
12760 #ifndef HAS_ISBLANK
12763 /* Get the list of all code points in Latin1
12764 * that are not ASCII blanks, and add them to
12765 * the running total */
12766 _invlist_subtract(PL_Latin1, ascii_source,
12768 _invlist_union(posixes, scratch_list, &posixes);
12769 SvREFCNT_dec_NN(scratch_list);
12776 continue; /* Go get next character */
12778 } /* end of namedclass \blah */
12780 /* Here, we have a single value. If 'range' is set, it is the ending
12781 * of a range--check its validity. Later, we will handle each
12782 * individual code point in the range. If 'range' isn't set, this
12783 * could be the beginning of a range, so check for that by looking
12784 * ahead to see if the next real character to be processed is the range
12785 * indicator--the minus sign */
12788 RExC_parse = regpatws(pRExC_state, RExC_parse,
12789 FALSE /* means don't recognize comments */);
12793 if (prevvalue > value) /* b-a */ {
12794 const int w = RExC_parse - rangebegin;
12795 Simple_vFAIL4("Invalid [] range \"%*.*s\"", w, w, rangebegin);
12796 range = 0; /* not a valid range */
12800 prevvalue = value; /* save the beginning of the potential range */
12801 if (! stop_at_1 /* Can't be a range if parsing just one thing */
12802 && *RExC_parse == '-')
12804 char* next_char_ptr = RExC_parse + 1;
12805 if (skip_white) { /* Get the next real char after the '-' */
12806 next_char_ptr = regpatws(pRExC_state,
12808 FALSE); /* means don't recognize
12812 /* If the '-' is at the end of the class (just before the ']',
12813 * it is a literal minus; otherwise it is a range */
12814 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
12815 RExC_parse = next_char_ptr;
12817 /* a bad range like \w-, [:word:]- ? */
12818 if (namedclass > OOB_NAMEDCLASS) {
12819 if (strict || ckWARN(WARN_REGEXP)) {
12821 RExC_parse >= rangebegin ?
12822 RExC_parse - rangebegin : 0;
12824 vFAIL4("False [] range \"%*.*s\"",
12829 "False [] range \"%*.*s\"",
12834 cp_list = add_cp_to_invlist(cp_list, '-');
12838 range = 1; /* yeah, it's a range! */
12839 continue; /* but do it the next time */
12844 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
12847 /* non-Latin1 code point implies unicode semantics. Must be set in
12848 * pass1 so is there for the whole of pass 2 */
12850 RExC_uni_semantics = 1;
12853 /* Ready to process either the single value, or the completed range.
12854 * For single-valued non-inverted ranges, we consider the possibility
12855 * of multi-char folds. (We made a conscious decision to not do this
12856 * for the other cases because it can often lead to non-intuitive
12857 * results. For example, you have the peculiar case that:
12858 * "s s" =~ /^[^\xDF]+$/i => Y
12859 * "ss" =~ /^[^\xDF]+$/i => N
12861 * See [perl #89750] */
12862 if (FOLD && allow_multi_folds && value == prevvalue) {
12863 if (value == LATIN_SMALL_LETTER_SHARP_S
12864 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
12867 /* Here <value> is indeed a multi-char fold. Get what it is */
12869 U8 foldbuf[UTF8_MAXBYTES_CASE];
12872 UV folded = _to_uni_fold_flags(
12877 | ((LOC) ? FOLD_FLAGS_LOCALE
12878 : (ASCII_FOLD_RESTRICTED)
12879 ? FOLD_FLAGS_NOMIX_ASCII
12883 /* Here, <folded> should be the first character of the
12884 * multi-char fold of <value>, with <foldbuf> containing the
12885 * whole thing. But, if this fold is not allowed (because of
12886 * the flags), <fold> will be the same as <value>, and should
12887 * be processed like any other character, so skip the special
12889 if (folded != value) {
12891 /* Skip if we are recursed, currently parsing the class
12892 * again. Otherwise add this character to the list of
12893 * multi-char folds. */
12894 if (! RExC_in_multi_char_class) {
12895 AV** this_array_ptr;
12897 STRLEN cp_count = utf8_length(foldbuf,
12898 foldbuf + foldlen);
12899 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
12901 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
12904 if (! multi_char_matches) {
12905 multi_char_matches = newAV();
12908 /* <multi_char_matches> is actually an array of arrays.
12909 * There will be one or two top-level elements: [2],
12910 * and/or [3]. The [2] element is an array, each
12911 * element thereof is a character which folds to two
12912 * characters; likewise for [3]. (Unicode guarantees a
12913 * maximum of 3 characters in any fold.) When we
12914 * rewrite the character class below, we will do so
12915 * such that the longest folds are written first, so
12916 * that it prefers the longest matching strings first.
12917 * This is done even if it turns out that any
12918 * quantifier is non-greedy, out of programmer
12919 * laziness. Tom Christiansen has agreed that this is
12920 * ok. This makes the test for the ligature 'ffi' come
12921 * before the test for 'ff' */
12922 if (av_exists(multi_char_matches, cp_count)) {
12923 this_array_ptr = (AV**) av_fetch(multi_char_matches,
12925 this_array = *this_array_ptr;
12928 this_array = newAV();
12929 av_store(multi_char_matches, cp_count,
12932 av_push(this_array, multi_fold);
12935 /* This element should not be processed further in this
12938 value = save_value;
12939 prevvalue = save_prevvalue;
12945 /* Deal with this element of the class */
12948 cp_list = _add_range_to_invlist(cp_list, prevvalue, value);
12950 SV* this_range = _new_invlist(1);
12951 _append_range_to_invlist(this_range, prevvalue, value);
12953 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
12954 * If this range was specified using something like 'i-j', we want
12955 * to include only the 'i' and the 'j', and not anything in
12956 * between, so exclude non-ASCII, non-alphabetics from it.
12957 * However, if the range was specified with something like
12958 * [\x89-\x91] or [\x89-j], all code points within it should be
12959 * included. literal_endpoint==2 means both ends of the range used
12960 * a literal character, not \x{foo} */
12961 if (literal_endpoint == 2
12962 && (prevvalue >= 'a' && value <= 'z')
12963 || (prevvalue >= 'A' && value <= 'Z'))
12965 _invlist_intersection(this_range, PL_Posix_ptrs[_CC_ALPHA],
12968 _invlist_union(cp_list, this_range, &cp_list);
12969 literal_endpoint = 0;
12973 range = 0; /* this range (if it was one) is done now */
12974 } /* End of loop through all the text within the brackets */
12976 /* If anything in the class expands to more than one character, we have to
12977 * deal with them by building up a substitute parse string, and recursively
12978 * calling reg() on it, instead of proceeding */
12979 if (multi_char_matches) {
12980 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
12983 char *save_end = RExC_end;
12984 char *save_parse = RExC_parse;
12985 bool first_time = TRUE; /* First multi-char occurrence doesn't get
12990 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
12991 because too confusing */
12993 sv_catpv(substitute_parse, "(?:");
12997 /* Look at the longest folds first */
12998 for (cp_count = av_len(multi_char_matches); cp_count > 0; cp_count--) {
13000 if (av_exists(multi_char_matches, cp_count)) {
13001 AV** this_array_ptr;
13004 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13006 while ((this_sequence = av_pop(*this_array_ptr)) !=
13009 if (! first_time) {
13010 sv_catpv(substitute_parse, "|");
13012 first_time = FALSE;
13014 sv_catpv(substitute_parse, SvPVX(this_sequence));
13019 /* If the character class contains anything else besides these
13020 * multi-character folds, have to include it in recursive parsing */
13021 if (element_count) {
13022 sv_catpv(substitute_parse, "|[");
13023 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
13024 sv_catpv(substitute_parse, "]");
13027 sv_catpv(substitute_parse, ")");
13030 /* This is a way to get the parse to skip forward a whole named
13031 * sequence instead of matching the 2nd character when it fails the
13033 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
13037 RExC_parse = SvPV(substitute_parse, len);
13038 RExC_end = RExC_parse + len;
13039 RExC_in_multi_char_class = 1;
13040 RExC_emit = (regnode *)orig_emit;
13042 ret = reg(pRExC_state, 1, ®_flags, depth+1);
13044 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
13046 RExC_parse = save_parse;
13047 RExC_end = save_end;
13048 RExC_in_multi_char_class = 0;
13049 SvREFCNT_dec_NN(multi_char_matches);
13053 /* If the character class contains only a single element, it may be
13054 * optimizable into another node type which is smaller and runs faster.
13055 * Check if this is the case for this class */
13056 if (element_count == 1 && ! ret_invlist) {
13060 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like \w or
13061 [:digit:] or \p{foo} */
13063 /* All named classes are mapped into POSIXish nodes, with its FLAG
13064 * argument giving which class it is */
13065 switch ((I32)namedclass) {
13066 case ANYOF_UNIPROP:
13069 /* These don't depend on the charset modifiers. They always
13070 * match under /u rules */
13071 case ANYOF_NHORIZWS:
13072 case ANYOF_HORIZWS:
13073 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
13076 case ANYOF_NVERTWS:
13081 /* The actual POSIXish node for all the rest depends on the
13082 * charset modifier. The ones in the first set depend only on
13083 * ASCII or, if available on this platform, locale */
13087 op = (LOC) ? POSIXL : POSIXA;
13098 /* under /a could be alpha */
13100 if (ASCII_RESTRICTED) {
13101 namedclass = ANYOF_ALPHA + (namedclass % 2);
13109 /* The rest have more possibilities depending on the charset.
13110 * We take advantage of the enum ordering of the charset
13111 * modifiers to get the exact node type, */
13113 op = POSIXD + get_regex_charset(RExC_flags);
13114 if (op > POSIXA) { /* /aa is same as /a */
13117 #ifndef HAS_ISBLANK
13119 && (namedclass == ANYOF_BLANK
13120 || namedclass == ANYOF_NBLANK))
13127 /* The odd numbered ones are the complements of the
13128 * next-lower even number one */
13129 if (namedclass % 2 == 1) {
13133 arg = namedclass_to_classnum(namedclass);
13137 else if (value == prevvalue) {
13139 /* Here, the class consists of just a single code point */
13142 if (! LOC && value == '\n') {
13143 op = REG_ANY; /* Optimize [^\n] */
13144 *flagp |= HASWIDTH|SIMPLE;
13148 else if (value < 256 || UTF) {
13150 /* Optimize a single value into an EXACTish node, but not if it
13151 * would require converting the pattern to UTF-8. */
13152 op = compute_EXACTish(pRExC_state);
13154 } /* Otherwise is a range */
13155 else if (! LOC) { /* locale could vary these */
13156 if (prevvalue == '0') {
13157 if (value == '9') {
13164 /* Here, we have changed <op> away from its initial value iff we found
13165 * an optimization */
13168 /* Throw away this ANYOF regnode, and emit the calculated one,
13169 * which should correspond to the beginning, not current, state of
13171 const char * cur_parse = RExC_parse;
13172 RExC_parse = (char *)orig_parse;
13176 /* To get locale nodes to not use the full ANYOF size would
13177 * require moving the code above that writes the portions
13178 * of it that aren't in other nodes to after this point.
13179 * e.g. ANYOF_CLASS_SET */
13180 RExC_size = orig_size;
13184 RExC_emit = (regnode *)orig_emit;
13185 if (PL_regkind[op] == POSIXD) {
13187 op += NPOSIXD - POSIXD;
13192 ret = reg_node(pRExC_state, op);
13194 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
13198 *flagp |= HASWIDTH|SIMPLE;
13200 else if (PL_regkind[op] == EXACT) {
13201 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13204 RExC_parse = (char *) cur_parse;
13206 SvREFCNT_dec(posixes);
13207 SvREFCNT_dec(cp_list);
13214 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
13216 /* If folding, we calculate all characters that could fold to or from the
13217 * ones already on the list */
13218 if (FOLD && cp_list) {
13219 UV start, end; /* End points of code point ranges */
13221 SV* fold_intersection = NULL;
13223 /* If the highest code point is within Latin1, we can use the
13224 * compiled-in Alphas list, and not have to go out to disk. This
13225 * yields two false positives, the masculine and feminine ordinal
13226 * indicators, which are weeded out below using the
13227 * IS_IN_SOME_FOLD_L1() macro */
13228 if (invlist_highest(cp_list) < 256) {
13229 _invlist_intersection(PL_L1Posix_ptrs[_CC_ALPHA], cp_list,
13230 &fold_intersection);
13234 /* Here, there are non-Latin1 code points, so we will have to go
13235 * fetch the list of all the characters that participate in folds
13237 if (! PL_utf8_foldable) {
13238 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13239 &PL_sv_undef, 1, 0);
13240 PL_utf8_foldable = _get_swash_invlist(swash);
13241 SvREFCNT_dec_NN(swash);
13244 /* This is a hash that for a particular fold gives all characters
13245 * that are involved in it */
13246 if (! PL_utf8_foldclosures) {
13248 /* If we were unable to find any folds, then we likely won't be
13249 * able to find the closures. So just create an empty list.
13250 * Folding will effectively be restricted to the non-Unicode
13251 * rules hard-coded into Perl. (This case happens legitimately
13252 * during compilation of Perl itself before the Unicode tables
13253 * are generated) */
13254 if (_invlist_len(PL_utf8_foldable) == 0) {
13255 PL_utf8_foldclosures = newHV();
13258 /* If the folds haven't been read in, call a fold function
13260 if (! PL_utf8_tofold) {
13261 U8 dummy[UTF8_MAXBYTES+1];
13263 /* This string is just a short named one above \xff */
13264 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
13265 assert(PL_utf8_tofold); /* Verify that worked */
13267 PL_utf8_foldclosures =
13268 _swash_inversion_hash(PL_utf8_tofold);
13272 /* Only the characters in this class that participate in folds need
13273 * be checked. Get the intersection of this class and all the
13274 * possible characters that are foldable. This can quickly narrow
13275 * down a large class */
13276 _invlist_intersection(PL_utf8_foldable, cp_list,
13277 &fold_intersection);
13280 /* Now look at the foldable characters in this class individually */
13281 invlist_iterinit(fold_intersection);
13282 while (invlist_iternext(fold_intersection, &start, &end)) {
13285 /* Locale folding for Latin1 characters is deferred until runtime */
13286 if (LOC && start < 256) {
13290 /* Look at every character in the range */
13291 for (j = start; j <= end; j++) {
13293 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
13299 /* We have the latin1 folding rules hard-coded here so that
13300 * an innocent-looking character class, like /[ks]/i won't
13301 * have to go out to disk to find the possible matches.
13302 * XXX It would be better to generate these via regen, in
13303 * case a new version of the Unicode standard adds new
13304 * mappings, though that is not really likely, and may be
13305 * caught by the default: case of the switch below. */
13307 if (IS_IN_SOME_FOLD_L1(j)) {
13309 /* ASCII is always matched; non-ASCII is matched only
13310 * under Unicode rules */
13311 if (isASCII(j) || AT_LEAST_UNI_SEMANTICS) {
13313 add_cp_to_invlist(cp_list, PL_fold_latin1[j]);
13317 add_cp_to_invlist(depends_list, PL_fold_latin1[j]);
13321 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
13322 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
13324 /* Certain Latin1 characters have matches outside
13325 * Latin1. To get here, <j> is one of those
13326 * characters. None of these matches is valid for
13327 * ASCII characters under /aa, which is why the 'if'
13328 * just above excludes those. These matches only
13329 * happen when the target string is utf8. The code
13330 * below adds the single fold closures for <j> to the
13331 * inversion list. */
13336 add_cp_to_invlist(cp_list, KELVIN_SIGN);
13340 cp_list = add_cp_to_invlist(cp_list,
13341 LATIN_SMALL_LETTER_LONG_S);
13344 cp_list = add_cp_to_invlist(cp_list,
13345 GREEK_CAPITAL_LETTER_MU);
13346 cp_list = add_cp_to_invlist(cp_list,
13347 GREEK_SMALL_LETTER_MU);
13349 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
13350 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
13352 add_cp_to_invlist(cp_list, ANGSTROM_SIGN);
13354 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
13355 cp_list = add_cp_to_invlist(cp_list,
13356 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
13358 case LATIN_SMALL_LETTER_SHARP_S:
13359 cp_list = add_cp_to_invlist(cp_list,
13360 LATIN_CAPITAL_LETTER_SHARP_S);
13362 case 'F': case 'f':
13363 case 'I': case 'i':
13364 case 'L': case 'l':
13365 case 'T': case 't':
13366 case 'A': case 'a':
13367 case 'H': case 'h':
13368 case 'J': case 'j':
13369 case 'N': case 'n':
13370 case 'W': case 'w':
13371 case 'Y': case 'y':
13372 /* These all are targets of multi-character
13373 * folds from code points that require UTF8 to
13374 * express, so they can't match unless the
13375 * target string is in UTF-8, so no action here
13376 * is necessary, as regexec.c properly handles
13377 * the general case for UTF-8 matching and
13378 * multi-char folds */
13381 /* Use deprecated warning to increase the
13382 * chances of this being output */
13383 ckWARN2regdep(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
13390 /* Here is an above Latin1 character. We don't have the rules
13391 * hard-coded for it. First, get its fold. This is the simple
13392 * fold, as the multi-character folds have been handled earlier
13393 * and separated out */
13394 _to_uni_fold_flags(j, foldbuf, &foldlen,
13396 ? FOLD_FLAGS_LOCALE
13397 : (ASCII_FOLD_RESTRICTED)
13398 ? FOLD_FLAGS_NOMIX_ASCII
13401 /* Single character fold of above Latin1. Add everything in
13402 * its fold closure to the list that this node should match.
13403 * The fold closures data structure is a hash with the keys
13404 * being the UTF-8 of every character that is folded to, like
13405 * 'k', and the values each an array of all code points that
13406 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
13407 * Multi-character folds are not included */
13408 if ((listp = hv_fetch(PL_utf8_foldclosures,
13409 (char *) foldbuf, foldlen, FALSE)))
13411 AV* list = (AV*) *listp;
13413 for (k = 0; k <= av_len(list); k++) {
13414 SV** c_p = av_fetch(list, k, FALSE);
13417 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
13421 /* /aa doesn't allow folds between ASCII and non-; /l
13422 * doesn't allow them between above and below 256 */
13423 if ((ASCII_FOLD_RESTRICTED
13424 && (isASCII(c) != isASCII(j)))
13425 || (LOC && ((c < 256) != (j < 256))))
13430 /* Folds involving non-ascii Latin1 characters
13431 * under /d are added to a separate list */
13432 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
13434 cp_list = add_cp_to_invlist(cp_list, c);
13437 depends_list = add_cp_to_invlist(depends_list, c);
13443 SvREFCNT_dec_NN(fold_intersection);
13446 /* And combine the result (if any) with any inversion list from posix
13447 * classes. The lists are kept separate up to now because we don't want to
13448 * fold the classes (folding of those is automatically handled by the swash
13449 * fetching code) */
13451 if (! DEPENDS_SEMANTICS) {
13453 _invlist_union(cp_list, posixes, &cp_list);
13454 SvREFCNT_dec_NN(posixes);
13461 /* Under /d, we put into a separate list the Latin1 things that
13462 * match only when the target string is utf8 */
13463 SV* nonascii_but_latin1_properties = NULL;
13464 _invlist_intersection(posixes, PL_Latin1,
13465 &nonascii_but_latin1_properties);
13466 _invlist_subtract(nonascii_but_latin1_properties, PL_ASCII,
13467 &nonascii_but_latin1_properties);
13468 _invlist_subtract(posixes, nonascii_but_latin1_properties,
13471 _invlist_union(cp_list, posixes, &cp_list);
13472 SvREFCNT_dec_NN(posixes);
13478 if (depends_list) {
13479 _invlist_union(depends_list, nonascii_but_latin1_properties,
13481 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
13484 depends_list = nonascii_but_latin1_properties;
13489 /* And combine the result (if any) with any inversion list from properties.
13490 * The lists are kept separate up to now so that we can distinguish the two
13491 * in regards to matching above-Unicode. A run-time warning is generated
13492 * if a Unicode property is matched against a non-Unicode code point. But,
13493 * we allow user-defined properties to match anything, without any warning,
13494 * and we also suppress the warning if there is a portion of the character
13495 * class that isn't a Unicode property, and which matches above Unicode, \W
13496 * or [\x{110000}] for example.
13497 * (Note that in this case, unlike the Posix one above, there is no
13498 * <depends_list>, because having a Unicode property forces Unicode
13501 bool warn_super = ! has_user_defined_property;
13504 /* If it matters to the final outcome, see if a non-property
13505 * component of the class matches above Unicode. If so, the
13506 * warning gets suppressed. This is true even if just a single
13507 * such code point is specified, as though not strictly correct if
13508 * another such code point is matched against, the fact that they
13509 * are using above-Unicode code points indicates they should know
13510 * the issues involved */
13512 bool non_prop_matches_above_Unicode =
13513 runtime_posix_matches_above_Unicode
13514 | (invlist_highest(cp_list) > PERL_UNICODE_MAX);
13516 non_prop_matches_above_Unicode =
13517 ! non_prop_matches_above_Unicode;
13519 warn_super = ! non_prop_matches_above_Unicode;
13522 _invlist_union(properties, cp_list, &cp_list);
13523 SvREFCNT_dec_NN(properties);
13526 cp_list = properties;
13530 OP(ret) = ANYOF_WARN_SUPER;
13534 /* Here, we have calculated what code points should be in the character
13537 * Now we can see about various optimizations. Fold calculation (which we
13538 * did above) needs to take place before inversion. Otherwise /[^k]/i
13539 * would invert to include K, which under /i would match k, which it
13540 * shouldn't. Therefore we can't invert folded locale now, as it won't be
13541 * folded until runtime */
13543 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
13544 * at compile time. Besides not inverting folded locale now, we can't
13545 * invert if there are things such as \w, which aren't known until runtime
13548 && ! (LOC && (FOLD || (ANYOF_FLAGS(ret) & ANYOF_CLASS)))
13550 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13552 _invlist_invert(cp_list);
13554 /* Any swash can't be used as-is, because we've inverted things */
13556 SvREFCNT_dec_NN(swash);
13560 /* Clear the invert flag since have just done it here */
13565 *ret_invlist = cp_list;
13567 /* Discard the generated node */
13569 RExC_size = orig_size;
13572 RExC_emit = orig_emit;
13577 /* If we didn't do folding, it's because some information isn't available
13578 * until runtime; set the run-time fold flag for these. (We don't have to
13579 * worry about properties folding, as that is taken care of by the swash
13583 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
13586 /* Some character classes are equivalent to other nodes. Such nodes take
13587 * up less room and generally fewer operations to execute than ANYOF nodes.
13588 * Above, we checked for and optimized into some such equivalents for
13589 * certain common classes that are easy to test. Getting to this point in
13590 * the code means that the class didn't get optimized there. Since this
13591 * code is only executed in Pass 2, it is too late to save space--it has
13592 * been allocated in Pass 1, and currently isn't given back. But turning
13593 * things into an EXACTish node can allow the optimizer to join it to any
13594 * adjacent such nodes. And if the class is equivalent to things like /./,
13595 * expensive run-time swashes can be avoided. Now that we have more
13596 * complete information, we can find things necessarily missed by the
13597 * earlier code. I (khw) am not sure how much to look for here. It would
13598 * be easy, but perhaps too slow, to check any candidates against all the
13599 * node types they could possibly match using _invlistEQ(). */
13604 && ! (ANYOF_FLAGS(ret) & ANYOF_CLASS)
13605 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13608 U8 op = END; /* The optimzation node-type */
13609 const char * cur_parse= RExC_parse;
13611 invlist_iterinit(cp_list);
13612 if (! invlist_iternext(cp_list, &start, &end)) {
13614 /* Here, the list is empty. This happens, for example, when a
13615 * Unicode property is the only thing in the character class, and
13616 * it doesn't match anything. (perluniprops.pod notes such
13619 *flagp |= HASWIDTH|SIMPLE;
13621 else if (start == end) { /* The range is a single code point */
13622 if (! invlist_iternext(cp_list, &start, &end)
13624 /* Don't do this optimization if it would require changing
13625 * the pattern to UTF-8 */
13626 && (start < 256 || UTF))
13628 /* Here, the list contains a single code point. Can optimize
13629 * into an EXACT node */
13638 /* A locale node under folding with one code point can be
13639 * an EXACTFL, as its fold won't be calculated until
13645 /* Here, we are generally folding, but there is only one
13646 * code point to match. If we have to, we use an EXACT
13647 * node, but it would be better for joining with adjacent
13648 * nodes in the optimization pass if we used the same
13649 * EXACTFish node that any such are likely to be. We can
13650 * do this iff the code point doesn't participate in any
13651 * folds. For example, an EXACTF of a colon is the same as
13652 * an EXACT one, since nothing folds to or from a colon. */
13654 if (IS_IN_SOME_FOLD_L1(value)) {
13659 if (! PL_utf8_foldable) {
13660 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13661 &PL_sv_undef, 1, 0);
13662 PL_utf8_foldable = _get_swash_invlist(swash);
13663 SvREFCNT_dec_NN(swash);
13665 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
13670 /* If we haven't found the node type, above, it means we
13671 * can use the prevailing one */
13673 op = compute_EXACTish(pRExC_state);
13678 else if (start == 0) {
13679 if (end == UV_MAX) {
13681 *flagp |= HASWIDTH|SIMPLE;
13684 else if (end == '\n' - 1
13685 && invlist_iternext(cp_list, &start, &end)
13686 && start == '\n' + 1 && end == UV_MAX)
13689 *flagp |= HASWIDTH|SIMPLE;
13693 invlist_iterfinish(cp_list);
13696 RExC_parse = (char *)orig_parse;
13697 RExC_emit = (regnode *)orig_emit;
13699 ret = reg_node(pRExC_state, op);
13701 RExC_parse = (char *)cur_parse;
13703 if (PL_regkind[op] == EXACT) {
13704 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13707 SvREFCNT_dec_NN(cp_list);
13712 /* Here, <cp_list> contains all the code points we can determine at
13713 * compile time that match under all conditions. Go through it, and
13714 * for things that belong in the bitmap, put them there, and delete from
13715 * <cp_list>. While we are at it, see if everything above 255 is in the
13716 * list, and if so, set a flag to speed up execution */
13717 ANYOF_BITMAP_ZERO(ret);
13720 /* This gets set if we actually need to modify things */
13721 bool change_invlist = FALSE;
13725 /* Start looking through <cp_list> */
13726 invlist_iterinit(cp_list);
13727 while (invlist_iternext(cp_list, &start, &end)) {
13731 if (end == UV_MAX && start <= 256) {
13732 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
13735 /* Quit if are above what we should change */
13740 change_invlist = TRUE;
13742 /* Set all the bits in the range, up to the max that we are doing */
13743 high = (end < 255) ? end : 255;
13744 for (i = start; i <= (int) high; i++) {
13745 if (! ANYOF_BITMAP_TEST(ret, i)) {
13746 ANYOF_BITMAP_SET(ret, i);
13752 invlist_iterfinish(cp_list);
13754 /* Done with loop; remove any code points that are in the bitmap from
13756 if (change_invlist) {
13757 _invlist_subtract(cp_list, PL_Latin1, &cp_list);
13760 /* If have completely emptied it, remove it completely */
13761 if (_invlist_len(cp_list) == 0) {
13762 SvREFCNT_dec_NN(cp_list);
13768 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
13771 /* Here, the bitmap has been populated with all the Latin1 code points that
13772 * always match. Can now add to the overall list those that match only
13773 * when the target string is UTF-8 (<depends_list>). */
13774 if (depends_list) {
13776 _invlist_union(cp_list, depends_list, &cp_list);
13777 SvREFCNT_dec_NN(depends_list);
13780 cp_list = depends_list;
13784 /* If there is a swash and more than one element, we can't use the swash in
13785 * the optimization below. */
13786 if (swash && element_count > 1) {
13787 SvREFCNT_dec_NN(swash);
13792 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13794 ARG_SET(ret, ANYOF_NONBITMAP_EMPTY);
13797 /* av[0] stores the character class description in its textual form:
13798 * used later (regexec.c:Perl_regclass_swash()) to initialize the
13799 * appropriate swash, and is also useful for dumping the regnode.
13800 * av[1] if NULL, is a placeholder to later contain the swash computed
13801 * from av[0]. But if no further computation need be done, the
13802 * swash is stored there now.
13803 * av[2] stores the cp_list inversion list for use in addition or
13804 * instead of av[0]; used only if av[1] is NULL
13805 * av[3] is set if any component of the class is from a user-defined
13806 * property; used only if av[1] is NULL */
13807 AV * const av = newAV();
13810 av_store(av, 0, (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13811 ? SvREFCNT_inc(listsv) : &PL_sv_undef);
13813 av_store(av, 1, swash);
13814 SvREFCNT_dec_NN(cp_list);
13817 av_store(av, 1, NULL);
13819 av_store(av, 2, cp_list);
13820 av_store(av, 3, newSVuv(has_user_defined_property));
13824 rv = newRV_noinc(MUTABLE_SV(av));
13825 n = add_data(pRExC_state, 1, "s");
13826 RExC_rxi->data->data[n] = (void*)rv;
13830 *flagp |= HASWIDTH|SIMPLE;
13833 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
13836 /* reg_skipcomment()
13838 Absorbs an /x style # comments from the input stream.
13839 Returns true if there is more text remaining in the stream.
13840 Will set the REG_SEEN_RUN_ON_COMMENT flag if the comment
13841 terminates the pattern without including a newline.
13843 Note its the callers responsibility to ensure that we are
13844 actually in /x mode
13849 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
13853 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
13855 while (RExC_parse < RExC_end)
13856 if (*RExC_parse++ == '\n') {
13861 /* we ran off the end of the pattern without ending
13862 the comment, so we have to add an \n when wrapping */
13863 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
13871 Advances the parse position, and optionally absorbs
13872 "whitespace" from the inputstream.
13874 Without /x "whitespace" means (?#...) style comments only,
13875 with /x this means (?#...) and # comments and whitespace proper.
13877 Returns the RExC_parse point from BEFORE the scan occurs.
13879 This is the /x friendly way of saying RExC_parse++.
13883 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
13885 char* const retval = RExC_parse++;
13887 PERL_ARGS_ASSERT_NEXTCHAR;
13890 if (RExC_end - RExC_parse >= 3
13891 && *RExC_parse == '('
13892 && RExC_parse[1] == '?'
13893 && RExC_parse[2] == '#')
13895 while (*RExC_parse != ')') {
13896 if (RExC_parse == RExC_end)
13897 FAIL("Sequence (?#... not terminated");
13903 if (RExC_flags & RXf_PMf_EXTENDED) {
13904 if (isSPACE(*RExC_parse)) {
13908 else if (*RExC_parse == '#') {
13909 if ( reg_skipcomment( pRExC_state ) )
13918 - reg_node - emit a node
13920 STATIC regnode * /* Location. */
13921 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
13925 regnode * const ret = RExC_emit;
13926 GET_RE_DEBUG_FLAGS_DECL;
13928 PERL_ARGS_ASSERT_REG_NODE;
13931 SIZE_ALIGN(RExC_size);
13935 if (RExC_emit >= RExC_emit_bound)
13936 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
13937 op, RExC_emit, RExC_emit_bound);
13939 NODE_ALIGN_FILL(ret);
13941 FILL_ADVANCE_NODE(ptr, op);
13942 #ifdef RE_TRACK_PATTERN_OFFSETS
13943 if (RExC_offsets) { /* MJD */
13944 MJD_OFFSET_DEBUG(("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
13945 "reg_node", __LINE__,
13947 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
13948 ? "Overwriting end of array!\n" : "OK",
13949 (UV)(RExC_emit - RExC_emit_start),
13950 (UV)(RExC_parse - RExC_start),
13951 (UV)RExC_offsets[0]));
13952 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
13960 - reganode - emit a node with an argument
13962 STATIC regnode * /* Location. */
13963 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
13967 regnode * const ret = RExC_emit;
13968 GET_RE_DEBUG_FLAGS_DECL;
13970 PERL_ARGS_ASSERT_REGANODE;
13973 SIZE_ALIGN(RExC_size);
13978 assert(2==regarglen[op]+1);
13980 Anything larger than this has to allocate the extra amount.
13981 If we changed this to be:
13983 RExC_size += (1 + regarglen[op]);
13985 then it wouldn't matter. Its not clear what side effect
13986 might come from that so its not done so far.
13991 if (RExC_emit >= RExC_emit_bound)
13992 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
13993 op, RExC_emit, RExC_emit_bound);
13995 NODE_ALIGN_FILL(ret);
13997 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
13998 #ifdef RE_TRACK_PATTERN_OFFSETS
13999 if (RExC_offsets) { /* MJD */
14000 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14004 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
14005 "Overwriting end of array!\n" : "OK",
14006 (UV)(RExC_emit - RExC_emit_start),
14007 (UV)(RExC_parse - RExC_start),
14008 (UV)RExC_offsets[0]));
14009 Set_Cur_Node_Offset;
14017 - reguni - emit (if appropriate) a Unicode character
14020 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
14024 PERL_ARGS_ASSERT_REGUNI;
14026 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
14030 - reginsert - insert an operator in front of already-emitted operand
14032 * Means relocating the operand.
14035 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
14041 const int offset = regarglen[(U8)op];
14042 const int size = NODE_STEP_REGNODE + offset;
14043 GET_RE_DEBUG_FLAGS_DECL;
14045 PERL_ARGS_ASSERT_REGINSERT;
14046 PERL_UNUSED_ARG(depth);
14047 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
14048 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
14057 if (RExC_open_parens) {
14059 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
14060 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
14061 if ( RExC_open_parens[paren] >= opnd ) {
14062 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
14063 RExC_open_parens[paren] += size;
14065 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
14067 if ( RExC_close_parens[paren] >= opnd ) {
14068 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
14069 RExC_close_parens[paren] += size;
14071 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
14076 while (src > opnd) {
14077 StructCopy(--src, --dst, regnode);
14078 #ifdef RE_TRACK_PATTERN_OFFSETS
14079 if (RExC_offsets) { /* MJD 20010112 */
14080 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
14084 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
14085 ? "Overwriting end of array!\n" : "OK",
14086 (UV)(src - RExC_emit_start),
14087 (UV)(dst - RExC_emit_start),
14088 (UV)RExC_offsets[0]));
14089 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
14090 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
14096 place = opnd; /* Op node, where operand used to be. */
14097 #ifdef RE_TRACK_PATTERN_OFFSETS
14098 if (RExC_offsets) { /* MJD */
14099 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14103 (UV)(place - RExC_emit_start) > RExC_offsets[0]
14104 ? "Overwriting end of array!\n" : "OK",
14105 (UV)(place - RExC_emit_start),
14106 (UV)(RExC_parse - RExC_start),
14107 (UV)RExC_offsets[0]));
14108 Set_Node_Offset(place, RExC_parse);
14109 Set_Node_Length(place, 1);
14112 src = NEXTOPER(place);
14113 FILL_ADVANCE_NODE(place, op);
14114 Zero(src, offset, regnode);
14118 - regtail - set the next-pointer at the end of a node chain of p to val.
14119 - SEE ALSO: regtail_study
14121 /* TODO: All three parms should be const */
14123 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14127 GET_RE_DEBUG_FLAGS_DECL;
14129 PERL_ARGS_ASSERT_REGTAIL;
14131 PERL_UNUSED_ARG(depth);
14137 /* Find last node. */
14140 regnode * const temp = regnext(scan);
14142 SV * const mysv=sv_newmortal();
14143 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
14144 regprop(RExC_rx, mysv, scan);
14145 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
14146 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
14147 (temp == NULL ? "->" : ""),
14148 (temp == NULL ? PL_reg_name[OP(val)] : "")
14156 if (reg_off_by_arg[OP(scan)]) {
14157 ARG_SET(scan, val - scan);
14160 NEXT_OFF(scan) = val - scan;
14166 - regtail_study - set the next-pointer at the end of a node chain of p to val.
14167 - Look for optimizable sequences at the same time.
14168 - currently only looks for EXACT chains.
14170 This is experimental code. The idea is to use this routine to perform
14171 in place optimizations on branches and groups as they are constructed,
14172 with the long term intention of removing optimization from study_chunk so
14173 that it is purely analytical.
14175 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
14176 to control which is which.
14179 /* TODO: All four parms should be const */
14182 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14187 #ifdef EXPERIMENTAL_INPLACESCAN
14190 GET_RE_DEBUG_FLAGS_DECL;
14192 PERL_ARGS_ASSERT_REGTAIL_STUDY;
14198 /* Find last node. */
14202 regnode * const temp = regnext(scan);
14203 #ifdef EXPERIMENTAL_INPLACESCAN
14204 if (PL_regkind[OP(scan)] == EXACT) {
14205 bool has_exactf_sharp_s; /* Unexamined in this routine */
14206 if (join_exact(pRExC_state,scan,&min, &has_exactf_sharp_s, 1,val,depth+1))
14211 switch (OP(scan)) {
14217 case EXACTFU_TRICKYFOLD:
14219 if( exact == PSEUDO )
14221 else if ( exact != OP(scan) )
14230 SV * const mysv=sv_newmortal();
14231 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
14232 regprop(RExC_rx, mysv, scan);
14233 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
14234 SvPV_nolen_const(mysv),
14235 REG_NODE_NUM(scan),
14236 PL_reg_name[exact]);
14243 SV * const mysv_val=sv_newmortal();
14244 DEBUG_PARSE_MSG("");
14245 regprop(RExC_rx, mysv_val, val);
14246 PerlIO_printf(Perl_debug_log, "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
14247 SvPV_nolen_const(mysv_val),
14248 (IV)REG_NODE_NUM(val),
14252 if (reg_off_by_arg[OP(scan)]) {
14253 ARG_SET(scan, val - scan);
14256 NEXT_OFF(scan) = val - scan;
14264 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
14268 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
14274 for (bit=0; bit<32; bit++) {
14275 if (flags & (1<<bit)) {
14276 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
14279 if (!set++ && lead)
14280 PerlIO_printf(Perl_debug_log, "%s",lead);
14281 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
14284 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
14285 if (!set++ && lead) {
14286 PerlIO_printf(Perl_debug_log, "%s",lead);
14289 case REGEX_UNICODE_CHARSET:
14290 PerlIO_printf(Perl_debug_log, "UNICODE");
14292 case REGEX_LOCALE_CHARSET:
14293 PerlIO_printf(Perl_debug_log, "LOCALE");
14295 case REGEX_ASCII_RESTRICTED_CHARSET:
14296 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
14298 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
14299 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
14302 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
14308 PerlIO_printf(Perl_debug_log, "\n");
14310 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14316 Perl_regdump(pTHX_ const regexp *r)
14320 SV * const sv = sv_newmortal();
14321 SV *dsv= sv_newmortal();
14322 RXi_GET_DECL(r,ri);
14323 GET_RE_DEBUG_FLAGS_DECL;
14325 PERL_ARGS_ASSERT_REGDUMP;
14327 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
14329 /* Header fields of interest. */
14330 if (r->anchored_substr) {
14331 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
14332 RE_SV_DUMPLEN(r->anchored_substr), 30);
14333 PerlIO_printf(Perl_debug_log,
14334 "anchored %s%s at %"IVdf" ",
14335 s, RE_SV_TAIL(r->anchored_substr),
14336 (IV)r->anchored_offset);
14337 } else if (r->anchored_utf8) {
14338 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
14339 RE_SV_DUMPLEN(r->anchored_utf8), 30);
14340 PerlIO_printf(Perl_debug_log,
14341 "anchored utf8 %s%s at %"IVdf" ",
14342 s, RE_SV_TAIL(r->anchored_utf8),
14343 (IV)r->anchored_offset);
14345 if (r->float_substr) {
14346 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
14347 RE_SV_DUMPLEN(r->float_substr), 30);
14348 PerlIO_printf(Perl_debug_log,
14349 "floating %s%s at %"IVdf"..%"UVuf" ",
14350 s, RE_SV_TAIL(r->float_substr),
14351 (IV)r->float_min_offset, (UV)r->float_max_offset);
14352 } else if (r->float_utf8) {
14353 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
14354 RE_SV_DUMPLEN(r->float_utf8), 30);
14355 PerlIO_printf(Perl_debug_log,
14356 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
14357 s, RE_SV_TAIL(r->float_utf8),
14358 (IV)r->float_min_offset, (UV)r->float_max_offset);
14360 if (r->check_substr || r->check_utf8)
14361 PerlIO_printf(Perl_debug_log,
14363 (r->check_substr == r->float_substr
14364 && r->check_utf8 == r->float_utf8
14365 ? "(checking floating" : "(checking anchored"));
14366 if (r->extflags & RXf_NOSCAN)
14367 PerlIO_printf(Perl_debug_log, " noscan");
14368 if (r->extflags & RXf_CHECK_ALL)
14369 PerlIO_printf(Perl_debug_log, " isall");
14370 if (r->check_substr || r->check_utf8)
14371 PerlIO_printf(Perl_debug_log, ") ");
14373 if (ri->regstclass) {
14374 regprop(r, sv, ri->regstclass);
14375 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
14377 if (r->extflags & RXf_ANCH) {
14378 PerlIO_printf(Perl_debug_log, "anchored");
14379 if (r->extflags & RXf_ANCH_BOL)
14380 PerlIO_printf(Perl_debug_log, "(BOL)");
14381 if (r->extflags & RXf_ANCH_MBOL)
14382 PerlIO_printf(Perl_debug_log, "(MBOL)");
14383 if (r->extflags & RXf_ANCH_SBOL)
14384 PerlIO_printf(Perl_debug_log, "(SBOL)");
14385 if (r->extflags & RXf_ANCH_GPOS)
14386 PerlIO_printf(Perl_debug_log, "(GPOS)");
14387 PerlIO_putc(Perl_debug_log, ' ');
14389 if (r->extflags & RXf_GPOS_SEEN)
14390 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
14391 if (r->intflags & PREGf_SKIP)
14392 PerlIO_printf(Perl_debug_log, "plus ");
14393 if (r->intflags & PREGf_IMPLICIT)
14394 PerlIO_printf(Perl_debug_log, "implicit ");
14395 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
14396 if (r->extflags & RXf_EVAL_SEEN)
14397 PerlIO_printf(Perl_debug_log, "with eval ");
14398 PerlIO_printf(Perl_debug_log, "\n");
14399 DEBUG_FLAGS_r(regdump_extflags("r->extflags: ",r->extflags));
14401 PERL_ARGS_ASSERT_REGDUMP;
14402 PERL_UNUSED_CONTEXT;
14403 PERL_UNUSED_ARG(r);
14404 #endif /* DEBUGGING */
14408 - regprop - printable representation of opcode
14410 #define EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags) \
14413 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]); \
14414 if (flags & ANYOF_INVERT) \
14415 /*make sure the invert info is in each */ \
14416 sv_catpvs(sv, "^"); \
14422 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
14428 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
14429 static const char * const anyofs[] = {
14430 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
14431 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
14432 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
14433 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
14434 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
14435 || _CC_VERTSPACE != 16
14436 #error Need to adjust order of anyofs[]
14473 RXi_GET_DECL(prog,progi);
14474 GET_RE_DEBUG_FLAGS_DECL;
14476 PERL_ARGS_ASSERT_REGPROP;
14480 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
14481 /* It would be nice to FAIL() here, but this may be called from
14482 regexec.c, and it would be hard to supply pRExC_state. */
14483 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(o), (int)REGNODE_MAX);
14484 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
14486 k = PL_regkind[OP(o)];
14489 sv_catpvs(sv, " ");
14490 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
14491 * is a crude hack but it may be the best for now since
14492 * we have no flag "this EXACTish node was UTF-8"
14494 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
14495 PERL_PV_ESCAPE_UNI_DETECT |
14496 PERL_PV_ESCAPE_NONASCII |
14497 PERL_PV_PRETTY_ELLIPSES |
14498 PERL_PV_PRETTY_LTGT |
14499 PERL_PV_PRETTY_NOCLEAR
14501 } else if (k == TRIE) {
14502 /* print the details of the trie in dumpuntil instead, as
14503 * progi->data isn't available here */
14504 const char op = OP(o);
14505 const U32 n = ARG(o);
14506 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
14507 (reg_ac_data *)progi->data->data[n] :
14509 const reg_trie_data * const trie
14510 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
14512 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
14513 DEBUG_TRIE_COMPILE_r(
14514 Perl_sv_catpvf(aTHX_ sv,
14515 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
14516 (UV)trie->startstate,
14517 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
14518 (UV)trie->wordcount,
14521 (UV)TRIE_CHARCOUNT(trie),
14522 (UV)trie->uniquecharcount
14525 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
14527 int rangestart = -1;
14528 U8* bitmap = IS_ANYOF_TRIE(op) ? (U8*)ANYOF_BITMAP(o) : (U8*)TRIE_BITMAP(trie);
14529 sv_catpvs(sv, "[");
14530 for (i = 0; i <= 256; i++) {
14531 if (i < 256 && BITMAP_TEST(bitmap,i)) {
14532 if (rangestart == -1)
14534 } else if (rangestart != -1) {
14535 if (i <= rangestart + 3)
14536 for (; rangestart < i; rangestart++)
14537 put_byte(sv, rangestart);
14539 put_byte(sv, rangestart);
14540 sv_catpvs(sv, "-");
14541 put_byte(sv, i - 1);
14546 sv_catpvs(sv, "]");
14549 } else if (k == CURLY) {
14550 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
14551 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
14552 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
14554 else if (k == WHILEM && o->flags) /* Ordinal/of */
14555 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
14556 else if (k == REF || k == OPEN || k == CLOSE || k == GROUPP || OP(o)==ACCEPT) {
14557 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
14558 if ( RXp_PAREN_NAMES(prog) ) {
14559 if ( k != REF || (OP(o) < NREF)) {
14560 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
14561 SV **name= av_fetch(list, ARG(o), 0 );
14563 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14566 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
14567 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
14568 I32 *nums=(I32*)SvPVX(sv_dat);
14569 SV **name= av_fetch(list, nums[0], 0 );
14572 for ( n=0; n<SvIVX(sv_dat); n++ ) {
14573 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
14574 (n ? "," : ""), (IV)nums[n]);
14576 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14580 } else if (k == GOSUB)
14581 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o)); /* Paren and offset */
14582 else if (k == VERB) {
14584 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
14585 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
14586 } else if (k == LOGICAL)
14587 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* 2: embedded, otherwise 1 */
14588 else if (k == ANYOF) {
14589 int i, rangestart = -1;
14590 const U8 flags = ANYOF_FLAGS(o);
14594 if (flags & ANYOF_LOCALE)
14595 sv_catpvs(sv, "{loc}");
14596 if (flags & ANYOF_LOC_FOLD)
14597 sv_catpvs(sv, "{i}");
14598 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
14599 if (flags & ANYOF_INVERT)
14600 sv_catpvs(sv, "^");
14602 /* output what the standard cp 0-255 bitmap matches */
14603 for (i = 0; i <= 256; i++) {
14604 if (i < 256 && ANYOF_BITMAP_TEST(o,i)) {
14605 if (rangestart == -1)
14607 } else if (rangestart != -1) {
14608 if (i <= rangestart + 3)
14609 for (; rangestart < i; rangestart++)
14610 put_byte(sv, rangestart);
14612 put_byte(sv, rangestart);
14613 sv_catpvs(sv, "-");
14614 put_byte(sv, i - 1);
14621 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14622 /* output any special charclass tests (used entirely under use locale) */
14623 if (ANYOF_CLASS_TEST_ANY_SET(o))
14624 for (i = 0; i < (int)(sizeof(anyofs)/sizeof(char*)); i++)
14625 if (ANYOF_CLASS_TEST(o,i)) {
14626 sv_catpv(sv, anyofs[i]);
14630 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14632 if (flags & ANYOF_NON_UTF8_LATIN1_ALL) {
14633 sv_catpvs(sv, "{non-utf8-latin1-all}");
14636 /* output information about the unicode matching */
14637 if (flags & ANYOF_UNICODE_ALL)
14638 sv_catpvs(sv, "{unicode_all}");
14639 else if (ANYOF_NONBITMAP(o))
14640 sv_catpvs(sv, "{unicode}");
14641 if (flags & ANYOF_NONBITMAP_NON_UTF8)
14642 sv_catpvs(sv, "{outside bitmap}");
14644 if (ANYOF_NONBITMAP(o)) {
14645 SV *lv; /* Set if there is something outside the bit map */
14646 SV * const sw = regclass_swash(prog, o, FALSE, &lv, NULL);
14647 bool byte_output = FALSE; /* If something in the bitmap has been
14650 if (lv && lv != &PL_sv_undef) {
14652 U8 s[UTF8_MAXBYTES_CASE+1];
14654 for (i = 0; i <= 256; i++) { /* Look at chars in bitmap */
14655 uvchr_to_utf8(s, i);
14658 && ! ANYOF_BITMAP_TEST(o, i) /* Don't duplicate
14662 && swash_fetch(sw, s, TRUE))
14664 if (rangestart == -1)
14666 } else if (rangestart != -1) {
14667 byte_output = TRUE;
14668 if (i <= rangestart + 3)
14669 for (; rangestart < i; rangestart++) {
14670 put_byte(sv, rangestart);
14673 put_byte(sv, rangestart);
14674 sv_catpvs(sv, "-");
14683 char *s = savesvpv(lv);
14684 char * const origs = s;
14686 while (*s && *s != '\n')
14690 const char * const t = ++s;
14693 sv_catpvs(sv, " ");
14699 /* Truncate very long output */
14700 if (s - origs > 256) {
14701 Perl_sv_catpvf(aTHX_ sv,
14703 (int) (s - origs - 1),
14709 else if (*s == '\t') {
14724 SvREFCNT_dec_NN(lv);
14728 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
14730 else if (k == POSIXD || k == NPOSIXD) {
14731 U8 index = FLAGS(o) * 2;
14732 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
14733 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
14736 sv_catpv(sv, anyofs[index]);
14739 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
14740 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
14742 PERL_UNUSED_CONTEXT;
14743 PERL_UNUSED_ARG(sv);
14744 PERL_UNUSED_ARG(o);
14745 PERL_UNUSED_ARG(prog);
14746 #endif /* DEBUGGING */
14750 Perl_re_intuit_string(pTHX_ REGEXP * const r)
14751 { /* Assume that RE_INTUIT is set */
14753 struct regexp *const prog = ReANY(r);
14754 GET_RE_DEBUG_FLAGS_DECL;
14756 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
14757 PERL_UNUSED_CONTEXT;
14761 const char * const s = SvPV_nolen_const(prog->check_substr
14762 ? prog->check_substr : prog->check_utf8);
14764 if (!PL_colorset) reginitcolors();
14765 PerlIO_printf(Perl_debug_log,
14766 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
14768 prog->check_substr ? "" : "utf8 ",
14769 PL_colors[5],PL_colors[0],
14772 (strlen(s) > 60 ? "..." : ""));
14775 return prog->check_substr ? prog->check_substr : prog->check_utf8;
14781 handles refcounting and freeing the perl core regexp structure. When
14782 it is necessary to actually free the structure the first thing it
14783 does is call the 'free' method of the regexp_engine associated to
14784 the regexp, allowing the handling of the void *pprivate; member
14785 first. (This routine is not overridable by extensions, which is why
14786 the extensions free is called first.)
14788 See regdupe and regdupe_internal if you change anything here.
14790 #ifndef PERL_IN_XSUB_RE
14792 Perl_pregfree(pTHX_ REGEXP *r)
14798 Perl_pregfree2(pTHX_ REGEXP *rx)
14801 struct regexp *const r = ReANY(rx);
14802 GET_RE_DEBUG_FLAGS_DECL;
14804 PERL_ARGS_ASSERT_PREGFREE2;
14806 if (r->mother_re) {
14807 ReREFCNT_dec(r->mother_re);
14809 CALLREGFREE_PVT(rx); /* free the private data */
14810 SvREFCNT_dec(RXp_PAREN_NAMES(r));
14811 Safefree(r->xpv_len_u.xpvlenu_pv);
14814 SvREFCNT_dec(r->anchored_substr);
14815 SvREFCNT_dec(r->anchored_utf8);
14816 SvREFCNT_dec(r->float_substr);
14817 SvREFCNT_dec(r->float_utf8);
14818 Safefree(r->substrs);
14820 RX_MATCH_COPY_FREE(rx);
14821 #ifdef PERL_ANY_COW
14822 SvREFCNT_dec(r->saved_copy);
14825 SvREFCNT_dec(r->qr_anoncv);
14826 rx->sv_u.svu_rx = 0;
14831 This is a hacky workaround to the structural issue of match results
14832 being stored in the regexp structure which is in turn stored in
14833 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
14834 could be PL_curpm in multiple contexts, and could require multiple
14835 result sets being associated with the pattern simultaneously, such
14836 as when doing a recursive match with (??{$qr})
14838 The solution is to make a lightweight copy of the regexp structure
14839 when a qr// is returned from the code executed by (??{$qr}) this
14840 lightweight copy doesn't actually own any of its data except for
14841 the starp/end and the actual regexp structure itself.
14847 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
14849 struct regexp *ret;
14850 struct regexp *const r = ReANY(rx);
14851 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
14853 PERL_ARGS_ASSERT_REG_TEMP_COPY;
14856 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
14858 SvOK_off((SV *)ret_x);
14860 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
14861 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
14862 made both spots point to the same regexp body.) */
14863 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
14864 assert(!SvPVX(ret_x));
14865 ret_x->sv_u.svu_rx = temp->sv_any;
14866 temp->sv_any = NULL;
14867 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
14868 SvREFCNT_dec_NN(temp);
14869 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
14870 ing below will not set it. */
14871 SvCUR_set(ret_x, SvCUR(rx));
14874 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
14875 sv_force_normal(sv) is called. */
14877 ret = ReANY(ret_x);
14879 SvFLAGS(ret_x) |= SvUTF8(rx);
14880 /* We share the same string buffer as the original regexp, on which we
14881 hold a reference count, incremented when mother_re is set below.
14882 The string pointer is copied here, being part of the regexp struct.
14884 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
14885 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
14887 const I32 npar = r->nparens+1;
14888 Newx(ret->offs, npar, regexp_paren_pair);
14889 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
14892 Newx(ret->substrs, 1, struct reg_substr_data);
14893 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
14895 SvREFCNT_inc_void(ret->anchored_substr);
14896 SvREFCNT_inc_void(ret->anchored_utf8);
14897 SvREFCNT_inc_void(ret->float_substr);
14898 SvREFCNT_inc_void(ret->float_utf8);
14900 /* check_substr and check_utf8, if non-NULL, point to either their
14901 anchored or float namesakes, and don't hold a second reference. */
14903 RX_MATCH_COPIED_off(ret_x);
14904 #ifdef PERL_ANY_COW
14905 ret->saved_copy = NULL;
14907 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
14908 SvREFCNT_inc_void(ret->qr_anoncv);
14914 /* regfree_internal()
14916 Free the private data in a regexp. This is overloadable by
14917 extensions. Perl takes care of the regexp structure in pregfree(),
14918 this covers the *pprivate pointer which technically perl doesn't
14919 know about, however of course we have to handle the
14920 regexp_internal structure when no extension is in use.
14922 Note this is called before freeing anything in the regexp
14927 Perl_regfree_internal(pTHX_ REGEXP * const rx)
14930 struct regexp *const r = ReANY(rx);
14931 RXi_GET_DECL(r,ri);
14932 GET_RE_DEBUG_FLAGS_DECL;
14934 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
14940 SV *dsv= sv_newmortal();
14941 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
14942 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
14943 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
14944 PL_colors[4],PL_colors[5],s);
14947 #ifdef RE_TRACK_PATTERN_OFFSETS
14949 Safefree(ri->u.offsets); /* 20010421 MJD */
14951 if (ri->code_blocks) {
14953 for (n = 0; n < ri->num_code_blocks; n++)
14954 SvREFCNT_dec(ri->code_blocks[n].src_regex);
14955 Safefree(ri->code_blocks);
14959 int n = ri->data->count;
14962 /* If you add a ->what type here, update the comment in regcomp.h */
14963 switch (ri->data->what[n]) {
14969 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
14972 Safefree(ri->data->data[n]);
14978 { /* Aho Corasick add-on structure for a trie node.
14979 Used in stclass optimization only */
14981 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
14983 refcount = --aho->refcount;
14986 PerlMemShared_free(aho->states);
14987 PerlMemShared_free(aho->fail);
14988 /* do this last!!!! */
14989 PerlMemShared_free(ri->data->data[n]);
14990 PerlMemShared_free(ri->regstclass);
14996 /* trie structure. */
14998 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
15000 refcount = --trie->refcount;
15003 PerlMemShared_free(trie->charmap);
15004 PerlMemShared_free(trie->states);
15005 PerlMemShared_free(trie->trans);
15007 PerlMemShared_free(trie->bitmap);
15009 PerlMemShared_free(trie->jump);
15010 PerlMemShared_free(trie->wordinfo);
15011 /* do this last!!!! */
15012 PerlMemShared_free(ri->data->data[n]);
15017 Perl_croak(aTHX_ "panic: regfree data code '%c'", ri->data->what[n]);
15020 Safefree(ri->data->what);
15021 Safefree(ri->data);
15027 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
15028 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
15029 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
15032 re_dup - duplicate a regexp.
15034 This routine is expected to clone a given regexp structure. It is only
15035 compiled under USE_ITHREADS.
15037 After all of the core data stored in struct regexp is duplicated
15038 the regexp_engine.dupe method is used to copy any private data
15039 stored in the *pprivate pointer. This allows extensions to handle
15040 any duplication it needs to do.
15042 See pregfree() and regfree_internal() if you change anything here.
15044 #if defined(USE_ITHREADS)
15045 #ifndef PERL_IN_XSUB_RE
15047 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
15051 const struct regexp *r = ReANY(sstr);
15052 struct regexp *ret = ReANY(dstr);
15054 PERL_ARGS_ASSERT_RE_DUP_GUTS;
15056 npar = r->nparens+1;
15057 Newx(ret->offs, npar, regexp_paren_pair);
15058 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15060 if (ret->substrs) {
15061 /* Do it this way to avoid reading from *r after the StructCopy().
15062 That way, if any of the sv_dup_inc()s dislodge *r from the L1
15063 cache, it doesn't matter. */
15064 const bool anchored = r->check_substr
15065 ? r->check_substr == r->anchored_substr
15066 : r->check_utf8 == r->anchored_utf8;
15067 Newx(ret->substrs, 1, struct reg_substr_data);
15068 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15070 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
15071 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
15072 ret->float_substr = sv_dup_inc(ret->float_substr, param);
15073 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
15075 /* check_substr and check_utf8, if non-NULL, point to either their
15076 anchored or float namesakes, and don't hold a second reference. */
15078 if (ret->check_substr) {
15080 assert(r->check_utf8 == r->anchored_utf8);
15081 ret->check_substr = ret->anchored_substr;
15082 ret->check_utf8 = ret->anchored_utf8;
15084 assert(r->check_substr == r->float_substr);
15085 assert(r->check_utf8 == r->float_utf8);
15086 ret->check_substr = ret->float_substr;
15087 ret->check_utf8 = ret->float_utf8;
15089 } else if (ret->check_utf8) {
15091 ret->check_utf8 = ret->anchored_utf8;
15093 ret->check_utf8 = ret->float_utf8;
15098 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
15099 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
15102 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
15104 if (RX_MATCH_COPIED(dstr))
15105 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
15107 ret->subbeg = NULL;
15108 #ifdef PERL_ANY_COW
15109 ret->saved_copy = NULL;
15112 /* Whether mother_re be set or no, we need to copy the string. We
15113 cannot refrain from copying it when the storage points directly to
15114 our mother regexp, because that's
15115 1: a buffer in a different thread
15116 2: something we no longer hold a reference on
15117 so we need to copy it locally. */
15118 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
15119 ret->mother_re = NULL;
15122 #endif /* PERL_IN_XSUB_RE */
15127 This is the internal complement to regdupe() which is used to copy
15128 the structure pointed to by the *pprivate pointer in the regexp.
15129 This is the core version of the extension overridable cloning hook.
15130 The regexp structure being duplicated will be copied by perl prior
15131 to this and will be provided as the regexp *r argument, however
15132 with the /old/ structures pprivate pointer value. Thus this routine
15133 may override any copying normally done by perl.
15135 It returns a pointer to the new regexp_internal structure.
15139 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
15142 struct regexp *const r = ReANY(rx);
15143 regexp_internal *reti;
15145 RXi_GET_DECL(r,ri);
15147 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
15151 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode), char, regexp_internal);
15152 Copy(ri->program, reti->program, len+1, regnode);
15154 reti->num_code_blocks = ri->num_code_blocks;
15155 if (ri->code_blocks) {
15157 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
15158 struct reg_code_block);
15159 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
15160 struct reg_code_block);
15161 for (n = 0; n < ri->num_code_blocks; n++)
15162 reti->code_blocks[n].src_regex = (REGEXP*)
15163 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
15166 reti->code_blocks = NULL;
15168 reti->regstclass = NULL;
15171 struct reg_data *d;
15172 const int count = ri->data->count;
15175 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
15176 char, struct reg_data);
15177 Newx(d->what, count, U8);
15180 for (i = 0; i < count; i++) {
15181 d->what[i] = ri->data->what[i];
15182 switch (d->what[i]) {
15183 /* see also regcomp.h and regfree_internal() */
15184 case 'a': /* actually an AV, but the dup function is identical. */
15188 case 'u': /* actually an HV, but the dup function is identical. */
15189 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
15192 /* This is cheating. */
15193 Newx(d->data[i], 1, struct regnode_charclass_class);
15194 StructCopy(ri->data->data[i], d->data[i],
15195 struct regnode_charclass_class);
15196 reti->regstclass = (regnode*)d->data[i];
15199 /* Trie stclasses are readonly and can thus be shared
15200 * without duplication. We free the stclass in pregfree
15201 * when the corresponding reg_ac_data struct is freed.
15203 reti->regstclass= ri->regstclass;
15207 ((reg_trie_data*)ri->data->data[i])->refcount++;
15212 d->data[i] = ri->data->data[i];
15215 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'", ri->data->what[i]);
15224 reti->name_list_idx = ri->name_list_idx;
15226 #ifdef RE_TRACK_PATTERN_OFFSETS
15227 if (ri->u.offsets) {
15228 Newx(reti->u.offsets, 2*len+1, U32);
15229 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
15232 SetProgLen(reti,len);
15235 return (void*)reti;
15238 #endif /* USE_ITHREADS */
15240 #ifndef PERL_IN_XSUB_RE
15243 - regnext - dig the "next" pointer out of a node
15246 Perl_regnext(pTHX_ regnode *p)
15254 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
15255 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(p), (int)REGNODE_MAX);
15258 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
15267 S_re_croak2(pTHX_ const char* pat1,const char* pat2,...)
15270 STRLEN l1 = strlen(pat1);
15271 STRLEN l2 = strlen(pat2);
15274 const char *message;
15276 PERL_ARGS_ASSERT_RE_CROAK2;
15282 Copy(pat1, buf, l1 , char);
15283 Copy(pat2, buf + l1, l2 , char);
15284 buf[l1 + l2] = '\n';
15285 buf[l1 + l2 + 1] = '\0';
15287 /* ANSI variant takes additional second argument */
15288 va_start(args, pat2);
15292 msv = vmess(buf, &args);
15294 message = SvPV_const(msv,l1);
15297 Copy(message, buf, l1 , char);
15298 buf[l1-1] = '\0'; /* Overwrite \n */
15299 Perl_croak(aTHX_ "%s", buf);
15302 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
15304 #ifndef PERL_IN_XSUB_RE
15306 Perl_save_re_context(pTHX)
15310 struct re_save_state *state;
15312 SAVEVPTR(PL_curcop);
15313 SSGROW(SAVESTACK_ALLOC_FOR_RE_SAVE_STATE + 1);
15315 state = (struct re_save_state *)(PL_savestack + PL_savestack_ix);
15316 PL_savestack_ix += SAVESTACK_ALLOC_FOR_RE_SAVE_STATE;
15317 SSPUSHUV(SAVEt_RE_STATE);
15319 Copy(&PL_reg_state, state, 1, struct re_save_state);
15321 PL_reg_oldsaved = NULL;
15322 PL_reg_oldsavedlen = 0;
15323 PL_reg_oldsavedoffset = 0;
15324 PL_reg_oldsavedcoffset = 0;
15325 PL_reg_maxiter = 0;
15326 PL_reg_leftiter = 0;
15327 PL_reg_poscache = NULL;
15328 PL_reg_poscache_size = 0;
15329 #ifdef PERL_ANY_COW
15333 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
15335 const REGEXP * const rx = PM_GETRE(PL_curpm);
15338 for (i = 1; i <= RX_NPARENS(rx); i++) {
15339 char digits[TYPE_CHARS(long)];
15340 const STRLEN len = my_snprintf(digits, sizeof(digits), "%lu", (long)i);
15341 GV *const *const gvp
15342 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
15345 GV * const gv = *gvp;
15346 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
15358 S_put_byte(pTHX_ SV *sv, int c)
15360 PERL_ARGS_ASSERT_PUT_BYTE;
15362 /* Our definition of isPRINT() ignores locales, so only bytes that are
15363 not part of UTF-8 are considered printable. I assume that the same
15364 holds for UTF-EBCDIC.
15365 Also, code point 255 is not printable in either (it's E0 in EBCDIC,
15366 which Wikipedia says:
15368 EO, or Eight Ones, is an 8-bit EBCDIC character code represented as all
15369 ones (binary 1111 1111, hexadecimal FF). It is similar, but not
15370 identical, to the ASCII delete (DEL) or rubout control character. ...
15371 it is typically mapped to hexadecimal code 9F, in order to provide a
15372 unique character mapping in both directions)
15374 So the old condition can be simplified to !isPRINT(c) */
15377 Perl_sv_catpvf(aTHX_ sv, "\\x%02x", c);
15380 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
15384 const char string = c;
15385 if (c == '-' || c == ']' || c == '\\' || c == '^')
15386 sv_catpvs(sv, "\\");
15387 sv_catpvn(sv, &string, 1);
15392 #define CLEAR_OPTSTART \
15393 if (optstart) STMT_START { \
15394 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
15398 #define DUMPUNTIL(b,e) CLEAR_OPTSTART; node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
15400 STATIC const regnode *
15401 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
15402 const regnode *last, const regnode *plast,
15403 SV* sv, I32 indent, U32 depth)
15406 U8 op = PSEUDO; /* Arbitrary non-END op. */
15407 const regnode *next;
15408 const regnode *optstart= NULL;
15410 RXi_GET_DECL(r,ri);
15411 GET_RE_DEBUG_FLAGS_DECL;
15413 PERL_ARGS_ASSERT_DUMPUNTIL;
15415 #ifdef DEBUG_DUMPUNTIL
15416 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
15417 last ? last-start : 0,plast ? plast-start : 0);
15420 if (plast && plast < last)
15423 while (PL_regkind[op] != END && (!last || node < last)) {
15424 /* While that wasn't END last time... */
15427 if (op == CLOSE || op == WHILEM)
15429 next = regnext((regnode *)node);
15432 if (OP(node) == OPTIMIZED) {
15433 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
15440 regprop(r, sv, node);
15441 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
15442 (int)(2*indent + 1), "", SvPVX_const(sv));
15444 if (OP(node) != OPTIMIZED) {
15445 if (next == NULL) /* Next ptr. */
15446 PerlIO_printf(Perl_debug_log, " (0)");
15447 else if (PL_regkind[(U8)op] == BRANCH && PL_regkind[OP(next)] != BRANCH )
15448 PerlIO_printf(Perl_debug_log, " (FAIL)");
15450 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
15451 (void)PerlIO_putc(Perl_debug_log, '\n');
15455 if (PL_regkind[(U8)op] == BRANCHJ) {
15458 const regnode *nnode = (OP(next) == LONGJMP
15459 ? regnext((regnode *)next)
15461 if (last && nnode > last)
15463 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
15466 else if (PL_regkind[(U8)op] == BRANCH) {
15468 DUMPUNTIL(NEXTOPER(node), next);
15470 else if ( PL_regkind[(U8)op] == TRIE ) {
15471 const regnode *this_trie = node;
15472 const char op = OP(node);
15473 const U32 n = ARG(node);
15474 const reg_ac_data * const ac = op>=AHOCORASICK ?
15475 (reg_ac_data *)ri->data->data[n] :
15477 const reg_trie_data * const trie =
15478 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
15480 AV *const trie_words = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
15482 const regnode *nextbranch= NULL;
15485 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
15486 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
15488 PerlIO_printf(Perl_debug_log, "%*s%s ",
15489 (int)(2*(indent+3)), "",
15490 elem_ptr ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr), SvCUR(*elem_ptr), 60,
15491 PL_colors[0], PL_colors[1],
15492 (SvUTF8(*elem_ptr) ? PERL_PV_ESCAPE_UNI : 0) |
15493 PERL_PV_PRETTY_ELLIPSES |
15494 PERL_PV_PRETTY_LTGT
15499 U16 dist= trie->jump[word_idx+1];
15500 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
15501 (UV)((dist ? this_trie + dist : next) - start));
15504 nextbranch= this_trie + trie->jump[0];
15505 DUMPUNTIL(this_trie + dist, nextbranch);
15507 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
15508 nextbranch= regnext((regnode *)nextbranch);
15510 PerlIO_printf(Perl_debug_log, "\n");
15513 if (last && next > last)
15518 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
15519 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
15520 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
15522 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
15524 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
15526 else if ( op == PLUS || op == STAR) {
15527 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
15529 else if (PL_regkind[(U8)op] == ANYOF) {
15530 /* arglen 1 + class block */
15531 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_CLASS)
15532 ? ANYOF_CLASS_SKIP : ANYOF_SKIP);
15533 node = NEXTOPER(node);
15535 else if (PL_regkind[(U8)op] == EXACT) {
15536 /* Literal string, where present. */
15537 node += NODE_SZ_STR(node) - 1;
15538 node = NEXTOPER(node);
15541 node = NEXTOPER(node);
15542 node += regarglen[(U8)op];
15544 if (op == CURLYX || op == OPEN)
15548 #ifdef DEBUG_DUMPUNTIL
15549 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
15554 #endif /* DEBUGGING */
15558 * c-indentation-style: bsd
15559 * c-basic-offset: 4
15560 * indent-tabs-mode: nil
15563 * ex: set ts=8 sts=4 sw=4 et: