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);
4876 /* see if there are any run-time code blocks in the pattern.
4877 * False positives are allowed */
4880 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
4881 char *pat, STRLEN plen)
4886 for (s = 0; s < plen; s++) {
4887 if (n < pRExC_state->num_code_blocks
4888 && s == pRExC_state->code_blocks[n].start)
4890 s = pRExC_state->code_blocks[n].end;
4894 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
4896 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
4898 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
4905 /* Handle run-time code blocks. We will already have compiled any direct
4906 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
4907 * copy of it, but with any literal code blocks blanked out and
4908 * appropriate chars escaped; then feed it into
4910 * eval "qr'modified_pattern'"
4914 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
4918 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
4920 * After eval_sv()-ing that, grab any new code blocks from the returned qr
4921 * and merge them with any code blocks of the original regexp.
4923 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
4924 * instead, just save the qr and return FALSE; this tells our caller that
4925 * the original pattern needs upgrading to utf8.
4929 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
4930 char *pat, STRLEN plen)
4934 GET_RE_DEBUG_FLAGS_DECL;
4936 if (pRExC_state->runtime_code_qr) {
4937 /* this is the second time we've been called; this should
4938 * only happen if the main pattern got upgraded to utf8
4939 * during compilation; re-use the qr we compiled first time
4940 * round (which should be utf8 too)
4942 qr = pRExC_state->runtime_code_qr;
4943 pRExC_state->runtime_code_qr = NULL;
4944 assert(RExC_utf8 && SvUTF8(qr));
4950 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
4954 /* determine how many extra chars we need for ' and \ escaping */
4955 for (s = 0; s < plen; s++) {
4956 if (pat[s] == '\'' || pat[s] == '\\')
4960 Newx(newpat, newlen, char);
4962 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
4964 for (s = 0; s < plen; s++) {
4965 if (n < pRExC_state->num_code_blocks
4966 && s == pRExC_state->code_blocks[n].start)
4968 /* blank out literal code block */
4969 assert(pat[s] == '(');
4970 while (s <= pRExC_state->code_blocks[n].end) {
4978 if (pat[s] == '\'' || pat[s] == '\\')
4983 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
4987 PerlIO_printf(Perl_debug_log,
4988 "%sre-parsing pattern for runtime code:%s %s\n",
4989 PL_colors[4],PL_colors[5],newpat);
4992 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
4998 PUSHSTACKi(PERLSI_REQUIRE);
4999 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
5000 * parsing qr''; normally only q'' does this. It also alters
5002 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
5003 SvREFCNT_dec_NN(sv);
5008 SV * const errsv = ERRSV;
5009 if (SvTRUE_NN(errsv))
5011 Safefree(pRExC_state->code_blocks);
5012 /* use croak_sv ? */
5013 Perl_croak_nocontext("%s", SvPV_nolen_const(errsv));
5016 assert(SvROK(qr_ref));
5018 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5019 /* the leaving below frees the tmp qr_ref.
5020 * Give qr a life of its own */
5028 if (!RExC_utf8 && SvUTF8(qr)) {
5029 /* first time through; the pattern got upgraded; save the
5030 * qr for the next time through */
5031 assert(!pRExC_state->runtime_code_qr);
5032 pRExC_state->runtime_code_qr = qr;
5037 /* extract any code blocks within the returned qr// */
5040 /* merge the main (r1) and run-time (r2) code blocks into one */
5042 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
5043 struct reg_code_block *new_block, *dst;
5044 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
5047 if (!r2->num_code_blocks) /* we guessed wrong */
5049 SvREFCNT_dec_NN(qr);
5054 r1->num_code_blocks + r2->num_code_blocks,
5055 struct reg_code_block);
5058 while ( i1 < r1->num_code_blocks
5059 || i2 < r2->num_code_blocks)
5061 struct reg_code_block *src;
5064 if (i1 == r1->num_code_blocks) {
5065 src = &r2->code_blocks[i2++];
5068 else if (i2 == r2->num_code_blocks)
5069 src = &r1->code_blocks[i1++];
5070 else if ( r1->code_blocks[i1].start
5071 < r2->code_blocks[i2].start)
5073 src = &r1->code_blocks[i1++];
5074 assert(src->end < r2->code_blocks[i2].start);
5077 assert( r1->code_blocks[i1].start
5078 > r2->code_blocks[i2].start);
5079 src = &r2->code_blocks[i2++];
5081 assert(src->end < r1->code_blocks[i1].start);
5084 assert(pat[src->start] == '(');
5085 assert(pat[src->end] == ')');
5086 dst->start = src->start;
5087 dst->end = src->end;
5088 dst->block = src->block;
5089 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
5093 r1->num_code_blocks += r2->num_code_blocks;
5094 Safefree(r1->code_blocks);
5095 r1->code_blocks = new_block;
5098 SvREFCNT_dec_NN(qr);
5104 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)
5106 /* This is the common code for setting up the floating and fixed length
5107 * string data extracted from Perl_re_op_compile() below. Returns a boolean
5108 * as to whether succeeded or not */
5112 if (! (longest_length
5113 || (eol /* Can't have SEOL and MULTI */
5114 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
5116 /* See comments for join_exact for why REG_SEEN_EXACTF_SHARP_S */
5117 || (RExC_seen & REG_SEEN_EXACTF_SHARP_S))
5122 /* copy the information about the longest from the reg_scan_data
5123 over to the program. */
5124 if (SvUTF8(sv_longest)) {
5125 *rx_utf8 = sv_longest;
5128 *rx_substr = sv_longest;
5131 /* end_shift is how many chars that must be matched that
5132 follow this item. We calculate it ahead of time as once the
5133 lookbehind offset is added in we lose the ability to correctly
5135 ml = minlen ? *(minlen) : (I32)longest_length;
5136 *rx_end_shift = ml - offset
5137 - longest_length + (SvTAIL(sv_longest) != 0)
5140 t = (eol/* Can't have SEOL and MULTI */
5141 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
5142 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
5148 * Perl_re_op_compile - the perl internal RE engine's function to compile a
5149 * regular expression into internal code.
5150 * The pattern may be passed either as:
5151 * a list of SVs (patternp plus pat_count)
5152 * a list of OPs (expr)
5153 * If both are passed, the SV list is used, but the OP list indicates
5154 * which SVs are actually pre-compiled code blocks
5156 * The SVs in the list have magic and qr overloading applied to them (and
5157 * the list may be modified in-place with replacement SVs in the latter
5160 * If the pattern hasn't changed from old_re, then old_re will be
5163 * eng is the current engine. If that engine has an op_comp method, then
5164 * handle directly (i.e. we assume that op_comp was us); otherwise, just
5165 * do the initial concatenation of arguments and pass on to the external
5168 * If is_bare_re is not null, set it to a boolean indicating whether the
5169 * arg list reduced (after overloading) to a single bare regex which has
5170 * been returned (i.e. /$qr/).
5172 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
5174 * pm_flags contains the PMf_* flags, typically based on those from the
5175 * pm_flags field of the related PMOP. Currently we're only interested in
5176 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
5178 * We can't allocate space until we know how big the compiled form will be,
5179 * but we can't compile it (and thus know how big it is) until we've got a
5180 * place to put the code. So we cheat: we compile it twice, once with code
5181 * generation turned off and size counting turned on, and once "for real".
5182 * This also means that we don't allocate space until we are sure that the
5183 * thing really will compile successfully, and we never have to move the
5184 * code and thus invalidate pointers into it. (Note that it has to be in
5185 * one piece because free() must be able to free it all.) [NB: not true in perl]
5187 * Beware that the optimization-preparation code in here knows about some
5188 * of the structure of the compiled regexp. [I'll say.]
5192 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
5193 OP *expr, const regexp_engine* eng, REGEXP *old_re,
5194 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
5199 regexp_internal *ri;
5208 SV *code_blocksv = NULL;
5209 SV** new_patternp = patternp;
5211 /* these are all flags - maybe they should be turned
5212 * into a single int with different bit masks */
5213 I32 sawlookahead = 0;
5216 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
5218 bool runtime_code = 0;
5220 RExC_state_t RExC_state;
5221 RExC_state_t * const pRExC_state = &RExC_state;
5222 #ifdef TRIE_STUDY_OPT
5224 RExC_state_t copyRExC_state;
5226 GET_RE_DEBUG_FLAGS_DECL;
5228 PERL_ARGS_ASSERT_RE_OP_COMPILE;
5230 DEBUG_r(if (!PL_colorset) reginitcolors());
5232 #ifndef PERL_IN_XSUB_RE
5233 /* Initialize these here instead of as-needed, as is quick and avoids
5234 * having to test them each time otherwise */
5235 if (! PL_AboveLatin1) {
5236 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
5237 PL_ASCII = _new_invlist_C_array(ASCII_invlist);
5238 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
5240 PL_L1Posix_ptrs[_CC_ALPHANUMERIC]
5241 = _new_invlist_C_array(L1PosixAlnum_invlist);
5242 PL_Posix_ptrs[_CC_ALPHANUMERIC]
5243 = _new_invlist_C_array(PosixAlnum_invlist);
5245 PL_L1Posix_ptrs[_CC_ALPHA]
5246 = _new_invlist_C_array(L1PosixAlpha_invlist);
5247 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(PosixAlpha_invlist);
5249 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(PosixBlank_invlist);
5250 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(XPosixBlank_invlist);
5252 /* Cased is the same as Alpha in the ASCII range */
5253 PL_L1Posix_ptrs[_CC_CASED] = _new_invlist_C_array(L1Cased_invlist);
5254 PL_Posix_ptrs[_CC_CASED] = _new_invlist_C_array(PosixAlpha_invlist);
5256 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(PosixCntrl_invlist);
5257 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(XPosixCntrl_invlist);
5259 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5260 PL_L1Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5262 PL_L1Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(L1PosixGraph_invlist);
5263 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(PosixGraph_invlist);
5265 PL_L1Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(L1PosixLower_invlist);
5266 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(PosixLower_invlist);
5268 PL_L1Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(L1PosixPrint_invlist);
5269 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(PosixPrint_invlist);
5271 PL_L1Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(L1PosixPunct_invlist);
5272 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(PosixPunct_invlist);
5274 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(PerlSpace_invlist);
5275 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(XPerlSpace_invlist);
5276 PL_Posix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(PosixSpace_invlist);
5277 PL_XPosix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(XPosixSpace_invlist);
5279 PL_L1Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(L1PosixUpper_invlist);
5280 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(PosixUpper_invlist);
5282 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(VertSpace_invlist);
5284 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(PosixWord_invlist);
5285 PL_L1Posix_ptrs[_CC_WORDCHAR]
5286 = _new_invlist_C_array(L1PosixWord_invlist);
5288 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(PosixXDigit_invlist);
5289 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(XPosixXDigit_invlist);
5291 PL_HasMultiCharFold = _new_invlist_C_array(_Perl_Multi_Char_Folds_invlist);
5295 pRExC_state->code_blocks = NULL;
5296 pRExC_state->num_code_blocks = 0;
5299 *is_bare_re = FALSE;
5301 if (expr && (expr->op_type == OP_LIST ||
5302 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
5303 /* allocate code_blocks if needed */
5307 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling)
5308 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
5309 ncode++; /* count of DO blocks */
5311 pRExC_state->num_code_blocks = ncode;
5312 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
5317 /* compile-time pattern with just OP_CONSTs and DO blocks */
5322 /* find how many CONSTs there are */
5325 if (expr->op_type == OP_CONST)
5328 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5329 if (o->op_type == OP_CONST)
5333 /* fake up an SV array */
5335 assert(!new_patternp);
5336 Newx(new_patternp, n, SV*);
5337 SAVEFREEPV(new_patternp);
5341 if (expr->op_type == OP_CONST)
5342 new_patternp[n] = cSVOPx_sv(expr);
5344 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5345 if (o->op_type == OP_CONST)
5346 new_patternp[n++] = cSVOPo_sv;
5352 /* concat args, handling magic, overloading etc */
5357 STRLEN orig_patlen = 0;
5359 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5360 "Assembling pattern from %d elements%s\n", pat_count,
5361 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5363 /* apply magic and RE overloading to each arg */
5364 for (svp = new_patternp; svp < new_patternp + pat_count; svp++) {
5367 if (SvROK(rx) && SvAMAGIC(rx)) {
5368 SV *sv = AMG_CALLunary(rx, regexp_amg);
5372 if (SvTYPE(sv) != SVt_REGEXP)
5373 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5379 if (pRExC_state->num_code_blocks) {
5380 if (expr->op_type == OP_CONST)
5383 o = cLISTOPx(expr)->op_first;
5384 assert( o->op_type == OP_PUSHMARK
5385 || (o->op_type == OP_NULL && o->op_targ == OP_PUSHMARK)
5386 || o->op_type == OP_PADRANGE);
5391 if (pat_count > 1) {
5393 pat = newSVpvn("", 0);
5396 /* determine if the pattern is going to be utf8 (needed
5397 * in advance to align code block indices correctly).
5398 * XXX This could fail to be detected for an arg with
5399 * overloading but not concat overloading; but the main effect
5400 * in this obscure case is to need a 'use re eval' for a
5401 * literal code block */
5402 for (svp = new_patternp; svp < new_patternp + pat_count; svp++) {
5408 /* process args, concat them if there are multiple ones,
5409 * and find any code block indexes */
5412 for (svp = new_patternp; svp < new_patternp + pat_count; svp++) {
5413 SV *sv, *msv = *svp;
5416 /* we make the assumption here that each op in the list of
5417 * op_siblings maps to one SV pushed onto the stack,
5418 * except for code blocks, with have both an OP_NULL and
5420 * This allows us to match up the list of SVs against the
5421 * list of OPs to find the next code block.
5423 * Note that PUSHMARK PADSV PADSV ..
5425 * PADRANGE NULL NULL ..
5426 * so the alignment still works. */
5428 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL)) {
5429 assert(n < pRExC_state->num_code_blocks);
5430 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5431 pRExC_state->code_blocks[n].block = o;
5432 pRExC_state->code_blocks[n].src_regex = NULL;
5435 o = o->op_sibling; /* skip CONST */
5441 /* try concatenation overload ... */
5442 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5443 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5446 /* overloading involved: all bets are off over literal
5447 * code. Pretend we haven't seen it */
5448 pRExC_state->num_code_blocks -= n;
5452 /* ... or failing that, try "" overload */
5453 while (SvAMAGIC(msv)
5454 && (sv = AMG_CALLunary(msv, string_amg))
5458 && SvRV(msv) == SvRV(sv))
5463 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5466 orig_patlen = SvCUR(pat);
5467 sv_catsv_nomg(pat, msv);
5473 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5476 /* extract any code blocks within any embedded qr//'s */
5477 if (rx && SvTYPE(rx) == SVt_REGEXP
5478 && RX_ENGINE((REGEXP*)rx)->op_comp)
5481 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5482 if (ri->num_code_blocks) {
5484 /* the presence of an embedded qr// with code means
5485 * we should always recompile: the text of the
5486 * qr// may not have changed, but it may be a
5487 * different closure than last time */
5489 Renew(pRExC_state->code_blocks,
5490 pRExC_state->num_code_blocks + ri->num_code_blocks,
5491 struct reg_code_block);
5492 pRExC_state->num_code_blocks += ri->num_code_blocks;
5493 for (i=0; i < ri->num_code_blocks; i++) {
5494 struct reg_code_block *src, *dst;
5495 STRLEN offset = orig_patlen
5496 + ReANY((REGEXP *)rx)->pre_prefix;
5497 assert(n < pRExC_state->num_code_blocks);
5498 src = &ri->code_blocks[i];
5499 dst = &pRExC_state->code_blocks[n];
5500 dst->start = src->start + offset;
5501 dst->end = src->end + offset;
5502 dst->block = src->block;
5503 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5515 /* handle bare (possibly after overloading) regex: foo =~ $re */
5520 if (SvTYPE(re) == SVt_REGEXP) {
5524 Safefree(pRExC_state->code_blocks);
5525 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5526 "Precompiled pattern%s\n",
5527 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5534 exp = SvPV_nomg(pat, plen);
5537 if (!eng->op_comp) {
5538 if ((SvUTF8(pat) && IN_BYTES)
5539 || SvGMAGICAL(pat) || SvAMAGIC(pat))
5541 /* make a temporary copy; either to convert to bytes,
5542 * or to avoid repeating get-magic / overloaded stringify */
5543 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
5544 (IN_BYTES ? 0 : SvUTF8(pat)));
5546 Safefree(pRExC_state->code_blocks);
5547 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
5550 /* ignore the utf8ness if the pattern is 0 length */
5551 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
5552 RExC_uni_semantics = 0;
5553 RExC_contains_locale = 0;
5554 pRExC_state->runtime_code_qr = NULL;
5557 SV *dsv= sv_newmortal();
5558 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
5559 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
5560 PL_colors[4],PL_colors[5],s);
5566 U8 *const src = (U8*)exp;
5569 STRLEN s = 0, d = 0;
5572 /* It's possible to write a regexp in ascii that represents Unicode
5573 codepoints outside of the byte range, such as via \x{100}. If we
5574 detect such a sequence we have to convert the entire pattern to utf8
5575 and then recompile, as our sizing calculation will have been based
5576 on 1 byte == 1 character, but we will need to use utf8 to encode
5577 at least some part of the pattern, and therefore must convert the whole
5580 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5581 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
5583 /* upgrade pattern to UTF8, and if there are code blocks,
5584 * recalculate the indices.
5585 * This is essentially an unrolled Perl_bytes_to_utf8() */
5587 Newx(dst, plen * 2 + 1, U8);
5590 const UV uv = NATIVE_TO_ASCII(src[s]);
5591 if (UNI_IS_INVARIANT(uv))
5592 dst[d] = (U8)UTF_TO_NATIVE(uv);
5594 dst[d++] = (U8)UTF8_EIGHT_BIT_HI(uv);
5595 dst[d] = (U8)UTF8_EIGHT_BIT_LO(uv);
5597 if (n < pRExC_state->num_code_blocks) {
5598 if (!do_end && pRExC_state->code_blocks[n].start == s) {
5599 pRExC_state->code_blocks[n].start = d;
5600 assert(dst[d] == '(');
5603 else if (do_end && pRExC_state->code_blocks[n].end == s) {
5604 pRExC_state->code_blocks[n].end = d;
5605 assert(dst[d] == ')');
5618 RExC_orig_utf8 = RExC_utf8 = 1;
5622 if ((pm_flags & PMf_USE_RE_EVAL)
5623 /* this second condition covers the non-regex literal case,
5624 * i.e. $foo =~ '(?{})'. */
5625 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
5627 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
5629 /* return old regex if pattern hasn't changed */
5630 /* XXX: note in the below we have to check the flags as well as the pattern.
5632 * Things get a touch tricky as we have to compare the utf8 flag independently
5633 * from the compile flags.
5638 && !!RX_UTF8(old_re) == !!RExC_utf8
5639 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
5640 && RX_PRECOMP(old_re)
5641 && RX_PRELEN(old_re) == plen
5642 && memEQ(RX_PRECOMP(old_re), exp, plen)
5643 && !runtime_code /* with runtime code, always recompile */ )
5645 Safefree(pRExC_state->code_blocks);
5649 rx_flags = orig_rx_flags;
5651 if (initial_charset == REGEX_LOCALE_CHARSET) {
5652 RExC_contains_locale = 1;
5654 else if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
5656 /* Set to use unicode semantics if the pattern is in utf8 and has the
5657 * 'depends' charset specified, as it means unicode when utf8 */
5658 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5662 RExC_flags = rx_flags;
5663 RExC_pm_flags = pm_flags;
5666 if (TAINTING_get && TAINT_get)
5667 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
5669 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
5670 /* whoops, we have a non-utf8 pattern, whilst run-time code
5671 * got compiled as utf8. Try again with a utf8 pattern */
5672 goto redo_first_pass;
5675 assert(!pRExC_state->runtime_code_qr);
5680 RExC_in_lookbehind = 0;
5681 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
5683 RExC_override_recoding = 0;
5684 RExC_in_multi_char_class = 0;
5686 /* First pass: determine size, legality. */
5694 RExC_emit = &PL_regdummy;
5695 RExC_whilem_seen = 0;
5696 RExC_open_parens = NULL;
5697 RExC_close_parens = NULL;
5699 RExC_paren_names = NULL;
5701 RExC_paren_name_list = NULL;
5703 RExC_recurse = NULL;
5704 RExC_recurse_count = 0;
5705 pRExC_state->code_index = 0;
5707 #if 0 /* REGC() is (currently) a NOP at the first pass.
5708 * Clever compilers notice this and complain. --jhi */
5709 REGC((U8)REG_MAGIC, (char*)RExC_emit);
5712 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
5714 RExC_lastparse=NULL;
5716 /* reg may croak on us, not giving us a chance to free
5717 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
5718 need it to survive as long as the regexp (qr/(?{})/).
5719 We must check that code_blocksv is not already set, because we may
5720 have jumped back to restart the sizing pass. */
5721 if (pRExC_state->code_blocks && !code_blocksv) {
5722 code_blocksv = newSV_type(SVt_PV);
5723 SAVEFREESV(code_blocksv);
5724 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
5725 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
5727 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5728 if (flags & RESTART_UTF8) {
5729 goto redo_first_pass;
5731 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#X", flags);
5734 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
5737 PerlIO_printf(Perl_debug_log,
5738 "Required size %"IVdf" nodes\n"
5739 "Starting second pass (creation)\n",
5742 RExC_lastparse=NULL;
5745 /* The first pass could have found things that force Unicode semantics */
5746 if ((RExC_utf8 || RExC_uni_semantics)
5747 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
5749 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5752 /* Small enough for pointer-storage convention?
5753 If extralen==0, this means that we will not need long jumps. */
5754 if (RExC_size >= 0x10000L && RExC_extralen)
5755 RExC_size += RExC_extralen;
5758 if (RExC_whilem_seen > 15)
5759 RExC_whilem_seen = 15;
5761 /* Allocate space and zero-initialize. Note, the two step process
5762 of zeroing when in debug mode, thus anything assigned has to
5763 happen after that */
5764 rx = (REGEXP*) newSV_type(SVt_REGEXP);
5766 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
5767 char, regexp_internal);
5768 if ( r == NULL || ri == NULL )
5769 FAIL("Regexp out of space");
5771 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
5772 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode), char);
5774 /* bulk initialize base fields with 0. */
5775 Zero(ri, sizeof(regexp_internal), char);
5778 /* non-zero initialization begins here */
5781 r->extflags = rx_flags;
5782 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
5784 if (pm_flags & PMf_IS_QR) {
5785 ri->code_blocks = pRExC_state->code_blocks;
5786 ri->num_code_blocks = pRExC_state->num_code_blocks;
5791 for (n = 0; n < pRExC_state->num_code_blocks; n++)
5792 if (pRExC_state->code_blocks[n].src_regex)
5793 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
5794 SAVEFREEPV(pRExC_state->code_blocks);
5798 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
5799 bool has_charset = (get_regex_charset(r->extflags) != REGEX_DEPENDS_CHARSET);
5801 /* The caret is output if there are any defaults: if not all the STD
5802 * flags are set, or if no character set specifier is needed */
5804 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
5806 bool has_runon = ((RExC_seen & REG_SEEN_RUN_ON_COMMENT)==REG_SEEN_RUN_ON_COMMENT);
5807 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
5808 >> RXf_PMf_STD_PMMOD_SHIFT);
5809 const char *fptr = STD_PAT_MODS; /*"msix"*/
5811 /* Allocate for the worst case, which is all the std flags are turned
5812 * on. If more precision is desired, we could do a population count of
5813 * the flags set. This could be done with a small lookup table, or by
5814 * shifting, masking and adding, or even, when available, assembly
5815 * language for a machine-language population count.
5816 * We never output a minus, as all those are defaults, so are
5817 * covered by the caret */
5818 const STRLEN wraplen = plen + has_p + has_runon
5819 + has_default /* If needs a caret */
5821 /* If needs a character set specifier */
5822 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
5823 + (sizeof(STD_PAT_MODS) - 1)
5824 + (sizeof("(?:)") - 1);
5826 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
5827 r->xpv_len_u.xpvlenu_pv = p;
5829 SvFLAGS(rx) |= SVf_UTF8;
5832 /* If a default, cover it using the caret */
5834 *p++= DEFAULT_PAT_MOD;
5838 const char* const name = get_regex_charset_name(r->extflags, &len);
5839 Copy(name, p, len, char);
5843 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
5846 while((ch = *fptr++)) {
5854 Copy(RExC_precomp, p, plen, char);
5855 assert ((RX_WRAPPED(rx) - p) < 16);
5856 r->pre_prefix = p - RX_WRAPPED(rx);
5862 SvCUR_set(rx, p - RX_WRAPPED(rx));
5866 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
5868 if (RExC_seen & REG_SEEN_RECURSE) {
5869 Newxz(RExC_open_parens, RExC_npar,regnode *);
5870 SAVEFREEPV(RExC_open_parens);
5871 Newxz(RExC_close_parens,RExC_npar,regnode *);
5872 SAVEFREEPV(RExC_close_parens);
5875 /* Useful during FAIL. */
5876 #ifdef RE_TRACK_PATTERN_OFFSETS
5877 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
5878 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
5879 "%s %"UVuf" bytes for offset annotations.\n",
5880 ri->u.offsets ? "Got" : "Couldn't get",
5881 (UV)((2*RExC_size+1) * sizeof(U32))));
5883 SetProgLen(ri,RExC_size);
5888 /* Second pass: emit code. */
5889 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
5890 RExC_pm_flags = pm_flags;
5895 RExC_emit_start = ri->program;
5896 RExC_emit = ri->program;
5897 RExC_emit_bound = ri->program + RExC_size + 1;
5898 pRExC_state->code_index = 0;
5900 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
5901 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5903 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#X", flags);
5905 /* XXXX To minimize changes to RE engine we always allocate
5906 3-units-long substrs field. */
5907 Newx(r->substrs, 1, struct reg_substr_data);
5908 if (RExC_recurse_count) {
5909 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
5910 SAVEFREEPV(RExC_recurse);
5914 r->minlen = minlen = sawlookahead = sawplus = sawopen = 0;
5915 Zero(r->substrs, 1, struct reg_substr_data);
5917 #ifdef TRIE_STUDY_OPT
5919 StructCopy(&zero_scan_data, &data, scan_data_t);
5920 copyRExC_state = RExC_state;
5923 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
5925 RExC_state = copyRExC_state;
5926 if (seen & REG_TOP_LEVEL_BRANCHES)
5927 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
5929 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES;
5930 StructCopy(&zero_scan_data, &data, scan_data_t);
5933 StructCopy(&zero_scan_data, &data, scan_data_t);
5936 /* Dig out information for optimizations. */
5937 r->extflags = RExC_flags; /* was pm_op */
5938 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
5941 SvUTF8_on(rx); /* Unicode in it? */
5942 ri->regstclass = NULL;
5943 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
5944 r->intflags |= PREGf_NAUGHTY;
5945 scan = ri->program + 1; /* First BRANCH. */
5947 /* testing for BRANCH here tells us whether there is "must appear"
5948 data in the pattern. If there is then we can use it for optimisations */
5949 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES)) { /* Only one top-level choice. */
5951 STRLEN longest_float_length, longest_fixed_length;
5952 struct regnode_charclass_class ch_class; /* pointed to by data */
5954 I32 last_close = 0; /* pointed to by data */
5955 regnode *first= scan;
5956 regnode *first_next= regnext(first);
5958 * Skip introductions and multiplicators >= 1
5959 * so that we can extract the 'meat' of the pattern that must
5960 * match in the large if() sequence following.
5961 * NOTE that EXACT is NOT covered here, as it is normally
5962 * picked up by the optimiser separately.
5964 * This is unfortunate as the optimiser isnt handling lookahead
5965 * properly currently.
5968 while ((OP(first) == OPEN && (sawopen = 1)) ||
5969 /* An OR of *one* alternative - should not happen now. */
5970 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
5971 /* for now we can't handle lookbehind IFMATCH*/
5972 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
5973 (OP(first) == PLUS) ||
5974 (OP(first) == MINMOD) ||
5975 /* An {n,m} with n>0 */
5976 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
5977 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
5980 * the only op that could be a regnode is PLUS, all the rest
5981 * will be regnode_1 or regnode_2.
5984 if (OP(first) == PLUS)
5987 first += regarglen[OP(first)];
5989 first = NEXTOPER(first);
5990 first_next= regnext(first);
5993 /* Starting-point info. */
5995 DEBUG_PEEP("first:",first,0);
5996 /* Ignore EXACT as we deal with it later. */
5997 if (PL_regkind[OP(first)] == EXACT) {
5998 if (OP(first) == EXACT)
5999 NOOP; /* Empty, get anchored substr later. */
6001 ri->regstclass = first;
6004 else if (PL_regkind[OP(first)] == TRIE &&
6005 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6008 /* this can happen only on restudy */
6009 if ( OP(first) == TRIE ) {
6010 struct regnode_1 *trieop = (struct regnode_1 *)
6011 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6012 StructCopy(first,trieop,struct regnode_1);
6013 trie_op=(regnode *)trieop;
6015 struct regnode_charclass *trieop = (struct regnode_charclass *)
6016 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6017 StructCopy(first,trieop,struct regnode_charclass);
6018 trie_op=(regnode *)trieop;
6021 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6022 ri->regstclass = trie_op;
6025 else if (REGNODE_SIMPLE(OP(first)))
6026 ri->regstclass = first;
6027 else if (PL_regkind[OP(first)] == BOUND ||
6028 PL_regkind[OP(first)] == NBOUND)
6029 ri->regstclass = first;
6030 else if (PL_regkind[OP(first)] == BOL) {
6031 r->extflags |= (OP(first) == MBOL
6033 : (OP(first) == SBOL
6036 first = NEXTOPER(first);
6039 else if (OP(first) == GPOS) {
6040 r->extflags |= RXf_ANCH_GPOS;
6041 first = NEXTOPER(first);
6044 else if ((!sawopen || !RExC_sawback) &&
6045 (OP(first) == STAR &&
6046 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6047 !(r->extflags & RXf_ANCH) && !pRExC_state->num_code_blocks)
6049 /* turn .* into ^.* with an implied $*=1 */
6051 (OP(NEXTOPER(first)) == REG_ANY)
6054 r->extflags |= type;
6055 r->intflags |= PREGf_IMPLICIT;
6056 first = NEXTOPER(first);
6059 if (sawplus && !sawlookahead && (!sawopen || !RExC_sawback)
6060 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6061 /* x+ must match at the 1st pos of run of x's */
6062 r->intflags |= PREGf_SKIP;
6064 /* Scan is after the zeroth branch, first is atomic matcher. */
6065 #ifdef TRIE_STUDY_OPT
6068 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6069 (IV)(first - scan + 1))
6073 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6074 (IV)(first - scan + 1))
6080 * If there's something expensive in the r.e., find the
6081 * longest literal string that must appear and make it the
6082 * regmust. Resolve ties in favor of later strings, since
6083 * the regstart check works with the beginning of the r.e.
6084 * and avoiding duplication strengthens checking. Not a
6085 * strong reason, but sufficient in the absence of others.
6086 * [Now we resolve ties in favor of the earlier string if
6087 * it happens that c_offset_min has been invalidated, since the
6088 * earlier string may buy us something the later one won't.]
6091 data.longest_fixed = newSVpvs("");
6092 data.longest_float = newSVpvs("");
6093 data.last_found = newSVpvs("");
6094 data.longest = &(data.longest_fixed);
6095 ENTER_with_name("study_chunk");
6096 SAVEFREESV(data.longest_fixed);
6097 SAVEFREESV(data.longest_float);
6098 SAVEFREESV(data.last_found);
6100 if (!ri->regstclass) {
6101 cl_init(pRExC_state, &ch_class);
6102 data.start_class = &ch_class;
6103 stclass_flag = SCF_DO_STCLASS_AND;
6104 } else /* XXXX Check for BOUND? */
6106 data.last_closep = &last_close;
6108 minlen = study_chunk(pRExC_state, &first, &minlen, &fake, scan + RExC_size, /* Up to end */
6109 &data, -1, NULL, NULL,
6110 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag,0);
6113 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6116 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6117 && data.last_start_min == 0 && data.last_end > 0
6118 && !RExC_seen_zerolen
6119 && !(RExC_seen & REG_SEEN_VERBARG)
6120 && (!(RExC_seen & REG_SEEN_GPOS) || (r->extflags & RXf_ANCH_GPOS)))
6121 r->extflags |= RXf_CHECK_ALL;
6122 scan_commit(pRExC_state, &data,&minlen,0);
6124 longest_float_length = CHR_SVLEN(data.longest_float);
6126 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6127 && data.offset_fixed == data.offset_float_min
6128 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6129 && S_setup_longest (aTHX_ pRExC_state,
6133 &(r->float_end_shift),
6134 data.lookbehind_float,
6135 data.offset_float_min,
6137 longest_float_length,
6138 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6139 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6141 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6142 r->float_max_offset = data.offset_float_max;
6143 if (data.offset_float_max < I32_MAX) /* Don't offset infinity */
6144 r->float_max_offset -= data.lookbehind_float;
6145 SvREFCNT_inc_simple_void_NN(data.longest_float);
6148 r->float_substr = r->float_utf8 = NULL;
6149 longest_float_length = 0;
6152 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6154 if (S_setup_longest (aTHX_ pRExC_state,
6156 &(r->anchored_utf8),
6157 &(r->anchored_substr),
6158 &(r->anchored_end_shift),
6159 data.lookbehind_fixed,
6162 longest_fixed_length,
6163 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6164 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6166 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6167 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6170 r->anchored_substr = r->anchored_utf8 = NULL;
6171 longest_fixed_length = 0;
6173 LEAVE_with_name("study_chunk");
6176 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6177 ri->regstclass = NULL;
6179 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6181 && ! TEST_SSC_EOS(data.start_class)
6182 && !cl_is_anything(data.start_class))
6184 const U32 n = add_data(pRExC_state, 1, "f");
6185 OP(data.start_class) = ANYOF_SYNTHETIC;
6187 Newx(RExC_rxi->data->data[n], 1,
6188 struct regnode_charclass_class);
6189 StructCopy(data.start_class,
6190 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6191 struct regnode_charclass_class);
6192 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6193 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6194 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6195 regprop(r, sv, (regnode*)data.start_class);
6196 PerlIO_printf(Perl_debug_log,
6197 "synthetic stclass \"%s\".\n",
6198 SvPVX_const(sv));});
6201 /* A temporary algorithm prefers floated substr to fixed one to dig more info. */
6202 if (longest_fixed_length > longest_float_length) {
6203 r->check_end_shift = r->anchored_end_shift;
6204 r->check_substr = r->anchored_substr;
6205 r->check_utf8 = r->anchored_utf8;
6206 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6207 if (r->extflags & RXf_ANCH_SINGLE)
6208 r->extflags |= RXf_NOSCAN;
6211 r->check_end_shift = r->float_end_shift;
6212 r->check_substr = r->float_substr;
6213 r->check_utf8 = r->float_utf8;
6214 r->check_offset_min = r->float_min_offset;
6215 r->check_offset_max = r->float_max_offset;
6217 /* XXXX Currently intuiting is not compatible with ANCH_GPOS.
6218 This should be changed ASAP! */
6219 if ((r->check_substr || r->check_utf8) && !(r->extflags & RXf_ANCH_GPOS)) {
6220 r->extflags |= RXf_USE_INTUIT;
6221 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6222 r->extflags |= RXf_INTUIT_TAIL;
6224 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6225 if ( (STRLEN)minlen < longest_float_length )
6226 minlen= longest_float_length;
6227 if ( (STRLEN)minlen < longest_fixed_length )
6228 minlen= longest_fixed_length;
6232 /* Several toplevels. Best we can is to set minlen. */
6234 struct regnode_charclass_class ch_class;
6237 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6239 scan = ri->program + 1;
6240 cl_init(pRExC_state, &ch_class);
6241 data.start_class = &ch_class;
6242 data.last_closep = &last_close;
6245 minlen = study_chunk(pRExC_state, &scan, &minlen, &fake, scan + RExC_size,
6246 &data, -1, NULL, NULL, SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS,0);
6248 CHECK_RESTUDY_GOTO_butfirst(NOOP);
6250 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
6251 = r->float_substr = r->float_utf8 = NULL;
6253 if (! TEST_SSC_EOS(data.start_class)
6254 && !cl_is_anything(data.start_class))
6256 const U32 n = add_data(pRExC_state, 1, "f");
6257 OP(data.start_class) = ANYOF_SYNTHETIC;
6259 Newx(RExC_rxi->data->data[n], 1,
6260 struct regnode_charclass_class);
6261 StructCopy(data.start_class,
6262 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6263 struct regnode_charclass_class);
6264 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6265 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6266 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
6267 regprop(r, sv, (regnode*)data.start_class);
6268 PerlIO_printf(Perl_debug_log,
6269 "synthetic stclass \"%s\".\n",
6270 SvPVX_const(sv));});
6274 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
6275 the "real" pattern. */
6277 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf"\n",
6278 (IV)minlen, (IV)r->minlen);
6280 r->minlenret = minlen;
6281 if (r->minlen < minlen)
6284 if (RExC_seen & REG_SEEN_GPOS)
6285 r->extflags |= RXf_GPOS_SEEN;
6286 if (RExC_seen & REG_SEEN_LOOKBEHIND)
6287 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the lookbehind */
6288 if (pRExC_state->num_code_blocks)
6289 r->extflags |= RXf_EVAL_SEEN;
6290 if (RExC_seen & REG_SEEN_CANY)
6291 r->extflags |= RXf_CANY_SEEN;
6292 if (RExC_seen & REG_SEEN_VERBARG)
6294 r->intflags |= PREGf_VERBARG_SEEN;
6295 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
6297 if (RExC_seen & REG_SEEN_CUTGROUP)
6298 r->intflags |= PREGf_CUTGROUP_SEEN;
6299 if (pm_flags & PMf_USE_RE_EVAL)
6300 r->intflags |= PREGf_USE_RE_EVAL;
6301 if (RExC_paren_names)
6302 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
6304 RXp_PAREN_NAMES(r) = NULL;
6307 regnode *first = ri->program + 1;
6309 regnode *next = NEXTOPER(first);
6312 if (PL_regkind[fop] == NOTHING && nop == END)
6313 r->extflags |= RXf_NULL;
6314 else if (PL_regkind[fop] == BOL && nop == END)
6315 r->extflags |= RXf_START_ONLY;
6316 else if (fop == PLUS && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE && OP(regnext(first)) == END)
6317 r->extflags |= RXf_WHITE;
6318 else if ( r->extflags & RXf_SPLIT && fop == EXACT && STR_LEN(first) == 1 && *(STRING(first)) == ' ' && OP(regnext(first)) == END )
6319 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
6323 if (RExC_paren_names) {
6324 ri->name_list_idx = add_data( pRExC_state, 1, "a" );
6325 ri->data->data[ri->name_list_idx] = (void*)SvREFCNT_inc(RExC_paren_name_list);
6328 ri->name_list_idx = 0;
6330 if (RExC_recurse_count) {
6331 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
6332 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
6333 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
6336 Newxz(r->offs, RExC_npar, regexp_paren_pair);
6337 /* assume we don't need to swap parens around before we match */
6340 PerlIO_printf(Perl_debug_log,"Final program:\n");
6343 #ifdef RE_TRACK_PATTERN_OFFSETS
6344 DEBUG_OFFSETS_r(if (ri->u.offsets) {
6345 const U32 len = ri->u.offsets[0];
6347 GET_RE_DEBUG_FLAGS_DECL;
6348 PerlIO_printf(Perl_debug_log, "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
6349 for (i = 1; i <= len; i++) {
6350 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
6351 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
6352 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
6354 PerlIO_printf(Perl_debug_log, "\n");
6359 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
6360 * by setting the regexp SV to readonly-only instead. If the
6361 * pattern's been recompiled, the USEDness should remain. */
6362 if (old_re && SvREADONLY(old_re))
6370 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
6373 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
6375 PERL_UNUSED_ARG(value);
6377 if (flags & RXapif_FETCH) {
6378 return reg_named_buff_fetch(rx, key, flags);
6379 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
6380 Perl_croak_no_modify();
6382 } else if (flags & RXapif_EXISTS) {
6383 return reg_named_buff_exists(rx, key, flags)
6386 } else if (flags & RXapif_REGNAMES) {
6387 return reg_named_buff_all(rx, flags);
6388 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
6389 return reg_named_buff_scalar(rx, flags);
6391 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
6397 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
6400 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
6401 PERL_UNUSED_ARG(lastkey);
6403 if (flags & RXapif_FIRSTKEY)
6404 return reg_named_buff_firstkey(rx, flags);
6405 else if (flags & RXapif_NEXTKEY)
6406 return reg_named_buff_nextkey(rx, flags);
6408 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter", (int)flags);
6414 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
6417 AV *retarray = NULL;
6419 struct regexp *const rx = ReANY(r);
6421 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
6423 if (flags & RXapif_ALL)
6426 if (rx && RXp_PAREN_NAMES(rx)) {
6427 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
6430 SV* sv_dat=HeVAL(he_str);
6431 I32 *nums=(I32*)SvPVX(sv_dat);
6432 for ( i=0; i<SvIVX(sv_dat); i++ ) {
6433 if ((I32)(rx->nparens) >= nums[i]
6434 && rx->offs[nums[i]].start != -1
6435 && rx->offs[nums[i]].end != -1)
6438 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
6443 ret = newSVsv(&PL_sv_undef);
6446 av_push(retarray, ret);
6449 return newRV_noinc(MUTABLE_SV(retarray));
6456 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
6459 struct regexp *const rx = ReANY(r);
6461 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
6463 if (rx && RXp_PAREN_NAMES(rx)) {
6464 if (flags & RXapif_ALL) {
6465 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
6467 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
6469 SvREFCNT_dec_NN(sv);
6481 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
6483 struct regexp *const rx = ReANY(r);
6485 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
6487 if ( rx && RXp_PAREN_NAMES(rx) ) {
6488 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
6490 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
6497 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
6499 struct regexp *const rx = ReANY(r);
6500 GET_RE_DEBUG_FLAGS_DECL;
6502 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
6504 if (rx && RXp_PAREN_NAMES(rx)) {
6505 HV *hv = RXp_PAREN_NAMES(rx);
6507 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6510 SV* sv_dat = HeVAL(temphe);
6511 I32 *nums = (I32*)SvPVX(sv_dat);
6512 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6513 if ((I32)(rx->lastparen) >= nums[i] &&
6514 rx->offs[nums[i]].start != -1 &&
6515 rx->offs[nums[i]].end != -1)
6521 if (parno || flags & RXapif_ALL) {
6522 return newSVhek(HeKEY_hek(temphe));
6530 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
6535 struct regexp *const rx = ReANY(r);
6537 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
6539 if (rx && RXp_PAREN_NAMES(rx)) {
6540 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
6541 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
6542 } else if (flags & RXapif_ONE) {
6543 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
6544 av = MUTABLE_AV(SvRV(ret));
6545 length = av_len(av);
6546 SvREFCNT_dec_NN(ret);
6547 return newSViv(length + 1);
6549 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar", (int)flags);
6553 return &PL_sv_undef;
6557 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
6559 struct regexp *const rx = ReANY(r);
6562 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
6564 if (rx && RXp_PAREN_NAMES(rx)) {
6565 HV *hv= RXp_PAREN_NAMES(rx);
6567 (void)hv_iterinit(hv);
6568 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6571 SV* sv_dat = HeVAL(temphe);
6572 I32 *nums = (I32*)SvPVX(sv_dat);
6573 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6574 if ((I32)(rx->lastparen) >= nums[i] &&
6575 rx->offs[nums[i]].start != -1 &&
6576 rx->offs[nums[i]].end != -1)
6582 if (parno || flags & RXapif_ALL) {
6583 av_push(av, newSVhek(HeKEY_hek(temphe)));
6588 return newRV_noinc(MUTABLE_SV(av));
6592 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
6595 struct regexp *const rx = ReANY(r);
6601 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
6603 if ( ( n == RX_BUFF_IDX_CARET_PREMATCH
6604 || n == RX_BUFF_IDX_CARET_FULLMATCH
6605 || n == RX_BUFF_IDX_CARET_POSTMATCH
6607 && !(rx->extflags & RXf_PMf_KEEPCOPY)
6614 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
6615 /* no need to distinguish between them any more */
6616 n = RX_BUFF_IDX_FULLMATCH;
6618 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
6619 && rx->offs[0].start != -1)
6621 /* $`, ${^PREMATCH} */
6622 i = rx->offs[0].start;
6626 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
6627 && rx->offs[0].end != -1)
6629 /* $', ${^POSTMATCH} */
6630 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
6631 i = rx->sublen + rx->suboffset - rx->offs[0].end;
6634 if ( 0 <= n && n <= (I32)rx->nparens &&
6635 (s1 = rx->offs[n].start) != -1 &&
6636 (t1 = rx->offs[n].end) != -1)
6638 /* $&, ${^MATCH}, $1 ... */
6640 s = rx->subbeg + s1 - rx->suboffset;
6645 assert(s >= rx->subbeg);
6646 assert(rx->sublen >= (s - rx->subbeg) + i );
6648 #if NO_TAINT_SUPPORT
6649 sv_setpvn(sv, s, i);
6651 const int oldtainted = TAINT_get;
6653 sv_setpvn(sv, s, i);
6654 TAINT_set(oldtainted);
6656 if ( (rx->extflags & RXf_CANY_SEEN)
6657 ? (RXp_MATCH_UTF8(rx)
6658 && (!i || is_utf8_string((U8*)s, i)))
6659 : (RXp_MATCH_UTF8(rx)) )
6666 if (RXp_MATCH_TAINTED(rx)) {
6667 if (SvTYPE(sv) >= SVt_PVMG) {
6668 MAGIC* const mg = SvMAGIC(sv);
6671 SvMAGIC_set(sv, mg->mg_moremagic);
6673 if ((mgt = SvMAGIC(sv))) {
6674 mg->mg_moremagic = mgt;
6675 SvMAGIC_set(sv, mg);
6686 sv_setsv(sv,&PL_sv_undef);
6692 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
6693 SV const * const value)
6695 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
6697 PERL_UNUSED_ARG(rx);
6698 PERL_UNUSED_ARG(paren);
6699 PERL_UNUSED_ARG(value);
6702 Perl_croak_no_modify();
6706 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
6709 struct regexp *const rx = ReANY(r);
6713 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
6715 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
6717 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
6718 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6722 case RX_BUFF_IDX_PREMATCH: /* $` */
6723 if (rx->offs[0].start != -1) {
6724 i = rx->offs[0].start;
6733 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
6734 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6736 case RX_BUFF_IDX_POSTMATCH: /* $' */
6737 if (rx->offs[0].end != -1) {
6738 i = rx->sublen - rx->offs[0].end;
6740 s1 = rx->offs[0].end;
6747 case RX_BUFF_IDX_CARET_FULLMATCH: /* ${^MATCH} */
6748 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6752 /* $& / ${^MATCH}, $1, $2, ... */
6754 if (paren <= (I32)rx->nparens &&
6755 (s1 = rx->offs[paren].start) != -1 &&
6756 (t1 = rx->offs[paren].end) != -1)
6762 if (ckWARN(WARN_UNINITIALIZED))
6763 report_uninit((const SV *)sv);
6768 if (i > 0 && RXp_MATCH_UTF8(rx)) {
6769 const char * const s = rx->subbeg - rx->suboffset + s1;
6774 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
6781 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
6783 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
6784 PERL_UNUSED_ARG(rx);
6788 return newSVpvs("Regexp");
6791 /* Scans the name of a named buffer from the pattern.
6792 * If flags is REG_RSN_RETURN_NULL returns null.
6793 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
6794 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
6795 * to the parsed name as looked up in the RExC_paren_names hash.
6796 * If there is an error throws a vFAIL().. type exception.
6799 #define REG_RSN_RETURN_NULL 0
6800 #define REG_RSN_RETURN_NAME 1
6801 #define REG_RSN_RETURN_DATA 2
6804 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
6806 char *name_start = RExC_parse;
6808 PERL_ARGS_ASSERT_REG_SCAN_NAME;
6810 if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
6811 /* skip IDFIRST by using do...while */
6814 RExC_parse += UTF8SKIP(RExC_parse);
6815 } while (isWORDCHAR_utf8((U8*)RExC_parse));
6819 } while (isWORDCHAR(*RExC_parse));
6821 RExC_parse++; /* so the <- from the vFAIL is after the offending character */
6822 vFAIL("Group name must start with a non-digit word character");
6826 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
6827 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
6828 if ( flags == REG_RSN_RETURN_NAME)
6830 else if (flags==REG_RSN_RETURN_DATA) {
6833 if ( ! sv_name ) /* should not happen*/
6834 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
6835 if (RExC_paren_names)
6836 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
6838 sv_dat = HeVAL(he_str);
6840 vFAIL("Reference to nonexistent named group");
6844 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
6845 (unsigned long) flags);
6847 assert(0); /* NOT REACHED */
6852 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
6853 int rem=(int)(RExC_end - RExC_parse); \
6862 if (RExC_lastparse!=RExC_parse) \
6863 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
6866 iscut ? "..." : "<" \
6869 PerlIO_printf(Perl_debug_log,"%16s",""); \
6872 num = RExC_size + 1; \
6874 num=REG_NODE_NUM(RExC_emit); \
6875 if (RExC_lastnum!=num) \
6876 PerlIO_printf(Perl_debug_log,"|%4d",num); \
6878 PerlIO_printf(Perl_debug_log,"|%4s",""); \
6879 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
6880 (int)((depth*2)), "", \
6884 RExC_lastparse=RExC_parse; \
6889 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
6890 DEBUG_PARSE_MSG((funcname)); \
6891 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
6893 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
6894 DEBUG_PARSE_MSG((funcname)); \
6895 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
6898 /* This section of code defines the inversion list object and its methods. The
6899 * interfaces are highly subject to change, so as much as possible is static to
6900 * this file. An inversion list is here implemented as a malloc'd C UV array
6901 * with some added info that is placed as UVs at the beginning in a header
6902 * portion. An inversion list for Unicode is an array of code points, sorted
6903 * by ordinal number. The zeroth element is the first code point in the list.
6904 * The 1th element is the first element beyond that not in the list. In other
6905 * words, the first range is
6906 * invlist[0]..(invlist[1]-1)
6907 * The other ranges follow. Thus every element whose index is divisible by two
6908 * marks the beginning of a range that is in the list, and every element not
6909 * divisible by two marks the beginning of a range not in the list. A single
6910 * element inversion list that contains the single code point N generally
6911 * consists of two elements
6914 * (The exception is when N is the highest representable value on the
6915 * machine, in which case the list containing just it would be a single
6916 * element, itself. By extension, if the last range in the list extends to
6917 * infinity, then the first element of that range will be in the inversion list
6918 * at a position that is divisible by two, and is the final element in the
6920 * Taking the complement (inverting) an inversion list is quite simple, if the
6921 * first element is 0, remove it; otherwise add a 0 element at the beginning.
6922 * This implementation reserves an element at the beginning of each inversion
6923 * list to contain 0 when the list contains 0, and contains 1 otherwise. The
6924 * actual beginning of the list is either that element if 0, or the next one if
6927 * More about inversion lists can be found in "Unicode Demystified"
6928 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
6929 * More will be coming when functionality is added later.
6931 * The inversion list data structure is currently implemented as an SV pointing
6932 * to an array of UVs that the SV thinks are bytes. This allows us to have an
6933 * array of UV whose memory management is automatically handled by the existing
6934 * facilities for SV's.
6936 * Some of the methods should always be private to the implementation, and some
6937 * should eventually be made public */
6939 /* The header definitions are in F<inline_invlist.c> */
6940 #define TO_INTERNAL_SIZE(x) (((x) + HEADER_LENGTH) * sizeof(UV))
6941 #define FROM_INTERNAL_SIZE(x) (((x)/ sizeof(UV)) - HEADER_LENGTH)
6943 #define INVLIST_INITIAL_LEN 10
6945 PERL_STATIC_INLINE UV*
6946 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
6948 /* Returns a pointer to the first element in the inversion list's array.
6949 * This is called upon initialization of an inversion list. Where the
6950 * array begins depends on whether the list has the code point U+0000
6951 * in it or not. The other parameter tells it whether the code that
6952 * follows this call is about to put a 0 in the inversion list or not.
6953 * The first element is either the element with 0, if 0, or the next one,
6956 UV* zero = get_invlist_zero_addr(invlist);
6958 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
6961 assert(! *_get_invlist_len_addr(invlist));
6963 /* 1^1 = 0; 1^0 = 1 */
6964 *zero = 1 ^ will_have_0;
6965 return zero + *zero;
6968 PERL_STATIC_INLINE UV*
6969 S_invlist_array(pTHX_ SV* const invlist)
6971 /* Returns the pointer to the inversion list's array. Every time the
6972 * length changes, this needs to be called in case malloc or realloc moved
6975 PERL_ARGS_ASSERT_INVLIST_ARRAY;
6977 /* Must not be empty. If these fail, you probably didn't check for <len>
6978 * being non-zero before trying to get the array */
6979 assert(*_get_invlist_len_addr(invlist));
6980 assert(*get_invlist_zero_addr(invlist) == 0
6981 || *get_invlist_zero_addr(invlist) == 1);
6983 /* The array begins either at the element reserved for zero if the
6984 * list contains 0 (that element will be set to 0), or otherwise the next
6985 * element (in which case the reserved element will be set to 1). */
6986 return (UV *) (get_invlist_zero_addr(invlist)
6987 + *get_invlist_zero_addr(invlist));
6990 PERL_STATIC_INLINE void
6991 S_invlist_set_len(pTHX_ SV* const invlist, const UV len)
6993 /* Sets the current number of elements stored in the inversion list */
6995 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
6997 *_get_invlist_len_addr(invlist) = len;
6999 assert(len <= SvLEN(invlist));
7001 SvCUR_set(invlist, TO_INTERNAL_SIZE(len));
7002 /* If the list contains U+0000, that element is part of the header,
7003 * and should not be counted as part of the array. It will contain
7004 * 0 in that case, and 1 otherwise. So we could flop 0=>1, 1=>0 and
7006 * SvCUR_set(invlist,
7007 * TO_INTERNAL_SIZE(len
7008 * - (*get_invlist_zero_addr(inv_list) ^ 1)));
7009 * But, this is only valid if len is not 0. The consequences of not doing
7010 * this is that the memory allocation code may think that 1 more UV is
7011 * being used than actually is, and so might do an unnecessary grow. That
7012 * seems worth not bothering to make this the precise amount.
7014 * Note that when inverting, SvCUR shouldn't change */
7017 PERL_STATIC_INLINE IV*
7018 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7020 /* Return the address of the UV that is reserved to hold the cached index
7023 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7025 return (IV *) (SvPVX(invlist) + (INVLIST_PREVIOUS_INDEX_OFFSET * sizeof (UV)));
7028 PERL_STATIC_INLINE IV
7029 S_invlist_previous_index(pTHX_ SV* const invlist)
7031 /* Returns cached index of previous search */
7033 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7035 return *get_invlist_previous_index_addr(invlist);
7038 PERL_STATIC_INLINE void
7039 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7041 /* Caches <index> for later retrieval */
7043 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7045 assert(index == 0 || index < (int) _invlist_len(invlist));
7047 *get_invlist_previous_index_addr(invlist) = index;
7050 PERL_STATIC_INLINE UV
7051 S_invlist_max(pTHX_ SV* const invlist)
7053 /* Returns the maximum number of elements storable in the inversion list's
7054 * array, without having to realloc() */
7056 PERL_ARGS_ASSERT_INVLIST_MAX;
7058 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7059 ? _invlist_len(invlist)
7060 : FROM_INTERNAL_SIZE(SvLEN(invlist));
7063 PERL_STATIC_INLINE UV*
7064 S_get_invlist_zero_addr(pTHX_ SV* invlist)
7066 /* Return the address of the UV that is reserved to hold 0 if the inversion
7067 * list contains 0. This has to be the last element of the heading, as the
7068 * list proper starts with either it if 0, or the next element if not.
7069 * (But we force it to contain either 0 or 1) */
7071 PERL_ARGS_ASSERT_GET_INVLIST_ZERO_ADDR;
7073 return (UV *) (SvPVX(invlist) + (INVLIST_ZERO_OFFSET * sizeof (UV)));
7076 #ifndef PERL_IN_XSUB_RE
7078 Perl__new_invlist(pTHX_ IV initial_size)
7081 /* Return a pointer to a newly constructed inversion list, with enough
7082 * space to store 'initial_size' elements. If that number is negative, a
7083 * system default is used instead */
7087 if (initial_size < 0) {
7088 initial_size = INVLIST_INITIAL_LEN;
7091 /* Allocate the initial space */
7092 new_list = newSV(TO_INTERNAL_SIZE(initial_size));
7093 invlist_set_len(new_list, 0);
7095 /* Force iterinit() to be used to get iteration to work */
7096 *get_invlist_iter_addr(new_list) = UV_MAX;
7098 /* This should force a segfault if a method doesn't initialize this
7100 *get_invlist_zero_addr(new_list) = UV_MAX;
7102 *get_invlist_previous_index_addr(new_list) = 0;
7103 *get_invlist_version_id_addr(new_list) = INVLIST_VERSION_ID;
7104 #if HEADER_LENGTH != 5
7105 # 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
7113 S__new_invlist_C_array(pTHX_ UV* list)
7115 /* Return a pointer to a newly constructed inversion list, initialized to
7116 * point to <list>, which has to be in the exact correct inversion list
7117 * form, including internal fields. Thus this is a dangerous routine that
7118 * should not be used in the wrong hands */
7120 SV* invlist = newSV_type(SVt_PV);
7122 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7124 SvPV_set(invlist, (char *) list);
7125 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7126 shouldn't touch it */
7127 SvCUR_set(invlist, TO_INTERNAL_SIZE(_invlist_len(invlist)));
7129 if (*get_invlist_version_id_addr(invlist) != INVLIST_VERSION_ID) {
7130 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7133 /* Initialize the iteration pointer.
7134 * XXX This could be done at compile time in charclass_invlists.h, but I
7135 * (khw) am not confident that the suffixes for specifying the C constant
7136 * UV_MAX are portable, e.g. 'ull' on a 32 bit machine that is configured
7137 * to use 64 bits; might need a Configure probe */
7138 invlist_iterfinish(invlist);
7144 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7146 /* Grow the maximum size of an inversion list */
7148 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7150 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max));
7153 PERL_STATIC_INLINE void
7154 S_invlist_trim(pTHX_ SV* const invlist)
7156 PERL_ARGS_ASSERT_INVLIST_TRIM;
7158 /* Change the length of the inversion list to how many entries it currently
7161 SvPV_shrink_to_cur((SV *) invlist);
7164 #define _invlist_union_complement_2nd(a, b, output) _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
7167 S__append_range_to_invlist(pTHX_ SV* const invlist, const UV start, const UV end)
7169 /* Subject to change or removal. Append the range from 'start' to 'end' at
7170 * the end of the inversion list. The range must be above any existing
7174 UV max = invlist_max(invlist);
7175 UV len = _invlist_len(invlist);
7177 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7179 if (len == 0) { /* Empty lists must be initialized */
7180 array = _invlist_array_init(invlist, start == 0);
7183 /* Here, the existing list is non-empty. The current max entry in the
7184 * list is generally the first value not in the set, except when the
7185 * set extends to the end of permissible values, in which case it is
7186 * the first entry in that final set, and so this call is an attempt to
7187 * append out-of-order */
7189 UV final_element = len - 1;
7190 array = invlist_array(invlist);
7191 if (array[final_element] > start
7192 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
7194 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",
7195 array[final_element], start,
7196 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
7199 /* Here, it is a legal append. If the new range begins with the first
7200 * value not in the set, it is extending the set, so the new first
7201 * value not in the set is one greater than the newly extended range.
7203 if (array[final_element] == start) {
7204 if (end != UV_MAX) {
7205 array[final_element] = end + 1;
7208 /* But if the end is the maximum representable on the machine,
7209 * just let the range that this would extend to have no end */
7210 invlist_set_len(invlist, len - 1);
7216 /* Here the new range doesn't extend any existing set. Add it */
7218 len += 2; /* Includes an element each for the start and end of range */
7220 /* If overflows the existing space, extend, which may cause the array to be
7223 invlist_extend(invlist, len);
7224 invlist_set_len(invlist, len); /* Have to set len here to avoid assert
7225 failure in invlist_array() */
7226 array = invlist_array(invlist);
7229 invlist_set_len(invlist, len);
7232 /* The next item on the list starts the range, the one after that is
7233 * one past the new range. */
7234 array[len - 2] = start;
7235 if (end != UV_MAX) {
7236 array[len - 1] = end + 1;
7239 /* But if the end is the maximum representable on the machine, just let
7240 * the range have no end */
7241 invlist_set_len(invlist, len - 1);
7245 #ifndef PERL_IN_XSUB_RE
7248 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
7250 /* Searches the inversion list for the entry that contains the input code
7251 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
7252 * return value is the index into the list's array of the range that
7257 IV high = _invlist_len(invlist);
7258 const IV highest_element = high - 1;
7261 PERL_ARGS_ASSERT__INVLIST_SEARCH;
7263 /* If list is empty, return failure. */
7268 /* (We can't get the array unless we know the list is non-empty) */
7269 array = invlist_array(invlist);
7271 mid = invlist_previous_index(invlist);
7272 assert(mid >=0 && mid <= highest_element);
7274 /* <mid> contains the cache of the result of the previous call to this
7275 * function (0 the first time). See if this call is for the same result,
7276 * or if it is for mid-1. This is under the theory that calls to this
7277 * function will often be for related code points that are near each other.
7278 * And benchmarks show that caching gives better results. We also test
7279 * here if the code point is within the bounds of the list. These tests
7280 * replace others that would have had to be made anyway to make sure that
7281 * the array bounds were not exceeded, and these give us extra information
7282 * at the same time */
7283 if (cp >= array[mid]) {
7284 if (cp >= array[highest_element]) {
7285 return highest_element;
7288 /* Here, array[mid] <= cp < array[highest_element]. This means that
7289 * the final element is not the answer, so can exclude it; it also
7290 * means that <mid> is not the final element, so can refer to 'mid + 1'
7292 if (cp < array[mid + 1]) {
7298 else { /* cp < aray[mid] */
7299 if (cp < array[0]) { /* Fail if outside the array */
7303 if (cp >= array[mid - 1]) {
7308 /* Binary search. What we are looking for is <i> such that
7309 * array[i] <= cp < array[i+1]
7310 * The loop below converges on the i+1. Note that there may not be an
7311 * (i+1)th element in the array, and things work nonetheless */
7312 while (low < high) {
7313 mid = (low + high) / 2;
7314 assert(mid <= highest_element);
7315 if (array[mid] <= cp) { /* cp >= array[mid] */
7318 /* We could do this extra test to exit the loop early.
7319 if (cp < array[low]) {
7324 else { /* cp < array[mid] */
7331 invlist_set_previous_index(invlist, high);
7336 Perl__invlist_populate_swatch(pTHX_ SV* const invlist, const UV start, const UV end, U8* swatch)
7338 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
7339 * but is used when the swash has an inversion list. This makes this much
7340 * faster, as it uses a binary search instead of a linear one. This is
7341 * intimately tied to that function, and perhaps should be in utf8.c,
7342 * except it is intimately tied to inversion lists as well. It assumes
7343 * that <swatch> is all 0's on input */
7346 const IV len = _invlist_len(invlist);
7350 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
7352 if (len == 0) { /* Empty inversion list */
7356 array = invlist_array(invlist);
7358 /* Find which element it is */
7359 i = _invlist_search(invlist, start);
7361 /* We populate from <start> to <end> */
7362 while (current < end) {
7365 /* The inversion list gives the results for every possible code point
7366 * after the first one in the list. Only those ranges whose index is
7367 * even are ones that the inversion list matches. For the odd ones,
7368 * and if the initial code point is not in the list, we have to skip
7369 * forward to the next element */
7370 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
7372 if (i >= len) { /* Finished if beyond the end of the array */
7376 if (current >= end) { /* Finished if beyond the end of what we
7378 if (LIKELY(end < UV_MAX)) {
7382 /* We get here when the upper bound is the maximum
7383 * representable on the machine, and we are looking for just
7384 * that code point. Have to special case it */
7386 goto join_end_of_list;
7389 assert(current >= start);
7391 /* The current range ends one below the next one, except don't go past
7394 upper = (i < len && array[i] < end) ? array[i] : end;
7396 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
7397 * for each code point in it */
7398 for (; current < upper; current++) {
7399 const STRLEN offset = (STRLEN)(current - start);
7400 swatch[offset >> 3] |= 1 << (offset & 7);
7405 /* Quit if at the end of the list */
7408 /* But first, have to deal with the highest possible code point on
7409 * the platform. The previous code assumes that <end> is one
7410 * beyond where we want to populate, but that is impossible at the
7411 * platform's infinity, so have to handle it specially */
7412 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
7414 const STRLEN offset = (STRLEN)(end - start);
7415 swatch[offset >> 3] |= 1 << (offset & 7);
7420 /* Advance to the next range, which will be for code points not in the
7429 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** output)
7431 /* Take the union of two inversion lists and point <output> to it. *output
7432 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7433 * the reference count to that list will be decremented. The first list,
7434 * <a>, may be NULL, in which case a copy of the second list is returned.
7435 * If <complement_b> is TRUE, the union is taken of the complement
7436 * (inversion) of <b> instead of b itself.
7438 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7439 * Richard Gillam, published by Addison-Wesley, and explained at some
7440 * length there. The preface says to incorporate its examples into your
7441 * code at your own risk.
7443 * The algorithm is like a merge sort.
7445 * XXX A potential performance improvement is to keep track as we go along
7446 * if only one of the inputs contributes to the result, meaning the other
7447 * is a subset of that one. In that case, we can skip the final copy and
7448 * return the larger of the input lists, but then outside code might need
7449 * to keep track of whether to free the input list or not */
7451 UV* array_a; /* a's array */
7453 UV len_a; /* length of a's array */
7456 SV* u; /* the resulting union */
7460 UV i_a = 0; /* current index into a's array */
7464 /* running count, as explained in the algorithm source book; items are
7465 * stopped accumulating and are output when the count changes to/from 0.
7466 * The count is incremented when we start a range that's in the set, and
7467 * decremented when we start a range that's not in the set. So its range
7468 * is 0 to 2. Only when the count is zero is something not in the set.
7472 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
7475 /* If either one is empty, the union is the other one */
7476 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
7483 *output = invlist_clone(b);
7485 _invlist_invert(*output);
7487 } /* else *output already = b; */
7490 else if ((len_b = _invlist_len(b)) == 0) {
7495 /* The complement of an empty list is a list that has everything in it,
7496 * so the union with <a> includes everything too */
7501 *output = _new_invlist(1);
7502 _append_range_to_invlist(*output, 0, UV_MAX);
7504 else if (*output != a) {
7505 *output = invlist_clone(a);
7507 /* else *output already = a; */
7511 /* Here both lists exist and are non-empty */
7512 array_a = invlist_array(a);
7513 array_b = invlist_array(b);
7515 /* If are to take the union of 'a' with the complement of b, set it
7516 * up so are looking at b's complement. */
7519 /* To complement, we invert: if the first element is 0, remove it. To
7520 * do this, we just pretend the array starts one later, and clear the
7521 * flag as we don't have to do anything else later */
7522 if (array_b[0] == 0) {
7525 complement_b = FALSE;
7529 /* But if the first element is not zero, we unshift a 0 before the
7530 * array. The data structure reserves a space for that 0 (which
7531 * should be a '1' right now), so physical shifting is unneeded,
7532 * but temporarily change that element to 0. Before exiting the
7533 * routine, we must restore the element to '1' */
7540 /* Size the union for the worst case: that the sets are completely
7542 u = _new_invlist(len_a + len_b);
7544 /* Will contain U+0000 if either component does */
7545 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
7546 || (len_b > 0 && array_b[0] == 0));
7548 /* Go through each list item by item, stopping when exhausted one of
7550 while (i_a < len_a && i_b < len_b) {
7551 UV cp; /* The element to potentially add to the union's array */
7552 bool cp_in_set; /* is it in the the input list's set or not */
7554 /* We need to take one or the other of the two inputs for the union.
7555 * Since we are merging two sorted lists, we take the smaller of the
7556 * next items. In case of a tie, we take the one that is in its set
7557 * first. If we took one not in the set first, it would decrement the
7558 * count, possibly to 0 which would cause it to be output as ending the
7559 * range, and the next time through we would take the same number, and
7560 * output it again as beginning the next range. By doing it the
7561 * opposite way, there is no possibility that the count will be
7562 * momentarily decremented to 0, and thus the two adjoining ranges will
7563 * be seamlessly merged. (In a tie and both are in the set or both not
7564 * in the set, it doesn't matter which we take first.) */
7565 if (array_a[i_a] < array_b[i_b]
7566 || (array_a[i_a] == array_b[i_b]
7567 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7569 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7573 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7574 cp = array_b[i_b++];
7577 /* Here, have chosen which of the two inputs to look at. Only output
7578 * if the running count changes to/from 0, which marks the
7579 * beginning/end of a range in that's in the set */
7582 array_u[i_u++] = cp;
7589 array_u[i_u++] = cp;
7594 /* Here, we are finished going through at least one of the lists, which
7595 * means there is something remaining in at most one. We check if the list
7596 * that hasn't been exhausted is positioned such that we are in the middle
7597 * of a range in its set or not. (i_a and i_b point to the element beyond
7598 * the one we care about.) If in the set, we decrement 'count'; if 0, there
7599 * is potentially more to output.
7600 * There are four cases:
7601 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
7602 * in the union is entirely from the non-exhausted set.
7603 * 2) Both were in their sets, count is 2. Nothing further should
7604 * be output, as everything that remains will be in the exhausted
7605 * list's set, hence in the union; decrementing to 1 but not 0 insures
7607 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
7608 * Nothing further should be output because the union includes
7609 * everything from the exhausted set. Not decrementing ensures that.
7610 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
7611 * decrementing to 0 insures that we look at the remainder of the
7612 * non-exhausted set */
7613 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7614 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7619 /* The final length is what we've output so far, plus what else is about to
7620 * be output. (If 'count' is non-zero, then the input list we exhausted
7621 * has everything remaining up to the machine's limit in its set, and hence
7622 * in the union, so there will be no further output. */
7625 /* At most one of the subexpressions will be non-zero */
7626 len_u += (len_a - i_a) + (len_b - i_b);
7629 /* Set result to final length, which can change the pointer to array_u, so
7631 if (len_u != _invlist_len(u)) {
7632 invlist_set_len(u, len_u);
7634 array_u = invlist_array(u);
7637 /* When 'count' is 0, the list that was exhausted (if one was shorter than
7638 * the other) ended with everything above it not in its set. That means
7639 * that the remaining part of the union is precisely the same as the
7640 * non-exhausted list, so can just copy it unchanged. (If both list were
7641 * exhausted at the same time, then the operations below will be both 0.)
7644 IV copy_count; /* At most one will have a non-zero copy count */
7645 if ((copy_count = len_a - i_a) > 0) {
7646 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
7648 else if ((copy_count = len_b - i_b) > 0) {
7649 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
7653 /* If we've changed b, restore it */
7658 /* We may be removing a reference to one of the inputs */
7659 if (a == *output || b == *output) {
7660 assert(! invlist_is_iterating(*output));
7661 SvREFCNT_dec_NN(*output);
7669 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** i)
7671 /* Take the intersection of two inversion lists and point <i> to it. *i
7672 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7673 * the reference count to that list will be decremented.
7674 * If <complement_b> is TRUE, the result will be the intersection of <a>
7675 * and the complement (or inversion) of <b> instead of <b> directly.
7677 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7678 * Richard Gillam, published by Addison-Wesley, and explained at some
7679 * length there. The preface says to incorporate its examples into your
7680 * code at your own risk. In fact, it had bugs
7682 * The algorithm is like a merge sort, and is essentially the same as the
7686 UV* array_a; /* a's array */
7688 UV len_a; /* length of a's array */
7691 SV* r; /* the resulting intersection */
7695 UV i_a = 0; /* current index into a's array */
7699 /* running count, as explained in the algorithm source book; items are
7700 * stopped accumulating and are output when the count changes to/from 2.
7701 * The count is incremented when we start a range that's in the set, and
7702 * decremented when we start a range that's not in the set. So its range
7703 * is 0 to 2. Only when the count is 2 is something in the intersection.
7707 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
7710 /* Special case if either one is empty */
7711 len_a = _invlist_len(a);
7712 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
7714 if (len_a != 0 && complement_b) {
7716 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
7717 * be empty. Here, also we are using 'b's complement, which hence
7718 * must be every possible code point. Thus the intersection is
7721 *i = invlist_clone(a);
7727 /* else *i is already 'a' */
7731 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
7732 * intersection must be empty */
7739 *i = _new_invlist(0);
7743 /* Here both lists exist and are non-empty */
7744 array_a = invlist_array(a);
7745 array_b = invlist_array(b);
7747 /* If are to take the intersection of 'a' with the complement of b, set it
7748 * up so are looking at b's complement. */
7751 /* To complement, we invert: if the first element is 0, remove it. To
7752 * do this, we just pretend the array starts one later, and clear the
7753 * flag as we don't have to do anything else later */
7754 if (array_b[0] == 0) {
7757 complement_b = FALSE;
7761 /* But if the first element is not zero, we unshift a 0 before the
7762 * array. The data structure reserves a space for that 0 (which
7763 * should be a '1' right now), so physical shifting is unneeded,
7764 * but temporarily change that element to 0. Before exiting the
7765 * routine, we must restore the element to '1' */
7772 /* Size the intersection for the worst case: that the intersection ends up
7773 * fragmenting everything to be completely disjoint */
7774 r= _new_invlist(len_a + len_b);
7776 /* Will contain U+0000 iff both components do */
7777 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
7778 && len_b > 0 && array_b[0] == 0);
7780 /* Go through each list item by item, stopping when exhausted one of
7782 while (i_a < len_a && i_b < len_b) {
7783 UV cp; /* The element to potentially add to the intersection's
7785 bool cp_in_set; /* Is it in the input list's set or not */
7787 /* We need to take one or the other of the two inputs for the
7788 * intersection. Since we are merging two sorted lists, we take the
7789 * smaller of the next items. In case of a tie, we take the one that
7790 * is not in its set first (a difference from the union algorithm). If
7791 * we took one in the set first, it would increment the count, possibly
7792 * to 2 which would cause it to be output as starting a range in the
7793 * intersection, and the next time through we would take that same
7794 * number, and output it again as ending the set. By doing it the
7795 * opposite of this, there is no possibility that the count will be
7796 * momentarily incremented to 2. (In a tie and both are in the set or
7797 * both not in the set, it doesn't matter which we take first.) */
7798 if (array_a[i_a] < array_b[i_b]
7799 || (array_a[i_a] == array_b[i_b]
7800 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7802 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7806 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7810 /* Here, have chosen which of the two inputs to look at. Only output
7811 * if the running count changes to/from 2, which marks the
7812 * beginning/end of a range that's in the intersection */
7816 array_r[i_r++] = cp;
7821 array_r[i_r++] = cp;
7827 /* Here, we are finished going through at least one of the lists, which
7828 * means there is something remaining in at most one. We check if the list
7829 * that has been exhausted is positioned such that we are in the middle
7830 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
7831 * the ones we care about.) There are four cases:
7832 * 1) Both weren't in their sets, count is 0, and remains 0. There's
7833 * nothing left in the intersection.
7834 * 2) Both were in their sets, count is 2 and perhaps is incremented to
7835 * above 2. What should be output is exactly that which is in the
7836 * non-exhausted set, as everything it has is also in the intersection
7837 * set, and everything it doesn't have can't be in the intersection
7838 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
7839 * gets incremented to 2. Like the previous case, the intersection is
7840 * everything that remains in the non-exhausted set.
7841 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
7842 * remains 1. And the intersection has nothing more. */
7843 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7844 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7849 /* The final length is what we've output so far plus what else is in the
7850 * intersection. At most one of the subexpressions below will be non-zero */
7853 len_r += (len_a - i_a) + (len_b - i_b);
7856 /* Set result to final length, which can change the pointer to array_r, so
7858 if (len_r != _invlist_len(r)) {
7859 invlist_set_len(r, len_r);
7861 array_r = invlist_array(r);
7864 /* Finish outputting any remaining */
7865 if (count >= 2) { /* At most one will have a non-zero copy count */
7867 if ((copy_count = len_a - i_a) > 0) {
7868 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
7870 else if ((copy_count = len_b - i_b) > 0) {
7871 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
7875 /* If we've changed b, restore it */
7880 /* We may be removing a reference to one of the inputs */
7881 if (a == *i || b == *i) {
7882 assert(! invlist_is_iterating(*i));
7883 SvREFCNT_dec_NN(*i);
7891 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
7893 /* Add the range from 'start' to 'end' inclusive to the inversion list's
7894 * set. A pointer to the inversion list is returned. This may actually be
7895 * a new list, in which case the passed in one has been destroyed. The
7896 * passed in inversion list can be NULL, in which case a new one is created
7897 * with just the one range in it */
7902 if (invlist == NULL) {
7903 invlist = _new_invlist(2);
7907 len = _invlist_len(invlist);
7910 /* If comes after the final entry actually in the list, can just append it
7913 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
7914 && start >= invlist_array(invlist)[len - 1]))
7916 _append_range_to_invlist(invlist, start, end);
7920 /* Here, can't just append things, create and return a new inversion list
7921 * which is the union of this range and the existing inversion list */
7922 range_invlist = _new_invlist(2);
7923 _append_range_to_invlist(range_invlist, start, end);
7925 _invlist_union(invlist, range_invlist, &invlist);
7927 /* The temporary can be freed */
7928 SvREFCNT_dec_NN(range_invlist);
7935 PERL_STATIC_INLINE SV*
7936 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
7937 return _add_range_to_invlist(invlist, cp, cp);
7940 #ifndef PERL_IN_XSUB_RE
7942 Perl__invlist_invert(pTHX_ SV* const invlist)
7944 /* Complement the input inversion list. This adds a 0 if the list didn't
7945 * have a zero; removes it otherwise. As described above, the data
7946 * structure is set up so that this is very efficient */
7948 UV* len_pos = _get_invlist_len_addr(invlist);
7950 PERL_ARGS_ASSERT__INVLIST_INVERT;
7952 assert(! invlist_is_iterating(invlist));
7954 /* The inverse of matching nothing is matching everything */
7955 if (*len_pos == 0) {
7956 _append_range_to_invlist(invlist, 0, UV_MAX);
7960 /* The exclusive or complents 0 to 1; and 1 to 0. If the result is 1, the
7961 * zero element was a 0, so it is being removed, so the length decrements
7962 * by 1; and vice-versa. SvCUR is unaffected */
7963 if (*get_invlist_zero_addr(invlist) ^= 1) {
7972 Perl__invlist_invert_prop(pTHX_ SV* const invlist)
7974 /* Complement the input inversion list (which must be a Unicode property,
7975 * all of which don't match above the Unicode maximum code point.) And
7976 * Perl has chosen to not have the inversion match above that either. This
7977 * adds a 0x110000 if the list didn't end with it, and removes it if it did
7983 PERL_ARGS_ASSERT__INVLIST_INVERT_PROP;
7985 _invlist_invert(invlist);
7987 len = _invlist_len(invlist);
7989 if (len != 0) { /* If empty do nothing */
7990 array = invlist_array(invlist);
7991 if (array[len - 1] != PERL_UNICODE_MAX + 1) {
7992 /* Add 0x110000. First, grow if necessary */
7994 if (invlist_max(invlist) < len) {
7995 invlist_extend(invlist, len);
7996 array = invlist_array(invlist);
7998 invlist_set_len(invlist, len);
7999 array[len - 1] = PERL_UNICODE_MAX + 1;
8001 else { /* Remove the 0x110000 */
8002 invlist_set_len(invlist, len - 1);
8010 PERL_STATIC_INLINE SV*
8011 S_invlist_clone(pTHX_ SV* const invlist)
8014 /* Return a new inversion list that is a copy of the input one, which is
8017 /* Need to allocate extra space to accommodate Perl's addition of a
8018 * trailing NUL to SvPV's, since it thinks they are always strings */
8019 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8020 STRLEN length = SvCUR(invlist);
8022 PERL_ARGS_ASSERT_INVLIST_CLONE;
8024 SvCUR_set(new_invlist, length); /* This isn't done automatically */
8025 Copy(SvPVX(invlist), SvPVX(new_invlist), length, char);
8030 PERL_STATIC_INLINE UV*
8031 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8033 /* Return the address of the UV that contains the current iteration
8036 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8038 return (UV *) (SvPVX(invlist) + (INVLIST_ITER_OFFSET * sizeof (UV)));
8041 PERL_STATIC_INLINE UV*
8042 S_get_invlist_version_id_addr(pTHX_ SV* invlist)
8044 /* Return the address of the UV that contains the version id. */
8046 PERL_ARGS_ASSERT_GET_INVLIST_VERSION_ID_ADDR;
8048 return (UV *) (SvPVX(invlist) + (INVLIST_VERSION_ID_OFFSET * sizeof (UV)));
8051 PERL_STATIC_INLINE void
8052 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8054 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8056 *get_invlist_iter_addr(invlist) = 0;
8059 PERL_STATIC_INLINE void
8060 S_invlist_iterfinish(pTHX_ SV* invlist)
8062 /* Terminate iterator for invlist. This is to catch development errors.
8063 * Any iteration that is interrupted before completed should call this
8064 * function. Functions that add code points anywhere else but to the end
8065 * of an inversion list assert that they are not in the middle of an
8066 * iteration. If they were, the addition would make the iteration
8067 * problematical: if the iteration hadn't reached the place where things
8068 * were being added, it would be ok */
8070 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8072 *get_invlist_iter_addr(invlist) = UV_MAX;
8076 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8078 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8079 * This call sets in <*start> and <*end>, the next range in <invlist>.
8080 * Returns <TRUE> if successful and the next call will return the next
8081 * range; <FALSE> if was already at the end of the list. If the latter,
8082 * <*start> and <*end> are unchanged, and the next call to this function
8083 * will start over at the beginning of the list */
8085 UV* pos = get_invlist_iter_addr(invlist);
8086 UV len = _invlist_len(invlist);
8089 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8092 *pos = UV_MAX; /* Force iterinit() to be required next time */
8096 array = invlist_array(invlist);
8098 *start = array[(*pos)++];
8104 *end = array[(*pos)++] - 1;
8110 PERL_STATIC_INLINE bool
8111 S_invlist_is_iterating(pTHX_ SV* const invlist)
8113 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8115 return *(get_invlist_iter_addr(invlist)) < UV_MAX;
8118 PERL_STATIC_INLINE UV
8119 S_invlist_highest(pTHX_ SV* const invlist)
8121 /* Returns the highest code point that matches an inversion list. This API
8122 * has an ambiguity, as it returns 0 under either the highest is actually
8123 * 0, or if the list is empty. If this distinction matters to you, check
8124 * for emptiness before calling this function */
8126 UV len = _invlist_len(invlist);
8129 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8135 array = invlist_array(invlist);
8137 /* The last element in the array in the inversion list always starts a
8138 * range that goes to infinity. That range may be for code points that are
8139 * matched in the inversion list, or it may be for ones that aren't
8140 * matched. In the latter case, the highest code point in the set is one
8141 * less than the beginning of this range; otherwise it is the final element
8142 * of this range: infinity */
8143 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8145 : array[len - 1] - 1;
8148 #ifndef PERL_IN_XSUB_RE
8150 Perl__invlist_contents(pTHX_ SV* const invlist)
8152 /* Get the contents of an inversion list into a string SV so that they can
8153 * be printed out. It uses the format traditionally done for debug tracing
8157 SV* output = newSVpvs("\n");
8159 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8161 assert(! invlist_is_iterating(invlist));
8163 invlist_iterinit(invlist);
8164 while (invlist_iternext(invlist, &start, &end)) {
8165 if (end == UV_MAX) {
8166 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8168 else if (end != start) {
8169 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8173 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
8181 #ifdef PERL_ARGS_ASSERT__INVLIST_DUMP
8183 Perl__invlist_dump(pTHX_ SV* const invlist, const char * const header)
8185 /* Dumps out the ranges in an inversion list. The string 'header'
8186 * if present is output on a line before the first range */
8190 PERL_ARGS_ASSERT__INVLIST_DUMP;
8192 if (header && strlen(header)) {
8193 PerlIO_printf(Perl_debug_log, "%s\n", header);
8195 if (invlist_is_iterating(invlist)) {
8196 PerlIO_printf(Perl_debug_log, "Can't dump because is in middle of iterating\n");
8200 invlist_iterinit(invlist);
8201 while (invlist_iternext(invlist, &start, &end)) {
8202 if (end == UV_MAX) {
8203 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. INFINITY\n", start);
8205 else if (end != start) {
8206 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. 0x%04"UVXf"\n",
8210 PerlIO_printf(Perl_debug_log, "0x%04"UVXf"\n", start);
8218 S__invlistEQ(pTHX_ SV* const a, SV* const b, bool complement_b)
8220 /* Return a boolean as to if the two passed in inversion lists are
8221 * identical. The final argument, if TRUE, says to take the complement of
8222 * the second inversion list before doing the comparison */
8224 UV* array_a = invlist_array(a);
8225 UV* array_b = invlist_array(b);
8226 UV len_a = _invlist_len(a);
8227 UV len_b = _invlist_len(b);
8229 UV i = 0; /* current index into the arrays */
8230 bool retval = TRUE; /* Assume are identical until proven otherwise */
8232 PERL_ARGS_ASSERT__INVLISTEQ;
8234 /* If are to compare 'a' with the complement of b, set it
8235 * up so are looking at b's complement. */
8238 /* The complement of nothing is everything, so <a> would have to have
8239 * just one element, starting at zero (ending at infinity) */
8241 return (len_a == 1 && array_a[0] == 0);
8243 else if (array_b[0] == 0) {
8245 /* Otherwise, to complement, we invert. Here, the first element is
8246 * 0, just remove it. To do this, we just pretend the array starts
8247 * one later, and clear the flag as we don't have to do anything
8252 complement_b = FALSE;
8256 /* But if the first element is not zero, we unshift a 0 before the
8257 * array. The data structure reserves a space for that 0 (which
8258 * should be a '1' right now), so physical shifting is unneeded,
8259 * but temporarily change that element to 0. Before exiting the
8260 * routine, we must restore the element to '1' */
8267 /* Make sure that the lengths are the same, as well as the final element
8268 * before looping through the remainder. (Thus we test the length, final,
8269 * and first elements right off the bat) */
8270 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
8273 else for (i = 0; i < len_a - 1; i++) {
8274 if (array_a[i] != array_b[i]) {
8287 #undef HEADER_LENGTH
8288 #undef INVLIST_INITIAL_LENGTH
8289 #undef TO_INTERNAL_SIZE
8290 #undef FROM_INTERNAL_SIZE
8291 #undef INVLIST_LEN_OFFSET
8292 #undef INVLIST_ZERO_OFFSET
8293 #undef INVLIST_ITER_OFFSET
8294 #undef INVLIST_VERSION_ID
8295 #undef INVLIST_PREVIOUS_INDEX_OFFSET
8297 /* End of inversion list object */
8300 S_parse_lparen_question_flags(pTHX_ struct RExC_state_t *pRExC_state)
8302 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
8303 * constructs, and updates RExC_flags with them. On input, RExC_parse
8304 * should point to the first flag; it is updated on output to point to the
8305 * final ')' or ':'. There needs to be at least one flag, or this will
8308 /* for (?g), (?gc), and (?o) warnings; warning
8309 about (?c) will warn about (?g) -- japhy */
8311 #define WASTED_O 0x01
8312 #define WASTED_G 0x02
8313 #define WASTED_C 0x04
8314 #define WASTED_GC (0x02|0x04)
8315 I32 wastedflags = 0x00;
8316 U32 posflags = 0, negflags = 0;
8317 U32 *flagsp = &posflags;
8318 char has_charset_modifier = '\0';
8320 bool has_use_defaults = FALSE;
8321 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
8323 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
8325 /* '^' as an initial flag sets certain defaults */
8326 if (UCHARAT(RExC_parse) == '^') {
8328 has_use_defaults = TRUE;
8329 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
8330 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
8331 ? REGEX_UNICODE_CHARSET
8332 : REGEX_DEPENDS_CHARSET);
8335 cs = get_regex_charset(RExC_flags);
8336 if (cs == REGEX_DEPENDS_CHARSET
8337 && (RExC_utf8 || RExC_uni_semantics))
8339 cs = REGEX_UNICODE_CHARSET;
8342 while (*RExC_parse) {
8343 /* && strchr("iogcmsx", *RExC_parse) */
8344 /* (?g), (?gc) and (?o) are useless here
8345 and must be globally applied -- japhy */
8346 switch (*RExC_parse) {
8348 /* Code for the imsx flags */
8349 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
8351 case LOCALE_PAT_MOD:
8352 if (has_charset_modifier) {
8353 goto excess_modifier;
8355 else if (flagsp == &negflags) {
8358 cs = REGEX_LOCALE_CHARSET;
8359 has_charset_modifier = LOCALE_PAT_MOD;
8360 RExC_contains_locale = 1;
8362 case UNICODE_PAT_MOD:
8363 if (has_charset_modifier) {
8364 goto excess_modifier;
8366 else if (flagsp == &negflags) {
8369 cs = REGEX_UNICODE_CHARSET;
8370 has_charset_modifier = UNICODE_PAT_MOD;
8372 case ASCII_RESTRICT_PAT_MOD:
8373 if (flagsp == &negflags) {
8376 if (has_charset_modifier) {
8377 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
8378 goto excess_modifier;
8380 /* Doubled modifier implies more restricted */
8381 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
8384 cs = REGEX_ASCII_RESTRICTED_CHARSET;
8386 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
8388 case DEPENDS_PAT_MOD:
8389 if (has_use_defaults) {
8390 goto fail_modifiers;
8392 else if (flagsp == &negflags) {
8395 else if (has_charset_modifier) {
8396 goto excess_modifier;
8399 /* The dual charset means unicode semantics if the
8400 * pattern (or target, not known until runtime) are
8401 * utf8, or something in the pattern indicates unicode
8403 cs = (RExC_utf8 || RExC_uni_semantics)
8404 ? REGEX_UNICODE_CHARSET
8405 : REGEX_DEPENDS_CHARSET;
8406 has_charset_modifier = DEPENDS_PAT_MOD;
8410 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
8411 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
8413 else if (has_charset_modifier == *(RExC_parse - 1)) {
8414 vFAIL2("Regexp modifier \"%c\" may not appear twice", *(RExC_parse - 1));
8417 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
8422 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"", *(RExC_parse - 1));
8424 case ONCE_PAT_MOD: /* 'o' */
8425 case GLOBAL_PAT_MOD: /* 'g' */
8426 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8427 const I32 wflagbit = *RExC_parse == 'o' ? WASTED_O : WASTED_G;
8428 if (! (wastedflags & wflagbit) ) {
8429 wastedflags |= wflagbit;
8432 "Useless (%s%c) - %suse /%c modifier",
8433 flagsp == &negflags ? "?-" : "?",
8435 flagsp == &negflags ? "don't " : "",
8442 case CONTINUE_PAT_MOD: /* 'c' */
8443 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8444 if (! (wastedflags & WASTED_C) ) {
8445 wastedflags |= WASTED_GC;
8448 "Useless (%sc) - %suse /gc modifier",
8449 flagsp == &negflags ? "?-" : "?",
8450 flagsp == &negflags ? "don't " : ""
8455 case KEEPCOPY_PAT_MOD: /* 'p' */
8456 if (flagsp == &negflags) {
8458 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
8460 *flagsp |= RXf_PMf_KEEPCOPY;
8464 /* A flag is a default iff it is following a minus, so
8465 * if there is a minus, it means will be trying to
8466 * re-specify a default which is an error */
8467 if (has_use_defaults || flagsp == &negflags) {
8468 goto fail_modifiers;
8471 wastedflags = 0; /* reset so (?g-c) warns twice */
8475 RExC_flags |= posflags;
8476 RExC_flags &= ~negflags;
8477 set_regex_charset(&RExC_flags, cs);
8483 vFAIL3("Sequence (%.*s...) not recognized",
8484 RExC_parse-seqstart, seqstart);
8493 - reg - regular expression, i.e. main body or parenthesized thing
8495 * Caller must absorb opening parenthesis.
8497 * Combining parenthesis handling with the base level of regular expression
8498 * is a trifle forced, but the need to tie the tails of the branches to what
8499 * follows makes it hard to avoid.
8501 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
8503 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
8505 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
8508 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
8509 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
8510 needs to be restarted.
8511 Otherwise would only return NULL if regbranch() returns NULL, which
8514 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
8515 /* paren: Parenthesized? 0=top, 1=(, inside: changed to letter. */
8518 regnode *ret; /* Will be the head of the group. */
8521 regnode *ender = NULL;
8524 U32 oregflags = RExC_flags;
8525 bool have_branch = 0;
8527 I32 freeze_paren = 0;
8528 I32 after_freeze = 0;
8530 char * parse_start = RExC_parse; /* MJD */
8531 char * const oregcomp_parse = RExC_parse;
8533 GET_RE_DEBUG_FLAGS_DECL;
8535 PERL_ARGS_ASSERT_REG;
8536 DEBUG_PARSE("reg ");
8538 *flagp = 0; /* Tentatively. */
8541 /* Make an OPEN node, if parenthesized. */
8543 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
8544 char *start_verb = RExC_parse;
8545 STRLEN verb_len = 0;
8546 char *start_arg = NULL;
8547 unsigned char op = 0;
8549 int internal_argval = 0; /* internal_argval is only useful if !argok */
8550 while ( *RExC_parse && *RExC_parse != ')' ) {
8551 if ( *RExC_parse == ':' ) {
8552 start_arg = RExC_parse + 1;
8558 verb_len = RExC_parse - start_verb;
8561 while ( *RExC_parse && *RExC_parse != ')' )
8563 if ( *RExC_parse != ')' )
8564 vFAIL("Unterminated verb pattern argument");
8565 if ( RExC_parse == start_arg )
8568 if ( *RExC_parse != ')' )
8569 vFAIL("Unterminated verb pattern");
8572 switch ( *start_verb ) {
8573 case 'A': /* (*ACCEPT) */
8574 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
8576 internal_argval = RExC_nestroot;
8579 case 'C': /* (*COMMIT) */
8580 if ( memEQs(start_verb,verb_len,"COMMIT") )
8583 case 'F': /* (*FAIL) */
8584 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
8589 case ':': /* (*:NAME) */
8590 case 'M': /* (*MARK:NAME) */
8591 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
8596 case 'P': /* (*PRUNE) */
8597 if ( memEQs(start_verb,verb_len,"PRUNE") )
8600 case 'S': /* (*SKIP) */
8601 if ( memEQs(start_verb,verb_len,"SKIP") )
8604 case 'T': /* (*THEN) */
8605 /* [19:06] <TimToady> :: is then */
8606 if ( memEQs(start_verb,verb_len,"THEN") ) {
8608 RExC_seen |= REG_SEEN_CUTGROUP;
8614 vFAIL3("Unknown verb pattern '%.*s'",
8615 verb_len, start_verb);
8618 if ( start_arg && internal_argval ) {
8619 vFAIL3("Verb pattern '%.*s' may not have an argument",
8620 verb_len, start_verb);
8621 } else if ( argok < 0 && !start_arg ) {
8622 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
8623 verb_len, start_verb);
8625 ret = reganode(pRExC_state, op, internal_argval);
8626 if ( ! internal_argval && ! SIZE_ONLY ) {
8628 SV *sv = newSVpvn( start_arg, RExC_parse - start_arg);
8629 ARG(ret) = add_data( pRExC_state, 1, "S" );
8630 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
8637 if (!internal_argval)
8638 RExC_seen |= REG_SEEN_VERBARG;
8639 } else if ( start_arg ) {
8640 vFAIL3("Verb pattern '%.*s' may not have an argument",
8641 verb_len, start_verb);
8643 ret = reg_node(pRExC_state, op);
8645 nextchar(pRExC_state);
8648 if (*RExC_parse == '?') { /* (?...) */
8649 bool is_logical = 0;
8650 const char * const seqstart = RExC_parse;
8653 paren = *RExC_parse++;
8654 ret = NULL; /* For look-ahead/behind. */
8657 case 'P': /* (?P...) variants for those used to PCRE/Python */
8658 paren = *RExC_parse++;
8659 if ( paren == '<') /* (?P<...>) named capture */
8661 else if (paren == '>') { /* (?P>name) named recursion */
8662 goto named_recursion;
8664 else if (paren == '=') { /* (?P=...) named backref */
8665 /* this pretty much dupes the code for \k<NAME> in regatom(), if
8666 you change this make sure you change that */
8667 char* name_start = RExC_parse;
8669 SV *sv_dat = reg_scan_name(pRExC_state,
8670 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8671 if (RExC_parse == name_start || *RExC_parse != ')')
8672 vFAIL2("Sequence %.3s... not terminated",parse_start);
8675 num = add_data( pRExC_state, 1, "S" );
8676 RExC_rxi->data->data[num]=(void*)sv_dat;
8677 SvREFCNT_inc_simple_void(sv_dat);
8680 ret = reganode(pRExC_state,
8683 : (ASCII_FOLD_RESTRICTED)
8685 : (AT_LEAST_UNI_SEMANTICS)
8693 Set_Node_Offset(ret, parse_start+1);
8694 Set_Node_Cur_Length(ret); /* MJD */
8696 nextchar(pRExC_state);
8700 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8702 case '<': /* (?<...) */
8703 if (*RExC_parse == '!')
8705 else if (*RExC_parse != '=')
8711 case '\'': /* (?'...') */
8712 name_start= RExC_parse;
8713 svname = reg_scan_name(pRExC_state,
8714 SIZE_ONLY ? /* reverse test from the others */
8715 REG_RSN_RETURN_NAME :
8716 REG_RSN_RETURN_NULL);
8717 if (RExC_parse == name_start) {
8719 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8722 if (*RExC_parse != paren)
8723 vFAIL2("Sequence (?%c... not terminated",
8724 paren=='>' ? '<' : paren);
8728 if (!svname) /* shouldn't happen */
8730 "panic: reg_scan_name returned NULL");
8731 if (!RExC_paren_names) {
8732 RExC_paren_names= newHV();
8733 sv_2mortal(MUTABLE_SV(RExC_paren_names));
8735 RExC_paren_name_list= newAV();
8736 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
8739 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
8741 sv_dat = HeVAL(he_str);
8743 /* croak baby croak */
8745 "panic: paren_name hash element allocation failed");
8746 } else if ( SvPOK(sv_dat) ) {
8747 /* (?|...) can mean we have dupes so scan to check
8748 its already been stored. Maybe a flag indicating
8749 we are inside such a construct would be useful,
8750 but the arrays are likely to be quite small, so
8751 for now we punt -- dmq */
8752 IV count = SvIV(sv_dat);
8753 I32 *pv = (I32*)SvPVX(sv_dat);
8755 for ( i = 0 ; i < count ; i++ ) {
8756 if ( pv[i] == RExC_npar ) {
8762 pv = (I32*)SvGROW(sv_dat, SvCUR(sv_dat) + sizeof(I32)+1);
8763 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
8764 pv[count] = RExC_npar;
8765 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
8768 (void)SvUPGRADE(sv_dat,SVt_PVNV);
8769 sv_setpvn(sv_dat, (char *)&(RExC_npar), sizeof(I32));
8771 SvIV_set(sv_dat, 1);
8774 /* Yes this does cause a memory leak in debugging Perls */
8775 if (!av_store(RExC_paren_name_list, RExC_npar, SvREFCNT_inc(svname)))
8776 SvREFCNT_dec_NN(svname);
8779 /*sv_dump(sv_dat);*/
8781 nextchar(pRExC_state);
8783 goto capturing_parens;
8785 RExC_seen |= REG_SEEN_LOOKBEHIND;
8786 RExC_in_lookbehind++;
8788 case '=': /* (?=...) */
8789 RExC_seen_zerolen++;
8791 case '!': /* (?!...) */
8792 RExC_seen_zerolen++;
8793 if (*RExC_parse == ')') {
8794 ret=reg_node(pRExC_state, OPFAIL);
8795 nextchar(pRExC_state);
8799 case '|': /* (?|...) */
8800 /* branch reset, behave like a (?:...) except that
8801 buffers in alternations share the same numbers */
8803 after_freeze = freeze_paren = RExC_npar;
8805 case ':': /* (?:...) */
8806 case '>': /* (?>...) */
8808 case '$': /* (?$...) */
8809 case '@': /* (?@...) */
8810 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
8812 case '0' : /* (?0) */
8813 case 'R' : /* (?R) */
8814 if (*RExC_parse != ')')
8815 FAIL("Sequence (?R) not terminated");
8816 ret = reg_node(pRExC_state, GOSTART);
8817 *flagp |= POSTPONED;
8818 nextchar(pRExC_state);
8821 { /* named and numeric backreferences */
8823 case '&': /* (?&NAME) */
8824 parse_start = RExC_parse - 1;
8827 SV *sv_dat = reg_scan_name(pRExC_state,
8828 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8829 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8831 goto gen_recurse_regop;
8832 assert(0); /* NOT REACHED */
8834 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8836 vFAIL("Illegal pattern");
8838 goto parse_recursion;
8840 case '-': /* (?-1) */
8841 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8842 RExC_parse--; /* rewind to let it be handled later */
8846 case '1': case '2': case '3': case '4': /* (?1) */
8847 case '5': case '6': case '7': case '8': case '9':
8850 num = atoi(RExC_parse);
8851 parse_start = RExC_parse - 1; /* MJD */
8852 if (*RExC_parse == '-')
8854 while (isDIGIT(*RExC_parse))
8856 if (*RExC_parse!=')')
8857 vFAIL("Expecting close bracket");
8860 if ( paren == '-' ) {
8862 Diagram of capture buffer numbering.
8863 Top line is the normal capture buffer numbers
8864 Bottom line is the negative indexing as from
8868 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
8872 num = RExC_npar + num;
8875 vFAIL("Reference to nonexistent group");
8877 } else if ( paren == '+' ) {
8878 num = RExC_npar + num - 1;
8881 ret = reganode(pRExC_state, GOSUB, num);
8883 if (num > (I32)RExC_rx->nparens) {
8885 vFAIL("Reference to nonexistent group");
8887 ARG2L_SET( ret, RExC_recurse_count++);
8889 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
8890 "Recurse #%"UVuf" to %"IVdf"\n", (UV)ARG(ret), (IV)ARG2L(ret)));
8894 RExC_seen |= REG_SEEN_RECURSE;
8895 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
8896 Set_Node_Offset(ret, parse_start); /* MJD */
8898 *flagp |= POSTPONED;
8899 nextchar(pRExC_state);
8901 } /* named and numeric backreferences */
8902 assert(0); /* NOT REACHED */
8904 case '?': /* (??...) */
8906 if (*RExC_parse != '{') {
8908 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8911 *flagp |= POSTPONED;
8912 paren = *RExC_parse++;
8914 case '{': /* (?{...}) */
8917 struct reg_code_block *cb;
8919 RExC_seen_zerolen++;
8921 if ( !pRExC_state->num_code_blocks
8922 || pRExC_state->code_index >= pRExC_state->num_code_blocks
8923 || pRExC_state->code_blocks[pRExC_state->code_index].start
8924 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
8927 if (RExC_pm_flags & PMf_USE_RE_EVAL)
8928 FAIL("panic: Sequence (?{...}): no code block found\n");
8929 FAIL("Eval-group not allowed at runtime, use re 'eval'");
8931 /* this is a pre-compiled code block (?{...}) */
8932 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
8933 RExC_parse = RExC_start + cb->end;
8936 if (cb->src_regex) {
8937 n = add_data(pRExC_state, 2, "rl");
8938 RExC_rxi->data->data[n] =
8939 (void*)SvREFCNT_inc((SV*)cb->src_regex);
8940 RExC_rxi->data->data[n+1] = (void*)o;
8943 n = add_data(pRExC_state, 1,
8944 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l");
8945 RExC_rxi->data->data[n] = (void*)o;
8948 pRExC_state->code_index++;
8949 nextchar(pRExC_state);
8953 ret = reg_node(pRExC_state, LOGICAL);
8954 eval = reganode(pRExC_state, EVAL, n);
8957 /* for later propagation into (??{}) return value */
8958 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
8960 REGTAIL(pRExC_state, ret, eval);
8961 /* deal with the length of this later - MJD */
8964 ret = reganode(pRExC_state, EVAL, n);
8965 Set_Node_Length(ret, RExC_parse - parse_start + 1);
8966 Set_Node_Offset(ret, parse_start);
8969 case '(': /* (?(?{...})...) and (?(?=...)...) */
8972 if (RExC_parse[0] == '?') { /* (?(?...)) */
8973 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
8974 || RExC_parse[1] == '<'
8975 || RExC_parse[1] == '{') { /* Lookahead or eval. */
8979 ret = reg_node(pRExC_state, LOGICAL);
8983 tail = reg(pRExC_state, 1, &flag, depth+1);
8984 if (flag & RESTART_UTF8) {
8985 *flagp = RESTART_UTF8;
8988 REGTAIL(pRExC_state, ret, tail);
8992 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
8993 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
8995 char ch = RExC_parse[0] == '<' ? '>' : '\'';
8996 char *name_start= RExC_parse++;
8998 SV *sv_dat=reg_scan_name(pRExC_state,
8999 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9000 if (RExC_parse == name_start || *RExC_parse != ch)
9001 vFAIL2("Sequence (?(%c... not terminated",
9002 (ch == '>' ? '<' : ch));
9005 num = add_data( pRExC_state, 1, "S" );
9006 RExC_rxi->data->data[num]=(void*)sv_dat;
9007 SvREFCNT_inc_simple_void(sv_dat);
9009 ret = reganode(pRExC_state,NGROUPP,num);
9010 goto insert_if_check_paren;
9012 else if (RExC_parse[0] == 'D' &&
9013 RExC_parse[1] == 'E' &&
9014 RExC_parse[2] == 'F' &&
9015 RExC_parse[3] == 'I' &&
9016 RExC_parse[4] == 'N' &&
9017 RExC_parse[5] == 'E')
9019 ret = reganode(pRExC_state,DEFINEP,0);
9022 goto insert_if_check_paren;
9024 else if (RExC_parse[0] == 'R') {
9027 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9028 parno = atoi(RExC_parse++);
9029 while (isDIGIT(*RExC_parse))
9031 } else if (RExC_parse[0] == '&') {
9034 sv_dat = reg_scan_name(pRExC_state,
9035 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9036 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9038 ret = reganode(pRExC_state,INSUBP,parno);
9039 goto insert_if_check_paren;
9041 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9044 parno = atoi(RExC_parse++);
9046 while (isDIGIT(*RExC_parse))
9048 ret = reganode(pRExC_state, GROUPP, parno);
9050 insert_if_check_paren:
9051 if ((c = *nextchar(pRExC_state)) != ')')
9052 vFAIL("Switch condition not recognized");
9054 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9055 br = regbranch(pRExC_state, &flags, 1,depth+1);
9057 if (flags & RESTART_UTF8) {
9058 *flagp = RESTART_UTF8;
9061 FAIL2("panic: regbranch returned NULL, flags=%#X",
9064 REGTAIL(pRExC_state, br, reganode(pRExC_state, LONGJMP, 0));
9065 c = *nextchar(pRExC_state);
9070 vFAIL("(?(DEFINE)....) does not allow branches");
9071 lastbr = reganode(pRExC_state, IFTHEN, 0); /* Fake one for optimizer. */
9072 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
9073 if (flags & RESTART_UTF8) {
9074 *flagp = RESTART_UTF8;
9077 FAIL2("panic: regbranch returned NULL, flags=%#X",
9080 REGTAIL(pRExC_state, ret, lastbr);
9083 c = *nextchar(pRExC_state);
9088 vFAIL("Switch (?(condition)... contains too many branches");
9089 ender = reg_node(pRExC_state, TAIL);
9090 REGTAIL(pRExC_state, br, ender);
9092 REGTAIL(pRExC_state, lastbr, ender);
9093 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
9096 REGTAIL(pRExC_state, ret, ender);
9097 RExC_size++; /* XXX WHY do we need this?!!
9098 For large programs it seems to be required
9099 but I can't figure out why. -- dmq*/
9103 vFAIL2("Unknown switch condition (?(%.2s", RExC_parse);
9106 case '[': /* (?[ ... ]) */
9107 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
9110 RExC_parse--; /* for vFAIL to print correctly */
9111 vFAIL("Sequence (? incomplete");
9113 default: /* e.g., (?i) */
9116 parse_lparen_question_flags(pRExC_state);
9117 if (UCHARAT(RExC_parse) != ':') {
9118 nextchar(pRExC_state);
9123 nextchar(pRExC_state);
9133 ret = reganode(pRExC_state, OPEN, parno);
9136 RExC_nestroot = parno;
9137 if (RExC_seen & REG_SEEN_RECURSE
9138 && !RExC_open_parens[parno-1])
9140 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9141 "Setting open paren #%"IVdf" to %d\n",
9142 (IV)parno, REG_NODE_NUM(ret)));
9143 RExC_open_parens[parno-1]= ret;
9146 Set_Node_Length(ret, 1); /* MJD */
9147 Set_Node_Offset(ret, RExC_parse); /* MJD */
9155 /* Pick up the branches, linking them together. */
9156 parse_start = RExC_parse; /* MJD */
9157 br = regbranch(pRExC_state, &flags, 1,depth+1);
9159 /* branch_len = (paren != 0); */
9162 if (flags & RESTART_UTF8) {
9163 *flagp = RESTART_UTF8;
9166 FAIL2("panic: regbranch returned NULL, flags=%#X", flags);
9168 if (*RExC_parse == '|') {
9169 if (!SIZE_ONLY && RExC_extralen) {
9170 reginsert(pRExC_state, BRANCHJ, br, depth+1);
9173 reginsert(pRExC_state, BRANCH, br, depth+1);
9174 Set_Node_Length(br, paren != 0);
9175 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
9179 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
9181 else if (paren == ':') {
9182 *flagp |= flags&SIMPLE;
9184 if (is_open) { /* Starts with OPEN. */
9185 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
9187 else if (paren != '?') /* Not Conditional */
9189 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9191 while (*RExC_parse == '|') {
9192 if (!SIZE_ONLY && RExC_extralen) {
9193 ender = reganode(pRExC_state, LONGJMP,0);
9194 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender); /* Append to the previous. */
9197 RExC_extralen += 2; /* Account for LONGJMP. */
9198 nextchar(pRExC_state);
9200 if (RExC_npar > after_freeze)
9201 after_freeze = RExC_npar;
9202 RExC_npar = freeze_paren;
9204 br = regbranch(pRExC_state, &flags, 0, depth+1);
9207 if (flags & RESTART_UTF8) {
9208 *flagp = RESTART_UTF8;
9211 FAIL2("panic: regbranch returned NULL, flags=%#X", flags);
9213 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
9215 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9218 if (have_branch || paren != ':') {
9219 /* Make a closing node, and hook it on the end. */
9222 ender = reg_node(pRExC_state, TAIL);
9225 ender = reganode(pRExC_state, CLOSE, parno);
9226 if (!SIZE_ONLY && RExC_seen & REG_SEEN_RECURSE) {
9227 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9228 "Setting close paren #%"IVdf" to %d\n",
9229 (IV)parno, REG_NODE_NUM(ender)));
9230 RExC_close_parens[parno-1]= ender;
9231 if (RExC_nestroot == parno)
9234 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
9235 Set_Node_Length(ender,1); /* MJD */
9241 *flagp &= ~HASWIDTH;
9244 ender = reg_node(pRExC_state, SUCCEED);
9247 ender = reg_node(pRExC_state, END);
9249 assert(!RExC_opend); /* there can only be one! */
9254 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9255 SV * const mysv_val1=sv_newmortal();
9256 SV * const mysv_val2=sv_newmortal();
9257 DEBUG_PARSE_MSG("lsbr");
9258 regprop(RExC_rx, mysv_val1, lastbr);
9259 regprop(RExC_rx, mysv_val2, ender);
9260 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9261 SvPV_nolen_const(mysv_val1),
9262 (IV)REG_NODE_NUM(lastbr),
9263 SvPV_nolen_const(mysv_val2),
9264 (IV)REG_NODE_NUM(ender),
9265 (IV)(ender - lastbr)
9268 REGTAIL(pRExC_state, lastbr, ender);
9270 if (have_branch && !SIZE_ONLY) {
9273 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
9275 /* Hook the tails of the branches to the closing node. */
9276 for (br = ret; br; br = regnext(br)) {
9277 const U8 op = PL_regkind[OP(br)];
9279 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
9280 if (OP(NEXTOPER(br)) != NOTHING || regnext(NEXTOPER(br)) != ender)
9283 else if (op == BRANCHJ) {
9284 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
9285 /* for now we always disable this optimisation * /
9286 if (OP(NEXTOPER(NEXTOPER(br))) != NOTHING || regnext(NEXTOPER(NEXTOPER(br))) != ender)
9292 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
9293 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9294 SV * const mysv_val1=sv_newmortal();
9295 SV * const mysv_val2=sv_newmortal();
9296 DEBUG_PARSE_MSG("NADA");
9297 regprop(RExC_rx, mysv_val1, ret);
9298 regprop(RExC_rx, mysv_val2, ender);
9299 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9300 SvPV_nolen_const(mysv_val1),
9301 (IV)REG_NODE_NUM(ret),
9302 SvPV_nolen_const(mysv_val2),
9303 (IV)REG_NODE_NUM(ender),
9308 if (OP(ender) == TAIL) {
9313 for ( opt= br + 1; opt < ender ; opt++ )
9315 NEXT_OFF(br)= ender - br;
9323 static const char parens[] = "=!<,>";
9325 if (paren && (p = strchr(parens, paren))) {
9326 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
9327 int flag = (p - parens) > 1;
9330 node = SUSPEND, flag = 0;
9331 reginsert(pRExC_state, node,ret, depth+1);
9332 Set_Node_Cur_Length(ret);
9333 Set_Node_Offset(ret, parse_start + 1);
9335 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
9339 /* Check for proper termination. */
9341 RExC_flags = oregflags;
9342 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
9343 RExC_parse = oregcomp_parse;
9344 vFAIL("Unmatched (");
9347 else if (!paren && RExC_parse < RExC_end) {
9348 if (*RExC_parse == ')') {
9350 vFAIL("Unmatched )");
9353 FAIL("Junk on end of regexp"); /* "Can't happen". */
9354 assert(0); /* NOTREACHED */
9357 if (RExC_in_lookbehind) {
9358 RExC_in_lookbehind--;
9360 if (after_freeze > RExC_npar)
9361 RExC_npar = after_freeze;
9366 - regbranch - one alternative of an | operator
9368 * Implements the concatenation operator.
9370 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9374 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
9378 regnode *chain = NULL;
9380 I32 flags = 0, c = 0;
9381 GET_RE_DEBUG_FLAGS_DECL;
9383 PERL_ARGS_ASSERT_REGBRANCH;
9385 DEBUG_PARSE("brnc");
9390 if (!SIZE_ONLY && RExC_extralen)
9391 ret = reganode(pRExC_state, BRANCHJ,0);
9393 ret = reg_node(pRExC_state, BRANCH);
9394 Set_Node_Length(ret, 1);
9398 if (!first && SIZE_ONLY)
9399 RExC_extralen += 1; /* BRANCHJ */
9401 *flagp = WORST; /* Tentatively. */
9404 nextchar(pRExC_state);
9405 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
9407 latest = regpiece(pRExC_state, &flags,depth+1);
9408 if (latest == NULL) {
9409 if (flags & TRYAGAIN)
9411 if (flags & RESTART_UTF8) {
9412 *flagp = RESTART_UTF8;
9415 FAIL2("panic: regpiece returned NULL, flags=%#X", flags);
9417 else if (ret == NULL)
9419 *flagp |= flags&(HASWIDTH|POSTPONED);
9420 if (chain == NULL) /* First piece. */
9421 *flagp |= flags&SPSTART;
9424 REGTAIL(pRExC_state, chain, latest);
9429 if (chain == NULL) { /* Loop ran zero times. */
9430 chain = reg_node(pRExC_state, NOTHING);
9435 *flagp |= flags&SIMPLE;
9442 - regpiece - something followed by possible [*+?]
9444 * Note that the branching code sequences used for ? and the general cases
9445 * of * and + are somewhat optimized: they use the same NOTHING node as
9446 * both the endmarker for their branch list and the body of the last branch.
9447 * It might seem that this node could be dispensed with entirely, but the
9448 * endmarker role is not redundant.
9450 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
9452 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9456 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
9463 const char * const origparse = RExC_parse;
9465 I32 max = REG_INFTY;
9466 #ifdef RE_TRACK_PATTERN_OFFSETS
9469 const char *maxpos = NULL;
9471 /* Save the original in case we change the emitted regop to a FAIL. */
9472 regnode * const orig_emit = RExC_emit;
9474 GET_RE_DEBUG_FLAGS_DECL;
9476 PERL_ARGS_ASSERT_REGPIECE;
9478 DEBUG_PARSE("piec");
9480 ret = regatom(pRExC_state, &flags,depth+1);
9482 if (flags & (TRYAGAIN|RESTART_UTF8))
9483 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
9485 FAIL2("panic: regatom returned NULL, flags=%#X", flags);
9491 if (op == '{' && regcurly(RExC_parse, FALSE)) {
9493 #ifdef RE_TRACK_PATTERN_OFFSETS
9494 parse_start = RExC_parse; /* MJD */
9496 next = RExC_parse + 1;
9497 while (isDIGIT(*next) || *next == ',') {
9506 if (*next == '}') { /* got one */
9510 min = atoi(RExC_parse);
9514 maxpos = RExC_parse;
9516 if (!max && *maxpos != '0')
9517 max = REG_INFTY; /* meaning "infinity" */
9518 else if (max >= REG_INFTY)
9519 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
9521 nextchar(pRExC_state);
9522 if (max < min) { /* If can't match, warn and optimize to fail
9525 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
9527 /* We can't back off the size because we have to reserve
9528 * enough space for all the things we are about to throw
9529 * away, but we can shrink it by the ammount we are about
9531 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
9534 RExC_emit = orig_emit;
9536 ret = reg_node(pRExC_state, OPFAIL);
9539 else if (max == 0) { /* replace {0} with a nothing node */
9541 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)NOTHING];
9544 RExC_emit = orig_emit;
9546 ret = reg_node(pRExC_state, NOTHING);
9551 if ((flags&SIMPLE)) {
9552 RExC_naughty += 2 + RExC_naughty / 2;
9553 reginsert(pRExC_state, CURLY, ret, depth+1);
9554 Set_Node_Offset(ret, parse_start+1); /* MJD */
9555 Set_Node_Cur_Length(ret);
9558 regnode * const w = reg_node(pRExC_state, WHILEM);
9561 REGTAIL(pRExC_state, ret, w);
9562 if (!SIZE_ONLY && RExC_extralen) {
9563 reginsert(pRExC_state, LONGJMP,ret, depth+1);
9564 reginsert(pRExC_state, NOTHING,ret, depth+1);
9565 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
9567 reginsert(pRExC_state, CURLYX,ret, depth+1);
9569 Set_Node_Offset(ret, parse_start+1);
9570 Set_Node_Length(ret,
9571 op == '{' ? (RExC_parse - parse_start) : 1);
9573 if (!SIZE_ONLY && RExC_extralen)
9574 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
9575 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
9577 RExC_whilem_seen++, RExC_extralen += 3;
9578 RExC_naughty += 4 + RExC_naughty; /* compound interest */
9587 ARG1_SET(ret, (U16)min);
9588 ARG2_SET(ret, (U16)max);
9600 #if 0 /* Now runtime fix should be reliable. */
9602 /* if this is reinstated, don't forget to put this back into perldiag:
9604 =item Regexp *+ operand could be empty at {#} in regex m/%s/
9606 (F) The part of the regexp subject to either the * or + quantifier
9607 could match an empty string. The {#} shows in the regular
9608 expression about where the problem was discovered.
9612 if (!(flags&HASWIDTH) && op != '?')
9613 vFAIL("Regexp *+ operand could be empty");
9616 #ifdef RE_TRACK_PATTERN_OFFSETS
9617 parse_start = RExC_parse;
9619 nextchar(pRExC_state);
9621 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
9623 if (op == '*' && (flags&SIMPLE)) {
9624 reginsert(pRExC_state, STAR, ret, depth+1);
9628 else if (op == '*') {
9632 else if (op == '+' && (flags&SIMPLE)) {
9633 reginsert(pRExC_state, PLUS, ret, depth+1);
9637 else if (op == '+') {
9641 else if (op == '?') {
9646 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
9647 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
9648 ckWARN3reg(RExC_parse,
9649 "%.*s matches null string many times",
9650 (int)(RExC_parse >= origparse ? RExC_parse - origparse : 0),
9652 (void)ReREFCNT_inc(RExC_rx_sv);
9655 if (RExC_parse < RExC_end && *RExC_parse == '?') {
9656 nextchar(pRExC_state);
9657 reginsert(pRExC_state, MINMOD, ret, depth+1);
9658 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
9660 #ifndef REG_ALLOW_MINMOD_SUSPEND
9663 if (RExC_parse < RExC_end && *RExC_parse == '+') {
9665 nextchar(pRExC_state);
9666 ender = reg_node(pRExC_state, SUCCEED);
9667 REGTAIL(pRExC_state, ret, ender);
9668 reginsert(pRExC_state, SUSPEND, ret, depth+1);
9670 ender = reg_node(pRExC_state, TAIL);
9671 REGTAIL(pRExC_state, ret, ender);
9675 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
9677 vFAIL("Nested quantifiers");
9684 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p, UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
9685 const bool strict /* Apply stricter parsing rules? */
9689 /* This is expected to be called by a parser routine that has recognized '\N'
9690 and needs to handle the rest. RExC_parse is expected to point at the first
9691 char following the N at the time of the call. On successful return,
9692 RExC_parse has been updated to point to just after the sequence identified
9693 by this routine, and <*flagp> has been updated.
9695 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
9698 \N may begin either a named sequence, or if outside a character class, mean
9699 to match a non-newline. For non single-quoted regexes, the tokenizer has
9700 attempted to decide which, and in the case of a named sequence, converted it
9701 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
9702 where c1... are the characters in the sequence. For single-quoted regexes,
9703 the tokenizer passes the \N sequence through unchanged; this code will not
9704 attempt to determine this nor expand those, instead raising a syntax error.
9705 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
9706 or there is no '}', it signals that this \N occurrence means to match a
9709 Only the \N{U+...} form should occur in a character class, for the same
9710 reason that '.' inside a character class means to just match a period: it
9711 just doesn't make sense.
9713 The function raises an error (via vFAIL), and doesn't return for various
9714 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
9715 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
9716 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
9717 only possible if node_p is non-NULL.
9720 If <valuep> is non-null, it means the caller can accept an input sequence
9721 consisting of a just a single code point; <*valuep> is set to that value
9722 if the input is such.
9724 If <node_p> is non-null it signifies that the caller can accept any other
9725 legal sequence (i.e., one that isn't just a single code point). <*node_p>
9727 1) \N means not-a-NL: points to a newly created REG_ANY node;
9728 2) \N{}: points to a new NOTHING node;
9729 3) otherwise: points to a new EXACT node containing the resolved
9731 Note that FALSE is returned for single code point sequences if <valuep> is
9735 char * endbrace; /* '}' following the name */
9737 char *endchar; /* Points to '.' or '}' ending cur char in the input
9739 bool has_multiple_chars; /* true if the input stream contains a sequence of
9740 more than one character */
9742 GET_RE_DEBUG_FLAGS_DECL;
9744 PERL_ARGS_ASSERT_GROK_BSLASH_N;
9748 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
9750 /* The [^\n] meaning of \N ignores spaces and comments under the /x
9751 * modifier. The other meaning does not */
9752 p = (RExC_flags & RXf_PMf_EXTENDED)
9753 ? regwhite( pRExC_state, RExC_parse )
9756 /* Disambiguate between \N meaning a named character versus \N meaning
9757 * [^\n]. The former is assumed when it can't be the latter. */
9758 if (*p != '{' || regcurly(p, FALSE)) {
9761 /* no bare \N in a charclass */
9762 if (in_char_class) {
9763 vFAIL("\\N in a character class must be a named character: \\N{...}");
9767 nextchar(pRExC_state);
9768 *node_p = reg_node(pRExC_state, REG_ANY);
9769 *flagp |= HASWIDTH|SIMPLE;
9772 Set_Node_Length(*node_p, 1); /* MJD */
9776 /* Here, we have decided it should be a named character or sequence */
9778 /* The test above made sure that the next real character is a '{', but
9779 * under the /x modifier, it could be separated by space (or a comment and
9780 * \n) and this is not allowed (for consistency with \x{...} and the
9781 * tokenizer handling of \N{NAME}). */
9782 if (*RExC_parse != '{') {
9783 vFAIL("Missing braces on \\N{}");
9786 RExC_parse++; /* Skip past the '{' */
9788 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
9789 || ! (endbrace == RExC_parse /* nothing between the {} */
9790 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below */
9791 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg) */
9793 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
9794 vFAIL("\\N{NAME} must be resolved by the lexer");
9797 if (endbrace == RExC_parse) { /* empty: \N{} */
9800 *node_p = reg_node(pRExC_state,NOTHING);
9802 else if (in_char_class) {
9803 if (SIZE_ONLY && in_char_class) {
9805 RExC_parse++; /* Position after the "}" */
9806 vFAIL("Zero length \\N{}");
9809 ckWARNreg(RExC_parse,
9810 "Ignoring zero length \\N{} in character class");
9818 nextchar(pRExC_state);
9822 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
9823 RExC_parse += 2; /* Skip past the 'U+' */
9825 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9827 /* Code points are separated by dots. If none, there is only one code
9828 * point, and is terminated by the brace */
9829 has_multiple_chars = (endchar < endbrace);
9831 if (valuep && (! has_multiple_chars || in_char_class)) {
9832 /* We only pay attention to the first char of
9833 multichar strings being returned in char classes. I kinda wonder
9834 if this makes sense as it does change the behaviour
9835 from earlier versions, OTOH that behaviour was broken
9836 as well. XXX Solution is to recharacterize as
9837 [rest-of-class]|multi1|multi2... */
9839 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
9840 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
9841 | PERL_SCAN_DISALLOW_PREFIX
9842 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
9844 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
9846 /* The tokenizer should have guaranteed validity, but it's possible to
9847 * bypass it by using single quoting, so check */
9848 if (length_of_hex == 0
9849 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
9851 RExC_parse += length_of_hex; /* Includes all the valid */
9852 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
9853 ? UTF8SKIP(RExC_parse)
9855 /* Guard against malformed utf8 */
9856 if (RExC_parse >= endchar) {
9857 RExC_parse = endchar;
9859 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9862 if (in_char_class && has_multiple_chars) {
9864 RExC_parse = endbrace;
9865 vFAIL("\\N{} in character class restricted to one character");
9868 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
9872 RExC_parse = endbrace + 1;
9874 else if (! node_p || ! has_multiple_chars) {
9876 /* Here, the input is legal, but not according to the caller's
9877 * options. We fail without advancing the parse, so that the
9878 * caller can try again */
9884 /* What is done here is to convert this to a sub-pattern of the form
9885 * (?:\x{char1}\x{char2}...)
9886 * and then call reg recursively. That way, it retains its atomicness,
9887 * while not having to worry about special handling that some code
9888 * points may have. toke.c has converted the original Unicode values
9889 * to native, so that we can just pass on the hex values unchanged. We
9890 * do have to set a flag to keep recoding from happening in the
9893 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
9895 char *orig_end = RExC_end;
9898 while (RExC_parse < endbrace) {
9900 /* Convert to notation the rest of the code understands */
9901 sv_catpv(substitute_parse, "\\x{");
9902 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
9903 sv_catpv(substitute_parse, "}");
9905 /* Point to the beginning of the next character in the sequence. */
9906 RExC_parse = endchar + 1;
9907 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9909 sv_catpv(substitute_parse, ")");
9911 RExC_parse = SvPV(substitute_parse, len);
9913 /* Don't allow empty number */
9915 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9917 RExC_end = RExC_parse + len;
9919 /* The values are Unicode, and therefore not subject to recoding */
9920 RExC_override_recoding = 1;
9922 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
9923 if (flags & RESTART_UTF8) {
9924 *flagp = RESTART_UTF8;
9927 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#X",
9930 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
9932 RExC_parse = endbrace;
9933 RExC_end = orig_end;
9934 RExC_override_recoding = 0;
9936 nextchar(pRExC_state);
9946 * It returns the code point in utf8 for the value in *encp.
9947 * value: a code value in the source encoding
9948 * encp: a pointer to an Encode object
9950 * If the result from Encode is not a single character,
9951 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
9954 S_reg_recode(pTHX_ const char value, SV **encp)
9957 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
9958 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
9959 const STRLEN newlen = SvCUR(sv);
9960 UV uv = UNICODE_REPLACEMENT;
9962 PERL_ARGS_ASSERT_REG_RECODE;
9966 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
9969 if (!newlen || numlen != newlen) {
9970 uv = UNICODE_REPLACEMENT;
9976 PERL_STATIC_INLINE U8
9977 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
9981 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
9987 op = get_regex_charset(RExC_flags);
9988 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
9989 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
9990 been, so there is no hole */
9996 PERL_STATIC_INLINE void
9997 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state, regnode *node, I32* flagp, STRLEN len, UV code_point)
9999 /* This knows the details about sizing an EXACTish node, setting flags for
10000 * it (by setting <*flagp>, and potentially populating it with a single
10003 * If <len> (the length in bytes) is non-zero, this function assumes that
10004 * the node has already been populated, and just does the sizing. In this
10005 * case <code_point> should be the final code point that has already been
10006 * placed into the node. This value will be ignored except that under some
10007 * circumstances <*flagp> is set based on it.
10009 * If <len> is zero, the function assumes that the node is to contain only
10010 * the single character given by <code_point> and calculates what <len>
10011 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10012 * additionally will populate the node's STRING with <code_point>, if <len>
10013 * is 0. In both cases <*flagp> is appropriately set
10015 * It knows that under FOLD, UTF characters and the Latin Sharp S must be
10016 * folded (the latter only when the rules indicate it can match 'ss') */
10018 bool len_passed_in = cBOOL(len != 0);
10019 U8 character[UTF8_MAXBYTES_CASE+1];
10021 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10023 if (! len_passed_in) {
10026 to_uni_fold(NATIVE_TO_UNI(code_point), character, &len);
10029 uvchr_to_utf8( character, code_point);
10030 len = UTF8SKIP(character);
10034 || code_point != LATIN_SMALL_LETTER_SHARP_S
10035 || ASCII_FOLD_RESTRICTED
10036 || ! AT_LEAST_UNI_SEMANTICS)
10038 *character = (U8) code_point;
10043 *(character + 1) = 's';
10049 RExC_size += STR_SZ(len);
10052 RExC_emit += STR_SZ(len);
10053 STR_LEN(node) = len;
10054 if (! len_passed_in) {
10055 Copy((char *) character, STRING(node), len, char);
10059 *flagp |= HASWIDTH;
10061 /* A single character node is SIMPLE, except for the special-cased SHARP S
10063 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
10064 && (code_point != LATIN_SMALL_LETTER_SHARP_S
10065 || ! FOLD || ! DEPENDS_SEMANTICS))
10072 - regatom - the lowest level
10074 Try to identify anything special at the start of the pattern. If there
10075 is, then handle it as required. This may involve generating a single regop,
10076 such as for an assertion; or it may involve recursing, such as to
10077 handle a () structure.
10079 If the string doesn't start with something special then we gobble up
10080 as much literal text as we can.
10082 Once we have been able to handle whatever type of thing started the
10083 sequence, we return.
10085 Note: we have to be careful with escapes, as they can be both literal
10086 and special, and in the case of \10 and friends, context determines which.
10088 A summary of the code structure is:
10090 switch (first_byte) {
10091 cases for each special:
10092 handle this special;
10095 switch (2nd byte) {
10096 cases for each unambiguous special:
10097 handle this special;
10099 cases for each ambigous special/literal:
10101 if (special) handle here
10103 default: // unambiguously literal:
10106 default: // is a literal char
10109 create EXACTish node for literal;
10110 while (more input and node isn't full) {
10111 switch (input_byte) {
10112 cases for each special;
10113 make sure parse pointer is set so that the next call to
10114 regatom will see this special first
10115 goto loopdone; // EXACTish node terminated by prev. char
10117 append char to EXACTISH node;
10119 get next input byte;
10123 return the generated node;
10125 Specifically there are two separate switches for handling
10126 escape sequences, with the one for handling literal escapes requiring
10127 a dummy entry for all of the special escapes that are actually handled
10130 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
10132 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10134 Otherwise does not return NULL.
10138 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10141 regnode *ret = NULL;
10143 char *parse_start = RExC_parse;
10147 GET_RE_DEBUG_FLAGS_DECL;
10149 *flagp = WORST; /* Tentatively. */
10151 DEBUG_PARSE("atom");
10153 PERL_ARGS_ASSERT_REGATOM;
10156 switch ((U8)*RExC_parse) {
10158 RExC_seen_zerolen++;
10159 nextchar(pRExC_state);
10160 if (RExC_flags & RXf_PMf_MULTILINE)
10161 ret = reg_node(pRExC_state, MBOL);
10162 else if (RExC_flags & RXf_PMf_SINGLELINE)
10163 ret = reg_node(pRExC_state, SBOL);
10165 ret = reg_node(pRExC_state, BOL);
10166 Set_Node_Length(ret, 1); /* MJD */
10169 nextchar(pRExC_state);
10171 RExC_seen_zerolen++;
10172 if (RExC_flags & RXf_PMf_MULTILINE)
10173 ret = reg_node(pRExC_state, MEOL);
10174 else if (RExC_flags & RXf_PMf_SINGLELINE)
10175 ret = reg_node(pRExC_state, SEOL);
10177 ret = reg_node(pRExC_state, EOL);
10178 Set_Node_Length(ret, 1); /* MJD */
10181 nextchar(pRExC_state);
10182 if (RExC_flags & RXf_PMf_SINGLELINE)
10183 ret = reg_node(pRExC_state, SANY);
10185 ret = reg_node(pRExC_state, REG_ANY);
10186 *flagp |= HASWIDTH|SIMPLE;
10188 Set_Node_Length(ret, 1); /* MJD */
10192 char * const oregcomp_parse = ++RExC_parse;
10193 ret = regclass(pRExC_state, flagp,depth+1,
10194 FALSE, /* means parse the whole char class */
10195 TRUE, /* allow multi-char folds */
10196 FALSE, /* don't silence non-portable warnings. */
10198 if (*RExC_parse != ']') {
10199 RExC_parse = oregcomp_parse;
10200 vFAIL("Unmatched [");
10203 if (*flagp & RESTART_UTF8)
10205 FAIL2("panic: regclass returned NULL to regatom, flags=%#X",
10208 nextchar(pRExC_state);
10209 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
10213 nextchar(pRExC_state);
10214 ret = reg(pRExC_state, 1, &flags,depth+1);
10216 if (flags & TRYAGAIN) {
10217 if (RExC_parse == RExC_end) {
10218 /* Make parent create an empty node if needed. */
10219 *flagp |= TRYAGAIN;
10224 if (flags & RESTART_UTF8) {
10225 *flagp = RESTART_UTF8;
10228 FAIL2("panic: reg returned NULL to regatom, flags=%#X", flags);
10230 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10234 if (flags & TRYAGAIN) {
10235 *flagp |= TRYAGAIN;
10238 vFAIL("Internal urp");
10239 /* Supposed to be caught earlier. */
10242 if (!regcurly(RExC_parse, FALSE)) {
10251 vFAIL("Quantifier follows nothing");
10256 This switch handles escape sequences that resolve to some kind
10257 of special regop and not to literal text. Escape sequnces that
10258 resolve to literal text are handled below in the switch marked
10261 Every entry in this switch *must* have a corresponding entry
10262 in the literal escape switch. However, the opposite is not
10263 required, as the default for this switch is to jump to the
10264 literal text handling code.
10266 switch ((U8)*++RExC_parse) {
10268 /* Special Escapes */
10270 RExC_seen_zerolen++;
10271 ret = reg_node(pRExC_state, SBOL);
10273 goto finish_meta_pat;
10275 ret = reg_node(pRExC_state, GPOS);
10276 RExC_seen |= REG_SEEN_GPOS;
10278 goto finish_meta_pat;
10280 RExC_seen_zerolen++;
10281 ret = reg_node(pRExC_state, KEEPS);
10283 /* XXX:dmq : disabling in-place substitution seems to
10284 * be necessary here to avoid cases of memory corruption, as
10285 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
10287 RExC_seen |= REG_SEEN_LOOKBEHIND;
10288 goto finish_meta_pat;
10290 ret = reg_node(pRExC_state, SEOL);
10292 RExC_seen_zerolen++; /* Do not optimize RE away */
10293 goto finish_meta_pat;
10295 ret = reg_node(pRExC_state, EOS);
10297 RExC_seen_zerolen++; /* Do not optimize RE away */
10298 goto finish_meta_pat;
10300 ret = reg_node(pRExC_state, CANY);
10301 RExC_seen |= REG_SEEN_CANY;
10302 *flagp |= HASWIDTH|SIMPLE;
10303 goto finish_meta_pat;
10305 ret = reg_node(pRExC_state, CLUMP);
10306 *flagp |= HASWIDTH;
10307 goto finish_meta_pat;
10313 arg = ANYOF_WORDCHAR;
10317 RExC_seen_zerolen++;
10318 RExC_seen |= REG_SEEN_LOOKBEHIND;
10319 op = BOUND + get_regex_charset(RExC_flags);
10320 if (op > BOUNDA) { /* /aa is same as /a */
10323 ret = reg_node(pRExC_state, op);
10324 FLAGS(ret) = get_regex_charset(RExC_flags);
10326 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10327 ckWARNdep(RExC_parse, "\"\\b{\" is deprecated; use \"\\b\\{\" or \"\\b[{]\" instead");
10329 goto finish_meta_pat;
10331 RExC_seen_zerolen++;
10332 RExC_seen |= REG_SEEN_LOOKBEHIND;
10333 op = NBOUND + get_regex_charset(RExC_flags);
10334 if (op > NBOUNDA) { /* /aa is same as /a */
10337 ret = reg_node(pRExC_state, op);
10338 FLAGS(ret) = get_regex_charset(RExC_flags);
10340 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10341 ckWARNdep(RExC_parse, "\"\\B{\" is deprecated; use \"\\B\\{\" or \"\\B[{]\" instead");
10343 goto finish_meta_pat;
10353 ret = reg_node(pRExC_state, LNBREAK);
10354 *flagp |= HASWIDTH|SIMPLE;
10355 goto finish_meta_pat;
10363 goto join_posix_op_known;
10369 arg = ANYOF_VERTWS;
10371 goto join_posix_op_known;
10381 op = POSIXD + get_regex_charset(RExC_flags);
10382 if (op > POSIXA) { /* /aa is same as /a */
10386 join_posix_op_known:
10389 op += NPOSIXD - POSIXD;
10392 ret = reg_node(pRExC_state, op);
10394 FLAGS(ret) = namedclass_to_classnum(arg);
10397 *flagp |= HASWIDTH|SIMPLE;
10401 nextchar(pRExC_state);
10402 Set_Node_Length(ret, 2); /* MJD */
10408 char* parse_start = RExC_parse - 2;
10413 ret = regclass(pRExC_state, flagp,depth+1,
10414 TRUE, /* means just parse this element */
10415 FALSE, /* don't allow multi-char folds */
10416 FALSE, /* don't silence non-portable warnings.
10417 It would be a bug if these returned
10420 /* regclass() can only return RESTART_UTF8 if multi-char folds
10423 FAIL2("panic: regclass returned NULL to regatom, flags=%#X",
10428 Set_Node_Offset(ret, parse_start + 2);
10429 Set_Node_Cur_Length(ret);
10430 nextchar(pRExC_state);
10434 /* Handle \N and \N{NAME} with multiple code points here and not
10435 * below because it can be multicharacter. join_exact() will join
10436 * them up later on. Also this makes sure that things like
10437 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
10438 * The options to the grok function call causes it to fail if the
10439 * sequence is just a single code point. We then go treat it as
10440 * just another character in the current EXACT node, and hence it
10441 * gets uniform treatment with all the other characters. The
10442 * special treatment for quantifiers is not needed for such single
10443 * character sequences */
10445 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
10446 FALSE /* not strict */ )) {
10447 if (*flagp & RESTART_UTF8)
10453 case 'k': /* Handle \k<NAME> and \k'NAME' */
10456 char ch= RExC_parse[1];
10457 if (ch != '<' && ch != '\'' && ch != '{') {
10459 vFAIL2("Sequence %.2s... not terminated",parse_start);
10461 /* this pretty much dupes the code for (?P=...) in reg(), if
10462 you change this make sure you change that */
10463 char* name_start = (RExC_parse += 2);
10465 SV *sv_dat = reg_scan_name(pRExC_state,
10466 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10467 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
10468 if (RExC_parse == name_start || *RExC_parse != ch)
10469 vFAIL2("Sequence %.3s... not terminated",parse_start);
10472 num = add_data( pRExC_state, 1, "S" );
10473 RExC_rxi->data->data[num]=(void*)sv_dat;
10474 SvREFCNT_inc_simple_void(sv_dat);
10478 ret = reganode(pRExC_state,
10481 : (ASCII_FOLD_RESTRICTED)
10483 : (AT_LEAST_UNI_SEMANTICS)
10489 *flagp |= HASWIDTH;
10491 /* override incorrect value set in reganode MJD */
10492 Set_Node_Offset(ret, parse_start+1);
10493 Set_Node_Cur_Length(ret); /* MJD */
10494 nextchar(pRExC_state);
10500 case '1': case '2': case '3': case '4':
10501 case '5': case '6': case '7': case '8': case '9':
10504 bool isg = *RExC_parse == 'g';
10509 if (*RExC_parse == '{') {
10513 if (*RExC_parse == '-') {
10517 if (hasbrace && !isDIGIT(*RExC_parse)) {
10518 if (isrel) RExC_parse--;
10520 goto parse_named_seq;
10522 num = atoi(RExC_parse);
10523 if (isg && num == 0)
10524 vFAIL("Reference to invalid group 0");
10526 num = RExC_npar - num;
10528 vFAIL("Reference to nonexistent or unclosed group");
10530 if (!isg && num > 9 && num >= RExC_npar)
10531 /* Probably a character specified in octal, e.g. \35 */
10534 char * const parse_start = RExC_parse - 1; /* MJD */
10535 while (isDIGIT(*RExC_parse))
10537 if (parse_start == RExC_parse - 1)
10538 vFAIL("Unterminated \\g... pattern");
10540 if (*RExC_parse != '}')
10541 vFAIL("Unterminated \\g{...} pattern");
10545 if (num > (I32)RExC_rx->nparens)
10546 vFAIL("Reference to nonexistent group");
10549 ret = reganode(pRExC_state,
10552 : (ASCII_FOLD_RESTRICTED)
10554 : (AT_LEAST_UNI_SEMANTICS)
10560 *flagp |= HASWIDTH;
10562 /* override incorrect value set in reganode MJD */
10563 Set_Node_Offset(ret, parse_start+1);
10564 Set_Node_Cur_Length(ret); /* MJD */
10566 nextchar(pRExC_state);
10571 if (RExC_parse >= RExC_end)
10572 FAIL("Trailing \\");
10575 /* Do not generate "unrecognized" warnings here, we fall
10576 back into the quick-grab loop below */
10583 if (RExC_flags & RXf_PMf_EXTENDED) {
10584 if ( reg_skipcomment( pRExC_state ) )
10591 parse_start = RExC_parse - 1;
10600 #define MAX_NODE_STRING_SIZE 127
10601 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
10603 U8 upper_parse = MAX_NODE_STRING_SIZE;
10606 bool next_is_quantifier;
10607 char * oldp = NULL;
10609 /* If a folding node contains only code points that don't
10610 * participate in folds, it can be changed into an EXACT node,
10611 * which allows the optimizer more things to look for */
10615 node_type = compute_EXACTish(pRExC_state);
10616 ret = reg_node(pRExC_state, node_type);
10618 /* In pass1, folded, we use a temporary buffer instead of the
10619 * actual node, as the node doesn't exist yet */
10620 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
10626 /* We do the EXACTFish to EXACT node only if folding, and not if in
10627 * locale, as whether a character folds or not isn't known until
10629 maybe_exact = FOLD && ! LOC;
10631 /* XXX The node can hold up to 255 bytes, yet this only goes to
10632 * 127. I (khw) do not know why. Keeping it somewhat less than
10633 * 255 allows us to not have to worry about overflow due to
10634 * converting to utf8 and fold expansion, but that value is
10635 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
10636 * split up by this limit into a single one using the real max of
10637 * 255. Even at 127, this breaks under rare circumstances. If
10638 * folding, we do not want to split a node at a character that is a
10639 * non-final in a multi-char fold, as an input string could just
10640 * happen to want to match across the node boundary. The join
10641 * would solve that problem if the join actually happens. But a
10642 * series of more than two nodes in a row each of 127 would cause
10643 * the first join to succeed to get to 254, but then there wouldn't
10644 * be room for the next one, which could at be one of those split
10645 * multi-char folds. I don't know of any fool-proof solution. One
10646 * could back off to end with only a code point that isn't such a
10647 * non-final, but it is possible for there not to be any in the
10649 for (p = RExC_parse - 1;
10650 len < upper_parse && p < RExC_end;
10655 if (RExC_flags & RXf_PMf_EXTENDED)
10656 p = regwhite( pRExC_state, p );
10667 /* Literal Escapes Switch
10669 This switch is meant to handle escape sequences that
10670 resolve to a literal character.
10672 Every escape sequence that represents something
10673 else, like an assertion or a char class, is handled
10674 in the switch marked 'Special Escapes' above in this
10675 routine, but also has an entry here as anything that
10676 isn't explicitly mentioned here will be treated as
10677 an unescaped equivalent literal.
10680 switch ((U8)*++p) {
10681 /* These are all the special escapes. */
10682 case 'A': /* Start assertion */
10683 case 'b': case 'B': /* Word-boundary assertion*/
10684 case 'C': /* Single char !DANGEROUS! */
10685 case 'd': case 'D': /* digit class */
10686 case 'g': case 'G': /* generic-backref, pos assertion */
10687 case 'h': case 'H': /* HORIZWS */
10688 case 'k': case 'K': /* named backref, keep marker */
10689 case 'p': case 'P': /* Unicode property */
10690 case 'R': /* LNBREAK */
10691 case 's': case 'S': /* space class */
10692 case 'v': case 'V': /* VERTWS */
10693 case 'w': case 'W': /* word class */
10694 case 'X': /* eXtended Unicode "combining character sequence" */
10695 case 'z': case 'Z': /* End of line/string assertion */
10699 /* Anything after here is an escape that resolves to a
10700 literal. (Except digits, which may or may not)
10706 case 'N': /* Handle a single-code point named character. */
10707 /* The options cause it to fail if a multiple code
10708 * point sequence. Handle those in the switch() above
10710 RExC_parse = p + 1;
10711 if (! grok_bslash_N(pRExC_state, NULL, &ender,
10712 flagp, depth, FALSE,
10713 FALSE /* not strict */ ))
10715 if (*flagp & RESTART_UTF8)
10716 FAIL("panic: grok_bslash_N set RESTART_UTF8");
10717 RExC_parse = p = oldp;
10721 if (ender > 0xff) {
10738 ender = ASCII_TO_NATIVE('\033');
10742 ender = ASCII_TO_NATIVE('\007');
10748 const char* error_msg;
10750 bool valid = grok_bslash_o(&p,
10753 TRUE, /* out warnings */
10754 FALSE, /* not strict */
10755 TRUE, /* Output warnings
10760 RExC_parse = p; /* going to die anyway; point
10761 to exact spot of failure */
10765 if (PL_encoding && ender < 0x100) {
10766 goto recode_encoding;
10768 if (ender > 0xff) {
10775 UV result = UV_MAX; /* initialize to erroneous
10777 const char* error_msg;
10779 bool valid = grok_bslash_x(&p,
10782 TRUE, /* out warnings */
10783 FALSE, /* not strict */
10784 TRUE, /* Output warnings
10789 RExC_parse = p; /* going to die anyway; point
10790 to exact spot of failure */
10795 if (PL_encoding && ender < 0x100) {
10796 goto recode_encoding;
10798 if (ender > 0xff) {
10805 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
10807 case '0': case '1': case '2': case '3':case '4':
10808 case '5': case '6': case '7':
10810 (isDIGIT(p[1]) && atoi(p) >= RExC_npar))
10812 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
10814 ender = grok_oct(p, &numlen, &flags, NULL);
10815 if (ender > 0xff) {
10819 if (SIZE_ONLY /* like \08, \178 */
10822 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
10824 reg_warn_non_literal_string(
10826 form_short_octal_warning(p, numlen));
10829 else { /* Not to be treated as an octal constant, go
10834 if (PL_encoding && ender < 0x100)
10835 goto recode_encoding;
10838 if (! RExC_override_recoding) {
10839 SV* enc = PL_encoding;
10840 ender = reg_recode((const char)(U8)ender, &enc);
10841 if (!enc && SIZE_ONLY)
10842 ckWARNreg(p, "Invalid escape in the specified encoding");
10848 FAIL("Trailing \\");
10851 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
10852 /* Include any { following the alpha to emphasize
10853 * that it could be part of an escape at some point
10855 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
10856 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
10858 goto normal_default;
10859 } /* End of switch on '\' */
10861 default: /* A literal character */
10864 && RExC_flags & RXf_PMf_EXTENDED
10865 && ckWARN(WARN_DEPRECATED)
10866 && is_PATWS_non_low(p, UTF))
10868 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
10869 "Escape literal pattern white space under /x");
10873 if (UTF8_IS_START(*p) && UTF) {
10875 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
10876 &numlen, UTF8_ALLOW_DEFAULT);
10882 } /* End of switch on the literal */
10884 /* Here, have looked at the literal character and <ender>
10885 * contains its ordinal, <p> points to the character after it
10888 if ( RExC_flags & RXf_PMf_EXTENDED)
10889 p = regwhite( pRExC_state, p );
10891 /* If the next thing is a quantifier, it applies to this
10892 * character only, which means that this character has to be in
10893 * its own node and can't just be appended to the string in an
10894 * existing node, so if there are already other characters in
10895 * the node, close the node with just them, and set up to do
10896 * this character again next time through, when it will be the
10897 * only thing in its new node */
10898 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
10906 /* See comments for join_exact() as to why we fold
10907 * this non-UTF at compile time */
10908 || (node_type == EXACTFU
10909 && ender == LATIN_SMALL_LETTER_SHARP_S))
10913 /* Prime the casefolded buffer. Locale rules, which
10914 * apply only to code points < 256, aren't known until
10915 * execution, so for them, just output the original
10916 * character using utf8. If we start to fold non-UTF
10917 * patterns, be sure to update join_exact() */
10918 if (LOC && ender < 256) {
10919 if (UNI_IS_INVARIANT(ender)) {
10923 *s = UTF8_TWO_BYTE_HI(ender);
10924 *(s + 1) = UTF8_TWO_BYTE_LO(ender);
10929 UV folded = _to_uni_fold_flags(
10934 | ((LOC) ? FOLD_FLAGS_LOCALE
10935 : (ASCII_FOLD_RESTRICTED)
10936 ? FOLD_FLAGS_NOMIX_ASCII
10940 /* If this node only contains non-folding code
10941 * points so far, see if this new one is also
10944 if (folded != ender) {
10945 maybe_exact = FALSE;
10948 /* Here the fold is the original; we have
10949 * to check further to see if anything
10951 if (! PL_utf8_foldable) {
10952 SV* swash = swash_init("utf8",
10954 &PL_sv_undef, 1, 0);
10956 _get_swash_invlist(swash);
10957 SvREFCNT_dec_NN(swash);
10959 if (_invlist_contains_cp(PL_utf8_foldable,
10962 maybe_exact = FALSE;
10970 /* The loop increments <len> each time, as all but this
10971 * path (and the one just below for UTF) through it add
10972 * a single byte to the EXACTish node. But this one
10973 * has changed len to be the correct final value, so
10974 * subtract one to cancel out the increment that
10976 len += foldlen - 1;
10979 *(s++) = (char) ender;
10980 maybe_exact &= ! IS_IN_SOME_FOLD_L1(ender);
10984 const STRLEN unilen = reguni(pRExC_state, ender, s);
10990 /* See comment just above for - 1 */
10994 REGC((char)ender, s++);
10997 if (next_is_quantifier) {
10999 /* Here, the next input is a quantifier, and to get here,
11000 * the current character is the only one in the node.
11001 * Also, here <len> doesn't include the final byte for this
11007 } /* End of loop through literal characters */
11009 /* Here we have either exhausted the input or ran out of room in
11010 * the node. (If we encountered a character that can't be in the
11011 * node, transfer is made directly to <loopdone>, and so we
11012 * wouldn't have fallen off the end of the loop.) In the latter
11013 * case, we artificially have to split the node into two, because
11014 * we just don't have enough space to hold everything. This
11015 * creates a problem if the final character participates in a
11016 * multi-character fold in the non-final position, as a match that
11017 * should have occurred won't, due to the way nodes are matched,
11018 * and our artificial boundary. So back off until we find a non-
11019 * problematic character -- one that isn't at the beginning or
11020 * middle of such a fold. (Either it doesn't participate in any
11021 * folds, or appears only in the final position of all the folds it
11022 * does participate in.) A better solution with far fewer false
11023 * positives, and that would fill the nodes more completely, would
11024 * be to actually have available all the multi-character folds to
11025 * test against, and to back-off only far enough to be sure that
11026 * this node isn't ending with a partial one. <upper_parse> is set
11027 * further below (if we need to reparse the node) to include just
11028 * up through that final non-problematic character that this code
11029 * identifies, so when it is set to less than the full node, we can
11030 * skip the rest of this */
11031 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
11033 const STRLEN full_len = len;
11035 assert(len >= MAX_NODE_STRING_SIZE);
11037 /* Here, <s> points to the final byte of the final character.
11038 * Look backwards through the string until find a non-
11039 * problematic character */
11043 /* These two have no multi-char folds to non-UTF characters
11045 if (ASCII_FOLD_RESTRICTED || LOC) {
11049 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
11053 if (! PL_NonL1NonFinalFold) {
11054 PL_NonL1NonFinalFold = _new_invlist_C_array(
11055 NonL1_Perl_Non_Final_Folds_invlist);
11058 /* Point to the first byte of the final character */
11059 s = (char *) utf8_hop((U8 *) s, -1);
11061 while (s >= s0) { /* Search backwards until find
11062 non-problematic char */
11063 if (UTF8_IS_INVARIANT(*s)) {
11065 /* There are no ascii characters that participate
11066 * in multi-char folds under /aa. In EBCDIC, the
11067 * non-ascii invariants are all control characters,
11068 * so don't ever participate in any folds. */
11069 if (ASCII_FOLD_RESTRICTED
11070 || ! IS_NON_FINAL_FOLD(*s))
11075 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
11077 /* No Latin1 characters participate in multi-char
11078 * folds under /l */
11080 || ! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_UNI(
11086 else if (! _invlist_contains_cp(
11087 PL_NonL1NonFinalFold,
11088 valid_utf8_to_uvchr((U8 *) s, NULL)))
11093 /* Here, the current character is problematic in that
11094 * it does occur in the non-final position of some
11095 * fold, so try the character before it, but have to
11096 * special case the very first byte in the string, so
11097 * we don't read outside the string */
11098 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
11099 } /* End of loop backwards through the string */
11101 /* If there were only problematic characters in the string,
11102 * <s> will point to before s0, in which case the length
11103 * should be 0, otherwise include the length of the
11104 * non-problematic character just found */
11105 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
11108 /* Here, have found the final character, if any, that is
11109 * non-problematic as far as ending the node without splitting
11110 * it across a potential multi-char fold. <len> contains the
11111 * number of bytes in the node up-to and including that
11112 * character, or is 0 if there is no such character, meaning
11113 * the whole node contains only problematic characters. In
11114 * this case, give up and just take the node as-is. We can't
11120 /* Here, the node does contain some characters that aren't
11121 * problematic. If one such is the final character in the
11122 * node, we are done */
11123 if (len == full_len) {
11126 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
11128 /* If the final character is problematic, but the
11129 * penultimate is not, back-off that last character to
11130 * later start a new node with it */
11135 /* Here, the final non-problematic character is earlier
11136 * in the input than the penultimate character. What we do
11137 * is reparse from the beginning, going up only as far as
11138 * this final ok one, thus guaranteeing that the node ends
11139 * in an acceptable character. The reason we reparse is
11140 * that we know how far in the character is, but we don't
11141 * know how to correlate its position with the input parse.
11142 * An alternate implementation would be to build that
11143 * correlation as we go along during the original parse,
11144 * but that would entail extra work for every node, whereas
11145 * this code gets executed only when the string is too
11146 * large for the node, and the final two characters are
11147 * problematic, an infrequent occurrence. Yet another
11148 * possible strategy would be to save the tail of the
11149 * string, and the next time regatom is called, initialize
11150 * with that. The problem with this is that unless you
11151 * back off one more character, you won't be guaranteed
11152 * regatom will get called again, unless regbranch,
11153 * regpiece ... are also changed. If you do back off that
11154 * extra character, so that there is input guaranteed to
11155 * force calling regatom, you can't handle the case where
11156 * just the first character in the node is acceptable. I
11157 * (khw) decided to try this method which doesn't have that
11158 * pitfall; if performance issues are found, we can do a
11159 * combination of the current approach plus that one */
11165 } /* End of verifying node ends with an appropriate char */
11167 loopdone: /* Jumped to when encounters something that shouldn't be in
11170 /* If 'maybe_exact' is still set here, means there are no
11171 * code points in the node that participate in folds */
11172 if (FOLD && maybe_exact) {
11176 /* I (khw) don't know if you can get here with zero length, but the
11177 * old code handled this situation by creating a zero-length EXACT
11178 * node. Might as well be NOTHING instead */
11183 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
11186 RExC_parse = p - 1;
11187 Set_Node_Cur_Length(ret); /* MJD */
11188 nextchar(pRExC_state);
11190 /* len is STRLEN which is unsigned, need to copy to signed */
11193 vFAIL("Internal disaster");
11196 } /* End of label 'defchar:' */
11198 } /* End of giant switch on input character */
11204 S_regwhite( RExC_state_t *pRExC_state, char *p )
11206 const char *e = RExC_end;
11208 PERL_ARGS_ASSERT_REGWHITE;
11213 else if (*p == '#') {
11216 if (*p++ == '\n') {
11222 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11231 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
11233 /* Returns the next non-pattern-white space, non-comment character (the
11234 * latter only if 'recognize_comment is true) in the string p, which is
11235 * ended by RExC_end. If there is no line break ending a comment,
11236 * RExC_seen has added the REG_SEEN_RUN_ON_COMMENT flag; */
11237 const char *e = RExC_end;
11239 PERL_ARGS_ASSERT_REGPATWS;
11243 if ((len = is_PATWS_safe(p, e, UTF))) {
11246 else if (recognize_comment && *p == '#') {
11250 if (is_LNBREAK_safe(p, e, UTF)) {
11256 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11264 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
11265 Character classes ([:foo:]) can also be negated ([:^foo:]).
11266 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
11267 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
11268 but trigger failures because they are currently unimplemented. */
11270 #define POSIXCC_DONE(c) ((c) == ':')
11271 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
11272 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
11274 PERL_STATIC_INLINE I32
11275 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
11278 I32 namedclass = OOB_NAMEDCLASS;
11280 PERL_ARGS_ASSERT_REGPPOSIXCC;
11282 if (value == '[' && RExC_parse + 1 < RExC_end &&
11283 /* I smell either [: or [= or [. -- POSIX has been here, right? */
11284 POSIXCC(UCHARAT(RExC_parse)))
11286 const char c = UCHARAT(RExC_parse);
11287 char* const s = RExC_parse++;
11289 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
11291 if (RExC_parse == RExC_end) {
11294 /* Try to give a better location for the error (than the end of
11295 * the string) by looking for the matching ']' */
11297 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
11300 vFAIL2("Unmatched '%c' in POSIX class", c);
11302 /* Grandfather lone [:, [=, [. */
11306 const char* const t = RExC_parse++; /* skip over the c */
11309 if (UCHARAT(RExC_parse) == ']') {
11310 const char *posixcc = s + 1;
11311 RExC_parse++; /* skip over the ending ] */
11314 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
11315 const I32 skip = t - posixcc;
11317 /* Initially switch on the length of the name. */
11320 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
11321 this is the Perl \w
11323 namedclass = ANYOF_WORDCHAR;
11326 /* Names all of length 5. */
11327 /* alnum alpha ascii blank cntrl digit graph lower
11328 print punct space upper */
11329 /* Offset 4 gives the best switch position. */
11330 switch (posixcc[4]) {
11332 if (memEQ(posixcc, "alph", 4)) /* alpha */
11333 namedclass = ANYOF_ALPHA;
11336 if (memEQ(posixcc, "spac", 4)) /* space */
11337 namedclass = ANYOF_PSXSPC;
11340 if (memEQ(posixcc, "grap", 4)) /* graph */
11341 namedclass = ANYOF_GRAPH;
11344 if (memEQ(posixcc, "asci", 4)) /* ascii */
11345 namedclass = ANYOF_ASCII;
11348 if (memEQ(posixcc, "blan", 4)) /* blank */
11349 namedclass = ANYOF_BLANK;
11352 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
11353 namedclass = ANYOF_CNTRL;
11356 if (memEQ(posixcc, "alnu", 4)) /* alnum */
11357 namedclass = ANYOF_ALPHANUMERIC;
11360 if (memEQ(posixcc, "lowe", 4)) /* lower */
11361 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
11362 else if (memEQ(posixcc, "uppe", 4)) /* upper */
11363 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
11366 if (memEQ(posixcc, "digi", 4)) /* digit */
11367 namedclass = ANYOF_DIGIT;
11368 else if (memEQ(posixcc, "prin", 4)) /* print */
11369 namedclass = ANYOF_PRINT;
11370 else if (memEQ(posixcc, "punc", 4)) /* punct */
11371 namedclass = ANYOF_PUNCT;
11376 if (memEQ(posixcc, "xdigit", 6))
11377 namedclass = ANYOF_XDIGIT;
11381 if (namedclass == OOB_NAMEDCLASS)
11382 Simple_vFAIL3("POSIX class [:%.*s:] unknown",
11385 /* The #defines are structured so each complement is +1 to
11386 * the normal one */
11390 assert (posixcc[skip] == ':');
11391 assert (posixcc[skip+1] == ']');
11392 } else if (!SIZE_ONLY) {
11393 /* [[=foo=]] and [[.foo.]] are still future. */
11395 /* adjust RExC_parse so the warning shows after
11396 the class closes */
11397 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
11399 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11402 /* Maternal grandfather:
11403 * "[:" ending in ":" but not in ":]" */
11405 vFAIL("Unmatched '[' in POSIX class");
11408 /* Grandfather lone [:, [=, [. */
11418 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
11420 /* This applies some heuristics at the current parse position (which should
11421 * be at a '[') to see if what follows might be intended to be a [:posix:]
11422 * class. It returns true if it really is a posix class, of course, but it
11423 * also can return true if it thinks that what was intended was a posix
11424 * class that didn't quite make it.
11426 * It will return true for
11428 * [:alphanumerics] (as long as the ] isn't followed immediately by a
11429 * ')' indicating the end of the (?[
11430 * [:any garbage including %^&$ punctuation:]
11432 * This is designed to be called only from S_handle_regex_sets; it could be
11433 * easily adapted to be called from the spot at the beginning of regclass()
11434 * that checks to see in a normal bracketed class if the surrounding []
11435 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
11436 * change long-standing behavior, so I (khw) didn't do that */
11437 char* p = RExC_parse + 1;
11438 char first_char = *p;
11440 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
11442 assert(*(p - 1) == '[');
11444 if (! POSIXCC(first_char)) {
11449 while (p < RExC_end && isWORDCHAR(*p)) p++;
11451 if (p >= RExC_end) {
11455 if (p - RExC_parse > 2 /* Got at least 1 word character */
11456 && (*p == first_char
11457 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
11462 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
11465 && p - RExC_parse > 2 /* [:] evaluates to colon;
11466 [::] is a bad posix class. */
11467 && first_char == *(p - 1));
11471 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist, I32 *flagp, U32 depth,
11472 char * const oregcomp_parse)
11474 /* Handle the (?[...]) construct to do set operations */
11477 UV start, end; /* End points of code point ranges */
11479 char *save_end, *save_parse;
11484 const bool save_fold = FOLD;
11486 GET_RE_DEBUG_FLAGS_DECL;
11488 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
11491 vFAIL("(?[...]) not valid in locale");
11493 RExC_uni_semantics = 1;
11495 /* This will return only an ANYOF regnode, or (unlikely) something smaller
11496 * (such as EXACT). Thus we can skip most everything if just sizing. We
11497 * call regclass to handle '[]' so as to not have to reinvent its parsing
11498 * rules here (throwing away the size it computes each time). And, we exit
11499 * upon an unescaped ']' that isn't one ending a regclass. To do both
11500 * these things, we need to realize that something preceded by a backslash
11501 * is escaped, so we have to keep track of backslashes */
11504 Perl_ck_warner_d(aTHX_
11505 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
11506 "The regex_sets feature is experimental" REPORT_LOCATION,
11507 (int) (RExC_parse - RExC_precomp) , RExC_precomp, RExC_parse);
11509 while (RExC_parse < RExC_end) {
11510 SV* current = NULL;
11511 RExC_parse = regpatws(pRExC_state, RExC_parse,
11512 TRUE); /* means recognize comments */
11513 switch (*RExC_parse) {
11517 /* Skip the next byte (which could cause us to end up in
11518 * the middle of a UTF-8 character, but since none of those
11519 * are confusable with anything we currently handle in this
11520 * switch (invariants all), it's safe. We'll just hit the
11521 * default: case next time and keep on incrementing until
11522 * we find one of the invariants we do handle. */
11527 /* If this looks like it is a [:posix:] class, leave the
11528 * parse pointer at the '[' to fool regclass() into
11529 * thinking it is part of a '[[:posix:]]'. That function
11530 * will use strict checking to force a syntax error if it
11531 * doesn't work out to a legitimate class */
11532 bool is_posix_class
11533 = could_it_be_a_POSIX_class(pRExC_state);
11534 if (! is_posix_class) {
11538 /* regclass() can only return RESTART_UTF8 if multi-char
11539 folds are allowed. */
11540 if (!regclass(pRExC_state, flagp,depth+1,
11541 is_posix_class, /* parse the whole char
11542 class only if not a
11544 FALSE, /* don't allow multi-char folds */
11545 TRUE, /* silence non-portable warnings. */
11547 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X",
11550 /* function call leaves parse pointing to the ']', except
11551 * if we faked it */
11552 if (is_posix_class) {
11556 SvREFCNT_dec(current); /* In case it returned something */
11562 if (RExC_parse < RExC_end
11563 && *RExC_parse == ')')
11565 node = reganode(pRExC_state, ANYOF, 0);
11566 RExC_size += ANYOF_SKIP;
11567 nextchar(pRExC_state);
11568 Set_Node_Length(node,
11569 RExC_parse - oregcomp_parse + 1); /* MJD */
11578 FAIL("Syntax error in (?[...])");
11581 /* Pass 2 only after this. Everything in this construct is a
11582 * metacharacter. Operands begin with either a '\' (for an escape
11583 * sequence), or a '[' for a bracketed character class. Any other
11584 * character should be an operator, or parenthesis for grouping. Both
11585 * types of operands are handled by calling regclass() to parse them. It
11586 * is called with a parameter to indicate to return the computed inversion
11587 * list. The parsing here is implemented via a stack. Each entry on the
11588 * stack is a single character representing one of the operators, or the
11589 * '('; or else a pointer to an operand inversion list. */
11591 #define IS_OPERAND(a) (! SvIOK(a))
11593 /* The stack starts empty. It is a syntax error if the first thing parsed
11594 * is a binary operator; everything else is pushed on the stack. When an
11595 * operand is parsed, the top of the stack is examined. If it is a binary
11596 * operator, the item before it should be an operand, and both are replaced
11597 * by the result of doing that operation on the new operand and the one on
11598 * the stack. Thus a sequence of binary operands is reduced to a single
11599 * one before the next one is parsed.
11601 * A unary operator may immediately follow a binary in the input, for
11604 * When an operand is parsed and the top of the stack is a unary operator,
11605 * the operation is performed, and then the stack is rechecked to see if
11606 * this new operand is part of a binary operation; if so, it is handled as
11609 * A '(' is simply pushed on the stack; it is valid only if the stack is
11610 * empty, or the top element of the stack is an operator or another '('
11611 * (for which the parenthesized expression will become an operand). By the
11612 * time the corresponding ')' is parsed everything in between should have
11613 * been parsed and evaluated to a single operand (or else is a syntax
11614 * error), and is handled as a regular operand */
11618 while (RExC_parse < RExC_end) {
11619 I32 top_index = av_tindex(stack);
11621 SV* current = NULL;
11623 /* Skip white space */
11624 RExC_parse = regpatws(pRExC_state, RExC_parse,
11625 TRUE); /* means recognize comments */
11626 if (RExC_parse >= RExC_end) {
11627 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
11629 if ((curchar = UCHARAT(RExC_parse)) == ']') {
11636 if (av_tindex(stack) >= 0 /* This makes sure that we can
11637 safely subtract 1 from
11638 RExC_parse in the next clause.
11639 If we have something on the
11640 stack, we have parsed something
11642 && UCHARAT(RExC_parse - 1) == '('
11643 && RExC_parse < RExC_end)
11645 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
11646 * This happens when we have some thing like
11648 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
11650 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
11652 * Here we would be handling the interpolated
11653 * '$thai_or_lao'. We handle this by a recursive call to
11654 * ourselves which returns the inversion list the
11655 * interpolated expression evaluates to. We use the flags
11656 * from the interpolated pattern. */
11657 U32 save_flags = RExC_flags;
11658 const char * const save_parse = ++RExC_parse;
11660 parse_lparen_question_flags(pRExC_state);
11662 if (RExC_parse == save_parse /* Makes sure there was at
11663 least one flag (or this
11664 embedding wasn't compiled)
11666 || RExC_parse >= RExC_end - 4
11667 || UCHARAT(RExC_parse) != ':'
11668 || UCHARAT(++RExC_parse) != '('
11669 || UCHARAT(++RExC_parse) != '?'
11670 || UCHARAT(++RExC_parse) != '[')
11673 /* In combination with the above, this moves the
11674 * pointer to the point just after the first erroneous
11675 * character (or if there are no flags, to where they
11676 * should have been) */
11677 if (RExC_parse >= RExC_end - 4) {
11678 RExC_parse = RExC_end;
11680 else if (RExC_parse != save_parse) {
11681 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11683 vFAIL("Expecting '(?flags:(?[...'");
11686 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
11687 depth+1, oregcomp_parse);
11689 /* Here, 'current' contains the embedded expression's
11690 * inversion list, and RExC_parse points to the trailing
11691 * ']'; the next character should be the ')' which will be
11692 * paired with the '(' that has been put on the stack, so
11693 * the whole embedded expression reduces to '(operand)' */
11696 RExC_flags = save_flags;
11697 goto handle_operand;
11702 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11703 vFAIL("Unexpected character");
11706 /* regclass() can only return RESTART_UTF8 if multi-char
11707 folds are allowed. */
11708 if (!regclass(pRExC_state, flagp,depth+1,
11709 TRUE, /* means parse just the next thing */
11710 FALSE, /* don't allow multi-char folds */
11711 FALSE, /* don't silence non-portable warnings. */
11713 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X",
11715 /* regclass() will return with parsing just the \ sequence,
11716 * leaving the parse pointer at the next thing to parse */
11718 goto handle_operand;
11720 case '[': /* Is a bracketed character class */
11722 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
11724 if (! is_posix_class) {
11728 /* regclass() can only return RESTART_UTF8 if multi-char
11729 folds are allowed. */
11730 if(!regclass(pRExC_state, flagp,depth+1,
11731 is_posix_class, /* parse the whole char class
11732 only if not a posix class */
11733 FALSE, /* don't allow multi-char folds */
11734 FALSE, /* don't silence non-portable warnings. */
11736 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X",
11738 /* function call leaves parse pointing to the ']', except if we
11740 if (is_posix_class) {
11744 goto handle_operand;
11753 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
11754 || ! IS_OPERAND(*top_ptr))
11757 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
11759 av_push(stack, newSVuv(curchar));
11763 av_push(stack, newSVuv(curchar));
11767 if (top_index >= 0) {
11768 top_ptr = av_fetch(stack, top_index, FALSE);
11770 if (IS_OPERAND(*top_ptr)) {
11772 vFAIL("Unexpected '(' with no preceding operator");
11775 av_push(stack, newSVuv(curchar));
11782 || ! (current = av_pop(stack))
11783 || ! IS_OPERAND(current)
11784 || ! (lparen = av_pop(stack))
11785 || IS_OPERAND(lparen)
11786 || SvUV(lparen) != '(')
11789 vFAIL("Unexpected ')'");
11792 SvREFCNT_dec_NN(lparen);
11799 /* Here, we have an operand to process, in 'current' */
11801 if (top_index < 0) { /* Just push if stack is empty */
11802 av_push(stack, current);
11805 SV* top = av_pop(stack);
11806 char current_operator;
11808 if (IS_OPERAND(top)) {
11809 vFAIL("Operand with no preceding operator");
11811 current_operator = (char) SvUV(top);
11812 switch (current_operator) {
11813 case '(': /* Push the '(' back on followed by the new
11815 av_push(stack, top);
11816 av_push(stack, current);
11817 SvREFCNT_inc(top); /* Counters the '_dec' done
11818 just after the 'break', so
11819 it doesn't get wrongly freed
11824 _invlist_invert(current);
11826 /* Unlike binary operators, the top of the stack,
11827 * now that this unary one has been popped off, may
11828 * legally be an operator, and we now have operand
11831 SvREFCNT_dec_NN(top);
11832 goto handle_operand;
11835 _invlist_intersection(av_pop(stack),
11838 av_push(stack, current);
11843 _invlist_union(av_pop(stack), current, ¤t);
11844 av_push(stack, current);
11848 _invlist_subtract(av_pop(stack), current, ¤t);
11849 av_push(stack, current);
11852 case '^': /* The union minus the intersection */
11858 element = av_pop(stack);
11859 _invlist_union(element, current, &u);
11860 _invlist_intersection(element, current, &i);
11861 _invlist_subtract(u, i, ¤t);
11862 av_push(stack, current);
11863 SvREFCNT_dec_NN(i);
11864 SvREFCNT_dec_NN(u);
11865 SvREFCNT_dec_NN(element);
11870 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
11872 SvREFCNT_dec_NN(top);
11876 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11879 if (av_tindex(stack) < 0 /* Was empty */
11880 || ((final = av_pop(stack)) == NULL)
11881 || ! IS_OPERAND(final)
11882 || av_tindex(stack) >= 0) /* More left on stack */
11884 vFAIL("Incomplete expression within '(?[ ])'");
11887 /* Here, 'final' is the resultant inversion list from evaluating the
11888 * expression. Return it if so requested */
11889 if (return_invlist) {
11890 *return_invlist = final;
11894 /* Otherwise generate a resultant node, based on 'final'. regclass() is
11895 * expecting a string of ranges and individual code points */
11896 invlist_iterinit(final);
11897 result_string = newSVpvs("");
11898 while (invlist_iternext(final, &start, &end)) {
11899 if (start == end) {
11900 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
11903 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
11908 save_parse = RExC_parse;
11909 RExC_parse = SvPV(result_string, len);
11910 save_end = RExC_end;
11911 RExC_end = RExC_parse + len;
11913 /* We turn off folding around the call, as the class we have constructed
11914 * already has all folding taken into consideration, and we don't want
11915 * regclass() to add to that */
11916 RExC_flags &= ~RXf_PMf_FOLD;
11917 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
11919 node = regclass(pRExC_state, flagp,depth+1,
11920 FALSE, /* means parse the whole char class */
11921 FALSE, /* don't allow multi-char folds */
11922 TRUE, /* silence non-portable warnings. The above may very
11923 well have generated non-portable code points, but
11924 they're valid on this machine */
11927 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
11930 RExC_flags |= RXf_PMf_FOLD;
11932 RExC_parse = save_parse + 1;
11933 RExC_end = save_end;
11934 SvREFCNT_dec_NN(final);
11935 SvREFCNT_dec_NN(result_string);
11936 SvREFCNT_dec_NN(stack);
11938 nextchar(pRExC_state);
11939 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
11944 /* The names of properties whose definitions are not known at compile time are
11945 * stored in this SV, after a constant heading. So if the length has been
11946 * changed since initialization, then there is a run-time definition. */
11947 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION (SvCUR(listsv) != initial_listsv_len)
11950 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
11951 const bool stop_at_1, /* Just parse the next thing, don't
11952 look for a full character class */
11953 bool allow_multi_folds,
11954 const bool silence_non_portable, /* Don't output warnings
11957 SV** ret_invlist) /* Return an inversion list, not a node */
11959 /* parse a bracketed class specification. Most of these will produce an
11960 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
11961 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
11962 * under /i with multi-character folds: it will be rewritten following the
11963 * paradigm of this example, where the <multi-fold>s are characters which
11964 * fold to multiple character sequences:
11965 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
11966 * gets effectively rewritten as:
11967 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
11968 * reg() gets called (recursively) on the rewritten version, and this
11969 * function will return what it constructs. (Actually the <multi-fold>s
11970 * aren't physically removed from the [abcdefghi], it's just that they are
11971 * ignored in the recursion by means of a flag:
11972 * <RExC_in_multi_char_class>.)
11974 * ANYOF nodes contain a bit map for the first 256 characters, with the
11975 * corresponding bit set if that character is in the list. For characters
11976 * above 255, a range list or swash is used. There are extra bits for \w,
11977 * etc. in locale ANYOFs, as what these match is not determinable at
11980 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
11981 * to be restarted. This can only happen if ret_invlist is non-NULL.
11985 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
11987 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
11990 IV namedclass = OOB_NAMEDCLASS;
11991 char *rangebegin = NULL;
11992 bool need_class = 0;
11994 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
11995 than just initialized. */
11996 SV* properties = NULL; /* Code points that match \p{} \P{} */
11997 SV* posixes = NULL; /* Code points that match classes like, [:word:],
11998 extended beyond the Latin1 range */
11999 UV element_count = 0; /* Number of distinct elements in the class.
12000 Optimizations may be possible if this is tiny */
12001 AV * multi_char_matches = NULL; /* Code points that fold to more than one
12002 character; used under /i */
12004 char * stop_ptr = RExC_end; /* where to stop parsing */
12005 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
12007 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
12009 /* Unicode properties are stored in a swash; this holds the current one
12010 * being parsed. If this swash is the only above-latin1 component of the
12011 * character class, an optimization is to pass it directly on to the
12012 * execution engine. Otherwise, it is set to NULL to indicate that there
12013 * are other things in the class that have to be dealt with at execution
12015 SV* swash = NULL; /* Code points that match \p{} \P{} */
12017 /* Set if a component of this character class is user-defined; just passed
12018 * on to the engine */
12019 bool has_user_defined_property = FALSE;
12021 /* inversion list of code points this node matches only when the target
12022 * string is in UTF-8. (Because is under /d) */
12023 SV* depends_list = NULL;
12025 /* inversion list of code points this node matches. For much of the
12026 * function, it includes only those that match regardless of the utf8ness
12027 * of the target string */
12028 SV* cp_list = NULL;
12031 /* In a range, counts how many 0-2 of the ends of it came from literals,
12032 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
12033 UV literal_endpoint = 0;
12035 bool invert = FALSE; /* Is this class to be complemented */
12037 /* Is there any thing like \W or [:^digit:] that matches above the legal
12038 * Unicode range? */
12039 bool runtime_posix_matches_above_Unicode = FALSE;
12041 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
12042 case we need to change the emitted regop to an EXACT. */
12043 const char * orig_parse = RExC_parse;
12044 const I32 orig_size = RExC_size;
12045 GET_RE_DEBUG_FLAGS_DECL;
12047 PERL_ARGS_ASSERT_REGCLASS;
12049 PERL_UNUSED_ARG(depth);
12052 DEBUG_PARSE("clas");
12054 /* Assume we are going to generate an ANYOF node. */
12055 ret = reganode(pRExC_state, ANYOF, 0);
12058 RExC_size += ANYOF_SKIP;
12059 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
12062 ANYOF_FLAGS(ret) = 0;
12064 RExC_emit += ANYOF_SKIP;
12066 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
12068 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
12069 initial_listsv_len = SvCUR(listsv);
12070 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
12074 RExC_parse = regpatws(pRExC_state, RExC_parse,
12075 FALSE /* means don't recognize comments */);
12078 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
12081 allow_multi_folds = FALSE;
12084 RExC_parse = regpatws(pRExC_state, RExC_parse,
12085 FALSE /* means don't recognize comments */);
12089 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
12090 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
12091 const char *s = RExC_parse;
12092 const char c = *s++;
12094 while (isWORDCHAR(*s))
12096 if (*s && c == *s && s[1] == ']') {
12097 SAVEFREESV(RExC_rx_sv);
12099 "POSIX syntax [%c %c] belongs inside character classes",
12101 (void)ReREFCNT_inc(RExC_rx_sv);
12105 /* If the caller wants us to just parse a single element, accomplish this
12106 * by faking the loop ending condition */
12107 if (stop_at_1 && RExC_end > RExC_parse) {
12108 stop_ptr = RExC_parse + 1;
12111 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
12112 if (UCHARAT(RExC_parse) == ']')
12113 goto charclassloop;
12117 if (RExC_parse >= stop_ptr) {
12122 RExC_parse = regpatws(pRExC_state, RExC_parse,
12123 FALSE /* means don't recognize comments */);
12126 if (UCHARAT(RExC_parse) == ']') {
12132 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
12133 save_value = value;
12134 save_prevvalue = prevvalue;
12137 rangebegin = RExC_parse;
12141 value = utf8n_to_uvchr((U8*)RExC_parse,
12142 RExC_end - RExC_parse,
12143 &numlen, UTF8_ALLOW_DEFAULT);
12144 RExC_parse += numlen;
12147 value = UCHARAT(RExC_parse++);
12150 && RExC_parse < RExC_end
12151 && POSIXCC(UCHARAT(RExC_parse)))
12153 namedclass = regpposixcc(pRExC_state, value, strict);
12155 else if (value == '\\') {
12157 value = utf8n_to_uvchr((U8*)RExC_parse,
12158 RExC_end - RExC_parse,
12159 &numlen, UTF8_ALLOW_DEFAULT);
12160 RExC_parse += numlen;
12163 value = UCHARAT(RExC_parse++);
12165 /* Some compilers cannot handle switching on 64-bit integer
12166 * values, therefore value cannot be an UV. Yes, this will
12167 * be a problem later if we want switch on Unicode.
12168 * A similar issue a little bit later when switching on
12169 * namedclass. --jhi */
12171 /* If the \ is escaping white space when white space is being
12172 * skipped, it means that that white space is wanted literally, and
12173 * is already in 'value'. Otherwise, need to translate the escape
12174 * into what it signifies. */
12175 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
12177 case 'w': namedclass = ANYOF_WORDCHAR; break;
12178 case 'W': namedclass = ANYOF_NWORDCHAR; break;
12179 case 's': namedclass = ANYOF_SPACE; break;
12180 case 'S': namedclass = ANYOF_NSPACE; break;
12181 case 'd': namedclass = ANYOF_DIGIT; break;
12182 case 'D': namedclass = ANYOF_NDIGIT; break;
12183 case 'v': namedclass = ANYOF_VERTWS; break;
12184 case 'V': namedclass = ANYOF_NVERTWS; break;
12185 case 'h': namedclass = ANYOF_HORIZWS; break;
12186 case 'H': namedclass = ANYOF_NHORIZWS; break;
12187 case 'N': /* Handle \N{NAME} in class */
12189 /* We only pay attention to the first char of
12190 multichar strings being returned. I kinda wonder
12191 if this makes sense as it does change the behaviour
12192 from earlier versions, OTOH that behaviour was broken
12194 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
12195 TRUE, /* => charclass */
12198 if (*flagp & RESTART_UTF8)
12199 FAIL("panic: grok_bslash_N set RESTART_UTF8");
12209 /* We will handle any undefined properties ourselves */
12210 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF;
12212 if (RExC_parse >= RExC_end)
12213 vFAIL2("Empty \\%c{}", (U8)value);
12214 if (*RExC_parse == '{') {
12215 const U8 c = (U8)value;
12216 e = strchr(RExC_parse++, '}');
12218 vFAIL2("Missing right brace on \\%c{}", c);
12219 while (isSPACE(UCHARAT(RExC_parse)))
12221 if (e == RExC_parse)
12222 vFAIL2("Empty \\%c{}", c);
12223 n = e - RExC_parse;
12224 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
12235 if (UCHARAT(RExC_parse) == '^') {
12238 /* toggle. (The rhs xor gets the single bit that
12239 * differs between P and p; the other xor inverts just
12241 value ^= 'P' ^ 'p';
12243 while (isSPACE(UCHARAT(RExC_parse))) {
12248 /* Try to get the definition of the property into
12249 * <invlist>. If /i is in effect, the effective property
12250 * will have its name be <__NAME_i>. The design is
12251 * discussed in commit
12252 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
12253 Newx(name, n + sizeof("_i__\n"), char);
12255 sprintf(name, "%s%.*s%s\n",
12256 (FOLD) ? "__" : "",
12262 /* Look up the property name, and get its swash and
12263 * inversion list, if the property is found */
12265 SvREFCNT_dec_NN(swash);
12267 swash = _core_swash_init("utf8", name, &PL_sv_undef,
12270 NULL, /* No inversion list */
12273 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
12275 SvREFCNT_dec_NN(swash);
12279 /* Here didn't find it. It could be a user-defined
12280 * property that will be available at run-time. If we
12281 * accept only compile-time properties, is an error;
12282 * otherwise add it to the list for run-time look up */
12284 RExC_parse = e + 1;
12285 vFAIL3("Property '%.*s' is unknown", (int) n, name);
12287 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s\n",
12288 (value == 'p' ? '+' : '!'),
12290 has_user_defined_property = TRUE;
12292 /* We don't know yet, so have to assume that the
12293 * property could match something in the Latin1 range,
12294 * hence something that isn't utf8. Note that this
12295 * would cause things in <depends_list> to match
12296 * inappropriately, except that any \p{}, including
12297 * this one forces Unicode semantics, which means there
12298 * is <no depends_list> */
12299 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
12303 /* Here, did get the swash and its inversion list. If
12304 * the swash is from a user-defined property, then this
12305 * whole character class should be regarded as such */
12306 has_user_defined_property =
12308 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY);
12310 /* Invert if asking for the complement */
12311 if (value == 'P') {
12312 _invlist_union_complement_2nd(properties,
12316 /* The swash can't be used as-is, because we've
12317 * inverted things; delay removing it to here after
12318 * have copied its invlist above */
12319 SvREFCNT_dec_NN(swash);
12323 _invlist_union(properties, invlist, &properties);
12328 RExC_parse = e + 1;
12329 namedclass = ANYOF_UNIPROP; /* no official name, but it's
12332 /* \p means they want Unicode semantics */
12333 RExC_uni_semantics = 1;
12336 case 'n': value = '\n'; break;
12337 case 'r': value = '\r'; break;
12338 case 't': value = '\t'; break;
12339 case 'f': value = '\f'; break;
12340 case 'b': value = '\b'; break;
12341 case 'e': value = ASCII_TO_NATIVE('\033');break;
12342 case 'a': value = ASCII_TO_NATIVE('\007');break;
12344 RExC_parse--; /* function expects to be pointed at the 'o' */
12346 const char* error_msg;
12347 bool valid = grok_bslash_o(&RExC_parse,
12350 SIZE_ONLY, /* warnings in pass
12353 silence_non_portable,
12359 if (PL_encoding && value < 0x100) {
12360 goto recode_encoding;
12364 RExC_parse--; /* function expects to be pointed at the 'x' */
12366 const char* error_msg;
12367 bool valid = grok_bslash_x(&RExC_parse,
12370 TRUE, /* Output warnings */
12372 silence_non_portable,
12378 if (PL_encoding && value < 0x100)
12379 goto recode_encoding;
12382 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
12384 case '0': case '1': case '2': case '3': case '4':
12385 case '5': case '6': case '7':
12387 /* Take 1-3 octal digits */
12388 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12389 numlen = (strict) ? 4 : 3;
12390 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
12391 RExC_parse += numlen;
12394 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12395 vFAIL("Need exactly 3 octal digits");
12397 else if (! SIZE_ONLY /* like \08, \178 */
12399 && RExC_parse < RExC_end
12400 && isDIGIT(*RExC_parse)
12401 && ckWARN(WARN_REGEXP))
12403 SAVEFREESV(RExC_rx_sv);
12404 reg_warn_non_literal_string(
12406 form_short_octal_warning(RExC_parse, numlen));
12407 (void)ReREFCNT_inc(RExC_rx_sv);
12410 if (PL_encoding && value < 0x100)
12411 goto recode_encoding;
12415 if (! RExC_override_recoding) {
12416 SV* enc = PL_encoding;
12417 value = reg_recode((const char)(U8)value, &enc);
12420 vFAIL("Invalid escape in the specified encoding");
12422 else if (SIZE_ONLY) {
12423 ckWARNreg(RExC_parse,
12424 "Invalid escape in the specified encoding");
12430 /* Allow \_ to not give an error */
12431 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
12433 vFAIL2("Unrecognized escape \\%c in character class",
12437 SAVEFREESV(RExC_rx_sv);
12438 ckWARN2reg(RExC_parse,
12439 "Unrecognized escape \\%c in character class passed through",
12441 (void)ReREFCNT_inc(RExC_rx_sv);
12445 } /* End of switch on char following backslash */
12446 } /* end of handling backslash escape sequences */
12449 literal_endpoint++;
12452 /* Here, we have the current token in 'value' */
12454 /* What matches in a locale is not known until runtime. This includes
12455 * what the Posix classes (like \w, [:space:]) match. Room must be
12456 * reserved (one time per class) to store such classes, either if Perl
12457 * is compiled so that locale nodes always should have this space, or
12458 * if there is such class info to be stored. The space will contain a
12459 * bit for each named class that is to be matched against. This isn't
12460 * needed for \p{} and pseudo-classes, as they are not affected by
12461 * locale, and hence are dealt with separately */
12464 && (ANYOF_LOCALE == ANYOF_CLASS
12465 || (namedclass > OOB_NAMEDCLASS && namedclass < ANYOF_MAX)))
12469 RExC_size += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12472 RExC_emit += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12473 ANYOF_CLASS_ZERO(ret);
12475 ANYOF_FLAGS(ret) |= ANYOF_CLASS;
12478 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
12480 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
12481 * literal, as is the character that began the false range, i.e.
12482 * the 'a' in the examples */
12485 const int w = (RExC_parse >= rangebegin)
12486 ? RExC_parse - rangebegin
12489 vFAIL4("False [] range \"%*.*s\"", w, w, rangebegin);
12492 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
12493 ckWARN4reg(RExC_parse,
12494 "False [] range \"%*.*s\"",
12496 (void)ReREFCNT_inc(RExC_rx_sv);
12497 cp_list = add_cp_to_invlist(cp_list, '-');
12498 cp_list = add_cp_to_invlist(cp_list, prevvalue);
12502 range = 0; /* this was not a true range */
12503 element_count += 2; /* So counts for three values */
12507 U8 classnum = namedclass_to_classnum(namedclass);
12508 if (namedclass >= ANYOF_MAX) { /* If a special class */
12509 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
12511 /* Here, should be \h, \H, \v, or \V. Neither /d nor
12512 * /l make a difference in what these match. There
12513 * would be problems if these characters had folds
12514 * other than themselves, as cp_list is subject to
12516 if (classnum != _CC_VERTSPACE) {
12517 assert( namedclass == ANYOF_HORIZWS
12518 || namedclass == ANYOF_NHORIZWS);
12520 /* It turns out that \h is just a synonym for
12522 classnum = _CC_BLANK;
12525 _invlist_union_maybe_complement_2nd(
12527 PL_XPosix_ptrs[classnum],
12528 cBOOL(namedclass % 2), /* Complement if odd
12529 (NHORIZWS, NVERTWS)
12534 else if (classnum == _CC_ASCII) {
12537 ANYOF_CLASS_SET(ret, namedclass);
12540 #endif /* Not isascii(); just use the hard-coded definition for it */
12541 _invlist_union_maybe_complement_2nd(
12544 cBOOL(namedclass % 2), /* Complement if odd
12548 else { /* Garden variety class */
12550 /* The ascii range inversion list */
12551 SV* ascii_source = PL_Posix_ptrs[classnum];
12553 /* The full Latin1 range inversion list */
12554 SV* l1_source = PL_L1Posix_ptrs[classnum];
12556 /* This code is structured into two major clauses. The
12557 * first is for classes whose complete definitions may not
12558 * already be known. It not, the Latin1 definition
12559 * (guaranteed to already known) is used plus code is
12560 * generated to load the rest at run-time (only if needed).
12561 * If the complete definition is known, it drops down to
12562 * the second clause, where the complete definition is
12565 if (classnum < _FIRST_NON_SWASH_CC) {
12567 /* Here, the class has a swash, which may or not
12568 * already be loaded */
12570 /* The name of the property to use to match the full
12571 * eXtended Unicode range swash for this character
12573 const char *Xname = swash_property_names[classnum];
12575 /* If returning the inversion list, we can't defer
12576 * getting this until runtime */
12577 if (ret_invlist && ! PL_utf8_swash_ptrs[classnum]) {
12578 PL_utf8_swash_ptrs[classnum] =
12579 _core_swash_init("utf8", Xname, &PL_sv_undef,
12582 NULL, /* No inversion list */
12583 NULL /* No flags */
12585 assert(PL_utf8_swash_ptrs[classnum]);
12587 if ( ! PL_utf8_swash_ptrs[classnum]) {
12588 if (namedclass % 2 == 0) { /* A non-complemented
12590 /* If not /a matching, there are code points we
12591 * don't know at compile time. Arrange for the
12592 * unknown matches to be loaded at run-time, if
12594 if (! AT_LEAST_ASCII_RESTRICTED) {
12595 Perl_sv_catpvf(aTHX_ listsv, "+utf8::%s\n",
12598 if (LOC) { /* Under locale, set run-time
12600 ANYOF_CLASS_SET(ret, namedclass);
12603 /* Add the current class's code points to
12604 * the running total */
12605 _invlist_union(posixes,
12606 (AT_LEAST_ASCII_RESTRICTED)
12612 else { /* A complemented class */
12613 if (AT_LEAST_ASCII_RESTRICTED) {
12614 /* Under /a should match everything above
12615 * ASCII, plus the complement of the set's
12617 _invlist_union_complement_2nd(posixes,
12622 /* Arrange for the unknown matches to be
12623 * loaded at run-time, if needed */
12624 Perl_sv_catpvf(aTHX_ listsv, "!utf8::%s\n",
12626 runtime_posix_matches_above_Unicode = TRUE;
12628 ANYOF_CLASS_SET(ret, namedclass);
12632 /* We want to match everything in
12633 * Latin1, except those things that
12634 * l1_source matches */
12635 SV* scratch_list = NULL;
12636 _invlist_subtract(PL_Latin1, l1_source,
12639 /* Add the list from this class to the
12642 posixes = scratch_list;
12645 _invlist_union(posixes,
12648 SvREFCNT_dec_NN(scratch_list);
12650 if (DEPENDS_SEMANTICS) {
12652 |= ANYOF_NON_UTF8_LATIN1_ALL;
12657 goto namedclass_done;
12660 /* Here, there is a swash loaded for the class. If no
12661 * inversion list for it yet, get it */
12662 if (! PL_XPosix_ptrs[classnum]) {
12663 PL_XPosix_ptrs[classnum]
12664 = _swash_to_invlist(PL_utf8_swash_ptrs[classnum]);
12668 /* Here there is an inversion list already loaded for the
12671 if (namedclass % 2 == 0) { /* A non-complemented class,
12672 like ANYOF_PUNCT */
12674 /* For non-locale, just add it to any existing list
12676 _invlist_union(posixes,
12677 (AT_LEAST_ASCII_RESTRICTED)
12679 : PL_XPosix_ptrs[classnum],
12682 else { /* Locale */
12683 SV* scratch_list = NULL;
12685 /* For above Latin1 code points, we use the full
12687 _invlist_intersection(PL_AboveLatin1,
12688 PL_XPosix_ptrs[classnum],
12690 /* And set the output to it, adding instead if
12691 * there already is an output. Checking if
12692 * 'posixes' is NULL first saves an extra clone.
12693 * Its reference count will be decremented at the
12694 * next union, etc, or if this is the only
12695 * instance, at the end of the routine */
12697 posixes = scratch_list;
12700 _invlist_union(posixes, scratch_list, &posixes);
12701 SvREFCNT_dec_NN(scratch_list);
12704 #ifndef HAS_ISBLANK
12705 if (namedclass != ANYOF_BLANK) {
12707 /* Set this class in the node for runtime
12709 ANYOF_CLASS_SET(ret, namedclass);
12710 #ifndef HAS_ISBLANK
12713 /* No isblank(), use the hard-coded ASCII-range
12714 * blanks, adding them to the running total. */
12716 _invlist_union(posixes, ascii_source, &posixes);
12721 else { /* A complemented class, like ANYOF_NPUNCT */
12723 _invlist_union_complement_2nd(
12725 (AT_LEAST_ASCII_RESTRICTED)
12727 : PL_XPosix_ptrs[classnum],
12729 /* Under /d, everything in the upper half of the
12730 * Latin1 range matches this complement */
12731 if (DEPENDS_SEMANTICS) {
12732 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL;
12735 else { /* Locale */
12736 SV* scratch_list = NULL;
12737 _invlist_subtract(PL_AboveLatin1,
12738 PL_XPosix_ptrs[classnum],
12741 posixes = scratch_list;
12744 _invlist_union(posixes, scratch_list, &posixes);
12745 SvREFCNT_dec_NN(scratch_list);
12747 #ifndef HAS_ISBLANK
12748 if (namedclass != ANYOF_NBLANK) {
12750 ANYOF_CLASS_SET(ret, namedclass);
12751 #ifndef HAS_ISBLANK
12754 /* Get the list of all code points in Latin1
12755 * that are not ASCII blanks, and add them to
12756 * the running total */
12757 _invlist_subtract(PL_Latin1, ascii_source,
12759 _invlist_union(posixes, scratch_list, &posixes);
12760 SvREFCNT_dec_NN(scratch_list);
12767 continue; /* Go get next character */
12769 } /* end of namedclass \blah */
12771 /* Here, we have a single value. If 'range' is set, it is the ending
12772 * of a range--check its validity. Later, we will handle each
12773 * individual code point in the range. If 'range' isn't set, this
12774 * could be the beginning of a range, so check for that by looking
12775 * ahead to see if the next real character to be processed is the range
12776 * indicator--the minus sign */
12779 RExC_parse = regpatws(pRExC_state, RExC_parse,
12780 FALSE /* means don't recognize comments */);
12784 if (prevvalue > value) /* b-a */ {
12785 const int w = RExC_parse - rangebegin;
12786 Simple_vFAIL4("Invalid [] range \"%*.*s\"", w, w, rangebegin);
12787 range = 0; /* not a valid range */
12791 prevvalue = value; /* save the beginning of the potential range */
12792 if (! stop_at_1 /* Can't be a range if parsing just one thing */
12793 && *RExC_parse == '-')
12795 char* next_char_ptr = RExC_parse + 1;
12796 if (skip_white) { /* Get the next real char after the '-' */
12797 next_char_ptr = regpatws(pRExC_state,
12799 FALSE); /* means don't recognize
12803 /* If the '-' is at the end of the class (just before the ']',
12804 * it is a literal minus; otherwise it is a range */
12805 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
12806 RExC_parse = next_char_ptr;
12808 /* a bad range like \w-, [:word:]- ? */
12809 if (namedclass > OOB_NAMEDCLASS) {
12810 if (strict || ckWARN(WARN_REGEXP)) {
12812 RExC_parse >= rangebegin ?
12813 RExC_parse - rangebegin : 0;
12815 vFAIL4("False [] range \"%*.*s\"",
12820 "False [] range \"%*.*s\"",
12825 cp_list = add_cp_to_invlist(cp_list, '-');
12829 range = 1; /* yeah, it's a range! */
12830 continue; /* but do it the next time */
12835 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
12838 /* non-Latin1 code point implies unicode semantics. Must be set in
12839 * pass1 so is there for the whole of pass 2 */
12841 RExC_uni_semantics = 1;
12844 /* Ready to process either the single value, or the completed range.
12845 * For single-valued non-inverted ranges, we consider the possibility
12846 * of multi-char folds. (We made a conscious decision to not do this
12847 * for the other cases because it can often lead to non-intuitive
12848 * results. For example, you have the peculiar case that:
12849 * "s s" =~ /^[^\xDF]+$/i => Y
12850 * "ss" =~ /^[^\xDF]+$/i => N
12852 * See [perl #89750] */
12853 if (FOLD && allow_multi_folds && value == prevvalue) {
12854 if (value == LATIN_SMALL_LETTER_SHARP_S
12855 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
12858 /* Here <value> is indeed a multi-char fold. Get what it is */
12860 U8 foldbuf[UTF8_MAXBYTES_CASE];
12863 UV folded = _to_uni_fold_flags(
12868 | ((LOC) ? FOLD_FLAGS_LOCALE
12869 : (ASCII_FOLD_RESTRICTED)
12870 ? FOLD_FLAGS_NOMIX_ASCII
12874 /* Here, <folded> should be the first character of the
12875 * multi-char fold of <value>, with <foldbuf> containing the
12876 * whole thing. But, if this fold is not allowed (because of
12877 * the flags), <fold> will be the same as <value>, and should
12878 * be processed like any other character, so skip the special
12880 if (folded != value) {
12882 /* Skip if we are recursed, currently parsing the class
12883 * again. Otherwise add this character to the list of
12884 * multi-char folds. */
12885 if (! RExC_in_multi_char_class) {
12886 AV** this_array_ptr;
12888 STRLEN cp_count = utf8_length(foldbuf,
12889 foldbuf + foldlen);
12890 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
12892 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
12895 if (! multi_char_matches) {
12896 multi_char_matches = newAV();
12899 /* <multi_char_matches> is actually an array of arrays.
12900 * There will be one or two top-level elements: [2],
12901 * and/or [3]. The [2] element is an array, each
12902 * element thereof is a character which folds to two
12903 * characters; likewise for [3]. (Unicode guarantees a
12904 * maximum of 3 characters in any fold.) When we
12905 * rewrite the character class below, we will do so
12906 * such that the longest folds are written first, so
12907 * that it prefers the longest matching strings first.
12908 * This is done even if it turns out that any
12909 * quantifier is non-greedy, out of programmer
12910 * laziness. Tom Christiansen has agreed that this is
12911 * ok. This makes the test for the ligature 'ffi' come
12912 * before the test for 'ff' */
12913 if (av_exists(multi_char_matches, cp_count)) {
12914 this_array_ptr = (AV**) av_fetch(multi_char_matches,
12916 this_array = *this_array_ptr;
12919 this_array = newAV();
12920 av_store(multi_char_matches, cp_count,
12923 av_push(this_array, multi_fold);
12926 /* This element should not be processed further in this
12929 value = save_value;
12930 prevvalue = save_prevvalue;
12936 /* Deal with this element of the class */
12939 cp_list = _add_range_to_invlist(cp_list, prevvalue, value);
12941 SV* this_range = _new_invlist(1);
12942 _append_range_to_invlist(this_range, prevvalue, value);
12944 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
12945 * If this range was specified using something like 'i-j', we want
12946 * to include only the 'i' and the 'j', and not anything in
12947 * between, so exclude non-ASCII, non-alphabetics from it.
12948 * However, if the range was specified with something like
12949 * [\x89-\x91] or [\x89-j], all code points within it should be
12950 * included. literal_endpoint==2 means both ends of the range used
12951 * a literal character, not \x{foo} */
12952 if (literal_endpoint == 2
12953 && (prevvalue >= 'a' && value <= 'z')
12954 || (prevvalue >= 'A' && value <= 'Z'))
12956 _invlist_intersection(this_range, PL_Posix_ptrs[_CC_ALPHA],
12959 _invlist_union(cp_list, this_range, &cp_list);
12960 literal_endpoint = 0;
12964 range = 0; /* this range (if it was one) is done now */
12965 } /* End of loop through all the text within the brackets */
12967 /* If anything in the class expands to more than one character, we have to
12968 * deal with them by building up a substitute parse string, and recursively
12969 * calling reg() on it, instead of proceeding */
12970 if (multi_char_matches) {
12971 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
12974 char *save_end = RExC_end;
12975 char *save_parse = RExC_parse;
12976 bool first_time = TRUE; /* First multi-char occurrence doesn't get
12981 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
12982 because too confusing */
12984 sv_catpv(substitute_parse, "(?:");
12988 /* Look at the longest folds first */
12989 for (cp_count = av_len(multi_char_matches); cp_count > 0; cp_count--) {
12991 if (av_exists(multi_char_matches, cp_count)) {
12992 AV** this_array_ptr;
12995 this_array_ptr = (AV**) av_fetch(multi_char_matches,
12997 while ((this_sequence = av_pop(*this_array_ptr)) !=
13000 if (! first_time) {
13001 sv_catpv(substitute_parse, "|");
13003 first_time = FALSE;
13005 sv_catpv(substitute_parse, SvPVX(this_sequence));
13010 /* If the character class contains anything else besides these
13011 * multi-character folds, have to include it in recursive parsing */
13012 if (element_count) {
13013 sv_catpv(substitute_parse, "|[");
13014 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
13015 sv_catpv(substitute_parse, "]");
13018 sv_catpv(substitute_parse, ")");
13021 /* This is a way to get the parse to skip forward a whole named
13022 * sequence instead of matching the 2nd character when it fails the
13024 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
13028 RExC_parse = SvPV(substitute_parse, len);
13029 RExC_end = RExC_parse + len;
13030 RExC_in_multi_char_class = 1;
13031 RExC_emit = (regnode *)orig_emit;
13033 ret = reg(pRExC_state, 1, ®_flags, depth+1);
13035 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
13037 RExC_parse = save_parse;
13038 RExC_end = save_end;
13039 RExC_in_multi_char_class = 0;
13040 SvREFCNT_dec_NN(multi_char_matches);
13044 /* If the character class contains only a single element, it may be
13045 * optimizable into another node type which is smaller and runs faster.
13046 * Check if this is the case for this class */
13047 if (element_count == 1 && ! ret_invlist) {
13051 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like \w or
13052 [:digit:] or \p{foo} */
13054 /* All named classes are mapped into POSIXish nodes, with its FLAG
13055 * argument giving which class it is */
13056 switch ((I32)namedclass) {
13057 case ANYOF_UNIPROP:
13060 /* These don't depend on the charset modifiers. They always
13061 * match under /u rules */
13062 case ANYOF_NHORIZWS:
13063 case ANYOF_HORIZWS:
13064 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
13067 case ANYOF_NVERTWS:
13072 /* The actual POSIXish node for all the rest depends on the
13073 * charset modifier. The ones in the first set depend only on
13074 * ASCII or, if available on this platform, locale */
13078 op = (LOC) ? POSIXL : POSIXA;
13089 /* under /a could be alpha */
13091 if (ASCII_RESTRICTED) {
13092 namedclass = ANYOF_ALPHA + (namedclass % 2);
13100 /* The rest have more possibilities depending on the charset.
13101 * We take advantage of the enum ordering of the charset
13102 * modifiers to get the exact node type, */
13104 op = POSIXD + get_regex_charset(RExC_flags);
13105 if (op > POSIXA) { /* /aa is same as /a */
13108 #ifndef HAS_ISBLANK
13110 && (namedclass == ANYOF_BLANK
13111 || namedclass == ANYOF_NBLANK))
13118 /* The odd numbered ones are the complements of the
13119 * next-lower even number one */
13120 if (namedclass % 2 == 1) {
13124 arg = namedclass_to_classnum(namedclass);
13128 else if (value == prevvalue) {
13130 /* Here, the class consists of just a single code point */
13133 if (! LOC && value == '\n') {
13134 op = REG_ANY; /* Optimize [^\n] */
13135 *flagp |= HASWIDTH|SIMPLE;
13139 else if (value < 256 || UTF) {
13141 /* Optimize a single value into an EXACTish node, but not if it
13142 * would require converting the pattern to UTF-8. */
13143 op = compute_EXACTish(pRExC_state);
13145 } /* Otherwise is a range */
13146 else if (! LOC) { /* locale could vary these */
13147 if (prevvalue == '0') {
13148 if (value == '9') {
13155 /* Here, we have changed <op> away from its initial value iff we found
13156 * an optimization */
13159 /* Throw away this ANYOF regnode, and emit the calculated one,
13160 * which should correspond to the beginning, not current, state of
13162 const char * cur_parse = RExC_parse;
13163 RExC_parse = (char *)orig_parse;
13167 /* To get locale nodes to not use the full ANYOF size would
13168 * require moving the code above that writes the portions
13169 * of it that aren't in other nodes to after this point.
13170 * e.g. ANYOF_CLASS_SET */
13171 RExC_size = orig_size;
13175 RExC_emit = (regnode *)orig_emit;
13176 if (PL_regkind[op] == POSIXD) {
13178 op += NPOSIXD - POSIXD;
13183 ret = reg_node(pRExC_state, op);
13185 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
13189 *flagp |= HASWIDTH|SIMPLE;
13191 else if (PL_regkind[op] == EXACT) {
13192 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13195 RExC_parse = (char *) cur_parse;
13197 SvREFCNT_dec(posixes);
13198 SvREFCNT_dec(cp_list);
13205 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
13207 /* If folding, we calculate all characters that could fold to or from the
13208 * ones already on the list */
13209 if (FOLD && cp_list) {
13210 UV start, end; /* End points of code point ranges */
13212 SV* fold_intersection = NULL;
13214 /* If the highest code point is within Latin1, we can use the
13215 * compiled-in Alphas list, and not have to go out to disk. This
13216 * yields two false positives, the masculine and feminine ordinal
13217 * indicators, which are weeded out below using the
13218 * IS_IN_SOME_FOLD_L1() macro */
13219 if (invlist_highest(cp_list) < 256) {
13220 _invlist_intersection(PL_L1Posix_ptrs[_CC_ALPHA], cp_list,
13221 &fold_intersection);
13225 /* Here, there are non-Latin1 code points, so we will have to go
13226 * fetch the list of all the characters that participate in folds
13228 if (! PL_utf8_foldable) {
13229 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13230 &PL_sv_undef, 1, 0);
13231 PL_utf8_foldable = _get_swash_invlist(swash);
13232 SvREFCNT_dec_NN(swash);
13235 /* This is a hash that for a particular fold gives all characters
13236 * that are involved in it */
13237 if (! PL_utf8_foldclosures) {
13239 /* If we were unable to find any folds, then we likely won't be
13240 * able to find the closures. So just create an empty list.
13241 * Folding will effectively be restricted to the non-Unicode
13242 * rules hard-coded into Perl. (This case happens legitimately
13243 * during compilation of Perl itself before the Unicode tables
13244 * are generated) */
13245 if (_invlist_len(PL_utf8_foldable) == 0) {
13246 PL_utf8_foldclosures = newHV();
13249 /* If the folds haven't been read in, call a fold function
13251 if (! PL_utf8_tofold) {
13252 U8 dummy[UTF8_MAXBYTES+1];
13254 /* This string is just a short named one above \xff */
13255 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
13256 assert(PL_utf8_tofold); /* Verify that worked */
13258 PL_utf8_foldclosures =
13259 _swash_inversion_hash(PL_utf8_tofold);
13263 /* Only the characters in this class that participate in folds need
13264 * be checked. Get the intersection of this class and all the
13265 * possible characters that are foldable. This can quickly narrow
13266 * down a large class */
13267 _invlist_intersection(PL_utf8_foldable, cp_list,
13268 &fold_intersection);
13271 /* Now look at the foldable characters in this class individually */
13272 invlist_iterinit(fold_intersection);
13273 while (invlist_iternext(fold_intersection, &start, &end)) {
13276 /* Locale folding for Latin1 characters is deferred until runtime */
13277 if (LOC && start < 256) {
13281 /* Look at every character in the range */
13282 for (j = start; j <= end; j++) {
13284 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
13290 /* We have the latin1 folding rules hard-coded here so that
13291 * an innocent-looking character class, like /[ks]/i won't
13292 * have to go out to disk to find the possible matches.
13293 * XXX It would be better to generate these via regen, in
13294 * case a new version of the Unicode standard adds new
13295 * mappings, though that is not really likely, and may be
13296 * caught by the default: case of the switch below. */
13298 if (IS_IN_SOME_FOLD_L1(j)) {
13300 /* ASCII is always matched; non-ASCII is matched only
13301 * under Unicode rules */
13302 if (isASCII(j) || AT_LEAST_UNI_SEMANTICS) {
13304 add_cp_to_invlist(cp_list, PL_fold_latin1[j]);
13308 add_cp_to_invlist(depends_list, PL_fold_latin1[j]);
13312 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
13313 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
13315 /* Certain Latin1 characters have matches outside
13316 * Latin1. To get here, <j> is one of those
13317 * characters. None of these matches is valid for
13318 * ASCII characters under /aa, which is why the 'if'
13319 * just above excludes those. These matches only
13320 * happen when the target string is utf8. The code
13321 * below adds the single fold closures for <j> to the
13322 * inversion list. */
13327 add_cp_to_invlist(cp_list, KELVIN_SIGN);
13331 cp_list = add_cp_to_invlist(cp_list,
13332 LATIN_SMALL_LETTER_LONG_S);
13335 cp_list = add_cp_to_invlist(cp_list,
13336 GREEK_CAPITAL_LETTER_MU);
13337 cp_list = add_cp_to_invlist(cp_list,
13338 GREEK_SMALL_LETTER_MU);
13340 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
13341 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
13343 add_cp_to_invlist(cp_list, ANGSTROM_SIGN);
13345 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
13346 cp_list = add_cp_to_invlist(cp_list,
13347 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
13349 case LATIN_SMALL_LETTER_SHARP_S:
13350 cp_list = add_cp_to_invlist(cp_list,
13351 LATIN_CAPITAL_LETTER_SHARP_S);
13353 case 'F': case 'f':
13354 case 'I': case 'i':
13355 case 'L': case 'l':
13356 case 'T': case 't':
13357 case 'A': case 'a':
13358 case 'H': case 'h':
13359 case 'J': case 'j':
13360 case 'N': case 'n':
13361 case 'W': case 'w':
13362 case 'Y': case 'y':
13363 /* These all are targets of multi-character
13364 * folds from code points that require UTF8 to
13365 * express, so they can't match unless the
13366 * target string is in UTF-8, so no action here
13367 * is necessary, as regexec.c properly handles
13368 * the general case for UTF-8 matching and
13369 * multi-char folds */
13372 /* Use deprecated warning to increase the
13373 * chances of this being output */
13374 ckWARN2regdep(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
13381 /* Here is an above Latin1 character. We don't have the rules
13382 * hard-coded for it. First, get its fold. This is the simple
13383 * fold, as the multi-character folds have been handled earlier
13384 * and separated out */
13385 _to_uni_fold_flags(j, foldbuf, &foldlen,
13387 ? FOLD_FLAGS_LOCALE
13388 : (ASCII_FOLD_RESTRICTED)
13389 ? FOLD_FLAGS_NOMIX_ASCII
13392 /* Single character fold of above Latin1. Add everything in
13393 * its fold closure to the list that this node should match.
13394 * The fold closures data structure is a hash with the keys
13395 * being the UTF-8 of every character that is folded to, like
13396 * 'k', and the values each an array of all code points that
13397 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
13398 * Multi-character folds are not included */
13399 if ((listp = hv_fetch(PL_utf8_foldclosures,
13400 (char *) foldbuf, foldlen, FALSE)))
13402 AV* list = (AV*) *listp;
13404 for (k = 0; k <= av_len(list); k++) {
13405 SV** c_p = av_fetch(list, k, FALSE);
13408 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
13412 /* /aa doesn't allow folds between ASCII and non-; /l
13413 * doesn't allow them between above and below 256 */
13414 if ((ASCII_FOLD_RESTRICTED
13415 && (isASCII(c) != isASCII(j)))
13416 || (LOC && ((c < 256) != (j < 256))))
13421 /* Folds involving non-ascii Latin1 characters
13422 * under /d are added to a separate list */
13423 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
13425 cp_list = add_cp_to_invlist(cp_list, c);
13428 depends_list = add_cp_to_invlist(depends_list, c);
13434 SvREFCNT_dec_NN(fold_intersection);
13437 /* And combine the result (if any) with any inversion list from posix
13438 * classes. The lists are kept separate up to now because we don't want to
13439 * fold the classes (folding of those is automatically handled by the swash
13440 * fetching code) */
13442 if (! DEPENDS_SEMANTICS) {
13444 _invlist_union(cp_list, posixes, &cp_list);
13445 SvREFCNT_dec_NN(posixes);
13452 /* Under /d, we put into a separate list the Latin1 things that
13453 * match only when the target string is utf8 */
13454 SV* nonascii_but_latin1_properties = NULL;
13455 _invlist_intersection(posixes, PL_Latin1,
13456 &nonascii_but_latin1_properties);
13457 _invlist_subtract(nonascii_but_latin1_properties, PL_ASCII,
13458 &nonascii_but_latin1_properties);
13459 _invlist_subtract(posixes, nonascii_but_latin1_properties,
13462 _invlist_union(cp_list, posixes, &cp_list);
13463 SvREFCNT_dec_NN(posixes);
13469 if (depends_list) {
13470 _invlist_union(depends_list, nonascii_but_latin1_properties,
13472 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
13475 depends_list = nonascii_but_latin1_properties;
13480 /* And combine the result (if any) with any inversion list from properties.
13481 * The lists are kept separate up to now so that we can distinguish the two
13482 * in regards to matching above-Unicode. A run-time warning is generated
13483 * if a Unicode property is matched against a non-Unicode code point. But,
13484 * we allow user-defined properties to match anything, without any warning,
13485 * and we also suppress the warning if there is a portion of the character
13486 * class that isn't a Unicode property, and which matches above Unicode, \W
13487 * or [\x{110000}] for example.
13488 * (Note that in this case, unlike the Posix one above, there is no
13489 * <depends_list>, because having a Unicode property forces Unicode
13492 bool warn_super = ! has_user_defined_property;
13495 /* If it matters to the final outcome, see if a non-property
13496 * component of the class matches above Unicode. If so, the
13497 * warning gets suppressed. This is true even if just a single
13498 * such code point is specified, as though not strictly correct if
13499 * another such code point is matched against, the fact that they
13500 * are using above-Unicode code points indicates they should know
13501 * the issues involved */
13503 bool non_prop_matches_above_Unicode =
13504 runtime_posix_matches_above_Unicode
13505 | (invlist_highest(cp_list) > PERL_UNICODE_MAX);
13507 non_prop_matches_above_Unicode =
13508 ! non_prop_matches_above_Unicode;
13510 warn_super = ! non_prop_matches_above_Unicode;
13513 _invlist_union(properties, cp_list, &cp_list);
13514 SvREFCNT_dec_NN(properties);
13517 cp_list = properties;
13521 OP(ret) = ANYOF_WARN_SUPER;
13525 /* Here, we have calculated what code points should be in the character
13528 * Now we can see about various optimizations. Fold calculation (which we
13529 * did above) needs to take place before inversion. Otherwise /[^k]/i
13530 * would invert to include K, which under /i would match k, which it
13531 * shouldn't. Therefore we can't invert folded locale now, as it won't be
13532 * folded until runtime */
13534 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
13535 * at compile time. Besides not inverting folded locale now, we can't
13536 * invert if there are things such as \w, which aren't known until runtime
13539 && ! (LOC && (FOLD || (ANYOF_FLAGS(ret) & ANYOF_CLASS)))
13541 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13543 _invlist_invert(cp_list);
13545 /* Any swash can't be used as-is, because we've inverted things */
13547 SvREFCNT_dec_NN(swash);
13551 /* Clear the invert flag since have just done it here */
13556 *ret_invlist = cp_list;
13558 /* Discard the generated node */
13560 RExC_size = orig_size;
13563 RExC_emit = orig_emit;
13568 /* If we didn't do folding, it's because some information isn't available
13569 * until runtime; set the run-time fold flag for these. (We don't have to
13570 * worry about properties folding, as that is taken care of by the swash
13574 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
13577 /* Some character classes are equivalent to other nodes. Such nodes take
13578 * up less room and generally fewer operations to execute than ANYOF nodes.
13579 * Above, we checked for and optimized into some such equivalents for
13580 * certain common classes that are easy to test. Getting to this point in
13581 * the code means that the class didn't get optimized there. Since this
13582 * code is only executed in Pass 2, it is too late to save space--it has
13583 * been allocated in Pass 1, and currently isn't given back. But turning
13584 * things into an EXACTish node can allow the optimizer to join it to any
13585 * adjacent such nodes. And if the class is equivalent to things like /./,
13586 * expensive run-time swashes can be avoided. Now that we have more
13587 * complete information, we can find things necessarily missed by the
13588 * earlier code. I (khw) am not sure how much to look for here. It would
13589 * be easy, but perhaps too slow, to check any candidates against all the
13590 * node types they could possibly match using _invlistEQ(). */
13595 && ! (ANYOF_FLAGS(ret) & ANYOF_CLASS)
13596 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13599 U8 op = END; /* The optimzation node-type */
13600 const char * cur_parse= RExC_parse;
13602 invlist_iterinit(cp_list);
13603 if (! invlist_iternext(cp_list, &start, &end)) {
13605 /* Here, the list is empty. This happens, for example, when a
13606 * Unicode property is the only thing in the character class, and
13607 * it doesn't match anything. (perluniprops.pod notes such
13610 *flagp |= HASWIDTH|SIMPLE;
13612 else if (start == end) { /* The range is a single code point */
13613 if (! invlist_iternext(cp_list, &start, &end)
13615 /* Don't do this optimization if it would require changing
13616 * the pattern to UTF-8 */
13617 && (start < 256 || UTF))
13619 /* Here, the list contains a single code point. Can optimize
13620 * into an EXACT node */
13629 /* A locale node under folding with one code point can be
13630 * an EXACTFL, as its fold won't be calculated until
13636 /* Here, we are generally folding, but there is only one
13637 * code point to match. If we have to, we use an EXACT
13638 * node, but it would be better for joining with adjacent
13639 * nodes in the optimization pass if we used the same
13640 * EXACTFish node that any such are likely to be. We can
13641 * do this iff the code point doesn't participate in any
13642 * folds. For example, an EXACTF of a colon is the same as
13643 * an EXACT one, since nothing folds to or from a colon. */
13645 if (IS_IN_SOME_FOLD_L1(value)) {
13650 if (! PL_utf8_foldable) {
13651 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13652 &PL_sv_undef, 1, 0);
13653 PL_utf8_foldable = _get_swash_invlist(swash);
13654 SvREFCNT_dec_NN(swash);
13656 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
13661 /* If we haven't found the node type, above, it means we
13662 * can use the prevailing one */
13664 op = compute_EXACTish(pRExC_state);
13669 else if (start == 0) {
13670 if (end == UV_MAX) {
13672 *flagp |= HASWIDTH|SIMPLE;
13675 else if (end == '\n' - 1
13676 && invlist_iternext(cp_list, &start, &end)
13677 && start == '\n' + 1 && end == UV_MAX)
13680 *flagp |= HASWIDTH|SIMPLE;
13684 invlist_iterfinish(cp_list);
13687 RExC_parse = (char *)orig_parse;
13688 RExC_emit = (regnode *)orig_emit;
13690 ret = reg_node(pRExC_state, op);
13692 RExC_parse = (char *)cur_parse;
13694 if (PL_regkind[op] == EXACT) {
13695 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13698 SvREFCNT_dec_NN(cp_list);
13703 /* Here, <cp_list> contains all the code points we can determine at
13704 * compile time that match under all conditions. Go through it, and
13705 * for things that belong in the bitmap, put them there, and delete from
13706 * <cp_list>. While we are at it, see if everything above 255 is in the
13707 * list, and if so, set a flag to speed up execution */
13708 ANYOF_BITMAP_ZERO(ret);
13711 /* This gets set if we actually need to modify things */
13712 bool change_invlist = FALSE;
13716 /* Start looking through <cp_list> */
13717 invlist_iterinit(cp_list);
13718 while (invlist_iternext(cp_list, &start, &end)) {
13722 if (end == UV_MAX && start <= 256) {
13723 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
13726 /* Quit if are above what we should change */
13731 change_invlist = TRUE;
13733 /* Set all the bits in the range, up to the max that we are doing */
13734 high = (end < 255) ? end : 255;
13735 for (i = start; i <= (int) high; i++) {
13736 if (! ANYOF_BITMAP_TEST(ret, i)) {
13737 ANYOF_BITMAP_SET(ret, i);
13743 invlist_iterfinish(cp_list);
13745 /* Done with loop; remove any code points that are in the bitmap from
13747 if (change_invlist) {
13748 _invlist_subtract(cp_list, PL_Latin1, &cp_list);
13751 /* If have completely emptied it, remove it completely */
13752 if (_invlist_len(cp_list) == 0) {
13753 SvREFCNT_dec_NN(cp_list);
13759 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
13762 /* Here, the bitmap has been populated with all the Latin1 code points that
13763 * always match. Can now add to the overall list those that match only
13764 * when the target string is UTF-8 (<depends_list>). */
13765 if (depends_list) {
13767 _invlist_union(cp_list, depends_list, &cp_list);
13768 SvREFCNT_dec_NN(depends_list);
13771 cp_list = depends_list;
13775 /* If there is a swash and more than one element, we can't use the swash in
13776 * the optimization below. */
13777 if (swash && element_count > 1) {
13778 SvREFCNT_dec_NN(swash);
13783 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13785 ARG_SET(ret, ANYOF_NONBITMAP_EMPTY);
13788 /* av[0] stores the character class description in its textual form:
13789 * used later (regexec.c:Perl_regclass_swash()) to initialize the
13790 * appropriate swash, and is also useful for dumping the regnode.
13791 * av[1] if NULL, is a placeholder to later contain the swash computed
13792 * from av[0]. But if no further computation need be done, the
13793 * swash is stored there now.
13794 * av[2] stores the cp_list inversion list for use in addition or
13795 * instead of av[0]; used only if av[1] is NULL
13796 * av[3] is set if any component of the class is from a user-defined
13797 * property; used only if av[1] is NULL */
13798 AV * const av = newAV();
13801 av_store(av, 0, (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13802 ? SvREFCNT_inc(listsv) : &PL_sv_undef);
13804 av_store(av, 1, swash);
13805 SvREFCNT_dec_NN(cp_list);
13808 av_store(av, 1, NULL);
13810 av_store(av, 2, cp_list);
13811 av_store(av, 3, newSVuv(has_user_defined_property));
13815 rv = newRV_noinc(MUTABLE_SV(av));
13816 n = add_data(pRExC_state, 1, "s");
13817 RExC_rxi->data->data[n] = (void*)rv;
13821 *flagp |= HASWIDTH|SIMPLE;
13824 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
13827 /* reg_skipcomment()
13829 Absorbs an /x style # comments from the input stream.
13830 Returns true if there is more text remaining in the stream.
13831 Will set the REG_SEEN_RUN_ON_COMMENT flag if the comment
13832 terminates the pattern without including a newline.
13834 Note its the callers responsibility to ensure that we are
13835 actually in /x mode
13840 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
13844 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
13846 while (RExC_parse < RExC_end)
13847 if (*RExC_parse++ == '\n') {
13852 /* we ran off the end of the pattern without ending
13853 the comment, so we have to add an \n when wrapping */
13854 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
13862 Advances the parse position, and optionally absorbs
13863 "whitespace" from the inputstream.
13865 Without /x "whitespace" means (?#...) style comments only,
13866 with /x this means (?#...) and # comments and whitespace proper.
13868 Returns the RExC_parse point from BEFORE the scan occurs.
13870 This is the /x friendly way of saying RExC_parse++.
13874 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
13876 char* const retval = RExC_parse++;
13878 PERL_ARGS_ASSERT_NEXTCHAR;
13881 if (RExC_end - RExC_parse >= 3
13882 && *RExC_parse == '('
13883 && RExC_parse[1] == '?'
13884 && RExC_parse[2] == '#')
13886 while (*RExC_parse != ')') {
13887 if (RExC_parse == RExC_end)
13888 FAIL("Sequence (?#... not terminated");
13894 if (RExC_flags & RXf_PMf_EXTENDED) {
13895 if (isSPACE(*RExC_parse)) {
13899 else if (*RExC_parse == '#') {
13900 if ( reg_skipcomment( pRExC_state ) )
13909 - reg_node - emit a node
13911 STATIC regnode * /* Location. */
13912 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
13916 regnode * const ret = RExC_emit;
13917 GET_RE_DEBUG_FLAGS_DECL;
13919 PERL_ARGS_ASSERT_REG_NODE;
13922 SIZE_ALIGN(RExC_size);
13926 if (RExC_emit >= RExC_emit_bound)
13927 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
13928 op, RExC_emit, RExC_emit_bound);
13930 NODE_ALIGN_FILL(ret);
13932 FILL_ADVANCE_NODE(ptr, op);
13933 #ifdef RE_TRACK_PATTERN_OFFSETS
13934 if (RExC_offsets) { /* MJD */
13935 MJD_OFFSET_DEBUG(("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
13936 "reg_node", __LINE__,
13938 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
13939 ? "Overwriting end of array!\n" : "OK",
13940 (UV)(RExC_emit - RExC_emit_start),
13941 (UV)(RExC_parse - RExC_start),
13942 (UV)RExC_offsets[0]));
13943 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
13951 - reganode - emit a node with an argument
13953 STATIC regnode * /* Location. */
13954 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
13958 regnode * const ret = RExC_emit;
13959 GET_RE_DEBUG_FLAGS_DECL;
13961 PERL_ARGS_ASSERT_REGANODE;
13964 SIZE_ALIGN(RExC_size);
13969 assert(2==regarglen[op]+1);
13971 Anything larger than this has to allocate the extra amount.
13972 If we changed this to be:
13974 RExC_size += (1 + regarglen[op]);
13976 then it wouldn't matter. Its not clear what side effect
13977 might come from that so its not done so far.
13982 if (RExC_emit >= RExC_emit_bound)
13983 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
13984 op, RExC_emit, RExC_emit_bound);
13986 NODE_ALIGN_FILL(ret);
13988 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
13989 #ifdef RE_TRACK_PATTERN_OFFSETS
13990 if (RExC_offsets) { /* MJD */
13991 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
13995 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
13996 "Overwriting end of array!\n" : "OK",
13997 (UV)(RExC_emit - RExC_emit_start),
13998 (UV)(RExC_parse - RExC_start),
13999 (UV)RExC_offsets[0]));
14000 Set_Cur_Node_Offset;
14008 - reguni - emit (if appropriate) a Unicode character
14011 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
14015 PERL_ARGS_ASSERT_REGUNI;
14017 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
14021 - reginsert - insert an operator in front of already-emitted operand
14023 * Means relocating the operand.
14026 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
14032 const int offset = regarglen[(U8)op];
14033 const int size = NODE_STEP_REGNODE + offset;
14034 GET_RE_DEBUG_FLAGS_DECL;
14036 PERL_ARGS_ASSERT_REGINSERT;
14037 PERL_UNUSED_ARG(depth);
14038 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
14039 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
14048 if (RExC_open_parens) {
14050 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
14051 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
14052 if ( RExC_open_parens[paren] >= opnd ) {
14053 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
14054 RExC_open_parens[paren] += size;
14056 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
14058 if ( RExC_close_parens[paren] >= opnd ) {
14059 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
14060 RExC_close_parens[paren] += size;
14062 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
14067 while (src > opnd) {
14068 StructCopy(--src, --dst, regnode);
14069 #ifdef RE_TRACK_PATTERN_OFFSETS
14070 if (RExC_offsets) { /* MJD 20010112 */
14071 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
14075 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
14076 ? "Overwriting end of array!\n" : "OK",
14077 (UV)(src - RExC_emit_start),
14078 (UV)(dst - RExC_emit_start),
14079 (UV)RExC_offsets[0]));
14080 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
14081 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
14087 place = opnd; /* Op node, where operand used to be. */
14088 #ifdef RE_TRACK_PATTERN_OFFSETS
14089 if (RExC_offsets) { /* MJD */
14090 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14094 (UV)(place - RExC_emit_start) > RExC_offsets[0]
14095 ? "Overwriting end of array!\n" : "OK",
14096 (UV)(place - RExC_emit_start),
14097 (UV)(RExC_parse - RExC_start),
14098 (UV)RExC_offsets[0]));
14099 Set_Node_Offset(place, RExC_parse);
14100 Set_Node_Length(place, 1);
14103 src = NEXTOPER(place);
14104 FILL_ADVANCE_NODE(place, op);
14105 Zero(src, offset, regnode);
14109 - regtail - set the next-pointer at the end of a node chain of p to val.
14110 - SEE ALSO: regtail_study
14112 /* TODO: All three parms should be const */
14114 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14118 GET_RE_DEBUG_FLAGS_DECL;
14120 PERL_ARGS_ASSERT_REGTAIL;
14122 PERL_UNUSED_ARG(depth);
14128 /* Find last node. */
14131 regnode * const temp = regnext(scan);
14133 SV * const mysv=sv_newmortal();
14134 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
14135 regprop(RExC_rx, mysv, scan);
14136 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
14137 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
14138 (temp == NULL ? "->" : ""),
14139 (temp == NULL ? PL_reg_name[OP(val)] : "")
14147 if (reg_off_by_arg[OP(scan)]) {
14148 ARG_SET(scan, val - scan);
14151 NEXT_OFF(scan) = val - scan;
14157 - regtail_study - set the next-pointer at the end of a node chain of p to val.
14158 - Look for optimizable sequences at the same time.
14159 - currently only looks for EXACT chains.
14161 This is experimental code. The idea is to use this routine to perform
14162 in place optimizations on branches and groups as they are constructed,
14163 with the long term intention of removing optimization from study_chunk so
14164 that it is purely analytical.
14166 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
14167 to control which is which.
14170 /* TODO: All four parms should be const */
14173 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14178 #ifdef EXPERIMENTAL_INPLACESCAN
14181 GET_RE_DEBUG_FLAGS_DECL;
14183 PERL_ARGS_ASSERT_REGTAIL_STUDY;
14189 /* Find last node. */
14193 regnode * const temp = regnext(scan);
14194 #ifdef EXPERIMENTAL_INPLACESCAN
14195 if (PL_regkind[OP(scan)] == EXACT) {
14196 bool has_exactf_sharp_s; /* Unexamined in this routine */
14197 if (join_exact(pRExC_state,scan,&min, &has_exactf_sharp_s, 1,val,depth+1))
14202 switch (OP(scan)) {
14208 case EXACTFU_TRICKYFOLD:
14210 if( exact == PSEUDO )
14212 else if ( exact != OP(scan) )
14221 SV * const mysv=sv_newmortal();
14222 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
14223 regprop(RExC_rx, mysv, scan);
14224 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
14225 SvPV_nolen_const(mysv),
14226 REG_NODE_NUM(scan),
14227 PL_reg_name[exact]);
14234 SV * const mysv_val=sv_newmortal();
14235 DEBUG_PARSE_MSG("");
14236 regprop(RExC_rx, mysv_val, val);
14237 PerlIO_printf(Perl_debug_log, "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
14238 SvPV_nolen_const(mysv_val),
14239 (IV)REG_NODE_NUM(val),
14243 if (reg_off_by_arg[OP(scan)]) {
14244 ARG_SET(scan, val - scan);
14247 NEXT_OFF(scan) = val - scan;
14255 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
14259 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
14265 for (bit=0; bit<32; bit++) {
14266 if (flags & (1<<bit)) {
14267 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
14270 if (!set++ && lead)
14271 PerlIO_printf(Perl_debug_log, "%s",lead);
14272 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
14275 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
14276 if (!set++ && lead) {
14277 PerlIO_printf(Perl_debug_log, "%s",lead);
14280 case REGEX_UNICODE_CHARSET:
14281 PerlIO_printf(Perl_debug_log, "UNICODE");
14283 case REGEX_LOCALE_CHARSET:
14284 PerlIO_printf(Perl_debug_log, "LOCALE");
14286 case REGEX_ASCII_RESTRICTED_CHARSET:
14287 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
14289 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
14290 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
14293 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
14299 PerlIO_printf(Perl_debug_log, "\n");
14301 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14307 Perl_regdump(pTHX_ const regexp *r)
14311 SV * const sv = sv_newmortal();
14312 SV *dsv= sv_newmortal();
14313 RXi_GET_DECL(r,ri);
14314 GET_RE_DEBUG_FLAGS_DECL;
14316 PERL_ARGS_ASSERT_REGDUMP;
14318 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
14320 /* Header fields of interest. */
14321 if (r->anchored_substr) {
14322 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
14323 RE_SV_DUMPLEN(r->anchored_substr), 30);
14324 PerlIO_printf(Perl_debug_log,
14325 "anchored %s%s at %"IVdf" ",
14326 s, RE_SV_TAIL(r->anchored_substr),
14327 (IV)r->anchored_offset);
14328 } else if (r->anchored_utf8) {
14329 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
14330 RE_SV_DUMPLEN(r->anchored_utf8), 30);
14331 PerlIO_printf(Perl_debug_log,
14332 "anchored utf8 %s%s at %"IVdf" ",
14333 s, RE_SV_TAIL(r->anchored_utf8),
14334 (IV)r->anchored_offset);
14336 if (r->float_substr) {
14337 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
14338 RE_SV_DUMPLEN(r->float_substr), 30);
14339 PerlIO_printf(Perl_debug_log,
14340 "floating %s%s at %"IVdf"..%"UVuf" ",
14341 s, RE_SV_TAIL(r->float_substr),
14342 (IV)r->float_min_offset, (UV)r->float_max_offset);
14343 } else if (r->float_utf8) {
14344 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
14345 RE_SV_DUMPLEN(r->float_utf8), 30);
14346 PerlIO_printf(Perl_debug_log,
14347 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
14348 s, RE_SV_TAIL(r->float_utf8),
14349 (IV)r->float_min_offset, (UV)r->float_max_offset);
14351 if (r->check_substr || r->check_utf8)
14352 PerlIO_printf(Perl_debug_log,
14354 (r->check_substr == r->float_substr
14355 && r->check_utf8 == r->float_utf8
14356 ? "(checking floating" : "(checking anchored"));
14357 if (r->extflags & RXf_NOSCAN)
14358 PerlIO_printf(Perl_debug_log, " noscan");
14359 if (r->extflags & RXf_CHECK_ALL)
14360 PerlIO_printf(Perl_debug_log, " isall");
14361 if (r->check_substr || r->check_utf8)
14362 PerlIO_printf(Perl_debug_log, ") ");
14364 if (ri->regstclass) {
14365 regprop(r, sv, ri->regstclass);
14366 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
14368 if (r->extflags & RXf_ANCH) {
14369 PerlIO_printf(Perl_debug_log, "anchored");
14370 if (r->extflags & RXf_ANCH_BOL)
14371 PerlIO_printf(Perl_debug_log, "(BOL)");
14372 if (r->extflags & RXf_ANCH_MBOL)
14373 PerlIO_printf(Perl_debug_log, "(MBOL)");
14374 if (r->extflags & RXf_ANCH_SBOL)
14375 PerlIO_printf(Perl_debug_log, "(SBOL)");
14376 if (r->extflags & RXf_ANCH_GPOS)
14377 PerlIO_printf(Perl_debug_log, "(GPOS)");
14378 PerlIO_putc(Perl_debug_log, ' ');
14380 if (r->extflags & RXf_GPOS_SEEN)
14381 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
14382 if (r->intflags & PREGf_SKIP)
14383 PerlIO_printf(Perl_debug_log, "plus ");
14384 if (r->intflags & PREGf_IMPLICIT)
14385 PerlIO_printf(Perl_debug_log, "implicit ");
14386 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
14387 if (r->extflags & RXf_EVAL_SEEN)
14388 PerlIO_printf(Perl_debug_log, "with eval ");
14389 PerlIO_printf(Perl_debug_log, "\n");
14390 DEBUG_FLAGS_r(regdump_extflags("r->extflags: ",r->extflags));
14392 PERL_ARGS_ASSERT_REGDUMP;
14393 PERL_UNUSED_CONTEXT;
14394 PERL_UNUSED_ARG(r);
14395 #endif /* DEBUGGING */
14399 - regprop - printable representation of opcode
14401 #define EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags) \
14404 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]); \
14405 if (flags & ANYOF_INVERT) \
14406 /*make sure the invert info is in each */ \
14407 sv_catpvs(sv, "^"); \
14413 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
14419 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
14420 static const char * const anyofs[] = {
14421 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
14422 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
14423 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
14424 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
14425 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
14426 || _CC_VERTSPACE != 16
14427 #error Need to adjust order of anyofs[]
14464 RXi_GET_DECL(prog,progi);
14465 GET_RE_DEBUG_FLAGS_DECL;
14467 PERL_ARGS_ASSERT_REGPROP;
14471 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
14472 /* It would be nice to FAIL() here, but this may be called from
14473 regexec.c, and it would be hard to supply pRExC_state. */
14474 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(o), (int)REGNODE_MAX);
14475 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
14477 k = PL_regkind[OP(o)];
14480 sv_catpvs(sv, " ");
14481 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
14482 * is a crude hack but it may be the best for now since
14483 * we have no flag "this EXACTish node was UTF-8"
14485 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
14486 PERL_PV_ESCAPE_UNI_DETECT |
14487 PERL_PV_ESCAPE_NONASCII |
14488 PERL_PV_PRETTY_ELLIPSES |
14489 PERL_PV_PRETTY_LTGT |
14490 PERL_PV_PRETTY_NOCLEAR
14492 } else if (k == TRIE) {
14493 /* print the details of the trie in dumpuntil instead, as
14494 * progi->data isn't available here */
14495 const char op = OP(o);
14496 const U32 n = ARG(o);
14497 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
14498 (reg_ac_data *)progi->data->data[n] :
14500 const reg_trie_data * const trie
14501 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
14503 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
14504 DEBUG_TRIE_COMPILE_r(
14505 Perl_sv_catpvf(aTHX_ sv,
14506 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
14507 (UV)trie->startstate,
14508 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
14509 (UV)trie->wordcount,
14512 (UV)TRIE_CHARCOUNT(trie),
14513 (UV)trie->uniquecharcount
14516 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
14518 int rangestart = -1;
14519 U8* bitmap = IS_ANYOF_TRIE(op) ? (U8*)ANYOF_BITMAP(o) : (U8*)TRIE_BITMAP(trie);
14520 sv_catpvs(sv, "[");
14521 for (i = 0; i <= 256; i++) {
14522 if (i < 256 && BITMAP_TEST(bitmap,i)) {
14523 if (rangestart == -1)
14525 } else if (rangestart != -1) {
14526 if (i <= rangestart + 3)
14527 for (; rangestart < i; rangestart++)
14528 put_byte(sv, rangestart);
14530 put_byte(sv, rangestart);
14531 sv_catpvs(sv, "-");
14532 put_byte(sv, i - 1);
14537 sv_catpvs(sv, "]");
14540 } else if (k == CURLY) {
14541 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
14542 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
14543 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
14545 else if (k == WHILEM && o->flags) /* Ordinal/of */
14546 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
14547 else if (k == REF || k == OPEN || k == CLOSE || k == GROUPP || OP(o)==ACCEPT) {
14548 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
14549 if ( RXp_PAREN_NAMES(prog) ) {
14550 if ( k != REF || (OP(o) < NREF)) {
14551 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
14552 SV **name= av_fetch(list, ARG(o), 0 );
14554 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14557 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
14558 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
14559 I32 *nums=(I32*)SvPVX(sv_dat);
14560 SV **name= av_fetch(list, nums[0], 0 );
14563 for ( n=0; n<SvIVX(sv_dat); n++ ) {
14564 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
14565 (n ? "," : ""), (IV)nums[n]);
14567 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14571 } else if (k == GOSUB)
14572 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o)); /* Paren and offset */
14573 else if (k == VERB) {
14575 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
14576 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
14577 } else if (k == LOGICAL)
14578 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* 2: embedded, otherwise 1 */
14579 else if (k == ANYOF) {
14580 int i, rangestart = -1;
14581 const U8 flags = ANYOF_FLAGS(o);
14585 if (flags & ANYOF_LOCALE)
14586 sv_catpvs(sv, "{loc}");
14587 if (flags & ANYOF_LOC_FOLD)
14588 sv_catpvs(sv, "{i}");
14589 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
14590 if (flags & ANYOF_INVERT)
14591 sv_catpvs(sv, "^");
14593 /* output what the standard cp 0-255 bitmap matches */
14594 for (i = 0; i <= 256; i++) {
14595 if (i < 256 && ANYOF_BITMAP_TEST(o,i)) {
14596 if (rangestart == -1)
14598 } else if (rangestart != -1) {
14599 if (i <= rangestart + 3)
14600 for (; rangestart < i; rangestart++)
14601 put_byte(sv, rangestart);
14603 put_byte(sv, rangestart);
14604 sv_catpvs(sv, "-");
14605 put_byte(sv, i - 1);
14612 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14613 /* output any special charclass tests (used entirely under use locale) */
14614 if (ANYOF_CLASS_TEST_ANY_SET(o))
14615 for (i = 0; i < (int)(sizeof(anyofs)/sizeof(char*)); i++)
14616 if (ANYOF_CLASS_TEST(o,i)) {
14617 sv_catpv(sv, anyofs[i]);
14621 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14623 if (flags & ANYOF_NON_UTF8_LATIN1_ALL) {
14624 sv_catpvs(sv, "{non-utf8-latin1-all}");
14627 /* output information about the unicode matching */
14628 if (flags & ANYOF_UNICODE_ALL)
14629 sv_catpvs(sv, "{unicode_all}");
14630 else if (ANYOF_NONBITMAP(o))
14631 sv_catpvs(sv, "{unicode}");
14632 if (flags & ANYOF_NONBITMAP_NON_UTF8)
14633 sv_catpvs(sv, "{outside bitmap}");
14635 if (ANYOF_NONBITMAP(o)) {
14636 SV *lv; /* Set if there is something outside the bit map */
14637 SV * const sw = regclass_swash(prog, o, FALSE, &lv, NULL);
14638 bool byte_output = FALSE; /* If something in the bitmap has been
14641 if (lv && lv != &PL_sv_undef) {
14643 U8 s[UTF8_MAXBYTES_CASE+1];
14645 for (i = 0; i <= 256; i++) { /* Look at chars in bitmap */
14646 uvchr_to_utf8(s, i);
14649 && ! ANYOF_BITMAP_TEST(o, i) /* Don't duplicate
14653 && swash_fetch(sw, s, TRUE))
14655 if (rangestart == -1)
14657 } else if (rangestart != -1) {
14658 byte_output = TRUE;
14659 if (i <= rangestart + 3)
14660 for (; rangestart < i; rangestart++) {
14661 put_byte(sv, rangestart);
14664 put_byte(sv, rangestart);
14665 sv_catpvs(sv, "-");
14674 char *s = savesvpv(lv);
14675 char * const origs = s;
14677 while (*s && *s != '\n')
14681 const char * const t = ++s;
14684 sv_catpvs(sv, " ");
14690 /* Truncate very long output */
14691 if (s - origs > 256) {
14692 Perl_sv_catpvf(aTHX_ sv,
14694 (int) (s - origs - 1),
14700 else if (*s == '\t') {
14715 SvREFCNT_dec_NN(lv);
14719 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
14721 else if (k == POSIXD || k == NPOSIXD) {
14722 U8 index = FLAGS(o) * 2;
14723 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
14724 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
14727 sv_catpv(sv, anyofs[index]);
14730 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
14731 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
14733 PERL_UNUSED_CONTEXT;
14734 PERL_UNUSED_ARG(sv);
14735 PERL_UNUSED_ARG(o);
14736 PERL_UNUSED_ARG(prog);
14737 #endif /* DEBUGGING */
14741 Perl_re_intuit_string(pTHX_ REGEXP * const r)
14742 { /* Assume that RE_INTUIT is set */
14744 struct regexp *const prog = ReANY(r);
14745 GET_RE_DEBUG_FLAGS_DECL;
14747 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
14748 PERL_UNUSED_CONTEXT;
14752 const char * const s = SvPV_nolen_const(prog->check_substr
14753 ? prog->check_substr : prog->check_utf8);
14755 if (!PL_colorset) reginitcolors();
14756 PerlIO_printf(Perl_debug_log,
14757 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
14759 prog->check_substr ? "" : "utf8 ",
14760 PL_colors[5],PL_colors[0],
14763 (strlen(s) > 60 ? "..." : ""));
14766 return prog->check_substr ? prog->check_substr : prog->check_utf8;
14772 handles refcounting and freeing the perl core regexp structure. When
14773 it is necessary to actually free the structure the first thing it
14774 does is call the 'free' method of the regexp_engine associated to
14775 the regexp, allowing the handling of the void *pprivate; member
14776 first. (This routine is not overridable by extensions, which is why
14777 the extensions free is called first.)
14779 See regdupe and regdupe_internal if you change anything here.
14781 #ifndef PERL_IN_XSUB_RE
14783 Perl_pregfree(pTHX_ REGEXP *r)
14789 Perl_pregfree2(pTHX_ REGEXP *rx)
14792 struct regexp *const r = ReANY(rx);
14793 GET_RE_DEBUG_FLAGS_DECL;
14795 PERL_ARGS_ASSERT_PREGFREE2;
14797 if (r->mother_re) {
14798 ReREFCNT_dec(r->mother_re);
14800 CALLREGFREE_PVT(rx); /* free the private data */
14801 SvREFCNT_dec(RXp_PAREN_NAMES(r));
14802 Safefree(r->xpv_len_u.xpvlenu_pv);
14805 SvREFCNT_dec(r->anchored_substr);
14806 SvREFCNT_dec(r->anchored_utf8);
14807 SvREFCNT_dec(r->float_substr);
14808 SvREFCNT_dec(r->float_utf8);
14809 Safefree(r->substrs);
14811 RX_MATCH_COPY_FREE(rx);
14812 #ifdef PERL_ANY_COW
14813 SvREFCNT_dec(r->saved_copy);
14816 SvREFCNT_dec(r->qr_anoncv);
14817 rx->sv_u.svu_rx = 0;
14822 This is a hacky workaround to the structural issue of match results
14823 being stored in the regexp structure which is in turn stored in
14824 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
14825 could be PL_curpm in multiple contexts, and could require multiple
14826 result sets being associated with the pattern simultaneously, such
14827 as when doing a recursive match with (??{$qr})
14829 The solution is to make a lightweight copy of the regexp structure
14830 when a qr// is returned from the code executed by (??{$qr}) this
14831 lightweight copy doesn't actually own any of its data except for
14832 the starp/end and the actual regexp structure itself.
14838 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
14840 struct regexp *ret;
14841 struct regexp *const r = ReANY(rx);
14842 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
14844 PERL_ARGS_ASSERT_REG_TEMP_COPY;
14847 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
14849 SvOK_off((SV *)ret_x);
14851 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
14852 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
14853 made both spots point to the same regexp body.) */
14854 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
14855 assert(!SvPVX(ret_x));
14856 ret_x->sv_u.svu_rx = temp->sv_any;
14857 temp->sv_any = NULL;
14858 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
14859 SvREFCNT_dec_NN(temp);
14860 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
14861 ing below will not set it. */
14862 SvCUR_set(ret_x, SvCUR(rx));
14865 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
14866 sv_force_normal(sv) is called. */
14868 ret = ReANY(ret_x);
14870 SvFLAGS(ret_x) |= SvUTF8(rx);
14871 /* We share the same string buffer as the original regexp, on which we
14872 hold a reference count, incremented when mother_re is set below.
14873 The string pointer is copied here, being part of the regexp struct.
14875 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
14876 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
14878 const I32 npar = r->nparens+1;
14879 Newx(ret->offs, npar, regexp_paren_pair);
14880 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
14883 Newx(ret->substrs, 1, struct reg_substr_data);
14884 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
14886 SvREFCNT_inc_void(ret->anchored_substr);
14887 SvREFCNT_inc_void(ret->anchored_utf8);
14888 SvREFCNT_inc_void(ret->float_substr);
14889 SvREFCNT_inc_void(ret->float_utf8);
14891 /* check_substr and check_utf8, if non-NULL, point to either their
14892 anchored or float namesakes, and don't hold a second reference. */
14894 RX_MATCH_COPIED_off(ret_x);
14895 #ifdef PERL_ANY_COW
14896 ret->saved_copy = NULL;
14898 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
14899 SvREFCNT_inc_void(ret->qr_anoncv);
14905 /* regfree_internal()
14907 Free the private data in a regexp. This is overloadable by
14908 extensions. Perl takes care of the regexp structure in pregfree(),
14909 this covers the *pprivate pointer which technically perl doesn't
14910 know about, however of course we have to handle the
14911 regexp_internal structure when no extension is in use.
14913 Note this is called before freeing anything in the regexp
14918 Perl_regfree_internal(pTHX_ REGEXP * const rx)
14921 struct regexp *const r = ReANY(rx);
14922 RXi_GET_DECL(r,ri);
14923 GET_RE_DEBUG_FLAGS_DECL;
14925 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
14931 SV *dsv= sv_newmortal();
14932 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
14933 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
14934 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
14935 PL_colors[4],PL_colors[5],s);
14938 #ifdef RE_TRACK_PATTERN_OFFSETS
14940 Safefree(ri->u.offsets); /* 20010421 MJD */
14942 if (ri->code_blocks) {
14944 for (n = 0; n < ri->num_code_blocks; n++)
14945 SvREFCNT_dec(ri->code_blocks[n].src_regex);
14946 Safefree(ri->code_blocks);
14950 int n = ri->data->count;
14953 /* If you add a ->what type here, update the comment in regcomp.h */
14954 switch (ri->data->what[n]) {
14960 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
14963 Safefree(ri->data->data[n]);
14969 { /* Aho Corasick add-on structure for a trie node.
14970 Used in stclass optimization only */
14972 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
14974 refcount = --aho->refcount;
14977 PerlMemShared_free(aho->states);
14978 PerlMemShared_free(aho->fail);
14979 /* do this last!!!! */
14980 PerlMemShared_free(ri->data->data[n]);
14981 PerlMemShared_free(ri->regstclass);
14987 /* trie structure. */
14989 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
14991 refcount = --trie->refcount;
14994 PerlMemShared_free(trie->charmap);
14995 PerlMemShared_free(trie->states);
14996 PerlMemShared_free(trie->trans);
14998 PerlMemShared_free(trie->bitmap);
15000 PerlMemShared_free(trie->jump);
15001 PerlMemShared_free(trie->wordinfo);
15002 /* do this last!!!! */
15003 PerlMemShared_free(ri->data->data[n]);
15008 Perl_croak(aTHX_ "panic: regfree data code '%c'", ri->data->what[n]);
15011 Safefree(ri->data->what);
15012 Safefree(ri->data);
15018 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
15019 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
15020 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
15023 re_dup - duplicate a regexp.
15025 This routine is expected to clone a given regexp structure. It is only
15026 compiled under USE_ITHREADS.
15028 After all of the core data stored in struct regexp is duplicated
15029 the regexp_engine.dupe method is used to copy any private data
15030 stored in the *pprivate pointer. This allows extensions to handle
15031 any duplication it needs to do.
15033 See pregfree() and regfree_internal() if you change anything here.
15035 #if defined(USE_ITHREADS)
15036 #ifndef PERL_IN_XSUB_RE
15038 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
15042 const struct regexp *r = ReANY(sstr);
15043 struct regexp *ret = ReANY(dstr);
15045 PERL_ARGS_ASSERT_RE_DUP_GUTS;
15047 npar = r->nparens+1;
15048 Newx(ret->offs, npar, regexp_paren_pair);
15049 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15051 if (ret->substrs) {
15052 /* Do it this way to avoid reading from *r after the StructCopy().
15053 That way, if any of the sv_dup_inc()s dislodge *r from the L1
15054 cache, it doesn't matter. */
15055 const bool anchored = r->check_substr
15056 ? r->check_substr == r->anchored_substr
15057 : r->check_utf8 == r->anchored_utf8;
15058 Newx(ret->substrs, 1, struct reg_substr_data);
15059 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15061 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
15062 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
15063 ret->float_substr = sv_dup_inc(ret->float_substr, param);
15064 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
15066 /* check_substr and check_utf8, if non-NULL, point to either their
15067 anchored or float namesakes, and don't hold a second reference. */
15069 if (ret->check_substr) {
15071 assert(r->check_utf8 == r->anchored_utf8);
15072 ret->check_substr = ret->anchored_substr;
15073 ret->check_utf8 = ret->anchored_utf8;
15075 assert(r->check_substr == r->float_substr);
15076 assert(r->check_utf8 == r->float_utf8);
15077 ret->check_substr = ret->float_substr;
15078 ret->check_utf8 = ret->float_utf8;
15080 } else if (ret->check_utf8) {
15082 ret->check_utf8 = ret->anchored_utf8;
15084 ret->check_utf8 = ret->float_utf8;
15089 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
15090 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
15093 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
15095 if (RX_MATCH_COPIED(dstr))
15096 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
15098 ret->subbeg = NULL;
15099 #ifdef PERL_ANY_COW
15100 ret->saved_copy = NULL;
15103 /* Whether mother_re be set or no, we need to copy the string. We
15104 cannot refrain from copying it when the storage points directly to
15105 our mother regexp, because that's
15106 1: a buffer in a different thread
15107 2: something we no longer hold a reference on
15108 so we need to copy it locally. */
15109 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
15110 ret->mother_re = NULL;
15113 #endif /* PERL_IN_XSUB_RE */
15118 This is the internal complement to regdupe() which is used to copy
15119 the structure pointed to by the *pprivate pointer in the regexp.
15120 This is the core version of the extension overridable cloning hook.
15121 The regexp structure being duplicated will be copied by perl prior
15122 to this and will be provided as the regexp *r argument, however
15123 with the /old/ structures pprivate pointer value. Thus this routine
15124 may override any copying normally done by perl.
15126 It returns a pointer to the new regexp_internal structure.
15130 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
15133 struct regexp *const r = ReANY(rx);
15134 regexp_internal *reti;
15136 RXi_GET_DECL(r,ri);
15138 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
15142 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode), char, regexp_internal);
15143 Copy(ri->program, reti->program, len+1, regnode);
15145 reti->num_code_blocks = ri->num_code_blocks;
15146 if (ri->code_blocks) {
15148 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
15149 struct reg_code_block);
15150 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
15151 struct reg_code_block);
15152 for (n = 0; n < ri->num_code_blocks; n++)
15153 reti->code_blocks[n].src_regex = (REGEXP*)
15154 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
15157 reti->code_blocks = NULL;
15159 reti->regstclass = NULL;
15162 struct reg_data *d;
15163 const int count = ri->data->count;
15166 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
15167 char, struct reg_data);
15168 Newx(d->what, count, U8);
15171 for (i = 0; i < count; i++) {
15172 d->what[i] = ri->data->what[i];
15173 switch (d->what[i]) {
15174 /* see also regcomp.h and regfree_internal() */
15175 case 'a': /* actually an AV, but the dup function is identical. */
15179 case 'u': /* actually an HV, but the dup function is identical. */
15180 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
15183 /* This is cheating. */
15184 Newx(d->data[i], 1, struct regnode_charclass_class);
15185 StructCopy(ri->data->data[i], d->data[i],
15186 struct regnode_charclass_class);
15187 reti->regstclass = (regnode*)d->data[i];
15190 /* Trie stclasses are readonly and can thus be shared
15191 * without duplication. We free the stclass in pregfree
15192 * when the corresponding reg_ac_data struct is freed.
15194 reti->regstclass= ri->regstclass;
15198 ((reg_trie_data*)ri->data->data[i])->refcount++;
15203 d->data[i] = ri->data->data[i];
15206 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'", ri->data->what[i]);
15215 reti->name_list_idx = ri->name_list_idx;
15217 #ifdef RE_TRACK_PATTERN_OFFSETS
15218 if (ri->u.offsets) {
15219 Newx(reti->u.offsets, 2*len+1, U32);
15220 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
15223 SetProgLen(reti,len);
15226 return (void*)reti;
15229 #endif /* USE_ITHREADS */
15231 #ifndef PERL_IN_XSUB_RE
15234 - regnext - dig the "next" pointer out of a node
15237 Perl_regnext(pTHX_ regnode *p)
15245 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
15246 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(p), (int)REGNODE_MAX);
15249 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
15258 S_re_croak2(pTHX_ const char* pat1,const char* pat2,...)
15261 STRLEN l1 = strlen(pat1);
15262 STRLEN l2 = strlen(pat2);
15265 const char *message;
15267 PERL_ARGS_ASSERT_RE_CROAK2;
15273 Copy(pat1, buf, l1 , char);
15274 Copy(pat2, buf + l1, l2 , char);
15275 buf[l1 + l2] = '\n';
15276 buf[l1 + l2 + 1] = '\0';
15278 /* ANSI variant takes additional second argument */
15279 va_start(args, pat2);
15283 msv = vmess(buf, &args);
15285 message = SvPV_const(msv,l1);
15288 Copy(message, buf, l1 , char);
15289 buf[l1-1] = '\0'; /* Overwrite \n */
15290 Perl_croak(aTHX_ "%s", buf);
15293 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
15295 #ifndef PERL_IN_XSUB_RE
15297 Perl_save_re_context(pTHX)
15301 struct re_save_state *state;
15303 SAVEVPTR(PL_curcop);
15304 SSGROW(SAVESTACK_ALLOC_FOR_RE_SAVE_STATE + 1);
15306 state = (struct re_save_state *)(PL_savestack + PL_savestack_ix);
15307 PL_savestack_ix += SAVESTACK_ALLOC_FOR_RE_SAVE_STATE;
15308 SSPUSHUV(SAVEt_RE_STATE);
15310 Copy(&PL_reg_state, state, 1, struct re_save_state);
15312 PL_reg_oldsaved = NULL;
15313 PL_reg_oldsavedlen = 0;
15314 PL_reg_oldsavedoffset = 0;
15315 PL_reg_oldsavedcoffset = 0;
15316 PL_reg_maxiter = 0;
15317 PL_reg_leftiter = 0;
15318 PL_reg_poscache = NULL;
15319 PL_reg_poscache_size = 0;
15320 #ifdef PERL_ANY_COW
15324 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
15326 const REGEXP * const rx = PM_GETRE(PL_curpm);
15329 for (i = 1; i <= RX_NPARENS(rx); i++) {
15330 char digits[TYPE_CHARS(long)];
15331 const STRLEN len = my_snprintf(digits, sizeof(digits), "%lu", (long)i);
15332 GV *const *const gvp
15333 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
15336 GV * const gv = *gvp;
15337 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
15349 S_put_byte(pTHX_ SV *sv, int c)
15351 PERL_ARGS_ASSERT_PUT_BYTE;
15353 /* Our definition of isPRINT() ignores locales, so only bytes that are
15354 not part of UTF-8 are considered printable. I assume that the same
15355 holds for UTF-EBCDIC.
15356 Also, code point 255 is not printable in either (it's E0 in EBCDIC,
15357 which Wikipedia says:
15359 EO, or Eight Ones, is an 8-bit EBCDIC character code represented as all
15360 ones (binary 1111 1111, hexadecimal FF). It is similar, but not
15361 identical, to the ASCII delete (DEL) or rubout control character. ...
15362 it is typically mapped to hexadecimal code 9F, in order to provide a
15363 unique character mapping in both directions)
15365 So the old condition can be simplified to !isPRINT(c) */
15368 Perl_sv_catpvf(aTHX_ sv, "\\x%02x", c);
15371 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
15375 const char string = c;
15376 if (c == '-' || c == ']' || c == '\\' || c == '^')
15377 sv_catpvs(sv, "\\");
15378 sv_catpvn(sv, &string, 1);
15383 #define CLEAR_OPTSTART \
15384 if (optstart) STMT_START { \
15385 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
15389 #define DUMPUNTIL(b,e) CLEAR_OPTSTART; node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
15391 STATIC const regnode *
15392 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
15393 const regnode *last, const regnode *plast,
15394 SV* sv, I32 indent, U32 depth)
15397 U8 op = PSEUDO; /* Arbitrary non-END op. */
15398 const regnode *next;
15399 const regnode *optstart= NULL;
15401 RXi_GET_DECL(r,ri);
15402 GET_RE_DEBUG_FLAGS_DECL;
15404 PERL_ARGS_ASSERT_DUMPUNTIL;
15406 #ifdef DEBUG_DUMPUNTIL
15407 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
15408 last ? last-start : 0,plast ? plast-start : 0);
15411 if (plast && plast < last)
15414 while (PL_regkind[op] != END && (!last || node < last)) {
15415 /* While that wasn't END last time... */
15418 if (op == CLOSE || op == WHILEM)
15420 next = regnext((regnode *)node);
15423 if (OP(node) == OPTIMIZED) {
15424 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
15431 regprop(r, sv, node);
15432 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
15433 (int)(2*indent + 1), "", SvPVX_const(sv));
15435 if (OP(node) != OPTIMIZED) {
15436 if (next == NULL) /* Next ptr. */
15437 PerlIO_printf(Perl_debug_log, " (0)");
15438 else if (PL_regkind[(U8)op] == BRANCH && PL_regkind[OP(next)] != BRANCH )
15439 PerlIO_printf(Perl_debug_log, " (FAIL)");
15441 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
15442 (void)PerlIO_putc(Perl_debug_log, '\n');
15446 if (PL_regkind[(U8)op] == BRANCHJ) {
15449 const regnode *nnode = (OP(next) == LONGJMP
15450 ? regnext((regnode *)next)
15452 if (last && nnode > last)
15454 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
15457 else if (PL_regkind[(U8)op] == BRANCH) {
15459 DUMPUNTIL(NEXTOPER(node), next);
15461 else if ( PL_regkind[(U8)op] == TRIE ) {
15462 const regnode *this_trie = node;
15463 const char op = OP(node);
15464 const U32 n = ARG(node);
15465 const reg_ac_data * const ac = op>=AHOCORASICK ?
15466 (reg_ac_data *)ri->data->data[n] :
15468 const reg_trie_data * const trie =
15469 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
15471 AV *const trie_words = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
15473 const regnode *nextbranch= NULL;
15476 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
15477 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
15479 PerlIO_printf(Perl_debug_log, "%*s%s ",
15480 (int)(2*(indent+3)), "",
15481 elem_ptr ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr), SvCUR(*elem_ptr), 60,
15482 PL_colors[0], PL_colors[1],
15483 (SvUTF8(*elem_ptr) ? PERL_PV_ESCAPE_UNI : 0) |
15484 PERL_PV_PRETTY_ELLIPSES |
15485 PERL_PV_PRETTY_LTGT
15490 U16 dist= trie->jump[word_idx+1];
15491 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
15492 (UV)((dist ? this_trie + dist : next) - start));
15495 nextbranch= this_trie + trie->jump[0];
15496 DUMPUNTIL(this_trie + dist, nextbranch);
15498 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
15499 nextbranch= regnext((regnode *)nextbranch);
15501 PerlIO_printf(Perl_debug_log, "\n");
15504 if (last && next > last)
15509 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
15510 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
15511 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
15513 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
15515 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
15517 else if ( op == PLUS || op == STAR) {
15518 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
15520 else if (PL_regkind[(U8)op] == ANYOF) {
15521 /* arglen 1 + class block */
15522 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_CLASS)
15523 ? ANYOF_CLASS_SKIP : ANYOF_SKIP);
15524 node = NEXTOPER(node);
15526 else if (PL_regkind[(U8)op] == EXACT) {
15527 /* Literal string, where present. */
15528 node += NODE_SZ_STR(node) - 1;
15529 node = NEXTOPER(node);
15532 node = NEXTOPER(node);
15533 node += regarglen[(U8)op];
15535 if (op == CURLYX || op == OPEN)
15539 #ifdef DEBUG_DUMPUNTIL
15540 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
15545 #endif /* DEBUGGING */
15549 * c-indentation-style: bsd
15550 * c-basic-offset: 4
15551 * indent-tabs-mode: nil
15554 * ex: set ts=8 sts=4 sw=4 et: