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, OP *expr,
4881 U32 pm_flags, char *pat, STRLEN plen)
4886 /* avoid infinitely recursing when we recompile the pattern parcelled up
4887 * as qr'...'. A single constant qr// string can't have have any
4888 * run-time component in it, and thus, no runtime code. (A non-qr
4889 * string, however, can, e.g. $x =~ '(?{})') */
4890 if ((pm_flags & PMf_IS_QR) && expr && expr->op_type == OP_CONST)
4893 for (s = 0; s < plen; s++) {
4894 if (n < pRExC_state->num_code_blocks
4895 && s == pRExC_state->code_blocks[n].start)
4897 s = pRExC_state->code_blocks[n].end;
4901 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
4903 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
4905 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
4912 /* Handle run-time code blocks. We will already have compiled any direct
4913 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
4914 * copy of it, but with any literal code blocks blanked out and
4915 * appropriate chars escaped; then feed it into
4917 * eval "qr'modified_pattern'"
4921 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
4925 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
4927 * After eval_sv()-ing that, grab any new code blocks from the returned qr
4928 * and merge them with any code blocks of the original regexp.
4930 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
4931 * instead, just save the qr and return FALSE; this tells our caller that
4932 * the original pattern needs upgrading to utf8.
4936 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
4937 char *pat, STRLEN plen)
4941 GET_RE_DEBUG_FLAGS_DECL;
4943 if (pRExC_state->runtime_code_qr) {
4944 /* this is the second time we've been called; this should
4945 * only happen if the main pattern got upgraded to utf8
4946 * during compilation; re-use the qr we compiled first time
4947 * round (which should be utf8 too)
4949 qr = pRExC_state->runtime_code_qr;
4950 pRExC_state->runtime_code_qr = NULL;
4951 assert(RExC_utf8 && SvUTF8(qr));
4957 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
4961 /* determine how many extra chars we need for ' and \ escaping */
4962 for (s = 0; s < plen; s++) {
4963 if (pat[s] == '\'' || pat[s] == '\\')
4967 Newx(newpat, newlen, char);
4969 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
4971 for (s = 0; s < plen; s++) {
4972 if (n < pRExC_state->num_code_blocks
4973 && s == pRExC_state->code_blocks[n].start)
4975 /* blank out literal code block */
4976 assert(pat[s] == '(');
4977 while (s <= pRExC_state->code_blocks[n].end) {
4985 if (pat[s] == '\'' || pat[s] == '\\')
4990 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
4994 PerlIO_printf(Perl_debug_log,
4995 "%sre-parsing pattern for runtime code:%s %s\n",
4996 PL_colors[4],PL_colors[5],newpat);
4999 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5005 PUSHSTACKi(PERLSI_REQUIRE);
5006 /* this causes the toker to collapse \\ into \ when parsing
5007 * qr''; normally only q'' does this. It also alters hints
5009 PL_reg_state.re_reparsing = TRUE;
5010 eval_sv(sv, G_SCALAR);
5011 SvREFCNT_dec_NN(sv);
5016 SV * const errsv = ERRSV;
5017 if (SvTRUE_NN(errsv))
5019 Safefree(pRExC_state->code_blocks);
5020 /* use croak_sv ? */
5021 Perl_croak_nocontext("%s", SvPV_nolen_const(errsv));
5024 assert(SvROK(qr_ref));
5026 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5027 /* the leaving below frees the tmp qr_ref.
5028 * Give qr a life of its own */
5036 if (!RExC_utf8 && SvUTF8(qr)) {
5037 /* first time through; the pattern got upgraded; save the
5038 * qr for the next time through */
5039 assert(!pRExC_state->runtime_code_qr);
5040 pRExC_state->runtime_code_qr = qr;
5045 /* extract any code blocks within the returned qr// */
5048 /* merge the main (r1) and run-time (r2) code blocks into one */
5050 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
5051 struct reg_code_block *new_block, *dst;
5052 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
5055 if (!r2->num_code_blocks) /* we guessed wrong */
5057 SvREFCNT_dec_NN(qr);
5062 r1->num_code_blocks + r2->num_code_blocks,
5063 struct reg_code_block);
5066 while ( i1 < r1->num_code_blocks
5067 || i2 < r2->num_code_blocks)
5069 struct reg_code_block *src;
5072 if (i1 == r1->num_code_blocks) {
5073 src = &r2->code_blocks[i2++];
5076 else if (i2 == r2->num_code_blocks)
5077 src = &r1->code_blocks[i1++];
5078 else if ( r1->code_blocks[i1].start
5079 < r2->code_blocks[i2].start)
5081 src = &r1->code_blocks[i1++];
5082 assert(src->end < r2->code_blocks[i2].start);
5085 assert( r1->code_blocks[i1].start
5086 > r2->code_blocks[i2].start);
5087 src = &r2->code_blocks[i2++];
5089 assert(src->end < r1->code_blocks[i1].start);
5092 assert(pat[src->start] == '(');
5093 assert(pat[src->end] == ')');
5094 dst->start = src->start;
5095 dst->end = src->end;
5096 dst->block = src->block;
5097 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
5101 r1->num_code_blocks += r2->num_code_blocks;
5102 Safefree(r1->code_blocks);
5103 r1->code_blocks = new_block;
5106 SvREFCNT_dec_NN(qr);
5112 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)
5114 /* This is the common code for setting up the floating and fixed length
5115 * string data extracted from Perlre_op_compile() below. Returns a boolean
5116 * as to whether succeeded or not */
5120 if (! (longest_length
5121 || (eol /* Can't have SEOL and MULTI */
5122 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
5124 /* See comments for join_exact for why REG_SEEN_EXACTF_SHARP_S */
5125 || (RExC_seen & REG_SEEN_EXACTF_SHARP_S))
5130 /* copy the information about the longest from the reg_scan_data
5131 over to the program. */
5132 if (SvUTF8(sv_longest)) {
5133 *rx_utf8 = sv_longest;
5136 *rx_substr = sv_longest;
5139 /* end_shift is how many chars that must be matched that
5140 follow this item. We calculate it ahead of time as once the
5141 lookbehind offset is added in we lose the ability to correctly
5143 ml = minlen ? *(minlen) : (I32)longest_length;
5144 *rx_end_shift = ml - offset
5145 - longest_length + (SvTAIL(sv_longest) != 0)
5148 t = (eol/* Can't have SEOL and MULTI */
5149 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
5150 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
5156 * Perl_re_op_compile - the perl internal RE engine's function to compile a
5157 * regular expression into internal code.
5158 * The pattern may be passed either as:
5159 * a list of SVs (patternp plus pat_count)
5160 * a list of OPs (expr)
5161 * If both are passed, the SV list is used, but the OP list indicates
5162 * which SVs are actually pre-compiled code blocks
5164 * The SVs in the list have magic and qr overloading applied to them (and
5165 * the list may be modified in-place with replacement SVs in the latter
5168 * If the pattern hasn't changed from old_re, then old_re will be
5171 * eng is the current engine. If that engine has an op_comp method, then
5172 * handle directly (i.e. we assume that op_comp was us); otherwise, just
5173 * do the initial concatenation of arguments and pass on to the external
5176 * If is_bare_re is not null, set it to a boolean indicating whether the
5177 * arg list reduced (after overloading) to a single bare regex which has
5178 * been returned (i.e. /$qr/).
5180 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
5182 * pm_flags contains the PMf_* flags, typically based on those from the
5183 * pm_flags field of the related PMOP. Currently we're only interested in
5184 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
5186 * We can't allocate space until we know how big the compiled form will be,
5187 * but we can't compile it (and thus know how big it is) until we've got a
5188 * place to put the code. So we cheat: we compile it twice, once with code
5189 * generation turned off and size counting turned on, and once "for real".
5190 * This also means that we don't allocate space until we are sure that the
5191 * thing really will compile successfully, and we never have to move the
5192 * code and thus invalidate pointers into it. (Note that it has to be in
5193 * one piece because free() must be able to free it all.) [NB: not true in perl]
5195 * Beware that the optimization-preparation code in here knows about some
5196 * of the structure of the compiled regexp. [I'll say.]
5200 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
5201 OP *expr, const regexp_engine* eng, REGEXP *old_re,
5202 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
5207 regexp_internal *ri;
5216 SV *code_blocksv = NULL;
5218 /* these are all flags - maybe they should be turned
5219 * into a single int with different bit masks */
5220 I32 sawlookahead = 0;
5223 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
5224 bool code_is_utf8 = 0;
5226 bool runtime_code = 0;
5228 RExC_state_t RExC_state;
5229 RExC_state_t * const pRExC_state = &RExC_state;
5230 #ifdef TRIE_STUDY_OPT
5232 RExC_state_t copyRExC_state;
5234 GET_RE_DEBUG_FLAGS_DECL;
5236 PERL_ARGS_ASSERT_RE_OP_COMPILE;
5238 DEBUG_r(if (!PL_colorset) reginitcolors());
5240 #ifndef PERL_IN_XSUB_RE
5241 /* Initialize these here instead of as-needed, as is quick and avoids
5242 * having to test them each time otherwise */
5243 if (! PL_AboveLatin1) {
5244 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
5245 PL_ASCII = _new_invlist_C_array(ASCII_invlist);
5246 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
5248 PL_L1Posix_ptrs[_CC_ALPHANUMERIC]
5249 = _new_invlist_C_array(L1PosixAlnum_invlist);
5250 PL_Posix_ptrs[_CC_ALPHANUMERIC]
5251 = _new_invlist_C_array(PosixAlnum_invlist);
5253 PL_L1Posix_ptrs[_CC_ALPHA]
5254 = _new_invlist_C_array(L1PosixAlpha_invlist);
5255 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(PosixAlpha_invlist);
5257 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(PosixBlank_invlist);
5258 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(XPosixBlank_invlist);
5260 /* Cased is the same as Alpha in the ASCII range */
5261 PL_L1Posix_ptrs[_CC_CASED] = _new_invlist_C_array(L1Cased_invlist);
5262 PL_Posix_ptrs[_CC_CASED] = _new_invlist_C_array(PosixAlpha_invlist);
5264 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(PosixCntrl_invlist);
5265 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(XPosixCntrl_invlist);
5267 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5268 PL_L1Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5270 PL_L1Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(L1PosixGraph_invlist);
5271 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(PosixGraph_invlist);
5273 PL_L1Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(L1PosixLower_invlist);
5274 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(PosixLower_invlist);
5276 PL_L1Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(L1PosixPrint_invlist);
5277 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(PosixPrint_invlist);
5279 PL_L1Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(L1PosixPunct_invlist);
5280 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(PosixPunct_invlist);
5282 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(PerlSpace_invlist);
5283 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(XPerlSpace_invlist);
5284 PL_Posix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(PosixSpace_invlist);
5285 PL_XPosix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(XPosixSpace_invlist);
5287 PL_L1Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(L1PosixUpper_invlist);
5288 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(PosixUpper_invlist);
5290 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(VertSpace_invlist);
5292 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(PosixWord_invlist);
5293 PL_L1Posix_ptrs[_CC_WORDCHAR]
5294 = _new_invlist_C_array(L1PosixWord_invlist);
5296 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(PosixXDigit_invlist);
5297 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(XPosixXDigit_invlist);
5299 PL_HasMultiCharFold = _new_invlist_C_array(_Perl_Multi_Char_Folds_invlist);
5303 pRExC_state->code_blocks = NULL;
5304 pRExC_state->num_code_blocks = 0;
5307 *is_bare_re = FALSE;
5309 if (expr && (expr->op_type == OP_LIST ||
5310 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
5312 /* is the source UTF8, and how many code blocks are there? */
5316 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5317 if (o->op_type == OP_CONST) {
5318 /* skip if we have SVs as well as OPs. In this case,
5319 * a) we decide utf8 based on SVs not OPs;
5320 * b) the current pad may not match that which the ops
5321 * were compiled in, so, so on threaded builds,
5322 * cSVOPo_sv would look in the wrong pad */
5323 if (!pat_count && SvUTF8(cSVOPo_sv))
5326 else if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
5327 /* count of DO blocks */
5331 pRExC_state->num_code_blocks = ncode;
5332 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
5337 /* handle a list of SVs */
5341 /* apply magic and RE overloading to each arg */
5342 for (svp = patternp; svp < patternp + pat_count; svp++) {
5345 if (SvROK(rx) && SvAMAGIC(rx)) {
5346 SV *sv = AMG_CALLunary(rx, regexp_amg);
5350 if (SvTYPE(sv) != SVt_REGEXP)
5351 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5357 if (pat_count > 1) {
5358 /* concat multiple args and find any code block indexes */
5363 STRLEN orig_patlen = 0;
5365 if (pRExC_state->num_code_blocks) {
5366 o = cLISTOPx(expr)->op_first;
5367 assert( o->op_type == OP_PUSHMARK
5368 || (o->op_type == OP_NULL && o->op_targ == OP_PUSHMARK)
5369 || o->op_type == OP_PADRANGE);
5373 pat = newSVpvn("", 0);
5376 /* determine if the pattern is going to be utf8 (needed
5377 * in advance to align code block indices correctly).
5378 * XXX This could fail to be detected for an arg with
5379 * overloading but not concat overloading; but the main effect
5380 * in this obscure case is to need a 'use re eval' for a
5381 * literal code block */
5382 for (svp = patternp; svp < patternp + pat_count; svp++) {
5389 for (svp = patternp; svp < patternp + pat_count; svp++) {
5390 SV *sv, *msv = *svp;
5393 /* we make the assumption here that each op in the list of
5394 * op_siblings maps to one SV pushed onto the stack,
5395 * except for code blocks, with have both an OP_NULL and
5397 * This allows us to match up the list of SVs against the
5398 * list of OPs to find the next code block.
5400 * Note that PUSHMARK PADSV PADSV ..
5402 * PADRANGE NULL NULL ..
5403 * so the alignment still works. */
5405 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL)) {
5406 assert(n < pRExC_state->num_code_blocks);
5407 pRExC_state->code_blocks[n].start = SvCUR(pat);
5408 pRExC_state->code_blocks[n].block = o;
5409 pRExC_state->code_blocks[n].src_regex = NULL;
5412 o = o->op_sibling; /* skip CONST */
5418 if ((SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5419 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5422 /* overloading involved: all bets are off over literal
5423 * code. Pretend we haven't seen it */
5424 pRExC_state->num_code_blocks -= n;
5430 while (SvAMAGIC(msv)
5431 && (sv = AMG_CALLunary(msv, string_amg))
5435 && SvRV(msv) == SvRV(sv))
5440 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5442 orig_patlen = SvCUR(pat);
5443 sv_catsv_nomg(pat, msv);
5446 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5449 /* extract any code blocks within any embedded qr//'s */
5450 if (rx && SvTYPE(rx) == SVt_REGEXP
5451 && RX_ENGINE((REGEXP*)rx)->op_comp)
5454 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5455 if (ri->num_code_blocks) {
5457 /* the presence of an embedded qr// with code means
5458 * we should always recompile: the text of the
5459 * qr// may not have changed, but it may be a
5460 * different closure than last time */
5462 Renew(pRExC_state->code_blocks,
5463 pRExC_state->num_code_blocks + ri->num_code_blocks,
5464 struct reg_code_block);
5465 pRExC_state->num_code_blocks += ri->num_code_blocks;
5466 for (i=0; i < ri->num_code_blocks; i++) {
5467 struct reg_code_block *src, *dst;
5468 STRLEN offset = orig_patlen
5469 + ReANY((REGEXP *)rx)->pre_prefix;
5470 assert(n < pRExC_state->num_code_blocks);
5471 src = &ri->code_blocks[i];
5472 dst = &pRExC_state->code_blocks[n];
5473 dst->start = src->start + offset;
5474 dst->end = src->end + offset;
5475 dst->block = src->block;
5476 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5490 while (SvAMAGIC(pat)
5491 && (sv = AMG_CALLunary(pat, string_amg))
5499 /* handle bare regex: foo =~ $re */
5504 if (SvTYPE(re) == SVt_REGEXP) {
5508 Safefree(pRExC_state->code_blocks);
5514 /* not a list of SVs, so must be a list of OPs */
5516 if (expr->op_type == OP_LIST) {
5521 pat = newSVpvn("", 0);
5526 /* given a list of CONSTs and DO blocks in expr, append all
5527 * the CONSTs to pat, and record the start and end of each
5528 * code block in code_blocks[] (each DO{} op is followed by an
5529 * OP_CONST containing the corresponding literal '(?{...})
5532 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5533 if (o->op_type == OP_CONST) {
5534 sv_catsv(pat, cSVOPo_sv);
5536 pRExC_state->code_blocks[i].end = SvCUR(pat)-1;
5540 else if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL)) {
5541 assert(i+1 < pRExC_state->num_code_blocks);
5542 pRExC_state->code_blocks[++i].start = SvCUR(pat);
5543 pRExC_state->code_blocks[i].block = o;
5544 pRExC_state->code_blocks[i].src_regex = NULL;
5550 assert(expr->op_type == OP_CONST);
5551 pat = cSVOPx_sv(expr);
5555 exp = SvPV_nomg(pat, plen);
5558 if (!eng->op_comp) {
5559 if ((SvUTF8(pat) && IN_BYTES)
5560 || SvGMAGICAL(pat) || SvAMAGIC(pat))
5562 /* make a temporary copy; either to convert to bytes,
5563 * or to avoid repeating get-magic / overloaded stringify */
5564 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
5565 (IN_BYTES ? 0 : SvUTF8(pat)));
5567 Safefree(pRExC_state->code_blocks);
5568 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
5571 /* ignore the utf8ness if the pattern is 0 length */
5572 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
5573 RExC_uni_semantics = 0;
5574 RExC_contains_locale = 0;
5575 pRExC_state->runtime_code_qr = NULL;
5578 SV *dsv= sv_newmortal();
5579 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
5580 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
5581 PL_colors[4],PL_colors[5],s);
5587 U8 *const src = (U8*)exp;
5590 STRLEN s = 0, d = 0;
5593 /* It's possible to write a regexp in ascii that represents Unicode
5594 codepoints outside of the byte range, such as via \x{100}. If we
5595 detect such a sequence we have to convert the entire pattern to utf8
5596 and then recompile, as our sizing calculation will have been based
5597 on 1 byte == 1 character, but we will need to use utf8 to encode
5598 at least some part of the pattern, and therefore must convert the whole
5601 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5602 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
5604 /* upgrade pattern to UTF8, and if there are code blocks,
5605 * recalculate the indices.
5606 * This is essentially an unrolled Perl_bytes_to_utf8() */
5608 Newx(dst, plen * 2 + 1, U8);
5611 const UV uv = NATIVE_TO_ASCII(src[s]);
5612 if (UNI_IS_INVARIANT(uv))
5613 dst[d] = (U8)UTF_TO_NATIVE(uv);
5615 dst[d++] = (U8)UTF8_EIGHT_BIT_HI(uv);
5616 dst[d] = (U8)UTF8_EIGHT_BIT_LO(uv);
5618 if (n < pRExC_state->num_code_blocks) {
5619 if (!do_end && pRExC_state->code_blocks[n].start == s) {
5620 pRExC_state->code_blocks[n].start = d;
5621 assert(dst[d] == '(');
5624 else if (do_end && pRExC_state->code_blocks[n].end == s) {
5625 pRExC_state->code_blocks[n].end = d;
5626 assert(dst[d] == ')');
5639 RExC_orig_utf8 = RExC_utf8 = 1;
5643 /* return old regex if pattern hasn't changed */
5647 && !!RX_UTF8(old_re) == !!RExC_utf8
5648 && RX_PRECOMP(old_re)
5649 && RX_PRELEN(old_re) == plen
5650 && memEQ(RX_PRECOMP(old_re), exp, plen))
5652 /* with runtime code, always recompile */
5653 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, expr, pm_flags,
5655 if (!runtime_code) {
5656 Safefree(pRExC_state->code_blocks);
5660 else if ((pm_flags & PMf_USE_RE_EVAL)
5661 /* this second condition covers the non-regex literal case,
5662 * i.e. $foo =~ '(?{})'. */
5663 || ( !PL_reg_state.re_reparsing && IN_PERL_COMPILETIME
5664 && (PL_hints & HINT_RE_EVAL))
5666 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, expr, pm_flags,
5669 rx_flags = orig_rx_flags;
5671 if (initial_charset == REGEX_LOCALE_CHARSET) {
5672 RExC_contains_locale = 1;
5674 else if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
5676 /* Set to use unicode semantics if the pattern is in utf8 and has the
5677 * 'depends' charset specified, as it means unicode when utf8 */
5678 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5682 RExC_flags = rx_flags;
5683 RExC_pm_flags = pm_flags;
5686 if (TAINTING_get && TAINT_get)
5687 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
5689 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
5690 /* whoops, we have a non-utf8 pattern, whilst run-time code
5691 * got compiled as utf8. Try again with a utf8 pattern */
5692 goto redo_first_pass;
5695 assert(!pRExC_state->runtime_code_qr);
5700 RExC_in_lookbehind = 0;
5701 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
5703 RExC_override_recoding = 0;
5704 RExC_in_multi_char_class = 0;
5706 /* First pass: determine size, legality. */
5714 RExC_emit = &PL_regdummy;
5715 RExC_whilem_seen = 0;
5716 RExC_open_parens = NULL;
5717 RExC_close_parens = NULL;
5719 RExC_paren_names = NULL;
5721 RExC_paren_name_list = NULL;
5723 RExC_recurse = NULL;
5724 RExC_recurse_count = 0;
5725 pRExC_state->code_index = 0;
5727 #if 0 /* REGC() is (currently) a NOP at the first pass.
5728 * Clever compilers notice this and complain. --jhi */
5729 REGC((U8)REG_MAGIC, (char*)RExC_emit);
5732 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
5734 RExC_lastparse=NULL;
5736 /* reg may croak on us, not giving us a chance to free
5737 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
5738 need it to survive as long as the regexp (qr/(?{})/).
5739 We must check that code_blocksv is not already set, because we may
5740 have jumped back to restart the sizing pass. */
5741 if (pRExC_state->code_blocks && !code_blocksv) {
5742 code_blocksv = newSV_type(SVt_PV);
5743 SAVEFREESV(code_blocksv);
5744 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
5745 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
5747 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5748 if (flags & RESTART_UTF8) {
5749 goto redo_first_pass;
5751 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#X", flags);
5754 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
5757 PerlIO_printf(Perl_debug_log,
5758 "Required size %"IVdf" nodes\n"
5759 "Starting second pass (creation)\n",
5762 RExC_lastparse=NULL;
5765 /* The first pass could have found things that force Unicode semantics */
5766 if ((RExC_utf8 || RExC_uni_semantics)
5767 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
5769 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5772 /* Small enough for pointer-storage convention?
5773 If extralen==0, this means that we will not need long jumps. */
5774 if (RExC_size >= 0x10000L && RExC_extralen)
5775 RExC_size += RExC_extralen;
5778 if (RExC_whilem_seen > 15)
5779 RExC_whilem_seen = 15;
5781 /* Allocate space and zero-initialize. Note, the two step process
5782 of zeroing when in debug mode, thus anything assigned has to
5783 happen after that */
5784 rx = (REGEXP*) newSV_type(SVt_REGEXP);
5786 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
5787 char, regexp_internal);
5788 if ( r == NULL || ri == NULL )
5789 FAIL("Regexp out of space");
5791 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
5792 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode), char);
5794 /* bulk initialize base fields with 0. */
5795 Zero(ri, sizeof(regexp_internal), char);
5798 /* non-zero initialization begins here */
5801 r->extflags = rx_flags;
5802 if (pm_flags & PMf_IS_QR) {
5803 ri->code_blocks = pRExC_state->code_blocks;
5804 ri->num_code_blocks = pRExC_state->num_code_blocks;
5809 for (n = 0; n < pRExC_state->num_code_blocks; n++)
5810 if (pRExC_state->code_blocks[n].src_regex)
5811 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
5812 SAVEFREEPV(pRExC_state->code_blocks);
5816 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
5817 bool has_charset = (get_regex_charset(r->extflags) != REGEX_DEPENDS_CHARSET);
5819 /* The caret is output if there are any defaults: if not all the STD
5820 * flags are set, or if no character set specifier is needed */
5822 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
5824 bool has_runon = ((RExC_seen & REG_SEEN_RUN_ON_COMMENT)==REG_SEEN_RUN_ON_COMMENT);
5825 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
5826 >> RXf_PMf_STD_PMMOD_SHIFT);
5827 const char *fptr = STD_PAT_MODS; /*"msix"*/
5829 /* Allocate for the worst case, which is all the std flags are turned
5830 * on. If more precision is desired, we could do a population count of
5831 * the flags set. This could be done with a small lookup table, or by
5832 * shifting, masking and adding, or even, when available, assembly
5833 * language for a machine-language population count.
5834 * We never output a minus, as all those are defaults, so are
5835 * covered by the caret */
5836 const STRLEN wraplen = plen + has_p + has_runon
5837 + has_default /* If needs a caret */
5839 /* If needs a character set specifier */
5840 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
5841 + (sizeof(STD_PAT_MODS) - 1)
5842 + (sizeof("(?:)") - 1);
5844 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
5845 r->xpv_len_u.xpvlenu_pv = p;
5847 SvFLAGS(rx) |= SVf_UTF8;
5850 /* If a default, cover it using the caret */
5852 *p++= DEFAULT_PAT_MOD;
5856 const char* const name = get_regex_charset_name(r->extflags, &len);
5857 Copy(name, p, len, char);
5861 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
5864 while((ch = *fptr++)) {
5872 Copy(RExC_precomp, p, plen, char);
5873 assert ((RX_WRAPPED(rx) - p) < 16);
5874 r->pre_prefix = p - RX_WRAPPED(rx);
5880 SvCUR_set(rx, p - RX_WRAPPED(rx));
5884 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
5886 if (RExC_seen & REG_SEEN_RECURSE) {
5887 Newxz(RExC_open_parens, RExC_npar,regnode *);
5888 SAVEFREEPV(RExC_open_parens);
5889 Newxz(RExC_close_parens,RExC_npar,regnode *);
5890 SAVEFREEPV(RExC_close_parens);
5893 /* Useful during FAIL. */
5894 #ifdef RE_TRACK_PATTERN_OFFSETS
5895 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
5896 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
5897 "%s %"UVuf" bytes for offset annotations.\n",
5898 ri->u.offsets ? "Got" : "Couldn't get",
5899 (UV)((2*RExC_size+1) * sizeof(U32))));
5901 SetProgLen(ri,RExC_size);
5906 /* Second pass: emit code. */
5907 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
5908 RExC_pm_flags = pm_flags;
5913 RExC_emit_start = ri->program;
5914 RExC_emit = ri->program;
5915 RExC_emit_bound = ri->program + RExC_size + 1;
5916 pRExC_state->code_index = 0;
5918 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
5919 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5921 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#X", flags);
5923 /* XXXX To minimize changes to RE engine we always allocate
5924 3-units-long substrs field. */
5925 Newx(r->substrs, 1, struct reg_substr_data);
5926 if (RExC_recurse_count) {
5927 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
5928 SAVEFREEPV(RExC_recurse);
5932 r->minlen = minlen = sawlookahead = sawplus = sawopen = 0;
5933 Zero(r->substrs, 1, struct reg_substr_data);
5935 #ifdef TRIE_STUDY_OPT
5937 StructCopy(&zero_scan_data, &data, scan_data_t);
5938 copyRExC_state = RExC_state;
5941 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
5943 RExC_state = copyRExC_state;
5944 if (seen & REG_TOP_LEVEL_BRANCHES)
5945 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
5947 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES;
5948 StructCopy(&zero_scan_data, &data, scan_data_t);
5951 StructCopy(&zero_scan_data, &data, scan_data_t);
5954 /* Dig out information for optimizations. */
5955 r->extflags = RExC_flags; /* was pm_op */
5956 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
5959 SvUTF8_on(rx); /* Unicode in it? */
5960 ri->regstclass = NULL;
5961 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
5962 r->intflags |= PREGf_NAUGHTY;
5963 scan = ri->program + 1; /* First BRANCH. */
5965 /* testing for BRANCH here tells us whether there is "must appear"
5966 data in the pattern. If there is then we can use it for optimisations */
5967 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES)) { /* Only one top-level choice. */
5969 STRLEN longest_float_length, longest_fixed_length;
5970 struct regnode_charclass_class ch_class; /* pointed to by data */
5972 I32 last_close = 0; /* pointed to by data */
5973 regnode *first= scan;
5974 regnode *first_next= regnext(first);
5976 * Skip introductions and multiplicators >= 1
5977 * so that we can extract the 'meat' of the pattern that must
5978 * match in the large if() sequence following.
5979 * NOTE that EXACT is NOT covered here, as it is normally
5980 * picked up by the optimiser separately.
5982 * This is unfortunate as the optimiser isnt handling lookahead
5983 * properly currently.
5986 while ((OP(first) == OPEN && (sawopen = 1)) ||
5987 /* An OR of *one* alternative - should not happen now. */
5988 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
5989 /* for now we can't handle lookbehind IFMATCH*/
5990 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
5991 (OP(first) == PLUS) ||
5992 (OP(first) == MINMOD) ||
5993 /* An {n,m} with n>0 */
5994 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
5995 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
5998 * the only op that could be a regnode is PLUS, all the rest
5999 * will be regnode_1 or regnode_2.
6002 if (OP(first) == PLUS)
6005 first += regarglen[OP(first)];
6007 first = NEXTOPER(first);
6008 first_next= regnext(first);
6011 /* Starting-point info. */
6013 DEBUG_PEEP("first:",first,0);
6014 /* Ignore EXACT as we deal with it later. */
6015 if (PL_regkind[OP(first)] == EXACT) {
6016 if (OP(first) == EXACT)
6017 NOOP; /* Empty, get anchored substr later. */
6019 ri->regstclass = first;
6022 else if (PL_regkind[OP(first)] == TRIE &&
6023 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6026 /* this can happen only on restudy */
6027 if ( OP(first) == TRIE ) {
6028 struct regnode_1 *trieop = (struct regnode_1 *)
6029 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6030 StructCopy(first,trieop,struct regnode_1);
6031 trie_op=(regnode *)trieop;
6033 struct regnode_charclass *trieop = (struct regnode_charclass *)
6034 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6035 StructCopy(first,trieop,struct regnode_charclass);
6036 trie_op=(regnode *)trieop;
6039 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6040 ri->regstclass = trie_op;
6043 else if (REGNODE_SIMPLE(OP(first)))
6044 ri->regstclass = first;
6045 else if (PL_regkind[OP(first)] == BOUND ||
6046 PL_regkind[OP(first)] == NBOUND)
6047 ri->regstclass = first;
6048 else if (PL_regkind[OP(first)] == BOL) {
6049 r->extflags |= (OP(first) == MBOL
6051 : (OP(first) == SBOL
6054 first = NEXTOPER(first);
6057 else if (OP(first) == GPOS) {
6058 r->extflags |= RXf_ANCH_GPOS;
6059 first = NEXTOPER(first);
6062 else if ((!sawopen || !RExC_sawback) &&
6063 (OP(first) == STAR &&
6064 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6065 !(r->extflags & RXf_ANCH) && !pRExC_state->num_code_blocks)
6067 /* turn .* into ^.* with an implied $*=1 */
6069 (OP(NEXTOPER(first)) == REG_ANY)
6072 r->extflags |= type;
6073 r->intflags |= PREGf_IMPLICIT;
6074 first = NEXTOPER(first);
6077 if (sawplus && !sawlookahead && (!sawopen || !RExC_sawback)
6078 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6079 /* x+ must match at the 1st pos of run of x's */
6080 r->intflags |= PREGf_SKIP;
6082 /* Scan is after the zeroth branch, first is atomic matcher. */
6083 #ifdef TRIE_STUDY_OPT
6086 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6087 (IV)(first - scan + 1))
6091 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6092 (IV)(first - scan + 1))
6098 * If there's something expensive in the r.e., find the
6099 * longest literal string that must appear and make it the
6100 * regmust. Resolve ties in favor of later strings, since
6101 * the regstart check works with the beginning of the r.e.
6102 * and avoiding duplication strengthens checking. Not a
6103 * strong reason, but sufficient in the absence of others.
6104 * [Now we resolve ties in favor of the earlier string if
6105 * it happens that c_offset_min has been invalidated, since the
6106 * earlier string may buy us something the later one won't.]
6109 data.longest_fixed = newSVpvs("");
6110 data.longest_float = newSVpvs("");
6111 data.last_found = newSVpvs("");
6112 data.longest = &(data.longest_fixed);
6113 ENTER_with_name("study_chunk");
6114 SAVEFREESV(data.longest_fixed);
6115 SAVEFREESV(data.longest_float);
6116 SAVEFREESV(data.last_found);
6118 if (!ri->regstclass) {
6119 cl_init(pRExC_state, &ch_class);
6120 data.start_class = &ch_class;
6121 stclass_flag = SCF_DO_STCLASS_AND;
6122 } else /* XXXX Check for BOUND? */
6124 data.last_closep = &last_close;
6126 minlen = study_chunk(pRExC_state, &first, &minlen, &fake, scan + RExC_size, /* Up to end */
6127 &data, -1, NULL, NULL,
6128 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag,0);
6131 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6134 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6135 && data.last_start_min == 0 && data.last_end > 0
6136 && !RExC_seen_zerolen
6137 && !(RExC_seen & REG_SEEN_VERBARG)
6138 && (!(RExC_seen & REG_SEEN_GPOS) || (r->extflags & RXf_ANCH_GPOS)))
6139 r->extflags |= RXf_CHECK_ALL;
6140 scan_commit(pRExC_state, &data,&minlen,0);
6142 longest_float_length = CHR_SVLEN(data.longest_float);
6144 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6145 && data.offset_fixed == data.offset_float_min
6146 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6147 && S_setup_longest (aTHX_ pRExC_state,
6151 &(r->float_end_shift),
6152 data.lookbehind_float,
6153 data.offset_float_min,
6155 longest_float_length,
6156 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6157 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6159 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6160 r->float_max_offset = data.offset_float_max;
6161 if (data.offset_float_max < I32_MAX) /* Don't offset infinity */
6162 r->float_max_offset -= data.lookbehind_float;
6163 SvREFCNT_inc_simple_void_NN(data.longest_float);
6166 r->float_substr = r->float_utf8 = NULL;
6167 longest_float_length = 0;
6170 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6172 if (S_setup_longest (aTHX_ pRExC_state,
6174 &(r->anchored_utf8),
6175 &(r->anchored_substr),
6176 &(r->anchored_end_shift),
6177 data.lookbehind_fixed,
6180 longest_fixed_length,
6181 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6182 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6184 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6185 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6188 r->anchored_substr = r->anchored_utf8 = NULL;
6189 longest_fixed_length = 0;
6191 LEAVE_with_name("study_chunk");
6194 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6195 ri->regstclass = NULL;
6197 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6199 && ! TEST_SSC_EOS(data.start_class)
6200 && !cl_is_anything(data.start_class))
6202 const U32 n = add_data(pRExC_state, 1, "f");
6203 OP(data.start_class) = ANYOF_SYNTHETIC;
6205 Newx(RExC_rxi->data->data[n], 1,
6206 struct regnode_charclass_class);
6207 StructCopy(data.start_class,
6208 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6209 struct regnode_charclass_class);
6210 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6211 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6212 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6213 regprop(r, sv, (regnode*)data.start_class);
6214 PerlIO_printf(Perl_debug_log,
6215 "synthetic stclass \"%s\".\n",
6216 SvPVX_const(sv));});
6219 /* A temporary algorithm prefers floated substr to fixed one to dig more info. */
6220 if (longest_fixed_length > longest_float_length) {
6221 r->check_end_shift = r->anchored_end_shift;
6222 r->check_substr = r->anchored_substr;
6223 r->check_utf8 = r->anchored_utf8;
6224 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6225 if (r->extflags & RXf_ANCH_SINGLE)
6226 r->extflags |= RXf_NOSCAN;
6229 r->check_end_shift = r->float_end_shift;
6230 r->check_substr = r->float_substr;
6231 r->check_utf8 = r->float_utf8;
6232 r->check_offset_min = r->float_min_offset;
6233 r->check_offset_max = r->float_max_offset;
6235 /* XXXX Currently intuiting is not compatible with ANCH_GPOS.
6236 This should be changed ASAP! */
6237 if ((r->check_substr || r->check_utf8) && !(r->extflags & RXf_ANCH_GPOS)) {
6238 r->extflags |= RXf_USE_INTUIT;
6239 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6240 r->extflags |= RXf_INTUIT_TAIL;
6242 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6243 if ( (STRLEN)minlen < longest_float_length )
6244 minlen= longest_float_length;
6245 if ( (STRLEN)minlen < longest_fixed_length )
6246 minlen= longest_fixed_length;
6250 /* Several toplevels. Best we can is to set minlen. */
6252 struct regnode_charclass_class ch_class;
6255 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6257 scan = ri->program + 1;
6258 cl_init(pRExC_state, &ch_class);
6259 data.start_class = &ch_class;
6260 data.last_closep = &last_close;
6263 minlen = study_chunk(pRExC_state, &scan, &minlen, &fake, scan + RExC_size,
6264 &data, -1, NULL, NULL, SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS,0);
6266 CHECK_RESTUDY_GOTO_butfirst(NOOP);
6268 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
6269 = r->float_substr = r->float_utf8 = NULL;
6271 if (! TEST_SSC_EOS(data.start_class)
6272 && !cl_is_anything(data.start_class))
6274 const U32 n = add_data(pRExC_state, 1, "f");
6275 OP(data.start_class) = ANYOF_SYNTHETIC;
6277 Newx(RExC_rxi->data->data[n], 1,
6278 struct regnode_charclass_class);
6279 StructCopy(data.start_class,
6280 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6281 struct regnode_charclass_class);
6282 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6283 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6284 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
6285 regprop(r, sv, (regnode*)data.start_class);
6286 PerlIO_printf(Perl_debug_log,
6287 "synthetic stclass \"%s\".\n",
6288 SvPVX_const(sv));});
6292 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
6293 the "real" pattern. */
6295 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf"\n",
6296 (IV)minlen, (IV)r->minlen);
6298 r->minlenret = minlen;
6299 if (r->minlen < minlen)
6302 if (RExC_seen & REG_SEEN_GPOS)
6303 r->extflags |= RXf_GPOS_SEEN;
6304 if (RExC_seen & REG_SEEN_LOOKBEHIND)
6305 r->extflags |= RXf_LOOKBEHIND_SEEN;
6306 if (pRExC_state->num_code_blocks)
6307 r->extflags |= RXf_EVAL_SEEN;
6308 if (RExC_seen & REG_SEEN_CANY)
6309 r->extflags |= RXf_CANY_SEEN;
6310 if (RExC_seen & REG_SEEN_VERBARG)
6312 r->intflags |= PREGf_VERBARG_SEEN;
6313 r->extflags |= RXf_MODIFIES_VARS;
6315 if (RExC_seen & REG_SEEN_CUTGROUP)
6316 r->intflags |= PREGf_CUTGROUP_SEEN;
6317 if (pm_flags & PMf_USE_RE_EVAL)
6318 r->intflags |= PREGf_USE_RE_EVAL;
6319 if (RExC_paren_names)
6320 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
6322 RXp_PAREN_NAMES(r) = NULL;
6324 #ifdef STUPID_PATTERN_CHECKS
6325 if (RX_PRELEN(rx) == 0)
6326 r->extflags |= RXf_NULL;
6327 if (RX_PRELEN(rx) == 3 && memEQ("\\s+", RX_PRECOMP(rx), 3))
6328 r->extflags |= RXf_WHITE;
6329 else if (RX_PRELEN(rx) == 1 && RXp_PRECOMP(rx)[0] == '^')
6330 r->extflags |= RXf_START_ONLY;
6333 regnode *first = ri->program + 1;
6336 if (PL_regkind[fop] == NOTHING && OP(NEXTOPER(first)) == END)
6337 r->extflags |= RXf_NULL;
6338 else if (PL_regkind[fop] == BOL && OP(NEXTOPER(first)) == END)
6339 r->extflags |= RXf_START_ONLY;
6340 else if (fop == PLUS && PL_regkind[OP(NEXTOPER(first))] == POSIXD && FLAGS(NEXTOPER(first)) == _CC_SPACE
6341 && OP(regnext(first)) == END)
6342 r->extflags |= RXf_WHITE;
6346 if (RExC_paren_names) {
6347 ri->name_list_idx = add_data( pRExC_state, 1, "a" );
6348 ri->data->data[ri->name_list_idx] = (void*)SvREFCNT_inc(RExC_paren_name_list);
6351 ri->name_list_idx = 0;
6353 if (RExC_recurse_count) {
6354 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
6355 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
6356 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
6359 Newxz(r->offs, RExC_npar, regexp_paren_pair);
6360 /* assume we don't need to swap parens around before we match */
6363 PerlIO_printf(Perl_debug_log,"Final program:\n");
6366 #ifdef RE_TRACK_PATTERN_OFFSETS
6367 DEBUG_OFFSETS_r(if (ri->u.offsets) {
6368 const U32 len = ri->u.offsets[0];
6370 GET_RE_DEBUG_FLAGS_DECL;
6371 PerlIO_printf(Perl_debug_log, "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
6372 for (i = 1; i <= len; i++) {
6373 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
6374 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
6375 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
6377 PerlIO_printf(Perl_debug_log, "\n");
6382 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
6383 * by setting the regexp SV to readonly-only instead. If the
6384 * pattern's been recompiled, the USEDness should remain. */
6385 if (old_re && SvREADONLY(old_re))
6393 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
6396 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
6398 PERL_UNUSED_ARG(value);
6400 if (flags & RXapif_FETCH) {
6401 return reg_named_buff_fetch(rx, key, flags);
6402 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
6403 Perl_croak_no_modify();
6405 } else if (flags & RXapif_EXISTS) {
6406 return reg_named_buff_exists(rx, key, flags)
6409 } else if (flags & RXapif_REGNAMES) {
6410 return reg_named_buff_all(rx, flags);
6411 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
6412 return reg_named_buff_scalar(rx, flags);
6414 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
6420 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
6423 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
6424 PERL_UNUSED_ARG(lastkey);
6426 if (flags & RXapif_FIRSTKEY)
6427 return reg_named_buff_firstkey(rx, flags);
6428 else if (flags & RXapif_NEXTKEY)
6429 return reg_named_buff_nextkey(rx, flags);
6431 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter", (int)flags);
6437 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
6440 AV *retarray = NULL;
6442 struct regexp *const rx = ReANY(r);
6444 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
6446 if (flags & RXapif_ALL)
6449 if (rx && RXp_PAREN_NAMES(rx)) {
6450 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
6453 SV* sv_dat=HeVAL(he_str);
6454 I32 *nums=(I32*)SvPVX(sv_dat);
6455 for ( i=0; i<SvIVX(sv_dat); i++ ) {
6456 if ((I32)(rx->nparens) >= nums[i]
6457 && rx->offs[nums[i]].start != -1
6458 && rx->offs[nums[i]].end != -1)
6461 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
6466 ret = newSVsv(&PL_sv_undef);
6469 av_push(retarray, ret);
6472 return newRV_noinc(MUTABLE_SV(retarray));
6479 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
6482 struct regexp *const rx = ReANY(r);
6484 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
6486 if (rx && RXp_PAREN_NAMES(rx)) {
6487 if (flags & RXapif_ALL) {
6488 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
6490 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
6492 SvREFCNT_dec_NN(sv);
6504 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
6506 struct regexp *const rx = ReANY(r);
6508 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
6510 if ( rx && RXp_PAREN_NAMES(rx) ) {
6511 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
6513 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
6520 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
6522 struct regexp *const rx = ReANY(r);
6523 GET_RE_DEBUG_FLAGS_DECL;
6525 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
6527 if (rx && RXp_PAREN_NAMES(rx)) {
6528 HV *hv = RXp_PAREN_NAMES(rx);
6530 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6533 SV* sv_dat = HeVAL(temphe);
6534 I32 *nums = (I32*)SvPVX(sv_dat);
6535 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6536 if ((I32)(rx->lastparen) >= nums[i] &&
6537 rx->offs[nums[i]].start != -1 &&
6538 rx->offs[nums[i]].end != -1)
6544 if (parno || flags & RXapif_ALL) {
6545 return newSVhek(HeKEY_hek(temphe));
6553 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
6558 struct regexp *const rx = ReANY(r);
6560 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
6562 if (rx && RXp_PAREN_NAMES(rx)) {
6563 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
6564 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
6565 } else if (flags & RXapif_ONE) {
6566 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
6567 av = MUTABLE_AV(SvRV(ret));
6568 length = av_len(av);
6569 SvREFCNT_dec_NN(ret);
6570 return newSViv(length + 1);
6572 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar", (int)flags);
6576 return &PL_sv_undef;
6580 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
6582 struct regexp *const rx = ReANY(r);
6585 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
6587 if (rx && RXp_PAREN_NAMES(rx)) {
6588 HV *hv= RXp_PAREN_NAMES(rx);
6590 (void)hv_iterinit(hv);
6591 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6594 SV* sv_dat = HeVAL(temphe);
6595 I32 *nums = (I32*)SvPVX(sv_dat);
6596 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6597 if ((I32)(rx->lastparen) >= nums[i] &&
6598 rx->offs[nums[i]].start != -1 &&
6599 rx->offs[nums[i]].end != -1)
6605 if (parno || flags & RXapif_ALL) {
6606 av_push(av, newSVhek(HeKEY_hek(temphe)));
6611 return newRV_noinc(MUTABLE_SV(av));
6615 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
6618 struct regexp *const rx = ReANY(r);
6624 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
6626 if ( ( n == RX_BUFF_IDX_CARET_PREMATCH
6627 || n == RX_BUFF_IDX_CARET_FULLMATCH
6628 || n == RX_BUFF_IDX_CARET_POSTMATCH
6630 && !(rx->extflags & RXf_PMf_KEEPCOPY)
6637 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
6638 /* no need to distinguish between them any more */
6639 n = RX_BUFF_IDX_FULLMATCH;
6641 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
6642 && rx->offs[0].start != -1)
6644 /* $`, ${^PREMATCH} */
6645 i = rx->offs[0].start;
6649 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
6650 && rx->offs[0].end != -1)
6652 /* $', ${^POSTMATCH} */
6653 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
6654 i = rx->sublen + rx->suboffset - rx->offs[0].end;
6657 if ( 0 <= n && n <= (I32)rx->nparens &&
6658 (s1 = rx->offs[n].start) != -1 &&
6659 (t1 = rx->offs[n].end) != -1)
6661 /* $&, ${^MATCH}, $1 ... */
6663 s = rx->subbeg + s1 - rx->suboffset;
6668 assert(s >= rx->subbeg);
6669 assert(rx->sublen >= (s - rx->subbeg) + i );
6671 #if NO_TAINT_SUPPORT
6672 sv_setpvn(sv, s, i);
6674 const int oldtainted = TAINT_get;
6676 sv_setpvn(sv, s, i);
6677 TAINT_set(oldtainted);
6679 if ( (rx->extflags & RXf_CANY_SEEN)
6680 ? (RXp_MATCH_UTF8(rx)
6681 && (!i || is_utf8_string((U8*)s, i)))
6682 : (RXp_MATCH_UTF8(rx)) )
6689 if (RXp_MATCH_TAINTED(rx)) {
6690 if (SvTYPE(sv) >= SVt_PVMG) {
6691 MAGIC* const mg = SvMAGIC(sv);
6694 SvMAGIC_set(sv, mg->mg_moremagic);
6696 if ((mgt = SvMAGIC(sv))) {
6697 mg->mg_moremagic = mgt;
6698 SvMAGIC_set(sv, mg);
6709 sv_setsv(sv,&PL_sv_undef);
6715 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
6716 SV const * const value)
6718 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
6720 PERL_UNUSED_ARG(rx);
6721 PERL_UNUSED_ARG(paren);
6722 PERL_UNUSED_ARG(value);
6725 Perl_croak_no_modify();
6729 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
6732 struct regexp *const rx = ReANY(r);
6736 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
6738 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
6740 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
6741 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6745 case RX_BUFF_IDX_PREMATCH: /* $` */
6746 if (rx->offs[0].start != -1) {
6747 i = rx->offs[0].start;
6756 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
6757 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6759 case RX_BUFF_IDX_POSTMATCH: /* $' */
6760 if (rx->offs[0].end != -1) {
6761 i = rx->sublen - rx->offs[0].end;
6763 s1 = rx->offs[0].end;
6770 case RX_BUFF_IDX_CARET_FULLMATCH: /* ${^MATCH} */
6771 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6775 /* $& / ${^MATCH}, $1, $2, ... */
6777 if (paren <= (I32)rx->nparens &&
6778 (s1 = rx->offs[paren].start) != -1 &&
6779 (t1 = rx->offs[paren].end) != -1)
6785 if (ckWARN(WARN_UNINITIALIZED))
6786 report_uninit((const SV *)sv);
6791 if (i > 0 && RXp_MATCH_UTF8(rx)) {
6792 const char * const s = rx->subbeg - rx->suboffset + s1;
6797 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
6804 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
6806 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
6807 PERL_UNUSED_ARG(rx);
6811 return newSVpvs("Regexp");
6814 /* Scans the name of a named buffer from the pattern.
6815 * If flags is REG_RSN_RETURN_NULL returns null.
6816 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
6817 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
6818 * to the parsed name as looked up in the RExC_paren_names hash.
6819 * If there is an error throws a vFAIL().. type exception.
6822 #define REG_RSN_RETURN_NULL 0
6823 #define REG_RSN_RETURN_NAME 1
6824 #define REG_RSN_RETURN_DATA 2
6827 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
6829 char *name_start = RExC_parse;
6831 PERL_ARGS_ASSERT_REG_SCAN_NAME;
6833 if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
6834 /* skip IDFIRST by using do...while */
6837 RExC_parse += UTF8SKIP(RExC_parse);
6838 } while (isWORDCHAR_utf8((U8*)RExC_parse));
6842 } while (isWORDCHAR(*RExC_parse));
6844 RExC_parse++; /* so the <- from the vFAIL is after the offending character */
6845 vFAIL("Group name must start with a non-digit word character");
6849 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
6850 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
6851 if ( flags == REG_RSN_RETURN_NAME)
6853 else if (flags==REG_RSN_RETURN_DATA) {
6856 if ( ! sv_name ) /* should not happen*/
6857 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
6858 if (RExC_paren_names)
6859 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
6861 sv_dat = HeVAL(he_str);
6863 vFAIL("Reference to nonexistent named group");
6867 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
6868 (unsigned long) flags);
6870 assert(0); /* NOT REACHED */
6875 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
6876 int rem=(int)(RExC_end - RExC_parse); \
6885 if (RExC_lastparse!=RExC_parse) \
6886 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
6889 iscut ? "..." : "<" \
6892 PerlIO_printf(Perl_debug_log,"%16s",""); \
6895 num = RExC_size + 1; \
6897 num=REG_NODE_NUM(RExC_emit); \
6898 if (RExC_lastnum!=num) \
6899 PerlIO_printf(Perl_debug_log,"|%4d",num); \
6901 PerlIO_printf(Perl_debug_log,"|%4s",""); \
6902 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
6903 (int)((depth*2)), "", \
6907 RExC_lastparse=RExC_parse; \
6912 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
6913 DEBUG_PARSE_MSG((funcname)); \
6914 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
6916 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
6917 DEBUG_PARSE_MSG((funcname)); \
6918 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
6921 /* This section of code defines the inversion list object and its methods. The
6922 * interfaces are highly subject to change, so as much as possible is static to
6923 * this file. An inversion list is here implemented as a malloc'd C UV array
6924 * with some added info that is placed as UVs at the beginning in a header
6925 * portion. An inversion list for Unicode is an array of code points, sorted
6926 * by ordinal number. The zeroth element is the first code point in the list.
6927 * The 1th element is the first element beyond that not in the list. In other
6928 * words, the first range is
6929 * invlist[0]..(invlist[1]-1)
6930 * The other ranges follow. Thus every element whose index is divisible by two
6931 * marks the beginning of a range that is in the list, and every element not
6932 * divisible by two marks the beginning of a range not in the list. A single
6933 * element inversion list that contains the single code point N generally
6934 * consists of two elements
6937 * (The exception is when N is the highest representable value on the
6938 * machine, in which case the list containing just it would be a single
6939 * element, itself. By extension, if the last range in the list extends to
6940 * infinity, then the first element of that range will be in the inversion list
6941 * at a position that is divisible by two, and is the final element in the
6943 * Taking the complement (inverting) an inversion list is quite simple, if the
6944 * first element is 0, remove it; otherwise add a 0 element at the beginning.
6945 * This implementation reserves an element at the beginning of each inversion
6946 * list to contain 0 when the list contains 0, and contains 1 otherwise. The
6947 * actual beginning of the list is either that element if 0, or the next one if
6950 * More about inversion lists can be found in "Unicode Demystified"
6951 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
6952 * More will be coming when functionality is added later.
6954 * The inversion list data structure is currently implemented as an SV pointing
6955 * to an array of UVs that the SV thinks are bytes. This allows us to have an
6956 * array of UV whose memory management is automatically handled by the existing
6957 * facilities for SV's.
6959 * Some of the methods should always be private to the implementation, and some
6960 * should eventually be made public */
6962 /* The header definitions are in F<inline_invlist.c> */
6963 #define TO_INTERNAL_SIZE(x) (((x) + HEADER_LENGTH) * sizeof(UV))
6964 #define FROM_INTERNAL_SIZE(x) (((x)/ sizeof(UV)) - HEADER_LENGTH)
6966 #define INVLIST_INITIAL_LEN 10
6968 PERL_STATIC_INLINE UV*
6969 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
6971 /* Returns a pointer to the first element in the inversion list's array.
6972 * This is called upon initialization of an inversion list. Where the
6973 * array begins depends on whether the list has the code point U+0000
6974 * in it or not. The other parameter tells it whether the code that
6975 * follows this call is about to put a 0 in the inversion list or not.
6976 * The first element is either the element with 0, if 0, or the next one,
6979 UV* zero = get_invlist_zero_addr(invlist);
6981 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
6984 assert(! *_get_invlist_len_addr(invlist));
6986 /* 1^1 = 0; 1^0 = 1 */
6987 *zero = 1 ^ will_have_0;
6988 return zero + *zero;
6991 PERL_STATIC_INLINE UV*
6992 S_invlist_array(pTHX_ SV* const invlist)
6994 /* Returns the pointer to the inversion list's array. Every time the
6995 * length changes, this needs to be called in case malloc or realloc moved
6998 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7000 /* Must not be empty. If these fail, you probably didn't check for <len>
7001 * being non-zero before trying to get the array */
7002 assert(*_get_invlist_len_addr(invlist));
7003 assert(*get_invlist_zero_addr(invlist) == 0
7004 || *get_invlist_zero_addr(invlist) == 1);
7006 /* The array begins either at the element reserved for zero if the
7007 * list contains 0 (that element will be set to 0), or otherwise the next
7008 * element (in which case the reserved element will be set to 1). */
7009 return (UV *) (get_invlist_zero_addr(invlist)
7010 + *get_invlist_zero_addr(invlist));
7013 PERL_STATIC_INLINE void
7014 S_invlist_set_len(pTHX_ SV* const invlist, const UV len)
7016 /* Sets the current number of elements stored in the inversion list */
7018 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7020 *_get_invlist_len_addr(invlist) = len;
7022 assert(len <= SvLEN(invlist));
7024 SvCUR_set(invlist, TO_INTERNAL_SIZE(len));
7025 /* If the list contains U+0000, that element is part of the header,
7026 * and should not be counted as part of the array. It will contain
7027 * 0 in that case, and 1 otherwise. So we could flop 0=>1, 1=>0 and
7029 * SvCUR_set(invlist,
7030 * TO_INTERNAL_SIZE(len
7031 * - (*get_invlist_zero_addr(inv_list) ^ 1)));
7032 * But, this is only valid if len is not 0. The consequences of not doing
7033 * this is that the memory allocation code may think that 1 more UV is
7034 * being used than actually is, and so might do an unnecessary grow. That
7035 * seems worth not bothering to make this the precise amount.
7037 * Note that when inverting, SvCUR shouldn't change */
7040 PERL_STATIC_INLINE IV*
7041 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7043 /* Return the address of the UV that is reserved to hold the cached index
7046 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7048 return (IV *) (SvPVX(invlist) + (INVLIST_PREVIOUS_INDEX_OFFSET * sizeof (UV)));
7051 PERL_STATIC_INLINE IV
7052 S_invlist_previous_index(pTHX_ SV* const invlist)
7054 /* Returns cached index of previous search */
7056 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7058 return *get_invlist_previous_index_addr(invlist);
7061 PERL_STATIC_INLINE void
7062 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7064 /* Caches <index> for later retrieval */
7066 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7068 assert(index == 0 || index < (int) _invlist_len(invlist));
7070 *get_invlist_previous_index_addr(invlist) = index;
7073 PERL_STATIC_INLINE UV
7074 S_invlist_max(pTHX_ SV* const invlist)
7076 /* Returns the maximum number of elements storable in the inversion list's
7077 * array, without having to realloc() */
7079 PERL_ARGS_ASSERT_INVLIST_MAX;
7081 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7082 ? _invlist_len(invlist)
7083 : FROM_INTERNAL_SIZE(SvLEN(invlist));
7086 PERL_STATIC_INLINE UV*
7087 S_get_invlist_zero_addr(pTHX_ SV* invlist)
7089 /* Return the address of the UV that is reserved to hold 0 if the inversion
7090 * list contains 0. This has to be the last element of the heading, as the
7091 * list proper starts with either it if 0, or the next element if not.
7092 * (But we force it to contain either 0 or 1) */
7094 PERL_ARGS_ASSERT_GET_INVLIST_ZERO_ADDR;
7096 return (UV *) (SvPVX(invlist) + (INVLIST_ZERO_OFFSET * sizeof (UV)));
7099 #ifndef PERL_IN_XSUB_RE
7101 Perl__new_invlist(pTHX_ IV initial_size)
7104 /* Return a pointer to a newly constructed inversion list, with enough
7105 * space to store 'initial_size' elements. If that number is negative, a
7106 * system default is used instead */
7110 if (initial_size < 0) {
7111 initial_size = INVLIST_INITIAL_LEN;
7114 /* Allocate the initial space */
7115 new_list = newSV(TO_INTERNAL_SIZE(initial_size));
7116 invlist_set_len(new_list, 0);
7118 /* Force iterinit() to be used to get iteration to work */
7119 *get_invlist_iter_addr(new_list) = UV_MAX;
7121 /* This should force a segfault if a method doesn't initialize this
7123 *get_invlist_zero_addr(new_list) = UV_MAX;
7125 *get_invlist_previous_index_addr(new_list) = 0;
7126 *get_invlist_version_id_addr(new_list) = INVLIST_VERSION_ID;
7127 #if HEADER_LENGTH != 5
7128 # 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
7136 S__new_invlist_C_array(pTHX_ UV* list)
7138 /* Return a pointer to a newly constructed inversion list, initialized to
7139 * point to <list>, which has to be in the exact correct inversion list
7140 * form, including internal fields. Thus this is a dangerous routine that
7141 * should not be used in the wrong hands */
7143 SV* invlist = newSV_type(SVt_PV);
7145 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7147 SvPV_set(invlist, (char *) list);
7148 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7149 shouldn't touch it */
7150 SvCUR_set(invlist, TO_INTERNAL_SIZE(_invlist_len(invlist)));
7152 if (*get_invlist_version_id_addr(invlist) != INVLIST_VERSION_ID) {
7153 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7156 /* Initialize the iteration pointer.
7157 * XXX This could be done at compile time in charclass_invlists.h, but I
7158 * (khw) am not confident that the suffixes for specifying the C constant
7159 * UV_MAX are portable, e.g. 'ull' on a 32 bit machine that is configured
7160 * to use 64 bits; might need a Configure probe */
7161 invlist_iterfinish(invlist);
7167 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7169 /* Grow the maximum size of an inversion list */
7171 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7173 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max));
7176 PERL_STATIC_INLINE void
7177 S_invlist_trim(pTHX_ SV* const invlist)
7179 PERL_ARGS_ASSERT_INVLIST_TRIM;
7181 /* Change the length of the inversion list to how many entries it currently
7184 SvPV_shrink_to_cur((SV *) invlist);
7187 #define _invlist_union_complement_2nd(a, b, output) _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
7190 S__append_range_to_invlist(pTHX_ SV* const invlist, const UV start, const UV end)
7192 /* Subject to change or removal. Append the range from 'start' to 'end' at
7193 * the end of the inversion list. The range must be above any existing
7197 UV max = invlist_max(invlist);
7198 UV len = _invlist_len(invlist);
7200 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7202 if (len == 0) { /* Empty lists must be initialized */
7203 array = _invlist_array_init(invlist, start == 0);
7206 /* Here, the existing list is non-empty. The current max entry in the
7207 * list is generally the first value not in the set, except when the
7208 * set extends to the end of permissible values, in which case it is
7209 * the first entry in that final set, and so this call is an attempt to
7210 * append out-of-order */
7212 UV final_element = len - 1;
7213 array = invlist_array(invlist);
7214 if (array[final_element] > start
7215 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
7217 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",
7218 array[final_element], start,
7219 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
7222 /* Here, it is a legal append. If the new range begins with the first
7223 * value not in the set, it is extending the set, so the new first
7224 * value not in the set is one greater than the newly extended range.
7226 if (array[final_element] == start) {
7227 if (end != UV_MAX) {
7228 array[final_element] = end + 1;
7231 /* But if the end is the maximum representable on the machine,
7232 * just let the range that this would extend to have no end */
7233 invlist_set_len(invlist, len - 1);
7239 /* Here the new range doesn't extend any existing set. Add it */
7241 len += 2; /* Includes an element each for the start and end of range */
7243 /* If overflows the existing space, extend, which may cause the array to be
7246 invlist_extend(invlist, len);
7247 invlist_set_len(invlist, len); /* Have to set len here to avoid assert
7248 failure in invlist_array() */
7249 array = invlist_array(invlist);
7252 invlist_set_len(invlist, len);
7255 /* The next item on the list starts the range, the one after that is
7256 * one past the new range. */
7257 array[len - 2] = start;
7258 if (end != UV_MAX) {
7259 array[len - 1] = end + 1;
7262 /* But if the end is the maximum representable on the machine, just let
7263 * the range have no end */
7264 invlist_set_len(invlist, len - 1);
7268 #ifndef PERL_IN_XSUB_RE
7271 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
7273 /* Searches the inversion list for the entry that contains the input code
7274 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
7275 * return value is the index into the list's array of the range that
7280 IV high = _invlist_len(invlist);
7281 const IV highest_element = high - 1;
7284 PERL_ARGS_ASSERT__INVLIST_SEARCH;
7286 /* If list is empty, return failure. */
7291 /* (We can't get the array unless we know the list is non-empty) */
7292 array = invlist_array(invlist);
7294 mid = invlist_previous_index(invlist);
7295 assert(mid >=0 && mid <= highest_element);
7297 /* <mid> contains the cache of the result of the previous call to this
7298 * function (0 the first time). See if this call is for the same result,
7299 * or if it is for mid-1. This is under the theory that calls to this
7300 * function will often be for related code points that are near each other.
7301 * And benchmarks show that caching gives better results. We also test
7302 * here if the code point is within the bounds of the list. These tests
7303 * replace others that would have had to be made anyway to make sure that
7304 * the array bounds were not exceeded, and these give us extra information
7305 * at the same time */
7306 if (cp >= array[mid]) {
7307 if (cp >= array[highest_element]) {
7308 return highest_element;
7311 /* Here, array[mid] <= cp < array[highest_element]. This means that
7312 * the final element is not the answer, so can exclude it; it also
7313 * means that <mid> is not the final element, so can refer to 'mid + 1'
7315 if (cp < array[mid + 1]) {
7321 else { /* cp < aray[mid] */
7322 if (cp < array[0]) { /* Fail if outside the array */
7326 if (cp >= array[mid - 1]) {
7331 /* Binary search. What we are looking for is <i> such that
7332 * array[i] <= cp < array[i+1]
7333 * The loop below converges on the i+1. Note that there may not be an
7334 * (i+1)th element in the array, and things work nonetheless */
7335 while (low < high) {
7336 mid = (low + high) / 2;
7337 assert(mid <= highest_element);
7338 if (array[mid] <= cp) { /* cp >= array[mid] */
7341 /* We could do this extra test to exit the loop early.
7342 if (cp < array[low]) {
7347 else { /* cp < array[mid] */
7354 invlist_set_previous_index(invlist, high);
7359 Perl__invlist_populate_swatch(pTHX_ SV* const invlist, const UV start, const UV end, U8* swatch)
7361 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
7362 * but is used when the swash has an inversion list. This makes this much
7363 * faster, as it uses a binary search instead of a linear one. This is
7364 * intimately tied to that function, and perhaps should be in utf8.c,
7365 * except it is intimately tied to inversion lists as well. It assumes
7366 * that <swatch> is all 0's on input */
7369 const IV len = _invlist_len(invlist);
7373 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
7375 if (len == 0) { /* Empty inversion list */
7379 array = invlist_array(invlist);
7381 /* Find which element it is */
7382 i = _invlist_search(invlist, start);
7384 /* We populate from <start> to <end> */
7385 while (current < end) {
7388 /* The inversion list gives the results for every possible code point
7389 * after the first one in the list. Only those ranges whose index is
7390 * even are ones that the inversion list matches. For the odd ones,
7391 * and if the initial code point is not in the list, we have to skip
7392 * forward to the next element */
7393 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
7395 if (i >= len) { /* Finished if beyond the end of the array */
7399 if (current >= end) { /* Finished if beyond the end of what we
7401 if (LIKELY(end < UV_MAX)) {
7405 /* We get here when the upper bound is the maximum
7406 * representable on the machine, and we are looking for just
7407 * that code point. Have to special case it */
7409 goto join_end_of_list;
7412 assert(current >= start);
7414 /* The current range ends one below the next one, except don't go past
7417 upper = (i < len && array[i] < end) ? array[i] : end;
7419 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
7420 * for each code point in it */
7421 for (; current < upper; current++) {
7422 const STRLEN offset = (STRLEN)(current - start);
7423 swatch[offset >> 3] |= 1 << (offset & 7);
7428 /* Quit if at the end of the list */
7431 /* But first, have to deal with the highest possible code point on
7432 * the platform. The previous code assumes that <end> is one
7433 * beyond where we want to populate, but that is impossible at the
7434 * platform's infinity, so have to handle it specially */
7435 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
7437 const STRLEN offset = (STRLEN)(end - start);
7438 swatch[offset >> 3] |= 1 << (offset & 7);
7443 /* Advance to the next range, which will be for code points not in the
7452 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** output)
7454 /* Take the union of two inversion lists and point <output> to it. *output
7455 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7456 * the reference count to that list will be decremented. The first list,
7457 * <a>, may be NULL, in which case a copy of the second list is returned.
7458 * If <complement_b> is TRUE, the union is taken of the complement
7459 * (inversion) of <b> instead of b itself.
7461 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7462 * Richard Gillam, published by Addison-Wesley, and explained at some
7463 * length there. The preface says to incorporate its examples into your
7464 * code at your own risk.
7466 * The algorithm is like a merge sort.
7468 * XXX A potential performance improvement is to keep track as we go along
7469 * if only one of the inputs contributes to the result, meaning the other
7470 * is a subset of that one. In that case, we can skip the final copy and
7471 * return the larger of the input lists, but then outside code might need
7472 * to keep track of whether to free the input list or not */
7474 UV* array_a; /* a's array */
7476 UV len_a; /* length of a's array */
7479 SV* u; /* the resulting union */
7483 UV i_a = 0; /* current index into a's array */
7487 /* running count, as explained in the algorithm source book; items are
7488 * stopped accumulating and are output when the count changes to/from 0.
7489 * The count is incremented when we start a range that's in the set, and
7490 * decremented when we start a range that's not in the set. So its range
7491 * is 0 to 2. Only when the count is zero is something not in the set.
7495 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
7498 /* If either one is empty, the union is the other one */
7499 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
7506 *output = invlist_clone(b);
7508 _invlist_invert(*output);
7510 } /* else *output already = b; */
7513 else if ((len_b = _invlist_len(b)) == 0) {
7518 /* The complement of an empty list is a list that has everything in it,
7519 * so the union with <a> includes everything too */
7524 *output = _new_invlist(1);
7525 _append_range_to_invlist(*output, 0, UV_MAX);
7527 else if (*output != a) {
7528 *output = invlist_clone(a);
7530 /* else *output already = a; */
7534 /* Here both lists exist and are non-empty */
7535 array_a = invlist_array(a);
7536 array_b = invlist_array(b);
7538 /* If are to take the union of 'a' with the complement of b, set it
7539 * up so are looking at b's complement. */
7542 /* To complement, we invert: if the first element is 0, remove it. To
7543 * do this, we just pretend the array starts one later, and clear the
7544 * flag as we don't have to do anything else later */
7545 if (array_b[0] == 0) {
7548 complement_b = FALSE;
7552 /* But if the first element is not zero, we unshift a 0 before the
7553 * array. The data structure reserves a space for that 0 (which
7554 * should be a '1' right now), so physical shifting is unneeded,
7555 * but temporarily change that element to 0. Before exiting the
7556 * routine, we must restore the element to '1' */
7563 /* Size the union for the worst case: that the sets are completely
7565 u = _new_invlist(len_a + len_b);
7567 /* Will contain U+0000 if either component does */
7568 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
7569 || (len_b > 0 && array_b[0] == 0));
7571 /* Go through each list item by item, stopping when exhausted one of
7573 while (i_a < len_a && i_b < len_b) {
7574 UV cp; /* The element to potentially add to the union's array */
7575 bool cp_in_set; /* is it in the the input list's set or not */
7577 /* We need to take one or the other of the two inputs for the union.
7578 * Since we are merging two sorted lists, we take the smaller of the
7579 * next items. In case of a tie, we take the one that is in its set
7580 * first. If we took one not in the set first, it would decrement the
7581 * count, possibly to 0 which would cause it to be output as ending the
7582 * range, and the next time through we would take the same number, and
7583 * output it again as beginning the next range. By doing it the
7584 * opposite way, there is no possibility that the count will be
7585 * momentarily decremented to 0, and thus the two adjoining ranges will
7586 * be seamlessly merged. (In a tie and both are in the set or both not
7587 * in the set, it doesn't matter which we take first.) */
7588 if (array_a[i_a] < array_b[i_b]
7589 || (array_a[i_a] == array_b[i_b]
7590 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7592 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7596 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7597 cp = array_b[i_b++];
7600 /* Here, have chosen which of the two inputs to look at. Only output
7601 * if the running count changes to/from 0, which marks the
7602 * beginning/end of a range in that's in the set */
7605 array_u[i_u++] = cp;
7612 array_u[i_u++] = cp;
7617 /* Here, we are finished going through at least one of the lists, which
7618 * means there is something remaining in at most one. We check if the list
7619 * that hasn't been exhausted is positioned such that we are in the middle
7620 * of a range in its set or not. (i_a and i_b point to the element beyond
7621 * the one we care about.) If in the set, we decrement 'count'; if 0, there
7622 * is potentially more to output.
7623 * There are four cases:
7624 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
7625 * in the union is entirely from the non-exhausted set.
7626 * 2) Both were in their sets, count is 2. Nothing further should
7627 * be output, as everything that remains will be in the exhausted
7628 * list's set, hence in the union; decrementing to 1 but not 0 insures
7630 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
7631 * Nothing further should be output because the union includes
7632 * everything from the exhausted set. Not decrementing ensures that.
7633 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
7634 * decrementing to 0 insures that we look at the remainder of the
7635 * non-exhausted set */
7636 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7637 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7642 /* The final length is what we've output so far, plus what else is about to
7643 * be output. (If 'count' is non-zero, then the input list we exhausted
7644 * has everything remaining up to the machine's limit in its set, and hence
7645 * in the union, so there will be no further output. */
7648 /* At most one of the subexpressions will be non-zero */
7649 len_u += (len_a - i_a) + (len_b - i_b);
7652 /* Set result to final length, which can change the pointer to array_u, so
7654 if (len_u != _invlist_len(u)) {
7655 invlist_set_len(u, len_u);
7657 array_u = invlist_array(u);
7660 /* When 'count' is 0, the list that was exhausted (if one was shorter than
7661 * the other) ended with everything above it not in its set. That means
7662 * that the remaining part of the union is precisely the same as the
7663 * non-exhausted list, so can just copy it unchanged. (If both list were
7664 * exhausted at the same time, then the operations below will be both 0.)
7667 IV copy_count; /* At most one will have a non-zero copy count */
7668 if ((copy_count = len_a - i_a) > 0) {
7669 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
7671 else if ((copy_count = len_b - i_b) > 0) {
7672 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
7676 /* If we've changed b, restore it */
7681 /* We may be removing a reference to one of the inputs */
7682 if (a == *output || b == *output) {
7683 assert(! invlist_is_iterating(*output));
7684 SvREFCNT_dec_NN(*output);
7692 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** i)
7694 /* Take the intersection of two inversion lists and point <i> to it. *i
7695 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7696 * the reference count to that list will be decremented.
7697 * If <complement_b> is TRUE, the result will be the intersection of <a>
7698 * and the complement (or inversion) of <b> instead of <b> directly.
7700 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7701 * Richard Gillam, published by Addison-Wesley, and explained at some
7702 * length there. The preface says to incorporate its examples into your
7703 * code at your own risk. In fact, it had bugs
7705 * The algorithm is like a merge sort, and is essentially the same as the
7709 UV* array_a; /* a's array */
7711 UV len_a; /* length of a's array */
7714 SV* r; /* the resulting intersection */
7718 UV i_a = 0; /* current index into a's array */
7722 /* running count, as explained in the algorithm source book; items are
7723 * stopped accumulating and are output when the count changes to/from 2.
7724 * The count is incremented when we start a range that's in the set, and
7725 * decremented when we start a range that's not in the set. So its range
7726 * is 0 to 2. Only when the count is 2 is something in the intersection.
7730 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
7733 /* Special case if either one is empty */
7734 len_a = _invlist_len(a);
7735 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
7737 if (len_a != 0 && complement_b) {
7739 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
7740 * be empty. Here, also we are using 'b's complement, which hence
7741 * must be every possible code point. Thus the intersection is
7744 *i = invlist_clone(a);
7750 /* else *i is already 'a' */
7754 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
7755 * intersection must be empty */
7762 *i = _new_invlist(0);
7766 /* Here both lists exist and are non-empty */
7767 array_a = invlist_array(a);
7768 array_b = invlist_array(b);
7770 /* If are to take the intersection of 'a' with the complement of b, set it
7771 * up so are looking at b's complement. */
7774 /* To complement, we invert: if the first element is 0, remove it. To
7775 * do this, we just pretend the array starts one later, and clear the
7776 * flag as we don't have to do anything else later */
7777 if (array_b[0] == 0) {
7780 complement_b = FALSE;
7784 /* But if the first element is not zero, we unshift a 0 before the
7785 * array. The data structure reserves a space for that 0 (which
7786 * should be a '1' right now), so physical shifting is unneeded,
7787 * but temporarily change that element to 0. Before exiting the
7788 * routine, we must restore the element to '1' */
7795 /* Size the intersection for the worst case: that the intersection ends up
7796 * fragmenting everything to be completely disjoint */
7797 r= _new_invlist(len_a + len_b);
7799 /* Will contain U+0000 iff both components do */
7800 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
7801 && len_b > 0 && array_b[0] == 0);
7803 /* Go through each list item by item, stopping when exhausted one of
7805 while (i_a < len_a && i_b < len_b) {
7806 UV cp; /* The element to potentially add to the intersection's
7808 bool cp_in_set; /* Is it in the input list's set or not */
7810 /* We need to take one or the other of the two inputs for the
7811 * intersection. Since we are merging two sorted lists, we take the
7812 * smaller of the next items. In case of a tie, we take the one that
7813 * is not in its set first (a difference from the union algorithm). If
7814 * we took one in the set first, it would increment the count, possibly
7815 * to 2 which would cause it to be output as starting a range in the
7816 * intersection, and the next time through we would take that same
7817 * number, and output it again as ending the set. By doing it the
7818 * opposite of this, there is no possibility that the count will be
7819 * momentarily incremented to 2. (In a tie and both are in the set or
7820 * both not in the set, it doesn't matter which we take first.) */
7821 if (array_a[i_a] < array_b[i_b]
7822 || (array_a[i_a] == array_b[i_b]
7823 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7825 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7829 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7833 /* Here, have chosen which of the two inputs to look at. Only output
7834 * if the running count changes to/from 2, which marks the
7835 * beginning/end of a range that's in the intersection */
7839 array_r[i_r++] = cp;
7844 array_r[i_r++] = cp;
7850 /* Here, we are finished going through at least one of the lists, which
7851 * means there is something remaining in at most one. We check if the list
7852 * that has been exhausted is positioned such that we are in the middle
7853 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
7854 * the ones we care about.) There are four cases:
7855 * 1) Both weren't in their sets, count is 0, and remains 0. There's
7856 * nothing left in the intersection.
7857 * 2) Both were in their sets, count is 2 and perhaps is incremented to
7858 * above 2. What should be output is exactly that which is in the
7859 * non-exhausted set, as everything it has is also in the intersection
7860 * set, and everything it doesn't have can't be in the intersection
7861 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
7862 * gets incremented to 2. Like the previous case, the intersection is
7863 * everything that remains in the non-exhausted set.
7864 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
7865 * remains 1. And the intersection has nothing more. */
7866 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7867 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7872 /* The final length is what we've output so far plus what else is in the
7873 * intersection. At most one of the subexpressions below will be non-zero */
7876 len_r += (len_a - i_a) + (len_b - i_b);
7879 /* Set result to final length, which can change the pointer to array_r, so
7881 if (len_r != _invlist_len(r)) {
7882 invlist_set_len(r, len_r);
7884 array_r = invlist_array(r);
7887 /* Finish outputting any remaining */
7888 if (count >= 2) { /* At most one will have a non-zero copy count */
7890 if ((copy_count = len_a - i_a) > 0) {
7891 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
7893 else if ((copy_count = len_b - i_b) > 0) {
7894 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
7898 /* If we've changed b, restore it */
7903 /* We may be removing a reference to one of the inputs */
7904 if (a == *i || b == *i) {
7905 assert(! invlist_is_iterating(*i));
7906 SvREFCNT_dec_NN(*i);
7914 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
7916 /* Add the range from 'start' to 'end' inclusive to the inversion list's
7917 * set. A pointer to the inversion list is returned. This may actually be
7918 * a new list, in which case the passed in one has been destroyed. The
7919 * passed in inversion list can be NULL, in which case a new one is created
7920 * with just the one range in it */
7925 if (invlist == NULL) {
7926 invlist = _new_invlist(2);
7930 len = _invlist_len(invlist);
7933 /* If comes after the final entry actually in the list, can just append it
7936 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
7937 && start >= invlist_array(invlist)[len - 1]))
7939 _append_range_to_invlist(invlist, start, end);
7943 /* Here, can't just append things, create and return a new inversion list
7944 * which is the union of this range and the existing inversion list */
7945 range_invlist = _new_invlist(2);
7946 _append_range_to_invlist(range_invlist, start, end);
7948 _invlist_union(invlist, range_invlist, &invlist);
7950 /* The temporary can be freed */
7951 SvREFCNT_dec_NN(range_invlist);
7958 PERL_STATIC_INLINE SV*
7959 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
7960 return _add_range_to_invlist(invlist, cp, cp);
7963 #ifndef PERL_IN_XSUB_RE
7965 Perl__invlist_invert(pTHX_ SV* const invlist)
7967 /* Complement the input inversion list. This adds a 0 if the list didn't
7968 * have a zero; removes it otherwise. As described above, the data
7969 * structure is set up so that this is very efficient */
7971 UV* len_pos = _get_invlist_len_addr(invlist);
7973 PERL_ARGS_ASSERT__INVLIST_INVERT;
7975 assert(! invlist_is_iterating(invlist));
7977 /* The inverse of matching nothing is matching everything */
7978 if (*len_pos == 0) {
7979 _append_range_to_invlist(invlist, 0, UV_MAX);
7983 /* The exclusive or complents 0 to 1; and 1 to 0. If the result is 1, the
7984 * zero element was a 0, so it is being removed, so the length decrements
7985 * by 1; and vice-versa. SvCUR is unaffected */
7986 if (*get_invlist_zero_addr(invlist) ^= 1) {
7995 Perl__invlist_invert_prop(pTHX_ SV* const invlist)
7997 /* Complement the input inversion list (which must be a Unicode property,
7998 * all of which don't match above the Unicode maximum code point.) And
7999 * Perl has chosen to not have the inversion match above that either. This
8000 * adds a 0x110000 if the list didn't end with it, and removes it if it did
8006 PERL_ARGS_ASSERT__INVLIST_INVERT_PROP;
8008 _invlist_invert(invlist);
8010 len = _invlist_len(invlist);
8012 if (len != 0) { /* If empty do nothing */
8013 array = invlist_array(invlist);
8014 if (array[len - 1] != PERL_UNICODE_MAX + 1) {
8015 /* Add 0x110000. First, grow if necessary */
8017 if (invlist_max(invlist) < len) {
8018 invlist_extend(invlist, len);
8019 array = invlist_array(invlist);
8021 invlist_set_len(invlist, len);
8022 array[len - 1] = PERL_UNICODE_MAX + 1;
8024 else { /* Remove the 0x110000 */
8025 invlist_set_len(invlist, len - 1);
8033 PERL_STATIC_INLINE SV*
8034 S_invlist_clone(pTHX_ SV* const invlist)
8037 /* Return a new inversion list that is a copy of the input one, which is
8040 /* Need to allocate extra space to accommodate Perl's addition of a
8041 * trailing NUL to SvPV's, since it thinks they are always strings */
8042 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8043 STRLEN length = SvCUR(invlist);
8045 PERL_ARGS_ASSERT_INVLIST_CLONE;
8047 SvCUR_set(new_invlist, length); /* This isn't done automatically */
8048 Copy(SvPVX(invlist), SvPVX(new_invlist), length, char);
8053 PERL_STATIC_INLINE UV*
8054 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8056 /* Return the address of the UV that contains the current iteration
8059 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8061 return (UV *) (SvPVX(invlist) + (INVLIST_ITER_OFFSET * sizeof (UV)));
8064 PERL_STATIC_INLINE UV*
8065 S_get_invlist_version_id_addr(pTHX_ SV* invlist)
8067 /* Return the address of the UV that contains the version id. */
8069 PERL_ARGS_ASSERT_GET_INVLIST_VERSION_ID_ADDR;
8071 return (UV *) (SvPVX(invlist) + (INVLIST_VERSION_ID_OFFSET * sizeof (UV)));
8074 PERL_STATIC_INLINE void
8075 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8077 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8079 *get_invlist_iter_addr(invlist) = 0;
8082 PERL_STATIC_INLINE void
8083 S_invlist_iterfinish(pTHX_ SV* invlist)
8085 /* Terminate iterator for invlist. This is to catch development errors.
8086 * Any iteration that is interrupted before completed should call this
8087 * function. Functions that add code points anywhere else but to the end
8088 * of an inversion list assert that they are not in the middle of an
8089 * iteration. If they were, the addition would make the iteration
8090 * problematical: if the iteration hadn't reached the place where things
8091 * were being added, it would be ok */
8093 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8095 *get_invlist_iter_addr(invlist) = UV_MAX;
8099 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8101 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8102 * This call sets in <*start> and <*end>, the next range in <invlist>.
8103 * Returns <TRUE> if successful and the next call will return the next
8104 * range; <FALSE> if was already at the end of the list. If the latter,
8105 * <*start> and <*end> are unchanged, and the next call to this function
8106 * will start over at the beginning of the list */
8108 UV* pos = get_invlist_iter_addr(invlist);
8109 UV len = _invlist_len(invlist);
8112 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8115 *pos = UV_MAX; /* Force iterinit() to be required next time */
8119 array = invlist_array(invlist);
8121 *start = array[(*pos)++];
8127 *end = array[(*pos)++] - 1;
8133 PERL_STATIC_INLINE bool
8134 S_invlist_is_iterating(pTHX_ SV* const invlist)
8136 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8138 return *(get_invlist_iter_addr(invlist)) < UV_MAX;
8141 PERL_STATIC_INLINE UV
8142 S_invlist_highest(pTHX_ SV* const invlist)
8144 /* Returns the highest code point that matches an inversion list. This API
8145 * has an ambiguity, as it returns 0 under either the highest is actually
8146 * 0, or if the list is empty. If this distinction matters to you, check
8147 * for emptiness before calling this function */
8149 UV len = _invlist_len(invlist);
8152 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8158 array = invlist_array(invlist);
8160 /* The last element in the array in the inversion list always starts a
8161 * range that goes to infinity. That range may be for code points that are
8162 * matched in the inversion list, or it may be for ones that aren't
8163 * matched. In the latter case, the highest code point in the set is one
8164 * less than the beginning of this range; otherwise it is the final element
8165 * of this range: infinity */
8166 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8168 : array[len - 1] - 1;
8171 #ifndef PERL_IN_XSUB_RE
8173 Perl__invlist_contents(pTHX_ SV* const invlist)
8175 /* Get the contents of an inversion list into a string SV so that they can
8176 * be printed out. It uses the format traditionally done for debug tracing
8180 SV* output = newSVpvs("\n");
8182 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8184 assert(! invlist_is_iterating(invlist));
8186 invlist_iterinit(invlist);
8187 while (invlist_iternext(invlist, &start, &end)) {
8188 if (end == UV_MAX) {
8189 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8191 else if (end != start) {
8192 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8196 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
8204 #ifdef PERL_ARGS_ASSERT__INVLIST_DUMP
8206 Perl__invlist_dump(pTHX_ SV* const invlist, const char * const header)
8208 /* Dumps out the ranges in an inversion list. The string 'header'
8209 * if present is output on a line before the first range */
8213 PERL_ARGS_ASSERT__INVLIST_DUMP;
8215 if (header && strlen(header)) {
8216 PerlIO_printf(Perl_debug_log, "%s\n", header);
8218 if (invlist_is_iterating(invlist)) {
8219 PerlIO_printf(Perl_debug_log, "Can't dump because is in middle of iterating\n");
8223 invlist_iterinit(invlist);
8224 while (invlist_iternext(invlist, &start, &end)) {
8225 if (end == UV_MAX) {
8226 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. INFINITY\n", start);
8228 else if (end != start) {
8229 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. 0x%04"UVXf"\n",
8233 PerlIO_printf(Perl_debug_log, "0x%04"UVXf"\n", start);
8241 S__invlistEQ(pTHX_ SV* const a, SV* const b, bool complement_b)
8243 /* Return a boolean as to if the two passed in inversion lists are
8244 * identical. The final argument, if TRUE, says to take the complement of
8245 * the second inversion list before doing the comparison */
8247 UV* array_a = invlist_array(a);
8248 UV* array_b = invlist_array(b);
8249 UV len_a = _invlist_len(a);
8250 UV len_b = _invlist_len(b);
8252 UV i = 0; /* current index into the arrays */
8253 bool retval = TRUE; /* Assume are identical until proven otherwise */
8255 PERL_ARGS_ASSERT__INVLISTEQ;
8257 /* If are to compare 'a' with the complement of b, set it
8258 * up so are looking at b's complement. */
8261 /* The complement of nothing is everything, so <a> would have to have
8262 * just one element, starting at zero (ending at infinity) */
8264 return (len_a == 1 && array_a[0] == 0);
8266 else if (array_b[0] == 0) {
8268 /* Otherwise, to complement, we invert. Here, the first element is
8269 * 0, just remove it. To do this, we just pretend the array starts
8270 * one later, and clear the flag as we don't have to do anything
8275 complement_b = FALSE;
8279 /* But if the first element is not zero, we unshift a 0 before the
8280 * array. The data structure reserves a space for that 0 (which
8281 * should be a '1' right now), so physical shifting is unneeded,
8282 * but temporarily change that element to 0. Before exiting the
8283 * routine, we must restore the element to '1' */
8290 /* Make sure that the lengths are the same, as well as the final element
8291 * before looping through the remainder. (Thus we test the length, final,
8292 * and first elements right off the bat) */
8293 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
8296 else for (i = 0; i < len_a - 1; i++) {
8297 if (array_a[i] != array_b[i]) {
8310 #undef HEADER_LENGTH
8311 #undef INVLIST_INITIAL_LENGTH
8312 #undef TO_INTERNAL_SIZE
8313 #undef FROM_INTERNAL_SIZE
8314 #undef INVLIST_LEN_OFFSET
8315 #undef INVLIST_ZERO_OFFSET
8316 #undef INVLIST_ITER_OFFSET
8317 #undef INVLIST_VERSION_ID
8318 #undef INVLIST_PREVIOUS_INDEX_OFFSET
8320 /* End of inversion list object */
8323 S_parse_lparen_question_flags(pTHX_ struct RExC_state_t *pRExC_state)
8325 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
8326 * constructs, and updates RExC_flags with them. On input, RExC_parse
8327 * should point to the first flag; it is updated on output to point to the
8328 * final ')' or ':'. There needs to be at least one flag, or this will
8331 /* for (?g), (?gc), and (?o) warnings; warning
8332 about (?c) will warn about (?g) -- japhy */
8334 #define WASTED_O 0x01
8335 #define WASTED_G 0x02
8336 #define WASTED_C 0x04
8337 #define WASTED_GC (0x02|0x04)
8338 I32 wastedflags = 0x00;
8339 U32 posflags = 0, negflags = 0;
8340 U32 *flagsp = &posflags;
8341 char has_charset_modifier = '\0';
8343 bool has_use_defaults = FALSE;
8344 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
8346 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
8348 /* '^' as an initial flag sets certain defaults */
8349 if (UCHARAT(RExC_parse) == '^') {
8351 has_use_defaults = TRUE;
8352 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
8353 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
8354 ? REGEX_UNICODE_CHARSET
8355 : REGEX_DEPENDS_CHARSET);
8358 cs = get_regex_charset(RExC_flags);
8359 if (cs == REGEX_DEPENDS_CHARSET
8360 && (RExC_utf8 || RExC_uni_semantics))
8362 cs = REGEX_UNICODE_CHARSET;
8365 while (*RExC_parse) {
8366 /* && strchr("iogcmsx", *RExC_parse) */
8367 /* (?g), (?gc) and (?o) are useless here
8368 and must be globally applied -- japhy */
8369 switch (*RExC_parse) {
8371 /* Code for the imsx flags */
8372 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
8374 case LOCALE_PAT_MOD:
8375 if (has_charset_modifier) {
8376 goto excess_modifier;
8378 else if (flagsp == &negflags) {
8381 cs = REGEX_LOCALE_CHARSET;
8382 has_charset_modifier = LOCALE_PAT_MOD;
8383 RExC_contains_locale = 1;
8385 case UNICODE_PAT_MOD:
8386 if (has_charset_modifier) {
8387 goto excess_modifier;
8389 else if (flagsp == &negflags) {
8392 cs = REGEX_UNICODE_CHARSET;
8393 has_charset_modifier = UNICODE_PAT_MOD;
8395 case ASCII_RESTRICT_PAT_MOD:
8396 if (flagsp == &negflags) {
8399 if (has_charset_modifier) {
8400 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
8401 goto excess_modifier;
8403 /* Doubled modifier implies more restricted */
8404 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
8407 cs = REGEX_ASCII_RESTRICTED_CHARSET;
8409 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
8411 case DEPENDS_PAT_MOD:
8412 if (has_use_defaults) {
8413 goto fail_modifiers;
8415 else if (flagsp == &negflags) {
8418 else if (has_charset_modifier) {
8419 goto excess_modifier;
8422 /* The dual charset means unicode semantics if the
8423 * pattern (or target, not known until runtime) are
8424 * utf8, or something in the pattern indicates unicode
8426 cs = (RExC_utf8 || RExC_uni_semantics)
8427 ? REGEX_UNICODE_CHARSET
8428 : REGEX_DEPENDS_CHARSET;
8429 has_charset_modifier = DEPENDS_PAT_MOD;
8433 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
8434 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
8436 else if (has_charset_modifier == *(RExC_parse - 1)) {
8437 vFAIL2("Regexp modifier \"%c\" may not appear twice", *(RExC_parse - 1));
8440 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
8445 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"", *(RExC_parse - 1));
8447 case ONCE_PAT_MOD: /* 'o' */
8448 case GLOBAL_PAT_MOD: /* 'g' */
8449 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8450 const I32 wflagbit = *RExC_parse == 'o' ? WASTED_O : WASTED_G;
8451 if (! (wastedflags & wflagbit) ) {
8452 wastedflags |= wflagbit;
8455 "Useless (%s%c) - %suse /%c modifier",
8456 flagsp == &negflags ? "?-" : "?",
8458 flagsp == &negflags ? "don't " : "",
8465 case CONTINUE_PAT_MOD: /* 'c' */
8466 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8467 if (! (wastedflags & WASTED_C) ) {
8468 wastedflags |= WASTED_GC;
8471 "Useless (%sc) - %suse /gc modifier",
8472 flagsp == &negflags ? "?-" : "?",
8473 flagsp == &negflags ? "don't " : ""
8478 case KEEPCOPY_PAT_MOD: /* 'p' */
8479 if (flagsp == &negflags) {
8481 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
8483 *flagsp |= RXf_PMf_KEEPCOPY;
8487 /* A flag is a default iff it is following a minus, so
8488 * if there is a minus, it means will be trying to
8489 * re-specify a default which is an error */
8490 if (has_use_defaults || flagsp == &negflags) {
8491 goto fail_modifiers;
8494 wastedflags = 0; /* reset so (?g-c) warns twice */
8498 RExC_flags |= posflags;
8499 RExC_flags &= ~negflags;
8500 set_regex_charset(&RExC_flags, cs);
8506 vFAIL3("Sequence (%.*s...) not recognized",
8507 RExC_parse-seqstart, seqstart);
8516 - reg - regular expression, i.e. main body or parenthesized thing
8518 * Caller must absorb opening parenthesis.
8520 * Combining parenthesis handling with the base level of regular expression
8521 * is a trifle forced, but the need to tie the tails of the branches to what
8522 * follows makes it hard to avoid.
8524 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
8526 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
8528 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
8531 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
8532 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
8533 needs to be restarted.
8534 Otherwise would only return NULL if regbranch() returns NULL, which
8537 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
8538 /* paren: Parenthesized? 0=top, 1=(, inside: changed to letter. */
8541 regnode *ret; /* Will be the head of the group. */
8544 regnode *ender = NULL;
8547 U32 oregflags = RExC_flags;
8548 bool have_branch = 0;
8550 I32 freeze_paren = 0;
8551 I32 after_freeze = 0;
8553 char * parse_start = RExC_parse; /* MJD */
8554 char * const oregcomp_parse = RExC_parse;
8556 GET_RE_DEBUG_FLAGS_DECL;
8558 PERL_ARGS_ASSERT_REG;
8559 DEBUG_PARSE("reg ");
8561 *flagp = 0; /* Tentatively. */
8564 /* Make an OPEN node, if parenthesized. */
8566 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
8567 char *start_verb = RExC_parse;
8568 STRLEN verb_len = 0;
8569 char *start_arg = NULL;
8570 unsigned char op = 0;
8572 int internal_argval = 0; /* internal_argval is only useful if !argok */
8573 while ( *RExC_parse && *RExC_parse != ')' ) {
8574 if ( *RExC_parse == ':' ) {
8575 start_arg = RExC_parse + 1;
8581 verb_len = RExC_parse - start_verb;
8584 while ( *RExC_parse && *RExC_parse != ')' )
8586 if ( *RExC_parse != ')' )
8587 vFAIL("Unterminated verb pattern argument");
8588 if ( RExC_parse == start_arg )
8591 if ( *RExC_parse != ')' )
8592 vFAIL("Unterminated verb pattern");
8595 switch ( *start_verb ) {
8596 case 'A': /* (*ACCEPT) */
8597 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
8599 internal_argval = RExC_nestroot;
8602 case 'C': /* (*COMMIT) */
8603 if ( memEQs(start_verb,verb_len,"COMMIT") )
8606 case 'F': /* (*FAIL) */
8607 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
8612 case ':': /* (*:NAME) */
8613 case 'M': /* (*MARK:NAME) */
8614 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
8619 case 'P': /* (*PRUNE) */
8620 if ( memEQs(start_verb,verb_len,"PRUNE") )
8623 case 'S': /* (*SKIP) */
8624 if ( memEQs(start_verb,verb_len,"SKIP") )
8627 case 'T': /* (*THEN) */
8628 /* [19:06] <TimToady> :: is then */
8629 if ( memEQs(start_verb,verb_len,"THEN") ) {
8631 RExC_seen |= REG_SEEN_CUTGROUP;
8637 vFAIL3("Unknown verb pattern '%.*s'",
8638 verb_len, start_verb);
8641 if ( start_arg && internal_argval ) {
8642 vFAIL3("Verb pattern '%.*s' may not have an argument",
8643 verb_len, start_verb);
8644 } else if ( argok < 0 && !start_arg ) {
8645 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
8646 verb_len, start_verb);
8648 ret = reganode(pRExC_state, op, internal_argval);
8649 if ( ! internal_argval && ! SIZE_ONLY ) {
8651 SV *sv = newSVpvn( start_arg, RExC_parse - start_arg);
8652 ARG(ret) = add_data( pRExC_state, 1, "S" );
8653 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
8660 if (!internal_argval)
8661 RExC_seen |= REG_SEEN_VERBARG;
8662 } else if ( start_arg ) {
8663 vFAIL3("Verb pattern '%.*s' may not have an argument",
8664 verb_len, start_verb);
8666 ret = reg_node(pRExC_state, op);
8668 nextchar(pRExC_state);
8671 if (*RExC_parse == '?') { /* (?...) */
8672 bool is_logical = 0;
8673 const char * const seqstart = RExC_parse;
8676 paren = *RExC_parse++;
8677 ret = NULL; /* For look-ahead/behind. */
8680 case 'P': /* (?P...) variants for those used to PCRE/Python */
8681 paren = *RExC_parse++;
8682 if ( paren == '<') /* (?P<...>) named capture */
8684 else if (paren == '>') { /* (?P>name) named recursion */
8685 goto named_recursion;
8687 else if (paren == '=') { /* (?P=...) named backref */
8688 /* this pretty much dupes the code for \k<NAME> in regatom(), if
8689 you change this make sure you change that */
8690 char* name_start = RExC_parse;
8692 SV *sv_dat = reg_scan_name(pRExC_state,
8693 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8694 if (RExC_parse == name_start || *RExC_parse != ')')
8695 vFAIL2("Sequence %.3s... not terminated",parse_start);
8698 num = add_data( pRExC_state, 1, "S" );
8699 RExC_rxi->data->data[num]=(void*)sv_dat;
8700 SvREFCNT_inc_simple_void(sv_dat);
8703 ret = reganode(pRExC_state,
8706 : (ASCII_FOLD_RESTRICTED)
8708 : (AT_LEAST_UNI_SEMANTICS)
8716 Set_Node_Offset(ret, parse_start+1);
8717 Set_Node_Cur_Length(ret); /* MJD */
8719 nextchar(pRExC_state);
8723 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8725 case '<': /* (?<...) */
8726 if (*RExC_parse == '!')
8728 else if (*RExC_parse != '=')
8734 case '\'': /* (?'...') */
8735 name_start= RExC_parse;
8736 svname = reg_scan_name(pRExC_state,
8737 SIZE_ONLY ? /* reverse test from the others */
8738 REG_RSN_RETURN_NAME :
8739 REG_RSN_RETURN_NULL);
8740 if (RExC_parse == name_start) {
8742 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8745 if (*RExC_parse != paren)
8746 vFAIL2("Sequence (?%c... not terminated",
8747 paren=='>' ? '<' : paren);
8751 if (!svname) /* shouldn't happen */
8753 "panic: reg_scan_name returned NULL");
8754 if (!RExC_paren_names) {
8755 RExC_paren_names= newHV();
8756 sv_2mortal(MUTABLE_SV(RExC_paren_names));
8758 RExC_paren_name_list= newAV();
8759 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
8762 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
8764 sv_dat = HeVAL(he_str);
8766 /* croak baby croak */
8768 "panic: paren_name hash element allocation failed");
8769 } else if ( SvPOK(sv_dat) ) {
8770 /* (?|...) can mean we have dupes so scan to check
8771 its already been stored. Maybe a flag indicating
8772 we are inside such a construct would be useful,
8773 but the arrays are likely to be quite small, so
8774 for now we punt -- dmq */
8775 IV count = SvIV(sv_dat);
8776 I32 *pv = (I32*)SvPVX(sv_dat);
8778 for ( i = 0 ; i < count ; i++ ) {
8779 if ( pv[i] == RExC_npar ) {
8785 pv = (I32*)SvGROW(sv_dat, SvCUR(sv_dat) + sizeof(I32)+1);
8786 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
8787 pv[count] = RExC_npar;
8788 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
8791 (void)SvUPGRADE(sv_dat,SVt_PVNV);
8792 sv_setpvn(sv_dat, (char *)&(RExC_npar), sizeof(I32));
8794 SvIV_set(sv_dat, 1);
8797 /* Yes this does cause a memory leak in debugging Perls */
8798 if (!av_store(RExC_paren_name_list, RExC_npar, SvREFCNT_inc(svname)))
8799 SvREFCNT_dec_NN(svname);
8802 /*sv_dump(sv_dat);*/
8804 nextchar(pRExC_state);
8806 goto capturing_parens;
8808 RExC_seen |= REG_SEEN_LOOKBEHIND;
8809 RExC_in_lookbehind++;
8811 case '=': /* (?=...) */
8812 RExC_seen_zerolen++;
8814 case '!': /* (?!...) */
8815 RExC_seen_zerolen++;
8816 if (*RExC_parse == ')') {
8817 ret=reg_node(pRExC_state, OPFAIL);
8818 nextchar(pRExC_state);
8822 case '|': /* (?|...) */
8823 /* branch reset, behave like a (?:...) except that
8824 buffers in alternations share the same numbers */
8826 after_freeze = freeze_paren = RExC_npar;
8828 case ':': /* (?:...) */
8829 case '>': /* (?>...) */
8831 case '$': /* (?$...) */
8832 case '@': /* (?@...) */
8833 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
8835 case '0' : /* (?0) */
8836 case 'R' : /* (?R) */
8837 if (*RExC_parse != ')')
8838 FAIL("Sequence (?R) not terminated");
8839 ret = reg_node(pRExC_state, GOSTART);
8840 *flagp |= POSTPONED;
8841 nextchar(pRExC_state);
8844 { /* named and numeric backreferences */
8846 case '&': /* (?&NAME) */
8847 parse_start = RExC_parse - 1;
8850 SV *sv_dat = reg_scan_name(pRExC_state,
8851 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8852 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8854 goto gen_recurse_regop;
8855 assert(0); /* NOT REACHED */
8857 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8859 vFAIL("Illegal pattern");
8861 goto parse_recursion;
8863 case '-': /* (?-1) */
8864 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8865 RExC_parse--; /* rewind to let it be handled later */
8869 case '1': case '2': case '3': case '4': /* (?1) */
8870 case '5': case '6': case '7': case '8': case '9':
8873 num = atoi(RExC_parse);
8874 parse_start = RExC_parse - 1; /* MJD */
8875 if (*RExC_parse == '-')
8877 while (isDIGIT(*RExC_parse))
8879 if (*RExC_parse!=')')
8880 vFAIL("Expecting close bracket");
8883 if ( paren == '-' ) {
8885 Diagram of capture buffer numbering.
8886 Top line is the normal capture buffer numbers
8887 Bottom line is the negative indexing as from
8891 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
8895 num = RExC_npar + num;
8898 vFAIL("Reference to nonexistent group");
8900 } else if ( paren == '+' ) {
8901 num = RExC_npar + num - 1;
8904 ret = reganode(pRExC_state, GOSUB, num);
8906 if (num > (I32)RExC_rx->nparens) {
8908 vFAIL("Reference to nonexistent group");
8910 ARG2L_SET( ret, RExC_recurse_count++);
8912 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
8913 "Recurse #%"UVuf" to %"IVdf"\n", (UV)ARG(ret), (IV)ARG2L(ret)));
8917 RExC_seen |= REG_SEEN_RECURSE;
8918 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
8919 Set_Node_Offset(ret, parse_start); /* MJD */
8921 *flagp |= POSTPONED;
8922 nextchar(pRExC_state);
8924 } /* named and numeric backreferences */
8925 assert(0); /* NOT REACHED */
8927 case '?': /* (??...) */
8929 if (*RExC_parse != '{') {
8931 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8934 *flagp |= POSTPONED;
8935 paren = *RExC_parse++;
8937 case '{': /* (?{...}) */
8940 struct reg_code_block *cb;
8942 RExC_seen_zerolen++;
8944 if ( !pRExC_state->num_code_blocks
8945 || pRExC_state->code_index >= pRExC_state->num_code_blocks
8946 || pRExC_state->code_blocks[pRExC_state->code_index].start
8947 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
8950 if (RExC_pm_flags & PMf_USE_RE_EVAL)
8951 FAIL("panic: Sequence (?{...}): no code block found\n");
8952 FAIL("Eval-group not allowed at runtime, use re 'eval'");
8954 /* this is a pre-compiled code block (?{...}) */
8955 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
8956 RExC_parse = RExC_start + cb->end;
8959 if (cb->src_regex) {
8960 n = add_data(pRExC_state, 2, "rl");
8961 RExC_rxi->data->data[n] =
8962 (void*)SvREFCNT_inc((SV*)cb->src_regex);
8963 RExC_rxi->data->data[n+1] = (void*)o;
8966 n = add_data(pRExC_state, 1,
8967 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l");
8968 RExC_rxi->data->data[n] = (void*)o;
8971 pRExC_state->code_index++;
8972 nextchar(pRExC_state);
8976 ret = reg_node(pRExC_state, LOGICAL);
8977 eval = reganode(pRExC_state, EVAL, n);
8980 /* for later propagation into (??{}) return value */
8981 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
8983 REGTAIL(pRExC_state, ret, eval);
8984 /* deal with the length of this later - MJD */
8987 ret = reganode(pRExC_state, EVAL, n);
8988 Set_Node_Length(ret, RExC_parse - parse_start + 1);
8989 Set_Node_Offset(ret, parse_start);
8992 case '(': /* (?(?{...})...) and (?(?=...)...) */
8995 if (RExC_parse[0] == '?') { /* (?(?...)) */
8996 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
8997 || RExC_parse[1] == '<'
8998 || RExC_parse[1] == '{') { /* Lookahead or eval. */
9002 ret = reg_node(pRExC_state, LOGICAL);
9006 tail = reg(pRExC_state, 1, &flag, depth+1);
9007 if (flag & RESTART_UTF8) {
9008 *flagp = RESTART_UTF8;
9011 REGTAIL(pRExC_state, ret, tail);
9015 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9016 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9018 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9019 char *name_start= RExC_parse++;
9021 SV *sv_dat=reg_scan_name(pRExC_state,
9022 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9023 if (RExC_parse == name_start || *RExC_parse != ch)
9024 vFAIL2("Sequence (?(%c... not terminated",
9025 (ch == '>' ? '<' : ch));
9028 num = add_data( pRExC_state, 1, "S" );
9029 RExC_rxi->data->data[num]=(void*)sv_dat;
9030 SvREFCNT_inc_simple_void(sv_dat);
9032 ret = reganode(pRExC_state,NGROUPP,num);
9033 goto insert_if_check_paren;
9035 else if (RExC_parse[0] == 'D' &&
9036 RExC_parse[1] == 'E' &&
9037 RExC_parse[2] == 'F' &&
9038 RExC_parse[3] == 'I' &&
9039 RExC_parse[4] == 'N' &&
9040 RExC_parse[5] == 'E')
9042 ret = reganode(pRExC_state,DEFINEP,0);
9045 goto insert_if_check_paren;
9047 else if (RExC_parse[0] == 'R') {
9050 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9051 parno = atoi(RExC_parse++);
9052 while (isDIGIT(*RExC_parse))
9054 } else if (RExC_parse[0] == '&') {
9057 sv_dat = reg_scan_name(pRExC_state,
9058 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9059 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9061 ret = reganode(pRExC_state,INSUBP,parno);
9062 goto insert_if_check_paren;
9064 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9067 parno = atoi(RExC_parse++);
9069 while (isDIGIT(*RExC_parse))
9071 ret = reganode(pRExC_state, GROUPP, parno);
9073 insert_if_check_paren:
9074 if ((c = *nextchar(pRExC_state)) != ')')
9075 vFAIL("Switch condition not recognized");
9077 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9078 br = regbranch(pRExC_state, &flags, 1,depth+1);
9080 if (flags & RESTART_UTF8) {
9081 *flagp = RESTART_UTF8;
9084 FAIL2("panic: regbranch returned NULL, flags=%#X",
9087 REGTAIL(pRExC_state, br, reganode(pRExC_state, LONGJMP, 0));
9088 c = *nextchar(pRExC_state);
9093 vFAIL("(?(DEFINE)....) does not allow branches");
9094 lastbr = reganode(pRExC_state, IFTHEN, 0); /* Fake one for optimizer. */
9095 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
9096 if (flags & RESTART_UTF8) {
9097 *flagp = RESTART_UTF8;
9100 FAIL2("panic: regbranch returned NULL, flags=%#X",
9103 REGTAIL(pRExC_state, ret, lastbr);
9106 c = *nextchar(pRExC_state);
9111 vFAIL("Switch (?(condition)... contains too many branches");
9112 ender = reg_node(pRExC_state, TAIL);
9113 REGTAIL(pRExC_state, br, ender);
9115 REGTAIL(pRExC_state, lastbr, ender);
9116 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
9119 REGTAIL(pRExC_state, ret, ender);
9120 RExC_size++; /* XXX WHY do we need this?!!
9121 For large programs it seems to be required
9122 but I can't figure out why. -- dmq*/
9126 vFAIL2("Unknown switch condition (?(%.2s", RExC_parse);
9129 case '[': /* (?[ ... ]) */
9130 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
9133 RExC_parse--; /* for vFAIL to print correctly */
9134 vFAIL("Sequence (? incomplete");
9136 default: /* e.g., (?i) */
9139 parse_lparen_question_flags(pRExC_state);
9140 if (UCHARAT(RExC_parse) != ':') {
9141 nextchar(pRExC_state);
9146 nextchar(pRExC_state);
9156 ret = reganode(pRExC_state, OPEN, parno);
9159 RExC_nestroot = parno;
9160 if (RExC_seen & REG_SEEN_RECURSE
9161 && !RExC_open_parens[parno-1])
9163 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9164 "Setting open paren #%"IVdf" to %d\n",
9165 (IV)parno, REG_NODE_NUM(ret)));
9166 RExC_open_parens[parno-1]= ret;
9169 Set_Node_Length(ret, 1); /* MJD */
9170 Set_Node_Offset(ret, RExC_parse); /* MJD */
9178 /* Pick up the branches, linking them together. */
9179 parse_start = RExC_parse; /* MJD */
9180 br = regbranch(pRExC_state, &flags, 1,depth+1);
9182 /* branch_len = (paren != 0); */
9185 if (flags & RESTART_UTF8) {
9186 *flagp = RESTART_UTF8;
9189 FAIL2("panic: regbranch returned NULL, flags=%#X", flags);
9191 if (*RExC_parse == '|') {
9192 if (!SIZE_ONLY && RExC_extralen) {
9193 reginsert(pRExC_state, BRANCHJ, br, depth+1);
9196 reginsert(pRExC_state, BRANCH, br, depth+1);
9197 Set_Node_Length(br, paren != 0);
9198 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
9202 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
9204 else if (paren == ':') {
9205 *flagp |= flags&SIMPLE;
9207 if (is_open) { /* Starts with OPEN. */
9208 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
9210 else if (paren != '?') /* Not Conditional */
9212 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9214 while (*RExC_parse == '|') {
9215 if (!SIZE_ONLY && RExC_extralen) {
9216 ender = reganode(pRExC_state, LONGJMP,0);
9217 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender); /* Append to the previous. */
9220 RExC_extralen += 2; /* Account for LONGJMP. */
9221 nextchar(pRExC_state);
9223 if (RExC_npar > after_freeze)
9224 after_freeze = RExC_npar;
9225 RExC_npar = freeze_paren;
9227 br = regbranch(pRExC_state, &flags, 0, depth+1);
9230 if (flags & RESTART_UTF8) {
9231 *flagp = RESTART_UTF8;
9234 FAIL2("panic: regbranch returned NULL, flags=%#X", flags);
9236 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
9238 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9241 if (have_branch || paren != ':') {
9242 /* Make a closing node, and hook it on the end. */
9245 ender = reg_node(pRExC_state, TAIL);
9248 ender = reganode(pRExC_state, CLOSE, parno);
9249 if (!SIZE_ONLY && RExC_seen & REG_SEEN_RECURSE) {
9250 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9251 "Setting close paren #%"IVdf" to %d\n",
9252 (IV)parno, REG_NODE_NUM(ender)));
9253 RExC_close_parens[parno-1]= ender;
9254 if (RExC_nestroot == parno)
9257 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
9258 Set_Node_Length(ender,1); /* MJD */
9264 *flagp &= ~HASWIDTH;
9267 ender = reg_node(pRExC_state, SUCCEED);
9270 ender = reg_node(pRExC_state, END);
9272 assert(!RExC_opend); /* there can only be one! */
9277 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9278 SV * const mysv_val1=sv_newmortal();
9279 SV * const mysv_val2=sv_newmortal();
9280 DEBUG_PARSE_MSG("lsbr");
9281 regprop(RExC_rx, mysv_val1, lastbr);
9282 regprop(RExC_rx, mysv_val2, ender);
9283 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9284 SvPV_nolen_const(mysv_val1),
9285 (IV)REG_NODE_NUM(lastbr),
9286 SvPV_nolen_const(mysv_val2),
9287 (IV)REG_NODE_NUM(ender),
9288 (IV)(ender - lastbr)
9291 REGTAIL(pRExC_state, lastbr, ender);
9293 if (have_branch && !SIZE_ONLY) {
9296 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
9298 /* Hook the tails of the branches to the closing node. */
9299 for (br = ret; br; br = regnext(br)) {
9300 const U8 op = PL_regkind[OP(br)];
9302 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
9303 if (OP(NEXTOPER(br)) != NOTHING || regnext(NEXTOPER(br)) != ender)
9306 else if (op == BRANCHJ) {
9307 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
9308 /* for now we always disable this optimisation * /
9309 if (OP(NEXTOPER(NEXTOPER(br))) != NOTHING || regnext(NEXTOPER(NEXTOPER(br))) != ender)
9315 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
9316 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9317 SV * const mysv_val1=sv_newmortal();
9318 SV * const mysv_val2=sv_newmortal();
9319 DEBUG_PARSE_MSG("NADA");
9320 regprop(RExC_rx, mysv_val1, ret);
9321 regprop(RExC_rx, mysv_val2, ender);
9322 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9323 SvPV_nolen_const(mysv_val1),
9324 (IV)REG_NODE_NUM(ret),
9325 SvPV_nolen_const(mysv_val2),
9326 (IV)REG_NODE_NUM(ender),
9331 if (OP(ender) == TAIL) {
9336 for ( opt= br + 1; opt < ender ; opt++ )
9338 NEXT_OFF(br)= ender - br;
9346 static const char parens[] = "=!<,>";
9348 if (paren && (p = strchr(parens, paren))) {
9349 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
9350 int flag = (p - parens) > 1;
9353 node = SUSPEND, flag = 0;
9354 reginsert(pRExC_state, node,ret, depth+1);
9355 Set_Node_Cur_Length(ret);
9356 Set_Node_Offset(ret, parse_start + 1);
9358 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
9362 /* Check for proper termination. */
9364 RExC_flags = oregflags;
9365 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
9366 RExC_parse = oregcomp_parse;
9367 vFAIL("Unmatched (");
9370 else if (!paren && RExC_parse < RExC_end) {
9371 if (*RExC_parse == ')') {
9373 vFAIL("Unmatched )");
9376 FAIL("Junk on end of regexp"); /* "Can't happen". */
9377 assert(0); /* NOTREACHED */
9380 if (RExC_in_lookbehind) {
9381 RExC_in_lookbehind--;
9383 if (after_freeze > RExC_npar)
9384 RExC_npar = after_freeze;
9389 - regbranch - one alternative of an | operator
9391 * Implements the concatenation operator.
9393 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9397 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
9401 regnode *chain = NULL;
9403 I32 flags = 0, c = 0;
9404 GET_RE_DEBUG_FLAGS_DECL;
9406 PERL_ARGS_ASSERT_REGBRANCH;
9408 DEBUG_PARSE("brnc");
9413 if (!SIZE_ONLY && RExC_extralen)
9414 ret = reganode(pRExC_state, BRANCHJ,0);
9416 ret = reg_node(pRExC_state, BRANCH);
9417 Set_Node_Length(ret, 1);
9421 if (!first && SIZE_ONLY)
9422 RExC_extralen += 1; /* BRANCHJ */
9424 *flagp = WORST; /* Tentatively. */
9427 nextchar(pRExC_state);
9428 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
9430 latest = regpiece(pRExC_state, &flags,depth+1);
9431 if (latest == NULL) {
9432 if (flags & TRYAGAIN)
9434 if (flags & RESTART_UTF8) {
9435 *flagp = RESTART_UTF8;
9438 FAIL2("panic: regpiece returned NULL, flags=%#X", flags);
9440 else if (ret == NULL)
9442 *flagp |= flags&(HASWIDTH|POSTPONED);
9443 if (chain == NULL) /* First piece. */
9444 *flagp |= flags&SPSTART;
9447 REGTAIL(pRExC_state, chain, latest);
9452 if (chain == NULL) { /* Loop ran zero times. */
9453 chain = reg_node(pRExC_state, NOTHING);
9458 *flagp |= flags&SIMPLE;
9465 - regpiece - something followed by possible [*+?]
9467 * Note that the branching code sequences used for ? and the general cases
9468 * of * and + are somewhat optimized: they use the same NOTHING node as
9469 * both the endmarker for their branch list and the body of the last branch.
9470 * It might seem that this node could be dispensed with entirely, but the
9471 * endmarker role is not redundant.
9473 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
9475 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9479 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
9486 const char * const origparse = RExC_parse;
9488 I32 max = REG_INFTY;
9489 #ifdef RE_TRACK_PATTERN_OFFSETS
9492 const char *maxpos = NULL;
9494 /* Save the original in case we change the emitted regop to a FAIL. */
9495 regnode * const orig_emit = RExC_emit;
9497 GET_RE_DEBUG_FLAGS_DECL;
9499 PERL_ARGS_ASSERT_REGPIECE;
9501 DEBUG_PARSE("piec");
9503 ret = regatom(pRExC_state, &flags,depth+1);
9505 if (flags & (TRYAGAIN|RESTART_UTF8))
9506 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
9508 FAIL2("panic: regatom returned NULL, flags=%#X", flags);
9514 if (op == '{' && regcurly(RExC_parse, FALSE)) {
9516 #ifdef RE_TRACK_PATTERN_OFFSETS
9517 parse_start = RExC_parse; /* MJD */
9519 next = RExC_parse + 1;
9520 while (isDIGIT(*next) || *next == ',') {
9529 if (*next == '}') { /* got one */
9533 min = atoi(RExC_parse);
9537 maxpos = RExC_parse;
9539 if (!max && *maxpos != '0')
9540 max = REG_INFTY; /* meaning "infinity" */
9541 else if (max >= REG_INFTY)
9542 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
9544 nextchar(pRExC_state);
9545 if (max < min) { /* If can't match, warn and optimize to fail
9548 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
9550 /* We can't back off the size because we have to reserve
9551 * enough space for all the things we are about to throw
9552 * away, but we can shrink it by the ammount we are about
9554 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
9557 RExC_emit = orig_emit;
9559 ret = reg_node(pRExC_state, OPFAIL);
9562 else if (max == 0) { /* replace {0} with a nothing node */
9564 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)NOTHING];
9567 RExC_emit = orig_emit;
9569 ret = reg_node(pRExC_state, NOTHING);
9574 if ((flags&SIMPLE)) {
9575 RExC_naughty += 2 + RExC_naughty / 2;
9576 reginsert(pRExC_state, CURLY, ret, depth+1);
9577 Set_Node_Offset(ret, parse_start+1); /* MJD */
9578 Set_Node_Cur_Length(ret);
9581 regnode * const w = reg_node(pRExC_state, WHILEM);
9584 REGTAIL(pRExC_state, ret, w);
9585 if (!SIZE_ONLY && RExC_extralen) {
9586 reginsert(pRExC_state, LONGJMP,ret, depth+1);
9587 reginsert(pRExC_state, NOTHING,ret, depth+1);
9588 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
9590 reginsert(pRExC_state, CURLYX,ret, depth+1);
9592 Set_Node_Offset(ret, parse_start+1);
9593 Set_Node_Length(ret,
9594 op == '{' ? (RExC_parse - parse_start) : 1);
9596 if (!SIZE_ONLY && RExC_extralen)
9597 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
9598 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
9600 RExC_whilem_seen++, RExC_extralen += 3;
9601 RExC_naughty += 4 + RExC_naughty; /* compound interest */
9610 ARG1_SET(ret, (U16)min);
9611 ARG2_SET(ret, (U16)max);
9623 #if 0 /* Now runtime fix should be reliable. */
9625 /* if this is reinstated, don't forget to put this back into perldiag:
9627 =item Regexp *+ operand could be empty at {#} in regex m/%s/
9629 (F) The part of the regexp subject to either the * or + quantifier
9630 could match an empty string. The {#} shows in the regular
9631 expression about where the problem was discovered.
9635 if (!(flags&HASWIDTH) && op != '?')
9636 vFAIL("Regexp *+ operand could be empty");
9639 #ifdef RE_TRACK_PATTERN_OFFSETS
9640 parse_start = RExC_parse;
9642 nextchar(pRExC_state);
9644 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
9646 if (op == '*' && (flags&SIMPLE)) {
9647 reginsert(pRExC_state, STAR, ret, depth+1);
9651 else if (op == '*') {
9655 else if (op == '+' && (flags&SIMPLE)) {
9656 reginsert(pRExC_state, PLUS, ret, depth+1);
9660 else if (op == '+') {
9664 else if (op == '?') {
9669 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
9670 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
9671 ckWARN3reg(RExC_parse,
9672 "%.*s matches null string many times",
9673 (int)(RExC_parse >= origparse ? RExC_parse - origparse : 0),
9675 (void)ReREFCNT_inc(RExC_rx_sv);
9678 if (RExC_parse < RExC_end && *RExC_parse == '?') {
9679 nextchar(pRExC_state);
9680 reginsert(pRExC_state, MINMOD, ret, depth+1);
9681 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
9683 #ifndef REG_ALLOW_MINMOD_SUSPEND
9686 if (RExC_parse < RExC_end && *RExC_parse == '+') {
9688 nextchar(pRExC_state);
9689 ender = reg_node(pRExC_state, SUCCEED);
9690 REGTAIL(pRExC_state, ret, ender);
9691 reginsert(pRExC_state, SUSPEND, ret, depth+1);
9693 ender = reg_node(pRExC_state, TAIL);
9694 REGTAIL(pRExC_state, ret, ender);
9698 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
9700 vFAIL("Nested quantifiers");
9707 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p, UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
9708 const bool strict /* Apply stricter parsing rules? */
9712 /* This is expected to be called by a parser routine that has recognized '\N'
9713 and needs to handle the rest. RExC_parse is expected to point at the first
9714 char following the N at the time of the call. On successful return,
9715 RExC_parse has been updated to point to just after the sequence identified
9716 by this routine, and <*flagp> has been updated.
9718 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
9721 \N may begin either a named sequence, or if outside a character class, mean
9722 to match a non-newline. For non single-quoted regexes, the tokenizer has
9723 attempted to decide which, and in the case of a named sequence, converted it
9724 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
9725 where c1... are the characters in the sequence. For single-quoted regexes,
9726 the tokenizer passes the \N sequence through unchanged; this code will not
9727 attempt to determine this nor expand those, instead raising a syntax error.
9728 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
9729 or there is no '}', it signals that this \N occurrence means to match a
9732 Only the \N{U+...} form should occur in a character class, for the same
9733 reason that '.' inside a character class means to just match a period: it
9734 just doesn't make sense.
9736 The function raises an error (via vFAIL), and doesn't return for various
9737 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
9738 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
9739 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
9740 only possible if node_p is non-NULL.
9743 If <valuep> is non-null, it means the caller can accept an input sequence
9744 consisting of a just a single code point; <*valuep> is set to that value
9745 if the input is such.
9747 If <node_p> is non-null it signifies that the caller can accept any other
9748 legal sequence (i.e., one that isn't just a single code point). <*node_p>
9750 1) \N means not-a-NL: points to a newly created REG_ANY node;
9751 2) \N{}: points to a new NOTHING node;
9752 3) otherwise: points to a new EXACT node containing the resolved
9754 Note that FALSE is returned for single code point sequences if <valuep> is
9758 char * endbrace; /* '}' following the name */
9760 char *endchar; /* Points to '.' or '}' ending cur char in the input
9762 bool has_multiple_chars; /* true if the input stream contains a sequence of
9763 more than one character */
9765 GET_RE_DEBUG_FLAGS_DECL;
9767 PERL_ARGS_ASSERT_GROK_BSLASH_N;
9771 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
9773 /* The [^\n] meaning of \N ignores spaces and comments under the /x
9774 * modifier. The other meaning does not */
9775 p = (RExC_flags & RXf_PMf_EXTENDED)
9776 ? regwhite( pRExC_state, RExC_parse )
9779 /* Disambiguate between \N meaning a named character versus \N meaning
9780 * [^\n]. The former is assumed when it can't be the latter. */
9781 if (*p != '{' || regcurly(p, FALSE)) {
9784 /* no bare \N in a charclass */
9785 if (in_char_class) {
9786 vFAIL("\\N in a character class must be a named character: \\N{...}");
9790 nextchar(pRExC_state);
9791 *node_p = reg_node(pRExC_state, REG_ANY);
9792 *flagp |= HASWIDTH|SIMPLE;
9795 Set_Node_Length(*node_p, 1); /* MJD */
9799 /* Here, we have decided it should be a named character or sequence */
9801 /* The test above made sure that the next real character is a '{', but
9802 * under the /x modifier, it could be separated by space (or a comment and
9803 * \n) and this is not allowed (for consistency with \x{...} and the
9804 * tokenizer handling of \N{NAME}). */
9805 if (*RExC_parse != '{') {
9806 vFAIL("Missing braces on \\N{}");
9809 RExC_parse++; /* Skip past the '{' */
9811 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
9812 || ! (endbrace == RExC_parse /* nothing between the {} */
9813 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below */
9814 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg) */
9816 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
9817 vFAIL("\\N{NAME} must be resolved by the lexer");
9820 if (endbrace == RExC_parse) { /* empty: \N{} */
9823 *node_p = reg_node(pRExC_state,NOTHING);
9825 else if (in_char_class) {
9826 if (SIZE_ONLY && in_char_class) {
9828 RExC_parse++; /* Position after the "}" */
9829 vFAIL("Zero length \\N{}");
9832 ckWARNreg(RExC_parse,
9833 "Ignoring zero length \\N{} in character class");
9841 nextchar(pRExC_state);
9845 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
9846 RExC_parse += 2; /* Skip past the 'U+' */
9848 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9850 /* Code points are separated by dots. If none, there is only one code
9851 * point, and is terminated by the brace */
9852 has_multiple_chars = (endchar < endbrace);
9854 if (valuep && (! has_multiple_chars || in_char_class)) {
9855 /* We only pay attention to the first char of
9856 multichar strings being returned in char classes. I kinda wonder
9857 if this makes sense as it does change the behaviour
9858 from earlier versions, OTOH that behaviour was broken
9859 as well. XXX Solution is to recharacterize as
9860 [rest-of-class]|multi1|multi2... */
9862 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
9863 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
9864 | PERL_SCAN_DISALLOW_PREFIX
9865 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
9867 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
9869 /* The tokenizer should have guaranteed validity, but it's possible to
9870 * bypass it by using single quoting, so check */
9871 if (length_of_hex == 0
9872 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
9874 RExC_parse += length_of_hex; /* Includes all the valid */
9875 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
9876 ? UTF8SKIP(RExC_parse)
9878 /* Guard against malformed utf8 */
9879 if (RExC_parse >= endchar) {
9880 RExC_parse = endchar;
9882 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9885 if (in_char_class && has_multiple_chars) {
9887 RExC_parse = endbrace;
9888 vFAIL("\\N{} in character class restricted to one character");
9891 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
9895 RExC_parse = endbrace + 1;
9897 else if (! node_p || ! has_multiple_chars) {
9899 /* Here, the input is legal, but not according to the caller's
9900 * options. We fail without advancing the parse, so that the
9901 * caller can try again */
9907 /* What is done here is to convert this to a sub-pattern of the form
9908 * (?:\x{char1}\x{char2}...)
9909 * and then call reg recursively. That way, it retains its atomicness,
9910 * while not having to worry about special handling that some code
9911 * points may have. toke.c has converted the original Unicode values
9912 * to native, so that we can just pass on the hex values unchanged. We
9913 * do have to set a flag to keep recoding from happening in the
9916 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
9918 char *orig_end = RExC_end;
9921 while (RExC_parse < endbrace) {
9923 /* Convert to notation the rest of the code understands */
9924 sv_catpv(substitute_parse, "\\x{");
9925 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
9926 sv_catpv(substitute_parse, "}");
9928 /* Point to the beginning of the next character in the sequence. */
9929 RExC_parse = endchar + 1;
9930 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9932 sv_catpv(substitute_parse, ")");
9934 RExC_parse = SvPV(substitute_parse, len);
9936 /* Don't allow empty number */
9938 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9940 RExC_end = RExC_parse + len;
9942 /* The values are Unicode, and therefore not subject to recoding */
9943 RExC_override_recoding = 1;
9945 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
9946 if (flags & RESTART_UTF8) {
9947 *flagp = RESTART_UTF8;
9950 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#X",
9953 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
9955 RExC_parse = endbrace;
9956 RExC_end = orig_end;
9957 RExC_override_recoding = 0;
9959 nextchar(pRExC_state);
9969 * It returns the code point in utf8 for the value in *encp.
9970 * value: a code value in the source encoding
9971 * encp: a pointer to an Encode object
9973 * If the result from Encode is not a single character,
9974 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
9977 S_reg_recode(pTHX_ const char value, SV **encp)
9980 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
9981 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
9982 const STRLEN newlen = SvCUR(sv);
9983 UV uv = UNICODE_REPLACEMENT;
9985 PERL_ARGS_ASSERT_REG_RECODE;
9989 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
9992 if (!newlen || numlen != newlen) {
9993 uv = UNICODE_REPLACEMENT;
9999 PERL_STATIC_INLINE U8
10000 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
10004 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
10010 op = get_regex_charset(RExC_flags);
10011 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
10012 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
10013 been, so there is no hole */
10016 return op + EXACTF;
10019 PERL_STATIC_INLINE void
10020 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state, regnode *node, I32* flagp, STRLEN len, UV code_point)
10022 /* This knows the details about sizing an EXACTish node, setting flags for
10023 * it (by setting <*flagp>, and potentially populating it with a single
10026 * If <len> (the length in bytes) is non-zero, this function assumes that
10027 * the node has already been populated, and just does the sizing. In this
10028 * case <code_point> should be the final code point that has already been
10029 * placed into the node. This value will be ignored except that under some
10030 * circumstances <*flagp> is set based on it.
10032 * If <len> is zero, the function assumes that the node is to contain only
10033 * the single character given by <code_point> and calculates what <len>
10034 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10035 * additionally will populate the node's STRING with <code_point>, if <len>
10036 * is 0. In both cases <*flagp> is appropriately set
10038 * It knows that under FOLD, UTF characters and the Latin Sharp S must be
10039 * folded (the latter only when the rules indicate it can match 'ss') */
10041 bool len_passed_in = cBOOL(len != 0);
10042 U8 character[UTF8_MAXBYTES_CASE+1];
10044 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10046 if (! len_passed_in) {
10049 to_uni_fold(NATIVE_TO_UNI(code_point), character, &len);
10052 uvchr_to_utf8( character, code_point);
10053 len = UTF8SKIP(character);
10057 || code_point != LATIN_SMALL_LETTER_SHARP_S
10058 || ASCII_FOLD_RESTRICTED
10059 || ! AT_LEAST_UNI_SEMANTICS)
10061 *character = (U8) code_point;
10066 *(character + 1) = 's';
10072 RExC_size += STR_SZ(len);
10075 RExC_emit += STR_SZ(len);
10076 STR_LEN(node) = len;
10077 if (! len_passed_in) {
10078 Copy((char *) character, STRING(node), len, char);
10082 *flagp |= HASWIDTH;
10084 /* A single character node is SIMPLE, except for the special-cased SHARP S
10086 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
10087 && (code_point != LATIN_SMALL_LETTER_SHARP_S
10088 || ! FOLD || ! DEPENDS_SEMANTICS))
10095 - regatom - the lowest level
10097 Try to identify anything special at the start of the pattern. If there
10098 is, then handle it as required. This may involve generating a single regop,
10099 such as for an assertion; or it may involve recursing, such as to
10100 handle a () structure.
10102 If the string doesn't start with something special then we gobble up
10103 as much literal text as we can.
10105 Once we have been able to handle whatever type of thing started the
10106 sequence, we return.
10108 Note: we have to be careful with escapes, as they can be both literal
10109 and special, and in the case of \10 and friends, context determines which.
10111 A summary of the code structure is:
10113 switch (first_byte) {
10114 cases for each special:
10115 handle this special;
10118 switch (2nd byte) {
10119 cases for each unambiguous special:
10120 handle this special;
10122 cases for each ambigous special/literal:
10124 if (special) handle here
10126 default: // unambiguously literal:
10129 default: // is a literal char
10132 create EXACTish node for literal;
10133 while (more input and node isn't full) {
10134 switch (input_byte) {
10135 cases for each special;
10136 make sure parse pointer is set so that the next call to
10137 regatom will see this special first
10138 goto loopdone; // EXACTish node terminated by prev. char
10140 append char to EXACTISH node;
10142 get next input byte;
10146 return the generated node;
10148 Specifically there are two separate switches for handling
10149 escape sequences, with the one for handling literal escapes requiring
10150 a dummy entry for all of the special escapes that are actually handled
10153 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
10155 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10157 Otherwise does not return NULL.
10161 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10164 regnode *ret = NULL;
10166 char *parse_start = RExC_parse;
10170 GET_RE_DEBUG_FLAGS_DECL;
10172 *flagp = WORST; /* Tentatively. */
10174 DEBUG_PARSE("atom");
10176 PERL_ARGS_ASSERT_REGATOM;
10179 switch ((U8)*RExC_parse) {
10181 RExC_seen_zerolen++;
10182 nextchar(pRExC_state);
10183 if (RExC_flags & RXf_PMf_MULTILINE)
10184 ret = reg_node(pRExC_state, MBOL);
10185 else if (RExC_flags & RXf_PMf_SINGLELINE)
10186 ret = reg_node(pRExC_state, SBOL);
10188 ret = reg_node(pRExC_state, BOL);
10189 Set_Node_Length(ret, 1); /* MJD */
10192 nextchar(pRExC_state);
10194 RExC_seen_zerolen++;
10195 if (RExC_flags & RXf_PMf_MULTILINE)
10196 ret = reg_node(pRExC_state, MEOL);
10197 else if (RExC_flags & RXf_PMf_SINGLELINE)
10198 ret = reg_node(pRExC_state, SEOL);
10200 ret = reg_node(pRExC_state, EOL);
10201 Set_Node_Length(ret, 1); /* MJD */
10204 nextchar(pRExC_state);
10205 if (RExC_flags & RXf_PMf_SINGLELINE)
10206 ret = reg_node(pRExC_state, SANY);
10208 ret = reg_node(pRExC_state, REG_ANY);
10209 *flagp |= HASWIDTH|SIMPLE;
10211 Set_Node_Length(ret, 1); /* MJD */
10215 char * const oregcomp_parse = ++RExC_parse;
10216 ret = regclass(pRExC_state, flagp,depth+1,
10217 FALSE, /* means parse the whole char class */
10218 TRUE, /* allow multi-char folds */
10219 FALSE, /* don't silence non-portable warnings. */
10221 if (*RExC_parse != ']') {
10222 RExC_parse = oregcomp_parse;
10223 vFAIL("Unmatched [");
10226 if (*flagp & RESTART_UTF8)
10228 FAIL2("panic: regclass returned NULL to regatom, flags=%#X",
10231 nextchar(pRExC_state);
10232 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
10236 nextchar(pRExC_state);
10237 ret = reg(pRExC_state, 1, &flags,depth+1);
10239 if (flags & TRYAGAIN) {
10240 if (RExC_parse == RExC_end) {
10241 /* Make parent create an empty node if needed. */
10242 *flagp |= TRYAGAIN;
10247 if (flags & RESTART_UTF8) {
10248 *flagp = RESTART_UTF8;
10251 FAIL2("panic: reg returned NULL to regatom, flags=%#X", flags);
10253 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10257 if (flags & TRYAGAIN) {
10258 *flagp |= TRYAGAIN;
10261 vFAIL("Internal urp");
10262 /* Supposed to be caught earlier. */
10265 if (!regcurly(RExC_parse, FALSE)) {
10274 vFAIL("Quantifier follows nothing");
10279 This switch handles escape sequences that resolve to some kind
10280 of special regop and not to literal text. Escape sequnces that
10281 resolve to literal text are handled below in the switch marked
10284 Every entry in this switch *must* have a corresponding entry
10285 in the literal escape switch. However, the opposite is not
10286 required, as the default for this switch is to jump to the
10287 literal text handling code.
10289 switch ((U8)*++RExC_parse) {
10291 /* Special Escapes */
10293 RExC_seen_zerolen++;
10294 ret = reg_node(pRExC_state, SBOL);
10296 goto finish_meta_pat;
10298 ret = reg_node(pRExC_state, GPOS);
10299 RExC_seen |= REG_SEEN_GPOS;
10301 goto finish_meta_pat;
10303 RExC_seen_zerolen++;
10304 ret = reg_node(pRExC_state, KEEPS);
10306 /* XXX:dmq : disabling in-place substitution seems to
10307 * be necessary here to avoid cases of memory corruption, as
10308 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
10310 RExC_seen |= REG_SEEN_LOOKBEHIND;
10311 goto finish_meta_pat;
10313 ret = reg_node(pRExC_state, SEOL);
10315 RExC_seen_zerolen++; /* Do not optimize RE away */
10316 goto finish_meta_pat;
10318 ret = reg_node(pRExC_state, EOS);
10320 RExC_seen_zerolen++; /* Do not optimize RE away */
10321 goto finish_meta_pat;
10323 ret = reg_node(pRExC_state, CANY);
10324 RExC_seen |= REG_SEEN_CANY;
10325 *flagp |= HASWIDTH|SIMPLE;
10326 goto finish_meta_pat;
10328 ret = reg_node(pRExC_state, CLUMP);
10329 *flagp |= HASWIDTH;
10330 goto finish_meta_pat;
10336 arg = ANYOF_WORDCHAR;
10340 RExC_seen_zerolen++;
10341 RExC_seen |= REG_SEEN_LOOKBEHIND;
10342 op = BOUND + get_regex_charset(RExC_flags);
10343 if (op > BOUNDA) { /* /aa is same as /a */
10346 ret = reg_node(pRExC_state, op);
10347 FLAGS(ret) = get_regex_charset(RExC_flags);
10349 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10350 ckWARNdep(RExC_parse, "\"\\b{\" is deprecated; use \"\\b\\{\" or \"\\b[{]\" instead");
10352 goto finish_meta_pat;
10354 RExC_seen_zerolen++;
10355 RExC_seen |= REG_SEEN_LOOKBEHIND;
10356 op = NBOUND + get_regex_charset(RExC_flags);
10357 if (op > NBOUNDA) { /* /aa is same as /a */
10360 ret = reg_node(pRExC_state, op);
10361 FLAGS(ret) = get_regex_charset(RExC_flags);
10363 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10364 ckWARNdep(RExC_parse, "\"\\B{\" is deprecated; use \"\\B\\{\" or \"\\B[{]\" instead");
10366 goto finish_meta_pat;
10376 ret = reg_node(pRExC_state, LNBREAK);
10377 *flagp |= HASWIDTH|SIMPLE;
10378 goto finish_meta_pat;
10386 goto join_posix_op_known;
10392 arg = ANYOF_VERTWS;
10394 goto join_posix_op_known;
10404 op = POSIXD + get_regex_charset(RExC_flags);
10405 if (op > POSIXA) { /* /aa is same as /a */
10409 join_posix_op_known:
10412 op += NPOSIXD - POSIXD;
10415 ret = reg_node(pRExC_state, op);
10417 FLAGS(ret) = namedclass_to_classnum(arg);
10420 *flagp |= HASWIDTH|SIMPLE;
10424 nextchar(pRExC_state);
10425 Set_Node_Length(ret, 2); /* MJD */
10431 char* parse_start = RExC_parse - 2;
10436 ret = regclass(pRExC_state, flagp,depth+1,
10437 TRUE, /* means just parse this element */
10438 FALSE, /* don't allow multi-char folds */
10439 FALSE, /* don't silence non-portable warnings.
10440 It would be a bug if these returned
10443 /* regclass() can only return RESTART_UTF8 if multi-char folds
10446 FAIL2("panic: regclass returned NULL to regatom, flags=%#X",
10451 Set_Node_Offset(ret, parse_start + 2);
10452 Set_Node_Cur_Length(ret);
10453 nextchar(pRExC_state);
10457 /* Handle \N and \N{NAME} with multiple code points here and not
10458 * below because it can be multicharacter. join_exact() will join
10459 * them up later on. Also this makes sure that things like
10460 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
10461 * The options to the grok function call causes it to fail if the
10462 * sequence is just a single code point. We then go treat it as
10463 * just another character in the current EXACT node, and hence it
10464 * gets uniform treatment with all the other characters. The
10465 * special treatment for quantifiers is not needed for such single
10466 * character sequences */
10468 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
10469 FALSE /* not strict */ )) {
10470 if (*flagp & RESTART_UTF8)
10476 case 'k': /* Handle \k<NAME> and \k'NAME' */
10479 char ch= RExC_parse[1];
10480 if (ch != '<' && ch != '\'' && ch != '{') {
10482 vFAIL2("Sequence %.2s... not terminated",parse_start);
10484 /* this pretty much dupes the code for (?P=...) in reg(), if
10485 you change this make sure you change that */
10486 char* name_start = (RExC_parse += 2);
10488 SV *sv_dat = reg_scan_name(pRExC_state,
10489 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10490 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
10491 if (RExC_parse == name_start || *RExC_parse != ch)
10492 vFAIL2("Sequence %.3s... not terminated",parse_start);
10495 num = add_data( pRExC_state, 1, "S" );
10496 RExC_rxi->data->data[num]=(void*)sv_dat;
10497 SvREFCNT_inc_simple_void(sv_dat);
10501 ret = reganode(pRExC_state,
10504 : (ASCII_FOLD_RESTRICTED)
10506 : (AT_LEAST_UNI_SEMANTICS)
10512 *flagp |= HASWIDTH;
10514 /* override incorrect value set in reganode MJD */
10515 Set_Node_Offset(ret, parse_start+1);
10516 Set_Node_Cur_Length(ret); /* MJD */
10517 nextchar(pRExC_state);
10523 case '1': case '2': case '3': case '4':
10524 case '5': case '6': case '7': case '8': case '9':
10527 bool isg = *RExC_parse == 'g';
10532 if (*RExC_parse == '{') {
10536 if (*RExC_parse == '-') {
10540 if (hasbrace && !isDIGIT(*RExC_parse)) {
10541 if (isrel) RExC_parse--;
10543 goto parse_named_seq;
10545 num = atoi(RExC_parse);
10546 if (isg && num == 0)
10547 vFAIL("Reference to invalid group 0");
10549 num = RExC_npar - num;
10551 vFAIL("Reference to nonexistent or unclosed group");
10553 if (!isg && num > 9 && num >= RExC_npar)
10554 /* Probably a character specified in octal, e.g. \35 */
10557 char * const parse_start = RExC_parse - 1; /* MJD */
10558 while (isDIGIT(*RExC_parse))
10560 if (parse_start == RExC_parse - 1)
10561 vFAIL("Unterminated \\g... pattern");
10563 if (*RExC_parse != '}')
10564 vFAIL("Unterminated \\g{...} pattern");
10568 if (num > (I32)RExC_rx->nparens)
10569 vFAIL("Reference to nonexistent group");
10572 ret = reganode(pRExC_state,
10575 : (ASCII_FOLD_RESTRICTED)
10577 : (AT_LEAST_UNI_SEMANTICS)
10583 *flagp |= HASWIDTH;
10585 /* override incorrect value set in reganode MJD */
10586 Set_Node_Offset(ret, parse_start+1);
10587 Set_Node_Cur_Length(ret); /* MJD */
10589 nextchar(pRExC_state);
10594 if (RExC_parse >= RExC_end)
10595 FAIL("Trailing \\");
10598 /* Do not generate "unrecognized" warnings here, we fall
10599 back into the quick-grab loop below */
10606 if (RExC_flags & RXf_PMf_EXTENDED) {
10607 if ( reg_skipcomment( pRExC_state ) )
10614 parse_start = RExC_parse - 1;
10623 #define MAX_NODE_STRING_SIZE 127
10624 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
10626 U8 upper_parse = MAX_NODE_STRING_SIZE;
10629 bool next_is_quantifier;
10630 char * oldp = NULL;
10632 /* If a folding node contains only code points that don't
10633 * participate in folds, it can be changed into an EXACT node,
10634 * which allows the optimizer more things to look for */
10638 node_type = compute_EXACTish(pRExC_state);
10639 ret = reg_node(pRExC_state, node_type);
10641 /* In pass1, folded, we use a temporary buffer instead of the
10642 * actual node, as the node doesn't exist yet */
10643 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
10649 /* We do the EXACTFish to EXACT node only if folding, and not if in
10650 * locale, as whether a character folds or not isn't known until
10652 maybe_exact = FOLD && ! LOC;
10654 /* XXX The node can hold up to 255 bytes, yet this only goes to
10655 * 127. I (khw) do not know why. Keeping it somewhat less than
10656 * 255 allows us to not have to worry about overflow due to
10657 * converting to utf8 and fold expansion, but that value is
10658 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
10659 * split up by this limit into a single one using the real max of
10660 * 255. Even at 127, this breaks under rare circumstances. If
10661 * folding, we do not want to split a node at a character that is a
10662 * non-final in a multi-char fold, as an input string could just
10663 * happen to want to match across the node boundary. The join
10664 * would solve that problem if the join actually happens. But a
10665 * series of more than two nodes in a row each of 127 would cause
10666 * the first join to succeed to get to 254, but then there wouldn't
10667 * be room for the next one, which could at be one of those split
10668 * multi-char folds. I don't know of any fool-proof solution. One
10669 * could back off to end with only a code point that isn't such a
10670 * non-final, but it is possible for there not to be any in the
10672 for (p = RExC_parse - 1;
10673 len < upper_parse && p < RExC_end;
10678 if (RExC_flags & RXf_PMf_EXTENDED)
10679 p = regwhite( pRExC_state, p );
10690 /* Literal Escapes Switch
10692 This switch is meant to handle escape sequences that
10693 resolve to a literal character.
10695 Every escape sequence that represents something
10696 else, like an assertion or a char class, is handled
10697 in the switch marked 'Special Escapes' above in this
10698 routine, but also has an entry here as anything that
10699 isn't explicitly mentioned here will be treated as
10700 an unescaped equivalent literal.
10703 switch ((U8)*++p) {
10704 /* These are all the special escapes. */
10705 case 'A': /* Start assertion */
10706 case 'b': case 'B': /* Word-boundary assertion*/
10707 case 'C': /* Single char !DANGEROUS! */
10708 case 'd': case 'D': /* digit class */
10709 case 'g': case 'G': /* generic-backref, pos assertion */
10710 case 'h': case 'H': /* HORIZWS */
10711 case 'k': case 'K': /* named backref, keep marker */
10712 case 'p': case 'P': /* Unicode property */
10713 case 'R': /* LNBREAK */
10714 case 's': case 'S': /* space class */
10715 case 'v': case 'V': /* VERTWS */
10716 case 'w': case 'W': /* word class */
10717 case 'X': /* eXtended Unicode "combining character sequence" */
10718 case 'z': case 'Z': /* End of line/string assertion */
10722 /* Anything after here is an escape that resolves to a
10723 literal. (Except digits, which may or may not)
10729 case 'N': /* Handle a single-code point named character. */
10730 /* The options cause it to fail if a multiple code
10731 * point sequence. Handle those in the switch() above
10733 RExC_parse = p + 1;
10734 if (! grok_bslash_N(pRExC_state, NULL, &ender,
10735 flagp, depth, FALSE,
10736 FALSE /* not strict */ ))
10738 if (*flagp & RESTART_UTF8)
10739 FAIL("panic: grok_bslash_N set RESTART_UTF8");
10740 RExC_parse = p = oldp;
10744 if (ender > 0xff) {
10761 ender = ASCII_TO_NATIVE('\033');
10765 ender = ASCII_TO_NATIVE('\007');
10771 const char* error_msg;
10773 bool valid = grok_bslash_o(&p,
10776 TRUE, /* out warnings */
10777 FALSE, /* not strict */
10778 TRUE, /* Output warnings
10783 RExC_parse = p; /* going to die anyway; point
10784 to exact spot of failure */
10788 if (PL_encoding && ender < 0x100) {
10789 goto recode_encoding;
10791 if (ender > 0xff) {
10798 UV result = UV_MAX; /* initialize to erroneous
10800 const char* error_msg;
10802 bool valid = grok_bslash_x(&p,
10805 TRUE, /* out warnings */
10806 FALSE, /* not strict */
10807 TRUE, /* Output warnings
10812 RExC_parse = p; /* going to die anyway; point
10813 to exact spot of failure */
10818 if (PL_encoding && ender < 0x100) {
10819 goto recode_encoding;
10821 if (ender > 0xff) {
10828 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
10830 case '0': case '1': case '2': case '3':case '4':
10831 case '5': case '6': case '7':
10833 (isDIGIT(p[1]) && atoi(p) >= RExC_npar))
10835 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
10837 ender = grok_oct(p, &numlen, &flags, NULL);
10838 if (ender > 0xff) {
10842 if (SIZE_ONLY /* like \08, \178 */
10845 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
10847 reg_warn_non_literal_string(
10849 form_short_octal_warning(p, numlen));
10852 else { /* Not to be treated as an octal constant, go
10857 if (PL_encoding && ender < 0x100)
10858 goto recode_encoding;
10861 if (! RExC_override_recoding) {
10862 SV* enc = PL_encoding;
10863 ender = reg_recode((const char)(U8)ender, &enc);
10864 if (!enc && SIZE_ONLY)
10865 ckWARNreg(p, "Invalid escape in the specified encoding");
10871 FAIL("Trailing \\");
10874 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
10875 /* Include any { following the alpha to emphasize
10876 * that it could be part of an escape at some point
10878 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
10879 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
10881 goto normal_default;
10882 } /* End of switch on '\' */
10884 default: /* A literal character */
10887 && RExC_flags & RXf_PMf_EXTENDED
10888 && ckWARN(WARN_DEPRECATED)
10889 && is_PATWS_non_low(p, UTF))
10891 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
10892 "Escape literal pattern white space under /x");
10896 if (UTF8_IS_START(*p) && UTF) {
10898 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
10899 &numlen, UTF8_ALLOW_DEFAULT);
10905 } /* End of switch on the literal */
10907 /* Here, have looked at the literal character and <ender>
10908 * contains its ordinal, <p> points to the character after it
10911 if ( RExC_flags & RXf_PMf_EXTENDED)
10912 p = regwhite( pRExC_state, p );
10914 /* If the next thing is a quantifier, it applies to this
10915 * character only, which means that this character has to be in
10916 * its own node and can't just be appended to the string in an
10917 * existing node, so if there are already other characters in
10918 * the node, close the node with just them, and set up to do
10919 * this character again next time through, when it will be the
10920 * only thing in its new node */
10921 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
10929 /* See comments for join_exact() as to why we fold
10930 * this non-UTF at compile time */
10931 || (node_type == EXACTFU
10932 && ender == LATIN_SMALL_LETTER_SHARP_S))
10936 /* Prime the casefolded buffer. Locale rules, which
10937 * apply only to code points < 256, aren't known until
10938 * execution, so for them, just output the original
10939 * character using utf8. If we start to fold non-UTF
10940 * patterns, be sure to update join_exact() */
10941 if (LOC && ender < 256) {
10942 if (UNI_IS_INVARIANT(ender)) {
10946 *s = UTF8_TWO_BYTE_HI(ender);
10947 *(s + 1) = UTF8_TWO_BYTE_LO(ender);
10952 UV folded = _to_uni_fold_flags(
10957 | ((LOC) ? FOLD_FLAGS_LOCALE
10958 : (ASCII_FOLD_RESTRICTED)
10959 ? FOLD_FLAGS_NOMIX_ASCII
10963 /* If this node only contains non-folding code
10964 * points so far, see if this new one is also
10967 if (folded != ender) {
10968 maybe_exact = FALSE;
10971 /* Here the fold is the original; we have
10972 * to check further to see if anything
10974 if (! PL_utf8_foldable) {
10975 SV* swash = swash_init("utf8",
10977 &PL_sv_undef, 1, 0);
10979 _get_swash_invlist(swash);
10980 SvREFCNT_dec_NN(swash);
10982 if (_invlist_contains_cp(PL_utf8_foldable,
10985 maybe_exact = FALSE;
10993 /* The loop increments <len> each time, as all but this
10994 * path (and the one just below for UTF) through it add
10995 * a single byte to the EXACTish node. But this one
10996 * has changed len to be the correct final value, so
10997 * subtract one to cancel out the increment that
10999 len += foldlen - 1;
11002 *(s++) = (char) ender;
11003 maybe_exact &= ! IS_IN_SOME_FOLD_L1(ender);
11007 const STRLEN unilen = reguni(pRExC_state, ender, s);
11013 /* See comment just above for - 1 */
11017 REGC((char)ender, s++);
11020 if (next_is_quantifier) {
11022 /* Here, the next input is a quantifier, and to get here,
11023 * the current character is the only one in the node.
11024 * Also, here <len> doesn't include the final byte for this
11030 } /* End of loop through literal characters */
11032 /* Here we have either exhausted the input or ran out of room in
11033 * the node. (If we encountered a character that can't be in the
11034 * node, transfer is made directly to <loopdone>, and so we
11035 * wouldn't have fallen off the end of the loop.) In the latter
11036 * case, we artificially have to split the node into two, because
11037 * we just don't have enough space to hold everything. This
11038 * creates a problem if the final character participates in a
11039 * multi-character fold in the non-final position, as a match that
11040 * should have occurred won't, due to the way nodes are matched,
11041 * and our artificial boundary. So back off until we find a non-
11042 * problematic character -- one that isn't at the beginning or
11043 * middle of such a fold. (Either it doesn't participate in any
11044 * folds, or appears only in the final position of all the folds it
11045 * does participate in.) A better solution with far fewer false
11046 * positives, and that would fill the nodes more completely, would
11047 * be to actually have available all the multi-character folds to
11048 * test against, and to back-off only far enough to be sure that
11049 * this node isn't ending with a partial one. <upper_parse> is set
11050 * further below (if we need to reparse the node) to include just
11051 * up through that final non-problematic character that this code
11052 * identifies, so when it is set to less than the full node, we can
11053 * skip the rest of this */
11054 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
11056 const STRLEN full_len = len;
11058 assert(len >= MAX_NODE_STRING_SIZE);
11060 /* Here, <s> points to the final byte of the final character.
11061 * Look backwards through the string until find a non-
11062 * problematic character */
11066 /* These two have no multi-char folds to non-UTF characters
11068 if (ASCII_FOLD_RESTRICTED || LOC) {
11072 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
11076 if (! PL_NonL1NonFinalFold) {
11077 PL_NonL1NonFinalFold = _new_invlist_C_array(
11078 NonL1_Perl_Non_Final_Folds_invlist);
11081 /* Point to the first byte of the final character */
11082 s = (char *) utf8_hop((U8 *) s, -1);
11084 while (s >= s0) { /* Search backwards until find
11085 non-problematic char */
11086 if (UTF8_IS_INVARIANT(*s)) {
11088 /* There are no ascii characters that participate
11089 * in multi-char folds under /aa. In EBCDIC, the
11090 * non-ascii invariants are all control characters,
11091 * so don't ever participate in any folds. */
11092 if (ASCII_FOLD_RESTRICTED
11093 || ! IS_NON_FINAL_FOLD(*s))
11098 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
11100 /* No Latin1 characters participate in multi-char
11101 * folds under /l */
11103 || ! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_UNI(
11109 else if (! _invlist_contains_cp(
11110 PL_NonL1NonFinalFold,
11111 valid_utf8_to_uvchr((U8 *) s, NULL)))
11116 /* Here, the current character is problematic in that
11117 * it does occur in the non-final position of some
11118 * fold, so try the character before it, but have to
11119 * special case the very first byte in the string, so
11120 * we don't read outside the string */
11121 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
11122 } /* End of loop backwards through the string */
11124 /* If there were only problematic characters in the string,
11125 * <s> will point to before s0, in which case the length
11126 * should be 0, otherwise include the length of the
11127 * non-problematic character just found */
11128 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
11131 /* Here, have found the final character, if any, that is
11132 * non-problematic as far as ending the node without splitting
11133 * it across a potential multi-char fold. <len> contains the
11134 * number of bytes in the node up-to and including that
11135 * character, or is 0 if there is no such character, meaning
11136 * the whole node contains only problematic characters. In
11137 * this case, give up and just take the node as-is. We can't
11143 /* Here, the node does contain some characters that aren't
11144 * problematic. If one such is the final character in the
11145 * node, we are done */
11146 if (len == full_len) {
11149 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
11151 /* If the final character is problematic, but the
11152 * penultimate is not, back-off that last character to
11153 * later start a new node with it */
11158 /* Here, the final non-problematic character is earlier
11159 * in the input than the penultimate character. What we do
11160 * is reparse from the beginning, going up only as far as
11161 * this final ok one, thus guaranteeing that the node ends
11162 * in an acceptable character. The reason we reparse is
11163 * that we know how far in the character is, but we don't
11164 * know how to correlate its position with the input parse.
11165 * An alternate implementation would be to build that
11166 * correlation as we go along during the original parse,
11167 * but that would entail extra work for every node, whereas
11168 * this code gets executed only when the string is too
11169 * large for the node, and the final two characters are
11170 * problematic, an infrequent occurrence. Yet another
11171 * possible strategy would be to save the tail of the
11172 * string, and the next time regatom is called, initialize
11173 * with that. The problem with this is that unless you
11174 * back off one more character, you won't be guaranteed
11175 * regatom will get called again, unless regbranch,
11176 * regpiece ... are also changed. If you do back off that
11177 * extra character, so that there is input guaranteed to
11178 * force calling regatom, you can't handle the case where
11179 * just the first character in the node is acceptable. I
11180 * (khw) decided to try this method which doesn't have that
11181 * pitfall; if performance issues are found, we can do a
11182 * combination of the current approach plus that one */
11188 } /* End of verifying node ends with an appropriate char */
11190 loopdone: /* Jumped to when encounters something that shouldn't be in
11193 /* If 'maybe_exact' is still set here, means there are no
11194 * code points in the node that participate in folds */
11195 if (FOLD && maybe_exact) {
11199 /* I (khw) don't know if you can get here with zero length, but the
11200 * old code handled this situation by creating a zero-length EXACT
11201 * node. Might as well be NOTHING instead */
11206 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
11209 RExC_parse = p - 1;
11210 Set_Node_Cur_Length(ret); /* MJD */
11211 nextchar(pRExC_state);
11213 /* len is STRLEN which is unsigned, need to copy to signed */
11216 vFAIL("Internal disaster");
11219 } /* End of label 'defchar:' */
11221 } /* End of giant switch on input character */
11227 S_regwhite( RExC_state_t *pRExC_state, char *p )
11229 const char *e = RExC_end;
11231 PERL_ARGS_ASSERT_REGWHITE;
11236 else if (*p == '#') {
11239 if (*p++ == '\n') {
11245 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11254 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
11256 /* Returns the next non-pattern-white space, non-comment character (the
11257 * latter only if 'recognize_comment is true) in the string p, which is
11258 * ended by RExC_end. If there is no line break ending a comment,
11259 * RExC_seen has added the REG_SEEN_RUN_ON_COMMENT flag; */
11260 const char *e = RExC_end;
11262 PERL_ARGS_ASSERT_REGPATWS;
11266 if ((len = is_PATWS_safe(p, e, UTF))) {
11269 else if (recognize_comment && *p == '#') {
11273 if (is_LNBREAK_safe(p, e, UTF)) {
11279 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11287 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
11288 Character classes ([:foo:]) can also be negated ([:^foo:]).
11289 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
11290 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
11291 but trigger failures because they are currently unimplemented. */
11293 #define POSIXCC_DONE(c) ((c) == ':')
11294 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
11295 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
11297 PERL_STATIC_INLINE I32
11298 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
11301 I32 namedclass = OOB_NAMEDCLASS;
11303 PERL_ARGS_ASSERT_REGPPOSIXCC;
11305 if (value == '[' && RExC_parse + 1 < RExC_end &&
11306 /* I smell either [: or [= or [. -- POSIX has been here, right? */
11307 POSIXCC(UCHARAT(RExC_parse)))
11309 const char c = UCHARAT(RExC_parse);
11310 char* const s = RExC_parse++;
11312 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
11314 if (RExC_parse == RExC_end) {
11317 /* Try to give a better location for the error (than the end of
11318 * the string) by looking for the matching ']' */
11320 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
11323 vFAIL2("Unmatched '%c' in POSIX class", c);
11325 /* Grandfather lone [:, [=, [. */
11329 const char* const t = RExC_parse++; /* skip over the c */
11332 if (UCHARAT(RExC_parse) == ']') {
11333 const char *posixcc = s + 1;
11334 RExC_parse++; /* skip over the ending ] */
11337 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
11338 const I32 skip = t - posixcc;
11340 /* Initially switch on the length of the name. */
11343 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
11344 this is the Perl \w
11346 namedclass = ANYOF_WORDCHAR;
11349 /* Names all of length 5. */
11350 /* alnum alpha ascii blank cntrl digit graph lower
11351 print punct space upper */
11352 /* Offset 4 gives the best switch position. */
11353 switch (posixcc[4]) {
11355 if (memEQ(posixcc, "alph", 4)) /* alpha */
11356 namedclass = ANYOF_ALPHA;
11359 if (memEQ(posixcc, "spac", 4)) /* space */
11360 namedclass = ANYOF_PSXSPC;
11363 if (memEQ(posixcc, "grap", 4)) /* graph */
11364 namedclass = ANYOF_GRAPH;
11367 if (memEQ(posixcc, "asci", 4)) /* ascii */
11368 namedclass = ANYOF_ASCII;
11371 if (memEQ(posixcc, "blan", 4)) /* blank */
11372 namedclass = ANYOF_BLANK;
11375 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
11376 namedclass = ANYOF_CNTRL;
11379 if (memEQ(posixcc, "alnu", 4)) /* alnum */
11380 namedclass = ANYOF_ALPHANUMERIC;
11383 if (memEQ(posixcc, "lowe", 4)) /* lower */
11384 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
11385 else if (memEQ(posixcc, "uppe", 4)) /* upper */
11386 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
11389 if (memEQ(posixcc, "digi", 4)) /* digit */
11390 namedclass = ANYOF_DIGIT;
11391 else if (memEQ(posixcc, "prin", 4)) /* print */
11392 namedclass = ANYOF_PRINT;
11393 else if (memEQ(posixcc, "punc", 4)) /* punct */
11394 namedclass = ANYOF_PUNCT;
11399 if (memEQ(posixcc, "xdigit", 6))
11400 namedclass = ANYOF_XDIGIT;
11404 if (namedclass == OOB_NAMEDCLASS)
11405 Simple_vFAIL3("POSIX class [:%.*s:] unknown",
11408 /* The #defines are structured so each complement is +1 to
11409 * the normal one */
11413 assert (posixcc[skip] == ':');
11414 assert (posixcc[skip+1] == ']');
11415 } else if (!SIZE_ONLY) {
11416 /* [[=foo=]] and [[.foo.]] are still future. */
11418 /* adjust RExC_parse so the warning shows after
11419 the class closes */
11420 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
11422 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11425 /* Maternal grandfather:
11426 * "[:" ending in ":" but not in ":]" */
11428 vFAIL("Unmatched '[' in POSIX class");
11431 /* Grandfather lone [:, [=, [. */
11441 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
11443 /* This applies some heuristics at the current parse position (which should
11444 * be at a '[') to see if what follows might be intended to be a [:posix:]
11445 * class. It returns true if it really is a posix class, of course, but it
11446 * also can return true if it thinks that what was intended was a posix
11447 * class that didn't quite make it.
11449 * It will return true for
11451 * [:alphanumerics] (as long as the ] isn't followed immediately by a
11452 * ')' indicating the end of the (?[
11453 * [:any garbage including %^&$ punctuation:]
11455 * This is designed to be called only from S_handle_regex_sets; it could be
11456 * easily adapted to be called from the spot at the beginning of regclass()
11457 * that checks to see in a normal bracketed class if the surrounding []
11458 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
11459 * change long-standing behavior, so I (khw) didn't do that */
11460 char* p = RExC_parse + 1;
11461 char first_char = *p;
11463 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
11465 assert(*(p - 1) == '[');
11467 if (! POSIXCC(first_char)) {
11472 while (p < RExC_end && isWORDCHAR(*p)) p++;
11474 if (p >= RExC_end) {
11478 if (p - RExC_parse > 2 /* Got at least 1 word character */
11479 && (*p == first_char
11480 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
11485 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
11488 && p - RExC_parse > 2 /* [:] evaluates to colon;
11489 [::] is a bad posix class. */
11490 && first_char == *(p - 1));
11494 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist, I32 *flagp, U32 depth,
11495 char * const oregcomp_parse)
11497 /* Handle the (?[...]) construct to do set operations */
11500 UV start, end; /* End points of code point ranges */
11502 char *save_end, *save_parse;
11507 const bool save_fold = FOLD;
11509 GET_RE_DEBUG_FLAGS_DECL;
11511 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
11514 vFAIL("(?[...]) not valid in locale");
11516 RExC_uni_semantics = 1;
11518 /* This will return only an ANYOF regnode, or (unlikely) something smaller
11519 * (such as EXACT). Thus we can skip most everything if just sizing. We
11520 * call regclass to handle '[]' so as to not have to reinvent its parsing
11521 * rules here (throwing away the size it computes each time). And, we exit
11522 * upon an unescaped ']' that isn't one ending a regclass. To do both
11523 * these things, we need to realize that something preceded by a backslash
11524 * is escaped, so we have to keep track of backslashes */
11527 Perl_ck_warner_d(aTHX_
11528 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
11529 "The regex_sets feature is experimental" REPORT_LOCATION,
11530 (int) (RExC_parse - RExC_precomp) , RExC_precomp, RExC_parse);
11532 while (RExC_parse < RExC_end) {
11533 SV* current = NULL;
11534 RExC_parse = regpatws(pRExC_state, RExC_parse,
11535 TRUE); /* means recognize comments */
11536 switch (*RExC_parse) {
11540 /* Skip the next byte (which could cause us to end up in
11541 * the middle of a UTF-8 character, but since none of those
11542 * are confusable with anything we currently handle in this
11543 * switch (invariants all), it's safe. We'll just hit the
11544 * default: case next time and keep on incrementing until
11545 * we find one of the invariants we do handle. */
11550 /* If this looks like it is a [:posix:] class, leave the
11551 * parse pointer at the '[' to fool regclass() into
11552 * thinking it is part of a '[[:posix:]]'. That function
11553 * will use strict checking to force a syntax error if it
11554 * doesn't work out to a legitimate class */
11555 bool is_posix_class
11556 = could_it_be_a_POSIX_class(pRExC_state);
11557 if (! is_posix_class) {
11561 /* regclass() can only return RESTART_UTF8 if multi-char
11562 folds are allowed. */
11563 if (!regclass(pRExC_state, flagp,depth+1,
11564 is_posix_class, /* parse the whole char
11565 class only if not a
11567 FALSE, /* don't allow multi-char folds */
11568 TRUE, /* silence non-portable warnings. */
11570 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X",
11573 /* function call leaves parse pointing to the ']', except
11574 * if we faked it */
11575 if (is_posix_class) {
11579 SvREFCNT_dec(current); /* In case it returned something */
11585 if (RExC_parse < RExC_end
11586 && *RExC_parse == ')')
11588 node = reganode(pRExC_state, ANYOF, 0);
11589 RExC_size += ANYOF_SKIP;
11590 nextchar(pRExC_state);
11591 Set_Node_Length(node,
11592 RExC_parse - oregcomp_parse + 1); /* MJD */
11601 FAIL("Syntax error in (?[...])");
11604 /* Pass 2 only after this. Everything in this construct is a
11605 * metacharacter. Operands begin with either a '\' (for an escape
11606 * sequence), or a '[' for a bracketed character class. Any other
11607 * character should be an operator, or parenthesis for grouping. Both
11608 * types of operands are handled by calling regclass() to parse them. It
11609 * is called with a parameter to indicate to return the computed inversion
11610 * list. The parsing here is implemented via a stack. Each entry on the
11611 * stack is a single character representing one of the operators, or the
11612 * '('; or else a pointer to an operand inversion list. */
11614 #define IS_OPERAND(a) (! SvIOK(a))
11616 /* The stack starts empty. It is a syntax error if the first thing parsed
11617 * is a binary operator; everything else is pushed on the stack. When an
11618 * operand is parsed, the top of the stack is examined. If it is a binary
11619 * operator, the item before it should be an operand, and both are replaced
11620 * by the result of doing that operation on the new operand and the one on
11621 * the stack. Thus a sequence of binary operands is reduced to a single
11622 * one before the next one is parsed.
11624 * A unary operator may immediately follow a binary in the input, for
11627 * When an operand is parsed and the top of the stack is a unary operator,
11628 * the operation is performed, and then the stack is rechecked to see if
11629 * this new operand is part of a binary operation; if so, it is handled as
11632 * A '(' is simply pushed on the stack; it is valid only if the stack is
11633 * empty, or the top element of the stack is an operator or another '('
11634 * (for which the parenthesized expression will become an operand). By the
11635 * time the corresponding ')' is parsed everything in between should have
11636 * been parsed and evaluated to a single operand (or else is a syntax
11637 * error), and is handled as a regular operand */
11641 while (RExC_parse < RExC_end) {
11642 I32 top_index = av_tindex(stack);
11644 SV* current = NULL;
11646 /* Skip white space */
11647 RExC_parse = regpatws(pRExC_state, RExC_parse,
11648 TRUE); /* means recognize comments */
11649 if (RExC_parse >= RExC_end) {
11650 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
11652 if ((curchar = UCHARAT(RExC_parse)) == ']') {
11659 if (av_tindex(stack) >= 0 /* This makes sure that we can
11660 safely subtract 1 from
11661 RExC_parse in the next clause.
11662 If we have something on the
11663 stack, we have parsed something
11665 && UCHARAT(RExC_parse - 1) == '('
11666 && RExC_parse < RExC_end)
11668 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
11669 * This happens when we have some thing like
11671 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
11673 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
11675 * Here we would be handling the interpolated
11676 * '$thai_or_lao'. We handle this by a recursive call to
11677 * ourselves which returns the inversion list the
11678 * interpolated expression evaluates to. We use the flags
11679 * from the interpolated pattern. */
11680 U32 save_flags = RExC_flags;
11681 const char * const save_parse = ++RExC_parse;
11683 parse_lparen_question_flags(pRExC_state);
11685 if (RExC_parse == save_parse /* Makes sure there was at
11686 least one flag (or this
11687 embedding wasn't compiled)
11689 || RExC_parse >= RExC_end - 4
11690 || UCHARAT(RExC_parse) != ':'
11691 || UCHARAT(++RExC_parse) != '('
11692 || UCHARAT(++RExC_parse) != '?'
11693 || UCHARAT(++RExC_parse) != '[')
11696 /* In combination with the above, this moves the
11697 * pointer to the point just after the first erroneous
11698 * character (or if there are no flags, to where they
11699 * should have been) */
11700 if (RExC_parse >= RExC_end - 4) {
11701 RExC_parse = RExC_end;
11703 else if (RExC_parse != save_parse) {
11704 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11706 vFAIL("Expecting '(?flags:(?[...'");
11709 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
11710 depth+1, oregcomp_parse);
11712 /* Here, 'current' contains the embedded expression's
11713 * inversion list, and RExC_parse points to the trailing
11714 * ']'; the next character should be the ')' which will be
11715 * paired with the '(' that has been put on the stack, so
11716 * the whole embedded expression reduces to '(operand)' */
11719 RExC_flags = save_flags;
11720 goto handle_operand;
11725 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11726 vFAIL("Unexpected character");
11729 /* regclass() can only return RESTART_UTF8 if multi-char
11730 folds are allowed. */
11731 if (!regclass(pRExC_state, flagp,depth+1,
11732 TRUE, /* means parse just the next thing */
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 /* regclass() will return with parsing just the \ sequence,
11739 * leaving the parse pointer at the next thing to parse */
11741 goto handle_operand;
11743 case '[': /* Is a bracketed character class */
11745 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
11747 if (! is_posix_class) {
11751 /* regclass() can only return RESTART_UTF8 if multi-char
11752 folds are allowed. */
11753 if(!regclass(pRExC_state, flagp,depth+1,
11754 is_posix_class, /* parse the whole char class
11755 only if not a posix class */
11756 FALSE, /* don't allow multi-char folds */
11757 FALSE, /* don't silence non-portable warnings. */
11759 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X",
11761 /* function call leaves parse pointing to the ']', except if we
11763 if (is_posix_class) {
11767 goto handle_operand;
11776 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
11777 || ! IS_OPERAND(*top_ptr))
11780 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
11782 av_push(stack, newSVuv(curchar));
11786 av_push(stack, newSVuv(curchar));
11790 if (top_index >= 0) {
11791 top_ptr = av_fetch(stack, top_index, FALSE);
11793 if (IS_OPERAND(*top_ptr)) {
11795 vFAIL("Unexpected '(' with no preceding operator");
11798 av_push(stack, newSVuv(curchar));
11805 || ! (current = av_pop(stack))
11806 || ! IS_OPERAND(current)
11807 || ! (lparen = av_pop(stack))
11808 || IS_OPERAND(lparen)
11809 || SvUV(lparen) != '(')
11812 vFAIL("Unexpected ')'");
11815 SvREFCNT_dec_NN(lparen);
11822 /* Here, we have an operand to process, in 'current' */
11824 if (top_index < 0) { /* Just push if stack is empty */
11825 av_push(stack, current);
11828 SV* top = av_pop(stack);
11829 char current_operator;
11831 if (IS_OPERAND(top)) {
11832 vFAIL("Operand with no preceding operator");
11834 current_operator = (char) SvUV(top);
11835 switch (current_operator) {
11836 case '(': /* Push the '(' back on followed by the new
11838 av_push(stack, top);
11839 av_push(stack, current);
11840 SvREFCNT_inc(top); /* Counters the '_dec' done
11841 just after the 'break', so
11842 it doesn't get wrongly freed
11847 _invlist_invert(current);
11849 /* Unlike binary operators, the top of the stack,
11850 * now that this unary one has been popped off, may
11851 * legally be an operator, and we now have operand
11854 SvREFCNT_dec_NN(top);
11855 goto handle_operand;
11858 _invlist_intersection(av_pop(stack),
11861 av_push(stack, current);
11866 _invlist_union(av_pop(stack), current, ¤t);
11867 av_push(stack, current);
11871 _invlist_subtract(av_pop(stack), current, ¤t);
11872 av_push(stack, current);
11875 case '^': /* The union minus the intersection */
11881 element = av_pop(stack);
11882 _invlist_union(element, current, &u);
11883 _invlist_intersection(element, current, &i);
11884 _invlist_subtract(u, i, ¤t);
11885 av_push(stack, current);
11886 SvREFCNT_dec_NN(i);
11887 SvREFCNT_dec_NN(u);
11888 SvREFCNT_dec_NN(element);
11893 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
11895 SvREFCNT_dec_NN(top);
11899 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11902 if (av_tindex(stack) < 0 /* Was empty */
11903 || ((final = av_pop(stack)) == NULL)
11904 || ! IS_OPERAND(final)
11905 || av_tindex(stack) >= 0) /* More left on stack */
11907 vFAIL("Incomplete expression within '(?[ ])'");
11910 /* Here, 'final' is the resultant inversion list from evaluating the
11911 * expression. Return it if so requested */
11912 if (return_invlist) {
11913 *return_invlist = final;
11917 /* Otherwise generate a resultant node, based on 'final'. regclass() is
11918 * expecting a string of ranges and individual code points */
11919 invlist_iterinit(final);
11920 result_string = newSVpvs("");
11921 while (invlist_iternext(final, &start, &end)) {
11922 if (start == end) {
11923 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
11926 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
11931 save_parse = RExC_parse;
11932 RExC_parse = SvPV(result_string, len);
11933 save_end = RExC_end;
11934 RExC_end = RExC_parse + len;
11936 /* We turn off folding around the call, as the class we have constructed
11937 * already has all folding taken into consideration, and we don't want
11938 * regclass() to add to that */
11939 RExC_flags &= ~RXf_PMf_FOLD;
11940 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
11942 node = regclass(pRExC_state, flagp,depth+1,
11943 FALSE, /* means parse the whole char class */
11944 FALSE, /* don't allow multi-char folds */
11945 TRUE, /* silence non-portable warnings. The above may very
11946 well have generated non-portable code points, but
11947 they're valid on this machine */
11950 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
11953 RExC_flags |= RXf_PMf_FOLD;
11955 RExC_parse = save_parse + 1;
11956 RExC_end = save_end;
11957 SvREFCNT_dec_NN(final);
11958 SvREFCNT_dec_NN(result_string);
11959 SvREFCNT_dec_NN(stack);
11961 nextchar(pRExC_state);
11962 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
11967 /* The names of properties whose definitions are not known at compile time are
11968 * stored in this SV, after a constant heading. So if the length has been
11969 * changed since initialization, then there is a run-time definition. */
11970 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION (SvCUR(listsv) != initial_listsv_len)
11973 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
11974 const bool stop_at_1, /* Just parse the next thing, don't
11975 look for a full character class */
11976 bool allow_multi_folds,
11977 const bool silence_non_portable, /* Don't output warnings
11980 SV** ret_invlist) /* Return an inversion list, not a node */
11982 /* parse a bracketed class specification. Most of these will produce an
11983 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
11984 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
11985 * under /i with multi-character folds: it will be rewritten following the
11986 * paradigm of this example, where the <multi-fold>s are characters which
11987 * fold to multiple character sequences:
11988 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
11989 * gets effectively rewritten as:
11990 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
11991 * reg() gets called (recursively) on the rewritten version, and this
11992 * function will return what it constructs. (Actually the <multi-fold>s
11993 * aren't physically removed from the [abcdefghi], it's just that they are
11994 * ignored in the recursion by means of a flag:
11995 * <RExC_in_multi_char_class>.)
11997 * ANYOF nodes contain a bit map for the first 256 characters, with the
11998 * corresponding bit set if that character is in the list. For characters
11999 * above 255, a range list or swash is used. There are extra bits for \w,
12000 * etc. in locale ANYOFs, as what these match is not determinable at
12003 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
12004 * to be restarted. This can only happen if ret_invlist is non-NULL.
12008 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
12010 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
12013 IV namedclass = OOB_NAMEDCLASS;
12014 char *rangebegin = NULL;
12015 bool need_class = 0;
12017 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
12018 than just initialized. */
12019 SV* properties = NULL; /* Code points that match \p{} \P{} */
12020 SV* posixes = NULL; /* Code points that match classes like, [:word:],
12021 extended beyond the Latin1 range */
12022 UV element_count = 0; /* Number of distinct elements in the class.
12023 Optimizations may be possible if this is tiny */
12024 AV * multi_char_matches = NULL; /* Code points that fold to more than one
12025 character; used under /i */
12027 char * stop_ptr = RExC_end; /* where to stop parsing */
12028 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
12030 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
12032 /* Unicode properties are stored in a swash; this holds the current one
12033 * being parsed. If this swash is the only above-latin1 component of the
12034 * character class, an optimization is to pass it directly on to the
12035 * execution engine. Otherwise, it is set to NULL to indicate that there
12036 * are other things in the class that have to be dealt with at execution
12038 SV* swash = NULL; /* Code points that match \p{} \P{} */
12040 /* Set if a component of this character class is user-defined; just passed
12041 * on to the engine */
12042 bool has_user_defined_property = FALSE;
12044 /* inversion list of code points this node matches only when the target
12045 * string is in UTF-8. (Because is under /d) */
12046 SV* depends_list = NULL;
12048 /* inversion list of code points this node matches. For much of the
12049 * function, it includes only those that match regardless of the utf8ness
12050 * of the target string */
12051 SV* cp_list = NULL;
12054 /* In a range, counts how many 0-2 of the ends of it came from literals,
12055 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
12056 UV literal_endpoint = 0;
12058 bool invert = FALSE; /* Is this class to be complemented */
12060 /* Is there any thing like \W or [:^digit:] that matches above the legal
12061 * Unicode range? */
12062 bool runtime_posix_matches_above_Unicode = FALSE;
12064 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
12065 case we need to change the emitted regop to an EXACT. */
12066 const char * orig_parse = RExC_parse;
12067 const I32 orig_size = RExC_size;
12068 GET_RE_DEBUG_FLAGS_DECL;
12070 PERL_ARGS_ASSERT_REGCLASS;
12072 PERL_UNUSED_ARG(depth);
12075 DEBUG_PARSE("clas");
12077 /* Assume we are going to generate an ANYOF node. */
12078 ret = reganode(pRExC_state, ANYOF, 0);
12081 RExC_size += ANYOF_SKIP;
12082 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
12085 ANYOF_FLAGS(ret) = 0;
12087 RExC_emit += ANYOF_SKIP;
12089 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
12091 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
12092 initial_listsv_len = SvCUR(listsv);
12093 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
12097 RExC_parse = regpatws(pRExC_state, RExC_parse,
12098 FALSE /* means don't recognize comments */);
12101 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
12104 allow_multi_folds = FALSE;
12107 RExC_parse = regpatws(pRExC_state, RExC_parse,
12108 FALSE /* means don't recognize comments */);
12112 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
12113 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
12114 const char *s = RExC_parse;
12115 const char c = *s++;
12117 while (isWORDCHAR(*s))
12119 if (*s && c == *s && s[1] == ']') {
12120 SAVEFREESV(RExC_rx_sv);
12122 "POSIX syntax [%c %c] belongs inside character classes",
12124 (void)ReREFCNT_inc(RExC_rx_sv);
12128 /* If the caller wants us to just parse a single element, accomplish this
12129 * by faking the loop ending condition */
12130 if (stop_at_1 && RExC_end > RExC_parse) {
12131 stop_ptr = RExC_parse + 1;
12134 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
12135 if (UCHARAT(RExC_parse) == ']')
12136 goto charclassloop;
12140 if (RExC_parse >= stop_ptr) {
12145 RExC_parse = regpatws(pRExC_state, RExC_parse,
12146 FALSE /* means don't recognize comments */);
12149 if (UCHARAT(RExC_parse) == ']') {
12155 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
12156 save_value = value;
12157 save_prevvalue = prevvalue;
12160 rangebegin = RExC_parse;
12164 value = utf8n_to_uvchr((U8*)RExC_parse,
12165 RExC_end - RExC_parse,
12166 &numlen, UTF8_ALLOW_DEFAULT);
12167 RExC_parse += numlen;
12170 value = UCHARAT(RExC_parse++);
12173 && RExC_parse < RExC_end
12174 && POSIXCC(UCHARAT(RExC_parse)))
12176 namedclass = regpposixcc(pRExC_state, value, strict);
12178 else if (value == '\\') {
12180 value = utf8n_to_uvchr((U8*)RExC_parse,
12181 RExC_end - RExC_parse,
12182 &numlen, UTF8_ALLOW_DEFAULT);
12183 RExC_parse += numlen;
12186 value = UCHARAT(RExC_parse++);
12188 /* Some compilers cannot handle switching on 64-bit integer
12189 * values, therefore value cannot be an UV. Yes, this will
12190 * be a problem later if we want switch on Unicode.
12191 * A similar issue a little bit later when switching on
12192 * namedclass. --jhi */
12194 /* If the \ is escaping white space when white space is being
12195 * skipped, it means that that white space is wanted literally, and
12196 * is already in 'value'. Otherwise, need to translate the escape
12197 * into what it signifies. */
12198 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
12200 case 'w': namedclass = ANYOF_WORDCHAR; break;
12201 case 'W': namedclass = ANYOF_NWORDCHAR; break;
12202 case 's': namedclass = ANYOF_SPACE; break;
12203 case 'S': namedclass = ANYOF_NSPACE; break;
12204 case 'd': namedclass = ANYOF_DIGIT; break;
12205 case 'D': namedclass = ANYOF_NDIGIT; break;
12206 case 'v': namedclass = ANYOF_VERTWS; break;
12207 case 'V': namedclass = ANYOF_NVERTWS; break;
12208 case 'h': namedclass = ANYOF_HORIZWS; break;
12209 case 'H': namedclass = ANYOF_NHORIZWS; break;
12210 case 'N': /* Handle \N{NAME} in class */
12212 /* We only pay attention to the first char of
12213 multichar strings being returned. I kinda wonder
12214 if this makes sense as it does change the behaviour
12215 from earlier versions, OTOH that behaviour was broken
12217 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
12218 TRUE, /* => charclass */
12221 if (*flagp & RESTART_UTF8)
12222 FAIL("panic: grok_bslash_N set RESTART_UTF8");
12232 /* We will handle any undefined properties ourselves */
12233 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF;
12235 if (RExC_parse >= RExC_end)
12236 vFAIL2("Empty \\%c{}", (U8)value);
12237 if (*RExC_parse == '{') {
12238 const U8 c = (U8)value;
12239 e = strchr(RExC_parse++, '}');
12241 vFAIL2("Missing right brace on \\%c{}", c);
12242 while (isSPACE(UCHARAT(RExC_parse)))
12244 if (e == RExC_parse)
12245 vFAIL2("Empty \\%c{}", c);
12246 n = e - RExC_parse;
12247 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
12258 if (UCHARAT(RExC_parse) == '^') {
12261 /* toggle. (The rhs xor gets the single bit that
12262 * differs between P and p; the other xor inverts just
12264 value ^= 'P' ^ 'p';
12266 while (isSPACE(UCHARAT(RExC_parse))) {
12271 /* Try to get the definition of the property into
12272 * <invlist>. If /i is in effect, the effective property
12273 * will have its name be <__NAME_i>. The design is
12274 * discussed in commit
12275 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
12276 Newx(name, n + sizeof("_i__\n"), char);
12278 sprintf(name, "%s%.*s%s\n",
12279 (FOLD) ? "__" : "",
12285 /* Look up the property name, and get its swash and
12286 * inversion list, if the property is found */
12288 SvREFCNT_dec_NN(swash);
12290 swash = _core_swash_init("utf8", name, &PL_sv_undef,
12293 NULL, /* No inversion list */
12296 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
12298 SvREFCNT_dec_NN(swash);
12302 /* Here didn't find it. It could be a user-defined
12303 * property that will be available at run-time. If we
12304 * accept only compile-time properties, is an error;
12305 * otherwise add it to the list for run-time look up */
12307 RExC_parse = e + 1;
12308 vFAIL3("Property '%.*s' is unknown", (int) n, name);
12310 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s\n",
12311 (value == 'p' ? '+' : '!'),
12313 has_user_defined_property = TRUE;
12315 /* We don't know yet, so have to assume that the
12316 * property could match something in the Latin1 range,
12317 * hence something that isn't utf8. Note that this
12318 * would cause things in <depends_list> to match
12319 * inappropriately, except that any \p{}, including
12320 * this one forces Unicode semantics, which means there
12321 * is <no depends_list> */
12322 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
12326 /* Here, did get the swash and its inversion list. If
12327 * the swash is from a user-defined property, then this
12328 * whole character class should be regarded as such */
12329 has_user_defined_property =
12331 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY);
12333 /* Invert if asking for the complement */
12334 if (value == 'P') {
12335 _invlist_union_complement_2nd(properties,
12339 /* The swash can't be used as-is, because we've
12340 * inverted things; delay removing it to here after
12341 * have copied its invlist above */
12342 SvREFCNT_dec_NN(swash);
12346 _invlist_union(properties, invlist, &properties);
12351 RExC_parse = e + 1;
12352 namedclass = ANYOF_UNIPROP; /* no official name, but it's
12355 /* \p means they want Unicode semantics */
12356 RExC_uni_semantics = 1;
12359 case 'n': value = '\n'; break;
12360 case 'r': value = '\r'; break;
12361 case 't': value = '\t'; break;
12362 case 'f': value = '\f'; break;
12363 case 'b': value = '\b'; break;
12364 case 'e': value = ASCII_TO_NATIVE('\033');break;
12365 case 'a': value = ASCII_TO_NATIVE('\007');break;
12367 RExC_parse--; /* function expects to be pointed at the 'o' */
12369 const char* error_msg;
12370 bool valid = grok_bslash_o(&RExC_parse,
12373 SIZE_ONLY, /* warnings in pass
12376 silence_non_portable,
12382 if (PL_encoding && value < 0x100) {
12383 goto recode_encoding;
12387 RExC_parse--; /* function expects to be pointed at the 'x' */
12389 const char* error_msg;
12390 bool valid = grok_bslash_x(&RExC_parse,
12393 TRUE, /* Output warnings */
12395 silence_non_portable,
12401 if (PL_encoding && value < 0x100)
12402 goto recode_encoding;
12405 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
12407 case '0': case '1': case '2': case '3': case '4':
12408 case '5': case '6': case '7':
12410 /* Take 1-3 octal digits */
12411 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12412 numlen = (strict) ? 4 : 3;
12413 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
12414 RExC_parse += numlen;
12417 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12418 vFAIL("Need exactly 3 octal digits");
12420 else if (! SIZE_ONLY /* like \08, \178 */
12422 && RExC_parse < RExC_end
12423 && isDIGIT(*RExC_parse)
12424 && ckWARN(WARN_REGEXP))
12426 SAVEFREESV(RExC_rx_sv);
12427 reg_warn_non_literal_string(
12429 form_short_octal_warning(RExC_parse, numlen));
12430 (void)ReREFCNT_inc(RExC_rx_sv);
12433 if (PL_encoding && value < 0x100)
12434 goto recode_encoding;
12438 if (! RExC_override_recoding) {
12439 SV* enc = PL_encoding;
12440 value = reg_recode((const char)(U8)value, &enc);
12443 vFAIL("Invalid escape in the specified encoding");
12445 else if (SIZE_ONLY) {
12446 ckWARNreg(RExC_parse,
12447 "Invalid escape in the specified encoding");
12453 /* Allow \_ to not give an error */
12454 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
12456 vFAIL2("Unrecognized escape \\%c in character class",
12460 SAVEFREESV(RExC_rx_sv);
12461 ckWARN2reg(RExC_parse,
12462 "Unrecognized escape \\%c in character class passed through",
12464 (void)ReREFCNT_inc(RExC_rx_sv);
12468 } /* End of switch on char following backslash */
12469 } /* end of handling backslash escape sequences */
12472 literal_endpoint++;
12475 /* Here, we have the current token in 'value' */
12477 /* What matches in a locale is not known until runtime. This includes
12478 * what the Posix classes (like \w, [:space:]) match. Room must be
12479 * reserved (one time per class) to store such classes, either if Perl
12480 * is compiled so that locale nodes always should have this space, or
12481 * if there is such class info to be stored. The space will contain a
12482 * bit for each named class that is to be matched against. This isn't
12483 * needed for \p{} and pseudo-classes, as they are not affected by
12484 * locale, and hence are dealt with separately */
12487 && (ANYOF_LOCALE == ANYOF_CLASS
12488 || (namedclass > OOB_NAMEDCLASS && namedclass < ANYOF_MAX)))
12492 RExC_size += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12495 RExC_emit += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12496 ANYOF_CLASS_ZERO(ret);
12498 ANYOF_FLAGS(ret) |= ANYOF_CLASS;
12501 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
12503 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
12504 * literal, as is the character that began the false range, i.e.
12505 * the 'a' in the examples */
12508 const int w = (RExC_parse >= rangebegin)
12509 ? RExC_parse - rangebegin
12512 vFAIL4("False [] range \"%*.*s\"", w, w, rangebegin);
12515 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
12516 ckWARN4reg(RExC_parse,
12517 "False [] range \"%*.*s\"",
12519 (void)ReREFCNT_inc(RExC_rx_sv);
12520 cp_list = add_cp_to_invlist(cp_list, '-');
12521 cp_list = add_cp_to_invlist(cp_list, prevvalue);
12525 range = 0; /* this was not a true range */
12526 element_count += 2; /* So counts for three values */
12530 U8 classnum = namedclass_to_classnum(namedclass);
12531 if (namedclass >= ANYOF_MAX) { /* If a special class */
12532 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
12534 /* Here, should be \h, \H, \v, or \V. Neither /d nor
12535 * /l make a difference in what these match. There
12536 * would be problems if these characters had folds
12537 * other than themselves, as cp_list is subject to
12539 if (classnum != _CC_VERTSPACE) {
12540 assert( namedclass == ANYOF_HORIZWS
12541 || namedclass == ANYOF_NHORIZWS);
12543 /* It turns out that \h is just a synonym for
12545 classnum = _CC_BLANK;
12548 _invlist_union_maybe_complement_2nd(
12550 PL_XPosix_ptrs[classnum],
12551 cBOOL(namedclass % 2), /* Complement if odd
12552 (NHORIZWS, NVERTWS)
12557 else if (classnum == _CC_ASCII) {
12560 ANYOF_CLASS_SET(ret, namedclass);
12563 #endif /* Not isascii(); just use the hard-coded definition for it */
12564 _invlist_union_maybe_complement_2nd(
12567 cBOOL(namedclass % 2), /* Complement if odd
12571 else { /* Garden variety class */
12573 /* The ascii range inversion list */
12574 SV* ascii_source = PL_Posix_ptrs[classnum];
12576 /* The full Latin1 range inversion list */
12577 SV* l1_source = PL_L1Posix_ptrs[classnum];
12579 /* This code is structured into two major clauses. The
12580 * first is for classes whose complete definitions may not
12581 * already be known. It not, the Latin1 definition
12582 * (guaranteed to already known) is used plus code is
12583 * generated to load the rest at run-time (only if needed).
12584 * If the complete definition is known, it drops down to
12585 * the second clause, where the complete definition is
12588 if (classnum < _FIRST_NON_SWASH_CC) {
12590 /* Here, the class has a swash, which may or not
12591 * already be loaded */
12593 /* The name of the property to use to match the full
12594 * eXtended Unicode range swash for this character
12596 const char *Xname = swash_property_names[classnum];
12598 /* If returning the inversion list, we can't defer
12599 * getting this until runtime */
12600 if (ret_invlist && ! PL_utf8_swash_ptrs[classnum]) {
12601 PL_utf8_swash_ptrs[classnum] =
12602 _core_swash_init("utf8", Xname, &PL_sv_undef,
12605 NULL, /* No inversion list */
12606 NULL /* No flags */
12608 assert(PL_utf8_swash_ptrs[classnum]);
12610 if ( ! PL_utf8_swash_ptrs[classnum]) {
12611 if (namedclass % 2 == 0) { /* A non-complemented
12613 /* If not /a matching, there are code points we
12614 * don't know at compile time. Arrange for the
12615 * unknown matches to be loaded at run-time, if
12617 if (! AT_LEAST_ASCII_RESTRICTED) {
12618 Perl_sv_catpvf(aTHX_ listsv, "+utf8::%s\n",
12621 if (LOC) { /* Under locale, set run-time
12623 ANYOF_CLASS_SET(ret, namedclass);
12626 /* Add the current class's code points to
12627 * the running total */
12628 _invlist_union(posixes,
12629 (AT_LEAST_ASCII_RESTRICTED)
12635 else { /* A complemented class */
12636 if (AT_LEAST_ASCII_RESTRICTED) {
12637 /* Under /a should match everything above
12638 * ASCII, plus the complement of the set's
12640 _invlist_union_complement_2nd(posixes,
12645 /* Arrange for the unknown matches to be
12646 * loaded at run-time, if needed */
12647 Perl_sv_catpvf(aTHX_ listsv, "!utf8::%s\n",
12649 runtime_posix_matches_above_Unicode = TRUE;
12651 ANYOF_CLASS_SET(ret, namedclass);
12655 /* We want to match everything in
12656 * Latin1, except those things that
12657 * l1_source matches */
12658 SV* scratch_list = NULL;
12659 _invlist_subtract(PL_Latin1, l1_source,
12662 /* Add the list from this class to the
12665 posixes = scratch_list;
12668 _invlist_union(posixes,
12671 SvREFCNT_dec_NN(scratch_list);
12673 if (DEPENDS_SEMANTICS) {
12675 |= ANYOF_NON_UTF8_LATIN1_ALL;
12680 goto namedclass_done;
12683 /* Here, there is a swash loaded for the class. If no
12684 * inversion list for it yet, get it */
12685 if (! PL_XPosix_ptrs[classnum]) {
12686 PL_XPosix_ptrs[classnum]
12687 = _swash_to_invlist(PL_utf8_swash_ptrs[classnum]);
12691 /* Here there is an inversion list already loaded for the
12694 if (namedclass % 2 == 0) { /* A non-complemented class,
12695 like ANYOF_PUNCT */
12697 /* For non-locale, just add it to any existing list
12699 _invlist_union(posixes,
12700 (AT_LEAST_ASCII_RESTRICTED)
12702 : PL_XPosix_ptrs[classnum],
12705 else { /* Locale */
12706 SV* scratch_list = NULL;
12708 /* For above Latin1 code points, we use the full
12710 _invlist_intersection(PL_AboveLatin1,
12711 PL_XPosix_ptrs[classnum],
12713 /* And set the output to it, adding instead if
12714 * there already is an output. Checking if
12715 * 'posixes' is NULL first saves an extra clone.
12716 * Its reference count will be decremented at the
12717 * next union, etc, or if this is the only
12718 * instance, at the end of the routine */
12720 posixes = scratch_list;
12723 _invlist_union(posixes, scratch_list, &posixes);
12724 SvREFCNT_dec_NN(scratch_list);
12727 #ifndef HAS_ISBLANK
12728 if (namedclass != ANYOF_BLANK) {
12730 /* Set this class in the node for runtime
12732 ANYOF_CLASS_SET(ret, namedclass);
12733 #ifndef HAS_ISBLANK
12736 /* No isblank(), use the hard-coded ASCII-range
12737 * blanks, adding them to the running total. */
12739 _invlist_union(posixes, ascii_source, &posixes);
12744 else { /* A complemented class, like ANYOF_NPUNCT */
12746 _invlist_union_complement_2nd(
12748 (AT_LEAST_ASCII_RESTRICTED)
12750 : PL_XPosix_ptrs[classnum],
12752 /* Under /d, everything in the upper half of the
12753 * Latin1 range matches this complement */
12754 if (DEPENDS_SEMANTICS) {
12755 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL;
12758 else { /* Locale */
12759 SV* scratch_list = NULL;
12760 _invlist_subtract(PL_AboveLatin1,
12761 PL_XPosix_ptrs[classnum],
12764 posixes = scratch_list;
12767 _invlist_union(posixes, scratch_list, &posixes);
12768 SvREFCNT_dec_NN(scratch_list);
12770 #ifndef HAS_ISBLANK
12771 if (namedclass != ANYOF_NBLANK) {
12773 ANYOF_CLASS_SET(ret, namedclass);
12774 #ifndef HAS_ISBLANK
12777 /* Get the list of all code points in Latin1
12778 * that are not ASCII blanks, and add them to
12779 * the running total */
12780 _invlist_subtract(PL_Latin1, ascii_source,
12782 _invlist_union(posixes, scratch_list, &posixes);
12783 SvREFCNT_dec_NN(scratch_list);
12790 continue; /* Go get next character */
12792 } /* end of namedclass \blah */
12794 /* Here, we have a single value. If 'range' is set, it is the ending
12795 * of a range--check its validity. Later, we will handle each
12796 * individual code point in the range. If 'range' isn't set, this
12797 * could be the beginning of a range, so check for that by looking
12798 * ahead to see if the next real character to be processed is the range
12799 * indicator--the minus sign */
12802 RExC_parse = regpatws(pRExC_state, RExC_parse,
12803 FALSE /* means don't recognize comments */);
12807 if (prevvalue > value) /* b-a */ {
12808 const int w = RExC_parse - rangebegin;
12809 Simple_vFAIL4("Invalid [] range \"%*.*s\"", w, w, rangebegin);
12810 range = 0; /* not a valid range */
12814 prevvalue = value; /* save the beginning of the potential range */
12815 if (! stop_at_1 /* Can't be a range if parsing just one thing */
12816 && *RExC_parse == '-')
12818 char* next_char_ptr = RExC_parse + 1;
12819 if (skip_white) { /* Get the next real char after the '-' */
12820 next_char_ptr = regpatws(pRExC_state,
12822 FALSE); /* means don't recognize
12826 /* If the '-' is at the end of the class (just before the ']',
12827 * it is a literal minus; otherwise it is a range */
12828 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
12829 RExC_parse = next_char_ptr;
12831 /* a bad range like \w-, [:word:]- ? */
12832 if (namedclass > OOB_NAMEDCLASS) {
12833 if (strict || ckWARN(WARN_REGEXP)) {
12835 RExC_parse >= rangebegin ?
12836 RExC_parse - rangebegin : 0;
12838 vFAIL4("False [] range \"%*.*s\"",
12843 "False [] range \"%*.*s\"",
12848 cp_list = add_cp_to_invlist(cp_list, '-');
12852 range = 1; /* yeah, it's a range! */
12853 continue; /* but do it the next time */
12858 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
12861 /* non-Latin1 code point implies unicode semantics. Must be set in
12862 * pass1 so is there for the whole of pass 2 */
12864 RExC_uni_semantics = 1;
12867 /* Ready to process either the single value, or the completed range.
12868 * For single-valued non-inverted ranges, we consider the possibility
12869 * of multi-char folds. (We made a conscious decision to not do this
12870 * for the other cases because it can often lead to non-intuitive
12871 * results. For example, you have the peculiar case that:
12872 * "s s" =~ /^[^\xDF]+$/i => Y
12873 * "ss" =~ /^[^\xDF]+$/i => N
12875 * See [perl #89750] */
12876 if (FOLD && allow_multi_folds && value == prevvalue) {
12877 if (value == LATIN_SMALL_LETTER_SHARP_S
12878 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
12881 /* Here <value> is indeed a multi-char fold. Get what it is */
12883 U8 foldbuf[UTF8_MAXBYTES_CASE];
12886 UV folded = _to_uni_fold_flags(
12891 | ((LOC) ? FOLD_FLAGS_LOCALE
12892 : (ASCII_FOLD_RESTRICTED)
12893 ? FOLD_FLAGS_NOMIX_ASCII
12897 /* Here, <folded> should be the first character of the
12898 * multi-char fold of <value>, with <foldbuf> containing the
12899 * whole thing. But, if this fold is not allowed (because of
12900 * the flags), <fold> will be the same as <value>, and should
12901 * be processed like any other character, so skip the special
12903 if (folded != value) {
12905 /* Skip if we are recursed, currently parsing the class
12906 * again. Otherwise add this character to the list of
12907 * multi-char folds. */
12908 if (! RExC_in_multi_char_class) {
12909 AV** this_array_ptr;
12911 STRLEN cp_count = utf8_length(foldbuf,
12912 foldbuf + foldlen);
12913 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
12915 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
12918 if (! multi_char_matches) {
12919 multi_char_matches = newAV();
12922 /* <multi_char_matches> is actually an array of arrays.
12923 * There will be one or two top-level elements: [2],
12924 * and/or [3]. The [2] element is an array, each
12925 * element thereof is a character which folds to two
12926 * characters; likewise for [3]. (Unicode guarantees a
12927 * maximum of 3 characters in any fold.) When we
12928 * rewrite the character class below, we will do so
12929 * such that the longest folds are written first, so
12930 * that it prefers the longest matching strings first.
12931 * This is done even if it turns out that any
12932 * quantifier is non-greedy, out of programmer
12933 * laziness. Tom Christiansen has agreed that this is
12934 * ok. This makes the test for the ligature 'ffi' come
12935 * before the test for 'ff' */
12936 if (av_exists(multi_char_matches, cp_count)) {
12937 this_array_ptr = (AV**) av_fetch(multi_char_matches,
12939 this_array = *this_array_ptr;
12942 this_array = newAV();
12943 av_store(multi_char_matches, cp_count,
12946 av_push(this_array, multi_fold);
12949 /* This element should not be processed further in this
12952 value = save_value;
12953 prevvalue = save_prevvalue;
12959 /* Deal with this element of the class */
12962 cp_list = _add_range_to_invlist(cp_list, prevvalue, value);
12964 SV* this_range = _new_invlist(1);
12965 _append_range_to_invlist(this_range, prevvalue, value);
12967 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
12968 * If this range was specified using something like 'i-j', we want
12969 * to include only the 'i' and the 'j', and not anything in
12970 * between, so exclude non-ASCII, non-alphabetics from it.
12971 * However, if the range was specified with something like
12972 * [\x89-\x91] or [\x89-j], all code points within it should be
12973 * included. literal_endpoint==2 means both ends of the range used
12974 * a literal character, not \x{foo} */
12975 if (literal_endpoint == 2
12976 && (prevvalue >= 'a' && value <= 'z')
12977 || (prevvalue >= 'A' && value <= 'Z'))
12979 _invlist_intersection(this_range, PL_Posix_ptrs[_CC_ALPHA],
12982 _invlist_union(cp_list, this_range, &cp_list);
12983 literal_endpoint = 0;
12987 range = 0; /* this range (if it was one) is done now */
12988 } /* End of loop through all the text within the brackets */
12990 /* If anything in the class expands to more than one character, we have to
12991 * deal with them by building up a substitute parse string, and recursively
12992 * calling reg() on it, instead of proceeding */
12993 if (multi_char_matches) {
12994 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
12997 char *save_end = RExC_end;
12998 char *save_parse = RExC_parse;
12999 bool first_time = TRUE; /* First multi-char occurrence doesn't get
13004 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
13005 because too confusing */
13007 sv_catpv(substitute_parse, "(?:");
13011 /* Look at the longest folds first */
13012 for (cp_count = av_len(multi_char_matches); cp_count > 0; cp_count--) {
13014 if (av_exists(multi_char_matches, cp_count)) {
13015 AV** this_array_ptr;
13018 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13020 while ((this_sequence = av_pop(*this_array_ptr)) !=
13023 if (! first_time) {
13024 sv_catpv(substitute_parse, "|");
13026 first_time = FALSE;
13028 sv_catpv(substitute_parse, SvPVX(this_sequence));
13033 /* If the character class contains anything else besides these
13034 * multi-character folds, have to include it in recursive parsing */
13035 if (element_count) {
13036 sv_catpv(substitute_parse, "|[");
13037 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
13038 sv_catpv(substitute_parse, "]");
13041 sv_catpv(substitute_parse, ")");
13044 /* This is a way to get the parse to skip forward a whole named
13045 * sequence instead of matching the 2nd character when it fails the
13047 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
13051 RExC_parse = SvPV(substitute_parse, len);
13052 RExC_end = RExC_parse + len;
13053 RExC_in_multi_char_class = 1;
13054 RExC_emit = (regnode *)orig_emit;
13056 ret = reg(pRExC_state, 1, ®_flags, depth+1);
13058 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
13060 RExC_parse = save_parse;
13061 RExC_end = save_end;
13062 RExC_in_multi_char_class = 0;
13063 SvREFCNT_dec_NN(multi_char_matches);
13067 /* If the character class contains only a single element, it may be
13068 * optimizable into another node type which is smaller and runs faster.
13069 * Check if this is the case for this class */
13070 if (element_count == 1 && ! ret_invlist) {
13074 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like \w or
13075 [:digit:] or \p{foo} */
13077 /* All named classes are mapped into POSIXish nodes, with its FLAG
13078 * argument giving which class it is */
13079 switch ((I32)namedclass) {
13080 case ANYOF_UNIPROP:
13083 /* These don't depend on the charset modifiers. They always
13084 * match under /u rules */
13085 case ANYOF_NHORIZWS:
13086 case ANYOF_HORIZWS:
13087 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
13090 case ANYOF_NVERTWS:
13095 /* The actual POSIXish node for all the rest depends on the
13096 * charset modifier. The ones in the first set depend only on
13097 * ASCII or, if available on this platform, locale */
13101 op = (LOC) ? POSIXL : POSIXA;
13112 /* under /a could be alpha */
13114 if (ASCII_RESTRICTED) {
13115 namedclass = ANYOF_ALPHA + (namedclass % 2);
13123 /* The rest have more possibilities depending on the charset.
13124 * We take advantage of the enum ordering of the charset
13125 * modifiers to get the exact node type, */
13127 op = POSIXD + get_regex_charset(RExC_flags);
13128 if (op > POSIXA) { /* /aa is same as /a */
13131 #ifndef HAS_ISBLANK
13133 && (namedclass == ANYOF_BLANK
13134 || namedclass == ANYOF_NBLANK))
13141 /* The odd numbered ones are the complements of the
13142 * next-lower even number one */
13143 if (namedclass % 2 == 1) {
13147 arg = namedclass_to_classnum(namedclass);
13151 else if (value == prevvalue) {
13153 /* Here, the class consists of just a single code point */
13156 if (! LOC && value == '\n') {
13157 op = REG_ANY; /* Optimize [^\n] */
13158 *flagp |= HASWIDTH|SIMPLE;
13162 else if (value < 256 || UTF) {
13164 /* Optimize a single value into an EXACTish node, but not if it
13165 * would require converting the pattern to UTF-8. */
13166 op = compute_EXACTish(pRExC_state);
13168 } /* Otherwise is a range */
13169 else if (! LOC) { /* locale could vary these */
13170 if (prevvalue == '0') {
13171 if (value == '9') {
13178 /* Here, we have changed <op> away from its initial value iff we found
13179 * an optimization */
13182 /* Throw away this ANYOF regnode, and emit the calculated one,
13183 * which should correspond to the beginning, not current, state of
13185 const char * cur_parse = RExC_parse;
13186 RExC_parse = (char *)orig_parse;
13190 /* To get locale nodes to not use the full ANYOF size would
13191 * require moving the code above that writes the portions
13192 * of it that aren't in other nodes to after this point.
13193 * e.g. ANYOF_CLASS_SET */
13194 RExC_size = orig_size;
13198 RExC_emit = (regnode *)orig_emit;
13199 if (PL_regkind[op] == POSIXD) {
13201 op += NPOSIXD - POSIXD;
13206 ret = reg_node(pRExC_state, op);
13208 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
13212 *flagp |= HASWIDTH|SIMPLE;
13214 else if (PL_regkind[op] == EXACT) {
13215 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13218 RExC_parse = (char *) cur_parse;
13220 SvREFCNT_dec(posixes);
13221 SvREFCNT_dec(cp_list);
13228 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
13230 /* If folding, we calculate all characters that could fold to or from the
13231 * ones already on the list */
13232 if (FOLD && cp_list) {
13233 UV start, end; /* End points of code point ranges */
13235 SV* fold_intersection = NULL;
13237 /* If the highest code point is within Latin1, we can use the
13238 * compiled-in Alphas list, and not have to go out to disk. This
13239 * yields two false positives, the masculine and feminine ordinal
13240 * indicators, which are weeded out below using the
13241 * IS_IN_SOME_FOLD_L1() macro */
13242 if (invlist_highest(cp_list) < 256) {
13243 _invlist_intersection(PL_L1Posix_ptrs[_CC_ALPHA], cp_list,
13244 &fold_intersection);
13248 /* Here, there are non-Latin1 code points, so we will have to go
13249 * fetch the list of all the characters that participate in folds
13251 if (! PL_utf8_foldable) {
13252 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13253 &PL_sv_undef, 1, 0);
13254 PL_utf8_foldable = _get_swash_invlist(swash);
13255 SvREFCNT_dec_NN(swash);
13258 /* This is a hash that for a particular fold gives all characters
13259 * that are involved in it */
13260 if (! PL_utf8_foldclosures) {
13262 /* If we were unable to find any folds, then we likely won't be
13263 * able to find the closures. So just create an empty list.
13264 * Folding will effectively be restricted to the non-Unicode
13265 * rules hard-coded into Perl. (This case happens legitimately
13266 * during compilation of Perl itself before the Unicode tables
13267 * are generated) */
13268 if (_invlist_len(PL_utf8_foldable) == 0) {
13269 PL_utf8_foldclosures = newHV();
13272 /* If the folds haven't been read in, call a fold function
13274 if (! PL_utf8_tofold) {
13275 U8 dummy[UTF8_MAXBYTES+1];
13277 /* This string is just a short named one above \xff */
13278 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
13279 assert(PL_utf8_tofold); /* Verify that worked */
13281 PL_utf8_foldclosures =
13282 _swash_inversion_hash(PL_utf8_tofold);
13286 /* Only the characters in this class that participate in folds need
13287 * be checked. Get the intersection of this class and all the
13288 * possible characters that are foldable. This can quickly narrow
13289 * down a large class */
13290 _invlist_intersection(PL_utf8_foldable, cp_list,
13291 &fold_intersection);
13294 /* Now look at the foldable characters in this class individually */
13295 invlist_iterinit(fold_intersection);
13296 while (invlist_iternext(fold_intersection, &start, &end)) {
13299 /* Locale folding for Latin1 characters is deferred until runtime */
13300 if (LOC && start < 256) {
13304 /* Look at every character in the range */
13305 for (j = start; j <= end; j++) {
13307 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
13313 /* We have the latin1 folding rules hard-coded here so that
13314 * an innocent-looking character class, like /[ks]/i won't
13315 * have to go out to disk to find the possible matches.
13316 * XXX It would be better to generate these via regen, in
13317 * case a new version of the Unicode standard adds new
13318 * mappings, though that is not really likely, and may be
13319 * caught by the default: case of the switch below. */
13321 if (IS_IN_SOME_FOLD_L1(j)) {
13323 /* ASCII is always matched; non-ASCII is matched only
13324 * under Unicode rules */
13325 if (isASCII(j) || AT_LEAST_UNI_SEMANTICS) {
13327 add_cp_to_invlist(cp_list, PL_fold_latin1[j]);
13331 add_cp_to_invlist(depends_list, PL_fold_latin1[j]);
13335 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
13336 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
13338 /* Certain Latin1 characters have matches outside
13339 * Latin1. To get here, <j> is one of those
13340 * characters. None of these matches is valid for
13341 * ASCII characters under /aa, which is why the 'if'
13342 * just above excludes those. These matches only
13343 * happen when the target string is utf8. The code
13344 * below adds the single fold closures for <j> to the
13345 * inversion list. */
13350 add_cp_to_invlist(cp_list, KELVIN_SIGN);
13354 cp_list = add_cp_to_invlist(cp_list,
13355 LATIN_SMALL_LETTER_LONG_S);
13358 cp_list = add_cp_to_invlist(cp_list,
13359 GREEK_CAPITAL_LETTER_MU);
13360 cp_list = add_cp_to_invlist(cp_list,
13361 GREEK_SMALL_LETTER_MU);
13363 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
13364 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
13366 add_cp_to_invlist(cp_list, ANGSTROM_SIGN);
13368 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
13369 cp_list = add_cp_to_invlist(cp_list,
13370 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
13372 case LATIN_SMALL_LETTER_SHARP_S:
13373 cp_list = add_cp_to_invlist(cp_list,
13374 LATIN_CAPITAL_LETTER_SHARP_S);
13376 case 'F': case 'f':
13377 case 'I': case 'i':
13378 case 'L': case 'l':
13379 case 'T': case 't':
13380 case 'A': case 'a':
13381 case 'H': case 'h':
13382 case 'J': case 'j':
13383 case 'N': case 'n':
13384 case 'W': case 'w':
13385 case 'Y': case 'y':
13386 /* These all are targets of multi-character
13387 * folds from code points that require UTF8 to
13388 * express, so they can't match unless the
13389 * target string is in UTF-8, so no action here
13390 * is necessary, as regexec.c properly handles
13391 * the general case for UTF-8 matching and
13392 * multi-char folds */
13395 /* Use deprecated warning to increase the
13396 * chances of this being output */
13397 ckWARN2regdep(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
13404 /* Here is an above Latin1 character. We don't have the rules
13405 * hard-coded for it. First, get its fold. This is the simple
13406 * fold, as the multi-character folds have been handled earlier
13407 * and separated out */
13408 _to_uni_fold_flags(j, foldbuf, &foldlen,
13410 ? FOLD_FLAGS_LOCALE
13411 : (ASCII_FOLD_RESTRICTED)
13412 ? FOLD_FLAGS_NOMIX_ASCII
13415 /* Single character fold of above Latin1. Add everything in
13416 * its fold closure to the list that this node should match.
13417 * The fold closures data structure is a hash with the keys
13418 * being the UTF-8 of every character that is folded to, like
13419 * 'k', and the values each an array of all code points that
13420 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
13421 * Multi-character folds are not included */
13422 if ((listp = hv_fetch(PL_utf8_foldclosures,
13423 (char *) foldbuf, foldlen, FALSE)))
13425 AV* list = (AV*) *listp;
13427 for (k = 0; k <= av_len(list); k++) {
13428 SV** c_p = av_fetch(list, k, FALSE);
13431 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
13435 /* /aa doesn't allow folds between ASCII and non-; /l
13436 * doesn't allow them between above and below 256 */
13437 if ((ASCII_FOLD_RESTRICTED
13438 && (isASCII(c) != isASCII(j)))
13439 || (LOC && ((c < 256) != (j < 256))))
13444 /* Folds involving non-ascii Latin1 characters
13445 * under /d are added to a separate list */
13446 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
13448 cp_list = add_cp_to_invlist(cp_list, c);
13451 depends_list = add_cp_to_invlist(depends_list, c);
13457 SvREFCNT_dec_NN(fold_intersection);
13460 /* And combine the result (if any) with any inversion list from posix
13461 * classes. The lists are kept separate up to now because we don't want to
13462 * fold the classes (folding of those is automatically handled by the swash
13463 * fetching code) */
13465 if (! DEPENDS_SEMANTICS) {
13467 _invlist_union(cp_list, posixes, &cp_list);
13468 SvREFCNT_dec_NN(posixes);
13475 /* Under /d, we put into a separate list the Latin1 things that
13476 * match only when the target string is utf8 */
13477 SV* nonascii_but_latin1_properties = NULL;
13478 _invlist_intersection(posixes, PL_Latin1,
13479 &nonascii_but_latin1_properties);
13480 _invlist_subtract(nonascii_but_latin1_properties, PL_ASCII,
13481 &nonascii_but_latin1_properties);
13482 _invlist_subtract(posixes, nonascii_but_latin1_properties,
13485 _invlist_union(cp_list, posixes, &cp_list);
13486 SvREFCNT_dec_NN(posixes);
13492 if (depends_list) {
13493 _invlist_union(depends_list, nonascii_but_latin1_properties,
13495 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
13498 depends_list = nonascii_but_latin1_properties;
13503 /* And combine the result (if any) with any inversion list from properties.
13504 * The lists are kept separate up to now so that we can distinguish the two
13505 * in regards to matching above-Unicode. A run-time warning is generated
13506 * if a Unicode property is matched against a non-Unicode code point. But,
13507 * we allow user-defined properties to match anything, without any warning,
13508 * and we also suppress the warning if there is a portion of the character
13509 * class that isn't a Unicode property, and which matches above Unicode, \W
13510 * or [\x{110000}] for example.
13511 * (Note that in this case, unlike the Posix one above, there is no
13512 * <depends_list>, because having a Unicode property forces Unicode
13515 bool warn_super = ! has_user_defined_property;
13518 /* If it matters to the final outcome, see if a non-property
13519 * component of the class matches above Unicode. If so, the
13520 * warning gets suppressed. This is true even if just a single
13521 * such code point is specified, as though not strictly correct if
13522 * another such code point is matched against, the fact that they
13523 * are using above-Unicode code points indicates they should know
13524 * the issues involved */
13526 bool non_prop_matches_above_Unicode =
13527 runtime_posix_matches_above_Unicode
13528 | (invlist_highest(cp_list) > PERL_UNICODE_MAX);
13530 non_prop_matches_above_Unicode =
13531 ! non_prop_matches_above_Unicode;
13533 warn_super = ! non_prop_matches_above_Unicode;
13536 _invlist_union(properties, cp_list, &cp_list);
13537 SvREFCNT_dec_NN(properties);
13540 cp_list = properties;
13544 OP(ret) = ANYOF_WARN_SUPER;
13548 /* Here, we have calculated what code points should be in the character
13551 * Now we can see about various optimizations. Fold calculation (which we
13552 * did above) needs to take place before inversion. Otherwise /[^k]/i
13553 * would invert to include K, which under /i would match k, which it
13554 * shouldn't. Therefore we can't invert folded locale now, as it won't be
13555 * folded until runtime */
13557 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
13558 * at compile time. Besides not inverting folded locale now, we can't
13559 * invert if there are things such as \w, which aren't known until runtime
13562 && ! (LOC && (FOLD || (ANYOF_FLAGS(ret) & ANYOF_CLASS)))
13564 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13566 _invlist_invert(cp_list);
13568 /* Any swash can't be used as-is, because we've inverted things */
13570 SvREFCNT_dec_NN(swash);
13574 /* Clear the invert flag since have just done it here */
13579 *ret_invlist = cp_list;
13581 /* Discard the generated node */
13583 RExC_size = orig_size;
13586 RExC_emit = orig_emit;
13591 /* If we didn't do folding, it's because some information isn't available
13592 * until runtime; set the run-time fold flag for these. (We don't have to
13593 * worry about properties folding, as that is taken care of by the swash
13597 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
13600 /* Some character classes are equivalent to other nodes. Such nodes take
13601 * up less room and generally fewer operations to execute than ANYOF nodes.
13602 * Above, we checked for and optimized into some such equivalents for
13603 * certain common classes that are easy to test. Getting to this point in
13604 * the code means that the class didn't get optimized there. Since this
13605 * code is only executed in Pass 2, it is too late to save space--it has
13606 * been allocated in Pass 1, and currently isn't given back. But turning
13607 * things into an EXACTish node can allow the optimizer to join it to any
13608 * adjacent such nodes. And if the class is equivalent to things like /./,
13609 * expensive run-time swashes can be avoided. Now that we have more
13610 * complete information, we can find things necessarily missed by the
13611 * earlier code. I (khw) am not sure how much to look for here. It would
13612 * be easy, but perhaps too slow, to check any candidates against all the
13613 * node types they could possibly match using _invlistEQ(). */
13618 && ! (ANYOF_FLAGS(ret) & ANYOF_CLASS)
13619 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13622 U8 op = END; /* The optimzation node-type */
13623 const char * cur_parse= RExC_parse;
13625 invlist_iterinit(cp_list);
13626 if (! invlist_iternext(cp_list, &start, &end)) {
13628 /* Here, the list is empty. This happens, for example, when a
13629 * Unicode property is the only thing in the character class, and
13630 * it doesn't match anything. (perluniprops.pod notes such
13633 *flagp |= HASWIDTH|SIMPLE;
13635 else if (start == end) { /* The range is a single code point */
13636 if (! invlist_iternext(cp_list, &start, &end)
13638 /* Don't do this optimization if it would require changing
13639 * the pattern to UTF-8 */
13640 && (start < 256 || UTF))
13642 /* Here, the list contains a single code point. Can optimize
13643 * into an EXACT node */
13652 /* A locale node under folding with one code point can be
13653 * an EXACTFL, as its fold won't be calculated until
13659 /* Here, we are generally folding, but there is only one
13660 * code point to match. If we have to, we use an EXACT
13661 * node, but it would be better for joining with adjacent
13662 * nodes in the optimization pass if we used the same
13663 * EXACTFish node that any such are likely to be. We can
13664 * do this iff the code point doesn't participate in any
13665 * folds. For example, an EXACTF of a colon is the same as
13666 * an EXACT one, since nothing folds to or from a colon. */
13668 if (IS_IN_SOME_FOLD_L1(value)) {
13673 if (! PL_utf8_foldable) {
13674 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13675 &PL_sv_undef, 1, 0);
13676 PL_utf8_foldable = _get_swash_invlist(swash);
13677 SvREFCNT_dec_NN(swash);
13679 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
13684 /* If we haven't found the node type, above, it means we
13685 * can use the prevailing one */
13687 op = compute_EXACTish(pRExC_state);
13692 else if (start == 0) {
13693 if (end == UV_MAX) {
13695 *flagp |= HASWIDTH|SIMPLE;
13698 else if (end == '\n' - 1
13699 && invlist_iternext(cp_list, &start, &end)
13700 && start == '\n' + 1 && end == UV_MAX)
13703 *flagp |= HASWIDTH|SIMPLE;
13707 invlist_iterfinish(cp_list);
13710 RExC_parse = (char *)orig_parse;
13711 RExC_emit = (regnode *)orig_emit;
13713 ret = reg_node(pRExC_state, op);
13715 RExC_parse = (char *)cur_parse;
13717 if (PL_regkind[op] == EXACT) {
13718 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13721 SvREFCNT_dec_NN(cp_list);
13726 /* Here, <cp_list> contains all the code points we can determine at
13727 * compile time that match under all conditions. Go through it, and
13728 * for things that belong in the bitmap, put them there, and delete from
13729 * <cp_list>. While we are at it, see if everything above 255 is in the
13730 * list, and if so, set a flag to speed up execution */
13731 ANYOF_BITMAP_ZERO(ret);
13734 /* This gets set if we actually need to modify things */
13735 bool change_invlist = FALSE;
13739 /* Start looking through <cp_list> */
13740 invlist_iterinit(cp_list);
13741 while (invlist_iternext(cp_list, &start, &end)) {
13745 if (end == UV_MAX && start <= 256) {
13746 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
13749 /* Quit if are above what we should change */
13754 change_invlist = TRUE;
13756 /* Set all the bits in the range, up to the max that we are doing */
13757 high = (end < 255) ? end : 255;
13758 for (i = start; i <= (int) high; i++) {
13759 if (! ANYOF_BITMAP_TEST(ret, i)) {
13760 ANYOF_BITMAP_SET(ret, i);
13766 invlist_iterfinish(cp_list);
13768 /* Done with loop; remove any code points that are in the bitmap from
13770 if (change_invlist) {
13771 _invlist_subtract(cp_list, PL_Latin1, &cp_list);
13774 /* If have completely emptied it, remove it completely */
13775 if (_invlist_len(cp_list) == 0) {
13776 SvREFCNT_dec_NN(cp_list);
13782 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
13785 /* Here, the bitmap has been populated with all the Latin1 code points that
13786 * always match. Can now add to the overall list those that match only
13787 * when the target string is UTF-8 (<depends_list>). */
13788 if (depends_list) {
13790 _invlist_union(cp_list, depends_list, &cp_list);
13791 SvREFCNT_dec_NN(depends_list);
13794 cp_list = depends_list;
13798 /* If there is a swash and more than one element, we can't use the swash in
13799 * the optimization below. */
13800 if (swash && element_count > 1) {
13801 SvREFCNT_dec_NN(swash);
13806 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13808 ARG_SET(ret, ANYOF_NONBITMAP_EMPTY);
13811 /* av[0] stores the character class description in its textual form:
13812 * used later (regexec.c:Perl_regclass_swash()) to initialize the
13813 * appropriate swash, and is also useful for dumping the regnode.
13814 * av[1] if NULL, is a placeholder to later contain the swash computed
13815 * from av[0]. But if no further computation need be done, the
13816 * swash is stored there now.
13817 * av[2] stores the cp_list inversion list for use in addition or
13818 * instead of av[0]; used only if av[1] is NULL
13819 * av[3] is set if any component of the class is from a user-defined
13820 * property; used only if av[1] is NULL */
13821 AV * const av = newAV();
13824 av_store(av, 0, (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13825 ? SvREFCNT_inc(listsv) : &PL_sv_undef);
13827 av_store(av, 1, swash);
13828 SvREFCNT_dec_NN(cp_list);
13831 av_store(av, 1, NULL);
13833 av_store(av, 2, cp_list);
13834 av_store(av, 3, newSVuv(has_user_defined_property));
13838 rv = newRV_noinc(MUTABLE_SV(av));
13839 n = add_data(pRExC_state, 1, "s");
13840 RExC_rxi->data->data[n] = (void*)rv;
13844 *flagp |= HASWIDTH|SIMPLE;
13847 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
13850 /* reg_skipcomment()
13852 Absorbs an /x style # comments from the input stream.
13853 Returns true if there is more text remaining in the stream.
13854 Will set the REG_SEEN_RUN_ON_COMMENT flag if the comment
13855 terminates the pattern without including a newline.
13857 Note its the callers responsibility to ensure that we are
13858 actually in /x mode
13863 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
13867 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
13869 while (RExC_parse < RExC_end)
13870 if (*RExC_parse++ == '\n') {
13875 /* we ran off the end of the pattern without ending
13876 the comment, so we have to add an \n when wrapping */
13877 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
13885 Advances the parse position, and optionally absorbs
13886 "whitespace" from the inputstream.
13888 Without /x "whitespace" means (?#...) style comments only,
13889 with /x this means (?#...) and # comments and whitespace proper.
13891 Returns the RExC_parse point from BEFORE the scan occurs.
13893 This is the /x friendly way of saying RExC_parse++.
13897 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
13899 char* const retval = RExC_parse++;
13901 PERL_ARGS_ASSERT_NEXTCHAR;
13904 if (RExC_end - RExC_parse >= 3
13905 && *RExC_parse == '('
13906 && RExC_parse[1] == '?'
13907 && RExC_parse[2] == '#')
13909 while (*RExC_parse != ')') {
13910 if (RExC_parse == RExC_end)
13911 FAIL("Sequence (?#... not terminated");
13917 if (RExC_flags & RXf_PMf_EXTENDED) {
13918 if (isSPACE(*RExC_parse)) {
13922 else if (*RExC_parse == '#') {
13923 if ( reg_skipcomment( pRExC_state ) )
13932 - reg_node - emit a node
13934 STATIC regnode * /* Location. */
13935 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
13939 regnode * const ret = RExC_emit;
13940 GET_RE_DEBUG_FLAGS_DECL;
13942 PERL_ARGS_ASSERT_REG_NODE;
13945 SIZE_ALIGN(RExC_size);
13949 if (RExC_emit >= RExC_emit_bound)
13950 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
13951 op, RExC_emit, RExC_emit_bound);
13953 NODE_ALIGN_FILL(ret);
13955 FILL_ADVANCE_NODE(ptr, op);
13956 #ifdef RE_TRACK_PATTERN_OFFSETS
13957 if (RExC_offsets) { /* MJD */
13958 MJD_OFFSET_DEBUG(("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
13959 "reg_node", __LINE__,
13961 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
13962 ? "Overwriting end of array!\n" : "OK",
13963 (UV)(RExC_emit - RExC_emit_start),
13964 (UV)(RExC_parse - RExC_start),
13965 (UV)RExC_offsets[0]));
13966 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
13974 - reganode - emit a node with an argument
13976 STATIC regnode * /* Location. */
13977 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
13981 regnode * const ret = RExC_emit;
13982 GET_RE_DEBUG_FLAGS_DECL;
13984 PERL_ARGS_ASSERT_REGANODE;
13987 SIZE_ALIGN(RExC_size);
13992 assert(2==regarglen[op]+1);
13994 Anything larger than this has to allocate the extra amount.
13995 If we changed this to be:
13997 RExC_size += (1 + regarglen[op]);
13999 then it wouldn't matter. Its not clear what side effect
14000 might come from that so its not done so far.
14005 if (RExC_emit >= RExC_emit_bound)
14006 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14007 op, RExC_emit, RExC_emit_bound);
14009 NODE_ALIGN_FILL(ret);
14011 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
14012 #ifdef RE_TRACK_PATTERN_OFFSETS
14013 if (RExC_offsets) { /* MJD */
14014 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14018 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
14019 "Overwriting end of array!\n" : "OK",
14020 (UV)(RExC_emit - RExC_emit_start),
14021 (UV)(RExC_parse - RExC_start),
14022 (UV)RExC_offsets[0]));
14023 Set_Cur_Node_Offset;
14031 - reguni - emit (if appropriate) a Unicode character
14034 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
14038 PERL_ARGS_ASSERT_REGUNI;
14040 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
14044 - reginsert - insert an operator in front of already-emitted operand
14046 * Means relocating the operand.
14049 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
14055 const int offset = regarglen[(U8)op];
14056 const int size = NODE_STEP_REGNODE + offset;
14057 GET_RE_DEBUG_FLAGS_DECL;
14059 PERL_ARGS_ASSERT_REGINSERT;
14060 PERL_UNUSED_ARG(depth);
14061 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
14062 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
14071 if (RExC_open_parens) {
14073 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
14074 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
14075 if ( RExC_open_parens[paren] >= opnd ) {
14076 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
14077 RExC_open_parens[paren] += size;
14079 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
14081 if ( RExC_close_parens[paren] >= opnd ) {
14082 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
14083 RExC_close_parens[paren] += size;
14085 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
14090 while (src > opnd) {
14091 StructCopy(--src, --dst, regnode);
14092 #ifdef RE_TRACK_PATTERN_OFFSETS
14093 if (RExC_offsets) { /* MJD 20010112 */
14094 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
14098 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
14099 ? "Overwriting end of array!\n" : "OK",
14100 (UV)(src - RExC_emit_start),
14101 (UV)(dst - RExC_emit_start),
14102 (UV)RExC_offsets[0]));
14103 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
14104 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
14110 place = opnd; /* Op node, where operand used to be. */
14111 #ifdef RE_TRACK_PATTERN_OFFSETS
14112 if (RExC_offsets) { /* MJD */
14113 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14117 (UV)(place - RExC_emit_start) > RExC_offsets[0]
14118 ? "Overwriting end of array!\n" : "OK",
14119 (UV)(place - RExC_emit_start),
14120 (UV)(RExC_parse - RExC_start),
14121 (UV)RExC_offsets[0]));
14122 Set_Node_Offset(place, RExC_parse);
14123 Set_Node_Length(place, 1);
14126 src = NEXTOPER(place);
14127 FILL_ADVANCE_NODE(place, op);
14128 Zero(src, offset, regnode);
14132 - regtail - set the next-pointer at the end of a node chain of p to val.
14133 - SEE ALSO: regtail_study
14135 /* TODO: All three parms should be const */
14137 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14141 GET_RE_DEBUG_FLAGS_DECL;
14143 PERL_ARGS_ASSERT_REGTAIL;
14145 PERL_UNUSED_ARG(depth);
14151 /* Find last node. */
14154 regnode * const temp = regnext(scan);
14156 SV * const mysv=sv_newmortal();
14157 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
14158 regprop(RExC_rx, mysv, scan);
14159 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
14160 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
14161 (temp == NULL ? "->" : ""),
14162 (temp == NULL ? PL_reg_name[OP(val)] : "")
14170 if (reg_off_by_arg[OP(scan)]) {
14171 ARG_SET(scan, val - scan);
14174 NEXT_OFF(scan) = val - scan;
14180 - regtail_study - set the next-pointer at the end of a node chain of p to val.
14181 - Look for optimizable sequences at the same time.
14182 - currently only looks for EXACT chains.
14184 This is experimental code. The idea is to use this routine to perform
14185 in place optimizations on branches and groups as they are constructed,
14186 with the long term intention of removing optimization from study_chunk so
14187 that it is purely analytical.
14189 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
14190 to control which is which.
14193 /* TODO: All four parms should be const */
14196 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14201 #ifdef EXPERIMENTAL_INPLACESCAN
14204 GET_RE_DEBUG_FLAGS_DECL;
14206 PERL_ARGS_ASSERT_REGTAIL_STUDY;
14212 /* Find last node. */
14216 regnode * const temp = regnext(scan);
14217 #ifdef EXPERIMENTAL_INPLACESCAN
14218 if (PL_regkind[OP(scan)] == EXACT) {
14219 bool has_exactf_sharp_s; /* Unexamined in this routine */
14220 if (join_exact(pRExC_state,scan,&min, &has_exactf_sharp_s, 1,val,depth+1))
14225 switch (OP(scan)) {
14231 case EXACTFU_TRICKYFOLD:
14233 if( exact == PSEUDO )
14235 else if ( exact != OP(scan) )
14244 SV * const mysv=sv_newmortal();
14245 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
14246 regprop(RExC_rx, mysv, scan);
14247 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
14248 SvPV_nolen_const(mysv),
14249 REG_NODE_NUM(scan),
14250 PL_reg_name[exact]);
14257 SV * const mysv_val=sv_newmortal();
14258 DEBUG_PARSE_MSG("");
14259 regprop(RExC_rx, mysv_val, val);
14260 PerlIO_printf(Perl_debug_log, "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
14261 SvPV_nolen_const(mysv_val),
14262 (IV)REG_NODE_NUM(val),
14266 if (reg_off_by_arg[OP(scan)]) {
14267 ARG_SET(scan, val - scan);
14270 NEXT_OFF(scan) = val - scan;
14278 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
14282 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
14288 for (bit=0; bit<32; bit++) {
14289 if (flags & (1<<bit)) {
14290 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
14293 if (!set++ && lead)
14294 PerlIO_printf(Perl_debug_log, "%s",lead);
14295 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
14298 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
14299 if (!set++ && lead) {
14300 PerlIO_printf(Perl_debug_log, "%s",lead);
14303 case REGEX_UNICODE_CHARSET:
14304 PerlIO_printf(Perl_debug_log, "UNICODE");
14306 case REGEX_LOCALE_CHARSET:
14307 PerlIO_printf(Perl_debug_log, "LOCALE");
14309 case REGEX_ASCII_RESTRICTED_CHARSET:
14310 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
14312 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
14313 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
14316 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
14322 PerlIO_printf(Perl_debug_log, "\n");
14324 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14330 Perl_regdump(pTHX_ const regexp *r)
14334 SV * const sv = sv_newmortal();
14335 SV *dsv= sv_newmortal();
14336 RXi_GET_DECL(r,ri);
14337 GET_RE_DEBUG_FLAGS_DECL;
14339 PERL_ARGS_ASSERT_REGDUMP;
14341 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
14343 /* Header fields of interest. */
14344 if (r->anchored_substr) {
14345 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
14346 RE_SV_DUMPLEN(r->anchored_substr), 30);
14347 PerlIO_printf(Perl_debug_log,
14348 "anchored %s%s at %"IVdf" ",
14349 s, RE_SV_TAIL(r->anchored_substr),
14350 (IV)r->anchored_offset);
14351 } else if (r->anchored_utf8) {
14352 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
14353 RE_SV_DUMPLEN(r->anchored_utf8), 30);
14354 PerlIO_printf(Perl_debug_log,
14355 "anchored utf8 %s%s at %"IVdf" ",
14356 s, RE_SV_TAIL(r->anchored_utf8),
14357 (IV)r->anchored_offset);
14359 if (r->float_substr) {
14360 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
14361 RE_SV_DUMPLEN(r->float_substr), 30);
14362 PerlIO_printf(Perl_debug_log,
14363 "floating %s%s at %"IVdf"..%"UVuf" ",
14364 s, RE_SV_TAIL(r->float_substr),
14365 (IV)r->float_min_offset, (UV)r->float_max_offset);
14366 } else if (r->float_utf8) {
14367 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
14368 RE_SV_DUMPLEN(r->float_utf8), 30);
14369 PerlIO_printf(Perl_debug_log,
14370 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
14371 s, RE_SV_TAIL(r->float_utf8),
14372 (IV)r->float_min_offset, (UV)r->float_max_offset);
14374 if (r->check_substr || r->check_utf8)
14375 PerlIO_printf(Perl_debug_log,
14377 (r->check_substr == r->float_substr
14378 && r->check_utf8 == r->float_utf8
14379 ? "(checking floating" : "(checking anchored"));
14380 if (r->extflags & RXf_NOSCAN)
14381 PerlIO_printf(Perl_debug_log, " noscan");
14382 if (r->extflags & RXf_CHECK_ALL)
14383 PerlIO_printf(Perl_debug_log, " isall");
14384 if (r->check_substr || r->check_utf8)
14385 PerlIO_printf(Perl_debug_log, ") ");
14387 if (ri->regstclass) {
14388 regprop(r, sv, ri->regstclass);
14389 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
14391 if (r->extflags & RXf_ANCH) {
14392 PerlIO_printf(Perl_debug_log, "anchored");
14393 if (r->extflags & RXf_ANCH_BOL)
14394 PerlIO_printf(Perl_debug_log, "(BOL)");
14395 if (r->extflags & RXf_ANCH_MBOL)
14396 PerlIO_printf(Perl_debug_log, "(MBOL)");
14397 if (r->extflags & RXf_ANCH_SBOL)
14398 PerlIO_printf(Perl_debug_log, "(SBOL)");
14399 if (r->extflags & RXf_ANCH_GPOS)
14400 PerlIO_printf(Perl_debug_log, "(GPOS)");
14401 PerlIO_putc(Perl_debug_log, ' ');
14403 if (r->extflags & RXf_GPOS_SEEN)
14404 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
14405 if (r->intflags & PREGf_SKIP)
14406 PerlIO_printf(Perl_debug_log, "plus ");
14407 if (r->intflags & PREGf_IMPLICIT)
14408 PerlIO_printf(Perl_debug_log, "implicit ");
14409 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
14410 if (r->extflags & RXf_EVAL_SEEN)
14411 PerlIO_printf(Perl_debug_log, "with eval ");
14412 PerlIO_printf(Perl_debug_log, "\n");
14413 DEBUG_FLAGS_r(regdump_extflags("r->extflags: ",r->extflags));
14415 PERL_ARGS_ASSERT_REGDUMP;
14416 PERL_UNUSED_CONTEXT;
14417 PERL_UNUSED_ARG(r);
14418 #endif /* DEBUGGING */
14422 - regprop - printable representation of opcode
14424 #define EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags) \
14427 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]); \
14428 if (flags & ANYOF_INVERT) \
14429 /*make sure the invert info is in each */ \
14430 sv_catpvs(sv, "^"); \
14436 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
14442 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
14443 static const char * const anyofs[] = {
14444 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
14445 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
14446 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
14447 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
14448 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
14449 || _CC_VERTSPACE != 16
14450 #error Need to adjust order of anyofs[]
14487 RXi_GET_DECL(prog,progi);
14488 GET_RE_DEBUG_FLAGS_DECL;
14490 PERL_ARGS_ASSERT_REGPROP;
14494 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
14495 /* It would be nice to FAIL() here, but this may be called from
14496 regexec.c, and it would be hard to supply pRExC_state. */
14497 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(o), (int)REGNODE_MAX);
14498 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
14500 k = PL_regkind[OP(o)];
14503 sv_catpvs(sv, " ");
14504 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
14505 * is a crude hack but it may be the best for now since
14506 * we have no flag "this EXACTish node was UTF-8"
14508 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
14509 PERL_PV_ESCAPE_UNI_DETECT |
14510 PERL_PV_ESCAPE_NONASCII |
14511 PERL_PV_PRETTY_ELLIPSES |
14512 PERL_PV_PRETTY_LTGT |
14513 PERL_PV_PRETTY_NOCLEAR
14515 } else if (k == TRIE) {
14516 /* print the details of the trie in dumpuntil instead, as
14517 * progi->data isn't available here */
14518 const char op = OP(o);
14519 const U32 n = ARG(o);
14520 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
14521 (reg_ac_data *)progi->data->data[n] :
14523 const reg_trie_data * const trie
14524 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
14526 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
14527 DEBUG_TRIE_COMPILE_r(
14528 Perl_sv_catpvf(aTHX_ sv,
14529 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
14530 (UV)trie->startstate,
14531 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
14532 (UV)trie->wordcount,
14535 (UV)TRIE_CHARCOUNT(trie),
14536 (UV)trie->uniquecharcount
14539 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
14541 int rangestart = -1;
14542 U8* bitmap = IS_ANYOF_TRIE(op) ? (U8*)ANYOF_BITMAP(o) : (U8*)TRIE_BITMAP(trie);
14543 sv_catpvs(sv, "[");
14544 for (i = 0; i <= 256; i++) {
14545 if (i < 256 && BITMAP_TEST(bitmap,i)) {
14546 if (rangestart == -1)
14548 } else if (rangestart != -1) {
14549 if (i <= rangestart + 3)
14550 for (; rangestart < i; rangestart++)
14551 put_byte(sv, rangestart);
14553 put_byte(sv, rangestart);
14554 sv_catpvs(sv, "-");
14555 put_byte(sv, i - 1);
14560 sv_catpvs(sv, "]");
14563 } else if (k == CURLY) {
14564 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
14565 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
14566 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
14568 else if (k == WHILEM && o->flags) /* Ordinal/of */
14569 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
14570 else if (k == REF || k == OPEN || k == CLOSE || k == GROUPP || OP(o)==ACCEPT) {
14571 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
14572 if ( RXp_PAREN_NAMES(prog) ) {
14573 if ( k != REF || (OP(o) < NREF)) {
14574 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
14575 SV **name= av_fetch(list, ARG(o), 0 );
14577 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14580 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
14581 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
14582 I32 *nums=(I32*)SvPVX(sv_dat);
14583 SV **name= av_fetch(list, nums[0], 0 );
14586 for ( n=0; n<SvIVX(sv_dat); n++ ) {
14587 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
14588 (n ? "," : ""), (IV)nums[n]);
14590 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14594 } else if (k == GOSUB)
14595 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o)); /* Paren and offset */
14596 else if (k == VERB) {
14598 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
14599 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
14600 } else if (k == LOGICAL)
14601 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* 2: embedded, otherwise 1 */
14602 else if (k == ANYOF) {
14603 int i, rangestart = -1;
14604 const U8 flags = ANYOF_FLAGS(o);
14608 if (flags & ANYOF_LOCALE)
14609 sv_catpvs(sv, "{loc}");
14610 if (flags & ANYOF_LOC_FOLD)
14611 sv_catpvs(sv, "{i}");
14612 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
14613 if (flags & ANYOF_INVERT)
14614 sv_catpvs(sv, "^");
14616 /* output what the standard cp 0-255 bitmap matches */
14617 for (i = 0; i <= 256; i++) {
14618 if (i < 256 && ANYOF_BITMAP_TEST(o,i)) {
14619 if (rangestart == -1)
14621 } else if (rangestart != -1) {
14622 if (i <= rangestart + 3)
14623 for (; rangestart < i; rangestart++)
14624 put_byte(sv, rangestart);
14626 put_byte(sv, rangestart);
14627 sv_catpvs(sv, "-");
14628 put_byte(sv, i - 1);
14635 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14636 /* output any special charclass tests (used entirely under use locale) */
14637 if (ANYOF_CLASS_TEST_ANY_SET(o))
14638 for (i = 0; i < (int)(sizeof(anyofs)/sizeof(char*)); i++)
14639 if (ANYOF_CLASS_TEST(o,i)) {
14640 sv_catpv(sv, anyofs[i]);
14644 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14646 if (flags & ANYOF_NON_UTF8_LATIN1_ALL) {
14647 sv_catpvs(sv, "{non-utf8-latin1-all}");
14650 /* output information about the unicode matching */
14651 if (flags & ANYOF_UNICODE_ALL)
14652 sv_catpvs(sv, "{unicode_all}");
14653 else if (ANYOF_NONBITMAP(o))
14654 sv_catpvs(sv, "{unicode}");
14655 if (flags & ANYOF_NONBITMAP_NON_UTF8)
14656 sv_catpvs(sv, "{outside bitmap}");
14658 if (ANYOF_NONBITMAP(o)) {
14659 SV *lv; /* Set if there is something outside the bit map */
14660 SV * const sw = regclass_swash(prog, o, FALSE, &lv, NULL);
14661 bool byte_output = FALSE; /* If something in the bitmap has been
14664 if (lv && lv != &PL_sv_undef) {
14666 U8 s[UTF8_MAXBYTES_CASE+1];
14668 for (i = 0; i <= 256; i++) { /* Look at chars in bitmap */
14669 uvchr_to_utf8(s, i);
14672 && ! ANYOF_BITMAP_TEST(o, i) /* Don't duplicate
14676 && swash_fetch(sw, s, TRUE))
14678 if (rangestart == -1)
14680 } else if (rangestart != -1) {
14681 byte_output = TRUE;
14682 if (i <= rangestart + 3)
14683 for (; rangestart < i; rangestart++) {
14684 put_byte(sv, rangestart);
14687 put_byte(sv, rangestart);
14688 sv_catpvs(sv, "-");
14697 char *s = savesvpv(lv);
14698 char * const origs = s;
14700 while (*s && *s != '\n')
14704 const char * const t = ++s;
14707 sv_catpvs(sv, " ");
14713 /* Truncate very long output */
14714 if (s - origs > 256) {
14715 Perl_sv_catpvf(aTHX_ sv,
14717 (int) (s - origs - 1),
14723 else if (*s == '\t') {
14738 SvREFCNT_dec_NN(lv);
14742 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
14744 else if (k == POSIXD || k == NPOSIXD) {
14745 U8 index = FLAGS(o) * 2;
14746 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
14747 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
14750 sv_catpv(sv, anyofs[index]);
14753 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
14754 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
14756 PERL_UNUSED_CONTEXT;
14757 PERL_UNUSED_ARG(sv);
14758 PERL_UNUSED_ARG(o);
14759 PERL_UNUSED_ARG(prog);
14760 #endif /* DEBUGGING */
14764 Perl_re_intuit_string(pTHX_ REGEXP * const r)
14765 { /* Assume that RE_INTUIT is set */
14767 struct regexp *const prog = ReANY(r);
14768 GET_RE_DEBUG_FLAGS_DECL;
14770 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
14771 PERL_UNUSED_CONTEXT;
14775 const char * const s = SvPV_nolen_const(prog->check_substr
14776 ? prog->check_substr : prog->check_utf8);
14778 if (!PL_colorset) reginitcolors();
14779 PerlIO_printf(Perl_debug_log,
14780 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
14782 prog->check_substr ? "" : "utf8 ",
14783 PL_colors[5],PL_colors[0],
14786 (strlen(s) > 60 ? "..." : ""));
14789 return prog->check_substr ? prog->check_substr : prog->check_utf8;
14795 handles refcounting and freeing the perl core regexp structure. When
14796 it is necessary to actually free the structure the first thing it
14797 does is call the 'free' method of the regexp_engine associated to
14798 the regexp, allowing the handling of the void *pprivate; member
14799 first. (This routine is not overridable by extensions, which is why
14800 the extensions free is called first.)
14802 See regdupe and regdupe_internal if you change anything here.
14804 #ifndef PERL_IN_XSUB_RE
14806 Perl_pregfree(pTHX_ REGEXP *r)
14812 Perl_pregfree2(pTHX_ REGEXP *rx)
14815 struct regexp *const r = ReANY(rx);
14816 GET_RE_DEBUG_FLAGS_DECL;
14818 PERL_ARGS_ASSERT_PREGFREE2;
14820 if (r->mother_re) {
14821 ReREFCNT_dec(r->mother_re);
14823 CALLREGFREE_PVT(rx); /* free the private data */
14824 SvREFCNT_dec(RXp_PAREN_NAMES(r));
14825 Safefree(r->xpv_len_u.xpvlenu_pv);
14828 SvREFCNT_dec(r->anchored_substr);
14829 SvREFCNT_dec(r->anchored_utf8);
14830 SvREFCNT_dec(r->float_substr);
14831 SvREFCNT_dec(r->float_utf8);
14832 Safefree(r->substrs);
14834 RX_MATCH_COPY_FREE(rx);
14835 #ifdef PERL_ANY_COW
14836 SvREFCNT_dec(r->saved_copy);
14839 SvREFCNT_dec(r->qr_anoncv);
14840 rx->sv_u.svu_rx = 0;
14845 This is a hacky workaround to the structural issue of match results
14846 being stored in the regexp structure which is in turn stored in
14847 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
14848 could be PL_curpm in multiple contexts, and could require multiple
14849 result sets being associated with the pattern simultaneously, such
14850 as when doing a recursive match with (??{$qr})
14852 The solution is to make a lightweight copy of the regexp structure
14853 when a qr// is returned from the code executed by (??{$qr}) this
14854 lightweight copy doesn't actually own any of its data except for
14855 the starp/end and the actual regexp structure itself.
14861 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
14863 struct regexp *ret;
14864 struct regexp *const r = ReANY(rx);
14865 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
14867 PERL_ARGS_ASSERT_REG_TEMP_COPY;
14870 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
14872 SvOK_off((SV *)ret_x);
14874 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
14875 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
14876 made both spots point to the same regexp body.) */
14877 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
14878 assert(!SvPVX(ret_x));
14879 ret_x->sv_u.svu_rx = temp->sv_any;
14880 temp->sv_any = NULL;
14881 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
14882 SvREFCNT_dec_NN(temp);
14883 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
14884 ing below will not set it. */
14885 SvCUR_set(ret_x, SvCUR(rx));
14888 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
14889 sv_force_normal(sv) is called. */
14891 ret = ReANY(ret_x);
14893 SvFLAGS(ret_x) |= SvUTF8(rx);
14894 /* We share the same string buffer as the original regexp, on which we
14895 hold a reference count, incremented when mother_re is set below.
14896 The string pointer is copied here, being part of the regexp struct.
14898 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
14899 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
14901 const I32 npar = r->nparens+1;
14902 Newx(ret->offs, npar, regexp_paren_pair);
14903 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
14906 Newx(ret->substrs, 1, struct reg_substr_data);
14907 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
14909 SvREFCNT_inc_void(ret->anchored_substr);
14910 SvREFCNT_inc_void(ret->anchored_utf8);
14911 SvREFCNT_inc_void(ret->float_substr);
14912 SvREFCNT_inc_void(ret->float_utf8);
14914 /* check_substr and check_utf8, if non-NULL, point to either their
14915 anchored or float namesakes, and don't hold a second reference. */
14917 RX_MATCH_COPIED_off(ret_x);
14918 #ifdef PERL_ANY_COW
14919 ret->saved_copy = NULL;
14921 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
14922 SvREFCNT_inc_void(ret->qr_anoncv);
14928 /* regfree_internal()
14930 Free the private data in a regexp. This is overloadable by
14931 extensions. Perl takes care of the regexp structure in pregfree(),
14932 this covers the *pprivate pointer which technically perl doesn't
14933 know about, however of course we have to handle the
14934 regexp_internal structure when no extension is in use.
14936 Note this is called before freeing anything in the regexp
14941 Perl_regfree_internal(pTHX_ REGEXP * const rx)
14944 struct regexp *const r = ReANY(rx);
14945 RXi_GET_DECL(r,ri);
14946 GET_RE_DEBUG_FLAGS_DECL;
14948 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
14954 SV *dsv= sv_newmortal();
14955 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
14956 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
14957 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
14958 PL_colors[4],PL_colors[5],s);
14961 #ifdef RE_TRACK_PATTERN_OFFSETS
14963 Safefree(ri->u.offsets); /* 20010421 MJD */
14965 if (ri->code_blocks) {
14967 for (n = 0; n < ri->num_code_blocks; n++)
14968 SvREFCNT_dec(ri->code_blocks[n].src_regex);
14969 Safefree(ri->code_blocks);
14973 int n = ri->data->count;
14976 /* If you add a ->what type here, update the comment in regcomp.h */
14977 switch (ri->data->what[n]) {
14983 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
14986 Safefree(ri->data->data[n]);
14992 { /* Aho Corasick add-on structure for a trie node.
14993 Used in stclass optimization only */
14995 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
14997 refcount = --aho->refcount;
15000 PerlMemShared_free(aho->states);
15001 PerlMemShared_free(aho->fail);
15002 /* do this last!!!! */
15003 PerlMemShared_free(ri->data->data[n]);
15004 PerlMemShared_free(ri->regstclass);
15010 /* trie structure. */
15012 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
15014 refcount = --trie->refcount;
15017 PerlMemShared_free(trie->charmap);
15018 PerlMemShared_free(trie->states);
15019 PerlMemShared_free(trie->trans);
15021 PerlMemShared_free(trie->bitmap);
15023 PerlMemShared_free(trie->jump);
15024 PerlMemShared_free(trie->wordinfo);
15025 /* do this last!!!! */
15026 PerlMemShared_free(ri->data->data[n]);
15031 Perl_croak(aTHX_ "panic: regfree data code '%c'", ri->data->what[n]);
15034 Safefree(ri->data->what);
15035 Safefree(ri->data);
15041 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
15042 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
15043 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
15046 re_dup - duplicate a regexp.
15048 This routine is expected to clone a given regexp structure. It is only
15049 compiled under USE_ITHREADS.
15051 After all of the core data stored in struct regexp is duplicated
15052 the regexp_engine.dupe method is used to copy any private data
15053 stored in the *pprivate pointer. This allows extensions to handle
15054 any duplication it needs to do.
15056 See pregfree() and regfree_internal() if you change anything here.
15058 #if defined(USE_ITHREADS)
15059 #ifndef PERL_IN_XSUB_RE
15061 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
15065 const struct regexp *r = ReANY(sstr);
15066 struct regexp *ret = ReANY(dstr);
15068 PERL_ARGS_ASSERT_RE_DUP_GUTS;
15070 npar = r->nparens+1;
15071 Newx(ret->offs, npar, regexp_paren_pair);
15072 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15074 if (ret->substrs) {
15075 /* Do it this way to avoid reading from *r after the StructCopy().
15076 That way, if any of the sv_dup_inc()s dislodge *r from the L1
15077 cache, it doesn't matter. */
15078 const bool anchored = r->check_substr
15079 ? r->check_substr == r->anchored_substr
15080 : r->check_utf8 == r->anchored_utf8;
15081 Newx(ret->substrs, 1, struct reg_substr_data);
15082 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15084 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
15085 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
15086 ret->float_substr = sv_dup_inc(ret->float_substr, param);
15087 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
15089 /* check_substr and check_utf8, if non-NULL, point to either their
15090 anchored or float namesakes, and don't hold a second reference. */
15092 if (ret->check_substr) {
15094 assert(r->check_utf8 == r->anchored_utf8);
15095 ret->check_substr = ret->anchored_substr;
15096 ret->check_utf8 = ret->anchored_utf8;
15098 assert(r->check_substr == r->float_substr);
15099 assert(r->check_utf8 == r->float_utf8);
15100 ret->check_substr = ret->float_substr;
15101 ret->check_utf8 = ret->float_utf8;
15103 } else if (ret->check_utf8) {
15105 ret->check_utf8 = ret->anchored_utf8;
15107 ret->check_utf8 = ret->float_utf8;
15112 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
15113 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
15116 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
15118 if (RX_MATCH_COPIED(dstr))
15119 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
15121 ret->subbeg = NULL;
15122 #ifdef PERL_ANY_COW
15123 ret->saved_copy = NULL;
15126 /* Whether mother_re be set or no, we need to copy the string. We
15127 cannot refrain from copying it when the storage points directly to
15128 our mother regexp, because that's
15129 1: a buffer in a different thread
15130 2: something we no longer hold a reference on
15131 so we need to copy it locally. */
15132 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
15133 ret->mother_re = NULL;
15136 #endif /* PERL_IN_XSUB_RE */
15141 This is the internal complement to regdupe() which is used to copy
15142 the structure pointed to by the *pprivate pointer in the regexp.
15143 This is the core version of the extension overridable cloning hook.
15144 The regexp structure being duplicated will be copied by perl prior
15145 to this and will be provided as the regexp *r argument, however
15146 with the /old/ structures pprivate pointer value. Thus this routine
15147 may override any copying normally done by perl.
15149 It returns a pointer to the new regexp_internal structure.
15153 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
15156 struct regexp *const r = ReANY(rx);
15157 regexp_internal *reti;
15159 RXi_GET_DECL(r,ri);
15161 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
15165 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode), char, regexp_internal);
15166 Copy(ri->program, reti->program, len+1, regnode);
15168 reti->num_code_blocks = ri->num_code_blocks;
15169 if (ri->code_blocks) {
15171 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
15172 struct reg_code_block);
15173 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
15174 struct reg_code_block);
15175 for (n = 0; n < ri->num_code_blocks; n++)
15176 reti->code_blocks[n].src_regex = (REGEXP*)
15177 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
15180 reti->code_blocks = NULL;
15182 reti->regstclass = NULL;
15185 struct reg_data *d;
15186 const int count = ri->data->count;
15189 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
15190 char, struct reg_data);
15191 Newx(d->what, count, U8);
15194 for (i = 0; i < count; i++) {
15195 d->what[i] = ri->data->what[i];
15196 switch (d->what[i]) {
15197 /* see also regcomp.h and regfree_internal() */
15198 case 'a': /* actually an AV, but the dup function is identical. */
15202 case 'u': /* actually an HV, but the dup function is identical. */
15203 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
15206 /* This is cheating. */
15207 Newx(d->data[i], 1, struct regnode_charclass_class);
15208 StructCopy(ri->data->data[i], d->data[i],
15209 struct regnode_charclass_class);
15210 reti->regstclass = (regnode*)d->data[i];
15213 /* Trie stclasses are readonly and can thus be shared
15214 * without duplication. We free the stclass in pregfree
15215 * when the corresponding reg_ac_data struct is freed.
15217 reti->regstclass= ri->regstclass;
15221 ((reg_trie_data*)ri->data->data[i])->refcount++;
15226 d->data[i] = ri->data->data[i];
15229 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'", ri->data->what[i]);
15238 reti->name_list_idx = ri->name_list_idx;
15240 #ifdef RE_TRACK_PATTERN_OFFSETS
15241 if (ri->u.offsets) {
15242 Newx(reti->u.offsets, 2*len+1, U32);
15243 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
15246 SetProgLen(reti,len);
15249 return (void*)reti;
15252 #endif /* USE_ITHREADS */
15254 #ifndef PERL_IN_XSUB_RE
15257 - regnext - dig the "next" pointer out of a node
15260 Perl_regnext(pTHX_ regnode *p)
15268 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
15269 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(p), (int)REGNODE_MAX);
15272 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
15281 S_re_croak2(pTHX_ const char* pat1,const char* pat2,...)
15284 STRLEN l1 = strlen(pat1);
15285 STRLEN l2 = strlen(pat2);
15288 const char *message;
15290 PERL_ARGS_ASSERT_RE_CROAK2;
15296 Copy(pat1, buf, l1 , char);
15297 Copy(pat2, buf + l1, l2 , char);
15298 buf[l1 + l2] = '\n';
15299 buf[l1 + l2 + 1] = '\0';
15301 /* ANSI variant takes additional second argument */
15302 va_start(args, pat2);
15306 msv = vmess(buf, &args);
15308 message = SvPV_const(msv,l1);
15311 Copy(message, buf, l1 , char);
15312 buf[l1-1] = '\0'; /* Overwrite \n */
15313 Perl_croak(aTHX_ "%s", buf);
15316 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
15318 #ifndef PERL_IN_XSUB_RE
15320 Perl_save_re_context(pTHX)
15324 struct re_save_state *state;
15326 SAVEVPTR(PL_curcop);
15327 SSGROW(SAVESTACK_ALLOC_FOR_RE_SAVE_STATE + 1);
15329 state = (struct re_save_state *)(PL_savestack + PL_savestack_ix);
15330 PL_savestack_ix += SAVESTACK_ALLOC_FOR_RE_SAVE_STATE;
15331 SSPUSHUV(SAVEt_RE_STATE);
15333 Copy(&PL_reg_state, state, 1, struct re_save_state);
15335 PL_reg_oldsaved = NULL;
15336 PL_reg_oldsavedlen = 0;
15337 PL_reg_oldsavedoffset = 0;
15338 PL_reg_oldsavedcoffset = 0;
15339 PL_reg_maxiter = 0;
15340 PL_reg_leftiter = 0;
15341 PL_reg_poscache = NULL;
15342 PL_reg_poscache_size = 0;
15343 #ifdef PERL_ANY_COW
15347 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
15349 const REGEXP * const rx = PM_GETRE(PL_curpm);
15352 for (i = 1; i <= RX_NPARENS(rx); i++) {
15353 char digits[TYPE_CHARS(long)];
15354 const STRLEN len = my_snprintf(digits, sizeof(digits), "%lu", (long)i);
15355 GV *const *const gvp
15356 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
15359 GV * const gv = *gvp;
15360 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
15372 S_put_byte(pTHX_ SV *sv, int c)
15374 PERL_ARGS_ASSERT_PUT_BYTE;
15376 /* Our definition of isPRINT() ignores locales, so only bytes that are
15377 not part of UTF-8 are considered printable. I assume that the same
15378 holds for UTF-EBCDIC.
15379 Also, code point 255 is not printable in either (it's E0 in EBCDIC,
15380 which Wikipedia says:
15382 EO, or Eight Ones, is an 8-bit EBCDIC character code represented as all
15383 ones (binary 1111 1111, hexadecimal FF). It is similar, but not
15384 identical, to the ASCII delete (DEL) or rubout control character. ...
15385 it is typically mapped to hexadecimal code 9F, in order to provide a
15386 unique character mapping in both directions)
15388 So the old condition can be simplified to !isPRINT(c) */
15391 Perl_sv_catpvf(aTHX_ sv, "\\x%02x", c);
15394 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
15398 const char string = c;
15399 if (c == '-' || c == ']' || c == '\\' || c == '^')
15400 sv_catpvs(sv, "\\");
15401 sv_catpvn(sv, &string, 1);
15406 #define CLEAR_OPTSTART \
15407 if (optstart) STMT_START { \
15408 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
15412 #define DUMPUNTIL(b,e) CLEAR_OPTSTART; node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
15414 STATIC const regnode *
15415 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
15416 const regnode *last, const regnode *plast,
15417 SV* sv, I32 indent, U32 depth)
15420 U8 op = PSEUDO; /* Arbitrary non-END op. */
15421 const regnode *next;
15422 const regnode *optstart= NULL;
15424 RXi_GET_DECL(r,ri);
15425 GET_RE_DEBUG_FLAGS_DECL;
15427 PERL_ARGS_ASSERT_DUMPUNTIL;
15429 #ifdef DEBUG_DUMPUNTIL
15430 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
15431 last ? last-start : 0,plast ? plast-start : 0);
15434 if (plast && plast < last)
15437 while (PL_regkind[op] != END && (!last || node < last)) {
15438 /* While that wasn't END last time... */
15441 if (op == CLOSE || op == WHILEM)
15443 next = regnext((regnode *)node);
15446 if (OP(node) == OPTIMIZED) {
15447 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
15454 regprop(r, sv, node);
15455 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
15456 (int)(2*indent + 1), "", SvPVX_const(sv));
15458 if (OP(node) != OPTIMIZED) {
15459 if (next == NULL) /* Next ptr. */
15460 PerlIO_printf(Perl_debug_log, " (0)");
15461 else if (PL_regkind[(U8)op] == BRANCH && PL_regkind[OP(next)] != BRANCH )
15462 PerlIO_printf(Perl_debug_log, " (FAIL)");
15464 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
15465 (void)PerlIO_putc(Perl_debug_log, '\n');
15469 if (PL_regkind[(U8)op] == BRANCHJ) {
15472 const regnode *nnode = (OP(next) == LONGJMP
15473 ? regnext((regnode *)next)
15475 if (last && nnode > last)
15477 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
15480 else if (PL_regkind[(U8)op] == BRANCH) {
15482 DUMPUNTIL(NEXTOPER(node), next);
15484 else if ( PL_regkind[(U8)op] == TRIE ) {
15485 const regnode *this_trie = node;
15486 const char op = OP(node);
15487 const U32 n = ARG(node);
15488 const reg_ac_data * const ac = op>=AHOCORASICK ?
15489 (reg_ac_data *)ri->data->data[n] :
15491 const reg_trie_data * const trie =
15492 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
15494 AV *const trie_words = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
15496 const regnode *nextbranch= NULL;
15499 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
15500 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
15502 PerlIO_printf(Perl_debug_log, "%*s%s ",
15503 (int)(2*(indent+3)), "",
15504 elem_ptr ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr), SvCUR(*elem_ptr), 60,
15505 PL_colors[0], PL_colors[1],
15506 (SvUTF8(*elem_ptr) ? PERL_PV_ESCAPE_UNI : 0) |
15507 PERL_PV_PRETTY_ELLIPSES |
15508 PERL_PV_PRETTY_LTGT
15513 U16 dist= trie->jump[word_idx+1];
15514 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
15515 (UV)((dist ? this_trie + dist : next) - start));
15518 nextbranch= this_trie + trie->jump[0];
15519 DUMPUNTIL(this_trie + dist, nextbranch);
15521 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
15522 nextbranch= regnext((regnode *)nextbranch);
15524 PerlIO_printf(Perl_debug_log, "\n");
15527 if (last && next > last)
15532 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
15533 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
15534 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
15536 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
15538 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
15540 else if ( op == PLUS || op == STAR) {
15541 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
15543 else if (PL_regkind[(U8)op] == ANYOF) {
15544 /* arglen 1 + class block */
15545 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_CLASS)
15546 ? ANYOF_CLASS_SKIP : ANYOF_SKIP);
15547 node = NEXTOPER(node);
15549 else if (PL_regkind[(U8)op] == EXACT) {
15550 /* Literal string, where present. */
15551 node += NODE_SZ_STR(node) - 1;
15552 node = NEXTOPER(node);
15555 node = NEXTOPER(node);
15556 node += regarglen[(U8)op];
15558 if (op == CURLYX || op == OPEN)
15562 #ifdef DEBUG_DUMPUNTIL
15563 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
15568 #endif /* DEBUGGING */
15572 * c-indentation-style: bsd
15573 * c-basic-offset: 4
15574 * indent-tabs-mode: nil
15577 * ex: set ts=8 sts=4 sw=4 et: