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, FOLD_FLAGS_FULL); \
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 and EXACTFA nodes when the
2690 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2691 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2692 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2693 * character in the pattern corresponds to at most a single character in
2694 * the target string. (And I do mean character, and not byte here, unlike
2695 * other parts of the documentation that have never been updated to
2696 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2697 * two character string 'ss'; in EXACTFA nodes it can match
2698 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2699 * instances where it is violated. I'm reluctant to try to change the
2700 * assumption, as the code involved is impenetrable to me (khw), so
2701 * instead the code here punts. This routine examines (when the pattern
2702 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2703 * boolean indicating whether or not the node contains a sharp s. When it
2704 * is true, the caller sets a flag that later causes the optimizer in this
2705 * file to not set values for the floating and fixed string lengths, and
2706 * thus avoids the optimizer code in regexec.c that makes the invalid
2707 * assumption. Thus, there is no optimization based on string lengths for
2708 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2709 * (The reason the assumption is wrong only in these two cases is that all
2710 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2711 * other folds to their expanded versions. We can't prefold sharp s to
2712 * 'ss' in EXACTF nodes because we don't know at compile time if it
2713 * actually matches 'ss' or not. It will match iff the target string is
2714 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2715 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2716 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2717 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2718 * require the pattern to be forced into UTF-8, the overhead of which we
2722 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2723 if (PL_regkind[OP(scan)] == EXACT) \
2724 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2727 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) {
2728 /* Merge several consecutive EXACTish nodes into one. */
2729 regnode *n = regnext(scan);
2731 regnode *next = scan + NODE_SZ_STR(scan);
2735 regnode *stop = scan;
2736 GET_RE_DEBUG_FLAGS_DECL;
2738 PERL_UNUSED_ARG(depth);
2741 PERL_ARGS_ASSERT_JOIN_EXACT;
2742 #ifndef EXPERIMENTAL_INPLACESCAN
2743 PERL_UNUSED_ARG(flags);
2744 PERL_UNUSED_ARG(val);
2746 DEBUG_PEEP("join",scan,depth);
2748 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2749 * EXACT ones that are mergeable to the current one. */
2751 && (PL_regkind[OP(n)] == NOTHING
2752 || (stringok && OP(n) == OP(scan)))
2754 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2757 if (OP(n) == TAIL || n > next)
2759 if (PL_regkind[OP(n)] == NOTHING) {
2760 DEBUG_PEEP("skip:",n,depth);
2761 NEXT_OFF(scan) += NEXT_OFF(n);
2762 next = n + NODE_STEP_REGNODE;
2769 else if (stringok) {
2770 const unsigned int oldl = STR_LEN(scan);
2771 regnode * const nnext = regnext(n);
2773 /* XXX I (khw) kind of doubt that this works on platforms where
2774 * U8_MAX is above 255 because of lots of other assumptions */
2775 /* Don't join if the sum can't fit into a single node */
2776 if (oldl + STR_LEN(n) > U8_MAX)
2779 DEBUG_PEEP("merg",n,depth);
2782 NEXT_OFF(scan) += NEXT_OFF(n);
2783 STR_LEN(scan) += STR_LEN(n);
2784 next = n + NODE_SZ_STR(n);
2785 /* Now we can overwrite *n : */
2786 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2794 #ifdef EXPERIMENTAL_INPLACESCAN
2795 if (flags && !NEXT_OFF(n)) {
2796 DEBUG_PEEP("atch", val, depth);
2797 if (reg_off_by_arg[OP(n)]) {
2798 ARG_SET(n, val - n);
2801 NEXT_OFF(n) = val - n;
2809 *has_exactf_sharp_s = FALSE;
2811 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2812 * can now analyze for sequences of problematic code points. (Prior to
2813 * this final joining, sequences could have been split over boundaries, and
2814 * hence missed). The sequences only happen in folding, hence for any
2815 * non-EXACT EXACTish node */
2816 if (OP(scan) != EXACT) {
2817 const U8 * const s0 = (U8*) STRING(scan);
2819 const U8 * const s_end = s0 + STR_LEN(scan);
2821 /* One pass is made over the node's string looking for all the
2822 * possibilities. to avoid some tests in the loop, there are two main
2823 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2827 /* Examine the string for a multi-character fold sequence. UTF-8
2828 * patterns have all characters pre-folded by the time this code is
2830 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2831 length sequence we are looking for is 2 */
2834 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2835 if (! len) { /* Not a multi-char fold: get next char */
2840 /* Nodes with 'ss' require special handling, except for EXACTFL
2841 * and EXACTFA for which there is no multi-char fold to this */
2842 if (len == 2 && *s == 's' && *(s+1) == 's'
2843 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2846 OP(scan) = EXACTFU_SS;
2849 else if (len == 6 /* len is the same in both ASCII and EBCDIC
2851 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2852 COMBINING_DIAERESIS_UTF8
2853 COMBINING_ACUTE_ACCENT_UTF8,
2855 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2856 COMBINING_DIAERESIS_UTF8
2857 COMBINING_ACUTE_ACCENT_UTF8,
2862 /* These two folds require special handling by trie's, so
2863 * change the node type to indicate this. If EXACTFA and
2864 * EXACTFL were ever to be handled by trie's, this would
2865 * have to be changed. If this node has already been
2866 * changed to EXACTFU_SS in this loop, leave it as is. (I
2867 * (khw) think it doesn't matter in regexec.c for UTF
2868 * patterns, but no need to change it */
2869 if (OP(scan) == EXACTFU) {
2870 OP(scan) = EXACTFU_TRICKYFOLD;
2874 else { /* Here is a generic multi-char fold. */
2875 const U8* multi_end = s + len;
2877 /* Count how many characters in it. In the case of /l and
2878 * /aa, no folds which contain ASCII code points are
2879 * allowed, so check for those, and skip if found. (In
2880 * EXACTFL, no folds are allowed to any Latin1 code point,
2881 * not just ASCII. But there aren't any of these
2882 * currently, nor ever likely, so don't take the time to
2883 * test for them. The code that generates the
2884 * is_MULTI_foo() macros croaks should one actually get put
2885 * into Unicode .) */
2886 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2887 count = utf8_length(s, multi_end);
2891 while (s < multi_end) {
2894 goto next_iteration;
2904 /* The delta is how long the sequence is minus 1 (1 is how long
2905 * the character that folds to the sequence is) */
2906 *min_subtract += count - 1;
2910 else if (OP(scan) == EXACTFA) {
2912 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2913 * fold to the ASCII range (and there are no existing ones in the
2914 * upper latin1 range). But, as outlined in the comments preceding
2915 * this function, we need to flag any occurrences of the sharp s */
2917 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2918 *has_exactf_sharp_s = TRUE;
2925 else if (OP(scan) != EXACTFL) {
2927 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2928 * multi-char folds that are all Latin1. (This code knows that
2929 * there are no current multi-char folds possible with EXACTFL,
2930 * relying on fold_grind.t to catch any errors if the very unlikely
2931 * event happens that some get added in future Unicode versions.)
2932 * As explained in the comments preceding this function, we look
2933 * also for the sharp s in EXACTF nodes; it can be in the final
2934 * position. Otherwise we can stop looking 1 byte earlier because
2935 * have to find at least two characters for a multi-fold */
2936 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2938 /* The below is perhaps overboard, but this allows us to save a
2939 * test each time through the loop at the expense of a mask. This
2940 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2941 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2942 * are 64. This uses an exclusive 'or' to find that bit and then
2943 * inverts it to form a mask, with just a single 0, in the bit
2944 * position where 'S' and 's' differ. */
2945 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2946 const U8 s_masked = 's' & S_or_s_mask;
2949 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2950 if (! len) { /* Not a multi-char fold. */
2951 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2953 *has_exactf_sharp_s = TRUE;
2960 && ((*s & S_or_s_mask) == s_masked)
2961 && ((*(s+1) & S_or_s_mask) == s_masked))
2964 /* EXACTF nodes need to know that the minimum length
2965 * changed so that a sharp s in the string can match this
2966 * ss in the pattern, but they remain EXACTF nodes, as they
2967 * won't match this unless the target string is is UTF-8,
2968 * which we don't know until runtime */
2969 if (OP(scan) != EXACTF) {
2970 OP(scan) = EXACTFU_SS;
2974 *min_subtract += len - 1;
2981 /* Allow dumping but overwriting the collection of skipped
2982 * ops and/or strings with fake optimized ops */
2983 n = scan + NODE_SZ_STR(scan);
2991 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2995 /* REx optimizer. Converts nodes into quicker variants "in place".
2996 Finds fixed substrings. */
2998 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2999 to the position after last scanned or to NULL. */
3001 #define INIT_AND_WITHP \
3002 assert(!and_withp); \
3003 Newx(and_withp,1,struct regnode_charclass_class); \
3004 SAVEFREEPV(and_withp)
3006 /* this is a chain of data about sub patterns we are processing that
3007 need to be handled separately/specially in study_chunk. Its so
3008 we can simulate recursion without losing state. */
3010 typedef struct scan_frame {
3011 regnode *last; /* last node to process in this frame */
3012 regnode *next; /* next node to process when last is reached */
3013 struct scan_frame *prev; /*previous frame*/
3014 I32 stop; /* what stopparen do we use */
3018 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3021 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3022 I32 *minlenp, I32 *deltap,
3027 struct regnode_charclass_class *and_withp,
3028 U32 flags, U32 depth)
3029 /* scanp: Start here (read-write). */
3030 /* deltap: Write maxlen-minlen here. */
3031 /* last: Stop before this one. */
3032 /* data: string data about the pattern */
3033 /* stopparen: treat close N as END */
3034 /* recursed: which subroutines have we recursed into */
3035 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3038 I32 min = 0; /* There must be at least this number of characters to match */
3040 regnode *scan = *scanp, *next;
3042 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3043 int is_inf_internal = 0; /* The studied chunk is infinite */
3044 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3045 scan_data_t data_fake;
3046 SV *re_trie_maxbuff = NULL;
3047 regnode *first_non_open = scan;
3048 I32 stopmin = I32_MAX;
3049 scan_frame *frame = NULL;
3050 GET_RE_DEBUG_FLAGS_DECL;
3052 PERL_ARGS_ASSERT_STUDY_CHUNK;
3055 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3059 while (first_non_open && OP(first_non_open) == OPEN)
3060 first_non_open=regnext(first_non_open);
3065 while ( scan && OP(scan) != END && scan < last ){
3066 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3067 node length to get a real minimum (because
3068 the folded version may be shorter) */
3069 bool has_exactf_sharp_s = FALSE;
3070 /* Peephole optimizer: */
3071 DEBUG_STUDYDATA("Peep:", data,depth);
3072 DEBUG_PEEP("Peep",scan,depth);
3074 /* Its not clear to khw or hv why this is done here, and not in the
3075 * clauses that deal with EXACT nodes. khw's guess is that it's
3076 * because of a previous design */
3077 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3079 /* Follow the next-chain of the current node and optimize
3080 away all the NOTHINGs from it. */
3081 if (OP(scan) != CURLYX) {
3082 const int max = (reg_off_by_arg[OP(scan)]
3084 /* I32 may be smaller than U16 on CRAYs! */
3085 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3086 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3090 /* Skip NOTHING and LONGJMP. */
3091 while ((n = regnext(n))
3092 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3093 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3094 && off + noff < max)
3096 if (reg_off_by_arg[OP(scan)])
3099 NEXT_OFF(scan) = off;
3104 /* The principal pseudo-switch. Cannot be a switch, since we
3105 look into several different things. */
3106 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3107 || OP(scan) == IFTHEN) {
3108 next = regnext(scan);
3110 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3112 if (OP(next) == code || code == IFTHEN) {
3113 /* NOTE - There is similar code to this block below for handling
3114 TRIE nodes on a re-study. If you change stuff here check there
3116 I32 max1 = 0, min1 = I32_MAX, num = 0;
3117 struct regnode_charclass_class accum;
3118 regnode * const startbranch=scan;
3120 if (flags & SCF_DO_SUBSTR)
3121 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3122 if (flags & SCF_DO_STCLASS)
3123 cl_init_zero(pRExC_state, &accum);
3125 while (OP(scan) == code) {
3126 I32 deltanext, minnext, f = 0, fake;
3127 struct regnode_charclass_class this_class;
3130 data_fake.flags = 0;
3132 data_fake.whilem_c = data->whilem_c;
3133 data_fake.last_closep = data->last_closep;
3136 data_fake.last_closep = &fake;
3138 data_fake.pos_delta = delta;
3139 next = regnext(scan);
3140 scan = NEXTOPER(scan);
3142 scan = NEXTOPER(scan);
3143 if (flags & SCF_DO_STCLASS) {
3144 cl_init(pRExC_state, &this_class);
3145 data_fake.start_class = &this_class;
3146 f = SCF_DO_STCLASS_AND;
3148 if (flags & SCF_WHILEM_VISITED_POS)
3149 f |= SCF_WHILEM_VISITED_POS;
3151 /* we suppose the run is continuous, last=next...*/
3152 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3154 stopparen, recursed, NULL, f,depth+1);
3157 if (deltanext == I32_MAX) {
3158 is_inf = is_inf_internal = 1;
3160 } else if (max1 < minnext + deltanext)
3161 max1 = minnext + deltanext;
3163 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3165 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3166 if ( stopmin > minnext)
3167 stopmin = min + min1;
3168 flags &= ~SCF_DO_SUBSTR;
3170 data->flags |= SCF_SEEN_ACCEPT;
3173 if (data_fake.flags & SF_HAS_EVAL)
3174 data->flags |= SF_HAS_EVAL;
3175 data->whilem_c = data_fake.whilem_c;
3177 if (flags & SCF_DO_STCLASS)
3178 cl_or(pRExC_state, &accum, &this_class);
3180 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3182 if (flags & SCF_DO_SUBSTR) {
3183 data->pos_min += min1;
3184 if (data->pos_delta >= I32_MAX - (max1 - min1))
3185 data->pos_delta = I32_MAX;
3187 data->pos_delta += max1 - min1;
3188 if (max1 != min1 || is_inf)
3189 data->longest = &(data->longest_float);
3192 if (delta == I32_MAX || I32_MAX - delta - (max1 - min1) < 0)
3195 delta += max1 - min1;
3196 if (flags & SCF_DO_STCLASS_OR) {
3197 cl_or(pRExC_state, data->start_class, &accum);
3199 cl_and(data->start_class, and_withp);
3200 flags &= ~SCF_DO_STCLASS;
3203 else if (flags & SCF_DO_STCLASS_AND) {
3205 cl_and(data->start_class, &accum);
3206 flags &= ~SCF_DO_STCLASS;
3209 /* Switch to OR mode: cache the old value of
3210 * data->start_class */
3212 StructCopy(data->start_class, and_withp,
3213 struct regnode_charclass_class);
3214 flags &= ~SCF_DO_STCLASS_AND;
3215 StructCopy(&accum, data->start_class,
3216 struct regnode_charclass_class);
3217 flags |= SCF_DO_STCLASS_OR;
3218 SET_SSC_EOS(data->start_class);
3222 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3225 Assuming this was/is a branch we are dealing with: 'scan' now
3226 points at the item that follows the branch sequence, whatever
3227 it is. We now start at the beginning of the sequence and look
3234 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3236 If we can find such a subsequence we need to turn the first
3237 element into a trie and then add the subsequent branch exact
3238 strings to the trie.
3242 1. patterns where the whole set of branches can be converted.
3244 2. patterns where only a subset can be converted.
3246 In case 1 we can replace the whole set with a single regop
3247 for the trie. In case 2 we need to keep the start and end
3250 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3251 becomes BRANCH TRIE; BRANCH X;
3253 There is an additional case, that being where there is a
3254 common prefix, which gets split out into an EXACT like node
3255 preceding the TRIE node.
3257 If x(1..n)==tail then we can do a simple trie, if not we make
3258 a "jump" trie, such that when we match the appropriate word
3259 we "jump" to the appropriate tail node. Essentially we turn
3260 a nested if into a case structure of sorts.
3265 if (!re_trie_maxbuff) {
3266 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3267 if (!SvIOK(re_trie_maxbuff))
3268 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3270 if ( SvIV(re_trie_maxbuff)>=0 ) {
3272 regnode *first = (regnode *)NULL;
3273 regnode *last = (regnode *)NULL;
3274 regnode *tail = scan;
3279 SV * const mysv = sv_newmortal(); /* for dumping */
3281 /* var tail is used because there may be a TAIL
3282 regop in the way. Ie, the exacts will point to the
3283 thing following the TAIL, but the last branch will
3284 point at the TAIL. So we advance tail. If we
3285 have nested (?:) we may have to move through several
3289 while ( OP( tail ) == TAIL ) {
3290 /* this is the TAIL generated by (?:) */
3291 tail = regnext( tail );
3295 DEBUG_TRIE_COMPILE_r({
3296 regprop(RExC_rx, mysv, tail );
3297 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3298 (int)depth * 2 + 2, "",
3299 "Looking for TRIE'able sequences. Tail node is: ",
3300 SvPV_nolen_const( mysv )
3306 Step through the branches
3307 cur represents each branch,
3308 noper is the first thing to be matched as part of that branch
3309 noper_next is the regnext() of that node.
3311 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3312 via a "jump trie" but we also support building with NOJUMPTRIE,
3313 which restricts the trie logic to structures like /FOO|BAR/.
3315 If noper is a trieable nodetype then the branch is a possible optimization
3316 target. If we are building under NOJUMPTRIE then we require that noper_next
3317 is the same as scan (our current position in the regex program).
3319 Once we have two or more consecutive such branches we can create a
3320 trie of the EXACT's contents and stitch it in place into the program.
3322 If the sequence represents all of the branches in the alternation we
3323 replace the entire thing with a single TRIE node.
3325 Otherwise when it is a subsequence we need to stitch it in place and
3326 replace only the relevant branches. This means the first branch has
3327 to remain as it is used by the alternation logic, and its next pointer,
3328 and needs to be repointed at the item on the branch chain following
3329 the last branch we have optimized away.
3331 This could be either a BRANCH, in which case the subsequence is internal,
3332 or it could be the item following the branch sequence in which case the
3333 subsequence is at the end (which does not necessarily mean the first node
3334 is the start of the alternation).
3336 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3339 ----------------+-----------
3343 EXACTFU_SS | EXACTFU
3344 EXACTFU_TRICKYFOLD | EXACTFU
3349 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3350 ( EXACT == (X) ) ? EXACT : \
3351 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3354 /* dont use tail as the end marker for this traverse */
3355 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3356 regnode * const noper = NEXTOPER( cur );
3357 U8 noper_type = OP( noper );
3358 U8 noper_trietype = TRIE_TYPE( noper_type );
3359 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3360 regnode * const noper_next = regnext( noper );
3361 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3362 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3365 DEBUG_TRIE_COMPILE_r({
3366 regprop(RExC_rx, mysv, cur);
3367 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3368 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3370 regprop(RExC_rx, mysv, noper);
3371 PerlIO_printf( Perl_debug_log, " -> %s",
3372 SvPV_nolen_const(mysv));
3375 regprop(RExC_rx, mysv, noper_next );
3376 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3377 SvPV_nolen_const(mysv));
3379 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3380 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3381 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3385 /* Is noper a trieable nodetype that can be merged with the
3386 * current trie (if there is one)? */
3390 ( noper_trietype == NOTHING)
3391 || ( trietype == NOTHING )
3392 || ( trietype == noper_trietype )
3395 && noper_next == tail
3399 /* Handle mergable triable node
3400 * Either we are the first node in a new trieable sequence,
3401 * in which case we do some bookkeeping, otherwise we update
3402 * the end pointer. */
3405 if ( noper_trietype == NOTHING ) {
3406 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3407 regnode * const noper_next = regnext( noper );
3408 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3409 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3412 if ( noper_next_trietype ) {
3413 trietype = noper_next_trietype;
3414 } else if (noper_next_type) {
3415 /* a NOTHING regop is 1 regop wide. We need at least two
3416 * for a trie so we can't merge this in */
3420 trietype = noper_trietype;
3423 if ( trietype == NOTHING )
3424 trietype = noper_trietype;
3429 } /* end handle mergable triable node */
3431 /* handle unmergable node -
3432 * noper may either be a triable node which can not be tried
3433 * together with the current trie, or a non triable node */
3435 /* If last is set and trietype is not NOTHING then we have found
3436 * at least two triable branch sequences in a row of a similar
3437 * trietype so we can turn them into a trie. If/when we
3438 * allow NOTHING to start a trie sequence this condition will be
3439 * required, and it isn't expensive so we leave it in for now. */
3440 if ( trietype && trietype != NOTHING )
3441 make_trie( pRExC_state,
3442 startbranch, first, cur, tail, count,
3443 trietype, depth+1 );
3444 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3448 && noper_next == tail
3451 /* noper is triable, so we can start a new trie sequence */
3454 trietype = noper_trietype;
3456 /* if we already saw a first but the current node is not triable then we have
3457 * to reset the first information. */
3462 } /* end handle unmergable node */
3463 } /* loop over branches */
3464 DEBUG_TRIE_COMPILE_r({
3465 regprop(RExC_rx, mysv, cur);
3466 PerlIO_printf( Perl_debug_log,
3467 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3468 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3471 if ( last && trietype ) {
3472 if ( trietype != NOTHING ) {
3473 /* the last branch of the sequence was part of a trie,
3474 * so we have to construct it here outside of the loop
3476 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3477 #ifdef TRIE_STUDY_OPT
3478 if ( ((made == MADE_EXACT_TRIE &&
3479 startbranch == first)
3480 || ( first_non_open == first )) &&
3482 flags |= SCF_TRIE_RESTUDY;
3483 if ( startbranch == first
3486 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3491 /* at this point we know whatever we have is a NOTHING sequence/branch
3492 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3494 if ( startbranch == first ) {
3496 /* the entire thing is a NOTHING sequence, something like this:
3497 * (?:|) So we can turn it into a plain NOTHING op. */
3498 DEBUG_TRIE_COMPILE_r({
3499 regprop(RExC_rx, mysv, cur);
3500 PerlIO_printf( Perl_debug_log,
3501 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3502 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3505 OP(startbranch)= NOTHING;
3506 NEXT_OFF(startbranch)= tail - startbranch;
3507 for ( opt= startbranch + 1; opt < tail ; opt++ )
3511 } /* end if ( last) */
3512 } /* TRIE_MAXBUF is non zero */
3517 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3518 scan = NEXTOPER(NEXTOPER(scan));
3519 } else /* single branch is optimized. */
3520 scan = NEXTOPER(scan);
3522 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3523 scan_frame *newframe = NULL;
3528 if (OP(scan) != SUSPEND) {
3529 /* set the pointer */
3530 if (OP(scan) == GOSUB) {
3532 RExC_recurse[ARG2L(scan)] = scan;
3533 start = RExC_open_parens[paren-1];
3534 end = RExC_close_parens[paren-1];
3537 start = RExC_rxi->program + 1;
3541 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3542 SAVEFREEPV(recursed);
3544 if (!PAREN_TEST(recursed,paren+1)) {
3545 PAREN_SET(recursed,paren+1);
3546 Newx(newframe,1,scan_frame);
3548 if (flags & SCF_DO_SUBSTR) {
3549 SCAN_COMMIT(pRExC_state,data,minlenp);
3550 data->longest = &(data->longest_float);
3552 is_inf = is_inf_internal = 1;
3553 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3554 cl_anything(pRExC_state, data->start_class);
3555 flags &= ~SCF_DO_STCLASS;
3558 Newx(newframe,1,scan_frame);
3561 end = regnext(scan);
3566 SAVEFREEPV(newframe);
3567 newframe->next = regnext(scan);
3568 newframe->last = last;
3569 newframe->stop = stopparen;
3570 newframe->prev = frame;
3580 else if (OP(scan) == EXACT) {
3581 I32 l = STR_LEN(scan);
3584 const U8 * const s = (U8*)STRING(scan);
3585 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3586 l = utf8_length(s, s + l);
3588 uc = *((U8*)STRING(scan));
3591 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3592 /* The code below prefers earlier match for fixed
3593 offset, later match for variable offset. */
3594 if (data->last_end == -1) { /* Update the start info. */
3595 data->last_start_min = data->pos_min;
3596 data->last_start_max = is_inf
3597 ? I32_MAX : data->pos_min + data->pos_delta;
3599 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3601 SvUTF8_on(data->last_found);
3603 SV * const sv = data->last_found;
3604 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3605 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3606 if (mg && mg->mg_len >= 0)
3607 mg->mg_len += utf8_length((U8*)STRING(scan),
3608 (U8*)STRING(scan)+STR_LEN(scan));
3610 data->last_end = data->pos_min + l;
3611 data->pos_min += l; /* As in the first entry. */
3612 data->flags &= ~SF_BEFORE_EOL;
3614 if (flags & SCF_DO_STCLASS_AND) {
3615 /* Check whether it is compatible with what we know already! */
3619 /* If compatible, we or it in below. It is compatible if is
3620 * in the bitmp and either 1) its bit or its fold is set, or 2)
3621 * it's for a locale. Even if there isn't unicode semantics
3622 * here, at runtime there may be because of matching against a
3623 * utf8 string, so accept a possible false positive for
3624 * latin1-range folds */
3626 (!(data->start_class->flags & ANYOF_LOCALE)
3627 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3628 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3629 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3634 ANYOF_CLASS_ZERO(data->start_class);
3635 ANYOF_BITMAP_ZERO(data->start_class);
3637 ANYOF_BITMAP_SET(data->start_class, uc);
3638 else if (uc >= 0x100) {
3641 /* Some Unicode code points fold to the Latin1 range; as
3642 * XXX temporary code, instead of figuring out if this is
3643 * one, just assume it is and set all the start class bits
3644 * that could be some such above 255 code point's fold
3645 * which will generate fals positives. As the code
3646 * elsewhere that does compute the fold settles down, it
3647 * can be extracted out and re-used here */
3648 for (i = 0; i < 256; i++){
3649 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3650 ANYOF_BITMAP_SET(data->start_class, i);
3654 CLEAR_SSC_EOS(data->start_class);
3656 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3658 else if (flags & SCF_DO_STCLASS_OR) {
3659 /* false positive possible if the class is case-folded */
3661 ANYOF_BITMAP_SET(data->start_class, uc);
3663 data->start_class->flags |= ANYOF_UNICODE_ALL;
3664 CLEAR_SSC_EOS(data->start_class);
3665 cl_and(data->start_class, and_withp);
3667 flags &= ~SCF_DO_STCLASS;
3669 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3670 I32 l = STR_LEN(scan);
3671 UV uc = *((U8*)STRING(scan));
3673 /* Search for fixed substrings supports EXACT only. */
3674 if (flags & SCF_DO_SUBSTR) {
3676 SCAN_COMMIT(pRExC_state, data, minlenp);
3679 const U8 * const s = (U8 *)STRING(scan);
3680 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3681 l = utf8_length(s, s + l);
3683 if (has_exactf_sharp_s) {
3684 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3686 min += l - min_subtract;
3688 delta += min_subtract;
3689 if (flags & SCF_DO_SUBSTR) {
3690 data->pos_min += l - min_subtract;
3691 if (data->pos_min < 0) {
3694 data->pos_delta += min_subtract;
3696 data->longest = &(data->longest_float);
3699 if (flags & SCF_DO_STCLASS_AND) {
3700 /* Check whether it is compatible with what we know already! */
3703 (!(data->start_class->flags & ANYOF_LOCALE)
3704 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3705 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3709 ANYOF_CLASS_ZERO(data->start_class);
3710 ANYOF_BITMAP_ZERO(data->start_class);
3712 ANYOF_BITMAP_SET(data->start_class, uc);
3713 CLEAR_SSC_EOS(data->start_class);
3714 if (OP(scan) == EXACTFL) {
3715 /* XXX This set is probably no longer necessary, and
3716 * probably wrong as LOCALE now is on in the initial
3718 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3722 /* Also set the other member of the fold pair. In case
3723 * that unicode semantics is called for at runtime, use
3724 * the full latin1 fold. (Can't do this for locale,
3725 * because not known until runtime) */
3726 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3728 /* All other (EXACTFL handled above) folds except under
3729 * /iaa that include s, S, and sharp_s also may include
3731 if (OP(scan) != EXACTFA) {
3732 if (uc == 's' || uc == 'S') {
3733 ANYOF_BITMAP_SET(data->start_class,
3734 LATIN_SMALL_LETTER_SHARP_S);
3736 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3737 ANYOF_BITMAP_SET(data->start_class, 's');
3738 ANYOF_BITMAP_SET(data->start_class, 'S');
3743 else if (uc >= 0x100) {
3745 for (i = 0; i < 256; i++){
3746 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3747 ANYOF_BITMAP_SET(data->start_class, i);
3752 else if (flags & SCF_DO_STCLASS_OR) {
3753 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3754 /* false positive possible if the class is case-folded.
3755 Assume that the locale settings are the same... */
3757 ANYOF_BITMAP_SET(data->start_class, uc);
3758 if (OP(scan) != EXACTFL) {
3760 /* And set the other member of the fold pair, but
3761 * can't do that in locale because not known until
3763 ANYOF_BITMAP_SET(data->start_class,
3764 PL_fold_latin1[uc]);
3766 /* All folds except under /iaa that include s, S,
3767 * and sharp_s also may include the others */
3768 if (OP(scan) != EXACTFA) {
3769 if (uc == 's' || uc == 'S') {
3770 ANYOF_BITMAP_SET(data->start_class,
3771 LATIN_SMALL_LETTER_SHARP_S);
3773 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3774 ANYOF_BITMAP_SET(data->start_class, 's');
3775 ANYOF_BITMAP_SET(data->start_class, 'S');
3780 CLEAR_SSC_EOS(data->start_class);
3782 cl_and(data->start_class, and_withp);
3784 flags &= ~SCF_DO_STCLASS;
3786 else if (REGNODE_VARIES(OP(scan))) {
3787 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3788 I32 f = flags, pos_before = 0;
3789 regnode * const oscan = scan;
3790 struct regnode_charclass_class this_class;
3791 struct regnode_charclass_class *oclass = NULL;
3792 I32 next_is_eval = 0;
3794 switch (PL_regkind[OP(scan)]) {
3795 case WHILEM: /* End of (?:...)* . */
3796 scan = NEXTOPER(scan);
3799 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3800 next = NEXTOPER(scan);
3801 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3803 maxcount = REG_INFTY;
3804 next = regnext(scan);
3805 scan = NEXTOPER(scan);
3809 if (flags & SCF_DO_SUBSTR)
3814 if (flags & SCF_DO_STCLASS) {
3816 maxcount = REG_INFTY;
3817 next = regnext(scan);
3818 scan = NEXTOPER(scan);
3821 is_inf = is_inf_internal = 1;
3822 scan = regnext(scan);
3823 if (flags & SCF_DO_SUBSTR) {
3824 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3825 data->longest = &(data->longest_float);
3827 goto optimize_curly_tail;
3829 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3830 && (scan->flags == stopparen))
3835 mincount = ARG1(scan);
3836 maxcount = ARG2(scan);
3838 next = regnext(scan);
3839 if (OP(scan) == CURLYX) {
3840 I32 lp = (data ? *(data->last_closep) : 0);
3841 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3843 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3844 next_is_eval = (OP(scan) == EVAL);
3846 if (flags & SCF_DO_SUBSTR) {
3847 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3848 pos_before = data->pos_min;
3852 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3854 data->flags |= SF_IS_INF;
3856 if (flags & SCF_DO_STCLASS) {
3857 cl_init(pRExC_state, &this_class);
3858 oclass = data->start_class;
3859 data->start_class = &this_class;
3860 f |= SCF_DO_STCLASS_AND;
3861 f &= ~SCF_DO_STCLASS_OR;
3863 /* Exclude from super-linear cache processing any {n,m}
3864 regops for which the combination of input pos and regex
3865 pos is not enough information to determine if a match
3868 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3869 regex pos at the \s*, the prospects for a match depend not
3870 only on the input position but also on how many (bar\s*)
3871 repeats into the {4,8} we are. */
3872 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3873 f &= ~SCF_WHILEM_VISITED_POS;
3875 /* This will finish on WHILEM, setting scan, or on NULL: */
3876 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3877 last, data, stopparen, recursed, NULL,
3879 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3881 if (flags & SCF_DO_STCLASS)
3882 data->start_class = oclass;
3883 if (mincount == 0 || minnext == 0) {
3884 if (flags & SCF_DO_STCLASS_OR) {
3885 cl_or(pRExC_state, data->start_class, &this_class);
3887 else if (flags & SCF_DO_STCLASS_AND) {
3888 /* Switch to OR mode: cache the old value of
3889 * data->start_class */
3891 StructCopy(data->start_class, and_withp,
3892 struct regnode_charclass_class);
3893 flags &= ~SCF_DO_STCLASS_AND;
3894 StructCopy(&this_class, data->start_class,
3895 struct regnode_charclass_class);
3896 flags |= SCF_DO_STCLASS_OR;
3897 SET_SSC_EOS(data->start_class);
3899 } else { /* Non-zero len */
3900 if (flags & SCF_DO_STCLASS_OR) {
3901 cl_or(pRExC_state, data->start_class, &this_class);
3902 cl_and(data->start_class, and_withp);
3904 else if (flags & SCF_DO_STCLASS_AND)
3905 cl_and(data->start_class, &this_class);
3906 flags &= ~SCF_DO_STCLASS;
3908 if (!scan) /* It was not CURLYX, but CURLY. */
3910 if ( /* ? quantifier ok, except for (?{ ... }) */
3911 (next_is_eval || !(mincount == 0 && maxcount == 1))
3912 && (minnext == 0) && (deltanext == 0)
3913 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3914 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3916 /* Fatal warnings may leak the regexp without this: */
3917 SAVEFREESV(RExC_rx_sv);
3918 ckWARNreg(RExC_parse,
3919 "Quantifier unexpected on zero-length expression");
3920 (void)ReREFCNT_inc(RExC_rx_sv);
3923 min += minnext * mincount;
3924 is_inf_internal |= deltanext == I32_MAX
3925 || (maxcount == REG_INFTY && minnext + deltanext > 0);
3926 is_inf |= is_inf_internal;
3930 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3932 /* Try powerful optimization CURLYX => CURLYN. */
3933 if ( OP(oscan) == CURLYX && data
3934 && data->flags & SF_IN_PAR
3935 && !(data->flags & SF_HAS_EVAL)
3936 && !deltanext && minnext == 1 ) {
3937 /* Try to optimize to CURLYN. */
3938 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3939 regnode * const nxt1 = nxt;
3946 if (!REGNODE_SIMPLE(OP(nxt))
3947 && !(PL_regkind[OP(nxt)] == EXACT
3948 && STR_LEN(nxt) == 1))
3954 if (OP(nxt) != CLOSE)
3956 if (RExC_open_parens) {
3957 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3958 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3960 /* Now we know that nxt2 is the only contents: */
3961 oscan->flags = (U8)ARG(nxt);
3963 OP(nxt1) = NOTHING; /* was OPEN. */
3966 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3967 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3968 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3969 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3970 OP(nxt + 1) = OPTIMIZED; /* was count. */
3971 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3976 /* Try optimization CURLYX => CURLYM. */
3977 if ( OP(oscan) == CURLYX && data
3978 && !(data->flags & SF_HAS_PAR)
3979 && !(data->flags & SF_HAS_EVAL)
3980 && !deltanext /* atom is fixed width */
3981 && minnext != 0 /* CURLYM can't handle zero width */
3982 && ! (RExC_seen & REG_SEEN_EXACTF_SHARP_S) /* Nor \xDF */
3984 /* XXXX How to optimize if data == 0? */
3985 /* Optimize to a simpler form. */
3986 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3990 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3991 && (OP(nxt2) != WHILEM))
3993 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3994 /* Need to optimize away parenths. */
3995 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
3996 /* Set the parenth number. */
3997 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3999 oscan->flags = (U8)ARG(nxt);
4000 if (RExC_open_parens) {
4001 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4002 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4004 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4005 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4008 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4009 OP(nxt + 1) = OPTIMIZED; /* was count. */
4010 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4011 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4014 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4015 regnode *nnxt = regnext(nxt1);
4017 if (reg_off_by_arg[OP(nxt1)])
4018 ARG_SET(nxt1, nxt2 - nxt1);
4019 else if (nxt2 - nxt1 < U16_MAX)
4020 NEXT_OFF(nxt1) = nxt2 - nxt1;
4022 OP(nxt) = NOTHING; /* Cannot beautify */
4027 /* Optimize again: */
4028 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4029 NULL, stopparen, recursed, NULL, 0,depth+1);
4034 else if ((OP(oscan) == CURLYX)
4035 && (flags & SCF_WHILEM_VISITED_POS)
4036 /* See the comment on a similar expression above.
4037 However, this time it's not a subexpression
4038 we care about, but the expression itself. */
4039 && (maxcount == REG_INFTY)
4040 && data && ++data->whilem_c < 16) {
4041 /* This stays as CURLYX, we can put the count/of pair. */
4042 /* Find WHILEM (as in regexec.c) */
4043 regnode *nxt = oscan + NEXT_OFF(oscan);
4045 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4047 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4048 | (RExC_whilem_seen << 4)); /* On WHILEM */
4050 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4052 if (flags & SCF_DO_SUBSTR) {
4053 SV *last_str = NULL;
4054 int counted = mincount != 0;
4056 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
4057 #if defined(SPARC64_GCC_WORKAROUND)
4060 const char *s = NULL;
4063 if (pos_before >= data->last_start_min)
4066 b = data->last_start_min;
4069 s = SvPV_const(data->last_found, l);
4070 old = b - data->last_start_min;
4073 I32 b = pos_before >= data->last_start_min
4074 ? pos_before : data->last_start_min;
4076 const char * const s = SvPV_const(data->last_found, l);
4077 I32 old = b - data->last_start_min;
4081 old = utf8_hop((U8*)s, old) - (U8*)s;
4083 /* Get the added string: */
4084 last_str = newSVpvn_utf8(s + old, l, UTF);
4085 if (deltanext == 0 && pos_before == b) {
4086 /* What was added is a constant string */
4088 SvGROW(last_str, (mincount * l) + 1);
4089 repeatcpy(SvPVX(last_str) + l,
4090 SvPVX_const(last_str), l, mincount - 1);
4091 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4092 /* Add additional parts. */
4093 SvCUR_set(data->last_found,
4094 SvCUR(data->last_found) - l);
4095 sv_catsv(data->last_found, last_str);
4097 SV * sv = data->last_found;
4099 SvUTF8(sv) && SvMAGICAL(sv) ?
4100 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4101 if (mg && mg->mg_len >= 0)
4102 mg->mg_len += CHR_SVLEN(last_str) - l;
4104 data->last_end += l * (mincount - 1);
4107 /* start offset must point into the last copy */
4108 data->last_start_min += minnext * (mincount - 1);
4109 data->last_start_max += is_inf ? I32_MAX
4110 : (maxcount - 1) * (minnext + data->pos_delta);
4113 /* It is counted once already... */
4114 data->pos_min += minnext * (mincount - counted);
4116 PerlIO_printf(Perl_debug_log, "counted=%d deltanext=%d I32_MAX=%d minnext=%d maxcount=%d mincount=%d\n",
4117 counted, deltanext, I32_MAX, minnext, maxcount, mincount);
4118 if (deltanext != I32_MAX)
4119 PerlIO_printf(Perl_debug_log, "LHS=%d RHS=%d\n", -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount, I32_MAX - data->pos_delta);
4121 if (deltanext == I32_MAX || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= I32_MAX - data->pos_delta)
4122 data->pos_delta = I32_MAX;
4124 data->pos_delta += - counted * deltanext +
4125 (minnext + deltanext) * maxcount - minnext * mincount;
4126 if (mincount != maxcount) {
4127 /* Cannot extend fixed substrings found inside
4129 SCAN_COMMIT(pRExC_state,data,minlenp);
4130 if (mincount && last_str) {
4131 SV * const sv = data->last_found;
4132 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4133 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4137 sv_setsv(sv, last_str);
4138 data->last_end = data->pos_min;
4139 data->last_start_min =
4140 data->pos_min - CHR_SVLEN(last_str);
4141 data->last_start_max = is_inf
4143 : data->pos_min + data->pos_delta
4144 - CHR_SVLEN(last_str);
4146 data->longest = &(data->longest_float);
4148 SvREFCNT_dec(last_str);
4150 if (data && (fl & SF_HAS_EVAL))
4151 data->flags |= SF_HAS_EVAL;
4152 optimize_curly_tail:
4153 if (OP(oscan) != CURLYX) {
4154 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4156 NEXT_OFF(oscan) += NEXT_OFF(next);
4159 default: /* REF, and CLUMP only? */
4160 if (flags & SCF_DO_SUBSTR) {
4161 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4162 data->longest = &(data->longest_float);
4164 is_inf = is_inf_internal = 1;
4165 if (flags & SCF_DO_STCLASS_OR)
4166 cl_anything(pRExC_state, data->start_class);
4167 flags &= ~SCF_DO_STCLASS;
4171 else if (OP(scan) == LNBREAK) {
4172 if (flags & SCF_DO_STCLASS) {
4174 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4175 if (flags & SCF_DO_STCLASS_AND) {
4176 for (value = 0; value < 256; value++)
4177 if (!is_VERTWS_cp(value))
4178 ANYOF_BITMAP_CLEAR(data->start_class, value);
4181 for (value = 0; value < 256; value++)
4182 if (is_VERTWS_cp(value))
4183 ANYOF_BITMAP_SET(data->start_class, value);
4185 if (flags & SCF_DO_STCLASS_OR)
4186 cl_and(data->start_class, and_withp);
4187 flags &= ~SCF_DO_STCLASS;
4190 delta++; /* Because of the 2 char string cr-lf */
4191 if (flags & SCF_DO_SUBSTR) {
4192 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4194 data->pos_delta += 1;
4195 data->longest = &(data->longest_float);
4198 else if (REGNODE_SIMPLE(OP(scan))) {
4201 if (flags & SCF_DO_SUBSTR) {
4202 SCAN_COMMIT(pRExC_state,data,minlenp);
4206 if (flags & SCF_DO_STCLASS) {
4208 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4210 /* Some of the logic below assumes that switching
4211 locale on will only add false positives. */
4212 switch (PL_regkind[OP(scan)]) {
4218 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan));
4221 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4222 cl_anything(pRExC_state, data->start_class);
4225 if (OP(scan) == SANY)
4227 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
4228 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
4229 || ANYOF_CLASS_TEST_ANY_SET(data->start_class));
4230 cl_anything(pRExC_state, data->start_class);
4232 if (flags & SCF_DO_STCLASS_AND || !value)
4233 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
4236 if (flags & SCF_DO_STCLASS_AND)
4237 cl_and(data->start_class,
4238 (struct regnode_charclass_class*)scan);
4240 cl_or(pRExC_state, data->start_class,
4241 (struct regnode_charclass_class*)scan);
4249 classnum = FLAGS(scan);
4250 if (flags & SCF_DO_STCLASS_AND) {
4251 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4252 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum) + 1);
4253 for (value = 0; value < loop_max; value++) {
4254 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4255 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4261 if (data->start_class->flags & ANYOF_LOCALE) {
4262 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum));
4266 /* Even if under locale, set the bits for non-locale
4267 * in case it isn't a true locale-node. This will
4268 * create false positives if it truly is locale */
4269 for (value = 0; value < loop_max; value++) {
4270 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4271 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4283 classnum = FLAGS(scan);
4284 if (flags & SCF_DO_STCLASS_AND) {
4285 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4286 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum));
4287 for (value = 0; value < loop_max; value++) {
4288 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4289 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4295 if (data->start_class->flags & ANYOF_LOCALE) {
4296 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum) + 1);
4300 /* Even if under locale, set the bits for non-locale in
4301 * case it isn't a true locale-node. This will create
4302 * false positives if it truly is locale */
4303 for (value = 0; value < loop_max; value++) {
4304 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4305 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4308 if (PL_regkind[OP(scan)] == NPOSIXD) {
4309 data->start_class->flags |= ANYOF_NON_UTF8_LATIN1_ALL;
4315 if (flags & SCF_DO_STCLASS_OR)
4316 cl_and(data->start_class, and_withp);
4317 flags &= ~SCF_DO_STCLASS;
4320 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
4321 data->flags |= (OP(scan) == MEOL
4324 SCAN_COMMIT(pRExC_state, data, minlenp);
4327 else if ( PL_regkind[OP(scan)] == BRANCHJ
4328 /* Lookbehind, or need to calculate parens/evals/stclass: */
4329 && (scan->flags || data || (flags & SCF_DO_STCLASS))
4330 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
4331 if ( OP(scan) == UNLESSM &&
4333 OP(NEXTOPER(NEXTOPER(scan))) == NOTHING &&
4334 OP(regnext(NEXTOPER(NEXTOPER(scan)))) == SUCCEED
4337 regnode *upto= regnext(scan);
4339 SV * const mysv_val=sv_newmortal();
4340 DEBUG_STUDYDATA("OPFAIL",data,depth);
4342 /*DEBUG_PARSE_MSG("opfail");*/
4343 regprop(RExC_rx, mysv_val, upto);
4344 PerlIO_printf(Perl_debug_log, "~ replace with OPFAIL pointed at %s (%"IVdf") offset %"IVdf"\n",
4345 SvPV_nolen_const(mysv_val),
4346 (IV)REG_NODE_NUM(upto),
4351 NEXT_OFF(scan) = upto - scan;
4352 for (opt= scan + 1; opt < upto ; opt++)
4353 OP(opt) = OPTIMIZED;
4357 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4358 || OP(scan) == UNLESSM )
4360 /* Negative Lookahead/lookbehind
4361 In this case we can't do fixed string optimisation.
4364 I32 deltanext, minnext, fake = 0;
4366 struct regnode_charclass_class intrnl;
4369 data_fake.flags = 0;
4371 data_fake.whilem_c = data->whilem_c;
4372 data_fake.last_closep = data->last_closep;
4375 data_fake.last_closep = &fake;
4376 data_fake.pos_delta = delta;
4377 if ( flags & SCF_DO_STCLASS && !scan->flags
4378 && OP(scan) == IFMATCH ) { /* Lookahead */
4379 cl_init(pRExC_state, &intrnl);
4380 data_fake.start_class = &intrnl;
4381 f |= SCF_DO_STCLASS_AND;
4383 if (flags & SCF_WHILEM_VISITED_POS)
4384 f |= SCF_WHILEM_VISITED_POS;
4385 next = regnext(scan);
4386 nscan = NEXTOPER(NEXTOPER(scan));
4387 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
4388 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
4391 FAIL("Variable length lookbehind not implemented");
4393 else if (minnext > (I32)U8_MAX) {
4394 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4396 scan->flags = (U8)minnext;
4399 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4401 if (data_fake.flags & SF_HAS_EVAL)
4402 data->flags |= SF_HAS_EVAL;
4403 data->whilem_c = data_fake.whilem_c;
4405 if (f & SCF_DO_STCLASS_AND) {
4406 if (flags & SCF_DO_STCLASS_OR) {
4407 /* OR before, AND after: ideally we would recurse with
4408 * data_fake to get the AND applied by study of the
4409 * remainder of the pattern, and then derecurse;
4410 * *** HACK *** for now just treat as "no information".
4411 * See [perl #56690].
4413 cl_init(pRExC_state, data->start_class);
4415 /* AND before and after: combine and continue */
4416 const int was = TEST_SSC_EOS(data->start_class);
4418 cl_and(data->start_class, &intrnl);
4420 SET_SSC_EOS(data->start_class);
4424 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4426 /* Positive Lookahead/lookbehind
4427 In this case we can do fixed string optimisation,
4428 but we must be careful about it. Note in the case of
4429 lookbehind the positions will be offset by the minimum
4430 length of the pattern, something we won't know about
4431 until after the recurse.
4433 I32 deltanext, fake = 0;
4435 struct regnode_charclass_class intrnl;
4437 /* We use SAVEFREEPV so that when the full compile
4438 is finished perl will clean up the allocated
4439 minlens when it's all done. This way we don't
4440 have to worry about freeing them when we know
4441 they wont be used, which would be a pain.
4444 Newx( minnextp, 1, I32 );
4445 SAVEFREEPV(minnextp);
4448 StructCopy(data, &data_fake, scan_data_t);
4449 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
4452 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
4453 data_fake.last_found=newSVsv(data->last_found);
4457 data_fake.last_closep = &fake;
4458 data_fake.flags = 0;
4459 data_fake.pos_delta = delta;
4461 data_fake.flags |= SF_IS_INF;
4462 if ( flags & SCF_DO_STCLASS && !scan->flags
4463 && OP(scan) == IFMATCH ) { /* Lookahead */
4464 cl_init(pRExC_state, &intrnl);
4465 data_fake.start_class = &intrnl;
4466 f |= SCF_DO_STCLASS_AND;
4468 if (flags & SCF_WHILEM_VISITED_POS)
4469 f |= SCF_WHILEM_VISITED_POS;
4470 next = regnext(scan);
4471 nscan = NEXTOPER(NEXTOPER(scan));
4473 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
4474 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
4477 FAIL("Variable length lookbehind not implemented");
4479 else if (*minnextp > (I32)U8_MAX) {
4480 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4482 scan->flags = (U8)*minnextp;
4487 if (f & SCF_DO_STCLASS_AND) {
4488 const int was = TEST_SSC_EOS(data.start_class);
4490 cl_and(data->start_class, &intrnl);
4492 SET_SSC_EOS(data->start_class);
4495 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4497 if (data_fake.flags & SF_HAS_EVAL)
4498 data->flags |= SF_HAS_EVAL;
4499 data->whilem_c = data_fake.whilem_c;
4500 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
4501 if (RExC_rx->minlen<*minnextp)
4502 RExC_rx->minlen=*minnextp;
4503 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
4504 SvREFCNT_dec_NN(data_fake.last_found);
4506 if ( data_fake.minlen_fixed != minlenp )
4508 data->offset_fixed= data_fake.offset_fixed;
4509 data->minlen_fixed= data_fake.minlen_fixed;
4510 data->lookbehind_fixed+= scan->flags;
4512 if ( data_fake.minlen_float != minlenp )
4514 data->minlen_float= data_fake.minlen_float;
4515 data->offset_float_min=data_fake.offset_float_min;
4516 data->offset_float_max=data_fake.offset_float_max;
4517 data->lookbehind_float+= scan->flags;
4524 else if (OP(scan) == OPEN) {
4525 if (stopparen != (I32)ARG(scan))
4528 else if (OP(scan) == CLOSE) {
4529 if (stopparen == (I32)ARG(scan)) {
4532 if ((I32)ARG(scan) == is_par) {
4533 next = regnext(scan);
4535 if ( next && (OP(next) != WHILEM) && next < last)
4536 is_par = 0; /* Disable optimization */
4539 *(data->last_closep) = ARG(scan);
4541 else if (OP(scan) == EVAL) {
4543 data->flags |= SF_HAS_EVAL;
4545 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
4546 if (flags & SCF_DO_SUBSTR) {
4547 SCAN_COMMIT(pRExC_state,data,minlenp);
4548 flags &= ~SCF_DO_SUBSTR;
4550 if (data && OP(scan)==ACCEPT) {
4551 data->flags |= SCF_SEEN_ACCEPT;
4556 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
4558 if (flags & SCF_DO_SUBSTR) {
4559 SCAN_COMMIT(pRExC_state,data,minlenp);
4560 data->longest = &(data->longest_float);
4562 is_inf = is_inf_internal = 1;
4563 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4564 cl_anything(pRExC_state, data->start_class);
4565 flags &= ~SCF_DO_STCLASS;
4567 else if (OP(scan) == GPOS) {
4568 if (!(RExC_rx->extflags & RXf_GPOS_FLOAT) &&
4569 !(delta || is_inf || (data && data->pos_delta)))
4571 if (!(RExC_rx->extflags & RXf_ANCH) && (flags & SCF_DO_SUBSTR))
4572 RExC_rx->extflags |= RXf_ANCH_GPOS;
4573 if (RExC_rx->gofs < (U32)min)
4574 RExC_rx->gofs = min;
4576 RExC_rx->extflags |= RXf_GPOS_FLOAT;
4580 #ifdef TRIE_STUDY_OPT
4581 #ifdef FULL_TRIE_STUDY
4582 else if (PL_regkind[OP(scan)] == TRIE) {
4583 /* NOTE - There is similar code to this block above for handling
4584 BRANCH nodes on the initial study. If you change stuff here
4586 regnode *trie_node= scan;
4587 regnode *tail= regnext(scan);
4588 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4589 I32 max1 = 0, min1 = I32_MAX;
4590 struct regnode_charclass_class accum;
4592 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
4593 SCAN_COMMIT(pRExC_state, data,minlenp); /* Cannot merge strings after this. */
4594 if (flags & SCF_DO_STCLASS)
4595 cl_init_zero(pRExC_state, &accum);
4601 const regnode *nextbranch= NULL;
4604 for ( word=1 ; word <= trie->wordcount ; word++)
4606 I32 deltanext=0, minnext=0, f = 0, fake;
4607 struct regnode_charclass_class this_class;
4609 data_fake.flags = 0;
4611 data_fake.whilem_c = data->whilem_c;
4612 data_fake.last_closep = data->last_closep;
4615 data_fake.last_closep = &fake;
4616 data_fake.pos_delta = delta;
4617 if (flags & SCF_DO_STCLASS) {
4618 cl_init(pRExC_state, &this_class);
4619 data_fake.start_class = &this_class;
4620 f = SCF_DO_STCLASS_AND;
4622 if (flags & SCF_WHILEM_VISITED_POS)
4623 f |= SCF_WHILEM_VISITED_POS;
4625 if (trie->jump[word]) {
4627 nextbranch = trie_node + trie->jump[0];
4628 scan= trie_node + trie->jump[word];
4629 /* We go from the jump point to the branch that follows
4630 it. Note this means we need the vestigal unused branches
4631 even though they arent otherwise used.
4633 minnext = study_chunk(pRExC_state, &scan, minlenp,
4634 &deltanext, (regnode *)nextbranch, &data_fake,
4635 stopparen, recursed, NULL, f,depth+1);
4637 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
4638 nextbranch= regnext((regnode*)nextbranch);
4640 if (min1 > (I32)(minnext + trie->minlen))
4641 min1 = minnext + trie->minlen;
4642 if (deltanext == I32_MAX) {
4643 is_inf = is_inf_internal = 1;
4645 } else if (max1 < (I32)(minnext + deltanext + trie->maxlen))
4646 max1 = minnext + deltanext + trie->maxlen;
4648 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4650 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4651 if ( stopmin > min + min1)
4652 stopmin = min + min1;
4653 flags &= ~SCF_DO_SUBSTR;
4655 data->flags |= SCF_SEEN_ACCEPT;
4658 if (data_fake.flags & SF_HAS_EVAL)
4659 data->flags |= SF_HAS_EVAL;
4660 data->whilem_c = data_fake.whilem_c;
4662 if (flags & SCF_DO_STCLASS)
4663 cl_or(pRExC_state, &accum, &this_class);
4666 if (flags & SCF_DO_SUBSTR) {
4667 data->pos_min += min1;
4668 data->pos_delta += max1 - min1;
4669 if (max1 != min1 || is_inf)
4670 data->longest = &(data->longest_float);
4673 delta += max1 - min1;
4674 if (flags & SCF_DO_STCLASS_OR) {
4675 cl_or(pRExC_state, data->start_class, &accum);
4677 cl_and(data->start_class, and_withp);
4678 flags &= ~SCF_DO_STCLASS;
4681 else if (flags & SCF_DO_STCLASS_AND) {
4683 cl_and(data->start_class, &accum);
4684 flags &= ~SCF_DO_STCLASS;
4687 /* Switch to OR mode: cache the old value of
4688 * data->start_class */
4690 StructCopy(data->start_class, and_withp,
4691 struct regnode_charclass_class);
4692 flags &= ~SCF_DO_STCLASS_AND;
4693 StructCopy(&accum, data->start_class,
4694 struct regnode_charclass_class);
4695 flags |= SCF_DO_STCLASS_OR;
4696 SET_SSC_EOS(data->start_class);
4703 else if (PL_regkind[OP(scan)] == TRIE) {
4704 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4707 min += trie->minlen;
4708 delta += (trie->maxlen - trie->minlen);
4709 flags &= ~SCF_DO_STCLASS; /* xxx */
4710 if (flags & SCF_DO_SUBSTR) {
4711 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4712 data->pos_min += trie->minlen;
4713 data->pos_delta += (trie->maxlen - trie->minlen);
4714 if (trie->maxlen != trie->minlen)
4715 data->longest = &(data->longest_float);
4717 if (trie->jump) /* no more substrings -- for now /grr*/
4718 flags &= ~SCF_DO_SUBSTR;
4720 #endif /* old or new */
4721 #endif /* TRIE_STUDY_OPT */
4723 /* Else: zero-length, ignore. */
4724 scan = regnext(scan);
4729 stopparen = frame->stop;
4730 frame = frame->prev;
4731 goto fake_study_recurse;
4736 DEBUG_STUDYDATA("pre-fin:",data,depth);
4739 *deltap = is_inf_internal ? I32_MAX : delta;
4740 if (flags & SCF_DO_SUBSTR && is_inf)
4741 data->pos_delta = I32_MAX - data->pos_min;
4742 if (is_par > (I32)U8_MAX)
4744 if (is_par && pars==1 && data) {
4745 data->flags |= SF_IN_PAR;
4746 data->flags &= ~SF_HAS_PAR;
4748 else if (pars && data) {
4749 data->flags |= SF_HAS_PAR;
4750 data->flags &= ~SF_IN_PAR;
4752 if (flags & SCF_DO_STCLASS_OR)
4753 cl_and(data->start_class, and_withp);
4754 if (flags & SCF_TRIE_RESTUDY)
4755 data->flags |= SCF_TRIE_RESTUDY;
4757 DEBUG_STUDYDATA("post-fin:",data,depth);
4759 return min < stopmin ? min : stopmin;
4763 S_add_data(RExC_state_t *pRExC_state, U32 n, const char *s)
4765 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
4767 PERL_ARGS_ASSERT_ADD_DATA;
4769 Renewc(RExC_rxi->data,
4770 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
4771 char, struct reg_data);
4773 Renew(RExC_rxi->data->what, count + n, U8);
4775 Newx(RExC_rxi->data->what, n, U8);
4776 RExC_rxi->data->count = count + n;
4777 Copy(s, RExC_rxi->data->what + count, n, U8);
4781 /*XXX: todo make this not included in a non debugging perl */
4782 #ifndef PERL_IN_XSUB_RE
4784 Perl_reginitcolors(pTHX)
4787 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
4789 char *t = savepv(s);
4793 t = strchr(t, '\t');
4799 PL_colors[i] = t = (char *)"";
4804 PL_colors[i++] = (char *)"";
4811 #ifdef TRIE_STUDY_OPT
4812 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
4815 (data.flags & SCF_TRIE_RESTUDY) \
4823 #define CHECK_RESTUDY_GOTO_butfirst
4827 * pregcomp - compile a regular expression into internal code
4829 * Decides which engine's compiler to call based on the hint currently in
4833 #ifndef PERL_IN_XSUB_RE
4835 /* return the currently in-scope regex engine (or the default if none) */
4837 regexp_engine const *
4838 Perl_current_re_engine(pTHX)
4842 if (IN_PERL_COMPILETIME) {
4843 HV * const table = GvHV(PL_hintgv);
4847 return &PL_core_reg_engine;
4848 ptr = hv_fetchs(table, "regcomp", FALSE);
4849 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
4850 return &PL_core_reg_engine;
4851 return INT2PTR(regexp_engine*,SvIV(*ptr));
4855 if (!PL_curcop->cop_hints_hash)
4856 return &PL_core_reg_engine;
4857 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
4858 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
4859 return &PL_core_reg_engine;
4860 return INT2PTR(regexp_engine*,SvIV(ptr));
4866 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
4869 regexp_engine const *eng = current_re_engine();
4870 GET_RE_DEBUG_FLAGS_DECL;
4872 PERL_ARGS_ASSERT_PREGCOMP;
4874 /* Dispatch a request to compile a regexp to correct regexp engine. */
4876 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
4879 return CALLREGCOMP_ENG(eng, pattern, flags);
4883 /* public(ish) entry point for the perl core's own regex compiling code.
4884 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
4885 * pattern rather than a list of OPs, and uses the internal engine rather
4886 * than the current one */
4889 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
4891 SV *pat = pattern; /* defeat constness! */
4892 PERL_ARGS_ASSERT_RE_COMPILE;
4893 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
4894 #ifdef PERL_IN_XSUB_RE
4897 &PL_core_reg_engine,
4899 NULL, NULL, rx_flags, 0);
4903 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
4904 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
4905 * point to the realloced string and length.
4907 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
4911 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
4912 char **pat_p, STRLEN *plen_p, int num_code_blocks)
4914 U8 *const src = (U8*)*pat_p;
4917 STRLEN s = 0, d = 0;
4919 GET_RE_DEBUG_FLAGS_DECL;
4921 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
4922 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
4924 Newx(dst, *plen_p * 2 + 1, U8);
4926 while (s < *plen_p) {
4927 const UV uv = NATIVE_TO_ASCII(src[s]);
4928 if (UNI_IS_INVARIANT(uv))
4929 dst[d] = (U8)UTF_TO_NATIVE(uv);
4931 dst[d++] = (U8)UTF8_EIGHT_BIT_HI(uv);
4932 dst[d] = (U8)UTF8_EIGHT_BIT_LO(uv);
4934 if (n < num_code_blocks) {
4935 if (!do_end && pRExC_state->code_blocks[n].start == s) {
4936 pRExC_state->code_blocks[n].start = d;
4937 assert(dst[d] == '(');
4940 else if (do_end && pRExC_state->code_blocks[n].end == s) {
4941 pRExC_state->code_blocks[n].end = d;
4942 assert(dst[d] == ')');
4952 *pat_p = (char*) dst;
4954 RExC_orig_utf8 = RExC_utf8 = 1;
4959 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
4960 * while recording any code block indices, and handling overloading,
4961 * nested qr// objects etc. If pat is null, it will allocate a new
4962 * string, or just return the first arg, if there's only one.
4964 * Returns the malloced/updated pat.
4965 * patternp and pat_count is the array of SVs to be concatted;
4966 * oplist is the optional list of ops that generated the SVs;
4967 * recompile_p is a pointer to a boolean that will be set if
4968 * the regex will need to be recompiled.
4969 * delim, if non-null is an SV that will be inserted between each element
4973 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
4974 SV *pat, SV ** const patternp, int pat_count,
4975 OP *oplist, bool *recompile_p, SV *delim)
4979 bool use_delim = FALSE;
4980 bool alloced = FALSE;
4982 /* if we know we have at least two args, create an empty string,
4983 * then concatenate args to that. For no args, return an empty string */
4984 if (!pat && pat_count != 1) {
4985 pat = newSVpvn("", 0);
4990 for (svp = patternp; svp < patternp + pat_count; svp++) {
4993 STRLEN orig_patlen = 0;
4995 SV *msv = use_delim ? delim : *svp;
4997 /* if we've got a delimiter, we go round the loop twice for each
4998 * svp slot (except the last), using the delimiter the second
5007 if (SvTYPE(msv) == SVt_PVAV) {
5008 /* we've encountered an interpolated array within
5009 * the pattern, e.g. /...@a..../. Expand the list of elements,
5010 * then recursively append elements.
5011 * The code in this block is based on S_pushav() */
5013 AV *const av = (AV*)msv;
5014 const I32 maxarg = AvFILL(av) + 1;
5018 assert(oplist->op_type == OP_PADAV
5019 || oplist->op_type == OP_RV2AV);
5020 oplist = oplist->op_sibling;;
5023 if (SvRMAGICAL(av)) {
5026 Newx(array, maxarg, SV*);
5028 for (i=0; i < (U32)maxarg; i++) {
5029 SV ** const svp = av_fetch(av, i, FALSE);
5030 array[i] = svp ? *svp : &PL_sv_undef;
5034 array = AvARRAY(av);
5036 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5037 array, maxarg, NULL, recompile_p,
5039 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5045 /* we make the assumption here that each op in the list of
5046 * op_siblings maps to one SV pushed onto the stack,
5047 * except for code blocks, with have both an OP_NULL and
5049 * This allows us to match up the list of SVs against the
5050 * list of OPs to find the next code block.
5052 * Note that PUSHMARK PADSV PADSV ..
5054 * PADRANGE PADSV PADSV ..
5055 * so the alignment still works. */
5058 if (oplist->op_type == OP_NULL
5059 && (oplist->op_flags & OPf_SPECIAL))
5061 assert(n < pRExC_state->num_code_blocks);
5062 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5063 pRExC_state->code_blocks[n].block = oplist;
5064 pRExC_state->code_blocks[n].src_regex = NULL;
5067 oplist = oplist->op_sibling; /* skip CONST */
5070 oplist = oplist->op_sibling;;
5073 /* apply magic and QR overloading to arg */
5076 if (SvROK(msv) && SvAMAGIC(msv)) {
5077 SV *sv = AMG_CALLunary(msv, regexp_amg);
5081 if (SvTYPE(sv) != SVt_REGEXP)
5082 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5087 /* try concatenation overload ... */
5088 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5089 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5092 /* overloading involved: all bets are off over literal
5093 * code. Pretend we haven't seen it */
5094 pRExC_state->num_code_blocks -= n;
5098 /* ... or failing that, try "" overload */
5099 while (SvAMAGIC(msv)
5100 && (sv = AMG_CALLunary(msv, string_amg))
5104 && SvRV(msv) == SvRV(sv))
5109 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5113 /* this is a partially unrolled
5114 * sv_catsv_nomg(pat, msv);
5115 * that allows us to adjust code block indices if
5118 char *dst = SvPV_force_nomg(pat, dlen);
5119 const char *src = SvPV_flags_const(msv, slen, 0);
5121 if (SvUTF8(msv) && !SvUTF8(pat)) {
5122 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
5123 sv_setpvn(pat, dst, dlen);
5126 sv_catpvn_nomg(pat, src, slen);
5133 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5136 /* extract any code blocks within any embedded qr//'s */
5137 if (rx && SvTYPE(rx) == SVt_REGEXP
5138 && RX_ENGINE((REGEXP*)rx)->op_comp)
5141 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5142 if (ri->num_code_blocks) {
5144 /* the presence of an embedded qr// with code means
5145 * we should always recompile: the text of the
5146 * qr// may not have changed, but it may be a
5147 * different closure than last time */
5149 Renew(pRExC_state->code_blocks,
5150 pRExC_state->num_code_blocks + ri->num_code_blocks,
5151 struct reg_code_block);
5152 pRExC_state->num_code_blocks += ri->num_code_blocks;
5154 for (i=0; i < ri->num_code_blocks; i++) {
5155 struct reg_code_block *src, *dst;
5156 STRLEN offset = orig_patlen
5157 + ReANY((REGEXP *)rx)->pre_prefix;
5158 assert(n < pRExC_state->num_code_blocks);
5159 src = &ri->code_blocks[i];
5160 dst = &pRExC_state->code_blocks[n];
5161 dst->start = src->start + offset;
5162 dst->end = src->end + offset;
5163 dst->block = src->block;
5164 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5173 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
5182 /* see if there are any run-time code blocks in the pattern.
5183 * False positives are allowed */
5186 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5187 char *pat, STRLEN plen)
5192 for (s = 0; s < plen; s++) {
5193 if (n < pRExC_state->num_code_blocks
5194 && s == pRExC_state->code_blocks[n].start)
5196 s = pRExC_state->code_blocks[n].end;
5200 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
5202 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
5204 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
5211 /* Handle run-time code blocks. We will already have compiled any direct
5212 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
5213 * copy of it, but with any literal code blocks blanked out and
5214 * appropriate chars escaped; then feed it into
5216 * eval "qr'modified_pattern'"
5220 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
5224 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
5226 * After eval_sv()-ing that, grab any new code blocks from the returned qr
5227 * and merge them with any code blocks of the original regexp.
5229 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
5230 * instead, just save the qr and return FALSE; this tells our caller that
5231 * the original pattern needs upgrading to utf8.
5235 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5236 char *pat, STRLEN plen)
5240 GET_RE_DEBUG_FLAGS_DECL;
5242 if (pRExC_state->runtime_code_qr) {
5243 /* this is the second time we've been called; this should
5244 * only happen if the main pattern got upgraded to utf8
5245 * during compilation; re-use the qr we compiled first time
5246 * round (which should be utf8 too)
5248 qr = pRExC_state->runtime_code_qr;
5249 pRExC_state->runtime_code_qr = NULL;
5250 assert(RExC_utf8 && SvUTF8(qr));
5256 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
5260 /* determine how many extra chars we need for ' and \ escaping */
5261 for (s = 0; s < plen; s++) {
5262 if (pat[s] == '\'' || pat[s] == '\\')
5266 Newx(newpat, newlen, char);
5268 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
5270 for (s = 0; s < plen; s++) {
5271 if (n < pRExC_state->num_code_blocks
5272 && s == pRExC_state->code_blocks[n].start)
5274 /* blank out literal code block */
5275 assert(pat[s] == '(');
5276 while (s <= pRExC_state->code_blocks[n].end) {
5284 if (pat[s] == '\'' || pat[s] == '\\')
5289 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
5293 PerlIO_printf(Perl_debug_log,
5294 "%sre-parsing pattern for runtime code:%s %s\n",
5295 PL_colors[4],PL_colors[5],newpat);
5298 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5304 PUSHSTACKi(PERLSI_REQUIRE);
5305 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
5306 * parsing qr''; normally only q'' does this. It also alters
5308 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
5309 SvREFCNT_dec_NN(sv);
5314 SV * const errsv = ERRSV;
5315 if (SvTRUE_NN(errsv))
5317 Safefree(pRExC_state->code_blocks);
5318 /* use croak_sv ? */
5319 Perl_croak_nocontext("%s", SvPV_nolen_const(errsv));
5322 assert(SvROK(qr_ref));
5324 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5325 /* the leaving below frees the tmp qr_ref.
5326 * Give qr a life of its own */
5334 if (!RExC_utf8 && SvUTF8(qr)) {
5335 /* first time through; the pattern got upgraded; save the
5336 * qr for the next time through */
5337 assert(!pRExC_state->runtime_code_qr);
5338 pRExC_state->runtime_code_qr = qr;
5343 /* extract any code blocks within the returned qr// */
5346 /* merge the main (r1) and run-time (r2) code blocks into one */
5348 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
5349 struct reg_code_block *new_block, *dst;
5350 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
5353 if (!r2->num_code_blocks) /* we guessed wrong */
5355 SvREFCNT_dec_NN(qr);
5360 r1->num_code_blocks + r2->num_code_blocks,
5361 struct reg_code_block);
5364 while ( i1 < r1->num_code_blocks
5365 || i2 < r2->num_code_blocks)
5367 struct reg_code_block *src;
5370 if (i1 == r1->num_code_blocks) {
5371 src = &r2->code_blocks[i2++];
5374 else if (i2 == r2->num_code_blocks)
5375 src = &r1->code_blocks[i1++];
5376 else if ( r1->code_blocks[i1].start
5377 < r2->code_blocks[i2].start)
5379 src = &r1->code_blocks[i1++];
5380 assert(src->end < r2->code_blocks[i2].start);
5383 assert( r1->code_blocks[i1].start
5384 > r2->code_blocks[i2].start);
5385 src = &r2->code_blocks[i2++];
5387 assert(src->end < r1->code_blocks[i1].start);
5390 assert(pat[src->start] == '(');
5391 assert(pat[src->end] == ')');
5392 dst->start = src->start;
5393 dst->end = src->end;
5394 dst->block = src->block;
5395 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
5399 r1->num_code_blocks += r2->num_code_blocks;
5400 Safefree(r1->code_blocks);
5401 r1->code_blocks = new_block;
5404 SvREFCNT_dec_NN(qr);
5410 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)
5412 /* This is the common code for setting up the floating and fixed length
5413 * string data extracted from Perl_re_op_compile() below. Returns a boolean
5414 * as to whether succeeded or not */
5418 if (! (longest_length
5419 || (eol /* Can't have SEOL and MULTI */
5420 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
5422 /* See comments for join_exact for why REG_SEEN_EXACTF_SHARP_S */
5423 || (RExC_seen & REG_SEEN_EXACTF_SHARP_S))
5428 /* copy the information about the longest from the reg_scan_data
5429 over to the program. */
5430 if (SvUTF8(sv_longest)) {
5431 *rx_utf8 = sv_longest;
5434 *rx_substr = sv_longest;
5437 /* end_shift is how many chars that must be matched that
5438 follow this item. We calculate it ahead of time as once the
5439 lookbehind offset is added in we lose the ability to correctly
5441 ml = minlen ? *(minlen) : (I32)longest_length;
5442 *rx_end_shift = ml - offset
5443 - longest_length + (SvTAIL(sv_longest) != 0)
5446 t = (eol/* Can't have SEOL and MULTI */
5447 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
5448 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
5454 * Perl_re_op_compile - the perl internal RE engine's function to compile a
5455 * regular expression into internal code.
5456 * The pattern may be passed either as:
5457 * a list of SVs (patternp plus pat_count)
5458 * a list of OPs (expr)
5459 * If both are passed, the SV list is used, but the OP list indicates
5460 * which SVs are actually pre-compiled code blocks
5462 * The SVs in the list have magic and qr overloading applied to them (and
5463 * the list may be modified in-place with replacement SVs in the latter
5466 * If the pattern hasn't changed from old_re, then old_re will be
5469 * eng is the current engine. If that engine has an op_comp method, then
5470 * handle directly (i.e. we assume that op_comp was us); otherwise, just
5471 * do the initial concatenation of arguments and pass on to the external
5474 * If is_bare_re is not null, set it to a boolean indicating whether the
5475 * arg list reduced (after overloading) to a single bare regex which has
5476 * been returned (i.e. /$qr/).
5478 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
5480 * pm_flags contains the PMf_* flags, typically based on those from the
5481 * pm_flags field of the related PMOP. Currently we're only interested in
5482 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
5484 * We can't allocate space until we know how big the compiled form will be,
5485 * but we can't compile it (and thus know how big it is) until we've got a
5486 * place to put the code. So we cheat: we compile it twice, once with code
5487 * generation turned off and size counting turned on, and once "for real".
5488 * This also means that we don't allocate space until we are sure that the
5489 * thing really will compile successfully, and we never have to move the
5490 * code and thus invalidate pointers into it. (Note that it has to be in
5491 * one piece because free() must be able to free it all.) [NB: not true in perl]
5493 * Beware that the optimization-preparation code in here knows about some
5494 * of the structure of the compiled regexp. [I'll say.]
5498 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
5499 OP *expr, const regexp_engine* eng, REGEXP *old_re,
5500 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
5505 regexp_internal *ri;
5513 SV *code_blocksv = NULL;
5514 SV** new_patternp = patternp;
5516 /* these are all flags - maybe they should be turned
5517 * into a single int with different bit masks */
5518 I32 sawlookahead = 0;
5521 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
5523 bool runtime_code = 0;
5525 RExC_state_t RExC_state;
5526 RExC_state_t * const pRExC_state = &RExC_state;
5527 #ifdef TRIE_STUDY_OPT
5529 RExC_state_t copyRExC_state;
5531 GET_RE_DEBUG_FLAGS_DECL;
5533 PERL_ARGS_ASSERT_RE_OP_COMPILE;
5535 DEBUG_r(if (!PL_colorset) reginitcolors());
5537 #ifndef PERL_IN_XSUB_RE
5538 /* Initialize these here instead of as-needed, as is quick and avoids
5539 * having to test them each time otherwise */
5540 if (! PL_AboveLatin1) {
5541 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
5542 PL_ASCII = _new_invlist_C_array(ASCII_invlist);
5543 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
5545 PL_L1Posix_ptrs[_CC_ALPHANUMERIC]
5546 = _new_invlist_C_array(L1PosixAlnum_invlist);
5547 PL_Posix_ptrs[_CC_ALPHANUMERIC]
5548 = _new_invlist_C_array(PosixAlnum_invlist);
5550 PL_L1Posix_ptrs[_CC_ALPHA]
5551 = _new_invlist_C_array(L1PosixAlpha_invlist);
5552 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(PosixAlpha_invlist);
5554 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(PosixBlank_invlist);
5555 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(XPosixBlank_invlist);
5557 /* Cased is the same as Alpha in the ASCII range */
5558 PL_L1Posix_ptrs[_CC_CASED] = _new_invlist_C_array(L1Cased_invlist);
5559 PL_Posix_ptrs[_CC_CASED] = _new_invlist_C_array(PosixAlpha_invlist);
5561 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(PosixCntrl_invlist);
5562 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(XPosixCntrl_invlist);
5564 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5565 PL_L1Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5567 PL_L1Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(L1PosixGraph_invlist);
5568 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(PosixGraph_invlist);
5570 PL_L1Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(L1PosixLower_invlist);
5571 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(PosixLower_invlist);
5573 PL_L1Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(L1PosixPrint_invlist);
5574 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(PosixPrint_invlist);
5576 PL_L1Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(L1PosixPunct_invlist);
5577 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(PosixPunct_invlist);
5579 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(PerlSpace_invlist);
5580 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(XPerlSpace_invlist);
5581 PL_Posix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(PosixSpace_invlist);
5582 PL_XPosix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(XPosixSpace_invlist);
5584 PL_L1Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(L1PosixUpper_invlist);
5585 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(PosixUpper_invlist);
5587 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(VertSpace_invlist);
5589 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(PosixWord_invlist);
5590 PL_L1Posix_ptrs[_CC_WORDCHAR]
5591 = _new_invlist_C_array(L1PosixWord_invlist);
5593 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(PosixXDigit_invlist);
5594 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(XPosixXDigit_invlist);
5596 PL_HasMultiCharFold = _new_invlist_C_array(_Perl_Multi_Char_Folds_invlist);
5600 pRExC_state->code_blocks = NULL;
5601 pRExC_state->num_code_blocks = 0;
5604 *is_bare_re = FALSE;
5606 if (expr && (expr->op_type == OP_LIST ||
5607 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
5608 /* allocate code_blocks if needed */
5612 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling)
5613 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
5614 ncode++; /* count of DO blocks */
5616 pRExC_state->num_code_blocks = ncode;
5617 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
5622 /* compile-time pattern with just OP_CONSTs and DO blocks */
5627 /* find how many CONSTs there are */
5630 if (expr->op_type == OP_CONST)
5633 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5634 if (o->op_type == OP_CONST)
5638 /* fake up an SV array */
5640 assert(!new_patternp);
5641 Newx(new_patternp, n, SV*);
5642 SAVEFREEPV(new_patternp);
5646 if (expr->op_type == OP_CONST)
5647 new_patternp[n] = cSVOPx_sv(expr);
5649 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5650 if (o->op_type == OP_CONST)
5651 new_patternp[n++] = cSVOPo_sv;
5656 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5657 "Assembling pattern from %d elements%s\n", pat_count,
5658 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5660 /* set expr to the first arg op */
5662 if (pRExC_state->num_code_blocks
5663 && expr->op_type != OP_CONST)
5665 expr = cLISTOPx(expr)->op_first;
5666 assert( expr->op_type == OP_PUSHMARK
5667 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
5668 || expr->op_type == OP_PADRANGE);
5669 expr = expr->op_sibling;
5672 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
5673 expr, &recompile, NULL);
5675 /* handle bare (possibly after overloading) regex: foo =~ $re */
5680 if (SvTYPE(re) == SVt_REGEXP) {
5684 Safefree(pRExC_state->code_blocks);
5685 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5686 "Precompiled pattern%s\n",
5687 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5693 exp = SvPV_nomg(pat, plen);
5695 if (!eng->op_comp) {
5696 if ((SvUTF8(pat) && IN_BYTES)
5697 || SvGMAGICAL(pat) || SvAMAGIC(pat))
5699 /* make a temporary copy; either to convert to bytes,
5700 * or to avoid repeating get-magic / overloaded stringify */
5701 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
5702 (IN_BYTES ? 0 : SvUTF8(pat)));
5704 Safefree(pRExC_state->code_blocks);
5705 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
5708 /* ignore the utf8ness if the pattern is 0 length */
5709 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
5710 RExC_uni_semantics = 0;
5711 RExC_contains_locale = 0;
5712 pRExC_state->runtime_code_qr = NULL;
5715 SV *dsv= sv_newmortal();
5716 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
5717 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
5718 PL_colors[4],PL_colors[5],s);
5722 /* we jump here if we upgrade the pattern to utf8 and have to
5725 if ((pm_flags & PMf_USE_RE_EVAL)
5726 /* this second condition covers the non-regex literal case,
5727 * i.e. $foo =~ '(?{})'. */
5728 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
5730 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
5732 /* return old regex if pattern hasn't changed */
5733 /* XXX: note in the below we have to check the flags as well as the pattern.
5735 * Things get a touch tricky as we have to compare the utf8 flag independently
5736 * from the compile flags.
5741 && !!RX_UTF8(old_re) == !!RExC_utf8
5742 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
5743 && RX_PRECOMP(old_re)
5744 && RX_PRELEN(old_re) == plen
5745 && memEQ(RX_PRECOMP(old_re), exp, plen)
5746 && !runtime_code /* with runtime code, always recompile */ )
5748 Safefree(pRExC_state->code_blocks);
5752 rx_flags = orig_rx_flags;
5754 if (initial_charset == REGEX_LOCALE_CHARSET) {
5755 RExC_contains_locale = 1;
5757 else if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
5759 /* Set to use unicode semantics if the pattern is in utf8 and has the
5760 * 'depends' charset specified, as it means unicode when utf8 */
5761 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5765 RExC_flags = rx_flags;
5766 RExC_pm_flags = pm_flags;
5769 if (TAINTING_get && TAINT_get)
5770 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
5772 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
5773 /* whoops, we have a non-utf8 pattern, whilst run-time code
5774 * got compiled as utf8. Try again with a utf8 pattern */
5775 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
5776 pRExC_state->num_code_blocks);
5777 goto redo_first_pass;
5780 assert(!pRExC_state->runtime_code_qr);
5785 RExC_in_lookbehind = 0;
5786 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
5788 RExC_override_recoding = 0;
5789 RExC_in_multi_char_class = 0;
5791 /* First pass: determine size, legality. */
5794 RExC_end = exp + plen;
5799 RExC_emit = &PL_regdummy;
5800 RExC_whilem_seen = 0;
5801 RExC_open_parens = NULL;
5802 RExC_close_parens = NULL;
5804 RExC_paren_names = NULL;
5806 RExC_paren_name_list = NULL;
5808 RExC_recurse = NULL;
5809 RExC_recurse_count = 0;
5810 pRExC_state->code_index = 0;
5812 #if 0 /* REGC() is (currently) a NOP at the first pass.
5813 * Clever compilers notice this and complain. --jhi */
5814 REGC((U8)REG_MAGIC, (char*)RExC_emit);
5817 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
5819 RExC_lastparse=NULL;
5821 /* reg may croak on us, not giving us a chance to free
5822 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
5823 need it to survive as long as the regexp (qr/(?{})/).
5824 We must check that code_blocksv is not already set, because we may
5825 have jumped back to restart the sizing pass. */
5826 if (pRExC_state->code_blocks && !code_blocksv) {
5827 code_blocksv = newSV_type(SVt_PV);
5828 SAVEFREESV(code_blocksv);
5829 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
5830 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
5832 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5833 /* It's possible to write a regexp in ascii that represents Unicode
5834 codepoints outside of the byte range, such as via \x{100}. If we
5835 detect such a sequence we have to convert the entire pattern to utf8
5836 and then recompile, as our sizing calculation will have been based
5837 on 1 byte == 1 character, but we will need to use utf8 to encode
5838 at least some part of the pattern, and therefore must convert the whole
5841 if (flags & RESTART_UTF8) {
5842 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
5843 pRExC_state->num_code_blocks);
5844 goto redo_first_pass;
5846 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
5849 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
5852 PerlIO_printf(Perl_debug_log,
5853 "Required size %"IVdf" nodes\n"
5854 "Starting second pass (creation)\n",
5857 RExC_lastparse=NULL;
5860 /* The first pass could have found things that force Unicode semantics */
5861 if ((RExC_utf8 || RExC_uni_semantics)
5862 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
5864 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5867 /* Small enough for pointer-storage convention?
5868 If extralen==0, this means that we will not need long jumps. */
5869 if (RExC_size >= 0x10000L && RExC_extralen)
5870 RExC_size += RExC_extralen;
5873 if (RExC_whilem_seen > 15)
5874 RExC_whilem_seen = 15;
5876 /* Allocate space and zero-initialize. Note, the two step process
5877 of zeroing when in debug mode, thus anything assigned has to
5878 happen after that */
5879 rx = (REGEXP*) newSV_type(SVt_REGEXP);
5881 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
5882 char, regexp_internal);
5883 if ( r == NULL || ri == NULL )
5884 FAIL("Regexp out of space");
5886 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
5887 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode), char);
5889 /* bulk initialize base fields with 0. */
5890 Zero(ri, sizeof(regexp_internal), char);
5893 /* non-zero initialization begins here */
5896 r->extflags = rx_flags;
5897 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
5899 if (pm_flags & PMf_IS_QR) {
5900 ri->code_blocks = pRExC_state->code_blocks;
5901 ri->num_code_blocks = pRExC_state->num_code_blocks;
5906 for (n = 0; n < pRExC_state->num_code_blocks; n++)
5907 if (pRExC_state->code_blocks[n].src_regex)
5908 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
5909 SAVEFREEPV(pRExC_state->code_blocks);
5913 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
5914 bool has_charset = (get_regex_charset(r->extflags) != REGEX_DEPENDS_CHARSET);
5916 /* The caret is output if there are any defaults: if not all the STD
5917 * flags are set, or if no character set specifier is needed */
5919 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
5921 bool has_runon = ((RExC_seen & REG_SEEN_RUN_ON_COMMENT)==REG_SEEN_RUN_ON_COMMENT);
5922 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
5923 >> RXf_PMf_STD_PMMOD_SHIFT);
5924 const char *fptr = STD_PAT_MODS; /*"msix"*/
5926 /* Allocate for the worst case, which is all the std flags are turned
5927 * on. If more precision is desired, we could do a population count of
5928 * the flags set. This could be done with a small lookup table, or by
5929 * shifting, masking and adding, or even, when available, assembly
5930 * language for a machine-language population count.
5931 * We never output a minus, as all those are defaults, so are
5932 * covered by the caret */
5933 const STRLEN wraplen = plen + has_p + has_runon
5934 + has_default /* If needs a caret */
5936 /* If needs a character set specifier */
5937 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
5938 + (sizeof(STD_PAT_MODS) - 1)
5939 + (sizeof("(?:)") - 1);
5941 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
5942 r->xpv_len_u.xpvlenu_pv = p;
5944 SvFLAGS(rx) |= SVf_UTF8;
5947 /* If a default, cover it using the caret */
5949 *p++= DEFAULT_PAT_MOD;
5953 const char* const name = get_regex_charset_name(r->extflags, &len);
5954 Copy(name, p, len, char);
5958 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
5961 while((ch = *fptr++)) {
5969 Copy(RExC_precomp, p, plen, char);
5970 assert ((RX_WRAPPED(rx) - p) < 16);
5971 r->pre_prefix = p - RX_WRAPPED(rx);
5977 SvCUR_set(rx, p - RX_WRAPPED(rx));
5981 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
5983 if (RExC_seen & REG_SEEN_RECURSE) {
5984 Newxz(RExC_open_parens, RExC_npar,regnode *);
5985 SAVEFREEPV(RExC_open_parens);
5986 Newxz(RExC_close_parens,RExC_npar,regnode *);
5987 SAVEFREEPV(RExC_close_parens);
5990 /* Useful during FAIL. */
5991 #ifdef RE_TRACK_PATTERN_OFFSETS
5992 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
5993 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
5994 "%s %"UVuf" bytes for offset annotations.\n",
5995 ri->u.offsets ? "Got" : "Couldn't get",
5996 (UV)((2*RExC_size+1) * sizeof(U32))));
5998 SetProgLen(ri,RExC_size);
6003 /* Second pass: emit code. */
6004 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
6005 RExC_pm_flags = pm_flags;
6007 RExC_end = exp + plen;
6010 RExC_emit_start = ri->program;
6011 RExC_emit = ri->program;
6012 RExC_emit_bound = ri->program + RExC_size + 1;
6013 pRExC_state->code_index = 0;
6015 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
6016 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6018 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#"UVxf"", (UV) flags);
6020 /* XXXX To minimize changes to RE engine we always allocate
6021 3-units-long substrs field. */
6022 Newx(r->substrs, 1, struct reg_substr_data);
6023 if (RExC_recurse_count) {
6024 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6025 SAVEFREEPV(RExC_recurse);
6029 r->minlen = minlen = sawlookahead = sawplus = sawopen = 0;
6030 Zero(r->substrs, 1, struct reg_substr_data);
6032 #ifdef TRIE_STUDY_OPT
6034 StructCopy(&zero_scan_data, &data, scan_data_t);
6035 copyRExC_state = RExC_state;
6038 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6040 RExC_state = copyRExC_state;
6041 if (seen & REG_TOP_LEVEL_BRANCHES)
6042 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
6044 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES;
6045 StructCopy(&zero_scan_data, &data, scan_data_t);
6048 StructCopy(&zero_scan_data, &data, scan_data_t);
6051 /* Dig out information for optimizations. */
6052 r->extflags = RExC_flags; /* was pm_op */
6053 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6056 SvUTF8_on(rx); /* Unicode in it? */
6057 ri->regstclass = NULL;
6058 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
6059 r->intflags |= PREGf_NAUGHTY;
6060 scan = ri->program + 1; /* First BRANCH. */
6062 /* testing for BRANCH here tells us whether there is "must appear"
6063 data in the pattern. If there is then we can use it for optimisations */
6064 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES)) { /* Only one top-level choice. */
6066 STRLEN longest_float_length, longest_fixed_length;
6067 struct regnode_charclass_class ch_class; /* pointed to by data */
6069 I32 last_close = 0; /* pointed to by data */
6070 regnode *first= scan;
6071 regnode *first_next= regnext(first);
6073 * Skip introductions and multiplicators >= 1
6074 * so that we can extract the 'meat' of the pattern that must
6075 * match in the large if() sequence following.
6076 * NOTE that EXACT is NOT covered here, as it is normally
6077 * picked up by the optimiser separately.
6079 * This is unfortunate as the optimiser isnt handling lookahead
6080 * properly currently.
6083 while ((OP(first) == OPEN && (sawopen = 1)) ||
6084 /* An OR of *one* alternative - should not happen now. */
6085 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6086 /* for now we can't handle lookbehind IFMATCH*/
6087 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6088 (OP(first) == PLUS) ||
6089 (OP(first) == MINMOD) ||
6090 /* An {n,m} with n>0 */
6091 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6092 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6095 * the only op that could be a regnode is PLUS, all the rest
6096 * will be regnode_1 or regnode_2.
6099 if (OP(first) == PLUS)
6102 first += regarglen[OP(first)];
6104 first = NEXTOPER(first);
6105 first_next= regnext(first);
6108 /* Starting-point info. */
6110 DEBUG_PEEP("first:",first,0);
6111 /* Ignore EXACT as we deal with it later. */
6112 if (PL_regkind[OP(first)] == EXACT) {
6113 if (OP(first) == EXACT)
6114 NOOP; /* Empty, get anchored substr later. */
6116 ri->regstclass = first;
6119 else if (PL_regkind[OP(first)] == TRIE &&
6120 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6123 /* this can happen only on restudy */
6124 if ( OP(first) == TRIE ) {
6125 struct regnode_1 *trieop = (struct regnode_1 *)
6126 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6127 StructCopy(first,trieop,struct regnode_1);
6128 trie_op=(regnode *)trieop;
6130 struct regnode_charclass *trieop = (struct regnode_charclass *)
6131 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6132 StructCopy(first,trieop,struct regnode_charclass);
6133 trie_op=(regnode *)trieop;
6136 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6137 ri->regstclass = trie_op;
6140 else if (REGNODE_SIMPLE(OP(first)))
6141 ri->regstclass = first;
6142 else if (PL_regkind[OP(first)] == BOUND ||
6143 PL_regkind[OP(first)] == NBOUND)
6144 ri->regstclass = first;
6145 else if (PL_regkind[OP(first)] == BOL) {
6146 r->extflags |= (OP(first) == MBOL
6148 : (OP(first) == SBOL
6151 first = NEXTOPER(first);
6154 else if (OP(first) == GPOS) {
6155 r->extflags |= RXf_ANCH_GPOS;
6156 first = NEXTOPER(first);
6159 else if ((!sawopen || !RExC_sawback) &&
6160 (OP(first) == STAR &&
6161 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6162 !(r->extflags & RXf_ANCH) && !pRExC_state->num_code_blocks)
6164 /* turn .* into ^.* with an implied $*=1 */
6166 (OP(NEXTOPER(first)) == REG_ANY)
6169 r->extflags |= type;
6170 r->intflags |= PREGf_IMPLICIT;
6171 first = NEXTOPER(first);
6174 if (sawplus && !sawlookahead && (!sawopen || !RExC_sawback)
6175 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6176 /* x+ must match at the 1st pos of run of x's */
6177 r->intflags |= PREGf_SKIP;
6179 /* Scan is after the zeroth branch, first is atomic matcher. */
6180 #ifdef TRIE_STUDY_OPT
6183 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6184 (IV)(first - scan + 1))
6188 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6189 (IV)(first - scan + 1))
6195 * If there's something expensive in the r.e., find the
6196 * longest literal string that must appear and make it the
6197 * regmust. Resolve ties in favor of later strings, since
6198 * the regstart check works with the beginning of the r.e.
6199 * and avoiding duplication strengthens checking. Not a
6200 * strong reason, but sufficient in the absence of others.
6201 * [Now we resolve ties in favor of the earlier string if
6202 * it happens that c_offset_min has been invalidated, since the
6203 * earlier string may buy us something the later one won't.]
6206 data.longest_fixed = newSVpvs("");
6207 data.longest_float = newSVpvs("");
6208 data.last_found = newSVpvs("");
6209 data.longest = &(data.longest_fixed);
6210 ENTER_with_name("study_chunk");
6211 SAVEFREESV(data.longest_fixed);
6212 SAVEFREESV(data.longest_float);
6213 SAVEFREESV(data.last_found);
6215 if (!ri->regstclass) {
6216 cl_init(pRExC_state, &ch_class);
6217 data.start_class = &ch_class;
6218 stclass_flag = SCF_DO_STCLASS_AND;
6219 } else /* XXXX Check for BOUND? */
6221 data.last_closep = &last_close;
6223 minlen = study_chunk(pRExC_state, &first, &minlen, &fake, scan + RExC_size, /* Up to end */
6224 &data, -1, NULL, NULL,
6225 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag,0);
6228 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6231 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6232 && data.last_start_min == 0 && data.last_end > 0
6233 && !RExC_seen_zerolen
6234 && !(RExC_seen & REG_SEEN_VERBARG)
6235 && (!(RExC_seen & REG_SEEN_GPOS) || (r->extflags & RXf_ANCH_GPOS)))
6236 r->extflags |= RXf_CHECK_ALL;
6237 scan_commit(pRExC_state, &data,&minlen,0);
6239 longest_float_length = CHR_SVLEN(data.longest_float);
6241 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6242 && data.offset_fixed == data.offset_float_min
6243 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6244 && S_setup_longest (aTHX_ pRExC_state,
6248 &(r->float_end_shift),
6249 data.lookbehind_float,
6250 data.offset_float_min,
6252 longest_float_length,
6253 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6254 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6256 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6257 r->float_max_offset = data.offset_float_max;
6258 if (data.offset_float_max < I32_MAX) /* Don't offset infinity */
6259 r->float_max_offset -= data.lookbehind_float;
6260 SvREFCNT_inc_simple_void_NN(data.longest_float);
6263 r->float_substr = r->float_utf8 = NULL;
6264 longest_float_length = 0;
6267 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6269 if (S_setup_longest (aTHX_ pRExC_state,
6271 &(r->anchored_utf8),
6272 &(r->anchored_substr),
6273 &(r->anchored_end_shift),
6274 data.lookbehind_fixed,
6277 longest_fixed_length,
6278 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6279 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6281 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6282 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6285 r->anchored_substr = r->anchored_utf8 = NULL;
6286 longest_fixed_length = 0;
6288 LEAVE_with_name("study_chunk");
6291 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6292 ri->regstclass = NULL;
6294 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6296 && ! TEST_SSC_EOS(data.start_class)
6297 && !cl_is_anything(data.start_class))
6299 const U32 n = add_data(pRExC_state, 1, "f");
6300 OP(data.start_class) = ANYOF_SYNTHETIC;
6302 Newx(RExC_rxi->data->data[n], 1,
6303 struct regnode_charclass_class);
6304 StructCopy(data.start_class,
6305 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6306 struct regnode_charclass_class);
6307 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6308 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6309 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6310 regprop(r, sv, (regnode*)data.start_class);
6311 PerlIO_printf(Perl_debug_log,
6312 "synthetic stclass \"%s\".\n",
6313 SvPVX_const(sv));});
6316 /* A temporary algorithm prefers floated substr to fixed one to dig more info. */
6317 if (longest_fixed_length > longest_float_length) {
6318 r->check_end_shift = r->anchored_end_shift;
6319 r->check_substr = r->anchored_substr;
6320 r->check_utf8 = r->anchored_utf8;
6321 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6322 if (r->extflags & RXf_ANCH_SINGLE)
6323 r->extflags |= RXf_NOSCAN;
6326 r->check_end_shift = r->float_end_shift;
6327 r->check_substr = r->float_substr;
6328 r->check_utf8 = r->float_utf8;
6329 r->check_offset_min = r->float_min_offset;
6330 r->check_offset_max = r->float_max_offset;
6332 /* XXXX Currently intuiting is not compatible with ANCH_GPOS.
6333 This should be changed ASAP! */
6334 if ((r->check_substr || r->check_utf8) && !(r->extflags & RXf_ANCH_GPOS)) {
6335 r->extflags |= RXf_USE_INTUIT;
6336 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6337 r->extflags |= RXf_INTUIT_TAIL;
6339 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6340 if ( (STRLEN)minlen < longest_float_length )
6341 minlen= longest_float_length;
6342 if ( (STRLEN)minlen < longest_fixed_length )
6343 minlen= longest_fixed_length;
6347 /* Several toplevels. Best we can is to set minlen. */
6349 struct regnode_charclass_class ch_class;
6352 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6354 scan = ri->program + 1;
6355 cl_init(pRExC_state, &ch_class);
6356 data.start_class = &ch_class;
6357 data.last_closep = &last_close;
6360 minlen = study_chunk(pRExC_state, &scan, &minlen, &fake, scan + RExC_size,
6361 &data, -1, NULL, NULL, SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS,0);
6363 CHECK_RESTUDY_GOTO_butfirst(NOOP);
6365 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
6366 = r->float_substr = r->float_utf8 = NULL;
6368 if (! TEST_SSC_EOS(data.start_class)
6369 && !cl_is_anything(data.start_class))
6371 const U32 n = add_data(pRExC_state, 1, "f");
6372 OP(data.start_class) = ANYOF_SYNTHETIC;
6374 Newx(RExC_rxi->data->data[n], 1,
6375 struct regnode_charclass_class);
6376 StructCopy(data.start_class,
6377 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6378 struct regnode_charclass_class);
6379 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6380 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6381 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
6382 regprop(r, sv, (regnode*)data.start_class);
6383 PerlIO_printf(Perl_debug_log,
6384 "synthetic stclass \"%s\".\n",
6385 SvPVX_const(sv));});
6389 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
6390 the "real" pattern. */
6392 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf"\n",
6393 (IV)minlen, (IV)r->minlen);
6395 r->minlenret = minlen;
6396 if (r->minlen < minlen)
6399 if (RExC_seen & REG_SEEN_GPOS)
6400 r->extflags |= RXf_GPOS_SEEN;
6401 if (RExC_seen & REG_SEEN_LOOKBEHIND)
6402 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the lookbehind */
6403 if (pRExC_state->num_code_blocks)
6404 r->extflags |= RXf_EVAL_SEEN;
6405 if (RExC_seen & REG_SEEN_CANY)
6406 r->extflags |= RXf_CANY_SEEN;
6407 if (RExC_seen & REG_SEEN_VERBARG)
6409 r->intflags |= PREGf_VERBARG_SEEN;
6410 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
6412 if (RExC_seen & REG_SEEN_CUTGROUP)
6413 r->intflags |= PREGf_CUTGROUP_SEEN;
6414 if (pm_flags & PMf_USE_RE_EVAL)
6415 r->intflags |= PREGf_USE_RE_EVAL;
6416 if (RExC_paren_names)
6417 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
6419 RXp_PAREN_NAMES(r) = NULL;
6422 regnode *first = ri->program + 1;
6424 regnode *next = NEXTOPER(first);
6427 if (PL_regkind[fop] == NOTHING && nop == END)
6428 r->extflags |= RXf_NULL;
6429 else if (PL_regkind[fop] == BOL && nop == END)
6430 r->extflags |= RXf_START_ONLY;
6431 else if (fop == PLUS && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE && OP(regnext(first)) == END)
6432 r->extflags |= RXf_WHITE;
6433 else if ( r->extflags & RXf_SPLIT && fop == EXACT && STR_LEN(first) == 1 && *(STRING(first)) == ' ' && OP(regnext(first)) == END )
6434 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
6438 if (RExC_paren_names) {
6439 ri->name_list_idx = add_data( pRExC_state, 1, "a" );
6440 ri->data->data[ri->name_list_idx] = (void*)SvREFCNT_inc(RExC_paren_name_list);
6443 ri->name_list_idx = 0;
6445 if (RExC_recurse_count) {
6446 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
6447 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
6448 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
6451 Newxz(r->offs, RExC_npar, regexp_paren_pair);
6452 /* assume we don't need to swap parens around before we match */
6455 PerlIO_printf(Perl_debug_log,"Final program:\n");
6458 #ifdef RE_TRACK_PATTERN_OFFSETS
6459 DEBUG_OFFSETS_r(if (ri->u.offsets) {
6460 const U32 len = ri->u.offsets[0];
6462 GET_RE_DEBUG_FLAGS_DECL;
6463 PerlIO_printf(Perl_debug_log, "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
6464 for (i = 1; i <= len; i++) {
6465 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
6466 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
6467 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
6469 PerlIO_printf(Perl_debug_log, "\n");
6474 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
6475 * by setting the regexp SV to readonly-only instead. If the
6476 * pattern's been recompiled, the USEDness should remain. */
6477 if (old_re && SvREADONLY(old_re))
6485 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
6488 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
6490 PERL_UNUSED_ARG(value);
6492 if (flags & RXapif_FETCH) {
6493 return reg_named_buff_fetch(rx, key, flags);
6494 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
6495 Perl_croak_no_modify();
6497 } else if (flags & RXapif_EXISTS) {
6498 return reg_named_buff_exists(rx, key, flags)
6501 } else if (flags & RXapif_REGNAMES) {
6502 return reg_named_buff_all(rx, flags);
6503 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
6504 return reg_named_buff_scalar(rx, flags);
6506 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
6512 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
6515 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
6516 PERL_UNUSED_ARG(lastkey);
6518 if (flags & RXapif_FIRSTKEY)
6519 return reg_named_buff_firstkey(rx, flags);
6520 else if (flags & RXapif_NEXTKEY)
6521 return reg_named_buff_nextkey(rx, flags);
6523 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter", (int)flags);
6529 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
6532 AV *retarray = NULL;
6534 struct regexp *const rx = ReANY(r);
6536 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
6538 if (flags & RXapif_ALL)
6541 if (rx && RXp_PAREN_NAMES(rx)) {
6542 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
6545 SV* sv_dat=HeVAL(he_str);
6546 I32 *nums=(I32*)SvPVX(sv_dat);
6547 for ( i=0; i<SvIVX(sv_dat); i++ ) {
6548 if ((I32)(rx->nparens) >= nums[i]
6549 && rx->offs[nums[i]].start != -1
6550 && rx->offs[nums[i]].end != -1)
6553 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
6558 ret = newSVsv(&PL_sv_undef);
6561 av_push(retarray, ret);
6564 return newRV_noinc(MUTABLE_SV(retarray));
6571 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
6574 struct regexp *const rx = ReANY(r);
6576 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
6578 if (rx && RXp_PAREN_NAMES(rx)) {
6579 if (flags & RXapif_ALL) {
6580 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
6582 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
6584 SvREFCNT_dec_NN(sv);
6596 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
6598 struct regexp *const rx = ReANY(r);
6600 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
6602 if ( rx && RXp_PAREN_NAMES(rx) ) {
6603 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
6605 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
6612 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
6614 struct regexp *const rx = ReANY(r);
6615 GET_RE_DEBUG_FLAGS_DECL;
6617 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
6619 if (rx && RXp_PAREN_NAMES(rx)) {
6620 HV *hv = RXp_PAREN_NAMES(rx);
6622 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6625 SV* sv_dat = HeVAL(temphe);
6626 I32 *nums = (I32*)SvPVX(sv_dat);
6627 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6628 if ((I32)(rx->lastparen) >= nums[i] &&
6629 rx->offs[nums[i]].start != -1 &&
6630 rx->offs[nums[i]].end != -1)
6636 if (parno || flags & RXapif_ALL) {
6637 return newSVhek(HeKEY_hek(temphe));
6645 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
6650 struct regexp *const rx = ReANY(r);
6652 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
6654 if (rx && RXp_PAREN_NAMES(rx)) {
6655 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
6656 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
6657 } else if (flags & RXapif_ONE) {
6658 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
6659 av = MUTABLE_AV(SvRV(ret));
6660 length = av_len(av);
6661 SvREFCNT_dec_NN(ret);
6662 return newSViv(length + 1);
6664 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar", (int)flags);
6668 return &PL_sv_undef;
6672 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
6674 struct regexp *const rx = ReANY(r);
6677 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
6679 if (rx && RXp_PAREN_NAMES(rx)) {
6680 HV *hv= RXp_PAREN_NAMES(rx);
6682 (void)hv_iterinit(hv);
6683 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6686 SV* sv_dat = HeVAL(temphe);
6687 I32 *nums = (I32*)SvPVX(sv_dat);
6688 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6689 if ((I32)(rx->lastparen) >= nums[i] &&
6690 rx->offs[nums[i]].start != -1 &&
6691 rx->offs[nums[i]].end != -1)
6697 if (parno || flags & RXapif_ALL) {
6698 av_push(av, newSVhek(HeKEY_hek(temphe)));
6703 return newRV_noinc(MUTABLE_SV(av));
6707 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
6710 struct regexp *const rx = ReANY(r);
6716 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
6718 if ( ( n == RX_BUFF_IDX_CARET_PREMATCH
6719 || n == RX_BUFF_IDX_CARET_FULLMATCH
6720 || n == RX_BUFF_IDX_CARET_POSTMATCH
6722 && !(rx->extflags & RXf_PMf_KEEPCOPY)
6729 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
6730 /* no need to distinguish between them any more */
6731 n = RX_BUFF_IDX_FULLMATCH;
6733 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
6734 && rx->offs[0].start != -1)
6736 /* $`, ${^PREMATCH} */
6737 i = rx->offs[0].start;
6741 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
6742 && rx->offs[0].end != -1)
6744 /* $', ${^POSTMATCH} */
6745 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
6746 i = rx->sublen + rx->suboffset - rx->offs[0].end;
6749 if ( 0 <= n && n <= (I32)rx->nparens &&
6750 (s1 = rx->offs[n].start) != -1 &&
6751 (t1 = rx->offs[n].end) != -1)
6753 /* $&, ${^MATCH}, $1 ... */
6755 s = rx->subbeg + s1 - rx->suboffset;
6760 assert(s >= rx->subbeg);
6761 assert(rx->sublen >= (s - rx->subbeg) + i );
6763 #if NO_TAINT_SUPPORT
6764 sv_setpvn(sv, s, i);
6766 const int oldtainted = TAINT_get;
6768 sv_setpvn(sv, s, i);
6769 TAINT_set(oldtainted);
6771 if ( (rx->extflags & RXf_CANY_SEEN)
6772 ? (RXp_MATCH_UTF8(rx)
6773 && (!i || is_utf8_string((U8*)s, i)))
6774 : (RXp_MATCH_UTF8(rx)) )
6781 if (RXp_MATCH_TAINTED(rx)) {
6782 if (SvTYPE(sv) >= SVt_PVMG) {
6783 MAGIC* const mg = SvMAGIC(sv);
6786 SvMAGIC_set(sv, mg->mg_moremagic);
6788 if ((mgt = SvMAGIC(sv))) {
6789 mg->mg_moremagic = mgt;
6790 SvMAGIC_set(sv, mg);
6801 sv_setsv(sv,&PL_sv_undef);
6807 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
6808 SV const * const value)
6810 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
6812 PERL_UNUSED_ARG(rx);
6813 PERL_UNUSED_ARG(paren);
6814 PERL_UNUSED_ARG(value);
6817 Perl_croak_no_modify();
6821 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
6824 struct regexp *const rx = ReANY(r);
6828 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
6830 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
6832 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
6833 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6837 case RX_BUFF_IDX_PREMATCH: /* $` */
6838 if (rx->offs[0].start != -1) {
6839 i = rx->offs[0].start;
6848 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
6849 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6851 case RX_BUFF_IDX_POSTMATCH: /* $' */
6852 if (rx->offs[0].end != -1) {
6853 i = rx->sublen - rx->offs[0].end;
6855 s1 = rx->offs[0].end;
6862 case RX_BUFF_IDX_CARET_FULLMATCH: /* ${^MATCH} */
6863 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6867 /* $& / ${^MATCH}, $1, $2, ... */
6869 if (paren <= (I32)rx->nparens &&
6870 (s1 = rx->offs[paren].start) != -1 &&
6871 (t1 = rx->offs[paren].end) != -1)
6877 if (ckWARN(WARN_UNINITIALIZED))
6878 report_uninit((const SV *)sv);
6883 if (i > 0 && RXp_MATCH_UTF8(rx)) {
6884 const char * const s = rx->subbeg - rx->suboffset + s1;
6889 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
6896 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
6898 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
6899 PERL_UNUSED_ARG(rx);
6903 return newSVpvs("Regexp");
6906 /* Scans the name of a named buffer from the pattern.
6907 * If flags is REG_RSN_RETURN_NULL returns null.
6908 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
6909 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
6910 * to the parsed name as looked up in the RExC_paren_names hash.
6911 * If there is an error throws a vFAIL().. type exception.
6914 #define REG_RSN_RETURN_NULL 0
6915 #define REG_RSN_RETURN_NAME 1
6916 #define REG_RSN_RETURN_DATA 2
6919 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
6921 char *name_start = RExC_parse;
6923 PERL_ARGS_ASSERT_REG_SCAN_NAME;
6925 if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
6926 /* skip IDFIRST by using do...while */
6929 RExC_parse += UTF8SKIP(RExC_parse);
6930 } while (isWORDCHAR_utf8((U8*)RExC_parse));
6934 } while (isWORDCHAR(*RExC_parse));
6936 RExC_parse++; /* so the <- from the vFAIL is after the offending character */
6937 vFAIL("Group name must start with a non-digit word character");
6941 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
6942 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
6943 if ( flags == REG_RSN_RETURN_NAME)
6945 else if (flags==REG_RSN_RETURN_DATA) {
6948 if ( ! sv_name ) /* should not happen*/
6949 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
6950 if (RExC_paren_names)
6951 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
6953 sv_dat = HeVAL(he_str);
6955 vFAIL("Reference to nonexistent named group");
6959 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
6960 (unsigned long) flags);
6962 assert(0); /* NOT REACHED */
6967 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
6968 int rem=(int)(RExC_end - RExC_parse); \
6977 if (RExC_lastparse!=RExC_parse) \
6978 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
6981 iscut ? "..." : "<" \
6984 PerlIO_printf(Perl_debug_log,"%16s",""); \
6987 num = RExC_size + 1; \
6989 num=REG_NODE_NUM(RExC_emit); \
6990 if (RExC_lastnum!=num) \
6991 PerlIO_printf(Perl_debug_log,"|%4d",num); \
6993 PerlIO_printf(Perl_debug_log,"|%4s",""); \
6994 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
6995 (int)((depth*2)), "", \
6999 RExC_lastparse=RExC_parse; \
7004 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
7005 DEBUG_PARSE_MSG((funcname)); \
7006 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
7008 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7009 DEBUG_PARSE_MSG((funcname)); \
7010 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7013 /* This section of code defines the inversion list object and its methods. The
7014 * interfaces are highly subject to change, so as much as possible is static to
7015 * this file. An inversion list is here implemented as a malloc'd C UV array
7016 * with some added info that is placed as UVs at the beginning in a header
7017 * portion. An inversion list for Unicode is an array of code points, sorted
7018 * by ordinal number. The zeroth element is the first code point in the list.
7019 * The 1th element is the first element beyond that not in the list. In other
7020 * words, the first range is
7021 * invlist[0]..(invlist[1]-1)
7022 * The other ranges follow. Thus every element whose index is divisible by two
7023 * marks the beginning of a range that is in the list, and every element not
7024 * divisible by two marks the beginning of a range not in the list. A single
7025 * element inversion list that contains the single code point N generally
7026 * consists of two elements
7029 * (The exception is when N is the highest representable value on the
7030 * machine, in which case the list containing just it would be a single
7031 * element, itself. By extension, if the last range in the list extends to
7032 * infinity, then the first element of that range will be in the inversion list
7033 * at a position that is divisible by two, and is the final element in the
7035 * Taking the complement (inverting) an inversion list is quite simple, if the
7036 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7037 * This implementation reserves an element at the beginning of each inversion
7038 * list to contain 0 when the list contains 0, and contains 1 otherwise. The
7039 * actual beginning of the list is either that element if 0, or the next one if
7042 * More about inversion lists can be found in "Unicode Demystified"
7043 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7044 * More will be coming when functionality is added later.
7046 * The inversion list data structure is currently implemented as an SV pointing
7047 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7048 * array of UV whose memory management is automatically handled by the existing
7049 * facilities for SV's.
7051 * Some of the methods should always be private to the implementation, and some
7052 * should eventually be made public */
7054 /* The header definitions are in F<inline_invlist.c> */
7055 #define TO_INTERNAL_SIZE(x) (((x) + HEADER_LENGTH) * sizeof(UV))
7056 #define FROM_INTERNAL_SIZE(x) (((x)/ sizeof(UV)) - HEADER_LENGTH)
7058 #define INVLIST_INITIAL_LEN 10
7060 PERL_STATIC_INLINE UV*
7061 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
7063 /* Returns a pointer to the first element in the inversion list's array.
7064 * This is called upon initialization of an inversion list. Where the
7065 * array begins depends on whether the list has the code point U+0000
7066 * in it or not. The other parameter tells it whether the code that
7067 * follows this call is about to put a 0 in the inversion list or not.
7068 * The first element is either the element with 0, if 0, or the next one,
7071 UV* zero = get_invlist_zero_addr(invlist);
7073 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7076 assert(! *_get_invlist_len_addr(invlist));
7078 /* 1^1 = 0; 1^0 = 1 */
7079 *zero = 1 ^ will_have_0;
7080 return zero + *zero;
7083 PERL_STATIC_INLINE UV*
7084 S_invlist_array(pTHX_ SV* const invlist)
7086 /* Returns the pointer to the inversion list's array. Every time the
7087 * length changes, this needs to be called in case malloc or realloc moved
7090 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7092 /* Must not be empty. If these fail, you probably didn't check for <len>
7093 * being non-zero before trying to get the array */
7094 assert(*_get_invlist_len_addr(invlist));
7095 assert(*get_invlist_zero_addr(invlist) == 0
7096 || *get_invlist_zero_addr(invlist) == 1);
7098 /* The array begins either at the element reserved for zero if the
7099 * list contains 0 (that element will be set to 0), or otherwise the next
7100 * element (in which case the reserved element will be set to 1). */
7101 return (UV *) (get_invlist_zero_addr(invlist)
7102 + *get_invlist_zero_addr(invlist));
7105 PERL_STATIC_INLINE void
7106 S_invlist_set_len(pTHX_ SV* const invlist, const UV len)
7108 /* Sets the current number of elements stored in the inversion list */
7110 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7112 *_get_invlist_len_addr(invlist) = len;
7114 assert(len <= SvLEN(invlist));
7116 SvCUR_set(invlist, TO_INTERNAL_SIZE(len));
7117 /* If the list contains U+0000, that element is part of the header,
7118 * and should not be counted as part of the array. It will contain
7119 * 0 in that case, and 1 otherwise. So we could flop 0=>1, 1=>0 and
7121 * SvCUR_set(invlist,
7122 * TO_INTERNAL_SIZE(len
7123 * - (*get_invlist_zero_addr(inv_list) ^ 1)));
7124 * But, this is only valid if len is not 0. The consequences of not doing
7125 * this is that the memory allocation code may think that 1 more UV is
7126 * being used than actually is, and so might do an unnecessary grow. That
7127 * seems worth not bothering to make this the precise amount.
7129 * Note that when inverting, SvCUR shouldn't change */
7132 PERL_STATIC_INLINE IV*
7133 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7135 /* Return the address of the UV that is reserved to hold the cached index
7138 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7140 return (IV *) (SvPVX(invlist) + (INVLIST_PREVIOUS_INDEX_OFFSET * sizeof (UV)));
7143 PERL_STATIC_INLINE IV
7144 S_invlist_previous_index(pTHX_ SV* const invlist)
7146 /* Returns cached index of previous search */
7148 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7150 return *get_invlist_previous_index_addr(invlist);
7153 PERL_STATIC_INLINE void
7154 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7156 /* Caches <index> for later retrieval */
7158 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7160 assert(index == 0 || index < (int) _invlist_len(invlist));
7162 *get_invlist_previous_index_addr(invlist) = index;
7165 PERL_STATIC_INLINE UV
7166 S_invlist_max(pTHX_ SV* const invlist)
7168 /* Returns the maximum number of elements storable in the inversion list's
7169 * array, without having to realloc() */
7171 PERL_ARGS_ASSERT_INVLIST_MAX;
7173 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7174 ? _invlist_len(invlist)
7175 : FROM_INTERNAL_SIZE(SvLEN(invlist));
7178 PERL_STATIC_INLINE UV*
7179 S_get_invlist_zero_addr(pTHX_ SV* invlist)
7181 /* Return the address of the UV that is reserved to hold 0 if the inversion
7182 * list contains 0. This has to be the last element of the heading, as the
7183 * list proper starts with either it if 0, or the next element if not.
7184 * (But we force it to contain either 0 or 1) */
7186 PERL_ARGS_ASSERT_GET_INVLIST_ZERO_ADDR;
7188 return (UV *) (SvPVX(invlist) + (INVLIST_ZERO_OFFSET * sizeof (UV)));
7191 #ifndef PERL_IN_XSUB_RE
7193 Perl__new_invlist(pTHX_ IV initial_size)
7196 /* Return a pointer to a newly constructed inversion list, with enough
7197 * space to store 'initial_size' elements. If that number is negative, a
7198 * system default is used instead */
7202 if (initial_size < 0) {
7203 initial_size = INVLIST_INITIAL_LEN;
7206 /* Allocate the initial space */
7207 new_list = newSV(TO_INTERNAL_SIZE(initial_size));
7208 invlist_set_len(new_list, 0);
7210 /* Force iterinit() to be used to get iteration to work */
7211 *get_invlist_iter_addr(new_list) = UV_MAX;
7213 /* This should force a segfault if a method doesn't initialize this
7215 *get_invlist_zero_addr(new_list) = UV_MAX;
7217 *get_invlist_previous_index_addr(new_list) = 0;
7218 *get_invlist_version_id_addr(new_list) = INVLIST_VERSION_ID;
7219 #if HEADER_LENGTH != 5
7220 # 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
7228 S__new_invlist_C_array(pTHX_ UV* list)
7230 /* Return a pointer to a newly constructed inversion list, initialized to
7231 * point to <list>, which has to be in the exact correct inversion list
7232 * form, including internal fields. Thus this is a dangerous routine that
7233 * should not be used in the wrong hands */
7235 SV* invlist = newSV_type(SVt_PV);
7237 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7239 SvPV_set(invlist, (char *) list);
7240 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7241 shouldn't touch it */
7242 SvCUR_set(invlist, TO_INTERNAL_SIZE(_invlist_len(invlist)));
7244 if (*get_invlist_version_id_addr(invlist) != INVLIST_VERSION_ID) {
7245 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7248 /* Initialize the iteration pointer.
7249 * XXX This could be done at compile time in charclass_invlists.h, but I
7250 * (khw) am not confident that the suffixes for specifying the C constant
7251 * UV_MAX are portable, e.g. 'ull' on a 32 bit machine that is configured
7252 * to use 64 bits; might need a Configure probe */
7253 invlist_iterfinish(invlist);
7259 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7261 /* Grow the maximum size of an inversion list */
7263 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7265 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max));
7268 PERL_STATIC_INLINE void
7269 S_invlist_trim(pTHX_ SV* const invlist)
7271 PERL_ARGS_ASSERT_INVLIST_TRIM;
7273 /* Change the length of the inversion list to how many entries it currently
7276 SvPV_shrink_to_cur((SV *) invlist);
7279 #define _invlist_union_complement_2nd(a, b, output) _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
7282 S__append_range_to_invlist(pTHX_ SV* const invlist, const UV start, const UV end)
7284 /* Subject to change or removal. Append the range from 'start' to 'end' at
7285 * the end of the inversion list. The range must be above any existing
7289 UV max = invlist_max(invlist);
7290 UV len = _invlist_len(invlist);
7292 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7294 if (len == 0) { /* Empty lists must be initialized */
7295 array = _invlist_array_init(invlist, start == 0);
7298 /* Here, the existing list is non-empty. The current max entry in the
7299 * list is generally the first value not in the set, except when the
7300 * set extends to the end of permissible values, in which case it is
7301 * the first entry in that final set, and so this call is an attempt to
7302 * append out-of-order */
7304 UV final_element = len - 1;
7305 array = invlist_array(invlist);
7306 if (array[final_element] > start
7307 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
7309 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",
7310 array[final_element], start,
7311 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
7314 /* Here, it is a legal append. If the new range begins with the first
7315 * value not in the set, it is extending the set, so the new first
7316 * value not in the set is one greater than the newly extended range.
7318 if (array[final_element] == start) {
7319 if (end != UV_MAX) {
7320 array[final_element] = end + 1;
7323 /* But if the end is the maximum representable on the machine,
7324 * just let the range that this would extend to have no end */
7325 invlist_set_len(invlist, len - 1);
7331 /* Here the new range doesn't extend any existing set. Add it */
7333 len += 2; /* Includes an element each for the start and end of range */
7335 /* If overflows the existing space, extend, which may cause the array to be
7338 invlist_extend(invlist, len);
7339 invlist_set_len(invlist, len); /* Have to set len here to avoid assert
7340 failure in invlist_array() */
7341 array = invlist_array(invlist);
7344 invlist_set_len(invlist, len);
7347 /* The next item on the list starts the range, the one after that is
7348 * one past the new range. */
7349 array[len - 2] = start;
7350 if (end != UV_MAX) {
7351 array[len - 1] = end + 1;
7354 /* But if the end is the maximum representable on the machine, just let
7355 * the range have no end */
7356 invlist_set_len(invlist, len - 1);
7360 #ifndef PERL_IN_XSUB_RE
7363 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
7365 /* Searches the inversion list for the entry that contains the input code
7366 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
7367 * return value is the index into the list's array of the range that
7372 IV high = _invlist_len(invlist);
7373 const IV highest_element = high - 1;
7376 PERL_ARGS_ASSERT__INVLIST_SEARCH;
7378 /* If list is empty, return failure. */
7383 /* (We can't get the array unless we know the list is non-empty) */
7384 array = invlist_array(invlist);
7386 mid = invlist_previous_index(invlist);
7387 assert(mid >=0 && mid <= highest_element);
7389 /* <mid> contains the cache of the result of the previous call to this
7390 * function (0 the first time). See if this call is for the same result,
7391 * or if it is for mid-1. This is under the theory that calls to this
7392 * function will often be for related code points that are near each other.
7393 * And benchmarks show that caching gives better results. We also test
7394 * here if the code point is within the bounds of the list. These tests
7395 * replace others that would have had to be made anyway to make sure that
7396 * the array bounds were not exceeded, and these give us extra information
7397 * at the same time */
7398 if (cp >= array[mid]) {
7399 if (cp >= array[highest_element]) {
7400 return highest_element;
7403 /* Here, array[mid] <= cp < array[highest_element]. This means that
7404 * the final element is not the answer, so can exclude it; it also
7405 * means that <mid> is not the final element, so can refer to 'mid + 1'
7407 if (cp < array[mid + 1]) {
7413 else { /* cp < aray[mid] */
7414 if (cp < array[0]) { /* Fail if outside the array */
7418 if (cp >= array[mid - 1]) {
7423 /* Binary search. What we are looking for is <i> such that
7424 * array[i] <= cp < array[i+1]
7425 * The loop below converges on the i+1. Note that there may not be an
7426 * (i+1)th element in the array, and things work nonetheless */
7427 while (low < high) {
7428 mid = (low + high) / 2;
7429 assert(mid <= highest_element);
7430 if (array[mid] <= cp) { /* cp >= array[mid] */
7433 /* We could do this extra test to exit the loop early.
7434 if (cp < array[low]) {
7439 else { /* cp < array[mid] */
7446 invlist_set_previous_index(invlist, high);
7451 Perl__invlist_populate_swatch(pTHX_ SV* const invlist, const UV start, const UV end, U8* swatch)
7453 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
7454 * but is used when the swash has an inversion list. This makes this much
7455 * faster, as it uses a binary search instead of a linear one. This is
7456 * intimately tied to that function, and perhaps should be in utf8.c,
7457 * except it is intimately tied to inversion lists as well. It assumes
7458 * that <swatch> is all 0's on input */
7461 const IV len = _invlist_len(invlist);
7465 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
7467 if (len == 0) { /* Empty inversion list */
7471 array = invlist_array(invlist);
7473 /* Find which element it is */
7474 i = _invlist_search(invlist, start);
7476 /* We populate from <start> to <end> */
7477 while (current < end) {
7480 /* The inversion list gives the results for every possible code point
7481 * after the first one in the list. Only those ranges whose index is
7482 * even are ones that the inversion list matches. For the odd ones,
7483 * and if the initial code point is not in the list, we have to skip
7484 * forward to the next element */
7485 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
7487 if (i >= len) { /* Finished if beyond the end of the array */
7491 if (current >= end) { /* Finished if beyond the end of what we
7493 if (LIKELY(end < UV_MAX)) {
7497 /* We get here when the upper bound is the maximum
7498 * representable on the machine, and we are looking for just
7499 * that code point. Have to special case it */
7501 goto join_end_of_list;
7504 assert(current >= start);
7506 /* The current range ends one below the next one, except don't go past
7509 upper = (i < len && array[i] < end) ? array[i] : end;
7511 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
7512 * for each code point in it */
7513 for (; current < upper; current++) {
7514 const STRLEN offset = (STRLEN)(current - start);
7515 swatch[offset >> 3] |= 1 << (offset & 7);
7520 /* Quit if at the end of the list */
7523 /* But first, have to deal with the highest possible code point on
7524 * the platform. The previous code assumes that <end> is one
7525 * beyond where we want to populate, but that is impossible at the
7526 * platform's infinity, so have to handle it specially */
7527 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
7529 const STRLEN offset = (STRLEN)(end - start);
7530 swatch[offset >> 3] |= 1 << (offset & 7);
7535 /* Advance to the next range, which will be for code points not in the
7544 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** output)
7546 /* Take the union of two inversion lists and point <output> to it. *output
7547 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7548 * the reference count to that list will be decremented. The first list,
7549 * <a>, may be NULL, in which case a copy of the second list is returned.
7550 * If <complement_b> is TRUE, the union is taken of the complement
7551 * (inversion) of <b> instead of b itself.
7553 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7554 * Richard Gillam, published by Addison-Wesley, and explained at some
7555 * length there. The preface says to incorporate its examples into your
7556 * code at your own risk.
7558 * The algorithm is like a merge sort.
7560 * XXX A potential performance improvement is to keep track as we go along
7561 * if only one of the inputs contributes to the result, meaning the other
7562 * is a subset of that one. In that case, we can skip the final copy and
7563 * return the larger of the input lists, but then outside code might need
7564 * to keep track of whether to free the input list or not */
7566 UV* array_a; /* a's array */
7568 UV len_a; /* length of a's array */
7571 SV* u; /* the resulting union */
7575 UV i_a = 0; /* current index into a's array */
7579 /* running count, as explained in the algorithm source book; items are
7580 * stopped accumulating and are output when the count changes to/from 0.
7581 * The count is incremented when we start a range that's in the set, and
7582 * decremented when we start a range that's not in the set. So its range
7583 * is 0 to 2. Only when the count is zero is something not in the set.
7587 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
7590 /* If either one is empty, the union is the other one */
7591 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
7598 *output = invlist_clone(b);
7600 _invlist_invert(*output);
7602 } /* else *output already = b; */
7605 else if ((len_b = _invlist_len(b)) == 0) {
7610 /* The complement of an empty list is a list that has everything in it,
7611 * so the union with <a> includes everything too */
7616 *output = _new_invlist(1);
7617 _append_range_to_invlist(*output, 0, UV_MAX);
7619 else if (*output != a) {
7620 *output = invlist_clone(a);
7622 /* else *output already = a; */
7626 /* Here both lists exist and are non-empty */
7627 array_a = invlist_array(a);
7628 array_b = invlist_array(b);
7630 /* If are to take the union of 'a' with the complement of b, set it
7631 * up so are looking at b's complement. */
7634 /* To complement, we invert: if the first element is 0, remove it. To
7635 * do this, we just pretend the array starts one later, and clear the
7636 * flag as we don't have to do anything else later */
7637 if (array_b[0] == 0) {
7640 complement_b = FALSE;
7644 /* But if the first element is not zero, we unshift a 0 before the
7645 * array. The data structure reserves a space for that 0 (which
7646 * should be a '1' right now), so physical shifting is unneeded,
7647 * but temporarily change that element to 0. Before exiting the
7648 * routine, we must restore the element to '1' */
7655 /* Size the union for the worst case: that the sets are completely
7657 u = _new_invlist(len_a + len_b);
7659 /* Will contain U+0000 if either component does */
7660 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
7661 || (len_b > 0 && array_b[0] == 0));
7663 /* Go through each list item by item, stopping when exhausted one of
7665 while (i_a < len_a && i_b < len_b) {
7666 UV cp; /* The element to potentially add to the union's array */
7667 bool cp_in_set; /* is it in the the input list's set or not */
7669 /* We need to take one or the other of the two inputs for the union.
7670 * Since we are merging two sorted lists, we take the smaller of the
7671 * next items. In case of a tie, we take the one that is in its set
7672 * first. If we took one not in the set first, it would decrement the
7673 * count, possibly to 0 which would cause it to be output as ending the
7674 * range, and the next time through we would take the same number, and
7675 * output it again as beginning the next range. By doing it the
7676 * opposite way, there is no possibility that the count will be
7677 * momentarily decremented to 0, and thus the two adjoining ranges will
7678 * be seamlessly merged. (In a tie and both are in the set or both not
7679 * in the set, it doesn't matter which we take first.) */
7680 if (array_a[i_a] < array_b[i_b]
7681 || (array_a[i_a] == array_b[i_b]
7682 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7684 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7688 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7689 cp = array_b[i_b++];
7692 /* Here, have chosen which of the two inputs to look at. Only output
7693 * if the running count changes to/from 0, which marks the
7694 * beginning/end of a range in that's in the set */
7697 array_u[i_u++] = cp;
7704 array_u[i_u++] = cp;
7709 /* Here, we are finished going through at least one of the lists, which
7710 * means there is something remaining in at most one. We check if the list
7711 * that hasn't been exhausted is positioned such that we are in the middle
7712 * of a range in its set or not. (i_a and i_b point to the element beyond
7713 * the one we care about.) If in the set, we decrement 'count'; if 0, there
7714 * is potentially more to output.
7715 * There are four cases:
7716 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
7717 * in the union is entirely from the non-exhausted set.
7718 * 2) Both were in their sets, count is 2. Nothing further should
7719 * be output, as everything that remains will be in the exhausted
7720 * list's set, hence in the union; decrementing to 1 but not 0 insures
7722 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
7723 * Nothing further should be output because the union includes
7724 * everything from the exhausted set. Not decrementing ensures that.
7725 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
7726 * decrementing to 0 insures that we look at the remainder of the
7727 * non-exhausted set */
7728 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7729 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7734 /* The final length is what we've output so far, plus what else is about to
7735 * be output. (If 'count' is non-zero, then the input list we exhausted
7736 * has everything remaining up to the machine's limit in its set, and hence
7737 * in the union, so there will be no further output. */
7740 /* At most one of the subexpressions will be non-zero */
7741 len_u += (len_a - i_a) + (len_b - i_b);
7744 /* Set result to final length, which can change the pointer to array_u, so
7746 if (len_u != _invlist_len(u)) {
7747 invlist_set_len(u, len_u);
7749 array_u = invlist_array(u);
7752 /* When 'count' is 0, the list that was exhausted (if one was shorter than
7753 * the other) ended with everything above it not in its set. That means
7754 * that the remaining part of the union is precisely the same as the
7755 * non-exhausted list, so can just copy it unchanged. (If both list were
7756 * exhausted at the same time, then the operations below will be both 0.)
7759 IV copy_count; /* At most one will have a non-zero copy count */
7760 if ((copy_count = len_a - i_a) > 0) {
7761 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
7763 else if ((copy_count = len_b - i_b) > 0) {
7764 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
7768 /* If we've changed b, restore it */
7773 /* We may be removing a reference to one of the inputs */
7774 if (a == *output || b == *output) {
7775 assert(! invlist_is_iterating(*output));
7776 SvREFCNT_dec_NN(*output);
7784 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** i)
7786 /* Take the intersection of two inversion lists and point <i> to it. *i
7787 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7788 * the reference count to that list will be decremented.
7789 * If <complement_b> is TRUE, the result will be the intersection of <a>
7790 * and the complement (or inversion) of <b> instead of <b> directly.
7792 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7793 * Richard Gillam, published by Addison-Wesley, and explained at some
7794 * length there. The preface says to incorporate its examples into your
7795 * code at your own risk. In fact, it had bugs
7797 * The algorithm is like a merge sort, and is essentially the same as the
7801 UV* array_a; /* a's array */
7803 UV len_a; /* length of a's array */
7806 SV* r; /* the resulting intersection */
7810 UV i_a = 0; /* current index into a's array */
7814 /* running count, as explained in the algorithm source book; items are
7815 * stopped accumulating and are output when the count changes to/from 2.
7816 * The count is incremented when we start a range that's in the set, and
7817 * decremented when we start a range that's not in the set. So its range
7818 * is 0 to 2. Only when the count is 2 is something in the intersection.
7822 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
7825 /* Special case if either one is empty */
7826 len_a = _invlist_len(a);
7827 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
7829 if (len_a != 0 && complement_b) {
7831 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
7832 * be empty. Here, also we are using 'b's complement, which hence
7833 * must be every possible code point. Thus the intersection is
7836 *i = invlist_clone(a);
7842 /* else *i is already 'a' */
7846 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
7847 * intersection must be empty */
7854 *i = _new_invlist(0);
7858 /* Here both lists exist and are non-empty */
7859 array_a = invlist_array(a);
7860 array_b = invlist_array(b);
7862 /* If are to take the intersection of 'a' with the complement of b, set it
7863 * up so are looking at b's complement. */
7866 /* To complement, we invert: if the first element is 0, remove it. To
7867 * do this, we just pretend the array starts one later, and clear the
7868 * flag as we don't have to do anything else later */
7869 if (array_b[0] == 0) {
7872 complement_b = FALSE;
7876 /* But if the first element is not zero, we unshift a 0 before the
7877 * array. The data structure reserves a space for that 0 (which
7878 * should be a '1' right now), so physical shifting is unneeded,
7879 * but temporarily change that element to 0. Before exiting the
7880 * routine, we must restore the element to '1' */
7887 /* Size the intersection for the worst case: that the intersection ends up
7888 * fragmenting everything to be completely disjoint */
7889 r= _new_invlist(len_a + len_b);
7891 /* Will contain U+0000 iff both components do */
7892 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
7893 && len_b > 0 && array_b[0] == 0);
7895 /* Go through each list item by item, stopping when exhausted one of
7897 while (i_a < len_a && i_b < len_b) {
7898 UV cp; /* The element to potentially add to the intersection's
7900 bool cp_in_set; /* Is it in the input list's set or not */
7902 /* We need to take one or the other of the two inputs for the
7903 * intersection. Since we are merging two sorted lists, we take the
7904 * smaller of the next items. In case of a tie, we take the one that
7905 * is not in its set first (a difference from the union algorithm). If
7906 * we took one in the set first, it would increment the count, possibly
7907 * to 2 which would cause it to be output as starting a range in the
7908 * intersection, and the next time through we would take that same
7909 * number, and output it again as ending the set. By doing it the
7910 * opposite of this, there is no possibility that the count will be
7911 * momentarily incremented to 2. (In a tie and both are in the set or
7912 * both not in the set, it doesn't matter which we take first.) */
7913 if (array_a[i_a] < array_b[i_b]
7914 || (array_a[i_a] == array_b[i_b]
7915 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7917 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7921 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7925 /* Here, have chosen which of the two inputs to look at. Only output
7926 * if the running count changes to/from 2, which marks the
7927 * beginning/end of a range that's in the intersection */
7931 array_r[i_r++] = cp;
7936 array_r[i_r++] = cp;
7942 /* Here, we are finished going through at least one of the lists, which
7943 * means there is something remaining in at most one. We check if the list
7944 * that has been exhausted is positioned such that we are in the middle
7945 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
7946 * the ones we care about.) There are four cases:
7947 * 1) Both weren't in their sets, count is 0, and remains 0. There's
7948 * nothing left in the intersection.
7949 * 2) Both were in their sets, count is 2 and perhaps is incremented to
7950 * above 2. What should be output is exactly that which is in the
7951 * non-exhausted set, as everything it has is also in the intersection
7952 * set, and everything it doesn't have can't be in the intersection
7953 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
7954 * gets incremented to 2. Like the previous case, the intersection is
7955 * everything that remains in the non-exhausted set.
7956 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
7957 * remains 1. And the intersection has nothing more. */
7958 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7959 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7964 /* The final length is what we've output so far plus what else is in the
7965 * intersection. At most one of the subexpressions below will be non-zero */
7968 len_r += (len_a - i_a) + (len_b - i_b);
7971 /* Set result to final length, which can change the pointer to array_r, so
7973 if (len_r != _invlist_len(r)) {
7974 invlist_set_len(r, len_r);
7976 array_r = invlist_array(r);
7979 /* Finish outputting any remaining */
7980 if (count >= 2) { /* At most one will have a non-zero copy count */
7982 if ((copy_count = len_a - i_a) > 0) {
7983 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
7985 else if ((copy_count = len_b - i_b) > 0) {
7986 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
7990 /* If we've changed b, restore it */
7995 /* We may be removing a reference to one of the inputs */
7996 if (a == *i || b == *i) {
7997 assert(! invlist_is_iterating(*i));
7998 SvREFCNT_dec_NN(*i);
8006 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
8008 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8009 * set. A pointer to the inversion list is returned. This may actually be
8010 * a new list, in which case the passed in one has been destroyed. The
8011 * passed in inversion list can be NULL, in which case a new one is created
8012 * with just the one range in it */
8017 if (invlist == NULL) {
8018 invlist = _new_invlist(2);
8022 len = _invlist_len(invlist);
8025 /* If comes after the final entry actually in the list, can just append it
8028 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
8029 && start >= invlist_array(invlist)[len - 1]))
8031 _append_range_to_invlist(invlist, start, end);
8035 /* Here, can't just append things, create and return a new inversion list
8036 * which is the union of this range and the existing inversion list */
8037 range_invlist = _new_invlist(2);
8038 _append_range_to_invlist(range_invlist, start, end);
8040 _invlist_union(invlist, range_invlist, &invlist);
8042 /* The temporary can be freed */
8043 SvREFCNT_dec_NN(range_invlist);
8050 PERL_STATIC_INLINE SV*
8051 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
8052 return _add_range_to_invlist(invlist, cp, cp);
8055 #ifndef PERL_IN_XSUB_RE
8057 Perl__invlist_invert(pTHX_ SV* const invlist)
8059 /* Complement the input inversion list. This adds a 0 if the list didn't
8060 * have a zero; removes it otherwise. As described above, the data
8061 * structure is set up so that this is very efficient */
8063 UV* len_pos = _get_invlist_len_addr(invlist);
8065 PERL_ARGS_ASSERT__INVLIST_INVERT;
8067 assert(! invlist_is_iterating(invlist));
8069 /* The inverse of matching nothing is matching everything */
8070 if (*len_pos == 0) {
8071 _append_range_to_invlist(invlist, 0, UV_MAX);
8075 /* The exclusive or complents 0 to 1; and 1 to 0. If the result is 1, the
8076 * zero element was a 0, so it is being removed, so the length decrements
8077 * by 1; and vice-versa. SvCUR is unaffected */
8078 if (*get_invlist_zero_addr(invlist) ^= 1) {
8087 Perl__invlist_invert_prop(pTHX_ SV* const invlist)
8089 /* Complement the input inversion list (which must be a Unicode property,
8090 * all of which don't match above the Unicode maximum code point.) And
8091 * Perl has chosen to not have the inversion match above that either. This
8092 * adds a 0x110000 if the list didn't end with it, and removes it if it did
8098 PERL_ARGS_ASSERT__INVLIST_INVERT_PROP;
8100 _invlist_invert(invlist);
8102 len = _invlist_len(invlist);
8104 if (len != 0) { /* If empty do nothing */
8105 array = invlist_array(invlist);
8106 if (array[len - 1] != PERL_UNICODE_MAX + 1) {
8107 /* Add 0x110000. First, grow if necessary */
8109 if (invlist_max(invlist) < len) {
8110 invlist_extend(invlist, len);
8111 array = invlist_array(invlist);
8113 invlist_set_len(invlist, len);
8114 array[len - 1] = PERL_UNICODE_MAX + 1;
8116 else { /* Remove the 0x110000 */
8117 invlist_set_len(invlist, len - 1);
8125 PERL_STATIC_INLINE SV*
8126 S_invlist_clone(pTHX_ SV* const invlist)
8129 /* Return a new inversion list that is a copy of the input one, which is
8132 /* Need to allocate extra space to accommodate Perl's addition of a
8133 * trailing NUL to SvPV's, since it thinks they are always strings */
8134 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8135 STRLEN length = SvCUR(invlist);
8137 PERL_ARGS_ASSERT_INVLIST_CLONE;
8139 SvCUR_set(new_invlist, length); /* This isn't done automatically */
8140 Copy(SvPVX(invlist), SvPVX(new_invlist), length, char);
8145 PERL_STATIC_INLINE UV*
8146 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8148 /* Return the address of the UV that contains the current iteration
8151 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8153 return (UV *) (SvPVX(invlist) + (INVLIST_ITER_OFFSET * sizeof (UV)));
8156 PERL_STATIC_INLINE UV*
8157 S_get_invlist_version_id_addr(pTHX_ SV* invlist)
8159 /* Return the address of the UV that contains the version id. */
8161 PERL_ARGS_ASSERT_GET_INVLIST_VERSION_ID_ADDR;
8163 return (UV *) (SvPVX(invlist) + (INVLIST_VERSION_ID_OFFSET * sizeof (UV)));
8166 PERL_STATIC_INLINE void
8167 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8169 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8171 *get_invlist_iter_addr(invlist) = 0;
8174 PERL_STATIC_INLINE void
8175 S_invlist_iterfinish(pTHX_ SV* invlist)
8177 /* Terminate iterator for invlist. This is to catch development errors.
8178 * Any iteration that is interrupted before completed should call this
8179 * function. Functions that add code points anywhere else but to the end
8180 * of an inversion list assert that they are not in the middle of an
8181 * iteration. If they were, the addition would make the iteration
8182 * problematical: if the iteration hadn't reached the place where things
8183 * were being added, it would be ok */
8185 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8187 *get_invlist_iter_addr(invlist) = UV_MAX;
8191 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8193 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8194 * This call sets in <*start> and <*end>, the next range in <invlist>.
8195 * Returns <TRUE> if successful and the next call will return the next
8196 * range; <FALSE> if was already at the end of the list. If the latter,
8197 * <*start> and <*end> are unchanged, and the next call to this function
8198 * will start over at the beginning of the list */
8200 UV* pos = get_invlist_iter_addr(invlist);
8201 UV len = _invlist_len(invlist);
8204 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8207 *pos = UV_MAX; /* Force iterinit() to be required next time */
8211 array = invlist_array(invlist);
8213 *start = array[(*pos)++];
8219 *end = array[(*pos)++] - 1;
8225 PERL_STATIC_INLINE bool
8226 S_invlist_is_iterating(pTHX_ SV* const invlist)
8228 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8230 return *(get_invlist_iter_addr(invlist)) < UV_MAX;
8233 PERL_STATIC_INLINE UV
8234 S_invlist_highest(pTHX_ SV* const invlist)
8236 /* Returns the highest code point that matches an inversion list. This API
8237 * has an ambiguity, as it returns 0 under either the highest is actually
8238 * 0, or if the list is empty. If this distinction matters to you, check
8239 * for emptiness before calling this function */
8241 UV len = _invlist_len(invlist);
8244 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8250 array = invlist_array(invlist);
8252 /* The last element in the array in the inversion list always starts a
8253 * range that goes to infinity. That range may be for code points that are
8254 * matched in the inversion list, or it may be for ones that aren't
8255 * matched. In the latter case, the highest code point in the set is one
8256 * less than the beginning of this range; otherwise it is the final element
8257 * of this range: infinity */
8258 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8260 : array[len - 1] - 1;
8263 #ifndef PERL_IN_XSUB_RE
8265 Perl__invlist_contents(pTHX_ SV* const invlist)
8267 /* Get the contents of an inversion list into a string SV so that they can
8268 * be printed out. It uses the format traditionally done for debug tracing
8272 SV* output = newSVpvs("\n");
8274 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8276 assert(! invlist_is_iterating(invlist));
8278 invlist_iterinit(invlist);
8279 while (invlist_iternext(invlist, &start, &end)) {
8280 if (end == UV_MAX) {
8281 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8283 else if (end != start) {
8284 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8288 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
8296 #ifdef PERL_ARGS_ASSERT__INVLIST_DUMP
8298 Perl__invlist_dump(pTHX_ SV* const invlist, const char * const header)
8300 /* Dumps out the ranges in an inversion list. The string 'header'
8301 * if present is output on a line before the first range */
8305 PERL_ARGS_ASSERT__INVLIST_DUMP;
8307 if (header && strlen(header)) {
8308 PerlIO_printf(Perl_debug_log, "%s\n", header);
8310 if (invlist_is_iterating(invlist)) {
8311 PerlIO_printf(Perl_debug_log, "Can't dump because is in middle of iterating\n");
8315 invlist_iterinit(invlist);
8316 while (invlist_iternext(invlist, &start, &end)) {
8317 if (end == UV_MAX) {
8318 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. INFINITY\n", start);
8320 else if (end != start) {
8321 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. 0x%04"UVXf"\n",
8325 PerlIO_printf(Perl_debug_log, "0x%04"UVXf"\n", start);
8333 S__invlistEQ(pTHX_ SV* const a, SV* const b, bool complement_b)
8335 /* Return a boolean as to if the two passed in inversion lists are
8336 * identical. The final argument, if TRUE, says to take the complement of
8337 * the second inversion list before doing the comparison */
8339 UV* array_a = invlist_array(a);
8340 UV* array_b = invlist_array(b);
8341 UV len_a = _invlist_len(a);
8342 UV len_b = _invlist_len(b);
8344 UV i = 0; /* current index into the arrays */
8345 bool retval = TRUE; /* Assume are identical until proven otherwise */
8347 PERL_ARGS_ASSERT__INVLISTEQ;
8349 /* If are to compare 'a' with the complement of b, set it
8350 * up so are looking at b's complement. */
8353 /* The complement of nothing is everything, so <a> would have to have
8354 * just one element, starting at zero (ending at infinity) */
8356 return (len_a == 1 && array_a[0] == 0);
8358 else if (array_b[0] == 0) {
8360 /* Otherwise, to complement, we invert. Here, the first element is
8361 * 0, just remove it. To do this, we just pretend the array starts
8362 * one later, and clear the flag as we don't have to do anything
8367 complement_b = FALSE;
8371 /* But if the first element is not zero, we unshift a 0 before the
8372 * array. The data structure reserves a space for that 0 (which
8373 * should be a '1' right now), so physical shifting is unneeded,
8374 * but temporarily change that element to 0. Before exiting the
8375 * routine, we must restore the element to '1' */
8382 /* Make sure that the lengths are the same, as well as the final element
8383 * before looping through the remainder. (Thus we test the length, final,
8384 * and first elements right off the bat) */
8385 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
8388 else for (i = 0; i < len_a - 1; i++) {
8389 if (array_a[i] != array_b[i]) {
8402 #undef HEADER_LENGTH
8403 #undef INVLIST_INITIAL_LENGTH
8404 #undef TO_INTERNAL_SIZE
8405 #undef FROM_INTERNAL_SIZE
8406 #undef INVLIST_LEN_OFFSET
8407 #undef INVLIST_ZERO_OFFSET
8408 #undef INVLIST_ITER_OFFSET
8409 #undef INVLIST_VERSION_ID
8410 #undef INVLIST_PREVIOUS_INDEX_OFFSET
8412 /* End of inversion list object */
8415 S_parse_lparen_question_flags(pTHX_ struct RExC_state_t *pRExC_state)
8417 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
8418 * constructs, and updates RExC_flags with them. On input, RExC_parse
8419 * should point to the first flag; it is updated on output to point to the
8420 * final ')' or ':'. There needs to be at least one flag, or this will
8423 /* for (?g), (?gc), and (?o) warnings; warning
8424 about (?c) will warn about (?g) -- japhy */
8426 #define WASTED_O 0x01
8427 #define WASTED_G 0x02
8428 #define WASTED_C 0x04
8429 #define WASTED_GC (WASTED_G|WASTED_C)
8430 I32 wastedflags = 0x00;
8431 U32 posflags = 0, negflags = 0;
8432 U32 *flagsp = &posflags;
8433 char has_charset_modifier = '\0';
8435 bool has_use_defaults = FALSE;
8436 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
8438 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
8440 /* '^' as an initial flag sets certain defaults */
8441 if (UCHARAT(RExC_parse) == '^') {
8443 has_use_defaults = TRUE;
8444 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
8445 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
8446 ? REGEX_UNICODE_CHARSET
8447 : REGEX_DEPENDS_CHARSET);
8450 cs = get_regex_charset(RExC_flags);
8451 if (cs == REGEX_DEPENDS_CHARSET
8452 && (RExC_utf8 || RExC_uni_semantics))
8454 cs = REGEX_UNICODE_CHARSET;
8457 while (*RExC_parse) {
8458 /* && strchr("iogcmsx", *RExC_parse) */
8459 /* (?g), (?gc) and (?o) are useless here
8460 and must be globally applied -- japhy */
8461 switch (*RExC_parse) {
8463 /* Code for the imsx flags */
8464 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
8466 case LOCALE_PAT_MOD:
8467 if (has_charset_modifier) {
8468 goto excess_modifier;
8470 else if (flagsp == &negflags) {
8473 cs = REGEX_LOCALE_CHARSET;
8474 has_charset_modifier = LOCALE_PAT_MOD;
8475 RExC_contains_locale = 1;
8477 case UNICODE_PAT_MOD:
8478 if (has_charset_modifier) {
8479 goto excess_modifier;
8481 else if (flagsp == &negflags) {
8484 cs = REGEX_UNICODE_CHARSET;
8485 has_charset_modifier = UNICODE_PAT_MOD;
8487 case ASCII_RESTRICT_PAT_MOD:
8488 if (flagsp == &negflags) {
8491 if (has_charset_modifier) {
8492 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
8493 goto excess_modifier;
8495 /* Doubled modifier implies more restricted */
8496 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
8499 cs = REGEX_ASCII_RESTRICTED_CHARSET;
8501 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
8503 case DEPENDS_PAT_MOD:
8504 if (has_use_defaults) {
8505 goto fail_modifiers;
8507 else if (flagsp == &negflags) {
8510 else if (has_charset_modifier) {
8511 goto excess_modifier;
8514 /* The dual charset means unicode semantics if the
8515 * pattern (or target, not known until runtime) are
8516 * utf8, or something in the pattern indicates unicode
8518 cs = (RExC_utf8 || RExC_uni_semantics)
8519 ? REGEX_UNICODE_CHARSET
8520 : REGEX_DEPENDS_CHARSET;
8521 has_charset_modifier = DEPENDS_PAT_MOD;
8525 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
8526 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
8528 else if (has_charset_modifier == *(RExC_parse - 1)) {
8529 vFAIL2("Regexp modifier \"%c\" may not appear twice", *(RExC_parse - 1));
8532 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
8537 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"", *(RExC_parse - 1));
8539 case ONCE_PAT_MOD: /* 'o' */
8540 case GLOBAL_PAT_MOD: /* 'g' */
8541 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8542 const I32 wflagbit = *RExC_parse == 'o' ? WASTED_O : WASTED_G;
8543 if (! (wastedflags & wflagbit) ) {
8544 wastedflags |= wflagbit;
8545 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
8548 "Useless (%s%c) - %suse /%c modifier",
8549 flagsp == &negflags ? "?-" : "?",
8551 flagsp == &negflags ? "don't " : "",
8558 case CONTINUE_PAT_MOD: /* 'c' */
8559 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8560 if (! (wastedflags & WASTED_C) ) {
8561 wastedflags |= WASTED_GC;
8562 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
8565 "Useless (%sc) - %suse /gc modifier",
8566 flagsp == &negflags ? "?-" : "?",
8567 flagsp == &negflags ? "don't " : ""
8572 case KEEPCOPY_PAT_MOD: /* 'p' */
8573 if (flagsp == &negflags) {
8575 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
8577 *flagsp |= RXf_PMf_KEEPCOPY;
8581 /* A flag is a default iff it is following a minus, so
8582 * if there is a minus, it means will be trying to
8583 * re-specify a default which is an error */
8584 if (has_use_defaults || flagsp == &negflags) {
8585 goto fail_modifiers;
8588 wastedflags = 0; /* reset so (?g-c) warns twice */
8592 RExC_flags |= posflags;
8593 RExC_flags &= ~negflags;
8594 set_regex_charset(&RExC_flags, cs);
8600 vFAIL3("Sequence (%.*s...) not recognized",
8601 RExC_parse-seqstart, seqstart);
8610 - reg - regular expression, i.e. main body or parenthesized thing
8612 * Caller must absorb opening parenthesis.
8614 * Combining parenthesis handling with the base level of regular expression
8615 * is a trifle forced, but the need to tie the tails of the branches to what
8616 * follows makes it hard to avoid.
8618 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
8620 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
8622 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
8625 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
8626 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
8627 needs to be restarted.
8628 Otherwise would only return NULL if regbranch() returns NULL, which
8631 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
8632 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
8633 * 2 is like 1, but indicates that nextchar() has been called to advance
8634 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
8635 * this flag alerts us to the need to check for that */
8638 regnode *ret; /* Will be the head of the group. */
8641 regnode *ender = NULL;
8644 U32 oregflags = RExC_flags;
8645 bool have_branch = 0;
8647 I32 freeze_paren = 0;
8648 I32 after_freeze = 0;
8650 char * parse_start = RExC_parse; /* MJD */
8651 char * const oregcomp_parse = RExC_parse;
8653 GET_RE_DEBUG_FLAGS_DECL;
8655 PERL_ARGS_ASSERT_REG;
8656 DEBUG_PARSE("reg ");
8658 *flagp = 0; /* Tentatively. */
8661 /* Make an OPEN node, if parenthesized. */
8664 /* Under /x, space and comments can be gobbled up between the '(' and
8665 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
8666 * intervening space, as the sequence is a token, and a token should be
8668 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
8670 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
8671 char *start_verb = RExC_parse;
8672 STRLEN verb_len = 0;
8673 char *start_arg = NULL;
8674 unsigned char op = 0;
8676 int internal_argval = 0; /* internal_argval is only useful if !argok */
8678 if (has_intervening_patws && SIZE_ONLY) {
8679 ckWARNregdep(RExC_parse + 1, "In '(*VERB...)', splitting the initial '(*' is deprecated");
8681 while ( *RExC_parse && *RExC_parse != ')' ) {
8682 if ( *RExC_parse == ':' ) {
8683 start_arg = RExC_parse + 1;
8689 verb_len = RExC_parse - start_verb;
8692 while ( *RExC_parse && *RExC_parse != ')' )
8694 if ( *RExC_parse != ')' )
8695 vFAIL("Unterminated verb pattern argument");
8696 if ( RExC_parse == start_arg )
8699 if ( *RExC_parse != ')' )
8700 vFAIL("Unterminated verb pattern");
8703 switch ( *start_verb ) {
8704 case 'A': /* (*ACCEPT) */
8705 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
8707 internal_argval = RExC_nestroot;
8710 case 'C': /* (*COMMIT) */
8711 if ( memEQs(start_verb,verb_len,"COMMIT") )
8714 case 'F': /* (*FAIL) */
8715 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
8720 case ':': /* (*:NAME) */
8721 case 'M': /* (*MARK:NAME) */
8722 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
8727 case 'P': /* (*PRUNE) */
8728 if ( memEQs(start_verb,verb_len,"PRUNE") )
8731 case 'S': /* (*SKIP) */
8732 if ( memEQs(start_verb,verb_len,"SKIP") )
8735 case 'T': /* (*THEN) */
8736 /* [19:06] <TimToady> :: is then */
8737 if ( memEQs(start_verb,verb_len,"THEN") ) {
8739 RExC_seen |= REG_SEEN_CUTGROUP;
8745 vFAIL3("Unknown verb pattern '%.*s'",
8746 verb_len, start_verb);
8749 if ( start_arg && internal_argval ) {
8750 vFAIL3("Verb pattern '%.*s' may not have an argument",
8751 verb_len, start_verb);
8752 } else if ( argok < 0 && !start_arg ) {
8753 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
8754 verb_len, start_verb);
8756 ret = reganode(pRExC_state, op, internal_argval);
8757 if ( ! internal_argval && ! SIZE_ONLY ) {
8759 SV *sv = newSVpvn( start_arg, RExC_parse - start_arg);
8760 ARG(ret) = add_data( pRExC_state, 1, "S" );
8761 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
8768 if (!internal_argval)
8769 RExC_seen |= REG_SEEN_VERBARG;
8770 } else if ( start_arg ) {
8771 vFAIL3("Verb pattern '%.*s' may not have an argument",
8772 verb_len, start_verb);
8774 ret = reg_node(pRExC_state, op);
8776 nextchar(pRExC_state);
8779 else if (*RExC_parse == '?') { /* (?...) */
8780 bool is_logical = 0;
8781 const char * const seqstart = RExC_parse;
8782 if (has_intervening_patws && SIZE_ONLY) {
8783 ckWARNregdep(RExC_parse + 1, "In '(?...)', splitting the initial '(?' is deprecated");
8787 paren = *RExC_parse++;
8788 ret = NULL; /* For look-ahead/behind. */
8791 case 'P': /* (?P...) variants for those used to PCRE/Python */
8792 paren = *RExC_parse++;
8793 if ( paren == '<') /* (?P<...>) named capture */
8795 else if (paren == '>') { /* (?P>name) named recursion */
8796 goto named_recursion;
8798 else if (paren == '=') { /* (?P=...) named backref */
8799 /* this pretty much dupes the code for \k<NAME> in regatom(), if
8800 you change this make sure you change that */
8801 char* name_start = RExC_parse;
8803 SV *sv_dat = reg_scan_name(pRExC_state,
8804 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8805 if (RExC_parse == name_start || *RExC_parse != ')')
8806 vFAIL2("Sequence %.3s... not terminated",parse_start);
8809 num = add_data( pRExC_state, 1, "S" );
8810 RExC_rxi->data->data[num]=(void*)sv_dat;
8811 SvREFCNT_inc_simple_void(sv_dat);
8814 ret = reganode(pRExC_state,
8817 : (ASCII_FOLD_RESTRICTED)
8819 : (AT_LEAST_UNI_SEMANTICS)
8827 Set_Node_Offset(ret, parse_start+1);
8828 Set_Node_Cur_Length(ret); /* MJD */
8830 nextchar(pRExC_state);
8834 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8836 case '<': /* (?<...) */
8837 if (*RExC_parse == '!')
8839 else if (*RExC_parse != '=')
8845 case '\'': /* (?'...') */
8846 name_start= RExC_parse;
8847 svname = reg_scan_name(pRExC_state,
8848 SIZE_ONLY ? /* reverse test from the others */
8849 REG_RSN_RETURN_NAME :
8850 REG_RSN_RETURN_NULL);
8851 if (RExC_parse == name_start) {
8853 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8856 if (*RExC_parse != paren)
8857 vFAIL2("Sequence (?%c... not terminated",
8858 paren=='>' ? '<' : paren);
8862 if (!svname) /* shouldn't happen */
8864 "panic: reg_scan_name returned NULL");
8865 if (!RExC_paren_names) {
8866 RExC_paren_names= newHV();
8867 sv_2mortal(MUTABLE_SV(RExC_paren_names));
8869 RExC_paren_name_list= newAV();
8870 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
8873 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
8875 sv_dat = HeVAL(he_str);
8877 /* croak baby croak */
8879 "panic: paren_name hash element allocation failed");
8880 } else if ( SvPOK(sv_dat) ) {
8881 /* (?|...) can mean we have dupes so scan to check
8882 its already been stored. Maybe a flag indicating
8883 we are inside such a construct would be useful,
8884 but the arrays are likely to be quite small, so
8885 for now we punt -- dmq */
8886 IV count = SvIV(sv_dat);
8887 I32 *pv = (I32*)SvPVX(sv_dat);
8889 for ( i = 0 ; i < count ; i++ ) {
8890 if ( pv[i] == RExC_npar ) {
8896 pv = (I32*)SvGROW(sv_dat, SvCUR(sv_dat) + sizeof(I32)+1);
8897 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
8898 pv[count] = RExC_npar;
8899 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
8902 (void)SvUPGRADE(sv_dat,SVt_PVNV);
8903 sv_setpvn(sv_dat, (char *)&(RExC_npar), sizeof(I32));
8905 SvIV_set(sv_dat, 1);
8908 /* Yes this does cause a memory leak in debugging Perls */
8909 if (!av_store(RExC_paren_name_list, RExC_npar, SvREFCNT_inc(svname)))
8910 SvREFCNT_dec_NN(svname);
8913 /*sv_dump(sv_dat);*/
8915 nextchar(pRExC_state);
8917 goto capturing_parens;
8919 RExC_seen |= REG_SEEN_LOOKBEHIND;
8920 RExC_in_lookbehind++;
8922 case '=': /* (?=...) */
8923 RExC_seen_zerolen++;
8925 case '!': /* (?!...) */
8926 RExC_seen_zerolen++;
8927 if (*RExC_parse == ')') {
8928 ret=reg_node(pRExC_state, OPFAIL);
8929 nextchar(pRExC_state);
8933 case '|': /* (?|...) */
8934 /* branch reset, behave like a (?:...) except that
8935 buffers in alternations share the same numbers */
8937 after_freeze = freeze_paren = RExC_npar;
8939 case ':': /* (?:...) */
8940 case '>': /* (?>...) */
8942 case '$': /* (?$...) */
8943 case '@': /* (?@...) */
8944 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
8946 case '#': /* (?#...) */
8947 /* XXX As soon as we disallow separating the '?' and '*' (by
8948 * spaces or (?#...) comment), it is believed that this case
8949 * will be unreachable and can be removed. See
8951 while (*RExC_parse && *RExC_parse != ')')
8953 if (*RExC_parse != ')')
8954 FAIL("Sequence (?#... not terminated");
8955 nextchar(pRExC_state);
8958 case '0' : /* (?0) */
8959 case 'R' : /* (?R) */
8960 if (*RExC_parse != ')')
8961 FAIL("Sequence (?R) not terminated");
8962 ret = reg_node(pRExC_state, GOSTART);
8963 *flagp |= POSTPONED;
8964 nextchar(pRExC_state);
8967 { /* named and numeric backreferences */
8969 case '&': /* (?&NAME) */
8970 parse_start = RExC_parse - 1;
8973 SV *sv_dat = reg_scan_name(pRExC_state,
8974 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8975 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8977 goto gen_recurse_regop;
8978 assert(0); /* NOT REACHED */
8980 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8982 vFAIL("Illegal pattern");
8984 goto parse_recursion;
8986 case '-': /* (?-1) */
8987 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8988 RExC_parse--; /* rewind to let it be handled later */
8992 case '1': case '2': case '3': case '4': /* (?1) */
8993 case '5': case '6': case '7': case '8': case '9':
8996 num = atoi(RExC_parse);
8997 parse_start = RExC_parse - 1; /* MJD */
8998 if (*RExC_parse == '-')
9000 while (isDIGIT(*RExC_parse))
9002 if (*RExC_parse!=')')
9003 vFAIL("Expecting close bracket");
9006 if ( paren == '-' ) {
9008 Diagram of capture buffer numbering.
9009 Top line is the normal capture buffer numbers
9010 Bottom line is the negative indexing as from
9014 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
9018 num = RExC_npar + num;
9021 vFAIL("Reference to nonexistent group");
9023 } else if ( paren == '+' ) {
9024 num = RExC_npar + num - 1;
9027 ret = reganode(pRExC_state, GOSUB, num);
9029 if (num > (I32)RExC_rx->nparens) {
9031 vFAIL("Reference to nonexistent group");
9033 ARG2L_SET( ret, RExC_recurse_count++);
9035 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9036 "Recurse #%"UVuf" to %"IVdf"\n", (UV)ARG(ret), (IV)ARG2L(ret)));
9040 RExC_seen |= REG_SEEN_RECURSE;
9041 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
9042 Set_Node_Offset(ret, parse_start); /* MJD */
9044 *flagp |= POSTPONED;
9045 nextchar(pRExC_state);
9047 } /* named and numeric backreferences */
9048 assert(0); /* NOT REACHED */
9050 case '?': /* (??...) */
9052 if (*RExC_parse != '{') {
9054 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
9057 *flagp |= POSTPONED;
9058 paren = *RExC_parse++;
9060 case '{': /* (?{...}) */
9063 struct reg_code_block *cb;
9065 RExC_seen_zerolen++;
9067 if ( !pRExC_state->num_code_blocks
9068 || pRExC_state->code_index >= pRExC_state->num_code_blocks
9069 || pRExC_state->code_blocks[pRExC_state->code_index].start
9070 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
9073 if (RExC_pm_flags & PMf_USE_RE_EVAL)
9074 FAIL("panic: Sequence (?{...}): no code block found\n");
9075 FAIL("Eval-group not allowed at runtime, use re 'eval'");
9077 /* this is a pre-compiled code block (?{...}) */
9078 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
9079 RExC_parse = RExC_start + cb->end;
9082 if (cb->src_regex) {
9083 n = add_data(pRExC_state, 2, "rl");
9084 RExC_rxi->data->data[n] =
9085 (void*)SvREFCNT_inc((SV*)cb->src_regex);
9086 RExC_rxi->data->data[n+1] = (void*)o;
9089 n = add_data(pRExC_state, 1,
9090 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l");
9091 RExC_rxi->data->data[n] = (void*)o;
9094 pRExC_state->code_index++;
9095 nextchar(pRExC_state);
9099 ret = reg_node(pRExC_state, LOGICAL);
9100 eval = reganode(pRExC_state, EVAL, n);
9103 /* for later propagation into (??{}) return value */
9104 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
9106 REGTAIL(pRExC_state, ret, eval);
9107 /* deal with the length of this later - MJD */
9110 ret = reganode(pRExC_state, EVAL, n);
9111 Set_Node_Length(ret, RExC_parse - parse_start + 1);
9112 Set_Node_Offset(ret, parse_start);
9115 case '(': /* (?(?{...})...) and (?(?=...)...) */
9118 if (RExC_parse[0] == '?') { /* (?(?...)) */
9119 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
9120 || RExC_parse[1] == '<'
9121 || RExC_parse[1] == '{') { /* Lookahead or eval. */
9125 ret = reg_node(pRExC_state, LOGICAL);
9129 tail = reg(pRExC_state, 1, &flag, depth+1);
9130 if (flag & RESTART_UTF8) {
9131 *flagp = RESTART_UTF8;
9134 REGTAIL(pRExC_state, ret, tail);
9138 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9139 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9141 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9142 char *name_start= RExC_parse++;
9144 SV *sv_dat=reg_scan_name(pRExC_state,
9145 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9146 if (RExC_parse == name_start || *RExC_parse != ch)
9147 vFAIL2("Sequence (?(%c... not terminated",
9148 (ch == '>' ? '<' : ch));
9151 num = add_data( pRExC_state, 1, "S" );
9152 RExC_rxi->data->data[num]=(void*)sv_dat;
9153 SvREFCNT_inc_simple_void(sv_dat);
9155 ret = reganode(pRExC_state,NGROUPP,num);
9156 goto insert_if_check_paren;
9158 else if (RExC_parse[0] == 'D' &&
9159 RExC_parse[1] == 'E' &&
9160 RExC_parse[2] == 'F' &&
9161 RExC_parse[3] == 'I' &&
9162 RExC_parse[4] == 'N' &&
9163 RExC_parse[5] == 'E')
9165 ret = reganode(pRExC_state,DEFINEP,0);
9168 goto insert_if_check_paren;
9170 else if (RExC_parse[0] == 'R') {
9173 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9174 parno = atoi(RExC_parse++);
9175 while (isDIGIT(*RExC_parse))
9177 } else if (RExC_parse[0] == '&') {
9180 sv_dat = reg_scan_name(pRExC_state,
9181 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9182 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9184 ret = reganode(pRExC_state,INSUBP,parno);
9185 goto insert_if_check_paren;
9187 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9190 parno = atoi(RExC_parse++);
9192 while (isDIGIT(*RExC_parse))
9194 ret = reganode(pRExC_state, GROUPP, parno);
9196 insert_if_check_paren:
9197 if ((c = *nextchar(pRExC_state)) != ')')
9198 vFAIL("Switch condition not recognized");
9200 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9201 br = regbranch(pRExC_state, &flags, 1,depth+1);
9203 if (flags & RESTART_UTF8) {
9204 *flagp = RESTART_UTF8;
9207 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9210 REGTAIL(pRExC_state, br, reganode(pRExC_state, LONGJMP, 0));
9211 c = *nextchar(pRExC_state);
9216 vFAIL("(?(DEFINE)....) does not allow branches");
9217 lastbr = reganode(pRExC_state, IFTHEN, 0); /* Fake one for optimizer. */
9218 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
9219 if (flags & RESTART_UTF8) {
9220 *flagp = RESTART_UTF8;
9223 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9226 REGTAIL(pRExC_state, ret, lastbr);
9229 c = *nextchar(pRExC_state);
9234 vFAIL("Switch (?(condition)... contains too many branches");
9235 ender = reg_node(pRExC_state, TAIL);
9236 REGTAIL(pRExC_state, br, ender);
9238 REGTAIL(pRExC_state, lastbr, ender);
9239 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
9242 REGTAIL(pRExC_state, ret, ender);
9243 RExC_size++; /* XXX WHY do we need this?!!
9244 For large programs it seems to be required
9245 but I can't figure out why. -- dmq*/
9249 vFAIL2("Unknown switch condition (?(%.2s", RExC_parse);
9252 case '[': /* (?[ ... ]) */
9253 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
9256 RExC_parse--; /* for vFAIL to print correctly */
9257 vFAIL("Sequence (? incomplete");
9259 default: /* e.g., (?i) */
9262 parse_lparen_question_flags(pRExC_state);
9263 if (UCHARAT(RExC_parse) != ':') {
9264 nextchar(pRExC_state);
9269 nextchar(pRExC_state);
9279 ret = reganode(pRExC_state, OPEN, parno);
9282 RExC_nestroot = parno;
9283 if (RExC_seen & REG_SEEN_RECURSE
9284 && !RExC_open_parens[parno-1])
9286 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9287 "Setting open paren #%"IVdf" to %d\n",
9288 (IV)parno, REG_NODE_NUM(ret)));
9289 RExC_open_parens[parno-1]= ret;
9292 Set_Node_Length(ret, 1); /* MJD */
9293 Set_Node_Offset(ret, RExC_parse); /* MJD */
9301 /* Pick up the branches, linking them together. */
9302 parse_start = RExC_parse; /* MJD */
9303 br = regbranch(pRExC_state, &flags, 1,depth+1);
9305 /* branch_len = (paren != 0); */
9308 if (flags & RESTART_UTF8) {
9309 *flagp = RESTART_UTF8;
9312 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
9314 if (*RExC_parse == '|') {
9315 if (!SIZE_ONLY && RExC_extralen) {
9316 reginsert(pRExC_state, BRANCHJ, br, depth+1);
9319 reginsert(pRExC_state, BRANCH, br, depth+1);
9320 Set_Node_Length(br, paren != 0);
9321 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
9325 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
9327 else if (paren == ':') {
9328 *flagp |= flags&SIMPLE;
9330 if (is_open) { /* Starts with OPEN. */
9331 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
9333 else if (paren != '?') /* Not Conditional */
9335 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9337 while (*RExC_parse == '|') {
9338 if (!SIZE_ONLY && RExC_extralen) {
9339 ender = reganode(pRExC_state, LONGJMP,0);
9340 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender); /* Append to the previous. */
9343 RExC_extralen += 2; /* Account for LONGJMP. */
9344 nextchar(pRExC_state);
9346 if (RExC_npar > after_freeze)
9347 after_freeze = RExC_npar;
9348 RExC_npar = freeze_paren;
9350 br = regbranch(pRExC_state, &flags, 0, depth+1);
9353 if (flags & RESTART_UTF8) {
9354 *flagp = RESTART_UTF8;
9357 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
9359 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
9361 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9364 if (have_branch || paren != ':') {
9365 /* Make a closing node, and hook it on the end. */
9368 ender = reg_node(pRExC_state, TAIL);
9371 ender = reganode(pRExC_state, CLOSE, parno);
9372 if (!SIZE_ONLY && RExC_seen & REG_SEEN_RECURSE) {
9373 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9374 "Setting close paren #%"IVdf" to %d\n",
9375 (IV)parno, REG_NODE_NUM(ender)));
9376 RExC_close_parens[parno-1]= ender;
9377 if (RExC_nestroot == parno)
9380 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
9381 Set_Node_Length(ender,1); /* MJD */
9387 *flagp &= ~HASWIDTH;
9390 ender = reg_node(pRExC_state, SUCCEED);
9393 ender = reg_node(pRExC_state, END);
9395 assert(!RExC_opend); /* there can only be one! */
9400 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9401 SV * const mysv_val1=sv_newmortal();
9402 SV * const mysv_val2=sv_newmortal();
9403 DEBUG_PARSE_MSG("lsbr");
9404 regprop(RExC_rx, mysv_val1, lastbr);
9405 regprop(RExC_rx, mysv_val2, ender);
9406 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9407 SvPV_nolen_const(mysv_val1),
9408 (IV)REG_NODE_NUM(lastbr),
9409 SvPV_nolen_const(mysv_val2),
9410 (IV)REG_NODE_NUM(ender),
9411 (IV)(ender - lastbr)
9414 REGTAIL(pRExC_state, lastbr, ender);
9416 if (have_branch && !SIZE_ONLY) {
9419 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
9421 /* Hook the tails of the branches to the closing node. */
9422 for (br = ret; br; br = regnext(br)) {
9423 const U8 op = PL_regkind[OP(br)];
9425 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
9426 if (OP(NEXTOPER(br)) != NOTHING || regnext(NEXTOPER(br)) != ender)
9429 else if (op == BRANCHJ) {
9430 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
9431 /* for now we always disable this optimisation * /
9432 if (OP(NEXTOPER(NEXTOPER(br))) != NOTHING || regnext(NEXTOPER(NEXTOPER(br))) != ender)
9438 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
9439 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9440 SV * const mysv_val1=sv_newmortal();
9441 SV * const mysv_val2=sv_newmortal();
9442 DEBUG_PARSE_MSG("NADA");
9443 regprop(RExC_rx, mysv_val1, ret);
9444 regprop(RExC_rx, mysv_val2, ender);
9445 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9446 SvPV_nolen_const(mysv_val1),
9447 (IV)REG_NODE_NUM(ret),
9448 SvPV_nolen_const(mysv_val2),
9449 (IV)REG_NODE_NUM(ender),
9454 if (OP(ender) == TAIL) {
9459 for ( opt= br + 1; opt < ender ; opt++ )
9461 NEXT_OFF(br)= ender - br;
9469 static const char parens[] = "=!<,>";
9471 if (paren && (p = strchr(parens, paren))) {
9472 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
9473 int flag = (p - parens) > 1;
9476 node = SUSPEND, flag = 0;
9477 reginsert(pRExC_state, node,ret, depth+1);
9478 Set_Node_Cur_Length(ret);
9479 Set_Node_Offset(ret, parse_start + 1);
9481 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
9485 /* Check for proper termination. */
9487 /* restore original flags, but keep (?p) */
9488 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
9489 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
9490 RExC_parse = oregcomp_parse;
9491 vFAIL("Unmatched (");
9494 else if (!paren && RExC_parse < RExC_end) {
9495 if (*RExC_parse == ')') {
9497 vFAIL("Unmatched )");
9500 FAIL("Junk on end of regexp"); /* "Can't happen". */
9501 assert(0); /* NOTREACHED */
9504 if (RExC_in_lookbehind) {
9505 RExC_in_lookbehind--;
9507 if (after_freeze > RExC_npar)
9508 RExC_npar = after_freeze;
9513 - regbranch - one alternative of an | operator
9515 * Implements the concatenation operator.
9517 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9521 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
9525 regnode *chain = NULL;
9527 I32 flags = 0, c = 0;
9528 GET_RE_DEBUG_FLAGS_DECL;
9530 PERL_ARGS_ASSERT_REGBRANCH;
9532 DEBUG_PARSE("brnc");
9537 if (!SIZE_ONLY && RExC_extralen)
9538 ret = reganode(pRExC_state, BRANCHJ,0);
9540 ret = reg_node(pRExC_state, BRANCH);
9541 Set_Node_Length(ret, 1);
9545 if (!first && SIZE_ONLY)
9546 RExC_extralen += 1; /* BRANCHJ */
9548 *flagp = WORST; /* Tentatively. */
9551 nextchar(pRExC_state);
9552 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
9554 latest = regpiece(pRExC_state, &flags,depth+1);
9555 if (latest == NULL) {
9556 if (flags & TRYAGAIN)
9558 if (flags & RESTART_UTF8) {
9559 *flagp = RESTART_UTF8;
9562 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
9564 else if (ret == NULL)
9566 *flagp |= flags&(HASWIDTH|POSTPONED);
9567 if (chain == NULL) /* First piece. */
9568 *flagp |= flags&SPSTART;
9571 REGTAIL(pRExC_state, chain, latest);
9576 if (chain == NULL) { /* Loop ran zero times. */
9577 chain = reg_node(pRExC_state, NOTHING);
9582 *flagp |= flags&SIMPLE;
9589 - regpiece - something followed by possible [*+?]
9591 * Note that the branching code sequences used for ? and the general cases
9592 * of * and + are somewhat optimized: they use the same NOTHING node as
9593 * both the endmarker for their branch list and the body of the last branch.
9594 * It might seem that this node could be dispensed with entirely, but the
9595 * endmarker role is not redundant.
9597 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
9599 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9603 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
9610 const char * const origparse = RExC_parse;
9612 I32 max = REG_INFTY;
9613 #ifdef RE_TRACK_PATTERN_OFFSETS
9616 const char *maxpos = NULL;
9618 /* Save the original in case we change the emitted regop to a FAIL. */
9619 regnode * const orig_emit = RExC_emit;
9621 GET_RE_DEBUG_FLAGS_DECL;
9623 PERL_ARGS_ASSERT_REGPIECE;
9625 DEBUG_PARSE("piec");
9627 ret = regatom(pRExC_state, &flags,depth+1);
9629 if (flags & (TRYAGAIN|RESTART_UTF8))
9630 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
9632 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
9638 if (op == '{' && regcurly(RExC_parse, FALSE)) {
9640 #ifdef RE_TRACK_PATTERN_OFFSETS
9641 parse_start = RExC_parse; /* MJD */
9643 next = RExC_parse + 1;
9644 while (isDIGIT(*next) || *next == ',') {
9653 if (*next == '}') { /* got one */
9657 min = atoi(RExC_parse);
9661 maxpos = RExC_parse;
9663 if (!max && *maxpos != '0')
9664 max = REG_INFTY; /* meaning "infinity" */
9665 else if (max >= REG_INFTY)
9666 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
9668 nextchar(pRExC_state);
9669 if (max < min) { /* If can't match, warn and optimize to fail
9672 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
9674 /* We can't back off the size because we have to reserve
9675 * enough space for all the things we are about to throw
9676 * away, but we can shrink it by the ammount we are about
9678 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
9681 RExC_emit = orig_emit;
9683 ret = reg_node(pRExC_state, OPFAIL);
9686 else if (max == 0) { /* replace {0} with a nothing node */
9688 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)NOTHING];
9691 RExC_emit = orig_emit;
9693 ret = reg_node(pRExC_state, NOTHING);
9698 if ((flags&SIMPLE)) {
9699 RExC_naughty += 2 + RExC_naughty / 2;
9700 reginsert(pRExC_state, CURLY, ret, depth+1);
9701 Set_Node_Offset(ret, parse_start+1); /* MJD */
9702 Set_Node_Cur_Length(ret);
9705 regnode * const w = reg_node(pRExC_state, WHILEM);
9708 REGTAIL(pRExC_state, ret, w);
9709 if (!SIZE_ONLY && RExC_extralen) {
9710 reginsert(pRExC_state, LONGJMP,ret, depth+1);
9711 reginsert(pRExC_state, NOTHING,ret, depth+1);
9712 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
9714 reginsert(pRExC_state, CURLYX,ret, depth+1);
9716 Set_Node_Offset(ret, parse_start+1);
9717 Set_Node_Length(ret,
9718 op == '{' ? (RExC_parse - parse_start) : 1);
9720 if (!SIZE_ONLY && RExC_extralen)
9721 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
9722 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
9724 RExC_whilem_seen++, RExC_extralen += 3;
9725 RExC_naughty += 4 + RExC_naughty; /* compound interest */
9734 ARG1_SET(ret, (U16)min);
9735 ARG2_SET(ret, (U16)max);
9747 #if 0 /* Now runtime fix should be reliable. */
9749 /* if this is reinstated, don't forget to put this back into perldiag:
9751 =item Regexp *+ operand could be empty at {#} in regex m/%s/
9753 (F) The part of the regexp subject to either the * or + quantifier
9754 could match an empty string. The {#} shows in the regular
9755 expression about where the problem was discovered.
9759 if (!(flags&HASWIDTH) && op != '?')
9760 vFAIL("Regexp *+ operand could be empty");
9763 #ifdef RE_TRACK_PATTERN_OFFSETS
9764 parse_start = RExC_parse;
9766 nextchar(pRExC_state);
9768 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
9770 if (op == '*' && (flags&SIMPLE)) {
9771 reginsert(pRExC_state, STAR, ret, depth+1);
9775 else if (op == '*') {
9779 else if (op == '+' && (flags&SIMPLE)) {
9780 reginsert(pRExC_state, PLUS, ret, depth+1);
9784 else if (op == '+') {
9788 else if (op == '?') {
9793 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
9794 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
9795 ckWARN3reg(RExC_parse,
9796 "%.*s matches null string many times",
9797 (int)(RExC_parse >= origparse ? RExC_parse - origparse : 0),
9799 (void)ReREFCNT_inc(RExC_rx_sv);
9802 if (RExC_parse < RExC_end && *RExC_parse == '?') {
9803 nextchar(pRExC_state);
9804 reginsert(pRExC_state, MINMOD, ret, depth+1);
9805 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
9807 #ifndef REG_ALLOW_MINMOD_SUSPEND
9810 if (RExC_parse < RExC_end && *RExC_parse == '+') {
9812 nextchar(pRExC_state);
9813 ender = reg_node(pRExC_state, SUCCEED);
9814 REGTAIL(pRExC_state, ret, ender);
9815 reginsert(pRExC_state, SUSPEND, ret, depth+1);
9817 ender = reg_node(pRExC_state, TAIL);
9818 REGTAIL(pRExC_state, ret, ender);
9822 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
9824 vFAIL("Nested quantifiers");
9831 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p, UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
9832 const bool strict /* Apply stricter parsing rules? */
9836 /* This is expected to be called by a parser routine that has recognized '\N'
9837 and needs to handle the rest. RExC_parse is expected to point at the first
9838 char following the N at the time of the call. On successful return,
9839 RExC_parse has been updated to point to just after the sequence identified
9840 by this routine, and <*flagp> has been updated.
9842 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
9845 \N may begin either a named sequence, or if outside a character class, mean
9846 to match a non-newline. For non single-quoted regexes, the tokenizer has
9847 attempted to decide which, and in the case of a named sequence, converted it
9848 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
9849 where c1... are the characters in the sequence. For single-quoted regexes,
9850 the tokenizer passes the \N sequence through unchanged; this code will not
9851 attempt to determine this nor expand those, instead raising a syntax error.
9852 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
9853 or there is no '}', it signals that this \N occurrence means to match a
9856 Only the \N{U+...} form should occur in a character class, for the same
9857 reason that '.' inside a character class means to just match a period: it
9858 just doesn't make sense.
9860 The function raises an error (via vFAIL), and doesn't return for various
9861 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
9862 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
9863 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
9864 only possible if node_p is non-NULL.
9867 If <valuep> is non-null, it means the caller can accept an input sequence
9868 consisting of a just a single code point; <*valuep> is set to that value
9869 if the input is such.
9871 If <node_p> is non-null it signifies that the caller can accept any other
9872 legal sequence (i.e., one that isn't just a single code point). <*node_p>
9874 1) \N means not-a-NL: points to a newly created REG_ANY node;
9875 2) \N{}: points to a new NOTHING node;
9876 3) otherwise: points to a new EXACT node containing the resolved
9878 Note that FALSE is returned for single code point sequences if <valuep> is
9882 char * endbrace; /* '}' following the name */
9884 char *endchar; /* Points to '.' or '}' ending cur char in the input
9886 bool has_multiple_chars; /* true if the input stream contains a sequence of
9887 more than one character */
9889 GET_RE_DEBUG_FLAGS_DECL;
9891 PERL_ARGS_ASSERT_GROK_BSLASH_N;
9895 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
9897 /* The [^\n] meaning of \N ignores spaces and comments under the /x
9898 * modifier. The other meaning does not */
9899 p = (RExC_flags & RXf_PMf_EXTENDED)
9900 ? regwhite( pRExC_state, RExC_parse )
9903 /* Disambiguate between \N meaning a named character versus \N meaning
9904 * [^\n]. The former is assumed when it can't be the latter. */
9905 if (*p != '{' || regcurly(p, FALSE)) {
9908 /* no bare \N in a charclass */
9909 if (in_char_class) {
9910 vFAIL("\\N in a character class must be a named character: \\N{...}");
9914 nextchar(pRExC_state);
9915 *node_p = reg_node(pRExC_state, REG_ANY);
9916 *flagp |= HASWIDTH|SIMPLE;
9919 Set_Node_Length(*node_p, 1); /* MJD */
9923 /* Here, we have decided it should be a named character or sequence */
9925 /* The test above made sure that the next real character is a '{', but
9926 * under the /x modifier, it could be separated by space (or a comment and
9927 * \n) and this is not allowed (for consistency with \x{...} and the
9928 * tokenizer handling of \N{NAME}). */
9929 if (*RExC_parse != '{') {
9930 vFAIL("Missing braces on \\N{}");
9933 RExC_parse++; /* Skip past the '{' */
9935 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
9936 || ! (endbrace == RExC_parse /* nothing between the {} */
9937 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below */
9938 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg) */
9940 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
9941 vFAIL("\\N{NAME} must be resolved by the lexer");
9944 if (endbrace == RExC_parse) { /* empty: \N{} */
9947 *node_p = reg_node(pRExC_state,NOTHING);
9949 else if (in_char_class) {
9950 if (SIZE_ONLY && in_char_class) {
9952 RExC_parse++; /* Position after the "}" */
9953 vFAIL("Zero length \\N{}");
9956 ckWARNreg(RExC_parse,
9957 "Ignoring zero length \\N{} in character class");
9965 nextchar(pRExC_state);
9969 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
9970 RExC_parse += 2; /* Skip past the 'U+' */
9972 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9974 /* Code points are separated by dots. If none, there is only one code
9975 * point, and is terminated by the brace */
9976 has_multiple_chars = (endchar < endbrace);
9978 if (valuep && (! has_multiple_chars || in_char_class)) {
9979 /* We only pay attention to the first char of
9980 multichar strings being returned in char classes. I kinda wonder
9981 if this makes sense as it does change the behaviour
9982 from earlier versions, OTOH that behaviour was broken
9983 as well. XXX Solution is to recharacterize as
9984 [rest-of-class]|multi1|multi2... */
9986 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
9987 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
9988 | PERL_SCAN_DISALLOW_PREFIX
9989 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
9991 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
9993 /* The tokenizer should have guaranteed validity, but it's possible to
9994 * bypass it by using single quoting, so check */
9995 if (length_of_hex == 0
9996 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
9998 RExC_parse += length_of_hex; /* Includes all the valid */
9999 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
10000 ? UTF8SKIP(RExC_parse)
10002 /* Guard against malformed utf8 */
10003 if (RExC_parse >= endchar) {
10004 RExC_parse = endchar;
10006 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10009 if (in_char_class && has_multiple_chars) {
10011 RExC_parse = endbrace;
10012 vFAIL("\\N{} in character class restricted to one character");
10015 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
10019 RExC_parse = endbrace + 1;
10021 else if (! node_p || ! has_multiple_chars) {
10023 /* Here, the input is legal, but not according to the caller's
10024 * options. We fail without advancing the parse, so that the
10025 * caller can try again */
10031 /* What is done here is to convert this to a sub-pattern of the form
10032 * (?:\x{char1}\x{char2}...)
10033 * and then call reg recursively. That way, it retains its atomicness,
10034 * while not having to worry about special handling that some code
10035 * points may have. toke.c has converted the original Unicode values
10036 * to native, so that we can just pass on the hex values unchanged. We
10037 * do have to set a flag to keep recoding from happening in the
10040 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
10042 char *orig_end = RExC_end;
10045 while (RExC_parse < endbrace) {
10047 /* Convert to notation the rest of the code understands */
10048 sv_catpv(substitute_parse, "\\x{");
10049 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
10050 sv_catpv(substitute_parse, "}");
10052 /* Point to the beginning of the next character in the sequence. */
10053 RExC_parse = endchar + 1;
10054 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10056 sv_catpv(substitute_parse, ")");
10058 RExC_parse = SvPV(substitute_parse, len);
10060 /* Don't allow empty number */
10062 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10064 RExC_end = RExC_parse + len;
10066 /* The values are Unicode, and therefore not subject to recoding */
10067 RExC_override_recoding = 1;
10069 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
10070 if (flags & RESTART_UTF8) {
10071 *flagp = RESTART_UTF8;
10074 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
10077 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10079 RExC_parse = endbrace;
10080 RExC_end = orig_end;
10081 RExC_override_recoding = 0;
10083 nextchar(pRExC_state);
10093 * It returns the code point in utf8 for the value in *encp.
10094 * value: a code value in the source encoding
10095 * encp: a pointer to an Encode object
10097 * If the result from Encode is not a single character,
10098 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
10101 S_reg_recode(pTHX_ const char value, SV **encp)
10104 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
10105 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
10106 const STRLEN newlen = SvCUR(sv);
10107 UV uv = UNICODE_REPLACEMENT;
10109 PERL_ARGS_ASSERT_REG_RECODE;
10113 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
10116 if (!newlen || numlen != newlen) {
10117 uv = UNICODE_REPLACEMENT;
10123 PERL_STATIC_INLINE U8
10124 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
10128 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
10134 op = get_regex_charset(RExC_flags);
10135 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
10136 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
10137 been, so there is no hole */
10140 return op + EXACTF;
10143 PERL_STATIC_INLINE void
10144 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state, regnode *node, I32* flagp, STRLEN len, UV code_point)
10146 /* This knows the details about sizing an EXACTish node, setting flags for
10147 * it (by setting <*flagp>, and potentially populating it with a single
10150 * If <len> (the length in bytes) is non-zero, this function assumes that
10151 * the node has already been populated, and just does the sizing. In this
10152 * case <code_point> should be the final code point that has already been
10153 * placed into the node. This value will be ignored except that under some
10154 * circumstances <*flagp> is set based on it.
10156 * If <len> is zero, the function assumes that the node is to contain only
10157 * the single character given by <code_point> and calculates what <len>
10158 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10159 * additionally will populate the node's STRING with <code_point>, if <len>
10160 * is 0. In both cases <*flagp> is appropriately set
10162 * It knows that under FOLD, the Latin Sharp S and UTF characters above
10163 * 255, must be folded (the former only when the rules indicate it can
10166 bool len_passed_in = cBOOL(len != 0);
10167 U8 character[UTF8_MAXBYTES_CASE+1];
10169 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10171 if (! len_passed_in) {
10173 if (FOLD && (! LOC || code_point > 255)) {
10174 _to_uni_fold_flags(NATIVE_TO_UNI(code_point),
10177 FOLD_FLAGS_FULL | ((LOC)
10178 ? FOLD_FLAGS_LOCALE
10179 : (ASCII_FOLD_RESTRICTED)
10180 ? FOLD_FLAGS_NOMIX_ASCII
10184 uvchr_to_utf8( character, code_point);
10185 len = UTF8SKIP(character);
10189 || code_point != LATIN_SMALL_LETTER_SHARP_S
10190 || ASCII_FOLD_RESTRICTED
10191 || ! AT_LEAST_UNI_SEMANTICS)
10193 *character = (U8) code_point;
10198 *(character + 1) = 's';
10204 RExC_size += STR_SZ(len);
10207 RExC_emit += STR_SZ(len);
10208 STR_LEN(node) = len;
10209 if (! len_passed_in) {
10210 Copy((char *) character, STRING(node), len, char);
10214 *flagp |= HASWIDTH;
10216 /* A single character node is SIMPLE, except for the special-cased SHARP S
10218 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
10219 && (code_point != LATIN_SMALL_LETTER_SHARP_S
10220 || ! FOLD || ! DEPENDS_SEMANTICS))
10227 - regatom - the lowest level
10229 Try to identify anything special at the start of the pattern. If there
10230 is, then handle it as required. This may involve generating a single regop,
10231 such as for an assertion; or it may involve recursing, such as to
10232 handle a () structure.
10234 If the string doesn't start with something special then we gobble up
10235 as much literal text as we can.
10237 Once we have been able to handle whatever type of thing started the
10238 sequence, we return.
10240 Note: we have to be careful with escapes, as they can be both literal
10241 and special, and in the case of \10 and friends, context determines which.
10243 A summary of the code structure is:
10245 switch (first_byte) {
10246 cases for each special:
10247 handle this special;
10250 switch (2nd byte) {
10251 cases for each unambiguous special:
10252 handle this special;
10254 cases for each ambigous special/literal:
10256 if (special) handle here
10258 default: // unambiguously literal:
10261 default: // is a literal char
10264 create EXACTish node for literal;
10265 while (more input and node isn't full) {
10266 switch (input_byte) {
10267 cases for each special;
10268 make sure parse pointer is set so that the next call to
10269 regatom will see this special first
10270 goto loopdone; // EXACTish node terminated by prev. char
10272 append char to EXACTISH node;
10274 get next input byte;
10278 return the generated node;
10280 Specifically there are two separate switches for handling
10281 escape sequences, with the one for handling literal escapes requiring
10282 a dummy entry for all of the special escapes that are actually handled
10285 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
10287 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10289 Otherwise does not return NULL.
10293 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10296 regnode *ret = NULL;
10298 char *parse_start = RExC_parse;
10302 GET_RE_DEBUG_FLAGS_DECL;
10304 *flagp = WORST; /* Tentatively. */
10306 DEBUG_PARSE("atom");
10308 PERL_ARGS_ASSERT_REGATOM;
10311 switch ((U8)*RExC_parse) {
10313 RExC_seen_zerolen++;
10314 nextchar(pRExC_state);
10315 if (RExC_flags & RXf_PMf_MULTILINE)
10316 ret = reg_node(pRExC_state, MBOL);
10317 else if (RExC_flags & RXf_PMf_SINGLELINE)
10318 ret = reg_node(pRExC_state, SBOL);
10320 ret = reg_node(pRExC_state, BOL);
10321 Set_Node_Length(ret, 1); /* MJD */
10324 nextchar(pRExC_state);
10326 RExC_seen_zerolen++;
10327 if (RExC_flags & RXf_PMf_MULTILINE)
10328 ret = reg_node(pRExC_state, MEOL);
10329 else if (RExC_flags & RXf_PMf_SINGLELINE)
10330 ret = reg_node(pRExC_state, SEOL);
10332 ret = reg_node(pRExC_state, EOL);
10333 Set_Node_Length(ret, 1); /* MJD */
10336 nextchar(pRExC_state);
10337 if (RExC_flags & RXf_PMf_SINGLELINE)
10338 ret = reg_node(pRExC_state, SANY);
10340 ret = reg_node(pRExC_state, REG_ANY);
10341 *flagp |= HASWIDTH|SIMPLE;
10343 Set_Node_Length(ret, 1); /* MJD */
10347 char * const oregcomp_parse = ++RExC_parse;
10348 ret = regclass(pRExC_state, flagp,depth+1,
10349 FALSE, /* means parse the whole char class */
10350 TRUE, /* allow multi-char folds */
10351 FALSE, /* don't silence non-portable warnings. */
10353 if (*RExC_parse != ']') {
10354 RExC_parse = oregcomp_parse;
10355 vFAIL("Unmatched [");
10358 if (*flagp & RESTART_UTF8)
10360 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
10363 nextchar(pRExC_state);
10364 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
10368 nextchar(pRExC_state);
10369 ret = reg(pRExC_state, 2, &flags,depth+1);
10371 if (flags & TRYAGAIN) {
10372 if (RExC_parse == RExC_end) {
10373 /* Make parent create an empty node if needed. */
10374 *flagp |= TRYAGAIN;
10379 if (flags & RESTART_UTF8) {
10380 *flagp = RESTART_UTF8;
10383 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"", (UV) flags);
10385 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10389 if (flags & TRYAGAIN) {
10390 *flagp |= TRYAGAIN;
10393 vFAIL("Internal urp");
10394 /* Supposed to be caught earlier. */
10397 if (!regcurly(RExC_parse, FALSE)) {
10406 vFAIL("Quantifier follows nothing");
10411 This switch handles escape sequences that resolve to some kind
10412 of special regop and not to literal text. Escape sequnces that
10413 resolve to literal text are handled below in the switch marked
10416 Every entry in this switch *must* have a corresponding entry
10417 in the literal escape switch. However, the opposite is not
10418 required, as the default for this switch is to jump to the
10419 literal text handling code.
10421 switch ((U8)*++RExC_parse) {
10423 /* Special Escapes */
10425 RExC_seen_zerolen++;
10426 ret = reg_node(pRExC_state, SBOL);
10428 goto finish_meta_pat;
10430 ret = reg_node(pRExC_state, GPOS);
10431 RExC_seen |= REG_SEEN_GPOS;
10433 goto finish_meta_pat;
10435 RExC_seen_zerolen++;
10436 ret = reg_node(pRExC_state, KEEPS);
10438 /* XXX:dmq : disabling in-place substitution seems to
10439 * be necessary here to avoid cases of memory corruption, as
10440 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
10442 RExC_seen |= REG_SEEN_LOOKBEHIND;
10443 goto finish_meta_pat;
10445 ret = reg_node(pRExC_state, SEOL);
10447 RExC_seen_zerolen++; /* Do not optimize RE away */
10448 goto finish_meta_pat;
10450 ret = reg_node(pRExC_state, EOS);
10452 RExC_seen_zerolen++; /* Do not optimize RE away */
10453 goto finish_meta_pat;
10455 ret = reg_node(pRExC_state, CANY);
10456 RExC_seen |= REG_SEEN_CANY;
10457 *flagp |= HASWIDTH|SIMPLE;
10458 goto finish_meta_pat;
10460 ret = reg_node(pRExC_state, CLUMP);
10461 *flagp |= HASWIDTH;
10462 goto finish_meta_pat;
10468 arg = ANYOF_WORDCHAR;
10472 RExC_seen_zerolen++;
10473 RExC_seen |= REG_SEEN_LOOKBEHIND;
10474 op = BOUND + get_regex_charset(RExC_flags);
10475 if (op > BOUNDA) { /* /aa is same as /a */
10478 ret = reg_node(pRExC_state, op);
10479 FLAGS(ret) = get_regex_charset(RExC_flags);
10481 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10482 ckWARNdep(RExC_parse, "\"\\b{\" is deprecated; use \"\\b\\{\" or \"\\b[{]\" instead");
10484 goto finish_meta_pat;
10486 RExC_seen_zerolen++;
10487 RExC_seen |= REG_SEEN_LOOKBEHIND;
10488 op = NBOUND + get_regex_charset(RExC_flags);
10489 if (op > NBOUNDA) { /* /aa is same as /a */
10492 ret = reg_node(pRExC_state, op);
10493 FLAGS(ret) = get_regex_charset(RExC_flags);
10495 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10496 ckWARNdep(RExC_parse, "\"\\B{\" is deprecated; use \"\\B\\{\" or \"\\B[{]\" instead");
10498 goto finish_meta_pat;
10508 ret = reg_node(pRExC_state, LNBREAK);
10509 *flagp |= HASWIDTH|SIMPLE;
10510 goto finish_meta_pat;
10518 goto join_posix_op_known;
10524 arg = ANYOF_VERTWS;
10526 goto join_posix_op_known;
10536 op = POSIXD + get_regex_charset(RExC_flags);
10537 if (op > POSIXA) { /* /aa is same as /a */
10541 join_posix_op_known:
10544 op += NPOSIXD - POSIXD;
10547 ret = reg_node(pRExC_state, op);
10549 FLAGS(ret) = namedclass_to_classnum(arg);
10552 *flagp |= HASWIDTH|SIMPLE;
10556 nextchar(pRExC_state);
10557 Set_Node_Length(ret, 2); /* MJD */
10563 char* parse_start = RExC_parse - 2;
10568 ret = regclass(pRExC_state, flagp,depth+1,
10569 TRUE, /* means just parse this element */
10570 FALSE, /* don't allow multi-char folds */
10571 FALSE, /* don't silence non-portable warnings.
10572 It would be a bug if these returned
10575 /* regclass() can only return RESTART_UTF8 if multi-char folds
10578 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
10583 Set_Node_Offset(ret, parse_start + 2);
10584 Set_Node_Cur_Length(ret);
10585 nextchar(pRExC_state);
10589 /* Handle \N and \N{NAME} with multiple code points here and not
10590 * below because it can be multicharacter. join_exact() will join
10591 * them up later on. Also this makes sure that things like
10592 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
10593 * The options to the grok function call causes it to fail if the
10594 * sequence is just a single code point. We then go treat it as
10595 * just another character in the current EXACT node, and hence it
10596 * gets uniform treatment with all the other characters. The
10597 * special treatment for quantifiers is not needed for such single
10598 * character sequences */
10600 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
10601 FALSE /* not strict */ )) {
10602 if (*flagp & RESTART_UTF8)
10608 case 'k': /* Handle \k<NAME> and \k'NAME' */
10611 char ch= RExC_parse[1];
10612 if (ch != '<' && ch != '\'' && ch != '{') {
10614 vFAIL2("Sequence %.2s... not terminated",parse_start);
10616 /* this pretty much dupes the code for (?P=...) in reg(), if
10617 you change this make sure you change that */
10618 char* name_start = (RExC_parse += 2);
10620 SV *sv_dat = reg_scan_name(pRExC_state,
10621 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10622 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
10623 if (RExC_parse == name_start || *RExC_parse != ch)
10624 vFAIL2("Sequence %.3s... not terminated",parse_start);
10627 num = add_data( pRExC_state, 1, "S" );
10628 RExC_rxi->data->data[num]=(void*)sv_dat;
10629 SvREFCNT_inc_simple_void(sv_dat);
10633 ret = reganode(pRExC_state,
10636 : (ASCII_FOLD_RESTRICTED)
10638 : (AT_LEAST_UNI_SEMANTICS)
10644 *flagp |= HASWIDTH;
10646 /* override incorrect value set in reganode MJD */
10647 Set_Node_Offset(ret, parse_start+1);
10648 Set_Node_Cur_Length(ret); /* MJD */
10649 nextchar(pRExC_state);
10655 case '1': case '2': case '3': case '4':
10656 case '5': case '6': case '7': case '8': case '9':
10659 bool isg = *RExC_parse == 'g';
10664 if (*RExC_parse == '{') {
10668 if (*RExC_parse == '-') {
10672 if (hasbrace && !isDIGIT(*RExC_parse)) {
10673 if (isrel) RExC_parse--;
10675 goto parse_named_seq;
10677 num = atoi(RExC_parse);
10678 if (isg && num == 0) {
10679 if (*RExC_parse == '0') {
10680 vFAIL("Reference to invalid group 0");
10683 vFAIL("Unterminated \\g... pattern");
10687 num = RExC_npar - num;
10689 vFAIL("Reference to nonexistent or unclosed group");
10691 if (!isg && num > 9 && num >= RExC_npar)
10692 /* Probably a character specified in octal, e.g. \35 */
10695 char * const parse_start = RExC_parse - 1; /* MJD */
10696 while (isDIGIT(*RExC_parse))
10699 if (*RExC_parse != '}')
10700 vFAIL("Unterminated \\g{...} pattern");
10704 if (num > (I32)RExC_rx->nparens)
10705 vFAIL("Reference to nonexistent group");
10708 ret = reganode(pRExC_state,
10711 : (ASCII_FOLD_RESTRICTED)
10713 : (AT_LEAST_UNI_SEMANTICS)
10719 *flagp |= HASWIDTH;
10721 /* override incorrect value set in reganode MJD */
10722 Set_Node_Offset(ret, parse_start+1);
10723 Set_Node_Cur_Length(ret); /* MJD */
10725 nextchar(pRExC_state);
10730 if (RExC_parse >= RExC_end)
10731 FAIL("Trailing \\");
10734 /* Do not generate "unrecognized" warnings here, we fall
10735 back into the quick-grab loop below */
10742 if (RExC_flags & RXf_PMf_EXTENDED) {
10743 if ( reg_skipcomment( pRExC_state ) )
10750 parse_start = RExC_parse - 1;
10759 #define MAX_NODE_STRING_SIZE 127
10760 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
10762 U8 upper_parse = MAX_NODE_STRING_SIZE;
10765 bool next_is_quantifier;
10766 char * oldp = NULL;
10768 /* If a folding node contains only code points that don't
10769 * participate in folds, it can be changed into an EXACT node,
10770 * which allows the optimizer more things to look for */
10774 node_type = compute_EXACTish(pRExC_state);
10775 ret = reg_node(pRExC_state, node_type);
10777 /* In pass1, folded, we use a temporary buffer instead of the
10778 * actual node, as the node doesn't exist yet */
10779 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
10785 /* We do the EXACTFish to EXACT node only if folding, and not if in
10786 * locale, as whether a character folds or not isn't known until
10788 maybe_exact = FOLD && ! LOC;
10790 /* XXX The node can hold up to 255 bytes, yet this only goes to
10791 * 127. I (khw) do not know why. Keeping it somewhat less than
10792 * 255 allows us to not have to worry about overflow due to
10793 * converting to utf8 and fold expansion, but that value is
10794 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
10795 * split up by this limit into a single one using the real max of
10796 * 255. Even at 127, this breaks under rare circumstances. If
10797 * folding, we do not want to split a node at a character that is a
10798 * non-final in a multi-char fold, as an input string could just
10799 * happen to want to match across the node boundary. The join
10800 * would solve that problem if the join actually happens. But a
10801 * series of more than two nodes in a row each of 127 would cause
10802 * the first join to succeed to get to 254, but then there wouldn't
10803 * be room for the next one, which could at be one of those split
10804 * multi-char folds. I don't know of any fool-proof solution. One
10805 * could back off to end with only a code point that isn't such a
10806 * non-final, but it is possible for there not to be any in the
10808 for (p = RExC_parse - 1;
10809 len < upper_parse && p < RExC_end;
10814 if (RExC_flags & RXf_PMf_EXTENDED)
10815 p = regwhite( pRExC_state, p );
10826 /* Literal Escapes Switch
10828 This switch is meant to handle escape sequences that
10829 resolve to a literal character.
10831 Every escape sequence that represents something
10832 else, like an assertion or a char class, is handled
10833 in the switch marked 'Special Escapes' above in this
10834 routine, but also has an entry here as anything that
10835 isn't explicitly mentioned here will be treated as
10836 an unescaped equivalent literal.
10839 switch ((U8)*++p) {
10840 /* These are all the special escapes. */
10841 case 'A': /* Start assertion */
10842 case 'b': case 'B': /* Word-boundary assertion*/
10843 case 'C': /* Single char !DANGEROUS! */
10844 case 'd': case 'D': /* digit class */
10845 case 'g': case 'G': /* generic-backref, pos assertion */
10846 case 'h': case 'H': /* HORIZWS */
10847 case 'k': case 'K': /* named backref, keep marker */
10848 case 'p': case 'P': /* Unicode property */
10849 case 'R': /* LNBREAK */
10850 case 's': case 'S': /* space class */
10851 case 'v': case 'V': /* VERTWS */
10852 case 'w': case 'W': /* word class */
10853 case 'X': /* eXtended Unicode "combining character sequence" */
10854 case 'z': case 'Z': /* End of line/string assertion */
10858 /* Anything after here is an escape that resolves to a
10859 literal. (Except digits, which may or may not)
10865 case 'N': /* Handle a single-code point named character. */
10866 /* The options cause it to fail if a multiple code
10867 * point sequence. Handle those in the switch() above
10869 RExC_parse = p + 1;
10870 if (! grok_bslash_N(pRExC_state, NULL, &ender,
10871 flagp, depth, FALSE,
10872 FALSE /* not strict */ ))
10874 if (*flagp & RESTART_UTF8)
10875 FAIL("panic: grok_bslash_N set RESTART_UTF8");
10876 RExC_parse = p = oldp;
10880 if (ender > 0xff) {
10897 ender = ASCII_TO_NATIVE('\033');
10901 ender = ASCII_TO_NATIVE('\007');
10907 const char* error_msg;
10909 bool valid = grok_bslash_o(&p,
10912 TRUE, /* out warnings */
10913 FALSE, /* not strict */
10914 TRUE, /* Output warnings
10919 RExC_parse = p; /* going to die anyway; point
10920 to exact spot of failure */
10924 if (PL_encoding && ender < 0x100) {
10925 goto recode_encoding;
10927 if (ender > 0xff) {
10934 UV result = UV_MAX; /* initialize to erroneous
10936 const char* error_msg;
10938 bool valid = grok_bslash_x(&p,
10941 TRUE, /* out warnings */
10942 FALSE, /* not strict */
10943 TRUE, /* Output warnings
10948 RExC_parse = p; /* going to die anyway; point
10949 to exact spot of failure */
10954 if (PL_encoding && ender < 0x100) {
10955 goto recode_encoding;
10957 if (ender > 0xff) {
10964 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
10966 case '0': case '1': case '2': case '3':case '4':
10967 case '5': case '6': case '7':
10969 (isDIGIT(p[1]) && atoi(p) >= RExC_npar))
10971 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
10973 ender = grok_oct(p, &numlen, &flags, NULL);
10974 if (ender > 0xff) {
10978 if (SIZE_ONLY /* like \08, \178 */
10981 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
10983 reg_warn_non_literal_string(
10985 form_short_octal_warning(p, numlen));
10988 else { /* Not to be treated as an octal constant, go
10993 if (PL_encoding && ender < 0x100)
10994 goto recode_encoding;
10997 if (! RExC_override_recoding) {
10998 SV* enc = PL_encoding;
10999 ender = reg_recode((const char)(U8)ender, &enc);
11000 if (!enc && SIZE_ONLY)
11001 ckWARNreg(p, "Invalid escape in the specified encoding");
11007 FAIL("Trailing \\");
11010 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
11011 /* Include any { following the alpha to emphasize
11012 * that it could be part of an escape at some point
11014 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
11015 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
11017 goto normal_default;
11018 } /* End of switch on '\' */
11020 default: /* A literal character */
11023 && RExC_flags & RXf_PMf_EXTENDED
11024 && ckWARN_d(WARN_DEPRECATED)
11025 && is_PATWS_non_low(p, UTF))
11027 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
11028 "Escape literal pattern white space under /x");
11032 if (UTF8_IS_START(*p) && UTF) {
11034 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
11035 &numlen, UTF8_ALLOW_DEFAULT);
11041 } /* End of switch on the literal */
11043 /* Here, have looked at the literal character and <ender>
11044 * contains its ordinal, <p> points to the character after it
11047 if ( RExC_flags & RXf_PMf_EXTENDED)
11048 p = regwhite( pRExC_state, p );
11050 /* If the next thing is a quantifier, it applies to this
11051 * character only, which means that this character has to be in
11052 * its own node and can't just be appended to the string in an
11053 * existing node, so if there are already other characters in
11054 * the node, close the node with just them, and set up to do
11055 * this character again next time through, when it will be the
11056 * only thing in its new node */
11057 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
11065 const STRLEN unilen = reguni(pRExC_state, ender, s);
11071 /* The loop increments <len> each time, as all but this
11072 * path (and one other) through it add a single byte to
11073 * the EXACTish node. But this one has changed len to
11074 * be the correct final value, so subtract one to
11075 * cancel out the increment that follows */
11079 REGC((char)ender, s++);
11084 /* See comments for join_exact() as to why we fold this
11085 * non-UTF at compile time */
11086 || (node_type == EXACTFU
11087 && ender == LATIN_SMALL_LETTER_SHARP_S)))
11089 *(s++) = (char) ender;
11090 maybe_exact &= ! IS_IN_SOME_FOLD_L1(ender);
11094 /* Prime the casefolded buffer. Locale rules, which apply
11095 * only to code points < 256, aren't known until execution,
11096 * so for them, just output the original character using
11097 * utf8. If we start to fold non-UTF patterns, be sure to
11098 * update join_exact() */
11099 if (LOC && ender < 256) {
11100 if (UNI_IS_INVARIANT(ender)) {
11104 *s = UTF8_TWO_BYTE_HI(ender);
11105 *(s + 1) = UTF8_TWO_BYTE_LO(ender);
11110 UV folded = _to_uni_fold_flags(
11115 | ((LOC) ? FOLD_FLAGS_LOCALE
11116 : (ASCII_FOLD_RESTRICTED)
11117 ? FOLD_FLAGS_NOMIX_ASCII
11121 /* If this node only contains non-folding code points
11122 * so far, see if this new one is also non-folding */
11124 if (folded != ender) {
11125 maybe_exact = FALSE;
11128 /* Here the fold is the original; we have
11129 * to check further to see if anything
11131 if (! PL_utf8_foldable) {
11132 SV* swash = swash_init("utf8",
11134 &PL_sv_undef, 1, 0);
11136 _get_swash_invlist(swash);
11137 SvREFCNT_dec_NN(swash);
11139 if (_invlist_contains_cp(PL_utf8_foldable,
11142 maybe_exact = FALSE;
11150 /* The loop increments <len> each time, as all but this
11151 * path (and one other) through it add a single byte to the
11152 * EXACTish node. But this one has changed len to be the
11153 * correct final value, so subtract one to cancel out the
11154 * increment that follows */
11155 len += foldlen - 1;
11158 if (next_is_quantifier) {
11160 /* Here, the next input is a quantifier, and to get here,
11161 * the current character is the only one in the node.
11162 * Also, here <len> doesn't include the final byte for this
11168 } /* End of loop through literal characters */
11170 /* Here we have either exhausted the input or ran out of room in
11171 * the node. (If we encountered a character that can't be in the
11172 * node, transfer is made directly to <loopdone>, and so we
11173 * wouldn't have fallen off the end of the loop.) In the latter
11174 * case, we artificially have to split the node into two, because
11175 * we just don't have enough space to hold everything. This
11176 * creates a problem if the final character participates in a
11177 * multi-character fold in the non-final position, as a match that
11178 * should have occurred won't, due to the way nodes are matched,
11179 * and our artificial boundary. So back off until we find a non-
11180 * problematic character -- one that isn't at the beginning or
11181 * middle of such a fold. (Either it doesn't participate in any
11182 * folds, or appears only in the final position of all the folds it
11183 * does participate in.) A better solution with far fewer false
11184 * positives, and that would fill the nodes more completely, would
11185 * be to actually have available all the multi-character folds to
11186 * test against, and to back-off only far enough to be sure that
11187 * this node isn't ending with a partial one. <upper_parse> is set
11188 * further below (if we need to reparse the node) to include just
11189 * up through that final non-problematic character that this code
11190 * identifies, so when it is set to less than the full node, we can
11191 * skip the rest of this */
11192 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
11194 const STRLEN full_len = len;
11196 assert(len >= MAX_NODE_STRING_SIZE);
11198 /* Here, <s> points to the final byte of the final character.
11199 * Look backwards through the string until find a non-
11200 * problematic character */
11204 /* These two have no multi-char folds to non-UTF characters
11206 if (ASCII_FOLD_RESTRICTED || LOC) {
11210 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
11214 if (! PL_NonL1NonFinalFold) {
11215 PL_NonL1NonFinalFold = _new_invlist_C_array(
11216 NonL1_Perl_Non_Final_Folds_invlist);
11219 /* Point to the first byte of the final character */
11220 s = (char *) utf8_hop((U8 *) s, -1);
11222 while (s >= s0) { /* Search backwards until find
11223 non-problematic char */
11224 if (UTF8_IS_INVARIANT(*s)) {
11226 /* There are no ascii characters that participate
11227 * in multi-char folds under /aa. In EBCDIC, the
11228 * non-ascii invariants are all control characters,
11229 * so don't ever participate in any folds. */
11230 if (ASCII_FOLD_RESTRICTED
11231 || ! IS_NON_FINAL_FOLD(*s))
11236 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
11238 /* No Latin1 characters participate in multi-char
11239 * folds under /l */
11241 || ! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_UNI(
11247 else if (! _invlist_contains_cp(
11248 PL_NonL1NonFinalFold,
11249 valid_utf8_to_uvchr((U8 *) s, NULL)))
11254 /* Here, the current character is problematic in that
11255 * it does occur in the non-final position of some
11256 * fold, so try the character before it, but have to
11257 * special case the very first byte in the string, so
11258 * we don't read outside the string */
11259 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
11260 } /* End of loop backwards through the string */
11262 /* If there were only problematic characters in the string,
11263 * <s> will point to before s0, in which case the length
11264 * should be 0, otherwise include the length of the
11265 * non-problematic character just found */
11266 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
11269 /* Here, have found the final character, if any, that is
11270 * non-problematic as far as ending the node without splitting
11271 * it across a potential multi-char fold. <len> contains the
11272 * number of bytes in the node up-to and including that
11273 * character, or is 0 if there is no such character, meaning
11274 * the whole node contains only problematic characters. In
11275 * this case, give up and just take the node as-is. We can't
11281 /* Here, the node does contain some characters that aren't
11282 * problematic. If one such is the final character in the
11283 * node, we are done */
11284 if (len == full_len) {
11287 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
11289 /* If the final character is problematic, but the
11290 * penultimate is not, back-off that last character to
11291 * later start a new node with it */
11296 /* Here, the final non-problematic character is earlier
11297 * in the input than the penultimate character. What we do
11298 * is reparse from the beginning, going up only as far as
11299 * this final ok one, thus guaranteeing that the node ends
11300 * in an acceptable character. The reason we reparse is
11301 * that we know how far in the character is, but we don't
11302 * know how to correlate its position with the input parse.
11303 * An alternate implementation would be to build that
11304 * correlation as we go along during the original parse,
11305 * but that would entail extra work for every node, whereas
11306 * this code gets executed only when the string is too
11307 * large for the node, and the final two characters are
11308 * problematic, an infrequent occurrence. Yet another
11309 * possible strategy would be to save the tail of the
11310 * string, and the next time regatom is called, initialize
11311 * with that. The problem with this is that unless you
11312 * back off one more character, you won't be guaranteed
11313 * regatom will get called again, unless regbranch,
11314 * regpiece ... are also changed. If you do back off that
11315 * extra character, so that there is input guaranteed to
11316 * force calling regatom, you can't handle the case where
11317 * just the first character in the node is acceptable. I
11318 * (khw) decided to try this method which doesn't have that
11319 * pitfall; if performance issues are found, we can do a
11320 * combination of the current approach plus that one */
11326 } /* End of verifying node ends with an appropriate char */
11328 loopdone: /* Jumped to when encounters something that shouldn't be in
11331 /* If 'maybe_exact' is still set here, means there are no
11332 * code points in the node that participate in folds */
11333 if (FOLD && maybe_exact) {
11337 /* I (khw) don't know if you can get here with zero length, but the
11338 * old code handled this situation by creating a zero-length EXACT
11339 * node. Might as well be NOTHING instead */
11344 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
11347 RExC_parse = p - 1;
11348 Set_Node_Cur_Length(ret); /* MJD */
11349 nextchar(pRExC_state);
11351 /* len is STRLEN which is unsigned, need to copy to signed */
11354 vFAIL("Internal disaster");
11357 } /* End of label 'defchar:' */
11359 } /* End of giant switch on input character */
11365 S_regwhite( RExC_state_t *pRExC_state, char *p )
11367 const char *e = RExC_end;
11369 PERL_ARGS_ASSERT_REGWHITE;
11374 else if (*p == '#') {
11377 if (*p++ == '\n') {
11383 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11392 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
11394 /* Returns the next non-pattern-white space, non-comment character (the
11395 * latter only if 'recognize_comment is true) in the string p, which is
11396 * ended by RExC_end. If there is no line break ending a comment,
11397 * RExC_seen has added the REG_SEEN_RUN_ON_COMMENT flag; */
11398 const char *e = RExC_end;
11400 PERL_ARGS_ASSERT_REGPATWS;
11404 if ((len = is_PATWS_safe(p, e, UTF))) {
11407 else if (recognize_comment && *p == '#') {
11411 if (is_LNBREAK_safe(p, e, UTF)) {
11417 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11425 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
11426 Character classes ([:foo:]) can also be negated ([:^foo:]).
11427 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
11428 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
11429 but trigger failures because they are currently unimplemented. */
11431 #define POSIXCC_DONE(c) ((c) == ':')
11432 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
11433 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
11435 PERL_STATIC_INLINE I32
11436 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
11439 I32 namedclass = OOB_NAMEDCLASS;
11441 PERL_ARGS_ASSERT_REGPPOSIXCC;
11443 if (value == '[' && RExC_parse + 1 < RExC_end &&
11444 /* I smell either [: or [= or [. -- POSIX has been here, right? */
11445 POSIXCC(UCHARAT(RExC_parse)))
11447 const char c = UCHARAT(RExC_parse);
11448 char* const s = RExC_parse++;
11450 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
11452 if (RExC_parse == RExC_end) {
11455 /* Try to give a better location for the error (than the end of
11456 * the string) by looking for the matching ']' */
11458 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
11461 vFAIL2("Unmatched '%c' in POSIX class", c);
11463 /* Grandfather lone [:, [=, [. */
11467 const char* const t = RExC_parse++; /* skip over the c */
11470 if (UCHARAT(RExC_parse) == ']') {
11471 const char *posixcc = s + 1;
11472 RExC_parse++; /* skip over the ending ] */
11475 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
11476 const I32 skip = t - posixcc;
11478 /* Initially switch on the length of the name. */
11481 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
11482 this is the Perl \w
11484 namedclass = ANYOF_WORDCHAR;
11487 /* Names all of length 5. */
11488 /* alnum alpha ascii blank cntrl digit graph lower
11489 print punct space upper */
11490 /* Offset 4 gives the best switch position. */
11491 switch (posixcc[4]) {
11493 if (memEQ(posixcc, "alph", 4)) /* alpha */
11494 namedclass = ANYOF_ALPHA;
11497 if (memEQ(posixcc, "spac", 4)) /* space */
11498 namedclass = ANYOF_PSXSPC;
11501 if (memEQ(posixcc, "grap", 4)) /* graph */
11502 namedclass = ANYOF_GRAPH;
11505 if (memEQ(posixcc, "asci", 4)) /* ascii */
11506 namedclass = ANYOF_ASCII;
11509 if (memEQ(posixcc, "blan", 4)) /* blank */
11510 namedclass = ANYOF_BLANK;
11513 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
11514 namedclass = ANYOF_CNTRL;
11517 if (memEQ(posixcc, "alnu", 4)) /* alnum */
11518 namedclass = ANYOF_ALPHANUMERIC;
11521 if (memEQ(posixcc, "lowe", 4)) /* lower */
11522 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
11523 else if (memEQ(posixcc, "uppe", 4)) /* upper */
11524 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
11527 if (memEQ(posixcc, "digi", 4)) /* digit */
11528 namedclass = ANYOF_DIGIT;
11529 else if (memEQ(posixcc, "prin", 4)) /* print */
11530 namedclass = ANYOF_PRINT;
11531 else if (memEQ(posixcc, "punc", 4)) /* punct */
11532 namedclass = ANYOF_PUNCT;
11537 if (memEQ(posixcc, "xdigit", 6))
11538 namedclass = ANYOF_XDIGIT;
11542 if (namedclass == OOB_NAMEDCLASS)
11543 Simple_vFAIL3("POSIX class [:%.*s:] unknown",
11546 /* The #defines are structured so each complement is +1 to
11547 * the normal one */
11551 assert (posixcc[skip] == ':');
11552 assert (posixcc[skip+1] == ']');
11553 } else if (!SIZE_ONLY) {
11554 /* [[=foo=]] and [[.foo.]] are still future. */
11556 /* adjust RExC_parse so the warning shows after
11557 the class closes */
11558 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
11560 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11563 /* Maternal grandfather:
11564 * "[:" ending in ":" but not in ":]" */
11566 vFAIL("Unmatched '[' in POSIX class");
11569 /* Grandfather lone [:, [=, [. */
11579 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
11581 /* This applies some heuristics at the current parse position (which should
11582 * be at a '[') to see if what follows might be intended to be a [:posix:]
11583 * class. It returns true if it really is a posix class, of course, but it
11584 * also can return true if it thinks that what was intended was a posix
11585 * class that didn't quite make it.
11587 * It will return true for
11589 * [:alphanumerics] (as long as the ] isn't followed immediately by a
11590 * ')' indicating the end of the (?[
11591 * [:any garbage including %^&$ punctuation:]
11593 * This is designed to be called only from S_handle_regex_sets; it could be
11594 * easily adapted to be called from the spot at the beginning of regclass()
11595 * that checks to see in a normal bracketed class if the surrounding []
11596 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
11597 * change long-standing behavior, so I (khw) didn't do that */
11598 char* p = RExC_parse + 1;
11599 char first_char = *p;
11601 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
11603 assert(*(p - 1) == '[');
11605 if (! POSIXCC(first_char)) {
11610 while (p < RExC_end && isWORDCHAR(*p)) p++;
11612 if (p >= RExC_end) {
11616 if (p - RExC_parse > 2 /* Got at least 1 word character */
11617 && (*p == first_char
11618 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
11623 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
11626 && p - RExC_parse > 2 /* [:] evaluates to colon;
11627 [::] is a bad posix class. */
11628 && first_char == *(p - 1));
11632 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist, I32 *flagp, U32 depth,
11633 char * const oregcomp_parse)
11635 /* Handle the (?[...]) construct to do set operations */
11638 UV start, end; /* End points of code point ranges */
11640 char *save_end, *save_parse;
11645 const bool save_fold = FOLD;
11647 GET_RE_DEBUG_FLAGS_DECL;
11649 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
11652 vFAIL("(?[...]) not valid in locale");
11654 RExC_uni_semantics = 1;
11656 /* This will return only an ANYOF regnode, or (unlikely) something smaller
11657 * (such as EXACT). Thus we can skip most everything if just sizing. We
11658 * call regclass to handle '[]' so as to not have to reinvent its parsing
11659 * rules here (throwing away the size it computes each time). And, we exit
11660 * upon an unescaped ']' that isn't one ending a regclass. To do both
11661 * these things, we need to realize that something preceded by a backslash
11662 * is escaped, so we have to keep track of backslashes */
11665 Perl_ck_warner_d(aTHX_
11666 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
11667 "The regex_sets feature is experimental" REPORT_LOCATION,
11668 (int) (RExC_parse - RExC_precomp) , RExC_precomp, RExC_parse);
11670 while (RExC_parse < RExC_end) {
11671 SV* current = NULL;
11672 RExC_parse = regpatws(pRExC_state, RExC_parse,
11673 TRUE); /* means recognize comments */
11674 switch (*RExC_parse) {
11678 /* Skip the next byte (which could cause us to end up in
11679 * the middle of a UTF-8 character, but since none of those
11680 * are confusable with anything we currently handle in this
11681 * switch (invariants all), it's safe. We'll just hit the
11682 * default: case next time and keep on incrementing until
11683 * we find one of the invariants we do handle. */
11688 /* If this looks like it is a [:posix:] class, leave the
11689 * parse pointer at the '[' to fool regclass() into
11690 * thinking it is part of a '[[:posix:]]'. That function
11691 * will use strict checking to force a syntax error if it
11692 * doesn't work out to a legitimate class */
11693 bool is_posix_class
11694 = could_it_be_a_POSIX_class(pRExC_state);
11695 if (! is_posix_class) {
11699 /* regclass() can only return RESTART_UTF8 if multi-char
11700 folds are allowed. */
11701 if (!regclass(pRExC_state, flagp,depth+1,
11702 is_posix_class, /* parse the whole char
11703 class only if not a
11705 FALSE, /* don't allow multi-char folds */
11706 TRUE, /* silence non-portable warnings. */
11708 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11711 /* function call leaves parse pointing to the ']', except
11712 * if we faked it */
11713 if (is_posix_class) {
11717 SvREFCNT_dec(current); /* In case it returned something */
11723 if (RExC_parse < RExC_end
11724 && *RExC_parse == ')')
11726 node = reganode(pRExC_state, ANYOF, 0);
11727 RExC_size += ANYOF_SKIP;
11728 nextchar(pRExC_state);
11729 Set_Node_Length(node,
11730 RExC_parse - oregcomp_parse + 1); /* MJD */
11739 FAIL("Syntax error in (?[...])");
11742 /* Pass 2 only after this. Everything in this construct is a
11743 * metacharacter. Operands begin with either a '\' (for an escape
11744 * sequence), or a '[' for a bracketed character class. Any other
11745 * character should be an operator, or parenthesis for grouping. Both
11746 * types of operands are handled by calling regclass() to parse them. It
11747 * is called with a parameter to indicate to return the computed inversion
11748 * list. The parsing here is implemented via a stack. Each entry on the
11749 * stack is a single character representing one of the operators, or the
11750 * '('; or else a pointer to an operand inversion list. */
11752 #define IS_OPERAND(a) (! SvIOK(a))
11754 /* The stack starts empty. It is a syntax error if the first thing parsed
11755 * is a binary operator; everything else is pushed on the stack. When an
11756 * operand is parsed, the top of the stack is examined. If it is a binary
11757 * operator, the item before it should be an operand, and both are replaced
11758 * by the result of doing that operation on the new operand and the one on
11759 * the stack. Thus a sequence of binary operands is reduced to a single
11760 * one before the next one is parsed.
11762 * A unary operator may immediately follow a binary in the input, for
11765 * When an operand is parsed and the top of the stack is a unary operator,
11766 * the operation is performed, and then the stack is rechecked to see if
11767 * this new operand is part of a binary operation; if so, it is handled as
11770 * A '(' is simply pushed on the stack; it is valid only if the stack is
11771 * empty, or the top element of the stack is an operator or another '('
11772 * (for which the parenthesized expression will become an operand). By the
11773 * time the corresponding ')' is parsed everything in between should have
11774 * been parsed and evaluated to a single operand (or else is a syntax
11775 * error), and is handled as a regular operand */
11779 while (RExC_parse < RExC_end) {
11780 I32 top_index = av_tindex(stack);
11782 SV* current = NULL;
11784 /* Skip white space */
11785 RExC_parse = regpatws(pRExC_state, RExC_parse,
11786 TRUE); /* means recognize comments */
11787 if (RExC_parse >= RExC_end) {
11788 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
11790 if ((curchar = UCHARAT(RExC_parse)) == ']') {
11797 if (av_tindex(stack) >= 0 /* This makes sure that we can
11798 safely subtract 1 from
11799 RExC_parse in the next clause.
11800 If we have something on the
11801 stack, we have parsed something
11803 && UCHARAT(RExC_parse - 1) == '('
11804 && RExC_parse < RExC_end)
11806 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
11807 * This happens when we have some thing like
11809 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
11811 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
11813 * Here we would be handling the interpolated
11814 * '$thai_or_lao'. We handle this by a recursive call to
11815 * ourselves which returns the inversion list the
11816 * interpolated expression evaluates to. We use the flags
11817 * from the interpolated pattern. */
11818 U32 save_flags = RExC_flags;
11819 const char * const save_parse = ++RExC_parse;
11821 parse_lparen_question_flags(pRExC_state);
11823 if (RExC_parse == save_parse /* Makes sure there was at
11824 least one flag (or this
11825 embedding wasn't compiled)
11827 || RExC_parse >= RExC_end - 4
11828 || UCHARAT(RExC_parse) != ':'
11829 || UCHARAT(++RExC_parse) != '('
11830 || UCHARAT(++RExC_parse) != '?'
11831 || UCHARAT(++RExC_parse) != '[')
11834 /* In combination with the above, this moves the
11835 * pointer to the point just after the first erroneous
11836 * character (or if there are no flags, to where they
11837 * should have been) */
11838 if (RExC_parse >= RExC_end - 4) {
11839 RExC_parse = RExC_end;
11841 else if (RExC_parse != save_parse) {
11842 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11844 vFAIL("Expecting '(?flags:(?[...'");
11847 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
11848 depth+1, oregcomp_parse);
11850 /* Here, 'current' contains the embedded expression's
11851 * inversion list, and RExC_parse points to the trailing
11852 * ']'; the next character should be the ')' which will be
11853 * paired with the '(' that has been put on the stack, so
11854 * the whole embedded expression reduces to '(operand)' */
11857 RExC_flags = save_flags;
11858 goto handle_operand;
11863 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11864 vFAIL("Unexpected character");
11867 /* regclass() can only return RESTART_UTF8 if multi-char
11868 folds are allowed. */
11869 if (!regclass(pRExC_state, flagp,depth+1,
11870 TRUE, /* means parse just the next thing */
11871 FALSE, /* don't allow multi-char folds */
11872 FALSE, /* don't silence non-portable warnings. */
11874 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11876 /* regclass() will return with parsing just the \ sequence,
11877 * leaving the parse pointer at the next thing to parse */
11879 goto handle_operand;
11881 case '[': /* Is a bracketed character class */
11883 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
11885 if (! is_posix_class) {
11889 /* regclass() can only return RESTART_UTF8 if multi-char
11890 folds are allowed. */
11891 if(!regclass(pRExC_state, flagp,depth+1,
11892 is_posix_class, /* parse the whole char class
11893 only if not a posix class */
11894 FALSE, /* don't allow multi-char folds */
11895 FALSE, /* don't silence non-portable warnings. */
11897 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11899 /* function call leaves parse pointing to the ']', except if we
11901 if (is_posix_class) {
11905 goto handle_operand;
11914 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
11915 || ! IS_OPERAND(*top_ptr))
11918 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
11920 av_push(stack, newSVuv(curchar));
11924 av_push(stack, newSVuv(curchar));
11928 if (top_index >= 0) {
11929 top_ptr = av_fetch(stack, top_index, FALSE);
11931 if (IS_OPERAND(*top_ptr)) {
11933 vFAIL("Unexpected '(' with no preceding operator");
11936 av_push(stack, newSVuv(curchar));
11943 || ! (current = av_pop(stack))
11944 || ! IS_OPERAND(current)
11945 || ! (lparen = av_pop(stack))
11946 || IS_OPERAND(lparen)
11947 || SvUV(lparen) != '(')
11950 vFAIL("Unexpected ')'");
11953 SvREFCNT_dec_NN(lparen);
11960 /* Here, we have an operand to process, in 'current' */
11962 if (top_index < 0) { /* Just push if stack is empty */
11963 av_push(stack, current);
11966 SV* top = av_pop(stack);
11967 char current_operator;
11969 if (IS_OPERAND(top)) {
11970 vFAIL("Operand with no preceding operator");
11972 current_operator = (char) SvUV(top);
11973 switch (current_operator) {
11974 case '(': /* Push the '(' back on followed by the new
11976 av_push(stack, top);
11977 av_push(stack, current);
11978 SvREFCNT_inc(top); /* Counters the '_dec' done
11979 just after the 'break', so
11980 it doesn't get wrongly freed
11985 _invlist_invert(current);
11987 /* Unlike binary operators, the top of the stack,
11988 * now that this unary one has been popped off, may
11989 * legally be an operator, and we now have operand
11992 SvREFCNT_dec_NN(top);
11993 goto handle_operand;
11996 _invlist_intersection(av_pop(stack),
11999 av_push(stack, current);
12004 _invlist_union(av_pop(stack), current, ¤t);
12005 av_push(stack, current);
12009 _invlist_subtract(av_pop(stack), current, ¤t);
12010 av_push(stack, current);
12013 case '^': /* The union minus the intersection */
12019 element = av_pop(stack);
12020 _invlist_union(element, current, &u);
12021 _invlist_intersection(element, current, &i);
12022 _invlist_subtract(u, i, ¤t);
12023 av_push(stack, current);
12024 SvREFCNT_dec_NN(i);
12025 SvREFCNT_dec_NN(u);
12026 SvREFCNT_dec_NN(element);
12031 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
12033 SvREFCNT_dec_NN(top);
12037 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12040 if (av_tindex(stack) < 0 /* Was empty */
12041 || ((final = av_pop(stack)) == NULL)
12042 || ! IS_OPERAND(final)
12043 || av_tindex(stack) >= 0) /* More left on stack */
12045 vFAIL("Incomplete expression within '(?[ ])'");
12048 /* Here, 'final' is the resultant inversion list from evaluating the
12049 * expression. Return it if so requested */
12050 if (return_invlist) {
12051 *return_invlist = final;
12055 /* Otherwise generate a resultant node, based on 'final'. regclass() is
12056 * expecting a string of ranges and individual code points */
12057 invlist_iterinit(final);
12058 result_string = newSVpvs("");
12059 while (invlist_iternext(final, &start, &end)) {
12060 if (start == end) {
12061 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
12064 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
12069 save_parse = RExC_parse;
12070 RExC_parse = SvPV(result_string, len);
12071 save_end = RExC_end;
12072 RExC_end = RExC_parse + len;
12074 /* We turn off folding around the call, as the class we have constructed
12075 * already has all folding taken into consideration, and we don't want
12076 * regclass() to add to that */
12077 RExC_flags &= ~RXf_PMf_FOLD;
12078 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
12080 node = regclass(pRExC_state, flagp,depth+1,
12081 FALSE, /* means parse the whole char class */
12082 FALSE, /* don't allow multi-char folds */
12083 TRUE, /* silence non-portable warnings. The above may very
12084 well have generated non-portable code points, but
12085 they're valid on this machine */
12088 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
12091 RExC_flags |= RXf_PMf_FOLD;
12093 RExC_parse = save_parse + 1;
12094 RExC_end = save_end;
12095 SvREFCNT_dec_NN(final);
12096 SvREFCNT_dec_NN(result_string);
12097 SvREFCNT_dec_NN(stack);
12099 nextchar(pRExC_state);
12100 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
12105 /* The names of properties whose definitions are not known at compile time are
12106 * stored in this SV, after a constant heading. So if the length has been
12107 * changed since initialization, then there is a run-time definition. */
12108 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION (SvCUR(listsv) != initial_listsv_len)
12111 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
12112 const bool stop_at_1, /* Just parse the next thing, don't
12113 look for a full character class */
12114 bool allow_multi_folds,
12115 const bool silence_non_portable, /* Don't output warnings
12118 SV** ret_invlist) /* Return an inversion list, not a node */
12120 /* parse a bracketed class specification. Most of these will produce an
12121 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
12122 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
12123 * under /i with multi-character folds: it will be rewritten following the
12124 * paradigm of this example, where the <multi-fold>s are characters which
12125 * fold to multiple character sequences:
12126 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
12127 * gets effectively rewritten as:
12128 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
12129 * reg() gets called (recursively) on the rewritten version, and this
12130 * function will return what it constructs. (Actually the <multi-fold>s
12131 * aren't physically removed from the [abcdefghi], it's just that they are
12132 * ignored in the recursion by means of a flag:
12133 * <RExC_in_multi_char_class>.)
12135 * ANYOF nodes contain a bit map for the first 256 characters, with the
12136 * corresponding bit set if that character is in the list. For characters
12137 * above 255, a range list or swash is used. There are extra bits for \w,
12138 * etc. in locale ANYOFs, as what these match is not determinable at
12141 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
12142 * to be restarted. This can only happen if ret_invlist is non-NULL.
12146 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
12148 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
12151 IV namedclass = OOB_NAMEDCLASS;
12152 char *rangebegin = NULL;
12153 bool need_class = 0;
12155 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
12156 than just initialized. */
12157 SV* properties = NULL; /* Code points that match \p{} \P{} */
12158 SV* posixes = NULL; /* Code points that match classes like, [:word:],
12159 extended beyond the Latin1 range */
12160 UV element_count = 0; /* Number of distinct elements in the class.
12161 Optimizations may be possible if this is tiny */
12162 AV * multi_char_matches = NULL; /* Code points that fold to more than one
12163 character; used under /i */
12165 char * stop_ptr = RExC_end; /* where to stop parsing */
12166 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
12168 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
12170 /* Unicode properties are stored in a swash; this holds the current one
12171 * being parsed. If this swash is the only above-latin1 component of the
12172 * character class, an optimization is to pass it directly on to the
12173 * execution engine. Otherwise, it is set to NULL to indicate that there
12174 * are other things in the class that have to be dealt with at execution
12176 SV* swash = NULL; /* Code points that match \p{} \P{} */
12178 /* Set if a component of this character class is user-defined; just passed
12179 * on to the engine */
12180 bool has_user_defined_property = FALSE;
12182 /* inversion list of code points this node matches only when the target
12183 * string is in UTF-8. (Because is under /d) */
12184 SV* depends_list = NULL;
12186 /* inversion list of code points this node matches. For much of the
12187 * function, it includes only those that match regardless of the utf8ness
12188 * of the target string */
12189 SV* cp_list = NULL;
12192 /* In a range, counts how many 0-2 of the ends of it came from literals,
12193 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
12194 UV literal_endpoint = 0;
12196 bool invert = FALSE; /* Is this class to be complemented */
12198 /* Is there any thing like \W or [:^digit:] that matches above the legal
12199 * Unicode range? */
12200 bool runtime_posix_matches_above_Unicode = FALSE;
12202 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
12203 case we need to change the emitted regop to an EXACT. */
12204 const char * orig_parse = RExC_parse;
12205 const I32 orig_size = RExC_size;
12206 GET_RE_DEBUG_FLAGS_DECL;
12208 PERL_ARGS_ASSERT_REGCLASS;
12210 PERL_UNUSED_ARG(depth);
12213 DEBUG_PARSE("clas");
12215 /* Assume we are going to generate an ANYOF node. */
12216 ret = reganode(pRExC_state, ANYOF, 0);
12219 RExC_size += ANYOF_SKIP;
12220 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
12223 ANYOF_FLAGS(ret) = 0;
12225 RExC_emit += ANYOF_SKIP;
12227 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
12229 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
12230 initial_listsv_len = SvCUR(listsv);
12231 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
12235 RExC_parse = regpatws(pRExC_state, RExC_parse,
12236 FALSE /* means don't recognize comments */);
12239 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
12242 allow_multi_folds = FALSE;
12245 RExC_parse = regpatws(pRExC_state, RExC_parse,
12246 FALSE /* means don't recognize comments */);
12250 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
12251 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
12252 const char *s = RExC_parse;
12253 const char c = *s++;
12255 while (isWORDCHAR(*s))
12257 if (*s && c == *s && s[1] == ']') {
12258 SAVEFREESV(RExC_rx_sv);
12260 "POSIX syntax [%c %c] belongs inside character classes",
12262 (void)ReREFCNT_inc(RExC_rx_sv);
12266 /* If the caller wants us to just parse a single element, accomplish this
12267 * by faking the loop ending condition */
12268 if (stop_at_1 && RExC_end > RExC_parse) {
12269 stop_ptr = RExC_parse + 1;
12272 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
12273 if (UCHARAT(RExC_parse) == ']')
12274 goto charclassloop;
12278 if (RExC_parse >= stop_ptr) {
12283 RExC_parse = regpatws(pRExC_state, RExC_parse,
12284 FALSE /* means don't recognize comments */);
12287 if (UCHARAT(RExC_parse) == ']') {
12293 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
12294 save_value = value;
12295 save_prevvalue = prevvalue;
12298 rangebegin = RExC_parse;
12302 value = utf8n_to_uvchr((U8*)RExC_parse,
12303 RExC_end - RExC_parse,
12304 &numlen, UTF8_ALLOW_DEFAULT);
12305 RExC_parse += numlen;
12308 value = UCHARAT(RExC_parse++);
12311 && RExC_parse < RExC_end
12312 && POSIXCC(UCHARAT(RExC_parse)))
12314 namedclass = regpposixcc(pRExC_state, value, strict);
12316 else if (value == '\\') {
12318 value = utf8n_to_uvchr((U8*)RExC_parse,
12319 RExC_end - RExC_parse,
12320 &numlen, UTF8_ALLOW_DEFAULT);
12321 RExC_parse += numlen;
12324 value = UCHARAT(RExC_parse++);
12326 /* Some compilers cannot handle switching on 64-bit integer
12327 * values, therefore value cannot be an UV. Yes, this will
12328 * be a problem later if we want switch on Unicode.
12329 * A similar issue a little bit later when switching on
12330 * namedclass. --jhi */
12332 /* If the \ is escaping white space when white space is being
12333 * skipped, it means that that white space is wanted literally, and
12334 * is already in 'value'. Otherwise, need to translate the escape
12335 * into what it signifies. */
12336 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
12338 case 'w': namedclass = ANYOF_WORDCHAR; break;
12339 case 'W': namedclass = ANYOF_NWORDCHAR; break;
12340 case 's': namedclass = ANYOF_SPACE; break;
12341 case 'S': namedclass = ANYOF_NSPACE; break;
12342 case 'd': namedclass = ANYOF_DIGIT; break;
12343 case 'D': namedclass = ANYOF_NDIGIT; break;
12344 case 'v': namedclass = ANYOF_VERTWS; break;
12345 case 'V': namedclass = ANYOF_NVERTWS; break;
12346 case 'h': namedclass = ANYOF_HORIZWS; break;
12347 case 'H': namedclass = ANYOF_NHORIZWS; break;
12348 case 'N': /* Handle \N{NAME} in class */
12350 /* We only pay attention to the first char of
12351 multichar strings being returned. I kinda wonder
12352 if this makes sense as it does change the behaviour
12353 from earlier versions, OTOH that behaviour was broken
12355 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
12356 TRUE, /* => charclass */
12359 if (*flagp & RESTART_UTF8)
12360 FAIL("panic: grok_bslash_N set RESTART_UTF8");
12370 /* We will handle any undefined properties ourselves */
12371 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF;
12373 if (RExC_parse >= RExC_end)
12374 vFAIL2("Empty \\%c{}", (U8)value);
12375 if (*RExC_parse == '{') {
12376 const U8 c = (U8)value;
12377 e = strchr(RExC_parse++, '}');
12379 vFAIL2("Missing right brace on \\%c{}", c);
12380 while (isSPACE(UCHARAT(RExC_parse)))
12382 if (e == RExC_parse)
12383 vFAIL2("Empty \\%c{}", c);
12384 n = e - RExC_parse;
12385 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
12396 if (UCHARAT(RExC_parse) == '^') {
12399 /* toggle. (The rhs xor gets the single bit that
12400 * differs between P and p; the other xor inverts just
12402 value ^= 'P' ^ 'p';
12404 while (isSPACE(UCHARAT(RExC_parse))) {
12409 /* Try to get the definition of the property into
12410 * <invlist>. If /i is in effect, the effective property
12411 * will have its name be <__NAME_i>. The design is
12412 * discussed in commit
12413 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
12414 Newx(name, n + sizeof("_i__\n"), char);
12416 sprintf(name, "%s%.*s%s\n",
12417 (FOLD) ? "__" : "",
12423 /* Look up the property name, and get its swash and
12424 * inversion list, if the property is found */
12426 SvREFCNT_dec_NN(swash);
12428 swash = _core_swash_init("utf8", name, &PL_sv_undef,
12431 NULL, /* No inversion list */
12434 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
12436 SvREFCNT_dec_NN(swash);
12440 /* Here didn't find it. It could be a user-defined
12441 * property that will be available at run-time. If we
12442 * accept only compile-time properties, is an error;
12443 * otherwise add it to the list for run-time look up */
12445 RExC_parse = e + 1;
12446 vFAIL3("Property '%.*s' is unknown", (int) n, name);
12448 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s\n",
12449 (value == 'p' ? '+' : '!'),
12451 has_user_defined_property = TRUE;
12453 /* We don't know yet, so have to assume that the
12454 * property could match something in the Latin1 range,
12455 * hence something that isn't utf8. Note that this
12456 * would cause things in <depends_list> to match
12457 * inappropriately, except that any \p{}, including
12458 * this one forces Unicode semantics, which means there
12459 * is <no depends_list> */
12460 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
12464 /* Here, did get the swash and its inversion list. If
12465 * the swash is from a user-defined property, then this
12466 * whole character class should be regarded as such */
12467 has_user_defined_property =
12469 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY);
12471 /* Invert if asking for the complement */
12472 if (value == 'P') {
12473 _invlist_union_complement_2nd(properties,
12477 /* The swash can't be used as-is, because we've
12478 * inverted things; delay removing it to here after
12479 * have copied its invlist above */
12480 SvREFCNT_dec_NN(swash);
12484 _invlist_union(properties, invlist, &properties);
12489 RExC_parse = e + 1;
12490 namedclass = ANYOF_UNIPROP; /* no official name, but it's
12493 /* \p means they want Unicode semantics */
12494 RExC_uni_semantics = 1;
12497 case 'n': value = '\n'; break;
12498 case 'r': value = '\r'; break;
12499 case 't': value = '\t'; break;
12500 case 'f': value = '\f'; break;
12501 case 'b': value = '\b'; break;
12502 case 'e': value = ASCII_TO_NATIVE('\033');break;
12503 case 'a': value = ASCII_TO_NATIVE('\007');break;
12505 RExC_parse--; /* function expects to be pointed at the 'o' */
12507 const char* error_msg;
12508 bool valid = grok_bslash_o(&RExC_parse,
12511 SIZE_ONLY, /* warnings in pass
12514 silence_non_portable,
12520 if (PL_encoding && value < 0x100) {
12521 goto recode_encoding;
12525 RExC_parse--; /* function expects to be pointed at the 'x' */
12527 const char* error_msg;
12528 bool valid = grok_bslash_x(&RExC_parse,
12531 TRUE, /* Output warnings */
12533 silence_non_portable,
12539 if (PL_encoding && value < 0x100)
12540 goto recode_encoding;
12543 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
12545 case '0': case '1': case '2': case '3': case '4':
12546 case '5': case '6': case '7':
12548 /* Take 1-3 octal digits */
12549 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12550 numlen = (strict) ? 4 : 3;
12551 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
12552 RExC_parse += numlen;
12555 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12556 vFAIL("Need exactly 3 octal digits");
12558 else if (! SIZE_ONLY /* like \08, \178 */
12560 && RExC_parse < RExC_end
12561 && isDIGIT(*RExC_parse)
12562 && ckWARN(WARN_REGEXP))
12564 SAVEFREESV(RExC_rx_sv);
12565 reg_warn_non_literal_string(
12567 form_short_octal_warning(RExC_parse, numlen));
12568 (void)ReREFCNT_inc(RExC_rx_sv);
12571 if (PL_encoding && value < 0x100)
12572 goto recode_encoding;
12576 if (! RExC_override_recoding) {
12577 SV* enc = PL_encoding;
12578 value = reg_recode((const char)(U8)value, &enc);
12581 vFAIL("Invalid escape in the specified encoding");
12583 else if (SIZE_ONLY) {
12584 ckWARNreg(RExC_parse,
12585 "Invalid escape in the specified encoding");
12591 /* Allow \_ to not give an error */
12592 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
12594 vFAIL2("Unrecognized escape \\%c in character class",
12598 SAVEFREESV(RExC_rx_sv);
12599 ckWARN2reg(RExC_parse,
12600 "Unrecognized escape \\%c in character class passed through",
12602 (void)ReREFCNT_inc(RExC_rx_sv);
12606 } /* End of switch on char following backslash */
12607 } /* end of handling backslash escape sequences */
12610 literal_endpoint++;
12613 /* Here, we have the current token in 'value' */
12615 /* What matches in a locale is not known until runtime. This includes
12616 * what the Posix classes (like \w, [:space:]) match. Room must be
12617 * reserved (one time per class) to store such classes, either if Perl
12618 * is compiled so that locale nodes always should have this space, or
12619 * if there is such class info to be stored. The space will contain a
12620 * bit for each named class that is to be matched against. This isn't
12621 * needed for \p{} and pseudo-classes, as they are not affected by
12622 * locale, and hence are dealt with separately */
12625 && (ANYOF_LOCALE == ANYOF_CLASS
12626 || (namedclass > OOB_NAMEDCLASS && namedclass < ANYOF_MAX)))
12630 RExC_size += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12633 RExC_emit += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12634 ANYOF_CLASS_ZERO(ret);
12636 ANYOF_FLAGS(ret) |= ANYOF_CLASS;
12639 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
12641 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
12642 * literal, as is the character that began the false range, i.e.
12643 * the 'a' in the examples */
12646 const int w = (RExC_parse >= rangebegin)
12647 ? RExC_parse - rangebegin
12650 vFAIL4("False [] range \"%*.*s\"", w, w, rangebegin);
12653 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
12654 ckWARN4reg(RExC_parse,
12655 "False [] range \"%*.*s\"",
12657 (void)ReREFCNT_inc(RExC_rx_sv);
12658 cp_list = add_cp_to_invlist(cp_list, '-');
12659 cp_list = add_cp_to_invlist(cp_list, prevvalue);
12663 range = 0; /* this was not a true range */
12664 element_count += 2; /* So counts for three values */
12668 U8 classnum = namedclass_to_classnum(namedclass);
12669 if (namedclass >= ANYOF_MAX) { /* If a special class */
12670 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
12672 /* Here, should be \h, \H, \v, or \V. Neither /d nor
12673 * /l make a difference in what these match. There
12674 * would be problems if these characters had folds
12675 * other than themselves, as cp_list is subject to
12677 if (classnum != _CC_VERTSPACE) {
12678 assert( namedclass == ANYOF_HORIZWS
12679 || namedclass == ANYOF_NHORIZWS);
12681 /* It turns out that \h is just a synonym for
12683 classnum = _CC_BLANK;
12686 _invlist_union_maybe_complement_2nd(
12688 PL_XPosix_ptrs[classnum],
12689 cBOOL(namedclass % 2), /* Complement if odd
12690 (NHORIZWS, NVERTWS)
12695 else if (classnum == _CC_ASCII) {
12698 ANYOF_CLASS_SET(ret, namedclass);
12701 #endif /* Not isascii(); just use the hard-coded definition for it */
12702 _invlist_union_maybe_complement_2nd(
12705 cBOOL(namedclass % 2), /* Complement if odd
12709 else { /* Garden variety class */
12711 /* The ascii range inversion list */
12712 SV* ascii_source = PL_Posix_ptrs[classnum];
12714 /* The full Latin1 range inversion list */
12715 SV* l1_source = PL_L1Posix_ptrs[classnum];
12717 /* This code is structured into two major clauses. The
12718 * first is for classes whose complete definitions may not
12719 * already be known. It not, the Latin1 definition
12720 * (guaranteed to already known) is used plus code is
12721 * generated to load the rest at run-time (only if needed).
12722 * If the complete definition is known, it drops down to
12723 * the second clause, where the complete definition is
12726 if (classnum < _FIRST_NON_SWASH_CC) {
12728 /* Here, the class has a swash, which may or not
12729 * already be loaded */
12731 /* The name of the property to use to match the full
12732 * eXtended Unicode range swash for this character
12734 const char *Xname = swash_property_names[classnum];
12736 /* If returning the inversion list, we can't defer
12737 * getting this until runtime */
12738 if (ret_invlist && ! PL_utf8_swash_ptrs[classnum]) {
12739 PL_utf8_swash_ptrs[classnum] =
12740 _core_swash_init("utf8", Xname, &PL_sv_undef,
12743 NULL, /* No inversion list */
12744 NULL /* No flags */
12746 assert(PL_utf8_swash_ptrs[classnum]);
12748 if ( ! PL_utf8_swash_ptrs[classnum]) {
12749 if (namedclass % 2 == 0) { /* A non-complemented
12751 /* If not /a matching, there are code points we
12752 * don't know at compile time. Arrange for the
12753 * unknown matches to be loaded at run-time, if
12755 if (! AT_LEAST_ASCII_RESTRICTED) {
12756 Perl_sv_catpvf(aTHX_ listsv, "+utf8::%s\n",
12759 if (LOC) { /* Under locale, set run-time
12761 ANYOF_CLASS_SET(ret, namedclass);
12764 /* Add the current class's code points to
12765 * the running total */
12766 _invlist_union(posixes,
12767 (AT_LEAST_ASCII_RESTRICTED)
12773 else { /* A complemented class */
12774 if (AT_LEAST_ASCII_RESTRICTED) {
12775 /* Under /a should match everything above
12776 * ASCII, plus the complement of the set's
12778 _invlist_union_complement_2nd(posixes,
12783 /* Arrange for the unknown matches to be
12784 * loaded at run-time, if needed */
12785 Perl_sv_catpvf(aTHX_ listsv, "!utf8::%s\n",
12787 runtime_posix_matches_above_Unicode = TRUE;
12789 ANYOF_CLASS_SET(ret, namedclass);
12793 /* We want to match everything in
12794 * Latin1, except those things that
12795 * l1_source matches */
12796 SV* scratch_list = NULL;
12797 _invlist_subtract(PL_Latin1, l1_source,
12800 /* Add the list from this class to the
12803 posixes = scratch_list;
12806 _invlist_union(posixes,
12809 SvREFCNT_dec_NN(scratch_list);
12811 if (DEPENDS_SEMANTICS) {
12813 |= ANYOF_NON_UTF8_LATIN1_ALL;
12818 goto namedclass_done;
12821 /* Here, there is a swash loaded for the class. If no
12822 * inversion list for it yet, get it */
12823 if (! PL_XPosix_ptrs[classnum]) {
12824 PL_XPosix_ptrs[classnum]
12825 = _swash_to_invlist(PL_utf8_swash_ptrs[classnum]);
12829 /* Here there is an inversion list already loaded for the
12832 if (namedclass % 2 == 0) { /* A non-complemented class,
12833 like ANYOF_PUNCT */
12835 /* For non-locale, just add it to any existing list
12837 _invlist_union(posixes,
12838 (AT_LEAST_ASCII_RESTRICTED)
12840 : PL_XPosix_ptrs[classnum],
12843 else { /* Locale */
12844 SV* scratch_list = NULL;
12846 /* For above Latin1 code points, we use the full
12848 _invlist_intersection(PL_AboveLatin1,
12849 PL_XPosix_ptrs[classnum],
12851 /* And set the output to it, adding instead if
12852 * there already is an output. Checking if
12853 * 'posixes' is NULL first saves an extra clone.
12854 * Its reference count will be decremented at the
12855 * next union, etc, or if this is the only
12856 * instance, at the end of the routine */
12858 posixes = scratch_list;
12861 _invlist_union(posixes, scratch_list, &posixes);
12862 SvREFCNT_dec_NN(scratch_list);
12865 #ifndef HAS_ISBLANK
12866 if (namedclass != ANYOF_BLANK) {
12868 /* Set this class in the node for runtime
12870 ANYOF_CLASS_SET(ret, namedclass);
12871 #ifndef HAS_ISBLANK
12874 /* No isblank(), use the hard-coded ASCII-range
12875 * blanks, adding them to the running total. */
12877 _invlist_union(posixes, ascii_source, &posixes);
12882 else { /* A complemented class, like ANYOF_NPUNCT */
12884 _invlist_union_complement_2nd(
12886 (AT_LEAST_ASCII_RESTRICTED)
12888 : PL_XPosix_ptrs[classnum],
12890 /* Under /d, everything in the upper half of the
12891 * Latin1 range matches this complement */
12892 if (DEPENDS_SEMANTICS) {
12893 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL;
12896 else { /* Locale */
12897 SV* scratch_list = NULL;
12898 _invlist_subtract(PL_AboveLatin1,
12899 PL_XPosix_ptrs[classnum],
12902 posixes = scratch_list;
12905 _invlist_union(posixes, scratch_list, &posixes);
12906 SvREFCNT_dec_NN(scratch_list);
12908 #ifndef HAS_ISBLANK
12909 if (namedclass != ANYOF_NBLANK) {
12911 ANYOF_CLASS_SET(ret, namedclass);
12912 #ifndef HAS_ISBLANK
12915 /* Get the list of all code points in Latin1
12916 * that are not ASCII blanks, and add them to
12917 * the running total */
12918 _invlist_subtract(PL_Latin1, ascii_source,
12920 _invlist_union(posixes, scratch_list, &posixes);
12921 SvREFCNT_dec_NN(scratch_list);
12928 continue; /* Go get next character */
12930 } /* end of namedclass \blah */
12932 /* Here, we have a single value. If 'range' is set, it is the ending
12933 * of a range--check its validity. Later, we will handle each
12934 * individual code point in the range. If 'range' isn't set, this
12935 * could be the beginning of a range, so check for that by looking
12936 * ahead to see if the next real character to be processed is the range
12937 * indicator--the minus sign */
12940 RExC_parse = regpatws(pRExC_state, RExC_parse,
12941 FALSE /* means don't recognize comments */);
12945 if (prevvalue > value) /* b-a */ {
12946 const int w = RExC_parse - rangebegin;
12947 Simple_vFAIL4("Invalid [] range \"%*.*s\"", w, w, rangebegin);
12948 range = 0; /* not a valid range */
12952 prevvalue = value; /* save the beginning of the potential range */
12953 if (! stop_at_1 /* Can't be a range if parsing just one thing */
12954 && *RExC_parse == '-')
12956 char* next_char_ptr = RExC_parse + 1;
12957 if (skip_white) { /* Get the next real char after the '-' */
12958 next_char_ptr = regpatws(pRExC_state,
12960 FALSE); /* means don't recognize
12964 /* If the '-' is at the end of the class (just before the ']',
12965 * it is a literal minus; otherwise it is a range */
12966 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
12967 RExC_parse = next_char_ptr;
12969 /* a bad range like \w-, [:word:]- ? */
12970 if (namedclass > OOB_NAMEDCLASS) {
12971 if (strict || ckWARN(WARN_REGEXP)) {
12973 RExC_parse >= rangebegin ?
12974 RExC_parse - rangebegin : 0;
12976 vFAIL4("False [] range \"%*.*s\"",
12981 "False [] range \"%*.*s\"",
12986 cp_list = add_cp_to_invlist(cp_list, '-');
12990 range = 1; /* yeah, it's a range! */
12991 continue; /* but do it the next time */
12996 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
12999 /* non-Latin1 code point implies unicode semantics. Must be set in
13000 * pass1 so is there for the whole of pass 2 */
13002 RExC_uni_semantics = 1;
13005 /* Ready to process either the single value, or the completed range.
13006 * For single-valued non-inverted ranges, we consider the possibility
13007 * of multi-char folds. (We made a conscious decision to not do this
13008 * for the other cases because it can often lead to non-intuitive
13009 * results. For example, you have the peculiar case that:
13010 * "s s" =~ /^[^\xDF]+$/i => Y
13011 * "ss" =~ /^[^\xDF]+$/i => N
13013 * See [perl #89750] */
13014 if (FOLD && allow_multi_folds && value == prevvalue) {
13015 if (value == LATIN_SMALL_LETTER_SHARP_S
13016 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
13019 /* Here <value> is indeed a multi-char fold. Get what it is */
13021 U8 foldbuf[UTF8_MAXBYTES_CASE];
13024 UV folded = _to_uni_fold_flags(
13029 | ((LOC) ? FOLD_FLAGS_LOCALE
13030 : (ASCII_FOLD_RESTRICTED)
13031 ? FOLD_FLAGS_NOMIX_ASCII
13035 /* Here, <folded> should be the first character of the
13036 * multi-char fold of <value>, with <foldbuf> containing the
13037 * whole thing. But, if this fold is not allowed (because of
13038 * the flags), <fold> will be the same as <value>, and should
13039 * be processed like any other character, so skip the special
13041 if (folded != value) {
13043 /* Skip if we are recursed, currently parsing the class
13044 * again. Otherwise add this character to the list of
13045 * multi-char folds. */
13046 if (! RExC_in_multi_char_class) {
13047 AV** this_array_ptr;
13049 STRLEN cp_count = utf8_length(foldbuf,
13050 foldbuf + foldlen);
13051 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
13053 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
13056 if (! multi_char_matches) {
13057 multi_char_matches = newAV();
13060 /* <multi_char_matches> is actually an array of arrays.
13061 * There will be one or two top-level elements: [2],
13062 * and/or [3]. The [2] element is an array, each
13063 * element thereof is a character which folds to TWO
13064 * characters; [3] is for folds to THREE characters.
13065 * (Unicode guarantees a maximum of 3 characters in any
13066 * fold.) When we rewrite the character class below,
13067 * we will do so such that the longest folds are
13068 * written first, so that it prefers the longest
13069 * matching strings first. This is done even if it
13070 * turns out that any quantifier is non-greedy, out of
13071 * programmer laziness. Tom Christiansen has agreed
13072 * that this is ok. This makes the test for the
13073 * ligature 'ffi' come before the test for 'ff' */
13074 if (av_exists(multi_char_matches, cp_count)) {
13075 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13077 this_array = *this_array_ptr;
13080 this_array = newAV();
13081 av_store(multi_char_matches, cp_count,
13084 av_push(this_array, multi_fold);
13087 /* This element should not be processed further in this
13090 value = save_value;
13091 prevvalue = save_prevvalue;
13097 /* Deal with this element of the class */
13100 cp_list = _add_range_to_invlist(cp_list, prevvalue, value);
13102 SV* this_range = _new_invlist(1);
13103 _append_range_to_invlist(this_range, prevvalue, value);
13105 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
13106 * If this range was specified using something like 'i-j', we want
13107 * to include only the 'i' and the 'j', and not anything in
13108 * between, so exclude non-ASCII, non-alphabetics from it.
13109 * However, if the range was specified with something like
13110 * [\x89-\x91] or [\x89-j], all code points within it should be
13111 * included. literal_endpoint==2 means both ends of the range used
13112 * a literal character, not \x{foo} */
13113 if (literal_endpoint == 2
13114 && (prevvalue >= 'a' && value <= 'z')
13115 || (prevvalue >= 'A' && value <= 'Z'))
13117 _invlist_intersection(this_range, PL_Posix_ptrs[_CC_ALPHA],
13120 _invlist_union(cp_list, this_range, &cp_list);
13121 literal_endpoint = 0;
13125 range = 0; /* this range (if it was one) is done now */
13126 } /* End of loop through all the text within the brackets */
13128 /* If anything in the class expands to more than one character, we have to
13129 * deal with them by building up a substitute parse string, and recursively
13130 * calling reg() on it, instead of proceeding */
13131 if (multi_char_matches) {
13132 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
13135 char *save_end = RExC_end;
13136 char *save_parse = RExC_parse;
13137 bool first_time = TRUE; /* First multi-char occurrence doesn't get
13142 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
13143 because too confusing */
13145 sv_catpv(substitute_parse, "(?:");
13149 /* Look at the longest folds first */
13150 for (cp_count = av_len(multi_char_matches); cp_count > 0; cp_count--) {
13152 if (av_exists(multi_char_matches, cp_count)) {
13153 AV** this_array_ptr;
13156 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13158 while ((this_sequence = av_pop(*this_array_ptr)) !=
13161 if (! first_time) {
13162 sv_catpv(substitute_parse, "|");
13164 first_time = FALSE;
13166 sv_catpv(substitute_parse, SvPVX(this_sequence));
13171 /* If the character class contains anything else besides these
13172 * multi-character folds, have to include it in recursive parsing */
13173 if (element_count) {
13174 sv_catpv(substitute_parse, "|[");
13175 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
13176 sv_catpv(substitute_parse, "]");
13179 sv_catpv(substitute_parse, ")");
13182 /* This is a way to get the parse to skip forward a whole named
13183 * sequence instead of matching the 2nd character when it fails the
13185 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
13189 RExC_parse = SvPV(substitute_parse, len);
13190 RExC_end = RExC_parse + len;
13191 RExC_in_multi_char_class = 1;
13192 RExC_emit = (regnode *)orig_emit;
13194 ret = reg(pRExC_state, 1, ®_flags, depth+1);
13196 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
13198 RExC_parse = save_parse;
13199 RExC_end = save_end;
13200 RExC_in_multi_char_class = 0;
13201 SvREFCNT_dec_NN(multi_char_matches);
13205 /* If the character class contains only a single element, it may be
13206 * optimizable into another node type which is smaller and runs faster.
13207 * Check if this is the case for this class */
13208 if (element_count == 1 && ! ret_invlist) {
13212 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like \w or
13213 [:digit:] or \p{foo} */
13215 /* All named classes are mapped into POSIXish nodes, with its FLAG
13216 * argument giving which class it is */
13217 switch ((I32)namedclass) {
13218 case ANYOF_UNIPROP:
13221 /* These don't depend on the charset modifiers. They always
13222 * match under /u rules */
13223 case ANYOF_NHORIZWS:
13224 case ANYOF_HORIZWS:
13225 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
13228 case ANYOF_NVERTWS:
13233 /* The actual POSIXish node for all the rest depends on the
13234 * charset modifier. The ones in the first set depend only on
13235 * ASCII or, if available on this platform, locale */
13239 op = (LOC) ? POSIXL : POSIXA;
13250 /* under /a could be alpha */
13252 if (ASCII_RESTRICTED) {
13253 namedclass = ANYOF_ALPHA + (namedclass % 2);
13261 /* The rest have more possibilities depending on the charset.
13262 * We take advantage of the enum ordering of the charset
13263 * modifiers to get the exact node type, */
13265 op = POSIXD + get_regex_charset(RExC_flags);
13266 if (op > POSIXA) { /* /aa is same as /a */
13269 #ifndef HAS_ISBLANK
13271 && (namedclass == ANYOF_BLANK
13272 || namedclass == ANYOF_NBLANK))
13279 /* The odd numbered ones are the complements of the
13280 * next-lower even number one */
13281 if (namedclass % 2 == 1) {
13285 arg = namedclass_to_classnum(namedclass);
13289 else if (value == prevvalue) {
13291 /* Here, the class consists of just a single code point */
13294 if (! LOC && value == '\n') {
13295 op = REG_ANY; /* Optimize [^\n] */
13296 *flagp |= HASWIDTH|SIMPLE;
13300 else if (value < 256 || UTF) {
13302 /* Optimize a single value into an EXACTish node, but not if it
13303 * would require converting the pattern to UTF-8. */
13304 op = compute_EXACTish(pRExC_state);
13306 } /* Otherwise is a range */
13307 else if (! LOC) { /* locale could vary these */
13308 if (prevvalue == '0') {
13309 if (value == '9') {
13316 /* Here, we have changed <op> away from its initial value iff we found
13317 * an optimization */
13320 /* Throw away this ANYOF regnode, and emit the calculated one,
13321 * which should correspond to the beginning, not current, state of
13323 const char * cur_parse = RExC_parse;
13324 RExC_parse = (char *)orig_parse;
13328 /* To get locale nodes to not use the full ANYOF size would
13329 * require moving the code above that writes the portions
13330 * of it that aren't in other nodes to after this point.
13331 * e.g. ANYOF_CLASS_SET */
13332 RExC_size = orig_size;
13336 RExC_emit = (regnode *)orig_emit;
13337 if (PL_regkind[op] == POSIXD) {
13339 op += NPOSIXD - POSIXD;
13344 ret = reg_node(pRExC_state, op);
13346 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
13350 *flagp |= HASWIDTH|SIMPLE;
13352 else if (PL_regkind[op] == EXACT) {
13353 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13356 RExC_parse = (char *) cur_parse;
13358 SvREFCNT_dec(posixes);
13359 SvREFCNT_dec(cp_list);
13366 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
13368 /* If folding, we calculate all characters that could fold to or from the
13369 * ones already on the list */
13370 if (FOLD && cp_list) {
13371 UV start, end; /* End points of code point ranges */
13373 SV* fold_intersection = NULL;
13375 /* If the highest code point is within Latin1, we can use the
13376 * compiled-in Alphas list, and not have to go out to disk. This
13377 * yields two false positives, the masculine and feminine ordinal
13378 * indicators, which are weeded out below using the
13379 * IS_IN_SOME_FOLD_L1() macro */
13380 if (invlist_highest(cp_list) < 256) {
13381 _invlist_intersection(PL_L1Posix_ptrs[_CC_ALPHA], cp_list,
13382 &fold_intersection);
13386 /* Here, there are non-Latin1 code points, so we will have to go
13387 * fetch the list of all the characters that participate in folds
13389 if (! PL_utf8_foldable) {
13390 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13391 &PL_sv_undef, 1, 0);
13392 PL_utf8_foldable = _get_swash_invlist(swash);
13393 SvREFCNT_dec_NN(swash);
13396 /* This is a hash that for a particular fold gives all characters
13397 * that are involved in it */
13398 if (! PL_utf8_foldclosures) {
13400 /* If we were unable to find any folds, then we likely won't be
13401 * able to find the closures. So just create an empty list.
13402 * Folding will effectively be restricted to the non-Unicode
13403 * rules hard-coded into Perl. (This case happens legitimately
13404 * during compilation of Perl itself before the Unicode tables
13405 * are generated) */
13406 if (_invlist_len(PL_utf8_foldable) == 0) {
13407 PL_utf8_foldclosures = newHV();
13410 /* If the folds haven't been read in, call a fold function
13412 if (! PL_utf8_tofold) {
13413 U8 dummy[UTF8_MAXBYTES+1];
13415 /* This string is just a short named one above \xff */
13416 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
13417 assert(PL_utf8_tofold); /* Verify that worked */
13419 PL_utf8_foldclosures =
13420 _swash_inversion_hash(PL_utf8_tofold);
13424 /* Only the characters in this class that participate in folds need
13425 * be checked. Get the intersection of this class and all the
13426 * possible characters that are foldable. This can quickly narrow
13427 * down a large class */
13428 _invlist_intersection(PL_utf8_foldable, cp_list,
13429 &fold_intersection);
13432 /* Now look at the foldable characters in this class individually */
13433 invlist_iterinit(fold_intersection);
13434 while (invlist_iternext(fold_intersection, &start, &end)) {
13437 /* Locale folding for Latin1 characters is deferred until runtime */
13438 if (LOC && start < 256) {
13442 /* Look at every character in the range */
13443 for (j = start; j <= end; j++) {
13445 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
13451 /* We have the latin1 folding rules hard-coded here so that
13452 * an innocent-looking character class, like /[ks]/i won't
13453 * have to go out to disk to find the possible matches.
13454 * XXX It would be better to generate these via regen, in
13455 * case a new version of the Unicode standard adds new
13456 * mappings, though that is not really likely, and may be
13457 * caught by the default: case of the switch below. */
13459 if (IS_IN_SOME_FOLD_L1(j)) {
13461 /* ASCII is always matched; non-ASCII is matched only
13462 * under Unicode rules */
13463 if (isASCII(j) || AT_LEAST_UNI_SEMANTICS) {
13465 add_cp_to_invlist(cp_list, PL_fold_latin1[j]);
13469 add_cp_to_invlist(depends_list, PL_fold_latin1[j]);
13473 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
13474 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
13476 /* Certain Latin1 characters have matches outside
13477 * Latin1. To get here, <j> is one of those
13478 * characters. None of these matches is valid for
13479 * ASCII characters under /aa, which is why the 'if'
13480 * just above excludes those. These matches only
13481 * happen when the target string is utf8. The code
13482 * below adds the single fold closures for <j> to the
13483 * inversion list. */
13488 add_cp_to_invlist(cp_list, KELVIN_SIGN);
13492 cp_list = add_cp_to_invlist(cp_list,
13493 LATIN_SMALL_LETTER_LONG_S);
13496 cp_list = add_cp_to_invlist(cp_list,
13497 GREEK_CAPITAL_LETTER_MU);
13498 cp_list = add_cp_to_invlist(cp_list,
13499 GREEK_SMALL_LETTER_MU);
13501 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
13502 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
13504 add_cp_to_invlist(cp_list, ANGSTROM_SIGN);
13506 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
13507 cp_list = add_cp_to_invlist(cp_list,
13508 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
13510 case LATIN_SMALL_LETTER_SHARP_S:
13511 cp_list = add_cp_to_invlist(cp_list,
13512 LATIN_CAPITAL_LETTER_SHARP_S);
13514 case 'F': case 'f':
13515 case 'I': case 'i':
13516 case 'L': case 'l':
13517 case 'T': case 't':
13518 case 'A': case 'a':
13519 case 'H': case 'h':
13520 case 'J': case 'j':
13521 case 'N': case 'n':
13522 case 'W': case 'w':
13523 case 'Y': case 'y':
13524 /* These all are targets of multi-character
13525 * folds from code points that require UTF8 to
13526 * express, so they can't match unless the
13527 * target string is in UTF-8, so no action here
13528 * is necessary, as regexec.c properly handles
13529 * the general case for UTF-8 matching and
13530 * multi-char folds */
13533 /* Use deprecated warning to increase the
13534 * chances of this being output */
13535 ckWARN2regdep(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
13542 /* Here is an above Latin1 character. We don't have the rules
13543 * hard-coded for it. First, get its fold. This is the simple
13544 * fold, as the multi-character folds have been handled earlier
13545 * and separated out */
13546 _to_uni_fold_flags(j, foldbuf, &foldlen,
13548 ? FOLD_FLAGS_LOCALE
13549 : (ASCII_FOLD_RESTRICTED)
13550 ? FOLD_FLAGS_NOMIX_ASCII
13553 /* Single character fold of above Latin1. Add everything in
13554 * its fold closure to the list that this node should match.
13555 * The fold closures data structure is a hash with the keys
13556 * being the UTF-8 of every character that is folded to, like
13557 * 'k', and the values each an array of all code points that
13558 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
13559 * Multi-character folds are not included */
13560 if ((listp = hv_fetch(PL_utf8_foldclosures,
13561 (char *) foldbuf, foldlen, FALSE)))
13563 AV* list = (AV*) *listp;
13565 for (k = 0; k <= av_len(list); k++) {
13566 SV** c_p = av_fetch(list, k, FALSE);
13569 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
13573 /* /aa doesn't allow folds between ASCII and non-; /l
13574 * doesn't allow them between above and below 256 */
13575 if ((ASCII_FOLD_RESTRICTED
13576 && (isASCII(c) != isASCII(j)))
13577 || (LOC && c < 256)) {
13581 /* Folds involving non-ascii Latin1 characters
13582 * under /d are added to a separate list */
13583 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
13585 cp_list = add_cp_to_invlist(cp_list, c);
13588 depends_list = add_cp_to_invlist(depends_list, c);
13594 SvREFCNT_dec_NN(fold_intersection);
13597 /* And combine the result (if any) with any inversion list from posix
13598 * classes. The lists are kept separate up to now because we don't want to
13599 * fold the classes (folding of those is automatically handled by the swash
13600 * fetching code) */
13602 if (! DEPENDS_SEMANTICS) {
13604 _invlist_union(cp_list, posixes, &cp_list);
13605 SvREFCNT_dec_NN(posixes);
13612 /* Under /d, we put into a separate list the Latin1 things that
13613 * match only when the target string is utf8 */
13614 SV* nonascii_but_latin1_properties = NULL;
13615 _invlist_intersection(posixes, PL_Latin1,
13616 &nonascii_but_latin1_properties);
13617 _invlist_subtract(nonascii_but_latin1_properties, PL_ASCII,
13618 &nonascii_but_latin1_properties);
13619 _invlist_subtract(posixes, nonascii_but_latin1_properties,
13622 _invlist_union(cp_list, posixes, &cp_list);
13623 SvREFCNT_dec_NN(posixes);
13629 if (depends_list) {
13630 _invlist_union(depends_list, nonascii_but_latin1_properties,
13632 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
13635 depends_list = nonascii_but_latin1_properties;
13640 /* And combine the result (if any) with any inversion list from properties.
13641 * The lists are kept separate up to now so that we can distinguish the two
13642 * in regards to matching above-Unicode. A run-time warning is generated
13643 * if a Unicode property is matched against a non-Unicode code point. But,
13644 * we allow user-defined properties to match anything, without any warning,
13645 * and we also suppress the warning if there is a portion of the character
13646 * class that isn't a Unicode property, and which matches above Unicode, \W
13647 * or [\x{110000}] for example.
13648 * (Note that in this case, unlike the Posix one above, there is no
13649 * <depends_list>, because having a Unicode property forces Unicode
13652 bool warn_super = ! has_user_defined_property;
13655 /* If it matters to the final outcome, see if a non-property
13656 * component of the class matches above Unicode. If so, the
13657 * warning gets suppressed. This is true even if just a single
13658 * such code point is specified, as though not strictly correct if
13659 * another such code point is matched against, the fact that they
13660 * are using above-Unicode code points indicates they should know
13661 * the issues involved */
13663 bool non_prop_matches_above_Unicode =
13664 runtime_posix_matches_above_Unicode
13665 | (invlist_highest(cp_list) > PERL_UNICODE_MAX);
13667 non_prop_matches_above_Unicode =
13668 ! non_prop_matches_above_Unicode;
13670 warn_super = ! non_prop_matches_above_Unicode;
13673 _invlist_union(properties, cp_list, &cp_list);
13674 SvREFCNT_dec_NN(properties);
13677 cp_list = properties;
13681 OP(ret) = ANYOF_WARN_SUPER;
13685 /* Here, we have calculated what code points should be in the character
13688 * Now we can see about various optimizations. Fold calculation (which we
13689 * did above) needs to take place before inversion. Otherwise /[^k]/i
13690 * would invert to include K, which under /i would match k, which it
13691 * shouldn't. Therefore we can't invert folded locale now, as it won't be
13692 * folded until runtime */
13694 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
13695 * at compile time. Besides not inverting folded locale now, we can't
13696 * invert if there are things such as \w, which aren't known until runtime
13699 && ! (LOC && (FOLD || (ANYOF_FLAGS(ret) & ANYOF_CLASS)))
13701 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13703 _invlist_invert(cp_list);
13705 /* Any swash can't be used as-is, because we've inverted things */
13707 SvREFCNT_dec_NN(swash);
13711 /* Clear the invert flag since have just done it here */
13716 *ret_invlist = cp_list;
13718 /* Discard the generated node */
13720 RExC_size = orig_size;
13723 RExC_emit = orig_emit;
13728 /* If we didn't do folding, it's because some information isn't available
13729 * until runtime; set the run-time fold flag for these. (We don't have to
13730 * worry about properties folding, as that is taken care of by the swash
13734 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
13737 /* Some character classes are equivalent to other nodes. Such nodes take
13738 * up less room and generally fewer operations to execute than ANYOF nodes.
13739 * Above, we checked for and optimized into some such equivalents for
13740 * certain common classes that are easy to test. Getting to this point in
13741 * the code means that the class didn't get optimized there. Since this
13742 * code is only executed in Pass 2, it is too late to save space--it has
13743 * been allocated in Pass 1, and currently isn't given back. But turning
13744 * things into an EXACTish node can allow the optimizer to join it to any
13745 * adjacent such nodes. And if the class is equivalent to things like /./,
13746 * expensive run-time swashes can be avoided. Now that we have more
13747 * complete information, we can find things necessarily missed by the
13748 * earlier code. I (khw) am not sure how much to look for here. It would
13749 * be easy, but perhaps too slow, to check any candidates against all the
13750 * node types they could possibly match using _invlistEQ(). */
13755 && ! (ANYOF_FLAGS(ret) & ANYOF_CLASS)
13756 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13759 U8 op = END; /* The optimzation node-type */
13760 const char * cur_parse= RExC_parse;
13762 invlist_iterinit(cp_list);
13763 if (! invlist_iternext(cp_list, &start, &end)) {
13765 /* Here, the list is empty. This happens, for example, when a
13766 * Unicode property is the only thing in the character class, and
13767 * it doesn't match anything. (perluniprops.pod notes such
13770 *flagp |= HASWIDTH|SIMPLE;
13772 else if (start == end) { /* The range is a single code point */
13773 if (! invlist_iternext(cp_list, &start, &end)
13775 /* Don't do this optimization if it would require changing
13776 * the pattern to UTF-8 */
13777 && (start < 256 || UTF))
13779 /* Here, the list contains a single code point. Can optimize
13780 * into an EXACT node */
13789 /* A locale node under folding with one code point can be
13790 * an EXACTFL, as its fold won't be calculated until
13796 /* Here, we are generally folding, but there is only one
13797 * code point to match. If we have to, we use an EXACT
13798 * node, but it would be better for joining with adjacent
13799 * nodes in the optimization pass if we used the same
13800 * EXACTFish node that any such are likely to be. We can
13801 * do this iff the code point doesn't participate in any
13802 * folds. For example, an EXACTF of a colon is the same as
13803 * an EXACT one, since nothing folds to or from a colon. */
13805 if (IS_IN_SOME_FOLD_L1(value)) {
13810 if (! PL_utf8_foldable) {
13811 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13812 &PL_sv_undef, 1, 0);
13813 PL_utf8_foldable = _get_swash_invlist(swash);
13814 SvREFCNT_dec_NN(swash);
13816 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
13821 /* If we haven't found the node type, above, it means we
13822 * can use the prevailing one */
13824 op = compute_EXACTish(pRExC_state);
13829 else if (start == 0) {
13830 if (end == UV_MAX) {
13832 *flagp |= HASWIDTH|SIMPLE;
13835 else if (end == '\n' - 1
13836 && invlist_iternext(cp_list, &start, &end)
13837 && start == '\n' + 1 && end == UV_MAX)
13840 *flagp |= HASWIDTH|SIMPLE;
13844 invlist_iterfinish(cp_list);
13847 RExC_parse = (char *)orig_parse;
13848 RExC_emit = (regnode *)orig_emit;
13850 ret = reg_node(pRExC_state, op);
13852 RExC_parse = (char *)cur_parse;
13854 if (PL_regkind[op] == EXACT) {
13855 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13858 SvREFCNT_dec_NN(cp_list);
13863 /* Here, <cp_list> contains all the code points we can determine at
13864 * compile time that match under all conditions. Go through it, and
13865 * for things that belong in the bitmap, put them there, and delete from
13866 * <cp_list>. While we are at it, see if everything above 255 is in the
13867 * list, and if so, set a flag to speed up execution */
13868 ANYOF_BITMAP_ZERO(ret);
13871 /* This gets set if we actually need to modify things */
13872 bool change_invlist = FALSE;
13876 /* Start looking through <cp_list> */
13877 invlist_iterinit(cp_list);
13878 while (invlist_iternext(cp_list, &start, &end)) {
13882 if (end == UV_MAX && start <= 256) {
13883 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
13886 /* Quit if are above what we should change */
13891 change_invlist = TRUE;
13893 /* Set all the bits in the range, up to the max that we are doing */
13894 high = (end < 255) ? end : 255;
13895 for (i = start; i <= (int) high; i++) {
13896 if (! ANYOF_BITMAP_TEST(ret, i)) {
13897 ANYOF_BITMAP_SET(ret, i);
13903 invlist_iterfinish(cp_list);
13905 /* Done with loop; remove any code points that are in the bitmap from
13907 if (change_invlist) {
13908 _invlist_subtract(cp_list, PL_Latin1, &cp_list);
13911 /* If have completely emptied it, remove it completely */
13912 if (_invlist_len(cp_list) == 0) {
13913 SvREFCNT_dec_NN(cp_list);
13919 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
13922 /* Here, the bitmap has been populated with all the Latin1 code points that
13923 * always match. Can now add to the overall list those that match only
13924 * when the target string is UTF-8 (<depends_list>). */
13925 if (depends_list) {
13927 _invlist_union(cp_list, depends_list, &cp_list);
13928 SvREFCNT_dec_NN(depends_list);
13931 cp_list = depends_list;
13935 /* If there is a swash and more than one element, we can't use the swash in
13936 * the optimization below. */
13937 if (swash && element_count > 1) {
13938 SvREFCNT_dec_NN(swash);
13943 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13945 ARG_SET(ret, ANYOF_NONBITMAP_EMPTY);
13948 /* av[0] stores the character class description in its textual form:
13949 * used later (regexec.c:Perl_regclass_swash()) to initialize the
13950 * appropriate swash, and is also useful for dumping the regnode.
13951 * av[1] if NULL, is a placeholder to later contain the swash computed
13952 * from av[0]. But if no further computation need be done, the
13953 * swash is stored there now.
13954 * av[2] stores the cp_list inversion list for use in addition or
13955 * instead of av[0]; used only if av[1] is NULL
13956 * av[3] is set if any component of the class is from a user-defined
13957 * property; used only if av[1] is NULL */
13958 AV * const av = newAV();
13961 av_store(av, 0, (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13962 ? SvREFCNT_inc(listsv) : &PL_sv_undef);
13964 av_store(av, 1, swash);
13965 SvREFCNT_dec_NN(cp_list);
13968 av_store(av, 1, NULL);
13970 av_store(av, 2, cp_list);
13971 av_store(av, 3, newSVuv(has_user_defined_property));
13975 rv = newRV_noinc(MUTABLE_SV(av));
13976 n = add_data(pRExC_state, 1, "s");
13977 RExC_rxi->data->data[n] = (void*)rv;
13981 *flagp |= HASWIDTH|SIMPLE;
13984 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
13987 /* reg_skipcomment()
13989 Absorbs an /x style # comments from the input stream.
13990 Returns true if there is more text remaining in the stream.
13991 Will set the REG_SEEN_RUN_ON_COMMENT flag if the comment
13992 terminates the pattern without including a newline.
13994 Note its the callers responsibility to ensure that we are
13995 actually in /x mode
14000 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
14004 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
14006 while (RExC_parse < RExC_end)
14007 if (*RExC_parse++ == '\n') {
14012 /* we ran off the end of the pattern without ending
14013 the comment, so we have to add an \n when wrapping */
14014 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
14022 Advances the parse position, and optionally absorbs
14023 "whitespace" from the inputstream.
14025 Without /x "whitespace" means (?#...) style comments only,
14026 with /x this means (?#...) and # comments and whitespace proper.
14028 Returns the RExC_parse point from BEFORE the scan occurs.
14030 This is the /x friendly way of saying RExC_parse++.
14034 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
14036 char* const retval = RExC_parse++;
14038 PERL_ARGS_ASSERT_NEXTCHAR;
14041 if (RExC_end - RExC_parse >= 3
14042 && *RExC_parse == '('
14043 && RExC_parse[1] == '?'
14044 && RExC_parse[2] == '#')
14046 while (*RExC_parse != ')') {
14047 if (RExC_parse == RExC_end)
14048 FAIL("Sequence (?#... not terminated");
14054 if (RExC_flags & RXf_PMf_EXTENDED) {
14055 if (isSPACE(*RExC_parse)) {
14059 else if (*RExC_parse == '#') {
14060 if ( reg_skipcomment( pRExC_state ) )
14069 - reg_node - emit a node
14071 STATIC regnode * /* Location. */
14072 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
14076 regnode * const ret = RExC_emit;
14077 GET_RE_DEBUG_FLAGS_DECL;
14079 PERL_ARGS_ASSERT_REG_NODE;
14082 SIZE_ALIGN(RExC_size);
14086 if (RExC_emit >= RExC_emit_bound)
14087 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14088 op, RExC_emit, RExC_emit_bound);
14090 NODE_ALIGN_FILL(ret);
14092 FILL_ADVANCE_NODE(ptr, op);
14093 #ifdef RE_TRACK_PATTERN_OFFSETS
14094 if (RExC_offsets) { /* MJD */
14095 MJD_OFFSET_DEBUG(("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
14096 "reg_node", __LINE__,
14098 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
14099 ? "Overwriting end of array!\n" : "OK",
14100 (UV)(RExC_emit - RExC_emit_start),
14101 (UV)(RExC_parse - RExC_start),
14102 (UV)RExC_offsets[0]));
14103 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
14111 - reganode - emit a node with an argument
14113 STATIC regnode * /* Location. */
14114 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
14118 regnode * const ret = RExC_emit;
14119 GET_RE_DEBUG_FLAGS_DECL;
14121 PERL_ARGS_ASSERT_REGANODE;
14124 SIZE_ALIGN(RExC_size);
14129 assert(2==regarglen[op]+1);
14131 Anything larger than this has to allocate the extra amount.
14132 If we changed this to be:
14134 RExC_size += (1 + regarglen[op]);
14136 then it wouldn't matter. Its not clear what side effect
14137 might come from that so its not done so far.
14142 if (RExC_emit >= RExC_emit_bound)
14143 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14144 op, RExC_emit, RExC_emit_bound);
14146 NODE_ALIGN_FILL(ret);
14148 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
14149 #ifdef RE_TRACK_PATTERN_OFFSETS
14150 if (RExC_offsets) { /* MJD */
14151 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14155 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
14156 "Overwriting end of array!\n" : "OK",
14157 (UV)(RExC_emit - RExC_emit_start),
14158 (UV)(RExC_parse - RExC_start),
14159 (UV)RExC_offsets[0]));
14160 Set_Cur_Node_Offset;
14168 - reguni - emit (if appropriate) a Unicode character
14171 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
14175 PERL_ARGS_ASSERT_REGUNI;
14177 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
14181 - reginsert - insert an operator in front of already-emitted operand
14183 * Means relocating the operand.
14186 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
14192 const int offset = regarglen[(U8)op];
14193 const int size = NODE_STEP_REGNODE + offset;
14194 GET_RE_DEBUG_FLAGS_DECL;
14196 PERL_ARGS_ASSERT_REGINSERT;
14197 PERL_UNUSED_ARG(depth);
14198 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
14199 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
14208 if (RExC_open_parens) {
14210 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
14211 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
14212 if ( RExC_open_parens[paren] >= opnd ) {
14213 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
14214 RExC_open_parens[paren] += size;
14216 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
14218 if ( RExC_close_parens[paren] >= opnd ) {
14219 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
14220 RExC_close_parens[paren] += size;
14222 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
14227 while (src > opnd) {
14228 StructCopy(--src, --dst, regnode);
14229 #ifdef RE_TRACK_PATTERN_OFFSETS
14230 if (RExC_offsets) { /* MJD 20010112 */
14231 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
14235 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
14236 ? "Overwriting end of array!\n" : "OK",
14237 (UV)(src - RExC_emit_start),
14238 (UV)(dst - RExC_emit_start),
14239 (UV)RExC_offsets[0]));
14240 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
14241 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
14247 place = opnd; /* Op node, where operand used to be. */
14248 #ifdef RE_TRACK_PATTERN_OFFSETS
14249 if (RExC_offsets) { /* MJD */
14250 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14254 (UV)(place - RExC_emit_start) > RExC_offsets[0]
14255 ? "Overwriting end of array!\n" : "OK",
14256 (UV)(place - RExC_emit_start),
14257 (UV)(RExC_parse - RExC_start),
14258 (UV)RExC_offsets[0]));
14259 Set_Node_Offset(place, RExC_parse);
14260 Set_Node_Length(place, 1);
14263 src = NEXTOPER(place);
14264 FILL_ADVANCE_NODE(place, op);
14265 Zero(src, offset, regnode);
14269 - regtail - set the next-pointer at the end of a node chain of p to val.
14270 - SEE ALSO: regtail_study
14272 /* TODO: All three parms should be const */
14274 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14278 GET_RE_DEBUG_FLAGS_DECL;
14280 PERL_ARGS_ASSERT_REGTAIL;
14282 PERL_UNUSED_ARG(depth);
14288 /* Find last node. */
14291 regnode * const temp = regnext(scan);
14293 SV * const mysv=sv_newmortal();
14294 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
14295 regprop(RExC_rx, mysv, scan);
14296 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
14297 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
14298 (temp == NULL ? "->" : ""),
14299 (temp == NULL ? PL_reg_name[OP(val)] : "")
14307 if (reg_off_by_arg[OP(scan)]) {
14308 ARG_SET(scan, val - scan);
14311 NEXT_OFF(scan) = val - scan;
14317 - regtail_study - set the next-pointer at the end of a node chain of p to val.
14318 - Look for optimizable sequences at the same time.
14319 - currently only looks for EXACT chains.
14321 This is experimental code. The idea is to use this routine to perform
14322 in place optimizations on branches and groups as they are constructed,
14323 with the long term intention of removing optimization from study_chunk so
14324 that it is purely analytical.
14326 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
14327 to control which is which.
14330 /* TODO: All four parms should be const */
14333 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14338 #ifdef EXPERIMENTAL_INPLACESCAN
14341 GET_RE_DEBUG_FLAGS_DECL;
14343 PERL_ARGS_ASSERT_REGTAIL_STUDY;
14349 /* Find last node. */
14353 regnode * const temp = regnext(scan);
14354 #ifdef EXPERIMENTAL_INPLACESCAN
14355 if (PL_regkind[OP(scan)] == EXACT) {
14356 bool has_exactf_sharp_s; /* Unexamined in this routine */
14357 if (join_exact(pRExC_state,scan,&min, &has_exactf_sharp_s, 1,val,depth+1))
14362 switch (OP(scan)) {
14368 case EXACTFU_TRICKYFOLD:
14370 if( exact == PSEUDO )
14372 else if ( exact != OP(scan) )
14381 SV * const mysv=sv_newmortal();
14382 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
14383 regprop(RExC_rx, mysv, scan);
14384 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
14385 SvPV_nolen_const(mysv),
14386 REG_NODE_NUM(scan),
14387 PL_reg_name[exact]);
14394 SV * const mysv_val=sv_newmortal();
14395 DEBUG_PARSE_MSG("");
14396 regprop(RExC_rx, mysv_val, val);
14397 PerlIO_printf(Perl_debug_log, "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
14398 SvPV_nolen_const(mysv_val),
14399 (IV)REG_NODE_NUM(val),
14403 if (reg_off_by_arg[OP(scan)]) {
14404 ARG_SET(scan, val - scan);
14407 NEXT_OFF(scan) = val - scan;
14415 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
14419 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
14425 for (bit=0; bit<32; bit++) {
14426 if (flags & (1<<bit)) {
14427 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
14430 if (!set++ && lead)
14431 PerlIO_printf(Perl_debug_log, "%s",lead);
14432 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
14435 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
14436 if (!set++ && lead) {
14437 PerlIO_printf(Perl_debug_log, "%s",lead);
14440 case REGEX_UNICODE_CHARSET:
14441 PerlIO_printf(Perl_debug_log, "UNICODE");
14443 case REGEX_LOCALE_CHARSET:
14444 PerlIO_printf(Perl_debug_log, "LOCALE");
14446 case REGEX_ASCII_RESTRICTED_CHARSET:
14447 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
14449 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
14450 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
14453 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
14459 PerlIO_printf(Perl_debug_log, "\n");
14461 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14467 Perl_regdump(pTHX_ const regexp *r)
14471 SV * const sv = sv_newmortal();
14472 SV *dsv= sv_newmortal();
14473 RXi_GET_DECL(r,ri);
14474 GET_RE_DEBUG_FLAGS_DECL;
14476 PERL_ARGS_ASSERT_REGDUMP;
14478 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
14480 /* Header fields of interest. */
14481 if (r->anchored_substr) {
14482 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
14483 RE_SV_DUMPLEN(r->anchored_substr), 30);
14484 PerlIO_printf(Perl_debug_log,
14485 "anchored %s%s at %"IVdf" ",
14486 s, RE_SV_TAIL(r->anchored_substr),
14487 (IV)r->anchored_offset);
14488 } else if (r->anchored_utf8) {
14489 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
14490 RE_SV_DUMPLEN(r->anchored_utf8), 30);
14491 PerlIO_printf(Perl_debug_log,
14492 "anchored utf8 %s%s at %"IVdf" ",
14493 s, RE_SV_TAIL(r->anchored_utf8),
14494 (IV)r->anchored_offset);
14496 if (r->float_substr) {
14497 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
14498 RE_SV_DUMPLEN(r->float_substr), 30);
14499 PerlIO_printf(Perl_debug_log,
14500 "floating %s%s at %"IVdf"..%"UVuf" ",
14501 s, RE_SV_TAIL(r->float_substr),
14502 (IV)r->float_min_offset, (UV)r->float_max_offset);
14503 } else if (r->float_utf8) {
14504 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
14505 RE_SV_DUMPLEN(r->float_utf8), 30);
14506 PerlIO_printf(Perl_debug_log,
14507 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
14508 s, RE_SV_TAIL(r->float_utf8),
14509 (IV)r->float_min_offset, (UV)r->float_max_offset);
14511 if (r->check_substr || r->check_utf8)
14512 PerlIO_printf(Perl_debug_log,
14514 (r->check_substr == r->float_substr
14515 && r->check_utf8 == r->float_utf8
14516 ? "(checking floating" : "(checking anchored"));
14517 if (r->extflags & RXf_NOSCAN)
14518 PerlIO_printf(Perl_debug_log, " noscan");
14519 if (r->extflags & RXf_CHECK_ALL)
14520 PerlIO_printf(Perl_debug_log, " isall");
14521 if (r->check_substr || r->check_utf8)
14522 PerlIO_printf(Perl_debug_log, ") ");
14524 if (ri->regstclass) {
14525 regprop(r, sv, ri->regstclass);
14526 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
14528 if (r->extflags & RXf_ANCH) {
14529 PerlIO_printf(Perl_debug_log, "anchored");
14530 if (r->extflags & RXf_ANCH_BOL)
14531 PerlIO_printf(Perl_debug_log, "(BOL)");
14532 if (r->extflags & RXf_ANCH_MBOL)
14533 PerlIO_printf(Perl_debug_log, "(MBOL)");
14534 if (r->extflags & RXf_ANCH_SBOL)
14535 PerlIO_printf(Perl_debug_log, "(SBOL)");
14536 if (r->extflags & RXf_ANCH_GPOS)
14537 PerlIO_printf(Perl_debug_log, "(GPOS)");
14538 PerlIO_putc(Perl_debug_log, ' ');
14540 if (r->extflags & RXf_GPOS_SEEN)
14541 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
14542 if (r->intflags & PREGf_SKIP)
14543 PerlIO_printf(Perl_debug_log, "plus ");
14544 if (r->intflags & PREGf_IMPLICIT)
14545 PerlIO_printf(Perl_debug_log, "implicit ");
14546 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
14547 if (r->extflags & RXf_EVAL_SEEN)
14548 PerlIO_printf(Perl_debug_log, "with eval ");
14549 PerlIO_printf(Perl_debug_log, "\n");
14550 DEBUG_FLAGS_r(regdump_extflags("r->extflags: ",r->extflags));
14552 PERL_ARGS_ASSERT_REGDUMP;
14553 PERL_UNUSED_CONTEXT;
14554 PERL_UNUSED_ARG(r);
14555 #endif /* DEBUGGING */
14559 - regprop - printable representation of opcode
14561 #define EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags) \
14564 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]); \
14565 if (flags & ANYOF_INVERT) \
14566 /*make sure the invert info is in each */ \
14567 sv_catpvs(sv, "^"); \
14573 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
14579 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
14580 static const char * const anyofs[] = {
14581 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
14582 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
14583 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
14584 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
14585 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
14586 || _CC_VERTSPACE != 16
14587 #error Need to adjust order of anyofs[]
14624 RXi_GET_DECL(prog,progi);
14625 GET_RE_DEBUG_FLAGS_DECL;
14627 PERL_ARGS_ASSERT_REGPROP;
14631 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
14632 /* It would be nice to FAIL() here, but this may be called from
14633 regexec.c, and it would be hard to supply pRExC_state. */
14634 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(o), (int)REGNODE_MAX);
14635 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
14637 k = PL_regkind[OP(o)];
14640 sv_catpvs(sv, " ");
14641 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
14642 * is a crude hack but it may be the best for now since
14643 * we have no flag "this EXACTish node was UTF-8"
14645 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
14646 PERL_PV_ESCAPE_UNI_DETECT |
14647 PERL_PV_ESCAPE_NONASCII |
14648 PERL_PV_PRETTY_ELLIPSES |
14649 PERL_PV_PRETTY_LTGT |
14650 PERL_PV_PRETTY_NOCLEAR
14652 } else if (k == TRIE) {
14653 /* print the details of the trie in dumpuntil instead, as
14654 * progi->data isn't available here */
14655 const char op = OP(o);
14656 const U32 n = ARG(o);
14657 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
14658 (reg_ac_data *)progi->data->data[n] :
14660 const reg_trie_data * const trie
14661 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
14663 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
14664 DEBUG_TRIE_COMPILE_r(
14665 Perl_sv_catpvf(aTHX_ sv,
14666 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
14667 (UV)trie->startstate,
14668 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
14669 (UV)trie->wordcount,
14672 (UV)TRIE_CHARCOUNT(trie),
14673 (UV)trie->uniquecharcount
14676 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
14678 int rangestart = -1;
14679 U8* bitmap = IS_ANYOF_TRIE(op) ? (U8*)ANYOF_BITMAP(o) : (U8*)TRIE_BITMAP(trie);
14680 sv_catpvs(sv, "[");
14681 for (i = 0; i <= 256; i++) {
14682 if (i < 256 && BITMAP_TEST(bitmap,i)) {
14683 if (rangestart == -1)
14685 } else if (rangestart != -1) {
14686 if (i <= rangestart + 3)
14687 for (; rangestart < i; rangestart++)
14688 put_byte(sv, rangestart);
14690 put_byte(sv, rangestart);
14691 sv_catpvs(sv, "-");
14692 put_byte(sv, i - 1);
14697 sv_catpvs(sv, "]");
14700 } else if (k == CURLY) {
14701 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
14702 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
14703 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
14705 else if (k == WHILEM && o->flags) /* Ordinal/of */
14706 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
14707 else if (k == REF || k == OPEN || k == CLOSE || k == GROUPP || OP(o)==ACCEPT) {
14708 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
14709 if ( RXp_PAREN_NAMES(prog) ) {
14710 if ( k != REF || (OP(o) < NREF)) {
14711 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
14712 SV **name= av_fetch(list, ARG(o), 0 );
14714 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14717 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
14718 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
14719 I32 *nums=(I32*)SvPVX(sv_dat);
14720 SV **name= av_fetch(list, nums[0], 0 );
14723 for ( n=0; n<SvIVX(sv_dat); n++ ) {
14724 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
14725 (n ? "," : ""), (IV)nums[n]);
14727 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14731 } else if (k == GOSUB)
14732 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o)); /* Paren and offset */
14733 else if (k == VERB) {
14735 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
14736 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
14737 } else if (k == LOGICAL)
14738 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* 2: embedded, otherwise 1 */
14739 else if (k == ANYOF) {
14740 int i, rangestart = -1;
14741 const U8 flags = ANYOF_FLAGS(o);
14745 if (flags & ANYOF_LOCALE)
14746 sv_catpvs(sv, "{loc}");
14747 if (flags & ANYOF_LOC_FOLD)
14748 sv_catpvs(sv, "{i}");
14749 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
14750 if (flags & ANYOF_INVERT)
14751 sv_catpvs(sv, "^");
14753 /* output what the standard cp 0-255 bitmap matches */
14754 for (i = 0; i <= 256; i++) {
14755 if (i < 256 && ANYOF_BITMAP_TEST(o,i)) {
14756 if (rangestart == -1)
14758 } else if (rangestart != -1) {
14759 if (i <= rangestart + 3)
14760 for (; rangestart < i; rangestart++)
14761 put_byte(sv, rangestart);
14763 put_byte(sv, rangestart);
14764 sv_catpvs(sv, "-");
14765 put_byte(sv, i - 1);
14772 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14773 /* output any special charclass tests (used entirely under use locale) */
14774 if (ANYOF_CLASS_TEST_ANY_SET(o))
14775 for (i = 0; i < (int)(sizeof(anyofs)/sizeof(char*)); i++)
14776 if (ANYOF_CLASS_TEST(o,i)) {
14777 sv_catpv(sv, anyofs[i]);
14781 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14783 if (flags & ANYOF_NON_UTF8_LATIN1_ALL) {
14784 sv_catpvs(sv, "{non-utf8-latin1-all}");
14787 /* output information about the unicode matching */
14788 if (flags & ANYOF_UNICODE_ALL)
14789 sv_catpvs(sv, "{unicode_all}");
14790 else if (ANYOF_NONBITMAP(o))
14791 sv_catpvs(sv, "{unicode}");
14792 if (flags & ANYOF_NONBITMAP_NON_UTF8)
14793 sv_catpvs(sv, "{outside bitmap}");
14795 if (ANYOF_NONBITMAP(o)) {
14796 SV *lv; /* Set if there is something outside the bit map */
14797 SV * const sw = regclass_swash(prog, o, FALSE, &lv, NULL);
14798 bool byte_output = FALSE; /* If something in the bitmap has been
14801 if (lv && lv != &PL_sv_undef) {
14803 U8 s[UTF8_MAXBYTES_CASE+1];
14805 for (i = 0; i <= 256; i++) { /* Look at chars in bitmap */
14806 uvchr_to_utf8(s, i);
14809 && ! ANYOF_BITMAP_TEST(o, i) /* Don't duplicate
14813 && swash_fetch(sw, s, TRUE))
14815 if (rangestart == -1)
14817 } else if (rangestart != -1) {
14818 byte_output = TRUE;
14819 if (i <= rangestart + 3)
14820 for (; rangestart < i; rangestart++) {
14821 put_byte(sv, rangestart);
14824 put_byte(sv, rangestart);
14825 sv_catpvs(sv, "-");
14834 char *s = savesvpv(lv);
14835 char * const origs = s;
14837 while (*s && *s != '\n')
14841 const char * const t = ++s;
14844 sv_catpvs(sv, " ");
14850 /* Truncate very long output */
14851 if (s - origs > 256) {
14852 Perl_sv_catpvf(aTHX_ sv,
14854 (int) (s - origs - 1),
14860 else if (*s == '\t') {
14875 SvREFCNT_dec_NN(lv);
14879 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
14881 else if (k == POSIXD || k == NPOSIXD) {
14882 U8 index = FLAGS(o) * 2;
14883 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
14884 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
14887 sv_catpv(sv, anyofs[index]);
14890 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
14891 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
14893 PERL_UNUSED_CONTEXT;
14894 PERL_UNUSED_ARG(sv);
14895 PERL_UNUSED_ARG(o);
14896 PERL_UNUSED_ARG(prog);
14897 #endif /* DEBUGGING */
14901 Perl_re_intuit_string(pTHX_ REGEXP * const r)
14902 { /* Assume that RE_INTUIT is set */
14904 struct regexp *const prog = ReANY(r);
14905 GET_RE_DEBUG_FLAGS_DECL;
14907 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
14908 PERL_UNUSED_CONTEXT;
14912 const char * const s = SvPV_nolen_const(prog->check_substr
14913 ? prog->check_substr : prog->check_utf8);
14915 if (!PL_colorset) reginitcolors();
14916 PerlIO_printf(Perl_debug_log,
14917 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
14919 prog->check_substr ? "" : "utf8 ",
14920 PL_colors[5],PL_colors[0],
14923 (strlen(s) > 60 ? "..." : ""));
14926 return prog->check_substr ? prog->check_substr : prog->check_utf8;
14932 handles refcounting and freeing the perl core regexp structure. When
14933 it is necessary to actually free the structure the first thing it
14934 does is call the 'free' method of the regexp_engine associated to
14935 the regexp, allowing the handling of the void *pprivate; member
14936 first. (This routine is not overridable by extensions, which is why
14937 the extensions free is called first.)
14939 See regdupe and regdupe_internal if you change anything here.
14941 #ifndef PERL_IN_XSUB_RE
14943 Perl_pregfree(pTHX_ REGEXP *r)
14949 Perl_pregfree2(pTHX_ REGEXP *rx)
14952 struct regexp *const r = ReANY(rx);
14953 GET_RE_DEBUG_FLAGS_DECL;
14955 PERL_ARGS_ASSERT_PREGFREE2;
14957 if (r->mother_re) {
14958 ReREFCNT_dec(r->mother_re);
14960 CALLREGFREE_PVT(rx); /* free the private data */
14961 SvREFCNT_dec(RXp_PAREN_NAMES(r));
14962 Safefree(r->xpv_len_u.xpvlenu_pv);
14965 SvREFCNT_dec(r->anchored_substr);
14966 SvREFCNT_dec(r->anchored_utf8);
14967 SvREFCNT_dec(r->float_substr);
14968 SvREFCNT_dec(r->float_utf8);
14969 Safefree(r->substrs);
14971 RX_MATCH_COPY_FREE(rx);
14972 #ifdef PERL_ANY_COW
14973 SvREFCNT_dec(r->saved_copy);
14976 SvREFCNT_dec(r->qr_anoncv);
14977 rx->sv_u.svu_rx = 0;
14982 This is a hacky workaround to the structural issue of match results
14983 being stored in the regexp structure which is in turn stored in
14984 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
14985 could be PL_curpm in multiple contexts, and could require multiple
14986 result sets being associated with the pattern simultaneously, such
14987 as when doing a recursive match with (??{$qr})
14989 The solution is to make a lightweight copy of the regexp structure
14990 when a qr// is returned from the code executed by (??{$qr}) this
14991 lightweight copy doesn't actually own any of its data except for
14992 the starp/end and the actual regexp structure itself.
14998 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
15000 struct regexp *ret;
15001 struct regexp *const r = ReANY(rx);
15002 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
15004 PERL_ARGS_ASSERT_REG_TEMP_COPY;
15007 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
15009 SvOK_off((SV *)ret_x);
15011 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
15012 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
15013 made both spots point to the same regexp body.) */
15014 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
15015 assert(!SvPVX(ret_x));
15016 ret_x->sv_u.svu_rx = temp->sv_any;
15017 temp->sv_any = NULL;
15018 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
15019 SvREFCNT_dec_NN(temp);
15020 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
15021 ing below will not set it. */
15022 SvCUR_set(ret_x, SvCUR(rx));
15025 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
15026 sv_force_normal(sv) is called. */
15028 ret = ReANY(ret_x);
15030 SvFLAGS(ret_x) |= SvUTF8(rx);
15031 /* We share the same string buffer as the original regexp, on which we
15032 hold a reference count, incremented when mother_re is set below.
15033 The string pointer is copied here, being part of the regexp struct.
15035 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
15036 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
15038 const I32 npar = r->nparens+1;
15039 Newx(ret->offs, npar, regexp_paren_pair);
15040 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15043 Newx(ret->substrs, 1, struct reg_substr_data);
15044 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15046 SvREFCNT_inc_void(ret->anchored_substr);
15047 SvREFCNT_inc_void(ret->anchored_utf8);
15048 SvREFCNT_inc_void(ret->float_substr);
15049 SvREFCNT_inc_void(ret->float_utf8);
15051 /* check_substr and check_utf8, if non-NULL, point to either their
15052 anchored or float namesakes, and don't hold a second reference. */
15054 RX_MATCH_COPIED_off(ret_x);
15055 #ifdef PERL_ANY_COW
15056 ret->saved_copy = NULL;
15058 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
15059 SvREFCNT_inc_void(ret->qr_anoncv);
15065 /* regfree_internal()
15067 Free the private data in a regexp. This is overloadable by
15068 extensions. Perl takes care of the regexp structure in pregfree(),
15069 this covers the *pprivate pointer which technically perl doesn't
15070 know about, however of course we have to handle the
15071 regexp_internal structure when no extension is in use.
15073 Note this is called before freeing anything in the regexp
15078 Perl_regfree_internal(pTHX_ REGEXP * const rx)
15081 struct regexp *const r = ReANY(rx);
15082 RXi_GET_DECL(r,ri);
15083 GET_RE_DEBUG_FLAGS_DECL;
15085 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
15091 SV *dsv= sv_newmortal();
15092 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
15093 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
15094 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
15095 PL_colors[4],PL_colors[5],s);
15098 #ifdef RE_TRACK_PATTERN_OFFSETS
15100 Safefree(ri->u.offsets); /* 20010421 MJD */
15102 if (ri->code_blocks) {
15104 for (n = 0; n < ri->num_code_blocks; n++)
15105 SvREFCNT_dec(ri->code_blocks[n].src_regex);
15106 Safefree(ri->code_blocks);
15110 int n = ri->data->count;
15113 /* If you add a ->what type here, update the comment in regcomp.h */
15114 switch (ri->data->what[n]) {
15120 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
15123 Safefree(ri->data->data[n]);
15129 { /* Aho Corasick add-on structure for a trie node.
15130 Used in stclass optimization only */
15132 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
15134 refcount = --aho->refcount;
15137 PerlMemShared_free(aho->states);
15138 PerlMemShared_free(aho->fail);
15139 /* do this last!!!! */
15140 PerlMemShared_free(ri->data->data[n]);
15141 PerlMemShared_free(ri->regstclass);
15147 /* trie structure. */
15149 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
15151 refcount = --trie->refcount;
15154 PerlMemShared_free(trie->charmap);
15155 PerlMemShared_free(trie->states);
15156 PerlMemShared_free(trie->trans);
15158 PerlMemShared_free(trie->bitmap);
15160 PerlMemShared_free(trie->jump);
15161 PerlMemShared_free(trie->wordinfo);
15162 /* do this last!!!! */
15163 PerlMemShared_free(ri->data->data[n]);
15168 Perl_croak(aTHX_ "panic: regfree data code '%c'", ri->data->what[n]);
15171 Safefree(ri->data->what);
15172 Safefree(ri->data);
15178 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
15179 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
15180 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
15183 re_dup - duplicate a regexp.
15185 This routine is expected to clone a given regexp structure. It is only
15186 compiled under USE_ITHREADS.
15188 After all of the core data stored in struct regexp is duplicated
15189 the regexp_engine.dupe method is used to copy any private data
15190 stored in the *pprivate pointer. This allows extensions to handle
15191 any duplication it needs to do.
15193 See pregfree() and regfree_internal() if you change anything here.
15195 #if defined(USE_ITHREADS)
15196 #ifndef PERL_IN_XSUB_RE
15198 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
15202 const struct regexp *r = ReANY(sstr);
15203 struct regexp *ret = ReANY(dstr);
15205 PERL_ARGS_ASSERT_RE_DUP_GUTS;
15207 npar = r->nparens+1;
15208 Newx(ret->offs, npar, regexp_paren_pair);
15209 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15211 if (ret->substrs) {
15212 /* Do it this way to avoid reading from *r after the StructCopy().
15213 That way, if any of the sv_dup_inc()s dislodge *r from the L1
15214 cache, it doesn't matter. */
15215 const bool anchored = r->check_substr
15216 ? r->check_substr == r->anchored_substr
15217 : r->check_utf8 == r->anchored_utf8;
15218 Newx(ret->substrs, 1, struct reg_substr_data);
15219 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15221 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
15222 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
15223 ret->float_substr = sv_dup_inc(ret->float_substr, param);
15224 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
15226 /* check_substr and check_utf8, if non-NULL, point to either their
15227 anchored or float namesakes, and don't hold a second reference. */
15229 if (ret->check_substr) {
15231 assert(r->check_utf8 == r->anchored_utf8);
15232 ret->check_substr = ret->anchored_substr;
15233 ret->check_utf8 = ret->anchored_utf8;
15235 assert(r->check_substr == r->float_substr);
15236 assert(r->check_utf8 == r->float_utf8);
15237 ret->check_substr = ret->float_substr;
15238 ret->check_utf8 = ret->float_utf8;
15240 } else if (ret->check_utf8) {
15242 ret->check_utf8 = ret->anchored_utf8;
15244 ret->check_utf8 = ret->float_utf8;
15249 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
15250 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
15253 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
15255 if (RX_MATCH_COPIED(dstr))
15256 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
15258 ret->subbeg = NULL;
15259 #ifdef PERL_ANY_COW
15260 ret->saved_copy = NULL;
15263 /* Whether mother_re be set or no, we need to copy the string. We
15264 cannot refrain from copying it when the storage points directly to
15265 our mother regexp, because that's
15266 1: a buffer in a different thread
15267 2: something we no longer hold a reference on
15268 so we need to copy it locally. */
15269 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
15270 ret->mother_re = NULL;
15273 #endif /* PERL_IN_XSUB_RE */
15278 This is the internal complement to regdupe() which is used to copy
15279 the structure pointed to by the *pprivate pointer in the regexp.
15280 This is the core version of the extension overridable cloning hook.
15281 The regexp structure being duplicated will be copied by perl prior
15282 to this and will be provided as the regexp *r argument, however
15283 with the /old/ structures pprivate pointer value. Thus this routine
15284 may override any copying normally done by perl.
15286 It returns a pointer to the new regexp_internal structure.
15290 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
15293 struct regexp *const r = ReANY(rx);
15294 regexp_internal *reti;
15296 RXi_GET_DECL(r,ri);
15298 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
15302 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode), char, regexp_internal);
15303 Copy(ri->program, reti->program, len+1, regnode);
15305 reti->num_code_blocks = ri->num_code_blocks;
15306 if (ri->code_blocks) {
15308 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
15309 struct reg_code_block);
15310 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
15311 struct reg_code_block);
15312 for (n = 0; n < ri->num_code_blocks; n++)
15313 reti->code_blocks[n].src_regex = (REGEXP*)
15314 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
15317 reti->code_blocks = NULL;
15319 reti->regstclass = NULL;
15322 struct reg_data *d;
15323 const int count = ri->data->count;
15326 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
15327 char, struct reg_data);
15328 Newx(d->what, count, U8);
15331 for (i = 0; i < count; i++) {
15332 d->what[i] = ri->data->what[i];
15333 switch (d->what[i]) {
15334 /* see also regcomp.h and regfree_internal() */
15335 case 'a': /* actually an AV, but the dup function is identical. */
15339 case 'u': /* actually an HV, but the dup function is identical. */
15340 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
15343 /* This is cheating. */
15344 Newx(d->data[i], 1, struct regnode_charclass_class);
15345 StructCopy(ri->data->data[i], d->data[i],
15346 struct regnode_charclass_class);
15347 reti->regstclass = (regnode*)d->data[i];
15350 /* Trie stclasses are readonly and can thus be shared
15351 * without duplication. We free the stclass in pregfree
15352 * when the corresponding reg_ac_data struct is freed.
15354 reti->regstclass= ri->regstclass;
15358 ((reg_trie_data*)ri->data->data[i])->refcount++;
15363 d->data[i] = ri->data->data[i];
15366 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'", ri->data->what[i]);
15375 reti->name_list_idx = ri->name_list_idx;
15377 #ifdef RE_TRACK_PATTERN_OFFSETS
15378 if (ri->u.offsets) {
15379 Newx(reti->u.offsets, 2*len+1, U32);
15380 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
15383 SetProgLen(reti,len);
15386 return (void*)reti;
15389 #endif /* USE_ITHREADS */
15391 #ifndef PERL_IN_XSUB_RE
15394 - regnext - dig the "next" pointer out of a node
15397 Perl_regnext(pTHX_ regnode *p)
15405 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
15406 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(p), (int)REGNODE_MAX);
15409 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
15418 S_re_croak2(pTHX_ const char* pat1,const char* pat2,...)
15421 STRLEN l1 = strlen(pat1);
15422 STRLEN l2 = strlen(pat2);
15425 const char *message;
15427 PERL_ARGS_ASSERT_RE_CROAK2;
15433 Copy(pat1, buf, l1 , char);
15434 Copy(pat2, buf + l1, l2 , char);
15435 buf[l1 + l2] = '\n';
15436 buf[l1 + l2 + 1] = '\0';
15438 /* ANSI variant takes additional second argument */
15439 va_start(args, pat2);
15443 msv = vmess(buf, &args);
15445 message = SvPV_const(msv,l1);
15448 Copy(message, buf, l1 , char);
15449 buf[l1-1] = '\0'; /* Overwrite \n */
15450 Perl_croak(aTHX_ "%s", buf);
15453 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
15455 #ifndef PERL_IN_XSUB_RE
15457 Perl_save_re_context(pTHX)
15461 struct re_save_state *state;
15463 SAVEVPTR(PL_curcop);
15464 SSGROW(SAVESTACK_ALLOC_FOR_RE_SAVE_STATE + 1);
15466 state = (struct re_save_state *)(PL_savestack + PL_savestack_ix);
15467 PL_savestack_ix += SAVESTACK_ALLOC_FOR_RE_SAVE_STATE;
15468 SSPUSHUV(SAVEt_RE_STATE);
15470 Copy(&PL_reg_state, state, 1, struct re_save_state);
15472 PL_reg_oldsaved = NULL;
15473 PL_reg_oldsavedlen = 0;
15474 PL_reg_oldsavedoffset = 0;
15475 PL_reg_oldsavedcoffset = 0;
15476 PL_reg_maxiter = 0;
15477 PL_reg_leftiter = 0;
15478 PL_reg_poscache = NULL;
15479 PL_reg_poscache_size = 0;
15480 #ifdef PERL_ANY_COW
15484 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
15486 const REGEXP * const rx = PM_GETRE(PL_curpm);
15489 for (i = 1; i <= RX_NPARENS(rx); i++) {
15490 char digits[TYPE_CHARS(long)];
15491 const STRLEN len = my_snprintf(digits, sizeof(digits), "%lu", (long)i);
15492 GV *const *const gvp
15493 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
15496 GV * const gv = *gvp;
15497 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
15509 S_put_byte(pTHX_ SV *sv, int c)
15511 PERL_ARGS_ASSERT_PUT_BYTE;
15513 /* Our definition of isPRINT() ignores locales, so only bytes that are
15514 not part of UTF-8 are considered printable. I assume that the same
15515 holds for UTF-EBCDIC.
15516 Also, code point 255 is not printable in either (it's E0 in EBCDIC,
15517 which Wikipedia says:
15519 EO, or Eight Ones, is an 8-bit EBCDIC character code represented as all
15520 ones (binary 1111 1111, hexadecimal FF). It is similar, but not
15521 identical, to the ASCII delete (DEL) or rubout control character. ...
15522 it is typically mapped to hexadecimal code 9F, in order to provide a
15523 unique character mapping in both directions)
15525 So the old condition can be simplified to !isPRINT(c) */
15528 Perl_sv_catpvf(aTHX_ sv, "\\x%02x", c);
15531 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
15535 const char string = c;
15536 if (c == '-' || c == ']' || c == '\\' || c == '^')
15537 sv_catpvs(sv, "\\");
15538 sv_catpvn(sv, &string, 1);
15543 #define CLEAR_OPTSTART \
15544 if (optstart) STMT_START { \
15545 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
15549 #define DUMPUNTIL(b,e) CLEAR_OPTSTART; node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
15551 STATIC const regnode *
15552 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
15553 const regnode *last, const regnode *plast,
15554 SV* sv, I32 indent, U32 depth)
15557 U8 op = PSEUDO; /* Arbitrary non-END op. */
15558 const regnode *next;
15559 const regnode *optstart= NULL;
15561 RXi_GET_DECL(r,ri);
15562 GET_RE_DEBUG_FLAGS_DECL;
15564 PERL_ARGS_ASSERT_DUMPUNTIL;
15566 #ifdef DEBUG_DUMPUNTIL
15567 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
15568 last ? last-start : 0,plast ? plast-start : 0);
15571 if (plast && plast < last)
15574 while (PL_regkind[op] != END && (!last || node < last)) {
15575 /* While that wasn't END last time... */
15578 if (op == CLOSE || op == WHILEM)
15580 next = regnext((regnode *)node);
15583 if (OP(node) == OPTIMIZED) {
15584 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
15591 regprop(r, sv, node);
15592 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
15593 (int)(2*indent + 1), "", SvPVX_const(sv));
15595 if (OP(node) != OPTIMIZED) {
15596 if (next == NULL) /* Next ptr. */
15597 PerlIO_printf(Perl_debug_log, " (0)");
15598 else if (PL_regkind[(U8)op] == BRANCH && PL_regkind[OP(next)] != BRANCH )
15599 PerlIO_printf(Perl_debug_log, " (FAIL)");
15601 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
15602 (void)PerlIO_putc(Perl_debug_log, '\n');
15606 if (PL_regkind[(U8)op] == BRANCHJ) {
15609 const regnode *nnode = (OP(next) == LONGJMP
15610 ? regnext((regnode *)next)
15612 if (last && nnode > last)
15614 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
15617 else if (PL_regkind[(U8)op] == BRANCH) {
15619 DUMPUNTIL(NEXTOPER(node), next);
15621 else if ( PL_regkind[(U8)op] == TRIE ) {
15622 const regnode *this_trie = node;
15623 const char op = OP(node);
15624 const U32 n = ARG(node);
15625 const reg_ac_data * const ac = op>=AHOCORASICK ?
15626 (reg_ac_data *)ri->data->data[n] :
15628 const reg_trie_data * const trie =
15629 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
15631 AV *const trie_words = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
15633 const regnode *nextbranch= NULL;
15636 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
15637 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
15639 PerlIO_printf(Perl_debug_log, "%*s%s ",
15640 (int)(2*(indent+3)), "",
15641 elem_ptr ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr), SvCUR(*elem_ptr), 60,
15642 PL_colors[0], PL_colors[1],
15643 (SvUTF8(*elem_ptr) ? PERL_PV_ESCAPE_UNI : 0) |
15644 PERL_PV_PRETTY_ELLIPSES |
15645 PERL_PV_PRETTY_LTGT
15650 U16 dist= trie->jump[word_idx+1];
15651 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
15652 (UV)((dist ? this_trie + dist : next) - start));
15655 nextbranch= this_trie + trie->jump[0];
15656 DUMPUNTIL(this_trie + dist, nextbranch);
15658 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
15659 nextbranch= regnext((regnode *)nextbranch);
15661 PerlIO_printf(Perl_debug_log, "\n");
15664 if (last && next > last)
15669 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
15670 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
15671 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
15673 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
15675 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
15677 else if ( op == PLUS || op == STAR) {
15678 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
15680 else if (PL_regkind[(U8)op] == ANYOF) {
15681 /* arglen 1 + class block */
15682 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_CLASS)
15683 ? ANYOF_CLASS_SKIP : ANYOF_SKIP);
15684 node = NEXTOPER(node);
15686 else if (PL_regkind[(U8)op] == EXACT) {
15687 /* Literal string, where present. */
15688 node += NODE_SZ_STR(node) - 1;
15689 node = NEXTOPER(node);
15692 node = NEXTOPER(node);
15693 node += regarglen[(U8)op];
15695 if (op == CURLYX || op == OPEN)
15699 #ifdef DEBUG_DUMPUNTIL
15700 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
15705 #endif /* DEBUGGING */
15709 * c-indentation-style: bsd
15710 * c-basic-offset: 4
15711 * indent-tabs-mode: nil
15714 * ex: set ts=8 sts=4 sw=4 et: