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; if = &emit_dummy,
130 implies compiling, so don't emit */
131 regnode emit_dummy; /* placeholder for emit to point to */
132 I32 naughty; /* How bad is this pattern? */
133 I32 sawback; /* Did we see \1, ...? */
135 I32 size; /* Code size. */
136 I32 npar; /* Capture buffer count, (OPEN). */
137 I32 cpar; /* Capture buffer count, (CLOSE). */
138 I32 nestroot; /* root parens we are in - used by accept */
141 regnode **open_parens; /* pointers to open parens */
142 regnode **close_parens; /* pointers to close parens */
143 regnode *opend; /* END node in program */
144 I32 utf8; /* whether the pattern is utf8 or not */
145 I32 orig_utf8; /* whether the pattern was originally in utf8 */
146 /* XXX use this for future optimisation of case
147 * where pattern must be upgraded to utf8. */
148 I32 uni_semantics; /* If a d charset modifier should use unicode
149 rules, even if the pattern is not in
151 HV *paren_names; /* Paren names */
153 regnode **recurse; /* Recurse regops */
154 I32 recurse_count; /* Number of recurse regops */
157 I32 override_recoding;
158 I32 in_multi_char_class;
159 struct reg_code_block *code_blocks; /* positions of literal (?{})
161 int num_code_blocks; /* size of code_blocks[] */
162 int code_index; /* next code_blocks[] slot */
164 char *starttry; /* -Dr: where regtry was called. */
165 #define RExC_starttry (pRExC_state->starttry)
167 SV *runtime_code_qr; /* qr with the runtime code blocks */
169 const char *lastparse;
171 AV *paren_name_list; /* idx -> name */
172 #define RExC_lastparse (pRExC_state->lastparse)
173 #define RExC_lastnum (pRExC_state->lastnum)
174 #define RExC_paren_name_list (pRExC_state->paren_name_list)
178 #define RExC_flags (pRExC_state->flags)
179 #define RExC_pm_flags (pRExC_state->pm_flags)
180 #define RExC_precomp (pRExC_state->precomp)
181 #define RExC_rx_sv (pRExC_state->rx_sv)
182 #define RExC_rx (pRExC_state->rx)
183 #define RExC_rxi (pRExC_state->rxi)
184 #define RExC_start (pRExC_state->start)
185 #define RExC_end (pRExC_state->end)
186 #define RExC_parse (pRExC_state->parse)
187 #define RExC_whilem_seen (pRExC_state->whilem_seen)
188 #ifdef RE_TRACK_PATTERN_OFFSETS
189 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
191 #define RExC_emit (pRExC_state->emit)
192 #define RExC_emit_dummy (pRExC_state->emit_dummy)
193 #define RExC_emit_start (pRExC_state->emit_start)
194 #define RExC_emit_bound (pRExC_state->emit_bound)
195 #define RExC_naughty (pRExC_state->naughty)
196 #define RExC_sawback (pRExC_state->sawback)
197 #define RExC_seen (pRExC_state->seen)
198 #define RExC_size (pRExC_state->size)
199 #define RExC_npar (pRExC_state->npar)
200 #define RExC_nestroot (pRExC_state->nestroot)
201 #define RExC_extralen (pRExC_state->extralen)
202 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
203 #define RExC_utf8 (pRExC_state->utf8)
204 #define RExC_uni_semantics (pRExC_state->uni_semantics)
205 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
206 #define RExC_open_parens (pRExC_state->open_parens)
207 #define RExC_close_parens (pRExC_state->close_parens)
208 #define RExC_opend (pRExC_state->opend)
209 #define RExC_paren_names (pRExC_state->paren_names)
210 #define RExC_recurse (pRExC_state->recurse)
211 #define RExC_recurse_count (pRExC_state->recurse_count)
212 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
213 #define RExC_contains_locale (pRExC_state->contains_locale)
214 #define RExC_override_recoding (pRExC_state->override_recoding)
215 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
218 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
219 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
220 ((*s) == '{' && regcurly(s, FALSE)))
223 #undef SPSTART /* dratted cpp namespace... */
226 * Flags to be passed up and down.
228 #define WORST 0 /* Worst case. */
229 #define HASWIDTH 0x01 /* Known to match non-null strings. */
231 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
232 * character. (There needs to be a case: in the switch statement in regexec.c
233 * for any node marked SIMPLE.) Note that this is not the same thing as
236 #define SPSTART 0x04 /* Starts with * or + */
237 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
238 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
239 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
241 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
243 /* whether trie related optimizations are enabled */
244 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
245 #define TRIE_STUDY_OPT
246 #define FULL_TRIE_STUDY
252 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
253 #define PBITVAL(paren) (1 << ((paren) & 7))
254 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
255 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
256 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
258 #define REQUIRE_UTF8 STMT_START { \
260 *flagp = RESTART_UTF8; \
265 /* This converts the named class defined in regcomp.h to its equivalent class
266 * number defined in handy.h. */
267 #define namedclass_to_classnum(class) ((int) ((class) / 2))
268 #define classnum_to_namedclass(classnum) ((classnum) * 2)
270 /* About scan_data_t.
272 During optimisation we recurse through the regexp program performing
273 various inplace (keyhole style) optimisations. In addition study_chunk
274 and scan_commit populate this data structure with information about
275 what strings MUST appear in the pattern. We look for the longest
276 string that must appear at a fixed location, and we look for the
277 longest string that may appear at a floating location. So for instance
282 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
283 strings (because they follow a .* construct). study_chunk will identify
284 both FOO and BAR as being the longest fixed and floating strings respectively.
286 The strings can be composites, for instance
290 will result in a composite fixed substring 'foo'.
292 For each string some basic information is maintained:
294 - offset or min_offset
295 This is the position the string must appear at, or not before.
296 It also implicitly (when combined with minlenp) tells us how many
297 characters must match before the string we are searching for.
298 Likewise when combined with minlenp and the length of the string it
299 tells us how many characters must appear after the string we have
303 Only used for floating strings. This is the rightmost point that
304 the string can appear at. If set to I32 max it indicates that the
305 string can occur infinitely far to the right.
308 A pointer to the minimum number of characters of the pattern that the
309 string was found inside. This is important as in the case of positive
310 lookahead or positive lookbehind we can have multiple patterns
315 The minimum length of the pattern overall is 3, the minimum length
316 of the lookahead part is 3, but the minimum length of the part that
317 will actually match is 1. So 'FOO's minimum length is 3, but the
318 minimum length for the F is 1. This is important as the minimum length
319 is used to determine offsets in front of and behind the string being
320 looked for. Since strings can be composites this is the length of the
321 pattern at the time it was committed with a scan_commit. Note that
322 the length is calculated by study_chunk, so that the minimum lengths
323 are not known until the full pattern has been compiled, thus the
324 pointer to the value.
328 In the case of lookbehind the string being searched for can be
329 offset past the start point of the final matching string.
330 If this value was just blithely removed from the min_offset it would
331 invalidate some of the calculations for how many chars must match
332 before or after (as they are derived from min_offset and minlen and
333 the length of the string being searched for).
334 When the final pattern is compiled and the data is moved from the
335 scan_data_t structure into the regexp structure the information
336 about lookbehind is factored in, with the information that would
337 have been lost precalculated in the end_shift field for the
340 The fields pos_min and pos_delta are used to store the minimum offset
341 and the delta to the maximum offset at the current point in the pattern.
345 typedef struct scan_data_t {
346 /*I32 len_min; unused */
347 /*I32 len_delta; unused */
351 I32 last_end; /* min value, <0 unless valid. */
354 SV **longest; /* Either &l_fixed, or &l_float. */
355 SV *longest_fixed; /* longest fixed string found in pattern */
356 I32 offset_fixed; /* offset where it starts */
357 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
358 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
359 SV *longest_float; /* longest floating string found in pattern */
360 I32 offset_float_min; /* earliest point in string it can appear */
361 I32 offset_float_max; /* latest point in string it can appear */
362 I32 *minlen_float; /* pointer to the minlen relevant to the string */
363 I32 lookbehind_float; /* is the position of the string modified by LB */
367 struct regnode_charclass_class *start_class;
371 * Forward declarations for pregcomp()'s friends.
374 static const scan_data_t zero_scan_data =
375 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
377 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
378 #define SF_BEFORE_SEOL 0x0001
379 #define SF_BEFORE_MEOL 0x0002
380 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
381 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
384 # define SF_FIX_SHIFT_EOL (0+2)
385 # define SF_FL_SHIFT_EOL (0+4)
387 # define SF_FIX_SHIFT_EOL (+2)
388 # define SF_FL_SHIFT_EOL (+4)
391 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
392 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
394 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
395 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
396 #define SF_IS_INF 0x0040
397 #define SF_HAS_PAR 0x0080
398 #define SF_IN_PAR 0x0100
399 #define SF_HAS_EVAL 0x0200
400 #define SCF_DO_SUBSTR 0x0400
401 #define SCF_DO_STCLASS_AND 0x0800
402 #define SCF_DO_STCLASS_OR 0x1000
403 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
404 #define SCF_WHILEM_VISITED_POS 0x2000
406 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
407 #define SCF_SEEN_ACCEPT 0x8000
408 #define SCF_TRIE_DOING_RESTUDY 0x10000
410 #define UTF cBOOL(RExC_utf8)
412 /* The enums for all these are ordered so things work out correctly */
413 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
414 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
415 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
416 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
417 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
418 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
419 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
421 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
423 #define OOB_NAMEDCLASS -1
425 /* There is no code point that is out-of-bounds, so this is problematic. But
426 * its only current use is to initialize a variable that is always set before
428 #define OOB_UNICODE 0xDEADBEEF
430 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
431 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
434 /* length of regex to show in messages that don't mark a position within */
435 #define RegexLengthToShowInErrorMessages 127
438 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
439 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
440 * op/pragma/warn/regcomp.
442 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
443 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
445 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
448 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
449 * arg. Show regex, up to a maximum length. If it's too long, chop and add
452 #define _FAIL(code) STMT_START { \
453 const char *ellipses = ""; \
454 IV len = RExC_end - RExC_precomp; \
457 SAVEFREESV(RExC_rx_sv); \
458 if (len > RegexLengthToShowInErrorMessages) { \
459 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
460 len = RegexLengthToShowInErrorMessages - 10; \
466 #define FAIL(msg) _FAIL( \
467 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
468 msg, (int)len, RExC_precomp, ellipses))
470 #define FAIL2(msg,arg) _FAIL( \
471 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
472 arg, (int)len, RExC_precomp, ellipses))
475 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
477 #define Simple_vFAIL(m) STMT_START { \
478 const IV offset = RExC_parse - RExC_precomp; \
479 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
480 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
484 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
486 #define vFAIL(m) STMT_START { \
488 SAVEFREESV(RExC_rx_sv); \
493 * Like Simple_vFAIL(), but accepts two arguments.
495 #define Simple_vFAIL2(m,a1) STMT_START { \
496 const IV offset = RExC_parse - RExC_precomp; \
497 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
498 (int)offset, RExC_precomp, RExC_precomp + offset); \
502 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
504 #define vFAIL2(m,a1) STMT_START { \
506 SAVEFREESV(RExC_rx_sv); \
507 Simple_vFAIL2(m, a1); \
512 * Like Simple_vFAIL(), but accepts three arguments.
514 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
515 const IV offset = RExC_parse - RExC_precomp; \
516 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
517 (int)offset, RExC_precomp, RExC_precomp + offset); \
521 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
523 #define vFAIL3(m,a1,a2) STMT_START { \
525 SAVEFREESV(RExC_rx_sv); \
526 Simple_vFAIL3(m, a1, a2); \
530 * Like Simple_vFAIL(), but accepts four arguments.
532 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
533 const IV offset = RExC_parse - RExC_precomp; \
534 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
535 (int)offset, RExC_precomp, RExC_precomp + offset); \
538 #define vFAIL4(m,a1,a2,a3) STMT_START { \
540 SAVEFREESV(RExC_rx_sv); \
541 Simple_vFAIL4(m, a1, a2, a3); \
544 /* m is not necessarily a "literal string", in this macro */
545 #define reg_warn_non_literal_string(loc, m) STMT_START { \
546 const IV offset = loc - RExC_precomp; \
547 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
548 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
551 #define ckWARNreg(loc,m) STMT_START { \
552 const IV offset = loc - RExC_precomp; \
553 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
554 (int)offset, RExC_precomp, RExC_precomp + offset); \
557 #define vWARN_dep(loc, m) STMT_START { \
558 const IV offset = loc - RExC_precomp; \
559 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
560 (int)offset, RExC_precomp, RExC_precomp + offset); \
563 #define ckWARNdep(loc,m) STMT_START { \
564 const IV offset = loc - RExC_precomp; \
565 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
567 (int)offset, RExC_precomp, RExC_precomp + offset); \
570 #define ckWARNregdep(loc,m) STMT_START { \
571 const IV offset = loc - RExC_precomp; \
572 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
574 (int)offset, RExC_precomp, RExC_precomp + offset); \
577 #define ckWARN2regdep(loc,m, a1) STMT_START { \
578 const IV offset = loc - RExC_precomp; \
579 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
581 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
584 #define ckWARN2reg(loc, m, a1) STMT_START { \
585 const IV offset = loc - RExC_precomp; \
586 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
587 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
590 #define vWARN3(loc, m, a1, a2) STMT_START { \
591 const IV offset = loc - RExC_precomp; \
592 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
593 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
596 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
597 const IV offset = loc - RExC_precomp; \
598 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
599 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
602 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
603 const IV offset = loc - RExC_precomp; \
604 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
605 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
608 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
609 const IV offset = loc - RExC_precomp; \
610 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
611 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
614 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
615 const IV offset = loc - RExC_precomp; \
616 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
617 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
621 /* Allow for side effects in s */
622 #define REGC(c,s) STMT_START { \
623 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
626 /* Macros for recording node offsets. 20001227 mjd@plover.com
627 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
628 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
629 * Element 0 holds the number n.
630 * Position is 1 indexed.
632 #ifndef RE_TRACK_PATTERN_OFFSETS
633 #define Set_Node_Offset_To_R(node,byte)
634 #define Set_Node_Offset(node,byte)
635 #define Set_Cur_Node_Offset
636 #define Set_Node_Length_To_R(node,len)
637 #define Set_Node_Length(node,len)
638 #define Set_Node_Cur_Length(node,start)
639 #define Node_Offset(n)
640 #define Node_Length(n)
641 #define Set_Node_Offset_Length(node,offset,len)
642 #define ProgLen(ri) ri->u.proglen
643 #define SetProgLen(ri,x) ri->u.proglen = x
645 #define ProgLen(ri) ri->u.offsets[0]
646 #define SetProgLen(ri,x) ri->u.offsets[0] = x
647 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
649 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
650 __LINE__, (int)(node), (int)(byte))); \
652 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
654 RExC_offsets[2*(node)-1] = (byte); \
659 #define Set_Node_Offset(node,byte) \
660 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
661 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
663 #define Set_Node_Length_To_R(node,len) STMT_START { \
665 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
666 __LINE__, (int)(node), (int)(len))); \
668 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
670 RExC_offsets[2*(node)] = (len); \
675 #define Set_Node_Length(node,len) \
676 Set_Node_Length_To_R((node)-RExC_emit_start, len)
677 #define Set_Node_Cur_Length(node, start) \
678 Set_Node_Length(node, RExC_parse - start)
680 /* Get offsets and lengths */
681 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
682 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
684 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
685 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
686 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
690 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
691 #define EXPERIMENTAL_INPLACESCAN
692 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
694 #define DEBUG_STUDYDATA(str,data,depth) \
695 DEBUG_OPTIMISE_MORE_r(if(data){ \
696 PerlIO_printf(Perl_debug_log, \
697 "%*s" str "Pos:%"IVdf"/%"IVdf \
698 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
699 (int)(depth)*2, "", \
700 (IV)((data)->pos_min), \
701 (IV)((data)->pos_delta), \
702 (UV)((data)->flags), \
703 (IV)((data)->whilem_c), \
704 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
705 is_inf ? "INF " : "" \
707 if ((data)->last_found) \
708 PerlIO_printf(Perl_debug_log, \
709 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
710 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
711 SvPVX_const((data)->last_found), \
712 (IV)((data)->last_end), \
713 (IV)((data)->last_start_min), \
714 (IV)((data)->last_start_max), \
715 ((data)->longest && \
716 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
717 SvPVX_const((data)->longest_fixed), \
718 (IV)((data)->offset_fixed), \
719 ((data)->longest && \
720 (data)->longest==&((data)->longest_float)) ? "*" : "", \
721 SvPVX_const((data)->longest_float), \
722 (IV)((data)->offset_float_min), \
723 (IV)((data)->offset_float_max) \
725 PerlIO_printf(Perl_debug_log,"\n"); \
728 /* Mark that we cannot extend a found fixed substring at this point.
729 Update the longest found anchored substring and the longest found
730 floating substrings if needed. */
733 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
735 const STRLEN l = CHR_SVLEN(data->last_found);
736 const STRLEN old_l = CHR_SVLEN(*data->longest);
737 GET_RE_DEBUG_FLAGS_DECL;
739 PERL_ARGS_ASSERT_SCAN_COMMIT;
741 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
742 SvSetMagicSV(*data->longest, data->last_found);
743 if (*data->longest == data->longest_fixed) {
744 data->offset_fixed = l ? data->last_start_min : data->pos_min;
745 if (data->flags & SF_BEFORE_EOL)
747 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
749 data->flags &= ~SF_FIX_BEFORE_EOL;
750 data->minlen_fixed=minlenp;
751 data->lookbehind_fixed=0;
753 else { /* *data->longest == data->longest_float */
754 data->offset_float_min = l ? data->last_start_min : data->pos_min;
755 data->offset_float_max = (l
756 ? data->last_start_max
757 : (data->pos_delta == I32_MAX ? I32_MAX : data->pos_min + data->pos_delta));
758 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
759 data->offset_float_max = I32_MAX;
760 if (data->flags & SF_BEFORE_EOL)
762 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
764 data->flags &= ~SF_FL_BEFORE_EOL;
765 data->minlen_float=minlenp;
766 data->lookbehind_float=0;
769 SvCUR_set(data->last_found, 0);
771 SV * const sv = data->last_found;
772 if (SvUTF8(sv) && SvMAGICAL(sv)) {
773 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
779 data->flags &= ~SF_BEFORE_EOL;
780 DEBUG_STUDYDATA("commit: ",data,0);
783 /* These macros set, clear and test whether the synthetic start class ('ssc',
784 * given by the parameter) matches an empty string (EOS). This uses the
785 * 'next_off' field in the node, to save a bit in the flags field. The ssc
786 * stands alone, so there is never a next_off, so this field is otherwise
787 * unused. The EOS information is used only for compilation, but theoretically
788 * it could be passed on to the execution code. This could be used to store
789 * more than one bit of information, but only this one is currently used. */
790 #define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
791 #define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
792 #define TEST_SSC_EOS(node) cBOOL((node)->next_off)
794 /* Can match anything (initialization) */
796 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
798 PERL_ARGS_ASSERT_CL_ANYTHING;
800 ANYOF_BITMAP_SETALL(cl);
801 cl->flags = ANYOF_UNICODE_ALL;
804 /* If any portion of the regex is to operate under locale rules,
805 * initialization includes it. The reason this isn't done for all regexes
806 * is that the optimizer was written under the assumption that locale was
807 * all-or-nothing. Given the complexity and lack of documentation in the
808 * optimizer, and that there are inadequate test cases for locale, so many
809 * parts of it may not work properly, it is safest to avoid locale unless
811 if (RExC_contains_locale) {
812 ANYOF_CLASS_SETALL(cl); /* /l uses class */
813 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
816 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
820 /* Can match anything (initialization) */
822 S_cl_is_anything(const struct regnode_charclass_class *cl)
826 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
828 for (value = 0; value < ANYOF_MAX; value += 2)
829 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
831 if (!(cl->flags & ANYOF_UNICODE_ALL))
833 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
838 /* Can match anything (initialization) */
840 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
842 PERL_ARGS_ASSERT_CL_INIT;
844 Zero(cl, 1, struct regnode_charclass_class);
846 cl_anything(pRExC_state, cl);
847 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
850 /* These two functions currently do the exact same thing */
851 #define cl_init_zero S_cl_init
853 /* 'AND' a given class with another one. Can create false positives. 'cl'
854 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
855 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
857 S_cl_and(struct regnode_charclass_class *cl,
858 const struct regnode_charclass_class *and_with)
860 PERL_ARGS_ASSERT_CL_AND;
862 assert(PL_regkind[and_with->type] == ANYOF);
864 /* I (khw) am not sure all these restrictions are necessary XXX */
865 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
866 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
867 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
868 && !(and_with->flags & ANYOF_LOC_FOLD)
869 && !(cl->flags & ANYOF_LOC_FOLD)) {
872 if (and_with->flags & ANYOF_INVERT)
873 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
874 cl->bitmap[i] &= ~and_with->bitmap[i];
876 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
877 cl->bitmap[i] &= and_with->bitmap[i];
878 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
880 if (and_with->flags & ANYOF_INVERT) {
882 /* Here, the and'ed node is inverted. Get the AND of the flags that
883 * aren't affected by the inversion. Those that are affected are
884 * handled individually below */
885 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
886 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
887 cl->flags |= affected_flags;
889 /* We currently don't know how to deal with things that aren't in the
890 * bitmap, but we know that the intersection is no greater than what
891 * is already in cl, so let there be false positives that get sorted
892 * out after the synthetic start class succeeds, and the node is
893 * matched for real. */
895 /* The inversion of these two flags indicate that the resulting
896 * intersection doesn't have them */
897 if (and_with->flags & ANYOF_UNICODE_ALL) {
898 cl->flags &= ~ANYOF_UNICODE_ALL;
900 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
901 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
904 else { /* and'd node is not inverted */
905 U8 outside_bitmap_but_not_utf8; /* Temp variable */
907 if (! ANYOF_NONBITMAP(and_with)) {
909 /* Here 'and_with' doesn't match anything outside the bitmap
910 * (except possibly ANYOF_UNICODE_ALL), which means the
911 * intersection can't either, except for ANYOF_UNICODE_ALL, in
912 * which case we don't know what the intersection is, but it's no
913 * greater than what cl already has, so can just leave it alone,
914 * with possible false positives */
915 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
916 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
917 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
920 else if (! ANYOF_NONBITMAP(cl)) {
922 /* Here, 'and_with' does match something outside the bitmap, and cl
923 * doesn't have a list of things to match outside the bitmap. If
924 * cl can match all code points above 255, the intersection will
925 * be those above-255 code points that 'and_with' matches. If cl
926 * can't match all Unicode code points, it means that it can't
927 * match anything outside the bitmap (since the 'if' that got us
928 * into this block tested for that), so we leave the bitmap empty.
930 if (cl->flags & ANYOF_UNICODE_ALL) {
931 ARG_SET(cl, ARG(and_with));
933 /* and_with's ARG may match things that don't require UTF8.
934 * And now cl's will too, in spite of this being an 'and'. See
935 * the comments below about the kludge */
936 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
940 /* Here, both 'and_with' and cl match something outside the
941 * bitmap. Currently we do not do the intersection, so just match
942 * whatever cl had at the beginning. */
946 /* Take the intersection of the two sets of flags. However, the
947 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
948 * kludge around the fact that this flag is not treated like the others
949 * which are initialized in cl_anything(). The way the optimizer works
950 * is that the synthetic start class (SSC) is initialized to match
951 * anything, and then the first time a real node is encountered, its
952 * values are AND'd with the SSC's with the result being the values of
953 * the real node. However, there are paths through the optimizer where
954 * the AND never gets called, so those initialized bits are set
955 * inappropriately, which is not usually a big deal, as they just cause
956 * false positives in the SSC, which will just mean a probably
957 * imperceptible slow down in execution. However this bit has a
958 * higher false positive consequence in that it can cause utf8.pm,
959 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
960 * bigger slowdown and also causes significant extra memory to be used.
961 * In order to prevent this, the code now takes a different tack. The
962 * bit isn't set unless some part of the regular expression needs it,
963 * but once set it won't get cleared. This means that these extra
964 * modules won't get loaded unless there was some path through the
965 * pattern that would have required them anyway, and so any false
966 * positives that occur by not ANDing them out when they could be
967 * aren't as severe as they would be if we treated this bit like all
969 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
970 & ANYOF_NONBITMAP_NON_UTF8;
971 cl->flags &= and_with->flags;
972 cl->flags |= outside_bitmap_but_not_utf8;
976 /* 'OR' a given class with another one. Can create false positives. 'cl'
977 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
978 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
980 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
982 PERL_ARGS_ASSERT_CL_OR;
984 if (or_with->flags & ANYOF_INVERT) {
986 /* Here, the or'd node is to be inverted. This means we take the
987 * complement of everything not in the bitmap, but currently we don't
988 * know what that is, so give up and match anything */
989 if (ANYOF_NONBITMAP(or_with)) {
990 cl_anything(pRExC_state, cl);
993 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
994 * <= (B1 | !B2) | (CL1 | !CL2)
995 * which is wasteful if CL2 is small, but we ignore CL2:
996 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
997 * XXXX Can we handle case-fold? Unclear:
998 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
999 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
1001 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1002 && !(or_with->flags & ANYOF_LOC_FOLD)
1003 && !(cl->flags & ANYOF_LOC_FOLD) ) {
1006 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1007 cl->bitmap[i] |= ~or_with->bitmap[i];
1008 } /* XXXX: logic is complicated otherwise */
1010 cl_anything(pRExC_state, cl);
1013 /* And, we can just take the union of the flags that aren't affected
1014 * by the inversion */
1015 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1017 /* For the remaining flags:
1018 ANYOF_UNICODE_ALL and inverted means to not match anything above
1019 255, which means that the union with cl should just be
1020 what cl has in it, so can ignore this flag
1021 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1022 is 127-255 to match them, but then invert that, so the
1023 union with cl should just be what cl has in it, so can
1026 } else { /* 'or_with' is not inverted */
1027 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1028 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1029 && (!(or_with->flags & ANYOF_LOC_FOLD)
1030 || (cl->flags & ANYOF_LOC_FOLD)) ) {
1033 /* OR char bitmap and class bitmap separately */
1034 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1035 cl->bitmap[i] |= or_with->bitmap[i];
1036 if (or_with->flags & ANYOF_CLASS) {
1037 ANYOF_CLASS_OR(or_with, cl);
1040 else { /* XXXX: logic is complicated, leave it along for a moment. */
1041 cl_anything(pRExC_state, cl);
1044 if (ANYOF_NONBITMAP(or_with)) {
1046 /* Use the added node's outside-the-bit-map match if there isn't a
1047 * conflict. If there is a conflict (both nodes match something
1048 * outside the bitmap, but what they match outside is not the same
1049 * pointer, and hence not easily compared until XXX we extend
1050 * inversion lists this far), give up and allow the start class to
1051 * match everything outside the bitmap. If that stuff is all above
1052 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1053 if (! ANYOF_NONBITMAP(cl)) {
1054 ARG_SET(cl, ARG(or_with));
1056 else if (ARG(cl) != ARG(or_with)) {
1058 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1059 cl_anything(pRExC_state, cl);
1062 cl->flags |= ANYOF_UNICODE_ALL;
1067 /* Take the union */
1068 cl->flags |= or_with->flags;
1072 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1073 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1074 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1075 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1080 dump_trie(trie,widecharmap,revcharmap)
1081 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1082 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1084 These routines dump out a trie in a somewhat readable format.
1085 The _interim_ variants are used for debugging the interim
1086 tables that are used to generate the final compressed
1087 representation which is what dump_trie expects.
1089 Part of the reason for their existence is to provide a form
1090 of documentation as to how the different representations function.
1095 Dumps the final compressed table form of the trie to Perl_debug_log.
1096 Used for debugging make_trie().
1100 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1101 AV *revcharmap, U32 depth)
1104 SV *sv=sv_newmortal();
1105 int colwidth= widecharmap ? 6 : 4;
1107 GET_RE_DEBUG_FLAGS_DECL;
1109 PERL_ARGS_ASSERT_DUMP_TRIE;
1111 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1112 (int)depth * 2 + 2,"",
1113 "Match","Base","Ofs" );
1115 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1116 SV ** const tmp = av_fetch( revcharmap, state, 0);
1118 PerlIO_printf( Perl_debug_log, "%*s",
1120 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1121 PL_colors[0], PL_colors[1],
1122 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1123 PERL_PV_ESCAPE_FIRSTCHAR
1128 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1129 (int)depth * 2 + 2,"");
1131 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1132 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1133 PerlIO_printf( Perl_debug_log, "\n");
1135 for( state = 1 ; state < trie->statecount ; state++ ) {
1136 const U32 base = trie->states[ state ].trans.base;
1138 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1140 if ( trie->states[ state ].wordnum ) {
1141 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1143 PerlIO_printf( Perl_debug_log, "%6s", "" );
1146 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1151 while( ( base + ofs < trie->uniquecharcount ) ||
1152 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1153 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1156 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1158 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1159 if ( ( base + ofs >= trie->uniquecharcount ) &&
1160 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1161 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1163 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1165 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1167 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1171 PerlIO_printf( Perl_debug_log, "]");
1174 PerlIO_printf( Perl_debug_log, "\n" );
1176 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1177 for (word=1; word <= trie->wordcount; word++) {
1178 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1179 (int)word, (int)(trie->wordinfo[word].prev),
1180 (int)(trie->wordinfo[word].len));
1182 PerlIO_printf(Perl_debug_log, "\n" );
1185 Dumps a fully constructed but uncompressed trie in list form.
1186 List tries normally only are used for construction when the number of
1187 possible chars (trie->uniquecharcount) is very high.
1188 Used for debugging make_trie().
1191 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1192 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1196 SV *sv=sv_newmortal();
1197 int colwidth= widecharmap ? 6 : 4;
1198 GET_RE_DEBUG_FLAGS_DECL;
1200 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1202 /* print out the table precompression. */
1203 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1204 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1205 "------:-----+-----------------\n" );
1207 for( state=1 ; state < next_alloc ; state ++ ) {
1210 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1211 (int)depth * 2 + 2,"", (UV)state );
1212 if ( ! trie->states[ state ].wordnum ) {
1213 PerlIO_printf( Perl_debug_log, "%5s| ","");
1215 PerlIO_printf( Perl_debug_log, "W%4x| ",
1216 trie->states[ state ].wordnum
1219 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1220 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1222 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1224 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1225 PL_colors[0], PL_colors[1],
1226 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1227 PERL_PV_ESCAPE_FIRSTCHAR
1229 TRIE_LIST_ITEM(state,charid).forid,
1230 (UV)TRIE_LIST_ITEM(state,charid).newstate
1233 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1234 (int)((depth * 2) + 14), "");
1237 PerlIO_printf( Perl_debug_log, "\n");
1242 Dumps a fully constructed but uncompressed trie in table form.
1243 This is the normal DFA style state transition table, with a few
1244 twists to facilitate compression later.
1245 Used for debugging make_trie().
1248 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1249 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1254 SV *sv=sv_newmortal();
1255 int colwidth= widecharmap ? 6 : 4;
1256 GET_RE_DEBUG_FLAGS_DECL;
1258 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1261 print out the table precompression so that we can do a visual check
1262 that they are identical.
1265 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1267 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1268 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1270 PerlIO_printf( Perl_debug_log, "%*s",
1272 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1273 PL_colors[0], PL_colors[1],
1274 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1275 PERL_PV_ESCAPE_FIRSTCHAR
1281 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1283 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1284 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1287 PerlIO_printf( Perl_debug_log, "\n" );
1289 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1291 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1292 (int)depth * 2 + 2,"",
1293 (UV)TRIE_NODENUM( state ) );
1295 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1296 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1298 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1300 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1302 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1303 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1305 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1306 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1314 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1315 startbranch: the first branch in the whole branch sequence
1316 first : start branch of sequence of branch-exact nodes.
1317 May be the same as startbranch
1318 last : Thing following the last branch.
1319 May be the same as tail.
1320 tail : item following the branch sequence
1321 count : words in the sequence
1322 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1323 depth : indent depth
1325 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1327 A trie is an N'ary tree where the branches are determined by digital
1328 decomposition of the key. IE, at the root node you look up the 1st character and
1329 follow that branch repeat until you find the end of the branches. Nodes can be
1330 marked as "accepting" meaning they represent a complete word. Eg:
1334 would convert into the following structure. Numbers represent states, letters
1335 following numbers represent valid transitions on the letter from that state, if
1336 the number is in square brackets it represents an accepting state, otherwise it
1337 will be in parenthesis.
1339 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1343 (1) +-i->(6)-+-s->[7]
1345 +-s->(3)-+-h->(4)-+-e->[5]
1347 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1349 This shows that when matching against the string 'hers' we will begin at state 1
1350 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1351 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1352 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1353 single traverse. We store a mapping from accepting to state to which word was
1354 matched, and then when we have multiple possibilities we try to complete the
1355 rest of the regex in the order in which they occured in the alternation.
1357 The only prior NFA like behaviour that would be changed by the TRIE support is
1358 the silent ignoring of duplicate alternations which are of the form:
1360 / (DUPE|DUPE) X? (?{ ... }) Y /x
1362 Thus EVAL blocks following a trie may be called a different number of times with
1363 and without the optimisation. With the optimisations dupes will be silently
1364 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1365 the following demonstrates:
1367 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1369 which prints out 'word' three times, but
1371 'words'=~/(word|word|word)(?{ print $1 })S/
1373 which doesnt print it out at all. This is due to other optimisations kicking in.
1375 Example of what happens on a structural level:
1377 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1379 1: CURLYM[1] {1,32767}(18)
1390 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1391 and should turn into:
1393 1: CURLYM[1] {1,32767}(18)
1395 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1403 Cases where tail != last would be like /(?foo|bar)baz/:
1413 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1414 and would end up looking like:
1417 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1424 d = uvuni_to_utf8_flags(d, uv, 0);
1426 is the recommended Unicode-aware way of saying
1431 #define TRIE_STORE_REVCHAR(val) \
1434 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1435 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1436 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1437 SvCUR_set(zlopp, kapow - flrbbbbb); \
1440 av_push(revcharmap, zlopp); \
1442 char ooooff = (char)val; \
1443 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1447 #define TRIE_READ_CHAR STMT_START { \
1450 /* if it is UTF then it is either already folded, or does not need folding */ \
1451 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1453 else if (folder == PL_fold_latin1) { \
1454 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1455 if ( foldlen > 0 ) { \
1456 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1462 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, FOLD_FLAGS_FULL); \
1463 skiplen = UNISKIP(uvc); \
1464 foldlen -= skiplen; \
1465 scan = foldbuf + skiplen; \
1468 /* raw data, will be folded later if needed */ \
1476 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1477 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1478 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1479 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1481 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1482 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1483 TRIE_LIST_CUR( state )++; \
1486 #define TRIE_LIST_NEW(state) STMT_START { \
1487 Newxz( trie->states[ state ].trans.list, \
1488 4, reg_trie_trans_le ); \
1489 TRIE_LIST_CUR( state ) = 1; \
1490 TRIE_LIST_LEN( state ) = 4; \
1493 #define TRIE_HANDLE_WORD(state) STMT_START { \
1494 U16 dupe= trie->states[ state ].wordnum; \
1495 regnode * const noper_next = regnext( noper ); \
1498 /* store the word for dumping */ \
1500 if (OP(noper) != NOTHING) \
1501 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1503 tmp = newSVpvn_utf8( "", 0, UTF ); \
1504 av_push( trie_words, tmp ); \
1508 trie->wordinfo[curword].prev = 0; \
1509 trie->wordinfo[curword].len = wordlen; \
1510 trie->wordinfo[curword].accept = state; \
1512 if ( noper_next < tail ) { \
1514 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1515 trie->jump[curword] = (U16)(noper_next - convert); \
1517 jumper = noper_next; \
1519 nextbranch= regnext(cur); \
1523 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1524 /* chain, so that when the bits of chain are later */\
1525 /* linked together, the dups appear in the chain */\
1526 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1527 trie->wordinfo[dupe].prev = curword; \
1529 /* we haven't inserted this word yet. */ \
1530 trie->states[ state ].wordnum = curword; \
1535 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1536 ( ( base + charid >= ucharcount \
1537 && base + charid < ubound \
1538 && state == trie->trans[ base - ucharcount + charid ].check \
1539 && trie->trans[ base - ucharcount + charid ].next ) \
1540 ? trie->trans[ base - ucharcount + charid ].next \
1541 : ( state==1 ? special : 0 ) \
1545 #define MADE_JUMP_TRIE 2
1546 #define MADE_EXACT_TRIE 4
1549 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1552 /* first pass, loop through and scan words */
1553 reg_trie_data *trie;
1554 HV *widecharmap = NULL;
1555 AV *revcharmap = newAV();
1557 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1562 regnode *jumper = NULL;
1563 regnode *nextbranch = NULL;
1564 regnode *convert = NULL;
1565 U32 *prev_states; /* temp array mapping each state to previous one */
1566 /* we just use folder as a flag in utf8 */
1567 const U8 * folder = NULL;
1570 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1571 AV *trie_words = NULL;
1572 /* along with revcharmap, this only used during construction but both are
1573 * useful during debugging so we store them in the struct when debugging.
1576 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1577 STRLEN trie_charcount=0;
1579 SV *re_trie_maxbuff;
1580 GET_RE_DEBUG_FLAGS_DECL;
1582 PERL_ARGS_ASSERT_MAKE_TRIE;
1584 PERL_UNUSED_ARG(depth);
1591 case EXACTFU_TRICKYFOLD:
1592 case EXACTFU: folder = PL_fold_latin1; break;
1593 case EXACTF: folder = PL_fold; break;
1594 case EXACTFL: folder = PL_fold_locale; break;
1595 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1598 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1600 trie->startstate = 1;
1601 trie->wordcount = word_count;
1602 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1603 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1605 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1606 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1607 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1610 trie_words = newAV();
1613 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1614 if (!SvIOK(re_trie_maxbuff)) {
1615 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1617 DEBUG_TRIE_COMPILE_r({
1618 PerlIO_printf( Perl_debug_log,
1619 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1620 (int)depth * 2 + 2, "",
1621 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1622 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1626 /* Find the node we are going to overwrite */
1627 if ( first == startbranch && OP( last ) != BRANCH ) {
1628 /* whole branch chain */
1631 /* branch sub-chain */
1632 convert = NEXTOPER( first );
1635 /* -- First loop and Setup --
1637 We first traverse the branches and scan each word to determine if it
1638 contains widechars, and how many unique chars there are, this is
1639 important as we have to build a table with at least as many columns as we
1642 We use an array of integers to represent the character codes 0..255
1643 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1644 native representation of the character value as the key and IV's for the
1647 *TODO* If we keep track of how many times each character is used we can
1648 remap the columns so that the table compression later on is more
1649 efficient in terms of memory by ensuring the most common value is in the
1650 middle and the least common are on the outside. IMO this would be better
1651 than a most to least common mapping as theres a decent chance the most
1652 common letter will share a node with the least common, meaning the node
1653 will not be compressible. With a middle is most common approach the worst
1654 case is when we have the least common nodes twice.
1658 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1659 regnode *noper = NEXTOPER( cur );
1660 const U8 *uc = (U8*)STRING( noper );
1661 const U8 *e = uc + STR_LEN( noper );
1663 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1665 const U8 *scan = (U8*)NULL;
1666 U32 wordlen = 0; /* required init */
1668 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1670 if (OP(noper) == NOTHING) {
1671 regnode *noper_next= regnext(noper);
1672 if (noper_next != tail && OP(noper_next) == flags) {
1674 uc= (U8*)STRING(noper);
1675 e= uc + STR_LEN(noper);
1676 trie->minlen= STR_LEN(noper);
1683 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1684 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1685 regardless of encoding */
1686 if (OP( noper ) == EXACTFU_SS) {
1687 /* false positives are ok, so just set this */
1688 TRIE_BITMAP_SET(trie,0xDF);
1691 for ( ; uc < e ; uc += len ) {
1692 TRIE_CHARCOUNT(trie)++;
1697 U8 folded= folder[ (U8) uvc ];
1698 if ( !trie->charmap[ folded ] ) {
1699 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1700 TRIE_STORE_REVCHAR( folded );
1703 if ( !trie->charmap[ uvc ] ) {
1704 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1705 TRIE_STORE_REVCHAR( uvc );
1708 /* store the codepoint in the bitmap, and its folded
1710 TRIE_BITMAP_SET(trie, uvc);
1712 /* store the folded codepoint */
1713 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1716 /* store first byte of utf8 representation of
1717 variant codepoints */
1718 if (! UNI_IS_INVARIANT(uvc)) {
1719 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1722 set_bit = 0; /* We've done our bit :-) */
1727 widecharmap = newHV();
1729 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1732 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1734 if ( !SvTRUE( *svpp ) ) {
1735 sv_setiv( *svpp, ++trie->uniquecharcount );
1736 TRIE_STORE_REVCHAR(uvc);
1740 if( cur == first ) {
1741 trie->minlen = chars;
1742 trie->maxlen = chars;
1743 } else if (chars < trie->minlen) {
1744 trie->minlen = chars;
1745 } else if (chars > trie->maxlen) {
1746 trie->maxlen = chars;
1748 if (OP( noper ) == EXACTFU_SS) {
1749 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1750 if (trie->minlen > 1)
1753 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1754 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1755 * - We assume that any such sequence might match a 2 byte string */
1756 if (trie->minlen > 2 )
1760 } /* end first pass */
1761 DEBUG_TRIE_COMPILE_r(
1762 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1763 (int)depth * 2 + 2,"",
1764 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1765 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1766 (int)trie->minlen, (int)trie->maxlen )
1770 We now know what we are dealing with in terms of unique chars and
1771 string sizes so we can calculate how much memory a naive
1772 representation using a flat table will take. If it's over a reasonable
1773 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1774 conservative but potentially much slower representation using an array
1777 At the end we convert both representations into the same compressed
1778 form that will be used in regexec.c for matching with. The latter
1779 is a form that cannot be used to construct with but has memory
1780 properties similar to the list form and access properties similar
1781 to the table form making it both suitable for fast searches and
1782 small enough that its feasable to store for the duration of a program.
1784 See the comment in the code where the compressed table is produced
1785 inplace from the flat tabe representation for an explanation of how
1786 the compression works.
1791 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1794 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1796 Second Pass -- Array Of Lists Representation
1798 Each state will be represented by a list of charid:state records
1799 (reg_trie_trans_le) the first such element holds the CUR and LEN
1800 points of the allocated array. (See defines above).
1802 We build the initial structure using the lists, and then convert
1803 it into the compressed table form which allows faster lookups
1804 (but cant be modified once converted).
1807 STRLEN transcount = 1;
1809 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1810 "%*sCompiling trie using list compiler\n",
1811 (int)depth * 2 + 2, ""));
1813 trie->states = (reg_trie_state *)
1814 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1815 sizeof(reg_trie_state) );
1819 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1821 regnode *noper = NEXTOPER( cur );
1822 U8 *uc = (U8*)STRING( noper );
1823 const U8 *e = uc + STR_LEN( noper );
1824 U32 state = 1; /* required init */
1825 U16 charid = 0; /* sanity init */
1826 U8 *scan = (U8*)NULL; /* sanity init */
1827 STRLEN foldlen = 0; /* required init */
1828 U32 wordlen = 0; /* required init */
1829 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1832 if (OP(noper) == NOTHING) {
1833 regnode *noper_next= regnext(noper);
1834 if (noper_next != tail && OP(noper_next) == flags) {
1836 uc= (U8*)STRING(noper);
1837 e= uc + STR_LEN(noper);
1841 if (OP(noper) != NOTHING) {
1842 for ( ; uc < e ; uc += len ) {
1847 charid = trie->charmap[ uvc ];
1849 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1853 charid=(U16)SvIV( *svpp );
1856 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1863 if ( !trie->states[ state ].trans.list ) {
1864 TRIE_LIST_NEW( state );
1866 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1867 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1868 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1873 newstate = next_alloc++;
1874 prev_states[newstate] = state;
1875 TRIE_LIST_PUSH( state, charid, newstate );
1880 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1884 TRIE_HANDLE_WORD(state);
1886 } /* end second pass */
1888 /* next alloc is the NEXT state to be allocated */
1889 trie->statecount = next_alloc;
1890 trie->states = (reg_trie_state *)
1891 PerlMemShared_realloc( trie->states,
1893 * sizeof(reg_trie_state) );
1895 /* and now dump it out before we compress it */
1896 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1897 revcharmap, next_alloc,
1901 trie->trans = (reg_trie_trans *)
1902 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1909 for( state=1 ; state < next_alloc ; state ++ ) {
1913 DEBUG_TRIE_COMPILE_MORE_r(
1914 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1918 if (trie->states[state].trans.list) {
1919 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1923 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1924 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1925 if ( forid < minid ) {
1927 } else if ( forid > maxid ) {
1931 if ( transcount < tp + maxid - minid + 1) {
1933 trie->trans = (reg_trie_trans *)
1934 PerlMemShared_realloc( trie->trans,
1936 * sizeof(reg_trie_trans) );
1937 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1939 base = trie->uniquecharcount + tp - minid;
1940 if ( maxid == minid ) {
1942 for ( ; zp < tp ; zp++ ) {
1943 if ( ! trie->trans[ zp ].next ) {
1944 base = trie->uniquecharcount + zp - minid;
1945 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1946 trie->trans[ zp ].check = state;
1952 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1953 trie->trans[ tp ].check = state;
1958 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1959 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1960 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1961 trie->trans[ tid ].check = state;
1963 tp += ( maxid - minid + 1 );
1965 Safefree(trie->states[ state ].trans.list);
1968 DEBUG_TRIE_COMPILE_MORE_r(
1969 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1972 trie->states[ state ].trans.base=base;
1974 trie->lasttrans = tp + 1;
1978 Second Pass -- Flat Table Representation.
1980 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1981 We know that we will need Charcount+1 trans at most to store the data
1982 (one row per char at worst case) So we preallocate both structures
1983 assuming worst case.
1985 We then construct the trie using only the .next slots of the entry
1988 We use the .check field of the first entry of the node temporarily to
1989 make compression both faster and easier by keeping track of how many non
1990 zero fields are in the node.
1992 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1995 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1996 number representing the first entry of the node, and state as a
1997 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1998 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1999 are 2 entrys per node. eg:
2007 The table is internally in the right hand, idx form. However as we also
2008 have to deal with the states array which is indexed by nodenum we have to
2009 use TRIE_NODENUM() to convert.
2012 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2013 "%*sCompiling trie using table compiler\n",
2014 (int)depth * 2 + 2, ""));
2016 trie->trans = (reg_trie_trans *)
2017 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2018 * trie->uniquecharcount + 1,
2019 sizeof(reg_trie_trans) );
2020 trie->states = (reg_trie_state *)
2021 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2022 sizeof(reg_trie_state) );
2023 next_alloc = trie->uniquecharcount + 1;
2026 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2028 regnode *noper = NEXTOPER( cur );
2029 const U8 *uc = (U8*)STRING( noper );
2030 const U8 *e = uc + STR_LEN( noper );
2032 U32 state = 1; /* required init */
2034 U16 charid = 0; /* sanity init */
2035 U32 accept_state = 0; /* sanity init */
2036 U8 *scan = (U8*)NULL; /* sanity init */
2038 STRLEN foldlen = 0; /* required init */
2039 U32 wordlen = 0; /* required init */
2041 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2043 if (OP(noper) == NOTHING) {
2044 regnode *noper_next= regnext(noper);
2045 if (noper_next != tail && OP(noper_next) == flags) {
2047 uc= (U8*)STRING(noper);
2048 e= uc + STR_LEN(noper);
2052 if ( OP(noper) != NOTHING ) {
2053 for ( ; uc < e ; uc += len ) {
2058 charid = trie->charmap[ uvc ];
2060 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2061 charid = svpp ? (U16)SvIV(*svpp) : 0;
2065 if ( !trie->trans[ state + charid ].next ) {
2066 trie->trans[ state + charid ].next = next_alloc;
2067 trie->trans[ state ].check++;
2068 prev_states[TRIE_NODENUM(next_alloc)]
2069 = TRIE_NODENUM(state);
2070 next_alloc += trie->uniquecharcount;
2072 state = trie->trans[ state + charid ].next;
2074 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2076 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2079 accept_state = TRIE_NODENUM( state );
2080 TRIE_HANDLE_WORD(accept_state);
2082 } /* end second pass */
2084 /* and now dump it out before we compress it */
2085 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2087 next_alloc, depth+1));
2091 * Inplace compress the table.*
2093 For sparse data sets the table constructed by the trie algorithm will
2094 be mostly 0/FAIL transitions or to put it another way mostly empty.
2095 (Note that leaf nodes will not contain any transitions.)
2097 This algorithm compresses the tables by eliminating most such
2098 transitions, at the cost of a modest bit of extra work during lookup:
2100 - Each states[] entry contains a .base field which indicates the
2101 index in the state[] array wheres its transition data is stored.
2103 - If .base is 0 there are no valid transitions from that node.
2105 - If .base is nonzero then charid is added to it to find an entry in
2108 -If trans[states[state].base+charid].check!=state then the
2109 transition is taken to be a 0/Fail transition. Thus if there are fail
2110 transitions at the front of the node then the .base offset will point
2111 somewhere inside the previous nodes data (or maybe even into a node
2112 even earlier), but the .check field determines if the transition is
2116 The following process inplace converts the table to the compressed
2117 table: We first do not compress the root node 1,and mark all its
2118 .check pointers as 1 and set its .base pointer as 1 as well. This
2119 allows us to do a DFA construction from the compressed table later,
2120 and ensures that any .base pointers we calculate later are greater
2123 - We set 'pos' to indicate the first entry of the second node.
2125 - We then iterate over the columns of the node, finding the first and
2126 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2127 and set the .check pointers accordingly, and advance pos
2128 appropriately and repreat for the next node. Note that when we copy
2129 the next pointers we have to convert them from the original
2130 NODEIDX form to NODENUM form as the former is not valid post
2133 - If a node has no transitions used we mark its base as 0 and do not
2134 advance the pos pointer.
2136 - If a node only has one transition we use a second pointer into the
2137 structure to fill in allocated fail transitions from other states.
2138 This pointer is independent of the main pointer and scans forward
2139 looking for null transitions that are allocated to a state. When it
2140 finds one it writes the single transition into the "hole". If the
2141 pointer doesnt find one the single transition is appended as normal.
2143 - Once compressed we can Renew/realloc the structures to release the
2146 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2147 specifically Fig 3.47 and the associated pseudocode.
2151 const U32 laststate = TRIE_NODENUM( next_alloc );
2154 trie->statecount = laststate;
2156 for ( state = 1 ; state < laststate ; state++ ) {
2158 const U32 stateidx = TRIE_NODEIDX( state );
2159 const U32 o_used = trie->trans[ stateidx ].check;
2160 U32 used = trie->trans[ stateidx ].check;
2161 trie->trans[ stateidx ].check = 0;
2163 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2164 if ( flag || trie->trans[ stateidx + charid ].next ) {
2165 if ( trie->trans[ stateidx + charid ].next ) {
2167 for ( ; zp < pos ; zp++ ) {
2168 if ( ! trie->trans[ zp ].next ) {
2172 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2173 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2174 trie->trans[ zp ].check = state;
2175 if ( ++zp > pos ) pos = zp;
2182 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2184 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2185 trie->trans[ pos ].check = state;
2190 trie->lasttrans = pos + 1;
2191 trie->states = (reg_trie_state *)
2192 PerlMemShared_realloc( trie->states, laststate
2193 * sizeof(reg_trie_state) );
2194 DEBUG_TRIE_COMPILE_MORE_r(
2195 PerlIO_printf( Perl_debug_log,
2196 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2197 (int)depth * 2 + 2,"",
2198 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2201 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2204 } /* end table compress */
2206 DEBUG_TRIE_COMPILE_MORE_r(
2207 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2208 (int)depth * 2 + 2, "",
2209 (UV)trie->statecount,
2210 (UV)trie->lasttrans)
2212 /* resize the trans array to remove unused space */
2213 trie->trans = (reg_trie_trans *)
2214 PerlMemShared_realloc( trie->trans, trie->lasttrans
2215 * sizeof(reg_trie_trans) );
2217 { /* Modify the program and insert the new TRIE node */
2218 U8 nodetype =(U8)(flags & 0xFF);
2222 regnode *optimize = NULL;
2223 #ifdef RE_TRACK_PATTERN_OFFSETS
2226 U32 mjd_nodelen = 0;
2227 #endif /* RE_TRACK_PATTERN_OFFSETS */
2228 #endif /* DEBUGGING */
2230 This means we convert either the first branch or the first Exact,
2231 depending on whether the thing following (in 'last') is a branch
2232 or not and whther first is the startbranch (ie is it a sub part of
2233 the alternation or is it the whole thing.)
2234 Assuming its a sub part we convert the EXACT otherwise we convert
2235 the whole branch sequence, including the first.
2237 /* Find the node we are going to overwrite */
2238 if ( first != startbranch || OP( last ) == BRANCH ) {
2239 /* branch sub-chain */
2240 NEXT_OFF( first ) = (U16)(last - first);
2241 #ifdef RE_TRACK_PATTERN_OFFSETS
2243 mjd_offset= Node_Offset((convert));
2244 mjd_nodelen= Node_Length((convert));
2247 /* whole branch chain */
2249 #ifdef RE_TRACK_PATTERN_OFFSETS
2252 const regnode *nop = NEXTOPER( convert );
2253 mjd_offset= Node_Offset((nop));
2254 mjd_nodelen= Node_Length((nop));
2258 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2259 (int)depth * 2 + 2, "",
2260 (UV)mjd_offset, (UV)mjd_nodelen)
2263 /* But first we check to see if there is a common prefix we can
2264 split out as an EXACT and put in front of the TRIE node. */
2265 trie->startstate= 1;
2266 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2268 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2272 const U32 base = trie->states[ state ].trans.base;
2274 if ( trie->states[state].wordnum )
2277 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2278 if ( ( base + ofs >= trie->uniquecharcount ) &&
2279 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2280 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2282 if ( ++count > 1 ) {
2283 SV **tmp = av_fetch( revcharmap, ofs, 0);
2284 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2285 if ( state == 1 ) break;
2287 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2289 PerlIO_printf(Perl_debug_log,
2290 "%*sNew Start State=%"UVuf" Class: [",
2291 (int)depth * 2 + 2, "",
2294 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2295 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2297 TRIE_BITMAP_SET(trie,*ch);
2299 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2301 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2305 TRIE_BITMAP_SET(trie,*ch);
2307 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2308 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2314 SV **tmp = av_fetch( revcharmap, idx, 0);
2316 char *ch = SvPV( *tmp, len );
2318 SV *sv=sv_newmortal();
2319 PerlIO_printf( Perl_debug_log,
2320 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2321 (int)depth * 2 + 2, "",
2323 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2324 PL_colors[0], PL_colors[1],
2325 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2326 PERL_PV_ESCAPE_FIRSTCHAR
2331 OP( convert ) = nodetype;
2332 str=STRING(convert);
2335 STR_LEN(convert) += len;
2341 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2346 trie->prefixlen = (state-1);
2348 regnode *n = convert+NODE_SZ_STR(convert);
2349 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2350 trie->startstate = state;
2351 trie->minlen -= (state - 1);
2352 trie->maxlen -= (state - 1);
2354 /* At least the UNICOS C compiler choked on this
2355 * being argument to DEBUG_r(), so let's just have
2358 #ifdef PERL_EXT_RE_BUILD
2364 regnode *fix = convert;
2365 U32 word = trie->wordcount;
2367 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2368 while( ++fix < n ) {
2369 Set_Node_Offset_Length(fix, 0, 0);
2372 SV ** const tmp = av_fetch( trie_words, word, 0 );
2374 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2375 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2377 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2385 NEXT_OFF(convert) = (U16)(tail - convert);
2386 DEBUG_r(optimize= n);
2392 if ( trie->maxlen ) {
2393 NEXT_OFF( convert ) = (U16)(tail - convert);
2394 ARG_SET( convert, data_slot );
2395 /* Store the offset to the first unabsorbed branch in
2396 jump[0], which is otherwise unused by the jump logic.
2397 We use this when dumping a trie and during optimisation. */
2399 trie->jump[0] = (U16)(nextbranch - convert);
2401 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2402 * and there is a bitmap
2403 * and the first "jump target" node we found leaves enough room
2404 * then convert the TRIE node into a TRIEC node, with the bitmap
2405 * embedded inline in the opcode - this is hypothetically faster.
2407 if ( !trie->states[trie->startstate].wordnum
2409 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2411 OP( convert ) = TRIEC;
2412 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2413 PerlMemShared_free(trie->bitmap);
2416 OP( convert ) = TRIE;
2418 /* store the type in the flags */
2419 convert->flags = nodetype;
2423 + regarglen[ OP( convert ) ];
2425 /* XXX We really should free up the resource in trie now,
2426 as we won't use them - (which resources?) dmq */
2428 /* needed for dumping*/
2429 DEBUG_r(if (optimize) {
2430 regnode *opt = convert;
2432 while ( ++opt < optimize) {
2433 Set_Node_Offset_Length(opt,0,0);
2436 Try to clean up some of the debris left after the
2439 while( optimize < jumper ) {
2440 mjd_nodelen += Node_Length((optimize));
2441 OP( optimize ) = OPTIMIZED;
2442 Set_Node_Offset_Length(optimize,0,0);
2445 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2447 } /* end node insert */
2448 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, convert);
2450 /* Finish populating the prev field of the wordinfo array. Walk back
2451 * from each accept state until we find another accept state, and if
2452 * so, point the first word's .prev field at the second word. If the
2453 * second already has a .prev field set, stop now. This will be the
2454 * case either if we've already processed that word's accept state,
2455 * or that state had multiple words, and the overspill words were
2456 * already linked up earlier.
2463 for (word=1; word <= trie->wordcount; word++) {
2465 if (trie->wordinfo[word].prev)
2467 state = trie->wordinfo[word].accept;
2469 state = prev_states[state];
2472 prev = trie->states[state].wordnum;
2476 trie->wordinfo[word].prev = prev;
2478 Safefree(prev_states);
2482 /* and now dump out the compressed format */
2483 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2485 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2487 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2488 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2490 SvREFCNT_dec_NN(revcharmap);
2494 : trie->startstate>1
2500 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2502 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2504 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2505 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2508 We find the fail state for each state in the trie, this state is the longest proper
2509 suffix of the current state's 'word' that is also a proper prefix of another word in our
2510 trie. State 1 represents the word '' and is thus the default fail state. This allows
2511 the DFA not to have to restart after its tried and failed a word at a given point, it
2512 simply continues as though it had been matching the other word in the first place.
2514 'abcdgu'=~/abcdefg|cdgu/
2515 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2516 fail, which would bring us to the state representing 'd' in the second word where we would
2517 try 'g' and succeed, proceeding to match 'cdgu'.
2519 /* add a fail transition */
2520 const U32 trie_offset = ARG(source);
2521 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2523 const U32 ucharcount = trie->uniquecharcount;
2524 const U32 numstates = trie->statecount;
2525 const U32 ubound = trie->lasttrans + ucharcount;
2529 U32 base = trie->states[ 1 ].trans.base;
2532 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2533 GET_RE_DEBUG_FLAGS_DECL;
2535 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2537 PERL_UNUSED_ARG(depth);
2541 ARG_SET( stclass, data_slot );
2542 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2543 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2544 aho->trie=trie_offset;
2545 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2546 Copy( trie->states, aho->states, numstates, reg_trie_state );
2547 Newxz( q, numstates, U32);
2548 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2551 /* initialize fail[0..1] to be 1 so that we always have
2552 a valid final fail state */
2553 fail[ 0 ] = fail[ 1 ] = 1;
2555 for ( charid = 0; charid < ucharcount ; charid++ ) {
2556 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2558 q[ q_write ] = newstate;
2559 /* set to point at the root */
2560 fail[ q[ q_write++ ] ]=1;
2563 while ( q_read < q_write) {
2564 const U32 cur = q[ q_read++ % numstates ];
2565 base = trie->states[ cur ].trans.base;
2567 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2568 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2570 U32 fail_state = cur;
2573 fail_state = fail[ fail_state ];
2574 fail_base = aho->states[ fail_state ].trans.base;
2575 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2577 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2578 fail[ ch_state ] = fail_state;
2579 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2581 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2583 q[ q_write++ % numstates] = ch_state;
2587 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2588 when we fail in state 1, this allows us to use the
2589 charclass scan to find a valid start char. This is based on the principle
2590 that theres a good chance the string being searched contains lots of stuff
2591 that cant be a start char.
2593 fail[ 0 ] = fail[ 1 ] = 0;
2594 DEBUG_TRIE_COMPILE_r({
2595 PerlIO_printf(Perl_debug_log,
2596 "%*sStclass Failtable (%"UVuf" states): 0",
2597 (int)(depth * 2), "", (UV)numstates
2599 for( q_read=1; q_read<numstates; q_read++ ) {
2600 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2602 PerlIO_printf(Perl_debug_log, "\n");
2605 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2610 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2611 * These need to be revisited when a newer toolchain becomes available.
2613 #if defined(__sparc64__) && defined(__GNUC__)
2614 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2615 # undef SPARC64_GCC_WORKAROUND
2616 # define SPARC64_GCC_WORKAROUND 1
2620 #define DEBUG_PEEP(str,scan,depth) \
2621 DEBUG_OPTIMISE_r({if (scan){ \
2622 SV * const mysv=sv_newmortal(); \
2623 regnode *Next = regnext(scan); \
2624 regprop(RExC_rx, mysv, scan); \
2625 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2626 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2627 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2631 /* The below joins as many adjacent EXACTish nodes as possible into a single
2632 * one. The regop may be changed if the node(s) contain certain sequences that
2633 * require special handling. The joining is only done if:
2634 * 1) there is room in the current conglomerated node to entirely contain the
2636 * 2) they are the exact same node type
2638 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2639 * these get optimized out
2641 * If a node is to match under /i (folded), the number of characters it matches
2642 * can be different than its character length if it contains a multi-character
2643 * fold. *min_subtract is set to the total delta of the input nodes.
2645 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2646 * and contains LATIN SMALL LETTER SHARP S
2648 * This is as good a place as any to discuss the design of handling these
2649 * multi-character fold sequences. It's been wrong in Perl for a very long
2650 * time. There are three code points in Unicode whose multi-character folds
2651 * were long ago discovered to mess things up. The previous designs for
2652 * dealing with these involved assigning a special node for them. This
2653 * approach doesn't work, as evidenced by this example:
2654 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2655 * Both these fold to "sss", but if the pattern is parsed to create a node that
2656 * would match just the \xDF, it won't be able to handle the case where a
2657 * successful match would have to cross the node's boundary. The new approach
2658 * that hopefully generally solves the problem generates an EXACTFU_SS node
2661 * It turns out that there are problems with all multi-character folds, and not
2662 * just these three. Now the code is general, for all such cases, but the
2663 * three still have some special handling. The approach taken is:
2664 * 1) This routine examines each EXACTFish node that could contain multi-
2665 * character fold sequences. It returns in *min_subtract how much to
2666 * subtract from the the actual length of the string to get a real minimum
2667 * match length; it is 0 if there are no multi-char folds. This delta is
2668 * used by the caller to adjust the min length of the match, and the delta
2669 * between min and max, so that the optimizer doesn't reject these
2670 * possibilities based on size constraints.
2671 * 2) Certain of these sequences require special handling by the trie code,
2672 * so, if found, this code changes the joined node type to special ops:
2673 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2674 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2675 * is used for an EXACTFU node that contains at least one "ss" sequence in
2676 * it. For non-UTF-8 patterns and strings, this is the only case where
2677 * there is a possible fold length change. That means that a regular
2678 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2679 * with length changes, and so can be processed faster. regexec.c takes
2680 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2681 * pre-folded by regcomp.c. This saves effort in regex matching.
2682 * However, the pre-folding isn't done for non-UTF8 patterns because the
2683 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2684 * down by forcing the pattern into UTF8 unless necessary. Also what
2685 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2686 * possibilities for the non-UTF8 patterns are quite simple, except for
2687 * the sharp s. All the ones that don't involve a UTF-8 target string are
2688 * members of a fold-pair, and arrays are set up for all of them so that
2689 * the other member of the pair can be found quickly. Code elsewhere in
2690 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2691 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2692 * described in the next item.
2693 * 4) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
2694 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2695 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2696 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2697 * character in the pattern corresponds to at most a single character in
2698 * the target string. (And I do mean character, and not byte here, unlike
2699 * other parts of the documentation that have never been updated to
2700 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2701 * two character string 'ss'; in EXACTFA nodes it can match
2702 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2703 * instances where it is violated. I'm reluctant to try to change the
2704 * assumption, as the code involved is impenetrable to me (khw), so
2705 * instead the code here punts. This routine examines (when the pattern
2706 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2707 * boolean indicating whether or not the node contains a sharp s. When it
2708 * is true, the caller sets a flag that later causes the optimizer in this
2709 * file to not set values for the floating and fixed string lengths, and
2710 * thus avoids the optimizer code in regexec.c that makes the invalid
2711 * assumption. Thus, there is no optimization based on string lengths for
2712 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2713 * (The reason the assumption is wrong only in these two cases is that all
2714 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2715 * other folds to their expanded versions. We can't prefold sharp s to
2716 * 'ss' in EXACTF nodes because we don't know at compile time if it
2717 * actually matches 'ss' or not. It will match iff the target string is
2718 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2719 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2720 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2721 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2722 * require the pattern to be forced into UTF-8, the overhead of which we
2726 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2727 if (PL_regkind[OP(scan)] == EXACT) \
2728 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2731 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) {
2732 /* Merge several consecutive EXACTish nodes into one. */
2733 regnode *n = regnext(scan);
2735 regnode *next = scan + NODE_SZ_STR(scan);
2739 regnode *stop = scan;
2740 GET_RE_DEBUG_FLAGS_DECL;
2742 PERL_UNUSED_ARG(depth);
2745 PERL_ARGS_ASSERT_JOIN_EXACT;
2746 #ifndef EXPERIMENTAL_INPLACESCAN
2747 PERL_UNUSED_ARG(flags);
2748 PERL_UNUSED_ARG(val);
2750 DEBUG_PEEP("join",scan,depth);
2752 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2753 * EXACT ones that are mergeable to the current one. */
2755 && (PL_regkind[OP(n)] == NOTHING
2756 || (stringok && OP(n) == OP(scan)))
2758 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2761 if (OP(n) == TAIL || n > next)
2763 if (PL_regkind[OP(n)] == NOTHING) {
2764 DEBUG_PEEP("skip:",n,depth);
2765 NEXT_OFF(scan) += NEXT_OFF(n);
2766 next = n + NODE_STEP_REGNODE;
2773 else if (stringok) {
2774 const unsigned int oldl = STR_LEN(scan);
2775 regnode * const nnext = regnext(n);
2777 /* XXX I (khw) kind of doubt that this works on platforms where
2778 * U8_MAX is above 255 because of lots of other assumptions */
2779 /* Don't join if the sum can't fit into a single node */
2780 if (oldl + STR_LEN(n) > U8_MAX)
2783 DEBUG_PEEP("merg",n,depth);
2786 NEXT_OFF(scan) += NEXT_OFF(n);
2787 STR_LEN(scan) += STR_LEN(n);
2788 next = n + NODE_SZ_STR(n);
2789 /* Now we can overwrite *n : */
2790 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2798 #ifdef EXPERIMENTAL_INPLACESCAN
2799 if (flags && !NEXT_OFF(n)) {
2800 DEBUG_PEEP("atch", val, depth);
2801 if (reg_off_by_arg[OP(n)]) {
2802 ARG_SET(n, val - n);
2805 NEXT_OFF(n) = val - n;
2813 *has_exactf_sharp_s = FALSE;
2815 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2816 * can now analyze for sequences of problematic code points. (Prior to
2817 * this final joining, sequences could have been split over boundaries, and
2818 * hence missed). The sequences only happen in folding, hence for any
2819 * non-EXACT EXACTish node */
2820 if (OP(scan) != EXACT) {
2821 const U8 * const s0 = (U8*) STRING(scan);
2823 const U8 * const s_end = s0 + STR_LEN(scan);
2825 /* One pass is made over the node's string looking for all the
2826 * possibilities. to avoid some tests in the loop, there are two main
2827 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2831 /* Examine the string for a multi-character fold sequence. UTF-8
2832 * patterns have all characters pre-folded by the time this code is
2834 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2835 length sequence we are looking for is 2 */
2838 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2839 if (! len) { /* Not a multi-char fold: get next char */
2844 /* Nodes with 'ss' require special handling, except for EXACTFL
2845 * and EXACTFA for which there is no multi-char fold to this */
2846 if (len == 2 && *s == 's' && *(s+1) == 's'
2847 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2850 OP(scan) = EXACTFU_SS;
2853 else if (len == 6 /* len is the same in both ASCII and EBCDIC
2855 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2856 COMBINING_DIAERESIS_UTF8
2857 COMBINING_ACUTE_ACCENT_UTF8,
2859 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2860 COMBINING_DIAERESIS_UTF8
2861 COMBINING_ACUTE_ACCENT_UTF8,
2866 /* These two folds require special handling by trie's, so
2867 * change the node type to indicate this. If EXACTFA and
2868 * EXACTFL were ever to be handled by trie's, this would
2869 * have to be changed. If this node has already been
2870 * changed to EXACTFU_SS in this loop, leave it as is. (I
2871 * (khw) think it doesn't matter in regexec.c for UTF
2872 * patterns, but no need to change it */
2873 if (OP(scan) == EXACTFU) {
2874 OP(scan) = EXACTFU_TRICKYFOLD;
2878 else { /* Here is a generic multi-char fold. */
2879 const U8* multi_end = s + len;
2881 /* Count how many characters in it. In the case of /l and
2882 * /aa, no folds which contain ASCII code points are
2883 * allowed, so check for those, and skip if found. (In
2884 * EXACTFL, no folds are allowed to any Latin1 code point,
2885 * not just ASCII. But there aren't any of these
2886 * currently, nor ever likely, so don't take the time to
2887 * test for them. The code that generates the
2888 * is_MULTI_foo() macros croaks should one actually get put
2889 * into Unicode .) */
2890 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2891 count = utf8_length(s, multi_end);
2895 while (s < multi_end) {
2898 goto next_iteration;
2908 /* The delta is how long the sequence is minus 1 (1 is how long
2909 * the character that folds to the sequence is) */
2910 *min_subtract += count - 1;
2914 else if (OP(scan) == EXACTFA) {
2916 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2917 * fold to the ASCII range (and there are no existing ones in the
2918 * upper latin1 range). But, as outlined in the comments preceding
2919 * this function, we need to flag any occurrences of the sharp s */
2921 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2922 *has_exactf_sharp_s = TRUE;
2929 else if (OP(scan) != EXACTFL) {
2931 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2932 * multi-char folds that are all Latin1. (This code knows that
2933 * there are no current multi-char folds possible with EXACTFL,
2934 * relying on fold_grind.t to catch any errors if the very unlikely
2935 * event happens that some get added in future Unicode versions.)
2936 * As explained in the comments preceding this function, we look
2937 * also for the sharp s in EXACTF nodes; it can be in the final
2938 * position. Otherwise we can stop looking 1 byte earlier because
2939 * have to find at least two characters for a multi-fold */
2940 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2942 /* The below is perhaps overboard, but this allows us to save a
2943 * test each time through the loop at the expense of a mask. This
2944 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2945 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2946 * are 64. This uses an exclusive 'or' to find that bit and then
2947 * inverts it to form a mask, with just a single 0, in the bit
2948 * position where 'S' and 's' differ. */
2949 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2950 const U8 s_masked = 's' & S_or_s_mask;
2953 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2954 if (! len) { /* Not a multi-char fold. */
2955 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2957 *has_exactf_sharp_s = TRUE;
2964 && ((*s & S_or_s_mask) == s_masked)
2965 && ((*(s+1) & S_or_s_mask) == s_masked))
2968 /* EXACTF nodes need to know that the minimum length
2969 * changed so that a sharp s in the string can match this
2970 * ss in the pattern, but they remain EXACTF nodes, as they
2971 * won't match this unless the target string is is UTF-8,
2972 * which we don't know until runtime */
2973 if (OP(scan) != EXACTF) {
2974 OP(scan) = EXACTFU_SS;
2978 *min_subtract += len - 1;
2985 /* Allow dumping but overwriting the collection of skipped
2986 * ops and/or strings with fake optimized ops */
2987 n = scan + NODE_SZ_STR(scan);
2995 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2999 /* REx optimizer. Converts nodes into quicker variants "in place".
3000 Finds fixed substrings. */
3002 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3003 to the position after last scanned or to NULL. */
3005 #define INIT_AND_WITHP \
3006 assert(!and_withp); \
3007 Newx(and_withp,1,struct regnode_charclass_class); \
3008 SAVEFREEPV(and_withp)
3010 /* this is a chain of data about sub patterns we are processing that
3011 need to be handled separately/specially in study_chunk. Its so
3012 we can simulate recursion without losing state. */
3014 typedef struct scan_frame {
3015 regnode *last; /* last node to process in this frame */
3016 regnode *next; /* next node to process when last is reached */
3017 struct scan_frame *prev; /*previous frame*/
3018 I32 stop; /* what stopparen do we use */
3022 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3025 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3026 I32 *minlenp, I32 *deltap,
3031 struct regnode_charclass_class *and_withp,
3032 U32 flags, U32 depth)
3033 /* scanp: Start here (read-write). */
3034 /* deltap: Write maxlen-minlen here. */
3035 /* last: Stop before this one. */
3036 /* data: string data about the pattern */
3037 /* stopparen: treat close N as END */
3038 /* recursed: which subroutines have we recursed into */
3039 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3042 I32 min = 0; /* There must be at least this number of characters to match */
3044 regnode *scan = *scanp, *next;
3046 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3047 int is_inf_internal = 0; /* The studied chunk is infinite */
3048 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3049 scan_data_t data_fake;
3050 SV *re_trie_maxbuff = NULL;
3051 regnode *first_non_open = scan;
3052 I32 stopmin = I32_MAX;
3053 scan_frame *frame = NULL;
3054 GET_RE_DEBUG_FLAGS_DECL;
3056 PERL_ARGS_ASSERT_STUDY_CHUNK;
3059 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3063 while (first_non_open && OP(first_non_open) == OPEN)
3064 first_non_open=regnext(first_non_open);
3069 while ( scan && OP(scan) != END && scan < last ){
3070 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3071 node length to get a real minimum (because
3072 the folded version may be shorter) */
3073 bool has_exactf_sharp_s = FALSE;
3074 /* Peephole optimizer: */
3075 DEBUG_STUDYDATA("Peep:", data,depth);
3076 DEBUG_PEEP("Peep",scan,depth);
3078 /* Its not clear to khw or hv why this is done here, and not in the
3079 * clauses that deal with EXACT nodes. khw's guess is that it's
3080 * because of a previous design */
3081 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3083 /* Follow the next-chain of the current node and optimize
3084 away all the NOTHINGs from it. */
3085 if (OP(scan) != CURLYX) {
3086 const int max = (reg_off_by_arg[OP(scan)]
3088 /* I32 may be smaller than U16 on CRAYs! */
3089 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3090 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3094 /* Skip NOTHING and LONGJMP. */
3095 while ((n = regnext(n))
3096 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3097 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3098 && off + noff < max)
3100 if (reg_off_by_arg[OP(scan)])
3103 NEXT_OFF(scan) = off;
3108 /* The principal pseudo-switch. Cannot be a switch, since we
3109 look into several different things. */
3110 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3111 || OP(scan) == IFTHEN) {
3112 next = regnext(scan);
3114 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3116 if (OP(next) == code || code == IFTHEN) {
3117 /* NOTE - There is similar code to this block below for handling
3118 TRIE nodes on a re-study. If you change stuff here check there
3120 I32 max1 = 0, min1 = I32_MAX, num = 0;
3121 struct regnode_charclass_class accum;
3122 regnode * const startbranch=scan;
3124 if (flags & SCF_DO_SUBSTR)
3125 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3126 if (flags & SCF_DO_STCLASS)
3127 cl_init_zero(pRExC_state, &accum);
3129 while (OP(scan) == code) {
3130 I32 deltanext, minnext, f = 0, fake;
3131 struct regnode_charclass_class this_class;
3134 data_fake.flags = 0;
3136 data_fake.whilem_c = data->whilem_c;
3137 data_fake.last_closep = data->last_closep;
3140 data_fake.last_closep = &fake;
3142 data_fake.pos_delta = delta;
3143 next = regnext(scan);
3144 scan = NEXTOPER(scan);
3146 scan = NEXTOPER(scan);
3147 if (flags & SCF_DO_STCLASS) {
3148 cl_init(pRExC_state, &this_class);
3149 data_fake.start_class = &this_class;
3150 f = SCF_DO_STCLASS_AND;
3152 if (flags & SCF_WHILEM_VISITED_POS)
3153 f |= SCF_WHILEM_VISITED_POS;
3155 /* we suppose the run is continuous, last=next...*/
3156 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3158 stopparen, recursed, NULL, f,depth+1);
3161 if (deltanext == I32_MAX) {
3162 is_inf = is_inf_internal = 1;
3164 } else if (max1 < minnext + deltanext)
3165 max1 = minnext + deltanext;
3167 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3169 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3170 if ( stopmin > minnext)
3171 stopmin = min + min1;
3172 flags &= ~SCF_DO_SUBSTR;
3174 data->flags |= SCF_SEEN_ACCEPT;
3177 if (data_fake.flags & SF_HAS_EVAL)
3178 data->flags |= SF_HAS_EVAL;
3179 data->whilem_c = data_fake.whilem_c;
3181 if (flags & SCF_DO_STCLASS)
3182 cl_or(pRExC_state, &accum, &this_class);
3184 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3186 if (flags & SCF_DO_SUBSTR) {
3187 data->pos_min += min1;
3188 if (data->pos_delta >= I32_MAX - (max1 - min1))
3189 data->pos_delta = I32_MAX;
3191 data->pos_delta += max1 - min1;
3192 if (max1 != min1 || is_inf)
3193 data->longest = &(data->longest_float);
3196 if (delta == I32_MAX || I32_MAX - delta - (max1 - min1) < 0)
3199 delta += max1 - min1;
3200 if (flags & SCF_DO_STCLASS_OR) {
3201 cl_or(pRExC_state, data->start_class, &accum);
3203 cl_and(data->start_class, and_withp);
3204 flags &= ~SCF_DO_STCLASS;
3207 else if (flags & SCF_DO_STCLASS_AND) {
3209 cl_and(data->start_class, &accum);
3210 flags &= ~SCF_DO_STCLASS;
3213 /* Switch to OR mode: cache the old value of
3214 * data->start_class */
3216 StructCopy(data->start_class, and_withp,
3217 struct regnode_charclass_class);
3218 flags &= ~SCF_DO_STCLASS_AND;
3219 StructCopy(&accum, data->start_class,
3220 struct regnode_charclass_class);
3221 flags |= SCF_DO_STCLASS_OR;
3222 SET_SSC_EOS(data->start_class);
3226 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3229 Assuming this was/is a branch we are dealing with: 'scan' now
3230 points at the item that follows the branch sequence, whatever
3231 it is. We now start at the beginning of the sequence and look
3238 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3240 If we can find such a subsequence we need to turn the first
3241 element into a trie and then add the subsequent branch exact
3242 strings to the trie.
3246 1. patterns where the whole set of branches can be converted.
3248 2. patterns where only a subset can be converted.
3250 In case 1 we can replace the whole set with a single regop
3251 for the trie. In case 2 we need to keep the start and end
3254 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3255 becomes BRANCH TRIE; BRANCH X;
3257 There is an additional case, that being where there is a
3258 common prefix, which gets split out into an EXACT like node
3259 preceding the TRIE node.
3261 If x(1..n)==tail then we can do a simple trie, if not we make
3262 a "jump" trie, such that when we match the appropriate word
3263 we "jump" to the appropriate tail node. Essentially we turn
3264 a nested if into a case structure of sorts.
3269 if (!re_trie_maxbuff) {
3270 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3271 if (!SvIOK(re_trie_maxbuff))
3272 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3274 if ( SvIV(re_trie_maxbuff)>=0 ) {
3276 regnode *first = (regnode *)NULL;
3277 regnode *last = (regnode *)NULL;
3278 regnode *tail = scan;
3283 SV * const mysv = sv_newmortal(); /* for dumping */
3285 /* var tail is used because there may be a TAIL
3286 regop in the way. Ie, the exacts will point to the
3287 thing following the TAIL, but the last branch will
3288 point at the TAIL. So we advance tail. If we
3289 have nested (?:) we may have to move through several
3293 while ( OP( tail ) == TAIL ) {
3294 /* this is the TAIL generated by (?:) */
3295 tail = regnext( tail );
3299 DEBUG_TRIE_COMPILE_r({
3300 regprop(RExC_rx, mysv, tail );
3301 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3302 (int)depth * 2 + 2, "",
3303 "Looking for TRIE'able sequences. Tail node is: ",
3304 SvPV_nolen_const( mysv )
3310 Step through the branches
3311 cur represents each branch,
3312 noper is the first thing to be matched as part of that branch
3313 noper_next is the regnext() of that node.
3315 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3316 via a "jump trie" but we also support building with NOJUMPTRIE,
3317 which restricts the trie logic to structures like /FOO|BAR/.
3319 If noper is a trieable nodetype then the branch is a possible optimization
3320 target. If we are building under NOJUMPTRIE then we require that noper_next
3321 is the same as scan (our current position in the regex program).
3323 Once we have two or more consecutive such branches we can create a
3324 trie of the EXACT's contents and stitch it in place into the program.
3326 If the sequence represents all of the branches in the alternation we
3327 replace the entire thing with a single TRIE node.
3329 Otherwise when it is a subsequence we need to stitch it in place and
3330 replace only the relevant branches. This means the first branch has
3331 to remain as it is used by the alternation logic, and its next pointer,
3332 and needs to be repointed at the item on the branch chain following
3333 the last branch we have optimized away.
3335 This could be either a BRANCH, in which case the subsequence is internal,
3336 or it could be the item following the branch sequence in which case the
3337 subsequence is at the end (which does not necessarily mean the first node
3338 is the start of the alternation).
3340 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3343 ----------------+-----------
3347 EXACTFU_SS | EXACTFU
3348 EXACTFU_TRICKYFOLD | EXACTFU
3353 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3354 ( EXACT == (X) ) ? EXACT : \
3355 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3358 /* dont use tail as the end marker for this traverse */
3359 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3360 regnode * const noper = NEXTOPER( cur );
3361 U8 noper_type = OP( noper );
3362 U8 noper_trietype = TRIE_TYPE( noper_type );
3363 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3364 regnode * const noper_next = regnext( noper );
3365 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3366 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3369 DEBUG_TRIE_COMPILE_r({
3370 regprop(RExC_rx, mysv, cur);
3371 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3372 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3374 regprop(RExC_rx, mysv, noper);
3375 PerlIO_printf( Perl_debug_log, " -> %s",
3376 SvPV_nolen_const(mysv));
3379 regprop(RExC_rx, mysv, noper_next );
3380 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3381 SvPV_nolen_const(mysv));
3383 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3384 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3385 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3389 /* Is noper a trieable nodetype that can be merged with the
3390 * current trie (if there is one)? */
3394 ( noper_trietype == NOTHING)
3395 || ( trietype == NOTHING )
3396 || ( trietype == noper_trietype )
3399 && noper_next == tail
3403 /* Handle mergable triable node
3404 * Either we are the first node in a new trieable sequence,
3405 * in which case we do some bookkeeping, otherwise we update
3406 * the end pointer. */
3409 if ( noper_trietype == NOTHING ) {
3410 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3411 regnode * const noper_next = regnext( noper );
3412 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3413 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3416 if ( noper_next_trietype ) {
3417 trietype = noper_next_trietype;
3418 } else if (noper_next_type) {
3419 /* a NOTHING regop is 1 regop wide. We need at least two
3420 * for a trie so we can't merge this in */
3424 trietype = noper_trietype;
3427 if ( trietype == NOTHING )
3428 trietype = noper_trietype;
3433 } /* end handle mergable triable node */
3435 /* handle unmergable node -
3436 * noper may either be a triable node which can not be tried
3437 * together with the current trie, or a non triable node */
3439 /* If last is set and trietype is not NOTHING then we have found
3440 * at least two triable branch sequences in a row of a similar
3441 * trietype so we can turn them into a trie. If/when we
3442 * allow NOTHING to start a trie sequence this condition will be
3443 * required, and it isn't expensive so we leave it in for now. */
3444 if ( trietype && trietype != NOTHING )
3445 make_trie( pRExC_state,
3446 startbranch, first, cur, tail, count,
3447 trietype, depth+1 );
3448 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3452 && noper_next == tail
3455 /* noper is triable, so we can start a new trie sequence */
3458 trietype = noper_trietype;
3460 /* if we already saw a first but the current node is not triable then we have
3461 * to reset the first information. */
3466 } /* end handle unmergable node */
3467 } /* loop over branches */
3468 DEBUG_TRIE_COMPILE_r({
3469 regprop(RExC_rx, mysv, cur);
3470 PerlIO_printf( Perl_debug_log,
3471 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3472 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3475 if ( last && trietype ) {
3476 if ( trietype != NOTHING ) {
3477 /* the last branch of the sequence was part of a trie,
3478 * so we have to construct it here outside of the loop
3480 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3481 #ifdef TRIE_STUDY_OPT
3482 if ( ((made == MADE_EXACT_TRIE &&
3483 startbranch == first)
3484 || ( first_non_open == first )) &&
3486 flags |= SCF_TRIE_RESTUDY;
3487 if ( startbranch == first
3490 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3495 /* at this point we know whatever we have is a NOTHING sequence/branch
3496 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3498 if ( startbranch == first ) {
3500 /* the entire thing is a NOTHING sequence, something like this:
3501 * (?:|) So we can turn it into a plain NOTHING op. */
3502 DEBUG_TRIE_COMPILE_r({
3503 regprop(RExC_rx, mysv, cur);
3504 PerlIO_printf( Perl_debug_log,
3505 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3506 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3509 OP(startbranch)= NOTHING;
3510 NEXT_OFF(startbranch)= tail - startbranch;
3511 for ( opt= startbranch + 1; opt < tail ; opt++ )
3515 } /* end if ( last) */
3516 } /* TRIE_MAXBUF is non zero */
3521 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3522 scan = NEXTOPER(NEXTOPER(scan));
3523 } else /* single branch is optimized. */
3524 scan = NEXTOPER(scan);
3526 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3527 scan_frame *newframe = NULL;
3532 if (OP(scan) != SUSPEND) {
3533 /* set the pointer */
3534 if (OP(scan) == GOSUB) {
3536 RExC_recurse[ARG2L(scan)] = scan;
3537 start = RExC_open_parens[paren-1];
3538 end = RExC_close_parens[paren-1];
3541 start = RExC_rxi->program + 1;
3545 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3546 SAVEFREEPV(recursed);
3548 if (!PAREN_TEST(recursed,paren+1)) {
3549 PAREN_SET(recursed,paren+1);
3550 Newx(newframe,1,scan_frame);
3552 if (flags & SCF_DO_SUBSTR) {
3553 SCAN_COMMIT(pRExC_state,data,minlenp);
3554 data->longest = &(data->longest_float);
3556 is_inf = is_inf_internal = 1;
3557 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3558 cl_anything(pRExC_state, data->start_class);
3559 flags &= ~SCF_DO_STCLASS;
3562 Newx(newframe,1,scan_frame);
3565 end = regnext(scan);
3570 SAVEFREEPV(newframe);
3571 newframe->next = regnext(scan);
3572 newframe->last = last;
3573 newframe->stop = stopparen;
3574 newframe->prev = frame;
3584 else if (OP(scan) == EXACT) {
3585 I32 l = STR_LEN(scan);
3588 const U8 * const s = (U8*)STRING(scan);
3589 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3590 l = utf8_length(s, s + l);
3592 uc = *((U8*)STRING(scan));
3595 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3596 /* The code below prefers earlier match for fixed
3597 offset, later match for variable offset. */
3598 if (data->last_end == -1) { /* Update the start info. */
3599 data->last_start_min = data->pos_min;
3600 data->last_start_max = is_inf
3601 ? I32_MAX : data->pos_min + data->pos_delta;
3603 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3605 SvUTF8_on(data->last_found);
3607 SV * const sv = data->last_found;
3608 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3609 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3610 if (mg && mg->mg_len >= 0)
3611 mg->mg_len += utf8_length((U8*)STRING(scan),
3612 (U8*)STRING(scan)+STR_LEN(scan));
3614 data->last_end = data->pos_min + l;
3615 data->pos_min += l; /* As in the first entry. */
3616 data->flags &= ~SF_BEFORE_EOL;
3618 if (flags & SCF_DO_STCLASS_AND) {
3619 /* Check whether it is compatible with what we know already! */
3623 /* If compatible, we or it in below. It is compatible if is
3624 * in the bitmp and either 1) its bit or its fold is set, or 2)
3625 * it's for a locale. Even if there isn't unicode semantics
3626 * here, at runtime there may be because of matching against a
3627 * utf8 string, so accept a possible false positive for
3628 * latin1-range folds */
3630 (!(data->start_class->flags & ANYOF_LOCALE)
3631 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3632 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3633 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3638 ANYOF_CLASS_ZERO(data->start_class);
3639 ANYOF_BITMAP_ZERO(data->start_class);
3641 ANYOF_BITMAP_SET(data->start_class, uc);
3642 else if (uc >= 0x100) {
3645 /* Some Unicode code points fold to the Latin1 range; as
3646 * XXX temporary code, instead of figuring out if this is
3647 * one, just assume it is and set all the start class bits
3648 * that could be some such above 255 code point's fold
3649 * which will generate fals positives. As the code
3650 * elsewhere that does compute the fold settles down, it
3651 * can be extracted out and re-used here */
3652 for (i = 0; i < 256; i++){
3653 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3654 ANYOF_BITMAP_SET(data->start_class, i);
3658 CLEAR_SSC_EOS(data->start_class);
3660 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3662 else if (flags & SCF_DO_STCLASS_OR) {
3663 /* false positive possible if the class is case-folded */
3665 ANYOF_BITMAP_SET(data->start_class, uc);
3667 data->start_class->flags |= ANYOF_UNICODE_ALL;
3668 CLEAR_SSC_EOS(data->start_class);
3669 cl_and(data->start_class, and_withp);
3671 flags &= ~SCF_DO_STCLASS;
3673 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3674 I32 l = STR_LEN(scan);
3675 UV uc = *((U8*)STRING(scan));
3677 /* Search for fixed substrings supports EXACT only. */
3678 if (flags & SCF_DO_SUBSTR) {
3680 SCAN_COMMIT(pRExC_state, data, minlenp);
3683 const U8 * const s = (U8 *)STRING(scan);
3684 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3685 l = utf8_length(s, s + l);
3687 if (has_exactf_sharp_s) {
3688 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3690 min += l - min_subtract;
3692 delta += min_subtract;
3693 if (flags & SCF_DO_SUBSTR) {
3694 data->pos_min += l - min_subtract;
3695 if (data->pos_min < 0) {
3698 data->pos_delta += min_subtract;
3700 data->longest = &(data->longest_float);
3703 if (flags & SCF_DO_STCLASS_AND) {
3704 /* Check whether it is compatible with what we know already! */
3707 (!(data->start_class->flags & ANYOF_LOCALE)
3708 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3709 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3713 ANYOF_CLASS_ZERO(data->start_class);
3714 ANYOF_BITMAP_ZERO(data->start_class);
3716 ANYOF_BITMAP_SET(data->start_class, uc);
3717 CLEAR_SSC_EOS(data->start_class);
3718 if (OP(scan) == EXACTFL) {
3719 /* XXX This set is probably no longer necessary, and
3720 * probably wrong as LOCALE now is on in the initial
3722 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3726 /* Also set the other member of the fold pair. In case
3727 * that unicode semantics is called for at runtime, use
3728 * the full latin1 fold. (Can't do this for locale,
3729 * because not known until runtime) */
3730 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3732 /* All other (EXACTFL handled above) folds except under
3733 * /iaa that include s, S, and sharp_s also may include
3735 if (OP(scan) != EXACTFA) {
3736 if (uc == 's' || uc == 'S') {
3737 ANYOF_BITMAP_SET(data->start_class,
3738 LATIN_SMALL_LETTER_SHARP_S);
3740 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3741 ANYOF_BITMAP_SET(data->start_class, 's');
3742 ANYOF_BITMAP_SET(data->start_class, 'S');
3747 else if (uc >= 0x100) {
3749 for (i = 0; i < 256; i++){
3750 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3751 ANYOF_BITMAP_SET(data->start_class, i);
3756 else if (flags & SCF_DO_STCLASS_OR) {
3757 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3758 /* false positive possible if the class is case-folded.
3759 Assume that the locale settings are the same... */
3761 ANYOF_BITMAP_SET(data->start_class, uc);
3762 if (OP(scan) != EXACTFL) {
3764 /* And set the other member of the fold pair, but
3765 * can't do that in locale because not known until
3767 ANYOF_BITMAP_SET(data->start_class,
3768 PL_fold_latin1[uc]);
3770 /* All folds except under /iaa that include s, S,
3771 * and sharp_s also may include the others */
3772 if (OP(scan) != EXACTFA) {
3773 if (uc == 's' || uc == 'S') {
3774 ANYOF_BITMAP_SET(data->start_class,
3775 LATIN_SMALL_LETTER_SHARP_S);
3777 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3778 ANYOF_BITMAP_SET(data->start_class, 's');
3779 ANYOF_BITMAP_SET(data->start_class, 'S');
3784 CLEAR_SSC_EOS(data->start_class);
3786 cl_and(data->start_class, and_withp);
3788 flags &= ~SCF_DO_STCLASS;
3790 else if (REGNODE_VARIES(OP(scan))) {
3791 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3792 I32 f = flags, pos_before = 0;
3793 regnode * const oscan = scan;
3794 struct regnode_charclass_class this_class;
3795 struct regnode_charclass_class *oclass = NULL;
3796 I32 next_is_eval = 0;
3798 switch (PL_regkind[OP(scan)]) {
3799 case WHILEM: /* End of (?:...)* . */
3800 scan = NEXTOPER(scan);
3803 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3804 next = NEXTOPER(scan);
3805 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3807 maxcount = REG_INFTY;
3808 next = regnext(scan);
3809 scan = NEXTOPER(scan);
3813 if (flags & SCF_DO_SUBSTR)
3818 if (flags & SCF_DO_STCLASS) {
3820 maxcount = REG_INFTY;
3821 next = regnext(scan);
3822 scan = NEXTOPER(scan);
3825 is_inf = is_inf_internal = 1;
3826 scan = regnext(scan);
3827 if (flags & SCF_DO_SUBSTR) {
3828 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3829 data->longest = &(data->longest_float);
3831 goto optimize_curly_tail;
3833 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3834 && (scan->flags == stopparen))
3839 mincount = ARG1(scan);
3840 maxcount = ARG2(scan);
3842 next = regnext(scan);
3843 if (OP(scan) == CURLYX) {
3844 I32 lp = (data ? *(data->last_closep) : 0);
3845 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3847 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3848 next_is_eval = (OP(scan) == EVAL);
3850 if (flags & SCF_DO_SUBSTR) {
3851 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3852 pos_before = data->pos_min;
3856 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3858 data->flags |= SF_IS_INF;
3860 if (flags & SCF_DO_STCLASS) {
3861 cl_init(pRExC_state, &this_class);
3862 oclass = data->start_class;
3863 data->start_class = &this_class;
3864 f |= SCF_DO_STCLASS_AND;
3865 f &= ~SCF_DO_STCLASS_OR;
3867 /* Exclude from super-linear cache processing any {n,m}
3868 regops for which the combination of input pos and regex
3869 pos is not enough information to determine if a match
3872 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3873 regex pos at the \s*, the prospects for a match depend not
3874 only on the input position but also on how many (bar\s*)
3875 repeats into the {4,8} we are. */
3876 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3877 f &= ~SCF_WHILEM_VISITED_POS;
3879 /* This will finish on WHILEM, setting scan, or on NULL: */
3880 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3881 last, data, stopparen, recursed, NULL,
3883 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3885 if (flags & SCF_DO_STCLASS)
3886 data->start_class = oclass;
3887 if (mincount == 0 || minnext == 0) {
3888 if (flags & SCF_DO_STCLASS_OR) {
3889 cl_or(pRExC_state, data->start_class, &this_class);
3891 else if (flags & SCF_DO_STCLASS_AND) {
3892 /* Switch to OR mode: cache the old value of
3893 * data->start_class */
3895 StructCopy(data->start_class, and_withp,
3896 struct regnode_charclass_class);
3897 flags &= ~SCF_DO_STCLASS_AND;
3898 StructCopy(&this_class, data->start_class,
3899 struct regnode_charclass_class);
3900 flags |= SCF_DO_STCLASS_OR;
3901 SET_SSC_EOS(data->start_class);
3903 } else { /* Non-zero len */
3904 if (flags & SCF_DO_STCLASS_OR) {
3905 cl_or(pRExC_state, data->start_class, &this_class);
3906 cl_and(data->start_class, and_withp);
3908 else if (flags & SCF_DO_STCLASS_AND)
3909 cl_and(data->start_class, &this_class);
3910 flags &= ~SCF_DO_STCLASS;
3912 if (!scan) /* It was not CURLYX, but CURLY. */
3914 if (!(flags & SCF_TRIE_DOING_RESTUDY)
3915 /* ? quantifier ok, except for (?{ ... }) */
3916 && (next_is_eval || !(mincount == 0 && maxcount == 1))
3917 && (minnext == 0) && (deltanext == 0)
3918 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3919 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3921 /* Fatal warnings may leak the regexp without this: */
3922 SAVEFREESV(RExC_rx_sv);
3923 ckWARNreg(RExC_parse,
3924 "Quantifier unexpected on zero-length expression");
3925 (void)ReREFCNT_inc(RExC_rx_sv);
3928 min += minnext * mincount;
3929 is_inf_internal |= deltanext == I32_MAX
3930 || (maxcount == REG_INFTY && minnext + deltanext > 0);
3931 is_inf |= is_inf_internal;
3935 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3937 /* Try powerful optimization CURLYX => CURLYN. */
3938 if ( OP(oscan) == CURLYX && data
3939 && data->flags & SF_IN_PAR
3940 && !(data->flags & SF_HAS_EVAL)
3941 && !deltanext && minnext == 1 ) {
3942 /* Try to optimize to CURLYN. */
3943 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3944 regnode * const nxt1 = nxt;
3951 if (!REGNODE_SIMPLE(OP(nxt))
3952 && !(PL_regkind[OP(nxt)] == EXACT
3953 && STR_LEN(nxt) == 1))
3959 if (OP(nxt) != CLOSE)
3961 if (RExC_open_parens) {
3962 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3963 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3965 /* Now we know that nxt2 is the only contents: */
3966 oscan->flags = (U8)ARG(nxt);
3968 OP(nxt1) = NOTHING; /* was OPEN. */
3971 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3972 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3973 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3974 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3975 OP(nxt + 1) = OPTIMIZED; /* was count. */
3976 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3981 /* Try optimization CURLYX => CURLYM. */
3982 if ( OP(oscan) == CURLYX && data
3983 && !(data->flags & SF_HAS_PAR)
3984 && !(data->flags & SF_HAS_EVAL)
3985 && !deltanext /* atom is fixed width */
3986 && minnext != 0 /* CURLYM can't handle zero width */
3987 && ! (RExC_seen & REG_SEEN_EXACTF_SHARP_S) /* Nor \xDF */
3989 /* XXXX How to optimize if data == 0? */
3990 /* Optimize to a simpler form. */
3991 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3995 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3996 && (OP(nxt2) != WHILEM))
3998 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3999 /* Need to optimize away parenths. */
4000 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
4001 /* Set the parenth number. */
4002 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
4004 oscan->flags = (U8)ARG(nxt);
4005 if (RExC_open_parens) {
4006 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4007 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4009 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4010 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4013 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4014 OP(nxt + 1) = OPTIMIZED; /* was count. */
4015 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4016 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4019 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4020 regnode *nnxt = regnext(nxt1);
4022 if (reg_off_by_arg[OP(nxt1)])
4023 ARG_SET(nxt1, nxt2 - nxt1);
4024 else if (nxt2 - nxt1 < U16_MAX)
4025 NEXT_OFF(nxt1) = nxt2 - nxt1;
4027 OP(nxt) = NOTHING; /* Cannot beautify */
4032 /* Optimize again: */
4033 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4034 NULL, stopparen, recursed, NULL, 0,depth+1);
4039 else if ((OP(oscan) == CURLYX)
4040 && (flags & SCF_WHILEM_VISITED_POS)
4041 /* See the comment on a similar expression above.
4042 However, this time it's not a subexpression
4043 we care about, but the expression itself. */
4044 && (maxcount == REG_INFTY)
4045 && data && ++data->whilem_c < 16) {
4046 /* This stays as CURLYX, we can put the count/of pair. */
4047 /* Find WHILEM (as in regexec.c) */
4048 regnode *nxt = oscan + NEXT_OFF(oscan);
4050 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4052 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4053 | (RExC_whilem_seen << 4)); /* On WHILEM */
4055 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4057 if (flags & SCF_DO_SUBSTR) {
4058 SV *last_str = NULL;
4059 int counted = mincount != 0;
4061 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
4062 #if defined(SPARC64_GCC_WORKAROUND)
4065 const char *s = NULL;
4068 if (pos_before >= data->last_start_min)
4071 b = data->last_start_min;
4074 s = SvPV_const(data->last_found, l);
4075 old = b - data->last_start_min;
4078 I32 b = pos_before >= data->last_start_min
4079 ? pos_before : data->last_start_min;
4081 const char * const s = SvPV_const(data->last_found, l);
4082 I32 old = b - data->last_start_min;
4086 old = utf8_hop((U8*)s, old) - (U8*)s;
4088 /* Get the added string: */
4089 last_str = newSVpvn_utf8(s + old, l, UTF);
4090 if (deltanext == 0 && pos_before == b) {
4091 /* What was added is a constant string */
4093 SvGROW(last_str, (mincount * l) + 1);
4094 repeatcpy(SvPVX(last_str) + l,
4095 SvPVX_const(last_str), l, mincount - 1);
4096 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4097 /* Add additional parts. */
4098 SvCUR_set(data->last_found,
4099 SvCUR(data->last_found) - l);
4100 sv_catsv(data->last_found, last_str);
4102 SV * sv = data->last_found;
4104 SvUTF8(sv) && SvMAGICAL(sv) ?
4105 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4106 if (mg && mg->mg_len >= 0)
4107 mg->mg_len += CHR_SVLEN(last_str) - l;
4109 data->last_end += l * (mincount - 1);
4112 /* start offset must point into the last copy */
4113 data->last_start_min += minnext * (mincount - 1);
4114 data->last_start_max += is_inf ? I32_MAX
4115 : (maxcount - 1) * (minnext + data->pos_delta);
4118 /* It is counted once already... */
4119 data->pos_min += minnext * (mincount - counted);
4121 PerlIO_printf(Perl_debug_log, "counted=%d deltanext=%d I32_MAX=%d minnext=%d maxcount=%d mincount=%d\n",
4122 counted, deltanext, I32_MAX, minnext, maxcount, mincount);
4123 if (deltanext != I32_MAX)
4124 PerlIO_printf(Perl_debug_log, "LHS=%d RHS=%d\n", -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount, I32_MAX - data->pos_delta);
4126 if (deltanext == I32_MAX || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= I32_MAX - data->pos_delta)
4127 data->pos_delta = I32_MAX;
4129 data->pos_delta += - counted * deltanext +
4130 (minnext + deltanext) * maxcount - minnext * mincount;
4131 if (mincount != maxcount) {
4132 /* Cannot extend fixed substrings found inside
4134 SCAN_COMMIT(pRExC_state,data,minlenp);
4135 if (mincount && last_str) {
4136 SV * const sv = data->last_found;
4137 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4138 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4142 sv_setsv(sv, last_str);
4143 data->last_end = data->pos_min;
4144 data->last_start_min =
4145 data->pos_min - CHR_SVLEN(last_str);
4146 data->last_start_max = is_inf
4148 : data->pos_min + data->pos_delta
4149 - CHR_SVLEN(last_str);
4151 data->longest = &(data->longest_float);
4153 SvREFCNT_dec(last_str);
4155 if (data && (fl & SF_HAS_EVAL))
4156 data->flags |= SF_HAS_EVAL;
4157 optimize_curly_tail:
4158 if (OP(oscan) != CURLYX) {
4159 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4161 NEXT_OFF(oscan) += NEXT_OFF(next);
4164 default: /* REF, and CLUMP only? */
4165 if (flags & SCF_DO_SUBSTR) {
4166 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4167 data->longest = &(data->longest_float);
4169 is_inf = is_inf_internal = 1;
4170 if (flags & SCF_DO_STCLASS_OR)
4171 cl_anything(pRExC_state, data->start_class);
4172 flags &= ~SCF_DO_STCLASS;
4176 else if (OP(scan) == LNBREAK) {
4177 if (flags & SCF_DO_STCLASS) {
4179 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4180 if (flags & SCF_DO_STCLASS_AND) {
4181 for (value = 0; value < 256; value++)
4182 if (!is_VERTWS_cp(value))
4183 ANYOF_BITMAP_CLEAR(data->start_class, value);
4186 for (value = 0; value < 256; value++)
4187 if (is_VERTWS_cp(value))
4188 ANYOF_BITMAP_SET(data->start_class, value);
4190 if (flags & SCF_DO_STCLASS_OR)
4191 cl_and(data->start_class, and_withp);
4192 flags &= ~SCF_DO_STCLASS;
4195 delta++; /* Because of the 2 char string cr-lf */
4196 if (flags & SCF_DO_SUBSTR) {
4197 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4199 data->pos_delta += 1;
4200 data->longest = &(data->longest_float);
4203 else if (REGNODE_SIMPLE(OP(scan))) {
4206 if (flags & SCF_DO_SUBSTR) {
4207 SCAN_COMMIT(pRExC_state,data,minlenp);
4211 if (flags & SCF_DO_STCLASS) {
4213 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4215 /* Some of the logic below assumes that switching
4216 locale on will only add false positives. */
4217 switch (PL_regkind[OP(scan)]) {
4223 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan));
4226 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4227 cl_anything(pRExC_state, data->start_class);
4230 if (OP(scan) == SANY)
4232 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
4233 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
4234 || ANYOF_CLASS_TEST_ANY_SET(data->start_class));
4235 cl_anything(pRExC_state, data->start_class);
4237 if (flags & SCF_DO_STCLASS_AND || !value)
4238 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
4241 if (flags & SCF_DO_STCLASS_AND)
4242 cl_and(data->start_class,
4243 (struct regnode_charclass_class*)scan);
4245 cl_or(pRExC_state, data->start_class,
4246 (struct regnode_charclass_class*)scan);
4254 classnum = FLAGS(scan);
4255 if (flags & SCF_DO_STCLASS_AND) {
4256 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4257 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum) + 1);
4258 for (value = 0; value < loop_max; value++) {
4259 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4260 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4266 if (data->start_class->flags & ANYOF_LOCALE) {
4267 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum));
4271 /* Even if under locale, set the bits for non-locale
4272 * in case it isn't a true locale-node. This will
4273 * create false positives if it truly is locale */
4274 for (value = 0; value < loop_max; value++) {
4275 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4276 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4288 classnum = FLAGS(scan);
4289 if (flags & SCF_DO_STCLASS_AND) {
4290 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4291 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum));
4292 for (value = 0; value < loop_max; value++) {
4293 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4294 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4300 if (data->start_class->flags & ANYOF_LOCALE) {
4301 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum) + 1);
4305 /* Even if under locale, set the bits for non-locale in
4306 * case it isn't a true locale-node. This will create
4307 * false positives if it truly is locale */
4308 for (value = 0; value < loop_max; value++) {
4309 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4310 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4313 if (PL_regkind[OP(scan)] == NPOSIXD) {
4314 data->start_class->flags |= ANYOF_NON_UTF8_LATIN1_ALL;
4320 if (flags & SCF_DO_STCLASS_OR)
4321 cl_and(data->start_class, and_withp);
4322 flags &= ~SCF_DO_STCLASS;
4325 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
4326 data->flags |= (OP(scan) == MEOL
4329 SCAN_COMMIT(pRExC_state, data, minlenp);
4332 else if ( PL_regkind[OP(scan)] == BRANCHJ
4333 /* Lookbehind, or need to calculate parens/evals/stclass: */
4334 && (scan->flags || data || (flags & SCF_DO_STCLASS))
4335 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
4336 if ( OP(scan) == UNLESSM &&
4338 OP(NEXTOPER(NEXTOPER(scan))) == NOTHING &&
4339 OP(regnext(NEXTOPER(NEXTOPER(scan)))) == SUCCEED
4342 regnode *upto= regnext(scan);
4344 SV * const mysv_val=sv_newmortal();
4345 DEBUG_STUDYDATA("OPFAIL",data,depth);
4347 /*DEBUG_PARSE_MSG("opfail");*/
4348 regprop(RExC_rx, mysv_val, upto);
4349 PerlIO_printf(Perl_debug_log, "~ replace with OPFAIL pointed at %s (%"IVdf") offset %"IVdf"\n",
4350 SvPV_nolen_const(mysv_val),
4351 (IV)REG_NODE_NUM(upto),
4356 NEXT_OFF(scan) = upto - scan;
4357 for (opt= scan + 1; opt < upto ; opt++)
4358 OP(opt) = OPTIMIZED;
4362 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4363 || OP(scan) == UNLESSM )
4365 /* Negative Lookahead/lookbehind
4366 In this case we can't do fixed string optimisation.
4369 I32 deltanext, minnext, fake = 0;
4371 struct regnode_charclass_class intrnl;
4374 data_fake.flags = 0;
4376 data_fake.whilem_c = data->whilem_c;
4377 data_fake.last_closep = data->last_closep;
4380 data_fake.last_closep = &fake;
4381 data_fake.pos_delta = delta;
4382 if ( flags & SCF_DO_STCLASS && !scan->flags
4383 && OP(scan) == IFMATCH ) { /* Lookahead */
4384 cl_init(pRExC_state, &intrnl);
4385 data_fake.start_class = &intrnl;
4386 f |= SCF_DO_STCLASS_AND;
4388 if (flags & SCF_WHILEM_VISITED_POS)
4389 f |= SCF_WHILEM_VISITED_POS;
4390 next = regnext(scan);
4391 nscan = NEXTOPER(NEXTOPER(scan));
4392 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
4393 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
4396 FAIL("Variable length lookbehind not implemented");
4398 else if (minnext > (I32)U8_MAX) {
4399 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4401 scan->flags = (U8)minnext;
4404 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4406 if (data_fake.flags & SF_HAS_EVAL)
4407 data->flags |= SF_HAS_EVAL;
4408 data->whilem_c = data_fake.whilem_c;
4410 if (f & SCF_DO_STCLASS_AND) {
4411 if (flags & SCF_DO_STCLASS_OR) {
4412 /* OR before, AND after: ideally we would recurse with
4413 * data_fake to get the AND applied by study of the
4414 * remainder of the pattern, and then derecurse;
4415 * *** HACK *** for now just treat as "no information".
4416 * See [perl #56690].
4418 cl_init(pRExC_state, data->start_class);
4420 /* AND before and after: combine and continue */
4421 const int was = TEST_SSC_EOS(data->start_class);
4423 cl_and(data->start_class, &intrnl);
4425 SET_SSC_EOS(data->start_class);
4429 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4431 /* Positive Lookahead/lookbehind
4432 In this case we can do fixed string optimisation,
4433 but we must be careful about it. Note in the case of
4434 lookbehind the positions will be offset by the minimum
4435 length of the pattern, something we won't know about
4436 until after the recurse.
4438 I32 deltanext, fake = 0;
4440 struct regnode_charclass_class intrnl;
4442 /* We use SAVEFREEPV so that when the full compile
4443 is finished perl will clean up the allocated
4444 minlens when it's all done. This way we don't
4445 have to worry about freeing them when we know
4446 they wont be used, which would be a pain.
4449 Newx( minnextp, 1, I32 );
4450 SAVEFREEPV(minnextp);
4453 StructCopy(data, &data_fake, scan_data_t);
4454 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
4457 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
4458 data_fake.last_found=newSVsv(data->last_found);
4462 data_fake.last_closep = &fake;
4463 data_fake.flags = 0;
4464 data_fake.pos_delta = delta;
4466 data_fake.flags |= SF_IS_INF;
4467 if ( flags & SCF_DO_STCLASS && !scan->flags
4468 && OP(scan) == IFMATCH ) { /* Lookahead */
4469 cl_init(pRExC_state, &intrnl);
4470 data_fake.start_class = &intrnl;
4471 f |= SCF_DO_STCLASS_AND;
4473 if (flags & SCF_WHILEM_VISITED_POS)
4474 f |= SCF_WHILEM_VISITED_POS;
4475 next = regnext(scan);
4476 nscan = NEXTOPER(NEXTOPER(scan));
4478 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
4479 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
4482 FAIL("Variable length lookbehind not implemented");
4484 else if (*minnextp > (I32)U8_MAX) {
4485 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4487 scan->flags = (U8)*minnextp;
4492 if (f & SCF_DO_STCLASS_AND) {
4493 const int was = TEST_SSC_EOS(data.start_class);
4495 cl_and(data->start_class, &intrnl);
4497 SET_SSC_EOS(data->start_class);
4500 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4502 if (data_fake.flags & SF_HAS_EVAL)
4503 data->flags |= SF_HAS_EVAL;
4504 data->whilem_c = data_fake.whilem_c;
4505 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
4506 if (RExC_rx->minlen<*minnextp)
4507 RExC_rx->minlen=*minnextp;
4508 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
4509 SvREFCNT_dec_NN(data_fake.last_found);
4511 if ( data_fake.minlen_fixed != minlenp )
4513 data->offset_fixed= data_fake.offset_fixed;
4514 data->minlen_fixed= data_fake.minlen_fixed;
4515 data->lookbehind_fixed+= scan->flags;
4517 if ( data_fake.minlen_float != minlenp )
4519 data->minlen_float= data_fake.minlen_float;
4520 data->offset_float_min=data_fake.offset_float_min;
4521 data->offset_float_max=data_fake.offset_float_max;
4522 data->lookbehind_float+= scan->flags;
4529 else if (OP(scan) == OPEN) {
4530 if (stopparen != (I32)ARG(scan))
4533 else if (OP(scan) == CLOSE) {
4534 if (stopparen == (I32)ARG(scan)) {
4537 if ((I32)ARG(scan) == is_par) {
4538 next = regnext(scan);
4540 if ( next && (OP(next) != WHILEM) && next < last)
4541 is_par = 0; /* Disable optimization */
4544 *(data->last_closep) = ARG(scan);
4546 else if (OP(scan) == EVAL) {
4548 data->flags |= SF_HAS_EVAL;
4550 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
4551 if (flags & SCF_DO_SUBSTR) {
4552 SCAN_COMMIT(pRExC_state,data,minlenp);
4553 flags &= ~SCF_DO_SUBSTR;
4555 if (data && OP(scan)==ACCEPT) {
4556 data->flags |= SCF_SEEN_ACCEPT;
4561 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
4563 if (flags & SCF_DO_SUBSTR) {
4564 SCAN_COMMIT(pRExC_state,data,minlenp);
4565 data->longest = &(data->longest_float);
4567 is_inf = is_inf_internal = 1;
4568 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4569 cl_anything(pRExC_state, data->start_class);
4570 flags &= ~SCF_DO_STCLASS;
4572 else if (OP(scan) == GPOS) {
4573 if (!(RExC_rx->extflags & RXf_GPOS_FLOAT) &&
4574 !(delta || is_inf || (data && data->pos_delta)))
4576 if (!(RExC_rx->extflags & RXf_ANCH) && (flags & SCF_DO_SUBSTR))
4577 RExC_rx->extflags |= RXf_ANCH_GPOS;
4578 if (RExC_rx->gofs < (U32)min)
4579 RExC_rx->gofs = min;
4581 RExC_rx->extflags |= RXf_GPOS_FLOAT;
4585 #ifdef TRIE_STUDY_OPT
4586 #ifdef FULL_TRIE_STUDY
4587 else if (PL_regkind[OP(scan)] == TRIE) {
4588 /* NOTE - There is similar code to this block above for handling
4589 BRANCH nodes on the initial study. If you change stuff here
4591 regnode *trie_node= scan;
4592 regnode *tail= regnext(scan);
4593 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4594 I32 max1 = 0, min1 = I32_MAX;
4595 struct regnode_charclass_class accum;
4597 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
4598 SCAN_COMMIT(pRExC_state, data,minlenp); /* Cannot merge strings after this. */
4599 if (flags & SCF_DO_STCLASS)
4600 cl_init_zero(pRExC_state, &accum);
4606 const regnode *nextbranch= NULL;
4609 for ( word=1 ; word <= trie->wordcount ; word++)
4611 I32 deltanext=0, minnext=0, f = 0, fake;
4612 struct regnode_charclass_class this_class;
4614 data_fake.flags = 0;
4616 data_fake.whilem_c = data->whilem_c;
4617 data_fake.last_closep = data->last_closep;
4620 data_fake.last_closep = &fake;
4621 data_fake.pos_delta = delta;
4622 if (flags & SCF_DO_STCLASS) {
4623 cl_init(pRExC_state, &this_class);
4624 data_fake.start_class = &this_class;
4625 f = SCF_DO_STCLASS_AND;
4627 if (flags & SCF_WHILEM_VISITED_POS)
4628 f |= SCF_WHILEM_VISITED_POS;
4630 if (trie->jump[word]) {
4632 nextbranch = trie_node + trie->jump[0];
4633 scan= trie_node + trie->jump[word];
4634 /* We go from the jump point to the branch that follows
4635 it. Note this means we need the vestigal unused branches
4636 even though they arent otherwise used.
4638 minnext = study_chunk(pRExC_state, &scan, minlenp,
4639 &deltanext, (regnode *)nextbranch, &data_fake,
4640 stopparen, recursed, NULL, f,depth+1);
4642 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
4643 nextbranch= regnext((regnode*)nextbranch);
4645 if (min1 > (I32)(minnext + trie->minlen))
4646 min1 = minnext + trie->minlen;
4647 if (deltanext == I32_MAX) {
4648 is_inf = is_inf_internal = 1;
4650 } else if (max1 < (I32)(minnext + deltanext + trie->maxlen))
4651 max1 = minnext + deltanext + trie->maxlen;
4653 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4655 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4656 if ( stopmin > min + min1)
4657 stopmin = min + min1;
4658 flags &= ~SCF_DO_SUBSTR;
4660 data->flags |= SCF_SEEN_ACCEPT;
4663 if (data_fake.flags & SF_HAS_EVAL)
4664 data->flags |= SF_HAS_EVAL;
4665 data->whilem_c = data_fake.whilem_c;
4667 if (flags & SCF_DO_STCLASS)
4668 cl_or(pRExC_state, &accum, &this_class);
4671 if (flags & SCF_DO_SUBSTR) {
4672 data->pos_min += min1;
4673 data->pos_delta += max1 - min1;
4674 if (max1 != min1 || is_inf)
4675 data->longest = &(data->longest_float);
4678 delta += max1 - min1;
4679 if (flags & SCF_DO_STCLASS_OR) {
4680 cl_or(pRExC_state, data->start_class, &accum);
4682 cl_and(data->start_class, and_withp);
4683 flags &= ~SCF_DO_STCLASS;
4686 else if (flags & SCF_DO_STCLASS_AND) {
4688 cl_and(data->start_class, &accum);
4689 flags &= ~SCF_DO_STCLASS;
4692 /* Switch to OR mode: cache the old value of
4693 * data->start_class */
4695 StructCopy(data->start_class, and_withp,
4696 struct regnode_charclass_class);
4697 flags &= ~SCF_DO_STCLASS_AND;
4698 StructCopy(&accum, data->start_class,
4699 struct regnode_charclass_class);
4700 flags |= SCF_DO_STCLASS_OR;
4701 SET_SSC_EOS(data->start_class);
4708 else if (PL_regkind[OP(scan)] == TRIE) {
4709 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4712 min += trie->minlen;
4713 delta += (trie->maxlen - trie->minlen);
4714 flags &= ~SCF_DO_STCLASS; /* xxx */
4715 if (flags & SCF_DO_SUBSTR) {
4716 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4717 data->pos_min += trie->minlen;
4718 data->pos_delta += (trie->maxlen - trie->minlen);
4719 if (trie->maxlen != trie->minlen)
4720 data->longest = &(data->longest_float);
4722 if (trie->jump) /* no more substrings -- for now /grr*/
4723 flags &= ~SCF_DO_SUBSTR;
4725 #endif /* old or new */
4726 #endif /* TRIE_STUDY_OPT */
4728 /* Else: zero-length, ignore. */
4729 scan = regnext(scan);
4734 stopparen = frame->stop;
4735 frame = frame->prev;
4736 goto fake_study_recurse;
4741 DEBUG_STUDYDATA("pre-fin:",data,depth);
4744 *deltap = is_inf_internal ? I32_MAX : delta;
4745 if (flags & SCF_DO_SUBSTR && is_inf)
4746 data->pos_delta = I32_MAX - data->pos_min;
4747 if (is_par > (I32)U8_MAX)
4749 if (is_par && pars==1 && data) {
4750 data->flags |= SF_IN_PAR;
4751 data->flags &= ~SF_HAS_PAR;
4753 else if (pars && data) {
4754 data->flags |= SF_HAS_PAR;
4755 data->flags &= ~SF_IN_PAR;
4757 if (flags & SCF_DO_STCLASS_OR)
4758 cl_and(data->start_class, and_withp);
4759 if (flags & SCF_TRIE_RESTUDY)
4760 data->flags |= SCF_TRIE_RESTUDY;
4762 DEBUG_STUDYDATA("post-fin:",data,depth);
4764 return min < stopmin ? min : stopmin;
4768 S_add_data(RExC_state_t *pRExC_state, U32 n, const char *s)
4770 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
4772 PERL_ARGS_ASSERT_ADD_DATA;
4774 Renewc(RExC_rxi->data,
4775 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
4776 char, struct reg_data);
4778 Renew(RExC_rxi->data->what, count + n, U8);
4780 Newx(RExC_rxi->data->what, n, U8);
4781 RExC_rxi->data->count = count + n;
4782 Copy(s, RExC_rxi->data->what + count, n, U8);
4786 /*XXX: todo make this not included in a non debugging perl */
4787 #ifndef PERL_IN_XSUB_RE
4789 Perl_reginitcolors(pTHX)
4792 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
4794 char *t = savepv(s);
4798 t = strchr(t, '\t');
4804 PL_colors[i] = t = (char *)"";
4809 PL_colors[i++] = (char *)"";
4816 #ifdef TRIE_STUDY_OPT
4817 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
4820 (data.flags & SCF_TRIE_RESTUDY) \
4828 #define CHECK_RESTUDY_GOTO_butfirst
4832 * pregcomp - compile a regular expression into internal code
4834 * Decides which engine's compiler to call based on the hint currently in
4838 #ifndef PERL_IN_XSUB_RE
4840 /* return the currently in-scope regex engine (or the default if none) */
4842 regexp_engine const *
4843 Perl_current_re_engine(pTHX)
4847 if (IN_PERL_COMPILETIME) {
4848 HV * const table = GvHV(PL_hintgv);
4852 return &reh_regexp_engine;
4853 ptr = hv_fetchs(table, "regcomp", FALSE);
4854 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
4855 return &reh_regexp_engine;
4856 return INT2PTR(regexp_engine*,SvIV(*ptr));
4860 if (!PL_curcop->cop_hints_hash)
4861 return &reh_regexp_engine;
4862 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
4863 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
4864 return &reh_regexp_engine;
4865 return INT2PTR(regexp_engine*,SvIV(ptr));
4871 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
4874 regexp_engine const *eng = current_re_engine();
4875 GET_RE_DEBUG_FLAGS_DECL;
4877 PERL_ARGS_ASSERT_PREGCOMP;
4879 /* Dispatch a request to compile a regexp to correct regexp engine. */
4881 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
4884 return CALLREGCOMP_ENG(eng, pattern, flags);
4888 /* public(ish) entry point for the perl core's own regex compiling code.
4889 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
4890 * pattern rather than a list of OPs, and uses the internal engine rather
4891 * than the current one */
4894 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
4896 SV *pat = pattern; /* defeat constness! */
4897 PERL_ARGS_ASSERT_RE_COMPILE;
4898 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
4899 #ifdef PERL_IN_XSUB_RE
4904 NULL, NULL, rx_flags, 0);
4908 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
4909 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
4910 * point to the realloced string and length.
4912 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
4916 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
4917 char **pat_p, STRLEN *plen_p, int num_code_blocks)
4919 U8 *const src = (U8*)*pat_p;
4922 STRLEN s = 0, d = 0;
4924 GET_RE_DEBUG_FLAGS_DECL;
4926 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
4927 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
4929 Newx(dst, *plen_p * 2 + 1, U8);
4931 while (s < *plen_p) {
4932 const UV uv = NATIVE_TO_ASCII(src[s]);
4933 if (UNI_IS_INVARIANT(uv))
4934 dst[d] = (U8)UTF_TO_NATIVE(uv);
4936 dst[d++] = (U8)UTF8_EIGHT_BIT_HI(uv);
4937 dst[d] = (U8)UTF8_EIGHT_BIT_LO(uv);
4939 if (n < num_code_blocks) {
4940 if (!do_end && pRExC_state->code_blocks[n].start == s) {
4941 pRExC_state->code_blocks[n].start = d;
4942 assert(dst[d] == '(');
4945 else if (do_end && pRExC_state->code_blocks[n].end == s) {
4946 pRExC_state->code_blocks[n].end = d;
4947 assert(dst[d] == ')');
4957 *pat_p = (char*) dst;
4959 RExC_orig_utf8 = RExC_utf8 = 1;
4964 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
4965 * while recording any code block indices, and handling overloading,
4966 * nested qr// objects etc. If pat is null, it will allocate a new
4967 * string, or just return the first arg, if there's only one.
4969 * Returns the malloced/updated pat.
4970 * patternp and pat_count is the array of SVs to be concatted;
4971 * oplist is the optional list of ops that generated the SVs;
4972 * recompile_p is a pointer to a boolean that will be set if
4973 * the regex will need to be recompiled.
4974 * delim, if non-null is an SV that will be inserted between each element
4978 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
4979 SV *pat, SV ** const patternp, int pat_count,
4980 OP *oplist, bool *recompile_p, SV *delim)
4984 bool use_delim = FALSE;
4985 bool alloced = FALSE;
4987 /* if we know we have at least two args, create an empty string,
4988 * then concatenate args to that. For no args, return an empty string */
4989 if (!pat && pat_count != 1) {
4990 pat = newSVpvn("", 0);
4995 for (svp = patternp; svp < patternp + pat_count; svp++) {
4998 STRLEN orig_patlen = 0;
5000 SV *msv = use_delim ? delim : *svp;
5002 /* if we've got a delimiter, we go round the loop twice for each
5003 * svp slot (except the last), using the delimiter the second
5012 if (SvTYPE(msv) == SVt_PVAV) {
5013 /* we've encountered an interpolated array within
5014 * the pattern, e.g. /...@a..../. Expand the list of elements,
5015 * then recursively append elements.
5016 * The code in this block is based on S_pushav() */
5018 AV *const av = (AV*)msv;
5019 const I32 maxarg = AvFILL(av) + 1;
5023 assert(oplist->op_type == OP_PADAV
5024 || oplist->op_type == OP_RV2AV);
5025 oplist = oplist->op_sibling;;
5028 if (SvRMAGICAL(av)) {
5031 Newx(array, maxarg, SV*);
5033 for (i=0; i < (U32)maxarg; i++) {
5034 SV ** const svp = av_fetch(av, i, FALSE);
5035 array[i] = svp ? *svp : &PL_sv_undef;
5039 array = AvARRAY(av);
5041 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5042 array, maxarg, NULL, recompile_p,
5044 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5050 /* we make the assumption here that each op in the list of
5051 * op_siblings maps to one SV pushed onto the stack,
5052 * except for code blocks, with have both an OP_NULL and
5054 * This allows us to match up the list of SVs against the
5055 * list of OPs to find the next code block.
5057 * Note that PUSHMARK PADSV PADSV ..
5059 * PADRANGE PADSV PADSV ..
5060 * so the alignment still works. */
5063 if (oplist->op_type == OP_NULL
5064 && (oplist->op_flags & OPf_SPECIAL))
5066 assert(n < pRExC_state->num_code_blocks);
5067 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5068 pRExC_state->code_blocks[n].block = oplist;
5069 pRExC_state->code_blocks[n].src_regex = NULL;
5072 oplist = oplist->op_sibling; /* skip CONST */
5075 oplist = oplist->op_sibling;;
5078 /* apply magic and QR overloading to arg */
5081 if (SvROK(msv) && SvAMAGIC(msv)) {
5082 SV *sv = AMG_CALLunary(msv, regexp_amg);
5086 if (SvTYPE(sv) != SVt_REGEXP)
5087 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5092 /* try concatenation overload ... */
5093 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5094 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5097 /* overloading involved: all bets are off over literal
5098 * code. Pretend we haven't seen it */
5099 pRExC_state->num_code_blocks -= n;
5103 /* ... or failing that, try "" overload */
5104 while (SvAMAGIC(msv)
5105 && (sv = AMG_CALLunary(msv, string_amg))
5109 && SvRV(msv) == SvRV(sv))
5114 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5118 /* this is a partially unrolled
5119 * sv_catsv_nomg(pat, msv);
5120 * that allows us to adjust code block indices if
5123 char *dst = SvPV_force_nomg(pat, dlen);
5125 if (SvUTF8(msv) && !SvUTF8(pat)) {
5126 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
5127 sv_setpvn(pat, dst, dlen);
5130 sv_catsv_nomg(pat, msv);
5137 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5140 /* extract any code blocks within any embedded qr//'s */
5141 if (rx && SvTYPE(rx) == SVt_REGEXP
5142 && RX_ENGINE((REGEXP*)rx)->op_comp)
5145 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5146 if (ri->num_code_blocks) {
5148 /* the presence of an embedded qr// with code means
5149 * we should always recompile: the text of the
5150 * qr// may not have changed, but it may be a
5151 * different closure than last time */
5153 Renew(pRExC_state->code_blocks,
5154 pRExC_state->num_code_blocks + ri->num_code_blocks,
5155 struct reg_code_block);
5156 pRExC_state->num_code_blocks += ri->num_code_blocks;
5158 for (i=0; i < ri->num_code_blocks; i++) {
5159 struct reg_code_block *src, *dst;
5160 STRLEN offset = orig_patlen
5161 + ReANY((REGEXP *)rx)->pre_prefix;
5162 assert(n < pRExC_state->num_code_blocks);
5163 src = &ri->code_blocks[i];
5164 dst = &pRExC_state->code_blocks[n];
5165 dst->start = src->start + offset;
5166 dst->end = src->end + offset;
5167 dst->block = src->block;
5168 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5177 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
5186 /* see if there are any run-time code blocks in the pattern.
5187 * False positives are allowed */
5190 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5191 char *pat, STRLEN plen)
5196 for (s = 0; s < plen; s++) {
5197 if (n < pRExC_state->num_code_blocks
5198 && s == pRExC_state->code_blocks[n].start)
5200 s = pRExC_state->code_blocks[n].end;
5204 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
5206 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
5208 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
5215 /* Handle run-time code blocks. We will already have compiled any direct
5216 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
5217 * copy of it, but with any literal code blocks blanked out and
5218 * appropriate chars escaped; then feed it into
5220 * eval "qr'modified_pattern'"
5224 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
5228 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
5230 * After eval_sv()-ing that, grab any new code blocks from the returned qr
5231 * and merge them with any code blocks of the original regexp.
5233 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
5234 * instead, just save the qr and return FALSE; this tells our caller that
5235 * the original pattern needs upgrading to utf8.
5239 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5240 char *pat, STRLEN plen)
5244 GET_RE_DEBUG_FLAGS_DECL;
5246 if (pRExC_state->runtime_code_qr) {
5247 /* this is the second time we've been called; this should
5248 * only happen if the main pattern got upgraded to utf8
5249 * during compilation; re-use the qr we compiled first time
5250 * round (which should be utf8 too)
5252 qr = pRExC_state->runtime_code_qr;
5253 pRExC_state->runtime_code_qr = NULL;
5254 assert(RExC_utf8 && SvUTF8(qr));
5260 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
5264 /* determine how many extra chars we need for ' and \ escaping */
5265 for (s = 0; s < plen; s++) {
5266 if (pat[s] == '\'' || pat[s] == '\\')
5270 Newx(newpat, newlen, char);
5272 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
5274 for (s = 0; s < plen; s++) {
5275 if (n < pRExC_state->num_code_blocks
5276 && s == pRExC_state->code_blocks[n].start)
5278 /* blank out literal code block */
5279 assert(pat[s] == '(');
5280 while (s <= pRExC_state->code_blocks[n].end) {
5288 if (pat[s] == '\'' || pat[s] == '\\')
5293 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
5297 PerlIO_printf(Perl_debug_log,
5298 "%sre-parsing pattern for runtime code:%s %s\n",
5299 PL_colors[4],PL_colors[5],newpat);
5302 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5308 PUSHSTACKi(PERLSI_REQUIRE);
5309 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
5310 * parsing qr''; normally only q'' does this. It also alters
5312 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
5313 SvREFCNT_dec_NN(sv);
5318 SV * const errsv = ERRSV;
5319 if (SvTRUE_NN(errsv))
5321 Safefree(pRExC_state->code_blocks);
5322 /* use croak_sv ? */
5323 Perl_croak_nocontext("%s", SvPV_nolen_const(errsv));
5326 assert(SvROK(qr_ref));
5328 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5329 /* the leaving below frees the tmp qr_ref.
5330 * Give qr a life of its own */
5338 if (!RExC_utf8 && SvUTF8(qr)) {
5339 /* first time through; the pattern got upgraded; save the
5340 * qr for the next time through */
5341 assert(!pRExC_state->runtime_code_qr);
5342 pRExC_state->runtime_code_qr = qr;
5347 /* extract any code blocks within the returned qr// */
5350 /* merge the main (r1) and run-time (r2) code blocks into one */
5352 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
5353 struct reg_code_block *new_block, *dst;
5354 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
5357 if (!r2->num_code_blocks) /* we guessed wrong */
5359 SvREFCNT_dec_NN(qr);
5364 r1->num_code_blocks + r2->num_code_blocks,
5365 struct reg_code_block);
5368 while ( i1 < r1->num_code_blocks
5369 || i2 < r2->num_code_blocks)
5371 struct reg_code_block *src;
5374 if (i1 == r1->num_code_blocks) {
5375 src = &r2->code_blocks[i2++];
5378 else if (i2 == r2->num_code_blocks)
5379 src = &r1->code_blocks[i1++];
5380 else if ( r1->code_blocks[i1].start
5381 < r2->code_blocks[i2].start)
5383 src = &r1->code_blocks[i1++];
5384 assert(src->end < r2->code_blocks[i2].start);
5387 assert( r1->code_blocks[i1].start
5388 > r2->code_blocks[i2].start);
5389 src = &r2->code_blocks[i2++];
5391 assert(src->end < r1->code_blocks[i1].start);
5394 assert(pat[src->start] == '(');
5395 assert(pat[src->end] == ')');
5396 dst->start = src->start;
5397 dst->end = src->end;
5398 dst->block = src->block;
5399 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
5403 r1->num_code_blocks += r2->num_code_blocks;
5404 Safefree(r1->code_blocks);
5405 r1->code_blocks = new_block;
5408 SvREFCNT_dec_NN(qr);
5414 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)
5416 /* This is the common code for setting up the floating and fixed length
5417 * string data extracted from Perl_re_op_compile() below. Returns a boolean
5418 * as to whether succeeded or not */
5422 if (! (longest_length
5423 || (eol /* Can't have SEOL and MULTI */
5424 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
5426 /* See comments for join_exact for why REG_SEEN_EXACTF_SHARP_S */
5427 || (RExC_seen & REG_SEEN_EXACTF_SHARP_S))
5432 /* copy the information about the longest from the reg_scan_data
5433 over to the program. */
5434 if (SvUTF8(sv_longest)) {
5435 *rx_utf8 = sv_longest;
5438 *rx_substr = sv_longest;
5441 /* end_shift is how many chars that must be matched that
5442 follow this item. We calculate it ahead of time as once the
5443 lookbehind offset is added in we lose the ability to correctly
5445 ml = minlen ? *(minlen) : (I32)longest_length;
5446 *rx_end_shift = ml - offset
5447 - longest_length + (SvTAIL(sv_longest) != 0)
5450 t = (eol/* Can't have SEOL and MULTI */
5451 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
5452 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
5458 * Perl_re_op_compile - the perl internal RE engine's function to compile a
5459 * regular expression into internal code.
5460 * The pattern may be passed either as:
5461 * a list of SVs (patternp plus pat_count)
5462 * a list of OPs (expr)
5463 * If both are passed, the SV list is used, but the OP list indicates
5464 * which SVs are actually pre-compiled code blocks
5466 * The SVs in the list have magic and qr overloading applied to them (and
5467 * the list may be modified in-place with replacement SVs in the latter
5470 * If the pattern hasn't changed from old_re, then old_re will be
5473 * eng is the current engine. If that engine has an op_comp method, then
5474 * handle directly (i.e. we assume that op_comp was us); otherwise, just
5475 * do the initial concatenation of arguments and pass on to the external
5478 * If is_bare_re is not null, set it to a boolean indicating whether the
5479 * arg list reduced (after overloading) to a single bare regex which has
5480 * been returned (i.e. /$qr/).
5482 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
5484 * pm_flags contains the PMf_* flags, typically based on those from the
5485 * pm_flags field of the related PMOP. Currently we're only interested in
5486 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
5488 * We can't allocate space until we know how big the compiled form will be,
5489 * but we can't compile it (and thus know how big it is) until we've got a
5490 * place to put the code. So we cheat: we compile it twice, once with code
5491 * generation turned off and size counting turned on, and once "for real".
5492 * This also means that we don't allocate space until we are sure that the
5493 * thing really will compile successfully, and we never have to move the
5494 * code and thus invalidate pointers into it. (Note that it has to be in
5495 * one piece because free() must be able to free it all.) [NB: not true in perl]
5497 * Beware that the optimization-preparation code in here knows about some
5498 * of the structure of the compiled regexp. [I'll say.]
5502 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
5503 OP *expr, const regexp_engine* eng, REGEXP *old_re,
5504 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
5509 regexp_internal *ri;
5517 SV *code_blocksv = NULL;
5518 SV** new_patternp = patternp;
5520 /* these are all flags - maybe they should be turned
5521 * into a single int with different bit masks */
5522 I32 sawlookahead = 0;
5525 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
5527 bool runtime_code = 0;
5529 RExC_state_t RExC_state;
5530 RExC_state_t * const pRExC_state = &RExC_state;
5531 #ifdef TRIE_STUDY_OPT
5533 RExC_state_t copyRExC_state;
5535 GET_RE_DEBUG_FLAGS_DECL;
5537 PERL_ARGS_ASSERT_RE_OP_COMPILE;
5539 DEBUG_r(if (!PL_colorset) reginitcolors());
5541 #ifndef PERL_IN_XSUB_RE
5542 /* Initialize these here instead of as-needed, as is quick and avoids
5543 * having to test them each time otherwise */
5544 if (! PL_AboveLatin1) {
5545 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
5546 PL_ASCII = _new_invlist_C_array(ASCII_invlist);
5547 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
5549 PL_L1Posix_ptrs[_CC_ALPHANUMERIC]
5550 = _new_invlist_C_array(L1PosixAlnum_invlist);
5551 PL_Posix_ptrs[_CC_ALPHANUMERIC]
5552 = _new_invlist_C_array(PosixAlnum_invlist);
5554 PL_L1Posix_ptrs[_CC_ALPHA]
5555 = _new_invlist_C_array(L1PosixAlpha_invlist);
5556 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(PosixAlpha_invlist);
5558 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(PosixBlank_invlist);
5559 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(XPosixBlank_invlist);
5561 /* Cased is the same as Alpha in the ASCII range */
5562 PL_L1Posix_ptrs[_CC_CASED] = _new_invlist_C_array(L1Cased_invlist);
5563 PL_Posix_ptrs[_CC_CASED] = _new_invlist_C_array(PosixAlpha_invlist);
5565 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(PosixCntrl_invlist);
5566 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(XPosixCntrl_invlist);
5568 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5569 PL_L1Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5571 PL_L1Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(L1PosixGraph_invlist);
5572 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(PosixGraph_invlist);
5574 PL_L1Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(L1PosixLower_invlist);
5575 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(PosixLower_invlist);
5577 PL_L1Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(L1PosixPrint_invlist);
5578 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(PosixPrint_invlist);
5580 PL_L1Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(L1PosixPunct_invlist);
5581 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(PosixPunct_invlist);
5583 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(PerlSpace_invlist);
5584 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(XPerlSpace_invlist);
5585 PL_Posix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(PosixSpace_invlist);
5586 PL_XPosix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(XPosixSpace_invlist);
5588 PL_L1Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(L1PosixUpper_invlist);
5589 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(PosixUpper_invlist);
5591 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(VertSpace_invlist);
5593 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(PosixWord_invlist);
5594 PL_L1Posix_ptrs[_CC_WORDCHAR]
5595 = _new_invlist_C_array(L1PosixWord_invlist);
5597 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(PosixXDigit_invlist);
5598 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(XPosixXDigit_invlist);
5600 PL_HasMultiCharFold = _new_invlist_C_array(_Perl_Multi_Char_Folds_invlist);
5604 pRExC_state->code_blocks = NULL;
5605 pRExC_state->num_code_blocks = 0;
5608 *is_bare_re = FALSE;
5610 if (expr && (expr->op_type == OP_LIST ||
5611 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
5612 /* allocate code_blocks if needed */
5616 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling)
5617 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
5618 ncode++; /* count of DO blocks */
5620 pRExC_state->num_code_blocks = ncode;
5621 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
5626 /* compile-time pattern with just OP_CONSTs and DO blocks */
5631 /* find how many CONSTs there are */
5634 if (expr->op_type == OP_CONST)
5637 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5638 if (o->op_type == OP_CONST)
5642 /* fake up an SV array */
5644 assert(!new_patternp);
5645 Newx(new_patternp, n, SV*);
5646 SAVEFREEPV(new_patternp);
5650 if (expr->op_type == OP_CONST)
5651 new_patternp[n] = cSVOPx_sv(expr);
5653 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5654 if (o->op_type == OP_CONST)
5655 new_patternp[n++] = cSVOPo_sv;
5660 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5661 "Assembling pattern from %d elements%s\n", pat_count,
5662 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5664 /* set expr to the first arg op */
5666 if (pRExC_state->num_code_blocks
5667 && expr->op_type != OP_CONST)
5669 expr = cLISTOPx(expr)->op_first;
5670 assert( expr->op_type == OP_PUSHMARK
5671 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
5672 || expr->op_type == OP_PADRANGE);
5673 expr = expr->op_sibling;
5676 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
5677 expr, &recompile, NULL);
5679 /* handle bare (possibly after overloading) regex: foo =~ $re */
5684 if (SvTYPE(re) == SVt_REGEXP) {
5688 Safefree(pRExC_state->code_blocks);
5689 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5690 "Precompiled pattern%s\n",
5691 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5697 exp = SvPV_nomg(pat, plen);
5699 if (!eng->op_comp) {
5700 if ((SvUTF8(pat) && IN_BYTES)
5701 || SvGMAGICAL(pat) || SvAMAGIC(pat))
5703 /* make a temporary copy; either to convert to bytes,
5704 * or to avoid repeating get-magic / overloaded stringify */
5705 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
5706 (IN_BYTES ? 0 : SvUTF8(pat)));
5708 Safefree(pRExC_state->code_blocks);
5709 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
5712 /* ignore the utf8ness if the pattern is 0 length */
5713 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
5714 RExC_uni_semantics = 0;
5715 RExC_contains_locale = 0;
5716 pRExC_state->runtime_code_qr = NULL;
5719 SV *dsv= sv_newmortal();
5720 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
5721 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
5722 PL_colors[4],PL_colors[5],s);
5726 /* we jump here if we upgrade the pattern to utf8 and have to
5729 if ((pm_flags & PMf_USE_RE_EVAL)
5730 /* this second condition covers the non-regex literal case,
5731 * i.e. $foo =~ '(?{})'. */
5732 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
5734 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
5736 /* return old regex if pattern hasn't changed */
5737 /* XXX: note in the below we have to check the flags as well as the pattern.
5739 * Things get a touch tricky as we have to compare the utf8 flag independently
5740 * from the compile flags.
5745 && !!RX_UTF8(old_re) == !!RExC_utf8
5746 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
5747 && RX_PRECOMP(old_re)
5748 && RX_PRELEN(old_re) == plen
5749 && memEQ(RX_PRECOMP(old_re), exp, plen)
5750 && !runtime_code /* with runtime code, always recompile */ )
5752 Safefree(pRExC_state->code_blocks);
5756 rx_flags = orig_rx_flags;
5758 if (initial_charset == REGEX_LOCALE_CHARSET) {
5759 RExC_contains_locale = 1;
5761 else if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
5763 /* Set to use unicode semantics if the pattern is in utf8 and has the
5764 * 'depends' charset specified, as it means unicode when utf8 */
5765 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5769 RExC_flags = rx_flags;
5770 RExC_pm_flags = pm_flags;
5773 if (TAINTING_get && TAINT_get)
5774 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
5776 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
5777 /* whoops, we have a non-utf8 pattern, whilst run-time code
5778 * got compiled as utf8. Try again with a utf8 pattern */
5779 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
5780 pRExC_state->num_code_blocks);
5781 goto redo_first_pass;
5784 assert(!pRExC_state->runtime_code_qr);
5789 RExC_in_lookbehind = 0;
5790 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
5792 RExC_override_recoding = 0;
5793 RExC_in_multi_char_class = 0;
5795 /* First pass: determine size, legality. */
5798 RExC_end = exp + plen;
5803 RExC_emit = &RExC_emit_dummy;
5804 RExC_whilem_seen = 0;
5805 RExC_open_parens = NULL;
5806 RExC_close_parens = NULL;
5808 RExC_paren_names = NULL;
5810 RExC_paren_name_list = NULL;
5812 RExC_recurse = NULL;
5813 RExC_recurse_count = 0;
5814 pRExC_state->code_index = 0;
5816 #if 0 /* REGC() is (currently) a NOP at the first pass.
5817 * Clever compilers notice this and complain. --jhi */
5818 REGC((U8)REG_MAGIC, (char*)RExC_emit);
5821 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
5823 RExC_lastparse=NULL;
5825 /* reg may croak on us, not giving us a chance to free
5826 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
5827 need it to survive as long as the regexp (qr/(?{})/).
5828 We must check that code_blocksv is not already set, because we may
5829 have jumped back to restart the sizing pass. */
5830 if (pRExC_state->code_blocks && !code_blocksv) {
5831 code_blocksv = newSV_type(SVt_PV);
5832 SAVEFREESV(code_blocksv);
5833 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
5834 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
5836 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5837 /* It's possible to write a regexp in ascii that represents Unicode
5838 codepoints outside of the byte range, such as via \x{100}. If we
5839 detect such a sequence we have to convert the entire pattern to utf8
5840 and then recompile, as our sizing calculation will have been based
5841 on 1 byte == 1 character, but we will need to use utf8 to encode
5842 at least some part of the pattern, and therefore must convert the whole
5845 if (flags & RESTART_UTF8) {
5846 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
5847 pRExC_state->num_code_blocks);
5848 goto redo_first_pass;
5850 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
5853 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
5856 PerlIO_printf(Perl_debug_log,
5857 "Required size %"IVdf" nodes\n"
5858 "Starting second pass (creation)\n",
5861 RExC_lastparse=NULL;
5864 /* The first pass could have found things that force Unicode semantics */
5865 if ((RExC_utf8 || RExC_uni_semantics)
5866 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
5868 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5871 /* Small enough for pointer-storage convention?
5872 If extralen==0, this means that we will not need long jumps. */
5873 if (RExC_size >= 0x10000L && RExC_extralen)
5874 RExC_size += RExC_extralen;
5877 if (RExC_whilem_seen > 15)
5878 RExC_whilem_seen = 15;
5880 /* Allocate space and zero-initialize. Note, the two step process
5881 of zeroing when in debug mode, thus anything assigned has to
5882 happen after that */
5883 rx = (REGEXP*) newSV_type(SVt_REGEXP);
5885 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
5886 char, regexp_internal);
5887 if ( r == NULL || ri == NULL )
5888 FAIL("Regexp out of space");
5890 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
5891 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode), char);
5893 /* bulk initialize base fields with 0. */
5894 Zero(ri, sizeof(regexp_internal), char);
5897 /* non-zero initialization begins here */
5900 r->extflags = rx_flags;
5901 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
5903 if (pm_flags & PMf_IS_QR) {
5904 ri->code_blocks = pRExC_state->code_blocks;
5905 ri->num_code_blocks = pRExC_state->num_code_blocks;
5910 for (n = 0; n < pRExC_state->num_code_blocks; n++)
5911 if (pRExC_state->code_blocks[n].src_regex)
5912 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
5913 SAVEFREEPV(pRExC_state->code_blocks);
5917 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
5918 bool has_charset = (get_regex_charset(r->extflags) != REGEX_DEPENDS_CHARSET);
5920 /* The caret is output if there are any defaults: if not all the STD
5921 * flags are set, or if no character set specifier is needed */
5923 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
5925 bool has_runon = ((RExC_seen & REG_SEEN_RUN_ON_COMMENT)==REG_SEEN_RUN_ON_COMMENT);
5926 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
5927 >> RXf_PMf_STD_PMMOD_SHIFT);
5928 const char *fptr = STD_PAT_MODS; /*"msix"*/
5930 /* Allocate for the worst case, which is all the std flags are turned
5931 * on. If more precision is desired, we could do a population count of
5932 * the flags set. This could be done with a small lookup table, or by
5933 * shifting, masking and adding, or even, when available, assembly
5934 * language for a machine-language population count.
5935 * We never output a minus, as all those are defaults, so are
5936 * covered by the caret */
5937 const STRLEN wraplen = plen + has_p + has_runon
5938 + has_default /* If needs a caret */
5940 /* If needs a character set specifier */
5941 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
5942 + (sizeof(STD_PAT_MODS) - 1)
5943 + (sizeof("(?:)") - 1);
5945 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
5946 r->xpv_len_u.xpvlenu_pv = p;
5948 SvFLAGS(rx) |= SVf_UTF8;
5951 /* If a default, cover it using the caret */
5953 *p++= DEFAULT_PAT_MOD;
5957 const char* const name = get_regex_charset_name(r->extflags, &len);
5958 Copy(name, p, len, char);
5962 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
5965 while((ch = *fptr++)) {
5973 Copy(RExC_precomp, p, plen, char);
5974 assert ((RX_WRAPPED(rx) - p) < 16);
5975 r->pre_prefix = p - RX_WRAPPED(rx);
5981 SvCUR_set(rx, p - RX_WRAPPED(rx));
5985 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
5987 if (RExC_seen & REG_SEEN_RECURSE) {
5988 Newxz(RExC_open_parens, RExC_npar,regnode *);
5989 SAVEFREEPV(RExC_open_parens);
5990 Newxz(RExC_close_parens,RExC_npar,regnode *);
5991 SAVEFREEPV(RExC_close_parens);
5994 /* Useful during FAIL. */
5995 #ifdef RE_TRACK_PATTERN_OFFSETS
5996 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
5997 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
5998 "%s %"UVuf" bytes for offset annotations.\n",
5999 ri->u.offsets ? "Got" : "Couldn't get",
6000 (UV)((2*RExC_size+1) * sizeof(U32))));
6002 SetProgLen(ri,RExC_size);
6006 REH_CALL_COMP_BEGIN_HOOK(pRExC_state->rx);
6008 /* Second pass: emit code. */
6009 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
6010 RExC_pm_flags = pm_flags;
6012 RExC_end = exp + plen;
6015 RExC_emit_start = ri->program;
6016 RExC_emit = ri->program;
6017 RExC_emit_bound = ri->program + RExC_size + 1;
6018 pRExC_state->code_index = 0;
6020 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
6021 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6023 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#"UVxf"", (UV) flags);
6025 /* XXXX To minimize changes to RE engine we always allocate
6026 3-units-long substrs field. */
6027 Newx(r->substrs, 1, struct reg_substr_data);
6028 if (RExC_recurse_count) {
6029 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6030 SAVEFREEPV(RExC_recurse);
6034 r->minlen = minlen = sawlookahead = sawplus = sawopen = 0;
6035 Zero(r->substrs, 1, struct reg_substr_data);
6037 #ifdef TRIE_STUDY_OPT
6039 StructCopy(&zero_scan_data, &data, scan_data_t);
6040 copyRExC_state = RExC_state;
6043 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6045 RExC_state = copyRExC_state;
6046 if (seen & REG_TOP_LEVEL_BRANCHES)
6047 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
6049 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES;
6050 StructCopy(&zero_scan_data, &data, scan_data_t);
6053 StructCopy(&zero_scan_data, &data, scan_data_t);
6056 /* Dig out information for optimizations. */
6057 r->extflags = RExC_flags; /* was pm_op */
6058 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6061 SvUTF8_on(rx); /* Unicode in it? */
6062 ri->regstclass = NULL;
6063 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
6064 r->intflags |= PREGf_NAUGHTY;
6065 scan = ri->program + 1; /* First BRANCH. */
6067 /* testing for BRANCH here tells us whether there is "must appear"
6068 data in the pattern. If there is then we can use it for optimisations */
6069 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES)) { /* Only one top-level choice. */
6071 STRLEN longest_float_length, longest_fixed_length;
6072 struct regnode_charclass_class ch_class; /* pointed to by data */
6074 I32 last_close = 0; /* pointed to by data */
6075 regnode *first= scan;
6076 regnode *first_next= regnext(first);
6078 * Skip introductions and multiplicators >= 1
6079 * so that we can extract the 'meat' of the pattern that must
6080 * match in the large if() sequence following.
6081 * NOTE that EXACT is NOT covered here, as it is normally
6082 * picked up by the optimiser separately.
6084 * This is unfortunate as the optimiser isnt handling lookahead
6085 * properly currently.
6088 while ((OP(first) == OPEN && (sawopen = 1)) ||
6089 /* An OR of *one* alternative - should not happen now. */
6090 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6091 /* for now we can't handle lookbehind IFMATCH*/
6092 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6093 (OP(first) == PLUS) ||
6094 (OP(first) == MINMOD) ||
6095 /* An {n,m} with n>0 */
6096 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6097 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6100 * the only op that could be a regnode is PLUS, all the rest
6101 * will be regnode_1 or regnode_2.
6104 if (OP(first) == PLUS)
6107 first += regarglen[OP(first)];
6109 first = NEXTOPER(first);
6110 first_next= regnext(first);
6113 /* Starting-point info. */
6115 DEBUG_PEEP("first:",first,0);
6116 /* Ignore EXACT as we deal with it later. */
6117 if (PL_regkind[OP(first)] == EXACT) {
6118 if (OP(first) == EXACT)
6119 NOOP; /* Empty, get anchored substr later. */
6121 ri->regstclass = first;
6124 else if (PL_regkind[OP(first)] == TRIE &&
6125 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6128 /* this can happen only on restudy */
6129 if ( OP(first) == TRIE ) {
6130 struct regnode_1 *trieop = (struct regnode_1 *)
6131 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6132 StructCopy(first,trieop,struct regnode_1);
6133 trie_op=(regnode *)trieop;
6135 struct regnode_charclass *trieop = (struct regnode_charclass *)
6136 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6137 StructCopy(first,trieop,struct regnode_charclass);
6138 trie_op=(regnode *)trieop;
6141 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6142 ri->regstclass = trie_op;
6145 else if (REGNODE_SIMPLE(OP(first)))
6146 ri->regstclass = first;
6147 else if (PL_regkind[OP(first)] == BOUND ||
6148 PL_regkind[OP(first)] == NBOUND)
6149 ri->regstclass = first;
6150 else if (PL_regkind[OP(first)] == BOL) {
6151 r->extflags |= (OP(first) == MBOL
6153 : (OP(first) == SBOL
6156 first = NEXTOPER(first);
6159 else if (OP(first) == GPOS) {
6160 r->extflags |= RXf_ANCH_GPOS;
6161 first = NEXTOPER(first);
6164 else if ((!sawopen || !RExC_sawback) &&
6165 (OP(first) == STAR &&
6166 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6167 !(r->extflags & RXf_ANCH) && !pRExC_state->num_code_blocks)
6169 /* turn .* into ^.* with an implied $*=1 */
6171 (OP(NEXTOPER(first)) == REG_ANY)
6174 r->extflags |= type;
6175 r->intflags |= PREGf_IMPLICIT;
6176 first = NEXTOPER(first);
6179 if (sawplus && !sawlookahead && (!sawopen || !RExC_sawback)
6180 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6181 /* x+ must match at the 1st pos of run of x's */
6182 r->intflags |= PREGf_SKIP;
6184 /* Scan is after the zeroth branch, first is atomic matcher. */
6185 #ifdef TRIE_STUDY_OPT
6188 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6189 (IV)(first - scan + 1))
6193 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6194 (IV)(first - scan + 1))
6200 * If there's something expensive in the r.e., find the
6201 * longest literal string that must appear and make it the
6202 * regmust. Resolve ties in favor of later strings, since
6203 * the regstart check works with the beginning of the r.e.
6204 * and avoiding duplication strengthens checking. Not a
6205 * strong reason, but sufficient in the absence of others.
6206 * [Now we resolve ties in favor of the earlier string if
6207 * it happens that c_offset_min has been invalidated, since the
6208 * earlier string may buy us something the later one won't.]
6211 data.longest_fixed = newSVpvs("");
6212 data.longest_float = newSVpvs("");
6213 data.last_found = newSVpvs("");
6214 data.longest = &(data.longest_fixed);
6215 ENTER_with_name("study_chunk");
6216 SAVEFREESV(data.longest_fixed);
6217 SAVEFREESV(data.longest_float);
6218 SAVEFREESV(data.last_found);
6220 if (!ri->regstclass) {
6221 cl_init(pRExC_state, &ch_class);
6222 data.start_class = &ch_class;
6223 stclass_flag = SCF_DO_STCLASS_AND;
6224 } else /* XXXX Check for BOUND? */
6226 data.last_closep = &last_close;
6228 minlen = study_chunk(pRExC_state, &first, &minlen, &fake, scan + RExC_size, /* Up to end */
6229 &data, -1, NULL, NULL,
6230 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
6231 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
6235 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6238 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6239 && data.last_start_min == 0 && data.last_end > 0
6240 && !RExC_seen_zerolen
6241 && !(RExC_seen & REG_SEEN_VERBARG)
6242 && (!(RExC_seen & REG_SEEN_GPOS) || (r->extflags & RXf_ANCH_GPOS)))
6243 r->extflags |= RXf_CHECK_ALL;
6244 scan_commit(pRExC_state, &data,&minlen,0);
6246 longest_float_length = CHR_SVLEN(data.longest_float);
6248 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6249 && data.offset_fixed == data.offset_float_min
6250 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6251 && S_setup_longest (aTHX_ pRExC_state,
6255 &(r->float_end_shift),
6256 data.lookbehind_float,
6257 data.offset_float_min,
6259 longest_float_length,
6260 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6261 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6263 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6264 r->float_max_offset = data.offset_float_max;
6265 if (data.offset_float_max < I32_MAX) /* Don't offset infinity */
6266 r->float_max_offset -= data.lookbehind_float;
6267 SvREFCNT_inc_simple_void_NN(data.longest_float);
6270 r->float_substr = r->float_utf8 = NULL;
6271 longest_float_length = 0;
6274 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6276 if (S_setup_longest (aTHX_ pRExC_state,
6278 &(r->anchored_utf8),
6279 &(r->anchored_substr),
6280 &(r->anchored_end_shift),
6281 data.lookbehind_fixed,
6284 longest_fixed_length,
6285 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6286 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6288 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6289 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6292 r->anchored_substr = r->anchored_utf8 = NULL;
6293 longest_fixed_length = 0;
6295 LEAVE_with_name("study_chunk");
6298 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6299 ri->regstclass = NULL;
6301 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6303 && ! TEST_SSC_EOS(data.start_class)
6304 && !cl_is_anything(data.start_class))
6306 const U32 n = add_data(pRExC_state, 1, "f");
6307 OP(data.start_class) = ANYOF_SYNTHETIC;
6309 Newx(RExC_rxi->data->data[n], 1,
6310 struct regnode_charclass_class);
6311 StructCopy(data.start_class,
6312 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6313 struct regnode_charclass_class);
6314 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6315 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6316 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6317 regprop(r, sv, (regnode*)data.start_class);
6318 PerlIO_printf(Perl_debug_log,
6319 "synthetic stclass \"%s\".\n",
6320 SvPVX_const(sv));});
6323 /* A temporary algorithm prefers floated substr to fixed one to dig more info. */
6324 if (longest_fixed_length > longest_float_length) {
6325 r->check_end_shift = r->anchored_end_shift;
6326 r->check_substr = r->anchored_substr;
6327 r->check_utf8 = r->anchored_utf8;
6328 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6329 if (r->extflags & RXf_ANCH_SINGLE)
6330 r->extflags |= RXf_NOSCAN;
6333 r->check_end_shift = r->float_end_shift;
6334 r->check_substr = r->float_substr;
6335 r->check_utf8 = r->float_utf8;
6336 r->check_offset_min = r->float_min_offset;
6337 r->check_offset_max = r->float_max_offset;
6339 /* XXXX Currently intuiting is not compatible with ANCH_GPOS.
6340 This should be changed ASAP! */
6341 if ((r->check_substr || r->check_utf8) && !(r->extflags & RXf_ANCH_GPOS)) {
6342 r->extflags |= RXf_USE_INTUIT;
6343 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6344 r->extflags |= RXf_INTUIT_TAIL;
6346 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6347 if ( (STRLEN)minlen < longest_float_length )
6348 minlen= longest_float_length;
6349 if ( (STRLEN)minlen < longest_fixed_length )
6350 minlen= longest_fixed_length;
6354 /* Several toplevels. Best we can is to set minlen. */
6356 struct regnode_charclass_class ch_class;
6359 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6361 scan = ri->program + 1;
6362 cl_init(pRExC_state, &ch_class);
6363 data.start_class = &ch_class;
6364 data.last_closep = &last_close;
6367 minlen = study_chunk(pRExC_state, &scan, &minlen, &fake, scan + RExC_size,
6368 &data, -1, NULL, NULL,
6369 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS
6370 |(restudied ? SCF_TRIE_DOING_RESTUDY : 0),
6373 CHECK_RESTUDY_GOTO_butfirst(NOOP);
6375 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
6376 = r->float_substr = r->float_utf8 = NULL;
6378 if (! TEST_SSC_EOS(data.start_class)
6379 && !cl_is_anything(data.start_class))
6381 const U32 n = add_data(pRExC_state, 1, "f");
6382 OP(data.start_class) = ANYOF_SYNTHETIC;
6384 Newx(RExC_rxi->data->data[n], 1,
6385 struct regnode_charclass_class);
6386 StructCopy(data.start_class,
6387 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6388 struct regnode_charclass_class);
6389 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6390 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6391 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
6392 regprop(r, sv, (regnode*)data.start_class);
6393 PerlIO_printf(Perl_debug_log,
6394 "synthetic stclass \"%s\".\n",
6395 SvPVX_const(sv));});
6399 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
6400 the "real" pattern. */
6402 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf"\n",
6403 (IV)minlen, (IV)r->minlen);
6405 r->minlenret = minlen;
6406 if (r->minlen < minlen)
6409 if (RExC_seen & REG_SEEN_GPOS)
6410 r->extflags |= RXf_GPOS_SEEN;
6411 if (RExC_seen & REG_SEEN_LOOKBEHIND)
6412 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the lookbehind */
6413 if (pRExC_state->num_code_blocks)
6414 r->extflags |= RXf_EVAL_SEEN;
6415 if (RExC_seen & REG_SEEN_CANY)
6416 r->extflags |= RXf_CANY_SEEN;
6417 if (RExC_seen & REG_SEEN_VERBARG)
6419 r->intflags |= PREGf_VERBARG_SEEN;
6420 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
6422 if (RExC_seen & REG_SEEN_CUTGROUP)
6423 r->intflags |= PREGf_CUTGROUP_SEEN;
6424 if (pm_flags & PMf_USE_RE_EVAL)
6425 r->intflags |= PREGf_USE_RE_EVAL;
6426 if (RExC_paren_names)
6427 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
6429 RXp_PAREN_NAMES(r) = NULL;
6432 regnode *first = ri->program + 1;
6434 regnode *next = NEXTOPER(first);
6437 if (PL_regkind[fop] == NOTHING && nop == END)
6438 r->extflags |= RXf_NULL;
6439 else if (PL_regkind[fop] == BOL && nop == END)
6440 r->extflags |= RXf_START_ONLY;
6441 else if (fop == PLUS && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE && OP(regnext(first)) == END)
6442 r->extflags |= RXf_WHITE;
6443 else if ( r->extflags & RXf_SPLIT && fop == EXACT && STR_LEN(first) == 1 && *(STRING(first)) == ' ' && OP(regnext(first)) == END )
6444 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
6448 if (RExC_paren_names) {
6449 ri->name_list_idx = add_data( pRExC_state, 1, "a" );
6450 ri->data->data[ri->name_list_idx] = (void*)SvREFCNT_inc(RExC_paren_name_list);
6453 ri->name_list_idx = 0;
6455 if (RExC_recurse_count) {
6456 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
6457 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
6458 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
6461 Newxz(r->offs, RExC_npar, regexp_paren_pair);
6462 /* assume we don't need to swap parens around before we match */
6465 PerlIO_printf(Perl_debug_log,"Final program:\n");
6468 #ifdef RE_TRACK_PATTERN_OFFSETS
6469 DEBUG_OFFSETS_r(if (ri->u.offsets) {
6470 const U32 len = ri->u.offsets[0];
6472 GET_RE_DEBUG_FLAGS_DECL;
6473 PerlIO_printf(Perl_debug_log, "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
6474 for (i = 1; i <= len; i++) {
6475 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
6476 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
6477 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
6479 PerlIO_printf(Perl_debug_log, "\n");
6484 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
6485 * by setting the regexp SV to readonly-only instead. If the
6486 * pattern's been recompiled, the USEDness should remain. */
6487 if (old_re && SvREADONLY(old_re))
6495 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
6498 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
6500 PERL_UNUSED_ARG(value);
6502 if (flags & RXapif_FETCH) {
6503 return reg_named_buff_fetch(rx, key, flags);
6504 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
6505 Perl_croak_no_modify();
6507 } else if (flags & RXapif_EXISTS) {
6508 return reg_named_buff_exists(rx, key, flags)
6511 } else if (flags & RXapif_REGNAMES) {
6512 return reg_named_buff_all(rx, flags);
6513 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
6514 return reg_named_buff_scalar(rx, flags);
6516 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
6522 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
6525 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
6526 PERL_UNUSED_ARG(lastkey);
6528 if (flags & RXapif_FIRSTKEY)
6529 return reg_named_buff_firstkey(rx, flags);
6530 else if (flags & RXapif_NEXTKEY)
6531 return reg_named_buff_nextkey(rx, flags);
6533 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter", (int)flags);
6539 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
6542 AV *retarray = NULL;
6544 struct regexp *const rx = ReANY(r);
6546 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
6548 if (flags & RXapif_ALL)
6551 if (rx && RXp_PAREN_NAMES(rx)) {
6552 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
6555 SV* sv_dat=HeVAL(he_str);
6556 I32 *nums=(I32*)SvPVX(sv_dat);
6557 for ( i=0; i<SvIVX(sv_dat); i++ ) {
6558 if ((I32)(rx->nparens) >= nums[i]
6559 && rx->offs[nums[i]].start != -1
6560 && rx->offs[nums[i]].end != -1)
6563 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
6568 ret = newSVsv(&PL_sv_undef);
6571 av_push(retarray, ret);
6574 return newRV_noinc(MUTABLE_SV(retarray));
6581 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
6584 struct regexp *const rx = ReANY(r);
6586 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
6588 if (rx && RXp_PAREN_NAMES(rx)) {
6589 if (flags & RXapif_ALL) {
6590 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
6592 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
6594 SvREFCNT_dec_NN(sv);
6606 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
6608 struct regexp *const rx = ReANY(r);
6610 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
6612 if ( rx && RXp_PAREN_NAMES(rx) ) {
6613 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
6615 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
6622 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
6624 struct regexp *const rx = ReANY(r);
6625 GET_RE_DEBUG_FLAGS_DECL;
6627 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
6629 if (rx && RXp_PAREN_NAMES(rx)) {
6630 HV *hv = RXp_PAREN_NAMES(rx);
6632 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6635 SV* sv_dat = HeVAL(temphe);
6636 I32 *nums = (I32*)SvPVX(sv_dat);
6637 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6638 if ((I32)(rx->lastparen) >= nums[i] &&
6639 rx->offs[nums[i]].start != -1 &&
6640 rx->offs[nums[i]].end != -1)
6646 if (parno || flags & RXapif_ALL) {
6647 return newSVhek(HeKEY_hek(temphe));
6655 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
6660 struct regexp *const rx = ReANY(r);
6662 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
6664 if (rx && RXp_PAREN_NAMES(rx)) {
6665 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
6666 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
6667 } else if (flags & RXapif_ONE) {
6668 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
6669 av = MUTABLE_AV(SvRV(ret));
6670 length = av_len(av);
6671 SvREFCNT_dec_NN(ret);
6672 return newSViv(length + 1);
6674 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar", (int)flags);
6678 return &PL_sv_undef;
6682 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
6684 struct regexp *const rx = ReANY(r);
6687 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
6689 if (rx && RXp_PAREN_NAMES(rx)) {
6690 HV *hv= RXp_PAREN_NAMES(rx);
6692 (void)hv_iterinit(hv);
6693 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6696 SV* sv_dat = HeVAL(temphe);
6697 I32 *nums = (I32*)SvPVX(sv_dat);
6698 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6699 if ((I32)(rx->lastparen) >= nums[i] &&
6700 rx->offs[nums[i]].start != -1 &&
6701 rx->offs[nums[i]].end != -1)
6707 if (parno || flags & RXapif_ALL) {
6708 av_push(av, newSVhek(HeKEY_hek(temphe)));
6713 return newRV_noinc(MUTABLE_SV(av));
6717 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
6720 struct regexp *const rx = ReANY(r);
6726 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
6728 if ( ( n == RX_BUFF_IDX_CARET_PREMATCH
6729 || n == RX_BUFF_IDX_CARET_FULLMATCH
6730 || n == RX_BUFF_IDX_CARET_POSTMATCH
6732 && !(rx->extflags & RXf_PMf_KEEPCOPY)
6739 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
6740 /* no need to distinguish between them any more */
6741 n = RX_BUFF_IDX_FULLMATCH;
6743 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
6744 && rx->offs[0].start != -1)
6746 /* $`, ${^PREMATCH} */
6747 i = rx->offs[0].start;
6751 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
6752 && rx->offs[0].end != -1)
6754 /* $', ${^POSTMATCH} */
6755 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
6756 i = rx->sublen + rx->suboffset - rx->offs[0].end;
6759 if ( 0 <= n && n <= (I32)rx->nparens &&
6760 (s1 = rx->offs[n].start) != -1 &&
6761 (t1 = rx->offs[n].end) != -1)
6763 /* $&, ${^MATCH}, $1 ... */
6765 s = rx->subbeg + s1 - rx->suboffset;
6770 assert(s >= rx->subbeg);
6771 assert(rx->sublen >= (s - rx->subbeg) + i );
6773 #if NO_TAINT_SUPPORT
6774 sv_setpvn(sv, s, i);
6776 const int oldtainted = TAINT_get;
6778 sv_setpvn(sv, s, i);
6779 TAINT_set(oldtainted);
6781 if ( (rx->extflags & RXf_CANY_SEEN)
6782 ? (RXp_MATCH_UTF8(rx)
6783 && (!i || is_utf8_string((U8*)s, i)))
6784 : (RXp_MATCH_UTF8(rx)) )
6791 if (RXp_MATCH_TAINTED(rx)) {
6792 if (SvTYPE(sv) >= SVt_PVMG) {
6793 MAGIC* const mg = SvMAGIC(sv);
6796 SvMAGIC_set(sv, mg->mg_moremagic);
6798 if ((mgt = SvMAGIC(sv))) {
6799 mg->mg_moremagic = mgt;
6800 SvMAGIC_set(sv, mg);
6811 sv_setsv(sv,&PL_sv_undef);
6817 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
6818 SV const * const value)
6820 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
6822 PERL_UNUSED_ARG(rx);
6823 PERL_UNUSED_ARG(paren);
6824 PERL_UNUSED_ARG(value);
6827 Perl_croak_no_modify();
6831 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
6834 struct regexp *const rx = ReANY(r);
6838 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
6840 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
6842 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
6843 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6847 case RX_BUFF_IDX_PREMATCH: /* $` */
6848 if (rx->offs[0].start != -1) {
6849 i = rx->offs[0].start;
6858 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
6859 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6861 case RX_BUFF_IDX_POSTMATCH: /* $' */
6862 if (rx->offs[0].end != -1) {
6863 i = rx->sublen - rx->offs[0].end;
6865 s1 = rx->offs[0].end;
6872 case RX_BUFF_IDX_CARET_FULLMATCH: /* ${^MATCH} */
6873 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6877 /* $& / ${^MATCH}, $1, $2, ... */
6879 if (paren <= (I32)rx->nparens &&
6880 (s1 = rx->offs[paren].start) != -1 &&
6881 (t1 = rx->offs[paren].end) != -1)
6887 if (ckWARN(WARN_UNINITIALIZED))
6888 report_uninit((const SV *)sv);
6893 if (i > 0 && RXp_MATCH_UTF8(rx)) {
6894 const char * const s = rx->subbeg - rx->suboffset + s1;
6899 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
6906 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
6908 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
6909 PERL_UNUSED_ARG(rx);
6913 return newSVpvs("Regexp");
6916 /* Scans the name of a named buffer from the pattern.
6917 * If flags is REG_RSN_RETURN_NULL returns null.
6918 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
6919 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
6920 * to the parsed name as looked up in the RExC_paren_names hash.
6921 * If there is an error throws a vFAIL().. type exception.
6924 #define REG_RSN_RETURN_NULL 0
6925 #define REG_RSN_RETURN_NAME 1
6926 #define REG_RSN_RETURN_DATA 2
6929 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
6931 char *name_start = RExC_parse;
6933 PERL_ARGS_ASSERT_REG_SCAN_NAME;
6935 if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
6936 /* skip IDFIRST by using do...while */
6939 RExC_parse += UTF8SKIP(RExC_parse);
6940 } while (isWORDCHAR_utf8((U8*)RExC_parse));
6944 } while (isWORDCHAR(*RExC_parse));
6946 RExC_parse++; /* so the <- from the vFAIL is after the offending character */
6947 vFAIL("Group name must start with a non-digit word character");
6951 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
6952 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
6953 if ( flags == REG_RSN_RETURN_NAME)
6955 else if (flags==REG_RSN_RETURN_DATA) {
6958 if ( ! sv_name ) /* should not happen*/
6959 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
6960 if (RExC_paren_names)
6961 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
6963 sv_dat = HeVAL(he_str);
6965 vFAIL("Reference to nonexistent named group");
6969 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
6970 (unsigned long) flags);
6972 assert(0); /* NOT REACHED */
6977 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
6978 int rem=(int)(RExC_end - RExC_parse); \
6987 if (RExC_lastparse!=RExC_parse) \
6988 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
6991 iscut ? "..." : "<" \
6994 PerlIO_printf(Perl_debug_log,"%16s",""); \
6997 num = RExC_size + 1; \
6999 num=REG_NODE_NUM(RExC_emit); \
7000 if (RExC_lastnum!=num) \
7001 PerlIO_printf(Perl_debug_log,"|%4d",num); \
7003 PerlIO_printf(Perl_debug_log,"|%4s",""); \
7004 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
7005 (int)((depth*2)), "", \
7009 RExC_lastparse=RExC_parse; \
7014 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
7015 DEBUG_PARSE_MSG((funcname)); \
7016 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
7018 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7019 DEBUG_PARSE_MSG((funcname)); \
7020 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7023 /* This section of code defines the inversion list object and its methods. The
7024 * interfaces are highly subject to change, so as much as possible is static to
7025 * this file. An inversion list is here implemented as a malloc'd C UV array
7026 * with some added info that is placed as UVs at the beginning in a header
7027 * portion. An inversion list for Unicode is an array of code points, sorted
7028 * by ordinal number. The zeroth element is the first code point in the list.
7029 * The 1th element is the first element beyond that not in the list. In other
7030 * words, the first range is
7031 * invlist[0]..(invlist[1]-1)
7032 * The other ranges follow. Thus every element whose index is divisible by two
7033 * marks the beginning of a range that is in the list, and every element not
7034 * divisible by two marks the beginning of a range not in the list. A single
7035 * element inversion list that contains the single code point N generally
7036 * consists of two elements
7039 * (The exception is when N is the highest representable value on the
7040 * machine, in which case the list containing just it would be a single
7041 * element, itself. By extension, if the last range in the list extends to
7042 * infinity, then the first element of that range will be in the inversion list
7043 * at a position that is divisible by two, and is the final element in the
7045 * Taking the complement (inverting) an inversion list is quite simple, if the
7046 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7047 * This implementation reserves an element at the beginning of each inversion
7048 * list to contain 0 when the list contains 0, and contains 1 otherwise. The
7049 * actual beginning of the list is either that element if 0, or the next one if
7052 * More about inversion lists can be found in "Unicode Demystified"
7053 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7054 * More will be coming when functionality is added later.
7056 * The inversion list data structure is currently implemented as an SV pointing
7057 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7058 * array of UV whose memory management is automatically handled by the existing
7059 * facilities for SV's.
7061 * Some of the methods should always be private to the implementation, and some
7062 * should eventually be made public */
7064 /* The header definitions are in F<inline_invlist.c> */
7065 #define TO_INTERNAL_SIZE(x) (((x) + HEADER_LENGTH) * sizeof(UV))
7066 #define FROM_INTERNAL_SIZE(x) (((x)/ sizeof(UV)) - HEADER_LENGTH)
7068 #define INVLIST_INITIAL_LEN 10
7070 PERL_STATIC_INLINE UV*
7071 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
7073 /* Returns a pointer to the first element in the inversion list's array.
7074 * This is called upon initialization of an inversion list. Where the
7075 * array begins depends on whether the list has the code point U+0000
7076 * in it or not. The other parameter tells it whether the code that
7077 * follows this call is about to put a 0 in the inversion list or not.
7078 * The first element is either the element with 0, if 0, or the next one,
7081 UV* zero = get_invlist_zero_addr(invlist);
7083 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7086 assert(! *_get_invlist_len_addr(invlist));
7088 /* 1^1 = 0; 1^0 = 1 */
7089 *zero = 1 ^ will_have_0;
7090 return zero + *zero;
7093 PERL_STATIC_INLINE UV*
7094 S_invlist_array(pTHX_ SV* const invlist)
7096 /* Returns the pointer to the inversion list's array. Every time the
7097 * length changes, this needs to be called in case malloc or realloc moved
7100 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7102 /* Must not be empty. If these fail, you probably didn't check for <len>
7103 * being non-zero before trying to get the array */
7104 assert(*_get_invlist_len_addr(invlist));
7105 assert(*get_invlist_zero_addr(invlist) == 0
7106 || *get_invlist_zero_addr(invlist) == 1);
7108 /* The array begins either at the element reserved for zero if the
7109 * list contains 0 (that element will be set to 0), or otherwise the next
7110 * element (in which case the reserved element will be set to 1). */
7111 return (UV *) (get_invlist_zero_addr(invlist)
7112 + *get_invlist_zero_addr(invlist));
7115 PERL_STATIC_INLINE void
7116 S_invlist_set_len(pTHX_ SV* const invlist, const UV len)
7118 /* Sets the current number of elements stored in the inversion list */
7120 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7122 *_get_invlist_len_addr(invlist) = len;
7124 assert(len <= SvLEN(invlist));
7126 SvCUR_set(invlist, TO_INTERNAL_SIZE(len));
7127 /* If the list contains U+0000, that element is part of the header,
7128 * and should not be counted as part of the array. It will contain
7129 * 0 in that case, and 1 otherwise. So we could flop 0=>1, 1=>0 and
7131 * SvCUR_set(invlist,
7132 * TO_INTERNAL_SIZE(len
7133 * - (*get_invlist_zero_addr(inv_list) ^ 1)));
7134 * But, this is only valid if len is not 0. The consequences of not doing
7135 * this is that the memory allocation code may think that 1 more UV is
7136 * being used than actually is, and so might do an unnecessary grow. That
7137 * seems worth not bothering to make this the precise amount.
7139 * Note that when inverting, SvCUR shouldn't change */
7142 PERL_STATIC_INLINE IV*
7143 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7145 /* Return the address of the UV that is reserved to hold the cached index
7148 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7150 return (IV *) (SvPVX(invlist) + (INVLIST_PREVIOUS_INDEX_OFFSET * sizeof (UV)));
7153 PERL_STATIC_INLINE IV
7154 S_invlist_previous_index(pTHX_ SV* const invlist)
7156 /* Returns cached index of previous search */
7158 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7160 return *get_invlist_previous_index_addr(invlist);
7163 PERL_STATIC_INLINE void
7164 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7166 /* Caches <index> for later retrieval */
7168 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7170 assert(index == 0 || index < (int) _invlist_len(invlist));
7172 *get_invlist_previous_index_addr(invlist) = index;
7175 PERL_STATIC_INLINE UV
7176 S_invlist_max(pTHX_ SV* const invlist)
7178 /* Returns the maximum number of elements storable in the inversion list's
7179 * array, without having to realloc() */
7181 PERL_ARGS_ASSERT_INVLIST_MAX;
7183 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7184 ? _invlist_len(invlist)
7185 : FROM_INTERNAL_SIZE(SvLEN(invlist));
7188 PERL_STATIC_INLINE UV*
7189 S_get_invlist_zero_addr(pTHX_ SV* invlist)
7191 /* Return the address of the UV that is reserved to hold 0 if the inversion
7192 * list contains 0. This has to be the last element of the heading, as the
7193 * list proper starts with either it if 0, or the next element if not.
7194 * (But we force it to contain either 0 or 1) */
7196 PERL_ARGS_ASSERT_GET_INVLIST_ZERO_ADDR;
7198 return (UV *) (SvPVX(invlist) + (INVLIST_ZERO_OFFSET * sizeof (UV)));
7201 #ifndef PERL_IN_XSUB_RE
7203 Perl__new_invlist(pTHX_ IV initial_size)
7206 /* Return a pointer to a newly constructed inversion list, with enough
7207 * space to store 'initial_size' elements. If that number is negative, a
7208 * system default is used instead */
7212 if (initial_size < 0) {
7213 initial_size = INVLIST_INITIAL_LEN;
7216 /* Allocate the initial space */
7217 new_list = newSV(TO_INTERNAL_SIZE(initial_size));
7218 invlist_set_len(new_list, 0);
7220 /* Force iterinit() to be used to get iteration to work */
7221 *get_invlist_iter_addr(new_list) = UV_MAX;
7223 /* This should force a segfault if a method doesn't initialize this
7225 *get_invlist_zero_addr(new_list) = UV_MAX;
7227 *get_invlist_previous_index_addr(new_list) = 0;
7228 *get_invlist_version_id_addr(new_list) = INVLIST_VERSION_ID;
7229 #if HEADER_LENGTH != 5
7230 # 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
7238 S__new_invlist_C_array(pTHX_ UV* list)
7240 /* Return a pointer to a newly constructed inversion list, initialized to
7241 * point to <list>, which has to be in the exact correct inversion list
7242 * form, including internal fields. Thus this is a dangerous routine that
7243 * should not be used in the wrong hands */
7245 SV* invlist = newSV_type(SVt_PV);
7247 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7249 SvPV_set(invlist, (char *) list);
7250 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7251 shouldn't touch it */
7252 SvCUR_set(invlist, TO_INTERNAL_SIZE(_invlist_len(invlist)));
7254 if (*get_invlist_version_id_addr(invlist) != INVLIST_VERSION_ID) {
7255 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7258 /* Initialize the iteration pointer.
7259 * XXX This could be done at compile time in charclass_invlists.h, but I
7260 * (khw) am not confident that the suffixes for specifying the C constant
7261 * UV_MAX are portable, e.g. 'ull' on a 32 bit machine that is configured
7262 * to use 64 bits; might need a Configure probe */
7263 invlist_iterfinish(invlist);
7269 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7271 /* Grow the maximum size of an inversion list */
7273 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7275 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max));
7278 PERL_STATIC_INLINE void
7279 S_invlist_trim(pTHX_ SV* const invlist)
7281 PERL_ARGS_ASSERT_INVLIST_TRIM;
7283 /* Change the length of the inversion list to how many entries it currently
7286 SvPV_shrink_to_cur((SV *) invlist);
7289 #define _invlist_union_complement_2nd(a, b, output) _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
7292 S__append_range_to_invlist(pTHX_ SV* const invlist, const UV start, const UV end)
7294 /* Subject to change or removal. Append the range from 'start' to 'end' at
7295 * the end of the inversion list. The range must be above any existing
7299 UV max = invlist_max(invlist);
7300 UV len = _invlist_len(invlist);
7302 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7304 if (len == 0) { /* Empty lists must be initialized */
7305 array = _invlist_array_init(invlist, start == 0);
7308 /* Here, the existing list is non-empty. The current max entry in the
7309 * list is generally the first value not in the set, except when the
7310 * set extends to the end of permissible values, in which case it is
7311 * the first entry in that final set, and so this call is an attempt to
7312 * append out-of-order */
7314 UV final_element = len - 1;
7315 array = invlist_array(invlist);
7316 if (array[final_element] > start
7317 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
7319 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",
7320 array[final_element], start,
7321 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
7324 /* Here, it is a legal append. If the new range begins with the first
7325 * value not in the set, it is extending the set, so the new first
7326 * value not in the set is one greater than the newly extended range.
7328 if (array[final_element] == start) {
7329 if (end != UV_MAX) {
7330 array[final_element] = end + 1;
7333 /* But if the end is the maximum representable on the machine,
7334 * just let the range that this would extend to have no end */
7335 invlist_set_len(invlist, len - 1);
7341 /* Here the new range doesn't extend any existing set. Add it */
7343 len += 2; /* Includes an element each for the start and end of range */
7345 /* If overflows the existing space, extend, which may cause the array to be
7348 invlist_extend(invlist, len);
7349 invlist_set_len(invlist, len); /* Have to set len here to avoid assert
7350 failure in invlist_array() */
7351 array = invlist_array(invlist);
7354 invlist_set_len(invlist, len);
7357 /* The next item on the list starts the range, the one after that is
7358 * one past the new range. */
7359 array[len - 2] = start;
7360 if (end != UV_MAX) {
7361 array[len - 1] = end + 1;
7364 /* But if the end is the maximum representable on the machine, just let
7365 * the range have no end */
7366 invlist_set_len(invlist, len - 1);
7370 #ifndef PERL_IN_XSUB_RE
7373 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
7375 /* Searches the inversion list for the entry that contains the input code
7376 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
7377 * return value is the index into the list's array of the range that
7382 IV high = _invlist_len(invlist);
7383 const IV highest_element = high - 1;
7386 PERL_ARGS_ASSERT__INVLIST_SEARCH;
7388 /* If list is empty, return failure. */
7393 /* (We can't get the array unless we know the list is non-empty) */
7394 array = invlist_array(invlist);
7396 mid = invlist_previous_index(invlist);
7397 assert(mid >=0 && mid <= highest_element);
7399 /* <mid> contains the cache of the result of the previous call to this
7400 * function (0 the first time). See if this call is for the same result,
7401 * or if it is for mid-1. This is under the theory that calls to this
7402 * function will often be for related code points that are near each other.
7403 * And benchmarks show that caching gives better results. We also test
7404 * here if the code point is within the bounds of the list. These tests
7405 * replace others that would have had to be made anyway to make sure that
7406 * the array bounds were not exceeded, and these give us extra information
7407 * at the same time */
7408 if (cp >= array[mid]) {
7409 if (cp >= array[highest_element]) {
7410 return highest_element;
7413 /* Here, array[mid] <= cp < array[highest_element]. This means that
7414 * the final element is not the answer, so can exclude it; it also
7415 * means that <mid> is not the final element, so can refer to 'mid + 1'
7417 if (cp < array[mid + 1]) {
7423 else { /* cp < aray[mid] */
7424 if (cp < array[0]) { /* Fail if outside the array */
7428 if (cp >= array[mid - 1]) {
7433 /* Binary search. What we are looking for is <i> such that
7434 * array[i] <= cp < array[i+1]
7435 * The loop below converges on the i+1. Note that there may not be an
7436 * (i+1)th element in the array, and things work nonetheless */
7437 while (low < high) {
7438 mid = (low + high) / 2;
7439 assert(mid <= highest_element);
7440 if (array[mid] <= cp) { /* cp >= array[mid] */
7443 /* We could do this extra test to exit the loop early.
7444 if (cp < array[low]) {
7449 else { /* cp < array[mid] */
7456 invlist_set_previous_index(invlist, high);
7461 Perl__invlist_populate_swatch(pTHX_ SV* const invlist, const UV start, const UV end, U8* swatch)
7463 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
7464 * but is used when the swash has an inversion list. This makes this much
7465 * faster, as it uses a binary search instead of a linear one. This is
7466 * intimately tied to that function, and perhaps should be in utf8.c,
7467 * except it is intimately tied to inversion lists as well. It assumes
7468 * that <swatch> is all 0's on input */
7471 const IV len = _invlist_len(invlist);
7475 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
7477 if (len == 0) { /* Empty inversion list */
7481 array = invlist_array(invlist);
7483 /* Find which element it is */
7484 i = _invlist_search(invlist, start);
7486 /* We populate from <start> to <end> */
7487 while (current < end) {
7490 /* The inversion list gives the results for every possible code point
7491 * after the first one in the list. Only those ranges whose index is
7492 * even are ones that the inversion list matches. For the odd ones,
7493 * and if the initial code point is not in the list, we have to skip
7494 * forward to the next element */
7495 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
7497 if (i >= len) { /* Finished if beyond the end of the array */
7501 if (current >= end) { /* Finished if beyond the end of what we
7503 if (LIKELY(end < UV_MAX)) {
7507 /* We get here when the upper bound is the maximum
7508 * representable on the machine, and we are looking for just
7509 * that code point. Have to special case it */
7511 goto join_end_of_list;
7514 assert(current >= start);
7516 /* The current range ends one below the next one, except don't go past
7519 upper = (i < len && array[i] < end) ? array[i] : end;
7521 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
7522 * for each code point in it */
7523 for (; current < upper; current++) {
7524 const STRLEN offset = (STRLEN)(current - start);
7525 swatch[offset >> 3] |= 1 << (offset & 7);
7530 /* Quit if at the end of the list */
7533 /* But first, have to deal with the highest possible code point on
7534 * the platform. The previous code assumes that <end> is one
7535 * beyond where we want to populate, but that is impossible at the
7536 * platform's infinity, so have to handle it specially */
7537 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
7539 const STRLEN offset = (STRLEN)(end - start);
7540 swatch[offset >> 3] |= 1 << (offset & 7);
7545 /* Advance to the next range, which will be for code points not in the
7554 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** output)
7556 /* Take the union of two inversion lists and point <output> to it. *output
7557 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7558 * the reference count to that list will be decremented. The first list,
7559 * <a>, may be NULL, in which case a copy of the second list is returned.
7560 * If <complement_b> is TRUE, the union is taken of the complement
7561 * (inversion) of <b> instead of b itself.
7563 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7564 * Richard Gillam, published by Addison-Wesley, and explained at some
7565 * length there. The preface says to incorporate its examples into your
7566 * code at your own risk.
7568 * The algorithm is like a merge sort.
7570 * XXX A potential performance improvement is to keep track as we go along
7571 * if only one of the inputs contributes to the result, meaning the other
7572 * is a subset of that one. In that case, we can skip the final copy and
7573 * return the larger of the input lists, but then outside code might need
7574 * to keep track of whether to free the input list or not */
7576 UV* array_a; /* a's array */
7578 UV len_a; /* length of a's array */
7581 SV* u; /* the resulting union */
7585 UV i_a = 0; /* current index into a's array */
7589 /* running count, as explained in the algorithm source book; items are
7590 * stopped accumulating and are output when the count changes to/from 0.
7591 * The count is incremented when we start a range that's in the set, and
7592 * decremented when we start a range that's not in the set. So its range
7593 * is 0 to 2. Only when the count is zero is something not in the set.
7597 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
7600 /* If either one is empty, the union is the other one */
7601 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
7608 *output = invlist_clone(b);
7610 _invlist_invert(*output);
7612 } /* else *output already = b; */
7615 else if ((len_b = _invlist_len(b)) == 0) {
7620 /* The complement of an empty list is a list that has everything in it,
7621 * so the union with <a> includes everything too */
7626 *output = _new_invlist(1);
7627 _append_range_to_invlist(*output, 0, UV_MAX);
7629 else if (*output != a) {
7630 *output = invlist_clone(a);
7632 /* else *output already = a; */
7636 /* Here both lists exist and are non-empty */
7637 array_a = invlist_array(a);
7638 array_b = invlist_array(b);
7640 /* If are to take the union of 'a' with the complement of b, set it
7641 * up so are looking at b's complement. */
7644 /* To complement, we invert: if the first element is 0, remove it. To
7645 * do this, we just pretend the array starts one later, and clear the
7646 * flag as we don't have to do anything else later */
7647 if (array_b[0] == 0) {
7650 complement_b = FALSE;
7654 /* But if the first element is not zero, we unshift a 0 before the
7655 * array. The data structure reserves a space for that 0 (which
7656 * should be a '1' right now), so physical shifting is unneeded,
7657 * but temporarily change that element to 0. Before exiting the
7658 * routine, we must restore the element to '1' */
7665 /* Size the union for the worst case: that the sets are completely
7667 u = _new_invlist(len_a + len_b);
7669 /* Will contain U+0000 if either component does */
7670 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
7671 || (len_b > 0 && array_b[0] == 0));
7673 /* Go through each list item by item, stopping when exhausted one of
7675 while (i_a < len_a && i_b < len_b) {
7676 UV cp; /* The element to potentially add to the union's array */
7677 bool cp_in_set; /* is it in the the input list's set or not */
7679 /* We need to take one or the other of the two inputs for the union.
7680 * Since we are merging two sorted lists, we take the smaller of the
7681 * next items. In case of a tie, we take the one that is in its set
7682 * first. If we took one not in the set first, it would decrement the
7683 * count, possibly to 0 which would cause it to be output as ending the
7684 * range, and the next time through we would take the same number, and
7685 * output it again as beginning the next range. By doing it the
7686 * opposite way, there is no possibility that the count will be
7687 * momentarily decremented to 0, and thus the two adjoining ranges will
7688 * be seamlessly merged. (In a tie and both are in the set or both not
7689 * in the set, it doesn't matter which we take first.) */
7690 if (array_a[i_a] < array_b[i_b]
7691 || (array_a[i_a] == array_b[i_b]
7692 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7694 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7698 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7699 cp = array_b[i_b++];
7702 /* Here, have chosen which of the two inputs to look at. Only output
7703 * if the running count changes to/from 0, which marks the
7704 * beginning/end of a range in that's in the set */
7707 array_u[i_u++] = cp;
7714 array_u[i_u++] = cp;
7719 /* Here, we are finished going through at least one of the lists, which
7720 * means there is something remaining in at most one. We check if the list
7721 * that hasn't been exhausted is positioned such that we are in the middle
7722 * of a range in its set or not. (i_a and i_b point to the element beyond
7723 * the one we care about.) If in the set, we decrement 'count'; if 0, there
7724 * is potentially more to output.
7725 * There are four cases:
7726 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
7727 * in the union is entirely from the non-exhausted set.
7728 * 2) Both were in their sets, count is 2. Nothing further should
7729 * be output, as everything that remains will be in the exhausted
7730 * list's set, hence in the union; decrementing to 1 but not 0 insures
7732 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
7733 * Nothing further should be output because the union includes
7734 * everything from the exhausted set. Not decrementing ensures that.
7735 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
7736 * decrementing to 0 insures that we look at the remainder of the
7737 * non-exhausted set */
7738 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7739 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7744 /* The final length is what we've output so far, plus what else is about to
7745 * be output. (If 'count' is non-zero, then the input list we exhausted
7746 * has everything remaining up to the machine's limit in its set, and hence
7747 * in the union, so there will be no further output. */
7750 /* At most one of the subexpressions will be non-zero */
7751 len_u += (len_a - i_a) + (len_b - i_b);
7754 /* Set result to final length, which can change the pointer to array_u, so
7756 if (len_u != _invlist_len(u)) {
7757 invlist_set_len(u, len_u);
7759 array_u = invlist_array(u);
7762 /* When 'count' is 0, the list that was exhausted (if one was shorter than
7763 * the other) ended with everything above it not in its set. That means
7764 * that the remaining part of the union is precisely the same as the
7765 * non-exhausted list, so can just copy it unchanged. (If both list were
7766 * exhausted at the same time, then the operations below will be both 0.)
7769 IV copy_count; /* At most one will have a non-zero copy count */
7770 if ((copy_count = len_a - i_a) > 0) {
7771 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
7773 else if ((copy_count = len_b - i_b) > 0) {
7774 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
7778 /* If we've changed b, restore it */
7783 /* We may be removing a reference to one of the inputs */
7784 if (a == *output || b == *output) {
7785 assert(! invlist_is_iterating(*output));
7786 SvREFCNT_dec_NN(*output);
7794 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** i)
7796 /* Take the intersection of two inversion lists and point <i> to it. *i
7797 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7798 * the reference count to that list will be decremented.
7799 * If <complement_b> is TRUE, the result will be the intersection of <a>
7800 * and the complement (or inversion) of <b> instead of <b> directly.
7802 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7803 * Richard Gillam, published by Addison-Wesley, and explained at some
7804 * length there. The preface says to incorporate its examples into your
7805 * code at your own risk. In fact, it had bugs
7807 * The algorithm is like a merge sort, and is essentially the same as the
7811 UV* array_a; /* a's array */
7813 UV len_a; /* length of a's array */
7816 SV* r; /* the resulting intersection */
7820 UV i_a = 0; /* current index into a's array */
7824 /* running count, as explained in the algorithm source book; items are
7825 * stopped accumulating and are output when the count changes to/from 2.
7826 * The count is incremented when we start a range that's in the set, and
7827 * decremented when we start a range that's not in the set. So its range
7828 * is 0 to 2. Only when the count is 2 is something in the intersection.
7832 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
7835 /* Special case if either one is empty */
7836 len_a = _invlist_len(a);
7837 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
7839 if (len_a != 0 && complement_b) {
7841 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
7842 * be empty. Here, also we are using 'b's complement, which hence
7843 * must be every possible code point. Thus the intersection is
7846 *i = invlist_clone(a);
7852 /* else *i is already 'a' */
7856 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
7857 * intersection must be empty */
7864 *i = _new_invlist(0);
7868 /* Here both lists exist and are non-empty */
7869 array_a = invlist_array(a);
7870 array_b = invlist_array(b);
7872 /* If are to take the intersection of 'a' with the complement of b, set it
7873 * up so are looking at b's complement. */
7876 /* To complement, we invert: if the first element is 0, remove it. To
7877 * do this, we just pretend the array starts one later, and clear the
7878 * flag as we don't have to do anything else later */
7879 if (array_b[0] == 0) {
7882 complement_b = FALSE;
7886 /* But if the first element is not zero, we unshift a 0 before the
7887 * array. The data structure reserves a space for that 0 (which
7888 * should be a '1' right now), so physical shifting is unneeded,
7889 * but temporarily change that element to 0. Before exiting the
7890 * routine, we must restore the element to '1' */
7897 /* Size the intersection for the worst case: that the intersection ends up
7898 * fragmenting everything to be completely disjoint */
7899 r= _new_invlist(len_a + len_b);
7901 /* Will contain U+0000 iff both components do */
7902 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
7903 && len_b > 0 && array_b[0] == 0);
7905 /* Go through each list item by item, stopping when exhausted one of
7907 while (i_a < len_a && i_b < len_b) {
7908 UV cp; /* The element to potentially add to the intersection's
7910 bool cp_in_set; /* Is it in the input list's set or not */
7912 /* We need to take one or the other of the two inputs for the
7913 * intersection. Since we are merging two sorted lists, we take the
7914 * smaller of the next items. In case of a tie, we take the one that
7915 * is not in its set first (a difference from the union algorithm). If
7916 * we took one in the set first, it would increment the count, possibly
7917 * to 2 which would cause it to be output as starting a range in the
7918 * intersection, and the next time through we would take that same
7919 * number, and output it again as ending the set. By doing it the
7920 * opposite of this, there is no possibility that the count will be
7921 * momentarily incremented to 2. (In a tie and both are in the set or
7922 * both not in the set, it doesn't matter which we take first.) */
7923 if (array_a[i_a] < array_b[i_b]
7924 || (array_a[i_a] == array_b[i_b]
7925 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7927 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7931 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7935 /* Here, have chosen which of the two inputs to look at. Only output
7936 * if the running count changes to/from 2, which marks the
7937 * beginning/end of a range that's in the intersection */
7941 array_r[i_r++] = cp;
7946 array_r[i_r++] = cp;
7952 /* Here, we are finished going through at least one of the lists, which
7953 * means there is something remaining in at most one. We check if the list
7954 * that has been exhausted is positioned such that we are in the middle
7955 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
7956 * the ones we care about.) There are four cases:
7957 * 1) Both weren't in their sets, count is 0, and remains 0. There's
7958 * nothing left in the intersection.
7959 * 2) Both were in their sets, count is 2 and perhaps is incremented to
7960 * above 2. What should be output is exactly that which is in the
7961 * non-exhausted set, as everything it has is also in the intersection
7962 * set, and everything it doesn't have can't be in the intersection
7963 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
7964 * gets incremented to 2. Like the previous case, the intersection is
7965 * everything that remains in the non-exhausted set.
7966 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
7967 * remains 1. And the intersection has nothing more. */
7968 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7969 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7974 /* The final length is what we've output so far plus what else is in the
7975 * intersection. At most one of the subexpressions below will be non-zero */
7978 len_r += (len_a - i_a) + (len_b - i_b);
7981 /* Set result to final length, which can change the pointer to array_r, so
7983 if (len_r != _invlist_len(r)) {
7984 invlist_set_len(r, len_r);
7986 array_r = invlist_array(r);
7989 /* Finish outputting any remaining */
7990 if (count >= 2) { /* At most one will have a non-zero copy count */
7992 if ((copy_count = len_a - i_a) > 0) {
7993 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
7995 else if ((copy_count = len_b - i_b) > 0) {
7996 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
8000 /* If we've changed b, restore it */
8005 /* We may be removing a reference to one of the inputs */
8006 if (a == *i || b == *i) {
8007 assert(! invlist_is_iterating(*i));
8008 SvREFCNT_dec_NN(*i);
8016 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
8018 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8019 * set. A pointer to the inversion list is returned. This may actually be
8020 * a new list, in which case the passed in one has been destroyed. The
8021 * passed in inversion list can be NULL, in which case a new one is created
8022 * with just the one range in it */
8027 if (invlist == NULL) {
8028 invlist = _new_invlist(2);
8032 len = _invlist_len(invlist);
8035 /* If comes after the final entry actually in the list, can just append it
8038 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
8039 && start >= invlist_array(invlist)[len - 1]))
8041 _append_range_to_invlist(invlist, start, end);
8045 /* Here, can't just append things, create and return a new inversion list
8046 * which is the union of this range and the existing inversion list */
8047 range_invlist = _new_invlist(2);
8048 _append_range_to_invlist(range_invlist, start, end);
8050 _invlist_union(invlist, range_invlist, &invlist);
8052 /* The temporary can be freed */
8053 SvREFCNT_dec_NN(range_invlist);
8060 PERL_STATIC_INLINE SV*
8061 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
8062 return _add_range_to_invlist(invlist, cp, cp);
8065 #ifndef PERL_IN_XSUB_RE
8067 Perl__invlist_invert(pTHX_ SV* const invlist)
8069 /* Complement the input inversion list. This adds a 0 if the list didn't
8070 * have a zero; removes it otherwise. As described above, the data
8071 * structure is set up so that this is very efficient */
8073 UV* len_pos = _get_invlist_len_addr(invlist);
8075 PERL_ARGS_ASSERT__INVLIST_INVERT;
8077 assert(! invlist_is_iterating(invlist));
8079 /* The inverse of matching nothing is matching everything */
8080 if (*len_pos == 0) {
8081 _append_range_to_invlist(invlist, 0, UV_MAX);
8085 /* The exclusive or complents 0 to 1; and 1 to 0. If the result is 1, the
8086 * zero element was a 0, so it is being removed, so the length decrements
8087 * by 1; and vice-versa. SvCUR is unaffected */
8088 if (*get_invlist_zero_addr(invlist) ^= 1) {
8097 Perl__invlist_invert_prop(pTHX_ SV* const invlist)
8099 /* Complement the input inversion list (which must be a Unicode property,
8100 * all of which don't match above the Unicode maximum code point.) And
8101 * Perl has chosen to not have the inversion match above that either. This
8102 * adds a 0x110000 if the list didn't end with it, and removes it if it did
8108 PERL_ARGS_ASSERT__INVLIST_INVERT_PROP;
8110 _invlist_invert(invlist);
8112 len = _invlist_len(invlist);
8114 if (len != 0) { /* If empty do nothing */
8115 array = invlist_array(invlist);
8116 if (array[len - 1] != PERL_UNICODE_MAX + 1) {
8117 /* Add 0x110000. First, grow if necessary */
8119 if (invlist_max(invlist) < len) {
8120 invlist_extend(invlist, len);
8121 array = invlist_array(invlist);
8123 invlist_set_len(invlist, len);
8124 array[len - 1] = PERL_UNICODE_MAX + 1;
8126 else { /* Remove the 0x110000 */
8127 invlist_set_len(invlist, len - 1);
8135 PERL_STATIC_INLINE SV*
8136 S_invlist_clone(pTHX_ SV* const invlist)
8139 /* Return a new inversion list that is a copy of the input one, which is
8142 /* Need to allocate extra space to accommodate Perl's addition of a
8143 * trailing NUL to SvPV's, since it thinks they are always strings */
8144 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8145 STRLEN length = SvCUR(invlist);
8147 PERL_ARGS_ASSERT_INVLIST_CLONE;
8149 SvCUR_set(new_invlist, length); /* This isn't done automatically */
8150 Copy(SvPVX(invlist), SvPVX(new_invlist), length, char);
8155 PERL_STATIC_INLINE UV*
8156 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8158 /* Return the address of the UV that contains the current iteration
8161 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8163 return (UV *) (SvPVX(invlist) + (INVLIST_ITER_OFFSET * sizeof (UV)));
8166 PERL_STATIC_INLINE UV*
8167 S_get_invlist_version_id_addr(pTHX_ SV* invlist)
8169 /* Return the address of the UV that contains the version id. */
8171 PERL_ARGS_ASSERT_GET_INVLIST_VERSION_ID_ADDR;
8173 return (UV *) (SvPVX(invlist) + (INVLIST_VERSION_ID_OFFSET * sizeof (UV)));
8176 PERL_STATIC_INLINE void
8177 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8179 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8181 *get_invlist_iter_addr(invlist) = 0;
8184 PERL_STATIC_INLINE void
8185 S_invlist_iterfinish(pTHX_ SV* invlist)
8187 /* Terminate iterator for invlist. This is to catch development errors.
8188 * Any iteration that is interrupted before completed should call this
8189 * function. Functions that add code points anywhere else but to the end
8190 * of an inversion list assert that they are not in the middle of an
8191 * iteration. If they were, the addition would make the iteration
8192 * problematical: if the iteration hadn't reached the place where things
8193 * were being added, it would be ok */
8195 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8197 *get_invlist_iter_addr(invlist) = UV_MAX;
8201 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8203 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8204 * This call sets in <*start> and <*end>, the next range in <invlist>.
8205 * Returns <TRUE> if successful and the next call will return the next
8206 * range; <FALSE> if was already at the end of the list. If the latter,
8207 * <*start> and <*end> are unchanged, and the next call to this function
8208 * will start over at the beginning of the list */
8210 UV* pos = get_invlist_iter_addr(invlist);
8211 UV len = _invlist_len(invlist);
8214 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8217 *pos = UV_MAX; /* Force iterinit() to be required next time */
8221 array = invlist_array(invlist);
8223 *start = array[(*pos)++];
8229 *end = array[(*pos)++] - 1;
8235 PERL_STATIC_INLINE bool
8236 S_invlist_is_iterating(pTHX_ SV* const invlist)
8238 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8240 return *(get_invlist_iter_addr(invlist)) < UV_MAX;
8243 PERL_STATIC_INLINE UV
8244 S_invlist_highest(pTHX_ SV* const invlist)
8246 /* Returns the highest code point that matches an inversion list. This API
8247 * has an ambiguity, as it returns 0 under either the highest is actually
8248 * 0, or if the list is empty. If this distinction matters to you, check
8249 * for emptiness before calling this function */
8251 UV len = _invlist_len(invlist);
8254 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8260 array = invlist_array(invlist);
8262 /* The last element in the array in the inversion list always starts a
8263 * range that goes to infinity. That range may be for code points that are
8264 * matched in the inversion list, or it may be for ones that aren't
8265 * matched. In the latter case, the highest code point in the set is one
8266 * less than the beginning of this range; otherwise it is the final element
8267 * of this range: infinity */
8268 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8270 : array[len - 1] - 1;
8273 #ifndef PERL_IN_XSUB_RE
8275 Perl__invlist_contents(pTHX_ SV* const invlist)
8277 /* Get the contents of an inversion list into a string SV so that they can
8278 * be printed out. It uses the format traditionally done for debug tracing
8282 SV* output = newSVpvs("\n");
8284 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8286 assert(! invlist_is_iterating(invlist));
8288 invlist_iterinit(invlist);
8289 while (invlist_iternext(invlist, &start, &end)) {
8290 if (end == UV_MAX) {
8291 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8293 else if (end != start) {
8294 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8298 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
8306 #ifdef PERL_ARGS_ASSERT__INVLIST_DUMP
8308 Perl__invlist_dump(pTHX_ SV* const invlist, const char * const header)
8310 /* Dumps out the ranges in an inversion list. The string 'header'
8311 * if present is output on a line before the first range */
8315 PERL_ARGS_ASSERT__INVLIST_DUMP;
8317 if (header && strlen(header)) {
8318 PerlIO_printf(Perl_debug_log, "%s\n", header);
8320 if (invlist_is_iterating(invlist)) {
8321 PerlIO_printf(Perl_debug_log, "Can't dump because is in middle of iterating\n");
8325 invlist_iterinit(invlist);
8326 while (invlist_iternext(invlist, &start, &end)) {
8327 if (end == UV_MAX) {
8328 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. INFINITY\n", start);
8330 else if (end != start) {
8331 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. 0x%04"UVXf"\n",
8335 PerlIO_printf(Perl_debug_log, "0x%04"UVXf"\n", start);
8343 S__invlistEQ(pTHX_ SV* const a, SV* const b, bool complement_b)
8345 /* Return a boolean as to if the two passed in inversion lists are
8346 * identical. The final argument, if TRUE, says to take the complement of
8347 * the second inversion list before doing the comparison */
8349 UV* array_a = invlist_array(a);
8350 UV* array_b = invlist_array(b);
8351 UV len_a = _invlist_len(a);
8352 UV len_b = _invlist_len(b);
8354 UV i = 0; /* current index into the arrays */
8355 bool retval = TRUE; /* Assume are identical until proven otherwise */
8357 PERL_ARGS_ASSERT__INVLISTEQ;
8359 /* If are to compare 'a' with the complement of b, set it
8360 * up so are looking at b's complement. */
8363 /* The complement of nothing is everything, so <a> would have to have
8364 * just one element, starting at zero (ending at infinity) */
8366 return (len_a == 1 && array_a[0] == 0);
8368 else if (array_b[0] == 0) {
8370 /* Otherwise, to complement, we invert. Here, the first element is
8371 * 0, just remove it. To do this, we just pretend the array starts
8372 * one later, and clear the flag as we don't have to do anything
8377 complement_b = FALSE;
8381 /* But if the first element is not zero, we unshift a 0 before the
8382 * array. The data structure reserves a space for that 0 (which
8383 * should be a '1' right now), so physical shifting is unneeded,
8384 * but temporarily change that element to 0. Before exiting the
8385 * routine, we must restore the element to '1' */
8392 /* Make sure that the lengths are the same, as well as the final element
8393 * before looping through the remainder. (Thus we test the length, final,
8394 * and first elements right off the bat) */
8395 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
8398 else for (i = 0; i < len_a - 1; i++) {
8399 if (array_a[i] != array_b[i]) {
8412 #undef HEADER_LENGTH
8413 #undef INVLIST_INITIAL_LENGTH
8414 #undef TO_INTERNAL_SIZE
8415 #undef FROM_INTERNAL_SIZE
8416 #undef INVLIST_LEN_OFFSET
8417 #undef INVLIST_ZERO_OFFSET
8418 #undef INVLIST_ITER_OFFSET
8419 #undef INVLIST_VERSION_ID
8420 #undef INVLIST_PREVIOUS_INDEX_OFFSET
8422 /* End of inversion list object */
8425 S_parse_lparen_question_flags(pTHX_ struct RExC_state_t *pRExC_state)
8427 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
8428 * constructs, and updates RExC_flags with them. On input, RExC_parse
8429 * should point to the first flag; it is updated on output to point to the
8430 * final ')' or ':'. There needs to be at least one flag, or this will
8433 /* for (?g), (?gc), and (?o) warnings; warning
8434 about (?c) will warn about (?g) -- japhy */
8436 #define WASTED_O 0x01
8437 #define WASTED_G 0x02
8438 #define WASTED_C 0x04
8439 #define WASTED_GC (WASTED_G|WASTED_C)
8440 I32 wastedflags = 0x00;
8441 U32 posflags = 0, negflags = 0;
8442 U32 *flagsp = &posflags;
8443 char has_charset_modifier = '\0';
8445 bool has_use_defaults = FALSE;
8446 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
8448 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
8450 /* '^' as an initial flag sets certain defaults */
8451 if (UCHARAT(RExC_parse) == '^') {
8453 has_use_defaults = TRUE;
8454 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
8455 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
8456 ? REGEX_UNICODE_CHARSET
8457 : REGEX_DEPENDS_CHARSET);
8460 cs = get_regex_charset(RExC_flags);
8461 if (cs == REGEX_DEPENDS_CHARSET
8462 && (RExC_utf8 || RExC_uni_semantics))
8464 cs = REGEX_UNICODE_CHARSET;
8467 while (*RExC_parse) {
8468 /* && strchr("iogcmsx", *RExC_parse) */
8469 /* (?g), (?gc) and (?o) are useless here
8470 and must be globally applied -- japhy */
8471 switch (*RExC_parse) {
8473 /* Code for the imsx flags */
8474 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
8476 case LOCALE_PAT_MOD:
8477 if (has_charset_modifier) {
8478 goto excess_modifier;
8480 else if (flagsp == &negflags) {
8483 cs = REGEX_LOCALE_CHARSET;
8484 has_charset_modifier = LOCALE_PAT_MOD;
8485 RExC_contains_locale = 1;
8487 case UNICODE_PAT_MOD:
8488 if (has_charset_modifier) {
8489 goto excess_modifier;
8491 else if (flagsp == &negflags) {
8494 cs = REGEX_UNICODE_CHARSET;
8495 has_charset_modifier = UNICODE_PAT_MOD;
8497 case ASCII_RESTRICT_PAT_MOD:
8498 if (flagsp == &negflags) {
8501 if (has_charset_modifier) {
8502 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
8503 goto excess_modifier;
8505 /* Doubled modifier implies more restricted */
8506 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
8509 cs = REGEX_ASCII_RESTRICTED_CHARSET;
8511 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
8513 case DEPENDS_PAT_MOD:
8514 if (has_use_defaults) {
8515 goto fail_modifiers;
8517 else if (flagsp == &negflags) {
8520 else if (has_charset_modifier) {
8521 goto excess_modifier;
8524 /* The dual charset means unicode semantics if the
8525 * pattern (or target, not known until runtime) are
8526 * utf8, or something in the pattern indicates unicode
8528 cs = (RExC_utf8 || RExC_uni_semantics)
8529 ? REGEX_UNICODE_CHARSET
8530 : REGEX_DEPENDS_CHARSET;
8531 has_charset_modifier = DEPENDS_PAT_MOD;
8535 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
8536 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
8538 else if (has_charset_modifier == *(RExC_parse - 1)) {
8539 vFAIL2("Regexp modifier \"%c\" may not appear twice", *(RExC_parse - 1));
8542 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
8547 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"", *(RExC_parse - 1));
8549 case ONCE_PAT_MOD: /* 'o' */
8550 case GLOBAL_PAT_MOD: /* 'g' */
8551 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8552 const I32 wflagbit = *RExC_parse == 'o' ? WASTED_O : WASTED_G;
8553 if (! (wastedflags & wflagbit) ) {
8554 wastedflags |= wflagbit;
8555 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
8558 "Useless (%s%c) - %suse /%c modifier",
8559 flagsp == &negflags ? "?-" : "?",
8561 flagsp == &negflags ? "don't " : "",
8568 case CONTINUE_PAT_MOD: /* 'c' */
8569 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8570 if (! (wastedflags & WASTED_C) ) {
8571 wastedflags |= WASTED_GC;
8572 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
8575 "Useless (%sc) - %suse /gc modifier",
8576 flagsp == &negflags ? "?-" : "?",
8577 flagsp == &negflags ? "don't " : ""
8582 case KEEPCOPY_PAT_MOD: /* 'p' */
8583 if (flagsp == &negflags) {
8585 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
8587 *flagsp |= RXf_PMf_KEEPCOPY;
8591 /* A flag is a default iff it is following a minus, so
8592 * if there is a minus, it means will be trying to
8593 * re-specify a default which is an error */
8594 if (has_use_defaults || flagsp == &negflags) {
8595 goto fail_modifiers;
8598 wastedflags = 0; /* reset so (?g-c) warns twice */
8602 RExC_flags |= posflags;
8603 RExC_flags &= ~negflags;
8604 set_regex_charset(&RExC_flags, cs);
8610 vFAIL3("Sequence (%.*s...) not recognized",
8611 RExC_parse-seqstart, seqstart);
8620 - reg - regular expression, i.e. main body or parenthesized thing
8622 * Caller must absorb opening parenthesis.
8624 * Combining parenthesis handling with the base level of regular expression
8625 * is a trifle forced, but the need to tie the tails of the branches to what
8626 * follows makes it hard to avoid.
8628 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
8630 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
8632 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
8635 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
8636 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
8637 needs to be restarted.
8638 Otherwise would only return NULL if regbranch() returns NULL, which
8641 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
8642 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
8643 * 2 is like 1, but indicates that nextchar() has been called to advance
8644 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
8645 * this flag alerts us to the need to check for that */
8648 regnode *ret; /* Will be the head of the group. */
8651 regnode *ender = NULL;
8654 U32 oregflags = RExC_flags;
8655 bool have_branch = 0;
8657 I32 freeze_paren = 0;
8658 I32 after_freeze = 0;
8660 char * parse_start = RExC_parse; /* MJD */
8661 char * const oregcomp_parse = RExC_parse;
8663 GET_RE_DEBUG_FLAGS_DECL;
8665 PERL_ARGS_ASSERT_REG;
8666 DEBUG_PARSE("reg ");
8668 *flagp = 0; /* Tentatively. */
8671 /* Make an OPEN node, if parenthesized. */
8674 /* Under /x, space and comments can be gobbled up between the '(' and
8675 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
8676 * intervening space, as the sequence is a token, and a token should be
8678 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
8680 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
8681 char *start_verb = RExC_parse;
8682 STRLEN verb_len = 0;
8683 char *start_arg = NULL;
8684 unsigned char op = 0;
8686 int internal_argval = 0; /* internal_argval is only useful if !argok */
8688 if (has_intervening_patws && SIZE_ONLY) {
8689 ckWARNregdep(RExC_parse + 1, "In '(*VERB...)', splitting the initial '(*' is deprecated");
8691 while ( *RExC_parse && *RExC_parse != ')' ) {
8692 if ( *RExC_parse == ':' ) {
8693 start_arg = RExC_parse + 1;
8699 verb_len = RExC_parse - start_verb;
8702 while ( *RExC_parse && *RExC_parse != ')' )
8704 if ( *RExC_parse != ')' )
8705 vFAIL("Unterminated verb pattern argument");
8706 if ( RExC_parse == start_arg )
8709 if ( *RExC_parse != ')' )
8710 vFAIL("Unterminated verb pattern");
8713 switch ( *start_verb ) {
8714 case 'A': /* (*ACCEPT) */
8715 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
8717 internal_argval = RExC_nestroot;
8720 case 'C': /* (*COMMIT) */
8721 if ( memEQs(start_verb,verb_len,"COMMIT") )
8724 case 'F': /* (*FAIL) */
8725 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
8730 case ':': /* (*:NAME) */
8731 case 'M': /* (*MARK:NAME) */
8732 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
8737 case 'P': /* (*PRUNE) */
8738 if ( memEQs(start_verb,verb_len,"PRUNE") )
8741 case 'S': /* (*SKIP) */
8742 if ( memEQs(start_verb,verb_len,"SKIP") )
8745 case 'T': /* (*THEN) */
8746 /* [19:06] <TimToady> :: is then */
8747 if ( memEQs(start_verb,verb_len,"THEN") ) {
8749 RExC_seen |= REG_SEEN_CUTGROUP;
8755 vFAIL3("Unknown verb pattern '%.*s'",
8756 verb_len, start_verb);
8759 if ( start_arg && internal_argval ) {
8760 vFAIL3("Verb pattern '%.*s' may not have an argument",
8761 verb_len, start_verb);
8762 } else if ( argok < 0 && !start_arg ) {
8763 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
8764 verb_len, start_verb);
8766 ret = reganode(pRExC_state, op, internal_argval);
8767 if ( ! internal_argval && ! SIZE_ONLY ) {
8769 SV *sv = newSVpvn( start_arg, RExC_parse - start_arg);
8770 ARG(ret) = add_data( pRExC_state, 1, "S" );
8771 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
8778 if (!internal_argval)
8779 RExC_seen |= REG_SEEN_VERBARG;
8780 } else if ( start_arg ) {
8781 vFAIL3("Verb pattern '%.*s' may not have an argument",
8782 verb_len, start_verb);
8784 ret = reg_node(pRExC_state, op);
8786 nextchar(pRExC_state);
8789 else if (*RExC_parse == '?') { /* (?...) */
8790 bool is_logical = 0;
8791 const char * const seqstart = RExC_parse;
8792 if (has_intervening_patws && SIZE_ONLY) {
8793 ckWARNregdep(RExC_parse + 1, "In '(?...)', splitting the initial '(?' is deprecated");
8797 paren = *RExC_parse++;
8798 ret = NULL; /* For look-ahead/behind. */
8801 case 'P': /* (?P...) variants for those used to PCRE/Python */
8802 paren = *RExC_parse++;
8803 if ( paren == '<') /* (?P<...>) named capture */
8805 else if (paren == '>') { /* (?P>name) named recursion */
8806 goto named_recursion;
8808 else if (paren == '=') { /* (?P=...) named backref */
8809 /* this pretty much dupes the code for \k<NAME> in regatom(), if
8810 you change this make sure you change that */
8811 char* name_start = RExC_parse;
8813 SV *sv_dat = reg_scan_name(pRExC_state,
8814 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8815 if (RExC_parse == name_start || *RExC_parse != ')')
8816 vFAIL2("Sequence %.3s... not terminated",parse_start);
8819 num = add_data( pRExC_state, 1, "S" );
8820 RExC_rxi->data->data[num]=(void*)sv_dat;
8821 SvREFCNT_inc_simple_void(sv_dat);
8824 ret = reganode(pRExC_state,
8827 : (ASCII_FOLD_RESTRICTED)
8829 : (AT_LEAST_UNI_SEMANTICS)
8837 Set_Node_Offset(ret, parse_start+1);
8838 Set_Node_Cur_Length(ret, parse_start);
8840 nextchar(pRExC_state);
8844 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8846 case '<': /* (?<...) */
8847 if (*RExC_parse == '!')
8849 else if (*RExC_parse != '=')
8855 case '\'': /* (?'...') */
8856 name_start= RExC_parse;
8857 svname = reg_scan_name(pRExC_state,
8858 SIZE_ONLY ? /* reverse test from the others */
8859 REG_RSN_RETURN_NAME :
8860 REG_RSN_RETURN_NULL);
8861 if (RExC_parse == name_start) {
8863 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8866 if (*RExC_parse != paren)
8867 vFAIL2("Sequence (?%c... not terminated",
8868 paren=='>' ? '<' : paren);
8872 if (!svname) /* shouldn't happen */
8874 "panic: reg_scan_name returned NULL");
8875 if (!RExC_paren_names) {
8876 RExC_paren_names= newHV();
8877 sv_2mortal(MUTABLE_SV(RExC_paren_names));
8879 RExC_paren_name_list= newAV();
8880 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
8883 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
8885 sv_dat = HeVAL(he_str);
8887 /* croak baby croak */
8889 "panic: paren_name hash element allocation failed");
8890 } else if ( SvPOK(sv_dat) ) {
8891 /* (?|...) can mean we have dupes so scan to check
8892 its already been stored. Maybe a flag indicating
8893 we are inside such a construct would be useful,
8894 but the arrays are likely to be quite small, so
8895 for now we punt -- dmq */
8896 IV count = SvIV(sv_dat);
8897 I32 *pv = (I32*)SvPVX(sv_dat);
8899 for ( i = 0 ; i < count ; i++ ) {
8900 if ( pv[i] == RExC_npar ) {
8906 pv = (I32*)SvGROW(sv_dat, SvCUR(sv_dat) + sizeof(I32)+1);
8907 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
8908 pv[count] = RExC_npar;
8909 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
8912 (void)SvUPGRADE(sv_dat,SVt_PVNV);
8913 sv_setpvn(sv_dat, (char *)&(RExC_npar), sizeof(I32));
8915 SvIV_set(sv_dat, 1);
8918 /* Yes this does cause a memory leak in debugging Perls */
8919 if (!av_store(RExC_paren_name_list, RExC_npar, SvREFCNT_inc(svname)))
8920 SvREFCNT_dec_NN(svname);
8923 /*sv_dump(sv_dat);*/
8925 nextchar(pRExC_state);
8927 goto capturing_parens;
8929 RExC_seen |= REG_SEEN_LOOKBEHIND;
8930 RExC_in_lookbehind++;
8932 case '=': /* (?=...) */
8933 RExC_seen_zerolen++;
8935 case '!': /* (?!...) */
8936 RExC_seen_zerolen++;
8937 if (*RExC_parse == ')') {
8938 ret=reg_node(pRExC_state, OPFAIL);
8939 nextchar(pRExC_state);
8943 case '|': /* (?|...) */
8944 /* branch reset, behave like a (?:...) except that
8945 buffers in alternations share the same numbers */
8947 after_freeze = freeze_paren = RExC_npar;
8949 case ':': /* (?:...) */
8950 case '>': /* (?>...) */
8952 case '$': /* (?$...) */
8953 case '@': /* (?@...) */
8954 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
8956 case '#': /* (?#...) */
8957 /* XXX As soon as we disallow separating the '?' and '*' (by
8958 * spaces or (?#...) comment), it is believed that this case
8959 * will be unreachable and can be removed. See
8961 while (*RExC_parse && *RExC_parse != ')')
8963 if (*RExC_parse != ')')
8964 FAIL("Sequence (?#... not terminated");
8965 nextchar(pRExC_state);
8968 case '0' : /* (?0) */
8969 case 'R' : /* (?R) */
8970 if (*RExC_parse != ')')
8971 FAIL("Sequence (?R) not terminated");
8972 ret = reg_node(pRExC_state, GOSTART);
8973 *flagp |= POSTPONED;
8974 nextchar(pRExC_state);
8977 { /* named and numeric backreferences */
8979 case '&': /* (?&NAME) */
8980 parse_start = RExC_parse - 1;
8983 SV *sv_dat = reg_scan_name(pRExC_state,
8984 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8985 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8987 goto gen_recurse_regop;
8988 assert(0); /* NOT REACHED */
8990 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8992 vFAIL("Illegal pattern");
8994 goto parse_recursion;
8996 case '-': /* (?-1) */
8997 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8998 RExC_parse--; /* rewind to let it be handled later */
9002 case '1': case '2': case '3': case '4': /* (?1) */
9003 case '5': case '6': case '7': case '8': case '9':
9006 num = atoi(RExC_parse);
9007 parse_start = RExC_parse - 1; /* MJD */
9008 if (*RExC_parse == '-')
9010 while (isDIGIT(*RExC_parse))
9012 if (*RExC_parse!=')')
9013 vFAIL("Expecting close bracket");
9016 if ( paren == '-' ) {
9018 Diagram of capture buffer numbering.
9019 Top line is the normal capture buffer numbers
9020 Bottom line is the negative indexing as from
9024 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
9028 num = RExC_npar + num;
9031 vFAIL("Reference to nonexistent group");
9033 } else if ( paren == '+' ) {
9034 num = RExC_npar + num - 1;
9037 ret = reganode(pRExC_state, GOSUB, num);
9039 if (num > (I32)RExC_rx->nparens) {
9041 vFAIL("Reference to nonexistent group");
9043 ARG2L_SET( ret, RExC_recurse_count++);
9045 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9046 "Recurse #%"UVuf" to %"IVdf"\n", (UV)ARG(ret), (IV)ARG2L(ret)));
9050 RExC_seen |= REG_SEEN_RECURSE;
9051 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
9052 Set_Node_Offset(ret, parse_start); /* MJD */
9054 *flagp |= POSTPONED;
9055 nextchar(pRExC_state);
9057 } /* named and numeric backreferences */
9058 assert(0); /* NOT REACHED */
9060 case '?': /* (??...) */
9062 if (*RExC_parse != '{') {
9064 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
9067 *flagp |= POSTPONED;
9068 paren = *RExC_parse++;
9070 case '{': /* (?{...}) */
9073 struct reg_code_block *cb;
9075 RExC_seen_zerolen++;
9077 if ( !pRExC_state->num_code_blocks
9078 || pRExC_state->code_index >= pRExC_state->num_code_blocks
9079 || pRExC_state->code_blocks[pRExC_state->code_index].start
9080 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
9083 if (RExC_pm_flags & PMf_USE_RE_EVAL)
9084 FAIL("panic: Sequence (?{...}): no code block found\n");
9085 FAIL("Eval-group not allowed at runtime, use re 'eval'");
9087 /* this is a pre-compiled code block (?{...}) */
9088 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
9089 RExC_parse = RExC_start + cb->end;
9092 if (cb->src_regex) {
9093 n = add_data(pRExC_state, 2, "rl");
9094 RExC_rxi->data->data[n] =
9095 (void*)SvREFCNT_inc((SV*)cb->src_regex);
9096 RExC_rxi->data->data[n+1] = (void*)o;
9099 n = add_data(pRExC_state, 1,
9100 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l");
9101 RExC_rxi->data->data[n] = (void*)o;
9104 pRExC_state->code_index++;
9105 nextchar(pRExC_state);
9109 ret = reg_node(pRExC_state, LOGICAL);
9110 eval = reganode(pRExC_state, EVAL, n);
9113 /* for later propagation into (??{}) return value */
9114 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
9116 REGTAIL(pRExC_state, ret, eval);
9117 /* deal with the length of this later - MJD */
9120 ret = reganode(pRExC_state, EVAL, n);
9121 Set_Node_Length(ret, RExC_parse - parse_start + 1);
9122 Set_Node_Offset(ret, parse_start);
9125 case '(': /* (?(?{...})...) and (?(?=...)...) */
9128 if (RExC_parse[0] == '?') { /* (?(?...)) */
9129 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
9130 || RExC_parse[1] == '<'
9131 || RExC_parse[1] == '{') { /* Lookahead or eval. */
9135 ret = reg_node(pRExC_state, LOGICAL);
9139 tail = reg(pRExC_state, 1, &flag, depth+1);
9140 if (flag & RESTART_UTF8) {
9141 *flagp = RESTART_UTF8;
9144 REGTAIL(pRExC_state, ret, tail);
9148 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9149 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9151 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9152 char *name_start= RExC_parse++;
9154 SV *sv_dat=reg_scan_name(pRExC_state,
9155 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9156 if (RExC_parse == name_start || *RExC_parse != ch)
9157 vFAIL2("Sequence (?(%c... not terminated",
9158 (ch == '>' ? '<' : ch));
9161 num = add_data( pRExC_state, 1, "S" );
9162 RExC_rxi->data->data[num]=(void*)sv_dat;
9163 SvREFCNT_inc_simple_void(sv_dat);
9165 ret = reganode(pRExC_state,NGROUPP,num);
9166 goto insert_if_check_paren;
9168 else if (RExC_parse[0] == 'D' &&
9169 RExC_parse[1] == 'E' &&
9170 RExC_parse[2] == 'F' &&
9171 RExC_parse[3] == 'I' &&
9172 RExC_parse[4] == 'N' &&
9173 RExC_parse[5] == 'E')
9175 ret = reganode(pRExC_state,DEFINEP,0);
9178 goto insert_if_check_paren;
9180 else if (RExC_parse[0] == 'R') {
9183 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9184 parno = atoi(RExC_parse++);
9185 while (isDIGIT(*RExC_parse))
9187 } else if (RExC_parse[0] == '&') {
9190 sv_dat = reg_scan_name(pRExC_state,
9191 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9192 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9194 ret = reganode(pRExC_state,INSUBP,parno);
9195 goto insert_if_check_paren;
9197 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9200 parno = atoi(RExC_parse++);
9202 while (isDIGIT(*RExC_parse))
9204 ret = reganode(pRExC_state, GROUPP, parno);
9206 insert_if_check_paren:
9207 if ((c = *nextchar(pRExC_state)) != ')')
9208 vFAIL("Switch condition not recognized");
9210 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9211 br = regbranch(pRExC_state, &flags, 1,depth+1);
9213 if (flags & RESTART_UTF8) {
9214 *flagp = RESTART_UTF8;
9217 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9220 REGTAIL(pRExC_state, br, reganode(pRExC_state, LONGJMP, 0));
9221 c = *nextchar(pRExC_state);
9226 vFAIL("(?(DEFINE)....) does not allow branches");
9227 lastbr = reganode(pRExC_state, IFTHEN, 0); /* Fake one for optimizer. */
9228 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
9229 if (flags & RESTART_UTF8) {
9230 *flagp = RESTART_UTF8;
9233 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9236 REGTAIL(pRExC_state, ret, lastbr);
9239 c = *nextchar(pRExC_state);
9244 vFAIL("Switch (?(condition)... contains too many branches");
9245 ender = reg_node(pRExC_state, TAIL);
9246 REGTAIL(pRExC_state, br, ender);
9248 REGTAIL(pRExC_state, lastbr, ender);
9249 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
9252 REGTAIL(pRExC_state, ret, ender);
9253 RExC_size++; /* XXX WHY do we need this?!!
9254 For large programs it seems to be required
9255 but I can't figure out why. -- dmq*/
9259 vFAIL2("Unknown switch condition (?(%.2s", RExC_parse);
9262 case '[': /* (?[ ... ]) */
9263 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
9266 RExC_parse--; /* for vFAIL to print correctly */
9267 vFAIL("Sequence (? incomplete");
9269 default: /* e.g., (?i) */
9272 parse_lparen_question_flags(pRExC_state);
9273 if (UCHARAT(RExC_parse) != ':') {
9274 nextchar(pRExC_state);
9279 nextchar(pRExC_state);
9289 ret = reganode(pRExC_state, OPEN, parno);
9292 RExC_nestroot = parno;
9293 if (RExC_seen & REG_SEEN_RECURSE
9294 && !RExC_open_parens[parno-1])
9296 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9297 "Setting open paren #%"IVdf" to %d\n",
9298 (IV)parno, REG_NODE_NUM(ret)));
9299 RExC_open_parens[parno-1]= ret;
9302 Set_Node_Length(ret, 1); /* MJD */
9303 Set_Node_Offset(ret, RExC_parse); /* MJD */
9311 /* Pick up the branches, linking them together. */
9312 parse_start = RExC_parse; /* MJD */
9313 br = regbranch(pRExC_state, &flags, 1,depth+1);
9315 /* branch_len = (paren != 0); */
9318 if (flags & RESTART_UTF8) {
9319 *flagp = RESTART_UTF8;
9322 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
9324 if (*RExC_parse == '|') {
9325 if (!SIZE_ONLY && RExC_extralen) {
9326 reginsert(pRExC_state, BRANCHJ, br, depth+1);
9329 reginsert(pRExC_state, BRANCH, br, depth+1);
9330 Set_Node_Length(br, paren != 0);
9331 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
9335 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
9337 else if (paren == ':') {
9338 *flagp |= flags&SIMPLE;
9340 if (is_open) { /* Starts with OPEN. */
9341 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
9343 else if (paren != '?') /* Not Conditional */
9345 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9347 while (*RExC_parse == '|') {
9348 if (!SIZE_ONLY && RExC_extralen) {
9349 ender = reganode(pRExC_state, LONGJMP,0);
9350 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender); /* Append to the previous. */
9353 RExC_extralen += 2; /* Account for LONGJMP. */
9354 nextchar(pRExC_state);
9356 if (RExC_npar > after_freeze)
9357 after_freeze = RExC_npar;
9358 RExC_npar = freeze_paren;
9360 br = regbranch(pRExC_state, &flags, 0, depth+1);
9363 if (flags & RESTART_UTF8) {
9364 *flagp = RESTART_UTF8;
9367 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
9369 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
9371 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9374 if (have_branch || paren != ':') {
9375 /* Make a closing node, and hook it on the end. */
9378 ender = reg_node(pRExC_state, TAIL);
9381 ender = reganode(pRExC_state, CLOSE, parno);
9382 if (!SIZE_ONLY && RExC_seen & REG_SEEN_RECURSE) {
9383 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9384 "Setting close paren #%"IVdf" to %d\n",
9385 (IV)parno, REG_NODE_NUM(ender)));
9386 RExC_close_parens[parno-1]= ender;
9387 if (RExC_nestroot == parno)
9390 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
9391 Set_Node_Length(ender,1); /* MJD */
9397 *flagp &= ~HASWIDTH;
9400 ender = reg_node(pRExC_state, SUCCEED);
9403 ender = reg_node(pRExC_state, END);
9405 assert(!RExC_opend); /* there can only be one! */
9410 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9411 SV * const mysv_val1=sv_newmortal();
9412 SV * const mysv_val2=sv_newmortal();
9413 DEBUG_PARSE_MSG("lsbr");
9414 regprop(RExC_rx, mysv_val1, lastbr);
9415 regprop(RExC_rx, mysv_val2, ender);
9416 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9417 SvPV_nolen_const(mysv_val1),
9418 (IV)REG_NODE_NUM(lastbr),
9419 SvPV_nolen_const(mysv_val2),
9420 (IV)REG_NODE_NUM(ender),
9421 (IV)(ender - lastbr)
9424 REGTAIL(pRExC_state, lastbr, ender);
9426 if (have_branch && !SIZE_ONLY) {
9429 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
9431 /* Hook the tails of the branches to the closing node. */
9432 for (br = ret; br; br = regnext(br)) {
9433 const U8 op = PL_regkind[OP(br)];
9435 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
9436 if (OP(NEXTOPER(br)) != NOTHING || regnext(NEXTOPER(br)) != ender)
9439 else if (op == BRANCHJ) {
9440 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
9441 /* for now we always disable this optimisation * /
9442 if (OP(NEXTOPER(NEXTOPER(br))) != NOTHING || regnext(NEXTOPER(NEXTOPER(br))) != ender)
9448 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
9449 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9450 SV * const mysv_val1=sv_newmortal();
9451 SV * const mysv_val2=sv_newmortal();
9452 DEBUG_PARSE_MSG("NADA");
9453 regprop(RExC_rx, mysv_val1, ret);
9454 regprop(RExC_rx, mysv_val2, ender);
9455 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9456 SvPV_nolen_const(mysv_val1),
9457 (IV)REG_NODE_NUM(ret),
9458 SvPV_nolen_const(mysv_val2),
9459 (IV)REG_NODE_NUM(ender),
9464 if (OP(ender) == TAIL) {
9469 for ( opt= br + 1; opt < ender ; opt++ )
9471 NEXT_OFF(br)= ender - br;
9479 static const char parens[] = "=!<,>";
9481 if (paren && (p = strchr(parens, paren))) {
9482 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
9483 int flag = (p - parens) > 1;
9486 node = SUSPEND, flag = 0;
9487 reginsert(pRExC_state, node,ret, depth+1);
9488 Set_Node_Cur_Length(ret, parse_start);
9489 Set_Node_Offset(ret, parse_start + 1);
9491 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
9495 /* Check for proper termination. */
9497 /* restore original flags, but keep (?p) */
9498 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
9499 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
9500 RExC_parse = oregcomp_parse;
9501 vFAIL("Unmatched (");
9504 else if (!paren && RExC_parse < RExC_end) {
9505 if (*RExC_parse == ')') {
9507 vFAIL("Unmatched )");
9510 FAIL("Junk on end of regexp"); /* "Can't happen". */
9511 assert(0); /* NOTREACHED */
9514 if (RExC_in_lookbehind) {
9515 RExC_in_lookbehind--;
9517 if (after_freeze > RExC_npar)
9518 RExC_npar = after_freeze;
9523 - regbranch - one alternative of an | operator
9525 * Implements the concatenation operator.
9527 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9531 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
9535 regnode *chain = NULL;
9537 I32 flags = 0, c = 0;
9538 GET_RE_DEBUG_FLAGS_DECL;
9540 PERL_ARGS_ASSERT_REGBRANCH;
9542 DEBUG_PARSE("brnc");
9547 if (!SIZE_ONLY && RExC_extralen)
9548 ret = reganode(pRExC_state, BRANCHJ,0);
9550 ret = reg_node(pRExC_state, BRANCH);
9551 Set_Node_Length(ret, 1);
9555 if (!first && SIZE_ONLY)
9556 RExC_extralen += 1; /* BRANCHJ */
9558 *flagp = WORST; /* Tentatively. */
9561 nextchar(pRExC_state);
9562 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
9564 latest = regpiece(pRExC_state, &flags,depth+1);
9565 if (latest == NULL) {
9566 if (flags & TRYAGAIN)
9568 if (flags & RESTART_UTF8) {
9569 *flagp = RESTART_UTF8;
9572 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
9574 else if (ret == NULL)
9576 *flagp |= flags&(HASWIDTH|POSTPONED);
9577 if (chain == NULL) /* First piece. */
9578 *flagp |= flags&SPSTART;
9581 REGTAIL(pRExC_state, chain, latest);
9586 if (chain == NULL) { /* Loop ran zero times. */
9587 chain = reg_node(pRExC_state, NOTHING);
9592 *flagp |= flags&SIMPLE;
9599 - regpiece - something followed by possible [*+?]
9601 * Note that the branching code sequences used for ? and the general cases
9602 * of * and + are somewhat optimized: they use the same NOTHING node as
9603 * both the endmarker for their branch list and the body of the last branch.
9604 * It might seem that this node could be dispensed with entirely, but the
9605 * endmarker role is not redundant.
9607 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
9609 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9613 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
9620 const char * const origparse = RExC_parse;
9622 I32 max = REG_INFTY;
9623 #ifdef RE_TRACK_PATTERN_OFFSETS
9626 const char *maxpos = NULL;
9628 /* Save the original in case we change the emitted regop to a FAIL. */
9629 regnode * const orig_emit = RExC_emit;
9631 GET_RE_DEBUG_FLAGS_DECL;
9633 PERL_ARGS_ASSERT_REGPIECE;
9635 DEBUG_PARSE("piec");
9637 ret = regatom(pRExC_state, &flags,depth+1);
9639 if (flags & (TRYAGAIN|RESTART_UTF8))
9640 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
9642 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
9648 if (op == '{' && regcurly(RExC_parse, FALSE)) {
9650 #ifdef RE_TRACK_PATTERN_OFFSETS
9651 parse_start = RExC_parse; /* MJD */
9653 next = RExC_parse + 1;
9654 while (isDIGIT(*next) || *next == ',') {
9663 if (*next == '}') { /* got one */
9667 min = atoi(RExC_parse);
9671 maxpos = RExC_parse;
9673 if (!max && *maxpos != '0')
9674 max = REG_INFTY; /* meaning "infinity" */
9675 else if (max >= REG_INFTY)
9676 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
9678 nextchar(pRExC_state);
9679 if (max < min) { /* If can't match, warn and optimize to fail
9682 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
9684 /* We can't back off the size because we have to reserve
9685 * enough space for all the things we are about to throw
9686 * away, but we can shrink it by the ammount we are about
9688 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
9691 RExC_emit = orig_emit;
9693 ret = reg_node(pRExC_state, OPFAIL);
9696 else if (max == 0) { /* replace {0} with a nothing node */
9698 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)NOTHING];
9701 RExC_emit = orig_emit;
9703 ret = reg_node(pRExC_state, NOTHING);
9705 /* But the quantifier includes any '?', the non-greedy
9706 * modifier, after the {}, [perl #118375]
9707 * Likewise the '+', the possessive modifier. They are mutually exclusive.
9709 if (RExC_parse < RExC_end && (*RExC_parse == '?' || *RExC_parse == '+') ) {
9710 nextchar(pRExC_state);
9716 if ((flags&SIMPLE)) {
9717 RExC_naughty += 2 + RExC_naughty / 2;
9718 reginsert(pRExC_state, CURLY, ret, depth+1);
9719 Set_Node_Offset(ret, parse_start+1); /* MJD */
9720 Set_Node_Cur_Length(ret, parse_start);
9723 regnode * const w = reg_node(pRExC_state, WHILEM);
9726 REGTAIL(pRExC_state, ret, w);
9727 if (!SIZE_ONLY && RExC_extralen) {
9728 reginsert(pRExC_state, LONGJMP,ret, depth+1);
9729 reginsert(pRExC_state, NOTHING,ret, depth+1);
9730 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
9732 reginsert(pRExC_state, CURLYX,ret, depth+1);
9734 Set_Node_Offset(ret, parse_start+1);
9735 Set_Node_Length(ret,
9736 op == '{' ? (RExC_parse - parse_start) : 1);
9738 if (!SIZE_ONLY && RExC_extralen)
9739 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
9740 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
9742 RExC_whilem_seen++, RExC_extralen += 3;
9743 RExC_naughty += 4 + RExC_naughty; /* compound interest */
9752 ARG1_SET(ret, (U16)min);
9753 ARG2_SET(ret, (U16)max);
9765 #if 0 /* Now runtime fix should be reliable. */
9767 /* if this is reinstated, don't forget to put this back into perldiag:
9769 =item Regexp *+ operand could be empty at {#} in regex m/%s/
9771 (F) The part of the regexp subject to either the * or + quantifier
9772 could match an empty string. The {#} shows in the regular
9773 expression about where the problem was discovered.
9777 if (!(flags&HASWIDTH) && op != '?')
9778 vFAIL("Regexp *+ operand could be empty");
9781 #ifdef RE_TRACK_PATTERN_OFFSETS
9782 parse_start = RExC_parse;
9784 nextchar(pRExC_state);
9786 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
9788 if (op == '*' && (flags&SIMPLE)) {
9789 reginsert(pRExC_state, STAR, ret, depth+1);
9793 else if (op == '*') {
9797 else if (op == '+' && (flags&SIMPLE)) {
9798 reginsert(pRExC_state, PLUS, ret, depth+1);
9802 else if (op == '+') {
9806 else if (op == '?') {
9811 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
9812 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
9813 ckWARN3reg(RExC_parse,
9814 "%.*s matches null string many times",
9815 (int)(RExC_parse >= origparse ? RExC_parse - origparse : 0),
9817 (void)ReREFCNT_inc(RExC_rx_sv);
9820 if (RExC_parse < RExC_end && *RExC_parse == '?') {
9821 nextchar(pRExC_state);
9822 reginsert(pRExC_state, MINMOD, ret, depth+1);
9823 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
9826 if (RExC_parse < RExC_end && *RExC_parse == '+') {
9828 nextchar(pRExC_state);
9829 ender = reg_node(pRExC_state, SUCCEED);
9830 REGTAIL(pRExC_state, ret, ender);
9831 reginsert(pRExC_state, SUSPEND, ret, depth+1);
9833 ender = reg_node(pRExC_state, TAIL);
9834 REGTAIL(pRExC_state, ret, ender);
9837 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
9839 vFAIL("Nested quantifiers");
9846 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p, UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
9847 const bool strict /* Apply stricter parsing rules? */
9851 /* This is expected to be called by a parser routine that has recognized '\N'
9852 and needs to handle the rest. RExC_parse is expected to point at the first
9853 char following the N at the time of the call. On successful return,
9854 RExC_parse has been updated to point to just after the sequence identified
9855 by this routine, and <*flagp> has been updated.
9857 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
9860 \N may begin either a named sequence, or if outside a character class, mean
9861 to match a non-newline. For non single-quoted regexes, the tokenizer has
9862 attempted to decide which, and in the case of a named sequence, converted it
9863 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
9864 where c1... are the characters in the sequence. For single-quoted regexes,
9865 the tokenizer passes the \N sequence through unchanged; this code will not
9866 attempt to determine this nor expand those, instead raising a syntax error.
9867 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
9868 or there is no '}', it signals that this \N occurrence means to match a
9871 Only the \N{U+...} form should occur in a character class, for the same
9872 reason that '.' inside a character class means to just match a period: it
9873 just doesn't make sense.
9875 The function raises an error (via vFAIL), and doesn't return for various
9876 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
9877 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
9878 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
9879 only possible if node_p is non-NULL.
9882 If <valuep> is non-null, it means the caller can accept an input sequence
9883 consisting of a just a single code point; <*valuep> is set to that value
9884 if the input is such.
9886 If <node_p> is non-null it signifies that the caller can accept any other
9887 legal sequence (i.e., one that isn't just a single code point). <*node_p>
9889 1) \N means not-a-NL: points to a newly created REG_ANY node;
9890 2) \N{}: points to a new NOTHING node;
9891 3) otherwise: points to a new EXACT node containing the resolved
9893 Note that FALSE is returned for single code point sequences if <valuep> is
9897 char * endbrace; /* '}' following the name */
9899 char *endchar; /* Points to '.' or '}' ending cur char in the input
9901 bool has_multiple_chars; /* true if the input stream contains a sequence of
9902 more than one character */
9904 GET_RE_DEBUG_FLAGS_DECL;
9906 PERL_ARGS_ASSERT_GROK_BSLASH_N;
9910 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
9912 /* The [^\n] meaning of \N ignores spaces and comments under the /x
9913 * modifier. The other meaning does not */
9914 p = (RExC_flags & RXf_PMf_EXTENDED)
9915 ? regwhite( pRExC_state, RExC_parse )
9918 /* Disambiguate between \N meaning a named character versus \N meaning
9919 * [^\n]. The former is assumed when it can't be the latter. */
9920 if (*p != '{' || regcurly(p, FALSE)) {
9923 /* no bare \N in a charclass */
9924 if (in_char_class) {
9925 vFAIL("\\N in a character class must be a named character: \\N{...}");
9929 nextchar(pRExC_state);
9930 *node_p = reg_node(pRExC_state, REG_ANY);
9931 *flagp |= HASWIDTH|SIMPLE;
9934 Set_Node_Length(*node_p, 1); /* MJD */
9938 /* Here, we have decided it should be a named character or sequence */
9940 /* The test above made sure that the next real character is a '{', but
9941 * under the /x modifier, it could be separated by space (or a comment and
9942 * \n) and this is not allowed (for consistency with \x{...} and the
9943 * tokenizer handling of \N{NAME}). */
9944 if (*RExC_parse != '{') {
9945 vFAIL("Missing braces on \\N{}");
9948 RExC_parse++; /* Skip past the '{' */
9950 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
9951 || ! (endbrace == RExC_parse /* nothing between the {} */
9952 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below */
9953 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg) */
9955 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
9956 vFAIL("\\N{NAME} must be resolved by the lexer");
9959 if (endbrace == RExC_parse) { /* empty: \N{} */
9962 *node_p = reg_node(pRExC_state,NOTHING);
9964 else if (in_char_class) {
9965 if (SIZE_ONLY && in_char_class) {
9967 RExC_parse++; /* Position after the "}" */
9968 vFAIL("Zero length \\N{}");
9971 ckWARNreg(RExC_parse,
9972 "Ignoring zero length \\N{} in character class");
9980 nextchar(pRExC_state);
9984 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
9985 RExC_parse += 2; /* Skip past the 'U+' */
9987 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9989 /* Code points are separated by dots. If none, there is only one code
9990 * point, and is terminated by the brace */
9991 has_multiple_chars = (endchar < endbrace);
9993 if (valuep && (! has_multiple_chars || in_char_class)) {
9994 /* We only pay attention to the first char of
9995 multichar strings being returned in char classes. I kinda wonder
9996 if this makes sense as it does change the behaviour
9997 from earlier versions, OTOH that behaviour was broken
9998 as well. XXX Solution is to recharacterize as
9999 [rest-of-class]|multi1|multi2... */
10001 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
10002 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
10003 | PERL_SCAN_DISALLOW_PREFIX
10004 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
10006 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
10008 /* The tokenizer should have guaranteed validity, but it's possible to
10009 * bypass it by using single quoting, so check */
10010 if (length_of_hex == 0
10011 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
10013 RExC_parse += length_of_hex; /* Includes all the valid */
10014 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
10015 ? UTF8SKIP(RExC_parse)
10017 /* Guard against malformed utf8 */
10018 if (RExC_parse >= endchar) {
10019 RExC_parse = endchar;
10021 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10024 if (in_char_class && has_multiple_chars) {
10026 RExC_parse = endbrace;
10027 vFAIL("\\N{} in character class restricted to one character");
10030 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
10034 RExC_parse = endbrace + 1;
10036 else if (! node_p || ! has_multiple_chars) {
10038 /* Here, the input is legal, but not according to the caller's
10039 * options. We fail without advancing the parse, so that the
10040 * caller can try again */
10046 /* What is done here is to convert this to a sub-pattern of the form
10047 * (?:\x{char1}\x{char2}...)
10048 * and then call reg recursively. That way, it retains its atomicness,
10049 * while not having to worry about special handling that some code
10050 * points may have. toke.c has converted the original Unicode values
10051 * to native, so that we can just pass on the hex values unchanged. We
10052 * do have to set a flag to keep recoding from happening in the
10055 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
10057 char *orig_end = RExC_end;
10060 while (RExC_parse < endbrace) {
10062 /* Convert to notation the rest of the code understands */
10063 sv_catpv(substitute_parse, "\\x{");
10064 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
10065 sv_catpv(substitute_parse, "}");
10067 /* Point to the beginning of the next character in the sequence. */
10068 RExC_parse = endchar + 1;
10069 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10071 sv_catpv(substitute_parse, ")");
10073 RExC_parse = SvPV(substitute_parse, len);
10075 /* Don't allow empty number */
10077 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10079 RExC_end = RExC_parse + len;
10081 /* The values are Unicode, and therefore not subject to recoding */
10082 RExC_override_recoding = 1;
10084 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
10085 if (flags & RESTART_UTF8) {
10086 *flagp = RESTART_UTF8;
10089 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
10092 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10094 RExC_parse = endbrace;
10095 RExC_end = orig_end;
10096 RExC_override_recoding = 0;
10098 nextchar(pRExC_state);
10108 * It returns the code point in utf8 for the value in *encp.
10109 * value: a code value in the source encoding
10110 * encp: a pointer to an Encode object
10112 * If the result from Encode is not a single character,
10113 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
10116 S_reg_recode(pTHX_ const char value, SV **encp)
10119 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
10120 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
10121 const STRLEN newlen = SvCUR(sv);
10122 UV uv = UNICODE_REPLACEMENT;
10124 PERL_ARGS_ASSERT_REG_RECODE;
10128 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
10131 if (!newlen || numlen != newlen) {
10132 uv = UNICODE_REPLACEMENT;
10138 PERL_STATIC_INLINE U8
10139 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
10143 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
10149 op = get_regex_charset(RExC_flags);
10150 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
10151 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
10152 been, so there is no hole */
10155 return op + EXACTF;
10158 PERL_STATIC_INLINE void
10159 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state, regnode *node, I32* flagp, STRLEN len, UV code_point)
10161 /* This knows the details about sizing an EXACTish node, setting flags for
10162 * it (by setting <*flagp>, and potentially populating it with a single
10165 * If <len> (the length in bytes) is non-zero, this function assumes that
10166 * the node has already been populated, and just does the sizing. In this
10167 * case <code_point> should be the final code point that has already been
10168 * placed into the node. This value will be ignored except that under some
10169 * circumstances <*flagp> is set based on it.
10171 * If <len> is zero, the function assumes that the node is to contain only
10172 * the single character given by <code_point> and calculates what <len>
10173 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10174 * additionally will populate the node's STRING with <code_point>, if <len>
10175 * is 0. In both cases <*flagp> is appropriately set
10177 * It knows that under FOLD, the Latin Sharp S and UTF characters above
10178 * 255, must be folded (the former only when the rules indicate it can
10181 bool len_passed_in = cBOOL(len != 0);
10182 U8 character[UTF8_MAXBYTES_CASE+1];
10184 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10186 if (! len_passed_in) {
10188 if (FOLD && (! LOC || code_point > 255)) {
10189 _to_uni_fold_flags(NATIVE_TO_UNI(code_point),
10192 FOLD_FLAGS_FULL | ((LOC)
10193 ? FOLD_FLAGS_LOCALE
10194 : (ASCII_FOLD_RESTRICTED)
10195 ? FOLD_FLAGS_NOMIX_ASCII
10199 uvchr_to_utf8( character, code_point);
10200 len = UTF8SKIP(character);
10204 || code_point != LATIN_SMALL_LETTER_SHARP_S
10205 || ASCII_FOLD_RESTRICTED
10206 || ! AT_LEAST_UNI_SEMANTICS)
10208 *character = (U8) code_point;
10213 *(character + 1) = 's';
10219 RExC_size += STR_SZ(len);
10222 RExC_emit += STR_SZ(len);
10223 STR_LEN(node) = len;
10224 if (! len_passed_in) {
10225 Copy((char *) character, STRING(node), len, char);
10229 *flagp |= HASWIDTH;
10231 /* A single character node is SIMPLE, except for the special-cased SHARP S
10233 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
10234 && (code_point != LATIN_SMALL_LETTER_SHARP_S
10235 || ! FOLD || ! DEPENDS_SEMANTICS))
10242 - regatom - the lowest level
10244 Try to identify anything special at the start of the pattern. If there
10245 is, then handle it as required. This may involve generating a single regop,
10246 such as for an assertion; or it may involve recursing, such as to
10247 handle a () structure.
10249 If the string doesn't start with something special then we gobble up
10250 as much literal text as we can.
10252 Once we have been able to handle whatever type of thing started the
10253 sequence, we return.
10255 Note: we have to be careful with escapes, as they can be both literal
10256 and special, and in the case of \10 and friends, context determines which.
10258 A summary of the code structure is:
10260 switch (first_byte) {
10261 cases for each special:
10262 handle this special;
10265 switch (2nd byte) {
10266 cases for each unambiguous special:
10267 handle this special;
10269 cases for each ambigous special/literal:
10271 if (special) handle here
10273 default: // unambiguously literal:
10276 default: // is a literal char
10279 create EXACTish node for literal;
10280 while (more input and node isn't full) {
10281 switch (input_byte) {
10282 cases for each special;
10283 make sure parse pointer is set so that the next call to
10284 regatom will see this special first
10285 goto loopdone; // EXACTish node terminated by prev. char
10287 append char to EXACTISH node;
10289 get next input byte;
10293 return the generated node;
10295 Specifically there are two separate switches for handling
10296 escape sequences, with the one for handling literal escapes requiring
10297 a dummy entry for all of the special escapes that are actually handled
10300 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
10302 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10304 Otherwise does not return NULL.
10308 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10311 regnode *ret = NULL;
10313 char *parse_start = RExC_parse;
10317 GET_RE_DEBUG_FLAGS_DECL;
10319 *flagp = WORST; /* Tentatively. */
10321 DEBUG_PARSE("atom");
10323 PERL_ARGS_ASSERT_REGATOM;
10326 switch ((U8)*RExC_parse) {
10328 RExC_seen_zerolen++;
10329 nextchar(pRExC_state);
10330 if (RExC_flags & RXf_PMf_MULTILINE)
10331 ret = reg_node(pRExC_state, MBOL);
10332 else if (RExC_flags & RXf_PMf_SINGLELINE)
10333 ret = reg_node(pRExC_state, SBOL);
10335 ret = reg_node(pRExC_state, BOL);
10336 Set_Node_Length(ret, 1); /* MJD */
10339 nextchar(pRExC_state);
10341 RExC_seen_zerolen++;
10342 if (RExC_flags & RXf_PMf_MULTILINE)
10343 ret = reg_node(pRExC_state, MEOL);
10344 else if (RExC_flags & RXf_PMf_SINGLELINE)
10345 ret = reg_node(pRExC_state, SEOL);
10347 ret = reg_node(pRExC_state, EOL);
10348 Set_Node_Length(ret, 1); /* MJD */
10351 nextchar(pRExC_state);
10352 if (RExC_flags & RXf_PMf_SINGLELINE)
10353 ret = reg_node(pRExC_state, SANY);
10355 ret = reg_node(pRExC_state, REG_ANY);
10356 *flagp |= HASWIDTH|SIMPLE;
10358 Set_Node_Length(ret, 1); /* MJD */
10362 char * const oregcomp_parse = ++RExC_parse;
10363 ret = regclass(pRExC_state, flagp,depth+1,
10364 FALSE, /* means parse the whole char class */
10365 TRUE, /* allow multi-char folds */
10366 FALSE, /* don't silence non-portable warnings. */
10368 if (*RExC_parse != ']') {
10369 RExC_parse = oregcomp_parse;
10370 vFAIL("Unmatched [");
10373 if (*flagp & RESTART_UTF8)
10375 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
10378 nextchar(pRExC_state);
10379 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
10383 nextchar(pRExC_state);
10384 ret = reg(pRExC_state, 2, &flags,depth+1);
10386 if (flags & TRYAGAIN) {
10387 if (RExC_parse == RExC_end) {
10388 /* Make parent create an empty node if needed. */
10389 *flagp |= TRYAGAIN;
10394 if (flags & RESTART_UTF8) {
10395 *flagp = RESTART_UTF8;
10398 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"", (UV) flags);
10400 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10404 if (flags & TRYAGAIN) {
10405 *flagp |= TRYAGAIN;
10408 vFAIL("Internal urp");
10409 /* Supposed to be caught earlier. */
10412 if (!regcurly(RExC_parse, FALSE)) {
10421 vFAIL("Quantifier follows nothing");
10426 This switch handles escape sequences that resolve to some kind
10427 of special regop and not to literal text. Escape sequnces that
10428 resolve to literal text are handled below in the switch marked
10431 Every entry in this switch *must* have a corresponding entry
10432 in the literal escape switch. However, the opposite is not
10433 required, as the default for this switch is to jump to the
10434 literal text handling code.
10436 switch ((U8)*++RExC_parse) {
10438 /* Special Escapes */
10440 RExC_seen_zerolen++;
10441 ret = reg_node(pRExC_state, SBOL);
10443 goto finish_meta_pat;
10445 ret = reg_node(pRExC_state, GPOS);
10446 RExC_seen |= REG_SEEN_GPOS;
10448 goto finish_meta_pat;
10450 RExC_seen_zerolen++;
10451 ret = reg_node(pRExC_state, KEEPS);
10453 /* XXX:dmq : disabling in-place substitution seems to
10454 * be necessary here to avoid cases of memory corruption, as
10455 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
10457 RExC_seen |= REG_SEEN_LOOKBEHIND;
10458 goto finish_meta_pat;
10460 ret = reg_node(pRExC_state, SEOL);
10462 RExC_seen_zerolen++; /* Do not optimize RE away */
10463 goto finish_meta_pat;
10465 ret = reg_node(pRExC_state, EOS);
10467 RExC_seen_zerolen++; /* Do not optimize RE away */
10468 goto finish_meta_pat;
10470 ret = reg_node(pRExC_state, CANY);
10471 RExC_seen |= REG_SEEN_CANY;
10472 *flagp |= HASWIDTH|SIMPLE;
10473 goto finish_meta_pat;
10475 ret = reg_node(pRExC_state, CLUMP);
10476 *flagp |= HASWIDTH;
10477 goto finish_meta_pat;
10483 arg = ANYOF_WORDCHAR;
10487 RExC_seen_zerolen++;
10488 RExC_seen |= REG_SEEN_LOOKBEHIND;
10489 op = BOUND + get_regex_charset(RExC_flags);
10490 if (op > BOUNDA) { /* /aa is same as /a */
10493 ret = reg_node(pRExC_state, op);
10494 FLAGS(ret) = get_regex_charset(RExC_flags);
10496 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10497 ckWARNdep(RExC_parse, "\"\\b{\" is deprecated; use \"\\b\\{\" or \"\\b[{]\" instead");
10499 goto finish_meta_pat;
10501 RExC_seen_zerolen++;
10502 RExC_seen |= REG_SEEN_LOOKBEHIND;
10503 op = NBOUND + get_regex_charset(RExC_flags);
10504 if (op > NBOUNDA) { /* /aa is same as /a */
10507 ret = reg_node(pRExC_state, op);
10508 FLAGS(ret) = get_regex_charset(RExC_flags);
10510 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10511 ckWARNdep(RExC_parse, "\"\\B{\" is deprecated; use \"\\B\\{\" or \"\\B[{]\" instead");
10513 goto finish_meta_pat;
10523 ret = reg_node(pRExC_state, LNBREAK);
10524 *flagp |= HASWIDTH|SIMPLE;
10525 goto finish_meta_pat;
10533 goto join_posix_op_known;
10539 arg = ANYOF_VERTWS;
10541 goto join_posix_op_known;
10551 op = POSIXD + get_regex_charset(RExC_flags);
10552 if (op > POSIXA) { /* /aa is same as /a */
10556 join_posix_op_known:
10559 op += NPOSIXD - POSIXD;
10562 ret = reg_node(pRExC_state, op);
10564 FLAGS(ret) = namedclass_to_classnum(arg);
10567 *flagp |= HASWIDTH|SIMPLE;
10571 nextchar(pRExC_state);
10572 Set_Node_Length(ret, 2); /* MJD */
10578 char* parse_start = RExC_parse - 2;
10583 ret = regclass(pRExC_state, flagp,depth+1,
10584 TRUE, /* means just parse this element */
10585 FALSE, /* don't allow multi-char folds */
10586 FALSE, /* don't silence non-portable warnings.
10587 It would be a bug if these returned
10590 /* regclass() can only return RESTART_UTF8 if multi-char folds
10593 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
10598 Set_Node_Offset(ret, parse_start + 2);
10599 Set_Node_Cur_Length(ret, parse_start);
10600 nextchar(pRExC_state);
10604 /* Handle \N and \N{NAME} with multiple code points here and not
10605 * below because it can be multicharacter. join_exact() will join
10606 * them up later on. Also this makes sure that things like
10607 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
10608 * The options to the grok function call causes it to fail if the
10609 * sequence is just a single code point. We then go treat it as
10610 * just another character in the current EXACT node, and hence it
10611 * gets uniform treatment with all the other characters. The
10612 * special treatment for quantifiers is not needed for such single
10613 * character sequences */
10615 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
10616 FALSE /* not strict */ )) {
10617 if (*flagp & RESTART_UTF8)
10623 case 'k': /* Handle \k<NAME> and \k'NAME' */
10626 char ch= RExC_parse[1];
10627 if (ch != '<' && ch != '\'' && ch != '{') {
10629 vFAIL2("Sequence %.2s... not terminated",parse_start);
10631 /* this pretty much dupes the code for (?P=...) in reg(), if
10632 you change this make sure you change that */
10633 char* name_start = (RExC_parse += 2);
10635 SV *sv_dat = reg_scan_name(pRExC_state,
10636 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10637 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
10638 if (RExC_parse == name_start || *RExC_parse != ch)
10639 vFAIL2("Sequence %.3s... not terminated",parse_start);
10642 num = add_data( pRExC_state, 1, "S" );
10643 RExC_rxi->data->data[num]=(void*)sv_dat;
10644 SvREFCNT_inc_simple_void(sv_dat);
10648 ret = reganode(pRExC_state,
10651 : (ASCII_FOLD_RESTRICTED)
10653 : (AT_LEAST_UNI_SEMANTICS)
10659 *flagp |= HASWIDTH;
10661 /* override incorrect value set in reganode MJD */
10662 Set_Node_Offset(ret, parse_start+1);
10663 Set_Node_Cur_Length(ret, parse_start);
10664 nextchar(pRExC_state);
10670 case '1': case '2': case '3': case '4':
10671 case '5': case '6': case '7': case '8': case '9':
10674 bool isg = *RExC_parse == 'g';
10679 if (*RExC_parse == '{') {
10683 if (*RExC_parse == '-') {
10687 if (hasbrace && !isDIGIT(*RExC_parse)) {
10688 if (isrel) RExC_parse--;
10690 goto parse_named_seq;
10692 num = atoi(RExC_parse);
10693 if (isg && num == 0) {
10694 if (*RExC_parse == '0') {
10695 vFAIL("Reference to invalid group 0");
10698 vFAIL("Unterminated \\g... pattern");
10702 num = RExC_npar - num;
10704 vFAIL("Reference to nonexistent or unclosed group");
10706 if (!isg && num > 9 && num >= RExC_npar)
10707 /* Probably a character specified in octal, e.g. \35 */
10710 #ifdef RE_TRACK_PATTERN_OFFSETS
10711 char * const parse_start = RExC_parse - 1; /* MJD */
10713 while (isDIGIT(*RExC_parse))
10716 if (*RExC_parse != '}')
10717 vFAIL("Unterminated \\g{...} pattern");
10721 if (num > (I32)RExC_rx->nparens)
10722 vFAIL("Reference to nonexistent group");
10725 ret = reganode(pRExC_state,
10728 : (ASCII_FOLD_RESTRICTED)
10730 : (AT_LEAST_UNI_SEMANTICS)
10736 *flagp |= HASWIDTH;
10738 /* override incorrect value set in reganode MJD */
10739 Set_Node_Offset(ret, parse_start+1);
10740 Set_Node_Cur_Length(ret, parse_start);
10742 nextchar(pRExC_state);
10747 if (RExC_parse >= RExC_end)
10748 FAIL("Trailing \\");
10751 /* Do not generate "unrecognized" warnings here, we fall
10752 back into the quick-grab loop below */
10759 if (RExC_flags & RXf_PMf_EXTENDED) {
10760 if ( reg_skipcomment( pRExC_state ) )
10767 parse_start = RExC_parse - 1;
10776 #define MAX_NODE_STRING_SIZE 127
10777 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
10779 U8 upper_parse = MAX_NODE_STRING_SIZE;
10782 bool next_is_quantifier;
10783 char * oldp = NULL;
10785 /* If a folding node contains only code points that don't
10786 * participate in folds, it can be changed into an EXACT node,
10787 * which allows the optimizer more things to look for */
10791 node_type = compute_EXACTish(pRExC_state);
10792 ret = reg_node(pRExC_state, node_type);
10794 /* In pass1, folded, we use a temporary buffer instead of the
10795 * actual node, as the node doesn't exist yet */
10796 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
10802 /* We do the EXACTFish to EXACT node only if folding, and not if in
10803 * locale, as whether a character folds or not isn't known until
10805 maybe_exact = FOLD && ! LOC;
10807 /* XXX The node can hold up to 255 bytes, yet this only goes to
10808 * 127. I (khw) do not know why. Keeping it somewhat less than
10809 * 255 allows us to not have to worry about overflow due to
10810 * converting to utf8 and fold expansion, but that value is
10811 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
10812 * split up by this limit into a single one using the real max of
10813 * 255. Even at 127, this breaks under rare circumstances. If
10814 * folding, we do not want to split a node at a character that is a
10815 * non-final in a multi-char fold, as an input string could just
10816 * happen to want to match across the node boundary. The join
10817 * would solve that problem if the join actually happens. But a
10818 * series of more than two nodes in a row each of 127 would cause
10819 * the first join to succeed to get to 254, but then there wouldn't
10820 * be room for the next one, which could at be one of those split
10821 * multi-char folds. I don't know of any fool-proof solution. One
10822 * could back off to end with only a code point that isn't such a
10823 * non-final, but it is possible for there not to be any in the
10825 for (p = RExC_parse - 1;
10826 len < upper_parse && p < RExC_end;
10831 if (RExC_flags & RXf_PMf_EXTENDED)
10832 p = regwhite( pRExC_state, p );
10843 /* Literal Escapes Switch
10845 This switch is meant to handle escape sequences that
10846 resolve to a literal character.
10848 Every escape sequence that represents something
10849 else, like an assertion or a char class, is handled
10850 in the switch marked 'Special Escapes' above in this
10851 routine, but also has an entry here as anything that
10852 isn't explicitly mentioned here will be treated as
10853 an unescaped equivalent literal.
10856 switch ((U8)*++p) {
10857 /* These are all the special escapes. */
10858 case 'A': /* Start assertion */
10859 case 'b': case 'B': /* Word-boundary assertion*/
10860 case 'C': /* Single char !DANGEROUS! */
10861 case 'd': case 'D': /* digit class */
10862 case 'g': case 'G': /* generic-backref, pos assertion */
10863 case 'h': case 'H': /* HORIZWS */
10864 case 'k': case 'K': /* named backref, keep marker */
10865 case 'p': case 'P': /* Unicode property */
10866 case 'R': /* LNBREAK */
10867 case 's': case 'S': /* space class */
10868 case 'v': case 'V': /* VERTWS */
10869 case 'w': case 'W': /* word class */
10870 case 'X': /* eXtended Unicode "combining character sequence" */
10871 case 'z': case 'Z': /* End of line/string assertion */
10875 /* Anything after here is an escape that resolves to a
10876 literal. (Except digits, which may or may not)
10882 case 'N': /* Handle a single-code point named character. */
10883 /* The options cause it to fail if a multiple code
10884 * point sequence. Handle those in the switch() above
10886 RExC_parse = p + 1;
10887 if (! grok_bslash_N(pRExC_state, NULL, &ender,
10888 flagp, depth, FALSE,
10889 FALSE /* not strict */ ))
10891 if (*flagp & RESTART_UTF8)
10892 FAIL("panic: grok_bslash_N set RESTART_UTF8");
10893 RExC_parse = p = oldp;
10897 if (ender > 0xff) {
10914 ender = ASCII_TO_NATIVE('\033');
10918 ender = ASCII_TO_NATIVE('\007');
10924 const char* error_msg;
10926 bool valid = grok_bslash_o(&p,
10929 TRUE, /* out warnings */
10930 FALSE, /* not strict */
10931 TRUE, /* Output warnings
10936 RExC_parse = p; /* going to die anyway; point
10937 to exact spot of failure */
10941 if (PL_encoding && ender < 0x100) {
10942 goto recode_encoding;
10944 if (ender > 0xff) {
10951 UV result = UV_MAX; /* initialize to erroneous
10953 const char* error_msg;
10955 bool valid = grok_bslash_x(&p,
10958 TRUE, /* out warnings */
10959 FALSE, /* not strict */
10960 TRUE, /* Output warnings
10965 RExC_parse = p; /* going to die anyway; point
10966 to exact spot of failure */
10971 if (PL_encoding && ender < 0x100) {
10972 goto recode_encoding;
10974 if (ender > 0xff) {
10981 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
10983 case '0': case '1': case '2': case '3':case '4':
10984 case '5': case '6': case '7':
10986 (isDIGIT(p[1]) && atoi(p) >= RExC_npar))
10988 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
10990 ender = grok_oct(p, &numlen, &flags, NULL);
10991 if (ender > 0xff) {
10995 if (SIZE_ONLY /* like \08, \178 */
10998 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
11000 reg_warn_non_literal_string(
11002 form_short_octal_warning(p, numlen));
11005 else { /* Not to be treated as an octal constant, go
11010 if (PL_encoding && ender < 0x100)
11011 goto recode_encoding;
11014 if (! RExC_override_recoding) {
11015 SV* enc = PL_encoding;
11016 ender = reg_recode((const char)(U8)ender, &enc);
11017 if (!enc && SIZE_ONLY)
11018 ckWARNreg(p, "Invalid escape in the specified encoding");
11024 FAIL("Trailing \\");
11027 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
11028 /* Include any { following the alpha to emphasize
11029 * that it could be part of an escape at some point
11031 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
11032 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
11034 goto normal_default;
11035 } /* End of switch on '\' */
11037 default: /* A literal character */
11040 && RExC_flags & RXf_PMf_EXTENDED
11041 && ckWARN_d(WARN_DEPRECATED)
11042 && is_PATWS_non_low(p, UTF))
11044 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
11045 "Escape literal pattern white space under /x");
11049 if (UTF8_IS_START(*p) && UTF) {
11051 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
11052 &numlen, UTF8_ALLOW_DEFAULT);
11058 } /* End of switch on the literal */
11060 /* Here, have looked at the literal character and <ender>
11061 * contains its ordinal, <p> points to the character after it
11064 if ( RExC_flags & RXf_PMf_EXTENDED)
11065 p = regwhite( pRExC_state, p );
11067 /* If the next thing is a quantifier, it applies to this
11068 * character only, which means that this character has to be in
11069 * its own node and can't just be appended to the string in an
11070 * existing node, so if there are already other characters in
11071 * the node, close the node with just them, and set up to do
11072 * this character again next time through, when it will be the
11073 * only thing in its new node */
11074 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
11082 const STRLEN unilen = reguni(pRExC_state, ender, s);
11088 /* The loop increments <len> each time, as all but this
11089 * path (and one other) through it add a single byte to
11090 * the EXACTish node. But this one has changed len to
11091 * be the correct final value, so subtract one to
11092 * cancel out the increment that follows */
11096 REGC((char)ender, s++);
11101 /* See comments for join_exact() as to why we fold this
11102 * non-UTF at compile time */
11103 || (node_type == EXACTFU
11104 && ender == LATIN_SMALL_LETTER_SHARP_S)))
11106 *(s++) = (char) ender;
11107 maybe_exact &= ! IS_IN_SOME_FOLD_L1(ender);
11111 /* Prime the casefolded buffer. Locale rules, which apply
11112 * only to code points < 256, aren't known until execution,
11113 * so for them, just output the original character using
11114 * utf8. If we start to fold non-UTF patterns, be sure to
11115 * update join_exact() */
11116 if (LOC && ender < 256) {
11117 if (UNI_IS_INVARIANT(ender)) {
11121 *s = UTF8_TWO_BYTE_HI(ender);
11122 *(s + 1) = UTF8_TWO_BYTE_LO(ender);
11127 UV folded = _to_uni_fold_flags(
11132 | ((LOC) ? FOLD_FLAGS_LOCALE
11133 : (ASCII_FOLD_RESTRICTED)
11134 ? FOLD_FLAGS_NOMIX_ASCII
11138 /* If this node only contains non-folding code points
11139 * so far, see if this new one is also non-folding */
11141 if (folded != ender) {
11142 maybe_exact = FALSE;
11145 /* Here the fold is the original; we have
11146 * to check further to see if anything
11148 if (! PL_utf8_foldable) {
11149 SV* swash = swash_init("utf8",
11151 &PL_sv_undef, 1, 0);
11153 _get_swash_invlist(swash);
11154 SvREFCNT_dec_NN(swash);
11156 if (_invlist_contains_cp(PL_utf8_foldable,
11159 maybe_exact = FALSE;
11167 /* The loop increments <len> each time, as all but this
11168 * path (and one other) through it add a single byte to the
11169 * EXACTish node. But this one has changed len to be the
11170 * correct final value, so subtract one to cancel out the
11171 * increment that follows */
11172 len += foldlen - 1;
11175 if (next_is_quantifier) {
11177 /* Here, the next input is a quantifier, and to get here,
11178 * the current character is the only one in the node.
11179 * Also, here <len> doesn't include the final byte for this
11185 } /* End of loop through literal characters */
11187 /* Here we have either exhausted the input or ran out of room in
11188 * the node. (If we encountered a character that can't be in the
11189 * node, transfer is made directly to <loopdone>, and so we
11190 * wouldn't have fallen off the end of the loop.) In the latter
11191 * case, we artificially have to split the node into two, because
11192 * we just don't have enough space to hold everything. This
11193 * creates a problem if the final character participates in a
11194 * multi-character fold in the non-final position, as a match that
11195 * should have occurred won't, due to the way nodes are matched,
11196 * and our artificial boundary. So back off until we find a non-
11197 * problematic character -- one that isn't at the beginning or
11198 * middle of such a fold. (Either it doesn't participate in any
11199 * folds, or appears only in the final position of all the folds it
11200 * does participate in.) A better solution with far fewer false
11201 * positives, and that would fill the nodes more completely, would
11202 * be to actually have available all the multi-character folds to
11203 * test against, and to back-off only far enough to be sure that
11204 * this node isn't ending with a partial one. <upper_parse> is set
11205 * further below (if we need to reparse the node) to include just
11206 * up through that final non-problematic character that this code
11207 * identifies, so when it is set to less than the full node, we can
11208 * skip the rest of this */
11209 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
11211 const STRLEN full_len = len;
11213 assert(len >= MAX_NODE_STRING_SIZE);
11215 /* Here, <s> points to the final byte of the final character.
11216 * Look backwards through the string until find a non-
11217 * problematic character */
11221 /* These two have no multi-char folds to non-UTF characters
11223 if (ASCII_FOLD_RESTRICTED || LOC) {
11227 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
11231 if (! PL_NonL1NonFinalFold) {
11232 PL_NonL1NonFinalFold = _new_invlist_C_array(
11233 NonL1_Perl_Non_Final_Folds_invlist);
11236 /* Point to the first byte of the final character */
11237 s = (char *) utf8_hop((U8 *) s, -1);
11239 while (s >= s0) { /* Search backwards until find
11240 non-problematic char */
11241 if (UTF8_IS_INVARIANT(*s)) {
11243 /* There are no ascii characters that participate
11244 * in multi-char folds under /aa. In EBCDIC, the
11245 * non-ascii invariants are all control characters,
11246 * so don't ever participate in any folds. */
11247 if (ASCII_FOLD_RESTRICTED
11248 || ! IS_NON_FINAL_FOLD(*s))
11253 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
11255 /* No Latin1 characters participate in multi-char
11256 * folds under /l */
11258 || ! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_UNI(
11264 else if (! _invlist_contains_cp(
11265 PL_NonL1NonFinalFold,
11266 valid_utf8_to_uvchr((U8 *) s, NULL)))
11271 /* Here, the current character is problematic in that
11272 * it does occur in the non-final position of some
11273 * fold, so try the character before it, but have to
11274 * special case the very first byte in the string, so
11275 * we don't read outside the string */
11276 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
11277 } /* End of loop backwards through the string */
11279 /* If there were only problematic characters in the string,
11280 * <s> will point to before s0, in which case the length
11281 * should be 0, otherwise include the length of the
11282 * non-problematic character just found */
11283 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
11286 /* Here, have found the final character, if any, that is
11287 * non-problematic as far as ending the node without splitting
11288 * it across a potential multi-char fold. <len> contains the
11289 * number of bytes in the node up-to and including that
11290 * character, or is 0 if there is no such character, meaning
11291 * the whole node contains only problematic characters. In
11292 * this case, give up and just take the node as-is. We can't
11298 /* Here, the node does contain some characters that aren't
11299 * problematic. If one such is the final character in the
11300 * node, we are done */
11301 if (len == full_len) {
11304 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
11306 /* If the final character is problematic, but the
11307 * penultimate is not, back-off that last character to
11308 * later start a new node with it */
11313 /* Here, the final non-problematic character is earlier
11314 * in the input than the penultimate character. What we do
11315 * is reparse from the beginning, going up only as far as
11316 * this final ok one, thus guaranteeing that the node ends
11317 * in an acceptable character. The reason we reparse is
11318 * that we know how far in the character is, but we don't
11319 * know how to correlate its position with the input parse.
11320 * An alternate implementation would be to build that
11321 * correlation as we go along during the original parse,
11322 * but that would entail extra work for every node, whereas
11323 * this code gets executed only when the string is too
11324 * large for the node, and the final two characters are
11325 * problematic, an infrequent occurrence. Yet another
11326 * possible strategy would be to save the tail of the
11327 * string, and the next time regatom is called, initialize
11328 * with that. The problem with this is that unless you
11329 * back off one more character, you won't be guaranteed
11330 * regatom will get called again, unless regbranch,
11331 * regpiece ... are also changed. If you do back off that
11332 * extra character, so that there is input guaranteed to
11333 * force calling regatom, you can't handle the case where
11334 * just the first character in the node is acceptable. I
11335 * (khw) decided to try this method which doesn't have that
11336 * pitfall; if performance issues are found, we can do a
11337 * combination of the current approach plus that one */
11343 } /* End of verifying node ends with an appropriate char */
11345 loopdone: /* Jumped to when encounters something that shouldn't be in
11348 /* I (khw) don't know if you can get here with zero length, but the
11349 * old code handled this situation by creating a zero-length EXACT
11350 * node. Might as well be NOTHING instead */
11356 /* If 'maybe_exact' is still set here, means there are no
11357 * code points in the node that participate in folds */
11358 if (FOLD && maybe_exact) {
11361 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
11364 RExC_parse = p - 1;
11365 Set_Node_Cur_Length(ret, parse_start);
11366 nextchar(pRExC_state);
11368 /* len is STRLEN which is unsigned, need to copy to signed */
11371 vFAIL("Internal disaster");
11374 } /* End of label 'defchar:' */
11376 } /* End of giant switch on input character */
11382 S_regwhite( RExC_state_t *pRExC_state, char *p )
11384 const char *e = RExC_end;
11386 PERL_ARGS_ASSERT_REGWHITE;
11391 else if (*p == '#') {
11394 if (*p++ == '\n') {
11400 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11409 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
11411 /* Returns the next non-pattern-white space, non-comment character (the
11412 * latter only if 'recognize_comment is true) in the string p, which is
11413 * ended by RExC_end. If there is no line break ending a comment,
11414 * RExC_seen has added the REG_SEEN_RUN_ON_COMMENT flag; */
11415 const char *e = RExC_end;
11417 PERL_ARGS_ASSERT_REGPATWS;
11421 if ((len = is_PATWS_safe(p, e, UTF))) {
11424 else if (recognize_comment && *p == '#') {
11428 if (is_LNBREAK_safe(p, e, UTF)) {
11434 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11442 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
11443 Character classes ([:foo:]) can also be negated ([:^foo:]).
11444 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
11445 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
11446 but trigger failures because they are currently unimplemented. */
11448 #define POSIXCC_DONE(c) ((c) == ':')
11449 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
11450 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
11452 PERL_STATIC_INLINE I32
11453 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
11456 I32 namedclass = OOB_NAMEDCLASS;
11458 PERL_ARGS_ASSERT_REGPPOSIXCC;
11460 if (value == '[' && RExC_parse + 1 < RExC_end &&
11461 /* I smell either [: or [= or [. -- POSIX has been here, right? */
11462 POSIXCC(UCHARAT(RExC_parse)))
11464 const char c = UCHARAT(RExC_parse);
11465 char* const s = RExC_parse++;
11467 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
11469 if (RExC_parse == RExC_end) {
11472 /* Try to give a better location for the error (than the end of
11473 * the string) by looking for the matching ']' */
11475 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
11478 vFAIL2("Unmatched '%c' in POSIX class", c);
11480 /* Grandfather lone [:, [=, [. */
11484 const char* const t = RExC_parse++; /* skip over the c */
11487 if (UCHARAT(RExC_parse) == ']') {
11488 const char *posixcc = s + 1;
11489 RExC_parse++; /* skip over the ending ] */
11492 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
11493 const I32 skip = t - posixcc;
11495 /* Initially switch on the length of the name. */
11498 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
11499 this is the Perl \w
11501 namedclass = ANYOF_WORDCHAR;
11504 /* Names all of length 5. */
11505 /* alnum alpha ascii blank cntrl digit graph lower
11506 print punct space upper */
11507 /* Offset 4 gives the best switch position. */
11508 switch (posixcc[4]) {
11510 if (memEQ(posixcc, "alph", 4)) /* alpha */
11511 namedclass = ANYOF_ALPHA;
11514 if (memEQ(posixcc, "spac", 4)) /* space */
11515 namedclass = ANYOF_PSXSPC;
11518 if (memEQ(posixcc, "grap", 4)) /* graph */
11519 namedclass = ANYOF_GRAPH;
11522 if (memEQ(posixcc, "asci", 4)) /* ascii */
11523 namedclass = ANYOF_ASCII;
11526 if (memEQ(posixcc, "blan", 4)) /* blank */
11527 namedclass = ANYOF_BLANK;
11530 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
11531 namedclass = ANYOF_CNTRL;
11534 if (memEQ(posixcc, "alnu", 4)) /* alnum */
11535 namedclass = ANYOF_ALPHANUMERIC;
11538 if (memEQ(posixcc, "lowe", 4)) /* lower */
11539 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
11540 else if (memEQ(posixcc, "uppe", 4)) /* upper */
11541 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
11544 if (memEQ(posixcc, "digi", 4)) /* digit */
11545 namedclass = ANYOF_DIGIT;
11546 else if (memEQ(posixcc, "prin", 4)) /* print */
11547 namedclass = ANYOF_PRINT;
11548 else if (memEQ(posixcc, "punc", 4)) /* punct */
11549 namedclass = ANYOF_PUNCT;
11554 if (memEQ(posixcc, "xdigit", 6))
11555 namedclass = ANYOF_XDIGIT;
11559 if (namedclass == OOB_NAMEDCLASS)
11560 Simple_vFAIL3("POSIX class [:%.*s:] unknown",
11563 /* The #defines are structured so each complement is +1 to
11564 * the normal one */
11568 assert (posixcc[skip] == ':');
11569 assert (posixcc[skip+1] == ']');
11570 } else if (!SIZE_ONLY) {
11571 /* [[=foo=]] and [[.foo.]] are still future. */
11573 /* adjust RExC_parse so the warning shows after
11574 the class closes */
11575 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
11577 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11580 /* Maternal grandfather:
11581 * "[:" ending in ":" but not in ":]" */
11583 vFAIL("Unmatched '[' in POSIX class");
11586 /* Grandfather lone [:, [=, [. */
11596 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
11598 /* This applies some heuristics at the current parse position (which should
11599 * be at a '[') to see if what follows might be intended to be a [:posix:]
11600 * class. It returns true if it really is a posix class, of course, but it
11601 * also can return true if it thinks that what was intended was a posix
11602 * class that didn't quite make it.
11604 * It will return true for
11606 * [:alphanumerics] (as long as the ] isn't followed immediately by a
11607 * ')' indicating the end of the (?[
11608 * [:any garbage including %^&$ punctuation:]
11610 * This is designed to be called only from S_handle_regex_sets; it could be
11611 * easily adapted to be called from the spot at the beginning of regclass()
11612 * that checks to see in a normal bracketed class if the surrounding []
11613 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
11614 * change long-standing behavior, so I (khw) didn't do that */
11615 char* p = RExC_parse + 1;
11616 char first_char = *p;
11618 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
11620 assert(*(p - 1) == '[');
11622 if (! POSIXCC(first_char)) {
11627 while (p < RExC_end && isWORDCHAR(*p)) p++;
11629 if (p >= RExC_end) {
11633 if (p - RExC_parse > 2 /* Got at least 1 word character */
11634 && (*p == first_char
11635 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
11640 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
11643 && p - RExC_parse > 2 /* [:] evaluates to colon;
11644 [::] is a bad posix class. */
11645 && first_char == *(p - 1));
11649 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist, I32 *flagp, U32 depth,
11650 char * const oregcomp_parse)
11652 /* Handle the (?[...]) construct to do set operations */
11655 UV start, end; /* End points of code point ranges */
11657 char *save_end, *save_parse;
11662 const bool save_fold = FOLD;
11664 GET_RE_DEBUG_FLAGS_DECL;
11666 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
11669 vFAIL("(?[...]) not valid in locale");
11671 RExC_uni_semantics = 1;
11673 /* This will return only an ANYOF regnode, or (unlikely) something smaller
11674 * (such as EXACT). Thus we can skip most everything if just sizing. We
11675 * call regclass to handle '[]' so as to not have to reinvent its parsing
11676 * rules here (throwing away the size it computes each time). And, we exit
11677 * upon an unescaped ']' that isn't one ending a regclass. To do both
11678 * these things, we need to realize that something preceded by a backslash
11679 * is escaped, so we have to keep track of backslashes */
11681 UV depth = 0; /* how many nested (?[...]) constructs */
11683 Perl_ck_warner_d(aTHX_
11684 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
11685 "The regex_sets feature is experimental" REPORT_LOCATION,
11686 (int) (RExC_parse - RExC_precomp) , RExC_precomp, RExC_parse);
11688 while (RExC_parse < RExC_end) {
11689 SV* current = NULL;
11690 RExC_parse = regpatws(pRExC_state, RExC_parse,
11691 TRUE); /* means recognize comments */
11692 switch (*RExC_parse) {
11694 if (RExC_parse[1] == '[') depth++, RExC_parse++;
11699 /* Skip the next byte (which could cause us to end up in
11700 * the middle of a UTF-8 character, but since none of those
11701 * are confusable with anything we currently handle in this
11702 * switch (invariants all), it's safe. We'll just hit the
11703 * default: case next time and keep on incrementing until
11704 * we find one of the invariants we do handle. */
11709 /* If this looks like it is a [:posix:] class, leave the
11710 * parse pointer at the '[' to fool regclass() into
11711 * thinking it is part of a '[[:posix:]]'. That function
11712 * will use strict checking to force a syntax error if it
11713 * doesn't work out to a legitimate class */
11714 bool is_posix_class
11715 = could_it_be_a_POSIX_class(pRExC_state);
11716 if (! is_posix_class) {
11720 /* regclass() can only return RESTART_UTF8 if multi-char
11721 folds are allowed. */
11722 if (!regclass(pRExC_state, flagp,depth+1,
11723 is_posix_class, /* parse the whole char
11724 class only if not a
11726 FALSE, /* don't allow multi-char folds */
11727 TRUE, /* silence non-portable warnings. */
11729 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11732 /* function call leaves parse pointing to the ']', except
11733 * if we faked it */
11734 if (is_posix_class) {
11738 SvREFCNT_dec(current); /* In case it returned something */
11743 if (depth--) break;
11745 if (RExC_parse < RExC_end
11746 && *RExC_parse == ')')
11748 node = reganode(pRExC_state, ANYOF, 0);
11749 RExC_size += ANYOF_SKIP;
11750 nextchar(pRExC_state);
11751 Set_Node_Length(node,
11752 RExC_parse - oregcomp_parse + 1); /* MJD */
11761 FAIL("Syntax error in (?[...])");
11764 /* Pass 2 only after this. Everything in this construct is a
11765 * metacharacter. Operands begin with either a '\' (for an escape
11766 * sequence), or a '[' for a bracketed character class. Any other
11767 * character should be an operator, or parenthesis for grouping. Both
11768 * types of operands are handled by calling regclass() to parse them. It
11769 * is called with a parameter to indicate to return the computed inversion
11770 * list. The parsing here is implemented via a stack. Each entry on the
11771 * stack is a single character representing one of the operators, or the
11772 * '('; or else a pointer to an operand inversion list. */
11774 #define IS_OPERAND(a) (! SvIOK(a))
11776 /* The stack starts empty. It is a syntax error if the first thing parsed
11777 * is a binary operator; everything else is pushed on the stack. When an
11778 * operand is parsed, the top of the stack is examined. If it is a binary
11779 * operator, the item before it should be an operand, and both are replaced
11780 * by the result of doing that operation on the new operand and the one on
11781 * the stack. Thus a sequence of binary operands is reduced to a single
11782 * one before the next one is parsed.
11784 * A unary operator may immediately follow a binary in the input, for
11787 * When an operand is parsed and the top of the stack is a unary operator,
11788 * the operation is performed, and then the stack is rechecked to see if
11789 * this new operand is part of a binary operation; if so, it is handled as
11792 * A '(' is simply pushed on the stack; it is valid only if the stack is
11793 * empty, or the top element of the stack is an operator or another '('
11794 * (for which the parenthesized expression will become an operand). By the
11795 * time the corresponding ')' is parsed everything in between should have
11796 * been parsed and evaluated to a single operand (or else is a syntax
11797 * error), and is handled as a regular operand */
11799 sv_2mortal((SV *)(stack = newAV()));
11801 while (RExC_parse < RExC_end) {
11802 I32 top_index = av_tindex(stack);
11804 SV* current = NULL;
11806 /* Skip white space */
11807 RExC_parse = regpatws(pRExC_state, RExC_parse,
11808 TRUE); /* means recognize comments */
11809 if (RExC_parse >= RExC_end) {
11810 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
11812 if ((curchar = UCHARAT(RExC_parse)) == ']') {
11819 if (av_tindex(stack) >= 0 /* This makes sure that we can
11820 safely subtract 1 from
11821 RExC_parse in the next clause.
11822 If we have something on the
11823 stack, we have parsed something
11825 && UCHARAT(RExC_parse - 1) == '('
11826 && RExC_parse < RExC_end)
11828 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
11829 * This happens when we have some thing like
11831 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
11833 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
11835 * Here we would be handling the interpolated
11836 * '$thai_or_lao'. We handle this by a recursive call to
11837 * ourselves which returns the inversion list the
11838 * interpolated expression evaluates to. We use the flags
11839 * from the interpolated pattern. */
11840 U32 save_flags = RExC_flags;
11841 const char * const save_parse = ++RExC_parse;
11843 parse_lparen_question_flags(pRExC_state);
11845 if (RExC_parse == save_parse /* Makes sure there was at
11846 least one flag (or this
11847 embedding wasn't compiled)
11849 || RExC_parse >= RExC_end - 4
11850 || UCHARAT(RExC_parse) != ':'
11851 || UCHARAT(++RExC_parse) != '('
11852 || UCHARAT(++RExC_parse) != '?'
11853 || UCHARAT(++RExC_parse) != '[')
11856 /* In combination with the above, this moves the
11857 * pointer to the point just after the first erroneous
11858 * character (or if there are no flags, to where they
11859 * should have been) */
11860 if (RExC_parse >= RExC_end - 4) {
11861 RExC_parse = RExC_end;
11863 else if (RExC_parse != save_parse) {
11864 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11866 vFAIL("Expecting '(?flags:(?[...'");
11869 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
11870 depth+1, oregcomp_parse);
11872 /* Here, 'current' contains the embedded expression's
11873 * inversion list, and RExC_parse points to the trailing
11874 * ']'; the next character should be the ')' which will be
11875 * paired with the '(' that has been put on the stack, so
11876 * the whole embedded expression reduces to '(operand)' */
11879 RExC_flags = save_flags;
11880 goto handle_operand;
11885 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11886 vFAIL("Unexpected character");
11889 /* regclass() can only return RESTART_UTF8 if multi-char
11890 folds are allowed. */
11891 if (!regclass(pRExC_state, flagp,depth+1,
11892 TRUE, /* means parse just the next thing */
11893 FALSE, /* don't allow multi-char folds */
11894 FALSE, /* don't silence non-portable warnings. */
11896 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11898 /* regclass() will return with parsing just the \ sequence,
11899 * leaving the parse pointer at the next thing to parse */
11901 goto handle_operand;
11903 case '[': /* Is a bracketed character class */
11905 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
11907 if (! is_posix_class) {
11911 /* regclass() can only return RESTART_UTF8 if multi-char
11912 folds are allowed. */
11913 if(!regclass(pRExC_state, flagp,depth+1,
11914 is_posix_class, /* parse the whole char class
11915 only if not a posix class */
11916 FALSE, /* don't allow multi-char folds */
11917 FALSE, /* don't silence non-portable warnings. */
11919 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11921 /* function call leaves parse pointing to the ']', except if we
11923 if (is_posix_class) {
11927 goto handle_operand;
11936 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
11937 || ! IS_OPERAND(*top_ptr))
11940 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
11942 av_push(stack, newSVuv(curchar));
11946 av_push(stack, newSVuv(curchar));
11950 if (top_index >= 0) {
11951 top_ptr = av_fetch(stack, top_index, FALSE);
11953 if (IS_OPERAND(*top_ptr)) {
11955 vFAIL("Unexpected '(' with no preceding operator");
11958 av_push(stack, newSVuv(curchar));
11965 || ! (current = av_pop(stack))
11966 || ! IS_OPERAND(current)
11967 || ! (lparen = av_pop(stack))
11968 || IS_OPERAND(lparen)
11969 || SvUV(lparen) != '(')
11971 SvREFCNT_dec(current);
11973 vFAIL("Unexpected ')'");
11976 SvREFCNT_dec_NN(lparen);
11983 /* Here, we have an operand to process, in 'current' */
11985 if (top_index < 0) { /* Just push if stack is empty */
11986 av_push(stack, current);
11989 SV* top = av_pop(stack);
11991 char current_operator;
11993 if (IS_OPERAND(top)) {
11994 SvREFCNT_dec_NN(top);
11995 SvREFCNT_dec_NN(current);
11996 vFAIL("Operand with no preceding operator");
11998 current_operator = (char) SvUV(top);
11999 switch (current_operator) {
12000 case '(': /* Push the '(' back on followed by the new
12002 av_push(stack, top);
12003 av_push(stack, current);
12004 SvREFCNT_inc(top); /* Counters the '_dec' done
12005 just after the 'break', so
12006 it doesn't get wrongly freed
12011 _invlist_invert(current);
12013 /* Unlike binary operators, the top of the stack,
12014 * now that this unary one has been popped off, may
12015 * legally be an operator, and we now have operand
12018 SvREFCNT_dec_NN(top);
12019 goto handle_operand;
12022 prev = av_pop(stack);
12023 _invlist_intersection(prev,
12026 av_push(stack, current);
12031 prev = av_pop(stack);
12032 _invlist_union(prev, current, ¤t);
12033 av_push(stack, current);
12037 prev = av_pop(stack);;
12038 _invlist_subtract(prev, current, ¤t);
12039 av_push(stack, current);
12042 case '^': /* The union minus the intersection */
12048 prev = av_pop(stack);
12049 _invlist_union(prev, current, &u);
12050 _invlist_intersection(prev, current, &i);
12051 /* _invlist_subtract will overwrite current
12052 without freeing what it already contains */
12054 _invlist_subtract(u, i, ¤t);
12055 av_push(stack, current);
12056 SvREFCNT_dec_NN(i);
12057 SvREFCNT_dec_NN(u);
12058 SvREFCNT_dec_NN(element);
12063 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
12065 SvREFCNT_dec_NN(top);
12066 SvREFCNT_dec(prev);
12070 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12073 if (av_tindex(stack) < 0 /* Was empty */
12074 || ((final = av_pop(stack)) == NULL)
12075 || ! IS_OPERAND(final)
12076 || av_tindex(stack) >= 0) /* More left on stack */
12078 vFAIL("Incomplete expression within '(?[ ])'");
12081 /* Here, 'final' is the resultant inversion list from evaluating the
12082 * expression. Return it if so requested */
12083 if (return_invlist) {
12084 *return_invlist = final;
12088 /* Otherwise generate a resultant node, based on 'final'. regclass() is
12089 * expecting a string of ranges and individual code points */
12090 invlist_iterinit(final);
12091 result_string = newSVpvs("");
12092 while (invlist_iternext(final, &start, &end)) {
12093 if (start == end) {
12094 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
12097 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
12102 save_parse = RExC_parse;
12103 RExC_parse = SvPV(result_string, len);
12104 save_end = RExC_end;
12105 RExC_end = RExC_parse + len;
12107 /* We turn off folding around the call, as the class we have constructed
12108 * already has all folding taken into consideration, and we don't want
12109 * regclass() to add to that */
12110 RExC_flags &= ~RXf_PMf_FOLD;
12111 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
12113 node = regclass(pRExC_state, flagp,depth+1,
12114 FALSE, /* means parse the whole char class */
12115 FALSE, /* don't allow multi-char folds */
12116 TRUE, /* silence non-portable warnings. The above may very
12117 well have generated non-portable code points, but
12118 they're valid on this machine */
12121 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
12124 RExC_flags |= RXf_PMf_FOLD;
12126 RExC_parse = save_parse + 1;
12127 RExC_end = save_end;
12128 SvREFCNT_dec_NN(final);
12129 SvREFCNT_dec_NN(result_string);
12131 nextchar(pRExC_state);
12132 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
12137 /* The names of properties whose definitions are not known at compile time are
12138 * stored in this SV, after a constant heading. So if the length has been
12139 * changed since initialization, then there is a run-time definition. */
12140 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION (SvCUR(listsv) != initial_listsv_len)
12143 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
12144 const bool stop_at_1, /* Just parse the next thing, don't
12145 look for a full character class */
12146 bool allow_multi_folds,
12147 const bool silence_non_portable, /* Don't output warnings
12150 SV** ret_invlist) /* Return an inversion list, not a node */
12152 /* parse a bracketed class specification. Most of these will produce an
12153 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
12154 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
12155 * under /i with multi-character folds: it will be rewritten following the
12156 * paradigm of this example, where the <multi-fold>s are characters which
12157 * fold to multiple character sequences:
12158 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
12159 * gets effectively rewritten as:
12160 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
12161 * reg() gets called (recursively) on the rewritten version, and this
12162 * function will return what it constructs. (Actually the <multi-fold>s
12163 * aren't physically removed from the [abcdefghi], it's just that they are
12164 * ignored in the recursion by means of a flag:
12165 * <RExC_in_multi_char_class>.)
12167 * ANYOF nodes contain a bit map for the first 256 characters, with the
12168 * corresponding bit set if that character is in the list. For characters
12169 * above 255, a range list or swash is used. There are extra bits for \w,
12170 * etc. in locale ANYOFs, as what these match is not determinable at
12173 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
12174 * to be restarted. This can only happen if ret_invlist is non-NULL.
12178 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
12180 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
12183 IV namedclass = OOB_NAMEDCLASS;
12184 char *rangebegin = NULL;
12185 bool need_class = 0;
12187 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
12188 than just initialized. */
12189 SV* properties = NULL; /* Code points that match \p{} \P{} */
12190 SV* posixes = NULL; /* Code points that match classes like, [:word:],
12191 extended beyond the Latin1 range */
12192 UV element_count = 0; /* Number of distinct elements in the class.
12193 Optimizations may be possible if this is tiny */
12194 AV * multi_char_matches = NULL; /* Code points that fold to more than one
12195 character; used under /i */
12197 char * stop_ptr = RExC_end; /* where to stop parsing */
12198 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
12200 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
12202 /* Unicode properties are stored in a swash; this holds the current one
12203 * being parsed. If this swash is the only above-latin1 component of the
12204 * character class, an optimization is to pass it directly on to the
12205 * execution engine. Otherwise, it is set to NULL to indicate that there
12206 * are other things in the class that have to be dealt with at execution
12208 SV* swash = NULL; /* Code points that match \p{} \P{} */
12210 /* Set if a component of this character class is user-defined; just passed
12211 * on to the engine */
12212 bool has_user_defined_property = FALSE;
12214 /* inversion list of code points this node matches only when the target
12215 * string is in UTF-8. (Because is under /d) */
12216 SV* depends_list = NULL;
12218 /* inversion list of code points this node matches. For much of the
12219 * function, it includes only those that match regardless of the utf8ness
12220 * of the target string */
12221 SV* cp_list = NULL;
12224 /* In a range, counts how many 0-2 of the ends of it came from literals,
12225 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
12226 UV literal_endpoint = 0;
12228 bool invert = FALSE; /* Is this class to be complemented */
12230 /* Is there any thing like \W or [:^digit:] that matches above the legal
12231 * Unicode range? */
12232 bool runtime_posix_matches_above_Unicode = FALSE;
12234 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
12235 case we need to change the emitted regop to an EXACT. */
12236 const char * orig_parse = RExC_parse;
12237 const I32 orig_size = RExC_size;
12238 GET_RE_DEBUG_FLAGS_DECL;
12240 PERL_ARGS_ASSERT_REGCLASS;
12242 PERL_UNUSED_ARG(depth);
12245 DEBUG_PARSE("clas");
12247 /* Assume we are going to generate an ANYOF node. */
12248 ret = reganode(pRExC_state, ANYOF, 0);
12251 RExC_size += ANYOF_SKIP;
12252 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
12255 ANYOF_FLAGS(ret) = 0;
12257 RExC_emit += ANYOF_SKIP;
12259 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
12261 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
12262 initial_listsv_len = SvCUR(listsv);
12263 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
12267 RExC_parse = regpatws(pRExC_state, RExC_parse,
12268 FALSE /* means don't recognize comments */);
12271 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
12274 allow_multi_folds = FALSE;
12277 RExC_parse = regpatws(pRExC_state, RExC_parse,
12278 FALSE /* means don't recognize comments */);
12282 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
12283 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
12284 const char *s = RExC_parse;
12285 const char c = *s++;
12287 while (isWORDCHAR(*s))
12289 if (*s && c == *s && s[1] == ']') {
12290 SAVEFREESV(RExC_rx_sv);
12292 "POSIX syntax [%c %c] belongs inside character classes",
12294 (void)ReREFCNT_inc(RExC_rx_sv);
12298 /* If the caller wants us to just parse a single element, accomplish this
12299 * by faking the loop ending condition */
12300 if (stop_at_1 && RExC_end > RExC_parse) {
12301 stop_ptr = RExC_parse + 1;
12304 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
12305 if (UCHARAT(RExC_parse) == ']')
12306 goto charclassloop;
12310 if (RExC_parse >= stop_ptr) {
12315 RExC_parse = regpatws(pRExC_state, RExC_parse,
12316 FALSE /* means don't recognize comments */);
12319 if (UCHARAT(RExC_parse) == ']') {
12325 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
12326 save_value = value;
12327 save_prevvalue = prevvalue;
12330 rangebegin = RExC_parse;
12334 value = utf8n_to_uvchr((U8*)RExC_parse,
12335 RExC_end - RExC_parse,
12336 &numlen, UTF8_ALLOW_DEFAULT);
12337 RExC_parse += numlen;
12340 value = UCHARAT(RExC_parse++);
12343 && RExC_parse < RExC_end
12344 && POSIXCC(UCHARAT(RExC_parse)))
12346 namedclass = regpposixcc(pRExC_state, value, strict);
12348 else if (value == '\\') {
12350 value = utf8n_to_uvchr((U8*)RExC_parse,
12351 RExC_end - RExC_parse,
12352 &numlen, UTF8_ALLOW_DEFAULT);
12353 RExC_parse += numlen;
12356 value = UCHARAT(RExC_parse++);
12358 /* Some compilers cannot handle switching on 64-bit integer
12359 * values, therefore value cannot be an UV. Yes, this will
12360 * be a problem later if we want switch on Unicode.
12361 * A similar issue a little bit later when switching on
12362 * namedclass. --jhi */
12364 /* If the \ is escaping white space when white space is being
12365 * skipped, it means that that white space is wanted literally, and
12366 * is already in 'value'. Otherwise, need to translate the escape
12367 * into what it signifies. */
12368 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
12370 case 'w': namedclass = ANYOF_WORDCHAR; break;
12371 case 'W': namedclass = ANYOF_NWORDCHAR; break;
12372 case 's': namedclass = ANYOF_SPACE; break;
12373 case 'S': namedclass = ANYOF_NSPACE; break;
12374 case 'd': namedclass = ANYOF_DIGIT; break;
12375 case 'D': namedclass = ANYOF_NDIGIT; break;
12376 case 'v': namedclass = ANYOF_VERTWS; break;
12377 case 'V': namedclass = ANYOF_NVERTWS; break;
12378 case 'h': namedclass = ANYOF_HORIZWS; break;
12379 case 'H': namedclass = ANYOF_NHORIZWS; break;
12380 case 'N': /* Handle \N{NAME} in class */
12382 /* We only pay attention to the first char of
12383 multichar strings being returned. I kinda wonder
12384 if this makes sense as it does change the behaviour
12385 from earlier versions, OTOH that behaviour was broken
12387 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
12388 TRUE, /* => charclass */
12391 if (*flagp & RESTART_UTF8)
12392 FAIL("panic: grok_bslash_N set RESTART_UTF8");
12402 /* We will handle any undefined properties ourselves */
12403 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF;
12405 if (RExC_parse >= RExC_end)
12406 vFAIL2("Empty \\%c{}", (U8)value);
12407 if (*RExC_parse == '{') {
12408 const U8 c = (U8)value;
12409 e = strchr(RExC_parse++, '}');
12411 vFAIL2("Missing right brace on \\%c{}", c);
12412 while (isSPACE(UCHARAT(RExC_parse)))
12414 if (e == RExC_parse)
12415 vFAIL2("Empty \\%c{}", c);
12416 n = e - RExC_parse;
12417 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
12428 if (UCHARAT(RExC_parse) == '^') {
12431 /* toggle. (The rhs xor gets the single bit that
12432 * differs between P and p; the other xor inverts just
12434 value ^= 'P' ^ 'p';
12436 while (isSPACE(UCHARAT(RExC_parse))) {
12441 /* Try to get the definition of the property into
12442 * <invlist>. If /i is in effect, the effective property
12443 * will have its name be <__NAME_i>. The design is
12444 * discussed in commit
12445 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
12446 Newx(name, n + sizeof("_i__\n"), char);
12448 sprintf(name, "%s%.*s%s\n",
12449 (FOLD) ? "__" : "",
12455 /* Look up the property name, and get its swash and
12456 * inversion list, if the property is found */
12458 SvREFCNT_dec_NN(swash);
12460 swash = _core_swash_init("utf8", name, &PL_sv_undef,
12463 NULL, /* No inversion list */
12466 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
12468 SvREFCNT_dec_NN(swash);
12472 /* Here didn't find it. It could be a user-defined
12473 * property that will be available at run-time. If we
12474 * accept only compile-time properties, is an error;
12475 * otherwise add it to the list for run-time look up */
12477 RExC_parse = e + 1;
12478 vFAIL3("Property '%.*s' is unknown", (int) n, name);
12480 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s\n",
12481 (value == 'p' ? '+' : '!'),
12483 has_user_defined_property = TRUE;
12485 /* We don't know yet, so have to assume that the
12486 * property could match something in the Latin1 range,
12487 * hence something that isn't utf8. Note that this
12488 * would cause things in <depends_list> to match
12489 * inappropriately, except that any \p{}, including
12490 * this one forces Unicode semantics, which means there
12491 * is <no depends_list> */
12492 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
12496 /* Here, did get the swash and its inversion list. If
12497 * the swash is from a user-defined property, then this
12498 * whole character class should be regarded as such */
12499 has_user_defined_property =
12501 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY);
12503 /* Invert if asking for the complement */
12504 if (value == 'P') {
12505 _invlist_union_complement_2nd(properties,
12509 /* The swash can't be used as-is, because we've
12510 * inverted things; delay removing it to here after
12511 * have copied its invlist above */
12512 SvREFCNT_dec_NN(swash);
12516 _invlist_union(properties, invlist, &properties);
12521 RExC_parse = e + 1;
12522 namedclass = ANYOF_UNIPROP; /* no official name, but it's
12525 /* \p means they want Unicode semantics */
12526 RExC_uni_semantics = 1;
12529 case 'n': value = '\n'; break;
12530 case 'r': value = '\r'; break;
12531 case 't': value = '\t'; break;
12532 case 'f': value = '\f'; break;
12533 case 'b': value = '\b'; break;
12534 case 'e': value = ASCII_TO_NATIVE('\033');break;
12535 case 'a': value = ASCII_TO_NATIVE('\007');break;
12537 RExC_parse--; /* function expects to be pointed at the 'o' */
12539 const char* error_msg;
12540 bool valid = grok_bslash_o(&RExC_parse,
12543 SIZE_ONLY, /* warnings in pass
12546 silence_non_portable,
12552 if (PL_encoding && value < 0x100) {
12553 goto recode_encoding;
12557 RExC_parse--; /* function expects to be pointed at the 'x' */
12559 const char* error_msg;
12560 bool valid = grok_bslash_x(&RExC_parse,
12563 TRUE, /* Output warnings */
12565 silence_non_portable,
12571 if (PL_encoding && value < 0x100)
12572 goto recode_encoding;
12575 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
12577 case '0': case '1': case '2': case '3': case '4':
12578 case '5': case '6': case '7':
12580 /* Take 1-3 octal digits */
12581 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12582 numlen = (strict) ? 4 : 3;
12583 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
12584 RExC_parse += numlen;
12587 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12588 vFAIL("Need exactly 3 octal digits");
12590 else if (! SIZE_ONLY /* like \08, \178 */
12592 && RExC_parse < RExC_end
12593 && isDIGIT(*RExC_parse)
12594 && ckWARN(WARN_REGEXP))
12596 SAVEFREESV(RExC_rx_sv);
12597 reg_warn_non_literal_string(
12599 form_short_octal_warning(RExC_parse, numlen));
12600 (void)ReREFCNT_inc(RExC_rx_sv);
12603 if (PL_encoding && value < 0x100)
12604 goto recode_encoding;
12608 if (! RExC_override_recoding) {
12609 SV* enc = PL_encoding;
12610 value = reg_recode((const char)(U8)value, &enc);
12613 vFAIL("Invalid escape in the specified encoding");
12615 else if (SIZE_ONLY) {
12616 ckWARNreg(RExC_parse,
12617 "Invalid escape in the specified encoding");
12623 /* Allow \_ to not give an error */
12624 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
12626 vFAIL2("Unrecognized escape \\%c in character class",
12630 SAVEFREESV(RExC_rx_sv);
12631 ckWARN2reg(RExC_parse,
12632 "Unrecognized escape \\%c in character class passed through",
12634 (void)ReREFCNT_inc(RExC_rx_sv);
12638 } /* End of switch on char following backslash */
12639 } /* end of handling backslash escape sequences */
12642 literal_endpoint++;
12645 /* Here, we have the current token in 'value' */
12647 /* What matches in a locale is not known until runtime. This includes
12648 * what the Posix classes (like \w, [:space:]) match. Room must be
12649 * reserved (one time per class) to store such classes, either if Perl
12650 * is compiled so that locale nodes always should have this space, or
12651 * if there is such class info to be stored. The space will contain a
12652 * bit for each named class that is to be matched against. This isn't
12653 * needed for \p{} and pseudo-classes, as they are not affected by
12654 * locale, and hence are dealt with separately */
12657 && (ANYOF_LOCALE == ANYOF_CLASS
12658 || (namedclass > OOB_NAMEDCLASS && namedclass < ANYOF_MAX)))
12662 RExC_size += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12665 RExC_emit += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12666 ANYOF_CLASS_ZERO(ret);
12668 ANYOF_FLAGS(ret) |= ANYOF_CLASS;
12671 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
12673 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
12674 * literal, as is the character that began the false range, i.e.
12675 * the 'a' in the examples */
12678 const int w = (RExC_parse >= rangebegin)
12679 ? RExC_parse - rangebegin
12682 vFAIL4("False [] range \"%*.*s\"", w, w, rangebegin);
12685 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
12686 ckWARN4reg(RExC_parse,
12687 "False [] range \"%*.*s\"",
12689 (void)ReREFCNT_inc(RExC_rx_sv);
12690 cp_list = add_cp_to_invlist(cp_list, '-');
12691 cp_list = add_cp_to_invlist(cp_list, prevvalue);
12695 range = 0; /* this was not a true range */
12696 element_count += 2; /* So counts for three values */
12700 U8 classnum = namedclass_to_classnum(namedclass);
12701 if (namedclass >= ANYOF_MAX) { /* If a special class */
12702 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
12704 /* Here, should be \h, \H, \v, or \V. Neither /d nor
12705 * /l make a difference in what these match. There
12706 * would be problems if these characters had folds
12707 * other than themselves, as cp_list is subject to
12709 if (classnum != _CC_VERTSPACE) {
12710 assert( namedclass == ANYOF_HORIZWS
12711 || namedclass == ANYOF_NHORIZWS);
12713 /* It turns out that \h is just a synonym for
12715 classnum = _CC_BLANK;
12718 _invlist_union_maybe_complement_2nd(
12720 PL_XPosix_ptrs[classnum],
12721 cBOOL(namedclass % 2), /* Complement if odd
12722 (NHORIZWS, NVERTWS)
12727 else if (classnum == _CC_ASCII) {
12730 ANYOF_CLASS_SET(ret, namedclass);
12733 #endif /* Not isascii(); just use the hard-coded definition for it */
12734 _invlist_union_maybe_complement_2nd(
12737 cBOOL(namedclass % 2), /* Complement if odd
12741 else { /* Garden variety class */
12743 /* The ascii range inversion list */
12744 SV* ascii_source = PL_Posix_ptrs[classnum];
12746 /* The full Latin1 range inversion list */
12747 SV* l1_source = PL_L1Posix_ptrs[classnum];
12749 /* This code is structured into two major clauses. The
12750 * first is for classes whose complete definitions may not
12751 * already be known. It not, the Latin1 definition
12752 * (guaranteed to already known) is used plus code is
12753 * generated to load the rest at run-time (only if needed).
12754 * If the complete definition is known, it drops down to
12755 * the second clause, where the complete definition is
12758 if (classnum < _FIRST_NON_SWASH_CC) {
12760 /* Here, the class has a swash, which may or not
12761 * already be loaded */
12763 /* The name of the property to use to match the full
12764 * eXtended Unicode range swash for this character
12766 const char *Xname = swash_property_names[classnum];
12768 /* If returning the inversion list, we can't defer
12769 * getting this until runtime */
12770 if (ret_invlist && ! PL_utf8_swash_ptrs[classnum]) {
12771 PL_utf8_swash_ptrs[classnum] =
12772 _core_swash_init("utf8", Xname, &PL_sv_undef,
12775 NULL, /* No inversion list */
12776 NULL /* No flags */
12778 assert(PL_utf8_swash_ptrs[classnum]);
12780 if ( ! PL_utf8_swash_ptrs[classnum]) {
12781 if (namedclass % 2 == 0) { /* A non-complemented
12783 /* If not /a matching, there are code points we
12784 * don't know at compile time. Arrange for the
12785 * unknown matches to be loaded at run-time, if
12787 if (! AT_LEAST_ASCII_RESTRICTED) {
12788 Perl_sv_catpvf(aTHX_ listsv, "+utf8::%s\n",
12791 if (LOC) { /* Under locale, set run-time
12793 ANYOF_CLASS_SET(ret, namedclass);
12796 /* Add the current class's code points to
12797 * the running total */
12798 _invlist_union(posixes,
12799 (AT_LEAST_ASCII_RESTRICTED)
12805 else { /* A complemented class */
12806 if (AT_LEAST_ASCII_RESTRICTED) {
12807 /* Under /a should match everything above
12808 * ASCII, plus the complement of the set's
12810 _invlist_union_complement_2nd(posixes,
12815 /* Arrange for the unknown matches to be
12816 * loaded at run-time, if needed */
12817 Perl_sv_catpvf(aTHX_ listsv, "!utf8::%s\n",
12819 runtime_posix_matches_above_Unicode = TRUE;
12821 ANYOF_CLASS_SET(ret, namedclass);
12825 /* We want to match everything in
12826 * Latin1, except those things that
12827 * l1_source matches */
12828 SV* scratch_list = NULL;
12829 _invlist_subtract(PL_Latin1, l1_source,
12832 /* Add the list from this class to the
12835 posixes = scratch_list;
12838 _invlist_union(posixes,
12841 SvREFCNT_dec_NN(scratch_list);
12843 if (DEPENDS_SEMANTICS) {
12845 |= ANYOF_NON_UTF8_LATIN1_ALL;
12850 goto namedclass_done;
12853 /* Here, there is a swash loaded for the class. If no
12854 * inversion list for it yet, get it */
12855 if (! PL_XPosix_ptrs[classnum]) {
12856 PL_XPosix_ptrs[classnum]
12857 = _swash_to_invlist(PL_utf8_swash_ptrs[classnum]);
12861 /* Here there is an inversion list already loaded for the
12864 if (namedclass % 2 == 0) { /* A non-complemented class,
12865 like ANYOF_PUNCT */
12867 /* For non-locale, just add it to any existing list
12869 _invlist_union(posixes,
12870 (AT_LEAST_ASCII_RESTRICTED)
12872 : PL_XPosix_ptrs[classnum],
12875 else { /* Locale */
12876 SV* scratch_list = NULL;
12878 /* For above Latin1 code points, we use the full
12880 _invlist_intersection(PL_AboveLatin1,
12881 PL_XPosix_ptrs[classnum],
12883 /* And set the output to it, adding instead if
12884 * there already is an output. Checking if
12885 * 'posixes' is NULL first saves an extra clone.
12886 * Its reference count will be decremented at the
12887 * next union, etc, or if this is the only
12888 * instance, at the end of the routine */
12890 posixes = scratch_list;
12893 _invlist_union(posixes, scratch_list, &posixes);
12894 SvREFCNT_dec_NN(scratch_list);
12897 #ifndef HAS_ISBLANK
12898 if (namedclass != ANYOF_BLANK) {
12900 /* Set this class in the node for runtime
12902 ANYOF_CLASS_SET(ret, namedclass);
12903 #ifndef HAS_ISBLANK
12906 /* No isblank(), use the hard-coded ASCII-range
12907 * blanks, adding them to the running total. */
12909 _invlist_union(posixes, ascii_source, &posixes);
12914 else { /* A complemented class, like ANYOF_NPUNCT */
12916 _invlist_union_complement_2nd(
12918 (AT_LEAST_ASCII_RESTRICTED)
12920 : PL_XPosix_ptrs[classnum],
12922 /* Under /d, everything in the upper half of the
12923 * Latin1 range matches this complement */
12924 if (DEPENDS_SEMANTICS) {
12925 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL;
12928 else { /* Locale */
12929 SV* scratch_list = NULL;
12930 _invlist_subtract(PL_AboveLatin1,
12931 PL_XPosix_ptrs[classnum],
12934 posixes = scratch_list;
12937 _invlist_union(posixes, scratch_list, &posixes);
12938 SvREFCNT_dec_NN(scratch_list);
12940 #ifndef HAS_ISBLANK
12941 if (namedclass != ANYOF_NBLANK) {
12943 ANYOF_CLASS_SET(ret, namedclass);
12944 #ifndef HAS_ISBLANK
12947 /* Get the list of all code points in Latin1
12948 * that are not ASCII blanks, and add them to
12949 * the running total */
12950 _invlist_subtract(PL_Latin1, ascii_source,
12952 _invlist_union(posixes, scratch_list, &posixes);
12953 SvREFCNT_dec_NN(scratch_list);
12960 continue; /* Go get next character */
12962 } /* end of namedclass \blah */
12964 /* Here, we have a single value. If 'range' is set, it is the ending
12965 * of a range--check its validity. Later, we will handle each
12966 * individual code point in the range. If 'range' isn't set, this
12967 * could be the beginning of a range, so check for that by looking
12968 * ahead to see if the next real character to be processed is the range
12969 * indicator--the minus sign */
12972 RExC_parse = regpatws(pRExC_state, RExC_parse,
12973 FALSE /* means don't recognize comments */);
12977 if (prevvalue > value) /* b-a */ {
12978 const int w = RExC_parse - rangebegin;
12979 Simple_vFAIL4("Invalid [] range \"%*.*s\"", w, w, rangebegin);
12980 range = 0; /* not a valid range */
12984 prevvalue = value; /* save the beginning of the potential range */
12985 if (! stop_at_1 /* Can't be a range if parsing just one thing */
12986 && *RExC_parse == '-')
12988 char* next_char_ptr = RExC_parse + 1;
12989 if (skip_white) { /* Get the next real char after the '-' */
12990 next_char_ptr = regpatws(pRExC_state,
12992 FALSE); /* means don't recognize
12996 /* If the '-' is at the end of the class (just before the ']',
12997 * it is a literal minus; otherwise it is a range */
12998 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
12999 RExC_parse = next_char_ptr;
13001 /* a bad range like \w-, [:word:]- ? */
13002 if (namedclass > OOB_NAMEDCLASS) {
13003 if (strict || ckWARN(WARN_REGEXP)) {
13005 RExC_parse >= rangebegin ?
13006 RExC_parse - rangebegin : 0;
13008 vFAIL4("False [] range \"%*.*s\"",
13013 "False [] range \"%*.*s\"",
13018 cp_list = add_cp_to_invlist(cp_list, '-');
13022 range = 1; /* yeah, it's a range! */
13023 continue; /* but do it the next time */
13028 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
13031 /* non-Latin1 code point implies unicode semantics. Must be set in
13032 * pass1 so is there for the whole of pass 2 */
13034 RExC_uni_semantics = 1;
13037 /* Ready to process either the single value, or the completed range.
13038 * For single-valued non-inverted ranges, we consider the possibility
13039 * of multi-char folds. (We made a conscious decision to not do this
13040 * for the other cases because it can often lead to non-intuitive
13041 * results. For example, you have the peculiar case that:
13042 * "s s" =~ /^[^\xDF]+$/i => Y
13043 * "ss" =~ /^[^\xDF]+$/i => N
13045 * See [perl #89750] */
13046 if (FOLD && allow_multi_folds && value == prevvalue) {
13047 if (value == LATIN_SMALL_LETTER_SHARP_S
13048 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
13051 /* Here <value> is indeed a multi-char fold. Get what it is */
13053 U8 foldbuf[UTF8_MAXBYTES_CASE];
13056 UV folded = _to_uni_fold_flags(
13061 | ((LOC) ? FOLD_FLAGS_LOCALE
13062 : (ASCII_FOLD_RESTRICTED)
13063 ? FOLD_FLAGS_NOMIX_ASCII
13067 /* Here, <folded> should be the first character of the
13068 * multi-char fold of <value>, with <foldbuf> containing the
13069 * whole thing. But, if this fold is not allowed (because of
13070 * the flags), <fold> will be the same as <value>, and should
13071 * be processed like any other character, so skip the special
13073 if (folded != value) {
13075 /* Skip if we are recursed, currently parsing the class
13076 * again. Otherwise add this character to the list of
13077 * multi-char folds. */
13078 if (! RExC_in_multi_char_class) {
13079 AV** this_array_ptr;
13081 STRLEN cp_count = utf8_length(foldbuf,
13082 foldbuf + foldlen);
13083 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
13085 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
13088 if (! multi_char_matches) {
13089 multi_char_matches = newAV();
13092 /* <multi_char_matches> is actually an array of arrays.
13093 * There will be one or two top-level elements: [2],
13094 * and/or [3]. The [2] element is an array, each
13095 * element thereof is a character which folds to TWO
13096 * characters; [3] is for folds to THREE characters.
13097 * (Unicode guarantees a maximum of 3 characters in any
13098 * fold.) When we rewrite the character class below,
13099 * we will do so such that the longest folds are
13100 * written first, so that it prefers the longest
13101 * matching strings first. This is done even if it
13102 * turns out that any quantifier is non-greedy, out of
13103 * programmer laziness. Tom Christiansen has agreed
13104 * that this is ok. This makes the test for the
13105 * ligature 'ffi' come before the test for 'ff' */
13106 if (av_exists(multi_char_matches, cp_count)) {
13107 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13109 this_array = *this_array_ptr;
13112 this_array = newAV();
13113 av_store(multi_char_matches, cp_count,
13116 av_push(this_array, multi_fold);
13119 /* This element should not be processed further in this
13122 value = save_value;
13123 prevvalue = save_prevvalue;
13129 /* Deal with this element of the class */
13132 cp_list = _add_range_to_invlist(cp_list, prevvalue, value);
13134 SV* this_range = _new_invlist(1);
13135 _append_range_to_invlist(this_range, prevvalue, value);
13137 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
13138 * If this range was specified using something like 'i-j', we want
13139 * to include only the 'i' and the 'j', and not anything in
13140 * between, so exclude non-ASCII, non-alphabetics from it.
13141 * However, if the range was specified with something like
13142 * [\x89-\x91] or [\x89-j], all code points within it should be
13143 * included. literal_endpoint==2 means both ends of the range used
13144 * a literal character, not \x{foo} */
13145 if (literal_endpoint == 2
13146 && (prevvalue >= 'a' && value <= 'z')
13147 || (prevvalue >= 'A' && value <= 'Z'))
13149 _invlist_intersection(this_range, PL_Posix_ptrs[_CC_ALPHA],
13152 _invlist_union(cp_list, this_range, &cp_list);
13153 literal_endpoint = 0;
13157 range = 0; /* this range (if it was one) is done now */
13158 } /* End of loop through all the text within the brackets */
13160 /* If anything in the class expands to more than one character, we have to
13161 * deal with them by building up a substitute parse string, and recursively
13162 * calling reg() on it, instead of proceeding */
13163 if (multi_char_matches) {
13164 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
13167 char *save_end = RExC_end;
13168 char *save_parse = RExC_parse;
13169 bool first_time = TRUE; /* First multi-char occurrence doesn't get
13174 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
13175 because too confusing */
13177 sv_catpv(substitute_parse, "(?:");
13181 /* Look at the longest folds first */
13182 for (cp_count = av_len(multi_char_matches); cp_count > 0; cp_count--) {
13184 if (av_exists(multi_char_matches, cp_count)) {
13185 AV** this_array_ptr;
13188 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13190 while ((this_sequence = av_pop(*this_array_ptr)) !=
13193 if (! first_time) {
13194 sv_catpv(substitute_parse, "|");
13196 first_time = FALSE;
13198 sv_catpv(substitute_parse, SvPVX(this_sequence));
13203 /* If the character class contains anything else besides these
13204 * multi-character folds, have to include it in recursive parsing */
13205 if (element_count) {
13206 sv_catpv(substitute_parse, "|[");
13207 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
13208 sv_catpv(substitute_parse, "]");
13211 sv_catpv(substitute_parse, ")");
13214 /* This is a way to get the parse to skip forward a whole named
13215 * sequence instead of matching the 2nd character when it fails the
13217 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
13221 RExC_parse = SvPV(substitute_parse, len);
13222 RExC_end = RExC_parse + len;
13223 RExC_in_multi_char_class = 1;
13224 RExC_emit = (regnode *)orig_emit;
13226 ret = reg(pRExC_state, 1, ®_flags, depth+1);
13228 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
13230 RExC_parse = save_parse;
13231 RExC_end = save_end;
13232 RExC_in_multi_char_class = 0;
13233 SvREFCNT_dec_NN(multi_char_matches);
13237 /* If the character class contains only a single element, it may be
13238 * optimizable into another node type which is smaller and runs faster.
13239 * Check if this is the case for this class */
13240 if (element_count == 1 && ! ret_invlist) {
13244 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like \w or
13245 [:digit:] or \p{foo} */
13247 /* All named classes are mapped into POSIXish nodes, with its FLAG
13248 * argument giving which class it is */
13249 switch ((I32)namedclass) {
13250 case ANYOF_UNIPROP:
13253 /* These don't depend on the charset modifiers. They always
13254 * match under /u rules */
13255 case ANYOF_NHORIZWS:
13256 case ANYOF_HORIZWS:
13257 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
13260 case ANYOF_NVERTWS:
13265 /* The actual POSIXish node for all the rest depends on the
13266 * charset modifier. The ones in the first set depend only on
13267 * ASCII or, if available on this platform, locale */
13271 op = (LOC) ? POSIXL : POSIXA;
13282 /* under /a could be alpha */
13284 if (ASCII_RESTRICTED) {
13285 namedclass = ANYOF_ALPHA + (namedclass % 2);
13293 /* The rest have more possibilities depending on the charset.
13294 * We take advantage of the enum ordering of the charset
13295 * modifiers to get the exact node type, */
13297 op = POSIXD + get_regex_charset(RExC_flags);
13298 if (op > POSIXA) { /* /aa is same as /a */
13301 #ifndef HAS_ISBLANK
13303 && (namedclass == ANYOF_BLANK
13304 || namedclass == ANYOF_NBLANK))
13311 /* The odd numbered ones are the complements of the
13312 * next-lower even number one */
13313 if (namedclass % 2 == 1) {
13317 arg = namedclass_to_classnum(namedclass);
13321 else if (value == prevvalue) {
13323 /* Here, the class consists of just a single code point */
13326 if (! LOC && value == '\n') {
13327 op = REG_ANY; /* Optimize [^\n] */
13328 *flagp |= HASWIDTH|SIMPLE;
13332 else if (value < 256 || UTF) {
13334 /* Optimize a single value into an EXACTish node, but not if it
13335 * would require converting the pattern to UTF-8. */
13336 op = compute_EXACTish(pRExC_state);
13338 } /* Otherwise is a range */
13339 else if (! LOC) { /* locale could vary these */
13340 if (prevvalue == '0') {
13341 if (value == '9') {
13348 /* Here, we have changed <op> away from its initial value iff we found
13349 * an optimization */
13352 /* Throw away this ANYOF regnode, and emit the calculated one,
13353 * which should correspond to the beginning, not current, state of
13355 const char * cur_parse = RExC_parse;
13356 RExC_parse = (char *)orig_parse;
13360 /* To get locale nodes to not use the full ANYOF size would
13361 * require moving the code above that writes the portions
13362 * of it that aren't in other nodes to after this point.
13363 * e.g. ANYOF_CLASS_SET */
13364 RExC_size = orig_size;
13368 RExC_emit = (regnode *)orig_emit;
13369 if (PL_regkind[op] == POSIXD) {
13371 op += NPOSIXD - POSIXD;
13376 ret = reg_node(pRExC_state, op);
13378 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
13382 *flagp |= HASWIDTH|SIMPLE;
13384 else if (PL_regkind[op] == EXACT) {
13385 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13388 RExC_parse = (char *) cur_parse;
13390 SvREFCNT_dec(posixes);
13391 SvREFCNT_dec(cp_list);
13398 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
13400 /* If folding, we calculate all characters that could fold to or from the
13401 * ones already on the list */
13402 if (FOLD && cp_list) {
13403 UV start, end; /* End points of code point ranges */
13405 SV* fold_intersection = NULL;
13407 /* If the highest code point is within Latin1, we can use the
13408 * compiled-in Alphas list, and not have to go out to disk. This
13409 * yields two false positives, the masculine and feminine ordinal
13410 * indicators, which are weeded out below using the
13411 * IS_IN_SOME_FOLD_L1() macro */
13412 if (invlist_highest(cp_list) < 256) {
13413 _invlist_intersection(PL_L1Posix_ptrs[_CC_ALPHA], cp_list,
13414 &fold_intersection);
13418 /* Here, there are non-Latin1 code points, so we will have to go
13419 * fetch the list of all the characters that participate in folds
13421 if (! PL_utf8_foldable) {
13422 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13423 &PL_sv_undef, 1, 0);
13424 PL_utf8_foldable = _get_swash_invlist(swash);
13425 SvREFCNT_dec_NN(swash);
13428 /* This is a hash that for a particular fold gives all characters
13429 * that are involved in it */
13430 if (! PL_utf8_foldclosures) {
13432 /* If we were unable to find any folds, then we likely won't be
13433 * able to find the closures. So just create an empty list.
13434 * Folding will effectively be restricted to the non-Unicode
13435 * rules hard-coded into Perl. (This case happens legitimately
13436 * during compilation of Perl itself before the Unicode tables
13437 * are generated) */
13438 if (_invlist_len(PL_utf8_foldable) == 0) {
13439 PL_utf8_foldclosures = newHV();
13442 /* If the folds haven't been read in, call a fold function
13444 if (! PL_utf8_tofold) {
13445 U8 dummy[UTF8_MAXBYTES+1];
13447 /* This string is just a short named one above \xff */
13448 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
13449 assert(PL_utf8_tofold); /* Verify that worked */
13451 PL_utf8_foldclosures =
13452 _swash_inversion_hash(PL_utf8_tofold);
13456 /* Only the characters in this class that participate in folds need
13457 * be checked. Get the intersection of this class and all the
13458 * possible characters that are foldable. This can quickly narrow
13459 * down a large class */
13460 _invlist_intersection(PL_utf8_foldable, cp_list,
13461 &fold_intersection);
13464 /* Now look at the foldable characters in this class individually */
13465 invlist_iterinit(fold_intersection);
13466 while (invlist_iternext(fold_intersection, &start, &end)) {
13469 /* Locale folding for Latin1 characters is deferred until runtime */
13470 if (LOC && start < 256) {
13474 /* Look at every character in the range */
13475 for (j = start; j <= end; j++) {
13477 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
13483 /* We have the latin1 folding rules hard-coded here so that
13484 * an innocent-looking character class, like /[ks]/i won't
13485 * have to go out to disk to find the possible matches.
13486 * XXX It would be better to generate these via regen, in
13487 * case a new version of the Unicode standard adds new
13488 * mappings, though that is not really likely, and may be
13489 * caught by the default: case of the switch below. */
13491 if (IS_IN_SOME_FOLD_L1(j)) {
13493 /* ASCII is always matched; non-ASCII is matched only
13494 * under Unicode rules */
13495 if (isASCII(j) || AT_LEAST_UNI_SEMANTICS) {
13497 add_cp_to_invlist(cp_list, PL_fold_latin1[j]);
13501 add_cp_to_invlist(depends_list, PL_fold_latin1[j]);
13505 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
13506 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
13508 /* Certain Latin1 characters have matches outside
13509 * Latin1. To get here, <j> is one of those
13510 * characters. None of these matches is valid for
13511 * ASCII characters under /aa, which is why the 'if'
13512 * just above excludes those. These matches only
13513 * happen when the target string is utf8. The code
13514 * below adds the single fold closures for <j> to the
13515 * inversion list. */
13520 add_cp_to_invlist(cp_list, KELVIN_SIGN);
13524 cp_list = add_cp_to_invlist(cp_list,
13525 LATIN_SMALL_LETTER_LONG_S);
13528 cp_list = add_cp_to_invlist(cp_list,
13529 GREEK_CAPITAL_LETTER_MU);
13530 cp_list = add_cp_to_invlist(cp_list,
13531 GREEK_SMALL_LETTER_MU);
13533 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
13534 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
13536 add_cp_to_invlist(cp_list, ANGSTROM_SIGN);
13538 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
13539 cp_list = add_cp_to_invlist(cp_list,
13540 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
13542 case LATIN_SMALL_LETTER_SHARP_S:
13543 cp_list = add_cp_to_invlist(cp_list,
13544 LATIN_CAPITAL_LETTER_SHARP_S);
13546 case 'F': case 'f':
13547 case 'I': case 'i':
13548 case 'L': case 'l':
13549 case 'T': case 't':
13550 case 'A': case 'a':
13551 case 'H': case 'h':
13552 case 'J': case 'j':
13553 case 'N': case 'n':
13554 case 'W': case 'w':
13555 case 'Y': case 'y':
13556 /* These all are targets of multi-character
13557 * folds from code points that require UTF8 to
13558 * express, so they can't match unless the
13559 * target string is in UTF-8, so no action here
13560 * is necessary, as regexec.c properly handles
13561 * the general case for UTF-8 matching and
13562 * multi-char folds */
13565 /* Use deprecated warning to increase the
13566 * chances of this being output */
13567 ckWARN2regdep(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
13574 /* Here is an above Latin1 character. We don't have the rules
13575 * hard-coded for it. First, get its fold. This is the simple
13576 * fold, as the multi-character folds have been handled earlier
13577 * and separated out */
13578 _to_uni_fold_flags(j, foldbuf, &foldlen,
13580 ? FOLD_FLAGS_LOCALE
13581 : (ASCII_FOLD_RESTRICTED)
13582 ? FOLD_FLAGS_NOMIX_ASCII
13585 /* Single character fold of above Latin1. Add everything in
13586 * its fold closure to the list that this node should match.
13587 * The fold closures data structure is a hash with the keys
13588 * being the UTF-8 of every character that is folded to, like
13589 * 'k', and the values each an array of all code points that
13590 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
13591 * Multi-character folds are not included */
13592 if ((listp = hv_fetch(PL_utf8_foldclosures,
13593 (char *) foldbuf, foldlen, FALSE)))
13595 AV* list = (AV*) *listp;
13597 for (k = 0; k <= av_len(list); k++) {
13598 SV** c_p = av_fetch(list, k, FALSE);
13601 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
13605 /* /aa doesn't allow folds between ASCII and non-; /l
13606 * doesn't allow them between above and below 256 */
13607 if ((ASCII_FOLD_RESTRICTED
13608 && (isASCII(c) != isASCII(j)))
13609 || (LOC && c < 256)) {
13613 /* Folds involving non-ascii Latin1 characters
13614 * under /d are added to a separate list */
13615 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
13617 cp_list = add_cp_to_invlist(cp_list, c);
13620 depends_list = add_cp_to_invlist(depends_list, c);
13626 SvREFCNT_dec_NN(fold_intersection);
13629 /* And combine the result (if any) with any inversion list from posix
13630 * classes. The lists are kept separate up to now because we don't want to
13631 * fold the classes (folding of those is automatically handled by the swash
13632 * fetching code) */
13634 if (! DEPENDS_SEMANTICS) {
13636 _invlist_union(cp_list, posixes, &cp_list);
13637 SvREFCNT_dec_NN(posixes);
13644 /* Under /d, we put into a separate list the Latin1 things that
13645 * match only when the target string is utf8 */
13646 SV* nonascii_but_latin1_properties = NULL;
13647 _invlist_intersection(posixes, PL_Latin1,
13648 &nonascii_but_latin1_properties);
13649 _invlist_subtract(nonascii_but_latin1_properties, PL_ASCII,
13650 &nonascii_but_latin1_properties);
13651 _invlist_subtract(posixes, nonascii_but_latin1_properties,
13654 _invlist_union(cp_list, posixes, &cp_list);
13655 SvREFCNT_dec_NN(posixes);
13661 if (depends_list) {
13662 _invlist_union(depends_list, nonascii_but_latin1_properties,
13664 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
13667 depends_list = nonascii_but_latin1_properties;
13672 /* And combine the result (if any) with any inversion list from properties.
13673 * The lists are kept separate up to now so that we can distinguish the two
13674 * in regards to matching above-Unicode. A run-time warning is generated
13675 * if a Unicode property is matched against a non-Unicode code point. But,
13676 * we allow user-defined properties to match anything, without any warning,
13677 * and we also suppress the warning if there is a portion of the character
13678 * class that isn't a Unicode property, and which matches above Unicode, \W
13679 * or [\x{110000}] for example.
13680 * (Note that in this case, unlike the Posix one above, there is no
13681 * <depends_list>, because having a Unicode property forces Unicode
13684 bool warn_super = ! has_user_defined_property;
13687 /* If it matters to the final outcome, see if a non-property
13688 * component of the class matches above Unicode. If so, the
13689 * warning gets suppressed. This is true even if just a single
13690 * such code point is specified, as though not strictly correct if
13691 * another such code point is matched against, the fact that they
13692 * are using above-Unicode code points indicates they should know
13693 * the issues involved */
13695 bool non_prop_matches_above_Unicode =
13696 runtime_posix_matches_above_Unicode
13697 | (invlist_highest(cp_list) > PERL_UNICODE_MAX);
13699 non_prop_matches_above_Unicode =
13700 ! non_prop_matches_above_Unicode;
13702 warn_super = ! non_prop_matches_above_Unicode;
13705 _invlist_union(properties, cp_list, &cp_list);
13706 SvREFCNT_dec_NN(properties);
13709 cp_list = properties;
13713 OP(ret) = ANYOF_WARN_SUPER;
13717 /* Here, we have calculated what code points should be in the character
13720 * Now we can see about various optimizations. Fold calculation (which we
13721 * did above) needs to take place before inversion. Otherwise /[^k]/i
13722 * would invert to include K, which under /i would match k, which it
13723 * shouldn't. Therefore we can't invert folded locale now, as it won't be
13724 * folded until runtime */
13726 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
13727 * at compile time. Besides not inverting folded locale now, we can't
13728 * invert if there are things such as \w, which aren't known until runtime
13731 && ! (LOC && (FOLD || (ANYOF_FLAGS(ret) & ANYOF_CLASS)))
13733 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13735 _invlist_invert(cp_list);
13737 /* Any swash can't be used as-is, because we've inverted things */
13739 SvREFCNT_dec_NN(swash);
13743 /* Clear the invert flag since have just done it here */
13748 *ret_invlist = cp_list;
13749 SvREFCNT_dec(swash);
13751 /* Discard the generated node */
13753 RExC_size = orig_size;
13756 RExC_emit = orig_emit;
13761 /* If we didn't do folding, it's because some information isn't available
13762 * until runtime; set the run-time fold flag for these. (We don't have to
13763 * worry about properties folding, as that is taken care of by the swash
13767 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
13770 /* Some character classes are equivalent to other nodes. Such nodes take
13771 * up less room and generally fewer operations to execute than ANYOF nodes.
13772 * Above, we checked for and optimized into some such equivalents for
13773 * certain common classes that are easy to test. Getting to this point in
13774 * the code means that the class didn't get optimized there. Since this
13775 * code is only executed in Pass 2, it is too late to save space--it has
13776 * been allocated in Pass 1, and currently isn't given back. But turning
13777 * things into an EXACTish node can allow the optimizer to join it to any
13778 * adjacent such nodes. And if the class is equivalent to things like /./,
13779 * expensive run-time swashes can be avoided. Now that we have more
13780 * complete information, we can find things necessarily missed by the
13781 * earlier code. I (khw) am not sure how much to look for here. It would
13782 * be easy, but perhaps too slow, to check any candidates against all the
13783 * node types they could possibly match using _invlistEQ(). */
13788 && ! (ANYOF_FLAGS(ret) & ANYOF_CLASS)
13789 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13792 U8 op = END; /* The optimzation node-type */
13793 const char * cur_parse= RExC_parse;
13795 invlist_iterinit(cp_list);
13796 if (! invlist_iternext(cp_list, &start, &end)) {
13798 /* Here, the list is empty. This happens, for example, when a
13799 * Unicode property is the only thing in the character class, and
13800 * it doesn't match anything. (perluniprops.pod notes such
13803 *flagp |= HASWIDTH|SIMPLE;
13805 else if (start == end) { /* The range is a single code point */
13806 if (! invlist_iternext(cp_list, &start, &end)
13808 /* Don't do this optimization if it would require changing
13809 * the pattern to UTF-8 */
13810 && (start < 256 || UTF))
13812 /* Here, the list contains a single code point. Can optimize
13813 * into an EXACT node */
13822 /* A locale node under folding with one code point can be
13823 * an EXACTFL, as its fold won't be calculated until
13829 /* Here, we are generally folding, but there is only one
13830 * code point to match. If we have to, we use an EXACT
13831 * node, but it would be better for joining with adjacent
13832 * nodes in the optimization pass if we used the same
13833 * EXACTFish node that any such are likely to be. We can
13834 * do this iff the code point doesn't participate in any
13835 * folds. For example, an EXACTF of a colon is the same as
13836 * an EXACT one, since nothing folds to or from a colon. */
13838 if (IS_IN_SOME_FOLD_L1(value)) {
13843 if (! PL_utf8_foldable) {
13844 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13845 &PL_sv_undef, 1, 0);
13846 PL_utf8_foldable = _get_swash_invlist(swash);
13847 SvREFCNT_dec_NN(swash);
13849 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
13854 /* If we haven't found the node type, above, it means we
13855 * can use the prevailing one */
13857 op = compute_EXACTish(pRExC_state);
13862 else if (start == 0) {
13863 if (end == UV_MAX) {
13865 *flagp |= HASWIDTH|SIMPLE;
13868 else if (end == '\n' - 1
13869 && invlist_iternext(cp_list, &start, &end)
13870 && start == '\n' + 1 && end == UV_MAX)
13873 *flagp |= HASWIDTH|SIMPLE;
13877 invlist_iterfinish(cp_list);
13880 RExC_parse = (char *)orig_parse;
13881 RExC_emit = (regnode *)orig_emit;
13883 ret = reg_node(pRExC_state, op);
13885 RExC_parse = (char *)cur_parse;
13887 if (PL_regkind[op] == EXACT) {
13888 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13891 SvREFCNT_dec_NN(cp_list);
13896 /* Here, <cp_list> contains all the code points we can determine at
13897 * compile time that match under all conditions. Go through it, and
13898 * for things that belong in the bitmap, put them there, and delete from
13899 * <cp_list>. While we are at it, see if everything above 255 is in the
13900 * list, and if so, set a flag to speed up execution */
13901 ANYOF_BITMAP_ZERO(ret);
13904 /* This gets set if we actually need to modify things */
13905 bool change_invlist = FALSE;
13909 /* Start looking through <cp_list> */
13910 invlist_iterinit(cp_list);
13911 while (invlist_iternext(cp_list, &start, &end)) {
13915 if (end == UV_MAX && start <= 256) {
13916 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
13919 /* Quit if are above what we should change */
13924 change_invlist = TRUE;
13926 /* Set all the bits in the range, up to the max that we are doing */
13927 high = (end < 255) ? end : 255;
13928 for (i = start; i <= (int) high; i++) {
13929 if (! ANYOF_BITMAP_TEST(ret, i)) {
13930 ANYOF_BITMAP_SET(ret, i);
13936 invlist_iterfinish(cp_list);
13938 /* Done with loop; remove any code points that are in the bitmap from
13940 if (change_invlist) {
13941 _invlist_subtract(cp_list, PL_Latin1, &cp_list);
13944 /* If have completely emptied it, remove it completely */
13945 if (_invlist_len(cp_list) == 0) {
13946 SvREFCNT_dec_NN(cp_list);
13952 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
13955 /* Here, the bitmap has been populated with all the Latin1 code points that
13956 * always match. Can now add to the overall list those that match only
13957 * when the target string is UTF-8 (<depends_list>). */
13958 if (depends_list) {
13960 _invlist_union(cp_list, depends_list, &cp_list);
13961 SvREFCNT_dec_NN(depends_list);
13964 cp_list = depends_list;
13968 /* If there is a swash and more than one element, we can't use the swash in
13969 * the optimization below. */
13970 if (swash && element_count > 1) {
13971 SvREFCNT_dec_NN(swash);
13976 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13978 ARG_SET(ret, ANYOF_NONBITMAP_EMPTY);
13981 /* av[0] stores the character class description in its textual form:
13982 * used later (regexec.c:Perl_regclass_swash()) to initialize the
13983 * appropriate swash, and is also useful for dumping the regnode.
13984 * av[1] if NULL, is a placeholder to later contain the swash computed
13985 * from av[0]. But if no further computation need be done, the
13986 * swash is stored there now.
13987 * av[2] stores the cp_list inversion list for use in addition or
13988 * instead of av[0]; used only if av[1] is NULL
13989 * av[3] is set if any component of the class is from a user-defined
13990 * property; used only if av[1] is NULL */
13991 AV * const av = newAV();
13994 av_store(av, 0, (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13995 ? SvREFCNT_inc(listsv) : &PL_sv_undef);
13997 av_store(av, 1, swash);
13998 SvREFCNT_dec_NN(cp_list);
14001 av_store(av, 1, NULL);
14003 av_store(av, 2, cp_list);
14004 av_store(av, 3, newSVuv(has_user_defined_property));
14008 rv = newRV_noinc(MUTABLE_SV(av));
14009 n = add_data(pRExC_state, 1, "s");
14010 RExC_rxi->data->data[n] = (void*)rv;
14014 *flagp |= HASWIDTH|SIMPLE;
14017 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
14020 /* reg_skipcomment()
14022 Absorbs an /x style # comments from the input stream.
14023 Returns true if there is more text remaining in the stream.
14024 Will set the REG_SEEN_RUN_ON_COMMENT flag if the comment
14025 terminates the pattern without including a newline.
14027 Note its the callers responsibility to ensure that we are
14028 actually in /x mode
14033 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
14037 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
14039 while (RExC_parse < RExC_end)
14040 if (*RExC_parse++ == '\n') {
14045 /* we ran off the end of the pattern without ending
14046 the comment, so we have to add an \n when wrapping */
14047 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
14055 Advances the parse position, and optionally absorbs
14056 "whitespace" from the inputstream.
14058 Without /x "whitespace" means (?#...) style comments only,
14059 with /x this means (?#...) and # comments and whitespace proper.
14061 Returns the RExC_parse point from BEFORE the scan occurs.
14063 This is the /x friendly way of saying RExC_parse++.
14067 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
14069 char* const retval = RExC_parse++;
14071 PERL_ARGS_ASSERT_NEXTCHAR;
14074 if (RExC_end - RExC_parse >= 3
14075 && *RExC_parse == '('
14076 && RExC_parse[1] == '?'
14077 && RExC_parse[2] == '#')
14079 while (*RExC_parse != ')') {
14080 if (RExC_parse == RExC_end)
14081 FAIL("Sequence (?#... not terminated");
14087 if (RExC_flags & RXf_PMf_EXTENDED) {
14088 if (isSPACE(*RExC_parse)) {
14092 else if (*RExC_parse == '#') {
14093 if ( reg_skipcomment( pRExC_state ) )
14102 - reg_node - emit a node
14104 STATIC regnode * /* Location. */
14105 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
14109 regnode * const ret = RExC_emit;
14110 GET_RE_DEBUG_FLAGS_DECL;
14112 PERL_ARGS_ASSERT_REG_NODE;
14115 SIZE_ALIGN(RExC_size);
14119 if (RExC_emit >= RExC_emit_bound)
14120 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14121 op, RExC_emit, RExC_emit_bound);
14123 NODE_ALIGN_FILL(ret);
14125 FILL_ADVANCE_NODE(ptr, op);
14126 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (ptr) - 1);
14127 #ifdef RE_TRACK_PATTERN_OFFSETS
14128 if (RExC_offsets) { /* MJD */
14129 MJD_OFFSET_DEBUG(("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
14130 "reg_node", __LINE__,
14132 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
14133 ? "Overwriting end of array!\n" : "OK",
14134 (UV)(RExC_emit - RExC_emit_start),
14135 (UV)(RExC_parse - RExC_start),
14136 (UV)RExC_offsets[0]));
14137 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
14145 - reganode - emit a node with an argument
14147 STATIC regnode * /* Location. */
14148 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
14152 regnode * const ret = RExC_emit;
14153 GET_RE_DEBUG_FLAGS_DECL;
14155 PERL_ARGS_ASSERT_REGANODE;
14158 SIZE_ALIGN(RExC_size);
14163 assert(2==regarglen[op]+1);
14165 Anything larger than this has to allocate the extra amount.
14166 If we changed this to be:
14168 RExC_size += (1 + regarglen[op]);
14170 then it wouldn't matter. Its not clear what side effect
14171 might come from that so its not done so far.
14176 if (RExC_emit >= RExC_emit_bound)
14177 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14178 op, RExC_emit, RExC_emit_bound);
14180 NODE_ALIGN_FILL(ret);
14182 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
14183 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (ptr) - 2);
14184 #ifdef RE_TRACK_PATTERN_OFFSETS
14185 if (RExC_offsets) { /* MJD */
14186 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14190 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
14191 "Overwriting end of array!\n" : "OK",
14192 (UV)(RExC_emit - RExC_emit_start),
14193 (UV)(RExC_parse - RExC_start),
14194 (UV)RExC_offsets[0]));
14195 Set_Cur_Node_Offset;
14203 - reguni - emit (if appropriate) a Unicode character
14206 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
14210 PERL_ARGS_ASSERT_REGUNI;
14212 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
14216 - reginsert - insert an operator in front of already-emitted operand
14218 * Means relocating the operand.
14221 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
14227 const int offset = regarglen[(U8)op];
14228 const int size = NODE_STEP_REGNODE + offset;
14229 GET_RE_DEBUG_FLAGS_DECL;
14231 PERL_ARGS_ASSERT_REGINSERT;
14232 PERL_UNUSED_ARG(depth);
14233 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
14234 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
14243 if (RExC_open_parens) {
14245 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
14246 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
14247 if ( RExC_open_parens[paren] >= opnd ) {
14248 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
14249 RExC_open_parens[paren] += size;
14251 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
14253 if ( RExC_close_parens[paren] >= opnd ) {
14254 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
14255 RExC_close_parens[paren] += size;
14257 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
14262 while (src > opnd) {
14263 StructCopy(--src, --dst, regnode);
14264 #ifdef RE_TRACK_PATTERN_OFFSETS
14265 if (RExC_offsets) { /* MJD 20010112 */
14266 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
14270 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
14271 ? "Overwriting end of array!\n" : "OK",
14272 (UV)(src - RExC_emit_start),
14273 (UV)(dst - RExC_emit_start),
14274 (UV)RExC_offsets[0]));
14275 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
14276 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
14282 place = opnd; /* Op node, where operand used to be. */
14283 #ifdef RE_TRACK_PATTERN_OFFSETS
14284 if (RExC_offsets) { /* MJD */
14285 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14289 (UV)(place - RExC_emit_start) > RExC_offsets[0]
14290 ? "Overwriting end of array!\n" : "OK",
14291 (UV)(place - RExC_emit_start),
14292 (UV)(RExC_parse - RExC_start),
14293 (UV)RExC_offsets[0]));
14294 Set_Node_Offset(place, RExC_parse);
14295 Set_Node_Length(place, 1);
14298 src = NEXTOPER(place);
14299 FILL_ADVANCE_NODE(place, op);
14300 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (place) - 1);
14301 Zero(src, offset, regnode);
14305 - regtail - set the next-pointer at the end of a node chain of p to val.
14306 - SEE ALSO: regtail_study
14308 /* TODO: All three parms should be const */
14310 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14314 GET_RE_DEBUG_FLAGS_DECL;
14316 PERL_ARGS_ASSERT_REGTAIL;
14318 PERL_UNUSED_ARG(depth);
14324 /* Find last node. */
14327 regnode * const temp = regnext(scan);
14329 SV * const mysv=sv_newmortal();
14330 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
14331 regprop(RExC_rx, mysv, scan);
14332 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
14333 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
14334 (temp == NULL ? "->" : ""),
14335 (temp == NULL ? PL_reg_name[OP(val)] : "")
14343 if (reg_off_by_arg[OP(scan)]) {
14344 ARG_SET(scan, val - scan);
14347 NEXT_OFF(scan) = val - scan;
14353 - regtail_study - set the next-pointer at the end of a node chain of p to val.
14354 - Look for optimizable sequences at the same time.
14355 - currently only looks for EXACT chains.
14357 This is experimental code. The idea is to use this routine to perform
14358 in place optimizations on branches and groups as they are constructed,
14359 with the long term intention of removing optimization from study_chunk so
14360 that it is purely analytical.
14362 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
14363 to control which is which.
14366 /* TODO: All four parms should be const */
14369 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14374 #ifdef EXPERIMENTAL_INPLACESCAN
14377 GET_RE_DEBUG_FLAGS_DECL;
14379 PERL_ARGS_ASSERT_REGTAIL_STUDY;
14385 /* Find last node. */
14389 regnode * const temp = regnext(scan);
14390 #ifdef EXPERIMENTAL_INPLACESCAN
14391 if (PL_regkind[OP(scan)] == EXACT) {
14392 bool has_exactf_sharp_s; /* Unexamined in this routine */
14393 if (join_exact(pRExC_state,scan,&min, &has_exactf_sharp_s, 1,val,depth+1))
14398 switch (OP(scan)) {
14404 case EXACTFU_TRICKYFOLD:
14406 if( exact == PSEUDO )
14408 else if ( exact != OP(scan) )
14417 SV * const mysv=sv_newmortal();
14418 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
14419 regprop(RExC_rx, mysv, scan);
14420 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
14421 SvPV_nolen_const(mysv),
14422 REG_NODE_NUM(scan),
14423 PL_reg_name[exact]);
14430 SV * const mysv_val=sv_newmortal();
14431 DEBUG_PARSE_MSG("");
14432 regprop(RExC_rx, mysv_val, val);
14433 PerlIO_printf(Perl_debug_log, "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
14434 SvPV_nolen_const(mysv_val),
14435 (IV)REG_NODE_NUM(val),
14439 if (reg_off_by_arg[OP(scan)]) {
14440 ARG_SET(scan, val - scan);
14443 NEXT_OFF(scan) = val - scan;
14451 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
14455 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
14461 for (bit=0; bit<32; bit++) {
14462 if (flags & (1<<bit)) {
14463 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
14466 if (!set++ && lead)
14467 PerlIO_printf(Perl_debug_log, "%s",lead);
14468 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
14471 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
14472 if (!set++ && lead) {
14473 PerlIO_printf(Perl_debug_log, "%s",lead);
14476 case REGEX_UNICODE_CHARSET:
14477 PerlIO_printf(Perl_debug_log, "UNICODE");
14479 case REGEX_LOCALE_CHARSET:
14480 PerlIO_printf(Perl_debug_log, "LOCALE");
14482 case REGEX_ASCII_RESTRICTED_CHARSET:
14483 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
14485 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
14486 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
14489 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
14495 PerlIO_printf(Perl_debug_log, "\n");
14497 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14503 Perl_regdump(pTHX_ const regexp *r)
14507 SV * const sv = sv_newmortal();
14508 SV *dsv= sv_newmortal();
14509 RXi_GET_DECL(r,ri);
14510 GET_RE_DEBUG_FLAGS_DECL;
14512 PERL_ARGS_ASSERT_REGDUMP;
14514 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
14516 /* Header fields of interest. */
14517 if (r->anchored_substr) {
14518 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
14519 RE_SV_DUMPLEN(r->anchored_substr), 30);
14520 PerlIO_printf(Perl_debug_log,
14521 "anchored %s%s at %"IVdf" ",
14522 s, RE_SV_TAIL(r->anchored_substr),
14523 (IV)r->anchored_offset);
14524 } else if (r->anchored_utf8) {
14525 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
14526 RE_SV_DUMPLEN(r->anchored_utf8), 30);
14527 PerlIO_printf(Perl_debug_log,
14528 "anchored utf8 %s%s at %"IVdf" ",
14529 s, RE_SV_TAIL(r->anchored_utf8),
14530 (IV)r->anchored_offset);
14532 if (r->float_substr) {
14533 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
14534 RE_SV_DUMPLEN(r->float_substr), 30);
14535 PerlIO_printf(Perl_debug_log,
14536 "floating %s%s at %"IVdf"..%"UVuf" ",
14537 s, RE_SV_TAIL(r->float_substr),
14538 (IV)r->float_min_offset, (UV)r->float_max_offset);
14539 } else if (r->float_utf8) {
14540 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
14541 RE_SV_DUMPLEN(r->float_utf8), 30);
14542 PerlIO_printf(Perl_debug_log,
14543 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
14544 s, RE_SV_TAIL(r->float_utf8),
14545 (IV)r->float_min_offset, (UV)r->float_max_offset);
14547 if (r->check_substr || r->check_utf8)
14548 PerlIO_printf(Perl_debug_log,
14550 (r->check_substr == r->float_substr
14551 && r->check_utf8 == r->float_utf8
14552 ? "(checking floating" : "(checking anchored"));
14553 if (r->extflags & RXf_NOSCAN)
14554 PerlIO_printf(Perl_debug_log, " noscan");
14555 if (r->extflags & RXf_CHECK_ALL)
14556 PerlIO_printf(Perl_debug_log, " isall");
14557 if (r->check_substr || r->check_utf8)
14558 PerlIO_printf(Perl_debug_log, ") ");
14560 if (ri->regstclass) {
14561 regprop(r, sv, ri->regstclass);
14562 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
14564 if (r->extflags & RXf_ANCH) {
14565 PerlIO_printf(Perl_debug_log, "anchored");
14566 if (r->extflags & RXf_ANCH_BOL)
14567 PerlIO_printf(Perl_debug_log, "(BOL)");
14568 if (r->extflags & RXf_ANCH_MBOL)
14569 PerlIO_printf(Perl_debug_log, "(MBOL)");
14570 if (r->extflags & RXf_ANCH_SBOL)
14571 PerlIO_printf(Perl_debug_log, "(SBOL)");
14572 if (r->extflags & RXf_ANCH_GPOS)
14573 PerlIO_printf(Perl_debug_log, "(GPOS)");
14574 PerlIO_putc(Perl_debug_log, ' ');
14576 if (r->extflags & RXf_GPOS_SEEN)
14577 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
14578 if (r->intflags & PREGf_SKIP)
14579 PerlIO_printf(Perl_debug_log, "plus ");
14580 if (r->intflags & PREGf_IMPLICIT)
14581 PerlIO_printf(Perl_debug_log, "implicit ");
14582 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
14583 if (r->extflags & RXf_EVAL_SEEN)
14584 PerlIO_printf(Perl_debug_log, "with eval ");
14585 PerlIO_printf(Perl_debug_log, "\n");
14586 DEBUG_FLAGS_r(regdump_extflags("r->extflags: ",r->extflags));
14588 PERL_ARGS_ASSERT_REGDUMP;
14589 PERL_UNUSED_CONTEXT;
14590 PERL_UNUSED_ARG(r);
14591 #endif /* DEBUGGING */
14595 - regprop - printable representation of opcode
14597 #define EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags) \
14600 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]); \
14601 if (flags & ANYOF_INVERT) \
14602 /*make sure the invert info is in each */ \
14603 sv_catpvs(sv, "^"); \
14609 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
14615 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
14616 static const char * const anyofs[] = {
14617 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
14618 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
14619 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
14620 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
14621 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
14622 || _CC_VERTSPACE != 16
14623 #error Need to adjust order of anyofs[]
14660 RXi_GET_DECL(prog,progi);
14661 GET_RE_DEBUG_FLAGS_DECL;
14663 PERL_ARGS_ASSERT_REGPROP;
14667 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
14668 /* It would be nice to FAIL() here, but this may be called from
14669 regexec.c, and it would be hard to supply pRExC_state. */
14670 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(o), (int)REGNODE_MAX);
14671 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
14673 k = PL_regkind[OP(o)];
14676 sv_catpvs(sv, " ");
14677 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
14678 * is a crude hack but it may be the best for now since
14679 * we have no flag "this EXACTish node was UTF-8"
14681 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
14682 PERL_PV_ESCAPE_UNI_DETECT |
14683 PERL_PV_ESCAPE_NONASCII |
14684 PERL_PV_PRETTY_ELLIPSES |
14685 PERL_PV_PRETTY_LTGT |
14686 PERL_PV_PRETTY_NOCLEAR
14688 } else if (k == TRIE) {
14689 /* print the details of the trie in dumpuntil instead, as
14690 * progi->data isn't available here */
14691 const char op = OP(o);
14692 const U32 n = ARG(o);
14693 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
14694 (reg_ac_data *)progi->data->data[n] :
14696 const reg_trie_data * const trie
14697 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
14699 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
14700 DEBUG_TRIE_COMPILE_r(
14701 Perl_sv_catpvf(aTHX_ sv,
14702 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
14703 (UV)trie->startstate,
14704 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
14705 (UV)trie->wordcount,
14708 (UV)TRIE_CHARCOUNT(trie),
14709 (UV)trie->uniquecharcount
14712 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
14714 int rangestart = -1;
14715 U8* bitmap = IS_ANYOF_TRIE(op) ? (U8*)ANYOF_BITMAP(o) : (U8*)TRIE_BITMAP(trie);
14716 sv_catpvs(sv, "[");
14717 for (i = 0; i <= 256; i++) {
14718 if (i < 256 && BITMAP_TEST(bitmap,i)) {
14719 if (rangestart == -1)
14721 } else if (rangestart != -1) {
14722 if (i <= rangestart + 3)
14723 for (; rangestart < i; rangestart++)
14724 put_byte(sv, rangestart);
14726 put_byte(sv, rangestart);
14727 sv_catpvs(sv, "-");
14728 put_byte(sv, i - 1);
14733 sv_catpvs(sv, "]");
14736 } else if (k == CURLY) {
14737 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
14738 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
14739 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
14741 else if (k == WHILEM && o->flags) /* Ordinal/of */
14742 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
14743 else if (k == REF || k == OPEN || k == CLOSE || k == GROUPP || OP(o)==ACCEPT) {
14744 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
14745 if ( RXp_PAREN_NAMES(prog) ) {
14746 if ( k != REF || (OP(o) < NREF)) {
14747 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
14748 SV **name= av_fetch(list, ARG(o), 0 );
14750 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14753 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
14754 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
14755 I32 *nums=(I32*)SvPVX(sv_dat);
14756 SV **name= av_fetch(list, nums[0], 0 );
14759 for ( n=0; n<SvIVX(sv_dat); n++ ) {
14760 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
14761 (n ? "," : ""), (IV)nums[n]);
14763 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14767 } else if (k == GOSUB)
14768 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o)); /* Paren and offset */
14769 else if (k == VERB) {
14771 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
14772 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
14773 } else if (k == LOGICAL)
14774 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* 2: embedded, otherwise 1 */
14775 else if (k == ANYOF) {
14776 int i, rangestart = -1;
14777 const U8 flags = ANYOF_FLAGS(o);
14781 if (flags & ANYOF_LOCALE)
14782 sv_catpvs(sv, "{loc}");
14783 if (flags & ANYOF_LOC_FOLD)
14784 sv_catpvs(sv, "{i}");
14785 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
14786 if (flags & ANYOF_INVERT)
14787 sv_catpvs(sv, "^");
14789 /* output what the standard cp 0-255 bitmap matches */
14790 for (i = 0; i <= 256; i++) {
14791 if (i < 256 && ANYOF_BITMAP_TEST(o,i)) {
14792 if (rangestart == -1)
14794 } else if (rangestart != -1) {
14795 if (i <= rangestart + 3)
14796 for (; rangestart < i; rangestart++)
14797 put_byte(sv, rangestart);
14799 put_byte(sv, rangestart);
14800 sv_catpvs(sv, "-");
14801 put_byte(sv, i - 1);
14808 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14809 /* output any special charclass tests (used entirely under use locale) */
14810 if (ANYOF_CLASS_TEST_ANY_SET(o))
14811 for (i = 0; i < (int)(sizeof(anyofs)/sizeof(char*)); i++)
14812 if (ANYOF_CLASS_TEST(o,i)) {
14813 sv_catpv(sv, anyofs[i]);
14817 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14819 if (flags & ANYOF_NON_UTF8_LATIN1_ALL) {
14820 sv_catpvs(sv, "{non-utf8-latin1-all}");
14823 /* output information about the unicode matching */
14824 if (flags & ANYOF_UNICODE_ALL)
14825 sv_catpvs(sv, "{unicode_all}");
14826 else if (ANYOF_NONBITMAP(o))
14827 sv_catpvs(sv, "{unicode}");
14828 if (flags & ANYOF_NONBITMAP_NON_UTF8)
14829 sv_catpvs(sv, "{outside bitmap}");
14831 if (ANYOF_NONBITMAP(o)) {
14832 SV *lv; /* Set if there is something outside the bit map */
14833 SV * const sw = regclass_swash(prog, o, FALSE, &lv, NULL);
14834 bool byte_output = FALSE; /* If something in the bitmap has been
14837 if (lv && lv != &PL_sv_undef) {
14839 U8 s[UTF8_MAXBYTES_CASE+1];
14841 for (i = 0; i <= 256; i++) { /* Look at chars in bitmap */
14842 uvchr_to_utf8(s, i);
14845 && ! ANYOF_BITMAP_TEST(o, i) /* Don't duplicate
14849 && swash_fetch(sw, s, TRUE))
14851 if (rangestart == -1)
14853 } else if (rangestart != -1) {
14854 byte_output = TRUE;
14855 if (i <= rangestart + 3)
14856 for (; rangestart < i; rangestart++) {
14857 put_byte(sv, rangestart);
14860 put_byte(sv, rangestart);
14861 sv_catpvs(sv, "-");
14870 char *s = savesvpv(lv);
14871 char * const origs = s;
14873 while (*s && *s != '\n')
14877 const char * const t = ++s;
14880 sv_catpvs(sv, " ");
14886 /* Truncate very long output */
14887 if (s - origs > 256) {
14888 Perl_sv_catpvf(aTHX_ sv,
14890 (int) (s - origs - 1),
14896 else if (*s == '\t') {
14911 SvREFCNT_dec_NN(lv);
14915 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
14917 else if (k == POSIXD || k == NPOSIXD) {
14918 U8 index = FLAGS(o) * 2;
14919 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
14920 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
14923 sv_catpv(sv, anyofs[index]);
14926 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
14927 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
14929 PERL_UNUSED_CONTEXT;
14930 PERL_UNUSED_ARG(sv);
14931 PERL_UNUSED_ARG(o);
14932 PERL_UNUSED_ARG(prog);
14933 #endif /* DEBUGGING */
14937 Perl_re_intuit_string(pTHX_ REGEXP * const r)
14938 { /* Assume that RE_INTUIT is set */
14940 struct regexp *const prog = ReANY(r);
14941 GET_RE_DEBUG_FLAGS_DECL;
14943 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
14944 PERL_UNUSED_CONTEXT;
14948 const char * const s = SvPV_nolen_const(prog->check_substr
14949 ? prog->check_substr : prog->check_utf8);
14951 if (!PL_colorset) reginitcolors();
14952 PerlIO_printf(Perl_debug_log,
14953 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
14955 prog->check_substr ? "" : "utf8 ",
14956 PL_colors[5],PL_colors[0],
14959 (strlen(s) > 60 ? "..." : ""));
14962 return prog->check_substr ? prog->check_substr : prog->check_utf8;
14968 handles refcounting and freeing the perl core regexp structure. When
14969 it is necessary to actually free the structure the first thing it
14970 does is call the 'free' method of the regexp_engine associated to
14971 the regexp, allowing the handling of the void *pprivate; member
14972 first. (This routine is not overridable by extensions, which is why
14973 the extensions free is called first.)
14975 See regdupe and regdupe_internal if you change anything here.
14977 #ifndef PERL_IN_XSUB_RE
14979 Perl_pregfree(pTHX_ REGEXP *r)
14985 Perl_pregfree2(pTHX_ REGEXP *rx)
14988 struct regexp *const r = ReANY(rx);
14989 GET_RE_DEBUG_FLAGS_DECL;
14991 PERL_ARGS_ASSERT_PREGFREE2;
14993 if (r->mother_re) {
14994 ReREFCNT_dec(r->mother_re);
14996 CALLREGFREE_PVT(rx); /* free the private data */
14997 SvREFCNT_dec(RXp_PAREN_NAMES(r));
14998 Safefree(r->xpv_len_u.xpvlenu_pv);
15001 SvREFCNT_dec(r->anchored_substr);
15002 SvREFCNT_dec(r->anchored_utf8);
15003 SvREFCNT_dec(r->float_substr);
15004 SvREFCNT_dec(r->float_utf8);
15005 Safefree(r->substrs);
15007 RX_MATCH_COPY_FREE(rx);
15008 #ifdef PERL_ANY_COW
15009 SvREFCNT_dec(r->saved_copy);
15012 SvREFCNT_dec(r->qr_anoncv);
15013 rx->sv_u.svu_rx = 0;
15018 This is a hacky workaround to the structural issue of match results
15019 being stored in the regexp structure which is in turn stored in
15020 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
15021 could be PL_curpm in multiple contexts, and could require multiple
15022 result sets being associated with the pattern simultaneously, such
15023 as when doing a recursive match with (??{$qr})
15025 The solution is to make a lightweight copy of the regexp structure
15026 when a qr// is returned from the code executed by (??{$qr}) this
15027 lightweight copy doesn't actually own any of its data except for
15028 the starp/end and the actual regexp structure itself.
15034 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
15036 struct regexp *ret;
15037 struct regexp *const r = ReANY(rx);
15038 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
15040 PERL_ARGS_ASSERT_REG_TEMP_COPY;
15043 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
15045 SvOK_off((SV *)ret_x);
15047 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
15048 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
15049 made both spots point to the same regexp body.) */
15050 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
15051 assert(!SvPVX(ret_x));
15052 ret_x->sv_u.svu_rx = temp->sv_any;
15053 temp->sv_any = NULL;
15054 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
15055 SvREFCNT_dec_NN(temp);
15056 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
15057 ing below will not set it. */
15058 SvCUR_set(ret_x, SvCUR(rx));
15061 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
15062 sv_force_normal(sv) is called. */
15064 ret = ReANY(ret_x);
15066 SvFLAGS(ret_x) |= SvUTF8(rx);
15067 /* We share the same string buffer as the original regexp, on which we
15068 hold a reference count, incremented when mother_re is set below.
15069 The string pointer is copied here, being part of the regexp struct.
15071 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
15072 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
15074 const I32 npar = r->nparens+1;
15075 Newx(ret->offs, npar, regexp_paren_pair);
15076 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15079 Newx(ret->substrs, 1, struct reg_substr_data);
15080 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15082 SvREFCNT_inc_void(ret->anchored_substr);
15083 SvREFCNT_inc_void(ret->anchored_utf8);
15084 SvREFCNT_inc_void(ret->float_substr);
15085 SvREFCNT_inc_void(ret->float_utf8);
15087 /* check_substr and check_utf8, if non-NULL, point to either their
15088 anchored or float namesakes, and don't hold a second reference. */
15090 RX_MATCH_COPIED_off(ret_x);
15091 #ifdef PERL_ANY_COW
15092 ret->saved_copy = NULL;
15094 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
15095 SvREFCNT_inc_void(ret->qr_anoncv);
15101 /* regfree_internal()
15103 Free the private data in a regexp. This is overloadable by
15104 extensions. Perl takes care of the regexp structure in pregfree(),
15105 this covers the *pprivate pointer which technically perl doesn't
15106 know about, however of course we have to handle the
15107 regexp_internal structure when no extension is in use.
15109 Note this is called before freeing anything in the regexp
15114 Perl_regfree_internal(pTHX_ REGEXP * const rx)
15117 struct regexp *const r = ReANY(rx);
15118 RXi_GET_DECL(r,ri);
15119 GET_RE_DEBUG_FLAGS_DECL;
15121 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
15127 SV *dsv= sv_newmortal();
15128 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
15129 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
15130 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
15131 PL_colors[4],PL_colors[5],s);
15134 #ifdef RE_TRACK_PATTERN_OFFSETS
15136 Safefree(ri->u.offsets); /* 20010421 MJD */
15138 if (ri->code_blocks) {
15140 for (n = 0; n < ri->num_code_blocks; n++)
15141 SvREFCNT_dec(ri->code_blocks[n].src_regex);
15142 Safefree(ri->code_blocks);
15146 int n = ri->data->count;
15149 /* If you add a ->what type here, update the comment in regcomp.h */
15150 switch (ri->data->what[n]) {
15156 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
15159 Safefree(ri->data->data[n]);
15165 { /* Aho Corasick add-on structure for a trie node.
15166 Used in stclass optimization only */
15168 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
15170 refcount = --aho->refcount;
15173 PerlMemShared_free(aho->states);
15174 PerlMemShared_free(aho->fail);
15175 /* do this last!!!! */
15176 PerlMemShared_free(ri->data->data[n]);
15177 PerlMemShared_free(ri->regstclass);
15183 /* trie structure. */
15185 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
15187 refcount = --trie->refcount;
15190 PerlMemShared_free(trie->charmap);
15191 PerlMemShared_free(trie->states);
15192 PerlMemShared_free(trie->trans);
15194 PerlMemShared_free(trie->bitmap);
15196 PerlMemShared_free(trie->jump);
15197 PerlMemShared_free(trie->wordinfo);
15198 /* do this last!!!! */
15199 PerlMemShared_free(ri->data->data[n]);
15204 Perl_croak(aTHX_ "panic: regfree data code '%c'", ri->data->what[n]);
15207 Safefree(ri->data->what);
15208 Safefree(ri->data);
15214 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
15215 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
15216 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
15219 re_dup - duplicate a regexp.
15221 This routine is expected to clone a given regexp structure. It is only
15222 compiled under USE_ITHREADS.
15224 After all of the core data stored in struct regexp is duplicated
15225 the regexp_engine.dupe method is used to copy any private data
15226 stored in the *pprivate pointer. This allows extensions to handle
15227 any duplication it needs to do.
15229 See pregfree() and regfree_internal() if you change anything here.
15231 #if defined(USE_ITHREADS)
15232 #ifndef PERL_IN_XSUB_RE
15234 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
15238 const struct regexp *r = ReANY(sstr);
15239 struct regexp *ret = ReANY(dstr);
15241 PERL_ARGS_ASSERT_RE_DUP_GUTS;
15243 npar = r->nparens+1;
15244 Newx(ret->offs, npar, regexp_paren_pair);
15245 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15247 if (ret->substrs) {
15248 /* Do it this way to avoid reading from *r after the StructCopy().
15249 That way, if any of the sv_dup_inc()s dislodge *r from the L1
15250 cache, it doesn't matter. */
15251 const bool anchored = r->check_substr
15252 ? r->check_substr == r->anchored_substr
15253 : r->check_utf8 == r->anchored_utf8;
15254 Newx(ret->substrs, 1, struct reg_substr_data);
15255 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15257 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
15258 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
15259 ret->float_substr = sv_dup_inc(ret->float_substr, param);
15260 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
15262 /* check_substr and check_utf8, if non-NULL, point to either their
15263 anchored or float namesakes, and don't hold a second reference. */
15265 if (ret->check_substr) {
15267 assert(r->check_utf8 == r->anchored_utf8);
15268 ret->check_substr = ret->anchored_substr;
15269 ret->check_utf8 = ret->anchored_utf8;
15271 assert(r->check_substr == r->float_substr);
15272 assert(r->check_utf8 == r->float_utf8);
15273 ret->check_substr = ret->float_substr;
15274 ret->check_utf8 = ret->float_utf8;
15276 } else if (ret->check_utf8) {
15278 ret->check_utf8 = ret->anchored_utf8;
15280 ret->check_utf8 = ret->float_utf8;
15285 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
15286 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
15289 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
15291 if (RX_MATCH_COPIED(dstr))
15292 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
15294 ret->subbeg = NULL;
15295 #ifdef PERL_ANY_COW
15296 ret->saved_copy = NULL;
15299 /* Whether mother_re be set or no, we need to copy the string. We
15300 cannot refrain from copying it when the storage points directly to
15301 our mother regexp, because that's
15302 1: a buffer in a different thread
15303 2: something we no longer hold a reference on
15304 so we need to copy it locally. */
15305 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
15306 ret->mother_re = NULL;
15309 #endif /* PERL_IN_XSUB_RE */
15314 This is the internal complement to regdupe() which is used to copy
15315 the structure pointed to by the *pprivate pointer in the regexp.
15316 This is the core version of the extension overridable cloning hook.
15317 The regexp structure being duplicated will be copied by perl prior
15318 to this and will be provided as the regexp *r argument, however
15319 with the /old/ structures pprivate pointer value. Thus this routine
15320 may override any copying normally done by perl.
15322 It returns a pointer to the new regexp_internal structure.
15326 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
15329 struct regexp *const r = ReANY(rx);
15330 regexp_internal *reti;
15332 RXi_GET_DECL(r,ri);
15334 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
15338 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode), char, regexp_internal);
15339 Copy(ri->program, reti->program, len+1, regnode);
15341 reti->num_code_blocks = ri->num_code_blocks;
15342 if (ri->code_blocks) {
15344 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
15345 struct reg_code_block);
15346 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
15347 struct reg_code_block);
15348 for (n = 0; n < ri->num_code_blocks; n++)
15349 reti->code_blocks[n].src_regex = (REGEXP*)
15350 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
15353 reti->code_blocks = NULL;
15355 reti->regstclass = NULL;
15358 struct reg_data *d;
15359 const int count = ri->data->count;
15362 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
15363 char, struct reg_data);
15364 Newx(d->what, count, U8);
15367 for (i = 0; i < count; i++) {
15368 d->what[i] = ri->data->what[i];
15369 switch (d->what[i]) {
15370 /* see also regcomp.h and regfree_internal() */
15371 case 'a': /* actually an AV, but the dup function is identical. */
15375 case 'u': /* actually an HV, but the dup function is identical. */
15376 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
15379 /* This is cheating. */
15380 Newx(d->data[i], 1, struct regnode_charclass_class);
15381 StructCopy(ri->data->data[i], d->data[i],
15382 struct regnode_charclass_class);
15383 reti->regstclass = (regnode*)d->data[i];
15386 /* Trie stclasses are readonly and can thus be shared
15387 * without duplication. We free the stclass in pregfree
15388 * when the corresponding reg_ac_data struct is freed.
15390 reti->regstclass= ri->regstclass;
15394 ((reg_trie_data*)ri->data->data[i])->refcount++;
15399 d->data[i] = ri->data->data[i];
15402 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'", ri->data->what[i]);
15411 reti->name_list_idx = ri->name_list_idx;
15413 #ifdef RE_TRACK_PATTERN_OFFSETS
15414 if (ri->u.offsets) {
15415 Newx(reti->u.offsets, 2*len+1, U32);
15416 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
15419 SetProgLen(reti,len);
15422 return (void*)reti;
15425 #endif /* USE_ITHREADS */
15427 #ifndef PERL_IN_XSUB_RE
15430 - regnext - dig the "next" pointer out of a node
15433 Perl_regnext(pTHX_ regnode *p)
15441 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
15442 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(p), (int)REGNODE_MAX);
15445 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
15454 S_re_croak2(pTHX_ const char* pat1,const char* pat2,...)
15457 STRLEN l1 = strlen(pat1);
15458 STRLEN l2 = strlen(pat2);
15461 const char *message;
15463 PERL_ARGS_ASSERT_RE_CROAK2;
15469 Copy(pat1, buf, l1 , char);
15470 Copy(pat2, buf + l1, l2 , char);
15471 buf[l1 + l2] = '\n';
15472 buf[l1 + l2 + 1] = '\0';
15474 /* ANSI variant takes additional second argument */
15475 va_start(args, pat2);
15479 msv = vmess(buf, &args);
15481 message = SvPV_const(msv,l1);
15484 Copy(message, buf, l1 , char);
15485 buf[l1-1] = '\0'; /* Overwrite \n */
15486 Perl_croak(aTHX_ "%s", buf);
15489 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
15491 #ifndef PERL_IN_XSUB_RE
15493 Perl_save_re_context(pTHX)
15497 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
15499 const REGEXP * const rx = PM_GETRE(PL_curpm);
15502 for (i = 1; i <= RX_NPARENS(rx); i++) {
15503 char digits[TYPE_CHARS(long)];
15504 const STRLEN len = my_snprintf(digits, sizeof(digits), "%lu", (long)i);
15505 GV *const *const gvp
15506 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
15509 GV * const gv = *gvp;
15510 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
15522 S_put_byte(pTHX_ SV *sv, int c)
15524 PERL_ARGS_ASSERT_PUT_BYTE;
15526 /* Our definition of isPRINT() ignores locales, so only bytes that are
15527 not part of UTF-8 are considered printable. I assume that the same
15528 holds for UTF-EBCDIC.
15529 Also, code point 255 is not printable in either (it's E0 in EBCDIC,
15530 which Wikipedia says:
15532 EO, or Eight Ones, is an 8-bit EBCDIC character code represented as all
15533 ones (binary 1111 1111, hexadecimal FF). It is similar, but not
15534 identical, to the ASCII delete (DEL) or rubout control character. ...
15535 it is typically mapped to hexadecimal code 9F, in order to provide a
15536 unique character mapping in both directions)
15538 So the old condition can be simplified to !isPRINT(c) */
15541 Perl_sv_catpvf(aTHX_ sv, "\\x%02x", c);
15544 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
15548 const char string = c;
15549 if (c == '-' || c == ']' || c == '\\' || c == '^')
15550 sv_catpvs(sv, "\\");
15551 sv_catpvn(sv, &string, 1);
15556 #define CLEAR_OPTSTART \
15557 if (optstart) STMT_START { \
15558 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
15562 #define DUMPUNTIL(b,e) CLEAR_OPTSTART; node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
15564 STATIC const regnode *
15565 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
15566 const regnode *last, const regnode *plast,
15567 SV* sv, I32 indent, U32 depth)
15570 U8 op = PSEUDO; /* Arbitrary non-END op. */
15571 const regnode *next;
15572 const regnode *optstart= NULL;
15574 RXi_GET_DECL(r,ri);
15575 GET_RE_DEBUG_FLAGS_DECL;
15577 PERL_ARGS_ASSERT_DUMPUNTIL;
15579 #ifdef DEBUG_DUMPUNTIL
15580 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
15581 last ? last-start : 0,plast ? plast-start : 0);
15584 if (plast && plast < last)
15587 while (PL_regkind[op] != END && (!last || node < last)) {
15588 /* While that wasn't END last time... */
15591 if (op == CLOSE || op == WHILEM)
15593 next = regnext((regnode *)node);
15596 if (OP(node) == OPTIMIZED) {
15597 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
15604 regprop(r, sv, node);
15605 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
15606 (int)(2*indent + 1), "", SvPVX_const(sv));
15608 if (OP(node) != OPTIMIZED) {
15609 if (next == NULL) /* Next ptr. */
15610 PerlIO_printf(Perl_debug_log, " (0)");
15611 else if (PL_regkind[(U8)op] == BRANCH && PL_regkind[OP(next)] != BRANCH )
15612 PerlIO_printf(Perl_debug_log, " (FAIL)");
15614 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
15615 (void)PerlIO_putc(Perl_debug_log, '\n');
15619 if (PL_regkind[(U8)op] == BRANCHJ) {
15622 const regnode *nnode = (OP(next) == LONGJMP
15623 ? regnext((regnode *)next)
15625 if (last && nnode > last)
15627 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
15630 else if (PL_regkind[(U8)op] == BRANCH) {
15632 DUMPUNTIL(NEXTOPER(node), next);
15634 else if ( PL_regkind[(U8)op] == TRIE ) {
15635 const regnode *this_trie = node;
15636 const char op = OP(node);
15637 const U32 n = ARG(node);
15638 const reg_ac_data * const ac = op>=AHOCORASICK ?
15639 (reg_ac_data *)ri->data->data[n] :
15641 const reg_trie_data * const trie =
15642 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
15644 AV *const trie_words = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
15646 const regnode *nextbranch= NULL;
15649 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
15650 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
15652 PerlIO_printf(Perl_debug_log, "%*s%s ",
15653 (int)(2*(indent+3)), "",
15654 elem_ptr ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr), SvCUR(*elem_ptr), 60,
15655 PL_colors[0], PL_colors[1],
15656 (SvUTF8(*elem_ptr) ? PERL_PV_ESCAPE_UNI : 0) |
15657 PERL_PV_PRETTY_ELLIPSES |
15658 PERL_PV_PRETTY_LTGT
15663 U16 dist= trie->jump[word_idx+1];
15664 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
15665 (UV)((dist ? this_trie + dist : next) - start));
15668 nextbranch= this_trie + trie->jump[0];
15669 DUMPUNTIL(this_trie + dist, nextbranch);
15671 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
15672 nextbranch= regnext((regnode *)nextbranch);
15674 PerlIO_printf(Perl_debug_log, "\n");
15677 if (last && next > last)
15682 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
15683 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
15684 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
15686 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
15688 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
15690 else if ( op == PLUS || op == STAR) {
15691 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
15693 else if (PL_regkind[(U8)op] == ANYOF) {
15694 /* arglen 1 + class block */
15695 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_CLASS)
15696 ? ANYOF_CLASS_SKIP : ANYOF_SKIP);
15697 node = NEXTOPER(node);
15699 else if (PL_regkind[(U8)op] == EXACT) {
15700 /* Literal string, where present. */
15701 node += NODE_SZ_STR(node) - 1;
15702 node = NEXTOPER(node);
15705 node = NEXTOPER(node);
15706 node += regarglen[(U8)op];
15708 if (op == CURLYX || op == OPEN)
15712 #ifdef DEBUG_DUMPUNTIL
15713 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
15718 #endif /* DEBUGGING */
15722 * c-indentation-style: bsd
15723 * c-basic-offset: 4
15724 * indent-tabs-mode: nil
15727 * ex: set ts=8 sts=4 sw=4 et: