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 */
2449 /* Finish populating the prev field of the wordinfo array. Walk back
2450 * from each accept state until we find another accept state, and if
2451 * so, point the first word's .prev field at the second word. If the
2452 * second already has a .prev field set, stop now. This will be the
2453 * case either if we've already processed that word's accept state,
2454 * or that state had multiple words, and the overspill words were
2455 * already linked up earlier.
2462 for (word=1; word <= trie->wordcount; word++) {
2464 if (trie->wordinfo[word].prev)
2466 state = trie->wordinfo[word].accept;
2468 state = prev_states[state];
2471 prev = trie->states[state].wordnum;
2475 trie->wordinfo[word].prev = prev;
2477 Safefree(prev_states);
2481 /* and now dump out the compressed format */
2482 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2484 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2486 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2487 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2489 SvREFCNT_dec_NN(revcharmap);
2493 : trie->startstate>1
2499 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2501 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2503 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2504 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2507 We find the fail state for each state in the trie, this state is the longest proper
2508 suffix of the current state's 'word' that is also a proper prefix of another word in our
2509 trie. State 1 represents the word '' and is thus the default fail state. This allows
2510 the DFA not to have to restart after its tried and failed a word at a given point, it
2511 simply continues as though it had been matching the other word in the first place.
2513 'abcdgu'=~/abcdefg|cdgu/
2514 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2515 fail, which would bring us to the state representing 'd' in the second word where we would
2516 try 'g' and succeed, proceeding to match 'cdgu'.
2518 /* add a fail transition */
2519 const U32 trie_offset = ARG(source);
2520 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2522 const U32 ucharcount = trie->uniquecharcount;
2523 const U32 numstates = trie->statecount;
2524 const U32 ubound = trie->lasttrans + ucharcount;
2528 U32 base = trie->states[ 1 ].trans.base;
2531 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2532 GET_RE_DEBUG_FLAGS_DECL;
2534 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2536 PERL_UNUSED_ARG(depth);
2540 ARG_SET( stclass, data_slot );
2541 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2542 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2543 aho->trie=trie_offset;
2544 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2545 Copy( trie->states, aho->states, numstates, reg_trie_state );
2546 Newxz( q, numstates, U32);
2547 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2550 /* initialize fail[0..1] to be 1 so that we always have
2551 a valid final fail state */
2552 fail[ 0 ] = fail[ 1 ] = 1;
2554 for ( charid = 0; charid < ucharcount ; charid++ ) {
2555 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2557 q[ q_write ] = newstate;
2558 /* set to point at the root */
2559 fail[ q[ q_write++ ] ]=1;
2562 while ( q_read < q_write) {
2563 const U32 cur = q[ q_read++ % numstates ];
2564 base = trie->states[ cur ].trans.base;
2566 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2567 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2569 U32 fail_state = cur;
2572 fail_state = fail[ fail_state ];
2573 fail_base = aho->states[ fail_state ].trans.base;
2574 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2576 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2577 fail[ ch_state ] = fail_state;
2578 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2580 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2582 q[ q_write++ % numstates] = ch_state;
2586 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2587 when we fail in state 1, this allows us to use the
2588 charclass scan to find a valid start char. This is based on the principle
2589 that theres a good chance the string being searched contains lots of stuff
2590 that cant be a start char.
2592 fail[ 0 ] = fail[ 1 ] = 0;
2593 DEBUG_TRIE_COMPILE_r({
2594 PerlIO_printf(Perl_debug_log,
2595 "%*sStclass Failtable (%"UVuf" states): 0",
2596 (int)(depth * 2), "", (UV)numstates
2598 for( q_read=1; q_read<numstates; q_read++ ) {
2599 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2601 PerlIO_printf(Perl_debug_log, "\n");
2604 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2609 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2610 * These need to be revisited when a newer toolchain becomes available.
2612 #if defined(__sparc64__) && defined(__GNUC__)
2613 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2614 # undef SPARC64_GCC_WORKAROUND
2615 # define SPARC64_GCC_WORKAROUND 1
2619 #define DEBUG_PEEP(str,scan,depth) \
2620 DEBUG_OPTIMISE_r({if (scan){ \
2621 SV * const mysv=sv_newmortal(); \
2622 regnode *Next = regnext(scan); \
2623 regprop(RExC_rx, mysv, scan); \
2624 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2625 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2626 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2630 /* The below joins as many adjacent EXACTish nodes as possible into a single
2631 * one. The regop may be changed if the node(s) contain certain sequences that
2632 * require special handling. The joining is only done if:
2633 * 1) there is room in the current conglomerated node to entirely contain the
2635 * 2) they are the exact same node type
2637 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2638 * these get optimized out
2640 * If a node is to match under /i (folded), the number of characters it matches
2641 * can be different than its character length if it contains a multi-character
2642 * fold. *min_subtract is set to the total delta of the input nodes.
2644 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2645 * and contains LATIN SMALL LETTER SHARP S
2647 * This is as good a place as any to discuss the design of handling these
2648 * multi-character fold sequences. It's been wrong in Perl for a very long
2649 * time. There are three code points in Unicode whose multi-character folds
2650 * were long ago discovered to mess things up. The previous designs for
2651 * dealing with these involved assigning a special node for them. This
2652 * approach doesn't work, as evidenced by this example:
2653 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2654 * Both these fold to "sss", but if the pattern is parsed to create a node that
2655 * would match just the \xDF, it won't be able to handle the case where a
2656 * successful match would have to cross the node's boundary. The new approach
2657 * that hopefully generally solves the problem generates an EXACTFU_SS node
2660 * It turns out that there are problems with all multi-character folds, and not
2661 * just these three. Now the code is general, for all such cases, but the
2662 * three still have some special handling. The approach taken is:
2663 * 1) This routine examines each EXACTFish node that could contain multi-
2664 * character fold sequences. It returns in *min_subtract how much to
2665 * subtract from the the actual length of the string to get a real minimum
2666 * match length; it is 0 if there are no multi-char folds. This delta is
2667 * used by the caller to adjust the min length of the match, and the delta
2668 * between min and max, so that the optimizer doesn't reject these
2669 * possibilities based on size constraints.
2670 * 2) Certain of these sequences require special handling by the trie code,
2671 * so, if found, this code changes the joined node type to special ops:
2672 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2673 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2674 * is used for an EXACTFU node that contains at least one "ss" sequence in
2675 * it. For non-UTF-8 patterns and strings, this is the only case where
2676 * there is a possible fold length change. That means that a regular
2677 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2678 * with length changes, and so can be processed faster. regexec.c takes
2679 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2680 * pre-folded by regcomp.c. This saves effort in regex matching.
2681 * However, the pre-folding isn't done for non-UTF8 patterns because the
2682 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2683 * down by forcing the pattern into UTF8 unless necessary. Also what
2684 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2685 * possibilities for the non-UTF8 patterns are quite simple, except for
2686 * the sharp s. All the ones that don't involve a UTF-8 target string are
2687 * members of a fold-pair, and arrays are set up for all of them so that
2688 * the other member of the pair can be found quickly. Code elsewhere in
2689 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2690 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2691 * described in the next item.
2692 * 4) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
2693 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2694 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2695 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2696 * character in the pattern corresponds to at most a single character in
2697 * the target string. (And I do mean character, and not byte here, unlike
2698 * other parts of the documentation that have never been updated to
2699 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2700 * two character string 'ss'; in EXACTFA nodes it can match
2701 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2702 * instances where it is violated. I'm reluctant to try to change the
2703 * assumption, as the code involved is impenetrable to me (khw), so
2704 * instead the code here punts. This routine examines (when the pattern
2705 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2706 * boolean indicating whether or not the node contains a sharp s. When it
2707 * is true, the caller sets a flag that later causes the optimizer in this
2708 * file to not set values for the floating and fixed string lengths, and
2709 * thus avoids the optimizer code in regexec.c that makes the invalid
2710 * assumption. Thus, there is no optimization based on string lengths for
2711 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2712 * (The reason the assumption is wrong only in these two cases is that all
2713 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2714 * other folds to their expanded versions. We can't prefold sharp s to
2715 * 'ss' in EXACTF nodes because we don't know at compile time if it
2716 * actually matches 'ss' or not. It will match iff the target string is
2717 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2718 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2719 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2720 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2721 * require the pattern to be forced into UTF-8, the overhead of which we
2725 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2726 if (PL_regkind[OP(scan)] == EXACT) \
2727 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2730 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) {
2731 /* Merge several consecutive EXACTish nodes into one. */
2732 regnode *n = regnext(scan);
2734 regnode *next = scan + NODE_SZ_STR(scan);
2738 regnode *stop = scan;
2739 GET_RE_DEBUG_FLAGS_DECL;
2741 PERL_UNUSED_ARG(depth);
2744 PERL_ARGS_ASSERT_JOIN_EXACT;
2745 #ifndef EXPERIMENTAL_INPLACESCAN
2746 PERL_UNUSED_ARG(flags);
2747 PERL_UNUSED_ARG(val);
2749 DEBUG_PEEP("join",scan,depth);
2751 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2752 * EXACT ones that are mergeable to the current one. */
2754 && (PL_regkind[OP(n)] == NOTHING
2755 || (stringok && OP(n) == OP(scan)))
2757 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2760 if (OP(n) == TAIL || n > next)
2762 if (PL_regkind[OP(n)] == NOTHING) {
2763 DEBUG_PEEP("skip:",n,depth);
2764 NEXT_OFF(scan) += NEXT_OFF(n);
2765 next = n + NODE_STEP_REGNODE;
2772 else if (stringok) {
2773 const unsigned int oldl = STR_LEN(scan);
2774 regnode * const nnext = regnext(n);
2776 /* XXX I (khw) kind of doubt that this works on platforms where
2777 * U8_MAX is above 255 because of lots of other assumptions */
2778 /* Don't join if the sum can't fit into a single node */
2779 if (oldl + STR_LEN(n) > U8_MAX)
2782 DEBUG_PEEP("merg",n,depth);
2785 NEXT_OFF(scan) += NEXT_OFF(n);
2786 STR_LEN(scan) += STR_LEN(n);
2787 next = n + NODE_SZ_STR(n);
2788 /* Now we can overwrite *n : */
2789 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2797 #ifdef EXPERIMENTAL_INPLACESCAN
2798 if (flags && !NEXT_OFF(n)) {
2799 DEBUG_PEEP("atch", val, depth);
2800 if (reg_off_by_arg[OP(n)]) {
2801 ARG_SET(n, val - n);
2804 NEXT_OFF(n) = val - n;
2812 *has_exactf_sharp_s = FALSE;
2814 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2815 * can now analyze for sequences of problematic code points. (Prior to
2816 * this final joining, sequences could have been split over boundaries, and
2817 * hence missed). The sequences only happen in folding, hence for any
2818 * non-EXACT EXACTish node */
2819 if (OP(scan) != EXACT) {
2820 const U8 * const s0 = (U8*) STRING(scan);
2822 const U8 * const s_end = s0 + STR_LEN(scan);
2824 /* One pass is made over the node's string looking for all the
2825 * possibilities. to avoid some tests in the loop, there are two main
2826 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2830 /* Examine the string for a multi-character fold sequence. UTF-8
2831 * patterns have all characters pre-folded by the time this code is
2833 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2834 length sequence we are looking for is 2 */
2837 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2838 if (! len) { /* Not a multi-char fold: get next char */
2843 /* Nodes with 'ss' require special handling, except for EXACTFL
2844 * and EXACTFA for which there is no multi-char fold to this */
2845 if (len == 2 && *s == 's' && *(s+1) == 's'
2846 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2849 OP(scan) = EXACTFU_SS;
2852 else if (len == 6 /* len is the same in both ASCII and EBCDIC
2854 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2855 COMBINING_DIAERESIS_UTF8
2856 COMBINING_ACUTE_ACCENT_UTF8,
2858 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2859 COMBINING_DIAERESIS_UTF8
2860 COMBINING_ACUTE_ACCENT_UTF8,
2865 /* These two folds require special handling by trie's, so
2866 * change the node type to indicate this. If EXACTFA and
2867 * EXACTFL were ever to be handled by trie's, this would
2868 * have to be changed. If this node has already been
2869 * changed to EXACTFU_SS in this loop, leave it as is. (I
2870 * (khw) think it doesn't matter in regexec.c for UTF
2871 * patterns, but no need to change it */
2872 if (OP(scan) == EXACTFU) {
2873 OP(scan) = EXACTFU_TRICKYFOLD;
2877 else { /* Here is a generic multi-char fold. */
2878 const U8* multi_end = s + len;
2880 /* Count how many characters in it. In the case of /l and
2881 * /aa, no folds which contain ASCII code points are
2882 * allowed, so check for those, and skip if found. (In
2883 * EXACTFL, no folds are allowed to any Latin1 code point,
2884 * not just ASCII. But there aren't any of these
2885 * currently, nor ever likely, so don't take the time to
2886 * test for them. The code that generates the
2887 * is_MULTI_foo() macros croaks should one actually get put
2888 * into Unicode .) */
2889 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2890 count = utf8_length(s, multi_end);
2894 while (s < multi_end) {
2897 goto next_iteration;
2907 /* The delta is how long the sequence is minus 1 (1 is how long
2908 * the character that folds to the sequence is) */
2909 *min_subtract += count - 1;
2913 else if (OP(scan) == EXACTFA) {
2915 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2916 * fold to the ASCII range (and there are no existing ones in the
2917 * upper latin1 range). But, as outlined in the comments preceding
2918 * this function, we need to flag any occurrences of the sharp s */
2920 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2921 *has_exactf_sharp_s = TRUE;
2928 else if (OP(scan) != EXACTFL) {
2930 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2931 * multi-char folds that are all Latin1. (This code knows that
2932 * there are no current multi-char folds possible with EXACTFL,
2933 * relying on fold_grind.t to catch any errors if the very unlikely
2934 * event happens that some get added in future Unicode versions.)
2935 * As explained in the comments preceding this function, we look
2936 * also for the sharp s in EXACTF nodes; it can be in the final
2937 * position. Otherwise we can stop looking 1 byte earlier because
2938 * have to find at least two characters for a multi-fold */
2939 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2941 /* The below is perhaps overboard, but this allows us to save a
2942 * test each time through the loop at the expense of a mask. This
2943 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2944 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2945 * are 64. This uses an exclusive 'or' to find that bit and then
2946 * inverts it to form a mask, with just a single 0, in the bit
2947 * position where 'S' and 's' differ. */
2948 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2949 const U8 s_masked = 's' & S_or_s_mask;
2952 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2953 if (! len) { /* Not a multi-char fold. */
2954 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2956 *has_exactf_sharp_s = TRUE;
2963 && ((*s & S_or_s_mask) == s_masked)
2964 && ((*(s+1) & S_or_s_mask) == s_masked))
2967 /* EXACTF nodes need to know that the minimum length
2968 * changed so that a sharp s in the string can match this
2969 * ss in the pattern, but they remain EXACTF nodes, as they
2970 * won't match this unless the target string is is UTF-8,
2971 * which we don't know until runtime */
2972 if (OP(scan) != EXACTF) {
2973 OP(scan) = EXACTFU_SS;
2977 *min_subtract += len - 1;
2984 /* Allow dumping but overwriting the collection of skipped
2985 * ops and/or strings with fake optimized ops */
2986 n = scan + NODE_SZ_STR(scan);
2994 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2998 /* REx optimizer. Converts nodes into quicker variants "in place".
2999 Finds fixed substrings. */
3001 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3002 to the position after last scanned or to NULL. */
3004 #define INIT_AND_WITHP \
3005 assert(!and_withp); \
3006 Newx(and_withp,1,struct regnode_charclass_class); \
3007 SAVEFREEPV(and_withp)
3009 /* this is a chain of data about sub patterns we are processing that
3010 need to be handled separately/specially in study_chunk. Its so
3011 we can simulate recursion without losing state. */
3013 typedef struct scan_frame {
3014 regnode *last; /* last node to process in this frame */
3015 regnode *next; /* next node to process when last is reached */
3016 struct scan_frame *prev; /*previous frame*/
3017 I32 stop; /* what stopparen do we use */
3021 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3024 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3025 I32 *minlenp, I32 *deltap,
3030 struct regnode_charclass_class *and_withp,
3031 U32 flags, U32 depth)
3032 /* scanp: Start here (read-write). */
3033 /* deltap: Write maxlen-minlen here. */
3034 /* last: Stop before this one. */
3035 /* data: string data about the pattern */
3036 /* stopparen: treat close N as END */
3037 /* recursed: which subroutines have we recursed into */
3038 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3041 I32 min = 0; /* There must be at least this number of characters to match */
3043 regnode *scan = *scanp, *next;
3045 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3046 int is_inf_internal = 0; /* The studied chunk is infinite */
3047 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3048 scan_data_t data_fake;
3049 SV *re_trie_maxbuff = NULL;
3050 regnode *first_non_open = scan;
3051 I32 stopmin = I32_MAX;
3052 scan_frame *frame = NULL;
3053 GET_RE_DEBUG_FLAGS_DECL;
3055 PERL_ARGS_ASSERT_STUDY_CHUNK;
3058 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3062 while (first_non_open && OP(first_non_open) == OPEN)
3063 first_non_open=regnext(first_non_open);
3068 while ( scan && OP(scan) != END && scan < last ){
3069 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3070 node length to get a real minimum (because
3071 the folded version may be shorter) */
3072 bool has_exactf_sharp_s = FALSE;
3073 /* Peephole optimizer: */
3074 DEBUG_STUDYDATA("Peep:", data,depth);
3075 DEBUG_PEEP("Peep",scan,depth);
3077 /* Its not clear to khw or hv why this is done here, and not in the
3078 * clauses that deal with EXACT nodes. khw's guess is that it's
3079 * because of a previous design */
3080 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3082 /* Follow the next-chain of the current node and optimize
3083 away all the NOTHINGs from it. */
3084 if (OP(scan) != CURLYX) {
3085 const int max = (reg_off_by_arg[OP(scan)]
3087 /* I32 may be smaller than U16 on CRAYs! */
3088 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3089 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3093 /* Skip NOTHING and LONGJMP. */
3094 while ((n = regnext(n))
3095 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3096 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3097 && off + noff < max)
3099 if (reg_off_by_arg[OP(scan)])
3102 NEXT_OFF(scan) = off;
3107 /* The principal pseudo-switch. Cannot be a switch, since we
3108 look into several different things. */
3109 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3110 || OP(scan) == IFTHEN) {
3111 next = regnext(scan);
3113 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3115 if (OP(next) == code || code == IFTHEN) {
3116 /* NOTE - There is similar code to this block below for handling
3117 TRIE nodes on a re-study. If you change stuff here check there
3119 I32 max1 = 0, min1 = I32_MAX, num = 0;
3120 struct regnode_charclass_class accum;
3121 regnode * const startbranch=scan;
3123 if (flags & SCF_DO_SUBSTR)
3124 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3125 if (flags & SCF_DO_STCLASS)
3126 cl_init_zero(pRExC_state, &accum);
3128 while (OP(scan) == code) {
3129 I32 deltanext, minnext, f = 0, fake;
3130 struct regnode_charclass_class this_class;
3133 data_fake.flags = 0;
3135 data_fake.whilem_c = data->whilem_c;
3136 data_fake.last_closep = data->last_closep;
3139 data_fake.last_closep = &fake;
3141 data_fake.pos_delta = delta;
3142 next = regnext(scan);
3143 scan = NEXTOPER(scan);
3145 scan = NEXTOPER(scan);
3146 if (flags & SCF_DO_STCLASS) {
3147 cl_init(pRExC_state, &this_class);
3148 data_fake.start_class = &this_class;
3149 f = SCF_DO_STCLASS_AND;
3151 if (flags & SCF_WHILEM_VISITED_POS)
3152 f |= SCF_WHILEM_VISITED_POS;
3154 /* we suppose the run is continuous, last=next...*/
3155 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3157 stopparen, recursed, NULL, f,depth+1);
3160 if (deltanext == I32_MAX) {
3161 is_inf = is_inf_internal = 1;
3163 } else if (max1 < minnext + deltanext)
3164 max1 = minnext + deltanext;
3166 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3168 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3169 if ( stopmin > minnext)
3170 stopmin = min + min1;
3171 flags &= ~SCF_DO_SUBSTR;
3173 data->flags |= SCF_SEEN_ACCEPT;
3176 if (data_fake.flags & SF_HAS_EVAL)
3177 data->flags |= SF_HAS_EVAL;
3178 data->whilem_c = data_fake.whilem_c;
3180 if (flags & SCF_DO_STCLASS)
3181 cl_or(pRExC_state, &accum, &this_class);
3183 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3185 if (flags & SCF_DO_SUBSTR) {
3186 data->pos_min += min1;
3187 if (data->pos_delta >= I32_MAX - (max1 - min1))
3188 data->pos_delta = I32_MAX;
3190 data->pos_delta += max1 - min1;
3191 if (max1 != min1 || is_inf)
3192 data->longest = &(data->longest_float);
3195 if (delta == I32_MAX || I32_MAX - delta - (max1 - min1) < 0)
3198 delta += max1 - min1;
3199 if (flags & SCF_DO_STCLASS_OR) {
3200 cl_or(pRExC_state, data->start_class, &accum);
3202 cl_and(data->start_class, and_withp);
3203 flags &= ~SCF_DO_STCLASS;
3206 else if (flags & SCF_DO_STCLASS_AND) {
3208 cl_and(data->start_class, &accum);
3209 flags &= ~SCF_DO_STCLASS;
3212 /* Switch to OR mode: cache the old value of
3213 * data->start_class */
3215 StructCopy(data->start_class, and_withp,
3216 struct regnode_charclass_class);
3217 flags &= ~SCF_DO_STCLASS_AND;
3218 StructCopy(&accum, data->start_class,
3219 struct regnode_charclass_class);
3220 flags |= SCF_DO_STCLASS_OR;
3221 SET_SSC_EOS(data->start_class);
3225 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3228 Assuming this was/is a branch we are dealing with: 'scan' now
3229 points at the item that follows the branch sequence, whatever
3230 it is. We now start at the beginning of the sequence and look
3237 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3239 If we can find such a subsequence we need to turn the first
3240 element into a trie and then add the subsequent branch exact
3241 strings to the trie.
3245 1. patterns where the whole set of branches can be converted.
3247 2. patterns where only a subset can be converted.
3249 In case 1 we can replace the whole set with a single regop
3250 for the trie. In case 2 we need to keep the start and end
3253 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3254 becomes BRANCH TRIE; BRANCH X;
3256 There is an additional case, that being where there is a
3257 common prefix, which gets split out into an EXACT like node
3258 preceding the TRIE node.
3260 If x(1..n)==tail then we can do a simple trie, if not we make
3261 a "jump" trie, such that when we match the appropriate word
3262 we "jump" to the appropriate tail node. Essentially we turn
3263 a nested if into a case structure of sorts.
3268 if (!re_trie_maxbuff) {
3269 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3270 if (!SvIOK(re_trie_maxbuff))
3271 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3273 if ( SvIV(re_trie_maxbuff)>=0 ) {
3275 regnode *first = (regnode *)NULL;
3276 regnode *last = (regnode *)NULL;
3277 regnode *tail = scan;
3282 SV * const mysv = sv_newmortal(); /* for dumping */
3284 /* var tail is used because there may be a TAIL
3285 regop in the way. Ie, the exacts will point to the
3286 thing following the TAIL, but the last branch will
3287 point at the TAIL. So we advance tail. If we
3288 have nested (?:) we may have to move through several
3292 while ( OP( tail ) == TAIL ) {
3293 /* this is the TAIL generated by (?:) */
3294 tail = regnext( tail );
3298 DEBUG_TRIE_COMPILE_r({
3299 regprop(RExC_rx, mysv, tail );
3300 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3301 (int)depth * 2 + 2, "",
3302 "Looking for TRIE'able sequences. Tail node is: ",
3303 SvPV_nolen_const( mysv )
3309 Step through the branches
3310 cur represents each branch,
3311 noper is the first thing to be matched as part of that branch
3312 noper_next is the regnext() of that node.
3314 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3315 via a "jump trie" but we also support building with NOJUMPTRIE,
3316 which restricts the trie logic to structures like /FOO|BAR/.
3318 If noper is a trieable nodetype then the branch is a possible optimization
3319 target. If we are building under NOJUMPTRIE then we require that noper_next
3320 is the same as scan (our current position in the regex program).
3322 Once we have two or more consecutive such branches we can create a
3323 trie of the EXACT's contents and stitch it in place into the program.
3325 If the sequence represents all of the branches in the alternation we
3326 replace the entire thing with a single TRIE node.
3328 Otherwise when it is a subsequence we need to stitch it in place and
3329 replace only the relevant branches. This means the first branch has
3330 to remain as it is used by the alternation logic, and its next pointer,
3331 and needs to be repointed at the item on the branch chain following
3332 the last branch we have optimized away.
3334 This could be either a BRANCH, in which case the subsequence is internal,
3335 or it could be the item following the branch sequence in which case the
3336 subsequence is at the end (which does not necessarily mean the first node
3337 is the start of the alternation).
3339 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3342 ----------------+-----------
3346 EXACTFU_SS | EXACTFU
3347 EXACTFU_TRICKYFOLD | EXACTFU
3352 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3353 ( EXACT == (X) ) ? EXACT : \
3354 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3357 /* dont use tail as the end marker for this traverse */
3358 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3359 regnode * const noper = NEXTOPER( cur );
3360 U8 noper_type = OP( noper );
3361 U8 noper_trietype = TRIE_TYPE( noper_type );
3362 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3363 regnode * const noper_next = regnext( noper );
3364 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3365 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3368 DEBUG_TRIE_COMPILE_r({
3369 regprop(RExC_rx, mysv, cur);
3370 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3371 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3373 regprop(RExC_rx, mysv, noper);
3374 PerlIO_printf( Perl_debug_log, " -> %s",
3375 SvPV_nolen_const(mysv));
3378 regprop(RExC_rx, mysv, noper_next );
3379 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3380 SvPV_nolen_const(mysv));
3382 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3383 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3384 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3388 /* Is noper a trieable nodetype that can be merged with the
3389 * current trie (if there is one)? */
3393 ( noper_trietype == NOTHING)
3394 || ( trietype == NOTHING )
3395 || ( trietype == noper_trietype )
3398 && noper_next == tail
3402 /* Handle mergable triable node
3403 * Either we are the first node in a new trieable sequence,
3404 * in which case we do some bookkeeping, otherwise we update
3405 * the end pointer. */
3408 if ( noper_trietype == NOTHING ) {
3409 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3410 regnode * const noper_next = regnext( noper );
3411 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3412 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3415 if ( noper_next_trietype ) {
3416 trietype = noper_next_trietype;
3417 } else if (noper_next_type) {
3418 /* a NOTHING regop is 1 regop wide. We need at least two
3419 * for a trie so we can't merge this in */
3423 trietype = noper_trietype;
3426 if ( trietype == NOTHING )
3427 trietype = noper_trietype;
3432 } /* end handle mergable triable node */
3434 /* handle unmergable node -
3435 * noper may either be a triable node which can not be tried
3436 * together with the current trie, or a non triable node */
3438 /* If last is set and trietype is not NOTHING then we have found
3439 * at least two triable branch sequences in a row of a similar
3440 * trietype so we can turn them into a trie. If/when we
3441 * allow NOTHING to start a trie sequence this condition will be
3442 * required, and it isn't expensive so we leave it in for now. */
3443 if ( trietype && trietype != NOTHING )
3444 make_trie( pRExC_state,
3445 startbranch, first, cur, tail, count,
3446 trietype, depth+1 );
3447 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3451 && noper_next == tail
3454 /* noper is triable, so we can start a new trie sequence */
3457 trietype = noper_trietype;
3459 /* if we already saw a first but the current node is not triable then we have
3460 * to reset the first information. */
3465 } /* end handle unmergable node */
3466 } /* loop over branches */
3467 DEBUG_TRIE_COMPILE_r({
3468 regprop(RExC_rx, mysv, cur);
3469 PerlIO_printf( Perl_debug_log,
3470 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3471 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3474 if ( last && trietype ) {
3475 if ( trietype != NOTHING ) {
3476 /* the last branch of the sequence was part of a trie,
3477 * so we have to construct it here outside of the loop
3479 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3480 #ifdef TRIE_STUDY_OPT
3481 if ( ((made == MADE_EXACT_TRIE &&
3482 startbranch == first)
3483 || ( first_non_open == first )) &&
3485 flags |= SCF_TRIE_RESTUDY;
3486 if ( startbranch == first
3489 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3494 /* at this point we know whatever we have is a NOTHING sequence/branch
3495 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3497 if ( startbranch == first ) {
3499 /* the entire thing is a NOTHING sequence, something like this:
3500 * (?:|) So we can turn it into a plain NOTHING op. */
3501 DEBUG_TRIE_COMPILE_r({
3502 regprop(RExC_rx, mysv, cur);
3503 PerlIO_printf( Perl_debug_log,
3504 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3505 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3508 OP(startbranch)= NOTHING;
3509 NEXT_OFF(startbranch)= tail - startbranch;
3510 for ( opt= startbranch + 1; opt < tail ; opt++ )
3514 } /* end if ( last) */
3515 } /* TRIE_MAXBUF is non zero */
3520 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3521 scan = NEXTOPER(NEXTOPER(scan));
3522 } else /* single branch is optimized. */
3523 scan = NEXTOPER(scan);
3525 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3526 scan_frame *newframe = NULL;
3531 if (OP(scan) != SUSPEND) {
3532 /* set the pointer */
3533 if (OP(scan) == GOSUB) {
3535 RExC_recurse[ARG2L(scan)] = scan;
3536 start = RExC_open_parens[paren-1];
3537 end = RExC_close_parens[paren-1];
3540 start = RExC_rxi->program + 1;
3544 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3545 SAVEFREEPV(recursed);
3547 if (!PAREN_TEST(recursed,paren+1)) {
3548 PAREN_SET(recursed,paren+1);
3549 Newx(newframe,1,scan_frame);
3551 if (flags & SCF_DO_SUBSTR) {
3552 SCAN_COMMIT(pRExC_state,data,minlenp);
3553 data->longest = &(data->longest_float);
3555 is_inf = is_inf_internal = 1;
3556 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3557 cl_anything(pRExC_state, data->start_class);
3558 flags &= ~SCF_DO_STCLASS;
3561 Newx(newframe,1,scan_frame);
3564 end = regnext(scan);
3569 SAVEFREEPV(newframe);
3570 newframe->next = regnext(scan);
3571 newframe->last = last;
3572 newframe->stop = stopparen;
3573 newframe->prev = frame;
3583 else if (OP(scan) == EXACT) {
3584 I32 l = STR_LEN(scan);
3587 const U8 * const s = (U8*)STRING(scan);
3588 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3589 l = utf8_length(s, s + l);
3591 uc = *((U8*)STRING(scan));
3594 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3595 /* The code below prefers earlier match for fixed
3596 offset, later match for variable offset. */
3597 if (data->last_end == -1) { /* Update the start info. */
3598 data->last_start_min = data->pos_min;
3599 data->last_start_max = is_inf
3600 ? I32_MAX : data->pos_min + data->pos_delta;
3602 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3604 SvUTF8_on(data->last_found);
3606 SV * const sv = data->last_found;
3607 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3608 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3609 if (mg && mg->mg_len >= 0)
3610 mg->mg_len += utf8_length((U8*)STRING(scan),
3611 (U8*)STRING(scan)+STR_LEN(scan));
3613 data->last_end = data->pos_min + l;
3614 data->pos_min += l; /* As in the first entry. */
3615 data->flags &= ~SF_BEFORE_EOL;
3617 if (flags & SCF_DO_STCLASS_AND) {
3618 /* Check whether it is compatible with what we know already! */
3622 /* If compatible, we or it in below. It is compatible if is
3623 * in the bitmp and either 1) its bit or its fold is set, or 2)
3624 * it's for a locale. Even if there isn't unicode semantics
3625 * here, at runtime there may be because of matching against a
3626 * utf8 string, so accept a possible false positive for
3627 * latin1-range folds */
3629 (!(data->start_class->flags & ANYOF_LOCALE)
3630 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3631 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3632 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3637 ANYOF_CLASS_ZERO(data->start_class);
3638 ANYOF_BITMAP_ZERO(data->start_class);
3640 ANYOF_BITMAP_SET(data->start_class, uc);
3641 else if (uc >= 0x100) {
3644 /* Some Unicode code points fold to the Latin1 range; as
3645 * XXX temporary code, instead of figuring out if this is
3646 * one, just assume it is and set all the start class bits
3647 * that could be some such above 255 code point's fold
3648 * which will generate fals positives. As the code
3649 * elsewhere that does compute the fold settles down, it
3650 * can be extracted out and re-used here */
3651 for (i = 0; i < 256; i++){
3652 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3653 ANYOF_BITMAP_SET(data->start_class, i);
3657 CLEAR_SSC_EOS(data->start_class);
3659 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3661 else if (flags & SCF_DO_STCLASS_OR) {
3662 /* false positive possible if the class is case-folded */
3664 ANYOF_BITMAP_SET(data->start_class, uc);
3666 data->start_class->flags |= ANYOF_UNICODE_ALL;
3667 CLEAR_SSC_EOS(data->start_class);
3668 cl_and(data->start_class, and_withp);
3670 flags &= ~SCF_DO_STCLASS;
3672 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3673 I32 l = STR_LEN(scan);
3674 UV uc = *((U8*)STRING(scan));
3676 /* Search for fixed substrings supports EXACT only. */
3677 if (flags & SCF_DO_SUBSTR) {
3679 SCAN_COMMIT(pRExC_state, data, minlenp);
3682 const U8 * const s = (U8 *)STRING(scan);
3683 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3684 l = utf8_length(s, s + l);
3686 if (has_exactf_sharp_s) {
3687 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3689 min += l - min_subtract;
3691 delta += min_subtract;
3692 if (flags & SCF_DO_SUBSTR) {
3693 data->pos_min += l - min_subtract;
3694 if (data->pos_min < 0) {
3697 data->pos_delta += min_subtract;
3699 data->longest = &(data->longest_float);
3702 if (flags & SCF_DO_STCLASS_AND) {
3703 /* Check whether it is compatible with what we know already! */
3706 (!(data->start_class->flags & ANYOF_LOCALE)
3707 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3708 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3712 ANYOF_CLASS_ZERO(data->start_class);
3713 ANYOF_BITMAP_ZERO(data->start_class);
3715 ANYOF_BITMAP_SET(data->start_class, uc);
3716 CLEAR_SSC_EOS(data->start_class);
3717 if (OP(scan) == EXACTFL) {
3718 /* XXX This set is probably no longer necessary, and
3719 * probably wrong as LOCALE now is on in the initial
3721 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3725 /* Also set the other member of the fold pair. In case
3726 * that unicode semantics is called for at runtime, use
3727 * the full latin1 fold. (Can't do this for locale,
3728 * because not known until runtime) */
3729 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3731 /* All other (EXACTFL handled above) folds except under
3732 * /iaa that include s, S, and sharp_s also may include
3734 if (OP(scan) != EXACTFA) {
3735 if (uc == 's' || uc == 'S') {
3736 ANYOF_BITMAP_SET(data->start_class,
3737 LATIN_SMALL_LETTER_SHARP_S);
3739 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3740 ANYOF_BITMAP_SET(data->start_class, 's');
3741 ANYOF_BITMAP_SET(data->start_class, 'S');
3746 else if (uc >= 0x100) {
3748 for (i = 0; i < 256; i++){
3749 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3750 ANYOF_BITMAP_SET(data->start_class, i);
3755 else if (flags & SCF_DO_STCLASS_OR) {
3756 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3757 /* false positive possible if the class is case-folded.
3758 Assume that the locale settings are the same... */
3760 ANYOF_BITMAP_SET(data->start_class, uc);
3761 if (OP(scan) != EXACTFL) {
3763 /* And set the other member of the fold pair, but
3764 * can't do that in locale because not known until
3766 ANYOF_BITMAP_SET(data->start_class,
3767 PL_fold_latin1[uc]);
3769 /* All folds except under /iaa that include s, S,
3770 * and sharp_s also may include the others */
3771 if (OP(scan) != EXACTFA) {
3772 if (uc == 's' || uc == 'S') {
3773 ANYOF_BITMAP_SET(data->start_class,
3774 LATIN_SMALL_LETTER_SHARP_S);
3776 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3777 ANYOF_BITMAP_SET(data->start_class, 's');
3778 ANYOF_BITMAP_SET(data->start_class, 'S');
3783 CLEAR_SSC_EOS(data->start_class);
3785 cl_and(data->start_class, and_withp);
3787 flags &= ~SCF_DO_STCLASS;
3789 else if (REGNODE_VARIES(OP(scan))) {
3790 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3791 I32 f = flags, pos_before = 0;
3792 regnode * const oscan = scan;
3793 struct regnode_charclass_class this_class;
3794 struct regnode_charclass_class *oclass = NULL;
3795 I32 next_is_eval = 0;
3797 switch (PL_regkind[OP(scan)]) {
3798 case WHILEM: /* End of (?:...)* . */
3799 scan = NEXTOPER(scan);
3802 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3803 next = NEXTOPER(scan);
3804 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3806 maxcount = REG_INFTY;
3807 next = regnext(scan);
3808 scan = NEXTOPER(scan);
3812 if (flags & SCF_DO_SUBSTR)
3817 if (flags & SCF_DO_STCLASS) {
3819 maxcount = REG_INFTY;
3820 next = regnext(scan);
3821 scan = NEXTOPER(scan);
3824 is_inf = is_inf_internal = 1;
3825 scan = regnext(scan);
3826 if (flags & SCF_DO_SUBSTR) {
3827 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3828 data->longest = &(data->longest_float);
3830 goto optimize_curly_tail;
3832 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3833 && (scan->flags == stopparen))
3838 mincount = ARG1(scan);
3839 maxcount = ARG2(scan);
3841 next = regnext(scan);
3842 if (OP(scan) == CURLYX) {
3843 I32 lp = (data ? *(data->last_closep) : 0);
3844 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3846 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3847 next_is_eval = (OP(scan) == EVAL);
3849 if (flags & SCF_DO_SUBSTR) {
3850 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3851 pos_before = data->pos_min;
3855 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3857 data->flags |= SF_IS_INF;
3859 if (flags & SCF_DO_STCLASS) {
3860 cl_init(pRExC_state, &this_class);
3861 oclass = data->start_class;
3862 data->start_class = &this_class;
3863 f |= SCF_DO_STCLASS_AND;
3864 f &= ~SCF_DO_STCLASS_OR;
3866 /* Exclude from super-linear cache processing any {n,m}
3867 regops for which the combination of input pos and regex
3868 pos is not enough information to determine if a match
3871 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3872 regex pos at the \s*, the prospects for a match depend not
3873 only on the input position but also on how many (bar\s*)
3874 repeats into the {4,8} we are. */
3875 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3876 f &= ~SCF_WHILEM_VISITED_POS;
3878 /* This will finish on WHILEM, setting scan, or on NULL: */
3879 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3880 last, data, stopparen, recursed, NULL,
3882 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3884 if (flags & SCF_DO_STCLASS)
3885 data->start_class = oclass;
3886 if (mincount == 0 || minnext == 0) {
3887 if (flags & SCF_DO_STCLASS_OR) {
3888 cl_or(pRExC_state, data->start_class, &this_class);
3890 else if (flags & SCF_DO_STCLASS_AND) {
3891 /* Switch to OR mode: cache the old value of
3892 * data->start_class */
3894 StructCopy(data->start_class, and_withp,
3895 struct regnode_charclass_class);
3896 flags &= ~SCF_DO_STCLASS_AND;
3897 StructCopy(&this_class, data->start_class,
3898 struct regnode_charclass_class);
3899 flags |= SCF_DO_STCLASS_OR;
3900 SET_SSC_EOS(data->start_class);
3902 } else { /* Non-zero len */
3903 if (flags & SCF_DO_STCLASS_OR) {
3904 cl_or(pRExC_state, data->start_class, &this_class);
3905 cl_and(data->start_class, and_withp);
3907 else if (flags & SCF_DO_STCLASS_AND)
3908 cl_and(data->start_class, &this_class);
3909 flags &= ~SCF_DO_STCLASS;
3911 if (!scan) /* It was not CURLYX, but CURLY. */
3913 if (!(flags & SCF_TRIE_DOING_RESTUDY)
3914 /* ? quantifier ok, except for (?{ ... }) */
3915 && (next_is_eval || !(mincount == 0 && maxcount == 1))
3916 && (minnext == 0) && (deltanext == 0)
3917 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3918 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3920 /* Fatal warnings may leak the regexp without this: */
3921 SAVEFREESV(RExC_rx_sv);
3922 ckWARNreg(RExC_parse,
3923 "Quantifier unexpected on zero-length expression");
3924 (void)ReREFCNT_inc(RExC_rx_sv);
3927 min += minnext * mincount;
3928 is_inf_internal |= deltanext == I32_MAX
3929 || (maxcount == REG_INFTY && minnext + deltanext > 0);
3930 is_inf |= is_inf_internal;
3934 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3936 /* Try powerful optimization CURLYX => CURLYN. */
3937 if ( OP(oscan) == CURLYX && data
3938 && data->flags & SF_IN_PAR
3939 && !(data->flags & SF_HAS_EVAL)
3940 && !deltanext && minnext == 1 ) {
3941 /* Try to optimize to CURLYN. */
3942 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3943 regnode * const nxt1 = nxt;
3950 if (!REGNODE_SIMPLE(OP(nxt))
3951 && !(PL_regkind[OP(nxt)] == EXACT
3952 && STR_LEN(nxt) == 1))
3958 if (OP(nxt) != CLOSE)
3960 if (RExC_open_parens) {
3961 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3962 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3964 /* Now we know that nxt2 is the only contents: */
3965 oscan->flags = (U8)ARG(nxt);
3967 OP(nxt1) = NOTHING; /* was OPEN. */
3970 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3971 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3972 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3973 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3974 OP(nxt + 1) = OPTIMIZED; /* was count. */
3975 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3980 /* Try optimization CURLYX => CURLYM. */
3981 if ( OP(oscan) == CURLYX && data
3982 && !(data->flags & SF_HAS_PAR)
3983 && !(data->flags & SF_HAS_EVAL)
3984 && !deltanext /* atom is fixed width */
3985 && minnext != 0 /* CURLYM can't handle zero width */
3986 && ! (RExC_seen & REG_SEEN_EXACTF_SHARP_S) /* Nor \xDF */
3988 /* XXXX How to optimize if data == 0? */
3989 /* Optimize to a simpler form. */
3990 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3994 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3995 && (OP(nxt2) != WHILEM))
3997 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3998 /* Need to optimize away parenths. */
3999 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
4000 /* Set the parenth number. */
4001 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
4003 oscan->flags = (U8)ARG(nxt);
4004 if (RExC_open_parens) {
4005 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4006 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4008 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4009 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4012 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4013 OP(nxt + 1) = OPTIMIZED; /* was count. */
4014 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4015 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4018 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4019 regnode *nnxt = regnext(nxt1);
4021 if (reg_off_by_arg[OP(nxt1)])
4022 ARG_SET(nxt1, nxt2 - nxt1);
4023 else if (nxt2 - nxt1 < U16_MAX)
4024 NEXT_OFF(nxt1) = nxt2 - nxt1;
4026 OP(nxt) = NOTHING; /* Cannot beautify */
4031 /* Optimize again: */
4032 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4033 NULL, stopparen, recursed, NULL, 0,depth+1);
4038 else if ((OP(oscan) == CURLYX)
4039 && (flags & SCF_WHILEM_VISITED_POS)
4040 /* See the comment on a similar expression above.
4041 However, this time it's not a subexpression
4042 we care about, but the expression itself. */
4043 && (maxcount == REG_INFTY)
4044 && data && ++data->whilem_c < 16) {
4045 /* This stays as CURLYX, we can put the count/of pair. */
4046 /* Find WHILEM (as in regexec.c) */
4047 regnode *nxt = oscan + NEXT_OFF(oscan);
4049 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4051 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4052 | (RExC_whilem_seen << 4)); /* On WHILEM */
4054 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4056 if (flags & SCF_DO_SUBSTR) {
4057 SV *last_str = NULL;
4058 int counted = mincount != 0;
4060 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
4061 #if defined(SPARC64_GCC_WORKAROUND)
4064 const char *s = NULL;
4067 if (pos_before >= data->last_start_min)
4070 b = data->last_start_min;
4073 s = SvPV_const(data->last_found, l);
4074 old = b - data->last_start_min;
4077 I32 b = pos_before >= data->last_start_min
4078 ? pos_before : data->last_start_min;
4080 const char * const s = SvPV_const(data->last_found, l);
4081 I32 old = b - data->last_start_min;
4085 old = utf8_hop((U8*)s, old) - (U8*)s;
4087 /* Get the added string: */
4088 last_str = newSVpvn_utf8(s + old, l, UTF);
4089 if (deltanext == 0 && pos_before == b) {
4090 /* What was added is a constant string */
4092 SvGROW(last_str, (mincount * l) + 1);
4093 repeatcpy(SvPVX(last_str) + l,
4094 SvPVX_const(last_str), l, mincount - 1);
4095 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4096 /* Add additional parts. */
4097 SvCUR_set(data->last_found,
4098 SvCUR(data->last_found) - l);
4099 sv_catsv(data->last_found, last_str);
4101 SV * sv = data->last_found;
4103 SvUTF8(sv) && SvMAGICAL(sv) ?
4104 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4105 if (mg && mg->mg_len >= 0)
4106 mg->mg_len += CHR_SVLEN(last_str) - l;
4108 data->last_end += l * (mincount - 1);
4111 /* start offset must point into the last copy */
4112 data->last_start_min += minnext * (mincount - 1);
4113 data->last_start_max += is_inf ? I32_MAX
4114 : (maxcount - 1) * (minnext + data->pos_delta);
4117 /* It is counted once already... */
4118 data->pos_min += minnext * (mincount - counted);
4120 PerlIO_printf(Perl_debug_log, "counted=%d deltanext=%d I32_MAX=%d minnext=%d maxcount=%d mincount=%d\n",
4121 counted, deltanext, I32_MAX, minnext, maxcount, mincount);
4122 if (deltanext != I32_MAX)
4123 PerlIO_printf(Perl_debug_log, "LHS=%d RHS=%d\n", -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount, I32_MAX - data->pos_delta);
4125 if (deltanext == I32_MAX || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= I32_MAX - data->pos_delta)
4126 data->pos_delta = I32_MAX;
4128 data->pos_delta += - counted * deltanext +
4129 (minnext + deltanext) * maxcount - minnext * mincount;
4130 if (mincount != maxcount) {
4131 /* Cannot extend fixed substrings found inside
4133 SCAN_COMMIT(pRExC_state,data,minlenp);
4134 if (mincount && last_str) {
4135 SV * const sv = data->last_found;
4136 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4137 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4141 sv_setsv(sv, last_str);
4142 data->last_end = data->pos_min;
4143 data->last_start_min =
4144 data->pos_min - CHR_SVLEN(last_str);
4145 data->last_start_max = is_inf
4147 : data->pos_min + data->pos_delta
4148 - CHR_SVLEN(last_str);
4150 data->longest = &(data->longest_float);
4152 SvREFCNT_dec(last_str);
4154 if (data && (fl & SF_HAS_EVAL))
4155 data->flags |= SF_HAS_EVAL;
4156 optimize_curly_tail:
4157 if (OP(oscan) != CURLYX) {
4158 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4160 NEXT_OFF(oscan) += NEXT_OFF(next);
4163 default: /* REF, and CLUMP only? */
4164 if (flags & SCF_DO_SUBSTR) {
4165 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4166 data->longest = &(data->longest_float);
4168 is_inf = is_inf_internal = 1;
4169 if (flags & SCF_DO_STCLASS_OR)
4170 cl_anything(pRExC_state, data->start_class);
4171 flags &= ~SCF_DO_STCLASS;
4175 else if (OP(scan) == LNBREAK) {
4176 if (flags & SCF_DO_STCLASS) {
4178 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4179 if (flags & SCF_DO_STCLASS_AND) {
4180 for (value = 0; value < 256; value++)
4181 if (!is_VERTWS_cp(value))
4182 ANYOF_BITMAP_CLEAR(data->start_class, value);
4185 for (value = 0; value < 256; value++)
4186 if (is_VERTWS_cp(value))
4187 ANYOF_BITMAP_SET(data->start_class, value);
4189 if (flags & SCF_DO_STCLASS_OR)
4190 cl_and(data->start_class, and_withp);
4191 flags &= ~SCF_DO_STCLASS;
4194 delta++; /* Because of the 2 char string cr-lf */
4195 if (flags & SCF_DO_SUBSTR) {
4196 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4198 data->pos_delta += 1;
4199 data->longest = &(data->longest_float);
4202 else if (REGNODE_SIMPLE(OP(scan))) {
4205 if (flags & SCF_DO_SUBSTR) {
4206 SCAN_COMMIT(pRExC_state,data,minlenp);
4210 if (flags & SCF_DO_STCLASS) {
4212 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4214 /* Some of the logic below assumes that switching
4215 locale on will only add false positives. */
4216 switch (PL_regkind[OP(scan)]) {
4222 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan));
4225 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4226 cl_anything(pRExC_state, data->start_class);
4229 if (OP(scan) == SANY)
4231 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
4232 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
4233 || ANYOF_CLASS_TEST_ANY_SET(data->start_class));
4234 cl_anything(pRExC_state, data->start_class);
4236 if (flags & SCF_DO_STCLASS_AND || !value)
4237 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
4240 if (flags & SCF_DO_STCLASS_AND)
4241 cl_and(data->start_class,
4242 (struct regnode_charclass_class*)scan);
4244 cl_or(pRExC_state, data->start_class,
4245 (struct regnode_charclass_class*)scan);
4253 classnum = FLAGS(scan);
4254 if (flags & SCF_DO_STCLASS_AND) {
4255 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4256 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum) + 1);
4257 for (value = 0; value < loop_max; value++) {
4258 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4259 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4265 if (data->start_class->flags & ANYOF_LOCALE) {
4266 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum));
4270 /* Even if under locale, set the bits for non-locale
4271 * in case it isn't a true locale-node. This will
4272 * create false positives if it truly is locale */
4273 for (value = 0; value < loop_max; value++) {
4274 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4275 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4287 classnum = FLAGS(scan);
4288 if (flags & SCF_DO_STCLASS_AND) {
4289 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4290 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum));
4291 for (value = 0; value < loop_max; value++) {
4292 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4293 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4299 if (data->start_class->flags & ANYOF_LOCALE) {
4300 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum) + 1);
4304 /* Even if under locale, set the bits for non-locale in
4305 * case it isn't a true locale-node. This will create
4306 * false positives if it truly is locale */
4307 for (value = 0; value < loop_max; value++) {
4308 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4309 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4312 if (PL_regkind[OP(scan)] == NPOSIXD) {
4313 data->start_class->flags |= ANYOF_NON_UTF8_LATIN1_ALL;
4319 if (flags & SCF_DO_STCLASS_OR)
4320 cl_and(data->start_class, and_withp);
4321 flags &= ~SCF_DO_STCLASS;
4324 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
4325 data->flags |= (OP(scan) == MEOL
4328 SCAN_COMMIT(pRExC_state, data, minlenp);
4331 else if ( PL_regkind[OP(scan)] == BRANCHJ
4332 /* Lookbehind, or need to calculate parens/evals/stclass: */
4333 && (scan->flags || data || (flags & SCF_DO_STCLASS))
4334 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
4335 if ( OP(scan) == UNLESSM &&
4337 OP(NEXTOPER(NEXTOPER(scan))) == NOTHING &&
4338 OP(regnext(NEXTOPER(NEXTOPER(scan)))) == SUCCEED
4341 regnode *upto= regnext(scan);
4343 SV * const mysv_val=sv_newmortal();
4344 DEBUG_STUDYDATA("OPFAIL",data,depth);
4346 /*DEBUG_PARSE_MSG("opfail");*/
4347 regprop(RExC_rx, mysv_val, upto);
4348 PerlIO_printf(Perl_debug_log, "~ replace with OPFAIL pointed at %s (%"IVdf") offset %"IVdf"\n",
4349 SvPV_nolen_const(mysv_val),
4350 (IV)REG_NODE_NUM(upto),
4355 NEXT_OFF(scan) = upto - scan;
4356 for (opt= scan + 1; opt < upto ; opt++)
4357 OP(opt) = OPTIMIZED;
4361 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4362 || OP(scan) == UNLESSM )
4364 /* Negative Lookahead/lookbehind
4365 In this case we can't do fixed string optimisation.
4368 I32 deltanext, minnext, fake = 0;
4370 struct regnode_charclass_class intrnl;
4373 data_fake.flags = 0;
4375 data_fake.whilem_c = data->whilem_c;
4376 data_fake.last_closep = data->last_closep;
4379 data_fake.last_closep = &fake;
4380 data_fake.pos_delta = delta;
4381 if ( flags & SCF_DO_STCLASS && !scan->flags
4382 && OP(scan) == IFMATCH ) { /* Lookahead */
4383 cl_init(pRExC_state, &intrnl);
4384 data_fake.start_class = &intrnl;
4385 f |= SCF_DO_STCLASS_AND;
4387 if (flags & SCF_WHILEM_VISITED_POS)
4388 f |= SCF_WHILEM_VISITED_POS;
4389 next = regnext(scan);
4390 nscan = NEXTOPER(NEXTOPER(scan));
4391 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
4392 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
4395 FAIL("Variable length lookbehind not implemented");
4397 else if (minnext > (I32)U8_MAX) {
4398 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4400 scan->flags = (U8)minnext;
4403 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4405 if (data_fake.flags & SF_HAS_EVAL)
4406 data->flags |= SF_HAS_EVAL;
4407 data->whilem_c = data_fake.whilem_c;
4409 if (f & SCF_DO_STCLASS_AND) {
4410 if (flags & SCF_DO_STCLASS_OR) {
4411 /* OR before, AND after: ideally we would recurse with
4412 * data_fake to get the AND applied by study of the
4413 * remainder of the pattern, and then derecurse;
4414 * *** HACK *** for now just treat as "no information".
4415 * See [perl #56690].
4417 cl_init(pRExC_state, data->start_class);
4419 /* AND before and after: combine and continue */
4420 const int was = TEST_SSC_EOS(data->start_class);
4422 cl_and(data->start_class, &intrnl);
4424 SET_SSC_EOS(data->start_class);
4428 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4430 /* Positive Lookahead/lookbehind
4431 In this case we can do fixed string optimisation,
4432 but we must be careful about it. Note in the case of
4433 lookbehind the positions will be offset by the minimum
4434 length of the pattern, something we won't know about
4435 until after the recurse.
4437 I32 deltanext, fake = 0;
4439 struct regnode_charclass_class intrnl;
4441 /* We use SAVEFREEPV so that when the full compile
4442 is finished perl will clean up the allocated
4443 minlens when it's all done. This way we don't
4444 have to worry about freeing them when we know
4445 they wont be used, which would be a pain.
4448 Newx( minnextp, 1, I32 );
4449 SAVEFREEPV(minnextp);
4452 StructCopy(data, &data_fake, scan_data_t);
4453 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
4456 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
4457 data_fake.last_found=newSVsv(data->last_found);
4461 data_fake.last_closep = &fake;
4462 data_fake.flags = 0;
4463 data_fake.pos_delta = delta;
4465 data_fake.flags |= SF_IS_INF;
4466 if ( flags & SCF_DO_STCLASS && !scan->flags
4467 && OP(scan) == IFMATCH ) { /* Lookahead */
4468 cl_init(pRExC_state, &intrnl);
4469 data_fake.start_class = &intrnl;
4470 f |= SCF_DO_STCLASS_AND;
4472 if (flags & SCF_WHILEM_VISITED_POS)
4473 f |= SCF_WHILEM_VISITED_POS;
4474 next = regnext(scan);
4475 nscan = NEXTOPER(NEXTOPER(scan));
4477 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
4478 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
4481 FAIL("Variable length lookbehind not implemented");
4483 else if (*minnextp > (I32)U8_MAX) {
4484 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4486 scan->flags = (U8)*minnextp;
4491 if (f & SCF_DO_STCLASS_AND) {
4492 const int was = TEST_SSC_EOS(data.start_class);
4494 cl_and(data->start_class, &intrnl);
4496 SET_SSC_EOS(data->start_class);
4499 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4501 if (data_fake.flags & SF_HAS_EVAL)
4502 data->flags |= SF_HAS_EVAL;
4503 data->whilem_c = data_fake.whilem_c;
4504 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
4505 if (RExC_rx->minlen<*minnextp)
4506 RExC_rx->minlen=*minnextp;
4507 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
4508 SvREFCNT_dec_NN(data_fake.last_found);
4510 if ( data_fake.minlen_fixed != minlenp )
4512 data->offset_fixed= data_fake.offset_fixed;
4513 data->minlen_fixed= data_fake.minlen_fixed;
4514 data->lookbehind_fixed+= scan->flags;
4516 if ( data_fake.minlen_float != minlenp )
4518 data->minlen_float= data_fake.minlen_float;
4519 data->offset_float_min=data_fake.offset_float_min;
4520 data->offset_float_max=data_fake.offset_float_max;
4521 data->lookbehind_float+= scan->flags;
4528 else if (OP(scan) == OPEN) {
4529 if (stopparen != (I32)ARG(scan))
4532 else if (OP(scan) == CLOSE) {
4533 if (stopparen == (I32)ARG(scan)) {
4536 if ((I32)ARG(scan) == is_par) {
4537 next = regnext(scan);
4539 if ( next && (OP(next) != WHILEM) && next < last)
4540 is_par = 0; /* Disable optimization */
4543 *(data->last_closep) = ARG(scan);
4545 else if (OP(scan) == EVAL) {
4547 data->flags |= SF_HAS_EVAL;
4549 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
4550 if (flags & SCF_DO_SUBSTR) {
4551 SCAN_COMMIT(pRExC_state,data,minlenp);
4552 flags &= ~SCF_DO_SUBSTR;
4554 if (data && OP(scan)==ACCEPT) {
4555 data->flags |= SCF_SEEN_ACCEPT;
4560 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
4562 if (flags & SCF_DO_SUBSTR) {
4563 SCAN_COMMIT(pRExC_state,data,minlenp);
4564 data->longest = &(data->longest_float);
4566 is_inf = is_inf_internal = 1;
4567 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4568 cl_anything(pRExC_state, data->start_class);
4569 flags &= ~SCF_DO_STCLASS;
4571 else if (OP(scan) == GPOS) {
4572 if (!(RExC_rx->extflags & RXf_GPOS_FLOAT) &&
4573 !(delta || is_inf || (data && data->pos_delta)))
4575 if (!(RExC_rx->extflags & RXf_ANCH) && (flags & SCF_DO_SUBSTR))
4576 RExC_rx->extflags |= RXf_ANCH_GPOS;
4577 if (RExC_rx->gofs < (U32)min)
4578 RExC_rx->gofs = min;
4580 RExC_rx->extflags |= RXf_GPOS_FLOAT;
4584 #ifdef TRIE_STUDY_OPT
4585 #ifdef FULL_TRIE_STUDY
4586 else if (PL_regkind[OP(scan)] == TRIE) {
4587 /* NOTE - There is similar code to this block above for handling
4588 BRANCH nodes on the initial study. If you change stuff here
4590 regnode *trie_node= scan;
4591 regnode *tail= regnext(scan);
4592 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4593 I32 max1 = 0, min1 = I32_MAX;
4594 struct regnode_charclass_class accum;
4596 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
4597 SCAN_COMMIT(pRExC_state, data,minlenp); /* Cannot merge strings after this. */
4598 if (flags & SCF_DO_STCLASS)
4599 cl_init_zero(pRExC_state, &accum);
4605 const regnode *nextbranch= NULL;
4608 for ( word=1 ; word <= trie->wordcount ; word++)
4610 I32 deltanext=0, minnext=0, f = 0, fake;
4611 struct regnode_charclass_class this_class;
4613 data_fake.flags = 0;
4615 data_fake.whilem_c = data->whilem_c;
4616 data_fake.last_closep = data->last_closep;
4619 data_fake.last_closep = &fake;
4620 data_fake.pos_delta = delta;
4621 if (flags & SCF_DO_STCLASS) {
4622 cl_init(pRExC_state, &this_class);
4623 data_fake.start_class = &this_class;
4624 f = SCF_DO_STCLASS_AND;
4626 if (flags & SCF_WHILEM_VISITED_POS)
4627 f |= SCF_WHILEM_VISITED_POS;
4629 if (trie->jump[word]) {
4631 nextbranch = trie_node + trie->jump[0];
4632 scan= trie_node + trie->jump[word];
4633 /* We go from the jump point to the branch that follows
4634 it. Note this means we need the vestigal unused branches
4635 even though they arent otherwise used.
4637 minnext = study_chunk(pRExC_state, &scan, minlenp,
4638 &deltanext, (regnode *)nextbranch, &data_fake,
4639 stopparen, recursed, NULL, f,depth+1);
4641 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
4642 nextbranch= regnext((regnode*)nextbranch);
4644 if (min1 > (I32)(minnext + trie->minlen))
4645 min1 = minnext + trie->minlen;
4646 if (deltanext == I32_MAX) {
4647 is_inf = is_inf_internal = 1;
4649 } else if (max1 < (I32)(minnext + deltanext + trie->maxlen))
4650 max1 = minnext + deltanext + trie->maxlen;
4652 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4654 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4655 if ( stopmin > min + min1)
4656 stopmin = min + min1;
4657 flags &= ~SCF_DO_SUBSTR;
4659 data->flags |= SCF_SEEN_ACCEPT;
4662 if (data_fake.flags & SF_HAS_EVAL)
4663 data->flags |= SF_HAS_EVAL;
4664 data->whilem_c = data_fake.whilem_c;
4666 if (flags & SCF_DO_STCLASS)
4667 cl_or(pRExC_state, &accum, &this_class);
4670 if (flags & SCF_DO_SUBSTR) {
4671 data->pos_min += min1;
4672 data->pos_delta += max1 - min1;
4673 if (max1 != min1 || is_inf)
4674 data->longest = &(data->longest_float);
4677 delta += max1 - min1;
4678 if (flags & SCF_DO_STCLASS_OR) {
4679 cl_or(pRExC_state, data->start_class, &accum);
4681 cl_and(data->start_class, and_withp);
4682 flags &= ~SCF_DO_STCLASS;
4685 else if (flags & SCF_DO_STCLASS_AND) {
4687 cl_and(data->start_class, &accum);
4688 flags &= ~SCF_DO_STCLASS;
4691 /* Switch to OR mode: cache the old value of
4692 * data->start_class */
4694 StructCopy(data->start_class, and_withp,
4695 struct regnode_charclass_class);
4696 flags &= ~SCF_DO_STCLASS_AND;
4697 StructCopy(&accum, data->start_class,
4698 struct regnode_charclass_class);
4699 flags |= SCF_DO_STCLASS_OR;
4700 SET_SSC_EOS(data->start_class);
4707 else if (PL_regkind[OP(scan)] == TRIE) {
4708 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4711 min += trie->minlen;
4712 delta += (trie->maxlen - trie->minlen);
4713 flags &= ~SCF_DO_STCLASS; /* xxx */
4714 if (flags & SCF_DO_SUBSTR) {
4715 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4716 data->pos_min += trie->minlen;
4717 data->pos_delta += (trie->maxlen - trie->minlen);
4718 if (trie->maxlen != trie->minlen)
4719 data->longest = &(data->longest_float);
4721 if (trie->jump) /* no more substrings -- for now /grr*/
4722 flags &= ~SCF_DO_SUBSTR;
4724 #endif /* old or new */
4725 #endif /* TRIE_STUDY_OPT */
4727 /* Else: zero-length, ignore. */
4728 scan = regnext(scan);
4733 stopparen = frame->stop;
4734 frame = frame->prev;
4735 goto fake_study_recurse;
4740 DEBUG_STUDYDATA("pre-fin:",data,depth);
4743 *deltap = is_inf_internal ? I32_MAX : delta;
4744 if (flags & SCF_DO_SUBSTR && is_inf)
4745 data->pos_delta = I32_MAX - data->pos_min;
4746 if (is_par > (I32)U8_MAX)
4748 if (is_par && pars==1 && data) {
4749 data->flags |= SF_IN_PAR;
4750 data->flags &= ~SF_HAS_PAR;
4752 else if (pars && data) {
4753 data->flags |= SF_HAS_PAR;
4754 data->flags &= ~SF_IN_PAR;
4756 if (flags & SCF_DO_STCLASS_OR)
4757 cl_and(data->start_class, and_withp);
4758 if (flags & SCF_TRIE_RESTUDY)
4759 data->flags |= SCF_TRIE_RESTUDY;
4761 DEBUG_STUDYDATA("post-fin:",data,depth);
4763 return min < stopmin ? min : stopmin;
4767 S_add_data(RExC_state_t *pRExC_state, U32 n, const char *s)
4769 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
4771 PERL_ARGS_ASSERT_ADD_DATA;
4773 Renewc(RExC_rxi->data,
4774 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
4775 char, struct reg_data);
4777 Renew(RExC_rxi->data->what, count + n, U8);
4779 Newx(RExC_rxi->data->what, n, U8);
4780 RExC_rxi->data->count = count + n;
4781 Copy(s, RExC_rxi->data->what + count, n, U8);
4785 /*XXX: todo make this not included in a non debugging perl */
4786 #ifndef PERL_IN_XSUB_RE
4788 Perl_reginitcolors(pTHX)
4791 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
4793 char *t = savepv(s);
4797 t = strchr(t, '\t');
4803 PL_colors[i] = t = (char *)"";
4808 PL_colors[i++] = (char *)"";
4815 #ifdef TRIE_STUDY_OPT
4816 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
4819 (data.flags & SCF_TRIE_RESTUDY) \
4827 #define CHECK_RESTUDY_GOTO_butfirst
4831 * pregcomp - compile a regular expression into internal code
4833 * Decides which engine's compiler to call based on the hint currently in
4837 #ifndef PERL_IN_XSUB_RE
4839 /* return the currently in-scope regex engine (or the default if none) */
4841 regexp_engine const *
4842 Perl_current_re_engine(pTHX)
4846 if (IN_PERL_COMPILETIME) {
4847 HV * const table = GvHV(PL_hintgv);
4851 return &PL_core_reg_engine;
4852 ptr = hv_fetchs(table, "regcomp", FALSE);
4853 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
4854 return &PL_core_reg_engine;
4855 return INT2PTR(regexp_engine*,SvIV(*ptr));
4859 if (!PL_curcop->cop_hints_hash)
4860 return &PL_core_reg_engine;
4861 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
4862 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
4863 return &PL_core_reg_engine;
4864 return INT2PTR(regexp_engine*,SvIV(ptr));
4870 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
4873 regexp_engine const *eng = current_re_engine();
4874 GET_RE_DEBUG_FLAGS_DECL;
4876 PERL_ARGS_ASSERT_PREGCOMP;
4878 /* Dispatch a request to compile a regexp to correct regexp engine. */
4880 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
4883 return CALLREGCOMP_ENG(eng, pattern, flags);
4887 /* public(ish) entry point for the perl core's own regex compiling code.
4888 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
4889 * pattern rather than a list of OPs, and uses the internal engine rather
4890 * than the current one */
4893 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
4895 SV *pat = pattern; /* defeat constness! */
4896 PERL_ARGS_ASSERT_RE_COMPILE;
4897 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
4898 #ifdef PERL_IN_XSUB_RE
4901 &PL_core_reg_engine,
4903 NULL, NULL, rx_flags, 0);
4907 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
4908 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
4909 * point to the realloced string and length.
4911 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
4915 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
4916 char **pat_p, STRLEN *plen_p, int num_code_blocks)
4918 U8 *const src = (U8*)*pat_p;
4921 STRLEN s = 0, d = 0;
4923 GET_RE_DEBUG_FLAGS_DECL;
4925 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
4926 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
4928 Newx(dst, *plen_p * 2 + 1, U8);
4930 while (s < *plen_p) {
4931 const UV uv = NATIVE_TO_ASCII(src[s]);
4932 if (UNI_IS_INVARIANT(uv))
4933 dst[d] = (U8)UTF_TO_NATIVE(uv);
4935 dst[d++] = (U8)UTF8_EIGHT_BIT_HI(uv);
4936 dst[d] = (U8)UTF8_EIGHT_BIT_LO(uv);
4938 if (n < num_code_blocks) {
4939 if (!do_end && pRExC_state->code_blocks[n].start == s) {
4940 pRExC_state->code_blocks[n].start = d;
4941 assert(dst[d] == '(');
4944 else if (do_end && pRExC_state->code_blocks[n].end == s) {
4945 pRExC_state->code_blocks[n].end = d;
4946 assert(dst[d] == ')');
4956 *pat_p = (char*) dst;
4958 RExC_orig_utf8 = RExC_utf8 = 1;
4963 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
4964 * while recording any code block indices, and handling overloading,
4965 * nested qr// objects etc. If pat is null, it will allocate a new
4966 * string, or just return the first arg, if there's only one.
4968 * Returns the malloced/updated pat.
4969 * patternp and pat_count is the array of SVs to be concatted;
4970 * oplist is the optional list of ops that generated the SVs;
4971 * recompile_p is a pointer to a boolean that will be set if
4972 * the regex will need to be recompiled.
4973 * delim, if non-null is an SV that will be inserted between each element
4977 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
4978 SV *pat, SV ** const patternp, int pat_count,
4979 OP *oplist, bool *recompile_p, SV *delim)
4983 bool use_delim = FALSE;
4984 bool alloced = FALSE;
4986 /* if we know we have at least two args, create an empty string,
4987 * then concatenate args to that. For no args, return an empty string */
4988 if (!pat && pat_count != 1) {
4989 pat = newSVpvn("", 0);
4994 for (svp = patternp; svp < patternp + pat_count; svp++) {
4997 STRLEN orig_patlen = 0;
4999 SV *msv = use_delim ? delim : *svp;
5001 /* if we've got a delimiter, we go round the loop twice for each
5002 * svp slot (except the last), using the delimiter the second
5011 if (SvTYPE(msv) == SVt_PVAV) {
5012 /* we've encountered an interpolated array within
5013 * the pattern, e.g. /...@a..../. Expand the list of elements,
5014 * then recursively append elements.
5015 * The code in this block is based on S_pushav() */
5017 AV *const av = (AV*)msv;
5018 const I32 maxarg = AvFILL(av) + 1;
5022 assert(oplist->op_type == OP_PADAV
5023 || oplist->op_type == OP_RV2AV);
5024 oplist = oplist->op_sibling;;
5027 if (SvRMAGICAL(av)) {
5030 Newx(array, maxarg, SV*);
5032 for (i=0; i < (U32)maxarg; i++) {
5033 SV ** const svp = av_fetch(av, i, FALSE);
5034 array[i] = svp ? *svp : &PL_sv_undef;
5038 array = AvARRAY(av);
5040 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5041 array, maxarg, NULL, recompile_p,
5043 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5049 /* we make the assumption here that each op in the list of
5050 * op_siblings maps to one SV pushed onto the stack,
5051 * except for code blocks, with have both an OP_NULL and
5053 * This allows us to match up the list of SVs against the
5054 * list of OPs to find the next code block.
5056 * Note that PUSHMARK PADSV PADSV ..
5058 * PADRANGE PADSV PADSV ..
5059 * so the alignment still works. */
5062 if (oplist->op_type == OP_NULL
5063 && (oplist->op_flags & OPf_SPECIAL))
5065 assert(n < pRExC_state->num_code_blocks);
5066 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5067 pRExC_state->code_blocks[n].block = oplist;
5068 pRExC_state->code_blocks[n].src_regex = NULL;
5071 oplist = oplist->op_sibling; /* skip CONST */
5074 oplist = oplist->op_sibling;;
5077 /* apply magic and QR overloading to arg */
5080 if (SvROK(msv) && SvAMAGIC(msv)) {
5081 SV *sv = AMG_CALLunary(msv, regexp_amg);
5085 if (SvTYPE(sv) != SVt_REGEXP)
5086 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5091 /* try concatenation overload ... */
5092 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5093 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5096 /* overloading involved: all bets are off over literal
5097 * code. Pretend we haven't seen it */
5098 pRExC_state->num_code_blocks -= n;
5102 /* ... or failing that, try "" overload */
5103 while (SvAMAGIC(msv)
5104 && (sv = AMG_CALLunary(msv, string_amg))
5108 && SvRV(msv) == SvRV(sv))
5113 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5117 /* this is a partially unrolled
5118 * sv_catsv_nomg(pat, msv);
5119 * that allows us to adjust code block indices if
5122 char *dst = SvPV_force_nomg(pat, dlen);
5124 if (SvUTF8(msv) && !SvUTF8(pat)) {
5125 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
5126 sv_setpvn(pat, dst, dlen);
5129 sv_catsv_nomg(pat, msv);
5136 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5139 /* extract any code blocks within any embedded qr//'s */
5140 if (rx && SvTYPE(rx) == SVt_REGEXP
5141 && RX_ENGINE((REGEXP*)rx)->op_comp)
5144 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5145 if (ri->num_code_blocks) {
5147 /* the presence of an embedded qr// with code means
5148 * we should always recompile: the text of the
5149 * qr// may not have changed, but it may be a
5150 * different closure than last time */
5152 Renew(pRExC_state->code_blocks,
5153 pRExC_state->num_code_blocks + ri->num_code_blocks,
5154 struct reg_code_block);
5155 pRExC_state->num_code_blocks += ri->num_code_blocks;
5157 for (i=0; i < ri->num_code_blocks; i++) {
5158 struct reg_code_block *src, *dst;
5159 STRLEN offset = orig_patlen
5160 + ReANY((REGEXP *)rx)->pre_prefix;
5161 assert(n < pRExC_state->num_code_blocks);
5162 src = &ri->code_blocks[i];
5163 dst = &pRExC_state->code_blocks[n];
5164 dst->start = src->start + offset;
5165 dst->end = src->end + offset;
5166 dst->block = src->block;
5167 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5176 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
5185 /* see if there are any run-time code blocks in the pattern.
5186 * False positives are allowed */
5189 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5190 char *pat, STRLEN plen)
5195 for (s = 0; s < plen; s++) {
5196 if (n < pRExC_state->num_code_blocks
5197 && s == pRExC_state->code_blocks[n].start)
5199 s = pRExC_state->code_blocks[n].end;
5203 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
5205 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
5207 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
5214 /* Handle run-time code blocks. We will already have compiled any direct
5215 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
5216 * copy of it, but with any literal code blocks blanked out and
5217 * appropriate chars escaped; then feed it into
5219 * eval "qr'modified_pattern'"
5223 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
5227 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
5229 * After eval_sv()-ing that, grab any new code blocks from the returned qr
5230 * and merge them with any code blocks of the original regexp.
5232 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
5233 * instead, just save the qr and return FALSE; this tells our caller that
5234 * the original pattern needs upgrading to utf8.
5238 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5239 char *pat, STRLEN plen)
5243 GET_RE_DEBUG_FLAGS_DECL;
5245 if (pRExC_state->runtime_code_qr) {
5246 /* this is the second time we've been called; this should
5247 * only happen if the main pattern got upgraded to utf8
5248 * during compilation; re-use the qr we compiled first time
5249 * round (which should be utf8 too)
5251 qr = pRExC_state->runtime_code_qr;
5252 pRExC_state->runtime_code_qr = NULL;
5253 assert(RExC_utf8 && SvUTF8(qr));
5259 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
5263 /* determine how many extra chars we need for ' and \ escaping */
5264 for (s = 0; s < plen; s++) {
5265 if (pat[s] == '\'' || pat[s] == '\\')
5269 Newx(newpat, newlen, char);
5271 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
5273 for (s = 0; s < plen; s++) {
5274 if (n < pRExC_state->num_code_blocks
5275 && s == pRExC_state->code_blocks[n].start)
5277 /* blank out literal code block */
5278 assert(pat[s] == '(');
5279 while (s <= pRExC_state->code_blocks[n].end) {
5287 if (pat[s] == '\'' || pat[s] == '\\')
5292 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
5296 PerlIO_printf(Perl_debug_log,
5297 "%sre-parsing pattern for runtime code:%s %s\n",
5298 PL_colors[4],PL_colors[5],newpat);
5301 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5307 PUSHSTACKi(PERLSI_REQUIRE);
5308 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
5309 * parsing qr''; normally only q'' does this. It also alters
5311 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
5312 SvREFCNT_dec_NN(sv);
5317 SV * const errsv = ERRSV;
5318 if (SvTRUE_NN(errsv))
5320 Safefree(pRExC_state->code_blocks);
5321 /* use croak_sv ? */
5322 Perl_croak_nocontext("%s", SvPV_nolen_const(errsv));
5325 assert(SvROK(qr_ref));
5327 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5328 /* the leaving below frees the tmp qr_ref.
5329 * Give qr a life of its own */
5337 if (!RExC_utf8 && SvUTF8(qr)) {
5338 /* first time through; the pattern got upgraded; save the
5339 * qr for the next time through */
5340 assert(!pRExC_state->runtime_code_qr);
5341 pRExC_state->runtime_code_qr = qr;
5346 /* extract any code blocks within the returned qr// */
5349 /* merge the main (r1) and run-time (r2) code blocks into one */
5351 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
5352 struct reg_code_block *new_block, *dst;
5353 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
5356 if (!r2->num_code_blocks) /* we guessed wrong */
5358 SvREFCNT_dec_NN(qr);
5363 r1->num_code_blocks + r2->num_code_blocks,
5364 struct reg_code_block);
5367 while ( i1 < r1->num_code_blocks
5368 || i2 < r2->num_code_blocks)
5370 struct reg_code_block *src;
5373 if (i1 == r1->num_code_blocks) {
5374 src = &r2->code_blocks[i2++];
5377 else if (i2 == r2->num_code_blocks)
5378 src = &r1->code_blocks[i1++];
5379 else if ( r1->code_blocks[i1].start
5380 < r2->code_blocks[i2].start)
5382 src = &r1->code_blocks[i1++];
5383 assert(src->end < r2->code_blocks[i2].start);
5386 assert( r1->code_blocks[i1].start
5387 > r2->code_blocks[i2].start);
5388 src = &r2->code_blocks[i2++];
5390 assert(src->end < r1->code_blocks[i1].start);
5393 assert(pat[src->start] == '(');
5394 assert(pat[src->end] == ')');
5395 dst->start = src->start;
5396 dst->end = src->end;
5397 dst->block = src->block;
5398 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
5402 r1->num_code_blocks += r2->num_code_blocks;
5403 Safefree(r1->code_blocks);
5404 r1->code_blocks = new_block;
5407 SvREFCNT_dec_NN(qr);
5413 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)
5415 /* This is the common code for setting up the floating and fixed length
5416 * string data extracted from Perl_re_op_compile() below. Returns a boolean
5417 * as to whether succeeded or not */
5421 if (! (longest_length
5422 || (eol /* Can't have SEOL and MULTI */
5423 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
5425 /* See comments for join_exact for why REG_SEEN_EXACTF_SHARP_S */
5426 || (RExC_seen & REG_SEEN_EXACTF_SHARP_S))
5431 /* copy the information about the longest from the reg_scan_data
5432 over to the program. */
5433 if (SvUTF8(sv_longest)) {
5434 *rx_utf8 = sv_longest;
5437 *rx_substr = sv_longest;
5440 /* end_shift is how many chars that must be matched that
5441 follow this item. We calculate it ahead of time as once the
5442 lookbehind offset is added in we lose the ability to correctly
5444 ml = minlen ? *(minlen) : (I32)longest_length;
5445 *rx_end_shift = ml - offset
5446 - longest_length + (SvTAIL(sv_longest) != 0)
5449 t = (eol/* Can't have SEOL and MULTI */
5450 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
5451 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
5457 * Perl_re_op_compile - the perl internal RE engine's function to compile a
5458 * regular expression into internal code.
5459 * The pattern may be passed either as:
5460 * a list of SVs (patternp plus pat_count)
5461 * a list of OPs (expr)
5462 * If both are passed, the SV list is used, but the OP list indicates
5463 * which SVs are actually pre-compiled code blocks
5465 * The SVs in the list have magic and qr overloading applied to them (and
5466 * the list may be modified in-place with replacement SVs in the latter
5469 * If the pattern hasn't changed from old_re, then old_re will be
5472 * eng is the current engine. If that engine has an op_comp method, then
5473 * handle directly (i.e. we assume that op_comp was us); otherwise, just
5474 * do the initial concatenation of arguments and pass on to the external
5477 * If is_bare_re is not null, set it to a boolean indicating whether the
5478 * arg list reduced (after overloading) to a single bare regex which has
5479 * been returned (i.e. /$qr/).
5481 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
5483 * pm_flags contains the PMf_* flags, typically based on those from the
5484 * pm_flags field of the related PMOP. Currently we're only interested in
5485 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
5487 * We can't allocate space until we know how big the compiled form will be,
5488 * but we can't compile it (and thus know how big it is) until we've got a
5489 * place to put the code. So we cheat: we compile it twice, once with code
5490 * generation turned off and size counting turned on, and once "for real".
5491 * This also means that we don't allocate space until we are sure that the
5492 * thing really will compile successfully, and we never have to move the
5493 * code and thus invalidate pointers into it. (Note that it has to be in
5494 * one piece because free() must be able to free it all.) [NB: not true in perl]
5496 * Beware that the optimization-preparation code in here knows about some
5497 * of the structure of the compiled regexp. [I'll say.]
5501 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
5502 OP *expr, const regexp_engine* eng, REGEXP *old_re,
5503 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
5508 regexp_internal *ri;
5516 SV *code_blocksv = NULL;
5517 SV** new_patternp = patternp;
5519 /* these are all flags - maybe they should be turned
5520 * into a single int with different bit masks */
5521 I32 sawlookahead = 0;
5526 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
5528 bool runtime_code = 0;
5530 RExC_state_t RExC_state;
5531 RExC_state_t * const pRExC_state = &RExC_state;
5532 #ifdef TRIE_STUDY_OPT
5534 RExC_state_t copyRExC_state;
5536 GET_RE_DEBUG_FLAGS_DECL;
5538 PERL_ARGS_ASSERT_RE_OP_COMPILE;
5540 DEBUG_r(if (!PL_colorset) reginitcolors());
5542 #ifndef PERL_IN_XSUB_RE
5543 /* Initialize these here instead of as-needed, as is quick and avoids
5544 * having to test them each time otherwise */
5545 if (! PL_AboveLatin1) {
5546 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
5547 PL_ASCII = _new_invlist_C_array(ASCII_invlist);
5548 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
5550 PL_L1Posix_ptrs[_CC_ALPHANUMERIC]
5551 = _new_invlist_C_array(L1PosixAlnum_invlist);
5552 PL_Posix_ptrs[_CC_ALPHANUMERIC]
5553 = _new_invlist_C_array(PosixAlnum_invlist);
5555 PL_L1Posix_ptrs[_CC_ALPHA]
5556 = _new_invlist_C_array(L1PosixAlpha_invlist);
5557 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(PosixAlpha_invlist);
5559 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(PosixBlank_invlist);
5560 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(XPosixBlank_invlist);
5562 /* Cased is the same as Alpha in the ASCII range */
5563 PL_L1Posix_ptrs[_CC_CASED] = _new_invlist_C_array(L1Cased_invlist);
5564 PL_Posix_ptrs[_CC_CASED] = _new_invlist_C_array(PosixAlpha_invlist);
5566 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(PosixCntrl_invlist);
5567 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(XPosixCntrl_invlist);
5569 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5570 PL_L1Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5572 PL_L1Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(L1PosixGraph_invlist);
5573 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(PosixGraph_invlist);
5575 PL_L1Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(L1PosixLower_invlist);
5576 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(PosixLower_invlist);
5578 PL_L1Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(L1PosixPrint_invlist);
5579 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(PosixPrint_invlist);
5581 PL_L1Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(L1PosixPunct_invlist);
5582 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(PosixPunct_invlist);
5584 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(PerlSpace_invlist);
5585 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(XPerlSpace_invlist);
5586 PL_Posix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(PosixSpace_invlist);
5587 PL_XPosix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(XPosixSpace_invlist);
5589 PL_L1Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(L1PosixUpper_invlist);
5590 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(PosixUpper_invlist);
5592 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(VertSpace_invlist);
5594 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(PosixWord_invlist);
5595 PL_L1Posix_ptrs[_CC_WORDCHAR]
5596 = _new_invlist_C_array(L1PosixWord_invlist);
5598 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(PosixXDigit_invlist);
5599 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(XPosixXDigit_invlist);
5601 PL_HasMultiCharFold = _new_invlist_C_array(_Perl_Multi_Char_Folds_invlist);
5605 pRExC_state->code_blocks = NULL;
5606 pRExC_state->num_code_blocks = 0;
5609 *is_bare_re = FALSE;
5611 if (expr && (expr->op_type == OP_LIST ||
5612 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
5613 /* allocate code_blocks if needed */
5617 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling)
5618 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
5619 ncode++; /* count of DO blocks */
5621 pRExC_state->num_code_blocks = ncode;
5622 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
5627 /* compile-time pattern with just OP_CONSTs and DO blocks */
5632 /* find how many CONSTs there are */
5635 if (expr->op_type == OP_CONST)
5638 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5639 if (o->op_type == OP_CONST)
5643 /* fake up an SV array */
5645 assert(!new_patternp);
5646 Newx(new_patternp, n, SV*);
5647 SAVEFREEPV(new_patternp);
5651 if (expr->op_type == OP_CONST)
5652 new_patternp[n] = cSVOPx_sv(expr);
5654 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5655 if (o->op_type == OP_CONST)
5656 new_patternp[n++] = cSVOPo_sv;
5661 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5662 "Assembling pattern from %d elements%s\n", pat_count,
5663 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5665 /* set expr to the first arg op */
5667 if (pRExC_state->num_code_blocks
5668 && expr->op_type != OP_CONST)
5670 expr = cLISTOPx(expr)->op_first;
5671 assert( expr->op_type == OP_PUSHMARK
5672 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
5673 || expr->op_type == OP_PADRANGE);
5674 expr = expr->op_sibling;
5677 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
5678 expr, &recompile, NULL);
5680 /* handle bare (possibly after overloading) regex: foo =~ $re */
5685 if (SvTYPE(re) == SVt_REGEXP) {
5689 Safefree(pRExC_state->code_blocks);
5690 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5691 "Precompiled pattern%s\n",
5692 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5698 exp = SvPV_nomg(pat, plen);
5700 if (!eng->op_comp) {
5701 if ((SvUTF8(pat) && IN_BYTES)
5702 || SvGMAGICAL(pat) || SvAMAGIC(pat))
5704 /* make a temporary copy; either to convert to bytes,
5705 * or to avoid repeating get-magic / overloaded stringify */
5706 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
5707 (IN_BYTES ? 0 : SvUTF8(pat)));
5709 Safefree(pRExC_state->code_blocks);
5710 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
5713 /* ignore the utf8ness if the pattern is 0 length */
5714 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
5715 RExC_uni_semantics = 0;
5716 RExC_contains_locale = 0;
5717 pRExC_state->runtime_code_qr = NULL;
5720 SV *dsv= sv_newmortal();
5721 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
5722 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
5723 PL_colors[4],PL_colors[5],s);
5727 /* we jump here if we upgrade the pattern to utf8 and have to
5730 if ((pm_flags & PMf_USE_RE_EVAL)
5731 /* this second condition covers the non-regex literal case,
5732 * i.e. $foo =~ '(?{})'. */
5733 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
5735 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
5737 /* return old regex if pattern hasn't changed */
5738 /* XXX: note in the below we have to check the flags as well as the pattern.
5740 * Things get a touch tricky as we have to compare the utf8 flag independently
5741 * from the compile flags.
5746 && !!RX_UTF8(old_re) == !!RExC_utf8
5747 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
5748 && RX_PRECOMP(old_re)
5749 && RX_PRELEN(old_re) == plen
5750 && memEQ(RX_PRECOMP(old_re), exp, plen)
5751 && !runtime_code /* with runtime code, always recompile */ )
5753 Safefree(pRExC_state->code_blocks);
5757 rx_flags = orig_rx_flags;
5759 if (initial_charset == REGEX_LOCALE_CHARSET) {
5760 RExC_contains_locale = 1;
5762 else if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
5764 /* Set to use unicode semantics if the pattern is in utf8 and has the
5765 * 'depends' charset specified, as it means unicode when utf8 */
5766 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5770 RExC_flags = rx_flags;
5771 RExC_pm_flags = pm_flags;
5774 if (TAINTING_get && TAINT_get)
5775 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
5777 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
5778 /* whoops, we have a non-utf8 pattern, whilst run-time code
5779 * got compiled as utf8. Try again with a utf8 pattern */
5780 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
5781 pRExC_state->num_code_blocks);
5782 goto redo_first_pass;
5785 assert(!pRExC_state->runtime_code_qr);
5790 RExC_in_lookbehind = 0;
5791 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
5793 RExC_override_recoding = 0;
5794 RExC_in_multi_char_class = 0;
5796 /* First pass: determine size, legality. */
5799 RExC_end = exp + plen;
5804 RExC_emit = &RExC_emit_dummy;
5805 RExC_whilem_seen = 0;
5806 RExC_open_parens = NULL;
5807 RExC_close_parens = NULL;
5809 RExC_paren_names = NULL;
5811 RExC_paren_name_list = NULL;
5813 RExC_recurse = NULL;
5814 RExC_recurse_count = 0;
5815 pRExC_state->code_index = 0;
5817 #if 0 /* REGC() is (currently) a NOP at the first pass.
5818 * Clever compilers notice this and complain. --jhi */
5819 REGC((U8)REG_MAGIC, (char*)RExC_emit);
5822 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
5824 RExC_lastparse=NULL;
5826 /* reg may croak on us, not giving us a chance to free
5827 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
5828 need it to survive as long as the regexp (qr/(?{})/).
5829 We must check that code_blocksv is not already set, because we may
5830 have jumped back to restart the sizing pass. */
5831 if (pRExC_state->code_blocks && !code_blocksv) {
5832 code_blocksv = newSV_type(SVt_PV);
5833 SAVEFREESV(code_blocksv);
5834 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
5835 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
5837 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5838 /* It's possible to write a regexp in ascii that represents Unicode
5839 codepoints outside of the byte range, such as via \x{100}. If we
5840 detect such a sequence we have to convert the entire pattern to utf8
5841 and then recompile, as our sizing calculation will have been based
5842 on 1 byte == 1 character, but we will need to use utf8 to encode
5843 at least some part of the pattern, and therefore must convert the whole
5846 if (flags & RESTART_UTF8) {
5847 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
5848 pRExC_state->num_code_blocks);
5849 goto redo_first_pass;
5851 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
5854 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
5857 PerlIO_printf(Perl_debug_log,
5858 "Required size %"IVdf" nodes\n"
5859 "Starting second pass (creation)\n",
5862 RExC_lastparse=NULL;
5865 /* The first pass could have found things that force Unicode semantics */
5866 if ((RExC_utf8 || RExC_uni_semantics)
5867 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
5869 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5872 /* Small enough for pointer-storage convention?
5873 If extralen==0, this means that we will not need long jumps. */
5874 if (RExC_size >= 0x10000L && RExC_extralen)
5875 RExC_size += RExC_extralen;
5878 if (RExC_whilem_seen > 15)
5879 RExC_whilem_seen = 15;
5881 /* Allocate space and zero-initialize. Note, the two step process
5882 of zeroing when in debug mode, thus anything assigned has to
5883 happen after that */
5884 rx = (REGEXP*) newSV_type(SVt_REGEXP);
5886 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
5887 char, regexp_internal);
5888 if ( r == NULL || ri == NULL )
5889 FAIL("Regexp out of space");
5891 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
5892 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode), char);
5894 /* bulk initialize base fields with 0. */
5895 Zero(ri, sizeof(regexp_internal), char);
5898 /* non-zero initialization begins here */
5901 r->extflags = rx_flags;
5902 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
5904 if (pm_flags & PMf_IS_QR) {
5905 ri->code_blocks = pRExC_state->code_blocks;
5906 ri->num_code_blocks = pRExC_state->num_code_blocks;
5911 for (n = 0; n < pRExC_state->num_code_blocks; n++)
5912 if (pRExC_state->code_blocks[n].src_regex)
5913 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
5914 SAVEFREEPV(pRExC_state->code_blocks);
5918 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
5919 bool has_charset = (get_regex_charset(r->extflags) != REGEX_DEPENDS_CHARSET);
5921 /* The caret is output if there are any defaults: if not all the STD
5922 * flags are set, or if no character set specifier is needed */
5924 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
5926 bool has_runon = ((RExC_seen & REG_SEEN_RUN_ON_COMMENT)==REG_SEEN_RUN_ON_COMMENT);
5927 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
5928 >> RXf_PMf_STD_PMMOD_SHIFT);
5929 const char *fptr = STD_PAT_MODS; /*"msix"*/
5931 /* Allocate for the worst case, which is all the std flags are turned
5932 * on. If more precision is desired, we could do a population count of
5933 * the flags set. This could be done with a small lookup table, or by
5934 * shifting, masking and adding, or even, when available, assembly
5935 * language for a machine-language population count.
5936 * We never output a minus, as all those are defaults, so are
5937 * covered by the caret */
5938 const STRLEN wraplen = plen + has_p + has_runon
5939 + has_default /* If needs a caret */
5941 /* If needs a character set specifier */
5942 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
5943 + (sizeof(STD_PAT_MODS) - 1)
5944 + (sizeof("(?:)") - 1);
5946 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
5947 r->xpv_len_u.xpvlenu_pv = p;
5949 SvFLAGS(rx) |= SVf_UTF8;
5952 /* If a default, cover it using the caret */
5954 *p++= DEFAULT_PAT_MOD;
5958 const char* const name = get_regex_charset_name(r->extflags, &len);
5959 Copy(name, p, len, char);
5963 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
5966 while((ch = *fptr++)) {
5974 Copy(RExC_precomp, p, plen, char);
5975 assert ((RX_WRAPPED(rx) - p) < 16);
5976 r->pre_prefix = p - RX_WRAPPED(rx);
5982 SvCUR_set(rx, p - RX_WRAPPED(rx));
5986 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
5988 if (RExC_seen & REG_SEEN_RECURSE) {
5989 Newxz(RExC_open_parens, RExC_npar,regnode *);
5990 SAVEFREEPV(RExC_open_parens);
5991 Newxz(RExC_close_parens,RExC_npar,regnode *);
5992 SAVEFREEPV(RExC_close_parens);
5995 /* Useful during FAIL. */
5996 #ifdef RE_TRACK_PATTERN_OFFSETS
5997 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
5998 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
5999 "%s %"UVuf" bytes for offset annotations.\n",
6000 ri->u.offsets ? "Got" : "Couldn't get",
6001 (UV)((2*RExC_size+1) * sizeof(U32))));
6003 SetProgLen(ri,RExC_size);
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 = sawminmod = 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.
6103 * (yves doesn't think this is true)
6105 if (OP(first) == PLUS)
6108 if (OP(first) == MINMOD)
6110 first += regarglen[OP(first)];
6112 first = NEXTOPER(first);
6113 first_next= regnext(first);
6116 /* Starting-point info. */
6118 DEBUG_PEEP("first:",first,0);
6119 /* Ignore EXACT as we deal with it later. */
6120 if (PL_regkind[OP(first)] == EXACT) {
6121 if (OP(first) == EXACT)
6122 NOOP; /* Empty, get anchored substr later. */
6124 ri->regstclass = first;
6127 else if (PL_regkind[OP(first)] == TRIE &&
6128 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6131 /* this can happen only on restudy */
6132 if ( OP(first) == TRIE ) {
6133 struct regnode_1 *trieop = (struct regnode_1 *)
6134 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6135 StructCopy(first,trieop,struct regnode_1);
6136 trie_op=(regnode *)trieop;
6138 struct regnode_charclass *trieop = (struct regnode_charclass *)
6139 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6140 StructCopy(first,trieop,struct regnode_charclass);
6141 trie_op=(regnode *)trieop;
6144 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6145 ri->regstclass = trie_op;
6148 else if (REGNODE_SIMPLE(OP(first)))
6149 ri->regstclass = first;
6150 else if (PL_regkind[OP(first)] == BOUND ||
6151 PL_regkind[OP(first)] == NBOUND)
6152 ri->regstclass = first;
6153 else if (PL_regkind[OP(first)] == BOL) {
6154 r->extflags |= (OP(first) == MBOL
6156 : (OP(first) == SBOL
6159 first = NEXTOPER(first);
6162 else if (OP(first) == GPOS) {
6163 r->extflags |= RXf_ANCH_GPOS;
6164 first = NEXTOPER(first);
6167 else if ((!sawopen || !RExC_sawback) &&
6168 (OP(first) == STAR &&
6169 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6170 !(r->extflags & RXf_ANCH) && !pRExC_state->num_code_blocks)
6172 /* turn .* into ^.* with an implied $*=1 */
6174 (OP(NEXTOPER(first)) == REG_ANY)
6177 r->extflags |= type;
6178 r->intflags |= PREGf_IMPLICIT;
6179 first = NEXTOPER(first);
6182 if (sawplus && !sawminmod && !sawlookahead && (!sawopen || !RExC_sawback)
6183 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6184 /* x+ must match at the 1st pos of run of x's */
6185 r->intflags |= PREGf_SKIP;
6187 /* Scan is after the zeroth branch, first is atomic matcher. */
6188 #ifdef TRIE_STUDY_OPT
6191 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6192 (IV)(first - scan + 1))
6196 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6197 (IV)(first - scan + 1))
6203 * If there's something expensive in the r.e., find the
6204 * longest literal string that must appear and make it the
6205 * regmust. Resolve ties in favor of later strings, since
6206 * the regstart check works with the beginning of the r.e.
6207 * and avoiding duplication strengthens checking. Not a
6208 * strong reason, but sufficient in the absence of others.
6209 * [Now we resolve ties in favor of the earlier string if
6210 * it happens that c_offset_min has been invalidated, since the
6211 * earlier string may buy us something the later one won't.]
6214 data.longest_fixed = newSVpvs("");
6215 data.longest_float = newSVpvs("");
6216 data.last_found = newSVpvs("");
6217 data.longest = &(data.longest_fixed);
6218 ENTER_with_name("study_chunk");
6219 SAVEFREESV(data.longest_fixed);
6220 SAVEFREESV(data.longest_float);
6221 SAVEFREESV(data.last_found);
6223 if (!ri->regstclass) {
6224 cl_init(pRExC_state, &ch_class);
6225 data.start_class = &ch_class;
6226 stclass_flag = SCF_DO_STCLASS_AND;
6227 } else /* XXXX Check for BOUND? */
6229 data.last_closep = &last_close;
6231 minlen = study_chunk(pRExC_state, &first, &minlen, &fake, scan + RExC_size, /* Up to end */
6232 &data, -1, NULL, NULL,
6233 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
6234 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
6238 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6241 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6242 && data.last_start_min == 0 && data.last_end > 0
6243 && !RExC_seen_zerolen
6244 && !(RExC_seen & REG_SEEN_VERBARG)
6245 && (!(RExC_seen & REG_SEEN_GPOS) || (r->extflags & RXf_ANCH_GPOS)))
6246 r->extflags |= RXf_CHECK_ALL;
6247 scan_commit(pRExC_state, &data,&minlen,0);
6249 longest_float_length = CHR_SVLEN(data.longest_float);
6251 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6252 && data.offset_fixed == data.offset_float_min
6253 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6254 && S_setup_longest (aTHX_ pRExC_state,
6258 &(r->float_end_shift),
6259 data.lookbehind_float,
6260 data.offset_float_min,
6262 longest_float_length,
6263 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6264 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6266 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6267 r->float_max_offset = data.offset_float_max;
6268 if (data.offset_float_max < I32_MAX) /* Don't offset infinity */
6269 r->float_max_offset -= data.lookbehind_float;
6270 SvREFCNT_inc_simple_void_NN(data.longest_float);
6273 r->float_substr = r->float_utf8 = NULL;
6274 longest_float_length = 0;
6277 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6279 if (S_setup_longest (aTHX_ pRExC_state,
6281 &(r->anchored_utf8),
6282 &(r->anchored_substr),
6283 &(r->anchored_end_shift),
6284 data.lookbehind_fixed,
6287 longest_fixed_length,
6288 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6289 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6291 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6292 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6295 r->anchored_substr = r->anchored_utf8 = NULL;
6296 longest_fixed_length = 0;
6298 LEAVE_with_name("study_chunk");
6301 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6302 ri->regstclass = NULL;
6304 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6306 && ! TEST_SSC_EOS(data.start_class)
6307 && !cl_is_anything(data.start_class))
6309 const U32 n = add_data(pRExC_state, 1, "f");
6310 OP(data.start_class) = ANYOF_SYNTHETIC;
6312 Newx(RExC_rxi->data->data[n], 1,
6313 struct regnode_charclass_class);
6314 StructCopy(data.start_class,
6315 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6316 struct regnode_charclass_class);
6317 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6318 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6319 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6320 regprop(r, sv, (regnode*)data.start_class);
6321 PerlIO_printf(Perl_debug_log,
6322 "synthetic stclass \"%s\".\n",
6323 SvPVX_const(sv));});
6326 /* A temporary algorithm prefers floated substr to fixed one to dig more info. */
6327 if (longest_fixed_length > longest_float_length) {
6328 r->check_end_shift = r->anchored_end_shift;
6329 r->check_substr = r->anchored_substr;
6330 r->check_utf8 = r->anchored_utf8;
6331 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6332 if (r->extflags & RXf_ANCH_SINGLE)
6333 r->extflags |= RXf_NOSCAN;
6336 r->check_end_shift = r->float_end_shift;
6337 r->check_substr = r->float_substr;
6338 r->check_utf8 = r->float_utf8;
6339 r->check_offset_min = r->float_min_offset;
6340 r->check_offset_max = r->float_max_offset;
6342 /* XXXX Currently intuiting is not compatible with ANCH_GPOS.
6343 This should be changed ASAP! */
6344 if ((r->check_substr || r->check_utf8) && !(r->extflags & RXf_ANCH_GPOS)) {
6345 r->extflags |= RXf_USE_INTUIT;
6346 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6347 r->extflags |= RXf_INTUIT_TAIL;
6349 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6350 if ( (STRLEN)minlen < longest_float_length )
6351 minlen= longest_float_length;
6352 if ( (STRLEN)minlen < longest_fixed_length )
6353 minlen= longest_fixed_length;
6357 /* Several toplevels. Best we can is to set minlen. */
6359 struct regnode_charclass_class ch_class;
6362 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6364 scan = ri->program + 1;
6365 cl_init(pRExC_state, &ch_class);
6366 data.start_class = &ch_class;
6367 data.last_closep = &last_close;
6370 minlen = study_chunk(pRExC_state, &scan, &minlen, &fake, scan + RExC_size,
6371 &data, -1, NULL, NULL,
6372 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS
6373 |(restudied ? SCF_TRIE_DOING_RESTUDY : 0),
6376 CHECK_RESTUDY_GOTO_butfirst(NOOP);
6378 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
6379 = r->float_substr = r->float_utf8 = NULL;
6381 if (! TEST_SSC_EOS(data.start_class)
6382 && !cl_is_anything(data.start_class))
6384 const U32 n = add_data(pRExC_state, 1, "f");
6385 OP(data.start_class) = ANYOF_SYNTHETIC;
6387 Newx(RExC_rxi->data->data[n], 1,
6388 struct regnode_charclass_class);
6389 StructCopy(data.start_class,
6390 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6391 struct regnode_charclass_class);
6392 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6393 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6394 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
6395 regprop(r, sv, (regnode*)data.start_class);
6396 PerlIO_printf(Perl_debug_log,
6397 "synthetic stclass \"%s\".\n",
6398 SvPVX_const(sv));});
6402 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
6403 the "real" pattern. */
6405 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf"\n",
6406 (IV)minlen, (IV)r->minlen);
6408 r->minlenret = minlen;
6409 if (r->minlen < minlen)
6412 if (RExC_seen & REG_SEEN_GPOS)
6413 r->extflags |= RXf_GPOS_SEEN;
6414 if (RExC_seen & REG_SEEN_LOOKBEHIND)
6415 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the lookbehind */
6416 if (pRExC_state->num_code_blocks)
6417 r->extflags |= RXf_EVAL_SEEN;
6418 if (RExC_seen & REG_SEEN_CANY)
6419 r->extflags |= RXf_CANY_SEEN;
6420 if (RExC_seen & REG_SEEN_VERBARG)
6422 r->intflags |= PREGf_VERBARG_SEEN;
6423 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
6425 if (RExC_seen & REG_SEEN_CUTGROUP)
6426 r->intflags |= PREGf_CUTGROUP_SEEN;
6427 if (pm_flags & PMf_USE_RE_EVAL)
6428 r->intflags |= PREGf_USE_RE_EVAL;
6429 if (RExC_paren_names)
6430 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
6432 RXp_PAREN_NAMES(r) = NULL;
6435 regnode *first = ri->program + 1;
6437 regnode *next = NEXTOPER(first);
6440 if (PL_regkind[fop] == NOTHING && nop == END)
6441 r->extflags |= RXf_NULL;
6442 else if (PL_regkind[fop] == BOL && nop == END)
6443 r->extflags |= RXf_START_ONLY;
6444 else if (fop == PLUS && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE && OP(regnext(first)) == END)
6445 r->extflags |= RXf_WHITE;
6446 else if ( r->extflags & RXf_SPLIT && fop == EXACT && STR_LEN(first) == 1 && *(STRING(first)) == ' ' && OP(regnext(first)) == END )
6447 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
6451 if (RExC_paren_names) {
6452 ri->name_list_idx = add_data( pRExC_state, 1, "a" );
6453 ri->data->data[ri->name_list_idx] = (void*)SvREFCNT_inc(RExC_paren_name_list);
6456 ri->name_list_idx = 0;
6458 if (RExC_recurse_count) {
6459 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
6460 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
6461 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
6464 Newxz(r->offs, RExC_npar, regexp_paren_pair);
6465 /* assume we don't need to swap parens around before we match */
6468 PerlIO_printf(Perl_debug_log,"Final program:\n");
6471 #ifdef RE_TRACK_PATTERN_OFFSETS
6472 DEBUG_OFFSETS_r(if (ri->u.offsets) {
6473 const U32 len = ri->u.offsets[0];
6475 GET_RE_DEBUG_FLAGS_DECL;
6476 PerlIO_printf(Perl_debug_log, "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
6477 for (i = 1; i <= len; i++) {
6478 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
6479 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
6480 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
6482 PerlIO_printf(Perl_debug_log, "\n");
6487 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
6488 * by setting the regexp SV to readonly-only instead. If the
6489 * pattern's been recompiled, the USEDness should remain. */
6490 if (old_re && SvREADONLY(old_re))
6498 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
6501 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
6503 PERL_UNUSED_ARG(value);
6505 if (flags & RXapif_FETCH) {
6506 return reg_named_buff_fetch(rx, key, flags);
6507 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
6508 Perl_croak_no_modify();
6510 } else if (flags & RXapif_EXISTS) {
6511 return reg_named_buff_exists(rx, key, flags)
6514 } else if (flags & RXapif_REGNAMES) {
6515 return reg_named_buff_all(rx, flags);
6516 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
6517 return reg_named_buff_scalar(rx, flags);
6519 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
6525 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
6528 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
6529 PERL_UNUSED_ARG(lastkey);
6531 if (flags & RXapif_FIRSTKEY)
6532 return reg_named_buff_firstkey(rx, flags);
6533 else if (flags & RXapif_NEXTKEY)
6534 return reg_named_buff_nextkey(rx, flags);
6536 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter", (int)flags);
6542 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
6545 AV *retarray = NULL;
6547 struct regexp *const rx = ReANY(r);
6549 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
6551 if (flags & RXapif_ALL)
6554 if (rx && RXp_PAREN_NAMES(rx)) {
6555 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
6558 SV* sv_dat=HeVAL(he_str);
6559 I32 *nums=(I32*)SvPVX(sv_dat);
6560 for ( i=0; i<SvIVX(sv_dat); i++ ) {
6561 if ((I32)(rx->nparens) >= nums[i]
6562 && rx->offs[nums[i]].start != -1
6563 && rx->offs[nums[i]].end != -1)
6566 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
6571 ret = newSVsv(&PL_sv_undef);
6574 av_push(retarray, ret);
6577 return newRV_noinc(MUTABLE_SV(retarray));
6584 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
6587 struct regexp *const rx = ReANY(r);
6589 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
6591 if (rx && RXp_PAREN_NAMES(rx)) {
6592 if (flags & RXapif_ALL) {
6593 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
6595 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
6597 SvREFCNT_dec_NN(sv);
6609 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
6611 struct regexp *const rx = ReANY(r);
6613 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
6615 if ( rx && RXp_PAREN_NAMES(rx) ) {
6616 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
6618 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
6625 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
6627 struct regexp *const rx = ReANY(r);
6628 GET_RE_DEBUG_FLAGS_DECL;
6630 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
6632 if (rx && RXp_PAREN_NAMES(rx)) {
6633 HV *hv = RXp_PAREN_NAMES(rx);
6635 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6638 SV* sv_dat = HeVAL(temphe);
6639 I32 *nums = (I32*)SvPVX(sv_dat);
6640 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6641 if ((I32)(rx->lastparen) >= nums[i] &&
6642 rx->offs[nums[i]].start != -1 &&
6643 rx->offs[nums[i]].end != -1)
6649 if (parno || flags & RXapif_ALL) {
6650 return newSVhek(HeKEY_hek(temphe));
6658 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
6663 struct regexp *const rx = ReANY(r);
6665 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
6667 if (rx && RXp_PAREN_NAMES(rx)) {
6668 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
6669 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
6670 } else if (flags & RXapif_ONE) {
6671 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
6672 av = MUTABLE_AV(SvRV(ret));
6673 length = av_len(av);
6674 SvREFCNT_dec_NN(ret);
6675 return newSViv(length + 1);
6677 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar", (int)flags);
6681 return &PL_sv_undef;
6685 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
6687 struct regexp *const rx = ReANY(r);
6690 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
6692 if (rx && RXp_PAREN_NAMES(rx)) {
6693 HV *hv= RXp_PAREN_NAMES(rx);
6695 (void)hv_iterinit(hv);
6696 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6699 SV* sv_dat = HeVAL(temphe);
6700 I32 *nums = (I32*)SvPVX(sv_dat);
6701 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6702 if ((I32)(rx->lastparen) >= nums[i] &&
6703 rx->offs[nums[i]].start != -1 &&
6704 rx->offs[nums[i]].end != -1)
6710 if (parno || flags & RXapif_ALL) {
6711 av_push(av, newSVhek(HeKEY_hek(temphe)));
6716 return newRV_noinc(MUTABLE_SV(av));
6720 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
6723 struct regexp *const rx = ReANY(r);
6729 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
6731 if ( ( n == RX_BUFF_IDX_CARET_PREMATCH
6732 || n == RX_BUFF_IDX_CARET_FULLMATCH
6733 || n == RX_BUFF_IDX_CARET_POSTMATCH
6735 && !(rx->extflags & RXf_PMf_KEEPCOPY)
6742 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
6743 /* no need to distinguish between them any more */
6744 n = RX_BUFF_IDX_FULLMATCH;
6746 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
6747 && rx->offs[0].start != -1)
6749 /* $`, ${^PREMATCH} */
6750 i = rx->offs[0].start;
6754 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
6755 && rx->offs[0].end != -1)
6757 /* $', ${^POSTMATCH} */
6758 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
6759 i = rx->sublen + rx->suboffset - rx->offs[0].end;
6762 if ( 0 <= n && n <= (I32)rx->nparens &&
6763 (s1 = rx->offs[n].start) != -1 &&
6764 (t1 = rx->offs[n].end) != -1)
6766 /* $&, ${^MATCH}, $1 ... */
6768 s = rx->subbeg + s1 - rx->suboffset;
6773 assert(s >= rx->subbeg);
6774 assert(rx->sublen >= (s - rx->subbeg) + i );
6776 #if NO_TAINT_SUPPORT
6777 sv_setpvn(sv, s, i);
6779 const int oldtainted = TAINT_get;
6781 sv_setpvn(sv, s, i);
6782 TAINT_set(oldtainted);
6784 if ( (rx->extflags & RXf_CANY_SEEN)
6785 ? (RXp_MATCH_UTF8(rx)
6786 && (!i || is_utf8_string((U8*)s, i)))
6787 : (RXp_MATCH_UTF8(rx)) )
6794 if (RXp_MATCH_TAINTED(rx)) {
6795 if (SvTYPE(sv) >= SVt_PVMG) {
6796 MAGIC* const mg = SvMAGIC(sv);
6799 SvMAGIC_set(sv, mg->mg_moremagic);
6801 if ((mgt = SvMAGIC(sv))) {
6802 mg->mg_moremagic = mgt;
6803 SvMAGIC_set(sv, mg);
6814 sv_setsv(sv,&PL_sv_undef);
6820 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
6821 SV const * const value)
6823 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
6825 PERL_UNUSED_ARG(rx);
6826 PERL_UNUSED_ARG(paren);
6827 PERL_UNUSED_ARG(value);
6830 Perl_croak_no_modify();
6834 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
6837 struct regexp *const rx = ReANY(r);
6841 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
6843 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
6845 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
6846 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6850 case RX_BUFF_IDX_PREMATCH: /* $` */
6851 if (rx->offs[0].start != -1) {
6852 i = rx->offs[0].start;
6861 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
6862 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6864 case RX_BUFF_IDX_POSTMATCH: /* $' */
6865 if (rx->offs[0].end != -1) {
6866 i = rx->sublen - rx->offs[0].end;
6868 s1 = rx->offs[0].end;
6875 case RX_BUFF_IDX_CARET_FULLMATCH: /* ${^MATCH} */
6876 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6880 /* $& / ${^MATCH}, $1, $2, ... */
6882 if (paren <= (I32)rx->nparens &&
6883 (s1 = rx->offs[paren].start) != -1 &&
6884 (t1 = rx->offs[paren].end) != -1)
6890 if (ckWARN(WARN_UNINITIALIZED))
6891 report_uninit((const SV *)sv);
6896 if (i > 0 && RXp_MATCH_UTF8(rx)) {
6897 const char * const s = rx->subbeg - rx->suboffset + s1;
6902 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
6909 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
6911 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
6912 PERL_UNUSED_ARG(rx);
6916 return newSVpvs("Regexp");
6919 /* Scans the name of a named buffer from the pattern.
6920 * If flags is REG_RSN_RETURN_NULL returns null.
6921 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
6922 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
6923 * to the parsed name as looked up in the RExC_paren_names hash.
6924 * If there is an error throws a vFAIL().. type exception.
6927 #define REG_RSN_RETURN_NULL 0
6928 #define REG_RSN_RETURN_NAME 1
6929 #define REG_RSN_RETURN_DATA 2
6932 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
6934 char *name_start = RExC_parse;
6936 PERL_ARGS_ASSERT_REG_SCAN_NAME;
6938 if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
6939 /* skip IDFIRST by using do...while */
6942 RExC_parse += UTF8SKIP(RExC_parse);
6943 } while (isWORDCHAR_utf8((U8*)RExC_parse));
6947 } while (isWORDCHAR(*RExC_parse));
6949 RExC_parse++; /* so the <- from the vFAIL is after the offending character */
6950 vFAIL("Group name must start with a non-digit word character");
6954 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
6955 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
6956 if ( flags == REG_RSN_RETURN_NAME)
6958 else if (flags==REG_RSN_RETURN_DATA) {
6961 if ( ! sv_name ) /* should not happen*/
6962 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
6963 if (RExC_paren_names)
6964 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
6966 sv_dat = HeVAL(he_str);
6968 vFAIL("Reference to nonexistent named group");
6972 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
6973 (unsigned long) flags);
6975 assert(0); /* NOT REACHED */
6980 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
6981 int rem=(int)(RExC_end - RExC_parse); \
6990 if (RExC_lastparse!=RExC_parse) \
6991 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
6994 iscut ? "..." : "<" \
6997 PerlIO_printf(Perl_debug_log,"%16s",""); \
7000 num = RExC_size + 1; \
7002 num=REG_NODE_NUM(RExC_emit); \
7003 if (RExC_lastnum!=num) \
7004 PerlIO_printf(Perl_debug_log,"|%4d",num); \
7006 PerlIO_printf(Perl_debug_log,"|%4s",""); \
7007 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
7008 (int)((depth*2)), "", \
7012 RExC_lastparse=RExC_parse; \
7017 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
7018 DEBUG_PARSE_MSG((funcname)); \
7019 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
7021 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7022 DEBUG_PARSE_MSG((funcname)); \
7023 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7026 /* This section of code defines the inversion list object and its methods. The
7027 * interfaces are highly subject to change, so as much as possible is static to
7028 * this file. An inversion list is here implemented as a malloc'd C UV array
7029 * with some added info that is placed as UVs at the beginning in a header
7030 * portion. An inversion list for Unicode is an array of code points, sorted
7031 * by ordinal number. The zeroth element is the first code point in the list.
7032 * The 1th element is the first element beyond that not in the list. In other
7033 * words, the first range is
7034 * invlist[0]..(invlist[1]-1)
7035 * The other ranges follow. Thus every element whose index is divisible by two
7036 * marks the beginning of a range that is in the list, and every element not
7037 * divisible by two marks the beginning of a range not in the list. A single
7038 * element inversion list that contains the single code point N generally
7039 * consists of two elements
7042 * (The exception is when N is the highest representable value on the
7043 * machine, in which case the list containing just it would be a single
7044 * element, itself. By extension, if the last range in the list extends to
7045 * infinity, then the first element of that range will be in the inversion list
7046 * at a position that is divisible by two, and is the final element in the
7048 * Taking the complement (inverting) an inversion list is quite simple, if the
7049 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7050 * This implementation reserves an element at the beginning of each inversion
7051 * list to contain 0 when the list contains 0, and contains 1 otherwise. The
7052 * actual beginning of the list is either that element if 0, or the next one if
7055 * More about inversion lists can be found in "Unicode Demystified"
7056 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7057 * More will be coming when functionality is added later.
7059 * The inversion list data structure is currently implemented as an SV pointing
7060 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7061 * array of UV whose memory management is automatically handled by the existing
7062 * facilities for SV's.
7064 * Some of the methods should always be private to the implementation, and some
7065 * should eventually be made public */
7067 /* The header definitions are in F<inline_invlist.c> */
7068 #define TO_INTERNAL_SIZE(x) (((x) + HEADER_LENGTH) * sizeof(UV))
7069 #define FROM_INTERNAL_SIZE(x) (((x)/ sizeof(UV)) - HEADER_LENGTH)
7071 #define INVLIST_INITIAL_LEN 10
7073 PERL_STATIC_INLINE UV*
7074 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
7076 /* Returns a pointer to the first element in the inversion list's array.
7077 * This is called upon initialization of an inversion list. Where the
7078 * array begins depends on whether the list has the code point U+0000
7079 * in it or not. The other parameter tells it whether the code that
7080 * follows this call is about to put a 0 in the inversion list or not.
7081 * The first element is either the element with 0, if 0, or the next one,
7084 UV* zero = get_invlist_zero_addr(invlist);
7086 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7089 assert(! *_get_invlist_len_addr(invlist));
7091 /* 1^1 = 0; 1^0 = 1 */
7092 *zero = 1 ^ will_have_0;
7093 return zero + *zero;
7096 PERL_STATIC_INLINE UV*
7097 S_invlist_array(pTHX_ SV* const invlist)
7099 /* Returns the pointer to the inversion list's array. Every time the
7100 * length changes, this needs to be called in case malloc or realloc moved
7103 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7105 /* Must not be empty. If these fail, you probably didn't check for <len>
7106 * being non-zero before trying to get the array */
7107 assert(*_get_invlist_len_addr(invlist));
7108 assert(*get_invlist_zero_addr(invlist) == 0
7109 || *get_invlist_zero_addr(invlist) == 1);
7111 /* The array begins either at the element reserved for zero if the
7112 * list contains 0 (that element will be set to 0), or otherwise the next
7113 * element (in which case the reserved element will be set to 1). */
7114 return (UV *) (get_invlist_zero_addr(invlist)
7115 + *get_invlist_zero_addr(invlist));
7118 PERL_STATIC_INLINE void
7119 S_invlist_set_len(pTHX_ SV* const invlist, const UV len)
7121 /* Sets the current number of elements stored in the inversion list */
7123 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7125 *_get_invlist_len_addr(invlist) = len;
7127 assert(len <= SvLEN(invlist));
7129 SvCUR_set(invlist, TO_INTERNAL_SIZE(len));
7130 /* If the list contains U+0000, that element is part of the header,
7131 * and should not be counted as part of the array. It will contain
7132 * 0 in that case, and 1 otherwise. So we could flop 0=>1, 1=>0 and
7134 * SvCUR_set(invlist,
7135 * TO_INTERNAL_SIZE(len
7136 * - (*get_invlist_zero_addr(inv_list) ^ 1)));
7137 * But, this is only valid if len is not 0. The consequences of not doing
7138 * this is that the memory allocation code may think that 1 more UV is
7139 * being used than actually is, and so might do an unnecessary grow. That
7140 * seems worth not bothering to make this the precise amount.
7142 * Note that when inverting, SvCUR shouldn't change */
7145 PERL_STATIC_INLINE IV*
7146 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7148 /* Return the address of the UV that is reserved to hold the cached index
7151 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7153 return (IV *) (SvPVX(invlist) + (INVLIST_PREVIOUS_INDEX_OFFSET * sizeof (UV)));
7156 PERL_STATIC_INLINE IV
7157 S_invlist_previous_index(pTHX_ SV* const invlist)
7159 /* Returns cached index of previous search */
7161 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7163 return *get_invlist_previous_index_addr(invlist);
7166 PERL_STATIC_INLINE void
7167 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7169 /* Caches <index> for later retrieval */
7171 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7173 assert(index == 0 || index < (int) _invlist_len(invlist));
7175 *get_invlist_previous_index_addr(invlist) = index;
7178 PERL_STATIC_INLINE UV
7179 S_invlist_max(pTHX_ SV* const invlist)
7181 /* Returns the maximum number of elements storable in the inversion list's
7182 * array, without having to realloc() */
7184 PERL_ARGS_ASSERT_INVLIST_MAX;
7186 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7187 ? _invlist_len(invlist)
7188 : FROM_INTERNAL_SIZE(SvLEN(invlist));
7191 PERL_STATIC_INLINE UV*
7192 S_get_invlist_zero_addr(pTHX_ SV* invlist)
7194 /* Return the address of the UV that is reserved to hold 0 if the inversion
7195 * list contains 0. This has to be the last element of the heading, as the
7196 * list proper starts with either it if 0, or the next element if not.
7197 * (But we force it to contain either 0 or 1) */
7199 PERL_ARGS_ASSERT_GET_INVLIST_ZERO_ADDR;
7201 return (UV *) (SvPVX(invlist) + (INVLIST_ZERO_OFFSET * sizeof (UV)));
7204 #ifndef PERL_IN_XSUB_RE
7206 Perl__new_invlist(pTHX_ IV initial_size)
7209 /* Return a pointer to a newly constructed inversion list, with enough
7210 * space to store 'initial_size' elements. If that number is negative, a
7211 * system default is used instead */
7215 if (initial_size < 0) {
7216 initial_size = INVLIST_INITIAL_LEN;
7219 /* Allocate the initial space */
7220 new_list = newSV(TO_INTERNAL_SIZE(initial_size));
7221 invlist_set_len(new_list, 0);
7223 /* Force iterinit() to be used to get iteration to work */
7224 *get_invlist_iter_addr(new_list) = UV_MAX;
7226 /* This should force a segfault if a method doesn't initialize this
7228 *get_invlist_zero_addr(new_list) = UV_MAX;
7230 *get_invlist_previous_index_addr(new_list) = 0;
7231 *get_invlist_version_id_addr(new_list) = INVLIST_VERSION_ID;
7232 #if HEADER_LENGTH != 5
7233 # 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
7241 S__new_invlist_C_array(pTHX_ UV* list)
7243 /* Return a pointer to a newly constructed inversion list, initialized to
7244 * point to <list>, which has to be in the exact correct inversion list
7245 * form, including internal fields. Thus this is a dangerous routine that
7246 * should not be used in the wrong hands */
7248 SV* invlist = newSV_type(SVt_PV);
7250 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7252 SvPV_set(invlist, (char *) list);
7253 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7254 shouldn't touch it */
7255 SvCUR_set(invlist, TO_INTERNAL_SIZE(_invlist_len(invlist)));
7257 if (*get_invlist_version_id_addr(invlist) != INVLIST_VERSION_ID) {
7258 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7261 /* Initialize the iteration pointer.
7262 * XXX This could be done at compile time in charclass_invlists.h, but I
7263 * (khw) am not confident that the suffixes for specifying the C constant
7264 * UV_MAX are portable, e.g. 'ull' on a 32 bit machine that is configured
7265 * to use 64 bits; might need a Configure probe */
7266 invlist_iterfinish(invlist);
7272 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7274 /* Grow the maximum size of an inversion list */
7276 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7278 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max));
7281 PERL_STATIC_INLINE void
7282 S_invlist_trim(pTHX_ SV* const invlist)
7284 PERL_ARGS_ASSERT_INVLIST_TRIM;
7286 /* Change the length of the inversion list to how many entries it currently
7289 SvPV_shrink_to_cur((SV *) invlist);
7292 #define _invlist_union_complement_2nd(a, b, output) _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
7295 S__append_range_to_invlist(pTHX_ SV* const invlist, const UV start, const UV end)
7297 /* Subject to change or removal. Append the range from 'start' to 'end' at
7298 * the end of the inversion list. The range must be above any existing
7302 UV max = invlist_max(invlist);
7303 UV len = _invlist_len(invlist);
7305 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7307 if (len == 0) { /* Empty lists must be initialized */
7308 array = _invlist_array_init(invlist, start == 0);
7311 /* Here, the existing list is non-empty. The current max entry in the
7312 * list is generally the first value not in the set, except when the
7313 * set extends to the end of permissible values, in which case it is
7314 * the first entry in that final set, and so this call is an attempt to
7315 * append out-of-order */
7317 UV final_element = len - 1;
7318 array = invlist_array(invlist);
7319 if (array[final_element] > start
7320 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
7322 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",
7323 array[final_element], start,
7324 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
7327 /* Here, it is a legal append. If the new range begins with the first
7328 * value not in the set, it is extending the set, so the new first
7329 * value not in the set is one greater than the newly extended range.
7331 if (array[final_element] == start) {
7332 if (end != UV_MAX) {
7333 array[final_element] = end + 1;
7336 /* But if the end is the maximum representable on the machine,
7337 * just let the range that this would extend to have no end */
7338 invlist_set_len(invlist, len - 1);
7344 /* Here the new range doesn't extend any existing set. Add it */
7346 len += 2; /* Includes an element each for the start and end of range */
7348 /* If overflows the existing space, extend, which may cause the array to be
7351 invlist_extend(invlist, len);
7352 invlist_set_len(invlist, len); /* Have to set len here to avoid assert
7353 failure in invlist_array() */
7354 array = invlist_array(invlist);
7357 invlist_set_len(invlist, len);
7360 /* The next item on the list starts the range, the one after that is
7361 * one past the new range. */
7362 array[len - 2] = start;
7363 if (end != UV_MAX) {
7364 array[len - 1] = end + 1;
7367 /* But if the end is the maximum representable on the machine, just let
7368 * the range have no end */
7369 invlist_set_len(invlist, len - 1);
7373 #ifndef PERL_IN_XSUB_RE
7376 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
7378 /* Searches the inversion list for the entry that contains the input code
7379 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
7380 * return value is the index into the list's array of the range that
7385 IV high = _invlist_len(invlist);
7386 const IV highest_element = high - 1;
7389 PERL_ARGS_ASSERT__INVLIST_SEARCH;
7391 /* If list is empty, return failure. */
7396 /* (We can't get the array unless we know the list is non-empty) */
7397 array = invlist_array(invlist);
7399 mid = invlist_previous_index(invlist);
7400 assert(mid >=0 && mid <= highest_element);
7402 /* <mid> contains the cache of the result of the previous call to this
7403 * function (0 the first time). See if this call is for the same result,
7404 * or if it is for mid-1. This is under the theory that calls to this
7405 * function will often be for related code points that are near each other.
7406 * And benchmarks show that caching gives better results. We also test
7407 * here if the code point is within the bounds of the list. These tests
7408 * replace others that would have had to be made anyway to make sure that
7409 * the array bounds were not exceeded, and these give us extra information
7410 * at the same time */
7411 if (cp >= array[mid]) {
7412 if (cp >= array[highest_element]) {
7413 return highest_element;
7416 /* Here, array[mid] <= cp < array[highest_element]. This means that
7417 * the final element is not the answer, so can exclude it; it also
7418 * means that <mid> is not the final element, so can refer to 'mid + 1'
7420 if (cp < array[mid + 1]) {
7426 else { /* cp < aray[mid] */
7427 if (cp < array[0]) { /* Fail if outside the array */
7431 if (cp >= array[mid - 1]) {
7436 /* Binary search. What we are looking for is <i> such that
7437 * array[i] <= cp < array[i+1]
7438 * The loop below converges on the i+1. Note that there may not be an
7439 * (i+1)th element in the array, and things work nonetheless */
7440 while (low < high) {
7441 mid = (low + high) / 2;
7442 assert(mid <= highest_element);
7443 if (array[mid] <= cp) { /* cp >= array[mid] */
7446 /* We could do this extra test to exit the loop early.
7447 if (cp < array[low]) {
7452 else { /* cp < array[mid] */
7459 invlist_set_previous_index(invlist, high);
7464 Perl__invlist_populate_swatch(pTHX_ SV* const invlist, const UV start, const UV end, U8* swatch)
7466 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
7467 * but is used when the swash has an inversion list. This makes this much
7468 * faster, as it uses a binary search instead of a linear one. This is
7469 * intimately tied to that function, and perhaps should be in utf8.c,
7470 * except it is intimately tied to inversion lists as well. It assumes
7471 * that <swatch> is all 0's on input */
7474 const IV len = _invlist_len(invlist);
7478 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
7480 if (len == 0) { /* Empty inversion list */
7484 array = invlist_array(invlist);
7486 /* Find which element it is */
7487 i = _invlist_search(invlist, start);
7489 /* We populate from <start> to <end> */
7490 while (current < end) {
7493 /* The inversion list gives the results for every possible code point
7494 * after the first one in the list. Only those ranges whose index is
7495 * even are ones that the inversion list matches. For the odd ones,
7496 * and if the initial code point is not in the list, we have to skip
7497 * forward to the next element */
7498 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
7500 if (i >= len) { /* Finished if beyond the end of the array */
7504 if (current >= end) { /* Finished if beyond the end of what we
7506 if (LIKELY(end < UV_MAX)) {
7510 /* We get here when the upper bound is the maximum
7511 * representable on the machine, and we are looking for just
7512 * that code point. Have to special case it */
7514 goto join_end_of_list;
7517 assert(current >= start);
7519 /* The current range ends one below the next one, except don't go past
7522 upper = (i < len && array[i] < end) ? array[i] : end;
7524 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
7525 * for each code point in it */
7526 for (; current < upper; current++) {
7527 const STRLEN offset = (STRLEN)(current - start);
7528 swatch[offset >> 3] |= 1 << (offset & 7);
7533 /* Quit if at the end of the list */
7536 /* But first, have to deal with the highest possible code point on
7537 * the platform. The previous code assumes that <end> is one
7538 * beyond where we want to populate, but that is impossible at the
7539 * platform's infinity, so have to handle it specially */
7540 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
7542 const STRLEN offset = (STRLEN)(end - start);
7543 swatch[offset >> 3] |= 1 << (offset & 7);
7548 /* Advance to the next range, which will be for code points not in the
7557 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** output)
7559 /* Take the union of two inversion lists and point <output> to it. *output
7560 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7561 * the reference count to that list will be decremented. The first list,
7562 * <a>, may be NULL, in which case a copy of the second list is returned.
7563 * If <complement_b> is TRUE, the union is taken of the complement
7564 * (inversion) of <b> instead of b itself.
7566 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7567 * Richard Gillam, published by Addison-Wesley, and explained at some
7568 * length there. The preface says to incorporate its examples into your
7569 * code at your own risk.
7571 * The algorithm is like a merge sort.
7573 * XXX A potential performance improvement is to keep track as we go along
7574 * if only one of the inputs contributes to the result, meaning the other
7575 * is a subset of that one. In that case, we can skip the final copy and
7576 * return the larger of the input lists, but then outside code might need
7577 * to keep track of whether to free the input list or not */
7579 UV* array_a; /* a's array */
7581 UV len_a; /* length of a's array */
7584 SV* u; /* the resulting union */
7588 UV i_a = 0; /* current index into a's array */
7592 /* running count, as explained in the algorithm source book; items are
7593 * stopped accumulating and are output when the count changes to/from 0.
7594 * The count is incremented when we start a range that's in the set, and
7595 * decremented when we start a range that's not in the set. So its range
7596 * is 0 to 2. Only when the count is zero is something not in the set.
7600 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
7603 /* If either one is empty, the union is the other one */
7604 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
7611 *output = invlist_clone(b);
7613 _invlist_invert(*output);
7615 } /* else *output already = b; */
7618 else if ((len_b = _invlist_len(b)) == 0) {
7623 /* The complement of an empty list is a list that has everything in it,
7624 * so the union with <a> includes everything too */
7629 *output = _new_invlist(1);
7630 _append_range_to_invlist(*output, 0, UV_MAX);
7632 else if (*output != a) {
7633 *output = invlist_clone(a);
7635 /* else *output already = a; */
7639 /* Here both lists exist and are non-empty */
7640 array_a = invlist_array(a);
7641 array_b = invlist_array(b);
7643 /* If are to take the union of 'a' with the complement of b, set it
7644 * up so are looking at b's complement. */
7647 /* To complement, we invert: if the first element is 0, remove it. To
7648 * do this, we just pretend the array starts one later, and clear the
7649 * flag as we don't have to do anything else later */
7650 if (array_b[0] == 0) {
7653 complement_b = FALSE;
7657 /* But if the first element is not zero, we unshift a 0 before the
7658 * array. The data structure reserves a space for that 0 (which
7659 * should be a '1' right now), so physical shifting is unneeded,
7660 * but temporarily change that element to 0. Before exiting the
7661 * routine, we must restore the element to '1' */
7668 /* Size the union for the worst case: that the sets are completely
7670 u = _new_invlist(len_a + len_b);
7672 /* Will contain U+0000 if either component does */
7673 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
7674 || (len_b > 0 && array_b[0] == 0));
7676 /* Go through each list item by item, stopping when exhausted one of
7678 while (i_a < len_a && i_b < len_b) {
7679 UV cp; /* The element to potentially add to the union's array */
7680 bool cp_in_set; /* is it in the the input list's set or not */
7682 /* We need to take one or the other of the two inputs for the union.
7683 * Since we are merging two sorted lists, we take the smaller of the
7684 * next items. In case of a tie, we take the one that is in its set
7685 * first. If we took one not in the set first, it would decrement the
7686 * count, possibly to 0 which would cause it to be output as ending the
7687 * range, and the next time through we would take the same number, and
7688 * output it again as beginning the next range. By doing it the
7689 * opposite way, there is no possibility that the count will be
7690 * momentarily decremented to 0, and thus the two adjoining ranges will
7691 * be seamlessly merged. (In a tie and both are in the set or both not
7692 * in the set, it doesn't matter which we take first.) */
7693 if (array_a[i_a] < array_b[i_b]
7694 || (array_a[i_a] == array_b[i_b]
7695 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7697 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7701 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7702 cp = array_b[i_b++];
7705 /* Here, have chosen which of the two inputs to look at. Only output
7706 * if the running count changes to/from 0, which marks the
7707 * beginning/end of a range in that's in the set */
7710 array_u[i_u++] = cp;
7717 array_u[i_u++] = cp;
7722 /* Here, we are finished going through at least one of the lists, which
7723 * means there is something remaining in at most one. We check if the list
7724 * that hasn't been exhausted is positioned such that we are in the middle
7725 * of a range in its set or not. (i_a and i_b point to the element beyond
7726 * the one we care about.) If in the set, we decrement 'count'; if 0, there
7727 * is potentially more to output.
7728 * There are four cases:
7729 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
7730 * in the union is entirely from the non-exhausted set.
7731 * 2) Both were in their sets, count is 2. Nothing further should
7732 * be output, as everything that remains will be in the exhausted
7733 * list's set, hence in the union; decrementing to 1 but not 0 insures
7735 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
7736 * Nothing further should be output because the union includes
7737 * everything from the exhausted set. Not decrementing ensures that.
7738 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
7739 * decrementing to 0 insures that we look at the remainder of the
7740 * non-exhausted set */
7741 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7742 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7747 /* The final length is what we've output so far, plus what else is about to
7748 * be output. (If 'count' is non-zero, then the input list we exhausted
7749 * has everything remaining up to the machine's limit in its set, and hence
7750 * in the union, so there will be no further output. */
7753 /* At most one of the subexpressions will be non-zero */
7754 len_u += (len_a - i_a) + (len_b - i_b);
7757 /* Set result to final length, which can change the pointer to array_u, so
7759 if (len_u != _invlist_len(u)) {
7760 invlist_set_len(u, len_u);
7762 array_u = invlist_array(u);
7765 /* When 'count' is 0, the list that was exhausted (if one was shorter than
7766 * the other) ended with everything above it not in its set. That means
7767 * that the remaining part of the union is precisely the same as the
7768 * non-exhausted list, so can just copy it unchanged. (If both list were
7769 * exhausted at the same time, then the operations below will be both 0.)
7772 IV copy_count; /* At most one will have a non-zero copy count */
7773 if ((copy_count = len_a - i_a) > 0) {
7774 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
7776 else if ((copy_count = len_b - i_b) > 0) {
7777 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
7781 /* If we've changed b, restore it */
7786 /* We may be removing a reference to one of the inputs */
7787 if (a == *output || b == *output) {
7788 assert(! invlist_is_iterating(*output));
7789 SvREFCNT_dec_NN(*output);
7797 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** i)
7799 /* Take the intersection of two inversion lists and point <i> to it. *i
7800 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7801 * the reference count to that list will be decremented.
7802 * If <complement_b> is TRUE, the result will be the intersection of <a>
7803 * and the complement (or inversion) of <b> instead of <b> directly.
7805 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7806 * Richard Gillam, published by Addison-Wesley, and explained at some
7807 * length there. The preface says to incorporate its examples into your
7808 * code at your own risk. In fact, it had bugs
7810 * The algorithm is like a merge sort, and is essentially the same as the
7814 UV* array_a; /* a's array */
7816 UV len_a; /* length of a's array */
7819 SV* r; /* the resulting intersection */
7823 UV i_a = 0; /* current index into a's array */
7827 /* running count, as explained in the algorithm source book; items are
7828 * stopped accumulating and are output when the count changes to/from 2.
7829 * The count is incremented when we start a range that's in the set, and
7830 * decremented when we start a range that's not in the set. So its range
7831 * is 0 to 2. Only when the count is 2 is something in the intersection.
7835 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
7838 /* Special case if either one is empty */
7839 len_a = _invlist_len(a);
7840 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
7842 if (len_a != 0 && complement_b) {
7844 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
7845 * be empty. Here, also we are using 'b's complement, which hence
7846 * must be every possible code point. Thus the intersection is
7849 *i = invlist_clone(a);
7855 /* else *i is already 'a' */
7859 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
7860 * intersection must be empty */
7867 *i = _new_invlist(0);
7871 /* Here both lists exist and are non-empty */
7872 array_a = invlist_array(a);
7873 array_b = invlist_array(b);
7875 /* If are to take the intersection of 'a' with the complement of b, set it
7876 * up so are looking at b's complement. */
7879 /* To complement, we invert: if the first element is 0, remove it. To
7880 * do this, we just pretend the array starts one later, and clear the
7881 * flag as we don't have to do anything else later */
7882 if (array_b[0] == 0) {
7885 complement_b = FALSE;
7889 /* But if the first element is not zero, we unshift a 0 before the
7890 * array. The data structure reserves a space for that 0 (which
7891 * should be a '1' right now), so physical shifting is unneeded,
7892 * but temporarily change that element to 0. Before exiting the
7893 * routine, we must restore the element to '1' */
7900 /* Size the intersection for the worst case: that the intersection ends up
7901 * fragmenting everything to be completely disjoint */
7902 r= _new_invlist(len_a + len_b);
7904 /* Will contain U+0000 iff both components do */
7905 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
7906 && len_b > 0 && array_b[0] == 0);
7908 /* Go through each list item by item, stopping when exhausted one of
7910 while (i_a < len_a && i_b < len_b) {
7911 UV cp; /* The element to potentially add to the intersection's
7913 bool cp_in_set; /* Is it in the input list's set or not */
7915 /* We need to take one or the other of the two inputs for the
7916 * intersection. Since we are merging two sorted lists, we take the
7917 * smaller of the next items. In case of a tie, we take the one that
7918 * is not in its set first (a difference from the union algorithm). If
7919 * we took one in the set first, it would increment the count, possibly
7920 * to 2 which would cause it to be output as starting a range in the
7921 * intersection, and the next time through we would take that same
7922 * number, and output it again as ending the set. By doing it the
7923 * opposite of this, there is no possibility that the count will be
7924 * momentarily incremented to 2. (In a tie and both are in the set or
7925 * both not in the set, it doesn't matter which we take first.) */
7926 if (array_a[i_a] < array_b[i_b]
7927 || (array_a[i_a] == array_b[i_b]
7928 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7930 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7934 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7938 /* Here, have chosen which of the two inputs to look at. Only output
7939 * if the running count changes to/from 2, which marks the
7940 * beginning/end of a range that's in the intersection */
7944 array_r[i_r++] = cp;
7949 array_r[i_r++] = cp;
7955 /* Here, we are finished going through at least one of the lists, which
7956 * means there is something remaining in at most one. We check if the list
7957 * that has been exhausted is positioned such that we are in the middle
7958 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
7959 * the ones we care about.) There are four cases:
7960 * 1) Both weren't in their sets, count is 0, and remains 0. There's
7961 * nothing left in the intersection.
7962 * 2) Both were in their sets, count is 2 and perhaps is incremented to
7963 * above 2. What should be output is exactly that which is in the
7964 * non-exhausted set, as everything it has is also in the intersection
7965 * set, and everything it doesn't have can't be in the intersection
7966 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
7967 * gets incremented to 2. Like the previous case, the intersection is
7968 * everything that remains in the non-exhausted set.
7969 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
7970 * remains 1. And the intersection has nothing more. */
7971 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7972 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7977 /* The final length is what we've output so far plus what else is in the
7978 * intersection. At most one of the subexpressions below will be non-zero */
7981 len_r += (len_a - i_a) + (len_b - i_b);
7984 /* Set result to final length, which can change the pointer to array_r, so
7986 if (len_r != _invlist_len(r)) {
7987 invlist_set_len(r, len_r);
7989 array_r = invlist_array(r);
7992 /* Finish outputting any remaining */
7993 if (count >= 2) { /* At most one will have a non-zero copy count */
7995 if ((copy_count = len_a - i_a) > 0) {
7996 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
7998 else if ((copy_count = len_b - i_b) > 0) {
7999 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
8003 /* If we've changed b, restore it */
8008 /* We may be removing a reference to one of the inputs */
8009 if (a == *i || b == *i) {
8010 assert(! invlist_is_iterating(*i));
8011 SvREFCNT_dec_NN(*i);
8019 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
8021 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8022 * set. A pointer to the inversion list is returned. This may actually be
8023 * a new list, in which case the passed in one has been destroyed. The
8024 * passed in inversion list can be NULL, in which case a new one is created
8025 * with just the one range in it */
8030 if (invlist == NULL) {
8031 invlist = _new_invlist(2);
8035 len = _invlist_len(invlist);
8038 /* If comes after the final entry actually in the list, can just append it
8041 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
8042 && start >= invlist_array(invlist)[len - 1]))
8044 _append_range_to_invlist(invlist, start, end);
8048 /* Here, can't just append things, create and return a new inversion list
8049 * which is the union of this range and the existing inversion list */
8050 range_invlist = _new_invlist(2);
8051 _append_range_to_invlist(range_invlist, start, end);
8053 _invlist_union(invlist, range_invlist, &invlist);
8055 /* The temporary can be freed */
8056 SvREFCNT_dec_NN(range_invlist);
8063 PERL_STATIC_INLINE SV*
8064 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
8065 return _add_range_to_invlist(invlist, cp, cp);
8068 #ifndef PERL_IN_XSUB_RE
8070 Perl__invlist_invert(pTHX_ SV* const invlist)
8072 /* Complement the input inversion list. This adds a 0 if the list didn't
8073 * have a zero; removes it otherwise. As described above, the data
8074 * structure is set up so that this is very efficient */
8076 UV* len_pos = _get_invlist_len_addr(invlist);
8078 PERL_ARGS_ASSERT__INVLIST_INVERT;
8080 assert(! invlist_is_iterating(invlist));
8082 /* The inverse of matching nothing is matching everything */
8083 if (*len_pos == 0) {
8084 _append_range_to_invlist(invlist, 0, UV_MAX);
8088 /* The exclusive or complents 0 to 1; and 1 to 0. If the result is 1, the
8089 * zero element was a 0, so it is being removed, so the length decrements
8090 * by 1; and vice-versa. SvCUR is unaffected */
8091 if (*get_invlist_zero_addr(invlist) ^= 1) {
8100 Perl__invlist_invert_prop(pTHX_ SV* const invlist)
8102 /* Complement the input inversion list (which must be a Unicode property,
8103 * all of which don't match above the Unicode maximum code point.) And
8104 * Perl has chosen to not have the inversion match above that either. This
8105 * adds a 0x110000 if the list didn't end with it, and removes it if it did
8111 PERL_ARGS_ASSERT__INVLIST_INVERT_PROP;
8113 _invlist_invert(invlist);
8115 len = _invlist_len(invlist);
8117 if (len != 0) { /* If empty do nothing */
8118 array = invlist_array(invlist);
8119 if (array[len - 1] != PERL_UNICODE_MAX + 1) {
8120 /* Add 0x110000. First, grow if necessary */
8122 if (invlist_max(invlist) < len) {
8123 invlist_extend(invlist, len);
8124 array = invlist_array(invlist);
8126 invlist_set_len(invlist, len);
8127 array[len - 1] = PERL_UNICODE_MAX + 1;
8129 else { /* Remove the 0x110000 */
8130 invlist_set_len(invlist, len - 1);
8138 PERL_STATIC_INLINE SV*
8139 S_invlist_clone(pTHX_ SV* const invlist)
8142 /* Return a new inversion list that is a copy of the input one, which is
8145 /* Need to allocate extra space to accommodate Perl's addition of a
8146 * trailing NUL to SvPV's, since it thinks they are always strings */
8147 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8148 STRLEN length = SvCUR(invlist);
8150 PERL_ARGS_ASSERT_INVLIST_CLONE;
8152 SvCUR_set(new_invlist, length); /* This isn't done automatically */
8153 Copy(SvPVX(invlist), SvPVX(new_invlist), length, char);
8158 PERL_STATIC_INLINE UV*
8159 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8161 /* Return the address of the UV that contains the current iteration
8164 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8166 return (UV *) (SvPVX(invlist) + (INVLIST_ITER_OFFSET * sizeof (UV)));
8169 PERL_STATIC_INLINE UV*
8170 S_get_invlist_version_id_addr(pTHX_ SV* invlist)
8172 /* Return the address of the UV that contains the version id. */
8174 PERL_ARGS_ASSERT_GET_INVLIST_VERSION_ID_ADDR;
8176 return (UV *) (SvPVX(invlist) + (INVLIST_VERSION_ID_OFFSET * sizeof (UV)));
8179 PERL_STATIC_INLINE void
8180 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8182 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8184 *get_invlist_iter_addr(invlist) = 0;
8187 PERL_STATIC_INLINE void
8188 S_invlist_iterfinish(pTHX_ SV* invlist)
8190 /* Terminate iterator for invlist. This is to catch development errors.
8191 * Any iteration that is interrupted before completed should call this
8192 * function. Functions that add code points anywhere else but to the end
8193 * of an inversion list assert that they are not in the middle of an
8194 * iteration. If they were, the addition would make the iteration
8195 * problematical: if the iteration hadn't reached the place where things
8196 * were being added, it would be ok */
8198 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8200 *get_invlist_iter_addr(invlist) = UV_MAX;
8204 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8206 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8207 * This call sets in <*start> and <*end>, the next range in <invlist>.
8208 * Returns <TRUE> if successful and the next call will return the next
8209 * range; <FALSE> if was already at the end of the list. If the latter,
8210 * <*start> and <*end> are unchanged, and the next call to this function
8211 * will start over at the beginning of the list */
8213 UV* pos = get_invlist_iter_addr(invlist);
8214 UV len = _invlist_len(invlist);
8217 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8220 *pos = UV_MAX; /* Force iterinit() to be required next time */
8224 array = invlist_array(invlist);
8226 *start = array[(*pos)++];
8232 *end = array[(*pos)++] - 1;
8238 PERL_STATIC_INLINE bool
8239 S_invlist_is_iterating(pTHX_ SV* const invlist)
8241 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8243 return *(get_invlist_iter_addr(invlist)) < UV_MAX;
8246 PERL_STATIC_INLINE UV
8247 S_invlist_highest(pTHX_ SV* const invlist)
8249 /* Returns the highest code point that matches an inversion list. This API
8250 * has an ambiguity, as it returns 0 under either the highest is actually
8251 * 0, or if the list is empty. If this distinction matters to you, check
8252 * for emptiness before calling this function */
8254 UV len = _invlist_len(invlist);
8257 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8263 array = invlist_array(invlist);
8265 /* The last element in the array in the inversion list always starts a
8266 * range that goes to infinity. That range may be for code points that are
8267 * matched in the inversion list, or it may be for ones that aren't
8268 * matched. In the latter case, the highest code point in the set is one
8269 * less than the beginning of this range; otherwise it is the final element
8270 * of this range: infinity */
8271 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8273 : array[len - 1] - 1;
8276 #ifndef PERL_IN_XSUB_RE
8278 Perl__invlist_contents(pTHX_ SV* const invlist)
8280 /* Get the contents of an inversion list into a string SV so that they can
8281 * be printed out. It uses the format traditionally done for debug tracing
8285 SV* output = newSVpvs("\n");
8287 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8289 assert(! invlist_is_iterating(invlist));
8291 invlist_iterinit(invlist);
8292 while (invlist_iternext(invlist, &start, &end)) {
8293 if (end == UV_MAX) {
8294 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8296 else if (end != start) {
8297 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8301 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
8309 #ifdef PERL_ARGS_ASSERT__INVLIST_DUMP
8311 Perl__invlist_dump(pTHX_ SV* const invlist, const char * const header)
8313 /* Dumps out the ranges in an inversion list. The string 'header'
8314 * if present is output on a line before the first range */
8318 PERL_ARGS_ASSERT__INVLIST_DUMP;
8320 if (header && strlen(header)) {
8321 PerlIO_printf(Perl_debug_log, "%s\n", header);
8323 if (invlist_is_iterating(invlist)) {
8324 PerlIO_printf(Perl_debug_log, "Can't dump because is in middle of iterating\n");
8328 invlist_iterinit(invlist);
8329 while (invlist_iternext(invlist, &start, &end)) {
8330 if (end == UV_MAX) {
8331 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. INFINITY\n", start);
8333 else if (end != start) {
8334 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. 0x%04"UVXf"\n",
8338 PerlIO_printf(Perl_debug_log, "0x%04"UVXf"\n", start);
8346 S__invlistEQ(pTHX_ SV* const a, SV* const b, bool complement_b)
8348 /* Return a boolean as to if the two passed in inversion lists are
8349 * identical. The final argument, if TRUE, says to take the complement of
8350 * the second inversion list before doing the comparison */
8352 UV* array_a = invlist_array(a);
8353 UV* array_b = invlist_array(b);
8354 UV len_a = _invlist_len(a);
8355 UV len_b = _invlist_len(b);
8357 UV i = 0; /* current index into the arrays */
8358 bool retval = TRUE; /* Assume are identical until proven otherwise */
8360 PERL_ARGS_ASSERT__INVLISTEQ;
8362 /* If are to compare 'a' with the complement of b, set it
8363 * up so are looking at b's complement. */
8366 /* The complement of nothing is everything, so <a> would have to have
8367 * just one element, starting at zero (ending at infinity) */
8369 return (len_a == 1 && array_a[0] == 0);
8371 else if (array_b[0] == 0) {
8373 /* Otherwise, to complement, we invert. Here, the first element is
8374 * 0, just remove it. To do this, we just pretend the array starts
8375 * one later, and clear the flag as we don't have to do anything
8380 complement_b = FALSE;
8384 /* But if the first element is not zero, we unshift a 0 before the
8385 * array. The data structure reserves a space for that 0 (which
8386 * should be a '1' right now), so physical shifting is unneeded,
8387 * but temporarily change that element to 0. Before exiting the
8388 * routine, we must restore the element to '1' */
8395 /* Make sure that the lengths are the same, as well as the final element
8396 * before looping through the remainder. (Thus we test the length, final,
8397 * and first elements right off the bat) */
8398 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
8401 else for (i = 0; i < len_a - 1; i++) {
8402 if (array_a[i] != array_b[i]) {
8415 #undef HEADER_LENGTH
8416 #undef INVLIST_INITIAL_LENGTH
8417 #undef TO_INTERNAL_SIZE
8418 #undef FROM_INTERNAL_SIZE
8419 #undef INVLIST_LEN_OFFSET
8420 #undef INVLIST_ZERO_OFFSET
8421 #undef INVLIST_ITER_OFFSET
8422 #undef INVLIST_VERSION_ID
8423 #undef INVLIST_PREVIOUS_INDEX_OFFSET
8425 /* End of inversion list object */
8428 S_parse_lparen_question_flags(pTHX_ struct RExC_state_t *pRExC_state)
8430 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
8431 * constructs, and updates RExC_flags with them. On input, RExC_parse
8432 * should point to the first flag; it is updated on output to point to the
8433 * final ')' or ':'. There needs to be at least one flag, or this will
8436 /* for (?g), (?gc), and (?o) warnings; warning
8437 about (?c) will warn about (?g) -- japhy */
8439 #define WASTED_O 0x01
8440 #define WASTED_G 0x02
8441 #define WASTED_C 0x04
8442 #define WASTED_GC (WASTED_G|WASTED_C)
8443 I32 wastedflags = 0x00;
8444 U32 posflags = 0, negflags = 0;
8445 U32 *flagsp = &posflags;
8446 char has_charset_modifier = '\0';
8448 bool has_use_defaults = FALSE;
8449 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
8451 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
8453 /* '^' as an initial flag sets certain defaults */
8454 if (UCHARAT(RExC_parse) == '^') {
8456 has_use_defaults = TRUE;
8457 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
8458 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
8459 ? REGEX_UNICODE_CHARSET
8460 : REGEX_DEPENDS_CHARSET);
8463 cs = get_regex_charset(RExC_flags);
8464 if (cs == REGEX_DEPENDS_CHARSET
8465 && (RExC_utf8 || RExC_uni_semantics))
8467 cs = REGEX_UNICODE_CHARSET;
8470 while (*RExC_parse) {
8471 /* && strchr("iogcmsx", *RExC_parse) */
8472 /* (?g), (?gc) and (?o) are useless here
8473 and must be globally applied -- japhy */
8474 switch (*RExC_parse) {
8476 /* Code for the imsx flags */
8477 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
8479 case LOCALE_PAT_MOD:
8480 if (has_charset_modifier) {
8481 goto excess_modifier;
8483 else if (flagsp == &negflags) {
8486 cs = REGEX_LOCALE_CHARSET;
8487 has_charset_modifier = LOCALE_PAT_MOD;
8488 RExC_contains_locale = 1;
8490 case UNICODE_PAT_MOD:
8491 if (has_charset_modifier) {
8492 goto excess_modifier;
8494 else if (flagsp == &negflags) {
8497 cs = REGEX_UNICODE_CHARSET;
8498 has_charset_modifier = UNICODE_PAT_MOD;
8500 case ASCII_RESTRICT_PAT_MOD:
8501 if (flagsp == &negflags) {
8504 if (has_charset_modifier) {
8505 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
8506 goto excess_modifier;
8508 /* Doubled modifier implies more restricted */
8509 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
8512 cs = REGEX_ASCII_RESTRICTED_CHARSET;
8514 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
8516 case DEPENDS_PAT_MOD:
8517 if (has_use_defaults) {
8518 goto fail_modifiers;
8520 else if (flagsp == &negflags) {
8523 else if (has_charset_modifier) {
8524 goto excess_modifier;
8527 /* The dual charset means unicode semantics if the
8528 * pattern (or target, not known until runtime) are
8529 * utf8, or something in the pattern indicates unicode
8531 cs = (RExC_utf8 || RExC_uni_semantics)
8532 ? REGEX_UNICODE_CHARSET
8533 : REGEX_DEPENDS_CHARSET;
8534 has_charset_modifier = DEPENDS_PAT_MOD;
8538 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
8539 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
8541 else if (has_charset_modifier == *(RExC_parse - 1)) {
8542 vFAIL2("Regexp modifier \"%c\" may not appear twice", *(RExC_parse - 1));
8545 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
8550 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"", *(RExC_parse - 1));
8552 case ONCE_PAT_MOD: /* 'o' */
8553 case GLOBAL_PAT_MOD: /* 'g' */
8554 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8555 const I32 wflagbit = *RExC_parse == 'o' ? WASTED_O : WASTED_G;
8556 if (! (wastedflags & wflagbit) ) {
8557 wastedflags |= wflagbit;
8558 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
8561 "Useless (%s%c) - %suse /%c modifier",
8562 flagsp == &negflags ? "?-" : "?",
8564 flagsp == &negflags ? "don't " : "",
8571 case CONTINUE_PAT_MOD: /* 'c' */
8572 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8573 if (! (wastedflags & WASTED_C) ) {
8574 wastedflags |= WASTED_GC;
8575 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
8578 "Useless (%sc) - %suse /gc modifier",
8579 flagsp == &negflags ? "?-" : "?",
8580 flagsp == &negflags ? "don't " : ""
8585 case KEEPCOPY_PAT_MOD: /* 'p' */
8586 if (flagsp == &negflags) {
8588 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
8590 *flagsp |= RXf_PMf_KEEPCOPY;
8594 /* A flag is a default iff it is following a minus, so
8595 * if there is a minus, it means will be trying to
8596 * re-specify a default which is an error */
8597 if (has_use_defaults || flagsp == &negflags) {
8598 goto fail_modifiers;
8601 wastedflags = 0; /* reset so (?g-c) warns twice */
8605 RExC_flags |= posflags;
8606 RExC_flags &= ~negflags;
8607 set_regex_charset(&RExC_flags, cs);
8613 vFAIL3("Sequence (%.*s...) not recognized",
8614 RExC_parse-seqstart, seqstart);
8623 - reg - regular expression, i.e. main body or parenthesized thing
8625 * Caller must absorb opening parenthesis.
8627 * Combining parenthesis handling with the base level of regular expression
8628 * is a trifle forced, but the need to tie the tails of the branches to what
8629 * follows makes it hard to avoid.
8631 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
8633 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
8635 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
8638 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
8639 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
8640 needs to be restarted.
8641 Otherwise would only return NULL if regbranch() returns NULL, which
8644 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
8645 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
8646 * 2 is like 1, but indicates that nextchar() has been called to advance
8647 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
8648 * this flag alerts us to the need to check for that */
8651 regnode *ret; /* Will be the head of the group. */
8654 regnode *ender = NULL;
8657 U32 oregflags = RExC_flags;
8658 bool have_branch = 0;
8660 I32 freeze_paren = 0;
8661 I32 after_freeze = 0;
8663 char * parse_start = RExC_parse; /* MJD */
8664 char * const oregcomp_parse = RExC_parse;
8666 GET_RE_DEBUG_FLAGS_DECL;
8668 PERL_ARGS_ASSERT_REG;
8669 DEBUG_PARSE("reg ");
8671 *flagp = 0; /* Tentatively. */
8674 /* Make an OPEN node, if parenthesized. */
8677 /* Under /x, space and comments can be gobbled up between the '(' and
8678 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
8679 * intervening space, as the sequence is a token, and a token should be
8681 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
8683 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
8684 char *start_verb = RExC_parse;
8685 STRLEN verb_len = 0;
8686 char *start_arg = NULL;
8687 unsigned char op = 0;
8689 int internal_argval = 0; /* internal_argval is only useful if !argok */
8691 if (has_intervening_patws && SIZE_ONLY) {
8692 ckWARNregdep(RExC_parse + 1, "In '(*VERB...)', splitting the initial '(*' is deprecated");
8694 while ( *RExC_parse && *RExC_parse != ')' ) {
8695 if ( *RExC_parse == ':' ) {
8696 start_arg = RExC_parse + 1;
8702 verb_len = RExC_parse - start_verb;
8705 while ( *RExC_parse && *RExC_parse != ')' )
8707 if ( *RExC_parse != ')' )
8708 vFAIL("Unterminated verb pattern argument");
8709 if ( RExC_parse == start_arg )
8712 if ( *RExC_parse != ')' )
8713 vFAIL("Unterminated verb pattern");
8716 switch ( *start_verb ) {
8717 case 'A': /* (*ACCEPT) */
8718 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
8720 internal_argval = RExC_nestroot;
8723 case 'C': /* (*COMMIT) */
8724 if ( memEQs(start_verb,verb_len,"COMMIT") )
8727 case 'F': /* (*FAIL) */
8728 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
8733 case ':': /* (*:NAME) */
8734 case 'M': /* (*MARK:NAME) */
8735 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
8740 case 'P': /* (*PRUNE) */
8741 if ( memEQs(start_verb,verb_len,"PRUNE") )
8744 case 'S': /* (*SKIP) */
8745 if ( memEQs(start_verb,verb_len,"SKIP") )
8748 case 'T': /* (*THEN) */
8749 /* [19:06] <TimToady> :: is then */
8750 if ( memEQs(start_verb,verb_len,"THEN") ) {
8752 RExC_seen |= REG_SEEN_CUTGROUP;
8758 vFAIL3("Unknown verb pattern '%.*s'",
8759 verb_len, start_verb);
8762 if ( start_arg && internal_argval ) {
8763 vFAIL3("Verb pattern '%.*s' may not have an argument",
8764 verb_len, start_verb);
8765 } else if ( argok < 0 && !start_arg ) {
8766 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
8767 verb_len, start_verb);
8769 ret = reganode(pRExC_state, op, internal_argval);
8770 if ( ! internal_argval && ! SIZE_ONLY ) {
8772 SV *sv = newSVpvn( start_arg, RExC_parse - start_arg);
8773 ARG(ret) = add_data( pRExC_state, 1, "S" );
8774 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
8781 if (!internal_argval)
8782 RExC_seen |= REG_SEEN_VERBARG;
8783 } else if ( start_arg ) {
8784 vFAIL3("Verb pattern '%.*s' may not have an argument",
8785 verb_len, start_verb);
8787 ret = reg_node(pRExC_state, op);
8789 nextchar(pRExC_state);
8792 else if (*RExC_parse == '?') { /* (?...) */
8793 bool is_logical = 0;
8794 const char * const seqstart = RExC_parse;
8795 if (has_intervening_patws && SIZE_ONLY) {
8796 ckWARNregdep(RExC_parse + 1, "In '(?...)', splitting the initial '(?' is deprecated");
8800 paren = *RExC_parse++;
8801 ret = NULL; /* For look-ahead/behind. */
8804 case 'P': /* (?P...) variants for those used to PCRE/Python */
8805 paren = *RExC_parse++;
8806 if ( paren == '<') /* (?P<...>) named capture */
8808 else if (paren == '>') { /* (?P>name) named recursion */
8809 goto named_recursion;
8811 else if (paren == '=') { /* (?P=...) named backref */
8812 /* this pretty much dupes the code for \k<NAME> in regatom(), if
8813 you change this make sure you change that */
8814 char* name_start = RExC_parse;
8816 SV *sv_dat = reg_scan_name(pRExC_state,
8817 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8818 if (RExC_parse == name_start || *RExC_parse != ')')
8819 vFAIL2("Sequence %.3s... not terminated",parse_start);
8822 num = add_data( pRExC_state, 1, "S" );
8823 RExC_rxi->data->data[num]=(void*)sv_dat;
8824 SvREFCNT_inc_simple_void(sv_dat);
8827 ret = reganode(pRExC_state,
8830 : (ASCII_FOLD_RESTRICTED)
8832 : (AT_LEAST_UNI_SEMANTICS)
8840 Set_Node_Offset(ret, parse_start+1);
8841 Set_Node_Cur_Length(ret, parse_start);
8843 nextchar(pRExC_state);
8847 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8849 case '<': /* (?<...) */
8850 if (*RExC_parse == '!')
8852 else if (*RExC_parse != '=')
8858 case '\'': /* (?'...') */
8859 name_start= RExC_parse;
8860 svname = reg_scan_name(pRExC_state,
8861 SIZE_ONLY ? /* reverse test from the others */
8862 REG_RSN_RETURN_NAME :
8863 REG_RSN_RETURN_NULL);
8864 if (RExC_parse == name_start) {
8866 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8869 if (*RExC_parse != paren)
8870 vFAIL2("Sequence (?%c... not terminated",
8871 paren=='>' ? '<' : paren);
8875 if (!svname) /* shouldn't happen */
8877 "panic: reg_scan_name returned NULL");
8878 if (!RExC_paren_names) {
8879 RExC_paren_names= newHV();
8880 sv_2mortal(MUTABLE_SV(RExC_paren_names));
8882 RExC_paren_name_list= newAV();
8883 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
8886 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
8888 sv_dat = HeVAL(he_str);
8890 /* croak baby croak */
8892 "panic: paren_name hash element allocation failed");
8893 } else if ( SvPOK(sv_dat) ) {
8894 /* (?|...) can mean we have dupes so scan to check
8895 its already been stored. Maybe a flag indicating
8896 we are inside such a construct would be useful,
8897 but the arrays are likely to be quite small, so
8898 for now we punt -- dmq */
8899 IV count = SvIV(sv_dat);
8900 I32 *pv = (I32*)SvPVX(sv_dat);
8902 for ( i = 0 ; i < count ; i++ ) {
8903 if ( pv[i] == RExC_npar ) {
8909 pv = (I32*)SvGROW(sv_dat, SvCUR(sv_dat) + sizeof(I32)+1);
8910 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
8911 pv[count] = RExC_npar;
8912 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
8915 (void)SvUPGRADE(sv_dat,SVt_PVNV);
8916 sv_setpvn(sv_dat, (char *)&(RExC_npar), sizeof(I32));
8918 SvIV_set(sv_dat, 1);
8921 /* Yes this does cause a memory leak in debugging Perls */
8922 if (!av_store(RExC_paren_name_list, RExC_npar, SvREFCNT_inc(svname)))
8923 SvREFCNT_dec_NN(svname);
8926 /*sv_dump(sv_dat);*/
8928 nextchar(pRExC_state);
8930 goto capturing_parens;
8932 RExC_seen |= REG_SEEN_LOOKBEHIND;
8933 RExC_in_lookbehind++;
8935 case '=': /* (?=...) */
8936 RExC_seen_zerolen++;
8938 case '!': /* (?!...) */
8939 RExC_seen_zerolen++;
8940 if (*RExC_parse == ')') {
8941 ret=reg_node(pRExC_state, OPFAIL);
8942 nextchar(pRExC_state);
8946 case '|': /* (?|...) */
8947 /* branch reset, behave like a (?:...) except that
8948 buffers in alternations share the same numbers */
8950 after_freeze = freeze_paren = RExC_npar;
8952 case ':': /* (?:...) */
8953 case '>': /* (?>...) */
8955 case '$': /* (?$...) */
8956 case '@': /* (?@...) */
8957 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
8959 case '#': /* (?#...) */
8960 /* XXX As soon as we disallow separating the '?' and '*' (by
8961 * spaces or (?#...) comment), it is believed that this case
8962 * will be unreachable and can be removed. See
8964 while (*RExC_parse && *RExC_parse != ')')
8966 if (*RExC_parse != ')')
8967 FAIL("Sequence (?#... not terminated");
8968 nextchar(pRExC_state);
8971 case '0' : /* (?0) */
8972 case 'R' : /* (?R) */
8973 if (*RExC_parse != ')')
8974 FAIL("Sequence (?R) not terminated");
8975 ret = reg_node(pRExC_state, GOSTART);
8976 *flagp |= POSTPONED;
8977 nextchar(pRExC_state);
8980 { /* named and numeric backreferences */
8982 case '&': /* (?&NAME) */
8983 parse_start = RExC_parse - 1;
8986 SV *sv_dat = reg_scan_name(pRExC_state,
8987 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8988 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8990 goto gen_recurse_regop;
8991 assert(0); /* NOT REACHED */
8993 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8995 vFAIL("Illegal pattern");
8997 goto parse_recursion;
8999 case '-': /* (?-1) */
9000 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
9001 RExC_parse--; /* rewind to let it be handled later */
9005 case '1': case '2': case '3': case '4': /* (?1) */
9006 case '5': case '6': case '7': case '8': case '9':
9009 num = atoi(RExC_parse);
9010 parse_start = RExC_parse - 1; /* MJD */
9011 if (*RExC_parse == '-')
9013 while (isDIGIT(*RExC_parse))
9015 if (*RExC_parse!=')')
9016 vFAIL("Expecting close bracket");
9019 if ( paren == '-' ) {
9021 Diagram of capture buffer numbering.
9022 Top line is the normal capture buffer numbers
9023 Bottom line is the negative indexing as from
9027 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
9031 num = RExC_npar + num;
9034 vFAIL("Reference to nonexistent group");
9036 } else if ( paren == '+' ) {
9037 num = RExC_npar + num - 1;
9040 ret = reganode(pRExC_state, GOSUB, num);
9042 if (num > (I32)RExC_rx->nparens) {
9044 vFAIL("Reference to nonexistent group");
9046 ARG2L_SET( ret, RExC_recurse_count++);
9048 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9049 "Recurse #%"UVuf" to %"IVdf"\n", (UV)ARG(ret), (IV)ARG2L(ret)));
9053 RExC_seen |= REG_SEEN_RECURSE;
9054 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
9055 Set_Node_Offset(ret, parse_start); /* MJD */
9057 *flagp |= POSTPONED;
9058 nextchar(pRExC_state);
9060 } /* named and numeric backreferences */
9061 assert(0); /* NOT REACHED */
9063 case '?': /* (??...) */
9065 if (*RExC_parse != '{') {
9067 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
9070 *flagp |= POSTPONED;
9071 paren = *RExC_parse++;
9073 case '{': /* (?{...}) */
9076 struct reg_code_block *cb;
9078 RExC_seen_zerolen++;
9080 if ( !pRExC_state->num_code_blocks
9081 || pRExC_state->code_index >= pRExC_state->num_code_blocks
9082 || pRExC_state->code_blocks[pRExC_state->code_index].start
9083 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
9086 if (RExC_pm_flags & PMf_USE_RE_EVAL)
9087 FAIL("panic: Sequence (?{...}): no code block found\n");
9088 FAIL("Eval-group not allowed at runtime, use re 'eval'");
9090 /* this is a pre-compiled code block (?{...}) */
9091 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
9092 RExC_parse = RExC_start + cb->end;
9095 if (cb->src_regex) {
9096 n = add_data(pRExC_state, 2, "rl");
9097 RExC_rxi->data->data[n] =
9098 (void*)SvREFCNT_inc((SV*)cb->src_regex);
9099 RExC_rxi->data->data[n+1] = (void*)o;
9102 n = add_data(pRExC_state, 1,
9103 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l");
9104 RExC_rxi->data->data[n] = (void*)o;
9107 pRExC_state->code_index++;
9108 nextchar(pRExC_state);
9112 ret = reg_node(pRExC_state, LOGICAL);
9113 eval = reganode(pRExC_state, EVAL, n);
9116 /* for later propagation into (??{}) return value */
9117 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
9119 REGTAIL(pRExC_state, ret, eval);
9120 /* deal with the length of this later - MJD */
9123 ret = reganode(pRExC_state, EVAL, n);
9124 Set_Node_Length(ret, RExC_parse - parse_start + 1);
9125 Set_Node_Offset(ret, parse_start);
9128 case '(': /* (?(?{...})...) and (?(?=...)...) */
9131 if (RExC_parse[0] == '?') { /* (?(?...)) */
9132 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
9133 || RExC_parse[1] == '<'
9134 || RExC_parse[1] == '{') { /* Lookahead or eval. */
9138 ret = reg_node(pRExC_state, LOGICAL);
9142 tail = reg(pRExC_state, 1, &flag, depth+1);
9143 if (flag & RESTART_UTF8) {
9144 *flagp = RESTART_UTF8;
9147 REGTAIL(pRExC_state, ret, tail);
9151 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9152 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9154 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9155 char *name_start= RExC_parse++;
9157 SV *sv_dat=reg_scan_name(pRExC_state,
9158 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9159 if (RExC_parse == name_start || *RExC_parse != ch)
9160 vFAIL2("Sequence (?(%c... not terminated",
9161 (ch == '>' ? '<' : ch));
9164 num = add_data( pRExC_state, 1, "S" );
9165 RExC_rxi->data->data[num]=(void*)sv_dat;
9166 SvREFCNT_inc_simple_void(sv_dat);
9168 ret = reganode(pRExC_state,NGROUPP,num);
9169 goto insert_if_check_paren;
9171 else if (RExC_parse[0] == 'D' &&
9172 RExC_parse[1] == 'E' &&
9173 RExC_parse[2] == 'F' &&
9174 RExC_parse[3] == 'I' &&
9175 RExC_parse[4] == 'N' &&
9176 RExC_parse[5] == 'E')
9178 ret = reganode(pRExC_state,DEFINEP,0);
9181 goto insert_if_check_paren;
9183 else if (RExC_parse[0] == 'R') {
9186 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9187 parno = atoi(RExC_parse++);
9188 while (isDIGIT(*RExC_parse))
9190 } else if (RExC_parse[0] == '&') {
9193 sv_dat = reg_scan_name(pRExC_state,
9194 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9195 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9197 ret = reganode(pRExC_state,INSUBP,parno);
9198 goto insert_if_check_paren;
9200 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9203 parno = atoi(RExC_parse++);
9205 while (isDIGIT(*RExC_parse))
9207 ret = reganode(pRExC_state, GROUPP, parno);
9209 insert_if_check_paren:
9210 if ((c = *nextchar(pRExC_state)) != ')')
9211 vFAIL("Switch condition not recognized");
9213 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9214 br = regbranch(pRExC_state, &flags, 1,depth+1);
9216 if (flags & RESTART_UTF8) {
9217 *flagp = RESTART_UTF8;
9220 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9223 REGTAIL(pRExC_state, br, reganode(pRExC_state, LONGJMP, 0));
9224 c = *nextchar(pRExC_state);
9229 vFAIL("(?(DEFINE)....) does not allow branches");
9230 lastbr = reganode(pRExC_state, IFTHEN, 0); /* Fake one for optimizer. */
9231 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
9232 if (flags & RESTART_UTF8) {
9233 *flagp = RESTART_UTF8;
9236 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9239 REGTAIL(pRExC_state, ret, lastbr);
9242 c = *nextchar(pRExC_state);
9247 vFAIL("Switch (?(condition)... contains too many branches");
9248 ender = reg_node(pRExC_state, TAIL);
9249 REGTAIL(pRExC_state, br, ender);
9251 REGTAIL(pRExC_state, lastbr, ender);
9252 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
9255 REGTAIL(pRExC_state, ret, ender);
9256 RExC_size++; /* XXX WHY do we need this?!!
9257 For large programs it seems to be required
9258 but I can't figure out why. -- dmq*/
9262 vFAIL2("Unknown switch condition (?(%.2s", RExC_parse);
9265 case '[': /* (?[ ... ]) */
9266 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
9269 RExC_parse--; /* for vFAIL to print correctly */
9270 vFAIL("Sequence (? incomplete");
9272 default: /* e.g., (?i) */
9275 parse_lparen_question_flags(pRExC_state);
9276 if (UCHARAT(RExC_parse) != ':') {
9277 nextchar(pRExC_state);
9282 nextchar(pRExC_state);
9292 ret = reganode(pRExC_state, OPEN, parno);
9295 RExC_nestroot = parno;
9296 if (RExC_seen & REG_SEEN_RECURSE
9297 && !RExC_open_parens[parno-1])
9299 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9300 "Setting open paren #%"IVdf" to %d\n",
9301 (IV)parno, REG_NODE_NUM(ret)));
9302 RExC_open_parens[parno-1]= ret;
9305 Set_Node_Length(ret, 1); /* MJD */
9306 Set_Node_Offset(ret, RExC_parse); /* MJD */
9314 /* Pick up the branches, linking them together. */
9315 parse_start = RExC_parse; /* MJD */
9316 br = regbranch(pRExC_state, &flags, 1,depth+1);
9318 /* branch_len = (paren != 0); */
9321 if (flags & RESTART_UTF8) {
9322 *flagp = RESTART_UTF8;
9325 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
9327 if (*RExC_parse == '|') {
9328 if (!SIZE_ONLY && RExC_extralen) {
9329 reginsert(pRExC_state, BRANCHJ, br, depth+1);
9332 reginsert(pRExC_state, BRANCH, br, depth+1);
9333 Set_Node_Length(br, paren != 0);
9334 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
9338 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
9340 else if (paren == ':') {
9341 *flagp |= flags&SIMPLE;
9343 if (is_open) { /* Starts with OPEN. */
9344 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
9346 else if (paren != '?') /* Not Conditional */
9348 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9350 while (*RExC_parse == '|') {
9351 if (!SIZE_ONLY && RExC_extralen) {
9352 ender = reganode(pRExC_state, LONGJMP,0);
9353 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender); /* Append to the previous. */
9356 RExC_extralen += 2; /* Account for LONGJMP. */
9357 nextchar(pRExC_state);
9359 if (RExC_npar > after_freeze)
9360 after_freeze = RExC_npar;
9361 RExC_npar = freeze_paren;
9363 br = regbranch(pRExC_state, &flags, 0, depth+1);
9366 if (flags & RESTART_UTF8) {
9367 *flagp = RESTART_UTF8;
9370 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
9372 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
9374 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9377 if (have_branch || paren != ':') {
9378 /* Make a closing node, and hook it on the end. */
9381 ender = reg_node(pRExC_state, TAIL);
9384 ender = reganode(pRExC_state, CLOSE, parno);
9385 if (!SIZE_ONLY && RExC_seen & REG_SEEN_RECURSE) {
9386 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9387 "Setting close paren #%"IVdf" to %d\n",
9388 (IV)parno, REG_NODE_NUM(ender)));
9389 RExC_close_parens[parno-1]= ender;
9390 if (RExC_nestroot == parno)
9393 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
9394 Set_Node_Length(ender,1); /* MJD */
9400 *flagp &= ~HASWIDTH;
9403 ender = reg_node(pRExC_state, SUCCEED);
9406 ender = reg_node(pRExC_state, END);
9408 assert(!RExC_opend); /* there can only be one! */
9413 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9414 SV * const mysv_val1=sv_newmortal();
9415 SV * const mysv_val2=sv_newmortal();
9416 DEBUG_PARSE_MSG("lsbr");
9417 regprop(RExC_rx, mysv_val1, lastbr);
9418 regprop(RExC_rx, mysv_val2, ender);
9419 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9420 SvPV_nolen_const(mysv_val1),
9421 (IV)REG_NODE_NUM(lastbr),
9422 SvPV_nolen_const(mysv_val2),
9423 (IV)REG_NODE_NUM(ender),
9424 (IV)(ender - lastbr)
9427 REGTAIL(pRExC_state, lastbr, ender);
9429 if (have_branch && !SIZE_ONLY) {
9432 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
9434 /* Hook the tails of the branches to the closing node. */
9435 for (br = ret; br; br = regnext(br)) {
9436 const U8 op = PL_regkind[OP(br)];
9438 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
9439 if (OP(NEXTOPER(br)) != NOTHING || regnext(NEXTOPER(br)) != ender)
9442 else if (op == BRANCHJ) {
9443 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
9444 /* for now we always disable this optimisation * /
9445 if (OP(NEXTOPER(NEXTOPER(br))) != NOTHING || regnext(NEXTOPER(NEXTOPER(br))) != ender)
9451 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
9452 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9453 SV * const mysv_val1=sv_newmortal();
9454 SV * const mysv_val2=sv_newmortal();
9455 DEBUG_PARSE_MSG("NADA");
9456 regprop(RExC_rx, mysv_val1, ret);
9457 regprop(RExC_rx, mysv_val2, ender);
9458 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9459 SvPV_nolen_const(mysv_val1),
9460 (IV)REG_NODE_NUM(ret),
9461 SvPV_nolen_const(mysv_val2),
9462 (IV)REG_NODE_NUM(ender),
9467 if (OP(ender) == TAIL) {
9472 for ( opt= br + 1; opt < ender ; opt++ )
9474 NEXT_OFF(br)= ender - br;
9482 static const char parens[] = "=!<,>";
9484 if (paren && (p = strchr(parens, paren))) {
9485 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
9486 int flag = (p - parens) > 1;
9489 node = SUSPEND, flag = 0;
9490 reginsert(pRExC_state, node,ret, depth+1);
9491 Set_Node_Cur_Length(ret, parse_start);
9492 Set_Node_Offset(ret, parse_start + 1);
9494 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
9498 /* Check for proper termination. */
9500 /* restore original flags, but keep (?p) */
9501 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
9502 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
9503 RExC_parse = oregcomp_parse;
9504 vFAIL("Unmatched (");
9507 else if (!paren && RExC_parse < RExC_end) {
9508 if (*RExC_parse == ')') {
9510 vFAIL("Unmatched )");
9513 FAIL("Junk on end of regexp"); /* "Can't happen". */
9514 assert(0); /* NOTREACHED */
9517 if (RExC_in_lookbehind) {
9518 RExC_in_lookbehind--;
9520 if (after_freeze > RExC_npar)
9521 RExC_npar = after_freeze;
9526 - regbranch - one alternative of an | operator
9528 * Implements the concatenation operator.
9530 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9534 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
9538 regnode *chain = NULL;
9540 I32 flags = 0, c = 0;
9541 GET_RE_DEBUG_FLAGS_DECL;
9543 PERL_ARGS_ASSERT_REGBRANCH;
9545 DEBUG_PARSE("brnc");
9550 if (!SIZE_ONLY && RExC_extralen)
9551 ret = reganode(pRExC_state, BRANCHJ,0);
9553 ret = reg_node(pRExC_state, BRANCH);
9554 Set_Node_Length(ret, 1);
9558 if (!first && SIZE_ONLY)
9559 RExC_extralen += 1; /* BRANCHJ */
9561 *flagp = WORST; /* Tentatively. */
9564 nextchar(pRExC_state);
9565 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
9567 latest = regpiece(pRExC_state, &flags,depth+1);
9568 if (latest == NULL) {
9569 if (flags & TRYAGAIN)
9571 if (flags & RESTART_UTF8) {
9572 *flagp = RESTART_UTF8;
9575 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
9577 else if (ret == NULL)
9579 *flagp |= flags&(HASWIDTH|POSTPONED);
9580 if (chain == NULL) /* First piece. */
9581 *flagp |= flags&SPSTART;
9584 REGTAIL(pRExC_state, chain, latest);
9589 if (chain == NULL) { /* Loop ran zero times. */
9590 chain = reg_node(pRExC_state, NOTHING);
9595 *flagp |= flags&SIMPLE;
9602 - regpiece - something followed by possible [*+?]
9604 * Note that the branching code sequences used for ? and the general cases
9605 * of * and + are somewhat optimized: they use the same NOTHING node as
9606 * both the endmarker for their branch list and the body of the last branch.
9607 * It might seem that this node could be dispensed with entirely, but the
9608 * endmarker role is not redundant.
9610 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
9612 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9616 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
9623 const char * const origparse = RExC_parse;
9625 I32 max = REG_INFTY;
9626 #ifdef RE_TRACK_PATTERN_OFFSETS
9629 const char *maxpos = NULL;
9631 /* Save the original in case we change the emitted regop to a FAIL. */
9632 regnode * const orig_emit = RExC_emit;
9634 GET_RE_DEBUG_FLAGS_DECL;
9636 PERL_ARGS_ASSERT_REGPIECE;
9638 DEBUG_PARSE("piec");
9640 ret = regatom(pRExC_state, &flags,depth+1);
9642 if (flags & (TRYAGAIN|RESTART_UTF8))
9643 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
9645 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
9651 if (op == '{' && regcurly(RExC_parse, FALSE)) {
9653 #ifdef RE_TRACK_PATTERN_OFFSETS
9654 parse_start = RExC_parse; /* MJD */
9656 next = RExC_parse + 1;
9657 while (isDIGIT(*next) || *next == ',') {
9666 if (*next == '}') { /* got one */
9670 min = atoi(RExC_parse);
9674 maxpos = RExC_parse;
9676 if (!max && *maxpos != '0')
9677 max = REG_INFTY; /* meaning "infinity" */
9678 else if (max >= REG_INFTY)
9679 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
9681 nextchar(pRExC_state);
9682 if (max < min) { /* If can't match, warn and optimize to fail
9685 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
9687 /* We can't back off the size because we have to reserve
9688 * enough space for all the things we are about to throw
9689 * away, but we can shrink it by the ammount we are about
9691 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
9694 RExC_emit = orig_emit;
9696 ret = reg_node(pRExC_state, OPFAIL);
9701 if ((flags&SIMPLE)) {
9702 RExC_naughty += 2 + RExC_naughty / 2;
9703 reginsert(pRExC_state, CURLY, ret, depth+1);
9704 Set_Node_Offset(ret, parse_start+1); /* MJD */
9705 Set_Node_Cur_Length(ret, parse_start);
9708 regnode * const w = reg_node(pRExC_state, WHILEM);
9711 REGTAIL(pRExC_state, ret, w);
9712 if (!SIZE_ONLY && RExC_extralen) {
9713 reginsert(pRExC_state, LONGJMP,ret, depth+1);
9714 reginsert(pRExC_state, NOTHING,ret, depth+1);
9715 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
9717 reginsert(pRExC_state, CURLYX,ret, depth+1);
9719 Set_Node_Offset(ret, parse_start+1);
9720 Set_Node_Length(ret,
9721 op == '{' ? (RExC_parse - parse_start) : 1);
9723 if (!SIZE_ONLY && RExC_extralen)
9724 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
9725 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
9727 RExC_whilem_seen++, RExC_extralen += 3;
9728 RExC_naughty += 4 + RExC_naughty; /* compound interest */
9737 ARG1_SET(ret, (U16)min);
9738 ARG2_SET(ret, (U16)max);
9750 #if 0 /* Now runtime fix should be reliable. */
9752 /* if this is reinstated, don't forget to put this back into perldiag:
9754 =item Regexp *+ operand could be empty at {#} in regex m/%s/
9756 (F) The part of the regexp subject to either the * or + quantifier
9757 could match an empty string. The {#} shows in the regular
9758 expression about where the problem was discovered.
9762 if (!(flags&HASWIDTH) && op != '?')
9763 vFAIL("Regexp *+ operand could be empty");
9766 #ifdef RE_TRACK_PATTERN_OFFSETS
9767 parse_start = RExC_parse;
9769 nextchar(pRExC_state);
9771 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
9773 if (op == '*' && (flags&SIMPLE)) {
9774 reginsert(pRExC_state, STAR, ret, depth+1);
9778 else if (op == '*') {
9782 else if (op == '+' && (flags&SIMPLE)) {
9783 reginsert(pRExC_state, PLUS, ret, depth+1);
9787 else if (op == '+') {
9791 else if (op == '?') {
9796 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
9797 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
9798 ckWARN3reg(RExC_parse,
9799 "%.*s matches null string many times",
9800 (int)(RExC_parse >= origparse ? RExC_parse - origparse : 0),
9802 (void)ReREFCNT_inc(RExC_rx_sv);
9805 if (RExC_parse < RExC_end && *RExC_parse == '?') {
9806 nextchar(pRExC_state);
9807 reginsert(pRExC_state, MINMOD, ret, depth+1);
9808 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
9811 if (RExC_parse < RExC_end && *RExC_parse == '+') {
9813 nextchar(pRExC_state);
9814 ender = reg_node(pRExC_state, SUCCEED);
9815 REGTAIL(pRExC_state, ret, ender);
9816 reginsert(pRExC_state, SUSPEND, ret, depth+1);
9818 ender = reg_node(pRExC_state, TAIL);
9819 REGTAIL(pRExC_state, ret, ender);
9822 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
9824 vFAIL("Nested quantifiers");
9831 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p, UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
9832 const bool strict /* Apply stricter parsing rules? */
9836 /* This is expected to be called by a parser routine that has recognized '\N'
9837 and needs to handle the rest. RExC_parse is expected to point at the first
9838 char following the N at the time of the call. On successful return,
9839 RExC_parse has been updated to point to just after the sequence identified
9840 by this routine, and <*flagp> has been updated.
9842 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
9845 \N may begin either a named sequence, or if outside a character class, mean
9846 to match a non-newline. For non single-quoted regexes, the tokenizer has
9847 attempted to decide which, and in the case of a named sequence, converted it
9848 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
9849 where c1... are the characters in the sequence. For single-quoted regexes,
9850 the tokenizer passes the \N sequence through unchanged; this code will not
9851 attempt to determine this nor expand those, instead raising a syntax error.
9852 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
9853 or there is no '}', it signals that this \N occurrence means to match a
9856 Only the \N{U+...} form should occur in a character class, for the same
9857 reason that '.' inside a character class means to just match a period: it
9858 just doesn't make sense.
9860 The function raises an error (via vFAIL), and doesn't return for various
9861 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
9862 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
9863 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
9864 only possible if node_p is non-NULL.
9867 If <valuep> is non-null, it means the caller can accept an input sequence
9868 consisting of a just a single code point; <*valuep> is set to that value
9869 if the input is such.
9871 If <node_p> is non-null it signifies that the caller can accept any other
9872 legal sequence (i.e., one that isn't just a single code point). <*node_p>
9874 1) \N means not-a-NL: points to a newly created REG_ANY node;
9875 2) \N{}: points to a new NOTHING node;
9876 3) otherwise: points to a new EXACT node containing the resolved
9878 Note that FALSE is returned for single code point sequences if <valuep> is
9882 char * endbrace; /* '}' following the name */
9884 char *endchar; /* Points to '.' or '}' ending cur char in the input
9886 bool has_multiple_chars; /* true if the input stream contains a sequence of
9887 more than one character */
9889 GET_RE_DEBUG_FLAGS_DECL;
9891 PERL_ARGS_ASSERT_GROK_BSLASH_N;
9895 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
9897 /* The [^\n] meaning of \N ignores spaces and comments under the /x
9898 * modifier. The other meaning does not */
9899 p = (RExC_flags & RXf_PMf_EXTENDED)
9900 ? regwhite( pRExC_state, RExC_parse )
9903 /* Disambiguate between \N meaning a named character versus \N meaning
9904 * [^\n]. The former is assumed when it can't be the latter. */
9905 if (*p != '{' || regcurly(p, FALSE)) {
9908 /* no bare \N in a charclass */
9909 if (in_char_class) {
9910 vFAIL("\\N in a character class must be a named character: \\N{...}");
9914 nextchar(pRExC_state);
9915 *node_p = reg_node(pRExC_state, REG_ANY);
9916 *flagp |= HASWIDTH|SIMPLE;
9919 Set_Node_Length(*node_p, 1); /* MJD */
9923 /* Here, we have decided it should be a named character or sequence */
9925 /* The test above made sure that the next real character is a '{', but
9926 * under the /x modifier, it could be separated by space (or a comment and
9927 * \n) and this is not allowed (for consistency with \x{...} and the
9928 * tokenizer handling of \N{NAME}). */
9929 if (*RExC_parse != '{') {
9930 vFAIL("Missing braces on \\N{}");
9933 RExC_parse++; /* Skip past the '{' */
9935 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
9936 || ! (endbrace == RExC_parse /* nothing between the {} */
9937 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below */
9938 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg) */
9940 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
9941 vFAIL("\\N{NAME} must be resolved by the lexer");
9944 if (endbrace == RExC_parse) { /* empty: \N{} */
9947 *node_p = reg_node(pRExC_state,NOTHING);
9949 else if (in_char_class) {
9950 if (SIZE_ONLY && in_char_class) {
9952 RExC_parse++; /* Position after the "}" */
9953 vFAIL("Zero length \\N{}");
9956 ckWARNreg(RExC_parse,
9957 "Ignoring zero length \\N{} in character class");
9965 nextchar(pRExC_state);
9969 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
9970 RExC_parse += 2; /* Skip past the 'U+' */
9972 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9974 /* Code points are separated by dots. If none, there is only one code
9975 * point, and is terminated by the brace */
9976 has_multiple_chars = (endchar < endbrace);
9978 if (valuep && (! has_multiple_chars || in_char_class)) {
9979 /* We only pay attention to the first char of
9980 multichar strings being returned in char classes. I kinda wonder
9981 if this makes sense as it does change the behaviour
9982 from earlier versions, OTOH that behaviour was broken
9983 as well. XXX Solution is to recharacterize as
9984 [rest-of-class]|multi1|multi2... */
9986 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
9987 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
9988 | PERL_SCAN_DISALLOW_PREFIX
9989 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
9991 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
9993 /* The tokenizer should have guaranteed validity, but it's possible to
9994 * bypass it by using single quoting, so check */
9995 if (length_of_hex == 0
9996 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
9998 RExC_parse += length_of_hex; /* Includes all the valid */
9999 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
10000 ? UTF8SKIP(RExC_parse)
10002 /* Guard against malformed utf8 */
10003 if (RExC_parse >= endchar) {
10004 RExC_parse = endchar;
10006 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10009 if (in_char_class && has_multiple_chars) {
10011 RExC_parse = endbrace;
10012 vFAIL("\\N{} in character class restricted to one character");
10015 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
10019 RExC_parse = endbrace + 1;
10021 else if (! node_p || ! has_multiple_chars) {
10023 /* Here, the input is legal, but not according to the caller's
10024 * options. We fail without advancing the parse, so that the
10025 * caller can try again */
10031 /* What is done here is to convert this to a sub-pattern of the form
10032 * (?:\x{char1}\x{char2}...)
10033 * and then call reg recursively. That way, it retains its atomicness,
10034 * while not having to worry about special handling that some code
10035 * points may have. toke.c has converted the original Unicode values
10036 * to native, so that we can just pass on the hex values unchanged. We
10037 * do have to set a flag to keep recoding from happening in the
10040 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
10042 char *orig_end = RExC_end;
10045 while (RExC_parse < endbrace) {
10047 /* Convert to notation the rest of the code understands */
10048 sv_catpv(substitute_parse, "\\x{");
10049 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
10050 sv_catpv(substitute_parse, "}");
10052 /* Point to the beginning of the next character in the sequence. */
10053 RExC_parse = endchar + 1;
10054 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10056 sv_catpv(substitute_parse, ")");
10058 RExC_parse = SvPV(substitute_parse, len);
10060 /* Don't allow empty number */
10062 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10064 RExC_end = RExC_parse + len;
10066 /* The values are Unicode, and therefore not subject to recoding */
10067 RExC_override_recoding = 1;
10069 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
10070 if (flags & RESTART_UTF8) {
10071 *flagp = RESTART_UTF8;
10074 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
10077 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10079 RExC_parse = endbrace;
10080 RExC_end = orig_end;
10081 RExC_override_recoding = 0;
10083 nextchar(pRExC_state);
10093 * It returns the code point in utf8 for the value in *encp.
10094 * value: a code value in the source encoding
10095 * encp: a pointer to an Encode object
10097 * If the result from Encode is not a single character,
10098 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
10101 S_reg_recode(pTHX_ const char value, SV **encp)
10104 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
10105 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
10106 const STRLEN newlen = SvCUR(sv);
10107 UV uv = UNICODE_REPLACEMENT;
10109 PERL_ARGS_ASSERT_REG_RECODE;
10113 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
10116 if (!newlen || numlen != newlen) {
10117 uv = UNICODE_REPLACEMENT;
10123 PERL_STATIC_INLINE U8
10124 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
10128 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
10134 op = get_regex_charset(RExC_flags);
10135 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
10136 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
10137 been, so there is no hole */
10140 return op + EXACTF;
10143 PERL_STATIC_INLINE void
10144 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state, regnode *node, I32* flagp, STRLEN len, UV code_point)
10146 /* This knows the details about sizing an EXACTish node, setting flags for
10147 * it (by setting <*flagp>, and potentially populating it with a single
10150 * If <len> (the length in bytes) is non-zero, this function assumes that
10151 * the node has already been populated, and just does the sizing. In this
10152 * case <code_point> should be the final code point that has already been
10153 * placed into the node. This value will be ignored except that under some
10154 * circumstances <*flagp> is set based on it.
10156 * If <len> is zero, the function assumes that the node is to contain only
10157 * the single character given by <code_point> and calculates what <len>
10158 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10159 * additionally will populate the node's STRING with <code_point>, if <len>
10160 * is 0. In both cases <*flagp> is appropriately set
10162 * It knows that under FOLD, the Latin Sharp S and UTF characters above
10163 * 255, must be folded (the former only when the rules indicate it can
10166 bool len_passed_in = cBOOL(len != 0);
10167 U8 character[UTF8_MAXBYTES_CASE+1];
10169 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10171 if (! len_passed_in) {
10173 if (FOLD && (! LOC || code_point > 255)) {
10174 _to_uni_fold_flags(NATIVE_TO_UNI(code_point),
10177 FOLD_FLAGS_FULL | ((LOC)
10178 ? FOLD_FLAGS_LOCALE
10179 : (ASCII_FOLD_RESTRICTED)
10180 ? FOLD_FLAGS_NOMIX_ASCII
10184 uvchr_to_utf8( character, code_point);
10185 len = UTF8SKIP(character);
10189 || code_point != LATIN_SMALL_LETTER_SHARP_S
10190 || ASCII_FOLD_RESTRICTED
10191 || ! AT_LEAST_UNI_SEMANTICS)
10193 *character = (U8) code_point;
10198 *(character + 1) = 's';
10204 RExC_size += STR_SZ(len);
10207 RExC_emit += STR_SZ(len);
10208 STR_LEN(node) = len;
10209 if (! len_passed_in) {
10210 Copy((char *) character, STRING(node), len, char);
10214 *flagp |= HASWIDTH;
10216 /* A single character node is SIMPLE, except for the special-cased SHARP S
10218 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
10219 && (code_point != LATIN_SMALL_LETTER_SHARP_S
10220 || ! FOLD || ! DEPENDS_SEMANTICS))
10227 - regatom - the lowest level
10229 Try to identify anything special at the start of the pattern. If there
10230 is, then handle it as required. This may involve generating a single regop,
10231 such as for an assertion; or it may involve recursing, such as to
10232 handle a () structure.
10234 If the string doesn't start with something special then we gobble up
10235 as much literal text as we can.
10237 Once we have been able to handle whatever type of thing started the
10238 sequence, we return.
10240 Note: we have to be careful with escapes, as they can be both literal
10241 and special, and in the case of \10 and friends, context determines which.
10243 A summary of the code structure is:
10245 switch (first_byte) {
10246 cases for each special:
10247 handle this special;
10250 switch (2nd byte) {
10251 cases for each unambiguous special:
10252 handle this special;
10254 cases for each ambigous special/literal:
10256 if (special) handle here
10258 default: // unambiguously literal:
10261 default: // is a literal char
10264 create EXACTish node for literal;
10265 while (more input and node isn't full) {
10266 switch (input_byte) {
10267 cases for each special;
10268 make sure parse pointer is set so that the next call to
10269 regatom will see this special first
10270 goto loopdone; // EXACTish node terminated by prev. char
10272 append char to EXACTISH node;
10274 get next input byte;
10278 return the generated node;
10280 Specifically there are two separate switches for handling
10281 escape sequences, with the one for handling literal escapes requiring
10282 a dummy entry for all of the special escapes that are actually handled
10285 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
10287 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10289 Otherwise does not return NULL.
10293 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10296 regnode *ret = NULL;
10298 char *parse_start = RExC_parse;
10302 GET_RE_DEBUG_FLAGS_DECL;
10304 *flagp = WORST; /* Tentatively. */
10306 DEBUG_PARSE("atom");
10308 PERL_ARGS_ASSERT_REGATOM;
10311 switch ((U8)*RExC_parse) {
10313 RExC_seen_zerolen++;
10314 nextchar(pRExC_state);
10315 if (RExC_flags & RXf_PMf_MULTILINE)
10316 ret = reg_node(pRExC_state, MBOL);
10317 else if (RExC_flags & RXf_PMf_SINGLELINE)
10318 ret = reg_node(pRExC_state, SBOL);
10320 ret = reg_node(pRExC_state, BOL);
10321 Set_Node_Length(ret, 1); /* MJD */
10324 nextchar(pRExC_state);
10326 RExC_seen_zerolen++;
10327 if (RExC_flags & RXf_PMf_MULTILINE)
10328 ret = reg_node(pRExC_state, MEOL);
10329 else if (RExC_flags & RXf_PMf_SINGLELINE)
10330 ret = reg_node(pRExC_state, SEOL);
10332 ret = reg_node(pRExC_state, EOL);
10333 Set_Node_Length(ret, 1); /* MJD */
10336 nextchar(pRExC_state);
10337 if (RExC_flags & RXf_PMf_SINGLELINE)
10338 ret = reg_node(pRExC_state, SANY);
10340 ret = reg_node(pRExC_state, REG_ANY);
10341 *flagp |= HASWIDTH|SIMPLE;
10343 Set_Node_Length(ret, 1); /* MJD */
10347 char * const oregcomp_parse = ++RExC_parse;
10348 ret = regclass(pRExC_state, flagp,depth+1,
10349 FALSE, /* means parse the whole char class */
10350 TRUE, /* allow multi-char folds */
10351 FALSE, /* don't silence non-portable warnings. */
10353 if (*RExC_parse != ']') {
10354 RExC_parse = oregcomp_parse;
10355 vFAIL("Unmatched [");
10358 if (*flagp & RESTART_UTF8)
10360 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
10363 nextchar(pRExC_state);
10364 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
10368 nextchar(pRExC_state);
10369 ret = reg(pRExC_state, 2, &flags,depth+1);
10371 if (flags & TRYAGAIN) {
10372 if (RExC_parse == RExC_end) {
10373 /* Make parent create an empty node if needed. */
10374 *flagp |= TRYAGAIN;
10379 if (flags & RESTART_UTF8) {
10380 *flagp = RESTART_UTF8;
10383 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"", (UV) flags);
10385 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10389 if (flags & TRYAGAIN) {
10390 *flagp |= TRYAGAIN;
10393 vFAIL("Internal urp");
10394 /* Supposed to be caught earlier. */
10397 if (!regcurly(RExC_parse, FALSE)) {
10406 vFAIL("Quantifier follows nothing");
10411 This switch handles escape sequences that resolve to some kind
10412 of special regop and not to literal text. Escape sequnces that
10413 resolve to literal text are handled below in the switch marked
10416 Every entry in this switch *must* have a corresponding entry
10417 in the literal escape switch. However, the opposite is not
10418 required, as the default for this switch is to jump to the
10419 literal text handling code.
10421 switch ((U8)*++RExC_parse) {
10423 /* Special Escapes */
10425 RExC_seen_zerolen++;
10426 ret = reg_node(pRExC_state, SBOL);
10428 goto finish_meta_pat;
10430 ret = reg_node(pRExC_state, GPOS);
10431 RExC_seen |= REG_SEEN_GPOS;
10433 goto finish_meta_pat;
10435 RExC_seen_zerolen++;
10436 ret = reg_node(pRExC_state, KEEPS);
10438 /* XXX:dmq : disabling in-place substitution seems to
10439 * be necessary here to avoid cases of memory corruption, as
10440 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
10442 RExC_seen |= REG_SEEN_LOOKBEHIND;
10443 goto finish_meta_pat;
10445 ret = reg_node(pRExC_state, SEOL);
10447 RExC_seen_zerolen++; /* Do not optimize RE away */
10448 goto finish_meta_pat;
10450 ret = reg_node(pRExC_state, EOS);
10452 RExC_seen_zerolen++; /* Do not optimize RE away */
10453 goto finish_meta_pat;
10455 ret = reg_node(pRExC_state, CANY);
10456 RExC_seen |= REG_SEEN_CANY;
10457 *flagp |= HASWIDTH|SIMPLE;
10458 goto finish_meta_pat;
10460 ret = reg_node(pRExC_state, CLUMP);
10461 *flagp |= HASWIDTH;
10462 goto finish_meta_pat;
10468 arg = ANYOF_WORDCHAR;
10472 RExC_seen_zerolen++;
10473 RExC_seen |= REG_SEEN_LOOKBEHIND;
10474 op = BOUND + get_regex_charset(RExC_flags);
10475 if (op > BOUNDA) { /* /aa is same as /a */
10478 ret = reg_node(pRExC_state, op);
10479 FLAGS(ret) = get_regex_charset(RExC_flags);
10481 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10482 ckWARNdep(RExC_parse, "\"\\b{\" is deprecated; use \"\\b\\{\" or \"\\b[{]\" instead");
10484 goto finish_meta_pat;
10486 RExC_seen_zerolen++;
10487 RExC_seen |= REG_SEEN_LOOKBEHIND;
10488 op = NBOUND + get_regex_charset(RExC_flags);
10489 if (op > NBOUNDA) { /* /aa is same as /a */
10492 ret = reg_node(pRExC_state, op);
10493 FLAGS(ret) = get_regex_charset(RExC_flags);
10495 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10496 ckWARNdep(RExC_parse, "\"\\B{\" is deprecated; use \"\\B\\{\" or \"\\B[{]\" instead");
10498 goto finish_meta_pat;
10508 ret = reg_node(pRExC_state, LNBREAK);
10509 *flagp |= HASWIDTH|SIMPLE;
10510 goto finish_meta_pat;
10518 goto join_posix_op_known;
10524 arg = ANYOF_VERTWS;
10526 goto join_posix_op_known;
10536 op = POSIXD + get_regex_charset(RExC_flags);
10537 if (op > POSIXA) { /* /aa is same as /a */
10541 join_posix_op_known:
10544 op += NPOSIXD - POSIXD;
10547 ret = reg_node(pRExC_state, op);
10549 FLAGS(ret) = namedclass_to_classnum(arg);
10552 *flagp |= HASWIDTH|SIMPLE;
10556 nextchar(pRExC_state);
10557 Set_Node_Length(ret, 2); /* MJD */
10563 char* parse_start = RExC_parse - 2;
10568 ret = regclass(pRExC_state, flagp,depth+1,
10569 TRUE, /* means just parse this element */
10570 FALSE, /* don't allow multi-char folds */
10571 FALSE, /* don't silence non-portable warnings.
10572 It would be a bug if these returned
10575 /* regclass() can only return RESTART_UTF8 if multi-char folds
10578 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
10583 Set_Node_Offset(ret, parse_start + 2);
10584 Set_Node_Cur_Length(ret, parse_start);
10585 nextchar(pRExC_state);
10589 /* Handle \N and \N{NAME} with multiple code points here and not
10590 * below because it can be multicharacter. join_exact() will join
10591 * them up later on. Also this makes sure that things like
10592 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
10593 * The options to the grok function call causes it to fail if the
10594 * sequence is just a single code point. We then go treat it as
10595 * just another character in the current EXACT node, and hence it
10596 * gets uniform treatment with all the other characters. The
10597 * special treatment for quantifiers is not needed for such single
10598 * character sequences */
10600 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
10601 FALSE /* not strict */ )) {
10602 if (*flagp & RESTART_UTF8)
10608 case 'k': /* Handle \k<NAME> and \k'NAME' */
10611 char ch= RExC_parse[1];
10612 if (ch != '<' && ch != '\'' && ch != '{') {
10614 vFAIL2("Sequence %.2s... not terminated",parse_start);
10616 /* this pretty much dupes the code for (?P=...) in reg(), if
10617 you change this make sure you change that */
10618 char* name_start = (RExC_parse += 2);
10620 SV *sv_dat = reg_scan_name(pRExC_state,
10621 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10622 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
10623 if (RExC_parse == name_start || *RExC_parse != ch)
10624 vFAIL2("Sequence %.3s... not terminated",parse_start);
10627 num = add_data( pRExC_state, 1, "S" );
10628 RExC_rxi->data->data[num]=(void*)sv_dat;
10629 SvREFCNT_inc_simple_void(sv_dat);
10633 ret = reganode(pRExC_state,
10636 : (ASCII_FOLD_RESTRICTED)
10638 : (AT_LEAST_UNI_SEMANTICS)
10644 *flagp |= HASWIDTH;
10646 /* override incorrect value set in reganode MJD */
10647 Set_Node_Offset(ret, parse_start+1);
10648 Set_Node_Cur_Length(ret, parse_start);
10649 nextchar(pRExC_state);
10655 case '1': case '2': case '3': case '4':
10656 case '5': case '6': case '7': case '8': case '9':
10659 bool isg = *RExC_parse == 'g';
10664 if (*RExC_parse == '{') {
10668 if (*RExC_parse == '-') {
10672 if (hasbrace && !isDIGIT(*RExC_parse)) {
10673 if (isrel) RExC_parse--;
10675 goto parse_named_seq;
10677 num = atoi(RExC_parse);
10678 if (isg && num == 0) {
10679 if (*RExC_parse == '0') {
10680 vFAIL("Reference to invalid group 0");
10683 vFAIL("Unterminated \\g... pattern");
10687 num = RExC_npar - num;
10689 vFAIL("Reference to nonexistent or unclosed group");
10691 if (!isg && num > 9 && num >= RExC_npar && *RExC_parse != '8' && *RExC_parse != '9')
10692 /* Probably a character specified in octal, e.g. \35 */
10695 #ifdef RE_TRACK_PATTERN_OFFSETS
10696 char * const parse_start = RExC_parse - 1; /* MJD */
10698 while (isDIGIT(*RExC_parse))
10701 if (*RExC_parse != '}')
10702 vFAIL("Unterminated \\g{...} pattern");
10706 if (num > (I32)RExC_rx->nparens)
10707 vFAIL("Reference to nonexistent group");
10710 ret = reganode(pRExC_state,
10713 : (ASCII_FOLD_RESTRICTED)
10715 : (AT_LEAST_UNI_SEMANTICS)
10721 *flagp |= HASWIDTH;
10723 /* override incorrect value set in reganode MJD */
10724 Set_Node_Offset(ret, parse_start+1);
10725 Set_Node_Cur_Length(ret, parse_start);
10727 nextchar(pRExC_state);
10732 if (RExC_parse >= RExC_end)
10733 FAIL("Trailing \\");
10736 /* Do not generate "unrecognized" warnings here, we fall
10737 back into the quick-grab loop below */
10744 if (RExC_flags & RXf_PMf_EXTENDED) {
10745 if ( reg_skipcomment( pRExC_state ) )
10752 parse_start = RExC_parse - 1;
10761 #define MAX_NODE_STRING_SIZE 127
10762 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
10764 U8 upper_parse = MAX_NODE_STRING_SIZE;
10767 bool next_is_quantifier;
10768 char * oldp = NULL;
10770 /* If a folding node contains only code points that don't
10771 * participate in folds, it can be changed into an EXACT node,
10772 * which allows the optimizer more things to look for */
10776 node_type = compute_EXACTish(pRExC_state);
10777 ret = reg_node(pRExC_state, node_type);
10779 /* In pass1, folded, we use a temporary buffer instead of the
10780 * actual node, as the node doesn't exist yet */
10781 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
10787 /* We do the EXACTFish to EXACT node only if folding, and not if in
10788 * locale, as whether a character folds or not isn't known until
10790 maybe_exact = FOLD && ! LOC;
10792 /* XXX The node can hold up to 255 bytes, yet this only goes to
10793 * 127. I (khw) do not know why. Keeping it somewhat less than
10794 * 255 allows us to not have to worry about overflow due to
10795 * converting to utf8 and fold expansion, but that value is
10796 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
10797 * split up by this limit into a single one using the real max of
10798 * 255. Even at 127, this breaks under rare circumstances. If
10799 * folding, we do not want to split a node at a character that is a
10800 * non-final in a multi-char fold, as an input string could just
10801 * happen to want to match across the node boundary. The join
10802 * would solve that problem if the join actually happens. But a
10803 * series of more than two nodes in a row each of 127 would cause
10804 * the first join to succeed to get to 254, but then there wouldn't
10805 * be room for the next one, which could at be one of those split
10806 * multi-char folds. I don't know of any fool-proof solution. One
10807 * could back off to end with only a code point that isn't such a
10808 * non-final, but it is possible for there not to be any in the
10810 for (p = RExC_parse - 1;
10811 len < upper_parse && p < RExC_end;
10816 if (RExC_flags & RXf_PMf_EXTENDED)
10817 p = regwhite( pRExC_state, p );
10828 /* Literal Escapes Switch
10830 This switch is meant to handle escape sequences that
10831 resolve to a literal character.
10833 Every escape sequence that represents something
10834 else, like an assertion or a char class, is handled
10835 in the switch marked 'Special Escapes' above in this
10836 routine, but also has an entry here as anything that
10837 isn't explicitly mentioned here will be treated as
10838 an unescaped equivalent literal.
10841 switch ((U8)*++p) {
10842 /* These are all the special escapes. */
10843 case 'A': /* Start assertion */
10844 case 'b': case 'B': /* Word-boundary assertion*/
10845 case 'C': /* Single char !DANGEROUS! */
10846 case 'd': case 'D': /* digit class */
10847 case 'g': case 'G': /* generic-backref, pos assertion */
10848 case 'h': case 'H': /* HORIZWS */
10849 case 'k': case 'K': /* named backref, keep marker */
10850 case 'p': case 'P': /* Unicode property */
10851 case 'R': /* LNBREAK */
10852 case 's': case 'S': /* space class */
10853 case 'v': case 'V': /* VERTWS */
10854 case 'w': case 'W': /* word class */
10855 case 'X': /* eXtended Unicode "combining character sequence" */
10856 case 'z': case 'Z': /* End of line/string assertion */
10860 /* Anything after here is an escape that resolves to a
10861 literal. (Except digits, which may or may not)
10867 case 'N': /* Handle a single-code point named character. */
10868 /* The options cause it to fail if a multiple code
10869 * point sequence. Handle those in the switch() above
10871 RExC_parse = p + 1;
10872 if (! grok_bslash_N(pRExC_state, NULL, &ender,
10873 flagp, depth, FALSE,
10874 FALSE /* not strict */ ))
10876 if (*flagp & RESTART_UTF8)
10877 FAIL("panic: grok_bslash_N set RESTART_UTF8");
10878 RExC_parse = p = oldp;
10882 if (ender > 0xff) {
10899 ender = ASCII_TO_NATIVE('\033');
10903 ender = ASCII_TO_NATIVE('\007');
10909 const char* error_msg;
10911 bool valid = grok_bslash_o(&p,
10914 TRUE, /* out warnings */
10915 FALSE, /* not strict */
10916 TRUE, /* Output warnings
10921 RExC_parse = p; /* going to die anyway; point
10922 to exact spot of failure */
10926 if (PL_encoding && ender < 0x100) {
10927 goto recode_encoding;
10929 if (ender > 0xff) {
10936 UV result = UV_MAX; /* initialize to erroneous
10938 const char* error_msg;
10940 bool valid = grok_bslash_x(&p,
10943 TRUE, /* out warnings */
10944 FALSE, /* not strict */
10945 TRUE, /* Output warnings
10950 RExC_parse = p; /* going to die anyway; point
10951 to exact spot of failure */
10956 if (PL_encoding && ender < 0x100) {
10957 goto recode_encoding;
10959 if (ender > 0xff) {
10966 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
10968 case '8': case '9': /* must be a backreference */
10971 case '1': case '2': case '3':case '4':
10972 case '5': case '6': case '7':
10973 /* When we parse backslash escapes there is ambiguity between
10974 * backreferences and octal escapes. Any escape from \1 - \9 is
10975 * a backreference, any multi-digit escape which does not start with
10976 * 0 and which when evaluated as decimal could refer to an already
10977 * parsed capture buffer is a backslash. Anything else is octal.
10979 * Note this implies that \118 could be interpreted as 118 OR as
10980 * "\11" . "8" depending on whether there were 118 capture buffers
10981 * defined already in the pattern.
10983 if ( !isDIGIT(p[1]) || atoi(p) <= RExC_npar )
10984 { /* Not to be treated as an octal constant, go
10991 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
10993 ender = grok_oct(p, &numlen, &flags, NULL);
10994 if (ender > 0xff) {
10998 if (SIZE_ONLY /* like \08, \178 */
11001 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
11003 reg_warn_non_literal_string(
11005 form_short_octal_warning(p, numlen));
11008 if (PL_encoding && ender < 0x100)
11009 goto recode_encoding;
11012 if (! RExC_override_recoding) {
11013 SV* enc = PL_encoding;
11014 ender = reg_recode((const char)(U8)ender, &enc);
11015 if (!enc && SIZE_ONLY)
11016 ckWARNreg(p, "Invalid escape in the specified encoding");
11022 FAIL("Trailing \\");
11025 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
11026 /* Include any { following the alpha to emphasize
11027 * that it could be part of an escape at some point
11029 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
11030 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
11032 goto normal_default;
11033 } /* End of switch on '\' */
11035 default: /* A literal character */
11038 && RExC_flags & RXf_PMf_EXTENDED
11039 && ckWARN_d(WARN_DEPRECATED)
11040 && is_PATWS_non_low(p, UTF))
11042 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
11043 "Escape literal pattern white space under /x");
11047 if (UTF8_IS_START(*p) && UTF) {
11049 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
11050 &numlen, UTF8_ALLOW_DEFAULT);
11056 } /* End of switch on the literal */
11058 /* Here, have looked at the literal character and <ender>
11059 * contains its ordinal, <p> points to the character after it
11062 if ( RExC_flags & RXf_PMf_EXTENDED)
11063 p = regwhite( pRExC_state, p );
11065 /* If the next thing is a quantifier, it applies to this
11066 * character only, which means that this character has to be in
11067 * its own node and can't just be appended to the string in an
11068 * existing node, so if there are already other characters in
11069 * the node, close the node with just them, and set up to do
11070 * this character again next time through, when it will be the
11071 * only thing in its new node */
11072 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
11080 const STRLEN unilen = reguni(pRExC_state, ender, s);
11086 /* The loop increments <len> each time, as all but this
11087 * path (and one other) through it add a single byte to
11088 * the EXACTish node. But this one has changed len to
11089 * be the correct final value, so subtract one to
11090 * cancel out the increment that follows */
11094 REGC((char)ender, s++);
11099 /* See comments for join_exact() as to why we fold this
11100 * non-UTF at compile time */
11101 || (node_type == EXACTFU
11102 && ender == LATIN_SMALL_LETTER_SHARP_S)))
11104 *(s++) = (char) ender;
11105 maybe_exact &= ! IS_IN_SOME_FOLD_L1(ender);
11109 /* Prime the casefolded buffer. Locale rules, which apply
11110 * only to code points < 256, aren't known until execution,
11111 * so for them, just output the original character using
11112 * utf8. If we start to fold non-UTF patterns, be sure to
11113 * update join_exact() */
11114 if (LOC && ender < 256) {
11115 if (UNI_IS_INVARIANT(ender)) {
11119 *s = UTF8_TWO_BYTE_HI(ender);
11120 *(s + 1) = UTF8_TWO_BYTE_LO(ender);
11125 UV folded = _to_uni_fold_flags(
11130 | ((LOC) ? FOLD_FLAGS_LOCALE
11131 : (ASCII_FOLD_RESTRICTED)
11132 ? FOLD_FLAGS_NOMIX_ASCII
11136 /* If this node only contains non-folding code points
11137 * so far, see if this new one is also non-folding */
11139 if (folded != ender) {
11140 maybe_exact = FALSE;
11143 /* Here the fold is the original; we have
11144 * to check further to see if anything
11146 if (! PL_utf8_foldable) {
11147 SV* swash = swash_init("utf8",
11149 &PL_sv_undef, 1, 0);
11151 _get_swash_invlist(swash);
11152 SvREFCNT_dec_NN(swash);
11154 if (_invlist_contains_cp(PL_utf8_foldable,
11157 maybe_exact = FALSE;
11165 /* The loop increments <len> each time, as all but this
11166 * path (and one other) through it add a single byte to the
11167 * EXACTish node. But this one has changed len to be the
11168 * correct final value, so subtract one to cancel out the
11169 * increment that follows */
11170 len += foldlen - 1;
11173 if (next_is_quantifier) {
11175 /* Here, the next input is a quantifier, and to get here,
11176 * the current character is the only one in the node.
11177 * Also, here <len> doesn't include the final byte for this
11183 } /* End of loop through literal characters */
11185 /* Here we have either exhausted the input or ran out of room in
11186 * the node. (If we encountered a character that can't be in the
11187 * node, transfer is made directly to <loopdone>, and so we
11188 * wouldn't have fallen off the end of the loop.) In the latter
11189 * case, we artificially have to split the node into two, because
11190 * we just don't have enough space to hold everything. This
11191 * creates a problem if the final character participates in a
11192 * multi-character fold in the non-final position, as a match that
11193 * should have occurred won't, due to the way nodes are matched,
11194 * and our artificial boundary. So back off until we find a non-
11195 * problematic character -- one that isn't at the beginning or
11196 * middle of such a fold. (Either it doesn't participate in any
11197 * folds, or appears only in the final position of all the folds it
11198 * does participate in.) A better solution with far fewer false
11199 * positives, and that would fill the nodes more completely, would
11200 * be to actually have available all the multi-character folds to
11201 * test against, and to back-off only far enough to be sure that
11202 * this node isn't ending with a partial one. <upper_parse> is set
11203 * further below (if we need to reparse the node) to include just
11204 * up through that final non-problematic character that this code
11205 * identifies, so when it is set to less than the full node, we can
11206 * skip the rest of this */
11207 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
11209 const STRLEN full_len = len;
11211 assert(len >= MAX_NODE_STRING_SIZE);
11213 /* Here, <s> points to the final byte of the final character.
11214 * Look backwards through the string until find a non-
11215 * problematic character */
11219 /* These two have no multi-char folds to non-UTF characters
11221 if (ASCII_FOLD_RESTRICTED || LOC) {
11225 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
11229 if (! PL_NonL1NonFinalFold) {
11230 PL_NonL1NonFinalFold = _new_invlist_C_array(
11231 NonL1_Perl_Non_Final_Folds_invlist);
11234 /* Point to the first byte of the final character */
11235 s = (char *) utf8_hop((U8 *) s, -1);
11237 while (s >= s0) { /* Search backwards until find
11238 non-problematic char */
11239 if (UTF8_IS_INVARIANT(*s)) {
11241 /* There are no ascii characters that participate
11242 * in multi-char folds under /aa. In EBCDIC, the
11243 * non-ascii invariants are all control characters,
11244 * so don't ever participate in any folds. */
11245 if (ASCII_FOLD_RESTRICTED
11246 || ! IS_NON_FINAL_FOLD(*s))
11251 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
11253 /* No Latin1 characters participate in multi-char
11254 * folds under /l */
11256 || ! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_UNI(
11262 else if (! _invlist_contains_cp(
11263 PL_NonL1NonFinalFold,
11264 valid_utf8_to_uvchr((U8 *) s, NULL)))
11269 /* Here, the current character is problematic in that
11270 * it does occur in the non-final position of some
11271 * fold, so try the character before it, but have to
11272 * special case the very first byte in the string, so
11273 * we don't read outside the string */
11274 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
11275 } /* End of loop backwards through the string */
11277 /* If there were only problematic characters in the string,
11278 * <s> will point to before s0, in which case the length
11279 * should be 0, otherwise include the length of the
11280 * non-problematic character just found */
11281 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
11284 /* Here, have found the final character, if any, that is
11285 * non-problematic as far as ending the node without splitting
11286 * it across a potential multi-char fold. <len> contains the
11287 * number of bytes in the node up-to and including that
11288 * character, or is 0 if there is no such character, meaning
11289 * the whole node contains only problematic characters. In
11290 * this case, give up and just take the node as-is. We can't
11296 /* Here, the node does contain some characters that aren't
11297 * problematic. If one such is the final character in the
11298 * node, we are done */
11299 if (len == full_len) {
11302 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
11304 /* If the final character is problematic, but the
11305 * penultimate is not, back-off that last character to
11306 * later start a new node with it */
11311 /* Here, the final non-problematic character is earlier
11312 * in the input than the penultimate character. What we do
11313 * is reparse from the beginning, going up only as far as
11314 * this final ok one, thus guaranteeing that the node ends
11315 * in an acceptable character. The reason we reparse is
11316 * that we know how far in the character is, but we don't
11317 * know how to correlate its position with the input parse.
11318 * An alternate implementation would be to build that
11319 * correlation as we go along during the original parse,
11320 * but that would entail extra work for every node, whereas
11321 * this code gets executed only when the string is too
11322 * large for the node, and the final two characters are
11323 * problematic, an infrequent occurrence. Yet another
11324 * possible strategy would be to save the tail of the
11325 * string, and the next time regatom is called, initialize
11326 * with that. The problem with this is that unless you
11327 * back off one more character, you won't be guaranteed
11328 * regatom will get called again, unless regbranch,
11329 * regpiece ... are also changed. If you do back off that
11330 * extra character, so that there is input guaranteed to
11331 * force calling regatom, you can't handle the case where
11332 * just the first character in the node is acceptable. I
11333 * (khw) decided to try this method which doesn't have that
11334 * pitfall; if performance issues are found, we can do a
11335 * combination of the current approach plus that one */
11341 } /* End of verifying node ends with an appropriate char */
11343 loopdone: /* Jumped to when encounters something that shouldn't be in
11346 /* I (khw) don't know if you can get here with zero length, but the
11347 * old code handled this situation by creating a zero-length EXACT
11348 * node. Might as well be NOTHING instead */
11354 /* If 'maybe_exact' is still set here, means there are no
11355 * code points in the node that participate in folds */
11356 if (FOLD && maybe_exact) {
11359 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
11362 RExC_parse = p - 1;
11363 Set_Node_Cur_Length(ret, parse_start);
11364 nextchar(pRExC_state);
11366 /* len is STRLEN which is unsigned, need to copy to signed */
11369 vFAIL("Internal disaster");
11372 } /* End of label 'defchar:' */
11374 } /* End of giant switch on input character */
11380 S_regwhite( RExC_state_t *pRExC_state, char *p )
11382 const char *e = RExC_end;
11384 PERL_ARGS_ASSERT_REGWHITE;
11389 else if (*p == '#') {
11392 if (*p++ == '\n') {
11398 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11407 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
11409 /* Returns the next non-pattern-white space, non-comment character (the
11410 * latter only if 'recognize_comment is true) in the string p, which is
11411 * ended by RExC_end. If there is no line break ending a comment,
11412 * RExC_seen has added the REG_SEEN_RUN_ON_COMMENT flag; */
11413 const char *e = RExC_end;
11415 PERL_ARGS_ASSERT_REGPATWS;
11419 if ((len = is_PATWS_safe(p, e, UTF))) {
11422 else if (recognize_comment && *p == '#') {
11426 if (is_LNBREAK_safe(p, e, UTF)) {
11432 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11440 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
11441 Character classes ([:foo:]) can also be negated ([:^foo:]).
11442 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
11443 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
11444 but trigger failures because they are currently unimplemented. */
11446 #define POSIXCC_DONE(c) ((c) == ':')
11447 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
11448 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
11450 PERL_STATIC_INLINE I32
11451 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
11454 I32 namedclass = OOB_NAMEDCLASS;
11456 PERL_ARGS_ASSERT_REGPPOSIXCC;
11458 if (value == '[' && RExC_parse + 1 < RExC_end &&
11459 /* I smell either [: or [= or [. -- POSIX has been here, right? */
11460 POSIXCC(UCHARAT(RExC_parse)))
11462 const char c = UCHARAT(RExC_parse);
11463 char* const s = RExC_parse++;
11465 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
11467 if (RExC_parse == RExC_end) {
11470 /* Try to give a better location for the error (than the end of
11471 * the string) by looking for the matching ']' */
11473 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
11476 vFAIL2("Unmatched '%c' in POSIX class", c);
11478 /* Grandfather lone [:, [=, [. */
11482 const char* const t = RExC_parse++; /* skip over the c */
11485 if (UCHARAT(RExC_parse) == ']') {
11486 const char *posixcc = s + 1;
11487 RExC_parse++; /* skip over the ending ] */
11490 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
11491 const I32 skip = t - posixcc;
11493 /* Initially switch on the length of the name. */
11496 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
11497 this is the Perl \w
11499 namedclass = ANYOF_WORDCHAR;
11502 /* Names all of length 5. */
11503 /* alnum alpha ascii blank cntrl digit graph lower
11504 print punct space upper */
11505 /* Offset 4 gives the best switch position. */
11506 switch (posixcc[4]) {
11508 if (memEQ(posixcc, "alph", 4)) /* alpha */
11509 namedclass = ANYOF_ALPHA;
11512 if (memEQ(posixcc, "spac", 4)) /* space */
11513 namedclass = ANYOF_PSXSPC;
11516 if (memEQ(posixcc, "grap", 4)) /* graph */
11517 namedclass = ANYOF_GRAPH;
11520 if (memEQ(posixcc, "asci", 4)) /* ascii */
11521 namedclass = ANYOF_ASCII;
11524 if (memEQ(posixcc, "blan", 4)) /* blank */
11525 namedclass = ANYOF_BLANK;
11528 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
11529 namedclass = ANYOF_CNTRL;
11532 if (memEQ(posixcc, "alnu", 4)) /* alnum */
11533 namedclass = ANYOF_ALPHANUMERIC;
11536 if (memEQ(posixcc, "lowe", 4)) /* lower */
11537 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
11538 else if (memEQ(posixcc, "uppe", 4)) /* upper */
11539 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
11542 if (memEQ(posixcc, "digi", 4)) /* digit */
11543 namedclass = ANYOF_DIGIT;
11544 else if (memEQ(posixcc, "prin", 4)) /* print */
11545 namedclass = ANYOF_PRINT;
11546 else if (memEQ(posixcc, "punc", 4)) /* punct */
11547 namedclass = ANYOF_PUNCT;
11552 if (memEQ(posixcc, "xdigit", 6))
11553 namedclass = ANYOF_XDIGIT;
11557 if (namedclass == OOB_NAMEDCLASS)
11558 Simple_vFAIL3("POSIX class [:%.*s:] unknown",
11561 /* The #defines are structured so each complement is +1 to
11562 * the normal one */
11566 assert (posixcc[skip] == ':');
11567 assert (posixcc[skip+1] == ']');
11568 } else if (!SIZE_ONLY) {
11569 /* [[=foo=]] and [[.foo.]] are still future. */
11571 /* adjust RExC_parse so the warning shows after
11572 the class closes */
11573 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
11575 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11578 /* Maternal grandfather:
11579 * "[:" ending in ":" but not in ":]" */
11581 vFAIL("Unmatched '[' in POSIX class");
11584 /* Grandfather lone [:, [=, [. */
11594 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
11596 /* This applies some heuristics at the current parse position (which should
11597 * be at a '[') to see if what follows might be intended to be a [:posix:]
11598 * class. It returns true if it really is a posix class, of course, but it
11599 * also can return true if it thinks that what was intended was a posix
11600 * class that didn't quite make it.
11602 * It will return true for
11604 * [:alphanumerics] (as long as the ] isn't followed immediately by a
11605 * ')' indicating the end of the (?[
11606 * [:any garbage including %^&$ punctuation:]
11608 * This is designed to be called only from S_handle_regex_sets; it could be
11609 * easily adapted to be called from the spot at the beginning of regclass()
11610 * that checks to see in a normal bracketed class if the surrounding []
11611 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
11612 * change long-standing behavior, so I (khw) didn't do that */
11613 char* p = RExC_parse + 1;
11614 char first_char = *p;
11616 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
11618 assert(*(p - 1) == '[');
11620 if (! POSIXCC(first_char)) {
11625 while (p < RExC_end && isWORDCHAR(*p)) p++;
11627 if (p >= RExC_end) {
11631 if (p - RExC_parse > 2 /* Got at least 1 word character */
11632 && (*p == first_char
11633 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
11638 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
11641 && p - RExC_parse > 2 /* [:] evaluates to colon;
11642 [::] is a bad posix class. */
11643 && first_char == *(p - 1));
11647 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist, I32 *flagp, U32 depth,
11648 char * const oregcomp_parse)
11650 /* Handle the (?[...]) construct to do set operations */
11653 UV start, end; /* End points of code point ranges */
11655 char *save_end, *save_parse;
11660 const bool save_fold = FOLD;
11662 GET_RE_DEBUG_FLAGS_DECL;
11664 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
11667 vFAIL("(?[...]) not valid in locale");
11669 RExC_uni_semantics = 1;
11671 /* This will return only an ANYOF regnode, or (unlikely) something smaller
11672 * (such as EXACT). Thus we can skip most everything if just sizing. We
11673 * call regclass to handle '[]' so as to not have to reinvent its parsing
11674 * rules here (throwing away the size it computes each time). And, we exit
11675 * upon an unescaped ']' that isn't one ending a regclass. To do both
11676 * these things, we need to realize that something preceded by a backslash
11677 * is escaped, so we have to keep track of backslashes */
11679 UV depth = 0; /* how many nested (?[...]) constructs */
11681 Perl_ck_warner_d(aTHX_
11682 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
11683 "The regex_sets feature is experimental" REPORT_LOCATION,
11684 (int) (RExC_parse - RExC_precomp) , RExC_precomp, RExC_parse);
11686 while (RExC_parse < RExC_end) {
11687 SV* current = NULL;
11688 RExC_parse = regpatws(pRExC_state, RExC_parse,
11689 TRUE); /* means recognize comments */
11690 switch (*RExC_parse) {
11692 if (RExC_parse[1] == '[') depth++, RExC_parse++;
11697 /* Skip the next byte (which could cause us to end up in
11698 * the middle of a UTF-8 character, but since none of those
11699 * are confusable with anything we currently handle in this
11700 * switch (invariants all), it's safe. We'll just hit the
11701 * default: case next time and keep on incrementing until
11702 * we find one of the invariants we do handle. */
11707 /* If this looks like it is a [:posix:] class, leave the
11708 * parse pointer at the '[' to fool regclass() into
11709 * thinking it is part of a '[[:posix:]]'. That function
11710 * will use strict checking to force a syntax error if it
11711 * doesn't work out to a legitimate class */
11712 bool is_posix_class
11713 = could_it_be_a_POSIX_class(pRExC_state);
11714 if (! is_posix_class) {
11718 /* regclass() can only return RESTART_UTF8 if multi-char
11719 folds are allowed. */
11720 if (!regclass(pRExC_state, flagp,depth+1,
11721 is_posix_class, /* parse the whole char
11722 class only if not a
11724 FALSE, /* don't allow multi-char folds */
11725 TRUE, /* silence non-portable warnings. */
11727 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11730 /* function call leaves parse pointing to the ']', except
11731 * if we faked it */
11732 if (is_posix_class) {
11736 SvREFCNT_dec(current); /* In case it returned something */
11741 if (depth--) break;
11743 if (RExC_parse < RExC_end
11744 && *RExC_parse == ')')
11746 node = reganode(pRExC_state, ANYOF, 0);
11747 RExC_size += ANYOF_SKIP;
11748 nextchar(pRExC_state);
11749 Set_Node_Length(node,
11750 RExC_parse - oregcomp_parse + 1); /* MJD */
11759 FAIL("Syntax error in (?[...])");
11762 /* Pass 2 only after this. Everything in this construct is a
11763 * metacharacter. Operands begin with either a '\' (for an escape
11764 * sequence), or a '[' for a bracketed character class. Any other
11765 * character should be an operator, or parenthesis for grouping. Both
11766 * types of operands are handled by calling regclass() to parse them. It
11767 * is called with a parameter to indicate to return the computed inversion
11768 * list. The parsing here is implemented via a stack. Each entry on the
11769 * stack is a single character representing one of the operators, or the
11770 * '('; or else a pointer to an operand inversion list. */
11772 #define IS_OPERAND(a) (! SvIOK(a))
11774 /* The stack starts empty. It is a syntax error if the first thing parsed
11775 * is a binary operator; everything else is pushed on the stack. When an
11776 * operand is parsed, the top of the stack is examined. If it is a binary
11777 * operator, the item before it should be an operand, and both are replaced
11778 * by the result of doing that operation on the new operand and the one on
11779 * the stack. Thus a sequence of binary operands is reduced to a single
11780 * one before the next one is parsed.
11782 * A unary operator may immediately follow a binary in the input, for
11785 * When an operand is parsed and the top of the stack is a unary operator,
11786 * the operation is performed, and then the stack is rechecked to see if
11787 * this new operand is part of a binary operation; if so, it is handled as
11790 * A '(' is simply pushed on the stack; it is valid only if the stack is
11791 * empty, or the top element of the stack is an operator or another '('
11792 * (for which the parenthesized expression will become an operand). By the
11793 * time the corresponding ')' is parsed everything in between should have
11794 * been parsed and evaluated to a single operand (or else is a syntax
11795 * error), and is handled as a regular operand */
11797 sv_2mortal((SV *)(stack = newAV()));
11799 while (RExC_parse < RExC_end) {
11800 I32 top_index = av_tindex(stack);
11802 SV* current = NULL;
11804 /* Skip white space */
11805 RExC_parse = regpatws(pRExC_state, RExC_parse,
11806 TRUE); /* means recognize comments */
11807 if (RExC_parse >= RExC_end) {
11808 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
11810 if ((curchar = UCHARAT(RExC_parse)) == ']') {
11817 if (av_tindex(stack) >= 0 /* This makes sure that we can
11818 safely subtract 1 from
11819 RExC_parse in the next clause.
11820 If we have something on the
11821 stack, we have parsed something
11823 && UCHARAT(RExC_parse - 1) == '('
11824 && RExC_parse < RExC_end)
11826 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
11827 * This happens when we have some thing like
11829 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
11831 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
11833 * Here we would be handling the interpolated
11834 * '$thai_or_lao'. We handle this by a recursive call to
11835 * ourselves which returns the inversion list the
11836 * interpolated expression evaluates to. We use the flags
11837 * from the interpolated pattern. */
11838 U32 save_flags = RExC_flags;
11839 const char * const save_parse = ++RExC_parse;
11841 parse_lparen_question_flags(pRExC_state);
11843 if (RExC_parse == save_parse /* Makes sure there was at
11844 least one flag (or this
11845 embedding wasn't compiled)
11847 || RExC_parse >= RExC_end - 4
11848 || UCHARAT(RExC_parse) != ':'
11849 || UCHARAT(++RExC_parse) != '('
11850 || UCHARAT(++RExC_parse) != '?'
11851 || UCHARAT(++RExC_parse) != '[')
11854 /* In combination with the above, this moves the
11855 * pointer to the point just after the first erroneous
11856 * character (or if there are no flags, to where they
11857 * should have been) */
11858 if (RExC_parse >= RExC_end - 4) {
11859 RExC_parse = RExC_end;
11861 else if (RExC_parse != save_parse) {
11862 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11864 vFAIL("Expecting '(?flags:(?[...'");
11867 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
11868 depth+1, oregcomp_parse);
11870 /* Here, 'current' contains the embedded expression's
11871 * inversion list, and RExC_parse points to the trailing
11872 * ']'; the next character should be the ')' which will be
11873 * paired with the '(' that has been put on the stack, so
11874 * the whole embedded expression reduces to '(operand)' */
11877 RExC_flags = save_flags;
11878 goto handle_operand;
11883 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11884 vFAIL("Unexpected character");
11887 /* regclass() can only return RESTART_UTF8 if multi-char
11888 folds are allowed. */
11889 if (!regclass(pRExC_state, flagp,depth+1,
11890 TRUE, /* means parse just the next thing */
11891 FALSE, /* don't allow multi-char folds */
11892 FALSE, /* don't silence non-portable warnings. */
11894 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11896 /* regclass() will return with parsing just the \ sequence,
11897 * leaving the parse pointer at the next thing to parse */
11899 goto handle_operand;
11901 case '[': /* Is a bracketed character class */
11903 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
11905 if (! is_posix_class) {
11909 /* regclass() can only return RESTART_UTF8 if multi-char
11910 folds are allowed. */
11911 if(!regclass(pRExC_state, flagp,depth+1,
11912 is_posix_class, /* parse the whole char class
11913 only if not a posix class */
11914 FALSE, /* don't allow multi-char folds */
11915 FALSE, /* don't silence non-portable warnings. */
11917 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11919 /* function call leaves parse pointing to the ']', except if we
11921 if (is_posix_class) {
11925 goto handle_operand;
11934 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
11935 || ! IS_OPERAND(*top_ptr))
11938 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
11940 av_push(stack, newSVuv(curchar));
11944 av_push(stack, newSVuv(curchar));
11948 if (top_index >= 0) {
11949 top_ptr = av_fetch(stack, top_index, FALSE);
11951 if (IS_OPERAND(*top_ptr)) {
11953 vFAIL("Unexpected '(' with no preceding operator");
11956 av_push(stack, newSVuv(curchar));
11963 || ! (current = av_pop(stack))
11964 || ! IS_OPERAND(current)
11965 || ! (lparen = av_pop(stack))
11966 || IS_OPERAND(lparen)
11967 || SvUV(lparen) != '(')
11969 SvREFCNT_dec(current);
11971 vFAIL("Unexpected ')'");
11974 SvREFCNT_dec_NN(lparen);
11981 /* Here, we have an operand to process, in 'current' */
11983 if (top_index < 0) { /* Just push if stack is empty */
11984 av_push(stack, current);
11987 SV* top = av_pop(stack);
11989 char current_operator;
11991 if (IS_OPERAND(top)) {
11992 SvREFCNT_dec_NN(top);
11993 SvREFCNT_dec_NN(current);
11994 vFAIL("Operand with no preceding operator");
11996 current_operator = (char) SvUV(top);
11997 switch (current_operator) {
11998 case '(': /* Push the '(' back on followed by the new
12000 av_push(stack, top);
12001 av_push(stack, current);
12002 SvREFCNT_inc(top); /* Counters the '_dec' done
12003 just after the 'break', so
12004 it doesn't get wrongly freed
12009 _invlist_invert(current);
12011 /* Unlike binary operators, the top of the stack,
12012 * now that this unary one has been popped off, may
12013 * legally be an operator, and we now have operand
12016 SvREFCNT_dec_NN(top);
12017 goto handle_operand;
12020 prev = av_pop(stack);
12021 _invlist_intersection(prev,
12024 av_push(stack, current);
12029 prev = av_pop(stack);
12030 _invlist_union(prev, current, ¤t);
12031 av_push(stack, current);
12035 prev = av_pop(stack);;
12036 _invlist_subtract(prev, current, ¤t);
12037 av_push(stack, current);
12040 case '^': /* The union minus the intersection */
12046 prev = av_pop(stack);
12047 _invlist_union(prev, current, &u);
12048 _invlist_intersection(prev, current, &i);
12049 /* _invlist_subtract will overwrite current
12050 without freeing what it already contains */
12052 _invlist_subtract(u, i, ¤t);
12053 av_push(stack, current);
12054 SvREFCNT_dec_NN(i);
12055 SvREFCNT_dec_NN(u);
12056 SvREFCNT_dec_NN(element);
12061 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
12063 SvREFCNT_dec_NN(top);
12064 SvREFCNT_dec(prev);
12068 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12071 if (av_tindex(stack) < 0 /* Was empty */
12072 || ((final = av_pop(stack)) == NULL)
12073 || ! IS_OPERAND(final)
12074 || av_tindex(stack) >= 0) /* More left on stack */
12076 vFAIL("Incomplete expression within '(?[ ])'");
12079 /* Here, 'final' is the resultant inversion list from evaluating the
12080 * expression. Return it if so requested */
12081 if (return_invlist) {
12082 *return_invlist = final;
12086 /* Otherwise generate a resultant node, based on 'final'. regclass() is
12087 * expecting a string of ranges and individual code points */
12088 invlist_iterinit(final);
12089 result_string = newSVpvs("");
12090 while (invlist_iternext(final, &start, &end)) {
12091 if (start == end) {
12092 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
12095 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
12100 save_parse = RExC_parse;
12101 RExC_parse = SvPV(result_string, len);
12102 save_end = RExC_end;
12103 RExC_end = RExC_parse + len;
12105 /* We turn off folding around the call, as the class we have constructed
12106 * already has all folding taken into consideration, and we don't want
12107 * regclass() to add to that */
12108 RExC_flags &= ~RXf_PMf_FOLD;
12109 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
12111 node = regclass(pRExC_state, flagp,depth+1,
12112 FALSE, /* means parse the whole char class */
12113 FALSE, /* don't allow multi-char folds */
12114 TRUE, /* silence non-portable warnings. The above may very
12115 well have generated non-portable code points, but
12116 they're valid on this machine */
12119 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
12122 RExC_flags |= RXf_PMf_FOLD;
12124 RExC_parse = save_parse + 1;
12125 RExC_end = save_end;
12126 SvREFCNT_dec_NN(final);
12127 SvREFCNT_dec_NN(result_string);
12129 nextchar(pRExC_state);
12130 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
12135 /* The names of properties whose definitions are not known at compile time are
12136 * stored in this SV, after a constant heading. So if the length has been
12137 * changed since initialization, then there is a run-time definition. */
12138 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION (SvCUR(listsv) != initial_listsv_len)
12141 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
12142 const bool stop_at_1, /* Just parse the next thing, don't
12143 look for a full character class */
12144 bool allow_multi_folds,
12145 const bool silence_non_portable, /* Don't output warnings
12148 SV** ret_invlist) /* Return an inversion list, not a node */
12150 /* parse a bracketed class specification. Most of these will produce an
12151 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
12152 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
12153 * under /i with multi-character folds: it will be rewritten following the
12154 * paradigm of this example, where the <multi-fold>s are characters which
12155 * fold to multiple character sequences:
12156 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
12157 * gets effectively rewritten as:
12158 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
12159 * reg() gets called (recursively) on the rewritten version, and this
12160 * function will return what it constructs. (Actually the <multi-fold>s
12161 * aren't physically removed from the [abcdefghi], it's just that they are
12162 * ignored in the recursion by means of a flag:
12163 * <RExC_in_multi_char_class>.)
12165 * ANYOF nodes contain a bit map for the first 256 characters, with the
12166 * corresponding bit set if that character is in the list. For characters
12167 * above 255, a range list or swash is used. There are extra bits for \w,
12168 * etc. in locale ANYOFs, as what these match is not determinable at
12171 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
12172 * to be restarted. This can only happen if ret_invlist is non-NULL.
12176 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
12178 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
12181 IV namedclass = OOB_NAMEDCLASS;
12182 char *rangebegin = NULL;
12183 bool need_class = 0;
12185 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
12186 than just initialized. */
12187 SV* properties = NULL; /* Code points that match \p{} \P{} */
12188 SV* posixes = NULL; /* Code points that match classes like, [:word:],
12189 extended beyond the Latin1 range */
12190 UV element_count = 0; /* Number of distinct elements in the class.
12191 Optimizations may be possible if this is tiny */
12192 AV * multi_char_matches = NULL; /* Code points that fold to more than one
12193 character; used under /i */
12195 char * stop_ptr = RExC_end; /* where to stop parsing */
12196 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
12198 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
12200 /* Unicode properties are stored in a swash; this holds the current one
12201 * being parsed. If this swash is the only above-latin1 component of the
12202 * character class, an optimization is to pass it directly on to the
12203 * execution engine. Otherwise, it is set to NULL to indicate that there
12204 * are other things in the class that have to be dealt with at execution
12206 SV* swash = NULL; /* Code points that match \p{} \P{} */
12208 /* Set if a component of this character class is user-defined; just passed
12209 * on to the engine */
12210 bool has_user_defined_property = FALSE;
12212 /* inversion list of code points this node matches only when the target
12213 * string is in UTF-8. (Because is under /d) */
12214 SV* depends_list = NULL;
12216 /* inversion list of code points this node matches. For much of the
12217 * function, it includes only those that match regardless of the utf8ness
12218 * of the target string */
12219 SV* cp_list = NULL;
12222 /* In a range, counts how many 0-2 of the ends of it came from literals,
12223 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
12224 UV literal_endpoint = 0;
12226 bool invert = FALSE; /* Is this class to be complemented */
12228 /* Is there any thing like \W or [:^digit:] that matches above the legal
12229 * Unicode range? */
12230 bool runtime_posix_matches_above_Unicode = FALSE;
12232 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
12233 case we need to change the emitted regop to an EXACT. */
12234 const char * orig_parse = RExC_parse;
12235 const I32 orig_size = RExC_size;
12236 GET_RE_DEBUG_FLAGS_DECL;
12238 PERL_ARGS_ASSERT_REGCLASS;
12240 PERL_UNUSED_ARG(depth);
12243 DEBUG_PARSE("clas");
12245 /* Assume we are going to generate an ANYOF node. */
12246 ret = reganode(pRExC_state, ANYOF, 0);
12249 RExC_size += ANYOF_SKIP;
12250 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
12253 ANYOF_FLAGS(ret) = 0;
12255 RExC_emit += ANYOF_SKIP;
12257 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
12259 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
12260 initial_listsv_len = SvCUR(listsv);
12261 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
12265 RExC_parse = regpatws(pRExC_state, RExC_parse,
12266 FALSE /* means don't recognize comments */);
12269 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
12272 allow_multi_folds = FALSE;
12275 RExC_parse = regpatws(pRExC_state, RExC_parse,
12276 FALSE /* means don't recognize comments */);
12280 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
12281 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
12282 const char *s = RExC_parse;
12283 const char c = *s++;
12285 while (isWORDCHAR(*s))
12287 if (*s && c == *s && s[1] == ']') {
12288 SAVEFREESV(RExC_rx_sv);
12290 "POSIX syntax [%c %c] belongs inside character classes",
12292 (void)ReREFCNT_inc(RExC_rx_sv);
12296 /* If the caller wants us to just parse a single element, accomplish this
12297 * by faking the loop ending condition */
12298 if (stop_at_1 && RExC_end > RExC_parse) {
12299 stop_ptr = RExC_parse + 1;
12302 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
12303 if (UCHARAT(RExC_parse) == ']')
12304 goto charclassloop;
12308 if (RExC_parse >= stop_ptr) {
12313 RExC_parse = regpatws(pRExC_state, RExC_parse,
12314 FALSE /* means don't recognize comments */);
12317 if (UCHARAT(RExC_parse) == ']') {
12323 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
12324 save_value = value;
12325 save_prevvalue = prevvalue;
12328 rangebegin = RExC_parse;
12332 value = utf8n_to_uvchr((U8*)RExC_parse,
12333 RExC_end - RExC_parse,
12334 &numlen, UTF8_ALLOW_DEFAULT);
12335 RExC_parse += numlen;
12338 value = UCHARAT(RExC_parse++);
12341 && RExC_parse < RExC_end
12342 && POSIXCC(UCHARAT(RExC_parse)))
12344 namedclass = regpposixcc(pRExC_state, value, strict);
12346 else if (value == '\\') {
12348 value = utf8n_to_uvchr((U8*)RExC_parse,
12349 RExC_end - RExC_parse,
12350 &numlen, UTF8_ALLOW_DEFAULT);
12351 RExC_parse += numlen;
12354 value = UCHARAT(RExC_parse++);
12356 /* Some compilers cannot handle switching on 64-bit integer
12357 * values, therefore value cannot be an UV. Yes, this will
12358 * be a problem later if we want switch on Unicode.
12359 * A similar issue a little bit later when switching on
12360 * namedclass. --jhi */
12362 /* If the \ is escaping white space when white space is being
12363 * skipped, it means that that white space is wanted literally, and
12364 * is already in 'value'. Otherwise, need to translate the escape
12365 * into what it signifies. */
12366 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
12368 case 'w': namedclass = ANYOF_WORDCHAR; break;
12369 case 'W': namedclass = ANYOF_NWORDCHAR; break;
12370 case 's': namedclass = ANYOF_SPACE; break;
12371 case 'S': namedclass = ANYOF_NSPACE; break;
12372 case 'd': namedclass = ANYOF_DIGIT; break;
12373 case 'D': namedclass = ANYOF_NDIGIT; break;
12374 case 'v': namedclass = ANYOF_VERTWS; break;
12375 case 'V': namedclass = ANYOF_NVERTWS; break;
12376 case 'h': namedclass = ANYOF_HORIZWS; break;
12377 case 'H': namedclass = ANYOF_NHORIZWS; break;
12378 case 'N': /* Handle \N{NAME} in class */
12380 /* We only pay attention to the first char of
12381 multichar strings being returned. I kinda wonder
12382 if this makes sense as it does change the behaviour
12383 from earlier versions, OTOH that behaviour was broken
12385 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
12386 TRUE, /* => charclass */
12389 if (*flagp & RESTART_UTF8)
12390 FAIL("panic: grok_bslash_N set RESTART_UTF8");
12400 /* We will handle any undefined properties ourselves */
12401 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF;
12403 if (RExC_parse >= RExC_end)
12404 vFAIL2("Empty \\%c{}", (U8)value);
12405 if (*RExC_parse == '{') {
12406 const U8 c = (U8)value;
12407 e = strchr(RExC_parse++, '}');
12409 vFAIL2("Missing right brace on \\%c{}", c);
12410 while (isSPACE(UCHARAT(RExC_parse)))
12412 if (e == RExC_parse)
12413 vFAIL2("Empty \\%c{}", c);
12414 n = e - RExC_parse;
12415 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
12426 if (UCHARAT(RExC_parse) == '^') {
12429 /* toggle. (The rhs xor gets the single bit that
12430 * differs between P and p; the other xor inverts just
12432 value ^= 'P' ^ 'p';
12434 while (isSPACE(UCHARAT(RExC_parse))) {
12439 /* Try to get the definition of the property into
12440 * <invlist>. If /i is in effect, the effective property
12441 * will have its name be <__NAME_i>. The design is
12442 * discussed in commit
12443 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
12444 Newx(name, n + sizeof("_i__\n"), char);
12446 sprintf(name, "%s%.*s%s\n",
12447 (FOLD) ? "__" : "",
12453 /* Look up the property name, and get its swash and
12454 * inversion list, if the property is found */
12456 SvREFCNT_dec_NN(swash);
12458 swash = _core_swash_init("utf8", name, &PL_sv_undef,
12461 NULL, /* No inversion list */
12464 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
12466 SvREFCNT_dec_NN(swash);
12470 /* Here didn't find it. It could be a user-defined
12471 * property that will be available at run-time. If we
12472 * accept only compile-time properties, is an error;
12473 * otherwise add it to the list for run-time look up */
12475 RExC_parse = e + 1;
12476 vFAIL3("Property '%.*s' is unknown", (int) n, name);
12478 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s\n",
12479 (value == 'p' ? '+' : '!'),
12481 has_user_defined_property = TRUE;
12483 /* We don't know yet, so have to assume that the
12484 * property could match something in the Latin1 range,
12485 * hence something that isn't utf8. Note that this
12486 * would cause things in <depends_list> to match
12487 * inappropriately, except that any \p{}, including
12488 * this one forces Unicode semantics, which means there
12489 * is <no depends_list> */
12490 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
12494 /* Here, did get the swash and its inversion list. If
12495 * the swash is from a user-defined property, then this
12496 * whole character class should be regarded as such */
12497 has_user_defined_property =
12499 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY);
12501 /* Invert if asking for the complement */
12502 if (value == 'P') {
12503 _invlist_union_complement_2nd(properties,
12507 /* The swash can't be used as-is, because we've
12508 * inverted things; delay removing it to here after
12509 * have copied its invlist above */
12510 SvREFCNT_dec_NN(swash);
12514 _invlist_union(properties, invlist, &properties);
12519 RExC_parse = e + 1;
12520 namedclass = ANYOF_UNIPROP; /* no official name, but it's
12523 /* \p means they want Unicode semantics */
12524 RExC_uni_semantics = 1;
12527 case 'n': value = '\n'; break;
12528 case 'r': value = '\r'; break;
12529 case 't': value = '\t'; break;
12530 case 'f': value = '\f'; break;
12531 case 'b': value = '\b'; break;
12532 case 'e': value = ASCII_TO_NATIVE('\033');break;
12533 case 'a': value = ASCII_TO_NATIVE('\007');break;
12535 RExC_parse--; /* function expects to be pointed at the 'o' */
12537 const char* error_msg;
12538 bool valid = grok_bslash_o(&RExC_parse,
12541 SIZE_ONLY, /* warnings in pass
12544 silence_non_portable,
12550 if (PL_encoding && value < 0x100) {
12551 goto recode_encoding;
12555 RExC_parse--; /* function expects to be pointed at the 'x' */
12557 const char* error_msg;
12558 bool valid = grok_bslash_x(&RExC_parse,
12561 TRUE, /* Output warnings */
12563 silence_non_portable,
12569 if (PL_encoding && value < 0x100)
12570 goto recode_encoding;
12573 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
12575 case '0': case '1': case '2': case '3': case '4':
12576 case '5': case '6': case '7':
12578 /* Take 1-3 octal digits */
12579 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12580 numlen = (strict) ? 4 : 3;
12581 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
12582 RExC_parse += numlen;
12585 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12586 vFAIL("Need exactly 3 octal digits");
12588 else if (! SIZE_ONLY /* like \08, \178 */
12590 && RExC_parse < RExC_end
12591 && isDIGIT(*RExC_parse)
12592 && ckWARN(WARN_REGEXP))
12594 SAVEFREESV(RExC_rx_sv);
12595 reg_warn_non_literal_string(
12597 form_short_octal_warning(RExC_parse, numlen));
12598 (void)ReREFCNT_inc(RExC_rx_sv);
12601 if (PL_encoding && value < 0x100)
12602 goto recode_encoding;
12606 if (! RExC_override_recoding) {
12607 SV* enc = PL_encoding;
12608 value = reg_recode((const char)(U8)value, &enc);
12611 vFAIL("Invalid escape in the specified encoding");
12613 else if (SIZE_ONLY) {
12614 ckWARNreg(RExC_parse,
12615 "Invalid escape in the specified encoding");
12621 /* Allow \_ to not give an error */
12622 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
12624 vFAIL2("Unrecognized escape \\%c in character class",
12628 SAVEFREESV(RExC_rx_sv);
12629 ckWARN2reg(RExC_parse,
12630 "Unrecognized escape \\%c in character class passed through",
12632 (void)ReREFCNT_inc(RExC_rx_sv);
12636 } /* End of switch on char following backslash */
12637 } /* end of handling backslash escape sequences */
12640 literal_endpoint++;
12643 /* Here, we have the current token in 'value' */
12645 /* What matches in a locale is not known until runtime. This includes
12646 * what the Posix classes (like \w, [:space:]) match. Room must be
12647 * reserved (one time per class) to store such classes, either if Perl
12648 * is compiled so that locale nodes always should have this space, or
12649 * if there is such class info to be stored. The space will contain a
12650 * bit for each named class that is to be matched against. This isn't
12651 * needed for \p{} and pseudo-classes, as they are not affected by
12652 * locale, and hence are dealt with separately */
12655 && (ANYOF_LOCALE == ANYOF_CLASS
12656 || (namedclass > OOB_NAMEDCLASS && namedclass < ANYOF_MAX)))
12660 RExC_size += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12663 RExC_emit += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12664 ANYOF_CLASS_ZERO(ret);
12666 ANYOF_FLAGS(ret) |= ANYOF_CLASS;
12669 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
12671 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
12672 * literal, as is the character that began the false range, i.e.
12673 * the 'a' in the examples */
12676 const int w = (RExC_parse >= rangebegin)
12677 ? RExC_parse - rangebegin
12680 vFAIL4("False [] range \"%*.*s\"", w, w, rangebegin);
12683 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
12684 ckWARN4reg(RExC_parse,
12685 "False [] range \"%*.*s\"",
12687 (void)ReREFCNT_inc(RExC_rx_sv);
12688 cp_list = add_cp_to_invlist(cp_list, '-');
12689 cp_list = add_cp_to_invlist(cp_list, prevvalue);
12693 range = 0; /* this was not a true range */
12694 element_count += 2; /* So counts for three values */
12698 U8 classnum = namedclass_to_classnum(namedclass);
12699 if (namedclass >= ANYOF_MAX) { /* If a special class */
12700 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
12702 /* Here, should be \h, \H, \v, or \V. Neither /d nor
12703 * /l make a difference in what these match. There
12704 * would be problems if these characters had folds
12705 * other than themselves, as cp_list is subject to
12707 if (classnum != _CC_VERTSPACE) {
12708 assert( namedclass == ANYOF_HORIZWS
12709 || namedclass == ANYOF_NHORIZWS);
12711 /* It turns out that \h is just a synonym for
12713 classnum = _CC_BLANK;
12716 _invlist_union_maybe_complement_2nd(
12718 PL_XPosix_ptrs[classnum],
12719 cBOOL(namedclass % 2), /* Complement if odd
12720 (NHORIZWS, NVERTWS)
12725 else if (classnum == _CC_ASCII) {
12728 ANYOF_CLASS_SET(ret, namedclass);
12731 #endif /* Not isascii(); just use the hard-coded definition for it */
12732 _invlist_union_maybe_complement_2nd(
12735 cBOOL(namedclass % 2), /* Complement if odd
12739 else { /* Garden variety class */
12741 /* The ascii range inversion list */
12742 SV* ascii_source = PL_Posix_ptrs[classnum];
12744 /* The full Latin1 range inversion list */
12745 SV* l1_source = PL_L1Posix_ptrs[classnum];
12747 /* This code is structured into two major clauses. The
12748 * first is for classes whose complete definitions may not
12749 * already be known. It not, the Latin1 definition
12750 * (guaranteed to already known) is used plus code is
12751 * generated to load the rest at run-time (only if needed).
12752 * If the complete definition is known, it drops down to
12753 * the second clause, where the complete definition is
12756 if (classnum < _FIRST_NON_SWASH_CC) {
12758 /* Here, the class has a swash, which may or not
12759 * already be loaded */
12761 /* The name of the property to use to match the full
12762 * eXtended Unicode range swash for this character
12764 const char *Xname = swash_property_names[classnum];
12766 /* If returning the inversion list, we can't defer
12767 * getting this until runtime */
12768 if (ret_invlist && ! PL_utf8_swash_ptrs[classnum]) {
12769 PL_utf8_swash_ptrs[classnum] =
12770 _core_swash_init("utf8", Xname, &PL_sv_undef,
12773 NULL, /* No inversion list */
12774 NULL /* No flags */
12776 assert(PL_utf8_swash_ptrs[classnum]);
12778 if ( ! PL_utf8_swash_ptrs[classnum]) {
12779 if (namedclass % 2 == 0) { /* A non-complemented
12781 /* If not /a matching, there are code points we
12782 * don't know at compile time. Arrange for the
12783 * unknown matches to be loaded at run-time, if
12785 if (! AT_LEAST_ASCII_RESTRICTED) {
12786 Perl_sv_catpvf(aTHX_ listsv, "+utf8::%s\n",
12789 if (LOC) { /* Under locale, set run-time
12791 ANYOF_CLASS_SET(ret, namedclass);
12794 /* Add the current class's code points to
12795 * the running total */
12796 _invlist_union(posixes,
12797 (AT_LEAST_ASCII_RESTRICTED)
12803 else { /* A complemented class */
12804 if (AT_LEAST_ASCII_RESTRICTED) {
12805 /* Under /a should match everything above
12806 * ASCII, plus the complement of the set's
12808 _invlist_union_complement_2nd(posixes,
12813 /* Arrange for the unknown matches to be
12814 * loaded at run-time, if needed */
12815 Perl_sv_catpvf(aTHX_ listsv, "!utf8::%s\n",
12817 runtime_posix_matches_above_Unicode = TRUE;
12819 ANYOF_CLASS_SET(ret, namedclass);
12823 /* We want to match everything in
12824 * Latin1, except those things that
12825 * l1_source matches */
12826 SV* scratch_list = NULL;
12827 _invlist_subtract(PL_Latin1, l1_source,
12830 /* Add the list from this class to the
12833 posixes = scratch_list;
12836 _invlist_union(posixes,
12839 SvREFCNT_dec_NN(scratch_list);
12841 if (DEPENDS_SEMANTICS) {
12843 |= ANYOF_NON_UTF8_LATIN1_ALL;
12848 goto namedclass_done;
12851 /* Here, there is a swash loaded for the class. If no
12852 * inversion list for it yet, get it */
12853 if (! PL_XPosix_ptrs[classnum]) {
12854 PL_XPosix_ptrs[classnum]
12855 = _swash_to_invlist(PL_utf8_swash_ptrs[classnum]);
12859 /* Here there is an inversion list already loaded for the
12862 if (namedclass % 2 == 0) { /* A non-complemented class,
12863 like ANYOF_PUNCT */
12865 /* For non-locale, just add it to any existing list
12867 _invlist_union(posixes,
12868 (AT_LEAST_ASCII_RESTRICTED)
12870 : PL_XPosix_ptrs[classnum],
12873 else { /* Locale */
12874 SV* scratch_list = NULL;
12876 /* For above Latin1 code points, we use the full
12878 _invlist_intersection(PL_AboveLatin1,
12879 PL_XPosix_ptrs[classnum],
12881 /* And set the output to it, adding instead if
12882 * there already is an output. Checking if
12883 * 'posixes' is NULL first saves an extra clone.
12884 * Its reference count will be decremented at the
12885 * next union, etc, or if this is the only
12886 * instance, at the end of the routine */
12888 posixes = scratch_list;
12891 _invlist_union(posixes, scratch_list, &posixes);
12892 SvREFCNT_dec_NN(scratch_list);
12895 #ifndef HAS_ISBLANK
12896 if (namedclass != ANYOF_BLANK) {
12898 /* Set this class in the node for runtime
12900 ANYOF_CLASS_SET(ret, namedclass);
12901 #ifndef HAS_ISBLANK
12904 /* No isblank(), use the hard-coded ASCII-range
12905 * blanks, adding them to the running total. */
12907 _invlist_union(posixes, ascii_source, &posixes);
12912 else { /* A complemented class, like ANYOF_NPUNCT */
12914 _invlist_union_complement_2nd(
12916 (AT_LEAST_ASCII_RESTRICTED)
12918 : PL_XPosix_ptrs[classnum],
12920 /* Under /d, everything in the upper half of the
12921 * Latin1 range matches this complement */
12922 if (DEPENDS_SEMANTICS) {
12923 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL;
12926 else { /* Locale */
12927 SV* scratch_list = NULL;
12928 _invlist_subtract(PL_AboveLatin1,
12929 PL_XPosix_ptrs[classnum],
12932 posixes = scratch_list;
12935 _invlist_union(posixes, scratch_list, &posixes);
12936 SvREFCNT_dec_NN(scratch_list);
12938 #ifndef HAS_ISBLANK
12939 if (namedclass != ANYOF_NBLANK) {
12941 ANYOF_CLASS_SET(ret, namedclass);
12942 #ifndef HAS_ISBLANK
12945 /* Get the list of all code points in Latin1
12946 * that are not ASCII blanks, and add them to
12947 * the running total */
12948 _invlist_subtract(PL_Latin1, ascii_source,
12950 _invlist_union(posixes, scratch_list, &posixes);
12951 SvREFCNT_dec_NN(scratch_list);
12958 continue; /* Go get next character */
12960 } /* end of namedclass \blah */
12962 /* Here, we have a single value. If 'range' is set, it is the ending
12963 * of a range--check its validity. Later, we will handle each
12964 * individual code point in the range. If 'range' isn't set, this
12965 * could be the beginning of a range, so check for that by looking
12966 * ahead to see if the next real character to be processed is the range
12967 * indicator--the minus sign */
12970 RExC_parse = regpatws(pRExC_state, RExC_parse,
12971 FALSE /* means don't recognize comments */);
12975 if (prevvalue > value) /* b-a */ {
12976 const int w = RExC_parse - rangebegin;
12977 Simple_vFAIL4("Invalid [] range \"%*.*s\"", w, w, rangebegin);
12978 range = 0; /* not a valid range */
12982 prevvalue = value; /* save the beginning of the potential range */
12983 if (! stop_at_1 /* Can't be a range if parsing just one thing */
12984 && *RExC_parse == '-')
12986 char* next_char_ptr = RExC_parse + 1;
12987 if (skip_white) { /* Get the next real char after the '-' */
12988 next_char_ptr = regpatws(pRExC_state,
12990 FALSE); /* means don't recognize
12994 /* If the '-' is at the end of the class (just before the ']',
12995 * it is a literal minus; otherwise it is a range */
12996 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
12997 RExC_parse = next_char_ptr;
12999 /* a bad range like \w-, [:word:]- ? */
13000 if (namedclass > OOB_NAMEDCLASS) {
13001 if (strict || ckWARN(WARN_REGEXP)) {
13003 RExC_parse >= rangebegin ?
13004 RExC_parse - rangebegin : 0;
13006 vFAIL4("False [] range \"%*.*s\"",
13011 "False [] range \"%*.*s\"",
13016 cp_list = add_cp_to_invlist(cp_list, '-');
13020 range = 1; /* yeah, it's a range! */
13021 continue; /* but do it the next time */
13026 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
13029 /* non-Latin1 code point implies unicode semantics. Must be set in
13030 * pass1 so is there for the whole of pass 2 */
13032 RExC_uni_semantics = 1;
13035 /* Ready to process either the single value, or the completed range.
13036 * For single-valued non-inverted ranges, we consider the possibility
13037 * of multi-char folds. (We made a conscious decision to not do this
13038 * for the other cases because it can often lead to non-intuitive
13039 * results. For example, you have the peculiar case that:
13040 * "s s" =~ /^[^\xDF]+$/i => Y
13041 * "ss" =~ /^[^\xDF]+$/i => N
13043 * See [perl #89750] */
13044 if (FOLD && allow_multi_folds && value == prevvalue) {
13045 if (value == LATIN_SMALL_LETTER_SHARP_S
13046 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
13049 /* Here <value> is indeed a multi-char fold. Get what it is */
13051 U8 foldbuf[UTF8_MAXBYTES_CASE];
13054 UV folded = _to_uni_fold_flags(
13059 | ((LOC) ? FOLD_FLAGS_LOCALE
13060 : (ASCII_FOLD_RESTRICTED)
13061 ? FOLD_FLAGS_NOMIX_ASCII
13065 /* Here, <folded> should be the first character of the
13066 * multi-char fold of <value>, with <foldbuf> containing the
13067 * whole thing. But, if this fold is not allowed (because of
13068 * the flags), <fold> will be the same as <value>, and should
13069 * be processed like any other character, so skip the special
13071 if (folded != value) {
13073 /* Skip if we are recursed, currently parsing the class
13074 * again. Otherwise add this character to the list of
13075 * multi-char folds. */
13076 if (! RExC_in_multi_char_class) {
13077 AV** this_array_ptr;
13079 STRLEN cp_count = utf8_length(foldbuf,
13080 foldbuf + foldlen);
13081 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
13083 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
13086 if (! multi_char_matches) {
13087 multi_char_matches = newAV();
13090 /* <multi_char_matches> is actually an array of arrays.
13091 * There will be one or two top-level elements: [2],
13092 * and/or [3]. The [2] element is an array, each
13093 * element thereof is a character which folds to TWO
13094 * characters; [3] is for folds to THREE characters.
13095 * (Unicode guarantees a maximum of 3 characters in any
13096 * fold.) When we rewrite the character class below,
13097 * we will do so such that the longest folds are
13098 * written first, so that it prefers the longest
13099 * matching strings first. This is done even if it
13100 * turns out that any quantifier is non-greedy, out of
13101 * programmer laziness. Tom Christiansen has agreed
13102 * that this is ok. This makes the test for the
13103 * ligature 'ffi' come before the test for 'ff' */
13104 if (av_exists(multi_char_matches, cp_count)) {
13105 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13107 this_array = *this_array_ptr;
13110 this_array = newAV();
13111 av_store(multi_char_matches, cp_count,
13114 av_push(this_array, multi_fold);
13117 /* This element should not be processed further in this
13120 value = save_value;
13121 prevvalue = save_prevvalue;
13127 /* Deal with this element of the class */
13130 cp_list = _add_range_to_invlist(cp_list, prevvalue, value);
13132 SV* this_range = _new_invlist(1);
13133 _append_range_to_invlist(this_range, prevvalue, value);
13135 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
13136 * If this range was specified using something like 'i-j', we want
13137 * to include only the 'i' and the 'j', and not anything in
13138 * between, so exclude non-ASCII, non-alphabetics from it.
13139 * However, if the range was specified with something like
13140 * [\x89-\x91] or [\x89-j], all code points within it should be
13141 * included. literal_endpoint==2 means both ends of the range used
13142 * a literal character, not \x{foo} */
13143 if (literal_endpoint == 2
13144 && (prevvalue >= 'a' && value <= 'z')
13145 || (prevvalue >= 'A' && value <= 'Z'))
13147 _invlist_intersection(this_range, PL_Posix_ptrs[_CC_ALPHA],
13150 _invlist_union(cp_list, this_range, &cp_list);
13151 literal_endpoint = 0;
13155 range = 0; /* this range (if it was one) is done now */
13156 } /* End of loop through all the text within the brackets */
13158 /* If anything in the class expands to more than one character, we have to
13159 * deal with them by building up a substitute parse string, and recursively
13160 * calling reg() on it, instead of proceeding */
13161 if (multi_char_matches) {
13162 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
13165 char *save_end = RExC_end;
13166 char *save_parse = RExC_parse;
13167 bool first_time = TRUE; /* First multi-char occurrence doesn't get
13172 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
13173 because too confusing */
13175 sv_catpv(substitute_parse, "(?:");
13179 /* Look at the longest folds first */
13180 for (cp_count = av_len(multi_char_matches); cp_count > 0; cp_count--) {
13182 if (av_exists(multi_char_matches, cp_count)) {
13183 AV** this_array_ptr;
13186 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13188 while ((this_sequence = av_pop(*this_array_ptr)) !=
13191 if (! first_time) {
13192 sv_catpv(substitute_parse, "|");
13194 first_time = FALSE;
13196 sv_catpv(substitute_parse, SvPVX(this_sequence));
13201 /* If the character class contains anything else besides these
13202 * multi-character folds, have to include it in recursive parsing */
13203 if (element_count) {
13204 sv_catpv(substitute_parse, "|[");
13205 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
13206 sv_catpv(substitute_parse, "]");
13209 sv_catpv(substitute_parse, ")");
13212 /* This is a way to get the parse to skip forward a whole named
13213 * sequence instead of matching the 2nd character when it fails the
13215 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
13219 RExC_parse = SvPV(substitute_parse, len);
13220 RExC_end = RExC_parse + len;
13221 RExC_in_multi_char_class = 1;
13222 RExC_emit = (regnode *)orig_emit;
13224 ret = reg(pRExC_state, 1, ®_flags, depth+1);
13226 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
13228 RExC_parse = save_parse;
13229 RExC_end = save_end;
13230 RExC_in_multi_char_class = 0;
13231 SvREFCNT_dec_NN(multi_char_matches);
13235 /* If the character class contains only a single element, it may be
13236 * optimizable into another node type which is smaller and runs faster.
13237 * Check if this is the case for this class */
13238 if (element_count == 1 && ! ret_invlist) {
13242 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like \w or
13243 [:digit:] or \p{foo} */
13245 /* All named classes are mapped into POSIXish nodes, with its FLAG
13246 * argument giving which class it is */
13247 switch ((I32)namedclass) {
13248 case ANYOF_UNIPROP:
13251 /* These don't depend on the charset modifiers. They always
13252 * match under /u rules */
13253 case ANYOF_NHORIZWS:
13254 case ANYOF_HORIZWS:
13255 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
13258 case ANYOF_NVERTWS:
13263 /* The actual POSIXish node for all the rest depends on the
13264 * charset modifier. The ones in the first set depend only on
13265 * ASCII or, if available on this platform, locale */
13269 op = (LOC) ? POSIXL : POSIXA;
13280 /* under /a could be alpha */
13282 if (ASCII_RESTRICTED) {
13283 namedclass = ANYOF_ALPHA + (namedclass % 2);
13291 /* The rest have more possibilities depending on the charset.
13292 * We take advantage of the enum ordering of the charset
13293 * modifiers to get the exact node type, */
13295 op = POSIXD + get_regex_charset(RExC_flags);
13296 if (op > POSIXA) { /* /aa is same as /a */
13299 #ifndef HAS_ISBLANK
13301 && (namedclass == ANYOF_BLANK
13302 || namedclass == ANYOF_NBLANK))
13309 /* The odd numbered ones are the complements of the
13310 * next-lower even number one */
13311 if (namedclass % 2 == 1) {
13315 arg = namedclass_to_classnum(namedclass);
13319 else if (value == prevvalue) {
13321 /* Here, the class consists of just a single code point */
13324 if (! LOC && value == '\n') {
13325 op = REG_ANY; /* Optimize [^\n] */
13326 *flagp |= HASWIDTH|SIMPLE;
13330 else if (value < 256 || UTF) {
13332 /* Optimize a single value into an EXACTish node, but not if it
13333 * would require converting the pattern to UTF-8. */
13334 op = compute_EXACTish(pRExC_state);
13336 } /* Otherwise is a range */
13337 else if (! LOC) { /* locale could vary these */
13338 if (prevvalue == '0') {
13339 if (value == '9') {
13346 /* Here, we have changed <op> away from its initial value iff we found
13347 * an optimization */
13350 /* Throw away this ANYOF regnode, and emit the calculated one,
13351 * which should correspond to the beginning, not current, state of
13353 const char * cur_parse = RExC_parse;
13354 RExC_parse = (char *)orig_parse;
13358 /* To get locale nodes to not use the full ANYOF size would
13359 * require moving the code above that writes the portions
13360 * of it that aren't in other nodes to after this point.
13361 * e.g. ANYOF_CLASS_SET */
13362 RExC_size = orig_size;
13366 RExC_emit = (regnode *)orig_emit;
13367 if (PL_regkind[op] == POSIXD) {
13369 op += NPOSIXD - POSIXD;
13374 ret = reg_node(pRExC_state, op);
13376 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
13380 *flagp |= HASWIDTH|SIMPLE;
13382 else if (PL_regkind[op] == EXACT) {
13383 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13386 RExC_parse = (char *) cur_parse;
13388 SvREFCNT_dec(posixes);
13389 SvREFCNT_dec(cp_list);
13396 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
13398 /* If folding, we calculate all characters that could fold to or from the
13399 * ones already on the list */
13400 if (FOLD && cp_list) {
13401 UV start, end; /* End points of code point ranges */
13403 SV* fold_intersection = NULL;
13405 /* If the highest code point is within Latin1, we can use the
13406 * compiled-in Alphas list, and not have to go out to disk. This
13407 * yields two false positives, the masculine and feminine ordinal
13408 * indicators, which are weeded out below using the
13409 * IS_IN_SOME_FOLD_L1() macro */
13410 if (invlist_highest(cp_list) < 256) {
13411 _invlist_intersection(PL_L1Posix_ptrs[_CC_ALPHA], cp_list,
13412 &fold_intersection);
13416 /* Here, there are non-Latin1 code points, so we will have to go
13417 * fetch the list of all the characters that participate in folds
13419 if (! PL_utf8_foldable) {
13420 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13421 &PL_sv_undef, 1, 0);
13422 PL_utf8_foldable = _get_swash_invlist(swash);
13423 SvREFCNT_dec_NN(swash);
13426 /* This is a hash that for a particular fold gives all characters
13427 * that are involved in it */
13428 if (! PL_utf8_foldclosures) {
13430 /* If we were unable to find any folds, then we likely won't be
13431 * able to find the closures. So just create an empty list.
13432 * Folding will effectively be restricted to the non-Unicode
13433 * rules hard-coded into Perl. (This case happens legitimately
13434 * during compilation of Perl itself before the Unicode tables
13435 * are generated) */
13436 if (_invlist_len(PL_utf8_foldable) == 0) {
13437 PL_utf8_foldclosures = newHV();
13440 /* If the folds haven't been read in, call a fold function
13442 if (! PL_utf8_tofold) {
13443 U8 dummy[UTF8_MAXBYTES+1];
13445 /* This string is just a short named one above \xff */
13446 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
13447 assert(PL_utf8_tofold); /* Verify that worked */
13449 PL_utf8_foldclosures =
13450 _swash_inversion_hash(PL_utf8_tofold);
13454 /* Only the characters in this class that participate in folds need
13455 * be checked. Get the intersection of this class and all the
13456 * possible characters that are foldable. This can quickly narrow
13457 * down a large class */
13458 _invlist_intersection(PL_utf8_foldable, cp_list,
13459 &fold_intersection);
13462 /* Now look at the foldable characters in this class individually */
13463 invlist_iterinit(fold_intersection);
13464 while (invlist_iternext(fold_intersection, &start, &end)) {
13467 /* Locale folding for Latin1 characters is deferred until runtime */
13468 if (LOC && start < 256) {
13472 /* Look at every character in the range */
13473 for (j = start; j <= end; j++) {
13475 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
13481 /* We have the latin1 folding rules hard-coded here so that
13482 * an innocent-looking character class, like /[ks]/i won't
13483 * have to go out to disk to find the possible matches.
13484 * XXX It would be better to generate these via regen, in
13485 * case a new version of the Unicode standard adds new
13486 * mappings, though that is not really likely, and may be
13487 * caught by the default: case of the switch below. */
13489 if (IS_IN_SOME_FOLD_L1(j)) {
13491 /* ASCII is always matched; non-ASCII is matched only
13492 * under Unicode rules */
13493 if (isASCII(j) || AT_LEAST_UNI_SEMANTICS) {
13495 add_cp_to_invlist(cp_list, PL_fold_latin1[j]);
13499 add_cp_to_invlist(depends_list, PL_fold_latin1[j]);
13503 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
13504 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
13506 /* Certain Latin1 characters have matches outside
13507 * Latin1. To get here, <j> is one of those
13508 * characters. None of these matches is valid for
13509 * ASCII characters under /aa, which is why the 'if'
13510 * just above excludes those. These matches only
13511 * happen when the target string is utf8. The code
13512 * below adds the single fold closures for <j> to the
13513 * inversion list. */
13518 add_cp_to_invlist(cp_list, KELVIN_SIGN);
13522 cp_list = add_cp_to_invlist(cp_list,
13523 LATIN_SMALL_LETTER_LONG_S);
13526 cp_list = add_cp_to_invlist(cp_list,
13527 GREEK_CAPITAL_LETTER_MU);
13528 cp_list = add_cp_to_invlist(cp_list,
13529 GREEK_SMALL_LETTER_MU);
13531 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
13532 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
13534 add_cp_to_invlist(cp_list, ANGSTROM_SIGN);
13536 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
13537 cp_list = add_cp_to_invlist(cp_list,
13538 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
13540 case LATIN_SMALL_LETTER_SHARP_S:
13541 cp_list = add_cp_to_invlist(cp_list,
13542 LATIN_CAPITAL_LETTER_SHARP_S);
13544 case 'F': case 'f':
13545 case 'I': case 'i':
13546 case 'L': case 'l':
13547 case 'T': case 't':
13548 case 'A': case 'a':
13549 case 'H': case 'h':
13550 case 'J': case 'j':
13551 case 'N': case 'n':
13552 case 'W': case 'w':
13553 case 'Y': case 'y':
13554 /* These all are targets of multi-character
13555 * folds from code points that require UTF8 to
13556 * express, so they can't match unless the
13557 * target string is in UTF-8, so no action here
13558 * is necessary, as regexec.c properly handles
13559 * the general case for UTF-8 matching and
13560 * multi-char folds */
13563 /* Use deprecated warning to increase the
13564 * chances of this being output */
13565 ckWARN2regdep(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
13572 /* Here is an above Latin1 character. We don't have the rules
13573 * hard-coded for it. First, get its fold. This is the simple
13574 * fold, as the multi-character folds have been handled earlier
13575 * and separated out */
13576 _to_uni_fold_flags(j, foldbuf, &foldlen,
13578 ? FOLD_FLAGS_LOCALE
13579 : (ASCII_FOLD_RESTRICTED)
13580 ? FOLD_FLAGS_NOMIX_ASCII
13583 /* Single character fold of above Latin1. Add everything in
13584 * its fold closure to the list that this node should match.
13585 * The fold closures data structure is a hash with the keys
13586 * being the UTF-8 of every character that is folded to, like
13587 * 'k', and the values each an array of all code points that
13588 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
13589 * Multi-character folds are not included */
13590 if ((listp = hv_fetch(PL_utf8_foldclosures,
13591 (char *) foldbuf, foldlen, FALSE)))
13593 AV* list = (AV*) *listp;
13595 for (k = 0; k <= av_len(list); k++) {
13596 SV** c_p = av_fetch(list, k, FALSE);
13599 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
13603 /* /aa doesn't allow folds between ASCII and non-; /l
13604 * doesn't allow them between above and below 256 */
13605 if ((ASCII_FOLD_RESTRICTED
13606 && (isASCII(c) != isASCII(j)))
13607 || (LOC && c < 256)) {
13611 /* Folds involving non-ascii Latin1 characters
13612 * under /d are added to a separate list */
13613 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
13615 cp_list = add_cp_to_invlist(cp_list, c);
13618 depends_list = add_cp_to_invlist(depends_list, c);
13624 SvREFCNT_dec_NN(fold_intersection);
13627 /* And combine the result (if any) with any inversion list from posix
13628 * classes. The lists are kept separate up to now because we don't want to
13629 * fold the classes (folding of those is automatically handled by the swash
13630 * fetching code) */
13632 if (! DEPENDS_SEMANTICS) {
13634 _invlist_union(cp_list, posixes, &cp_list);
13635 SvREFCNT_dec_NN(posixes);
13642 /* Under /d, we put into a separate list the Latin1 things that
13643 * match only when the target string is utf8 */
13644 SV* nonascii_but_latin1_properties = NULL;
13645 _invlist_intersection(posixes, PL_Latin1,
13646 &nonascii_but_latin1_properties);
13647 _invlist_subtract(nonascii_but_latin1_properties, PL_ASCII,
13648 &nonascii_but_latin1_properties);
13649 _invlist_subtract(posixes, nonascii_but_latin1_properties,
13652 _invlist_union(cp_list, posixes, &cp_list);
13653 SvREFCNT_dec_NN(posixes);
13659 if (depends_list) {
13660 _invlist_union(depends_list, nonascii_but_latin1_properties,
13662 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
13665 depends_list = nonascii_but_latin1_properties;
13670 /* And combine the result (if any) with any inversion list from properties.
13671 * The lists are kept separate up to now so that we can distinguish the two
13672 * in regards to matching above-Unicode. A run-time warning is generated
13673 * if a Unicode property is matched against a non-Unicode code point. But,
13674 * we allow user-defined properties to match anything, without any warning,
13675 * and we also suppress the warning if there is a portion of the character
13676 * class that isn't a Unicode property, and which matches above Unicode, \W
13677 * or [\x{110000}] for example.
13678 * (Note that in this case, unlike the Posix one above, there is no
13679 * <depends_list>, because having a Unicode property forces Unicode
13682 bool warn_super = ! has_user_defined_property;
13685 /* If it matters to the final outcome, see if a non-property
13686 * component of the class matches above Unicode. If so, the
13687 * warning gets suppressed. This is true even if just a single
13688 * such code point is specified, as though not strictly correct if
13689 * another such code point is matched against, the fact that they
13690 * are using above-Unicode code points indicates they should know
13691 * the issues involved */
13693 bool non_prop_matches_above_Unicode =
13694 runtime_posix_matches_above_Unicode
13695 | (invlist_highest(cp_list) > PERL_UNICODE_MAX);
13697 non_prop_matches_above_Unicode =
13698 ! non_prop_matches_above_Unicode;
13700 warn_super = ! non_prop_matches_above_Unicode;
13703 _invlist_union(properties, cp_list, &cp_list);
13704 SvREFCNT_dec_NN(properties);
13707 cp_list = properties;
13711 OP(ret) = ANYOF_WARN_SUPER;
13715 /* Here, we have calculated what code points should be in the character
13718 * Now we can see about various optimizations. Fold calculation (which we
13719 * did above) needs to take place before inversion. Otherwise /[^k]/i
13720 * would invert to include K, which under /i would match k, which it
13721 * shouldn't. Therefore we can't invert folded locale now, as it won't be
13722 * folded until runtime */
13724 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
13725 * at compile time. Besides not inverting folded locale now, we can't
13726 * invert if there are things such as \w, which aren't known until runtime
13729 && ! (LOC && (FOLD || (ANYOF_FLAGS(ret) & ANYOF_CLASS)))
13731 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13733 _invlist_invert(cp_list);
13735 /* Any swash can't be used as-is, because we've inverted things */
13737 SvREFCNT_dec_NN(swash);
13741 /* Clear the invert flag since have just done it here */
13746 *ret_invlist = cp_list;
13747 SvREFCNT_dec(swash);
13749 /* Discard the generated node */
13751 RExC_size = orig_size;
13754 RExC_emit = orig_emit;
13759 /* If we didn't do folding, it's because some information isn't available
13760 * until runtime; set the run-time fold flag for these. (We don't have to
13761 * worry about properties folding, as that is taken care of by the swash
13765 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
13768 /* Some character classes are equivalent to other nodes. Such nodes take
13769 * up less room and generally fewer operations to execute than ANYOF nodes.
13770 * Above, we checked for and optimized into some such equivalents for
13771 * certain common classes that are easy to test. Getting to this point in
13772 * the code means that the class didn't get optimized there. Since this
13773 * code is only executed in Pass 2, it is too late to save space--it has
13774 * been allocated in Pass 1, and currently isn't given back. But turning
13775 * things into an EXACTish node can allow the optimizer to join it to any
13776 * adjacent such nodes. And if the class is equivalent to things like /./,
13777 * expensive run-time swashes can be avoided. Now that we have more
13778 * complete information, we can find things necessarily missed by the
13779 * earlier code. I (khw) am not sure how much to look for here. It would
13780 * be easy, but perhaps too slow, to check any candidates against all the
13781 * node types they could possibly match using _invlistEQ(). */
13786 && ! (ANYOF_FLAGS(ret) & ANYOF_CLASS)
13787 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13790 U8 op = END; /* The optimzation node-type */
13791 const char * cur_parse= RExC_parse;
13793 invlist_iterinit(cp_list);
13794 if (! invlist_iternext(cp_list, &start, &end)) {
13796 /* Here, the list is empty. This happens, for example, when a
13797 * Unicode property is the only thing in the character class, and
13798 * it doesn't match anything. (perluniprops.pod notes such
13801 *flagp |= HASWIDTH|SIMPLE;
13803 else if (start == end) { /* The range is a single code point */
13804 if (! invlist_iternext(cp_list, &start, &end)
13806 /* Don't do this optimization if it would require changing
13807 * the pattern to UTF-8 */
13808 && (start < 256 || UTF))
13810 /* Here, the list contains a single code point. Can optimize
13811 * into an EXACT node */
13820 /* A locale node under folding with one code point can be
13821 * an EXACTFL, as its fold won't be calculated until
13827 /* Here, we are generally folding, but there is only one
13828 * code point to match. If we have to, we use an EXACT
13829 * node, but it would be better for joining with adjacent
13830 * nodes in the optimization pass if we used the same
13831 * EXACTFish node that any such are likely to be. We can
13832 * do this iff the code point doesn't participate in any
13833 * folds. For example, an EXACTF of a colon is the same as
13834 * an EXACT one, since nothing folds to or from a colon. */
13836 if (IS_IN_SOME_FOLD_L1(value)) {
13841 if (! PL_utf8_foldable) {
13842 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13843 &PL_sv_undef, 1, 0);
13844 PL_utf8_foldable = _get_swash_invlist(swash);
13845 SvREFCNT_dec_NN(swash);
13847 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
13852 /* If we haven't found the node type, above, it means we
13853 * can use the prevailing one */
13855 op = compute_EXACTish(pRExC_state);
13860 else if (start == 0) {
13861 if (end == UV_MAX) {
13863 *flagp |= HASWIDTH|SIMPLE;
13866 else if (end == '\n' - 1
13867 && invlist_iternext(cp_list, &start, &end)
13868 && start == '\n' + 1 && end == UV_MAX)
13871 *flagp |= HASWIDTH|SIMPLE;
13875 invlist_iterfinish(cp_list);
13878 RExC_parse = (char *)orig_parse;
13879 RExC_emit = (regnode *)orig_emit;
13881 ret = reg_node(pRExC_state, op);
13883 RExC_parse = (char *)cur_parse;
13885 if (PL_regkind[op] == EXACT) {
13886 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13889 SvREFCNT_dec_NN(cp_list);
13894 /* Here, <cp_list> contains all the code points we can determine at
13895 * compile time that match under all conditions. Go through it, and
13896 * for things that belong in the bitmap, put them there, and delete from
13897 * <cp_list>. While we are at it, see if everything above 255 is in the
13898 * list, and if so, set a flag to speed up execution */
13899 ANYOF_BITMAP_ZERO(ret);
13902 /* This gets set if we actually need to modify things */
13903 bool change_invlist = FALSE;
13907 /* Start looking through <cp_list> */
13908 invlist_iterinit(cp_list);
13909 while (invlist_iternext(cp_list, &start, &end)) {
13913 if (end == UV_MAX && start <= 256) {
13914 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
13917 /* Quit if are above what we should change */
13922 change_invlist = TRUE;
13924 /* Set all the bits in the range, up to the max that we are doing */
13925 high = (end < 255) ? end : 255;
13926 for (i = start; i <= (int) high; i++) {
13927 if (! ANYOF_BITMAP_TEST(ret, i)) {
13928 ANYOF_BITMAP_SET(ret, i);
13934 invlist_iterfinish(cp_list);
13936 /* Done with loop; remove any code points that are in the bitmap from
13938 if (change_invlist) {
13939 _invlist_subtract(cp_list, PL_Latin1, &cp_list);
13942 /* If have completely emptied it, remove it completely */
13943 if (_invlist_len(cp_list) == 0) {
13944 SvREFCNT_dec_NN(cp_list);
13950 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
13953 /* Here, the bitmap has been populated with all the Latin1 code points that
13954 * always match. Can now add to the overall list those that match only
13955 * when the target string is UTF-8 (<depends_list>). */
13956 if (depends_list) {
13958 _invlist_union(cp_list, depends_list, &cp_list);
13959 SvREFCNT_dec_NN(depends_list);
13962 cp_list = depends_list;
13966 /* If there is a swash and more than one element, we can't use the swash in
13967 * the optimization below. */
13968 if (swash && element_count > 1) {
13969 SvREFCNT_dec_NN(swash);
13974 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13976 ARG_SET(ret, ANYOF_NONBITMAP_EMPTY);
13979 /* av[0] stores the character class description in its textual form:
13980 * used later (regexec.c:Perl_regclass_swash()) to initialize the
13981 * appropriate swash, and is also useful for dumping the regnode.
13982 * av[1] if NULL, is a placeholder to later contain the swash computed
13983 * from av[0]. But if no further computation need be done, the
13984 * swash is stored there now.
13985 * av[2] stores the cp_list inversion list for use in addition or
13986 * instead of av[0]; used only if av[1] is NULL
13987 * av[3] is set if any component of the class is from a user-defined
13988 * property; used only if av[1] is NULL */
13989 AV * const av = newAV();
13992 av_store(av, 0, (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13993 ? SvREFCNT_inc(listsv) : &PL_sv_undef);
13995 av_store(av, 1, swash);
13996 SvREFCNT_dec_NN(cp_list);
13999 av_store(av, 1, NULL);
14001 av_store(av, 2, cp_list);
14002 av_store(av, 3, newSVuv(has_user_defined_property));
14006 rv = newRV_noinc(MUTABLE_SV(av));
14007 n = add_data(pRExC_state, 1, "s");
14008 RExC_rxi->data->data[n] = (void*)rv;
14012 *flagp |= HASWIDTH|SIMPLE;
14015 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
14018 /* reg_skipcomment()
14020 Absorbs an /x style # comments from the input stream.
14021 Returns true if there is more text remaining in the stream.
14022 Will set the REG_SEEN_RUN_ON_COMMENT flag if the comment
14023 terminates the pattern without including a newline.
14025 Note its the callers responsibility to ensure that we are
14026 actually in /x mode
14031 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
14035 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
14037 while (RExC_parse < RExC_end)
14038 if (*RExC_parse++ == '\n') {
14043 /* we ran off the end of the pattern without ending
14044 the comment, so we have to add an \n when wrapping */
14045 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
14053 Advances the parse position, and optionally absorbs
14054 "whitespace" from the inputstream.
14056 Without /x "whitespace" means (?#...) style comments only,
14057 with /x this means (?#...) and # comments and whitespace proper.
14059 Returns the RExC_parse point from BEFORE the scan occurs.
14061 This is the /x friendly way of saying RExC_parse++.
14065 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
14067 char* const retval = RExC_parse++;
14069 PERL_ARGS_ASSERT_NEXTCHAR;
14072 if (RExC_end - RExC_parse >= 3
14073 && *RExC_parse == '('
14074 && RExC_parse[1] == '?'
14075 && RExC_parse[2] == '#')
14077 while (*RExC_parse != ')') {
14078 if (RExC_parse == RExC_end)
14079 FAIL("Sequence (?#... not terminated");
14085 if (RExC_flags & RXf_PMf_EXTENDED) {
14086 if (isSPACE(*RExC_parse)) {
14090 else if (*RExC_parse == '#') {
14091 if ( reg_skipcomment( pRExC_state ) )
14100 - reg_node - emit a node
14102 STATIC regnode * /* Location. */
14103 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
14107 regnode * const ret = RExC_emit;
14108 GET_RE_DEBUG_FLAGS_DECL;
14110 PERL_ARGS_ASSERT_REG_NODE;
14113 SIZE_ALIGN(RExC_size);
14117 if (RExC_emit >= RExC_emit_bound)
14118 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14119 op, RExC_emit, RExC_emit_bound);
14121 NODE_ALIGN_FILL(ret);
14123 FILL_ADVANCE_NODE(ptr, op);
14124 #ifdef RE_TRACK_PATTERN_OFFSETS
14125 if (RExC_offsets) { /* MJD */
14126 MJD_OFFSET_DEBUG(("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
14127 "reg_node", __LINE__,
14129 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
14130 ? "Overwriting end of array!\n" : "OK",
14131 (UV)(RExC_emit - RExC_emit_start),
14132 (UV)(RExC_parse - RExC_start),
14133 (UV)RExC_offsets[0]));
14134 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
14142 - reganode - emit a node with an argument
14144 STATIC regnode * /* Location. */
14145 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
14149 regnode * const ret = RExC_emit;
14150 GET_RE_DEBUG_FLAGS_DECL;
14152 PERL_ARGS_ASSERT_REGANODE;
14155 SIZE_ALIGN(RExC_size);
14160 assert(2==regarglen[op]+1);
14162 Anything larger than this has to allocate the extra amount.
14163 If we changed this to be:
14165 RExC_size += (1 + regarglen[op]);
14167 then it wouldn't matter. Its not clear what side effect
14168 might come from that so its not done so far.
14173 if (RExC_emit >= RExC_emit_bound)
14174 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14175 op, RExC_emit, RExC_emit_bound);
14177 NODE_ALIGN_FILL(ret);
14179 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
14180 #ifdef RE_TRACK_PATTERN_OFFSETS
14181 if (RExC_offsets) { /* MJD */
14182 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14186 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
14187 "Overwriting end of array!\n" : "OK",
14188 (UV)(RExC_emit - RExC_emit_start),
14189 (UV)(RExC_parse - RExC_start),
14190 (UV)RExC_offsets[0]));
14191 Set_Cur_Node_Offset;
14199 - reguni - emit (if appropriate) a Unicode character
14202 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
14206 PERL_ARGS_ASSERT_REGUNI;
14208 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
14212 - reginsert - insert an operator in front of already-emitted operand
14214 * Means relocating the operand.
14217 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
14223 const int offset = regarglen[(U8)op];
14224 const int size = NODE_STEP_REGNODE + offset;
14225 GET_RE_DEBUG_FLAGS_DECL;
14227 PERL_ARGS_ASSERT_REGINSERT;
14228 PERL_UNUSED_ARG(depth);
14229 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
14230 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
14239 if (RExC_open_parens) {
14241 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
14242 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
14243 if ( RExC_open_parens[paren] >= opnd ) {
14244 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
14245 RExC_open_parens[paren] += size;
14247 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
14249 if ( RExC_close_parens[paren] >= opnd ) {
14250 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
14251 RExC_close_parens[paren] += size;
14253 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
14258 while (src > opnd) {
14259 StructCopy(--src, --dst, regnode);
14260 #ifdef RE_TRACK_PATTERN_OFFSETS
14261 if (RExC_offsets) { /* MJD 20010112 */
14262 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
14266 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
14267 ? "Overwriting end of array!\n" : "OK",
14268 (UV)(src - RExC_emit_start),
14269 (UV)(dst - RExC_emit_start),
14270 (UV)RExC_offsets[0]));
14271 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
14272 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
14278 place = opnd; /* Op node, where operand used to be. */
14279 #ifdef RE_TRACK_PATTERN_OFFSETS
14280 if (RExC_offsets) { /* MJD */
14281 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14285 (UV)(place - RExC_emit_start) > RExC_offsets[0]
14286 ? "Overwriting end of array!\n" : "OK",
14287 (UV)(place - RExC_emit_start),
14288 (UV)(RExC_parse - RExC_start),
14289 (UV)RExC_offsets[0]));
14290 Set_Node_Offset(place, RExC_parse);
14291 Set_Node_Length(place, 1);
14294 src = NEXTOPER(place);
14295 FILL_ADVANCE_NODE(place, op);
14296 Zero(src, offset, regnode);
14300 - regtail - set the next-pointer at the end of a node chain of p to val.
14301 - SEE ALSO: regtail_study
14303 /* TODO: All three parms should be const */
14305 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14309 GET_RE_DEBUG_FLAGS_DECL;
14311 PERL_ARGS_ASSERT_REGTAIL;
14313 PERL_UNUSED_ARG(depth);
14319 /* Find last node. */
14322 regnode * const temp = regnext(scan);
14324 SV * const mysv=sv_newmortal();
14325 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
14326 regprop(RExC_rx, mysv, scan);
14327 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
14328 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
14329 (temp == NULL ? "->" : ""),
14330 (temp == NULL ? PL_reg_name[OP(val)] : "")
14338 if (reg_off_by_arg[OP(scan)]) {
14339 ARG_SET(scan, val - scan);
14342 NEXT_OFF(scan) = val - scan;
14348 - regtail_study - set the next-pointer at the end of a node chain of p to val.
14349 - Look for optimizable sequences at the same time.
14350 - currently only looks for EXACT chains.
14352 This is experimental code. The idea is to use this routine to perform
14353 in place optimizations on branches and groups as they are constructed,
14354 with the long term intention of removing optimization from study_chunk so
14355 that it is purely analytical.
14357 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
14358 to control which is which.
14361 /* TODO: All four parms should be const */
14364 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14369 #ifdef EXPERIMENTAL_INPLACESCAN
14372 GET_RE_DEBUG_FLAGS_DECL;
14374 PERL_ARGS_ASSERT_REGTAIL_STUDY;
14380 /* Find last node. */
14384 regnode * const temp = regnext(scan);
14385 #ifdef EXPERIMENTAL_INPLACESCAN
14386 if (PL_regkind[OP(scan)] == EXACT) {
14387 bool has_exactf_sharp_s; /* Unexamined in this routine */
14388 if (join_exact(pRExC_state,scan,&min, &has_exactf_sharp_s, 1,val,depth+1))
14393 switch (OP(scan)) {
14399 case EXACTFU_TRICKYFOLD:
14401 if( exact == PSEUDO )
14403 else if ( exact != OP(scan) )
14412 SV * const mysv=sv_newmortal();
14413 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
14414 regprop(RExC_rx, mysv, scan);
14415 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
14416 SvPV_nolen_const(mysv),
14417 REG_NODE_NUM(scan),
14418 PL_reg_name[exact]);
14425 SV * const mysv_val=sv_newmortal();
14426 DEBUG_PARSE_MSG("");
14427 regprop(RExC_rx, mysv_val, val);
14428 PerlIO_printf(Perl_debug_log, "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
14429 SvPV_nolen_const(mysv_val),
14430 (IV)REG_NODE_NUM(val),
14434 if (reg_off_by_arg[OP(scan)]) {
14435 ARG_SET(scan, val - scan);
14438 NEXT_OFF(scan) = val - scan;
14446 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
14451 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
14456 for (bit=0; bit<32; bit++) {
14457 if (flags & (1<<bit)) {
14458 if (!set++ && lead)
14459 PerlIO_printf(Perl_debug_log, "%s",lead);
14460 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_intflags_name[bit]);
14465 PerlIO_printf(Perl_debug_log, "\n");
14467 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14472 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
14478 for (bit=0; bit<32; bit++) {
14479 if (flags & (1<<bit)) {
14480 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
14483 if (!set++ && lead)
14484 PerlIO_printf(Perl_debug_log, "%s",lead);
14485 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
14488 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
14489 if (!set++ && lead) {
14490 PerlIO_printf(Perl_debug_log, "%s",lead);
14493 case REGEX_UNICODE_CHARSET:
14494 PerlIO_printf(Perl_debug_log, "UNICODE");
14496 case REGEX_LOCALE_CHARSET:
14497 PerlIO_printf(Perl_debug_log, "LOCALE");
14499 case REGEX_ASCII_RESTRICTED_CHARSET:
14500 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
14502 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
14503 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
14506 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
14512 PerlIO_printf(Perl_debug_log, "\n");
14514 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14520 Perl_regdump(pTHX_ const regexp *r)
14524 SV * const sv = sv_newmortal();
14525 SV *dsv= sv_newmortal();
14526 RXi_GET_DECL(r,ri);
14527 GET_RE_DEBUG_FLAGS_DECL;
14529 PERL_ARGS_ASSERT_REGDUMP;
14531 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
14533 /* Header fields of interest. */
14534 if (r->anchored_substr) {
14535 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
14536 RE_SV_DUMPLEN(r->anchored_substr), 30);
14537 PerlIO_printf(Perl_debug_log,
14538 "anchored %s%s at %"IVdf" ",
14539 s, RE_SV_TAIL(r->anchored_substr),
14540 (IV)r->anchored_offset);
14541 } else if (r->anchored_utf8) {
14542 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
14543 RE_SV_DUMPLEN(r->anchored_utf8), 30);
14544 PerlIO_printf(Perl_debug_log,
14545 "anchored utf8 %s%s at %"IVdf" ",
14546 s, RE_SV_TAIL(r->anchored_utf8),
14547 (IV)r->anchored_offset);
14549 if (r->float_substr) {
14550 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
14551 RE_SV_DUMPLEN(r->float_substr), 30);
14552 PerlIO_printf(Perl_debug_log,
14553 "floating %s%s at %"IVdf"..%"UVuf" ",
14554 s, RE_SV_TAIL(r->float_substr),
14555 (IV)r->float_min_offset, (UV)r->float_max_offset);
14556 } else if (r->float_utf8) {
14557 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
14558 RE_SV_DUMPLEN(r->float_utf8), 30);
14559 PerlIO_printf(Perl_debug_log,
14560 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
14561 s, RE_SV_TAIL(r->float_utf8),
14562 (IV)r->float_min_offset, (UV)r->float_max_offset);
14564 if (r->check_substr || r->check_utf8)
14565 PerlIO_printf(Perl_debug_log,
14567 (r->check_substr == r->float_substr
14568 && r->check_utf8 == r->float_utf8
14569 ? "(checking floating" : "(checking anchored"));
14570 if (r->extflags & RXf_NOSCAN)
14571 PerlIO_printf(Perl_debug_log, " noscan");
14572 if (r->extflags & RXf_CHECK_ALL)
14573 PerlIO_printf(Perl_debug_log, " isall");
14574 if (r->check_substr || r->check_utf8)
14575 PerlIO_printf(Perl_debug_log, ") ");
14577 if (ri->regstclass) {
14578 regprop(r, sv, ri->regstclass);
14579 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
14581 if (r->extflags & RXf_ANCH) {
14582 PerlIO_printf(Perl_debug_log, "anchored");
14583 if (r->extflags & RXf_ANCH_BOL)
14584 PerlIO_printf(Perl_debug_log, "(BOL)");
14585 if (r->extflags & RXf_ANCH_MBOL)
14586 PerlIO_printf(Perl_debug_log, "(MBOL)");
14587 if (r->extflags & RXf_ANCH_SBOL)
14588 PerlIO_printf(Perl_debug_log, "(SBOL)");
14589 if (r->extflags & RXf_ANCH_GPOS)
14590 PerlIO_printf(Perl_debug_log, "(GPOS)");
14591 PerlIO_putc(Perl_debug_log, ' ');
14593 if (r->extflags & RXf_GPOS_SEEN)
14594 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
14595 if (r->intflags & PREGf_SKIP)
14596 PerlIO_printf(Perl_debug_log, "plus ");
14597 if (r->intflags & PREGf_IMPLICIT)
14598 PerlIO_printf(Perl_debug_log, "implicit ");
14599 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
14600 if (r->extflags & RXf_EVAL_SEEN)
14601 PerlIO_printf(Perl_debug_log, "with eval ");
14602 PerlIO_printf(Perl_debug_log, "\n");
14604 regdump_extflags("r->extflags: ",r->extflags);
14605 regdump_intflags("r->intflags: ",r->intflags);
14608 PERL_ARGS_ASSERT_REGDUMP;
14609 PERL_UNUSED_CONTEXT;
14610 PERL_UNUSED_ARG(r);
14611 #endif /* DEBUGGING */
14615 - regprop - printable representation of opcode
14617 #define EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags) \
14620 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]); \
14621 if (flags & ANYOF_INVERT) \
14622 /*make sure the invert info is in each */ \
14623 sv_catpvs(sv, "^"); \
14629 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
14635 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
14636 static const char * const anyofs[] = {
14637 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
14638 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
14639 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
14640 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
14641 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
14642 || _CC_VERTSPACE != 16
14643 #error Need to adjust order of anyofs[]
14680 RXi_GET_DECL(prog,progi);
14681 GET_RE_DEBUG_FLAGS_DECL;
14683 PERL_ARGS_ASSERT_REGPROP;
14687 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
14688 /* It would be nice to FAIL() here, but this may be called from
14689 regexec.c, and it would be hard to supply pRExC_state. */
14690 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(o), (int)REGNODE_MAX);
14691 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
14693 k = PL_regkind[OP(o)];
14696 sv_catpvs(sv, " ");
14697 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
14698 * is a crude hack but it may be the best for now since
14699 * we have no flag "this EXACTish node was UTF-8"
14701 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
14702 PERL_PV_ESCAPE_UNI_DETECT |
14703 PERL_PV_ESCAPE_NONASCII |
14704 PERL_PV_PRETTY_ELLIPSES |
14705 PERL_PV_PRETTY_LTGT |
14706 PERL_PV_PRETTY_NOCLEAR
14708 } else if (k == TRIE) {
14709 /* print the details of the trie in dumpuntil instead, as
14710 * progi->data isn't available here */
14711 const char op = OP(o);
14712 const U32 n = ARG(o);
14713 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
14714 (reg_ac_data *)progi->data->data[n] :
14716 const reg_trie_data * const trie
14717 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
14719 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
14720 DEBUG_TRIE_COMPILE_r(
14721 Perl_sv_catpvf(aTHX_ sv,
14722 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
14723 (UV)trie->startstate,
14724 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
14725 (UV)trie->wordcount,
14728 (UV)TRIE_CHARCOUNT(trie),
14729 (UV)trie->uniquecharcount
14732 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
14734 int rangestart = -1;
14735 U8* bitmap = IS_ANYOF_TRIE(op) ? (U8*)ANYOF_BITMAP(o) : (U8*)TRIE_BITMAP(trie);
14736 sv_catpvs(sv, "[");
14737 for (i = 0; i <= 256; i++) {
14738 if (i < 256 && BITMAP_TEST(bitmap,i)) {
14739 if (rangestart == -1)
14741 } else if (rangestart != -1) {
14742 if (i <= rangestart + 3)
14743 for (; rangestart < i; rangestart++)
14744 put_byte(sv, rangestart);
14746 put_byte(sv, rangestart);
14747 sv_catpvs(sv, "-");
14748 put_byte(sv, i - 1);
14753 sv_catpvs(sv, "]");
14756 } else if (k == CURLY) {
14757 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
14758 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
14759 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
14761 else if (k == WHILEM && o->flags) /* Ordinal/of */
14762 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
14763 else if (k == REF || k == OPEN || k == CLOSE || k == GROUPP || OP(o)==ACCEPT) {
14764 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
14765 if ( RXp_PAREN_NAMES(prog) ) {
14766 if ( k != REF || (OP(o) < NREF)) {
14767 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
14768 SV **name= av_fetch(list, ARG(o), 0 );
14770 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14773 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
14774 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
14775 I32 *nums=(I32*)SvPVX(sv_dat);
14776 SV **name= av_fetch(list, nums[0], 0 );
14779 for ( n=0; n<SvIVX(sv_dat); n++ ) {
14780 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
14781 (n ? "," : ""), (IV)nums[n]);
14783 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14787 } else if (k == GOSUB)
14788 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o)); /* Paren and offset */
14789 else if (k == VERB) {
14791 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
14792 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
14793 } else if (k == LOGICAL)
14794 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* 2: embedded, otherwise 1 */
14795 else if (k == ANYOF) {
14796 int i, rangestart = -1;
14797 const U8 flags = ANYOF_FLAGS(o);
14801 if (flags & ANYOF_LOCALE)
14802 sv_catpvs(sv, "{loc}");
14803 if (flags & ANYOF_LOC_FOLD)
14804 sv_catpvs(sv, "{i}");
14805 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
14806 if (flags & ANYOF_INVERT)
14807 sv_catpvs(sv, "^");
14809 /* output what the standard cp 0-255 bitmap matches */
14810 for (i = 0; i <= 256; i++) {
14811 if (i < 256 && ANYOF_BITMAP_TEST(o,i)) {
14812 if (rangestart == -1)
14814 } else if (rangestart != -1) {
14815 if (i <= rangestart + 3)
14816 for (; rangestart < i; rangestart++)
14817 put_byte(sv, rangestart);
14819 put_byte(sv, rangestart);
14820 sv_catpvs(sv, "-");
14821 put_byte(sv, i - 1);
14828 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14829 /* output any special charclass tests (used entirely under use locale) */
14830 if (ANYOF_CLASS_TEST_ANY_SET(o))
14831 for (i = 0; i < (int)(sizeof(anyofs)/sizeof(char*)); i++)
14832 if (ANYOF_CLASS_TEST(o,i)) {
14833 sv_catpv(sv, anyofs[i]);
14837 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14839 if (flags & ANYOF_NON_UTF8_LATIN1_ALL) {
14840 sv_catpvs(sv, "{non-utf8-latin1-all}");
14843 /* output information about the unicode matching */
14844 if (flags & ANYOF_UNICODE_ALL)
14845 sv_catpvs(sv, "{unicode_all}");
14846 else if (ANYOF_NONBITMAP(o))
14847 sv_catpvs(sv, "{unicode}");
14848 if (flags & ANYOF_NONBITMAP_NON_UTF8)
14849 sv_catpvs(sv, "{outside bitmap}");
14851 if (ANYOF_NONBITMAP(o)) {
14852 SV *lv; /* Set if there is something outside the bit map */
14853 SV * const sw = regclass_swash(prog, o, FALSE, &lv, NULL);
14854 bool byte_output = FALSE; /* If something in the bitmap has been
14857 if (lv && lv != &PL_sv_undef) {
14859 U8 s[UTF8_MAXBYTES_CASE+1];
14861 for (i = 0; i <= 256; i++) { /* Look at chars in bitmap */
14862 uvchr_to_utf8(s, i);
14865 && ! ANYOF_BITMAP_TEST(o, i) /* Don't duplicate
14869 && swash_fetch(sw, s, TRUE))
14871 if (rangestart == -1)
14873 } else if (rangestart != -1) {
14874 byte_output = TRUE;
14875 if (i <= rangestart + 3)
14876 for (; rangestart < i; rangestart++) {
14877 put_byte(sv, rangestart);
14880 put_byte(sv, rangestart);
14881 sv_catpvs(sv, "-");
14890 char *s = savesvpv(lv);
14891 char * const origs = s;
14893 while (*s && *s != '\n')
14897 const char * const t = ++s;
14900 sv_catpvs(sv, " ");
14906 /* Truncate very long output */
14907 if (s - origs > 256) {
14908 Perl_sv_catpvf(aTHX_ sv,
14910 (int) (s - origs - 1),
14916 else if (*s == '\t') {
14931 SvREFCNT_dec_NN(lv);
14935 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
14937 else if (k == POSIXD || k == NPOSIXD) {
14938 U8 index = FLAGS(o) * 2;
14939 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
14940 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
14943 sv_catpv(sv, anyofs[index]);
14946 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
14947 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
14949 PERL_UNUSED_CONTEXT;
14950 PERL_UNUSED_ARG(sv);
14951 PERL_UNUSED_ARG(o);
14952 PERL_UNUSED_ARG(prog);
14953 #endif /* DEBUGGING */
14957 Perl_re_intuit_string(pTHX_ REGEXP * const r)
14958 { /* Assume that RE_INTUIT is set */
14960 struct regexp *const prog = ReANY(r);
14961 GET_RE_DEBUG_FLAGS_DECL;
14963 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
14964 PERL_UNUSED_CONTEXT;
14968 const char * const s = SvPV_nolen_const(prog->check_substr
14969 ? prog->check_substr : prog->check_utf8);
14971 if (!PL_colorset) reginitcolors();
14972 PerlIO_printf(Perl_debug_log,
14973 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
14975 prog->check_substr ? "" : "utf8 ",
14976 PL_colors[5],PL_colors[0],
14979 (strlen(s) > 60 ? "..." : ""));
14982 return prog->check_substr ? prog->check_substr : prog->check_utf8;
14988 handles refcounting and freeing the perl core regexp structure. When
14989 it is necessary to actually free the structure the first thing it
14990 does is call the 'free' method of the regexp_engine associated to
14991 the regexp, allowing the handling of the void *pprivate; member
14992 first. (This routine is not overridable by extensions, which is why
14993 the extensions free is called first.)
14995 See regdupe and regdupe_internal if you change anything here.
14997 #ifndef PERL_IN_XSUB_RE
14999 Perl_pregfree(pTHX_ REGEXP *r)
15005 Perl_pregfree2(pTHX_ REGEXP *rx)
15008 struct regexp *const r = ReANY(rx);
15009 GET_RE_DEBUG_FLAGS_DECL;
15011 PERL_ARGS_ASSERT_PREGFREE2;
15013 if (r->mother_re) {
15014 ReREFCNT_dec(r->mother_re);
15016 CALLREGFREE_PVT(rx); /* free the private data */
15017 SvREFCNT_dec(RXp_PAREN_NAMES(r));
15018 Safefree(r->xpv_len_u.xpvlenu_pv);
15021 SvREFCNT_dec(r->anchored_substr);
15022 SvREFCNT_dec(r->anchored_utf8);
15023 SvREFCNT_dec(r->float_substr);
15024 SvREFCNT_dec(r->float_utf8);
15025 Safefree(r->substrs);
15027 RX_MATCH_COPY_FREE(rx);
15028 #ifdef PERL_ANY_COW
15029 SvREFCNT_dec(r->saved_copy);
15032 SvREFCNT_dec(r->qr_anoncv);
15033 rx->sv_u.svu_rx = 0;
15038 This is a hacky workaround to the structural issue of match results
15039 being stored in the regexp structure which is in turn stored in
15040 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
15041 could be PL_curpm in multiple contexts, and could require multiple
15042 result sets being associated with the pattern simultaneously, such
15043 as when doing a recursive match with (??{$qr})
15045 The solution is to make a lightweight copy of the regexp structure
15046 when a qr// is returned from the code executed by (??{$qr}) this
15047 lightweight copy doesn't actually own any of its data except for
15048 the starp/end and the actual regexp structure itself.
15054 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
15056 struct regexp *ret;
15057 struct regexp *const r = ReANY(rx);
15058 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
15060 PERL_ARGS_ASSERT_REG_TEMP_COPY;
15063 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
15065 SvOK_off((SV *)ret_x);
15067 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
15068 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
15069 made both spots point to the same regexp body.) */
15070 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
15071 assert(!SvPVX(ret_x));
15072 ret_x->sv_u.svu_rx = temp->sv_any;
15073 temp->sv_any = NULL;
15074 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
15075 SvREFCNT_dec_NN(temp);
15076 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
15077 ing below will not set it. */
15078 SvCUR_set(ret_x, SvCUR(rx));
15081 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
15082 sv_force_normal(sv) is called. */
15084 ret = ReANY(ret_x);
15086 SvFLAGS(ret_x) |= SvUTF8(rx);
15087 /* We share the same string buffer as the original regexp, on which we
15088 hold a reference count, incremented when mother_re is set below.
15089 The string pointer is copied here, being part of the regexp struct.
15091 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
15092 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
15094 const I32 npar = r->nparens+1;
15095 Newx(ret->offs, npar, regexp_paren_pair);
15096 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15099 Newx(ret->substrs, 1, struct reg_substr_data);
15100 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15102 SvREFCNT_inc_void(ret->anchored_substr);
15103 SvREFCNT_inc_void(ret->anchored_utf8);
15104 SvREFCNT_inc_void(ret->float_substr);
15105 SvREFCNT_inc_void(ret->float_utf8);
15107 /* check_substr and check_utf8, if non-NULL, point to either their
15108 anchored or float namesakes, and don't hold a second reference. */
15110 RX_MATCH_COPIED_off(ret_x);
15111 #ifdef PERL_ANY_COW
15112 ret->saved_copy = NULL;
15114 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
15115 SvREFCNT_inc_void(ret->qr_anoncv);
15121 /* regfree_internal()
15123 Free the private data in a regexp. This is overloadable by
15124 extensions. Perl takes care of the regexp structure in pregfree(),
15125 this covers the *pprivate pointer which technically perl doesn't
15126 know about, however of course we have to handle the
15127 regexp_internal structure when no extension is in use.
15129 Note this is called before freeing anything in the regexp
15134 Perl_regfree_internal(pTHX_ REGEXP * const rx)
15137 struct regexp *const r = ReANY(rx);
15138 RXi_GET_DECL(r,ri);
15139 GET_RE_DEBUG_FLAGS_DECL;
15141 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
15147 SV *dsv= sv_newmortal();
15148 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
15149 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
15150 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
15151 PL_colors[4],PL_colors[5],s);
15154 #ifdef RE_TRACK_PATTERN_OFFSETS
15156 Safefree(ri->u.offsets); /* 20010421 MJD */
15158 if (ri->code_blocks) {
15160 for (n = 0; n < ri->num_code_blocks; n++)
15161 SvREFCNT_dec(ri->code_blocks[n].src_regex);
15162 Safefree(ri->code_blocks);
15166 int n = ri->data->count;
15169 /* If you add a ->what type here, update the comment in regcomp.h */
15170 switch (ri->data->what[n]) {
15176 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
15179 Safefree(ri->data->data[n]);
15185 { /* Aho Corasick add-on structure for a trie node.
15186 Used in stclass optimization only */
15188 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
15190 refcount = --aho->refcount;
15193 PerlMemShared_free(aho->states);
15194 PerlMemShared_free(aho->fail);
15195 /* do this last!!!! */
15196 PerlMemShared_free(ri->data->data[n]);
15197 PerlMemShared_free(ri->regstclass);
15203 /* trie structure. */
15205 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
15207 refcount = --trie->refcount;
15210 PerlMemShared_free(trie->charmap);
15211 PerlMemShared_free(trie->states);
15212 PerlMemShared_free(trie->trans);
15214 PerlMemShared_free(trie->bitmap);
15216 PerlMemShared_free(trie->jump);
15217 PerlMemShared_free(trie->wordinfo);
15218 /* do this last!!!! */
15219 PerlMemShared_free(ri->data->data[n]);
15224 Perl_croak(aTHX_ "panic: regfree data code '%c'", ri->data->what[n]);
15227 Safefree(ri->data->what);
15228 Safefree(ri->data);
15234 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
15235 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
15236 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
15239 re_dup - duplicate a regexp.
15241 This routine is expected to clone a given regexp structure. It is only
15242 compiled under USE_ITHREADS.
15244 After all of the core data stored in struct regexp is duplicated
15245 the regexp_engine.dupe method is used to copy any private data
15246 stored in the *pprivate pointer. This allows extensions to handle
15247 any duplication it needs to do.
15249 See pregfree() and regfree_internal() if you change anything here.
15251 #if defined(USE_ITHREADS)
15252 #ifndef PERL_IN_XSUB_RE
15254 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
15258 const struct regexp *r = ReANY(sstr);
15259 struct regexp *ret = ReANY(dstr);
15261 PERL_ARGS_ASSERT_RE_DUP_GUTS;
15263 npar = r->nparens+1;
15264 Newx(ret->offs, npar, regexp_paren_pair);
15265 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15267 if (ret->substrs) {
15268 /* Do it this way to avoid reading from *r after the StructCopy().
15269 That way, if any of the sv_dup_inc()s dislodge *r from the L1
15270 cache, it doesn't matter. */
15271 const bool anchored = r->check_substr
15272 ? r->check_substr == r->anchored_substr
15273 : r->check_utf8 == r->anchored_utf8;
15274 Newx(ret->substrs, 1, struct reg_substr_data);
15275 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15277 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
15278 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
15279 ret->float_substr = sv_dup_inc(ret->float_substr, param);
15280 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
15282 /* check_substr and check_utf8, if non-NULL, point to either their
15283 anchored or float namesakes, and don't hold a second reference. */
15285 if (ret->check_substr) {
15287 assert(r->check_utf8 == r->anchored_utf8);
15288 ret->check_substr = ret->anchored_substr;
15289 ret->check_utf8 = ret->anchored_utf8;
15291 assert(r->check_substr == r->float_substr);
15292 assert(r->check_utf8 == r->float_utf8);
15293 ret->check_substr = ret->float_substr;
15294 ret->check_utf8 = ret->float_utf8;
15296 } else if (ret->check_utf8) {
15298 ret->check_utf8 = ret->anchored_utf8;
15300 ret->check_utf8 = ret->float_utf8;
15305 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
15306 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
15309 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
15311 if (RX_MATCH_COPIED(dstr))
15312 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
15314 ret->subbeg = NULL;
15315 #ifdef PERL_ANY_COW
15316 ret->saved_copy = NULL;
15319 /* Whether mother_re be set or no, we need to copy the string. We
15320 cannot refrain from copying it when the storage points directly to
15321 our mother regexp, because that's
15322 1: a buffer in a different thread
15323 2: something we no longer hold a reference on
15324 so we need to copy it locally. */
15325 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
15326 ret->mother_re = NULL;
15329 #endif /* PERL_IN_XSUB_RE */
15334 This is the internal complement to regdupe() which is used to copy
15335 the structure pointed to by the *pprivate pointer in the regexp.
15336 This is the core version of the extension overridable cloning hook.
15337 The regexp structure being duplicated will be copied by perl prior
15338 to this and will be provided as the regexp *r argument, however
15339 with the /old/ structures pprivate pointer value. Thus this routine
15340 may override any copying normally done by perl.
15342 It returns a pointer to the new regexp_internal structure.
15346 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
15349 struct regexp *const r = ReANY(rx);
15350 regexp_internal *reti;
15352 RXi_GET_DECL(r,ri);
15354 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
15358 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode), char, regexp_internal);
15359 Copy(ri->program, reti->program, len+1, regnode);
15361 reti->num_code_blocks = ri->num_code_blocks;
15362 if (ri->code_blocks) {
15364 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
15365 struct reg_code_block);
15366 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
15367 struct reg_code_block);
15368 for (n = 0; n < ri->num_code_blocks; n++)
15369 reti->code_blocks[n].src_regex = (REGEXP*)
15370 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
15373 reti->code_blocks = NULL;
15375 reti->regstclass = NULL;
15378 struct reg_data *d;
15379 const int count = ri->data->count;
15382 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
15383 char, struct reg_data);
15384 Newx(d->what, count, U8);
15387 for (i = 0; i < count; i++) {
15388 d->what[i] = ri->data->what[i];
15389 switch (d->what[i]) {
15390 /* see also regcomp.h and regfree_internal() */
15391 case 'a': /* actually an AV, but the dup function is identical. */
15395 case 'u': /* actually an HV, but the dup function is identical. */
15396 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
15399 /* This is cheating. */
15400 Newx(d->data[i], 1, struct regnode_charclass_class);
15401 StructCopy(ri->data->data[i], d->data[i],
15402 struct regnode_charclass_class);
15403 reti->regstclass = (regnode*)d->data[i];
15406 /* Trie stclasses are readonly and can thus be shared
15407 * without duplication. We free the stclass in pregfree
15408 * when the corresponding reg_ac_data struct is freed.
15410 reti->regstclass= ri->regstclass;
15414 ((reg_trie_data*)ri->data->data[i])->refcount++;
15419 d->data[i] = ri->data->data[i];
15422 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'", ri->data->what[i]);
15431 reti->name_list_idx = ri->name_list_idx;
15433 #ifdef RE_TRACK_PATTERN_OFFSETS
15434 if (ri->u.offsets) {
15435 Newx(reti->u.offsets, 2*len+1, U32);
15436 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
15439 SetProgLen(reti,len);
15442 return (void*)reti;
15445 #endif /* USE_ITHREADS */
15447 #ifndef PERL_IN_XSUB_RE
15450 - regnext - dig the "next" pointer out of a node
15453 Perl_regnext(pTHX_ regnode *p)
15461 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
15462 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(p), (int)REGNODE_MAX);
15465 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
15474 S_re_croak2(pTHX_ const char* pat1,const char* pat2,...)
15477 STRLEN l1 = strlen(pat1);
15478 STRLEN l2 = strlen(pat2);
15481 const char *message;
15483 PERL_ARGS_ASSERT_RE_CROAK2;
15489 Copy(pat1, buf, l1 , char);
15490 Copy(pat2, buf + l1, l2 , char);
15491 buf[l1 + l2] = '\n';
15492 buf[l1 + l2 + 1] = '\0';
15494 /* ANSI variant takes additional second argument */
15495 va_start(args, pat2);
15499 msv = vmess(buf, &args);
15501 message = SvPV_const(msv,l1);
15504 Copy(message, buf, l1 , char);
15505 buf[l1-1] = '\0'; /* Overwrite \n */
15506 Perl_croak(aTHX_ "%s", buf);
15509 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
15511 #ifndef PERL_IN_XSUB_RE
15513 Perl_save_re_context(pTHX)
15517 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
15519 const REGEXP * const rx = PM_GETRE(PL_curpm);
15522 for (i = 1; i <= RX_NPARENS(rx); i++) {
15523 char digits[TYPE_CHARS(long)];
15524 const STRLEN len = my_snprintf(digits, sizeof(digits), "%lu", (long)i);
15525 GV *const *const gvp
15526 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
15529 GV * const gv = *gvp;
15530 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
15542 S_put_byte(pTHX_ SV *sv, int c)
15544 PERL_ARGS_ASSERT_PUT_BYTE;
15546 /* Our definition of isPRINT() ignores locales, so only bytes that are
15547 not part of UTF-8 are considered printable. I assume that the same
15548 holds for UTF-EBCDIC.
15549 Also, code point 255 is not printable in either (it's E0 in EBCDIC,
15550 which Wikipedia says:
15552 EO, or Eight Ones, is an 8-bit EBCDIC character code represented as all
15553 ones (binary 1111 1111, hexadecimal FF). It is similar, but not
15554 identical, to the ASCII delete (DEL) or rubout control character. ...
15555 it is typically mapped to hexadecimal code 9F, in order to provide a
15556 unique character mapping in both directions)
15558 So the old condition can be simplified to !isPRINT(c) */
15561 Perl_sv_catpvf(aTHX_ sv, "\\x%02x", c);
15564 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
15568 const char string = c;
15569 if (c == '-' || c == ']' || c == '\\' || c == '^')
15570 sv_catpvs(sv, "\\");
15571 sv_catpvn(sv, &string, 1);
15576 #define CLEAR_OPTSTART \
15577 if (optstart) STMT_START { \
15578 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
15582 #define DUMPUNTIL(b,e) CLEAR_OPTSTART; node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
15584 STATIC const regnode *
15585 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
15586 const regnode *last, const regnode *plast,
15587 SV* sv, I32 indent, U32 depth)
15590 U8 op = PSEUDO; /* Arbitrary non-END op. */
15591 const regnode *next;
15592 const regnode *optstart= NULL;
15594 RXi_GET_DECL(r,ri);
15595 GET_RE_DEBUG_FLAGS_DECL;
15597 PERL_ARGS_ASSERT_DUMPUNTIL;
15599 #ifdef DEBUG_DUMPUNTIL
15600 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
15601 last ? last-start : 0,plast ? plast-start : 0);
15604 if (plast && plast < last)
15607 while (PL_regkind[op] != END && (!last || node < last)) {
15608 /* While that wasn't END last time... */
15611 if (op == CLOSE || op == WHILEM)
15613 next = regnext((regnode *)node);
15616 if (OP(node) == OPTIMIZED) {
15617 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
15624 regprop(r, sv, node);
15625 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
15626 (int)(2*indent + 1), "", SvPVX_const(sv));
15628 if (OP(node) != OPTIMIZED) {
15629 if (next == NULL) /* Next ptr. */
15630 PerlIO_printf(Perl_debug_log, " (0)");
15631 else if (PL_regkind[(U8)op] == BRANCH && PL_regkind[OP(next)] != BRANCH )
15632 PerlIO_printf(Perl_debug_log, " (FAIL)");
15634 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
15635 (void)PerlIO_putc(Perl_debug_log, '\n');
15639 if (PL_regkind[(U8)op] == BRANCHJ) {
15642 const regnode *nnode = (OP(next) == LONGJMP
15643 ? regnext((regnode *)next)
15645 if (last && nnode > last)
15647 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
15650 else if (PL_regkind[(U8)op] == BRANCH) {
15652 DUMPUNTIL(NEXTOPER(node), next);
15654 else if ( PL_regkind[(U8)op] == TRIE ) {
15655 const regnode *this_trie = node;
15656 const char op = OP(node);
15657 const U32 n = ARG(node);
15658 const reg_ac_data * const ac = op>=AHOCORASICK ?
15659 (reg_ac_data *)ri->data->data[n] :
15661 const reg_trie_data * const trie =
15662 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
15664 AV *const trie_words = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
15666 const regnode *nextbranch= NULL;
15669 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
15670 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
15672 PerlIO_printf(Perl_debug_log, "%*s%s ",
15673 (int)(2*(indent+3)), "",
15674 elem_ptr ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr), SvCUR(*elem_ptr), 60,
15675 PL_colors[0], PL_colors[1],
15676 (SvUTF8(*elem_ptr) ? PERL_PV_ESCAPE_UNI : 0) |
15677 PERL_PV_PRETTY_ELLIPSES |
15678 PERL_PV_PRETTY_LTGT
15683 U16 dist= trie->jump[word_idx+1];
15684 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
15685 (UV)((dist ? this_trie + dist : next) - start));
15688 nextbranch= this_trie + trie->jump[0];
15689 DUMPUNTIL(this_trie + dist, nextbranch);
15691 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
15692 nextbranch= regnext((regnode *)nextbranch);
15694 PerlIO_printf(Perl_debug_log, "\n");
15697 if (last && next > last)
15702 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
15703 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
15704 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
15706 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
15708 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
15710 else if ( op == PLUS || op == STAR) {
15711 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
15713 else if (PL_regkind[(U8)op] == ANYOF) {
15714 /* arglen 1 + class block */
15715 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_CLASS)
15716 ? ANYOF_CLASS_SKIP : ANYOF_SKIP);
15717 node = NEXTOPER(node);
15719 else if (PL_regkind[(U8)op] == EXACT) {
15720 /* Literal string, where present. */
15721 node += NODE_SZ_STR(node) - 1;
15722 node = NEXTOPER(node);
15725 node = NEXTOPER(node);
15726 node += regarglen[(U8)op];
15728 if (op == CURLYX || op == OPEN)
15732 #ifdef DEBUG_DUMPUNTIL
15733 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
15738 #endif /* DEBUGGING */
15742 * c-indentation-style: bsd
15743 * c-basic-offset: 4
15744 * indent-tabs-mode: nil
15747 * ex: set ts=8 sts=4 sw=4 et: