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 ckWARN2reg_d(loc,m, a1) STMT_START { \
578 const IV offset = loc - RExC_precomp; \
579 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
581 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
584 #define ckWARN2reg(loc, m, a1) STMT_START { \
585 const IV offset = loc - RExC_precomp; \
586 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
587 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
590 #define vWARN3(loc, m, a1, a2) STMT_START { \
591 const IV offset = loc - RExC_precomp; \
592 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
593 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
596 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
597 const IV offset = loc - RExC_precomp; \
598 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
599 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
602 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
603 const IV offset = loc - RExC_precomp; \
604 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
605 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
608 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
609 const IV offset = loc - RExC_precomp; \
610 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
611 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
614 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
615 const IV offset = loc - RExC_precomp; \
616 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
617 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
621 /* Allow for side effects in s */
622 #define REGC(c,s) STMT_START { \
623 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
626 /* Macros for recording node offsets. 20001227 mjd@plover.com
627 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
628 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
629 * Element 0 holds the number n.
630 * Position is 1 indexed.
632 #ifndef RE_TRACK_PATTERN_OFFSETS
633 #define Set_Node_Offset_To_R(node,byte)
634 #define Set_Node_Offset(node,byte)
635 #define Set_Cur_Node_Offset
636 #define Set_Node_Length_To_R(node,len)
637 #define Set_Node_Length(node,len)
638 #define Set_Node_Cur_Length(node,start)
639 #define Node_Offset(n)
640 #define Node_Length(n)
641 #define Set_Node_Offset_Length(node,offset,len)
642 #define ProgLen(ri) ri->u.proglen
643 #define SetProgLen(ri,x) ri->u.proglen = x
645 #define ProgLen(ri) ri->u.offsets[0]
646 #define SetProgLen(ri,x) ri->u.offsets[0] = x
647 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
649 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
650 __LINE__, (int)(node), (int)(byte))); \
652 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
654 RExC_offsets[2*(node)-1] = (byte); \
659 #define Set_Node_Offset(node,byte) \
660 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
661 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
663 #define Set_Node_Length_To_R(node,len) STMT_START { \
665 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
666 __LINE__, (int)(node), (int)(len))); \
668 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
670 RExC_offsets[2*(node)] = (len); \
675 #define Set_Node_Length(node,len) \
676 Set_Node_Length_To_R((node)-RExC_emit_start, len)
677 #define Set_Node_Cur_Length(node, start) \
678 Set_Node_Length(node, RExC_parse - start)
680 /* Get offsets and lengths */
681 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
682 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
684 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
685 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
686 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
690 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
691 #define EXPERIMENTAL_INPLACESCAN
692 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
694 #define DEBUG_STUDYDATA(str,data,depth) \
695 DEBUG_OPTIMISE_MORE_r(if(data){ \
696 PerlIO_printf(Perl_debug_log, \
697 "%*s" str "Pos:%"IVdf"/%"IVdf \
698 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
699 (int)(depth)*2, "", \
700 (IV)((data)->pos_min), \
701 (IV)((data)->pos_delta), \
702 (UV)((data)->flags), \
703 (IV)((data)->whilem_c), \
704 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
705 is_inf ? "INF " : "" \
707 if ((data)->last_found) \
708 PerlIO_printf(Perl_debug_log, \
709 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
710 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
711 SvPVX_const((data)->last_found), \
712 (IV)((data)->last_end), \
713 (IV)((data)->last_start_min), \
714 (IV)((data)->last_start_max), \
715 ((data)->longest && \
716 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
717 SvPVX_const((data)->longest_fixed), \
718 (IV)((data)->offset_fixed), \
719 ((data)->longest && \
720 (data)->longest==&((data)->longest_float)) ? "*" : "", \
721 SvPVX_const((data)->longest_float), \
722 (IV)((data)->offset_float_min), \
723 (IV)((data)->offset_float_max) \
725 PerlIO_printf(Perl_debug_log,"\n"); \
728 /* Mark that we cannot extend a found fixed substring at this point.
729 Update the longest found anchored substring and the longest found
730 floating substrings if needed. */
733 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
735 const STRLEN l = CHR_SVLEN(data->last_found);
736 const STRLEN old_l = CHR_SVLEN(*data->longest);
737 GET_RE_DEBUG_FLAGS_DECL;
739 PERL_ARGS_ASSERT_SCAN_COMMIT;
741 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
742 SvSetMagicSV(*data->longest, data->last_found);
743 if (*data->longest == data->longest_fixed) {
744 data->offset_fixed = l ? data->last_start_min : data->pos_min;
745 if (data->flags & SF_BEFORE_EOL)
747 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
749 data->flags &= ~SF_FIX_BEFORE_EOL;
750 data->minlen_fixed=minlenp;
751 data->lookbehind_fixed=0;
753 else { /* *data->longest == data->longest_float */
754 data->offset_float_min = l ? data->last_start_min : data->pos_min;
755 data->offset_float_max = (l
756 ? data->last_start_max
757 : (data->pos_delta == I32_MAX ? I32_MAX : data->pos_min + data->pos_delta));
758 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
759 data->offset_float_max = I32_MAX;
760 if (data->flags & SF_BEFORE_EOL)
762 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
764 data->flags &= ~SF_FL_BEFORE_EOL;
765 data->minlen_float=minlenp;
766 data->lookbehind_float=0;
769 SvCUR_set(data->last_found, 0);
771 SV * const sv = data->last_found;
772 if (SvUTF8(sv) && SvMAGICAL(sv)) {
773 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
779 data->flags &= ~SF_BEFORE_EOL;
780 DEBUG_STUDYDATA("commit: ",data,0);
783 /* These macros set, clear and test whether the synthetic start class ('ssc',
784 * given by the parameter) matches an empty string (EOS). This uses the
785 * 'next_off' field in the node, to save a bit in the flags field. The ssc
786 * stands alone, so there is never a next_off, so this field is otherwise
787 * unused. The EOS information is used only for compilation, but theoretically
788 * it could be passed on to the execution code. This could be used to store
789 * more than one bit of information, but only this one is currently used. */
790 #define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
791 #define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
792 #define TEST_SSC_EOS(node) cBOOL((node)->next_off)
794 /* Can match anything (initialization) */
796 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
798 PERL_ARGS_ASSERT_CL_ANYTHING;
800 ANYOF_BITMAP_SETALL(cl);
801 cl->flags = ANYOF_UNICODE_ALL;
804 /* If any portion of the regex is to operate under locale rules,
805 * initialization includes it. The reason this isn't done for all regexes
806 * is that the optimizer was written under the assumption that locale was
807 * all-or-nothing. Given the complexity and lack of documentation in the
808 * optimizer, and that there are inadequate test cases for locale, so many
809 * parts of it may not work properly, it is safest to avoid locale unless
811 if (RExC_contains_locale) {
812 ANYOF_CLASS_SETALL(cl); /* /l uses class */
813 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
816 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
820 /* Can match anything (initialization) */
822 S_cl_is_anything(const struct regnode_charclass_class *cl)
826 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
828 for (value = 0; value < ANYOF_MAX; value += 2)
829 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
831 if (!(cl->flags & ANYOF_UNICODE_ALL))
833 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
838 /* Can match anything (initialization) */
840 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
842 PERL_ARGS_ASSERT_CL_INIT;
844 Zero(cl, 1, struct regnode_charclass_class);
846 cl_anything(pRExC_state, cl);
847 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
850 /* These two functions currently do the exact same thing */
851 #define cl_init_zero S_cl_init
853 /* 'AND' a given class with another one. Can create false positives. 'cl'
854 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
855 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
857 S_cl_and(struct regnode_charclass_class *cl,
858 const struct regnode_charclass_class *and_with)
860 PERL_ARGS_ASSERT_CL_AND;
862 assert(PL_regkind[and_with->type] == ANYOF);
864 /* I (khw) am not sure all these restrictions are necessary XXX */
865 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
866 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
867 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
868 && !(and_with->flags & ANYOF_LOC_FOLD)
869 && !(cl->flags & ANYOF_LOC_FOLD)) {
872 if (and_with->flags & ANYOF_INVERT)
873 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
874 cl->bitmap[i] &= ~and_with->bitmap[i];
876 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
877 cl->bitmap[i] &= and_with->bitmap[i];
878 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
880 if (and_with->flags & ANYOF_INVERT) {
882 /* Here, the and'ed node is inverted. Get the AND of the flags that
883 * aren't affected by the inversion. Those that are affected are
884 * handled individually below */
885 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
886 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
887 cl->flags |= affected_flags;
889 /* We currently don't know how to deal with things that aren't in the
890 * bitmap, but we know that the intersection is no greater than what
891 * is already in cl, so let there be false positives that get sorted
892 * out after the synthetic start class succeeds, and the node is
893 * matched for real. */
895 /* The inversion of these two flags indicate that the resulting
896 * intersection doesn't have them */
897 if (and_with->flags & ANYOF_UNICODE_ALL) {
898 cl->flags &= ~ANYOF_UNICODE_ALL;
900 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
901 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
904 else { /* and'd node is not inverted */
905 U8 outside_bitmap_but_not_utf8; /* Temp variable */
907 if (! ANYOF_NONBITMAP(and_with)) {
909 /* Here 'and_with' doesn't match anything outside the bitmap
910 * (except possibly ANYOF_UNICODE_ALL), which means the
911 * intersection can't either, except for ANYOF_UNICODE_ALL, in
912 * which case we don't know what the intersection is, but it's no
913 * greater than what cl already has, so can just leave it alone,
914 * with possible false positives */
915 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
916 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
917 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
920 else if (! ANYOF_NONBITMAP(cl)) {
922 /* Here, 'and_with' does match something outside the bitmap, and cl
923 * doesn't have a list of things to match outside the bitmap. If
924 * cl can match all code points above 255, the intersection will
925 * be those above-255 code points that 'and_with' matches. If cl
926 * can't match all Unicode code points, it means that it can't
927 * match anything outside the bitmap (since the 'if' that got us
928 * into this block tested for that), so we leave the bitmap empty.
930 if (cl->flags & ANYOF_UNICODE_ALL) {
931 ARG_SET(cl, ARG(and_with));
933 /* and_with's ARG may match things that don't require UTF8.
934 * And now cl's will too, in spite of this being an 'and'. See
935 * the comments below about the kludge */
936 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
940 /* Here, both 'and_with' and cl match something outside the
941 * bitmap. Currently we do not do the intersection, so just match
942 * whatever cl had at the beginning. */
946 /* Take the intersection of the two sets of flags. However, the
947 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
948 * kludge around the fact that this flag is not treated like the others
949 * which are initialized in cl_anything(). The way the optimizer works
950 * is that the synthetic start class (SSC) is initialized to match
951 * anything, and then the first time a real node is encountered, its
952 * values are AND'd with the SSC's with the result being the values of
953 * the real node. However, there are paths through the optimizer where
954 * the AND never gets called, so those initialized bits are set
955 * inappropriately, which is not usually a big deal, as they just cause
956 * false positives in the SSC, which will just mean a probably
957 * imperceptible slow down in execution. However this bit has a
958 * higher false positive consequence in that it can cause utf8.pm,
959 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
960 * bigger slowdown and also causes significant extra memory to be used.
961 * In order to prevent this, the code now takes a different tack. The
962 * bit isn't set unless some part of the regular expression needs it,
963 * but once set it won't get cleared. This means that these extra
964 * modules won't get loaded unless there was some path through the
965 * pattern that would have required them anyway, and so any false
966 * positives that occur by not ANDing them out when they could be
967 * aren't as severe as they would be if we treated this bit like all
969 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
970 & ANYOF_NONBITMAP_NON_UTF8;
971 cl->flags &= and_with->flags;
972 cl->flags |= outside_bitmap_but_not_utf8;
976 /* 'OR' a given class with another one. Can create false positives. 'cl'
977 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
978 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
980 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
982 PERL_ARGS_ASSERT_CL_OR;
984 if (or_with->flags & ANYOF_INVERT) {
986 /* Here, the or'd node is to be inverted. This means we take the
987 * complement of everything not in the bitmap, but currently we don't
988 * know what that is, so give up and match anything */
989 if (ANYOF_NONBITMAP(or_with)) {
990 cl_anything(pRExC_state, cl);
993 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
994 * <= (B1 | !B2) | (CL1 | !CL2)
995 * which is wasteful if CL2 is small, but we ignore CL2:
996 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
997 * XXXX Can we handle case-fold? Unclear:
998 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
999 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
1001 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1002 && !(or_with->flags & ANYOF_LOC_FOLD)
1003 && !(cl->flags & ANYOF_LOC_FOLD) ) {
1006 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1007 cl->bitmap[i] |= ~or_with->bitmap[i];
1008 } /* XXXX: logic is complicated otherwise */
1010 cl_anything(pRExC_state, cl);
1013 /* And, we can just take the union of the flags that aren't affected
1014 * by the inversion */
1015 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1017 /* For the remaining flags:
1018 ANYOF_UNICODE_ALL and inverted means to not match anything above
1019 255, which means that the union with cl should just be
1020 what cl has in it, so can ignore this flag
1021 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1022 is 127-255 to match them, but then invert that, so the
1023 union with cl should just be what cl has in it, so can
1026 } else { /* 'or_with' is not inverted */
1027 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1028 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1029 && (!(or_with->flags & ANYOF_LOC_FOLD)
1030 || (cl->flags & ANYOF_LOC_FOLD)) ) {
1033 /* OR char bitmap and class bitmap separately */
1034 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1035 cl->bitmap[i] |= or_with->bitmap[i];
1036 if (or_with->flags & ANYOF_CLASS) {
1037 ANYOF_CLASS_OR(or_with, cl);
1040 else { /* XXXX: logic is complicated, leave it along for a moment. */
1041 cl_anything(pRExC_state, cl);
1044 if (ANYOF_NONBITMAP(or_with)) {
1046 /* Use the added node's outside-the-bit-map match if there isn't a
1047 * conflict. If there is a conflict (both nodes match something
1048 * outside the bitmap, but what they match outside is not the same
1049 * pointer, and hence not easily compared until XXX we extend
1050 * inversion lists this far), give up and allow the start class to
1051 * match everything outside the bitmap. If that stuff is all above
1052 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1053 if (! ANYOF_NONBITMAP(cl)) {
1054 ARG_SET(cl, ARG(or_with));
1056 else if (ARG(cl) != ARG(or_with)) {
1058 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1059 cl_anything(pRExC_state, cl);
1062 cl->flags |= ANYOF_UNICODE_ALL;
1067 /* Take the union */
1068 cl->flags |= or_with->flags;
1072 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1073 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1074 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1075 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1080 dump_trie(trie,widecharmap,revcharmap)
1081 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1082 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1084 These routines dump out a trie in a somewhat readable format.
1085 The _interim_ variants are used for debugging the interim
1086 tables that are used to generate the final compressed
1087 representation which is what dump_trie expects.
1089 Part of the reason for their existence is to provide a form
1090 of documentation as to how the different representations function.
1095 Dumps the final compressed table form of the trie to Perl_debug_log.
1096 Used for debugging make_trie().
1100 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1101 AV *revcharmap, U32 depth)
1104 SV *sv=sv_newmortal();
1105 int colwidth= widecharmap ? 6 : 4;
1107 GET_RE_DEBUG_FLAGS_DECL;
1109 PERL_ARGS_ASSERT_DUMP_TRIE;
1111 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1112 (int)depth * 2 + 2,"",
1113 "Match","Base","Ofs" );
1115 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1116 SV ** const tmp = av_fetch( revcharmap, state, 0);
1118 PerlIO_printf( Perl_debug_log, "%*s",
1120 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1121 PL_colors[0], PL_colors[1],
1122 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1123 PERL_PV_ESCAPE_FIRSTCHAR
1128 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1129 (int)depth * 2 + 2,"");
1131 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1132 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1133 PerlIO_printf( Perl_debug_log, "\n");
1135 for( state = 1 ; state < trie->statecount ; state++ ) {
1136 const U32 base = trie->states[ state ].trans.base;
1138 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1140 if ( trie->states[ state ].wordnum ) {
1141 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1143 PerlIO_printf( Perl_debug_log, "%6s", "" );
1146 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1151 while( ( base + ofs < trie->uniquecharcount ) ||
1152 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1153 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1156 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1158 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1159 if ( ( base + ofs >= trie->uniquecharcount ) &&
1160 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1161 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1163 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1165 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1167 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1171 PerlIO_printf( Perl_debug_log, "]");
1174 PerlIO_printf( Perl_debug_log, "\n" );
1176 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1177 for (word=1; word <= trie->wordcount; word++) {
1178 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1179 (int)word, (int)(trie->wordinfo[word].prev),
1180 (int)(trie->wordinfo[word].len));
1182 PerlIO_printf(Perl_debug_log, "\n" );
1185 Dumps a fully constructed but uncompressed trie in list form.
1186 List tries normally only are used for construction when the number of
1187 possible chars (trie->uniquecharcount) is very high.
1188 Used for debugging make_trie().
1191 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1192 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1196 SV *sv=sv_newmortal();
1197 int colwidth= widecharmap ? 6 : 4;
1198 GET_RE_DEBUG_FLAGS_DECL;
1200 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1202 /* print out the table precompression. */
1203 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1204 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1205 "------:-----+-----------------\n" );
1207 for( state=1 ; state < next_alloc ; state ++ ) {
1210 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1211 (int)depth * 2 + 2,"", (UV)state );
1212 if ( ! trie->states[ state ].wordnum ) {
1213 PerlIO_printf( Perl_debug_log, "%5s| ","");
1215 PerlIO_printf( Perl_debug_log, "W%4x| ",
1216 trie->states[ state ].wordnum
1219 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1220 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1222 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1224 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1225 PL_colors[0], PL_colors[1],
1226 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1227 PERL_PV_ESCAPE_FIRSTCHAR
1229 TRIE_LIST_ITEM(state,charid).forid,
1230 (UV)TRIE_LIST_ITEM(state,charid).newstate
1233 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1234 (int)((depth * 2) + 14), "");
1237 PerlIO_printf( Perl_debug_log, "\n");
1242 Dumps a fully constructed but uncompressed trie in table form.
1243 This is the normal DFA style state transition table, with a few
1244 twists to facilitate compression later.
1245 Used for debugging make_trie().
1248 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1249 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1254 SV *sv=sv_newmortal();
1255 int colwidth= widecharmap ? 6 : 4;
1256 GET_RE_DEBUG_FLAGS_DECL;
1258 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1261 print out the table precompression so that we can do a visual check
1262 that they are identical.
1265 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1267 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1268 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1270 PerlIO_printf( Perl_debug_log, "%*s",
1272 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1273 PL_colors[0], PL_colors[1],
1274 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1275 PERL_PV_ESCAPE_FIRSTCHAR
1281 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1283 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1284 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1287 PerlIO_printf( Perl_debug_log, "\n" );
1289 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1291 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1292 (int)depth * 2 + 2,"",
1293 (UV)TRIE_NODENUM( state ) );
1295 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1296 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1298 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1300 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1302 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1303 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1305 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1306 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1314 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1315 startbranch: the first branch in the whole branch sequence
1316 first : start branch of sequence of branch-exact nodes.
1317 May be the same as startbranch
1318 last : Thing following the last branch.
1319 May be the same as tail.
1320 tail : item following the branch sequence
1321 count : words in the sequence
1322 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1323 depth : indent depth
1325 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1327 A trie is an N'ary tree where the branches are determined by digital
1328 decomposition of the key. IE, at the root node you look up the 1st character and
1329 follow that branch repeat until you find the end of the branches. Nodes can be
1330 marked as "accepting" meaning they represent a complete word. Eg:
1334 would convert into the following structure. Numbers represent states, letters
1335 following numbers represent valid transitions on the letter from that state, if
1336 the number is in square brackets it represents an accepting state, otherwise it
1337 will be in parenthesis.
1339 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1343 (1) +-i->(6)-+-s->[7]
1345 +-s->(3)-+-h->(4)-+-e->[5]
1347 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1349 This shows that when matching against the string 'hers' we will begin at state 1
1350 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1351 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1352 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1353 single traverse. We store a mapping from accepting to state to which word was
1354 matched, and then when we have multiple possibilities we try to complete the
1355 rest of the regex in the order in which they occured in the alternation.
1357 The only prior NFA like behaviour that would be changed by the TRIE support is
1358 the silent ignoring of duplicate alternations which are of the form:
1360 / (DUPE|DUPE) X? (?{ ... }) Y /x
1362 Thus EVAL blocks following a trie may be called a different number of times with
1363 and without the optimisation. With the optimisations dupes will be silently
1364 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1365 the following demonstrates:
1367 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1369 which prints out 'word' three times, but
1371 'words'=~/(word|word|word)(?{ print $1 })S/
1373 which doesnt print it out at all. This is due to other optimisations kicking in.
1375 Example of what happens on a structural level:
1377 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1379 1: CURLYM[1] {1,32767}(18)
1390 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1391 and should turn into:
1393 1: CURLYM[1] {1,32767}(18)
1395 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1403 Cases where tail != last would be like /(?foo|bar)baz/:
1413 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1414 and would end up looking like:
1417 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1424 d = uvuni_to_utf8_flags(d, uv, 0);
1426 is the recommended Unicode-aware way of saying
1431 #define TRIE_STORE_REVCHAR(val) \
1434 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1435 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1436 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1437 SvCUR_set(zlopp, kapow - flrbbbbb); \
1440 av_push(revcharmap, zlopp); \
1442 char ooooff = (char)val; \
1443 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1447 #define TRIE_READ_CHAR STMT_START { \
1450 /* if it is UTF then it is either already folded, or does not need folding */ \
1451 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1453 else if (folder == PL_fold_latin1) { \
1454 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1455 if ( foldlen > 0 ) { \
1456 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1462 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, FOLD_FLAGS_FULL); \
1463 skiplen = UNISKIP(uvc); \
1464 foldlen -= skiplen; \
1465 scan = foldbuf + skiplen; \
1468 /* raw data, will be folded later if needed */ \
1476 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1477 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1478 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1479 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1481 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1482 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1483 TRIE_LIST_CUR( state )++; \
1486 #define TRIE_LIST_NEW(state) STMT_START { \
1487 Newxz( trie->states[ state ].trans.list, \
1488 4, reg_trie_trans_le ); \
1489 TRIE_LIST_CUR( state ) = 1; \
1490 TRIE_LIST_LEN( state ) = 4; \
1493 #define TRIE_HANDLE_WORD(state) STMT_START { \
1494 U16 dupe= trie->states[ state ].wordnum; \
1495 regnode * const noper_next = regnext( noper ); \
1498 /* store the word for dumping */ \
1500 if (OP(noper) != NOTHING) \
1501 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1503 tmp = newSVpvn_utf8( "", 0, UTF ); \
1504 av_push( trie_words, tmp ); \
1508 trie->wordinfo[curword].prev = 0; \
1509 trie->wordinfo[curword].len = wordlen; \
1510 trie->wordinfo[curword].accept = state; \
1512 if ( noper_next < tail ) { \
1514 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1515 trie->jump[curword] = (U16)(noper_next - convert); \
1517 jumper = noper_next; \
1519 nextbranch= regnext(cur); \
1523 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1524 /* chain, so that when the bits of chain are later */\
1525 /* linked together, the dups appear in the chain */\
1526 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1527 trie->wordinfo[dupe].prev = curword; \
1529 /* we haven't inserted this word yet. */ \
1530 trie->states[ state ].wordnum = curword; \
1535 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1536 ( ( base + charid >= ucharcount \
1537 && base + charid < ubound \
1538 && state == trie->trans[ base - ucharcount + charid ].check \
1539 && trie->trans[ base - ucharcount + charid ].next ) \
1540 ? trie->trans[ base - ucharcount + charid ].next \
1541 : ( state==1 ? special : 0 ) \
1545 #define MADE_JUMP_TRIE 2
1546 #define MADE_EXACT_TRIE 4
1549 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1552 /* first pass, loop through and scan words */
1553 reg_trie_data *trie;
1554 HV *widecharmap = NULL;
1555 AV *revcharmap = newAV();
1557 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1562 regnode *jumper = NULL;
1563 regnode *nextbranch = NULL;
1564 regnode *convert = NULL;
1565 U32 *prev_states; /* temp array mapping each state to previous one */
1566 /* we just use folder as a flag in utf8 */
1567 const U8 * folder = NULL;
1570 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1571 AV *trie_words = NULL;
1572 /* along with revcharmap, this only used during construction but both are
1573 * useful during debugging so we store them in the struct when debugging.
1576 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1577 STRLEN trie_charcount=0;
1579 SV *re_trie_maxbuff;
1580 GET_RE_DEBUG_FLAGS_DECL;
1582 PERL_ARGS_ASSERT_MAKE_TRIE;
1584 PERL_UNUSED_ARG(depth);
1591 case EXACTFU_TRICKYFOLD:
1592 case EXACTFU: folder = PL_fold_latin1; break;
1593 case EXACTF: folder = PL_fold; break;
1594 case EXACTFL: folder = PL_fold_locale; break;
1595 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1598 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1600 trie->startstate = 1;
1601 trie->wordcount = word_count;
1602 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1603 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1605 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1606 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1607 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1610 trie_words = newAV();
1613 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1614 if (!SvIOK(re_trie_maxbuff)) {
1615 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1617 DEBUG_TRIE_COMPILE_r({
1618 PerlIO_printf( Perl_debug_log,
1619 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1620 (int)depth * 2 + 2, "",
1621 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1622 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1626 /* Find the node we are going to overwrite */
1627 if ( first == startbranch && OP( last ) != BRANCH ) {
1628 /* whole branch chain */
1631 /* branch sub-chain */
1632 convert = NEXTOPER( first );
1635 /* -- First loop and Setup --
1637 We first traverse the branches and scan each word to determine if it
1638 contains widechars, and how many unique chars there are, this is
1639 important as we have to build a table with at least as many columns as we
1642 We use an array of integers to represent the character codes 0..255
1643 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1644 native representation of the character value as the key and IV's for the
1647 *TODO* If we keep track of how many times each character is used we can
1648 remap the columns so that the table compression later on is more
1649 efficient in terms of memory by ensuring the most common value is in the
1650 middle and the least common are on the outside. IMO this would be better
1651 than a most to least common mapping as theres a decent chance the most
1652 common letter will share a node with the least common, meaning the node
1653 will not be compressible. With a middle is most common approach the worst
1654 case is when we have the least common nodes twice.
1658 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1659 regnode *noper = NEXTOPER( cur );
1660 const U8 *uc = (U8*)STRING( noper );
1661 const U8 *e = uc + STR_LEN( noper );
1663 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1665 const U8 *scan = (U8*)NULL;
1666 U32 wordlen = 0; /* required init */
1668 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1670 if (OP(noper) == NOTHING) {
1671 regnode *noper_next= regnext(noper);
1672 if (noper_next != tail && OP(noper_next) == flags) {
1674 uc= (U8*)STRING(noper);
1675 e= uc + STR_LEN(noper);
1676 trie->minlen= STR_LEN(noper);
1683 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1684 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1685 regardless of encoding */
1686 if (OP( noper ) == EXACTFU_SS) {
1687 /* false positives are ok, so just set this */
1688 TRIE_BITMAP_SET(trie,0xDF);
1691 for ( ; uc < e ; uc += len ) {
1692 TRIE_CHARCOUNT(trie)++;
1697 U8 folded= folder[ (U8) uvc ];
1698 if ( !trie->charmap[ folded ] ) {
1699 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1700 TRIE_STORE_REVCHAR( folded );
1703 if ( !trie->charmap[ uvc ] ) {
1704 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1705 TRIE_STORE_REVCHAR( uvc );
1708 /* store the codepoint in the bitmap, and its folded
1710 TRIE_BITMAP_SET(trie, uvc);
1712 /* store the folded codepoint */
1713 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1716 /* store first byte of utf8 representation of
1717 variant codepoints */
1718 if (! UNI_IS_INVARIANT(uvc)) {
1719 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1722 set_bit = 0; /* We've done our bit :-) */
1727 widecharmap = newHV();
1729 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1732 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1734 if ( !SvTRUE( *svpp ) ) {
1735 sv_setiv( *svpp, ++trie->uniquecharcount );
1736 TRIE_STORE_REVCHAR(uvc);
1740 if( cur == first ) {
1741 trie->minlen = chars;
1742 trie->maxlen = chars;
1743 } else if (chars < trie->minlen) {
1744 trie->minlen = chars;
1745 } else if (chars > trie->maxlen) {
1746 trie->maxlen = chars;
1748 if (OP( noper ) == EXACTFU_SS) {
1749 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1750 if (trie->minlen > 1)
1753 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1754 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1755 * - We assume that any such sequence might match a 2 byte string */
1756 if (trie->minlen > 2 )
1760 } /* end first pass */
1761 DEBUG_TRIE_COMPILE_r(
1762 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1763 (int)depth * 2 + 2,"",
1764 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1765 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1766 (int)trie->minlen, (int)trie->maxlen )
1770 We now know what we are dealing with in terms of unique chars and
1771 string sizes so we can calculate how much memory a naive
1772 representation using a flat table will take. If it's over a reasonable
1773 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1774 conservative but potentially much slower representation using an array
1777 At the end we convert both representations into the same compressed
1778 form that will be used in regexec.c for matching with. The latter
1779 is a form that cannot be used to construct with but has memory
1780 properties similar to the list form and access properties similar
1781 to the table form making it both suitable for fast searches and
1782 small enough that its feasable to store for the duration of a program.
1784 See the comment in the code where the compressed table is produced
1785 inplace from the flat tabe representation for an explanation of how
1786 the compression works.
1791 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1794 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1796 Second Pass -- Array Of Lists Representation
1798 Each state will be represented by a list of charid:state records
1799 (reg_trie_trans_le) the first such element holds the CUR and LEN
1800 points of the allocated array. (See defines above).
1802 We build the initial structure using the lists, and then convert
1803 it into the compressed table form which allows faster lookups
1804 (but cant be modified once converted).
1807 STRLEN transcount = 1;
1809 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1810 "%*sCompiling trie using list compiler\n",
1811 (int)depth * 2 + 2, ""));
1813 trie->states = (reg_trie_state *)
1814 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1815 sizeof(reg_trie_state) );
1819 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1821 regnode *noper = NEXTOPER( cur );
1822 U8 *uc = (U8*)STRING( noper );
1823 const U8 *e = uc + STR_LEN( noper );
1824 U32 state = 1; /* required init */
1825 U16 charid = 0; /* sanity init */
1826 U8 *scan = (U8*)NULL; /* sanity init */
1827 STRLEN foldlen = 0; /* required init */
1828 U32 wordlen = 0; /* required init */
1829 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1832 if (OP(noper) == NOTHING) {
1833 regnode *noper_next= regnext(noper);
1834 if (noper_next != tail && OP(noper_next) == flags) {
1836 uc= (U8*)STRING(noper);
1837 e= uc + STR_LEN(noper);
1841 if (OP(noper) != NOTHING) {
1842 for ( ; uc < e ; uc += len ) {
1847 charid = trie->charmap[ uvc ];
1849 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1853 charid=(U16)SvIV( *svpp );
1856 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1863 if ( !trie->states[ state ].trans.list ) {
1864 TRIE_LIST_NEW( state );
1866 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1867 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1868 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1873 newstate = next_alloc++;
1874 prev_states[newstate] = state;
1875 TRIE_LIST_PUSH( state, charid, newstate );
1880 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1884 TRIE_HANDLE_WORD(state);
1886 } /* end second pass */
1888 /* next alloc is the NEXT state to be allocated */
1889 trie->statecount = next_alloc;
1890 trie->states = (reg_trie_state *)
1891 PerlMemShared_realloc( trie->states,
1893 * sizeof(reg_trie_state) );
1895 /* and now dump it out before we compress it */
1896 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1897 revcharmap, next_alloc,
1901 trie->trans = (reg_trie_trans *)
1902 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1909 for( state=1 ; state < next_alloc ; state ++ ) {
1913 DEBUG_TRIE_COMPILE_MORE_r(
1914 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1918 if (trie->states[state].trans.list) {
1919 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1923 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1924 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1925 if ( forid < minid ) {
1927 } else if ( forid > maxid ) {
1931 if ( transcount < tp + maxid - minid + 1) {
1933 trie->trans = (reg_trie_trans *)
1934 PerlMemShared_realloc( trie->trans,
1936 * sizeof(reg_trie_trans) );
1937 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1939 base = trie->uniquecharcount + tp - minid;
1940 if ( maxid == minid ) {
1942 for ( ; zp < tp ; zp++ ) {
1943 if ( ! trie->trans[ zp ].next ) {
1944 base = trie->uniquecharcount + zp - minid;
1945 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1946 trie->trans[ zp ].check = state;
1952 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1953 trie->trans[ tp ].check = state;
1958 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1959 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1960 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1961 trie->trans[ tid ].check = state;
1963 tp += ( maxid - minid + 1 );
1965 Safefree(trie->states[ state ].trans.list);
1968 DEBUG_TRIE_COMPILE_MORE_r(
1969 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1972 trie->states[ state ].trans.base=base;
1974 trie->lasttrans = tp + 1;
1978 Second Pass -- Flat Table Representation.
1980 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1981 We know that we will need Charcount+1 trans at most to store the data
1982 (one row per char at worst case) So we preallocate both structures
1983 assuming worst case.
1985 We then construct the trie using only the .next slots of the entry
1988 We use the .check field of the first entry of the node temporarily to
1989 make compression both faster and easier by keeping track of how many non
1990 zero fields are in the node.
1992 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1995 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1996 number representing the first entry of the node, and state as a
1997 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1998 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1999 are 2 entrys per node. eg:
2007 The table is internally in the right hand, idx form. However as we also
2008 have to deal with the states array which is indexed by nodenum we have to
2009 use TRIE_NODENUM() to convert.
2012 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2013 "%*sCompiling trie using table compiler\n",
2014 (int)depth * 2 + 2, ""));
2016 trie->trans = (reg_trie_trans *)
2017 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2018 * trie->uniquecharcount + 1,
2019 sizeof(reg_trie_trans) );
2020 trie->states = (reg_trie_state *)
2021 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2022 sizeof(reg_trie_state) );
2023 next_alloc = trie->uniquecharcount + 1;
2026 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2028 regnode *noper = NEXTOPER( cur );
2029 const U8 *uc = (U8*)STRING( noper );
2030 const U8 *e = uc + STR_LEN( noper );
2032 U32 state = 1; /* required init */
2034 U16 charid = 0; /* sanity init */
2035 U32 accept_state = 0; /* sanity init */
2036 U8 *scan = (U8*)NULL; /* sanity init */
2038 STRLEN foldlen = 0; /* required init */
2039 U32 wordlen = 0; /* required init */
2041 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2043 if (OP(noper) == NOTHING) {
2044 regnode *noper_next= regnext(noper);
2045 if (noper_next != tail && OP(noper_next) == flags) {
2047 uc= (U8*)STRING(noper);
2048 e= uc + STR_LEN(noper);
2052 if ( OP(noper) != NOTHING ) {
2053 for ( ; uc < e ; uc += len ) {
2058 charid = trie->charmap[ uvc ];
2060 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2061 charid = svpp ? (U16)SvIV(*svpp) : 0;
2065 if ( !trie->trans[ state + charid ].next ) {
2066 trie->trans[ state + charid ].next = next_alloc;
2067 trie->trans[ state ].check++;
2068 prev_states[TRIE_NODENUM(next_alloc)]
2069 = TRIE_NODENUM(state);
2070 next_alloc += trie->uniquecharcount;
2072 state = trie->trans[ state + charid ].next;
2074 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2076 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2079 accept_state = TRIE_NODENUM( state );
2080 TRIE_HANDLE_WORD(accept_state);
2082 } /* end second pass */
2084 /* and now dump it out before we compress it */
2085 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2087 next_alloc, depth+1));
2091 * Inplace compress the table.*
2093 For sparse data sets the table constructed by the trie algorithm will
2094 be mostly 0/FAIL transitions or to put it another way mostly empty.
2095 (Note that leaf nodes will not contain any transitions.)
2097 This algorithm compresses the tables by eliminating most such
2098 transitions, at the cost of a modest bit of extra work during lookup:
2100 - Each states[] entry contains a .base field which indicates the
2101 index in the state[] array wheres its transition data is stored.
2103 - If .base is 0 there are no valid transitions from that node.
2105 - If .base is nonzero then charid is added to it to find an entry in
2108 -If trans[states[state].base+charid].check!=state then the
2109 transition is taken to be a 0/Fail transition. Thus if there are fail
2110 transitions at the front of the node then the .base offset will point
2111 somewhere inside the previous nodes data (or maybe even into a node
2112 even earlier), but the .check field determines if the transition is
2116 The following process inplace converts the table to the compressed
2117 table: We first do not compress the root node 1,and mark all its
2118 .check pointers as 1 and set its .base pointer as 1 as well. This
2119 allows us to do a DFA construction from the compressed table later,
2120 and ensures that any .base pointers we calculate later are greater
2123 - We set 'pos' to indicate the first entry of the second node.
2125 - We then iterate over the columns of the node, finding the first and
2126 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2127 and set the .check pointers accordingly, and advance pos
2128 appropriately and repreat for the next node. Note that when we copy
2129 the next pointers we have to convert them from the original
2130 NODEIDX form to NODENUM form as the former is not valid post
2133 - If a node has no transitions used we mark its base as 0 and do not
2134 advance the pos pointer.
2136 - If a node only has one transition we use a second pointer into the
2137 structure to fill in allocated fail transitions from other states.
2138 This pointer is independent of the main pointer and scans forward
2139 looking for null transitions that are allocated to a state. When it
2140 finds one it writes the single transition into the "hole". If the
2141 pointer doesnt find one the single transition is appended as normal.
2143 - Once compressed we can Renew/realloc the structures to release the
2146 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2147 specifically Fig 3.47 and the associated pseudocode.
2151 const U32 laststate = TRIE_NODENUM( next_alloc );
2154 trie->statecount = laststate;
2156 for ( state = 1 ; state < laststate ; state++ ) {
2158 const U32 stateidx = TRIE_NODEIDX( state );
2159 const U32 o_used = trie->trans[ stateidx ].check;
2160 U32 used = trie->trans[ stateidx ].check;
2161 trie->trans[ stateidx ].check = 0;
2163 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2164 if ( flag || trie->trans[ stateidx + charid ].next ) {
2165 if ( trie->trans[ stateidx + charid ].next ) {
2167 for ( ; zp < pos ; zp++ ) {
2168 if ( ! trie->trans[ zp ].next ) {
2172 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2173 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2174 trie->trans[ zp ].check = state;
2175 if ( ++zp > pos ) pos = zp;
2182 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2184 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2185 trie->trans[ pos ].check = state;
2190 trie->lasttrans = pos + 1;
2191 trie->states = (reg_trie_state *)
2192 PerlMemShared_realloc( trie->states, laststate
2193 * sizeof(reg_trie_state) );
2194 DEBUG_TRIE_COMPILE_MORE_r(
2195 PerlIO_printf( Perl_debug_log,
2196 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2197 (int)depth * 2 + 2,"",
2198 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2201 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2204 } /* end table compress */
2206 DEBUG_TRIE_COMPILE_MORE_r(
2207 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2208 (int)depth * 2 + 2, "",
2209 (UV)trie->statecount,
2210 (UV)trie->lasttrans)
2212 /* resize the trans array to remove unused space */
2213 trie->trans = (reg_trie_trans *)
2214 PerlMemShared_realloc( trie->trans, trie->lasttrans
2215 * sizeof(reg_trie_trans) );
2217 { /* Modify the program and insert the new TRIE node */
2218 U8 nodetype =(U8)(flags & 0xFF);
2222 regnode *optimize = NULL;
2223 #ifdef RE_TRACK_PATTERN_OFFSETS
2226 U32 mjd_nodelen = 0;
2227 #endif /* RE_TRACK_PATTERN_OFFSETS */
2228 #endif /* DEBUGGING */
2230 This means we convert either the first branch or the first Exact,
2231 depending on whether the thing following (in 'last') is a branch
2232 or not and whther first is the startbranch (ie is it a sub part of
2233 the alternation or is it the whole thing.)
2234 Assuming its a sub part we convert the EXACT otherwise we convert
2235 the whole branch sequence, including the first.
2237 /* Find the node we are going to overwrite */
2238 if ( first != startbranch || OP( last ) == BRANCH ) {
2239 /* branch sub-chain */
2240 NEXT_OFF( first ) = (U16)(last - first);
2241 #ifdef RE_TRACK_PATTERN_OFFSETS
2243 mjd_offset= Node_Offset((convert));
2244 mjd_nodelen= Node_Length((convert));
2247 /* whole branch chain */
2249 #ifdef RE_TRACK_PATTERN_OFFSETS
2252 const regnode *nop = NEXTOPER( convert );
2253 mjd_offset= Node_Offset((nop));
2254 mjd_nodelen= Node_Length((nop));
2258 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2259 (int)depth * 2 + 2, "",
2260 (UV)mjd_offset, (UV)mjd_nodelen)
2263 /* But first we check to see if there is a common prefix we can
2264 split out as an EXACT and put in front of the TRIE node. */
2265 trie->startstate= 1;
2266 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2268 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2272 const U32 base = trie->states[ state ].trans.base;
2274 if ( trie->states[state].wordnum )
2277 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2278 if ( ( base + ofs >= trie->uniquecharcount ) &&
2279 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2280 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2282 if ( ++count > 1 ) {
2283 SV **tmp = av_fetch( revcharmap, ofs, 0);
2284 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2285 if ( state == 1 ) break;
2287 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2289 PerlIO_printf(Perl_debug_log,
2290 "%*sNew Start State=%"UVuf" Class: [",
2291 (int)depth * 2 + 2, "",
2294 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2295 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2297 TRIE_BITMAP_SET(trie,*ch);
2299 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2301 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2305 TRIE_BITMAP_SET(trie,*ch);
2307 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2308 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2314 SV **tmp = av_fetch( revcharmap, idx, 0);
2316 char *ch = SvPV( *tmp, len );
2318 SV *sv=sv_newmortal();
2319 PerlIO_printf( Perl_debug_log,
2320 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2321 (int)depth * 2 + 2, "",
2323 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2324 PL_colors[0], PL_colors[1],
2325 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2326 PERL_PV_ESCAPE_FIRSTCHAR
2331 OP( convert ) = nodetype;
2332 str=STRING(convert);
2335 STR_LEN(convert) += len;
2341 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2346 trie->prefixlen = (state-1);
2348 regnode *n = convert+NODE_SZ_STR(convert);
2349 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2350 trie->startstate = state;
2351 trie->minlen -= (state - 1);
2352 trie->maxlen -= (state - 1);
2354 /* At least the UNICOS C compiler choked on this
2355 * being argument to DEBUG_r(), so let's just have
2358 #ifdef PERL_EXT_RE_BUILD
2364 regnode *fix = convert;
2365 U32 word = trie->wordcount;
2367 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2368 while( ++fix < n ) {
2369 Set_Node_Offset_Length(fix, 0, 0);
2372 SV ** const tmp = av_fetch( trie_words, word, 0 );
2374 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2375 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2377 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2385 NEXT_OFF(convert) = (U16)(tail - convert);
2386 DEBUG_r(optimize= n);
2392 if ( trie->maxlen ) {
2393 NEXT_OFF( convert ) = (U16)(tail - convert);
2394 ARG_SET( convert, data_slot );
2395 /* Store the offset to the first unabsorbed branch in
2396 jump[0], which is otherwise unused by the jump logic.
2397 We use this when dumping a trie and during optimisation. */
2399 trie->jump[0] = (U16)(nextbranch - convert);
2401 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2402 * and there is a bitmap
2403 * and the first "jump target" node we found leaves enough room
2404 * then convert the TRIE node into a TRIEC node, with the bitmap
2405 * embedded inline in the opcode - this is hypothetically faster.
2407 if ( !trie->states[trie->startstate].wordnum
2409 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2411 OP( convert ) = TRIEC;
2412 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2413 PerlMemShared_free(trie->bitmap);
2416 OP( convert ) = TRIE;
2418 /* store the type in the flags */
2419 convert->flags = nodetype;
2423 + regarglen[ OP( convert ) ];
2425 /* XXX We really should free up the resource in trie now,
2426 as we won't use them - (which resources?) dmq */
2428 /* needed for dumping*/
2429 DEBUG_r(if (optimize) {
2430 regnode *opt = convert;
2432 while ( ++opt < optimize) {
2433 Set_Node_Offset_Length(opt,0,0);
2436 Try to clean up some of the debris left after the
2439 while( optimize < jumper ) {
2440 mjd_nodelen += Node_Length((optimize));
2441 OP( optimize ) = OPTIMIZED;
2442 Set_Node_Offset_Length(optimize,0,0);
2445 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2447 } /* end node insert */
2448 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, convert);
2450 /* Finish populating the prev field of the wordinfo array. Walk back
2451 * from each accept state until we find another accept state, and if
2452 * so, point the first word's .prev field at the second word. If the
2453 * second already has a .prev field set, stop now. This will be the
2454 * case either if we've already processed that word's accept state,
2455 * or that state had multiple words, and the overspill words were
2456 * already linked up earlier.
2463 for (word=1; word <= trie->wordcount; word++) {
2465 if (trie->wordinfo[word].prev)
2467 state = trie->wordinfo[word].accept;
2469 state = prev_states[state];
2472 prev = trie->states[state].wordnum;
2476 trie->wordinfo[word].prev = prev;
2478 Safefree(prev_states);
2482 /* and now dump out the compressed format */
2483 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2485 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2487 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2488 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2490 SvREFCNT_dec_NN(revcharmap);
2494 : trie->startstate>1
2500 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2502 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2504 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2505 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2508 We find the fail state for each state in the trie, this state is the longest proper
2509 suffix of the current state's 'word' that is also a proper prefix of another word in our
2510 trie. State 1 represents the word '' and is thus the default fail state. This allows
2511 the DFA not to have to restart after its tried and failed a word at a given point, it
2512 simply continues as though it had been matching the other word in the first place.
2514 'abcdgu'=~/abcdefg|cdgu/
2515 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2516 fail, which would bring us to the state representing 'd' in the second word where we would
2517 try 'g' and succeed, proceeding to match 'cdgu'.
2519 /* add a fail transition */
2520 const U32 trie_offset = ARG(source);
2521 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2523 const U32 ucharcount = trie->uniquecharcount;
2524 const U32 numstates = trie->statecount;
2525 const U32 ubound = trie->lasttrans + ucharcount;
2529 U32 base = trie->states[ 1 ].trans.base;
2532 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2533 GET_RE_DEBUG_FLAGS_DECL;
2535 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2537 PERL_UNUSED_ARG(depth);
2541 ARG_SET( stclass, data_slot );
2542 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2543 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2544 aho->trie=trie_offset;
2545 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2546 Copy( trie->states, aho->states, numstates, reg_trie_state );
2547 Newxz( q, numstates, U32);
2548 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2551 /* initialize fail[0..1] to be 1 so that we always have
2552 a valid final fail state */
2553 fail[ 0 ] = fail[ 1 ] = 1;
2555 for ( charid = 0; charid < ucharcount ; charid++ ) {
2556 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2558 q[ q_write ] = newstate;
2559 /* set to point at the root */
2560 fail[ q[ q_write++ ] ]=1;
2563 while ( q_read < q_write) {
2564 const U32 cur = q[ q_read++ % numstates ];
2565 base = trie->states[ cur ].trans.base;
2567 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2568 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2570 U32 fail_state = cur;
2573 fail_state = fail[ fail_state ];
2574 fail_base = aho->states[ fail_state ].trans.base;
2575 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2577 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2578 fail[ ch_state ] = fail_state;
2579 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2581 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2583 q[ q_write++ % numstates] = ch_state;
2587 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2588 when we fail in state 1, this allows us to use the
2589 charclass scan to find a valid start char. This is based on the principle
2590 that theres a good chance the string being searched contains lots of stuff
2591 that cant be a start char.
2593 fail[ 0 ] = fail[ 1 ] = 0;
2594 DEBUG_TRIE_COMPILE_r({
2595 PerlIO_printf(Perl_debug_log,
2596 "%*sStclass Failtable (%"UVuf" states): 0",
2597 (int)(depth * 2), "", (UV)numstates
2599 for( q_read=1; q_read<numstates; q_read++ ) {
2600 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2602 PerlIO_printf(Perl_debug_log, "\n");
2605 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2610 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2611 * These need to be revisited when a newer toolchain becomes available.
2613 #if defined(__sparc64__) && defined(__GNUC__)
2614 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2615 # undef SPARC64_GCC_WORKAROUND
2616 # define SPARC64_GCC_WORKAROUND 1
2620 #define DEBUG_PEEP(str,scan,depth) \
2621 DEBUG_OPTIMISE_r({if (scan){ \
2622 SV * const mysv=sv_newmortal(); \
2623 regnode *Next = regnext(scan); \
2624 regprop(RExC_rx, mysv, scan); \
2625 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2626 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2627 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2631 /* The below joins as many adjacent EXACTish nodes as possible into a single
2632 * one. The regop may be changed if the node(s) contain certain sequences that
2633 * require special handling. The joining is only done if:
2634 * 1) there is room in the current conglomerated node to entirely contain the
2636 * 2) they are the exact same node type
2638 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2639 * these get optimized out
2641 * If a node is to match under /i (folded), the number of characters it matches
2642 * can be different than its character length if it contains a multi-character
2643 * fold. *min_subtract is set to the total delta of the input nodes.
2645 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2646 * and contains LATIN SMALL LETTER SHARP S
2648 * This is as good a place as any to discuss the design of handling these
2649 * multi-character fold sequences. It's been wrong in Perl for a very long
2650 * time. There are three code points in Unicode whose multi-character folds
2651 * were long ago discovered to mess things up. The previous designs for
2652 * dealing with these involved assigning a special node for them. This
2653 * approach doesn't work, as evidenced by this example:
2654 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2655 * Both these fold to "sss", but if the pattern is parsed to create a node that
2656 * would match just the \xDF, it won't be able to handle the case where a
2657 * successful match would have to cross the node's boundary. The new approach
2658 * that hopefully generally solves the problem generates an EXACTFU_SS node
2661 * It turns out that there are problems with all multi-character folds, and not
2662 * just these three. Now the code is general, for all such cases, but the
2663 * three still have some special handling. The approach taken is:
2664 * 1) This routine examines each EXACTFish node that could contain multi-
2665 * character fold sequences. It returns in *min_subtract how much to
2666 * subtract from the the actual length of the string to get a real minimum
2667 * match length; it is 0 if there are no multi-char folds. This delta is
2668 * used by the caller to adjust the min length of the match, and the delta
2669 * between min and max, so that the optimizer doesn't reject these
2670 * possibilities based on size constraints.
2671 * 2) Certain of these sequences require special handling by the trie code,
2672 * so, if found, this code changes the joined node type to special ops:
2673 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2674 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2675 * is used for an EXACTFU node that contains at least one "ss" sequence in
2676 * it. For non-UTF-8 patterns and strings, this is the only case where
2677 * there is a possible fold length change. That means that a regular
2678 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2679 * with length changes, and so can be processed faster. regexec.c takes
2680 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2681 * pre-folded by regcomp.c. This saves effort in regex matching.
2682 * However, the pre-folding isn't done for non-UTF8 patterns because the
2683 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2684 * down by forcing the pattern into UTF8 unless necessary. Also what
2685 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2686 * possibilities for the non-UTF8 patterns are quite simple, except for
2687 * the sharp s. All the ones that don't involve a UTF-8 target string are
2688 * members of a fold-pair, and arrays are set up for all of them so that
2689 * the other member of the pair can be found quickly. Code elsewhere in
2690 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2691 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2692 * described in the next item.
2693 * 4) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
2694 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2695 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2696 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2697 * character in the pattern corresponds to at most a single character in
2698 * the target string. (And I do mean character, and not byte here, unlike
2699 * other parts of the documentation that have never been updated to
2700 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2701 * two character string 'ss'; in EXACTFA nodes it can match
2702 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2703 * instances where it is violated. I'm reluctant to try to change the
2704 * assumption, as the code involved is impenetrable to me (khw), so
2705 * instead the code here punts. This routine examines (when the pattern
2706 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2707 * boolean indicating whether or not the node contains a sharp s. When it
2708 * is true, the caller sets a flag that later causes the optimizer in this
2709 * file to not set values for the floating and fixed string lengths, and
2710 * thus avoids the optimizer code in regexec.c that makes the invalid
2711 * assumption. Thus, there is no optimization based on string lengths for
2712 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2713 * (The reason the assumption is wrong only in these two cases is that all
2714 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2715 * other folds to their expanded versions. We can't prefold sharp s to
2716 * 'ss' in EXACTF nodes because we don't know at compile time if it
2717 * actually matches 'ss' or not. It will match iff the target string is
2718 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2719 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2720 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2721 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2722 * require the pattern to be forced into UTF-8, the overhead of which we
2726 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2727 if (PL_regkind[OP(scan)] == EXACT) \
2728 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2731 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, UV *min_subtract, bool *has_exactf_sharp_s, U32 flags,regnode *val, U32 depth) {
2732 /* Merge several consecutive EXACTish nodes into one. */
2733 regnode *n = regnext(scan);
2735 regnode *next = scan + NODE_SZ_STR(scan);
2739 regnode *stop = scan;
2740 GET_RE_DEBUG_FLAGS_DECL;
2742 PERL_UNUSED_ARG(depth);
2745 PERL_ARGS_ASSERT_JOIN_EXACT;
2746 #ifndef EXPERIMENTAL_INPLACESCAN
2747 PERL_UNUSED_ARG(flags);
2748 PERL_UNUSED_ARG(val);
2750 DEBUG_PEEP("join",scan,depth);
2752 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2753 * EXACT ones that are mergeable to the current one. */
2755 && (PL_regkind[OP(n)] == NOTHING
2756 || (stringok && OP(n) == OP(scan)))
2758 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2761 if (OP(n) == TAIL || n > next)
2763 if (PL_regkind[OP(n)] == NOTHING) {
2764 DEBUG_PEEP("skip:",n,depth);
2765 NEXT_OFF(scan) += NEXT_OFF(n);
2766 next = n + NODE_STEP_REGNODE;
2773 else if (stringok) {
2774 const unsigned int oldl = STR_LEN(scan);
2775 regnode * const nnext = regnext(n);
2777 /* XXX I (khw) kind of doubt that this works on platforms where
2778 * U8_MAX is above 255 because of lots of other assumptions */
2779 /* Don't join if the sum can't fit into a single node */
2780 if (oldl + STR_LEN(n) > U8_MAX)
2783 DEBUG_PEEP("merg",n,depth);
2786 NEXT_OFF(scan) += NEXT_OFF(n);
2787 STR_LEN(scan) += STR_LEN(n);
2788 next = n + NODE_SZ_STR(n);
2789 /* Now we can overwrite *n : */
2790 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2798 #ifdef EXPERIMENTAL_INPLACESCAN
2799 if (flags && !NEXT_OFF(n)) {
2800 DEBUG_PEEP("atch", val, depth);
2801 if (reg_off_by_arg[OP(n)]) {
2802 ARG_SET(n, val - n);
2805 NEXT_OFF(n) = val - n;
2813 *has_exactf_sharp_s = FALSE;
2815 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2816 * can now analyze for sequences of problematic code points. (Prior to
2817 * this final joining, sequences could have been split over boundaries, and
2818 * hence missed). The sequences only happen in folding, hence for any
2819 * non-EXACT EXACTish node */
2820 if (OP(scan) != EXACT) {
2821 const U8 * const s0 = (U8*) STRING(scan);
2823 const U8 * const s_end = s0 + STR_LEN(scan);
2825 /* One pass is made over the node's string looking for all the
2826 * possibilities. to avoid some tests in the loop, there are two main
2827 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2831 /* Examine the string for a multi-character fold sequence. UTF-8
2832 * patterns have all characters pre-folded by the time this code is
2834 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2835 length sequence we are looking for is 2 */
2838 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2839 if (! len) { /* Not a multi-char fold: get next char */
2844 /* Nodes with 'ss' require special handling, except for EXACTFL
2845 * and EXACTFA for which there is no multi-char fold to this */
2846 if (len == 2 && *s == 's' && *(s+1) == 's'
2847 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2850 OP(scan) = EXACTFU_SS;
2853 else if (len == 6 /* len is the same in both ASCII and EBCDIC
2855 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2856 COMBINING_DIAERESIS_UTF8
2857 COMBINING_ACUTE_ACCENT_UTF8,
2859 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2860 COMBINING_DIAERESIS_UTF8
2861 COMBINING_ACUTE_ACCENT_UTF8,
2866 /* These two folds require special handling by trie's, so
2867 * change the node type to indicate this. If EXACTFA and
2868 * EXACTFL were ever to be handled by trie's, this would
2869 * have to be changed. If this node has already been
2870 * changed to EXACTFU_SS in this loop, leave it as is. (I
2871 * (khw) think it doesn't matter in regexec.c for UTF
2872 * patterns, but no need to change it */
2873 if (OP(scan) == EXACTFU) {
2874 OP(scan) = EXACTFU_TRICKYFOLD;
2878 else { /* Here is a generic multi-char fold. */
2879 const U8* multi_end = s + len;
2881 /* Count how many characters in it. In the case of /l and
2882 * /aa, no folds which contain ASCII code points are
2883 * allowed, so check for those, and skip if found. (In
2884 * EXACTFL, no folds are allowed to any Latin1 code point,
2885 * not just ASCII. But there aren't any of these
2886 * currently, nor ever likely, so don't take the time to
2887 * test for them. The code that generates the
2888 * is_MULTI_foo() macros croaks should one actually get put
2889 * into Unicode .) */
2890 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2891 count = utf8_length(s, multi_end);
2895 while (s < multi_end) {
2898 goto next_iteration;
2908 /* The delta is how long the sequence is minus 1 (1 is how long
2909 * the character that folds to the sequence is) */
2910 *min_subtract += count - 1;
2914 else if (OP(scan) == EXACTFA) {
2916 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2917 * fold to the ASCII range (and there are no existing ones in the
2918 * upper latin1 range). But, as outlined in the comments preceding
2919 * this function, we need to flag any occurrences of the sharp s */
2921 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2922 *has_exactf_sharp_s = TRUE;
2929 else if (OP(scan) != EXACTFL) {
2931 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2932 * multi-char folds that are all Latin1. (This code knows that
2933 * there are no current multi-char folds possible with EXACTFL,
2934 * relying on fold_grind.t to catch any errors if the very unlikely
2935 * event happens that some get added in future Unicode versions.)
2936 * As explained in the comments preceding this function, we look
2937 * also for the sharp s in EXACTF nodes; it can be in the final
2938 * position. Otherwise we can stop looking 1 byte earlier because
2939 * have to find at least two characters for a multi-fold */
2940 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2942 /* The below is perhaps overboard, but this allows us to save a
2943 * test each time through the loop at the expense of a mask. This
2944 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2945 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2946 * are 64. This uses an exclusive 'or' to find that bit and then
2947 * inverts it to form a mask, with just a single 0, in the bit
2948 * position where 'S' and 's' differ. */
2949 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2950 const U8 s_masked = 's' & S_or_s_mask;
2953 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2954 if (! len) { /* Not a multi-char fold. */
2955 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2957 *has_exactf_sharp_s = TRUE;
2964 && ((*s & S_or_s_mask) == s_masked)
2965 && ((*(s+1) & S_or_s_mask) == s_masked))
2968 /* EXACTF nodes need to know that the minimum length
2969 * changed so that a sharp s in the string can match this
2970 * ss in the pattern, but they remain EXACTF nodes, as they
2971 * won't match this unless the target string is is UTF-8,
2972 * which we don't know until runtime */
2973 if (OP(scan) != EXACTF) {
2974 OP(scan) = EXACTFU_SS;
2978 *min_subtract += len - 1;
2985 /* Allow dumping but overwriting the collection of skipped
2986 * ops and/or strings with fake optimized ops */
2987 n = scan + NODE_SZ_STR(scan);
2995 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2999 /* REx optimizer. Converts nodes into quicker variants "in place".
3000 Finds fixed substrings. */
3002 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3003 to the position after last scanned or to NULL. */
3005 #define INIT_AND_WITHP \
3006 assert(!and_withp); \
3007 Newx(and_withp,1,struct regnode_charclass_class); \
3008 SAVEFREEPV(and_withp)
3010 /* this is a chain of data about sub patterns we are processing that
3011 need to be handled separately/specially in study_chunk. Its so
3012 we can simulate recursion without losing state. */
3014 typedef struct scan_frame {
3015 regnode *last; /* last node to process in this frame */
3016 regnode *next; /* next node to process when last is reached */
3017 struct scan_frame *prev; /*previous frame*/
3018 I32 stop; /* what stopparen do we use */
3022 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3025 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3026 I32 *minlenp, I32 *deltap,
3031 struct regnode_charclass_class *and_withp,
3032 U32 flags, U32 depth)
3033 /* scanp: Start here (read-write). */
3034 /* deltap: Write maxlen-minlen here. */
3035 /* last: Stop before this one. */
3036 /* data: string data about the pattern */
3037 /* stopparen: treat close N as END */
3038 /* recursed: which subroutines have we recursed into */
3039 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3042 I32 min = 0; /* There must be at least this number of characters to match */
3044 regnode *scan = *scanp, *next;
3046 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3047 int is_inf_internal = 0; /* The studied chunk is infinite */
3048 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3049 scan_data_t data_fake;
3050 SV *re_trie_maxbuff = NULL;
3051 regnode *first_non_open = scan;
3052 I32 stopmin = I32_MAX;
3053 scan_frame *frame = NULL;
3054 GET_RE_DEBUG_FLAGS_DECL;
3056 PERL_ARGS_ASSERT_STUDY_CHUNK;
3059 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3063 while (first_non_open && OP(first_non_open) == OPEN)
3064 first_non_open=regnext(first_non_open);
3069 while ( scan && OP(scan) != END && scan < last ){
3070 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3071 node length to get a real minimum (because
3072 the folded version may be shorter) */
3073 bool has_exactf_sharp_s = FALSE;
3074 /* Peephole optimizer: */
3075 DEBUG_STUDYDATA("Peep:", data,depth);
3076 DEBUG_PEEP("Peep",scan,depth);
3078 /* Its not clear to khw or hv why this is done here, and not in the
3079 * clauses that deal with EXACT nodes. khw's guess is that it's
3080 * because of a previous design */
3081 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3083 /* Follow the next-chain of the current node and optimize
3084 away all the NOTHINGs from it. */
3085 if (OP(scan) != CURLYX) {
3086 const int max = (reg_off_by_arg[OP(scan)]
3088 /* I32 may be smaller than U16 on CRAYs! */
3089 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3090 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3094 /* Skip NOTHING and LONGJMP. */
3095 while ((n = regnext(n))
3096 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3097 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3098 && off + noff < max)
3100 if (reg_off_by_arg[OP(scan)])
3103 NEXT_OFF(scan) = off;
3108 /* The principal pseudo-switch. Cannot be a switch, since we
3109 look into several different things. */
3110 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3111 || OP(scan) == IFTHEN) {
3112 next = regnext(scan);
3114 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3116 if (OP(next) == code || code == IFTHEN) {
3117 /* NOTE - There is similar code to this block below for handling
3118 TRIE nodes on a re-study. If you change stuff here check there
3120 I32 max1 = 0, min1 = I32_MAX, num = 0;
3121 struct regnode_charclass_class accum;
3122 regnode * const startbranch=scan;
3124 if (flags & SCF_DO_SUBSTR)
3125 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3126 if (flags & SCF_DO_STCLASS)
3127 cl_init_zero(pRExC_state, &accum);
3129 while (OP(scan) == code) {
3130 I32 deltanext, minnext, f = 0, fake;
3131 struct regnode_charclass_class this_class;
3134 data_fake.flags = 0;
3136 data_fake.whilem_c = data->whilem_c;
3137 data_fake.last_closep = data->last_closep;
3140 data_fake.last_closep = &fake;
3142 data_fake.pos_delta = delta;
3143 next = regnext(scan);
3144 scan = NEXTOPER(scan);
3146 scan = NEXTOPER(scan);
3147 if (flags & SCF_DO_STCLASS) {
3148 cl_init(pRExC_state, &this_class);
3149 data_fake.start_class = &this_class;
3150 f = SCF_DO_STCLASS_AND;
3152 if (flags & SCF_WHILEM_VISITED_POS)
3153 f |= SCF_WHILEM_VISITED_POS;
3155 /* we suppose the run is continuous, last=next...*/
3156 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3158 stopparen, recursed, NULL, f,depth+1);
3161 if (deltanext == I32_MAX) {
3162 is_inf = is_inf_internal = 1;
3164 } else if (max1 < minnext + deltanext)
3165 max1 = minnext + deltanext;
3167 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3169 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3170 if ( stopmin > minnext)
3171 stopmin = min + min1;
3172 flags &= ~SCF_DO_SUBSTR;
3174 data->flags |= SCF_SEEN_ACCEPT;
3177 if (data_fake.flags & SF_HAS_EVAL)
3178 data->flags |= SF_HAS_EVAL;
3179 data->whilem_c = data_fake.whilem_c;
3181 if (flags & SCF_DO_STCLASS)
3182 cl_or(pRExC_state, &accum, &this_class);
3184 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3186 if (flags & SCF_DO_SUBSTR) {
3187 data->pos_min += min1;
3188 if (data->pos_delta >= I32_MAX - (max1 - min1))
3189 data->pos_delta = I32_MAX;
3191 data->pos_delta += max1 - min1;
3192 if (max1 != min1 || is_inf)
3193 data->longest = &(data->longest_float);
3196 if (delta == I32_MAX || I32_MAX - delta - (max1 - min1) < 0)
3199 delta += max1 - min1;
3200 if (flags & SCF_DO_STCLASS_OR) {
3201 cl_or(pRExC_state, data->start_class, &accum);
3203 cl_and(data->start_class, and_withp);
3204 flags &= ~SCF_DO_STCLASS;
3207 else if (flags & SCF_DO_STCLASS_AND) {
3209 cl_and(data->start_class, &accum);
3210 flags &= ~SCF_DO_STCLASS;
3213 /* Switch to OR mode: cache the old value of
3214 * data->start_class */
3216 StructCopy(data->start_class, and_withp,
3217 struct regnode_charclass_class);
3218 flags &= ~SCF_DO_STCLASS_AND;
3219 StructCopy(&accum, data->start_class,
3220 struct regnode_charclass_class);
3221 flags |= SCF_DO_STCLASS_OR;
3222 SET_SSC_EOS(data->start_class);
3226 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3229 Assuming this was/is a branch we are dealing with: 'scan' now
3230 points at the item that follows the branch sequence, whatever
3231 it is. We now start at the beginning of the sequence and look
3238 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3240 If we can find such a subsequence we need to turn the first
3241 element into a trie and then add the subsequent branch exact
3242 strings to the trie.
3246 1. patterns where the whole set of branches can be converted.
3248 2. patterns where only a subset can be converted.
3250 In case 1 we can replace the whole set with a single regop
3251 for the trie. In case 2 we need to keep the start and end
3254 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3255 becomes BRANCH TRIE; BRANCH X;
3257 There is an additional case, that being where there is a
3258 common prefix, which gets split out into an EXACT like node
3259 preceding the TRIE node.
3261 If x(1..n)==tail then we can do a simple trie, if not we make
3262 a "jump" trie, such that when we match the appropriate word
3263 we "jump" to the appropriate tail node. Essentially we turn
3264 a nested if into a case structure of sorts.
3269 if (!re_trie_maxbuff) {
3270 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3271 if (!SvIOK(re_trie_maxbuff))
3272 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3274 if ( SvIV(re_trie_maxbuff)>=0 ) {
3276 regnode *first = (regnode *)NULL;
3277 regnode *last = (regnode *)NULL;
3278 regnode *tail = scan;
3283 SV * const mysv = sv_newmortal(); /* for dumping */
3285 /* var tail is used because there may be a TAIL
3286 regop in the way. Ie, the exacts will point to the
3287 thing following the TAIL, but the last branch will
3288 point at the TAIL. So we advance tail. If we
3289 have nested (?:) we may have to move through several
3293 while ( OP( tail ) == TAIL ) {
3294 /* this is the TAIL generated by (?:) */
3295 tail = regnext( tail );
3299 DEBUG_TRIE_COMPILE_r({
3300 regprop(RExC_rx, mysv, tail );
3301 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3302 (int)depth * 2 + 2, "",
3303 "Looking for TRIE'able sequences. Tail node is: ",
3304 SvPV_nolen_const( mysv )
3310 Step through the branches
3311 cur represents each branch,
3312 noper is the first thing to be matched as part of that branch
3313 noper_next is the regnext() of that node.
3315 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3316 via a "jump trie" but we also support building with NOJUMPTRIE,
3317 which restricts the trie logic to structures like /FOO|BAR/.
3319 If noper is a trieable nodetype then the branch is a possible optimization
3320 target. If we are building under NOJUMPTRIE then we require that noper_next
3321 is the same as scan (our current position in the regex program).
3323 Once we have two or more consecutive such branches we can create a
3324 trie of the EXACT's contents and stitch it in place into the program.
3326 If the sequence represents all of the branches in the alternation we
3327 replace the entire thing with a single TRIE node.
3329 Otherwise when it is a subsequence we need to stitch it in place and
3330 replace only the relevant branches. This means the first branch has
3331 to remain as it is used by the alternation logic, and its next pointer,
3332 and needs to be repointed at the item on the branch chain following
3333 the last branch we have optimized away.
3335 This could be either a BRANCH, in which case the subsequence is internal,
3336 or it could be the item following the branch sequence in which case the
3337 subsequence is at the end (which does not necessarily mean the first node
3338 is the start of the alternation).
3340 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3343 ----------------+-----------
3347 EXACTFU_SS | EXACTFU
3348 EXACTFU_TRICKYFOLD | EXACTFU
3353 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3354 ( EXACT == (X) ) ? EXACT : \
3355 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3358 /* dont use tail as the end marker for this traverse */
3359 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3360 regnode * const noper = NEXTOPER( cur );
3361 U8 noper_type = OP( noper );
3362 U8 noper_trietype = TRIE_TYPE( noper_type );
3363 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3364 regnode * const noper_next = regnext( noper );
3365 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3366 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3369 DEBUG_TRIE_COMPILE_r({
3370 regprop(RExC_rx, mysv, cur);
3371 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3372 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3374 regprop(RExC_rx, mysv, noper);
3375 PerlIO_printf( Perl_debug_log, " -> %s",
3376 SvPV_nolen_const(mysv));
3379 regprop(RExC_rx, mysv, noper_next );
3380 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3381 SvPV_nolen_const(mysv));
3383 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3384 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3385 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3389 /* Is noper a trieable nodetype that can be merged with the
3390 * current trie (if there is one)? */
3394 ( noper_trietype == NOTHING)
3395 || ( trietype == NOTHING )
3396 || ( trietype == noper_trietype )
3399 && noper_next == tail
3403 /* Handle mergable triable node
3404 * Either we are the first node in a new trieable sequence,
3405 * in which case we do some bookkeeping, otherwise we update
3406 * the end pointer. */
3409 if ( noper_trietype == NOTHING ) {
3410 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3411 regnode * const noper_next = regnext( noper );
3412 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3413 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3416 if ( noper_next_trietype ) {
3417 trietype = noper_next_trietype;
3418 } else if (noper_next_type) {
3419 /* a NOTHING regop is 1 regop wide. We need at least two
3420 * for a trie so we can't merge this in */
3424 trietype = noper_trietype;
3427 if ( trietype == NOTHING )
3428 trietype = noper_trietype;
3433 } /* end handle mergable triable node */
3435 /* handle unmergable node -
3436 * noper may either be a triable node which can not be tried
3437 * together with the current trie, or a non triable node */
3439 /* If last is set and trietype is not NOTHING then we have found
3440 * at least two triable branch sequences in a row of a similar
3441 * trietype so we can turn them into a trie. If/when we
3442 * allow NOTHING to start a trie sequence this condition will be
3443 * required, and it isn't expensive so we leave it in for now. */
3444 if ( trietype && trietype != NOTHING )
3445 make_trie( pRExC_state,
3446 startbranch, first, cur, tail, count,
3447 trietype, depth+1 );
3448 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3452 && noper_next == tail
3455 /* noper is triable, so we can start a new trie sequence */
3458 trietype = noper_trietype;
3460 /* if we already saw a first but the current node is not triable then we have
3461 * to reset the first information. */
3466 } /* end handle unmergable node */
3467 } /* loop over branches */
3468 DEBUG_TRIE_COMPILE_r({
3469 regprop(RExC_rx, mysv, cur);
3470 PerlIO_printf( Perl_debug_log,
3471 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3472 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3475 if ( last && trietype ) {
3476 if ( trietype != NOTHING ) {
3477 /* the last branch of the sequence was part of a trie,
3478 * so we have to construct it here outside of the loop
3480 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3481 #ifdef TRIE_STUDY_OPT
3482 if ( ((made == MADE_EXACT_TRIE &&
3483 startbranch == first)
3484 || ( first_non_open == first )) &&
3486 flags |= SCF_TRIE_RESTUDY;
3487 if ( startbranch == first
3490 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3495 /* at this point we know whatever we have is a NOTHING sequence/branch
3496 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3498 if ( startbranch == first ) {
3500 /* the entire thing is a NOTHING sequence, something like this:
3501 * (?:|) So we can turn it into a plain NOTHING op. */
3502 DEBUG_TRIE_COMPILE_r({
3503 regprop(RExC_rx, mysv, cur);
3504 PerlIO_printf( Perl_debug_log,
3505 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3506 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3509 OP(startbranch)= NOTHING;
3510 NEXT_OFF(startbranch)= tail - startbranch;
3511 for ( opt= startbranch + 1; opt < tail ; opt++ )
3515 } /* end if ( last) */
3516 } /* TRIE_MAXBUF is non zero */
3521 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3522 scan = NEXTOPER(NEXTOPER(scan));
3523 } else /* single branch is optimized. */
3524 scan = NEXTOPER(scan);
3526 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3527 scan_frame *newframe = NULL;
3532 if (OP(scan) != SUSPEND) {
3533 /* set the pointer */
3534 if (OP(scan) == GOSUB) {
3536 RExC_recurse[ARG2L(scan)] = scan;
3537 start = RExC_open_parens[paren-1];
3538 end = RExC_close_parens[paren-1];
3541 start = RExC_rxi->program + 1;
3545 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3546 SAVEFREEPV(recursed);
3548 if (!PAREN_TEST(recursed,paren+1)) {
3549 PAREN_SET(recursed,paren+1);
3550 Newx(newframe,1,scan_frame);
3552 if (flags & SCF_DO_SUBSTR) {
3553 SCAN_COMMIT(pRExC_state,data,minlenp);
3554 data->longest = &(data->longest_float);
3556 is_inf = is_inf_internal = 1;
3557 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3558 cl_anything(pRExC_state, data->start_class);
3559 flags &= ~SCF_DO_STCLASS;
3562 Newx(newframe,1,scan_frame);
3565 end = regnext(scan);
3570 SAVEFREEPV(newframe);
3571 newframe->next = regnext(scan);
3572 newframe->last = last;
3573 newframe->stop = stopparen;
3574 newframe->prev = frame;
3584 else if (OP(scan) == EXACT) {
3585 I32 l = STR_LEN(scan);
3588 const U8 * const s = (U8*)STRING(scan);
3589 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3590 l = utf8_length(s, s + l);
3592 uc = *((U8*)STRING(scan));
3595 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3596 /* The code below prefers earlier match for fixed
3597 offset, later match for variable offset. */
3598 if (data->last_end == -1) { /* Update the start info. */
3599 data->last_start_min = data->pos_min;
3600 data->last_start_max = is_inf
3601 ? I32_MAX : data->pos_min + data->pos_delta;
3603 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3605 SvUTF8_on(data->last_found);
3607 SV * const sv = data->last_found;
3608 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3609 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3610 if (mg && mg->mg_len >= 0)
3611 mg->mg_len += utf8_length((U8*)STRING(scan),
3612 (U8*)STRING(scan)+STR_LEN(scan));
3614 data->last_end = data->pos_min + l;
3615 data->pos_min += l; /* As in the first entry. */
3616 data->flags &= ~SF_BEFORE_EOL;
3618 if (flags & SCF_DO_STCLASS_AND) {
3619 /* Check whether it is compatible with what we know already! */
3623 /* If compatible, we or it in below. It is compatible if is
3624 * in the bitmp and either 1) its bit or its fold is set, or 2)
3625 * it's for a locale. Even if there isn't unicode semantics
3626 * here, at runtime there may be because of matching against a
3627 * utf8 string, so accept a possible false positive for
3628 * latin1-range folds */
3630 (!(data->start_class->flags & ANYOF_LOCALE)
3631 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3632 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3633 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3638 ANYOF_CLASS_ZERO(data->start_class);
3639 ANYOF_BITMAP_ZERO(data->start_class);
3641 ANYOF_BITMAP_SET(data->start_class, uc);
3642 else if (uc >= 0x100) {
3645 /* Some Unicode code points fold to the Latin1 range; as
3646 * XXX temporary code, instead of figuring out if this is
3647 * one, just assume it is and set all the start class bits
3648 * that could be some such above 255 code point's fold
3649 * which will generate fals positives. As the code
3650 * elsewhere that does compute the fold settles down, it
3651 * can be extracted out and re-used here */
3652 for (i = 0; i < 256; i++){
3653 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3654 ANYOF_BITMAP_SET(data->start_class, i);
3658 CLEAR_SSC_EOS(data->start_class);
3660 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3662 else if (flags & SCF_DO_STCLASS_OR) {
3663 /* false positive possible if the class is case-folded */
3665 ANYOF_BITMAP_SET(data->start_class, uc);
3667 data->start_class->flags |= ANYOF_UNICODE_ALL;
3668 CLEAR_SSC_EOS(data->start_class);
3669 cl_and(data->start_class, and_withp);
3671 flags &= ~SCF_DO_STCLASS;
3673 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3674 I32 l = STR_LEN(scan);
3675 UV uc = *((U8*)STRING(scan));
3677 /* Search for fixed substrings supports EXACT only. */
3678 if (flags & SCF_DO_SUBSTR) {
3680 SCAN_COMMIT(pRExC_state, data, minlenp);
3683 const U8 * const s = (U8 *)STRING(scan);
3684 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3685 l = utf8_length(s, s + l);
3687 if (has_exactf_sharp_s) {
3688 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3690 min += l - min_subtract;
3692 delta += min_subtract;
3693 if (flags & SCF_DO_SUBSTR) {
3694 data->pos_min += l - min_subtract;
3695 if (data->pos_min < 0) {
3698 data->pos_delta += min_subtract;
3700 data->longest = &(data->longest_float);
3703 if (flags & SCF_DO_STCLASS_AND) {
3704 /* Check whether it is compatible with what we know already! */
3707 (!(data->start_class->flags & ANYOF_LOCALE)
3708 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3709 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3713 ANYOF_CLASS_ZERO(data->start_class);
3714 ANYOF_BITMAP_ZERO(data->start_class);
3716 ANYOF_BITMAP_SET(data->start_class, uc);
3717 CLEAR_SSC_EOS(data->start_class);
3718 if (OP(scan) == EXACTFL) {
3719 /* XXX This set is probably no longer necessary, and
3720 * probably wrong as LOCALE now is on in the initial
3722 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3726 /* Also set the other member of the fold pair. In case
3727 * that unicode semantics is called for at runtime, use
3728 * the full latin1 fold. (Can't do this for locale,
3729 * because not known until runtime) */
3730 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3732 /* All other (EXACTFL handled above) folds except under
3733 * /iaa that include s, S, and sharp_s also may include
3735 if (OP(scan) != EXACTFA) {
3736 if (uc == 's' || uc == 'S') {
3737 ANYOF_BITMAP_SET(data->start_class,
3738 LATIN_SMALL_LETTER_SHARP_S);
3740 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3741 ANYOF_BITMAP_SET(data->start_class, 's');
3742 ANYOF_BITMAP_SET(data->start_class, 'S');
3747 else if (uc >= 0x100) {
3749 for (i = 0; i < 256; i++){
3750 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3751 ANYOF_BITMAP_SET(data->start_class, i);
3756 else if (flags & SCF_DO_STCLASS_OR) {
3757 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3758 /* false positive possible if the class is case-folded.
3759 Assume that the locale settings are the same... */
3761 ANYOF_BITMAP_SET(data->start_class, uc);
3762 if (OP(scan) != EXACTFL) {
3764 /* And set the other member of the fold pair, but
3765 * can't do that in locale because not known until
3767 ANYOF_BITMAP_SET(data->start_class,
3768 PL_fold_latin1[uc]);
3770 /* All folds except under /iaa that include s, S,
3771 * and sharp_s also may include the others */
3772 if (OP(scan) != EXACTFA) {
3773 if (uc == 's' || uc == 'S') {
3774 ANYOF_BITMAP_SET(data->start_class,
3775 LATIN_SMALL_LETTER_SHARP_S);
3777 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3778 ANYOF_BITMAP_SET(data->start_class, 's');
3779 ANYOF_BITMAP_SET(data->start_class, 'S');
3784 CLEAR_SSC_EOS(data->start_class);
3786 cl_and(data->start_class, and_withp);
3788 flags &= ~SCF_DO_STCLASS;
3790 else if (REGNODE_VARIES(OP(scan))) {
3791 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3792 I32 f = flags, pos_before = 0;
3793 regnode * const oscan = scan;
3794 struct regnode_charclass_class this_class;
3795 struct regnode_charclass_class *oclass = NULL;
3796 I32 next_is_eval = 0;
3798 switch (PL_regkind[OP(scan)]) {
3799 case WHILEM: /* End of (?:...)* . */
3800 scan = NEXTOPER(scan);
3803 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3804 next = NEXTOPER(scan);
3805 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3807 maxcount = REG_INFTY;
3808 next = regnext(scan);
3809 scan = NEXTOPER(scan);
3813 if (flags & SCF_DO_SUBSTR)
3818 if (flags & SCF_DO_STCLASS) {
3820 maxcount = REG_INFTY;
3821 next = regnext(scan);
3822 scan = NEXTOPER(scan);
3825 is_inf = is_inf_internal = 1;
3826 scan = regnext(scan);
3827 if (flags & SCF_DO_SUBSTR) {
3828 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3829 data->longest = &(data->longest_float);
3831 goto optimize_curly_tail;
3833 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3834 && (scan->flags == stopparen))
3839 mincount = ARG1(scan);
3840 maxcount = ARG2(scan);
3842 next = regnext(scan);
3843 if (OP(scan) == CURLYX) {
3844 I32 lp = (data ? *(data->last_closep) : 0);
3845 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3847 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3848 next_is_eval = (OP(scan) == EVAL);
3850 if (flags & SCF_DO_SUBSTR) {
3851 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3852 pos_before = data->pos_min;
3856 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3858 data->flags |= SF_IS_INF;
3860 if (flags & SCF_DO_STCLASS) {
3861 cl_init(pRExC_state, &this_class);
3862 oclass = data->start_class;
3863 data->start_class = &this_class;
3864 f |= SCF_DO_STCLASS_AND;
3865 f &= ~SCF_DO_STCLASS_OR;
3867 /* Exclude from super-linear cache processing any {n,m}
3868 regops for which the combination of input pos and regex
3869 pos is not enough information to determine if a match
3872 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3873 regex pos at the \s*, the prospects for a match depend not
3874 only on the input position but also on how many (bar\s*)
3875 repeats into the {4,8} we are. */
3876 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3877 f &= ~SCF_WHILEM_VISITED_POS;
3879 /* This will finish on WHILEM, setting scan, or on NULL: */
3880 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3881 last, data, stopparen, recursed, NULL,
3883 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3885 if (flags & SCF_DO_STCLASS)
3886 data->start_class = oclass;
3887 if (mincount == 0 || minnext == 0) {
3888 if (flags & SCF_DO_STCLASS_OR) {
3889 cl_or(pRExC_state, data->start_class, &this_class);
3891 else if (flags & SCF_DO_STCLASS_AND) {
3892 /* Switch to OR mode: cache the old value of
3893 * data->start_class */
3895 StructCopy(data->start_class, and_withp,
3896 struct regnode_charclass_class);
3897 flags &= ~SCF_DO_STCLASS_AND;
3898 StructCopy(&this_class, data->start_class,
3899 struct regnode_charclass_class);
3900 flags |= SCF_DO_STCLASS_OR;
3901 SET_SSC_EOS(data->start_class);
3903 } else { /* Non-zero len */
3904 if (flags & SCF_DO_STCLASS_OR) {
3905 cl_or(pRExC_state, data->start_class, &this_class);
3906 cl_and(data->start_class, and_withp);
3908 else if (flags & SCF_DO_STCLASS_AND)
3909 cl_and(data->start_class, &this_class);
3910 flags &= ~SCF_DO_STCLASS;
3912 if (!scan) /* It was not CURLYX, but CURLY. */
3914 if (!(flags & SCF_TRIE_DOING_RESTUDY)
3915 /* ? quantifier ok, except for (?{ ... }) */
3916 && (next_is_eval || !(mincount == 0 && maxcount == 1))
3917 && (minnext == 0) && (deltanext == 0)
3918 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3919 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3921 /* Fatal warnings may leak the regexp without this: */
3922 SAVEFREESV(RExC_rx_sv);
3923 ckWARNreg(RExC_parse,
3924 "Quantifier unexpected on zero-length expression");
3925 (void)ReREFCNT_inc(RExC_rx_sv);
3928 min += minnext * mincount;
3929 is_inf_internal |= deltanext == I32_MAX
3930 || (maxcount == REG_INFTY && minnext + deltanext > 0);
3931 is_inf |= is_inf_internal;
3935 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3937 /* Try powerful optimization CURLYX => CURLYN. */
3938 if ( OP(oscan) == CURLYX && data
3939 && data->flags & SF_IN_PAR
3940 && !(data->flags & SF_HAS_EVAL)
3941 && !deltanext && minnext == 1 ) {
3942 /* Try to optimize to CURLYN. */
3943 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3944 regnode * const nxt1 = nxt;
3951 if (!REGNODE_SIMPLE(OP(nxt))
3952 && !(PL_regkind[OP(nxt)] == EXACT
3953 && STR_LEN(nxt) == 1))
3959 if (OP(nxt) != CLOSE)
3961 if (RExC_open_parens) {
3962 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3963 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3965 /* Now we know that nxt2 is the only contents: */
3966 oscan->flags = (U8)ARG(nxt);
3968 OP(nxt1) = NOTHING; /* was OPEN. */
3971 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3972 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3973 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3974 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3975 OP(nxt + 1) = OPTIMIZED; /* was count. */
3976 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3981 /* Try optimization CURLYX => CURLYM. */
3982 if ( OP(oscan) == CURLYX && data
3983 && !(data->flags & SF_HAS_PAR)
3984 && !(data->flags & SF_HAS_EVAL)
3985 && !deltanext /* atom is fixed width */
3986 && minnext != 0 /* CURLYM can't handle zero width */
3987 && ! (RExC_seen & REG_SEEN_EXACTF_SHARP_S) /* Nor \xDF */
3989 /* XXXX How to optimize if data == 0? */
3990 /* Optimize to a simpler form. */
3991 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3995 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3996 && (OP(nxt2) != WHILEM))
3998 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3999 /* Need to optimize away parenths. */
4000 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
4001 /* Set the parenth number. */
4002 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
4004 oscan->flags = (U8)ARG(nxt);
4005 if (RExC_open_parens) {
4006 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4007 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4009 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4010 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4013 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4014 OP(nxt + 1) = OPTIMIZED; /* was count. */
4015 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4016 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4019 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4020 regnode *nnxt = regnext(nxt1);
4022 if (reg_off_by_arg[OP(nxt1)])
4023 ARG_SET(nxt1, nxt2 - nxt1);
4024 else if (nxt2 - nxt1 < U16_MAX)
4025 NEXT_OFF(nxt1) = nxt2 - nxt1;
4027 OP(nxt) = NOTHING; /* Cannot beautify */
4032 /* Optimize again: */
4033 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4034 NULL, stopparen, recursed, NULL, 0,depth+1);
4039 else if ((OP(oscan) == CURLYX)
4040 && (flags & SCF_WHILEM_VISITED_POS)
4041 /* See the comment on a similar expression above.
4042 However, this time it's not a subexpression
4043 we care about, but the expression itself. */
4044 && (maxcount == REG_INFTY)
4045 && data && ++data->whilem_c < 16) {
4046 /* This stays as CURLYX, we can put the count/of pair. */
4047 /* Find WHILEM (as in regexec.c) */
4048 regnode *nxt = oscan + NEXT_OFF(oscan);
4050 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4052 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4053 | (RExC_whilem_seen << 4)); /* On WHILEM */
4055 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4057 if (flags & SCF_DO_SUBSTR) {
4058 SV *last_str = NULL;
4059 int counted = mincount != 0;
4061 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
4062 #if defined(SPARC64_GCC_WORKAROUND)
4065 const char *s = NULL;
4068 if (pos_before >= data->last_start_min)
4071 b = data->last_start_min;
4074 s = SvPV_const(data->last_found, l);
4075 old = b - data->last_start_min;
4078 I32 b = pos_before >= data->last_start_min
4079 ? pos_before : data->last_start_min;
4081 const char * const s = SvPV_const(data->last_found, l);
4082 I32 old = b - data->last_start_min;
4086 old = utf8_hop((U8*)s, old) - (U8*)s;
4088 /* Get the added string: */
4089 last_str = newSVpvn_utf8(s + old, l, UTF);
4090 if (deltanext == 0 && pos_before == b) {
4091 /* What was added is a constant string */
4093 SvGROW(last_str, (mincount * l) + 1);
4094 repeatcpy(SvPVX(last_str) + l,
4095 SvPVX_const(last_str), l, mincount - 1);
4096 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4097 /* Add additional parts. */
4098 SvCUR_set(data->last_found,
4099 SvCUR(data->last_found) - l);
4100 sv_catsv(data->last_found, last_str);
4102 SV * sv = data->last_found;
4104 SvUTF8(sv) && SvMAGICAL(sv) ?
4105 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4106 if (mg && mg->mg_len >= 0)
4107 mg->mg_len += CHR_SVLEN(last_str) - l;
4109 data->last_end += l * (mincount - 1);
4112 /* start offset must point into the last copy */
4113 data->last_start_min += minnext * (mincount - 1);
4114 data->last_start_max += is_inf ? I32_MAX
4115 : (maxcount - 1) * (minnext + data->pos_delta);
4118 /* It is counted once already... */
4119 data->pos_min += minnext * (mincount - counted);
4121 PerlIO_printf(Perl_debug_log, "counted=%d deltanext=%d I32_MAX=%d minnext=%d maxcount=%d mincount=%d\n",
4122 counted, deltanext, I32_MAX, minnext, maxcount, mincount);
4123 if (deltanext != I32_MAX)
4124 PerlIO_printf(Perl_debug_log, "LHS=%d RHS=%d\n", -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount, I32_MAX - data->pos_delta);
4126 if (deltanext == I32_MAX || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= I32_MAX - data->pos_delta)
4127 data->pos_delta = I32_MAX;
4129 data->pos_delta += - counted * deltanext +
4130 (minnext + deltanext) * maxcount - minnext * mincount;
4131 if (mincount != maxcount) {
4132 /* Cannot extend fixed substrings found inside
4134 SCAN_COMMIT(pRExC_state,data,minlenp);
4135 if (mincount && last_str) {
4136 SV * const sv = data->last_found;
4137 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4138 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4142 sv_setsv(sv, last_str);
4143 data->last_end = data->pos_min;
4144 data->last_start_min =
4145 data->pos_min - CHR_SVLEN(last_str);
4146 data->last_start_max = is_inf
4148 : data->pos_min + data->pos_delta
4149 - CHR_SVLEN(last_str);
4151 data->longest = &(data->longest_float);
4153 SvREFCNT_dec(last_str);
4155 if (data && (fl & SF_HAS_EVAL))
4156 data->flags |= SF_HAS_EVAL;
4157 optimize_curly_tail:
4158 if (OP(oscan) != CURLYX) {
4159 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4161 NEXT_OFF(oscan) += NEXT_OFF(next);
4164 default: /* REF, and CLUMP only? */
4165 if (flags & SCF_DO_SUBSTR) {
4166 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4167 data->longest = &(data->longest_float);
4169 is_inf = is_inf_internal = 1;
4170 if (flags & SCF_DO_STCLASS_OR)
4171 cl_anything(pRExC_state, data->start_class);
4172 flags &= ~SCF_DO_STCLASS;
4176 else if (OP(scan) == LNBREAK) {
4177 if (flags & SCF_DO_STCLASS) {
4179 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4180 if (flags & SCF_DO_STCLASS_AND) {
4181 for (value = 0; value < 256; value++)
4182 if (!is_VERTWS_cp(value))
4183 ANYOF_BITMAP_CLEAR(data->start_class, value);
4186 for (value = 0; value < 256; value++)
4187 if (is_VERTWS_cp(value))
4188 ANYOF_BITMAP_SET(data->start_class, value);
4190 if (flags & SCF_DO_STCLASS_OR)
4191 cl_and(data->start_class, and_withp);
4192 flags &= ~SCF_DO_STCLASS;
4195 delta++; /* Because of the 2 char string cr-lf */
4196 if (flags & SCF_DO_SUBSTR) {
4197 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4199 data->pos_delta += 1;
4200 data->longest = &(data->longest_float);
4203 else if (REGNODE_SIMPLE(OP(scan))) {
4206 if (flags & SCF_DO_SUBSTR) {
4207 SCAN_COMMIT(pRExC_state,data,minlenp);
4211 if (flags & SCF_DO_STCLASS) {
4213 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4215 /* Some of the logic below assumes that switching
4216 locale on will only add false positives. */
4217 switch (PL_regkind[OP(scan)]) {
4223 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan));
4226 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4227 cl_anything(pRExC_state, data->start_class);
4230 if (OP(scan) == SANY)
4232 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
4233 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
4234 || ANYOF_CLASS_TEST_ANY_SET(data->start_class));
4235 cl_anything(pRExC_state, data->start_class);
4237 if (flags & SCF_DO_STCLASS_AND || !value)
4238 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
4241 if (flags & SCF_DO_STCLASS_AND)
4242 cl_and(data->start_class,
4243 (struct regnode_charclass_class*)scan);
4245 cl_or(pRExC_state, data->start_class,
4246 (struct regnode_charclass_class*)scan);
4254 classnum = FLAGS(scan);
4255 if (flags & SCF_DO_STCLASS_AND) {
4256 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4257 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum) + 1);
4258 for (value = 0; value < loop_max; value++) {
4259 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4260 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4266 if (data->start_class->flags & ANYOF_LOCALE) {
4267 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum));
4271 /* Even if under locale, set the bits for non-locale
4272 * in case it isn't a true locale-node. This will
4273 * create false positives if it truly is locale */
4274 for (value = 0; value < loop_max; value++) {
4275 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4276 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4288 classnum = FLAGS(scan);
4289 if (flags & SCF_DO_STCLASS_AND) {
4290 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4291 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum));
4292 for (value = 0; value < loop_max; value++) {
4293 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4294 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4300 if (data->start_class->flags & ANYOF_LOCALE) {
4301 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum) + 1);
4305 /* Even if under locale, set the bits for non-locale in
4306 * case it isn't a true locale-node. This will create
4307 * false positives if it truly is locale */
4308 for (value = 0; value < loop_max; value++) {
4309 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4310 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4313 if (PL_regkind[OP(scan)] == NPOSIXD) {
4314 data->start_class->flags |= ANYOF_NON_UTF8_LATIN1_ALL;
4320 if (flags & SCF_DO_STCLASS_OR)
4321 cl_and(data->start_class, and_withp);
4322 flags &= ~SCF_DO_STCLASS;
4325 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
4326 data->flags |= (OP(scan) == MEOL
4329 SCAN_COMMIT(pRExC_state, data, minlenp);
4332 else if ( PL_regkind[OP(scan)] == BRANCHJ
4333 /* Lookbehind, or need to calculate parens/evals/stclass: */
4334 && (scan->flags || data || (flags & SCF_DO_STCLASS))
4335 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
4336 if ( OP(scan) == UNLESSM &&
4338 OP(NEXTOPER(NEXTOPER(scan))) == NOTHING &&
4339 OP(regnext(NEXTOPER(NEXTOPER(scan)))) == SUCCEED
4342 regnode *upto= regnext(scan);
4344 SV * const mysv_val=sv_newmortal();
4345 DEBUG_STUDYDATA("OPFAIL",data,depth);
4347 /*DEBUG_PARSE_MSG("opfail");*/
4348 regprop(RExC_rx, mysv_val, upto);
4349 PerlIO_printf(Perl_debug_log, "~ replace with OPFAIL pointed at %s (%"IVdf") offset %"IVdf"\n",
4350 SvPV_nolen_const(mysv_val),
4351 (IV)REG_NODE_NUM(upto),
4356 NEXT_OFF(scan) = upto - scan;
4357 for (opt= scan + 1; opt < upto ; opt++)
4358 OP(opt) = OPTIMIZED;
4362 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4363 || OP(scan) == UNLESSM )
4365 /* Negative Lookahead/lookbehind
4366 In this case we can't do fixed string optimisation.
4369 I32 deltanext, minnext, fake = 0;
4371 struct regnode_charclass_class intrnl;
4374 data_fake.flags = 0;
4376 data_fake.whilem_c = data->whilem_c;
4377 data_fake.last_closep = data->last_closep;
4380 data_fake.last_closep = &fake;
4381 data_fake.pos_delta = delta;
4382 if ( flags & SCF_DO_STCLASS && !scan->flags
4383 && OP(scan) == IFMATCH ) { /* Lookahead */
4384 cl_init(pRExC_state, &intrnl);
4385 data_fake.start_class = &intrnl;
4386 f |= SCF_DO_STCLASS_AND;
4388 if (flags & SCF_WHILEM_VISITED_POS)
4389 f |= SCF_WHILEM_VISITED_POS;
4390 next = regnext(scan);
4391 nscan = NEXTOPER(NEXTOPER(scan));
4392 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
4393 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
4396 FAIL("Variable length lookbehind not implemented");
4398 else if (minnext > (I32)U8_MAX) {
4399 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4401 scan->flags = (U8)minnext;
4404 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4406 if (data_fake.flags & SF_HAS_EVAL)
4407 data->flags |= SF_HAS_EVAL;
4408 data->whilem_c = data_fake.whilem_c;
4410 if (f & SCF_DO_STCLASS_AND) {
4411 if (flags & SCF_DO_STCLASS_OR) {
4412 /* OR before, AND after: ideally we would recurse with
4413 * data_fake to get the AND applied by study of the
4414 * remainder of the pattern, and then derecurse;
4415 * *** HACK *** for now just treat as "no information".
4416 * See [perl #56690].
4418 cl_init(pRExC_state, data->start_class);
4420 /* AND before and after: combine and continue */
4421 const int was = TEST_SSC_EOS(data->start_class);
4423 cl_and(data->start_class, &intrnl);
4425 SET_SSC_EOS(data->start_class);
4429 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4431 /* Positive Lookahead/lookbehind
4432 In this case we can do fixed string optimisation,
4433 but we must be careful about it. Note in the case of
4434 lookbehind the positions will be offset by the minimum
4435 length of the pattern, something we won't know about
4436 until after the recurse.
4438 I32 deltanext, fake = 0;
4440 struct regnode_charclass_class intrnl;
4442 /* We use SAVEFREEPV so that when the full compile
4443 is finished perl will clean up the allocated
4444 minlens when it's all done. This way we don't
4445 have to worry about freeing them when we know
4446 they wont be used, which would be a pain.
4449 Newx( minnextp, 1, I32 );
4450 SAVEFREEPV(minnextp);
4453 StructCopy(data, &data_fake, scan_data_t);
4454 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
4457 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
4458 data_fake.last_found=newSVsv(data->last_found);
4462 data_fake.last_closep = &fake;
4463 data_fake.flags = 0;
4464 data_fake.pos_delta = delta;
4466 data_fake.flags |= SF_IS_INF;
4467 if ( flags & SCF_DO_STCLASS && !scan->flags
4468 && OP(scan) == IFMATCH ) { /* Lookahead */
4469 cl_init(pRExC_state, &intrnl);
4470 data_fake.start_class = &intrnl;
4471 f |= SCF_DO_STCLASS_AND;
4473 if (flags & SCF_WHILEM_VISITED_POS)
4474 f |= SCF_WHILEM_VISITED_POS;
4475 next = regnext(scan);
4476 nscan = NEXTOPER(NEXTOPER(scan));
4478 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
4479 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
4482 FAIL("Variable length lookbehind not implemented");
4484 else if (*minnextp > (I32)U8_MAX) {
4485 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4487 scan->flags = (U8)*minnextp;
4492 if (f & SCF_DO_STCLASS_AND) {
4493 const int was = TEST_SSC_EOS(data.start_class);
4495 cl_and(data->start_class, &intrnl);
4497 SET_SSC_EOS(data->start_class);
4500 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4502 if (data_fake.flags & SF_HAS_EVAL)
4503 data->flags |= SF_HAS_EVAL;
4504 data->whilem_c = data_fake.whilem_c;
4505 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
4506 if (RExC_rx->minlen<*minnextp)
4507 RExC_rx->minlen=*minnextp;
4508 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
4509 SvREFCNT_dec_NN(data_fake.last_found);
4511 if ( data_fake.minlen_fixed != minlenp )
4513 data->offset_fixed= data_fake.offset_fixed;
4514 data->minlen_fixed= data_fake.minlen_fixed;
4515 data->lookbehind_fixed+= scan->flags;
4517 if ( data_fake.minlen_float != minlenp )
4519 data->minlen_float= data_fake.minlen_float;
4520 data->offset_float_min=data_fake.offset_float_min;
4521 data->offset_float_max=data_fake.offset_float_max;
4522 data->lookbehind_float+= scan->flags;
4529 else if (OP(scan) == OPEN) {
4530 if (stopparen != (I32)ARG(scan))
4533 else if (OP(scan) == CLOSE) {
4534 if (stopparen == (I32)ARG(scan)) {
4537 if ((I32)ARG(scan) == is_par) {
4538 next = regnext(scan);
4540 if ( next && (OP(next) != WHILEM) && next < last)
4541 is_par = 0; /* Disable optimization */
4544 *(data->last_closep) = ARG(scan);
4546 else if (OP(scan) == EVAL) {
4548 data->flags |= SF_HAS_EVAL;
4550 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
4551 if (flags & SCF_DO_SUBSTR) {
4552 SCAN_COMMIT(pRExC_state,data,minlenp);
4553 flags &= ~SCF_DO_SUBSTR;
4555 if (data && OP(scan)==ACCEPT) {
4556 data->flags |= SCF_SEEN_ACCEPT;
4561 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
4563 if (flags & SCF_DO_SUBSTR) {
4564 SCAN_COMMIT(pRExC_state,data,minlenp);
4565 data->longest = &(data->longest_float);
4567 is_inf = is_inf_internal = 1;
4568 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4569 cl_anything(pRExC_state, data->start_class);
4570 flags &= ~SCF_DO_STCLASS;
4572 else if (OP(scan) == GPOS) {
4573 if (!(RExC_rx->extflags & RXf_GPOS_FLOAT) &&
4574 !(delta || is_inf || (data && data->pos_delta)))
4576 if (!(RExC_rx->extflags & RXf_ANCH) && (flags & SCF_DO_SUBSTR))
4577 RExC_rx->extflags |= RXf_ANCH_GPOS;
4578 if (RExC_rx->gofs < (U32)min)
4579 RExC_rx->gofs = min;
4581 RExC_rx->extflags |= RXf_GPOS_FLOAT;
4585 #ifdef TRIE_STUDY_OPT
4586 #ifdef FULL_TRIE_STUDY
4587 else if (PL_regkind[OP(scan)] == TRIE) {
4588 /* NOTE - There is similar code to this block above for handling
4589 BRANCH nodes on the initial study. If you change stuff here
4591 regnode *trie_node= scan;
4592 regnode *tail= regnext(scan);
4593 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4594 I32 max1 = 0, min1 = I32_MAX;
4595 struct regnode_charclass_class accum;
4597 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
4598 SCAN_COMMIT(pRExC_state, data,minlenp); /* Cannot merge strings after this. */
4599 if (flags & SCF_DO_STCLASS)
4600 cl_init_zero(pRExC_state, &accum);
4606 const regnode *nextbranch= NULL;
4609 for ( word=1 ; word <= trie->wordcount ; word++)
4611 I32 deltanext=0, minnext=0, f = 0, fake;
4612 struct regnode_charclass_class this_class;
4614 data_fake.flags = 0;
4616 data_fake.whilem_c = data->whilem_c;
4617 data_fake.last_closep = data->last_closep;
4620 data_fake.last_closep = &fake;
4621 data_fake.pos_delta = delta;
4622 if (flags & SCF_DO_STCLASS) {
4623 cl_init(pRExC_state, &this_class);
4624 data_fake.start_class = &this_class;
4625 f = SCF_DO_STCLASS_AND;
4627 if (flags & SCF_WHILEM_VISITED_POS)
4628 f |= SCF_WHILEM_VISITED_POS;
4630 if (trie->jump[word]) {
4632 nextbranch = trie_node + trie->jump[0];
4633 scan= trie_node + trie->jump[word];
4634 /* We go from the jump point to the branch that follows
4635 it. Note this means we need the vestigal unused branches
4636 even though they arent otherwise used.
4638 minnext = study_chunk(pRExC_state, &scan, minlenp,
4639 &deltanext, (regnode *)nextbranch, &data_fake,
4640 stopparen, recursed, NULL, f,depth+1);
4642 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
4643 nextbranch= regnext((regnode*)nextbranch);
4645 if (min1 > (I32)(minnext + trie->minlen))
4646 min1 = minnext + trie->minlen;
4647 if (deltanext == I32_MAX) {
4648 is_inf = is_inf_internal = 1;
4650 } else if (max1 < (I32)(minnext + deltanext + trie->maxlen))
4651 max1 = minnext + deltanext + trie->maxlen;
4653 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4655 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4656 if ( stopmin > min + min1)
4657 stopmin = min + min1;
4658 flags &= ~SCF_DO_SUBSTR;
4660 data->flags |= SCF_SEEN_ACCEPT;
4663 if (data_fake.flags & SF_HAS_EVAL)
4664 data->flags |= SF_HAS_EVAL;
4665 data->whilem_c = data_fake.whilem_c;
4667 if (flags & SCF_DO_STCLASS)
4668 cl_or(pRExC_state, &accum, &this_class);
4671 if (flags & SCF_DO_SUBSTR) {
4672 data->pos_min += min1;
4673 data->pos_delta += max1 - min1;
4674 if (max1 != min1 || is_inf)
4675 data->longest = &(data->longest_float);
4678 delta += max1 - min1;
4679 if (flags & SCF_DO_STCLASS_OR) {
4680 cl_or(pRExC_state, data->start_class, &accum);
4682 cl_and(data->start_class, and_withp);
4683 flags &= ~SCF_DO_STCLASS;
4686 else if (flags & SCF_DO_STCLASS_AND) {
4688 cl_and(data->start_class, &accum);
4689 flags &= ~SCF_DO_STCLASS;
4692 /* Switch to OR mode: cache the old value of
4693 * data->start_class */
4695 StructCopy(data->start_class, and_withp,
4696 struct regnode_charclass_class);
4697 flags &= ~SCF_DO_STCLASS_AND;
4698 StructCopy(&accum, data->start_class,
4699 struct regnode_charclass_class);
4700 flags |= SCF_DO_STCLASS_OR;
4701 SET_SSC_EOS(data->start_class);
4708 else if (PL_regkind[OP(scan)] == TRIE) {
4709 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4712 min += trie->minlen;
4713 delta += (trie->maxlen - trie->minlen);
4714 flags &= ~SCF_DO_STCLASS; /* xxx */
4715 if (flags & SCF_DO_SUBSTR) {
4716 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4717 data->pos_min += trie->minlen;
4718 data->pos_delta += (trie->maxlen - trie->minlen);
4719 if (trie->maxlen != trie->minlen)
4720 data->longest = &(data->longest_float);
4722 if (trie->jump) /* no more substrings -- for now /grr*/
4723 flags &= ~SCF_DO_SUBSTR;
4725 #endif /* old or new */
4726 #endif /* TRIE_STUDY_OPT */
4728 /* Else: zero-length, ignore. */
4729 scan = regnext(scan);
4734 stopparen = frame->stop;
4735 frame = frame->prev;
4736 goto fake_study_recurse;
4741 DEBUG_STUDYDATA("pre-fin:",data,depth);
4744 *deltap = is_inf_internal ? I32_MAX : delta;
4745 if (flags & SCF_DO_SUBSTR && is_inf)
4746 data->pos_delta = I32_MAX - data->pos_min;
4747 if (is_par > (I32)U8_MAX)
4749 if (is_par && pars==1 && data) {
4750 data->flags |= SF_IN_PAR;
4751 data->flags &= ~SF_HAS_PAR;
4753 else if (pars && data) {
4754 data->flags |= SF_HAS_PAR;
4755 data->flags &= ~SF_IN_PAR;
4757 if (flags & SCF_DO_STCLASS_OR)
4758 cl_and(data->start_class, and_withp);
4759 if (flags & SCF_TRIE_RESTUDY)
4760 data->flags |= SCF_TRIE_RESTUDY;
4762 DEBUG_STUDYDATA("post-fin:",data,depth);
4764 return min < stopmin ? min : stopmin;
4768 S_add_data(RExC_state_t *pRExC_state, U32 n, const char *s)
4770 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
4772 PERL_ARGS_ASSERT_ADD_DATA;
4774 Renewc(RExC_rxi->data,
4775 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
4776 char, struct reg_data);
4778 Renew(RExC_rxi->data->what, count + n, U8);
4780 Newx(RExC_rxi->data->what, n, U8);
4781 RExC_rxi->data->count = count + n;
4782 Copy(s, RExC_rxi->data->what + count, n, U8);
4786 /*XXX: todo make this not included in a non debugging perl */
4787 #ifndef PERL_IN_XSUB_RE
4789 Perl_reginitcolors(pTHX)
4792 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
4794 char *t = savepv(s);
4798 t = strchr(t, '\t');
4804 PL_colors[i] = t = (char *)"";
4809 PL_colors[i++] = (char *)"";
4816 #ifdef TRIE_STUDY_OPT
4817 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
4820 (data.flags & SCF_TRIE_RESTUDY) \
4828 #define CHECK_RESTUDY_GOTO_butfirst
4832 * pregcomp - compile a regular expression into internal code
4834 * Decides which engine's compiler to call based on the hint currently in
4838 #ifndef PERL_IN_XSUB_RE
4840 /* return the currently in-scope regex engine (or the default if none) */
4842 regexp_engine const *
4843 Perl_current_re_engine(pTHX)
4847 if (IN_PERL_COMPILETIME) {
4848 HV * const table = GvHV(PL_hintgv);
4852 return &reh_regexp_engine;
4853 ptr = hv_fetchs(table, "regcomp", FALSE);
4854 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
4855 return &reh_regexp_engine;
4856 return INT2PTR(regexp_engine*,SvIV(*ptr));
4860 if (!PL_curcop->cop_hints_hash)
4861 return &reh_regexp_engine;
4862 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
4863 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
4864 return &reh_regexp_engine;
4865 return INT2PTR(regexp_engine*,SvIV(ptr));
4871 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
4874 regexp_engine const *eng = current_re_engine();
4875 GET_RE_DEBUG_FLAGS_DECL;
4877 PERL_ARGS_ASSERT_PREGCOMP;
4879 /* Dispatch a request to compile a regexp to correct regexp engine. */
4881 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
4884 return CALLREGCOMP_ENG(eng, pattern, flags);
4888 /* public(ish) entry point for the perl core's own regex compiling code.
4889 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
4890 * pattern rather than a list of OPs, and uses the internal engine rather
4891 * than the current one */
4894 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
4896 SV *pat = pattern; /* defeat constness! */
4897 PERL_ARGS_ASSERT_RE_COMPILE;
4898 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
4899 #ifdef PERL_IN_XSUB_RE
4904 NULL, NULL, rx_flags, 0);
4908 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
4909 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
4910 * point to the realloced string and length.
4912 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
4916 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
4917 char **pat_p, STRLEN *plen_p, int num_code_blocks)
4919 U8 *const src = (U8*)*pat_p;
4922 STRLEN s = 0, d = 0;
4924 GET_RE_DEBUG_FLAGS_DECL;
4926 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
4927 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
4929 Newx(dst, *plen_p * 2 + 1, U8);
4931 while (s < *plen_p) {
4932 const UV uv = NATIVE_TO_ASCII(src[s]);
4933 if (UNI_IS_INVARIANT(uv))
4934 dst[d] = (U8)UTF_TO_NATIVE(uv);
4936 dst[d++] = (U8)UTF8_EIGHT_BIT_HI(uv);
4937 dst[d] = (U8)UTF8_EIGHT_BIT_LO(uv);
4939 if (n < num_code_blocks) {
4940 if (!do_end && pRExC_state->code_blocks[n].start == s) {
4941 pRExC_state->code_blocks[n].start = d;
4942 assert(dst[d] == '(');
4945 else if (do_end && pRExC_state->code_blocks[n].end == s) {
4946 pRExC_state->code_blocks[n].end = d;
4947 assert(dst[d] == ')');
4957 *pat_p = (char*) dst;
4959 RExC_orig_utf8 = RExC_utf8 = 1;
4964 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
4965 * while recording any code block indices, and handling overloading,
4966 * nested qr// objects etc. If pat is null, it will allocate a new
4967 * string, or just return the first arg, if there's only one.
4969 * Returns the malloced/updated pat.
4970 * patternp and pat_count is the array of SVs to be concatted;
4971 * oplist is the optional list of ops that generated the SVs;
4972 * recompile_p is a pointer to a boolean that will be set if
4973 * the regex will need to be recompiled.
4974 * delim, if non-null is an SV that will be inserted between each element
4978 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
4979 SV *pat, SV ** const patternp, int pat_count,
4980 OP *oplist, bool *recompile_p, SV *delim)
4984 bool use_delim = FALSE;
4985 bool alloced = FALSE;
4987 /* if we know we have at least two args, create an empty string,
4988 * then concatenate args to that. For no args, return an empty string */
4989 if (!pat && pat_count != 1) {
4990 pat = newSVpvn("", 0);
4995 for (svp = patternp; svp < patternp + pat_count; svp++) {
4998 STRLEN orig_patlen = 0;
5000 SV *msv = use_delim ? delim : *svp;
5002 /* if we've got a delimiter, we go round the loop twice for each
5003 * svp slot (except the last), using the delimiter the second
5012 if (SvTYPE(msv) == SVt_PVAV) {
5013 /* we've encountered an interpolated array within
5014 * the pattern, e.g. /...@a..../. Expand the list of elements,
5015 * then recursively append elements.
5016 * The code in this block is based on S_pushav() */
5018 AV *const av = (AV*)msv;
5019 const I32 maxarg = AvFILL(av) + 1;
5023 assert(oplist->op_type == OP_PADAV
5024 || oplist->op_type == OP_RV2AV);
5025 oplist = oplist->op_sibling;;
5028 if (SvRMAGICAL(av)) {
5031 Newx(array, maxarg, SV*);
5033 for (i=0; i < (U32)maxarg; i++) {
5034 SV ** const svp = av_fetch(av, i, FALSE);
5035 array[i] = svp ? *svp : &PL_sv_undef;
5039 array = AvARRAY(av);
5041 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5042 array, maxarg, NULL, recompile_p,
5044 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5050 /* we make the assumption here that each op in the list of
5051 * op_siblings maps to one SV pushed onto the stack,
5052 * except for code blocks, with have both an OP_NULL and
5054 * This allows us to match up the list of SVs against the
5055 * list of OPs to find the next code block.
5057 * Note that PUSHMARK PADSV PADSV ..
5059 * PADRANGE PADSV PADSV ..
5060 * so the alignment still works. */
5063 if (oplist->op_type == OP_NULL
5064 && (oplist->op_flags & OPf_SPECIAL))
5066 assert(n < pRExC_state->num_code_blocks);
5067 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5068 pRExC_state->code_blocks[n].block = oplist;
5069 pRExC_state->code_blocks[n].src_regex = NULL;
5072 oplist = oplist->op_sibling; /* skip CONST */
5075 oplist = oplist->op_sibling;;
5078 /* apply magic and QR overloading to arg */
5081 if (SvROK(msv) && SvAMAGIC(msv)) {
5082 SV *sv = AMG_CALLunary(msv, regexp_amg);
5086 if (SvTYPE(sv) != SVt_REGEXP)
5087 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5092 /* try concatenation overload ... */
5093 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5094 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5097 /* overloading involved: all bets are off over literal
5098 * code. Pretend we haven't seen it */
5099 pRExC_state->num_code_blocks -= n;
5103 /* ... or failing that, try "" overload */
5104 while (SvAMAGIC(msv)
5105 && (sv = AMG_CALLunary(msv, string_amg))
5109 && SvRV(msv) == SvRV(sv))
5114 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5118 /* this is a partially unrolled
5119 * sv_catsv_nomg(pat, msv);
5120 * that allows us to adjust code block indices if
5123 char *dst = SvPV_force_nomg(pat, dlen);
5125 if (SvUTF8(msv) && !SvUTF8(pat)) {
5126 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
5127 sv_setpvn(pat, dst, dlen);
5130 sv_catsv_nomg(pat, msv);
5137 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5140 /* extract any code blocks within any embedded qr//'s */
5141 if (rx && SvTYPE(rx) == SVt_REGEXP
5142 && RX_ENGINE((REGEXP*)rx)->op_comp)
5145 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5146 if (ri->num_code_blocks) {
5148 /* the presence of an embedded qr// with code means
5149 * we should always recompile: the text of the
5150 * qr// may not have changed, but it may be a
5151 * different closure than last time */
5153 Renew(pRExC_state->code_blocks,
5154 pRExC_state->num_code_blocks + ri->num_code_blocks,
5155 struct reg_code_block);
5156 pRExC_state->num_code_blocks += ri->num_code_blocks;
5158 for (i=0; i < ri->num_code_blocks; i++) {
5159 struct reg_code_block *src, *dst;
5160 STRLEN offset = orig_patlen
5161 + ReANY((REGEXP *)rx)->pre_prefix;
5162 assert(n < pRExC_state->num_code_blocks);
5163 src = &ri->code_blocks[i];
5164 dst = &pRExC_state->code_blocks[n];
5165 dst->start = src->start + offset;
5166 dst->end = src->end + offset;
5167 dst->block = src->block;
5168 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5177 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
5186 /* see if there are any run-time code blocks in the pattern.
5187 * False positives are allowed */
5190 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5191 char *pat, STRLEN plen)
5196 for (s = 0; s < plen; s++) {
5197 if (n < pRExC_state->num_code_blocks
5198 && s == pRExC_state->code_blocks[n].start)
5200 s = pRExC_state->code_blocks[n].end;
5204 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
5206 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
5208 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
5215 /* Handle run-time code blocks. We will already have compiled any direct
5216 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
5217 * copy of it, but with any literal code blocks blanked out and
5218 * appropriate chars escaped; then feed it into
5220 * eval "qr'modified_pattern'"
5224 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
5228 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
5230 * After eval_sv()-ing that, grab any new code blocks from the returned qr
5231 * and merge them with any code blocks of the original regexp.
5233 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
5234 * instead, just save the qr and return FALSE; this tells our caller that
5235 * the original pattern needs upgrading to utf8.
5239 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5240 char *pat, STRLEN plen)
5244 GET_RE_DEBUG_FLAGS_DECL;
5246 if (pRExC_state->runtime_code_qr) {
5247 /* this is the second time we've been called; this should
5248 * only happen if the main pattern got upgraded to utf8
5249 * during compilation; re-use the qr we compiled first time
5250 * round (which should be utf8 too)
5252 qr = pRExC_state->runtime_code_qr;
5253 pRExC_state->runtime_code_qr = NULL;
5254 assert(RExC_utf8 && SvUTF8(qr));
5260 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
5264 /* determine how many extra chars we need for ' and \ escaping */
5265 for (s = 0; s < plen; s++) {
5266 if (pat[s] == '\'' || pat[s] == '\\')
5270 Newx(newpat, newlen, char);
5272 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
5274 for (s = 0; s < plen; s++) {
5275 if (n < pRExC_state->num_code_blocks
5276 && s == pRExC_state->code_blocks[n].start)
5278 /* blank out literal code block */
5279 assert(pat[s] == '(');
5280 while (s <= pRExC_state->code_blocks[n].end) {
5288 if (pat[s] == '\'' || pat[s] == '\\')
5293 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
5297 PerlIO_printf(Perl_debug_log,
5298 "%sre-parsing pattern for runtime code:%s %s\n",
5299 PL_colors[4],PL_colors[5],newpat);
5302 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5308 PUSHSTACKi(PERLSI_REQUIRE);
5309 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
5310 * parsing qr''; normally only q'' does this. It also alters
5312 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
5313 SvREFCNT_dec_NN(sv);
5318 SV * const errsv = ERRSV;
5319 if (SvTRUE_NN(errsv))
5321 Safefree(pRExC_state->code_blocks);
5322 /* use croak_sv ? */
5323 Perl_croak_nocontext("%s", SvPV_nolen_const(errsv));
5326 assert(SvROK(qr_ref));
5328 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5329 /* the leaving below frees the tmp qr_ref.
5330 * Give qr a life of its own */
5338 if (!RExC_utf8 && SvUTF8(qr)) {
5339 /* first time through; the pattern got upgraded; save the
5340 * qr for the next time through */
5341 assert(!pRExC_state->runtime_code_qr);
5342 pRExC_state->runtime_code_qr = qr;
5347 /* extract any code blocks within the returned qr// */
5350 /* merge the main (r1) and run-time (r2) code blocks into one */
5352 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
5353 struct reg_code_block *new_block, *dst;
5354 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
5357 if (!r2->num_code_blocks) /* we guessed wrong */
5359 SvREFCNT_dec_NN(qr);
5364 r1->num_code_blocks + r2->num_code_blocks,
5365 struct reg_code_block);
5368 while ( i1 < r1->num_code_blocks
5369 || i2 < r2->num_code_blocks)
5371 struct reg_code_block *src;
5374 if (i1 == r1->num_code_blocks) {
5375 src = &r2->code_blocks[i2++];
5378 else if (i2 == r2->num_code_blocks)
5379 src = &r1->code_blocks[i1++];
5380 else if ( r1->code_blocks[i1].start
5381 < r2->code_blocks[i2].start)
5383 src = &r1->code_blocks[i1++];
5384 assert(src->end < r2->code_blocks[i2].start);
5387 assert( r1->code_blocks[i1].start
5388 > r2->code_blocks[i2].start);
5389 src = &r2->code_blocks[i2++];
5391 assert(src->end < r1->code_blocks[i1].start);
5394 assert(pat[src->start] == '(');
5395 assert(pat[src->end] == ')');
5396 dst->start = src->start;
5397 dst->end = src->end;
5398 dst->block = src->block;
5399 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
5403 r1->num_code_blocks += r2->num_code_blocks;
5404 Safefree(r1->code_blocks);
5405 r1->code_blocks = new_block;
5408 SvREFCNT_dec_NN(qr);
5414 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest, SV** rx_utf8, SV** rx_substr, I32* rx_end_shift, I32 lookbehind, I32 offset, I32 *minlen, STRLEN longest_length, bool eol, bool meol)
5416 /* This is the common code for setting up the floating and fixed length
5417 * string data extracted from Perl_re_op_compile() below. Returns a boolean
5418 * as to whether succeeded or not */
5422 if (! (longest_length
5423 || (eol /* Can't have SEOL and MULTI */
5424 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
5426 /* See comments for join_exact for why REG_SEEN_EXACTF_SHARP_S */
5427 || (RExC_seen & REG_SEEN_EXACTF_SHARP_S))
5432 /* copy the information about the longest from the reg_scan_data
5433 over to the program. */
5434 if (SvUTF8(sv_longest)) {
5435 *rx_utf8 = sv_longest;
5438 *rx_substr = sv_longest;
5441 /* end_shift is how many chars that must be matched that
5442 follow this item. We calculate it ahead of time as once the
5443 lookbehind offset is added in we lose the ability to correctly
5445 ml = minlen ? *(minlen) : (I32)longest_length;
5446 *rx_end_shift = ml - offset
5447 - longest_length + (SvTAIL(sv_longest) != 0)
5450 t = (eol/* Can't have SEOL and MULTI */
5451 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
5452 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
5458 * Perl_re_op_compile - the perl internal RE engine's function to compile a
5459 * regular expression into internal code.
5460 * The pattern may be passed either as:
5461 * a list of SVs (patternp plus pat_count)
5462 * a list of OPs (expr)
5463 * If both are passed, the SV list is used, but the OP list indicates
5464 * which SVs are actually pre-compiled code blocks
5466 * The SVs in the list have magic and qr overloading applied to them (and
5467 * the list may be modified in-place with replacement SVs in the latter
5470 * If the pattern hasn't changed from old_re, then old_re will be
5473 * eng is the current engine. If that engine has an op_comp method, then
5474 * handle directly (i.e. we assume that op_comp was us); otherwise, just
5475 * do the initial concatenation of arguments and pass on to the external
5478 * If is_bare_re is not null, set it to a boolean indicating whether the
5479 * arg list reduced (after overloading) to a single bare regex which has
5480 * been returned (i.e. /$qr/).
5482 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
5484 * pm_flags contains the PMf_* flags, typically based on those from the
5485 * pm_flags field of the related PMOP. Currently we're only interested in
5486 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
5488 * We can't allocate space until we know how big the compiled form will be,
5489 * but we can't compile it (and thus know how big it is) until we've got a
5490 * place to put the code. So we cheat: we compile it twice, once with code
5491 * generation turned off and size counting turned on, and once "for real".
5492 * This also means that we don't allocate space until we are sure that the
5493 * thing really will compile successfully, and we never have to move the
5494 * code and thus invalidate pointers into it. (Note that it has to be in
5495 * one piece because free() must be able to free it all.) [NB: not true in perl]
5497 * Beware that the optimization-preparation code in here knows about some
5498 * of the structure of the compiled regexp. [I'll say.]
5502 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
5503 OP *expr, const regexp_engine* eng, REGEXP *old_re,
5504 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
5509 regexp_internal *ri;
5517 SV *code_blocksv = NULL;
5518 SV** new_patternp = patternp;
5520 /* these are all flags - maybe they should be turned
5521 * into a single int with different bit masks */
5522 I32 sawlookahead = 0;
5527 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
5529 bool runtime_code = 0;
5531 RExC_state_t RExC_state;
5532 RExC_state_t * const pRExC_state = &RExC_state;
5533 #ifdef TRIE_STUDY_OPT
5535 RExC_state_t copyRExC_state;
5537 GET_RE_DEBUG_FLAGS_DECL;
5539 PERL_ARGS_ASSERT_RE_OP_COMPILE;
5541 DEBUG_r(if (!PL_colorset) reginitcolors());
5543 #ifndef PERL_IN_XSUB_RE
5544 /* Initialize these here instead of as-needed, as is quick and avoids
5545 * having to test them each time otherwise */
5546 if (! PL_AboveLatin1) {
5547 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
5548 PL_ASCII = _new_invlist_C_array(ASCII_invlist);
5549 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
5551 PL_L1Posix_ptrs[_CC_ALPHANUMERIC]
5552 = _new_invlist_C_array(L1PosixAlnum_invlist);
5553 PL_Posix_ptrs[_CC_ALPHANUMERIC]
5554 = _new_invlist_C_array(PosixAlnum_invlist);
5556 PL_L1Posix_ptrs[_CC_ALPHA]
5557 = _new_invlist_C_array(L1PosixAlpha_invlist);
5558 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(PosixAlpha_invlist);
5560 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(PosixBlank_invlist);
5561 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(XPosixBlank_invlist);
5563 /* Cased is the same as Alpha in the ASCII range */
5564 PL_L1Posix_ptrs[_CC_CASED] = _new_invlist_C_array(L1Cased_invlist);
5565 PL_Posix_ptrs[_CC_CASED] = _new_invlist_C_array(PosixAlpha_invlist);
5567 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(PosixCntrl_invlist);
5568 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(XPosixCntrl_invlist);
5570 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5571 PL_L1Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5573 PL_L1Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(L1PosixGraph_invlist);
5574 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(PosixGraph_invlist);
5576 PL_L1Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(L1PosixLower_invlist);
5577 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(PosixLower_invlist);
5579 PL_L1Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(L1PosixPrint_invlist);
5580 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(PosixPrint_invlist);
5582 PL_L1Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(L1PosixPunct_invlist);
5583 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(PosixPunct_invlist);
5585 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(PerlSpace_invlist);
5586 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(XPerlSpace_invlist);
5587 PL_Posix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(PosixSpace_invlist);
5588 PL_XPosix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(XPosixSpace_invlist);
5590 PL_L1Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(L1PosixUpper_invlist);
5591 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(PosixUpper_invlist);
5593 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(VertSpace_invlist);
5595 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(PosixWord_invlist);
5596 PL_L1Posix_ptrs[_CC_WORDCHAR]
5597 = _new_invlist_C_array(L1PosixWord_invlist);
5599 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(PosixXDigit_invlist);
5600 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(XPosixXDigit_invlist);
5602 PL_HasMultiCharFold = _new_invlist_C_array(_Perl_Multi_Char_Folds_invlist);
5606 pRExC_state->code_blocks = NULL;
5607 pRExC_state->num_code_blocks = 0;
5610 *is_bare_re = FALSE;
5612 if (expr && (expr->op_type == OP_LIST ||
5613 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
5614 /* allocate code_blocks if needed */
5618 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling)
5619 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
5620 ncode++; /* count of DO blocks */
5622 pRExC_state->num_code_blocks = ncode;
5623 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
5628 /* compile-time pattern with just OP_CONSTs and DO blocks */
5633 /* find how many CONSTs there are */
5636 if (expr->op_type == OP_CONST)
5639 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5640 if (o->op_type == OP_CONST)
5644 /* fake up an SV array */
5646 assert(!new_patternp);
5647 Newx(new_patternp, n, SV*);
5648 SAVEFREEPV(new_patternp);
5652 if (expr->op_type == OP_CONST)
5653 new_patternp[n] = cSVOPx_sv(expr);
5655 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5656 if (o->op_type == OP_CONST)
5657 new_patternp[n++] = cSVOPo_sv;
5662 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5663 "Assembling pattern from %d elements%s\n", pat_count,
5664 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5666 /* set expr to the first arg op */
5668 if (pRExC_state->num_code_blocks
5669 && expr->op_type != OP_CONST)
5671 expr = cLISTOPx(expr)->op_first;
5672 assert( expr->op_type == OP_PUSHMARK
5673 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
5674 || expr->op_type == OP_PADRANGE);
5675 expr = expr->op_sibling;
5678 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
5679 expr, &recompile, NULL);
5681 /* handle bare (possibly after overloading) regex: foo =~ $re */
5686 if (SvTYPE(re) == SVt_REGEXP) {
5690 Safefree(pRExC_state->code_blocks);
5691 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5692 "Precompiled pattern%s\n",
5693 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5699 exp = SvPV_nomg(pat, plen);
5701 if (!eng->op_comp) {
5702 if ((SvUTF8(pat) && IN_BYTES)
5703 || SvGMAGICAL(pat) || SvAMAGIC(pat))
5705 /* make a temporary copy; either to convert to bytes,
5706 * or to avoid repeating get-magic / overloaded stringify */
5707 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
5708 (IN_BYTES ? 0 : SvUTF8(pat)));
5710 Safefree(pRExC_state->code_blocks);
5711 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
5714 /* ignore the utf8ness if the pattern is 0 length */
5715 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
5716 RExC_uni_semantics = 0;
5717 RExC_contains_locale = 0;
5718 pRExC_state->runtime_code_qr = NULL;
5721 SV *dsv= sv_newmortal();
5722 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
5723 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
5724 PL_colors[4],PL_colors[5],s);
5728 /* we jump here if we upgrade the pattern to utf8 and have to
5731 if ((pm_flags & PMf_USE_RE_EVAL)
5732 /* this second condition covers the non-regex literal case,
5733 * i.e. $foo =~ '(?{})'. */
5734 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
5736 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
5738 /* return old regex if pattern hasn't changed */
5739 /* XXX: note in the below we have to check the flags as well as the pattern.
5741 * Things get a touch tricky as we have to compare the utf8 flag independently
5742 * from the compile flags.
5747 && !!RX_UTF8(old_re) == !!RExC_utf8
5748 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
5749 && RX_PRECOMP(old_re)
5750 && RX_PRELEN(old_re) == plen
5751 && memEQ(RX_PRECOMP(old_re), exp, plen)
5752 && !runtime_code /* with runtime code, always recompile */ )
5754 Safefree(pRExC_state->code_blocks);
5758 rx_flags = orig_rx_flags;
5760 if (initial_charset == REGEX_LOCALE_CHARSET) {
5761 RExC_contains_locale = 1;
5763 else if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
5765 /* Set to use unicode semantics if the pattern is in utf8 and has the
5766 * 'depends' charset specified, as it means unicode when utf8 */
5767 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5771 RExC_flags = rx_flags;
5772 RExC_pm_flags = pm_flags;
5775 if (TAINTING_get && TAINT_get)
5776 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
5778 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
5779 /* whoops, we have a non-utf8 pattern, whilst run-time code
5780 * got compiled as utf8. Try again with a utf8 pattern */
5781 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
5782 pRExC_state->num_code_blocks);
5783 goto redo_first_pass;
5786 assert(!pRExC_state->runtime_code_qr);
5791 RExC_in_lookbehind = 0;
5792 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
5794 RExC_override_recoding = 0;
5795 RExC_in_multi_char_class = 0;
5797 /* First pass: determine size, legality. */
5800 RExC_end = exp + plen;
5805 RExC_emit = &RExC_emit_dummy;
5806 RExC_whilem_seen = 0;
5807 RExC_open_parens = NULL;
5808 RExC_close_parens = NULL;
5810 RExC_paren_names = NULL;
5812 RExC_paren_name_list = NULL;
5814 RExC_recurse = NULL;
5815 RExC_recurse_count = 0;
5816 pRExC_state->code_index = 0;
5818 #if 0 /* REGC() is (currently) a NOP at the first pass.
5819 * Clever compilers notice this and complain. --jhi */
5820 REGC((U8)REG_MAGIC, (char*)RExC_emit);
5823 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
5825 RExC_lastparse=NULL;
5827 /* reg may croak on us, not giving us a chance to free
5828 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
5829 need it to survive as long as the regexp (qr/(?{})/).
5830 We must check that code_blocksv is not already set, because we may
5831 have jumped back to restart the sizing pass. */
5832 if (pRExC_state->code_blocks && !code_blocksv) {
5833 code_blocksv = newSV_type(SVt_PV);
5834 SAVEFREESV(code_blocksv);
5835 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
5836 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
5838 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5839 /* It's possible to write a regexp in ascii that represents Unicode
5840 codepoints outside of the byte range, such as via \x{100}. If we
5841 detect such a sequence we have to convert the entire pattern to utf8
5842 and then recompile, as our sizing calculation will have been based
5843 on 1 byte == 1 character, but we will need to use utf8 to encode
5844 at least some part of the pattern, and therefore must convert the whole
5847 if (flags & RESTART_UTF8) {
5848 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
5849 pRExC_state->num_code_blocks);
5850 goto redo_first_pass;
5852 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
5855 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
5858 PerlIO_printf(Perl_debug_log,
5859 "Required size %"IVdf" nodes\n"
5860 "Starting second pass (creation)\n",
5863 RExC_lastparse=NULL;
5866 /* The first pass could have found things that force Unicode semantics */
5867 if ((RExC_utf8 || RExC_uni_semantics)
5868 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
5870 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5873 /* Small enough for pointer-storage convention?
5874 If extralen==0, this means that we will not need long jumps. */
5875 if (RExC_size >= 0x10000L && RExC_extralen)
5876 RExC_size += RExC_extralen;
5879 if (RExC_whilem_seen > 15)
5880 RExC_whilem_seen = 15;
5882 /* Allocate space and zero-initialize. Note, the two step process
5883 of zeroing when in debug mode, thus anything assigned has to
5884 happen after that */
5885 rx = (REGEXP*) newSV_type(SVt_REGEXP);
5887 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
5888 char, regexp_internal);
5889 if ( r == NULL || ri == NULL )
5890 FAIL("Regexp out of space");
5892 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
5893 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode), char);
5895 /* bulk initialize base fields with 0. */
5896 Zero(ri, sizeof(regexp_internal), char);
5899 /* non-zero initialization begins here */
5902 r->extflags = rx_flags;
5903 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
5905 if (pm_flags & PMf_IS_QR) {
5906 ri->code_blocks = pRExC_state->code_blocks;
5907 ri->num_code_blocks = pRExC_state->num_code_blocks;
5912 for (n = 0; n < pRExC_state->num_code_blocks; n++)
5913 if (pRExC_state->code_blocks[n].src_regex)
5914 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
5915 SAVEFREEPV(pRExC_state->code_blocks);
5919 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
5920 bool has_charset = (get_regex_charset(r->extflags) != REGEX_DEPENDS_CHARSET);
5922 /* The caret is output if there are any defaults: if not all the STD
5923 * flags are set, or if no character set specifier is needed */
5925 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
5927 bool has_runon = ((RExC_seen & REG_SEEN_RUN_ON_COMMENT)==REG_SEEN_RUN_ON_COMMENT);
5928 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
5929 >> RXf_PMf_STD_PMMOD_SHIFT);
5930 const char *fptr = STD_PAT_MODS; /*"msix"*/
5932 /* Allocate for the worst case, which is all the std flags are turned
5933 * on. If more precision is desired, we could do a population count of
5934 * the flags set. This could be done with a small lookup table, or by
5935 * shifting, masking and adding, or even, when available, assembly
5936 * language for a machine-language population count.
5937 * We never output a minus, as all those are defaults, so are
5938 * covered by the caret */
5939 const STRLEN wraplen = plen + has_p + has_runon
5940 + has_default /* If needs a caret */
5942 /* If needs a character set specifier */
5943 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
5944 + (sizeof(STD_PAT_MODS) - 1)
5945 + (sizeof("(?:)") - 1);
5947 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
5948 r->xpv_len_u.xpvlenu_pv = p;
5950 SvFLAGS(rx) |= SVf_UTF8;
5953 /* If a default, cover it using the caret */
5955 *p++= DEFAULT_PAT_MOD;
5959 const char* const name = get_regex_charset_name(r->extflags, &len);
5960 Copy(name, p, len, char);
5964 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
5967 while((ch = *fptr++)) {
5975 Copy(RExC_precomp, p, plen, char);
5976 assert ((RX_WRAPPED(rx) - p) < 16);
5977 r->pre_prefix = p - RX_WRAPPED(rx);
5983 SvCUR_set(rx, p - RX_WRAPPED(rx));
5987 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
5989 if (RExC_seen & REG_SEEN_RECURSE) {
5990 Newxz(RExC_open_parens, RExC_npar,regnode *);
5991 SAVEFREEPV(RExC_open_parens);
5992 Newxz(RExC_close_parens,RExC_npar,regnode *);
5993 SAVEFREEPV(RExC_close_parens);
5996 /* Useful during FAIL. */
5997 #ifdef RE_TRACK_PATTERN_OFFSETS
5998 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
5999 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
6000 "%s %"UVuf" bytes for offset annotations.\n",
6001 ri->u.offsets ? "Got" : "Couldn't get",
6002 (UV)((2*RExC_size+1) * sizeof(U32))));
6004 SetProgLen(ri,RExC_size);
6008 REH_CALL_COMP_BEGIN_HOOK(pRExC_state->rx);
6010 /* Second pass: emit code. */
6011 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
6012 RExC_pm_flags = pm_flags;
6014 RExC_end = exp + plen;
6017 RExC_emit_start = ri->program;
6018 RExC_emit = ri->program;
6019 RExC_emit_bound = ri->program + RExC_size + 1;
6020 pRExC_state->code_index = 0;
6022 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
6023 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6025 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#"UVxf"", (UV) flags);
6027 /* XXXX To minimize changes to RE engine we always allocate
6028 3-units-long substrs field. */
6029 Newx(r->substrs, 1, struct reg_substr_data);
6030 if (RExC_recurse_count) {
6031 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6032 SAVEFREEPV(RExC_recurse);
6036 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
6037 Zero(r->substrs, 1, struct reg_substr_data);
6039 #ifdef TRIE_STUDY_OPT
6041 StructCopy(&zero_scan_data, &data, scan_data_t);
6042 copyRExC_state = RExC_state;
6045 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6047 RExC_state = copyRExC_state;
6048 if (seen & REG_TOP_LEVEL_BRANCHES)
6049 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
6051 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES;
6052 StructCopy(&zero_scan_data, &data, scan_data_t);
6055 StructCopy(&zero_scan_data, &data, scan_data_t);
6058 /* Dig out information for optimizations. */
6059 r->extflags = RExC_flags; /* was pm_op */
6060 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6063 SvUTF8_on(rx); /* Unicode in it? */
6064 ri->regstclass = NULL;
6065 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
6066 r->intflags |= PREGf_NAUGHTY;
6067 scan = ri->program + 1; /* First BRANCH. */
6069 /* testing for BRANCH here tells us whether there is "must appear"
6070 data in the pattern. If there is then we can use it for optimisations */
6071 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES)) { /* Only one top-level choice. */
6073 STRLEN longest_float_length, longest_fixed_length;
6074 struct regnode_charclass_class ch_class; /* pointed to by data */
6076 I32 last_close = 0; /* pointed to by data */
6077 regnode *first= scan;
6078 regnode *first_next= regnext(first);
6080 * Skip introductions and multiplicators >= 1
6081 * so that we can extract the 'meat' of the pattern that must
6082 * match in the large if() sequence following.
6083 * NOTE that EXACT is NOT covered here, as it is normally
6084 * picked up by the optimiser separately.
6086 * This is unfortunate as the optimiser isnt handling lookahead
6087 * properly currently.
6090 while ((OP(first) == OPEN && (sawopen = 1)) ||
6091 /* An OR of *one* alternative - should not happen now. */
6092 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6093 /* for now we can't handle lookbehind IFMATCH*/
6094 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6095 (OP(first) == PLUS) ||
6096 (OP(first) == MINMOD) ||
6097 /* An {n,m} with n>0 */
6098 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6099 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6102 * the only op that could be a regnode is PLUS, all the rest
6103 * will be regnode_1 or regnode_2.
6105 * (yves doesn't think this is true)
6107 if (OP(first) == PLUS)
6110 if (OP(first) == MINMOD)
6112 first += regarglen[OP(first)];
6114 first = NEXTOPER(first);
6115 first_next= regnext(first);
6118 /* Starting-point info. */
6120 DEBUG_PEEP("first:",first,0);
6121 /* Ignore EXACT as we deal with it later. */
6122 if (PL_regkind[OP(first)] == EXACT) {
6123 if (OP(first) == EXACT)
6124 NOOP; /* Empty, get anchored substr later. */
6126 ri->regstclass = first;
6129 else if (PL_regkind[OP(first)] == TRIE &&
6130 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6133 /* this can happen only on restudy */
6134 if ( OP(first) == TRIE ) {
6135 struct regnode_1 *trieop = (struct regnode_1 *)
6136 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6137 StructCopy(first,trieop,struct regnode_1);
6138 trie_op=(regnode *)trieop;
6140 struct regnode_charclass *trieop = (struct regnode_charclass *)
6141 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6142 StructCopy(first,trieop,struct regnode_charclass);
6143 trie_op=(regnode *)trieop;
6146 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6147 ri->regstclass = trie_op;
6150 else if (REGNODE_SIMPLE(OP(first)))
6151 ri->regstclass = first;
6152 else if (PL_regkind[OP(first)] == BOUND ||
6153 PL_regkind[OP(first)] == NBOUND)
6154 ri->regstclass = first;
6155 else if (PL_regkind[OP(first)] == BOL) {
6156 r->extflags |= (OP(first) == MBOL
6158 : (OP(first) == SBOL
6161 first = NEXTOPER(first);
6164 else if (OP(first) == GPOS) {
6165 r->extflags |= RXf_ANCH_GPOS;
6166 first = NEXTOPER(first);
6169 else if ((!sawopen || !RExC_sawback) &&
6170 (OP(first) == STAR &&
6171 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6172 !(r->extflags & RXf_ANCH) && !pRExC_state->num_code_blocks)
6174 /* turn .* into ^.* with an implied $*=1 */
6176 (OP(NEXTOPER(first)) == REG_ANY)
6179 r->extflags |= type;
6180 r->intflags |= PREGf_IMPLICIT;
6181 first = NEXTOPER(first);
6184 if (sawplus && !sawminmod && !sawlookahead && (!sawopen || !RExC_sawback)
6185 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6186 /* x+ must match at the 1st pos of run of x's */
6187 r->intflags |= PREGf_SKIP;
6189 /* Scan is after the zeroth branch, first is atomic matcher. */
6190 #ifdef TRIE_STUDY_OPT
6193 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6194 (IV)(first - scan + 1))
6198 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6199 (IV)(first - scan + 1))
6205 * If there's something expensive in the r.e., find the
6206 * longest literal string that must appear and make it the
6207 * regmust. Resolve ties in favor of later strings, since
6208 * the regstart check works with the beginning of the r.e.
6209 * and avoiding duplication strengthens checking. Not a
6210 * strong reason, but sufficient in the absence of others.
6211 * [Now we resolve ties in favor of the earlier string if
6212 * it happens that c_offset_min has been invalidated, since the
6213 * earlier string may buy us something the later one won't.]
6216 data.longest_fixed = newSVpvs("");
6217 data.longest_float = newSVpvs("");
6218 data.last_found = newSVpvs("");
6219 data.longest = &(data.longest_fixed);
6220 ENTER_with_name("study_chunk");
6221 SAVEFREESV(data.longest_fixed);
6222 SAVEFREESV(data.longest_float);
6223 SAVEFREESV(data.last_found);
6225 if (!ri->regstclass) {
6226 cl_init(pRExC_state, &ch_class);
6227 data.start_class = &ch_class;
6228 stclass_flag = SCF_DO_STCLASS_AND;
6229 } else /* XXXX Check for BOUND? */
6231 data.last_closep = &last_close;
6233 minlen = study_chunk(pRExC_state, &first, &minlen, &fake, scan + RExC_size, /* Up to end */
6234 &data, -1, NULL, NULL,
6235 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
6236 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
6240 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6243 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6244 && data.last_start_min == 0 && data.last_end > 0
6245 && !RExC_seen_zerolen
6246 && !(RExC_seen & REG_SEEN_VERBARG)
6247 && (!(RExC_seen & REG_SEEN_GPOS) || (r->extflags & RXf_ANCH_GPOS)))
6248 r->extflags |= RXf_CHECK_ALL;
6249 scan_commit(pRExC_state, &data,&minlen,0);
6251 longest_float_length = CHR_SVLEN(data.longest_float);
6253 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6254 && data.offset_fixed == data.offset_float_min
6255 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6256 && S_setup_longest (aTHX_ pRExC_state,
6260 &(r->float_end_shift),
6261 data.lookbehind_float,
6262 data.offset_float_min,
6264 longest_float_length,
6265 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6266 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6268 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6269 r->float_max_offset = data.offset_float_max;
6270 if (data.offset_float_max < I32_MAX) /* Don't offset infinity */
6271 r->float_max_offset -= data.lookbehind_float;
6272 SvREFCNT_inc_simple_void_NN(data.longest_float);
6275 r->float_substr = r->float_utf8 = NULL;
6276 longest_float_length = 0;
6279 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6281 if (S_setup_longest (aTHX_ pRExC_state,
6283 &(r->anchored_utf8),
6284 &(r->anchored_substr),
6285 &(r->anchored_end_shift),
6286 data.lookbehind_fixed,
6289 longest_fixed_length,
6290 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6291 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6293 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6294 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6297 r->anchored_substr = r->anchored_utf8 = NULL;
6298 longest_fixed_length = 0;
6300 LEAVE_with_name("study_chunk");
6303 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6304 ri->regstclass = NULL;
6306 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6308 && ! TEST_SSC_EOS(data.start_class)
6309 && !cl_is_anything(data.start_class))
6311 const U32 n = add_data(pRExC_state, 1, "f");
6312 OP(data.start_class) = ANYOF_SYNTHETIC;
6314 Newx(RExC_rxi->data->data[n], 1,
6315 struct regnode_charclass_class);
6316 StructCopy(data.start_class,
6317 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6318 struct regnode_charclass_class);
6319 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6320 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6321 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6322 regprop(r, sv, (regnode*)data.start_class);
6323 PerlIO_printf(Perl_debug_log,
6324 "synthetic stclass \"%s\".\n",
6325 SvPVX_const(sv));});
6328 /* A temporary algorithm prefers floated substr to fixed one to dig more info. */
6329 if (longest_fixed_length > longest_float_length) {
6330 r->check_end_shift = r->anchored_end_shift;
6331 r->check_substr = r->anchored_substr;
6332 r->check_utf8 = r->anchored_utf8;
6333 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6334 if (r->extflags & RXf_ANCH_SINGLE)
6335 r->extflags |= RXf_NOSCAN;
6338 r->check_end_shift = r->float_end_shift;
6339 r->check_substr = r->float_substr;
6340 r->check_utf8 = r->float_utf8;
6341 r->check_offset_min = r->float_min_offset;
6342 r->check_offset_max = r->float_max_offset;
6344 /* XXXX Currently intuiting is not compatible with ANCH_GPOS.
6345 This should be changed ASAP! */
6346 if ((r->check_substr || r->check_utf8) && !(r->extflags & RXf_ANCH_GPOS)) {
6347 r->extflags |= RXf_USE_INTUIT;
6348 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6349 r->extflags |= RXf_INTUIT_TAIL;
6351 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6352 if ( (STRLEN)minlen < longest_float_length )
6353 minlen= longest_float_length;
6354 if ( (STRLEN)minlen < longest_fixed_length )
6355 minlen= longest_fixed_length;
6359 /* Several toplevels. Best we can is to set minlen. */
6361 struct regnode_charclass_class ch_class;
6364 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6366 scan = ri->program + 1;
6367 cl_init(pRExC_state, &ch_class);
6368 data.start_class = &ch_class;
6369 data.last_closep = &last_close;
6372 minlen = study_chunk(pRExC_state, &scan, &minlen, &fake, scan + RExC_size,
6373 &data, -1, NULL, NULL,
6374 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS
6375 |(restudied ? SCF_TRIE_DOING_RESTUDY : 0),
6378 CHECK_RESTUDY_GOTO_butfirst(NOOP);
6380 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
6381 = r->float_substr = r->float_utf8 = NULL;
6383 if (! TEST_SSC_EOS(data.start_class)
6384 && !cl_is_anything(data.start_class))
6386 const U32 n = add_data(pRExC_state, 1, "f");
6387 OP(data.start_class) = ANYOF_SYNTHETIC;
6389 Newx(RExC_rxi->data->data[n], 1,
6390 struct regnode_charclass_class);
6391 StructCopy(data.start_class,
6392 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6393 struct regnode_charclass_class);
6394 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6395 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6396 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
6397 regprop(r, sv, (regnode*)data.start_class);
6398 PerlIO_printf(Perl_debug_log,
6399 "synthetic stclass \"%s\".\n",
6400 SvPVX_const(sv));});
6404 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
6405 the "real" pattern. */
6407 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf"\n",
6408 (IV)minlen, (IV)r->minlen);
6410 r->minlenret = minlen;
6411 if (r->minlen < minlen)
6414 if (RExC_seen & REG_SEEN_GPOS)
6415 r->extflags |= RXf_GPOS_SEEN;
6416 if (RExC_seen & REG_SEEN_LOOKBEHIND)
6417 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the lookbehind */
6418 if (pRExC_state->num_code_blocks)
6419 r->extflags |= RXf_EVAL_SEEN;
6420 if (RExC_seen & REG_SEEN_CANY)
6421 r->extflags |= RXf_CANY_SEEN;
6422 if (RExC_seen & REG_SEEN_VERBARG)
6424 r->intflags |= PREGf_VERBARG_SEEN;
6425 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
6427 if (RExC_seen & REG_SEEN_CUTGROUP)
6428 r->intflags |= PREGf_CUTGROUP_SEEN;
6429 if (pm_flags & PMf_USE_RE_EVAL)
6430 r->intflags |= PREGf_USE_RE_EVAL;
6431 if (RExC_paren_names)
6432 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
6434 RXp_PAREN_NAMES(r) = NULL;
6437 regnode *first = ri->program + 1;
6439 regnode *next = NEXTOPER(first);
6442 if (PL_regkind[fop] == NOTHING && nop == END)
6443 r->extflags |= RXf_NULL;
6444 else if (PL_regkind[fop] == BOL && nop == END)
6445 r->extflags |= RXf_START_ONLY;
6446 else if (fop == PLUS && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE && OP(regnext(first)) == END)
6447 r->extflags |= RXf_WHITE;
6448 else if ( r->extflags & RXf_SPLIT && fop == EXACT && STR_LEN(first) == 1 && *(STRING(first)) == ' ' && OP(regnext(first)) == END )
6449 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
6453 if (RExC_paren_names) {
6454 ri->name_list_idx = add_data( pRExC_state, 1, "a" );
6455 ri->data->data[ri->name_list_idx] = (void*)SvREFCNT_inc(RExC_paren_name_list);
6458 ri->name_list_idx = 0;
6460 if (RExC_recurse_count) {
6461 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
6462 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
6463 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
6466 Newxz(r->offs, RExC_npar, regexp_paren_pair);
6467 /* assume we don't need to swap parens around before we match */
6470 PerlIO_printf(Perl_debug_log,"Final program:\n");
6473 #ifdef RE_TRACK_PATTERN_OFFSETS
6474 DEBUG_OFFSETS_r(if (ri->u.offsets) {
6475 const U32 len = ri->u.offsets[0];
6477 GET_RE_DEBUG_FLAGS_DECL;
6478 PerlIO_printf(Perl_debug_log, "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
6479 for (i = 1; i <= len; i++) {
6480 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
6481 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
6482 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
6484 PerlIO_printf(Perl_debug_log, "\n");
6489 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
6490 * by setting the regexp SV to readonly-only instead. If the
6491 * pattern's been recompiled, the USEDness should remain. */
6492 if (old_re && SvREADONLY(old_re))
6500 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
6503 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
6505 PERL_UNUSED_ARG(value);
6507 if (flags & RXapif_FETCH) {
6508 return reg_named_buff_fetch(rx, key, flags);
6509 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
6510 Perl_croak_no_modify();
6512 } else if (flags & RXapif_EXISTS) {
6513 return reg_named_buff_exists(rx, key, flags)
6516 } else if (flags & RXapif_REGNAMES) {
6517 return reg_named_buff_all(rx, flags);
6518 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
6519 return reg_named_buff_scalar(rx, flags);
6521 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
6527 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
6530 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
6531 PERL_UNUSED_ARG(lastkey);
6533 if (flags & RXapif_FIRSTKEY)
6534 return reg_named_buff_firstkey(rx, flags);
6535 else if (flags & RXapif_NEXTKEY)
6536 return reg_named_buff_nextkey(rx, flags);
6538 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter", (int)flags);
6544 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
6547 AV *retarray = NULL;
6549 struct regexp *const rx = ReANY(r);
6551 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
6553 if (flags & RXapif_ALL)
6556 if (rx && RXp_PAREN_NAMES(rx)) {
6557 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
6560 SV* sv_dat=HeVAL(he_str);
6561 I32 *nums=(I32*)SvPVX(sv_dat);
6562 for ( i=0; i<SvIVX(sv_dat); i++ ) {
6563 if ((I32)(rx->nparens) >= nums[i]
6564 && rx->offs[nums[i]].start != -1
6565 && rx->offs[nums[i]].end != -1)
6568 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
6573 ret = newSVsv(&PL_sv_undef);
6576 av_push(retarray, ret);
6579 return newRV_noinc(MUTABLE_SV(retarray));
6586 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
6589 struct regexp *const rx = ReANY(r);
6591 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
6593 if (rx && RXp_PAREN_NAMES(rx)) {
6594 if (flags & RXapif_ALL) {
6595 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
6597 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
6599 SvREFCNT_dec_NN(sv);
6611 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
6613 struct regexp *const rx = ReANY(r);
6615 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
6617 if ( rx && RXp_PAREN_NAMES(rx) ) {
6618 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
6620 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
6627 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
6629 struct regexp *const rx = ReANY(r);
6630 GET_RE_DEBUG_FLAGS_DECL;
6632 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
6634 if (rx && RXp_PAREN_NAMES(rx)) {
6635 HV *hv = RXp_PAREN_NAMES(rx);
6637 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6640 SV* sv_dat = HeVAL(temphe);
6641 I32 *nums = (I32*)SvPVX(sv_dat);
6642 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6643 if ((I32)(rx->lastparen) >= nums[i] &&
6644 rx->offs[nums[i]].start != -1 &&
6645 rx->offs[nums[i]].end != -1)
6651 if (parno || flags & RXapif_ALL) {
6652 return newSVhek(HeKEY_hek(temphe));
6660 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
6665 struct regexp *const rx = ReANY(r);
6667 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
6669 if (rx && RXp_PAREN_NAMES(rx)) {
6670 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
6671 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
6672 } else if (flags & RXapif_ONE) {
6673 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
6674 av = MUTABLE_AV(SvRV(ret));
6675 length = av_len(av);
6676 SvREFCNT_dec_NN(ret);
6677 return newSViv(length + 1);
6679 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar", (int)flags);
6683 return &PL_sv_undef;
6687 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
6689 struct regexp *const rx = ReANY(r);
6692 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
6694 if (rx && RXp_PAREN_NAMES(rx)) {
6695 HV *hv= RXp_PAREN_NAMES(rx);
6697 (void)hv_iterinit(hv);
6698 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6701 SV* sv_dat = HeVAL(temphe);
6702 I32 *nums = (I32*)SvPVX(sv_dat);
6703 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6704 if ((I32)(rx->lastparen) >= nums[i] &&
6705 rx->offs[nums[i]].start != -1 &&
6706 rx->offs[nums[i]].end != -1)
6712 if (parno || flags & RXapif_ALL) {
6713 av_push(av, newSVhek(HeKEY_hek(temphe)));
6718 return newRV_noinc(MUTABLE_SV(av));
6722 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
6725 struct regexp *const rx = ReANY(r);
6731 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
6733 if ( ( n == RX_BUFF_IDX_CARET_PREMATCH
6734 || n == RX_BUFF_IDX_CARET_FULLMATCH
6735 || n == RX_BUFF_IDX_CARET_POSTMATCH
6737 && !(rx->extflags & RXf_PMf_KEEPCOPY)
6744 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
6745 /* no need to distinguish between them any more */
6746 n = RX_BUFF_IDX_FULLMATCH;
6748 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
6749 && rx->offs[0].start != -1)
6751 /* $`, ${^PREMATCH} */
6752 i = rx->offs[0].start;
6756 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
6757 && rx->offs[0].end != -1)
6759 /* $', ${^POSTMATCH} */
6760 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
6761 i = rx->sublen + rx->suboffset - rx->offs[0].end;
6764 if ( 0 <= n && n <= (I32)rx->nparens &&
6765 (s1 = rx->offs[n].start) != -1 &&
6766 (t1 = rx->offs[n].end) != -1)
6768 /* $&, ${^MATCH}, $1 ... */
6770 s = rx->subbeg + s1 - rx->suboffset;
6775 assert(s >= rx->subbeg);
6776 assert(rx->sublen >= (s - rx->subbeg) + i );
6778 #if NO_TAINT_SUPPORT
6779 sv_setpvn(sv, s, i);
6781 const int oldtainted = TAINT_get;
6783 sv_setpvn(sv, s, i);
6784 TAINT_set(oldtainted);
6786 if ( (rx->extflags & RXf_CANY_SEEN)
6787 ? (RXp_MATCH_UTF8(rx)
6788 && (!i || is_utf8_string((U8*)s, i)))
6789 : (RXp_MATCH_UTF8(rx)) )
6796 if (RXp_MATCH_TAINTED(rx)) {
6797 if (SvTYPE(sv) >= SVt_PVMG) {
6798 MAGIC* const mg = SvMAGIC(sv);
6801 SvMAGIC_set(sv, mg->mg_moremagic);
6803 if ((mgt = SvMAGIC(sv))) {
6804 mg->mg_moremagic = mgt;
6805 SvMAGIC_set(sv, mg);
6816 sv_setsv(sv,&PL_sv_undef);
6822 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
6823 SV const * const value)
6825 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
6827 PERL_UNUSED_ARG(rx);
6828 PERL_UNUSED_ARG(paren);
6829 PERL_UNUSED_ARG(value);
6832 Perl_croak_no_modify();
6836 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
6839 struct regexp *const rx = ReANY(r);
6843 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
6845 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
6847 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
6848 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6852 case RX_BUFF_IDX_PREMATCH: /* $` */
6853 if (rx->offs[0].start != -1) {
6854 i = rx->offs[0].start;
6863 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
6864 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6866 case RX_BUFF_IDX_POSTMATCH: /* $' */
6867 if (rx->offs[0].end != -1) {
6868 i = rx->sublen - rx->offs[0].end;
6870 s1 = rx->offs[0].end;
6877 case RX_BUFF_IDX_CARET_FULLMATCH: /* ${^MATCH} */
6878 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6882 /* $& / ${^MATCH}, $1, $2, ... */
6884 if (paren <= (I32)rx->nparens &&
6885 (s1 = rx->offs[paren].start) != -1 &&
6886 (t1 = rx->offs[paren].end) != -1)
6892 if (ckWARN(WARN_UNINITIALIZED))
6893 report_uninit((const SV *)sv);
6898 if (i > 0 && RXp_MATCH_UTF8(rx)) {
6899 const char * const s = rx->subbeg - rx->suboffset + s1;
6904 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
6911 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
6913 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
6914 PERL_UNUSED_ARG(rx);
6918 return newSVpvs("Regexp");
6921 /* Scans the name of a named buffer from the pattern.
6922 * If flags is REG_RSN_RETURN_NULL returns null.
6923 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
6924 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
6925 * to the parsed name as looked up in the RExC_paren_names hash.
6926 * If there is an error throws a vFAIL().. type exception.
6929 #define REG_RSN_RETURN_NULL 0
6930 #define REG_RSN_RETURN_NAME 1
6931 #define REG_RSN_RETURN_DATA 2
6934 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
6936 char *name_start = RExC_parse;
6938 PERL_ARGS_ASSERT_REG_SCAN_NAME;
6940 if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
6941 /* skip IDFIRST by using do...while */
6944 RExC_parse += UTF8SKIP(RExC_parse);
6945 } while (isWORDCHAR_utf8((U8*)RExC_parse));
6949 } while (isWORDCHAR(*RExC_parse));
6951 RExC_parse++; /* so the <- from the vFAIL is after the offending character */
6952 vFAIL("Group name must start with a non-digit word character");
6956 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
6957 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
6958 if ( flags == REG_RSN_RETURN_NAME)
6960 else if (flags==REG_RSN_RETURN_DATA) {
6963 if ( ! sv_name ) /* should not happen*/
6964 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
6965 if (RExC_paren_names)
6966 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
6968 sv_dat = HeVAL(he_str);
6970 vFAIL("Reference to nonexistent named group");
6974 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
6975 (unsigned long) flags);
6977 assert(0); /* NOT REACHED */
6982 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
6983 int rem=(int)(RExC_end - RExC_parse); \
6992 if (RExC_lastparse!=RExC_parse) \
6993 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
6996 iscut ? "..." : "<" \
6999 PerlIO_printf(Perl_debug_log,"%16s",""); \
7002 num = RExC_size + 1; \
7004 num=REG_NODE_NUM(RExC_emit); \
7005 if (RExC_lastnum!=num) \
7006 PerlIO_printf(Perl_debug_log,"|%4d",num); \
7008 PerlIO_printf(Perl_debug_log,"|%4s",""); \
7009 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
7010 (int)((depth*2)), "", \
7014 RExC_lastparse=RExC_parse; \
7019 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
7020 DEBUG_PARSE_MSG((funcname)); \
7021 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
7023 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7024 DEBUG_PARSE_MSG((funcname)); \
7025 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7028 /* This section of code defines the inversion list object and its methods. The
7029 * interfaces are highly subject to change, so as much as possible is static to
7030 * this file. An inversion list is here implemented as a malloc'd C UV array
7031 * as an SVt_INVLIST scalar.
7033 * An inversion list for Unicode is an array of code points, sorted by ordinal
7034 * number. The zeroth element is the first code point in the list. The 1th
7035 * element is the first element beyond that not in the list. In other words,
7036 * the first range is
7037 * invlist[0]..(invlist[1]-1)
7038 * The other ranges follow. Thus every element whose index is divisible by two
7039 * marks the beginning of a range that is in the list, and every element not
7040 * divisible by two marks the beginning of a range not in the list. A single
7041 * element inversion list that contains the single code point N generally
7042 * consists of two elements
7045 * (The exception is when N is the highest representable value on the
7046 * machine, in which case the list containing just it would be a single
7047 * element, itself. By extension, if the last range in the list extends to
7048 * infinity, then the first element of that range will be in the inversion list
7049 * at a position that is divisible by two, and is the final element in the
7051 * Taking the complement (inverting) an inversion list is quite simple, if the
7052 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7053 * This implementation reserves an element at the beginning of each inversion
7054 * list to always contain 0; there is an additional flag in the header which
7055 * indicates if the list begins at the 0, or is offset to begin at the next
7058 * More about inversion lists can be found in "Unicode Demystified"
7059 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7060 * More will be coming when functionality is added later.
7062 * The inversion list data structure is currently implemented as an SV pointing
7063 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7064 * array of UV whose memory management is automatically handled by the existing
7065 * facilities for SV's.
7067 * Some of the methods should always be private to the implementation, and some
7068 * should eventually be made public */
7070 /* The header definitions are in F<inline_invlist.c> */
7072 PERL_STATIC_INLINE UV*
7073 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
7075 /* Returns a pointer to the first element in the inversion list's array.
7076 * This is called upon initialization of an inversion list. Where the
7077 * array begins depends on whether the list has the code point U+0000 in it
7078 * or not. The other parameter tells it whether the code that follows this
7079 * call is about to put a 0 in the inversion list or not. The first
7080 * element is either the element reserved for 0, if TRUE, or the element
7081 * after it, if FALSE */
7083 bool* offset = get_invlist_offset_addr(invlist);
7084 UV* zero_addr = (UV *) SvPVX(invlist);
7086 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7089 assert(! _invlist_len(invlist));
7093 /* 1^1 = 0; 1^0 = 1 */
7094 *offset = 1 ^ will_have_0;
7095 return zero_addr + *offset;
7098 PERL_STATIC_INLINE UV*
7099 S_invlist_array(pTHX_ SV* const invlist)
7101 /* Returns the pointer to the inversion list's array. Every time the
7102 * length changes, this needs to be called in case malloc or realloc moved
7105 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7107 /* Must not be empty. If these fail, you probably didn't check for <len>
7108 * being non-zero before trying to get the array */
7109 assert(_invlist_len(invlist));
7111 /* The very first element always contains zero, The array begins either
7112 * there, or if the inversion list is offset, at the element after it.
7113 * The offset header field determines which; it contains 0 or 1 to indicate
7114 * how much additionally to add */
7115 assert(0 == *(SvPVX(invlist)));
7116 return ((UV *) SvPVX(invlist) + *get_invlist_offset_addr(invlist));
7119 PERL_STATIC_INLINE void
7120 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
7122 /* Sets the current number of elements stored in the inversion list.
7123 * Updates SvCUR correspondingly */
7125 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7130 : TO_INTERNAL_SIZE(len + offset));
7131 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
7134 PERL_STATIC_INLINE IV*
7135 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7137 /* Return the address of the IV that is reserved to hold the cached index
7140 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7142 return &(((XINVLIST*) SvANY(invlist))->prev_index);
7145 PERL_STATIC_INLINE IV
7146 S_invlist_previous_index(pTHX_ SV* const invlist)
7148 /* Returns cached index of previous search */
7150 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7152 return *get_invlist_previous_index_addr(invlist);
7155 PERL_STATIC_INLINE void
7156 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7158 /* Caches <index> for later retrieval */
7160 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7162 assert(index == 0 || index < (int) _invlist_len(invlist));
7164 *get_invlist_previous_index_addr(invlist) = index;
7167 PERL_STATIC_INLINE UV
7168 S_invlist_max(pTHX_ SV* const invlist)
7170 /* Returns the maximum number of elements storable in the inversion list's
7171 * array, without having to realloc() */
7173 PERL_ARGS_ASSERT_INVLIST_MAX;
7175 /* Assumes worst case, in which the 0 element is not counted in the
7176 * inversion list, so subtracts 1 for that */
7177 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7178 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
7179 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
7182 #ifndef PERL_IN_XSUB_RE
7184 Perl__new_invlist(pTHX_ IV initial_size)
7187 /* Return a pointer to a newly constructed inversion list, with enough
7188 * space to store 'initial_size' elements. If that number is negative, a
7189 * system default is used instead */
7193 if (initial_size < 0) {
7197 /* Allocate the initial space */
7198 new_list = newSV_type(SVt_INVLIST);
7200 /* First 1 is in case the zero element isn't in the list; second 1 is for
7202 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
7203 invlist_set_len(new_list, 0, 0);
7205 /* Force iterinit() to be used to get iteration to work */
7206 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
7208 *get_invlist_previous_index_addr(new_list) = 0;
7215 S__new_invlist_C_array(pTHX_ const UV* const list)
7217 /* Return a pointer to a newly constructed inversion list, initialized to
7218 * point to <list>, which has to be in the exact correct inversion list
7219 * form, including internal fields. Thus this is a dangerous routine that
7220 * should not be used in the wrong hands. The passed in 'list' contains
7221 * several header fields at the beginning that are not part of the
7222 * inversion list body proper */
7224 const STRLEN length = (STRLEN) list[0];
7225 const UV version_id = list[1];
7226 const bool offset = cBOOL(list[2]);
7227 #define HEADER_LENGTH 3
7228 /* If any of the above changes in any way, you must change HEADER_LENGTH
7229 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
7230 * perl -E 'say int(rand 2**31-1)'
7232 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
7233 data structure type, so that one being
7234 passed in can be validated to be an
7235 inversion list of the correct vintage.
7238 SV* invlist = newSV_type(SVt_INVLIST);
7240 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7242 if (version_id != INVLIST_VERSION_ID) {
7243 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7246 /* The generated array passed in includes header elements that aren't part
7247 * of the list proper, so start it just after them */
7248 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
7250 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7251 shouldn't touch it */
7253 *(get_invlist_offset_addr(invlist)) = offset;
7255 /* The 'length' passed to us is the physical number of elements in the
7256 * inversion list. But if there is an offset the logical number is one
7258 invlist_set_len(invlist, length - offset, offset);
7260 invlist_set_previous_index(invlist, 0);
7262 /* Initialize the iteration pointer. */
7263 invlist_iterfinish(invlist);
7269 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7271 /* Grow the maximum size of an inversion list */
7273 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7275 /* Add one to account for the zero element at the beginning which may not
7276 * be counted by the calling parameters */
7277 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
7280 PERL_STATIC_INLINE void
7281 S_invlist_trim(pTHX_ SV* const invlist)
7283 PERL_ARGS_ASSERT_INVLIST_TRIM;
7285 /* Change the length of the inversion list to how many entries it currently
7287 SvPV_shrink_to_cur((SV *) invlist);
7290 #define _invlist_union_complement_2nd(a, b, output) _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
7293 S__append_range_to_invlist(pTHX_ SV* const invlist, const UV start, const UV end)
7295 /* Subject to change or removal. Append the range from 'start' to 'end' at
7296 * the end of the inversion list. The range must be above any existing
7300 UV max = invlist_max(invlist);
7301 UV len = _invlist_len(invlist);
7304 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7306 if (len == 0) { /* Empty lists must be initialized */
7307 offset = start != 0;
7308 array = _invlist_array_init(invlist, ! offset);
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 offset = *get_invlist_offset_addr(invlist);
7332 if (array[final_element] == start) {
7333 if (end != UV_MAX) {
7334 array[final_element] = end + 1;
7337 /* But if the end is the maximum representable on the machine,
7338 * just let the range that this would extend to have no end */
7339 invlist_set_len(invlist, len - 1, offset);
7345 /* Here the new range doesn't extend any existing set. Add it */
7347 len += 2; /* Includes an element each for the start and end of range */
7349 /* If wll overflow the existing space, extend, which may cause the array to
7352 invlist_extend(invlist, len);
7354 /* Have to set len here to avoid assert failure in invlist_array() */
7355 invlist_set_len(invlist, len, offset);
7357 array = invlist_array(invlist);
7360 invlist_set_len(invlist, len, offset);
7363 /* The next item on the list starts the range, the one after that is
7364 * one past the new range. */
7365 array[len - 2] = start;
7366 if (end != UV_MAX) {
7367 array[len - 1] = end + 1;
7370 /* But if the end is the maximum representable on the machine, just let
7371 * the range have no end */
7372 invlist_set_len(invlist, len - 1, offset);
7376 #ifndef PERL_IN_XSUB_RE
7379 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
7381 /* Searches the inversion list for the entry that contains the input code
7382 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
7383 * return value is the index into the list's array of the range that
7388 IV high = _invlist_len(invlist);
7389 const IV highest_element = high - 1;
7392 PERL_ARGS_ASSERT__INVLIST_SEARCH;
7394 /* If list is empty, return failure. */
7399 /* (We can't get the array unless we know the list is non-empty) */
7400 array = invlist_array(invlist);
7402 mid = invlist_previous_index(invlist);
7403 assert(mid >=0 && mid <= highest_element);
7405 /* <mid> contains the cache of the result of the previous call to this
7406 * function (0 the first time). See if this call is for the same result,
7407 * or if it is for mid-1. This is under the theory that calls to this
7408 * function will often be for related code points that are near each other.
7409 * And benchmarks show that caching gives better results. We also test
7410 * here if the code point is within the bounds of the list. These tests
7411 * replace others that would have had to be made anyway to make sure that
7412 * the array bounds were not exceeded, and these give us extra information
7413 * at the same time */
7414 if (cp >= array[mid]) {
7415 if (cp >= array[highest_element]) {
7416 return highest_element;
7419 /* Here, array[mid] <= cp < array[highest_element]. This means that
7420 * the final element is not the answer, so can exclude it; it also
7421 * means that <mid> is not the final element, so can refer to 'mid + 1'
7423 if (cp < array[mid + 1]) {
7429 else { /* cp < aray[mid] */
7430 if (cp < array[0]) { /* Fail if outside the array */
7434 if (cp >= array[mid - 1]) {
7439 /* Binary search. What we are looking for is <i> such that
7440 * array[i] <= cp < array[i+1]
7441 * The loop below converges on the i+1. Note that there may not be an
7442 * (i+1)th element in the array, and things work nonetheless */
7443 while (low < high) {
7444 mid = (low + high) / 2;
7445 assert(mid <= highest_element);
7446 if (array[mid] <= cp) { /* cp >= array[mid] */
7449 /* We could do this extra test to exit the loop early.
7450 if (cp < array[low]) {
7455 else { /* cp < array[mid] */
7462 invlist_set_previous_index(invlist, high);
7467 Perl__invlist_populate_swatch(pTHX_ SV* const invlist, const UV start, const UV end, U8* swatch)
7469 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
7470 * but is used when the swash has an inversion list. This makes this much
7471 * faster, as it uses a binary search instead of a linear one. This is
7472 * intimately tied to that function, and perhaps should be in utf8.c,
7473 * except it is intimately tied to inversion lists as well. It assumes
7474 * that <swatch> is all 0's on input */
7477 const IV len = _invlist_len(invlist);
7481 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
7483 if (len == 0) { /* Empty inversion list */
7487 array = invlist_array(invlist);
7489 /* Find which element it is */
7490 i = _invlist_search(invlist, start);
7492 /* We populate from <start> to <end> */
7493 while (current < end) {
7496 /* The inversion list gives the results for every possible code point
7497 * after the first one in the list. Only those ranges whose index is
7498 * even are ones that the inversion list matches. For the odd ones,
7499 * and if the initial code point is not in the list, we have to skip
7500 * forward to the next element */
7501 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
7503 if (i >= len) { /* Finished if beyond the end of the array */
7507 if (current >= end) { /* Finished if beyond the end of what we
7509 if (LIKELY(end < UV_MAX)) {
7513 /* We get here when the upper bound is the maximum
7514 * representable on the machine, and we are looking for just
7515 * that code point. Have to special case it */
7517 goto join_end_of_list;
7520 assert(current >= start);
7522 /* The current range ends one below the next one, except don't go past
7525 upper = (i < len && array[i] < end) ? array[i] : end;
7527 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
7528 * for each code point in it */
7529 for (; current < upper; current++) {
7530 const STRLEN offset = (STRLEN)(current - start);
7531 swatch[offset >> 3] |= 1 << (offset & 7);
7536 /* Quit if at the end of the list */
7539 /* But first, have to deal with the highest possible code point on
7540 * the platform. The previous code assumes that <end> is one
7541 * beyond where we want to populate, but that is impossible at the
7542 * platform's infinity, so have to handle it specially */
7543 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
7545 const STRLEN offset = (STRLEN)(end - start);
7546 swatch[offset >> 3] |= 1 << (offset & 7);
7551 /* Advance to the next range, which will be for code points not in the
7560 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, const bool complement_b, SV** output)
7562 /* Take the union of two inversion lists and point <output> to it. *output
7563 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7564 * the reference count to that list will be decremented. The first list,
7565 * <a>, may be NULL, in which case a copy of the second list is returned.
7566 * If <complement_b> is TRUE, the union is taken of the complement
7567 * (inversion) of <b> instead of b itself.
7569 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7570 * Richard Gillam, published by Addison-Wesley, and explained at some
7571 * length there. The preface says to incorporate its examples into your
7572 * code at your own risk.
7574 * The algorithm is like a merge sort.
7576 * XXX A potential performance improvement is to keep track as we go along
7577 * if only one of the inputs contributes to the result, meaning the other
7578 * is a subset of that one. In that case, we can skip the final copy and
7579 * return the larger of the input lists, but then outside code might need
7580 * to keep track of whether to free the input list or not */
7582 const UV* array_a; /* a's array */
7584 UV len_a; /* length of a's array */
7587 SV* u; /* the resulting union */
7591 UV i_a = 0; /* current index into a's array */
7595 /* running count, as explained in the algorithm source book; items are
7596 * stopped accumulating and are output when the count changes to/from 0.
7597 * The count is incremented when we start a range that's in the set, and
7598 * decremented when we start a range that's not in the set. So its range
7599 * is 0 to 2. Only when the count is zero is something not in the set.
7603 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
7606 /* If either one is empty, the union is the other one */
7607 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
7614 *output = invlist_clone(b);
7616 _invlist_invert(*output);
7618 } /* else *output already = b; */
7621 else if ((len_b = _invlist_len(b)) == 0) {
7626 /* The complement of an empty list is a list that has everything in it,
7627 * so the union with <a> includes everything too */
7632 *output = _new_invlist(1);
7633 _append_range_to_invlist(*output, 0, UV_MAX);
7635 else if (*output != a) {
7636 *output = invlist_clone(a);
7638 /* else *output already = a; */
7642 /* Here both lists exist and are non-empty */
7643 array_a = invlist_array(a);
7644 array_b = invlist_array(b);
7646 /* If are to take the union of 'a' with the complement of b, set it
7647 * up so are looking at b's complement. */
7650 /* To complement, we invert: if the first element is 0, remove it. To
7651 * do this, we just pretend the array starts one later */
7652 if (array_b[0] == 0) {
7658 /* But if the first element is not zero, we pretend the list starts
7659 * at the 0 that is always stored immediately before the array. */
7665 /* Size the union for the worst case: that the sets are completely
7667 u = _new_invlist(len_a + len_b);
7669 /* Will contain U+0000 if either component does */
7670 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
7671 || (len_b > 0 && array_b[0] == 0));
7673 /* Go through each list item by item, stopping when exhausted one of
7675 while (i_a < len_a && i_b < len_b) {
7676 UV cp; /* The element to potentially add to the union's array */
7677 bool cp_in_set; /* is it in the the input list's set or not */
7679 /* We need to take one or the other of the two inputs for the union.
7680 * Since we are merging two sorted lists, we take the smaller of the
7681 * next items. In case of a tie, we take the one that is in its set
7682 * first. If we took one not in the set first, it would decrement the
7683 * count, possibly to 0 which would cause it to be output as ending the
7684 * range, and the next time through we would take the same number, and
7685 * output it again as beginning the next range. By doing it the
7686 * opposite way, there is no possibility that the count will be
7687 * momentarily decremented to 0, and thus the two adjoining ranges will
7688 * be seamlessly merged. (In a tie and both are in the set or both not
7689 * in the set, it doesn't matter which we take first.) */
7690 if (array_a[i_a] < array_b[i_b]
7691 || (array_a[i_a] == array_b[i_b]
7692 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7694 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7698 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7699 cp = array_b[i_b++];
7702 /* Here, have chosen which of the two inputs to look at. Only output
7703 * if the running count changes to/from 0, which marks the
7704 * beginning/end of a range in that's in the set */
7707 array_u[i_u++] = cp;
7714 array_u[i_u++] = cp;
7719 /* Here, we are finished going through at least one of the lists, which
7720 * means there is something remaining in at most one. We check if the list
7721 * that hasn't been exhausted is positioned such that we are in the middle
7722 * of a range in its set or not. (i_a and i_b point to the element beyond
7723 * the one we care about.) If in the set, we decrement 'count'; if 0, there
7724 * is potentially more to output.
7725 * There are four cases:
7726 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
7727 * in the union is entirely from the non-exhausted set.
7728 * 2) Both were in their sets, count is 2. Nothing further should
7729 * be output, as everything that remains will be in the exhausted
7730 * list's set, hence in the union; decrementing to 1 but not 0 insures
7732 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
7733 * Nothing further should be output because the union includes
7734 * everything from the exhausted set. Not decrementing ensures that.
7735 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
7736 * decrementing to 0 insures that we look at the remainder of the
7737 * non-exhausted set */
7738 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7739 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7744 /* The final length is what we've output so far, plus what else is about to
7745 * be output. (If 'count' is non-zero, then the input list we exhausted
7746 * has everything remaining up to the machine's limit in its set, and hence
7747 * in the union, so there will be no further output. */
7750 /* At most one of the subexpressions will be non-zero */
7751 len_u += (len_a - i_a) + (len_b - i_b);
7754 /* Set result to final length, which can change the pointer to array_u, so
7756 if (len_u != _invlist_len(u)) {
7757 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
7759 array_u = invlist_array(u);
7762 /* When 'count' is 0, the list that was exhausted (if one was shorter than
7763 * the other) ended with everything above it not in its set. That means
7764 * that the remaining part of the union is precisely the same as the
7765 * non-exhausted list, so can just copy it unchanged. (If both list were
7766 * exhausted at the same time, then the operations below will be both 0.)
7769 IV copy_count; /* At most one will have a non-zero copy count */
7770 if ((copy_count = len_a - i_a) > 0) {
7771 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
7773 else if ((copy_count = len_b - i_b) > 0) {
7774 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
7778 /* We may be removing a reference to one of the inputs */
7779 if (a == *output || b == *output) {
7780 assert(! invlist_is_iterating(*output));
7781 SvREFCNT_dec_NN(*output);
7789 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, const bool complement_b, SV** i)
7791 /* Take the intersection of two inversion lists and point <i> to it. *i
7792 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7793 * the reference count to that list will be decremented.
7794 * If <complement_b> is TRUE, the result will be the intersection of <a>
7795 * and the complement (or inversion) of <b> instead of <b> directly.
7797 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7798 * Richard Gillam, published by Addison-Wesley, and explained at some
7799 * length there. The preface says to incorporate its examples into your
7800 * code at your own risk. In fact, it had bugs
7802 * The algorithm is like a merge sort, and is essentially the same as the
7806 const UV* array_a; /* a's array */
7808 UV len_a; /* length of a's array */
7811 SV* r; /* the resulting intersection */
7815 UV i_a = 0; /* current index into a's array */
7819 /* running count, as explained in the algorithm source book; items are
7820 * stopped accumulating and are output when the count changes to/from 2.
7821 * The count is incremented when we start a range that's in the set, and
7822 * decremented when we start a range that's not in the set. So its range
7823 * is 0 to 2. Only when the count is 2 is something in the intersection.
7827 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
7830 /* Special case if either one is empty */
7831 len_a = _invlist_len(a);
7832 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
7834 if (len_a != 0 && complement_b) {
7836 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
7837 * be empty. Here, also we are using 'b's complement, which hence
7838 * must be every possible code point. Thus the intersection is
7841 *i = invlist_clone(a);
7847 /* else *i is already 'a' */
7851 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
7852 * intersection must be empty */
7859 *i = _new_invlist(0);
7863 /* Here both lists exist and are non-empty */
7864 array_a = invlist_array(a);
7865 array_b = invlist_array(b);
7867 /* If are to take the intersection of 'a' with the complement of b, set it
7868 * up so are looking at b's complement. */
7871 /* To complement, we invert: if the first element is 0, remove it. To
7872 * do this, we just pretend the array starts one later */
7873 if (array_b[0] == 0) {
7879 /* But if the first element is not zero, we pretend the list starts
7880 * at the 0 that is always stored immediately before the array. */
7886 /* Size the intersection for the worst case: that the intersection ends up
7887 * fragmenting everything to be completely disjoint */
7888 r= _new_invlist(len_a + len_b);
7890 /* Will contain U+0000 iff both components do */
7891 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
7892 && len_b > 0 && array_b[0] == 0);
7894 /* Go through each list item by item, stopping when exhausted one of
7896 while (i_a < len_a && i_b < len_b) {
7897 UV cp; /* The element to potentially add to the intersection's
7899 bool cp_in_set; /* Is it in the input list's set or not */
7901 /* We need to take one or the other of the two inputs for the
7902 * intersection. Since we are merging two sorted lists, we take the
7903 * smaller of the next items. In case of a tie, we take the one that
7904 * is not in its set first (a difference from the union algorithm). If
7905 * we took one in the set first, it would increment the count, possibly
7906 * to 2 which would cause it to be output as starting a range in the
7907 * intersection, and the next time through we would take that same
7908 * number, and output it again as ending the set. By doing it the
7909 * opposite of this, there is no possibility that the count will be
7910 * momentarily incremented to 2. (In a tie and both are in the set or
7911 * both not in the set, it doesn't matter which we take first.) */
7912 if (array_a[i_a] < array_b[i_b]
7913 || (array_a[i_a] == array_b[i_b]
7914 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7916 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7920 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7924 /* Here, have chosen which of the two inputs to look at. Only output
7925 * if the running count changes to/from 2, which marks the
7926 * beginning/end of a range that's in the intersection */
7930 array_r[i_r++] = cp;
7935 array_r[i_r++] = cp;
7941 /* Here, we are finished going through at least one of the lists, which
7942 * means there is something remaining in at most one. We check if the list
7943 * that has been exhausted is positioned such that we are in the middle
7944 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
7945 * the ones we care about.) There are four cases:
7946 * 1) Both weren't in their sets, count is 0, and remains 0. There's
7947 * nothing left in the intersection.
7948 * 2) Both were in their sets, count is 2 and perhaps is incremented to
7949 * above 2. What should be output is exactly that which is in the
7950 * non-exhausted set, as everything it has is also in the intersection
7951 * set, and everything it doesn't have can't be in the intersection
7952 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
7953 * gets incremented to 2. Like the previous case, the intersection is
7954 * everything that remains in the non-exhausted set.
7955 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
7956 * remains 1. And the intersection has nothing more. */
7957 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7958 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7963 /* The final length is what we've output so far plus what else is in the
7964 * intersection. At most one of the subexpressions below will be non-zero */
7967 len_r += (len_a - i_a) + (len_b - i_b);
7970 /* Set result to final length, which can change the pointer to array_r, so
7972 if (len_r != _invlist_len(r)) {
7973 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
7975 array_r = invlist_array(r);
7978 /* Finish outputting any remaining */
7979 if (count >= 2) { /* At most one will have a non-zero copy count */
7981 if ((copy_count = len_a - i_a) > 0) {
7982 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
7984 else if ((copy_count = len_b - i_b) > 0) {
7985 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
7989 /* We may be removing a reference to one of the inputs */
7990 if (a == *i || b == *i) {
7991 assert(! invlist_is_iterating(*i));
7992 SvREFCNT_dec_NN(*i);
8000 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
8002 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8003 * set. A pointer to the inversion list is returned. This may actually be
8004 * a new list, in which case the passed in one has been destroyed. The
8005 * passed in inversion list can be NULL, in which case a new one is created
8006 * with just the one range in it */
8011 if (invlist == NULL) {
8012 invlist = _new_invlist(2);
8016 len = _invlist_len(invlist);
8019 /* If comes after the final entry actually in the list, can just append it
8022 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
8023 && start >= invlist_array(invlist)[len - 1]))
8025 _append_range_to_invlist(invlist, start, end);
8029 /* Here, can't just append things, create and return a new inversion list
8030 * which is the union of this range and the existing inversion list */
8031 range_invlist = _new_invlist(2);
8032 _append_range_to_invlist(range_invlist, start, end);
8034 _invlist_union(invlist, range_invlist, &invlist);
8036 /* The temporary can be freed */
8037 SvREFCNT_dec_NN(range_invlist);
8044 PERL_STATIC_INLINE SV*
8045 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
8046 return _add_range_to_invlist(invlist, cp, cp);
8049 #ifndef PERL_IN_XSUB_RE
8051 Perl__invlist_invert(pTHX_ SV* const invlist)
8053 /* Complement the input inversion list. This adds a 0 if the list didn't
8054 * have a zero; removes it otherwise. As described above, the data
8055 * structure is set up so that this is very efficient */
8057 PERL_ARGS_ASSERT__INVLIST_INVERT;
8059 assert(! invlist_is_iterating(invlist));
8061 /* The inverse of matching nothing is matching everything */
8062 if (_invlist_len(invlist) == 0) {
8063 _append_range_to_invlist(invlist, 0, UV_MAX);
8067 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
8071 Perl__invlist_invert_prop(pTHX_ SV* const invlist)
8073 /* Complement the input inversion list (which must be a Unicode property,
8074 * all of which don't match above the Unicode maximum code point.) And
8075 * Perl has chosen to not have the inversion match above that either. This
8076 * adds a 0x110000 if the list didn't end with it, and removes it if it did
8082 PERL_ARGS_ASSERT__INVLIST_INVERT_PROP;
8084 _invlist_invert(invlist);
8086 len = _invlist_len(invlist);
8088 if (len != 0) { /* If empty do nothing */
8089 array = invlist_array(invlist);
8090 if (array[len - 1] != PERL_UNICODE_MAX + 1) {
8091 /* Add 0x110000. First, grow if necessary */
8093 if (invlist_max(invlist) < len) {
8094 invlist_extend(invlist, len);
8095 array = invlist_array(invlist);
8097 invlist_set_len(invlist, len, *get_invlist_offset_addr(invlist));
8098 array[len - 1] = PERL_UNICODE_MAX + 1;
8100 else { /* Remove the 0x110000 */
8101 invlist_set_len(invlist, len - 1, *get_invlist_offset_addr(invlist));
8109 PERL_STATIC_INLINE SV*
8110 S_invlist_clone(pTHX_ SV* const invlist)
8113 /* Return a new inversion list that is a copy of the input one, which is
8116 /* Need to allocate extra space to accommodate Perl's addition of a
8117 * trailing NUL to SvPV's, since it thinks they are always strings */
8118 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8119 STRLEN physical_length = SvCUR(invlist);
8120 bool offset = *(get_invlist_offset_addr(invlist));
8122 PERL_ARGS_ASSERT_INVLIST_CLONE;
8124 *(get_invlist_offset_addr(new_invlist)) = offset;
8125 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
8126 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
8131 PERL_STATIC_INLINE STRLEN*
8132 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8134 /* Return the address of the UV that contains the current iteration
8137 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8139 return &(((XINVLIST*) SvANY(invlist))->iterator);
8142 PERL_STATIC_INLINE void
8143 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8145 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8147 *get_invlist_iter_addr(invlist) = 0;
8150 PERL_STATIC_INLINE void
8151 S_invlist_iterfinish(pTHX_ SV* invlist)
8153 /* Terminate iterator for invlist. This is to catch development errors.
8154 * Any iteration that is interrupted before completed should call this
8155 * function. Functions that add code points anywhere else but to the end
8156 * of an inversion list assert that they are not in the middle of an
8157 * iteration. If they were, the addition would make the iteration
8158 * problematical: if the iteration hadn't reached the place where things
8159 * were being added, it would be ok */
8161 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8163 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
8167 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8169 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8170 * This call sets in <*start> and <*end>, the next range in <invlist>.
8171 * Returns <TRUE> if successful and the next call will return the next
8172 * range; <FALSE> if was already at the end of the list. If the latter,
8173 * <*start> and <*end> are unchanged, and the next call to this function
8174 * will start over at the beginning of the list */
8176 STRLEN* pos = get_invlist_iter_addr(invlist);
8177 UV len = _invlist_len(invlist);
8180 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8183 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
8187 array = invlist_array(invlist);
8189 *start = array[(*pos)++];
8195 *end = array[(*pos)++] - 1;
8201 PERL_STATIC_INLINE bool
8202 S_invlist_is_iterating(pTHX_ SV* const invlist)
8204 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8206 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8209 PERL_STATIC_INLINE UV
8210 S_invlist_highest(pTHX_ SV* const invlist)
8212 /* Returns the highest code point that matches an inversion list. This API
8213 * has an ambiguity, as it returns 0 under either the highest is actually
8214 * 0, or if the list is empty. If this distinction matters to you, check
8215 * for emptiness before calling this function */
8217 UV len = _invlist_len(invlist);
8220 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8226 array = invlist_array(invlist);
8228 /* The last element in the array in the inversion list always starts a
8229 * range that goes to infinity. That range may be for code points that are
8230 * matched in the inversion list, or it may be for ones that aren't
8231 * matched. In the latter case, the highest code point in the set is one
8232 * less than the beginning of this range; otherwise it is the final element
8233 * of this range: infinity */
8234 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8236 : array[len - 1] - 1;
8239 #ifndef PERL_IN_XSUB_RE
8241 Perl__invlist_contents(pTHX_ SV* const invlist)
8243 /* Get the contents of an inversion list into a string SV so that they can
8244 * be printed out. It uses the format traditionally done for debug tracing
8248 SV* output = newSVpvs("\n");
8250 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8252 assert(! invlist_is_iterating(invlist));
8254 invlist_iterinit(invlist);
8255 while (invlist_iternext(invlist, &start, &end)) {
8256 if (end == UV_MAX) {
8257 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8259 else if (end != start) {
8260 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8264 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
8272 #ifdef PERL_ARGS_ASSERT__INVLIST_DUMP
8274 Perl__invlist_dump(pTHX_ SV* const invlist, const char * const header)
8276 /* Dumps out the ranges in an inversion list. The string 'header'
8277 * if present is output on a line before the first range */
8281 PERL_ARGS_ASSERT__INVLIST_DUMP;
8283 if (header && strlen(header)) {
8284 PerlIO_printf(Perl_debug_log, "%s\n", header);
8286 if (invlist_is_iterating(invlist)) {
8287 PerlIO_printf(Perl_debug_log, "Can't dump because is in middle of iterating\n");
8291 invlist_iterinit(invlist);
8292 while (invlist_iternext(invlist, &start, &end)) {
8293 if (end == UV_MAX) {
8294 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. INFINITY\n", start);
8296 else if (end != start) {
8297 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. 0x%04"UVXf"\n",
8301 PerlIO_printf(Perl_debug_log, "0x%04"UVXf"\n", start);
8309 S__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
8311 /* Return a boolean as to if the two passed in inversion lists are
8312 * identical. The final argument, if TRUE, says to take the complement of
8313 * the second inversion list before doing the comparison */
8315 const UV* array_a = invlist_array(a);
8316 const UV* array_b = invlist_array(b);
8317 UV len_a = _invlist_len(a);
8318 UV len_b = _invlist_len(b);
8320 UV i = 0; /* current index into the arrays */
8321 bool retval = TRUE; /* Assume are identical until proven otherwise */
8323 PERL_ARGS_ASSERT__INVLISTEQ;
8325 /* If are to compare 'a' with the complement of b, set it
8326 * up so are looking at b's complement. */
8329 /* The complement of nothing is everything, so <a> would have to have
8330 * just one element, starting at zero (ending at infinity) */
8332 return (len_a == 1 && array_a[0] == 0);
8334 else if (array_b[0] == 0) {
8336 /* Otherwise, to complement, we invert. Here, the first element is
8337 * 0, just remove it. To do this, we just pretend the array starts
8345 /* But if the first element is not zero, we pretend the list starts
8346 * at the 0 that is always stored immediately before the array. */
8353 /* Make sure that the lengths are the same, as well as the final element
8354 * before looping through the remainder. (Thus we test the length, final,
8355 * and first elements right off the bat) */
8356 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
8359 else for (i = 0; i < len_a - 1; i++) {
8360 if (array_a[i] != array_b[i]) {
8370 #undef HEADER_LENGTH
8371 #undef TO_INTERNAL_SIZE
8372 #undef FROM_INTERNAL_SIZE
8373 #undef INVLIST_VERSION_ID
8375 /* End of inversion list object */
8378 S_parse_lparen_question_flags(pTHX_ struct RExC_state_t *pRExC_state)
8380 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
8381 * constructs, and updates RExC_flags with them. On input, RExC_parse
8382 * should point to the first flag; it is updated on output to point to the
8383 * final ')' or ':'. There needs to be at least one flag, or this will
8386 /* for (?g), (?gc), and (?o) warnings; warning
8387 about (?c) will warn about (?g) -- japhy */
8389 #define WASTED_O 0x01
8390 #define WASTED_G 0x02
8391 #define WASTED_C 0x04
8392 #define WASTED_GC (WASTED_G|WASTED_C)
8393 I32 wastedflags = 0x00;
8394 U32 posflags = 0, negflags = 0;
8395 U32 *flagsp = &posflags;
8396 char has_charset_modifier = '\0';
8398 bool has_use_defaults = FALSE;
8399 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
8401 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
8403 /* '^' as an initial flag sets certain defaults */
8404 if (UCHARAT(RExC_parse) == '^') {
8406 has_use_defaults = TRUE;
8407 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
8408 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
8409 ? REGEX_UNICODE_CHARSET
8410 : REGEX_DEPENDS_CHARSET);
8413 cs = get_regex_charset(RExC_flags);
8414 if (cs == REGEX_DEPENDS_CHARSET
8415 && (RExC_utf8 || RExC_uni_semantics))
8417 cs = REGEX_UNICODE_CHARSET;
8420 while (*RExC_parse) {
8421 /* && strchr("iogcmsx", *RExC_parse) */
8422 /* (?g), (?gc) and (?o) are useless here
8423 and must be globally applied -- japhy */
8424 switch (*RExC_parse) {
8426 /* Code for the imsx flags */
8427 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
8429 case LOCALE_PAT_MOD:
8430 if (has_charset_modifier) {
8431 goto excess_modifier;
8433 else if (flagsp == &negflags) {
8436 cs = REGEX_LOCALE_CHARSET;
8437 has_charset_modifier = LOCALE_PAT_MOD;
8438 RExC_contains_locale = 1;
8440 case UNICODE_PAT_MOD:
8441 if (has_charset_modifier) {
8442 goto excess_modifier;
8444 else if (flagsp == &negflags) {
8447 cs = REGEX_UNICODE_CHARSET;
8448 has_charset_modifier = UNICODE_PAT_MOD;
8450 case ASCII_RESTRICT_PAT_MOD:
8451 if (flagsp == &negflags) {
8454 if (has_charset_modifier) {
8455 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
8456 goto excess_modifier;
8458 /* Doubled modifier implies more restricted */
8459 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
8462 cs = REGEX_ASCII_RESTRICTED_CHARSET;
8464 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
8466 case DEPENDS_PAT_MOD:
8467 if (has_use_defaults) {
8468 goto fail_modifiers;
8470 else if (flagsp == &negflags) {
8473 else if (has_charset_modifier) {
8474 goto excess_modifier;
8477 /* The dual charset means unicode semantics if the
8478 * pattern (or target, not known until runtime) are
8479 * utf8, or something in the pattern indicates unicode
8481 cs = (RExC_utf8 || RExC_uni_semantics)
8482 ? REGEX_UNICODE_CHARSET
8483 : REGEX_DEPENDS_CHARSET;
8484 has_charset_modifier = DEPENDS_PAT_MOD;
8488 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
8489 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
8491 else if (has_charset_modifier == *(RExC_parse - 1)) {
8492 vFAIL2("Regexp modifier \"%c\" may not appear twice", *(RExC_parse - 1));
8495 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
8500 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"", *(RExC_parse - 1));
8502 case ONCE_PAT_MOD: /* 'o' */
8503 case GLOBAL_PAT_MOD: /* 'g' */
8504 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8505 const I32 wflagbit = *RExC_parse == 'o' ? WASTED_O : WASTED_G;
8506 if (! (wastedflags & wflagbit) ) {
8507 wastedflags |= wflagbit;
8508 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
8511 "Useless (%s%c) - %suse /%c modifier",
8512 flagsp == &negflags ? "?-" : "?",
8514 flagsp == &negflags ? "don't " : "",
8521 case CONTINUE_PAT_MOD: /* 'c' */
8522 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8523 if (! (wastedflags & WASTED_C) ) {
8524 wastedflags |= WASTED_GC;
8525 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
8528 "Useless (%sc) - %suse /gc modifier",
8529 flagsp == &negflags ? "?-" : "?",
8530 flagsp == &negflags ? "don't " : ""
8535 case KEEPCOPY_PAT_MOD: /* 'p' */
8536 if (flagsp == &negflags) {
8538 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
8540 *flagsp |= RXf_PMf_KEEPCOPY;
8544 /* A flag is a default iff it is following a minus, so
8545 * if there is a minus, it means will be trying to
8546 * re-specify a default which is an error */
8547 if (has_use_defaults || flagsp == &negflags) {
8548 goto fail_modifiers;
8551 wastedflags = 0; /* reset so (?g-c) warns twice */
8555 RExC_flags |= posflags;
8556 RExC_flags &= ~negflags;
8557 set_regex_charset(&RExC_flags, cs);
8563 vFAIL3("Sequence (%.*s...) not recognized",
8564 RExC_parse-seqstart, seqstart);
8573 - reg - regular expression, i.e. main body or parenthesized thing
8575 * Caller must absorb opening parenthesis.
8577 * Combining parenthesis handling with the base level of regular expression
8578 * is a trifle forced, but the need to tie the tails of the branches to what
8579 * follows makes it hard to avoid.
8581 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
8583 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
8585 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
8588 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
8589 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
8590 needs to be restarted.
8591 Otherwise would only return NULL if regbranch() returns NULL, which
8594 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
8595 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
8596 * 2 is like 1, but indicates that nextchar() has been called to advance
8597 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
8598 * this flag alerts us to the need to check for that */
8601 regnode *ret; /* Will be the head of the group. */
8604 regnode *ender = NULL;
8607 U32 oregflags = RExC_flags;
8608 bool have_branch = 0;
8610 I32 freeze_paren = 0;
8611 I32 after_freeze = 0;
8613 char * parse_start = RExC_parse; /* MJD */
8614 char * const oregcomp_parse = RExC_parse;
8616 GET_RE_DEBUG_FLAGS_DECL;
8618 PERL_ARGS_ASSERT_REG;
8619 DEBUG_PARSE("reg ");
8621 *flagp = 0; /* Tentatively. */
8624 /* Make an OPEN node, if parenthesized. */
8627 /* Under /x, space and comments can be gobbled up between the '(' and
8628 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
8629 * intervening space, as the sequence is a token, and a token should be
8631 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
8633 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
8634 char *start_verb = RExC_parse;
8635 STRLEN verb_len = 0;
8636 char *start_arg = NULL;
8637 unsigned char op = 0;
8639 int internal_argval = 0; /* internal_argval is only useful if !argok */
8641 if (has_intervening_patws && SIZE_ONLY) {
8642 ckWARNregdep(RExC_parse + 1, "In '(*VERB...)', splitting the initial '(*' is deprecated");
8644 while ( *RExC_parse && *RExC_parse != ')' ) {
8645 if ( *RExC_parse == ':' ) {
8646 start_arg = RExC_parse + 1;
8652 verb_len = RExC_parse - start_verb;
8655 while ( *RExC_parse && *RExC_parse != ')' )
8657 if ( *RExC_parse != ')' )
8658 vFAIL("Unterminated verb pattern argument");
8659 if ( RExC_parse == start_arg )
8662 if ( *RExC_parse != ')' )
8663 vFAIL("Unterminated verb pattern");
8666 switch ( *start_verb ) {
8667 case 'A': /* (*ACCEPT) */
8668 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
8670 internal_argval = RExC_nestroot;
8673 case 'C': /* (*COMMIT) */
8674 if ( memEQs(start_verb,verb_len,"COMMIT") )
8677 case 'F': /* (*FAIL) */
8678 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
8683 case ':': /* (*:NAME) */
8684 case 'M': /* (*MARK:NAME) */
8685 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
8690 case 'P': /* (*PRUNE) */
8691 if ( memEQs(start_verb,verb_len,"PRUNE") )
8694 case 'S': /* (*SKIP) */
8695 if ( memEQs(start_verb,verb_len,"SKIP") )
8698 case 'T': /* (*THEN) */
8699 /* [19:06] <TimToady> :: is then */
8700 if ( memEQs(start_verb,verb_len,"THEN") ) {
8702 RExC_seen |= REG_SEEN_CUTGROUP;
8708 vFAIL3("Unknown verb pattern '%.*s'",
8709 verb_len, start_verb);
8712 if ( start_arg && internal_argval ) {
8713 vFAIL3("Verb pattern '%.*s' may not have an argument",
8714 verb_len, start_verb);
8715 } else if ( argok < 0 && !start_arg ) {
8716 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
8717 verb_len, start_verb);
8719 ret = reganode(pRExC_state, op, internal_argval);
8720 if ( ! internal_argval && ! SIZE_ONLY ) {
8722 SV *sv = newSVpvn( start_arg, RExC_parse - start_arg);
8723 ARG(ret) = add_data( pRExC_state, 1, "S" );
8724 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
8731 if (!internal_argval)
8732 RExC_seen |= REG_SEEN_VERBARG;
8733 } else if ( start_arg ) {
8734 vFAIL3("Verb pattern '%.*s' may not have an argument",
8735 verb_len, start_verb);
8737 ret = reg_node(pRExC_state, op);
8739 nextchar(pRExC_state);
8742 else if (*RExC_parse == '?') { /* (?...) */
8743 bool is_logical = 0;
8744 const char * const seqstart = RExC_parse;
8745 if (has_intervening_patws && SIZE_ONLY) {
8746 ckWARNregdep(RExC_parse + 1, "In '(?...)', splitting the initial '(?' is deprecated");
8750 paren = *RExC_parse++;
8751 ret = NULL; /* For look-ahead/behind. */
8754 case 'P': /* (?P...) variants for those used to PCRE/Python */
8755 paren = *RExC_parse++;
8756 if ( paren == '<') /* (?P<...>) named capture */
8758 else if (paren == '>') { /* (?P>name) named recursion */
8759 goto named_recursion;
8761 else if (paren == '=') { /* (?P=...) named backref */
8762 /* this pretty much dupes the code for \k<NAME> in regatom(), if
8763 you change this make sure you change that */
8764 char* name_start = RExC_parse;
8766 SV *sv_dat = reg_scan_name(pRExC_state,
8767 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8768 if (RExC_parse == name_start || *RExC_parse != ')')
8769 vFAIL2("Sequence %.3s... not terminated",parse_start);
8772 num = add_data( pRExC_state, 1, "S" );
8773 RExC_rxi->data->data[num]=(void*)sv_dat;
8774 SvREFCNT_inc_simple_void(sv_dat);
8777 ret = reganode(pRExC_state,
8780 : (ASCII_FOLD_RESTRICTED)
8782 : (AT_LEAST_UNI_SEMANTICS)
8790 Set_Node_Offset(ret, parse_start+1);
8791 Set_Node_Cur_Length(ret, parse_start);
8793 nextchar(pRExC_state);
8797 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8799 case '<': /* (?<...) */
8800 if (*RExC_parse == '!')
8802 else if (*RExC_parse != '=')
8808 case '\'': /* (?'...') */
8809 name_start= RExC_parse;
8810 svname = reg_scan_name(pRExC_state,
8811 SIZE_ONLY ? /* reverse test from the others */
8812 REG_RSN_RETURN_NAME :
8813 REG_RSN_RETURN_NULL);
8814 if (RExC_parse == name_start) {
8816 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8819 if (*RExC_parse != paren)
8820 vFAIL2("Sequence (?%c... not terminated",
8821 paren=='>' ? '<' : paren);
8825 if (!svname) /* shouldn't happen */
8827 "panic: reg_scan_name returned NULL");
8828 if (!RExC_paren_names) {
8829 RExC_paren_names= newHV();
8830 sv_2mortal(MUTABLE_SV(RExC_paren_names));
8832 RExC_paren_name_list= newAV();
8833 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
8836 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
8838 sv_dat = HeVAL(he_str);
8840 /* croak baby croak */
8842 "panic: paren_name hash element allocation failed");
8843 } else if ( SvPOK(sv_dat) ) {
8844 /* (?|...) can mean we have dupes so scan to check
8845 its already been stored. Maybe a flag indicating
8846 we are inside such a construct would be useful,
8847 but the arrays are likely to be quite small, so
8848 for now we punt -- dmq */
8849 IV count = SvIV(sv_dat);
8850 I32 *pv = (I32*)SvPVX(sv_dat);
8852 for ( i = 0 ; i < count ; i++ ) {
8853 if ( pv[i] == RExC_npar ) {
8859 pv = (I32*)SvGROW(sv_dat, SvCUR(sv_dat) + sizeof(I32)+1);
8860 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
8861 pv[count] = RExC_npar;
8862 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
8865 (void)SvUPGRADE(sv_dat,SVt_PVNV);
8866 sv_setpvn(sv_dat, (char *)&(RExC_npar), sizeof(I32));
8868 SvIV_set(sv_dat, 1);
8871 /* Yes this does cause a memory leak in debugging Perls */
8872 if (!av_store(RExC_paren_name_list, RExC_npar, SvREFCNT_inc(svname)))
8873 SvREFCNT_dec_NN(svname);
8876 /*sv_dump(sv_dat);*/
8878 nextchar(pRExC_state);
8880 goto capturing_parens;
8882 RExC_seen |= REG_SEEN_LOOKBEHIND;
8883 RExC_in_lookbehind++;
8885 case '=': /* (?=...) */
8886 RExC_seen_zerolen++;
8888 case '!': /* (?!...) */
8889 RExC_seen_zerolen++;
8890 if (*RExC_parse == ')') {
8891 ret=reg_node(pRExC_state, OPFAIL);
8892 nextchar(pRExC_state);
8896 case '|': /* (?|...) */
8897 /* branch reset, behave like a (?:...) except that
8898 buffers in alternations share the same numbers */
8900 after_freeze = freeze_paren = RExC_npar;
8902 case ':': /* (?:...) */
8903 case '>': /* (?>...) */
8905 case '$': /* (?$...) */
8906 case '@': /* (?@...) */
8907 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
8909 case '#': /* (?#...) */
8910 /* XXX As soon as we disallow separating the '?' and '*' (by
8911 * spaces or (?#...) comment), it is believed that this case
8912 * will be unreachable and can be removed. See
8914 while (*RExC_parse && *RExC_parse != ')')
8916 if (*RExC_parse != ')')
8917 FAIL("Sequence (?#... not terminated");
8918 nextchar(pRExC_state);
8921 case '0' : /* (?0) */
8922 case 'R' : /* (?R) */
8923 if (*RExC_parse != ')')
8924 FAIL("Sequence (?R) not terminated");
8925 ret = reg_node(pRExC_state, GOSTART);
8926 *flagp |= POSTPONED;
8927 nextchar(pRExC_state);
8930 { /* named and numeric backreferences */
8932 case '&': /* (?&NAME) */
8933 parse_start = RExC_parse - 1;
8936 SV *sv_dat = reg_scan_name(pRExC_state,
8937 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8938 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8940 goto gen_recurse_regop;
8941 assert(0); /* NOT REACHED */
8943 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8945 vFAIL("Illegal pattern");
8947 goto parse_recursion;
8949 case '-': /* (?-1) */
8950 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8951 RExC_parse--; /* rewind to let it be handled later */
8955 case '1': case '2': case '3': case '4': /* (?1) */
8956 case '5': case '6': case '7': case '8': case '9':
8959 num = atoi(RExC_parse);
8960 parse_start = RExC_parse - 1; /* MJD */
8961 if (*RExC_parse == '-')
8963 while (isDIGIT(*RExC_parse))
8965 if (*RExC_parse!=')')
8966 vFAIL("Expecting close bracket");
8969 if ( paren == '-' ) {
8971 Diagram of capture buffer numbering.
8972 Top line is the normal capture buffer numbers
8973 Bottom line is the negative indexing as from
8977 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
8981 num = RExC_npar + num;
8984 vFAIL("Reference to nonexistent group");
8986 } else if ( paren == '+' ) {
8987 num = RExC_npar + num - 1;
8990 ret = reganode(pRExC_state, GOSUB, num);
8992 if (num > (I32)RExC_rx->nparens) {
8994 vFAIL("Reference to nonexistent group");
8996 ARG2L_SET( ret, RExC_recurse_count++);
8998 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
8999 "Recurse #%"UVuf" to %"IVdf"\n", (UV)ARG(ret), (IV)ARG2L(ret)));
9003 RExC_seen |= REG_SEEN_RECURSE;
9004 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
9005 Set_Node_Offset(ret, parse_start); /* MJD */
9007 *flagp |= POSTPONED;
9008 nextchar(pRExC_state);
9010 } /* named and numeric backreferences */
9011 assert(0); /* NOT REACHED */
9013 case '?': /* (??...) */
9015 if (*RExC_parse != '{') {
9017 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
9020 *flagp |= POSTPONED;
9021 paren = *RExC_parse++;
9023 case '{': /* (?{...}) */
9026 struct reg_code_block *cb;
9028 RExC_seen_zerolen++;
9030 if ( !pRExC_state->num_code_blocks
9031 || pRExC_state->code_index >= pRExC_state->num_code_blocks
9032 || pRExC_state->code_blocks[pRExC_state->code_index].start
9033 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
9036 if (RExC_pm_flags & PMf_USE_RE_EVAL)
9037 FAIL("panic: Sequence (?{...}): no code block found\n");
9038 FAIL("Eval-group not allowed at runtime, use re 'eval'");
9040 /* this is a pre-compiled code block (?{...}) */
9041 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
9042 RExC_parse = RExC_start + cb->end;
9045 if (cb->src_regex) {
9046 n = add_data(pRExC_state, 2, "rl");
9047 RExC_rxi->data->data[n] =
9048 (void*)SvREFCNT_inc((SV*)cb->src_regex);
9049 RExC_rxi->data->data[n+1] = (void*)o;
9052 n = add_data(pRExC_state, 1,
9053 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l");
9054 RExC_rxi->data->data[n] = (void*)o;
9057 pRExC_state->code_index++;
9058 nextchar(pRExC_state);
9062 ret = reg_node(pRExC_state, LOGICAL);
9063 eval = reganode(pRExC_state, EVAL, n);
9066 /* for later propagation into (??{}) return value */
9067 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
9069 REGTAIL(pRExC_state, ret, eval);
9070 /* deal with the length of this later - MJD */
9073 ret = reganode(pRExC_state, EVAL, n);
9074 Set_Node_Length(ret, RExC_parse - parse_start + 1);
9075 Set_Node_Offset(ret, parse_start);
9078 case '(': /* (?(?{...})...) and (?(?=...)...) */
9081 if (RExC_parse[0] == '?') { /* (?(?...)) */
9082 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
9083 || RExC_parse[1] == '<'
9084 || RExC_parse[1] == '{') { /* Lookahead or eval. */
9088 ret = reg_node(pRExC_state, LOGICAL);
9092 tail = reg(pRExC_state, 1, &flag, depth+1);
9093 if (flag & RESTART_UTF8) {
9094 *flagp = RESTART_UTF8;
9097 REGTAIL(pRExC_state, ret, tail);
9101 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9102 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9104 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9105 char *name_start= RExC_parse++;
9107 SV *sv_dat=reg_scan_name(pRExC_state,
9108 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9109 if (RExC_parse == name_start || *RExC_parse != ch)
9110 vFAIL2("Sequence (?(%c... not terminated",
9111 (ch == '>' ? '<' : ch));
9114 num = add_data( pRExC_state, 1, "S" );
9115 RExC_rxi->data->data[num]=(void*)sv_dat;
9116 SvREFCNT_inc_simple_void(sv_dat);
9118 ret = reganode(pRExC_state,NGROUPP,num);
9119 goto insert_if_check_paren;
9121 else if (RExC_parse[0] == 'D' &&
9122 RExC_parse[1] == 'E' &&
9123 RExC_parse[2] == 'F' &&
9124 RExC_parse[3] == 'I' &&
9125 RExC_parse[4] == 'N' &&
9126 RExC_parse[5] == 'E')
9128 ret = reganode(pRExC_state,DEFINEP,0);
9131 goto insert_if_check_paren;
9133 else if (RExC_parse[0] == 'R') {
9136 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9137 parno = atoi(RExC_parse++);
9138 while (isDIGIT(*RExC_parse))
9140 } else if (RExC_parse[0] == '&') {
9143 sv_dat = reg_scan_name(pRExC_state,
9144 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9145 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9147 ret = reganode(pRExC_state,INSUBP,parno);
9148 goto insert_if_check_paren;
9150 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9153 parno = atoi(RExC_parse++);
9155 while (isDIGIT(*RExC_parse))
9157 ret = reganode(pRExC_state, GROUPP, parno);
9159 insert_if_check_paren:
9160 if ((c = *nextchar(pRExC_state)) != ')')
9161 vFAIL("Switch condition not recognized");
9163 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9164 br = regbranch(pRExC_state, &flags, 1,depth+1);
9166 if (flags & RESTART_UTF8) {
9167 *flagp = RESTART_UTF8;
9170 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9173 REGTAIL(pRExC_state, br, reganode(pRExC_state, LONGJMP, 0));
9174 c = *nextchar(pRExC_state);
9179 vFAIL("(?(DEFINE)....) does not allow branches");
9180 lastbr = reganode(pRExC_state, IFTHEN, 0); /* Fake one for optimizer. */
9181 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
9182 if (flags & RESTART_UTF8) {
9183 *flagp = RESTART_UTF8;
9186 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9189 REGTAIL(pRExC_state, ret, lastbr);
9192 c = *nextchar(pRExC_state);
9197 vFAIL("Switch (?(condition)... contains too many branches");
9198 ender = reg_node(pRExC_state, TAIL);
9199 REGTAIL(pRExC_state, br, ender);
9201 REGTAIL(pRExC_state, lastbr, ender);
9202 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
9205 REGTAIL(pRExC_state, ret, ender);
9206 RExC_size++; /* XXX WHY do we need this?!!
9207 For large programs it seems to be required
9208 but I can't figure out why. -- dmq*/
9212 vFAIL2("Unknown switch condition (?(%.2s", RExC_parse);
9215 case '[': /* (?[ ... ]) */
9216 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
9219 RExC_parse--; /* for vFAIL to print correctly */
9220 vFAIL("Sequence (? incomplete");
9222 default: /* e.g., (?i) */
9225 parse_lparen_question_flags(pRExC_state);
9226 if (UCHARAT(RExC_parse) != ':') {
9227 nextchar(pRExC_state);
9232 nextchar(pRExC_state);
9242 ret = reganode(pRExC_state, OPEN, parno);
9245 RExC_nestroot = parno;
9246 if (RExC_seen & REG_SEEN_RECURSE
9247 && !RExC_open_parens[parno-1])
9249 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9250 "Setting open paren #%"IVdf" to %d\n",
9251 (IV)parno, REG_NODE_NUM(ret)));
9252 RExC_open_parens[parno-1]= ret;
9255 Set_Node_Length(ret, 1); /* MJD */
9256 Set_Node_Offset(ret, RExC_parse); /* MJD */
9264 /* Pick up the branches, linking them together. */
9265 parse_start = RExC_parse; /* MJD */
9266 br = regbranch(pRExC_state, &flags, 1,depth+1);
9268 /* branch_len = (paren != 0); */
9271 if (flags & RESTART_UTF8) {
9272 *flagp = RESTART_UTF8;
9275 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
9277 if (*RExC_parse == '|') {
9278 if (!SIZE_ONLY && RExC_extralen) {
9279 reginsert(pRExC_state, BRANCHJ, br, depth+1);
9282 reginsert(pRExC_state, BRANCH, br, depth+1);
9283 Set_Node_Length(br, paren != 0);
9284 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
9288 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
9290 else if (paren == ':') {
9291 *flagp |= flags&SIMPLE;
9293 if (is_open) { /* Starts with OPEN. */
9294 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
9296 else if (paren != '?') /* Not Conditional */
9298 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9300 while (*RExC_parse == '|') {
9301 if (!SIZE_ONLY && RExC_extralen) {
9302 ender = reganode(pRExC_state, LONGJMP,0);
9303 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender); /* Append to the previous. */
9306 RExC_extralen += 2; /* Account for LONGJMP. */
9307 nextchar(pRExC_state);
9309 if (RExC_npar > after_freeze)
9310 after_freeze = RExC_npar;
9311 RExC_npar = freeze_paren;
9313 br = regbranch(pRExC_state, &flags, 0, depth+1);
9316 if (flags & RESTART_UTF8) {
9317 *flagp = RESTART_UTF8;
9320 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
9322 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
9324 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9327 if (have_branch || paren != ':') {
9328 /* Make a closing node, and hook it on the end. */
9331 ender = reg_node(pRExC_state, TAIL);
9334 ender = reganode(pRExC_state, CLOSE, parno);
9335 if (!SIZE_ONLY && RExC_seen & REG_SEEN_RECURSE) {
9336 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9337 "Setting close paren #%"IVdf" to %d\n",
9338 (IV)parno, REG_NODE_NUM(ender)));
9339 RExC_close_parens[parno-1]= ender;
9340 if (RExC_nestroot == parno)
9343 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
9344 Set_Node_Length(ender,1); /* MJD */
9350 *flagp &= ~HASWIDTH;
9353 ender = reg_node(pRExC_state, SUCCEED);
9356 ender = reg_node(pRExC_state, END);
9358 assert(!RExC_opend); /* there can only be one! */
9363 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9364 SV * const mysv_val1=sv_newmortal();
9365 SV * const mysv_val2=sv_newmortal();
9366 DEBUG_PARSE_MSG("lsbr");
9367 regprop(RExC_rx, mysv_val1, lastbr);
9368 regprop(RExC_rx, mysv_val2, ender);
9369 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9370 SvPV_nolen_const(mysv_val1),
9371 (IV)REG_NODE_NUM(lastbr),
9372 SvPV_nolen_const(mysv_val2),
9373 (IV)REG_NODE_NUM(ender),
9374 (IV)(ender - lastbr)
9377 REGTAIL(pRExC_state, lastbr, ender);
9379 if (have_branch && !SIZE_ONLY) {
9382 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
9384 /* Hook the tails of the branches to the closing node. */
9385 for (br = ret; br; br = regnext(br)) {
9386 const U8 op = PL_regkind[OP(br)];
9388 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
9389 if (OP(NEXTOPER(br)) != NOTHING || regnext(NEXTOPER(br)) != ender)
9392 else if (op == BRANCHJ) {
9393 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
9394 /* for now we always disable this optimisation * /
9395 if (OP(NEXTOPER(NEXTOPER(br))) != NOTHING || regnext(NEXTOPER(NEXTOPER(br))) != ender)
9401 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
9402 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9403 SV * const mysv_val1=sv_newmortal();
9404 SV * const mysv_val2=sv_newmortal();
9405 DEBUG_PARSE_MSG("NADA");
9406 regprop(RExC_rx, mysv_val1, ret);
9407 regprop(RExC_rx, mysv_val2, ender);
9408 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9409 SvPV_nolen_const(mysv_val1),
9410 (IV)REG_NODE_NUM(ret),
9411 SvPV_nolen_const(mysv_val2),
9412 (IV)REG_NODE_NUM(ender),
9417 if (OP(ender) == TAIL) {
9422 for ( opt= br + 1; opt < ender ; opt++ )
9424 NEXT_OFF(br)= ender - br;
9432 static const char parens[] = "=!<,>";
9434 if (paren && (p = strchr(parens, paren))) {
9435 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
9436 int flag = (p - parens) > 1;
9439 node = SUSPEND, flag = 0;
9440 reginsert(pRExC_state, node,ret, depth+1);
9441 Set_Node_Cur_Length(ret, parse_start);
9442 Set_Node_Offset(ret, parse_start + 1);
9444 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
9448 /* Check for proper termination. */
9450 /* restore original flags, but keep (?p) */
9451 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
9452 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
9453 RExC_parse = oregcomp_parse;
9454 vFAIL("Unmatched (");
9457 else if (!paren && RExC_parse < RExC_end) {
9458 if (*RExC_parse == ')') {
9460 vFAIL("Unmatched )");
9463 FAIL("Junk on end of regexp"); /* "Can't happen". */
9464 assert(0); /* NOTREACHED */
9467 if (RExC_in_lookbehind) {
9468 RExC_in_lookbehind--;
9470 if (after_freeze > RExC_npar)
9471 RExC_npar = after_freeze;
9476 - regbranch - one alternative of an | operator
9478 * Implements the concatenation operator.
9480 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9484 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
9488 regnode *chain = NULL;
9490 I32 flags = 0, c = 0;
9491 GET_RE_DEBUG_FLAGS_DECL;
9493 PERL_ARGS_ASSERT_REGBRANCH;
9495 DEBUG_PARSE("brnc");
9500 if (!SIZE_ONLY && RExC_extralen)
9501 ret = reganode(pRExC_state, BRANCHJ,0);
9503 ret = reg_node(pRExC_state, BRANCH);
9504 Set_Node_Length(ret, 1);
9508 if (!first && SIZE_ONLY)
9509 RExC_extralen += 1; /* BRANCHJ */
9511 *flagp = WORST; /* Tentatively. */
9514 nextchar(pRExC_state);
9515 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
9517 latest = regpiece(pRExC_state, &flags,depth+1);
9518 if (latest == NULL) {
9519 if (flags & TRYAGAIN)
9521 if (flags & RESTART_UTF8) {
9522 *flagp = RESTART_UTF8;
9525 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
9527 else if (ret == NULL)
9529 *flagp |= flags&(HASWIDTH|POSTPONED);
9530 if (chain == NULL) /* First piece. */
9531 *flagp |= flags&SPSTART;
9534 REGTAIL(pRExC_state, chain, latest);
9539 if (chain == NULL) { /* Loop ran zero times. */
9540 chain = reg_node(pRExC_state, NOTHING);
9545 *flagp |= flags&SIMPLE;
9552 - regpiece - something followed by possible [*+?]
9554 * Note that the branching code sequences used for ? and the general cases
9555 * of * and + are somewhat optimized: they use the same NOTHING node as
9556 * both the endmarker for their branch list and the body of the last branch.
9557 * It might seem that this node could be dispensed with entirely, but the
9558 * endmarker role is not redundant.
9560 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
9562 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9566 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
9573 const char * const origparse = RExC_parse;
9575 I32 max = REG_INFTY;
9576 #ifdef RE_TRACK_PATTERN_OFFSETS
9579 const char *maxpos = NULL;
9581 /* Save the original in case we change the emitted regop to a FAIL. */
9582 regnode * const orig_emit = RExC_emit;
9584 GET_RE_DEBUG_FLAGS_DECL;
9586 PERL_ARGS_ASSERT_REGPIECE;
9588 DEBUG_PARSE("piec");
9590 ret = regatom(pRExC_state, &flags,depth+1);
9592 if (flags & (TRYAGAIN|RESTART_UTF8))
9593 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
9595 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
9601 if (op == '{' && regcurly(RExC_parse, FALSE)) {
9603 #ifdef RE_TRACK_PATTERN_OFFSETS
9604 parse_start = RExC_parse; /* MJD */
9606 next = RExC_parse + 1;
9607 while (isDIGIT(*next) || *next == ',') {
9616 if (*next == '}') { /* got one */
9620 min = atoi(RExC_parse);
9624 maxpos = RExC_parse;
9626 if (!max && *maxpos != '0')
9627 max = REG_INFTY; /* meaning "infinity" */
9628 else if (max >= REG_INFTY)
9629 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
9631 nextchar(pRExC_state);
9632 if (max < min) { /* If can't match, warn and optimize to fail
9635 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
9637 /* We can't back off the size because we have to reserve
9638 * enough space for all the things we are about to throw
9639 * away, but we can shrink it by the ammount we are about
9641 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
9644 RExC_emit = orig_emit;
9646 ret = reg_node(pRExC_state, OPFAIL);
9651 if ((flags&SIMPLE)) {
9652 RExC_naughty += 2 + RExC_naughty / 2;
9653 reginsert(pRExC_state, CURLY, ret, depth+1);
9654 Set_Node_Offset(ret, parse_start+1); /* MJD */
9655 Set_Node_Cur_Length(ret, parse_start);
9658 regnode * const w = reg_node(pRExC_state, WHILEM);
9661 REGTAIL(pRExC_state, ret, w);
9662 if (!SIZE_ONLY && RExC_extralen) {
9663 reginsert(pRExC_state, LONGJMP,ret, depth+1);
9664 reginsert(pRExC_state, NOTHING,ret, depth+1);
9665 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
9667 reginsert(pRExC_state, CURLYX,ret, depth+1);
9669 Set_Node_Offset(ret, parse_start+1);
9670 Set_Node_Length(ret,
9671 op == '{' ? (RExC_parse - parse_start) : 1);
9673 if (!SIZE_ONLY && RExC_extralen)
9674 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
9675 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
9677 RExC_whilem_seen++, RExC_extralen += 3;
9678 RExC_naughty += 4 + RExC_naughty; /* compound interest */
9687 ARG1_SET(ret, (U16)min);
9688 ARG2_SET(ret, (U16)max);
9700 #if 0 /* Now runtime fix should be reliable. */
9702 /* if this is reinstated, don't forget to put this back into perldiag:
9704 =item Regexp *+ operand could be empty at {#} in regex m/%s/
9706 (F) The part of the regexp subject to either the * or + quantifier
9707 could match an empty string. The {#} shows in the regular
9708 expression about where the problem was discovered.
9712 if (!(flags&HASWIDTH) && op != '?')
9713 vFAIL("Regexp *+ operand could be empty");
9716 #ifdef RE_TRACK_PATTERN_OFFSETS
9717 parse_start = RExC_parse;
9719 nextchar(pRExC_state);
9721 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
9723 if (op == '*' && (flags&SIMPLE)) {
9724 reginsert(pRExC_state, STAR, ret, depth+1);
9728 else if (op == '*') {
9732 else if (op == '+' && (flags&SIMPLE)) {
9733 reginsert(pRExC_state, PLUS, ret, depth+1);
9737 else if (op == '+') {
9741 else if (op == '?') {
9746 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
9747 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
9748 ckWARN3reg(RExC_parse,
9749 "%.*s matches null string many times",
9750 (int)(RExC_parse >= origparse ? RExC_parse - origparse : 0),
9752 (void)ReREFCNT_inc(RExC_rx_sv);
9755 if (RExC_parse < RExC_end && *RExC_parse == '?') {
9756 nextchar(pRExC_state);
9757 reginsert(pRExC_state, MINMOD, ret, depth+1);
9758 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
9761 if (RExC_parse < RExC_end && *RExC_parse == '+') {
9763 nextchar(pRExC_state);
9764 ender = reg_node(pRExC_state, SUCCEED);
9765 REGTAIL(pRExC_state, ret, ender);
9766 reginsert(pRExC_state, SUSPEND, ret, depth+1);
9768 ender = reg_node(pRExC_state, TAIL);
9769 REGTAIL(pRExC_state, ret, ender);
9772 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
9774 vFAIL("Nested quantifiers");
9781 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p, UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
9782 const bool strict /* Apply stricter parsing rules? */
9786 /* This is expected to be called by a parser routine that has recognized '\N'
9787 and needs to handle the rest. RExC_parse is expected to point at the first
9788 char following the N at the time of the call. On successful return,
9789 RExC_parse has been updated to point to just after the sequence identified
9790 by this routine, and <*flagp> has been updated.
9792 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
9795 \N may begin either a named sequence, or if outside a character class, mean
9796 to match a non-newline. For non single-quoted regexes, the tokenizer has
9797 attempted to decide which, and in the case of a named sequence, converted it
9798 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
9799 where c1... are the characters in the sequence. For single-quoted regexes,
9800 the tokenizer passes the \N sequence through unchanged; this code will not
9801 attempt to determine this nor expand those, instead raising a syntax error.
9802 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
9803 or there is no '}', it signals that this \N occurrence means to match a
9806 Only the \N{U+...} form should occur in a character class, for the same
9807 reason that '.' inside a character class means to just match a period: it
9808 just doesn't make sense.
9810 The function raises an error (via vFAIL), and doesn't return for various
9811 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
9812 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
9813 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
9814 only possible if node_p is non-NULL.
9817 If <valuep> is non-null, it means the caller can accept an input sequence
9818 consisting of a just a single code point; <*valuep> is set to that value
9819 if the input is such.
9821 If <node_p> is non-null it signifies that the caller can accept any other
9822 legal sequence (i.e., one that isn't just a single code point). <*node_p>
9824 1) \N means not-a-NL: points to a newly created REG_ANY node;
9825 2) \N{}: points to a new NOTHING node;
9826 3) otherwise: points to a new EXACT node containing the resolved
9828 Note that FALSE is returned for single code point sequences if <valuep> is
9832 char * endbrace; /* '}' following the name */
9834 char *endchar; /* Points to '.' or '}' ending cur char in the input
9836 bool has_multiple_chars; /* true if the input stream contains a sequence of
9837 more than one character */
9839 GET_RE_DEBUG_FLAGS_DECL;
9841 PERL_ARGS_ASSERT_GROK_BSLASH_N;
9845 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
9847 /* The [^\n] meaning of \N ignores spaces and comments under the /x
9848 * modifier. The other meaning does not */
9849 p = (RExC_flags & RXf_PMf_EXTENDED)
9850 ? regwhite( pRExC_state, RExC_parse )
9853 /* Disambiguate between \N meaning a named character versus \N meaning
9854 * [^\n]. The former is assumed when it can't be the latter. */
9855 if (*p != '{' || regcurly(p, FALSE)) {
9858 /* no bare \N in a charclass */
9859 if (in_char_class) {
9860 vFAIL("\\N in a character class must be a named character: \\N{...}");
9864 nextchar(pRExC_state);
9865 *node_p = reg_node(pRExC_state, REG_ANY);
9866 *flagp |= HASWIDTH|SIMPLE;
9869 Set_Node_Length(*node_p, 1); /* MJD */
9873 /* Here, we have decided it should be a named character or sequence */
9875 /* The test above made sure that the next real character is a '{', but
9876 * under the /x modifier, it could be separated by space (or a comment and
9877 * \n) and this is not allowed (for consistency with \x{...} and the
9878 * tokenizer handling of \N{NAME}). */
9879 if (*RExC_parse != '{') {
9880 vFAIL("Missing braces on \\N{}");
9883 RExC_parse++; /* Skip past the '{' */
9885 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
9886 || ! (endbrace == RExC_parse /* nothing between the {} */
9887 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below */
9888 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg) */
9890 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
9891 vFAIL("\\N{NAME} must be resolved by the lexer");
9894 if (endbrace == RExC_parse) { /* empty: \N{} */
9897 *node_p = reg_node(pRExC_state,NOTHING);
9899 else if (in_char_class) {
9900 if (SIZE_ONLY && in_char_class) {
9902 RExC_parse++; /* Position after the "}" */
9903 vFAIL("Zero length \\N{}");
9906 ckWARNreg(RExC_parse,
9907 "Ignoring zero length \\N{} in character class");
9915 nextchar(pRExC_state);
9919 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
9920 RExC_parse += 2; /* Skip past the 'U+' */
9922 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9924 /* Code points are separated by dots. If none, there is only one code
9925 * point, and is terminated by the brace */
9926 has_multiple_chars = (endchar < endbrace);
9928 if (valuep && (! has_multiple_chars || in_char_class)) {
9929 /* We only pay attention to the first char of
9930 multichar strings being returned in char classes. I kinda wonder
9931 if this makes sense as it does change the behaviour
9932 from earlier versions, OTOH that behaviour was broken
9933 as well. XXX Solution is to recharacterize as
9934 [rest-of-class]|multi1|multi2... */
9936 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
9937 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
9938 | PERL_SCAN_DISALLOW_PREFIX
9939 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
9941 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
9943 /* The tokenizer should have guaranteed validity, but it's possible to
9944 * bypass it by using single quoting, so check */
9945 if (length_of_hex == 0
9946 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
9948 RExC_parse += length_of_hex; /* Includes all the valid */
9949 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
9950 ? UTF8SKIP(RExC_parse)
9952 /* Guard against malformed utf8 */
9953 if (RExC_parse >= endchar) {
9954 RExC_parse = endchar;
9956 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9959 if (in_char_class && has_multiple_chars) {
9961 RExC_parse = endbrace;
9962 vFAIL("\\N{} in character class restricted to one character");
9965 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
9969 RExC_parse = endbrace + 1;
9971 else if (! node_p || ! has_multiple_chars) {
9973 /* Here, the input is legal, but not according to the caller's
9974 * options. We fail without advancing the parse, so that the
9975 * caller can try again */
9981 /* What is done here is to convert this to a sub-pattern of the form
9982 * (?:\x{char1}\x{char2}...)
9983 * and then call reg recursively. That way, it retains its atomicness,
9984 * while not having to worry about special handling that some code
9985 * points may have. toke.c has converted the original Unicode values
9986 * to native, so that we can just pass on the hex values unchanged. We
9987 * do have to set a flag to keep recoding from happening in the
9990 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
9992 char *orig_end = RExC_end;
9995 while (RExC_parse < endbrace) {
9997 /* Convert to notation the rest of the code understands */
9998 sv_catpv(substitute_parse, "\\x{");
9999 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
10000 sv_catpv(substitute_parse, "}");
10002 /* Point to the beginning of the next character in the sequence. */
10003 RExC_parse = endchar + 1;
10004 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10006 sv_catpv(substitute_parse, ")");
10008 RExC_parse = SvPV(substitute_parse, len);
10010 /* Don't allow empty number */
10012 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10014 RExC_end = RExC_parse + len;
10016 /* The values are Unicode, and therefore not subject to recoding */
10017 RExC_override_recoding = 1;
10019 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
10020 if (flags & RESTART_UTF8) {
10021 *flagp = RESTART_UTF8;
10024 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
10027 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10029 RExC_parse = endbrace;
10030 RExC_end = orig_end;
10031 RExC_override_recoding = 0;
10033 nextchar(pRExC_state);
10043 * It returns the code point in utf8 for the value in *encp.
10044 * value: a code value in the source encoding
10045 * encp: a pointer to an Encode object
10047 * If the result from Encode is not a single character,
10048 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
10051 S_reg_recode(pTHX_ const char value, SV **encp)
10054 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
10055 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
10056 const STRLEN newlen = SvCUR(sv);
10057 UV uv = UNICODE_REPLACEMENT;
10059 PERL_ARGS_ASSERT_REG_RECODE;
10063 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
10066 if (!newlen || numlen != newlen) {
10067 uv = UNICODE_REPLACEMENT;
10073 PERL_STATIC_INLINE U8
10074 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
10078 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
10084 op = get_regex_charset(RExC_flags);
10085 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
10086 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
10087 been, so there is no hole */
10090 return op + EXACTF;
10093 PERL_STATIC_INLINE void
10094 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state, regnode *node, I32* flagp, STRLEN len, UV code_point)
10096 /* This knows the details about sizing an EXACTish node, setting flags for
10097 * it (by setting <*flagp>, and potentially populating it with a single
10100 * If <len> (the length in bytes) is non-zero, this function assumes that
10101 * the node has already been populated, and just does the sizing. In this
10102 * case <code_point> should be the final code point that has already been
10103 * placed into the node. This value will be ignored except that under some
10104 * circumstances <*flagp> is set based on it.
10106 * If <len> is zero, the function assumes that the node is to contain only
10107 * the single character given by <code_point> and calculates what <len>
10108 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10109 * additionally will populate the node's STRING with <code_point>, if <len>
10110 * is 0. In both cases <*flagp> is appropriately set
10112 * It knows that under FOLD, the Latin Sharp S and UTF characters above
10113 * 255, must be folded (the former only when the rules indicate it can
10116 bool len_passed_in = cBOOL(len != 0);
10117 U8 character[UTF8_MAXBYTES_CASE+1];
10119 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10121 if (! len_passed_in) {
10123 if (FOLD && (! LOC || code_point > 255)) {
10124 _to_uni_fold_flags(NATIVE_TO_UNI(code_point),
10127 FOLD_FLAGS_FULL | ((LOC)
10128 ? FOLD_FLAGS_LOCALE
10129 : (ASCII_FOLD_RESTRICTED)
10130 ? FOLD_FLAGS_NOMIX_ASCII
10134 uvchr_to_utf8( character, code_point);
10135 len = UTF8SKIP(character);
10139 || code_point != LATIN_SMALL_LETTER_SHARP_S
10140 || ASCII_FOLD_RESTRICTED
10141 || ! AT_LEAST_UNI_SEMANTICS)
10143 *character = (U8) code_point;
10148 *(character + 1) = 's';
10154 RExC_size += STR_SZ(len);
10157 RExC_emit += STR_SZ(len);
10158 STR_LEN(node) = len;
10159 if (! len_passed_in) {
10160 Copy((char *) character, STRING(node), len, char);
10164 *flagp |= HASWIDTH;
10166 /* A single character node is SIMPLE, except for the special-cased SHARP S
10168 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
10169 && (code_point != LATIN_SMALL_LETTER_SHARP_S
10170 || ! FOLD || ! DEPENDS_SEMANTICS))
10177 - regatom - the lowest level
10179 Try to identify anything special at the start of the pattern. If there
10180 is, then handle it as required. This may involve generating a single regop,
10181 such as for an assertion; or it may involve recursing, such as to
10182 handle a () structure.
10184 If the string doesn't start with something special then we gobble up
10185 as much literal text as we can.
10187 Once we have been able to handle whatever type of thing started the
10188 sequence, we return.
10190 Note: we have to be careful with escapes, as they can be both literal
10191 and special, and in the case of \10 and friends, context determines which.
10193 A summary of the code structure is:
10195 switch (first_byte) {
10196 cases for each special:
10197 handle this special;
10200 switch (2nd byte) {
10201 cases for each unambiguous special:
10202 handle this special;
10204 cases for each ambigous special/literal:
10206 if (special) handle here
10208 default: // unambiguously literal:
10211 default: // is a literal char
10214 create EXACTish node for literal;
10215 while (more input and node isn't full) {
10216 switch (input_byte) {
10217 cases for each special;
10218 make sure parse pointer is set so that the next call to
10219 regatom will see this special first
10220 goto loopdone; // EXACTish node terminated by prev. char
10222 append char to EXACTISH node;
10224 get next input byte;
10228 return the generated node;
10230 Specifically there are two separate switches for handling
10231 escape sequences, with the one for handling literal escapes requiring
10232 a dummy entry for all of the special escapes that are actually handled
10235 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
10237 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10239 Otherwise does not return NULL.
10243 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10246 regnode *ret = NULL;
10248 char *parse_start = RExC_parse;
10252 GET_RE_DEBUG_FLAGS_DECL;
10254 *flagp = WORST; /* Tentatively. */
10256 DEBUG_PARSE("atom");
10258 PERL_ARGS_ASSERT_REGATOM;
10261 switch ((U8)*RExC_parse) {
10263 RExC_seen_zerolen++;
10264 nextchar(pRExC_state);
10265 if (RExC_flags & RXf_PMf_MULTILINE)
10266 ret = reg_node(pRExC_state, MBOL);
10267 else if (RExC_flags & RXf_PMf_SINGLELINE)
10268 ret = reg_node(pRExC_state, SBOL);
10270 ret = reg_node(pRExC_state, BOL);
10271 Set_Node_Length(ret, 1); /* MJD */
10274 nextchar(pRExC_state);
10276 RExC_seen_zerolen++;
10277 if (RExC_flags & RXf_PMf_MULTILINE)
10278 ret = reg_node(pRExC_state, MEOL);
10279 else if (RExC_flags & RXf_PMf_SINGLELINE)
10280 ret = reg_node(pRExC_state, SEOL);
10282 ret = reg_node(pRExC_state, EOL);
10283 Set_Node_Length(ret, 1); /* MJD */
10286 nextchar(pRExC_state);
10287 if (RExC_flags & RXf_PMf_SINGLELINE)
10288 ret = reg_node(pRExC_state, SANY);
10290 ret = reg_node(pRExC_state, REG_ANY);
10291 *flagp |= HASWIDTH|SIMPLE;
10293 Set_Node_Length(ret, 1); /* MJD */
10297 char * const oregcomp_parse = ++RExC_parse;
10298 ret = regclass(pRExC_state, flagp,depth+1,
10299 FALSE, /* means parse the whole char class */
10300 TRUE, /* allow multi-char folds */
10301 FALSE, /* don't silence non-portable warnings. */
10303 if (*RExC_parse != ']') {
10304 RExC_parse = oregcomp_parse;
10305 vFAIL("Unmatched [");
10308 if (*flagp & RESTART_UTF8)
10310 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
10313 nextchar(pRExC_state);
10314 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
10318 nextchar(pRExC_state);
10319 ret = reg(pRExC_state, 2, &flags,depth+1);
10321 if (flags & TRYAGAIN) {
10322 if (RExC_parse == RExC_end) {
10323 /* Make parent create an empty node if needed. */
10324 *flagp |= TRYAGAIN;
10329 if (flags & RESTART_UTF8) {
10330 *flagp = RESTART_UTF8;
10333 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"", (UV) flags);
10335 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10339 if (flags & TRYAGAIN) {
10340 *flagp |= TRYAGAIN;
10343 vFAIL("Internal urp");
10344 /* Supposed to be caught earlier. */
10347 if (!regcurly(RExC_parse, FALSE)) {
10356 vFAIL("Quantifier follows nothing");
10361 This switch handles escape sequences that resolve to some kind
10362 of special regop and not to literal text. Escape sequnces that
10363 resolve to literal text are handled below in the switch marked
10366 Every entry in this switch *must* have a corresponding entry
10367 in the literal escape switch. However, the opposite is not
10368 required, as the default for this switch is to jump to the
10369 literal text handling code.
10371 switch ((U8)*++RExC_parse) {
10373 /* Special Escapes */
10375 RExC_seen_zerolen++;
10376 ret = reg_node(pRExC_state, SBOL);
10378 goto finish_meta_pat;
10380 ret = reg_node(pRExC_state, GPOS);
10381 RExC_seen |= REG_SEEN_GPOS;
10383 goto finish_meta_pat;
10385 RExC_seen_zerolen++;
10386 ret = reg_node(pRExC_state, KEEPS);
10388 /* XXX:dmq : disabling in-place substitution seems to
10389 * be necessary here to avoid cases of memory corruption, as
10390 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
10392 RExC_seen |= REG_SEEN_LOOKBEHIND;
10393 goto finish_meta_pat;
10395 ret = reg_node(pRExC_state, SEOL);
10397 RExC_seen_zerolen++; /* Do not optimize RE away */
10398 goto finish_meta_pat;
10400 ret = reg_node(pRExC_state, EOS);
10402 RExC_seen_zerolen++; /* Do not optimize RE away */
10403 goto finish_meta_pat;
10405 ret = reg_node(pRExC_state, CANY);
10406 RExC_seen |= REG_SEEN_CANY;
10407 *flagp |= HASWIDTH|SIMPLE;
10408 goto finish_meta_pat;
10410 ret = reg_node(pRExC_state, CLUMP);
10411 *flagp |= HASWIDTH;
10412 goto finish_meta_pat;
10418 arg = ANYOF_WORDCHAR;
10422 RExC_seen_zerolen++;
10423 RExC_seen |= REG_SEEN_LOOKBEHIND;
10424 op = BOUND + get_regex_charset(RExC_flags);
10425 if (op > BOUNDA) { /* /aa is same as /a */
10428 ret = reg_node(pRExC_state, op);
10429 FLAGS(ret) = get_regex_charset(RExC_flags);
10431 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10432 ckWARNdep(RExC_parse, "\"\\b{\" is deprecated; use \"\\b\\{\" or \"\\b[{]\" instead");
10434 goto finish_meta_pat;
10436 RExC_seen_zerolen++;
10437 RExC_seen |= REG_SEEN_LOOKBEHIND;
10438 op = NBOUND + get_regex_charset(RExC_flags);
10439 if (op > NBOUNDA) { /* /aa is same as /a */
10442 ret = reg_node(pRExC_state, op);
10443 FLAGS(ret) = get_regex_charset(RExC_flags);
10445 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10446 ckWARNdep(RExC_parse, "\"\\B{\" is deprecated; use \"\\B\\{\" or \"\\B[{]\" instead");
10448 goto finish_meta_pat;
10458 ret = reg_node(pRExC_state, LNBREAK);
10459 *flagp |= HASWIDTH|SIMPLE;
10460 goto finish_meta_pat;
10468 goto join_posix_op_known;
10474 arg = ANYOF_VERTWS;
10476 goto join_posix_op_known;
10486 op = POSIXD + get_regex_charset(RExC_flags);
10487 if (op > POSIXA) { /* /aa is same as /a */
10491 join_posix_op_known:
10494 op += NPOSIXD - POSIXD;
10497 ret = reg_node(pRExC_state, op);
10499 FLAGS(ret) = namedclass_to_classnum(arg);
10502 *flagp |= HASWIDTH|SIMPLE;
10506 nextchar(pRExC_state);
10507 Set_Node_Length(ret, 2); /* MJD */
10513 char* parse_start = RExC_parse - 2;
10518 ret = regclass(pRExC_state, flagp,depth+1,
10519 TRUE, /* means just parse this element */
10520 FALSE, /* don't allow multi-char folds */
10521 FALSE, /* don't silence non-portable warnings.
10522 It would be a bug if these returned
10525 /* regclass() can only return RESTART_UTF8 if multi-char folds
10528 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
10533 Set_Node_Offset(ret, parse_start + 2);
10534 Set_Node_Cur_Length(ret, parse_start);
10535 nextchar(pRExC_state);
10539 /* Handle \N and \N{NAME} with multiple code points here and not
10540 * below because it can be multicharacter. join_exact() will join
10541 * them up later on. Also this makes sure that things like
10542 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
10543 * The options to the grok function call causes it to fail if the
10544 * sequence is just a single code point. We then go treat it as
10545 * just another character in the current EXACT node, and hence it
10546 * gets uniform treatment with all the other characters. The
10547 * special treatment for quantifiers is not needed for such single
10548 * character sequences */
10550 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
10551 FALSE /* not strict */ )) {
10552 if (*flagp & RESTART_UTF8)
10558 case 'k': /* Handle \k<NAME> and \k'NAME' */
10561 char ch= RExC_parse[1];
10562 if (ch != '<' && ch != '\'' && ch != '{') {
10564 vFAIL2("Sequence %.2s... not terminated",parse_start);
10566 /* this pretty much dupes the code for (?P=...) in reg(), if
10567 you change this make sure you change that */
10568 char* name_start = (RExC_parse += 2);
10570 SV *sv_dat = reg_scan_name(pRExC_state,
10571 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10572 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
10573 if (RExC_parse == name_start || *RExC_parse != ch)
10574 vFAIL2("Sequence %.3s... not terminated",parse_start);
10577 num = add_data( pRExC_state, 1, "S" );
10578 RExC_rxi->data->data[num]=(void*)sv_dat;
10579 SvREFCNT_inc_simple_void(sv_dat);
10583 ret = reganode(pRExC_state,
10586 : (ASCII_FOLD_RESTRICTED)
10588 : (AT_LEAST_UNI_SEMANTICS)
10594 *flagp |= HASWIDTH;
10596 /* override incorrect value set in reganode MJD */
10597 Set_Node_Offset(ret, parse_start+1);
10598 Set_Node_Cur_Length(ret, parse_start);
10599 nextchar(pRExC_state);
10605 case '1': case '2': case '3': case '4':
10606 case '5': case '6': case '7': case '8': case '9':
10609 bool isg = *RExC_parse == 'g';
10614 if (*RExC_parse == '{') {
10618 if (*RExC_parse == '-') {
10622 if (hasbrace && !isDIGIT(*RExC_parse)) {
10623 if (isrel) RExC_parse--;
10625 goto parse_named_seq;
10627 num = atoi(RExC_parse);
10628 if (isg && num == 0) {
10629 if (*RExC_parse == '0') {
10630 vFAIL("Reference to invalid group 0");
10633 vFAIL("Unterminated \\g... pattern");
10637 num = RExC_npar - num;
10639 vFAIL("Reference to nonexistent or unclosed group");
10641 if (!isg && num > 9 && num >= RExC_npar && *RExC_parse != '8' && *RExC_parse != '9')
10642 /* Probably a character specified in octal, e.g. \35 */
10645 #ifdef RE_TRACK_PATTERN_OFFSETS
10646 char * const parse_start = RExC_parse - 1; /* MJD */
10648 while (isDIGIT(*RExC_parse))
10651 if (*RExC_parse != '}')
10652 vFAIL("Unterminated \\g{...} pattern");
10656 if (num > (I32)RExC_rx->nparens)
10657 vFAIL("Reference to nonexistent group");
10660 ret = reganode(pRExC_state,
10663 : (ASCII_FOLD_RESTRICTED)
10665 : (AT_LEAST_UNI_SEMANTICS)
10671 *flagp |= HASWIDTH;
10673 /* override incorrect value set in reganode MJD */
10674 Set_Node_Offset(ret, parse_start+1);
10675 Set_Node_Cur_Length(ret, parse_start);
10677 nextchar(pRExC_state);
10682 if (RExC_parse >= RExC_end)
10683 FAIL("Trailing \\");
10686 /* Do not generate "unrecognized" warnings here, we fall
10687 back into the quick-grab loop below */
10694 if (RExC_flags & RXf_PMf_EXTENDED) {
10695 if ( reg_skipcomment( pRExC_state ) )
10702 parse_start = RExC_parse - 1;
10711 #define MAX_NODE_STRING_SIZE 127
10712 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
10714 U8 upper_parse = MAX_NODE_STRING_SIZE;
10717 bool next_is_quantifier;
10718 char * oldp = NULL;
10720 /* If a folding node contains only code points that don't
10721 * participate in folds, it can be changed into an EXACT node,
10722 * which allows the optimizer more things to look for */
10726 node_type = compute_EXACTish(pRExC_state);
10727 ret = reg_node(pRExC_state, node_type);
10729 /* In pass1, folded, we use a temporary buffer instead of the
10730 * actual node, as the node doesn't exist yet */
10731 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
10737 /* We do the EXACTFish to EXACT node only if folding, and not if in
10738 * locale, as whether a character folds or not isn't known until
10740 maybe_exact = FOLD && ! LOC;
10742 /* XXX The node can hold up to 255 bytes, yet this only goes to
10743 * 127. I (khw) do not know why. Keeping it somewhat less than
10744 * 255 allows us to not have to worry about overflow due to
10745 * converting to utf8 and fold expansion, but that value is
10746 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
10747 * split up by this limit into a single one using the real max of
10748 * 255. Even at 127, this breaks under rare circumstances. If
10749 * folding, we do not want to split a node at a character that is a
10750 * non-final in a multi-char fold, as an input string could just
10751 * happen to want to match across the node boundary. The join
10752 * would solve that problem if the join actually happens. But a
10753 * series of more than two nodes in a row each of 127 would cause
10754 * the first join to succeed to get to 254, but then there wouldn't
10755 * be room for the next one, which could at be one of those split
10756 * multi-char folds. I don't know of any fool-proof solution. One
10757 * could back off to end with only a code point that isn't such a
10758 * non-final, but it is possible for there not to be any in the
10760 for (p = RExC_parse - 1;
10761 len < upper_parse && p < RExC_end;
10766 if (RExC_flags & RXf_PMf_EXTENDED)
10767 p = regwhite( pRExC_state, p );
10778 /* Literal Escapes Switch
10780 This switch is meant to handle escape sequences that
10781 resolve to a literal character.
10783 Every escape sequence that represents something
10784 else, like an assertion or a char class, is handled
10785 in the switch marked 'Special Escapes' above in this
10786 routine, but also has an entry here as anything that
10787 isn't explicitly mentioned here will be treated as
10788 an unescaped equivalent literal.
10791 switch ((U8)*++p) {
10792 /* These are all the special escapes. */
10793 case 'A': /* Start assertion */
10794 case 'b': case 'B': /* Word-boundary assertion*/
10795 case 'C': /* Single char !DANGEROUS! */
10796 case 'd': case 'D': /* digit class */
10797 case 'g': case 'G': /* generic-backref, pos assertion */
10798 case 'h': case 'H': /* HORIZWS */
10799 case 'k': case 'K': /* named backref, keep marker */
10800 case 'p': case 'P': /* Unicode property */
10801 case 'R': /* LNBREAK */
10802 case 's': case 'S': /* space class */
10803 case 'v': case 'V': /* VERTWS */
10804 case 'w': case 'W': /* word class */
10805 case 'X': /* eXtended Unicode "combining character sequence" */
10806 case 'z': case 'Z': /* End of line/string assertion */
10810 /* Anything after here is an escape that resolves to a
10811 literal. (Except digits, which may or may not)
10817 case 'N': /* Handle a single-code point named character. */
10818 /* The options cause it to fail if a multiple code
10819 * point sequence. Handle those in the switch() above
10821 RExC_parse = p + 1;
10822 if (! grok_bslash_N(pRExC_state, NULL, &ender,
10823 flagp, depth, FALSE,
10824 FALSE /* not strict */ ))
10826 if (*flagp & RESTART_UTF8)
10827 FAIL("panic: grok_bslash_N set RESTART_UTF8");
10828 RExC_parse = p = oldp;
10832 if (ender > 0xff) {
10849 ender = ASCII_TO_NATIVE('\033');
10853 ender = ASCII_TO_NATIVE('\007');
10859 const char* error_msg;
10861 bool valid = grok_bslash_o(&p,
10864 TRUE, /* out warnings */
10865 FALSE, /* not strict */
10866 TRUE, /* Output warnings
10871 RExC_parse = p; /* going to die anyway; point
10872 to exact spot of failure */
10876 if (PL_encoding && ender < 0x100) {
10877 goto recode_encoding;
10879 if (ender > 0xff) {
10886 UV result = UV_MAX; /* initialize to erroneous
10888 const char* error_msg;
10890 bool valid = grok_bslash_x(&p,
10893 TRUE, /* out warnings */
10894 FALSE, /* not strict */
10895 TRUE, /* Output warnings
10900 RExC_parse = p; /* going to die anyway; point
10901 to exact spot of failure */
10906 if (PL_encoding && ender < 0x100) {
10907 goto recode_encoding;
10909 if (ender > 0xff) {
10916 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
10918 case '8': case '9': /* must be a backreference */
10921 case '1': case '2': case '3':case '4':
10922 case '5': case '6': case '7':
10923 /* When we parse backslash escapes there is ambiguity between
10924 * backreferences and octal escapes. Any escape from \1 - \9 is
10925 * a backreference, any multi-digit escape which does not start with
10926 * 0 and which when evaluated as decimal could refer to an already
10927 * parsed capture buffer is a backslash. Anything else is octal.
10929 * Note this implies that \118 could be interpreted as 118 OR as
10930 * "\11" . "8" depending on whether there were 118 capture buffers
10931 * defined already in the pattern.
10933 if ( !isDIGIT(p[1]) || atoi(p) <= RExC_npar )
10934 { /* Not to be treated as an octal constant, go
10941 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
10943 ender = grok_oct(p, &numlen, &flags, NULL);
10944 if (ender > 0xff) {
10948 if (SIZE_ONLY /* like \08, \178 */
10951 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
10953 reg_warn_non_literal_string(
10955 form_short_octal_warning(p, numlen));
10958 if (PL_encoding && ender < 0x100)
10959 goto recode_encoding;
10962 if (! RExC_override_recoding) {
10963 SV* enc = PL_encoding;
10964 ender = reg_recode((const char)(U8)ender, &enc);
10965 if (!enc && SIZE_ONLY)
10966 ckWARNreg(p, "Invalid escape in the specified encoding");
10972 FAIL("Trailing \\");
10975 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
10976 /* Include any { following the alpha to emphasize
10977 * that it could be part of an escape at some point
10979 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
10980 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
10982 goto normal_default;
10983 } /* End of switch on '\' */
10985 default: /* A literal character */
10988 && RExC_flags & RXf_PMf_EXTENDED
10989 && ckWARN_d(WARN_DEPRECATED)
10990 && is_PATWS_non_low(p, UTF))
10992 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
10993 "Escape literal pattern white space under /x");
10997 if (UTF8_IS_START(*p) && UTF) {
10999 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
11000 &numlen, UTF8_ALLOW_DEFAULT);
11006 } /* End of switch on the literal */
11008 /* Here, have looked at the literal character and <ender>
11009 * contains its ordinal, <p> points to the character after it
11012 if ( RExC_flags & RXf_PMf_EXTENDED)
11013 p = regwhite( pRExC_state, p );
11015 /* If the next thing is a quantifier, it applies to this
11016 * character only, which means that this character has to be in
11017 * its own node and can't just be appended to the string in an
11018 * existing node, so if there are already other characters in
11019 * the node, close the node with just them, and set up to do
11020 * this character again next time through, when it will be the
11021 * only thing in its new node */
11022 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
11030 const STRLEN unilen = reguni(pRExC_state, ender, s);
11036 /* The loop increments <len> each time, as all but this
11037 * path (and one other) through it add a single byte to
11038 * the EXACTish node. But this one has changed len to
11039 * be the correct final value, so subtract one to
11040 * cancel out the increment that follows */
11044 REGC((char)ender, s++);
11049 /* See comments for join_exact() as to why we fold this
11050 * non-UTF at compile time */
11051 || (node_type == EXACTFU
11052 && ender == LATIN_SMALL_LETTER_SHARP_S)))
11054 *(s++) = (char) ender;
11055 maybe_exact &= ! IS_IN_SOME_FOLD_L1(ender);
11059 /* Prime the casefolded buffer. Locale rules, which apply
11060 * only to code points < 256, aren't known until execution,
11061 * so for them, just output the original character using
11062 * utf8. If we start to fold non-UTF patterns, be sure to
11063 * update join_exact() */
11064 if (LOC && ender < 256) {
11065 if (UNI_IS_INVARIANT(ender)) {
11069 *s = UTF8_TWO_BYTE_HI(ender);
11070 *(s + 1) = UTF8_TWO_BYTE_LO(ender);
11075 UV folded = _to_uni_fold_flags(
11080 | ((LOC) ? FOLD_FLAGS_LOCALE
11081 : (ASCII_FOLD_RESTRICTED)
11082 ? FOLD_FLAGS_NOMIX_ASCII
11086 /* If this node only contains non-folding code points
11087 * so far, see if this new one is also non-folding */
11089 if (folded != ender) {
11090 maybe_exact = FALSE;
11093 /* Here the fold is the original; we have
11094 * to check further to see if anything
11096 if (! PL_utf8_foldable) {
11097 SV* swash = swash_init("utf8",
11099 &PL_sv_undef, 1, 0);
11101 _get_swash_invlist(swash);
11102 SvREFCNT_dec_NN(swash);
11104 if (_invlist_contains_cp(PL_utf8_foldable,
11107 maybe_exact = FALSE;
11115 /* The loop increments <len> each time, as all but this
11116 * path (and one other) through it add a single byte to the
11117 * EXACTish node. But this one has changed len to be the
11118 * correct final value, so subtract one to cancel out the
11119 * increment that follows */
11120 len += foldlen - 1;
11123 if (next_is_quantifier) {
11125 /* Here, the next input is a quantifier, and to get here,
11126 * the current character is the only one in the node.
11127 * Also, here <len> doesn't include the final byte for this
11133 } /* End of loop through literal characters */
11135 /* Here we have either exhausted the input or ran out of room in
11136 * the node. (If we encountered a character that can't be in the
11137 * node, transfer is made directly to <loopdone>, and so we
11138 * wouldn't have fallen off the end of the loop.) In the latter
11139 * case, we artificially have to split the node into two, because
11140 * we just don't have enough space to hold everything. This
11141 * creates a problem if the final character participates in a
11142 * multi-character fold in the non-final position, as a match that
11143 * should have occurred won't, due to the way nodes are matched,
11144 * and our artificial boundary. So back off until we find a non-
11145 * problematic character -- one that isn't at the beginning or
11146 * middle of such a fold. (Either it doesn't participate in any
11147 * folds, or appears only in the final position of all the folds it
11148 * does participate in.) A better solution with far fewer false
11149 * positives, and that would fill the nodes more completely, would
11150 * be to actually have available all the multi-character folds to
11151 * test against, and to back-off only far enough to be sure that
11152 * this node isn't ending with a partial one. <upper_parse> is set
11153 * further below (if we need to reparse the node) to include just
11154 * up through that final non-problematic character that this code
11155 * identifies, so when it is set to less than the full node, we can
11156 * skip the rest of this */
11157 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
11159 const STRLEN full_len = len;
11161 assert(len >= MAX_NODE_STRING_SIZE);
11163 /* Here, <s> points to the final byte of the final character.
11164 * Look backwards through the string until find a non-
11165 * problematic character */
11169 /* These two have no multi-char folds to non-UTF characters
11171 if (ASCII_FOLD_RESTRICTED || LOC) {
11175 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
11179 if (! PL_NonL1NonFinalFold) {
11180 PL_NonL1NonFinalFold = _new_invlist_C_array(
11181 NonL1_Perl_Non_Final_Folds_invlist);
11184 /* Point to the first byte of the final character */
11185 s = (char *) utf8_hop((U8 *) s, -1);
11187 while (s >= s0) { /* Search backwards until find
11188 non-problematic char */
11189 if (UTF8_IS_INVARIANT(*s)) {
11191 /* There are no ascii characters that participate
11192 * in multi-char folds under /aa. In EBCDIC, the
11193 * non-ascii invariants are all control characters,
11194 * so don't ever participate in any folds. */
11195 if (ASCII_FOLD_RESTRICTED
11196 || ! IS_NON_FINAL_FOLD(*s))
11201 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
11203 /* No Latin1 characters participate in multi-char
11204 * folds under /l */
11206 || ! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_UNI(
11212 else if (! _invlist_contains_cp(
11213 PL_NonL1NonFinalFold,
11214 valid_utf8_to_uvchr((U8 *) s, NULL)))
11219 /* Here, the current character is problematic in that
11220 * it does occur in the non-final position of some
11221 * fold, so try the character before it, but have to
11222 * special case the very first byte in the string, so
11223 * we don't read outside the string */
11224 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
11225 } /* End of loop backwards through the string */
11227 /* If there were only problematic characters in the string,
11228 * <s> will point to before s0, in which case the length
11229 * should be 0, otherwise include the length of the
11230 * non-problematic character just found */
11231 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
11234 /* Here, have found the final character, if any, that is
11235 * non-problematic as far as ending the node without splitting
11236 * it across a potential multi-char fold. <len> contains the
11237 * number of bytes in the node up-to and including that
11238 * character, or is 0 if there is no such character, meaning
11239 * the whole node contains only problematic characters. In
11240 * this case, give up and just take the node as-is. We can't
11246 /* Here, the node does contain some characters that aren't
11247 * problematic. If one such is the final character in the
11248 * node, we are done */
11249 if (len == full_len) {
11252 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
11254 /* If the final character is problematic, but the
11255 * penultimate is not, back-off that last character to
11256 * later start a new node with it */
11261 /* Here, the final non-problematic character is earlier
11262 * in the input than the penultimate character. What we do
11263 * is reparse from the beginning, going up only as far as
11264 * this final ok one, thus guaranteeing that the node ends
11265 * in an acceptable character. The reason we reparse is
11266 * that we know how far in the character is, but we don't
11267 * know how to correlate its position with the input parse.
11268 * An alternate implementation would be to build that
11269 * correlation as we go along during the original parse,
11270 * but that would entail extra work for every node, whereas
11271 * this code gets executed only when the string is too
11272 * large for the node, and the final two characters are
11273 * problematic, an infrequent occurrence. Yet another
11274 * possible strategy would be to save the tail of the
11275 * string, and the next time regatom is called, initialize
11276 * with that. The problem with this is that unless you
11277 * back off one more character, you won't be guaranteed
11278 * regatom will get called again, unless regbranch,
11279 * regpiece ... are also changed. If you do back off that
11280 * extra character, so that there is input guaranteed to
11281 * force calling regatom, you can't handle the case where
11282 * just the first character in the node is acceptable. I
11283 * (khw) decided to try this method which doesn't have that
11284 * pitfall; if performance issues are found, we can do a
11285 * combination of the current approach plus that one */
11291 } /* End of verifying node ends with an appropriate char */
11293 loopdone: /* Jumped to when encounters something that shouldn't be in
11296 /* I (khw) don't know if you can get here with zero length, but the
11297 * old code handled this situation by creating a zero-length EXACT
11298 * node. Might as well be NOTHING instead */
11304 /* If 'maybe_exact' is still set here, means there are no
11305 * code points in the node that participate in folds */
11306 if (FOLD && maybe_exact) {
11309 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
11312 RExC_parse = p - 1;
11313 Set_Node_Cur_Length(ret, parse_start);
11314 nextchar(pRExC_state);
11316 /* len is STRLEN which is unsigned, need to copy to signed */
11319 vFAIL("Internal disaster");
11322 } /* End of label 'defchar:' */
11324 } /* End of giant switch on input character */
11330 S_regwhite( RExC_state_t *pRExC_state, char *p )
11332 const char *e = RExC_end;
11334 PERL_ARGS_ASSERT_REGWHITE;
11339 else if (*p == '#') {
11342 if (*p++ == '\n') {
11348 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11357 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
11359 /* Returns the next non-pattern-white space, non-comment character (the
11360 * latter only if 'recognize_comment is true) in the string p, which is
11361 * ended by RExC_end. If there is no line break ending a comment,
11362 * RExC_seen has added the REG_SEEN_RUN_ON_COMMENT flag; */
11363 const char *e = RExC_end;
11365 PERL_ARGS_ASSERT_REGPATWS;
11369 if ((len = is_PATWS_safe(p, e, UTF))) {
11372 else if (recognize_comment && *p == '#') {
11376 if (is_LNBREAK_safe(p, e, UTF)) {
11382 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11390 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
11391 Character classes ([:foo:]) can also be negated ([:^foo:]).
11392 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
11393 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
11394 but trigger failures because they are currently unimplemented. */
11396 #define POSIXCC_DONE(c) ((c) == ':')
11397 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
11398 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
11400 PERL_STATIC_INLINE I32
11401 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
11404 I32 namedclass = OOB_NAMEDCLASS;
11406 PERL_ARGS_ASSERT_REGPPOSIXCC;
11408 if (value == '[' && RExC_parse + 1 < RExC_end &&
11409 /* I smell either [: or [= or [. -- POSIX has been here, right? */
11410 POSIXCC(UCHARAT(RExC_parse)))
11412 const char c = UCHARAT(RExC_parse);
11413 char* const s = RExC_parse++;
11415 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
11417 if (RExC_parse == RExC_end) {
11420 /* Try to give a better location for the error (than the end of
11421 * the string) by looking for the matching ']' */
11423 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
11426 vFAIL2("Unmatched '%c' in POSIX class", c);
11428 /* Grandfather lone [:, [=, [. */
11432 const char* const t = RExC_parse++; /* skip over the c */
11435 if (UCHARAT(RExC_parse) == ']') {
11436 const char *posixcc = s + 1;
11437 RExC_parse++; /* skip over the ending ] */
11440 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
11441 const I32 skip = t - posixcc;
11443 /* Initially switch on the length of the name. */
11446 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
11447 this is the Perl \w
11449 namedclass = ANYOF_WORDCHAR;
11452 /* Names all of length 5. */
11453 /* alnum alpha ascii blank cntrl digit graph lower
11454 print punct space upper */
11455 /* Offset 4 gives the best switch position. */
11456 switch (posixcc[4]) {
11458 if (memEQ(posixcc, "alph", 4)) /* alpha */
11459 namedclass = ANYOF_ALPHA;
11462 if (memEQ(posixcc, "spac", 4)) /* space */
11463 namedclass = ANYOF_PSXSPC;
11466 if (memEQ(posixcc, "grap", 4)) /* graph */
11467 namedclass = ANYOF_GRAPH;
11470 if (memEQ(posixcc, "asci", 4)) /* ascii */
11471 namedclass = ANYOF_ASCII;
11474 if (memEQ(posixcc, "blan", 4)) /* blank */
11475 namedclass = ANYOF_BLANK;
11478 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
11479 namedclass = ANYOF_CNTRL;
11482 if (memEQ(posixcc, "alnu", 4)) /* alnum */
11483 namedclass = ANYOF_ALPHANUMERIC;
11486 if (memEQ(posixcc, "lowe", 4)) /* lower */
11487 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
11488 else if (memEQ(posixcc, "uppe", 4)) /* upper */
11489 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
11492 if (memEQ(posixcc, "digi", 4)) /* digit */
11493 namedclass = ANYOF_DIGIT;
11494 else if (memEQ(posixcc, "prin", 4)) /* print */
11495 namedclass = ANYOF_PRINT;
11496 else if (memEQ(posixcc, "punc", 4)) /* punct */
11497 namedclass = ANYOF_PUNCT;
11502 if (memEQ(posixcc, "xdigit", 6))
11503 namedclass = ANYOF_XDIGIT;
11507 if (namedclass == OOB_NAMEDCLASS)
11508 Simple_vFAIL3("POSIX class [:%.*s:] unknown",
11511 /* The #defines are structured so each complement is +1 to
11512 * the normal one */
11516 assert (posixcc[skip] == ':');
11517 assert (posixcc[skip+1] == ']');
11518 } else if (!SIZE_ONLY) {
11519 /* [[=foo=]] and [[.foo.]] are still future. */
11521 /* adjust RExC_parse so the warning shows after
11522 the class closes */
11523 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
11525 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11528 /* Maternal grandfather:
11529 * "[:" ending in ":" but not in ":]" */
11531 vFAIL("Unmatched '[' in POSIX class");
11534 /* Grandfather lone [:, [=, [. */
11544 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
11546 /* This applies some heuristics at the current parse position (which should
11547 * be at a '[') to see if what follows might be intended to be a [:posix:]
11548 * class. It returns true if it really is a posix class, of course, but it
11549 * also can return true if it thinks that what was intended was a posix
11550 * class that didn't quite make it.
11552 * It will return true for
11554 * [:alphanumerics] (as long as the ] isn't followed immediately by a
11555 * ')' indicating the end of the (?[
11556 * [:any garbage including %^&$ punctuation:]
11558 * This is designed to be called only from S_handle_regex_sets; it could be
11559 * easily adapted to be called from the spot at the beginning of regclass()
11560 * that checks to see in a normal bracketed class if the surrounding []
11561 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
11562 * change long-standing behavior, so I (khw) didn't do that */
11563 char* p = RExC_parse + 1;
11564 char first_char = *p;
11566 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
11568 assert(*(p - 1) == '[');
11570 if (! POSIXCC(first_char)) {
11575 while (p < RExC_end && isWORDCHAR(*p)) p++;
11577 if (p >= RExC_end) {
11581 if (p - RExC_parse > 2 /* Got at least 1 word character */
11582 && (*p == first_char
11583 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
11588 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
11591 && p - RExC_parse > 2 /* [:] evaluates to colon;
11592 [::] is a bad posix class. */
11593 && first_char == *(p - 1));
11597 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist, I32 *flagp, U32 depth,
11598 char * const oregcomp_parse)
11600 /* Handle the (?[...]) construct to do set operations */
11603 UV start, end; /* End points of code point ranges */
11605 char *save_end, *save_parse;
11610 const bool save_fold = FOLD;
11612 GET_RE_DEBUG_FLAGS_DECL;
11614 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
11617 vFAIL("(?[...]) not valid in locale");
11619 RExC_uni_semantics = 1;
11621 /* This will return only an ANYOF regnode, or (unlikely) something smaller
11622 * (such as EXACT). Thus we can skip most everything if just sizing. We
11623 * call regclass to handle '[]' so as to not have to reinvent its parsing
11624 * rules here (throwing away the size it computes each time). And, we exit
11625 * upon an unescaped ']' that isn't one ending a regclass. To do both
11626 * these things, we need to realize that something preceded by a backslash
11627 * is escaped, so we have to keep track of backslashes */
11629 UV depth = 0; /* how many nested (?[...]) constructs */
11631 Perl_ck_warner_d(aTHX_
11632 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
11633 "The regex_sets feature is experimental" REPORT_LOCATION,
11634 (int) (RExC_parse - RExC_precomp) , RExC_precomp, RExC_parse);
11636 while (RExC_parse < RExC_end) {
11637 SV* current = NULL;
11638 RExC_parse = regpatws(pRExC_state, RExC_parse,
11639 TRUE); /* means recognize comments */
11640 switch (*RExC_parse) {
11642 if (RExC_parse[1] == '[') depth++, RExC_parse++;
11647 /* Skip the next byte (which could cause us to end up in
11648 * the middle of a UTF-8 character, but since none of those
11649 * are confusable with anything we currently handle in this
11650 * switch (invariants all), it's safe. We'll just hit the
11651 * default: case next time and keep on incrementing until
11652 * we find one of the invariants we do handle. */
11657 /* If this looks like it is a [:posix:] class, leave the
11658 * parse pointer at the '[' to fool regclass() into
11659 * thinking it is part of a '[[:posix:]]'. That function
11660 * will use strict checking to force a syntax error if it
11661 * doesn't work out to a legitimate class */
11662 bool is_posix_class
11663 = could_it_be_a_POSIX_class(pRExC_state);
11664 if (! is_posix_class) {
11668 /* regclass() can only return RESTART_UTF8 if multi-char
11669 folds are allowed. */
11670 if (!regclass(pRExC_state, flagp,depth+1,
11671 is_posix_class, /* parse the whole char
11672 class only if not a
11674 FALSE, /* don't allow multi-char folds */
11675 TRUE, /* silence non-portable warnings. */
11677 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11680 /* function call leaves parse pointing to the ']', except
11681 * if we faked it */
11682 if (is_posix_class) {
11686 SvREFCNT_dec(current); /* In case it returned something */
11691 if (depth--) break;
11693 if (RExC_parse < RExC_end
11694 && *RExC_parse == ')')
11696 node = reganode(pRExC_state, ANYOF, 0);
11697 RExC_size += ANYOF_SKIP;
11698 nextchar(pRExC_state);
11699 Set_Node_Length(node,
11700 RExC_parse - oregcomp_parse + 1); /* MJD */
11709 FAIL("Syntax error in (?[...])");
11712 /* Pass 2 only after this. Everything in this construct is a
11713 * metacharacter. Operands begin with either a '\' (for an escape
11714 * sequence), or a '[' for a bracketed character class. Any other
11715 * character should be an operator, or parenthesis for grouping. Both
11716 * types of operands are handled by calling regclass() to parse them. It
11717 * is called with a parameter to indicate to return the computed inversion
11718 * list. The parsing here is implemented via a stack. Each entry on the
11719 * stack is a single character representing one of the operators, or the
11720 * '('; or else a pointer to an operand inversion list. */
11722 #define IS_OPERAND(a) (! SvIOK(a))
11724 /* The stack starts empty. It is a syntax error if the first thing parsed
11725 * is a binary operator; everything else is pushed on the stack. When an
11726 * operand is parsed, the top of the stack is examined. If it is a binary
11727 * operator, the item before it should be an operand, and both are replaced
11728 * by the result of doing that operation on the new operand and the one on
11729 * the stack. Thus a sequence of binary operands is reduced to a single
11730 * one before the next one is parsed.
11732 * A unary operator may immediately follow a binary in the input, for
11735 * When an operand is parsed and the top of the stack is a unary operator,
11736 * the operation is performed, and then the stack is rechecked to see if
11737 * this new operand is part of a binary operation; if so, it is handled as
11740 * A '(' is simply pushed on the stack; it is valid only if the stack is
11741 * empty, or the top element of the stack is an operator or another '('
11742 * (for which the parenthesized expression will become an operand). By the
11743 * time the corresponding ')' is parsed everything in between should have
11744 * been parsed and evaluated to a single operand (or else is a syntax
11745 * error), and is handled as a regular operand */
11747 sv_2mortal((SV *)(stack = newAV()));
11749 while (RExC_parse < RExC_end) {
11750 I32 top_index = av_tindex(stack);
11752 SV* current = NULL;
11754 /* Skip white space */
11755 RExC_parse = regpatws(pRExC_state, RExC_parse,
11756 TRUE); /* means recognize comments */
11757 if (RExC_parse >= RExC_end) {
11758 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
11760 if ((curchar = UCHARAT(RExC_parse)) == ']') {
11767 if (av_tindex(stack) >= 0 /* This makes sure that we can
11768 safely subtract 1 from
11769 RExC_parse in the next clause.
11770 If we have something on the
11771 stack, we have parsed something
11773 && UCHARAT(RExC_parse - 1) == '('
11774 && RExC_parse < RExC_end)
11776 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
11777 * This happens when we have some thing like
11779 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
11781 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
11783 * Here we would be handling the interpolated
11784 * '$thai_or_lao'. We handle this by a recursive call to
11785 * ourselves which returns the inversion list the
11786 * interpolated expression evaluates to. We use the flags
11787 * from the interpolated pattern. */
11788 U32 save_flags = RExC_flags;
11789 const char * const save_parse = ++RExC_parse;
11791 parse_lparen_question_flags(pRExC_state);
11793 if (RExC_parse == save_parse /* Makes sure there was at
11794 least one flag (or this
11795 embedding wasn't compiled)
11797 || RExC_parse >= RExC_end - 4
11798 || UCHARAT(RExC_parse) != ':'
11799 || UCHARAT(++RExC_parse) != '('
11800 || UCHARAT(++RExC_parse) != '?'
11801 || UCHARAT(++RExC_parse) != '[')
11804 /* In combination with the above, this moves the
11805 * pointer to the point just after the first erroneous
11806 * character (or if there are no flags, to where they
11807 * should have been) */
11808 if (RExC_parse >= RExC_end - 4) {
11809 RExC_parse = RExC_end;
11811 else if (RExC_parse != save_parse) {
11812 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11814 vFAIL("Expecting '(?flags:(?[...'");
11817 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
11818 depth+1, oregcomp_parse);
11820 /* Here, 'current' contains the embedded expression's
11821 * inversion list, and RExC_parse points to the trailing
11822 * ']'; the next character should be the ')' which will be
11823 * paired with the '(' that has been put on the stack, so
11824 * the whole embedded expression reduces to '(operand)' */
11827 RExC_flags = save_flags;
11828 goto handle_operand;
11833 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11834 vFAIL("Unexpected character");
11837 /* regclass() can only return RESTART_UTF8 if multi-char
11838 folds are allowed. */
11839 if (!regclass(pRExC_state, flagp,depth+1,
11840 TRUE, /* means parse just the next thing */
11841 FALSE, /* don't allow multi-char folds */
11842 FALSE, /* don't silence non-portable warnings. */
11844 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11846 /* regclass() will return with parsing just the \ sequence,
11847 * leaving the parse pointer at the next thing to parse */
11849 goto handle_operand;
11851 case '[': /* Is a bracketed character class */
11853 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
11855 if (! is_posix_class) {
11859 /* regclass() can only return RESTART_UTF8 if multi-char
11860 folds are allowed. */
11861 if(!regclass(pRExC_state, flagp,depth+1,
11862 is_posix_class, /* parse the whole char class
11863 only if not a posix class */
11864 FALSE, /* don't allow multi-char folds */
11865 FALSE, /* don't silence non-portable warnings. */
11867 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11869 /* function call leaves parse pointing to the ']', except if we
11871 if (is_posix_class) {
11875 goto handle_operand;
11884 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
11885 || ! IS_OPERAND(*top_ptr))
11888 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
11890 av_push(stack, newSVuv(curchar));
11894 av_push(stack, newSVuv(curchar));
11898 if (top_index >= 0) {
11899 top_ptr = av_fetch(stack, top_index, FALSE);
11901 if (IS_OPERAND(*top_ptr)) {
11903 vFAIL("Unexpected '(' with no preceding operator");
11906 av_push(stack, newSVuv(curchar));
11913 || ! (current = av_pop(stack))
11914 || ! IS_OPERAND(current)
11915 || ! (lparen = av_pop(stack))
11916 || IS_OPERAND(lparen)
11917 || SvUV(lparen) != '(')
11919 SvREFCNT_dec(current);
11921 vFAIL("Unexpected ')'");
11924 SvREFCNT_dec_NN(lparen);
11931 /* Here, we have an operand to process, in 'current' */
11933 if (top_index < 0) { /* Just push if stack is empty */
11934 av_push(stack, current);
11937 SV* top = av_pop(stack);
11939 char current_operator;
11941 if (IS_OPERAND(top)) {
11942 SvREFCNT_dec_NN(top);
11943 SvREFCNT_dec_NN(current);
11944 vFAIL("Operand with no preceding operator");
11946 current_operator = (char) SvUV(top);
11947 switch (current_operator) {
11948 case '(': /* Push the '(' back on followed by the new
11950 av_push(stack, top);
11951 av_push(stack, current);
11952 SvREFCNT_inc(top); /* Counters the '_dec' done
11953 just after the 'break', so
11954 it doesn't get wrongly freed
11959 _invlist_invert(current);
11961 /* Unlike binary operators, the top of the stack,
11962 * now that this unary one has been popped off, may
11963 * legally be an operator, and we now have operand
11966 SvREFCNT_dec_NN(top);
11967 goto handle_operand;
11970 prev = av_pop(stack);
11971 _invlist_intersection(prev,
11974 av_push(stack, current);
11979 prev = av_pop(stack);
11980 _invlist_union(prev, current, ¤t);
11981 av_push(stack, current);
11985 prev = av_pop(stack);;
11986 _invlist_subtract(prev, current, ¤t);
11987 av_push(stack, current);
11990 case '^': /* The union minus the intersection */
11996 prev = av_pop(stack);
11997 _invlist_union(prev, current, &u);
11998 _invlist_intersection(prev, current, &i);
11999 /* _invlist_subtract will overwrite current
12000 without freeing what it already contains */
12002 _invlist_subtract(u, i, ¤t);
12003 av_push(stack, current);
12004 SvREFCNT_dec_NN(i);
12005 SvREFCNT_dec_NN(u);
12006 SvREFCNT_dec_NN(element);
12011 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
12013 SvREFCNT_dec_NN(top);
12014 SvREFCNT_dec(prev);
12018 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12021 if (av_tindex(stack) < 0 /* Was empty */
12022 || ((final = av_pop(stack)) == NULL)
12023 || ! IS_OPERAND(final)
12024 || av_tindex(stack) >= 0) /* More left on stack */
12026 vFAIL("Incomplete expression within '(?[ ])'");
12029 /* Here, 'final' is the resultant inversion list from evaluating the
12030 * expression. Return it if so requested */
12031 if (return_invlist) {
12032 *return_invlist = final;
12036 /* Otherwise generate a resultant node, based on 'final'. regclass() is
12037 * expecting a string of ranges and individual code points */
12038 invlist_iterinit(final);
12039 result_string = newSVpvs("");
12040 while (invlist_iternext(final, &start, &end)) {
12041 if (start == end) {
12042 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
12045 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
12050 save_parse = RExC_parse;
12051 RExC_parse = SvPV(result_string, len);
12052 save_end = RExC_end;
12053 RExC_end = RExC_parse + len;
12055 /* We turn off folding around the call, as the class we have constructed
12056 * already has all folding taken into consideration, and we don't want
12057 * regclass() to add to that */
12058 RExC_flags &= ~RXf_PMf_FOLD;
12059 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
12061 node = regclass(pRExC_state, flagp,depth+1,
12062 FALSE, /* means parse the whole char class */
12063 FALSE, /* don't allow multi-char folds */
12064 TRUE, /* silence non-portable warnings. The above may very
12065 well have generated non-portable code points, but
12066 they're valid on this machine */
12069 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
12072 RExC_flags |= RXf_PMf_FOLD;
12074 RExC_parse = save_parse + 1;
12075 RExC_end = save_end;
12076 SvREFCNT_dec_NN(final);
12077 SvREFCNT_dec_NN(result_string);
12079 nextchar(pRExC_state);
12080 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
12085 /* The names of properties whose definitions are not known at compile time are
12086 * stored in this SV, after a constant heading. So if the length has been
12087 * changed since initialization, then there is a run-time definition. */
12088 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION (SvCUR(listsv) != initial_listsv_len)
12091 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
12092 const bool stop_at_1, /* Just parse the next thing, don't
12093 look for a full character class */
12094 bool allow_multi_folds,
12095 const bool silence_non_portable, /* Don't output warnings
12098 SV** ret_invlist) /* Return an inversion list, not a node */
12100 /* parse a bracketed class specification. Most of these will produce an
12101 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
12102 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
12103 * under /i with multi-character folds: it will be rewritten following the
12104 * paradigm of this example, where the <multi-fold>s are characters which
12105 * fold to multiple character sequences:
12106 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
12107 * gets effectively rewritten as:
12108 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
12109 * reg() gets called (recursively) on the rewritten version, and this
12110 * function will return what it constructs. (Actually the <multi-fold>s
12111 * aren't physically removed from the [abcdefghi], it's just that they are
12112 * ignored in the recursion by means of a flag:
12113 * <RExC_in_multi_char_class>.)
12115 * ANYOF nodes contain a bit map for the first 256 characters, with the
12116 * corresponding bit set if that character is in the list. For characters
12117 * above 255, a range list or swash is used. There are extra bits for \w,
12118 * etc. in locale ANYOFs, as what these match is not determinable at
12121 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
12122 * to be restarted. This can only happen if ret_invlist is non-NULL.
12126 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
12128 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
12131 IV namedclass = OOB_NAMEDCLASS;
12132 char *rangebegin = NULL;
12133 bool need_class = 0;
12135 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
12136 than just initialized. */
12137 SV* properties = NULL; /* Code points that match \p{} \P{} */
12138 SV* posixes = NULL; /* Code points that match classes like, [:word:],
12139 extended beyond the Latin1 range */
12140 UV element_count = 0; /* Number of distinct elements in the class.
12141 Optimizations may be possible if this is tiny */
12142 AV * multi_char_matches = NULL; /* Code points that fold to more than one
12143 character; used under /i */
12145 char * stop_ptr = RExC_end; /* where to stop parsing */
12146 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
12148 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
12150 /* Unicode properties are stored in a swash; this holds the current one
12151 * being parsed. If this swash is the only above-latin1 component of the
12152 * character class, an optimization is to pass it directly on to the
12153 * execution engine. Otherwise, it is set to NULL to indicate that there
12154 * are other things in the class that have to be dealt with at execution
12156 SV* swash = NULL; /* Code points that match \p{} \P{} */
12158 /* Set if a component of this character class is user-defined; just passed
12159 * on to the engine */
12160 bool has_user_defined_property = FALSE;
12162 /* inversion list of code points this node matches only when the target
12163 * string is in UTF-8. (Because is under /d) */
12164 SV* depends_list = NULL;
12166 /* inversion list of code points this node matches. For much of the
12167 * function, it includes only those that match regardless of the utf8ness
12168 * of the target string */
12169 SV* cp_list = NULL;
12172 /* In a range, counts how many 0-2 of the ends of it came from literals,
12173 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
12174 UV literal_endpoint = 0;
12176 bool invert = FALSE; /* Is this class to be complemented */
12178 /* Is there any thing like \W or [:^digit:] that matches above the legal
12179 * Unicode range? */
12180 bool runtime_posix_matches_above_Unicode = FALSE;
12182 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
12183 case we need to change the emitted regop to an EXACT. */
12184 const char * orig_parse = RExC_parse;
12185 const I32 orig_size = RExC_size;
12186 GET_RE_DEBUG_FLAGS_DECL;
12188 PERL_ARGS_ASSERT_REGCLASS;
12190 PERL_UNUSED_ARG(depth);
12193 DEBUG_PARSE("clas");
12195 /* Assume we are going to generate an ANYOF node. */
12196 ret = reganode(pRExC_state, ANYOF, 0);
12199 RExC_size += ANYOF_SKIP;
12200 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
12203 ANYOF_FLAGS(ret) = 0;
12205 RExC_emit += ANYOF_SKIP;
12207 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
12209 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
12210 initial_listsv_len = SvCUR(listsv);
12211 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
12215 RExC_parse = regpatws(pRExC_state, RExC_parse,
12216 FALSE /* means don't recognize comments */);
12219 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
12222 allow_multi_folds = FALSE;
12225 RExC_parse = regpatws(pRExC_state, RExC_parse,
12226 FALSE /* means don't recognize comments */);
12230 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
12231 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
12232 const char *s = RExC_parse;
12233 const char c = *s++;
12235 while (isWORDCHAR(*s))
12237 if (*s && c == *s && s[1] == ']') {
12238 SAVEFREESV(RExC_rx_sv);
12240 "POSIX syntax [%c %c] belongs inside character classes",
12242 (void)ReREFCNT_inc(RExC_rx_sv);
12246 /* If the caller wants us to just parse a single element, accomplish this
12247 * by faking the loop ending condition */
12248 if (stop_at_1 && RExC_end > RExC_parse) {
12249 stop_ptr = RExC_parse + 1;
12252 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
12253 if (UCHARAT(RExC_parse) == ']')
12254 goto charclassloop;
12258 if (RExC_parse >= stop_ptr) {
12263 RExC_parse = regpatws(pRExC_state, RExC_parse,
12264 FALSE /* means don't recognize comments */);
12267 if (UCHARAT(RExC_parse) == ']') {
12273 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
12274 save_value = value;
12275 save_prevvalue = prevvalue;
12278 rangebegin = RExC_parse;
12282 value = utf8n_to_uvchr((U8*)RExC_parse,
12283 RExC_end - RExC_parse,
12284 &numlen, UTF8_ALLOW_DEFAULT);
12285 RExC_parse += numlen;
12288 value = UCHARAT(RExC_parse++);
12291 && RExC_parse < RExC_end
12292 && POSIXCC(UCHARAT(RExC_parse)))
12294 namedclass = regpposixcc(pRExC_state, value, strict);
12296 else if (value == '\\') {
12298 value = utf8n_to_uvchr((U8*)RExC_parse,
12299 RExC_end - RExC_parse,
12300 &numlen, UTF8_ALLOW_DEFAULT);
12301 RExC_parse += numlen;
12304 value = UCHARAT(RExC_parse++);
12306 /* Some compilers cannot handle switching on 64-bit integer
12307 * values, therefore value cannot be an UV. Yes, this will
12308 * be a problem later if we want switch on Unicode.
12309 * A similar issue a little bit later when switching on
12310 * namedclass. --jhi */
12312 /* If the \ is escaping white space when white space is being
12313 * skipped, it means that that white space is wanted literally, and
12314 * is already in 'value'. Otherwise, need to translate the escape
12315 * into what it signifies. */
12316 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
12318 case 'w': namedclass = ANYOF_WORDCHAR; break;
12319 case 'W': namedclass = ANYOF_NWORDCHAR; break;
12320 case 's': namedclass = ANYOF_SPACE; break;
12321 case 'S': namedclass = ANYOF_NSPACE; break;
12322 case 'd': namedclass = ANYOF_DIGIT; break;
12323 case 'D': namedclass = ANYOF_NDIGIT; break;
12324 case 'v': namedclass = ANYOF_VERTWS; break;
12325 case 'V': namedclass = ANYOF_NVERTWS; break;
12326 case 'h': namedclass = ANYOF_HORIZWS; break;
12327 case 'H': namedclass = ANYOF_NHORIZWS; break;
12328 case 'N': /* Handle \N{NAME} in class */
12330 /* We only pay attention to the first char of
12331 multichar strings being returned. I kinda wonder
12332 if this makes sense as it does change the behaviour
12333 from earlier versions, OTOH that behaviour was broken
12335 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
12336 TRUE, /* => charclass */
12339 if (*flagp & RESTART_UTF8)
12340 FAIL("panic: grok_bslash_N set RESTART_UTF8");
12350 /* We will handle any undefined properties ourselves */
12351 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF;
12353 if (RExC_parse >= RExC_end)
12354 vFAIL2("Empty \\%c{}", (U8)value);
12355 if (*RExC_parse == '{') {
12356 const U8 c = (U8)value;
12357 e = strchr(RExC_parse++, '}');
12359 vFAIL2("Missing right brace on \\%c{}", c);
12360 while (isSPACE(UCHARAT(RExC_parse)))
12362 if (e == RExC_parse)
12363 vFAIL2("Empty \\%c{}", c);
12364 n = e - RExC_parse;
12365 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
12376 if (UCHARAT(RExC_parse) == '^') {
12379 /* toggle. (The rhs xor gets the single bit that
12380 * differs between P and p; the other xor inverts just
12382 value ^= 'P' ^ 'p';
12384 while (isSPACE(UCHARAT(RExC_parse))) {
12389 /* Try to get the definition of the property into
12390 * <invlist>. If /i is in effect, the effective property
12391 * will have its name be <__NAME_i>. The design is
12392 * discussed in commit
12393 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
12394 Newx(name, n + sizeof("_i__\n"), char);
12396 sprintf(name, "%s%.*s%s\n",
12397 (FOLD) ? "__" : "",
12403 /* Look up the property name, and get its swash and
12404 * inversion list, if the property is found */
12406 SvREFCNT_dec_NN(swash);
12408 swash = _core_swash_init("utf8", name, &PL_sv_undef,
12411 NULL, /* No inversion list */
12414 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
12416 SvREFCNT_dec_NN(swash);
12420 /* Here didn't find it. It could be a user-defined
12421 * property that will be available at run-time. If we
12422 * accept only compile-time properties, is an error;
12423 * otherwise add it to the list for run-time look up */
12425 RExC_parse = e + 1;
12426 vFAIL3("Property '%.*s' is unknown", (int) n, name);
12428 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s\n",
12429 (value == 'p' ? '+' : '!'),
12431 has_user_defined_property = TRUE;
12433 /* We don't know yet, so have to assume that the
12434 * property could match something in the Latin1 range,
12435 * hence something that isn't utf8. Note that this
12436 * would cause things in <depends_list> to match
12437 * inappropriately, except that any \p{}, including
12438 * this one forces Unicode semantics, which means there
12439 * is <no depends_list> */
12440 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
12444 /* Here, did get the swash and its inversion list. If
12445 * the swash is from a user-defined property, then this
12446 * whole character class should be regarded as such */
12447 has_user_defined_property =
12449 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY);
12451 /* Invert if asking for the complement */
12452 if (value == 'P') {
12453 _invlist_union_complement_2nd(properties,
12457 /* The swash can't be used as-is, because we've
12458 * inverted things; delay removing it to here after
12459 * have copied its invlist above */
12460 SvREFCNT_dec_NN(swash);
12464 _invlist_union(properties, invlist, &properties);
12469 RExC_parse = e + 1;
12470 namedclass = ANYOF_UNIPROP; /* no official name, but it's
12473 /* \p means they want Unicode semantics */
12474 RExC_uni_semantics = 1;
12477 case 'n': value = '\n'; break;
12478 case 'r': value = '\r'; break;
12479 case 't': value = '\t'; break;
12480 case 'f': value = '\f'; break;
12481 case 'b': value = '\b'; break;
12482 case 'e': value = ASCII_TO_NATIVE('\033');break;
12483 case 'a': value = ASCII_TO_NATIVE('\007');break;
12485 RExC_parse--; /* function expects to be pointed at the 'o' */
12487 const char* error_msg;
12488 bool valid = grok_bslash_o(&RExC_parse,
12491 SIZE_ONLY, /* warnings in pass
12494 silence_non_portable,
12500 if (PL_encoding && value < 0x100) {
12501 goto recode_encoding;
12505 RExC_parse--; /* function expects to be pointed at the 'x' */
12507 const char* error_msg;
12508 bool valid = grok_bslash_x(&RExC_parse,
12511 TRUE, /* Output warnings */
12513 silence_non_portable,
12519 if (PL_encoding && value < 0x100)
12520 goto recode_encoding;
12523 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
12525 case '0': case '1': case '2': case '3': case '4':
12526 case '5': case '6': case '7':
12528 /* Take 1-3 octal digits */
12529 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12530 numlen = (strict) ? 4 : 3;
12531 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
12532 RExC_parse += numlen;
12535 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12536 vFAIL("Need exactly 3 octal digits");
12538 else if (! SIZE_ONLY /* like \08, \178 */
12540 && RExC_parse < RExC_end
12541 && isDIGIT(*RExC_parse)
12542 && ckWARN(WARN_REGEXP))
12544 SAVEFREESV(RExC_rx_sv);
12545 reg_warn_non_literal_string(
12547 form_short_octal_warning(RExC_parse, numlen));
12548 (void)ReREFCNT_inc(RExC_rx_sv);
12551 if (PL_encoding && value < 0x100)
12552 goto recode_encoding;
12556 if (! RExC_override_recoding) {
12557 SV* enc = PL_encoding;
12558 value = reg_recode((const char)(U8)value, &enc);
12561 vFAIL("Invalid escape in the specified encoding");
12563 else if (SIZE_ONLY) {
12564 ckWARNreg(RExC_parse,
12565 "Invalid escape in the specified encoding");
12571 /* Allow \_ to not give an error */
12572 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
12574 vFAIL2("Unrecognized escape \\%c in character class",
12578 SAVEFREESV(RExC_rx_sv);
12579 ckWARN2reg(RExC_parse,
12580 "Unrecognized escape \\%c in character class passed through",
12582 (void)ReREFCNT_inc(RExC_rx_sv);
12586 } /* End of switch on char following backslash */
12587 } /* end of handling backslash escape sequences */
12590 literal_endpoint++;
12593 /* Here, we have the current token in 'value' */
12595 /* What matches in a locale is not known until runtime. This includes
12596 * what the Posix classes (like \w, [:space:]) match. Room must be
12597 * reserved (one time per class) to store such classes, either if Perl
12598 * is compiled so that locale nodes always should have this space, or
12599 * if there is such class info to be stored. The space will contain a
12600 * bit for each named class that is to be matched against. This isn't
12601 * needed for \p{} and pseudo-classes, as they are not affected by
12602 * locale, and hence are dealt with separately */
12605 && (ANYOF_LOCALE == ANYOF_CLASS
12606 || (namedclass > OOB_NAMEDCLASS && namedclass < ANYOF_MAX)))
12610 RExC_size += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12613 RExC_emit += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12614 ANYOF_CLASS_ZERO(ret);
12616 ANYOF_FLAGS(ret) |= ANYOF_CLASS;
12619 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
12621 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
12622 * literal, as is the character that began the false range, i.e.
12623 * the 'a' in the examples */
12626 const int w = (RExC_parse >= rangebegin)
12627 ? RExC_parse - rangebegin
12630 vFAIL4("False [] range \"%*.*s\"", w, w, rangebegin);
12633 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
12634 ckWARN4reg(RExC_parse,
12635 "False [] range \"%*.*s\"",
12637 (void)ReREFCNT_inc(RExC_rx_sv);
12638 cp_list = add_cp_to_invlist(cp_list, '-');
12639 cp_list = add_cp_to_invlist(cp_list, prevvalue);
12643 range = 0; /* this was not a true range */
12644 element_count += 2; /* So counts for three values */
12648 U8 classnum = namedclass_to_classnum(namedclass);
12649 if (namedclass >= ANYOF_MAX) { /* If a special class */
12650 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
12652 /* Here, should be \h, \H, \v, or \V. Neither /d nor
12653 * /l make a difference in what these match. There
12654 * would be problems if these characters had folds
12655 * other than themselves, as cp_list is subject to
12657 if (classnum != _CC_VERTSPACE) {
12658 assert( namedclass == ANYOF_HORIZWS
12659 || namedclass == ANYOF_NHORIZWS);
12661 /* It turns out that \h is just a synonym for
12663 classnum = _CC_BLANK;
12666 _invlist_union_maybe_complement_2nd(
12668 PL_XPosix_ptrs[classnum],
12669 cBOOL(namedclass % 2), /* Complement if odd
12670 (NHORIZWS, NVERTWS)
12675 else if (classnum == _CC_ASCII) {
12678 ANYOF_CLASS_SET(ret, namedclass);
12681 #endif /* Not isascii(); just use the hard-coded definition for it */
12682 _invlist_union_maybe_complement_2nd(
12685 cBOOL(namedclass % 2), /* Complement if odd
12689 else { /* Garden variety class */
12691 /* The ascii range inversion list */
12692 SV* ascii_source = PL_Posix_ptrs[classnum];
12694 /* The full Latin1 range inversion list */
12695 SV* l1_source = PL_L1Posix_ptrs[classnum];
12697 /* This code is structured into two major clauses. The
12698 * first is for classes whose complete definitions may not
12699 * already be known. It not, the Latin1 definition
12700 * (guaranteed to already known) is used plus code is
12701 * generated to load the rest at run-time (only if needed).
12702 * If the complete definition is known, it drops down to
12703 * the second clause, where the complete definition is
12706 if (classnum < _FIRST_NON_SWASH_CC) {
12708 /* Here, the class has a swash, which may or not
12709 * already be loaded */
12711 /* The name of the property to use to match the full
12712 * eXtended Unicode range swash for this character
12714 const char *Xname = swash_property_names[classnum];
12716 /* If returning the inversion list, we can't defer
12717 * getting this until runtime */
12718 if (ret_invlist && ! PL_utf8_swash_ptrs[classnum]) {
12719 PL_utf8_swash_ptrs[classnum] =
12720 _core_swash_init("utf8", Xname, &PL_sv_undef,
12723 NULL, /* No inversion list */
12724 NULL /* No flags */
12726 assert(PL_utf8_swash_ptrs[classnum]);
12728 if ( ! PL_utf8_swash_ptrs[classnum]) {
12729 if (namedclass % 2 == 0) { /* A non-complemented
12731 /* If not /a matching, there are code points we
12732 * don't know at compile time. Arrange for the
12733 * unknown matches to be loaded at run-time, if
12735 if (! AT_LEAST_ASCII_RESTRICTED) {
12736 Perl_sv_catpvf(aTHX_ listsv, "+utf8::%s\n",
12739 if (LOC) { /* Under locale, set run-time
12741 ANYOF_CLASS_SET(ret, namedclass);
12744 /* Add the current class's code points to
12745 * the running total */
12746 _invlist_union(posixes,
12747 (AT_LEAST_ASCII_RESTRICTED)
12753 else { /* A complemented class */
12754 if (AT_LEAST_ASCII_RESTRICTED) {
12755 /* Under /a should match everything above
12756 * ASCII, plus the complement of the set's
12758 _invlist_union_complement_2nd(posixes,
12763 /* Arrange for the unknown matches to be
12764 * loaded at run-time, if needed */
12765 Perl_sv_catpvf(aTHX_ listsv, "!utf8::%s\n",
12767 runtime_posix_matches_above_Unicode = TRUE;
12769 ANYOF_CLASS_SET(ret, namedclass);
12773 /* We want to match everything in
12774 * Latin1, except those things that
12775 * l1_source matches */
12776 SV* scratch_list = NULL;
12777 _invlist_subtract(PL_Latin1, l1_source,
12780 /* Add the list from this class to the
12783 posixes = scratch_list;
12786 _invlist_union(posixes,
12789 SvREFCNT_dec_NN(scratch_list);
12791 if (DEPENDS_SEMANTICS) {
12793 |= ANYOF_NON_UTF8_LATIN1_ALL;
12798 goto namedclass_done;
12801 /* Here, there is a swash loaded for the class. If no
12802 * inversion list for it yet, get it */
12803 if (! PL_XPosix_ptrs[classnum]) {
12804 PL_XPosix_ptrs[classnum]
12805 = _swash_to_invlist(PL_utf8_swash_ptrs[classnum]);
12809 /* Here there is an inversion list already loaded for the
12812 if (namedclass % 2 == 0) { /* A non-complemented class,
12813 like ANYOF_PUNCT */
12815 /* For non-locale, just add it to any existing list
12817 _invlist_union(posixes,
12818 (AT_LEAST_ASCII_RESTRICTED)
12820 : PL_XPosix_ptrs[classnum],
12823 else { /* Locale */
12824 SV* scratch_list = NULL;
12826 /* For above Latin1 code points, we use the full
12828 _invlist_intersection(PL_AboveLatin1,
12829 PL_XPosix_ptrs[classnum],
12831 /* And set the output to it, adding instead if
12832 * there already is an output. Checking if
12833 * 'posixes' is NULL first saves an extra clone.
12834 * Its reference count will be decremented at the
12835 * next union, etc, or if this is the only
12836 * instance, at the end of the routine */
12838 posixes = scratch_list;
12841 _invlist_union(posixes, scratch_list, &posixes);
12842 SvREFCNT_dec_NN(scratch_list);
12845 #ifndef HAS_ISBLANK
12846 if (namedclass != ANYOF_BLANK) {
12848 /* Set this class in the node for runtime
12850 ANYOF_CLASS_SET(ret, namedclass);
12851 #ifndef HAS_ISBLANK
12854 /* No isblank(), use the hard-coded ASCII-range
12855 * blanks, adding them to the running total. */
12857 _invlist_union(posixes, ascii_source, &posixes);
12862 else { /* A complemented class, like ANYOF_NPUNCT */
12864 _invlist_union_complement_2nd(
12866 (AT_LEAST_ASCII_RESTRICTED)
12868 : PL_XPosix_ptrs[classnum],
12870 /* Under /d, everything in the upper half of the
12871 * Latin1 range matches this complement */
12872 if (DEPENDS_SEMANTICS) {
12873 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL;
12876 else { /* Locale */
12877 SV* scratch_list = NULL;
12878 _invlist_subtract(PL_AboveLatin1,
12879 PL_XPosix_ptrs[classnum],
12882 posixes = scratch_list;
12885 _invlist_union(posixes, scratch_list, &posixes);
12886 SvREFCNT_dec_NN(scratch_list);
12888 #ifndef HAS_ISBLANK
12889 if (namedclass != ANYOF_NBLANK) {
12891 ANYOF_CLASS_SET(ret, namedclass);
12892 #ifndef HAS_ISBLANK
12895 /* Get the list of all code points in Latin1
12896 * that are not ASCII blanks, and add them to
12897 * the running total */
12898 _invlist_subtract(PL_Latin1, ascii_source,
12900 _invlist_union(posixes, scratch_list, &posixes);
12901 SvREFCNT_dec_NN(scratch_list);
12908 continue; /* Go get next character */
12910 } /* end of namedclass \blah */
12912 /* Here, we have a single value. If 'range' is set, it is the ending
12913 * of a range--check its validity. Later, we will handle each
12914 * individual code point in the range. If 'range' isn't set, this
12915 * could be the beginning of a range, so check for that by looking
12916 * ahead to see if the next real character to be processed is the range
12917 * indicator--the minus sign */
12920 RExC_parse = regpatws(pRExC_state, RExC_parse,
12921 FALSE /* means don't recognize comments */);
12925 if (prevvalue > value) /* b-a */ {
12926 const int w = RExC_parse - rangebegin;
12927 Simple_vFAIL4("Invalid [] range \"%*.*s\"", w, w, rangebegin);
12928 range = 0; /* not a valid range */
12932 prevvalue = value; /* save the beginning of the potential range */
12933 if (! stop_at_1 /* Can't be a range if parsing just one thing */
12934 && *RExC_parse == '-')
12936 char* next_char_ptr = RExC_parse + 1;
12937 if (skip_white) { /* Get the next real char after the '-' */
12938 next_char_ptr = regpatws(pRExC_state,
12940 FALSE); /* means don't recognize
12944 /* If the '-' is at the end of the class (just before the ']',
12945 * it is a literal minus; otherwise it is a range */
12946 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
12947 RExC_parse = next_char_ptr;
12949 /* a bad range like \w-, [:word:]- ? */
12950 if (namedclass > OOB_NAMEDCLASS) {
12951 if (strict || ckWARN(WARN_REGEXP)) {
12953 RExC_parse >= rangebegin ?
12954 RExC_parse - rangebegin : 0;
12956 vFAIL4("False [] range \"%*.*s\"",
12961 "False [] range \"%*.*s\"",
12966 cp_list = add_cp_to_invlist(cp_list, '-');
12970 range = 1; /* yeah, it's a range! */
12971 continue; /* but do it the next time */
12976 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
12979 /* non-Latin1 code point implies unicode semantics. Must be set in
12980 * pass1 so is there for the whole of pass 2 */
12982 RExC_uni_semantics = 1;
12985 /* Ready to process either the single value, or the completed range.
12986 * For single-valued non-inverted ranges, we consider the possibility
12987 * of multi-char folds. (We made a conscious decision to not do this
12988 * for the other cases because it can often lead to non-intuitive
12989 * results. For example, you have the peculiar case that:
12990 * "s s" =~ /^[^\xDF]+$/i => Y
12991 * "ss" =~ /^[^\xDF]+$/i => N
12993 * See [perl #89750] */
12994 if (FOLD && allow_multi_folds && value == prevvalue) {
12995 if (value == LATIN_SMALL_LETTER_SHARP_S
12996 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
12999 /* Here <value> is indeed a multi-char fold. Get what it is */
13001 U8 foldbuf[UTF8_MAXBYTES_CASE];
13004 UV folded = _to_uni_fold_flags(
13009 | ((LOC) ? FOLD_FLAGS_LOCALE
13010 : (ASCII_FOLD_RESTRICTED)
13011 ? FOLD_FLAGS_NOMIX_ASCII
13015 /* Here, <folded> should be the first character of the
13016 * multi-char fold of <value>, with <foldbuf> containing the
13017 * whole thing. But, if this fold is not allowed (because of
13018 * the flags), <fold> will be the same as <value>, and should
13019 * be processed like any other character, so skip the special
13021 if (folded != value) {
13023 /* Skip if we are recursed, currently parsing the class
13024 * again. Otherwise add this character to the list of
13025 * multi-char folds. */
13026 if (! RExC_in_multi_char_class) {
13027 AV** this_array_ptr;
13029 STRLEN cp_count = utf8_length(foldbuf,
13030 foldbuf + foldlen);
13031 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
13033 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
13036 if (! multi_char_matches) {
13037 multi_char_matches = newAV();
13040 /* <multi_char_matches> is actually an array of arrays.
13041 * There will be one or two top-level elements: [2],
13042 * and/or [3]. The [2] element is an array, each
13043 * element thereof is a character which folds to TWO
13044 * characters; [3] is for folds to THREE characters.
13045 * (Unicode guarantees a maximum of 3 characters in any
13046 * fold.) When we rewrite the character class below,
13047 * we will do so such that the longest folds are
13048 * written first, so that it prefers the longest
13049 * matching strings first. This is done even if it
13050 * turns out that any quantifier is non-greedy, out of
13051 * programmer laziness. Tom Christiansen has agreed
13052 * that this is ok. This makes the test for the
13053 * ligature 'ffi' come before the test for 'ff' */
13054 if (av_exists(multi_char_matches, cp_count)) {
13055 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13057 this_array = *this_array_ptr;
13060 this_array = newAV();
13061 av_store(multi_char_matches, cp_count,
13064 av_push(this_array, multi_fold);
13067 /* This element should not be processed further in this
13070 value = save_value;
13071 prevvalue = save_prevvalue;
13077 /* Deal with this element of the class */
13080 cp_list = _add_range_to_invlist(cp_list, prevvalue, value);
13082 SV* this_range = _new_invlist(1);
13083 _append_range_to_invlist(this_range, prevvalue, value);
13085 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
13086 * If this range was specified using something like 'i-j', we want
13087 * to include only the 'i' and the 'j', and not anything in
13088 * between, so exclude non-ASCII, non-alphabetics from it.
13089 * However, if the range was specified with something like
13090 * [\x89-\x91] or [\x89-j], all code points within it should be
13091 * included. literal_endpoint==2 means both ends of the range used
13092 * a literal character, not \x{foo} */
13093 if (literal_endpoint == 2
13094 && (prevvalue >= 'a' && value <= 'z')
13095 || (prevvalue >= 'A' && value <= 'Z'))
13097 _invlist_intersection(this_range, PL_Posix_ptrs[_CC_ALPHA],
13100 _invlist_union(cp_list, this_range, &cp_list);
13101 literal_endpoint = 0;
13105 range = 0; /* this range (if it was one) is done now */
13106 } /* End of loop through all the text within the brackets */
13108 /* If anything in the class expands to more than one character, we have to
13109 * deal with them by building up a substitute parse string, and recursively
13110 * calling reg() on it, instead of proceeding */
13111 if (multi_char_matches) {
13112 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
13115 char *save_end = RExC_end;
13116 char *save_parse = RExC_parse;
13117 bool first_time = TRUE; /* First multi-char occurrence doesn't get
13122 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
13123 because too confusing */
13125 sv_catpv(substitute_parse, "(?:");
13129 /* Look at the longest folds first */
13130 for (cp_count = av_len(multi_char_matches); cp_count > 0; cp_count--) {
13132 if (av_exists(multi_char_matches, cp_count)) {
13133 AV** this_array_ptr;
13136 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13138 while ((this_sequence = av_pop(*this_array_ptr)) !=
13141 if (! first_time) {
13142 sv_catpv(substitute_parse, "|");
13144 first_time = FALSE;
13146 sv_catpv(substitute_parse, SvPVX(this_sequence));
13151 /* If the character class contains anything else besides these
13152 * multi-character folds, have to include it in recursive parsing */
13153 if (element_count) {
13154 sv_catpv(substitute_parse, "|[");
13155 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
13156 sv_catpv(substitute_parse, "]");
13159 sv_catpv(substitute_parse, ")");
13162 /* This is a way to get the parse to skip forward a whole named
13163 * sequence instead of matching the 2nd character when it fails the
13165 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
13169 RExC_parse = SvPV(substitute_parse, len);
13170 RExC_end = RExC_parse + len;
13171 RExC_in_multi_char_class = 1;
13172 RExC_emit = (regnode *)orig_emit;
13174 ret = reg(pRExC_state, 1, ®_flags, depth+1);
13176 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
13178 RExC_parse = save_parse;
13179 RExC_end = save_end;
13180 RExC_in_multi_char_class = 0;
13181 SvREFCNT_dec_NN(multi_char_matches);
13185 /* If the character class contains only a single element, it may be
13186 * optimizable into another node type which is smaller and runs faster.
13187 * Check if this is the case for this class */
13188 if (element_count == 1 && ! ret_invlist) {
13192 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like \w or
13193 [:digit:] or \p{foo} */
13195 /* All named classes are mapped into POSIXish nodes, with its FLAG
13196 * argument giving which class it is */
13197 switch ((I32)namedclass) {
13198 case ANYOF_UNIPROP:
13201 /* These don't depend on the charset modifiers. They always
13202 * match under /u rules */
13203 case ANYOF_NHORIZWS:
13204 case ANYOF_HORIZWS:
13205 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
13208 case ANYOF_NVERTWS:
13213 /* The actual POSIXish node for all the rest depends on the
13214 * charset modifier. The ones in the first set depend only on
13215 * ASCII or, if available on this platform, locale */
13219 op = (LOC) ? POSIXL : POSIXA;
13230 /* under /a could be alpha */
13232 if (ASCII_RESTRICTED) {
13233 namedclass = ANYOF_ALPHA + (namedclass % 2);
13241 /* The rest have more possibilities depending on the charset.
13242 * We take advantage of the enum ordering of the charset
13243 * modifiers to get the exact node type, */
13245 op = POSIXD + get_regex_charset(RExC_flags);
13246 if (op > POSIXA) { /* /aa is same as /a */
13249 #ifndef HAS_ISBLANK
13251 && (namedclass == ANYOF_BLANK
13252 || namedclass == ANYOF_NBLANK))
13259 /* The odd numbered ones are the complements of the
13260 * next-lower even number one */
13261 if (namedclass % 2 == 1) {
13265 arg = namedclass_to_classnum(namedclass);
13269 else if (value == prevvalue) {
13271 /* Here, the class consists of just a single code point */
13274 if (! LOC && value == '\n') {
13275 op = REG_ANY; /* Optimize [^\n] */
13276 *flagp |= HASWIDTH|SIMPLE;
13280 else if (value < 256 || UTF) {
13282 /* Optimize a single value into an EXACTish node, but not if it
13283 * would require converting the pattern to UTF-8. */
13284 op = compute_EXACTish(pRExC_state);
13286 } /* Otherwise is a range */
13287 else if (! LOC) { /* locale could vary these */
13288 if (prevvalue == '0') {
13289 if (value == '9') {
13296 /* Here, we have changed <op> away from its initial value iff we found
13297 * an optimization */
13300 /* Throw away this ANYOF regnode, and emit the calculated one,
13301 * which should correspond to the beginning, not current, state of
13303 const char * cur_parse = RExC_parse;
13304 RExC_parse = (char *)orig_parse;
13308 /* To get locale nodes to not use the full ANYOF size would
13309 * require moving the code above that writes the portions
13310 * of it that aren't in other nodes to after this point.
13311 * e.g. ANYOF_CLASS_SET */
13312 RExC_size = orig_size;
13316 RExC_emit = (regnode *)orig_emit;
13317 if (PL_regkind[op] == POSIXD) {
13319 op += NPOSIXD - POSIXD;
13324 ret = reg_node(pRExC_state, op);
13326 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
13330 *flagp |= HASWIDTH|SIMPLE;
13332 else if (PL_regkind[op] == EXACT) {
13333 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13336 RExC_parse = (char *) cur_parse;
13338 SvREFCNT_dec(posixes);
13339 SvREFCNT_dec(cp_list);
13346 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
13348 /* If folding, we calculate all characters that could fold to or from the
13349 * ones already on the list */
13350 if (FOLD && cp_list) {
13351 UV start, end; /* End points of code point ranges */
13353 SV* fold_intersection = NULL;
13355 /* If the highest code point is within Latin1, we can use the
13356 * compiled-in Alphas list, and not have to go out to disk. This
13357 * yields two false positives, the masculine and feminine ordinal
13358 * indicators, which are weeded out below using the
13359 * IS_IN_SOME_FOLD_L1() macro */
13360 if (invlist_highest(cp_list) < 256) {
13361 _invlist_intersection(PL_L1Posix_ptrs[_CC_ALPHA], cp_list,
13362 &fold_intersection);
13366 /* Here, there are non-Latin1 code points, so we will have to go
13367 * fetch the list of all the characters that participate in folds
13369 if (! PL_utf8_foldable) {
13370 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13371 &PL_sv_undef, 1, 0);
13372 PL_utf8_foldable = _get_swash_invlist(swash);
13373 SvREFCNT_dec_NN(swash);
13376 /* This is a hash that for a particular fold gives all characters
13377 * that are involved in it */
13378 if (! PL_utf8_foldclosures) {
13380 /* If we were unable to find any folds, then we likely won't be
13381 * able to find the closures. So just create an empty list.
13382 * Folding will effectively be restricted to the non-Unicode
13383 * rules hard-coded into Perl. (This case happens legitimately
13384 * during compilation of Perl itself before the Unicode tables
13385 * are generated) */
13386 if (_invlist_len(PL_utf8_foldable) == 0) {
13387 PL_utf8_foldclosures = newHV();
13390 /* If the folds haven't been read in, call a fold function
13392 if (! PL_utf8_tofold) {
13393 U8 dummy[UTF8_MAXBYTES+1];
13395 /* This string is just a short named one above \xff */
13396 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
13397 assert(PL_utf8_tofold); /* Verify that worked */
13399 PL_utf8_foldclosures =
13400 _swash_inversion_hash(PL_utf8_tofold);
13404 /* Only the characters in this class that participate in folds need
13405 * be checked. Get the intersection of this class and all the
13406 * possible characters that are foldable. This can quickly narrow
13407 * down a large class */
13408 _invlist_intersection(PL_utf8_foldable, cp_list,
13409 &fold_intersection);
13412 /* Now look at the foldable characters in this class individually */
13413 invlist_iterinit(fold_intersection);
13414 while (invlist_iternext(fold_intersection, &start, &end)) {
13417 /* Locale folding for Latin1 characters is deferred until runtime */
13418 if (LOC && start < 256) {
13422 /* Look at every character in the range */
13423 for (j = start; j <= end; j++) {
13425 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
13431 /* We have the latin1 folding rules hard-coded here so that
13432 * an innocent-looking character class, like /[ks]/i won't
13433 * have to go out to disk to find the possible matches.
13434 * XXX It would be better to generate these via regen, in
13435 * case a new version of the Unicode standard adds new
13436 * mappings, though that is not really likely, and may be
13437 * caught by the default: case of the switch below. */
13439 if (IS_IN_SOME_FOLD_L1(j)) {
13441 /* ASCII is always matched; non-ASCII is matched only
13442 * under Unicode rules */
13443 if (isASCII(j) || AT_LEAST_UNI_SEMANTICS) {
13445 add_cp_to_invlist(cp_list, PL_fold_latin1[j]);
13449 add_cp_to_invlist(depends_list, PL_fold_latin1[j]);
13453 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
13454 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
13456 /* Certain Latin1 characters have matches outside
13457 * Latin1. To get here, <j> is one of those
13458 * characters. None of these matches is valid for
13459 * ASCII characters under /aa, which is why the 'if'
13460 * just above excludes those. These matches only
13461 * happen when the target string is utf8. The code
13462 * below adds the single fold closures for <j> to the
13463 * inversion list. */
13468 add_cp_to_invlist(cp_list, KELVIN_SIGN);
13472 cp_list = add_cp_to_invlist(cp_list,
13473 LATIN_SMALL_LETTER_LONG_S);
13476 cp_list = add_cp_to_invlist(cp_list,
13477 GREEK_CAPITAL_LETTER_MU);
13478 cp_list = add_cp_to_invlist(cp_list,
13479 GREEK_SMALL_LETTER_MU);
13481 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
13482 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
13484 add_cp_to_invlist(cp_list, ANGSTROM_SIGN);
13486 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
13487 cp_list = add_cp_to_invlist(cp_list,
13488 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
13490 case LATIN_SMALL_LETTER_SHARP_S:
13491 cp_list = add_cp_to_invlist(cp_list,
13492 LATIN_CAPITAL_LETTER_SHARP_S);
13494 case 'F': case 'f':
13495 case 'I': case 'i':
13496 case 'L': case 'l':
13497 case 'T': case 't':
13498 case 'A': case 'a':
13499 case 'H': case 'h':
13500 case 'J': case 'j':
13501 case 'N': case 'n':
13502 case 'W': case 'w':
13503 case 'Y': case 'y':
13504 /* These all are targets of multi-character
13505 * folds from code points that require UTF8 to
13506 * express, so they can't match unless the
13507 * target string is in UTF-8, so no action here
13508 * is necessary, as regexec.c properly handles
13509 * the general case for UTF-8 matching and
13510 * multi-char folds */
13513 /* Use deprecated warning to increase the
13514 * chances of this being output */
13515 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
13522 /* Here is an above Latin1 character. We don't have the rules
13523 * hard-coded for it. First, get its fold. This is the simple
13524 * fold, as the multi-character folds have been handled earlier
13525 * and separated out */
13526 _to_uni_fold_flags(j, foldbuf, &foldlen,
13528 ? FOLD_FLAGS_LOCALE
13529 : (ASCII_FOLD_RESTRICTED)
13530 ? FOLD_FLAGS_NOMIX_ASCII
13533 /* Single character fold of above Latin1. Add everything in
13534 * its fold closure to the list that this node should match.
13535 * The fold closures data structure is a hash with the keys
13536 * being the UTF-8 of every character that is folded to, like
13537 * 'k', and the values each an array of all code points that
13538 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
13539 * Multi-character folds are not included */
13540 if ((listp = hv_fetch(PL_utf8_foldclosures,
13541 (char *) foldbuf, foldlen, FALSE)))
13543 AV* list = (AV*) *listp;
13545 for (k = 0; k <= av_len(list); k++) {
13546 SV** c_p = av_fetch(list, k, FALSE);
13549 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
13553 /* /aa doesn't allow folds between ASCII and non-; /l
13554 * doesn't allow them between above and below 256 */
13555 if ((ASCII_FOLD_RESTRICTED
13556 && (isASCII(c) != isASCII(j)))
13557 || (LOC && c < 256)) {
13561 /* Folds involving non-ascii Latin1 characters
13562 * under /d are added to a separate list */
13563 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
13565 cp_list = add_cp_to_invlist(cp_list, c);
13568 depends_list = add_cp_to_invlist(depends_list, c);
13574 SvREFCNT_dec_NN(fold_intersection);
13577 /* And combine the result (if any) with any inversion list from posix
13578 * classes. The lists are kept separate up to now because we don't want to
13579 * fold the classes (folding of those is automatically handled by the swash
13580 * fetching code) */
13582 if (! DEPENDS_SEMANTICS) {
13584 _invlist_union(cp_list, posixes, &cp_list);
13585 SvREFCNT_dec_NN(posixes);
13592 /* Under /d, we put into a separate list the Latin1 things that
13593 * match only when the target string is utf8 */
13594 SV* nonascii_but_latin1_properties = NULL;
13595 _invlist_intersection(posixes, PL_Latin1,
13596 &nonascii_but_latin1_properties);
13597 _invlist_subtract(nonascii_but_latin1_properties, PL_ASCII,
13598 &nonascii_but_latin1_properties);
13599 _invlist_subtract(posixes, nonascii_but_latin1_properties,
13602 _invlist_union(cp_list, posixes, &cp_list);
13603 SvREFCNT_dec_NN(posixes);
13609 if (depends_list) {
13610 _invlist_union(depends_list, nonascii_but_latin1_properties,
13612 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
13615 depends_list = nonascii_but_latin1_properties;
13620 /* And combine the result (if any) with any inversion list from properties.
13621 * The lists are kept separate up to now so that we can distinguish the two
13622 * in regards to matching above-Unicode. A run-time warning is generated
13623 * if a Unicode property is matched against a non-Unicode code point. But,
13624 * we allow user-defined properties to match anything, without any warning,
13625 * and we also suppress the warning if there is a portion of the character
13626 * class that isn't a Unicode property, and which matches above Unicode, \W
13627 * or [\x{110000}] for example.
13628 * (Note that in this case, unlike the Posix one above, there is no
13629 * <depends_list>, because having a Unicode property forces Unicode
13632 bool warn_super = ! has_user_defined_property;
13635 /* If it matters to the final outcome, see if a non-property
13636 * component of the class matches above Unicode. If so, the
13637 * warning gets suppressed. This is true even if just a single
13638 * such code point is specified, as though not strictly correct if
13639 * another such code point is matched against, the fact that they
13640 * are using above-Unicode code points indicates they should know
13641 * the issues involved */
13643 bool non_prop_matches_above_Unicode =
13644 runtime_posix_matches_above_Unicode
13645 | (invlist_highest(cp_list) > PERL_UNICODE_MAX);
13647 non_prop_matches_above_Unicode =
13648 ! non_prop_matches_above_Unicode;
13650 warn_super = ! non_prop_matches_above_Unicode;
13653 _invlist_union(properties, cp_list, &cp_list);
13654 SvREFCNT_dec_NN(properties);
13657 cp_list = properties;
13661 OP(ret) = ANYOF_WARN_SUPER;
13665 /* Here, we have calculated what code points should be in the character
13668 * Now we can see about various optimizations. Fold calculation (which we
13669 * did above) needs to take place before inversion. Otherwise /[^k]/i
13670 * would invert to include K, which under /i would match k, which it
13671 * shouldn't. Therefore we can't invert folded locale now, as it won't be
13672 * folded until runtime */
13674 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
13675 * at compile time. Besides not inverting folded locale now, we can't
13676 * invert if there are things such as \w, which aren't known until runtime
13679 && ! (LOC && (FOLD || (ANYOF_FLAGS(ret) & ANYOF_CLASS)))
13681 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13683 _invlist_invert(cp_list);
13685 /* Any swash can't be used as-is, because we've inverted things */
13687 SvREFCNT_dec_NN(swash);
13691 /* Clear the invert flag since have just done it here */
13696 *ret_invlist = cp_list;
13697 SvREFCNT_dec(swash);
13699 /* Discard the generated node */
13701 RExC_size = orig_size;
13704 RExC_emit = orig_emit;
13709 /* If we didn't do folding, it's because some information isn't available
13710 * until runtime; set the run-time fold flag for these. (We don't have to
13711 * worry about properties folding, as that is taken care of by the swash
13715 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
13718 /* Some character classes are equivalent to other nodes. Such nodes take
13719 * up less room and generally fewer operations to execute than ANYOF nodes.
13720 * Above, we checked for and optimized into some such equivalents for
13721 * certain common classes that are easy to test. Getting to this point in
13722 * the code means that the class didn't get optimized there. Since this
13723 * code is only executed in Pass 2, it is too late to save space--it has
13724 * been allocated in Pass 1, and currently isn't given back. But turning
13725 * things into an EXACTish node can allow the optimizer to join it to any
13726 * adjacent such nodes. And if the class is equivalent to things like /./,
13727 * expensive run-time swashes can be avoided. Now that we have more
13728 * complete information, we can find things necessarily missed by the
13729 * earlier code. I (khw) am not sure how much to look for here. It would
13730 * be easy, but perhaps too slow, to check any candidates against all the
13731 * node types they could possibly match using _invlistEQ(). */
13736 && ! (ANYOF_FLAGS(ret) & ANYOF_CLASS)
13737 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13740 U8 op = END; /* The optimzation node-type */
13741 const char * cur_parse= RExC_parse;
13743 invlist_iterinit(cp_list);
13744 if (! invlist_iternext(cp_list, &start, &end)) {
13746 /* Here, the list is empty. This happens, for example, when a
13747 * Unicode property is the only thing in the character class, and
13748 * it doesn't match anything. (perluniprops.pod notes such
13751 *flagp |= HASWIDTH|SIMPLE;
13753 else if (start == end) { /* The range is a single code point */
13754 if (! invlist_iternext(cp_list, &start, &end)
13756 /* Don't do this optimization if it would require changing
13757 * the pattern to UTF-8 */
13758 && (start < 256 || UTF))
13760 /* Here, the list contains a single code point. Can optimize
13761 * into an EXACT node */
13770 /* A locale node under folding with one code point can be
13771 * an EXACTFL, as its fold won't be calculated until
13777 /* Here, we are generally folding, but there is only one
13778 * code point to match. If we have to, we use an EXACT
13779 * node, but it would be better for joining with adjacent
13780 * nodes in the optimization pass if we used the same
13781 * EXACTFish node that any such are likely to be. We can
13782 * do this iff the code point doesn't participate in any
13783 * folds. For example, an EXACTF of a colon is the same as
13784 * an EXACT one, since nothing folds to or from a colon. */
13786 if (IS_IN_SOME_FOLD_L1(value)) {
13791 if (! PL_utf8_foldable) {
13792 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13793 &PL_sv_undef, 1, 0);
13794 PL_utf8_foldable = _get_swash_invlist(swash);
13795 SvREFCNT_dec_NN(swash);
13797 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
13802 /* If we haven't found the node type, above, it means we
13803 * can use the prevailing one */
13805 op = compute_EXACTish(pRExC_state);
13810 else if (start == 0) {
13811 if (end == UV_MAX) {
13813 *flagp |= HASWIDTH|SIMPLE;
13816 else if (end == '\n' - 1
13817 && invlist_iternext(cp_list, &start, &end)
13818 && start == '\n' + 1 && end == UV_MAX)
13821 *flagp |= HASWIDTH|SIMPLE;
13825 invlist_iterfinish(cp_list);
13828 RExC_parse = (char *)orig_parse;
13829 RExC_emit = (regnode *)orig_emit;
13831 ret = reg_node(pRExC_state, op);
13833 RExC_parse = (char *)cur_parse;
13835 if (PL_regkind[op] == EXACT) {
13836 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13839 SvREFCNT_dec_NN(cp_list);
13844 /* Here, <cp_list> contains all the code points we can determine at
13845 * compile time that match under all conditions. Go through it, and
13846 * for things that belong in the bitmap, put them there, and delete from
13847 * <cp_list>. While we are at it, see if everything above 255 is in the
13848 * list, and if so, set a flag to speed up execution */
13849 ANYOF_BITMAP_ZERO(ret);
13852 /* This gets set if we actually need to modify things */
13853 bool change_invlist = FALSE;
13857 /* Start looking through <cp_list> */
13858 invlist_iterinit(cp_list);
13859 while (invlist_iternext(cp_list, &start, &end)) {
13863 if (end == UV_MAX && start <= 256) {
13864 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
13867 /* Quit if are above what we should change */
13872 change_invlist = TRUE;
13874 /* Set all the bits in the range, up to the max that we are doing */
13875 high = (end < 255) ? end : 255;
13876 for (i = start; i <= (int) high; i++) {
13877 if (! ANYOF_BITMAP_TEST(ret, i)) {
13878 ANYOF_BITMAP_SET(ret, i);
13884 invlist_iterfinish(cp_list);
13886 /* Done with loop; remove any code points that are in the bitmap from
13888 if (change_invlist) {
13889 _invlist_subtract(cp_list, PL_Latin1, &cp_list);
13892 /* If have completely emptied it, remove it completely */
13893 if (_invlist_len(cp_list) == 0) {
13894 SvREFCNT_dec_NN(cp_list);
13900 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
13903 /* Here, the bitmap has been populated with all the Latin1 code points that
13904 * always match. Can now add to the overall list those that match only
13905 * when the target string is UTF-8 (<depends_list>). */
13906 if (depends_list) {
13908 _invlist_union(cp_list, depends_list, &cp_list);
13909 SvREFCNT_dec_NN(depends_list);
13912 cp_list = depends_list;
13916 /* If there is a swash and more than one element, we can't use the swash in
13917 * the optimization below. */
13918 if (swash && element_count > 1) {
13919 SvREFCNT_dec_NN(swash);
13924 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13926 ARG_SET(ret, ANYOF_NONBITMAP_EMPTY);
13929 /* av[0] stores the character class description in its textual form:
13930 * used later (regexec.c:Perl_regclass_swash()) to initialize the
13931 * appropriate swash, and is also useful for dumping the regnode.
13932 * av[1] if NULL, is a placeholder to later contain the swash computed
13933 * from av[0]. But if no further computation need be done, the
13934 * swash is stored there now.
13935 * av[2] stores the cp_list inversion list for use in addition or
13936 * instead of av[0]; used only if av[1] is NULL
13937 * av[3] is set if any component of the class is from a user-defined
13938 * property; used only if av[1] is NULL */
13939 AV * const av = newAV();
13942 av_store(av, 0, (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13943 ? SvREFCNT_inc(listsv) : &PL_sv_undef);
13945 av_store(av, 1, swash);
13946 SvREFCNT_dec_NN(cp_list);
13949 av_store(av, 1, NULL);
13951 av_store(av, 2, cp_list);
13952 av_store(av, 3, newSVuv(has_user_defined_property));
13956 rv = newRV_noinc(MUTABLE_SV(av));
13957 n = add_data(pRExC_state, 1, "s");
13958 RExC_rxi->data->data[n] = (void*)rv;
13962 *flagp |= HASWIDTH|SIMPLE;
13965 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
13968 /* reg_skipcomment()
13970 Absorbs an /x style # comments from the input stream.
13971 Returns true if there is more text remaining in the stream.
13972 Will set the REG_SEEN_RUN_ON_COMMENT flag if the comment
13973 terminates the pattern without including a newline.
13975 Note its the callers responsibility to ensure that we are
13976 actually in /x mode
13981 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
13985 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
13987 while (RExC_parse < RExC_end)
13988 if (*RExC_parse++ == '\n') {
13993 /* we ran off the end of the pattern without ending
13994 the comment, so we have to add an \n when wrapping */
13995 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
14003 Advances the parse position, and optionally absorbs
14004 "whitespace" from the inputstream.
14006 Without /x "whitespace" means (?#...) style comments only,
14007 with /x this means (?#...) and # comments and whitespace proper.
14009 Returns the RExC_parse point from BEFORE the scan occurs.
14011 This is the /x friendly way of saying RExC_parse++.
14015 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
14017 char* const retval = RExC_parse++;
14019 PERL_ARGS_ASSERT_NEXTCHAR;
14022 if (RExC_end - RExC_parse >= 3
14023 && *RExC_parse == '('
14024 && RExC_parse[1] == '?'
14025 && RExC_parse[2] == '#')
14027 while (*RExC_parse != ')') {
14028 if (RExC_parse == RExC_end)
14029 FAIL("Sequence (?#... not terminated");
14035 if (RExC_flags & RXf_PMf_EXTENDED) {
14036 if (isSPACE(*RExC_parse)) {
14040 else if (*RExC_parse == '#') {
14041 if ( reg_skipcomment( pRExC_state ) )
14050 - reg_node - emit a node
14052 STATIC regnode * /* Location. */
14053 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
14057 regnode * const ret = RExC_emit;
14058 GET_RE_DEBUG_FLAGS_DECL;
14060 PERL_ARGS_ASSERT_REG_NODE;
14063 SIZE_ALIGN(RExC_size);
14067 if (RExC_emit >= RExC_emit_bound)
14068 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14069 op, RExC_emit, RExC_emit_bound);
14071 NODE_ALIGN_FILL(ret);
14073 FILL_ADVANCE_NODE(ptr, op);
14074 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (ptr) - 1);
14075 #ifdef RE_TRACK_PATTERN_OFFSETS
14076 if (RExC_offsets) { /* MJD */
14077 MJD_OFFSET_DEBUG(("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
14078 "reg_node", __LINE__,
14080 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
14081 ? "Overwriting end of array!\n" : "OK",
14082 (UV)(RExC_emit - RExC_emit_start),
14083 (UV)(RExC_parse - RExC_start),
14084 (UV)RExC_offsets[0]));
14085 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
14093 - reganode - emit a node with an argument
14095 STATIC regnode * /* Location. */
14096 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
14100 regnode * const ret = RExC_emit;
14101 GET_RE_DEBUG_FLAGS_DECL;
14103 PERL_ARGS_ASSERT_REGANODE;
14106 SIZE_ALIGN(RExC_size);
14111 assert(2==regarglen[op]+1);
14113 Anything larger than this has to allocate the extra amount.
14114 If we changed this to be:
14116 RExC_size += (1 + regarglen[op]);
14118 then it wouldn't matter. Its not clear what side effect
14119 might come from that so its not done so far.
14124 if (RExC_emit >= RExC_emit_bound)
14125 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14126 op, RExC_emit, RExC_emit_bound);
14128 NODE_ALIGN_FILL(ret);
14130 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
14131 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (ptr) - 2);
14132 #ifdef RE_TRACK_PATTERN_OFFSETS
14133 if (RExC_offsets) { /* MJD */
14134 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14138 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
14139 "Overwriting end of array!\n" : "OK",
14140 (UV)(RExC_emit - RExC_emit_start),
14141 (UV)(RExC_parse - RExC_start),
14142 (UV)RExC_offsets[0]));
14143 Set_Cur_Node_Offset;
14151 - reguni - emit (if appropriate) a Unicode character
14154 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
14158 PERL_ARGS_ASSERT_REGUNI;
14160 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
14164 - reginsert - insert an operator in front of already-emitted operand
14166 * Means relocating the operand.
14169 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
14175 const int offset = regarglen[(U8)op];
14176 const int size = NODE_STEP_REGNODE + offset;
14177 GET_RE_DEBUG_FLAGS_DECL;
14179 PERL_ARGS_ASSERT_REGINSERT;
14180 PERL_UNUSED_ARG(depth);
14181 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
14182 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
14191 if (RExC_open_parens) {
14193 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
14194 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
14195 if ( RExC_open_parens[paren] >= opnd ) {
14196 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
14197 RExC_open_parens[paren] += size;
14199 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
14201 if ( RExC_close_parens[paren] >= opnd ) {
14202 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
14203 RExC_close_parens[paren] += size;
14205 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
14210 while (src > opnd) {
14211 StructCopy(--src, --dst, regnode);
14212 #ifdef RE_TRACK_PATTERN_OFFSETS
14213 if (RExC_offsets) { /* MJD 20010112 */
14214 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
14218 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
14219 ? "Overwriting end of array!\n" : "OK",
14220 (UV)(src - RExC_emit_start),
14221 (UV)(dst - RExC_emit_start),
14222 (UV)RExC_offsets[0]));
14223 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
14224 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
14230 place = opnd; /* Op node, where operand used to be. */
14231 #ifdef RE_TRACK_PATTERN_OFFSETS
14232 if (RExC_offsets) { /* MJD */
14233 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14237 (UV)(place - RExC_emit_start) > RExC_offsets[0]
14238 ? "Overwriting end of array!\n" : "OK",
14239 (UV)(place - RExC_emit_start),
14240 (UV)(RExC_parse - RExC_start),
14241 (UV)RExC_offsets[0]));
14242 Set_Node_Offset(place, RExC_parse);
14243 Set_Node_Length(place, 1);
14246 src = NEXTOPER(place);
14247 FILL_ADVANCE_NODE(place, op);
14248 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (place) - 1);
14249 Zero(src, offset, regnode);
14253 - regtail - set the next-pointer at the end of a node chain of p to val.
14254 - SEE ALSO: regtail_study
14256 /* TODO: All three parms should be const */
14258 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14262 GET_RE_DEBUG_FLAGS_DECL;
14264 PERL_ARGS_ASSERT_REGTAIL;
14266 PERL_UNUSED_ARG(depth);
14272 /* Find last node. */
14275 regnode * const temp = regnext(scan);
14277 SV * const mysv=sv_newmortal();
14278 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
14279 regprop(RExC_rx, mysv, scan);
14280 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
14281 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
14282 (temp == NULL ? "->" : ""),
14283 (temp == NULL ? PL_reg_name[OP(val)] : "")
14291 if (reg_off_by_arg[OP(scan)]) {
14292 ARG_SET(scan, val - scan);
14295 NEXT_OFF(scan) = val - scan;
14301 - regtail_study - set the next-pointer at the end of a node chain of p to val.
14302 - Look for optimizable sequences at the same time.
14303 - currently only looks for EXACT chains.
14305 This is experimental code. The idea is to use this routine to perform
14306 in place optimizations on branches and groups as they are constructed,
14307 with the long term intention of removing optimization from study_chunk so
14308 that it is purely analytical.
14310 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
14311 to control which is which.
14314 /* TODO: All four parms should be const */
14317 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14322 #ifdef EXPERIMENTAL_INPLACESCAN
14325 GET_RE_DEBUG_FLAGS_DECL;
14327 PERL_ARGS_ASSERT_REGTAIL_STUDY;
14333 /* Find last node. */
14337 regnode * const temp = regnext(scan);
14338 #ifdef EXPERIMENTAL_INPLACESCAN
14339 if (PL_regkind[OP(scan)] == EXACT) {
14340 bool has_exactf_sharp_s; /* Unexamined in this routine */
14341 if (join_exact(pRExC_state,scan,&min, &has_exactf_sharp_s, 1,val,depth+1))
14346 switch (OP(scan)) {
14352 case EXACTFU_TRICKYFOLD:
14354 if( exact == PSEUDO )
14356 else if ( exact != OP(scan) )
14365 SV * const mysv=sv_newmortal();
14366 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
14367 regprop(RExC_rx, mysv, scan);
14368 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
14369 SvPV_nolen_const(mysv),
14370 REG_NODE_NUM(scan),
14371 PL_reg_name[exact]);
14378 SV * const mysv_val=sv_newmortal();
14379 DEBUG_PARSE_MSG("");
14380 regprop(RExC_rx, mysv_val, val);
14381 PerlIO_printf(Perl_debug_log, "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
14382 SvPV_nolen_const(mysv_val),
14383 (IV)REG_NODE_NUM(val),
14387 if (reg_off_by_arg[OP(scan)]) {
14388 ARG_SET(scan, val - scan);
14391 NEXT_OFF(scan) = val - scan;
14399 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
14404 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
14409 for (bit=0; bit<32; bit++) {
14410 if (flags & (1<<bit)) {
14411 if (!set++ && lead)
14412 PerlIO_printf(Perl_debug_log, "%s",lead);
14413 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_intflags_name[bit]);
14418 PerlIO_printf(Perl_debug_log, "\n");
14420 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14425 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
14431 for (bit=0; bit<32; bit++) {
14432 if (flags & (1<<bit)) {
14433 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
14436 if (!set++ && lead)
14437 PerlIO_printf(Perl_debug_log, "%s",lead);
14438 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
14441 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
14442 if (!set++ && lead) {
14443 PerlIO_printf(Perl_debug_log, "%s",lead);
14446 case REGEX_UNICODE_CHARSET:
14447 PerlIO_printf(Perl_debug_log, "UNICODE");
14449 case REGEX_LOCALE_CHARSET:
14450 PerlIO_printf(Perl_debug_log, "LOCALE");
14452 case REGEX_ASCII_RESTRICTED_CHARSET:
14453 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
14455 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
14456 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
14459 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
14465 PerlIO_printf(Perl_debug_log, "\n");
14467 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14473 Perl_regdump(pTHX_ const regexp *r)
14477 SV * const sv = sv_newmortal();
14478 SV *dsv= sv_newmortal();
14479 RXi_GET_DECL(r,ri);
14480 GET_RE_DEBUG_FLAGS_DECL;
14482 PERL_ARGS_ASSERT_REGDUMP;
14484 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
14486 /* Header fields of interest. */
14487 if (r->anchored_substr) {
14488 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
14489 RE_SV_DUMPLEN(r->anchored_substr), 30);
14490 PerlIO_printf(Perl_debug_log,
14491 "anchored %s%s at %"IVdf" ",
14492 s, RE_SV_TAIL(r->anchored_substr),
14493 (IV)r->anchored_offset);
14494 } else if (r->anchored_utf8) {
14495 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
14496 RE_SV_DUMPLEN(r->anchored_utf8), 30);
14497 PerlIO_printf(Perl_debug_log,
14498 "anchored utf8 %s%s at %"IVdf" ",
14499 s, RE_SV_TAIL(r->anchored_utf8),
14500 (IV)r->anchored_offset);
14502 if (r->float_substr) {
14503 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
14504 RE_SV_DUMPLEN(r->float_substr), 30);
14505 PerlIO_printf(Perl_debug_log,
14506 "floating %s%s at %"IVdf"..%"UVuf" ",
14507 s, RE_SV_TAIL(r->float_substr),
14508 (IV)r->float_min_offset, (UV)r->float_max_offset);
14509 } else if (r->float_utf8) {
14510 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
14511 RE_SV_DUMPLEN(r->float_utf8), 30);
14512 PerlIO_printf(Perl_debug_log,
14513 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
14514 s, RE_SV_TAIL(r->float_utf8),
14515 (IV)r->float_min_offset, (UV)r->float_max_offset);
14517 if (r->check_substr || r->check_utf8)
14518 PerlIO_printf(Perl_debug_log,
14520 (r->check_substr == r->float_substr
14521 && r->check_utf8 == r->float_utf8
14522 ? "(checking floating" : "(checking anchored"));
14523 if (r->extflags & RXf_NOSCAN)
14524 PerlIO_printf(Perl_debug_log, " noscan");
14525 if (r->extflags & RXf_CHECK_ALL)
14526 PerlIO_printf(Perl_debug_log, " isall");
14527 if (r->check_substr || r->check_utf8)
14528 PerlIO_printf(Perl_debug_log, ") ");
14530 if (ri->regstclass) {
14531 regprop(r, sv, ri->regstclass);
14532 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
14534 if (r->extflags & RXf_ANCH) {
14535 PerlIO_printf(Perl_debug_log, "anchored");
14536 if (r->extflags & RXf_ANCH_BOL)
14537 PerlIO_printf(Perl_debug_log, "(BOL)");
14538 if (r->extflags & RXf_ANCH_MBOL)
14539 PerlIO_printf(Perl_debug_log, "(MBOL)");
14540 if (r->extflags & RXf_ANCH_SBOL)
14541 PerlIO_printf(Perl_debug_log, "(SBOL)");
14542 if (r->extflags & RXf_ANCH_GPOS)
14543 PerlIO_printf(Perl_debug_log, "(GPOS)");
14544 PerlIO_putc(Perl_debug_log, ' ');
14546 if (r->extflags & RXf_GPOS_SEEN)
14547 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
14548 if (r->intflags & PREGf_SKIP)
14549 PerlIO_printf(Perl_debug_log, "plus ");
14550 if (r->intflags & PREGf_IMPLICIT)
14551 PerlIO_printf(Perl_debug_log, "implicit ");
14552 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
14553 if (r->extflags & RXf_EVAL_SEEN)
14554 PerlIO_printf(Perl_debug_log, "with eval ");
14555 PerlIO_printf(Perl_debug_log, "\n");
14557 regdump_extflags("r->extflags: ",r->extflags);
14558 regdump_intflags("r->intflags: ",r->intflags);
14561 PERL_ARGS_ASSERT_REGDUMP;
14562 PERL_UNUSED_CONTEXT;
14563 PERL_UNUSED_ARG(r);
14564 #endif /* DEBUGGING */
14568 - regprop - printable representation of opcode
14570 #define EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags) \
14573 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]); \
14574 if (flags & ANYOF_INVERT) \
14575 /*make sure the invert info is in each */ \
14576 sv_catpvs(sv, "^"); \
14582 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
14588 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
14589 static const char * const anyofs[] = {
14590 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
14591 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
14592 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
14593 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
14594 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
14595 || _CC_VERTSPACE != 16
14596 #error Need to adjust order of anyofs[]
14633 RXi_GET_DECL(prog,progi);
14634 GET_RE_DEBUG_FLAGS_DECL;
14636 PERL_ARGS_ASSERT_REGPROP;
14640 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
14641 /* It would be nice to FAIL() here, but this may be called from
14642 regexec.c, and it would be hard to supply pRExC_state. */
14643 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(o), (int)REGNODE_MAX);
14644 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
14646 k = PL_regkind[OP(o)];
14649 sv_catpvs(sv, " ");
14650 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
14651 * is a crude hack but it may be the best for now since
14652 * we have no flag "this EXACTish node was UTF-8"
14654 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
14655 PERL_PV_ESCAPE_UNI_DETECT |
14656 PERL_PV_ESCAPE_NONASCII |
14657 PERL_PV_PRETTY_ELLIPSES |
14658 PERL_PV_PRETTY_LTGT |
14659 PERL_PV_PRETTY_NOCLEAR
14661 } else if (k == TRIE) {
14662 /* print the details of the trie in dumpuntil instead, as
14663 * progi->data isn't available here */
14664 const char op = OP(o);
14665 const U32 n = ARG(o);
14666 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
14667 (reg_ac_data *)progi->data->data[n] :
14669 const reg_trie_data * const trie
14670 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
14672 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
14673 DEBUG_TRIE_COMPILE_r(
14674 Perl_sv_catpvf(aTHX_ sv,
14675 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
14676 (UV)trie->startstate,
14677 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
14678 (UV)trie->wordcount,
14681 (UV)TRIE_CHARCOUNT(trie),
14682 (UV)trie->uniquecharcount
14685 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
14687 int rangestart = -1;
14688 U8* bitmap = IS_ANYOF_TRIE(op) ? (U8*)ANYOF_BITMAP(o) : (U8*)TRIE_BITMAP(trie);
14689 sv_catpvs(sv, "[");
14690 for (i = 0; i <= 256; i++) {
14691 if (i < 256 && BITMAP_TEST(bitmap,i)) {
14692 if (rangestart == -1)
14694 } else if (rangestart != -1) {
14695 if (i <= rangestart + 3)
14696 for (; rangestart < i; rangestart++)
14697 put_byte(sv, rangestart);
14699 put_byte(sv, rangestart);
14700 sv_catpvs(sv, "-");
14701 put_byte(sv, i - 1);
14706 sv_catpvs(sv, "]");
14709 } else if (k == CURLY) {
14710 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
14711 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
14712 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
14714 else if (k == WHILEM && o->flags) /* Ordinal/of */
14715 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
14716 else if (k == REF || k == OPEN || k == CLOSE || k == GROUPP || OP(o)==ACCEPT) {
14717 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
14718 if ( RXp_PAREN_NAMES(prog) ) {
14719 if ( k != REF || (OP(o) < NREF)) {
14720 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
14721 SV **name= av_fetch(list, ARG(o), 0 );
14723 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14726 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
14727 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
14728 I32 *nums=(I32*)SvPVX(sv_dat);
14729 SV **name= av_fetch(list, nums[0], 0 );
14732 for ( n=0; n<SvIVX(sv_dat); n++ ) {
14733 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
14734 (n ? "," : ""), (IV)nums[n]);
14736 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14740 } else if (k == GOSUB)
14741 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o)); /* Paren and offset */
14742 else if (k == VERB) {
14744 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
14745 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
14746 } else if (k == LOGICAL)
14747 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* 2: embedded, otherwise 1 */
14748 else if (k == ANYOF) {
14749 int i, rangestart = -1;
14750 const U8 flags = ANYOF_FLAGS(o);
14754 if (flags & ANYOF_LOCALE)
14755 sv_catpvs(sv, "{loc}");
14756 if (flags & ANYOF_LOC_FOLD)
14757 sv_catpvs(sv, "{i}");
14758 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
14759 if (flags & ANYOF_INVERT)
14760 sv_catpvs(sv, "^");
14762 /* output what the standard cp 0-255 bitmap matches */
14763 for (i = 0; i <= 256; i++) {
14764 if (i < 256 && ANYOF_BITMAP_TEST(o,i)) {
14765 if (rangestart == -1)
14767 } else if (rangestart != -1) {
14768 if (i <= rangestart + 3)
14769 for (; rangestart < i; rangestart++)
14770 put_byte(sv, rangestart);
14772 put_byte(sv, rangestart);
14773 sv_catpvs(sv, "-");
14774 put_byte(sv, i - 1);
14781 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14782 /* output any special charclass tests (used entirely under use locale) */
14783 if (ANYOF_CLASS_TEST_ANY_SET(o))
14784 for (i = 0; i < (int)(sizeof(anyofs)/sizeof(char*)); i++)
14785 if (ANYOF_CLASS_TEST(o,i)) {
14786 sv_catpv(sv, anyofs[i]);
14790 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14792 if (flags & ANYOF_NON_UTF8_LATIN1_ALL) {
14793 sv_catpvs(sv, "{non-utf8-latin1-all}");
14796 /* output information about the unicode matching */
14797 if (flags & ANYOF_UNICODE_ALL)
14798 sv_catpvs(sv, "{unicode_all}");
14799 else if (ANYOF_NONBITMAP(o))
14800 sv_catpvs(sv, "{unicode}");
14801 if (flags & ANYOF_NONBITMAP_NON_UTF8)
14802 sv_catpvs(sv, "{outside bitmap}");
14804 if (ANYOF_NONBITMAP(o)) {
14805 SV *lv; /* Set if there is something outside the bit map */
14806 SV * const sw = regclass_swash(prog, o, FALSE, &lv, NULL);
14807 bool byte_output = FALSE; /* If something in the bitmap has been
14810 if (lv && lv != &PL_sv_undef) {
14812 U8 s[UTF8_MAXBYTES_CASE+1];
14814 for (i = 0; i <= 256; i++) { /* Look at chars in bitmap */
14815 uvchr_to_utf8(s, i);
14818 && ! ANYOF_BITMAP_TEST(o, i) /* Don't duplicate
14822 && swash_fetch(sw, s, TRUE))
14824 if (rangestart == -1)
14826 } else if (rangestart != -1) {
14827 byte_output = TRUE;
14828 if (i <= rangestart + 3)
14829 for (; rangestart < i; rangestart++) {
14830 put_byte(sv, rangestart);
14833 put_byte(sv, rangestart);
14834 sv_catpvs(sv, "-");
14843 char *s = savesvpv(lv);
14844 char * const origs = s;
14846 while (*s && *s != '\n')
14850 const char * const t = ++s;
14853 sv_catpvs(sv, " ");
14859 /* Truncate very long output */
14860 if (s - origs > 256) {
14861 Perl_sv_catpvf(aTHX_ sv,
14863 (int) (s - origs - 1),
14869 else if (*s == '\t') {
14884 SvREFCNT_dec_NN(lv);
14888 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
14890 else if (k == POSIXD || k == NPOSIXD) {
14891 U8 index = FLAGS(o) * 2;
14892 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
14893 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
14896 sv_catpv(sv, anyofs[index]);
14899 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
14900 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
14902 PERL_UNUSED_CONTEXT;
14903 PERL_UNUSED_ARG(sv);
14904 PERL_UNUSED_ARG(o);
14905 PERL_UNUSED_ARG(prog);
14906 #endif /* DEBUGGING */
14910 Perl_re_intuit_string(pTHX_ REGEXP * const r)
14911 { /* Assume that RE_INTUIT is set */
14913 struct regexp *const prog = ReANY(r);
14914 GET_RE_DEBUG_FLAGS_DECL;
14916 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
14917 PERL_UNUSED_CONTEXT;
14921 const char * const s = SvPV_nolen_const(prog->check_substr
14922 ? prog->check_substr : prog->check_utf8);
14924 if (!PL_colorset) reginitcolors();
14925 PerlIO_printf(Perl_debug_log,
14926 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
14928 prog->check_substr ? "" : "utf8 ",
14929 PL_colors[5],PL_colors[0],
14932 (strlen(s) > 60 ? "..." : ""));
14935 return prog->check_substr ? prog->check_substr : prog->check_utf8;
14941 handles refcounting and freeing the perl core regexp structure. When
14942 it is necessary to actually free the structure the first thing it
14943 does is call the 'free' method of the regexp_engine associated to
14944 the regexp, allowing the handling of the void *pprivate; member
14945 first. (This routine is not overridable by extensions, which is why
14946 the extensions free is called first.)
14948 See regdupe and regdupe_internal if you change anything here.
14950 #ifndef PERL_IN_XSUB_RE
14952 Perl_pregfree(pTHX_ REGEXP *r)
14958 Perl_pregfree2(pTHX_ REGEXP *rx)
14961 struct regexp *const r = ReANY(rx);
14962 GET_RE_DEBUG_FLAGS_DECL;
14964 PERL_ARGS_ASSERT_PREGFREE2;
14966 if (r->mother_re) {
14967 ReREFCNT_dec(r->mother_re);
14969 CALLREGFREE_PVT(rx); /* free the private data */
14970 SvREFCNT_dec(RXp_PAREN_NAMES(r));
14971 Safefree(r->xpv_len_u.xpvlenu_pv);
14974 SvREFCNT_dec(r->anchored_substr);
14975 SvREFCNT_dec(r->anchored_utf8);
14976 SvREFCNT_dec(r->float_substr);
14977 SvREFCNT_dec(r->float_utf8);
14978 Safefree(r->substrs);
14980 RX_MATCH_COPY_FREE(rx);
14981 #ifdef PERL_ANY_COW
14982 SvREFCNT_dec(r->saved_copy);
14985 SvREFCNT_dec(r->qr_anoncv);
14986 rx->sv_u.svu_rx = 0;
14991 This is a hacky workaround to the structural issue of match results
14992 being stored in the regexp structure which is in turn stored in
14993 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
14994 could be PL_curpm in multiple contexts, and could require multiple
14995 result sets being associated with the pattern simultaneously, such
14996 as when doing a recursive match with (??{$qr})
14998 The solution is to make a lightweight copy of the regexp structure
14999 when a qr// is returned from the code executed by (??{$qr}) this
15000 lightweight copy doesn't actually own any of its data except for
15001 the starp/end and the actual regexp structure itself.
15007 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
15009 struct regexp *ret;
15010 struct regexp *const r = ReANY(rx);
15011 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
15013 PERL_ARGS_ASSERT_REG_TEMP_COPY;
15016 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
15018 SvOK_off((SV *)ret_x);
15020 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
15021 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
15022 made both spots point to the same regexp body.) */
15023 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
15024 assert(!SvPVX(ret_x));
15025 ret_x->sv_u.svu_rx = temp->sv_any;
15026 temp->sv_any = NULL;
15027 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
15028 SvREFCNT_dec_NN(temp);
15029 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
15030 ing below will not set it. */
15031 SvCUR_set(ret_x, SvCUR(rx));
15034 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
15035 sv_force_normal(sv) is called. */
15037 ret = ReANY(ret_x);
15039 SvFLAGS(ret_x) |= SvUTF8(rx);
15040 /* We share the same string buffer as the original regexp, on which we
15041 hold a reference count, incremented when mother_re is set below.
15042 The string pointer is copied here, being part of the regexp struct.
15044 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
15045 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
15047 const I32 npar = r->nparens+1;
15048 Newx(ret->offs, npar, regexp_paren_pair);
15049 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15052 Newx(ret->substrs, 1, struct reg_substr_data);
15053 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15055 SvREFCNT_inc_void(ret->anchored_substr);
15056 SvREFCNT_inc_void(ret->anchored_utf8);
15057 SvREFCNT_inc_void(ret->float_substr);
15058 SvREFCNT_inc_void(ret->float_utf8);
15060 /* check_substr and check_utf8, if non-NULL, point to either their
15061 anchored or float namesakes, and don't hold a second reference. */
15063 RX_MATCH_COPIED_off(ret_x);
15064 #ifdef PERL_ANY_COW
15065 ret->saved_copy = NULL;
15067 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
15068 SvREFCNT_inc_void(ret->qr_anoncv);
15074 /* regfree_internal()
15076 Free the private data in a regexp. This is overloadable by
15077 extensions. Perl takes care of the regexp structure in pregfree(),
15078 this covers the *pprivate pointer which technically perl doesn't
15079 know about, however of course we have to handle the
15080 regexp_internal structure when no extension is in use.
15082 Note this is called before freeing anything in the regexp
15087 Perl_regfree_internal(pTHX_ REGEXP * const rx)
15090 struct regexp *const r = ReANY(rx);
15091 RXi_GET_DECL(r,ri);
15092 GET_RE_DEBUG_FLAGS_DECL;
15094 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
15100 SV *dsv= sv_newmortal();
15101 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
15102 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
15103 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
15104 PL_colors[4],PL_colors[5],s);
15107 #ifdef RE_TRACK_PATTERN_OFFSETS
15109 Safefree(ri->u.offsets); /* 20010421 MJD */
15111 if (ri->code_blocks) {
15113 for (n = 0; n < ri->num_code_blocks; n++)
15114 SvREFCNT_dec(ri->code_blocks[n].src_regex);
15115 Safefree(ri->code_blocks);
15119 int n = ri->data->count;
15122 /* If you add a ->what type here, update the comment in regcomp.h */
15123 switch (ri->data->what[n]) {
15129 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
15132 Safefree(ri->data->data[n]);
15138 { /* Aho Corasick add-on structure for a trie node.
15139 Used in stclass optimization only */
15141 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
15143 refcount = --aho->refcount;
15146 PerlMemShared_free(aho->states);
15147 PerlMemShared_free(aho->fail);
15148 /* do this last!!!! */
15149 PerlMemShared_free(ri->data->data[n]);
15150 PerlMemShared_free(ri->regstclass);
15156 /* trie structure. */
15158 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
15160 refcount = --trie->refcount;
15163 PerlMemShared_free(trie->charmap);
15164 PerlMemShared_free(trie->states);
15165 PerlMemShared_free(trie->trans);
15167 PerlMemShared_free(trie->bitmap);
15169 PerlMemShared_free(trie->jump);
15170 PerlMemShared_free(trie->wordinfo);
15171 /* do this last!!!! */
15172 PerlMemShared_free(ri->data->data[n]);
15177 Perl_croak(aTHX_ "panic: regfree data code '%c'", ri->data->what[n]);
15180 Safefree(ri->data->what);
15181 Safefree(ri->data);
15187 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
15188 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
15189 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
15192 re_dup - duplicate a regexp.
15194 This routine is expected to clone a given regexp structure. It is only
15195 compiled under USE_ITHREADS.
15197 After all of the core data stored in struct regexp is duplicated
15198 the regexp_engine.dupe method is used to copy any private data
15199 stored in the *pprivate pointer. This allows extensions to handle
15200 any duplication it needs to do.
15202 See pregfree() and regfree_internal() if you change anything here.
15204 #if defined(USE_ITHREADS)
15205 #ifndef PERL_IN_XSUB_RE
15207 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
15211 const struct regexp *r = ReANY(sstr);
15212 struct regexp *ret = ReANY(dstr);
15214 PERL_ARGS_ASSERT_RE_DUP_GUTS;
15216 npar = r->nparens+1;
15217 Newx(ret->offs, npar, regexp_paren_pair);
15218 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15220 if (ret->substrs) {
15221 /* Do it this way to avoid reading from *r after the StructCopy().
15222 That way, if any of the sv_dup_inc()s dislodge *r from the L1
15223 cache, it doesn't matter. */
15224 const bool anchored = r->check_substr
15225 ? r->check_substr == r->anchored_substr
15226 : r->check_utf8 == r->anchored_utf8;
15227 Newx(ret->substrs, 1, struct reg_substr_data);
15228 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15230 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
15231 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
15232 ret->float_substr = sv_dup_inc(ret->float_substr, param);
15233 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
15235 /* check_substr and check_utf8, if non-NULL, point to either their
15236 anchored or float namesakes, and don't hold a second reference. */
15238 if (ret->check_substr) {
15240 assert(r->check_utf8 == r->anchored_utf8);
15241 ret->check_substr = ret->anchored_substr;
15242 ret->check_utf8 = ret->anchored_utf8;
15244 assert(r->check_substr == r->float_substr);
15245 assert(r->check_utf8 == r->float_utf8);
15246 ret->check_substr = ret->float_substr;
15247 ret->check_utf8 = ret->float_utf8;
15249 } else if (ret->check_utf8) {
15251 ret->check_utf8 = ret->anchored_utf8;
15253 ret->check_utf8 = ret->float_utf8;
15258 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
15259 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
15262 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
15264 if (RX_MATCH_COPIED(dstr))
15265 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
15267 ret->subbeg = NULL;
15268 #ifdef PERL_ANY_COW
15269 ret->saved_copy = NULL;
15272 /* Whether mother_re be set or no, we need to copy the string. We
15273 cannot refrain from copying it when the storage points directly to
15274 our mother regexp, because that's
15275 1: a buffer in a different thread
15276 2: something we no longer hold a reference on
15277 so we need to copy it locally. */
15278 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
15279 ret->mother_re = NULL;
15282 #endif /* PERL_IN_XSUB_RE */
15287 This is the internal complement to regdupe() which is used to copy
15288 the structure pointed to by the *pprivate pointer in the regexp.
15289 This is the core version of the extension overridable cloning hook.
15290 The regexp structure being duplicated will be copied by perl prior
15291 to this and will be provided as the regexp *r argument, however
15292 with the /old/ structures pprivate pointer value. Thus this routine
15293 may override any copying normally done by perl.
15295 It returns a pointer to the new regexp_internal structure.
15299 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
15302 struct regexp *const r = ReANY(rx);
15303 regexp_internal *reti;
15305 RXi_GET_DECL(r,ri);
15307 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
15311 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode), char, regexp_internal);
15312 Copy(ri->program, reti->program, len+1, regnode);
15314 reti->num_code_blocks = ri->num_code_blocks;
15315 if (ri->code_blocks) {
15317 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
15318 struct reg_code_block);
15319 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
15320 struct reg_code_block);
15321 for (n = 0; n < ri->num_code_blocks; n++)
15322 reti->code_blocks[n].src_regex = (REGEXP*)
15323 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
15326 reti->code_blocks = NULL;
15328 reti->regstclass = NULL;
15331 struct reg_data *d;
15332 const int count = ri->data->count;
15335 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
15336 char, struct reg_data);
15337 Newx(d->what, count, U8);
15340 for (i = 0; i < count; i++) {
15341 d->what[i] = ri->data->what[i];
15342 switch (d->what[i]) {
15343 /* see also regcomp.h and regfree_internal() */
15344 case 'a': /* actually an AV, but the dup function is identical. */
15348 case 'u': /* actually an HV, but the dup function is identical. */
15349 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
15352 /* This is cheating. */
15353 Newx(d->data[i], 1, struct regnode_charclass_class);
15354 StructCopy(ri->data->data[i], d->data[i],
15355 struct regnode_charclass_class);
15356 reti->regstclass = (regnode*)d->data[i];
15359 /* Trie stclasses are readonly and can thus be shared
15360 * without duplication. We free the stclass in pregfree
15361 * when the corresponding reg_ac_data struct is freed.
15363 reti->regstclass= ri->regstclass;
15367 ((reg_trie_data*)ri->data->data[i])->refcount++;
15372 d->data[i] = ri->data->data[i];
15375 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'", ri->data->what[i]);
15384 reti->name_list_idx = ri->name_list_idx;
15386 #ifdef RE_TRACK_PATTERN_OFFSETS
15387 if (ri->u.offsets) {
15388 Newx(reti->u.offsets, 2*len+1, U32);
15389 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
15392 SetProgLen(reti,len);
15395 return (void*)reti;
15398 #endif /* USE_ITHREADS */
15400 #ifndef PERL_IN_XSUB_RE
15403 - regnext - dig the "next" pointer out of a node
15406 Perl_regnext(pTHX_ regnode *p)
15414 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
15415 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(p), (int)REGNODE_MAX);
15418 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
15427 S_re_croak2(pTHX_ const char* pat1,const char* pat2,...)
15430 STRLEN l1 = strlen(pat1);
15431 STRLEN l2 = strlen(pat2);
15434 const char *message;
15436 PERL_ARGS_ASSERT_RE_CROAK2;
15442 Copy(pat1, buf, l1 , char);
15443 Copy(pat2, buf + l1, l2 , char);
15444 buf[l1 + l2] = '\n';
15445 buf[l1 + l2 + 1] = '\0';
15447 /* ANSI variant takes additional second argument */
15448 va_start(args, pat2);
15452 msv = vmess(buf, &args);
15454 message = SvPV_const(msv,l1);
15457 Copy(message, buf, l1 , char);
15458 buf[l1-1] = '\0'; /* Overwrite \n */
15459 Perl_croak(aTHX_ "%s", buf);
15462 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
15464 #ifndef PERL_IN_XSUB_RE
15466 Perl_save_re_context(pTHX)
15470 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
15472 const REGEXP * const rx = PM_GETRE(PL_curpm);
15475 for (i = 1; i <= RX_NPARENS(rx); i++) {
15476 char digits[TYPE_CHARS(long)];
15477 const STRLEN len = my_snprintf(digits, sizeof(digits), "%lu", (long)i);
15478 GV *const *const gvp
15479 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
15482 GV * const gv = *gvp;
15483 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
15495 S_put_byte(pTHX_ SV *sv, int c)
15497 PERL_ARGS_ASSERT_PUT_BYTE;
15499 /* Our definition of isPRINT() ignores locales, so only bytes that are
15500 not part of UTF-8 are considered printable. I assume that the same
15501 holds for UTF-EBCDIC.
15502 Also, code point 255 is not printable in either (it's E0 in EBCDIC,
15503 which Wikipedia says:
15505 EO, or Eight Ones, is an 8-bit EBCDIC character code represented as all
15506 ones (binary 1111 1111, hexadecimal FF). It is similar, but not
15507 identical, to the ASCII delete (DEL) or rubout control character. ...
15508 it is typically mapped to hexadecimal code 9F, in order to provide a
15509 unique character mapping in both directions)
15511 So the old condition can be simplified to !isPRINT(c) */
15514 Perl_sv_catpvf(aTHX_ sv, "\\x%02x", c);
15517 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
15521 const char string = c;
15522 if (c == '-' || c == ']' || c == '\\' || c == '^')
15523 sv_catpvs(sv, "\\");
15524 sv_catpvn(sv, &string, 1);
15529 #define CLEAR_OPTSTART \
15530 if (optstart) STMT_START { \
15531 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
15535 #define DUMPUNTIL(b,e) CLEAR_OPTSTART; node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
15537 STATIC const regnode *
15538 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
15539 const regnode *last, const regnode *plast,
15540 SV* sv, I32 indent, U32 depth)
15543 U8 op = PSEUDO; /* Arbitrary non-END op. */
15544 const regnode *next;
15545 const regnode *optstart= NULL;
15547 RXi_GET_DECL(r,ri);
15548 GET_RE_DEBUG_FLAGS_DECL;
15550 PERL_ARGS_ASSERT_DUMPUNTIL;
15552 #ifdef DEBUG_DUMPUNTIL
15553 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
15554 last ? last-start : 0,plast ? plast-start : 0);
15557 if (plast && plast < last)
15560 while (PL_regkind[op] != END && (!last || node < last)) {
15561 /* While that wasn't END last time... */
15564 if (op == CLOSE || op == WHILEM)
15566 next = regnext((regnode *)node);
15569 if (OP(node) == OPTIMIZED) {
15570 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
15577 regprop(r, sv, node);
15578 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
15579 (int)(2*indent + 1), "", SvPVX_const(sv));
15581 if (OP(node) != OPTIMIZED) {
15582 if (next == NULL) /* Next ptr. */
15583 PerlIO_printf(Perl_debug_log, " (0)");
15584 else if (PL_regkind[(U8)op] == BRANCH && PL_regkind[OP(next)] != BRANCH )
15585 PerlIO_printf(Perl_debug_log, " (FAIL)");
15587 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
15588 (void)PerlIO_putc(Perl_debug_log, '\n');
15592 if (PL_regkind[(U8)op] == BRANCHJ) {
15595 const regnode *nnode = (OP(next) == LONGJMP
15596 ? regnext((regnode *)next)
15598 if (last && nnode > last)
15600 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
15603 else if (PL_regkind[(U8)op] == BRANCH) {
15605 DUMPUNTIL(NEXTOPER(node), next);
15607 else if ( PL_regkind[(U8)op] == TRIE ) {
15608 const regnode *this_trie = node;
15609 const char op = OP(node);
15610 const U32 n = ARG(node);
15611 const reg_ac_data * const ac = op>=AHOCORASICK ?
15612 (reg_ac_data *)ri->data->data[n] :
15614 const reg_trie_data * const trie =
15615 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
15617 AV *const trie_words = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
15619 const regnode *nextbranch= NULL;
15622 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
15623 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
15625 PerlIO_printf(Perl_debug_log, "%*s%s ",
15626 (int)(2*(indent+3)), "",
15627 elem_ptr ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr), SvCUR(*elem_ptr), 60,
15628 PL_colors[0], PL_colors[1],
15629 (SvUTF8(*elem_ptr) ? PERL_PV_ESCAPE_UNI : 0) |
15630 PERL_PV_PRETTY_ELLIPSES |
15631 PERL_PV_PRETTY_LTGT
15636 U16 dist= trie->jump[word_idx+1];
15637 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
15638 (UV)((dist ? this_trie + dist : next) - start));
15641 nextbranch= this_trie + trie->jump[0];
15642 DUMPUNTIL(this_trie + dist, nextbranch);
15644 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
15645 nextbranch= regnext((regnode *)nextbranch);
15647 PerlIO_printf(Perl_debug_log, "\n");
15650 if (last && next > last)
15655 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
15656 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
15657 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
15659 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
15661 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
15663 else if ( op == PLUS || op == STAR) {
15664 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
15666 else if (PL_regkind[(U8)op] == ANYOF) {
15667 /* arglen 1 + class block */
15668 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_CLASS)
15669 ? ANYOF_CLASS_SKIP : ANYOF_SKIP);
15670 node = NEXTOPER(node);
15672 else if (PL_regkind[(U8)op] == EXACT) {
15673 /* Literal string, where present. */
15674 node += NODE_SZ_STR(node) - 1;
15675 node = NEXTOPER(node);
15678 node = NEXTOPER(node);
15679 node += regarglen[(U8)op];
15681 if (op == CURLYX || op == OPEN)
15685 #ifdef DEBUG_DUMPUNTIL
15686 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
15691 #endif /* DEBUGGING */
15695 * c-indentation-style: bsd
15696 * c-basic-offset: 4
15697 * indent-tabs-mode: nil
15700 * ex: set ts=8 sts=4 sw=4 et: