5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
84 extern const struct regexp_engine my_reg_engine;
89 #include "dquote_static.c"
90 #include "charclass_invlists.h"
91 #include "inline_invlist.c"
92 #include "unicode_constants.h"
94 #define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
96 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
103 # if defined(BUGGY_MSC6)
104 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
105 # pragma optimize("a",off)
106 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
107 # pragma optimize("w",on )
108 # endif /* BUGGY_MSC6 */
112 #define STATIC static
116 typedef struct RExC_state_t {
117 U32 flags; /* RXf_* are we folding, multilining? */
118 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
119 char *precomp; /* uncompiled string. */
120 REGEXP *rx_sv; /* The SV that is the regexp. */
121 regexp *rx; /* perl core regexp structure */
122 regexp_internal *rxi; /* internal data for regexp object pprivate field */
123 char *start; /* Start of input for compile */
124 char *end; /* End of input for compile */
125 char *parse; /* Input-scan pointer. */
126 I32 whilem_seen; /* number of WHILEM in this expr */
127 regnode *emit_start; /* Start of emitted-code area */
128 regnode *emit_bound; /* First regnode outside of the allocated space */
129 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
130 I32 naughty; /* How bad is this pattern? */
131 I32 sawback; /* Did we see \1, ...? */
133 I32 size; /* Code size. */
134 I32 npar; /* Capture buffer count, (OPEN). */
135 I32 cpar; /* Capture buffer count, (CLOSE). */
136 I32 nestroot; /* root parens we are in - used by accept */
139 regnode **open_parens; /* pointers to open parens */
140 regnode **close_parens; /* pointers to close parens */
141 regnode *opend; /* END node in program */
142 I32 utf8; /* whether the pattern is utf8 or not */
143 I32 orig_utf8; /* whether the pattern was originally in utf8 */
144 /* XXX use this for future optimisation of case
145 * where pattern must be upgraded to utf8. */
146 I32 uni_semantics; /* If a d charset modifier should use unicode
147 rules, even if the pattern is not in
149 HV *paren_names; /* Paren names */
151 regnode **recurse; /* Recurse regops */
152 I32 recurse_count; /* Number of recurse regops */
155 I32 override_recoding;
156 I32 in_multi_char_class;
157 struct reg_code_block *code_blocks; /* positions of literal (?{})
159 int num_code_blocks; /* size of code_blocks[] */
160 int code_index; /* next code_blocks[] slot */
162 char *starttry; /* -Dr: where regtry was called. */
163 #define RExC_starttry (pRExC_state->starttry)
165 SV *runtime_code_qr; /* qr with the runtime code blocks */
167 const char *lastparse;
169 AV *paren_name_list; /* idx -> name */
170 #define RExC_lastparse (pRExC_state->lastparse)
171 #define RExC_lastnum (pRExC_state->lastnum)
172 #define RExC_paren_name_list (pRExC_state->paren_name_list)
176 #define RExC_flags (pRExC_state->flags)
177 #define RExC_pm_flags (pRExC_state->pm_flags)
178 #define RExC_precomp (pRExC_state->precomp)
179 #define RExC_rx_sv (pRExC_state->rx_sv)
180 #define RExC_rx (pRExC_state->rx)
181 #define RExC_rxi (pRExC_state->rxi)
182 #define RExC_start (pRExC_state->start)
183 #define RExC_end (pRExC_state->end)
184 #define RExC_parse (pRExC_state->parse)
185 #define RExC_whilem_seen (pRExC_state->whilem_seen)
186 #ifdef RE_TRACK_PATTERN_OFFSETS
187 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
189 #define RExC_emit (pRExC_state->emit)
190 #define RExC_emit_start (pRExC_state->emit_start)
191 #define RExC_emit_bound (pRExC_state->emit_bound)
192 #define RExC_naughty (pRExC_state->naughty)
193 #define RExC_sawback (pRExC_state->sawback)
194 #define RExC_seen (pRExC_state->seen)
195 #define RExC_size (pRExC_state->size)
196 #define RExC_npar (pRExC_state->npar)
197 #define RExC_nestroot (pRExC_state->nestroot)
198 #define RExC_extralen (pRExC_state->extralen)
199 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
200 #define RExC_utf8 (pRExC_state->utf8)
201 #define RExC_uni_semantics (pRExC_state->uni_semantics)
202 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
203 #define RExC_open_parens (pRExC_state->open_parens)
204 #define RExC_close_parens (pRExC_state->close_parens)
205 #define RExC_opend (pRExC_state->opend)
206 #define RExC_paren_names (pRExC_state->paren_names)
207 #define RExC_recurse (pRExC_state->recurse)
208 #define RExC_recurse_count (pRExC_state->recurse_count)
209 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
210 #define RExC_contains_locale (pRExC_state->contains_locale)
211 #define RExC_override_recoding (pRExC_state->override_recoding)
212 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
215 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
216 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
217 ((*s) == '{' && regcurly(s, FALSE)))
220 #undef SPSTART /* dratted cpp namespace... */
223 * Flags to be passed up and down.
225 #define WORST 0 /* Worst case. */
226 #define HASWIDTH 0x01 /* Known to match non-null strings. */
228 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
229 * character. (There needs to be a case: in the switch statement in regexec.c
230 * for any node marked SIMPLE.) Note that this is not the same thing as
233 #define SPSTART 0x04 /* Starts with * or + */
234 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
235 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
236 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
238 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
240 /* whether trie related optimizations are enabled */
241 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
242 #define TRIE_STUDY_OPT
243 #define FULL_TRIE_STUDY
249 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
250 #define PBITVAL(paren) (1 << ((paren) & 7))
251 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
252 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
253 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
255 #define REQUIRE_UTF8 STMT_START { \
257 *flagp = RESTART_UTF8; \
262 /* This converts the named class defined in regcomp.h to its equivalent class
263 * number defined in handy.h. */
264 #define namedclass_to_classnum(class) ((int) ((class) / 2))
265 #define classnum_to_namedclass(classnum) ((classnum) * 2)
267 /* About scan_data_t.
269 During optimisation we recurse through the regexp program performing
270 various inplace (keyhole style) optimisations. In addition study_chunk
271 and scan_commit populate this data structure with information about
272 what strings MUST appear in the pattern. We look for the longest
273 string that must appear at a fixed location, and we look for the
274 longest string that may appear at a floating location. So for instance
279 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
280 strings (because they follow a .* construct). study_chunk will identify
281 both FOO and BAR as being the longest fixed and floating strings respectively.
283 The strings can be composites, for instance
287 will result in a composite fixed substring 'foo'.
289 For each string some basic information is maintained:
291 - offset or min_offset
292 This is the position the string must appear at, or not before.
293 It also implicitly (when combined with minlenp) tells us how many
294 characters must match before the string we are searching for.
295 Likewise when combined with minlenp and the length of the string it
296 tells us how many characters must appear after the string we have
300 Only used for floating strings. This is the rightmost point that
301 the string can appear at. If set to I32 max it indicates that the
302 string can occur infinitely far to the right.
305 A pointer to the minimum number of characters of the pattern that the
306 string was found inside. This is important as in the case of positive
307 lookahead or positive lookbehind we can have multiple patterns
312 The minimum length of the pattern overall is 3, the minimum length
313 of the lookahead part is 3, but the minimum length of the part that
314 will actually match is 1. So 'FOO's minimum length is 3, but the
315 minimum length for the F is 1. This is important as the minimum length
316 is used to determine offsets in front of and behind the string being
317 looked for. Since strings can be composites this is the length of the
318 pattern at the time it was committed with a scan_commit. Note that
319 the length is calculated by study_chunk, so that the minimum lengths
320 are not known until the full pattern has been compiled, thus the
321 pointer to the value.
325 In the case of lookbehind the string being searched for can be
326 offset past the start point of the final matching string.
327 If this value was just blithely removed from the min_offset it would
328 invalidate some of the calculations for how many chars must match
329 before or after (as they are derived from min_offset and minlen and
330 the length of the string being searched for).
331 When the final pattern is compiled and the data is moved from the
332 scan_data_t structure into the regexp structure the information
333 about lookbehind is factored in, with the information that would
334 have been lost precalculated in the end_shift field for the
337 The fields pos_min and pos_delta are used to store the minimum offset
338 and the delta to the maximum offset at the current point in the pattern.
342 typedef struct scan_data_t {
343 /*I32 len_min; unused */
344 /*I32 len_delta; unused */
348 I32 last_end; /* min value, <0 unless valid. */
351 SV **longest; /* Either &l_fixed, or &l_float. */
352 SV *longest_fixed; /* longest fixed string found in pattern */
353 I32 offset_fixed; /* offset where it starts */
354 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
355 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
356 SV *longest_float; /* longest floating string found in pattern */
357 I32 offset_float_min; /* earliest point in string it can appear */
358 I32 offset_float_max; /* latest point in string it can appear */
359 I32 *minlen_float; /* pointer to the minlen relevant to the string */
360 I32 lookbehind_float; /* is the position of the string modified by LB */
364 struct regnode_charclass_class *start_class;
368 * Forward declarations for pregcomp()'s friends.
371 static const scan_data_t zero_scan_data =
372 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
374 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
375 #define SF_BEFORE_SEOL 0x0001
376 #define SF_BEFORE_MEOL 0x0002
377 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
378 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
381 # define SF_FIX_SHIFT_EOL (0+2)
382 # define SF_FL_SHIFT_EOL (0+4)
384 # define SF_FIX_SHIFT_EOL (+2)
385 # define SF_FL_SHIFT_EOL (+4)
388 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
389 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
391 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
392 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
393 #define SF_IS_INF 0x0040
394 #define SF_HAS_PAR 0x0080
395 #define SF_IN_PAR 0x0100
396 #define SF_HAS_EVAL 0x0200
397 #define SCF_DO_SUBSTR 0x0400
398 #define SCF_DO_STCLASS_AND 0x0800
399 #define SCF_DO_STCLASS_OR 0x1000
400 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
401 #define SCF_WHILEM_VISITED_POS 0x2000
403 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
404 #define SCF_SEEN_ACCEPT 0x8000
406 #define UTF cBOOL(RExC_utf8)
408 /* The enums for all these are ordered so things work out correctly */
409 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
410 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
411 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
412 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
413 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
414 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
415 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
417 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
419 #define OOB_NAMEDCLASS -1
421 /* There is no code point that is out-of-bounds, so this is problematic. But
422 * its only current use is to initialize a variable that is always set before
424 #define OOB_UNICODE 0xDEADBEEF
426 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
427 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
430 /* length of regex to show in messages that don't mark a position within */
431 #define RegexLengthToShowInErrorMessages 127
434 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
435 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
436 * op/pragma/warn/regcomp.
438 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
439 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
441 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
444 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
445 * arg. Show regex, up to a maximum length. If it's too long, chop and add
448 #define _FAIL(code) STMT_START { \
449 const char *ellipses = ""; \
450 IV len = RExC_end - RExC_precomp; \
453 SAVEFREESV(RExC_rx_sv); \
454 if (len > RegexLengthToShowInErrorMessages) { \
455 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
456 len = RegexLengthToShowInErrorMessages - 10; \
462 #define FAIL(msg) _FAIL( \
463 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
464 msg, (int)len, RExC_precomp, ellipses))
466 #define FAIL2(msg,arg) _FAIL( \
467 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
468 arg, (int)len, RExC_precomp, ellipses))
471 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
473 #define Simple_vFAIL(m) STMT_START { \
474 const IV offset = RExC_parse - RExC_precomp; \
475 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
476 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
480 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
482 #define vFAIL(m) STMT_START { \
484 SAVEFREESV(RExC_rx_sv); \
489 * Like Simple_vFAIL(), but accepts two arguments.
491 #define Simple_vFAIL2(m,a1) STMT_START { \
492 const IV offset = RExC_parse - RExC_precomp; \
493 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
494 (int)offset, RExC_precomp, RExC_precomp + offset); \
498 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
500 #define vFAIL2(m,a1) STMT_START { \
502 SAVEFREESV(RExC_rx_sv); \
503 Simple_vFAIL2(m, a1); \
508 * Like Simple_vFAIL(), but accepts three arguments.
510 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
511 const IV offset = RExC_parse - RExC_precomp; \
512 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
513 (int)offset, RExC_precomp, RExC_precomp + offset); \
517 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
519 #define vFAIL3(m,a1,a2) STMT_START { \
521 SAVEFREESV(RExC_rx_sv); \
522 Simple_vFAIL3(m, a1, a2); \
526 * Like Simple_vFAIL(), but accepts four arguments.
528 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
529 const IV offset = RExC_parse - RExC_precomp; \
530 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
531 (int)offset, RExC_precomp, RExC_precomp + offset); \
534 #define vFAIL4(m,a1,a2,a3) STMT_START { \
536 SAVEFREESV(RExC_rx_sv); \
537 Simple_vFAIL4(m, a1, a2, a3); \
540 /* m is not necessarily a "literal string", in this macro */
541 #define reg_warn_non_literal_string(loc, m) STMT_START { \
542 const IV offset = loc - RExC_precomp; \
543 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
544 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
547 #define ckWARNreg(loc,m) STMT_START { \
548 const IV offset = loc - RExC_precomp; \
549 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
550 (int)offset, RExC_precomp, RExC_precomp + offset); \
553 #define vWARN_dep(loc, m) STMT_START { \
554 const IV offset = loc - RExC_precomp; \
555 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
556 (int)offset, RExC_precomp, RExC_precomp + offset); \
559 #define ckWARNdep(loc,m) STMT_START { \
560 const IV offset = loc - RExC_precomp; \
561 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
563 (int)offset, RExC_precomp, RExC_precomp + offset); \
566 #define ckWARNregdep(loc,m) STMT_START { \
567 const IV offset = loc - RExC_precomp; \
568 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
570 (int)offset, RExC_precomp, RExC_precomp + offset); \
573 #define ckWARN2regdep(loc,m, a1) STMT_START { \
574 const IV offset = loc - RExC_precomp; \
575 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
577 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
580 #define ckWARN2reg(loc, m, a1) STMT_START { \
581 const IV offset = loc - RExC_precomp; \
582 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
583 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
586 #define vWARN3(loc, m, a1, a2) STMT_START { \
587 const IV offset = loc - RExC_precomp; \
588 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
589 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
592 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
593 const IV offset = loc - RExC_precomp; \
594 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
595 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
598 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
599 const IV offset = loc - RExC_precomp; \
600 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
601 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
604 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
605 const IV offset = loc - RExC_precomp; \
606 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
607 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
610 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
611 const IV offset = loc - RExC_precomp; \
612 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
613 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
617 /* Allow for side effects in s */
618 #define REGC(c,s) STMT_START { \
619 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
622 /* Macros for recording node offsets. 20001227 mjd@plover.com
623 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
624 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
625 * Element 0 holds the number n.
626 * Position is 1 indexed.
628 #ifndef RE_TRACK_PATTERN_OFFSETS
629 #define Set_Node_Offset_To_R(node,byte)
630 #define Set_Node_Offset(node,byte)
631 #define Set_Cur_Node_Offset
632 #define Set_Node_Length_To_R(node,len)
633 #define Set_Node_Length(node,len)
634 #define Set_Node_Cur_Length(node)
635 #define Node_Offset(n)
636 #define Node_Length(n)
637 #define Set_Node_Offset_Length(node,offset,len)
638 #define ProgLen(ri) ri->u.proglen
639 #define SetProgLen(ri,x) ri->u.proglen = x
641 #define ProgLen(ri) ri->u.offsets[0]
642 #define SetProgLen(ri,x) ri->u.offsets[0] = x
643 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
645 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
646 __LINE__, (int)(node), (int)(byte))); \
648 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
650 RExC_offsets[2*(node)-1] = (byte); \
655 #define Set_Node_Offset(node,byte) \
656 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
657 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
659 #define Set_Node_Length_To_R(node,len) STMT_START { \
661 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
662 __LINE__, (int)(node), (int)(len))); \
664 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
666 RExC_offsets[2*(node)] = (len); \
671 #define Set_Node_Length(node,len) \
672 Set_Node_Length_To_R((node)-RExC_emit_start, len)
673 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
674 #define Set_Node_Cur_Length(node) \
675 Set_Node_Length(node, RExC_parse - parse_start)
677 /* Get offsets and lengths */
678 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
679 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
681 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
682 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
683 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
687 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
688 #define EXPERIMENTAL_INPLACESCAN
689 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
691 #define DEBUG_STUDYDATA(str,data,depth) \
692 DEBUG_OPTIMISE_MORE_r(if(data){ \
693 PerlIO_printf(Perl_debug_log, \
694 "%*s" str "Pos:%"IVdf"/%"IVdf \
695 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
696 (int)(depth)*2, "", \
697 (IV)((data)->pos_min), \
698 (IV)((data)->pos_delta), \
699 (UV)((data)->flags), \
700 (IV)((data)->whilem_c), \
701 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
702 is_inf ? "INF " : "" \
704 if ((data)->last_found) \
705 PerlIO_printf(Perl_debug_log, \
706 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
707 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
708 SvPVX_const((data)->last_found), \
709 (IV)((data)->last_end), \
710 (IV)((data)->last_start_min), \
711 (IV)((data)->last_start_max), \
712 ((data)->longest && \
713 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
714 SvPVX_const((data)->longest_fixed), \
715 (IV)((data)->offset_fixed), \
716 ((data)->longest && \
717 (data)->longest==&((data)->longest_float)) ? "*" : "", \
718 SvPVX_const((data)->longest_float), \
719 (IV)((data)->offset_float_min), \
720 (IV)((data)->offset_float_max) \
722 PerlIO_printf(Perl_debug_log,"\n"); \
725 /* Mark that we cannot extend a found fixed substring at this point.
726 Update the longest found anchored substring and the longest found
727 floating substrings if needed. */
730 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
732 const STRLEN l = CHR_SVLEN(data->last_found);
733 const STRLEN old_l = CHR_SVLEN(*data->longest);
734 GET_RE_DEBUG_FLAGS_DECL;
736 PERL_ARGS_ASSERT_SCAN_COMMIT;
738 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
739 SvSetMagicSV(*data->longest, data->last_found);
740 if (*data->longest == data->longest_fixed) {
741 data->offset_fixed = l ? data->last_start_min : data->pos_min;
742 if (data->flags & SF_BEFORE_EOL)
744 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
746 data->flags &= ~SF_FIX_BEFORE_EOL;
747 data->minlen_fixed=minlenp;
748 data->lookbehind_fixed=0;
750 else { /* *data->longest == data->longest_float */
751 data->offset_float_min = l ? data->last_start_min : data->pos_min;
752 data->offset_float_max = (l
753 ? data->last_start_max
754 : (data->pos_delta == I32_MAX ? I32_MAX : data->pos_min + data->pos_delta));
755 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
756 data->offset_float_max = I32_MAX;
757 if (data->flags & SF_BEFORE_EOL)
759 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
761 data->flags &= ~SF_FL_BEFORE_EOL;
762 data->minlen_float=minlenp;
763 data->lookbehind_float=0;
766 SvCUR_set(data->last_found, 0);
768 SV * const sv = data->last_found;
769 if (SvUTF8(sv) && SvMAGICAL(sv)) {
770 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
776 data->flags &= ~SF_BEFORE_EOL;
777 DEBUG_STUDYDATA("commit: ",data,0);
780 /* These macros set, clear and test whether the synthetic start class ('ssc',
781 * given by the parameter) matches an empty string (EOS). This uses the
782 * 'next_off' field in the node, to save a bit in the flags field. The ssc
783 * stands alone, so there is never a next_off, so this field is otherwise
784 * unused. The EOS information is used only for compilation, but theoretically
785 * it could be passed on to the execution code. This could be used to store
786 * more than one bit of information, but only this one is currently used. */
787 #define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
788 #define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
789 #define TEST_SSC_EOS(node) cBOOL((node)->next_off)
791 /* Can match anything (initialization) */
793 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
795 PERL_ARGS_ASSERT_CL_ANYTHING;
797 ANYOF_BITMAP_SETALL(cl);
798 cl->flags = ANYOF_UNICODE_ALL;
801 /* If any portion of the regex is to operate under locale rules,
802 * initialization includes it. The reason this isn't done for all regexes
803 * is that the optimizer was written under the assumption that locale was
804 * all-or-nothing. Given the complexity and lack of documentation in the
805 * optimizer, and that there are inadequate test cases for locale, so many
806 * parts of it may not work properly, it is safest to avoid locale unless
808 if (RExC_contains_locale) {
809 ANYOF_CLASS_SETALL(cl); /* /l uses class */
810 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
813 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
817 /* Can match anything (initialization) */
819 S_cl_is_anything(const struct regnode_charclass_class *cl)
823 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
825 for (value = 0; value < ANYOF_MAX; value += 2)
826 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
828 if (!(cl->flags & ANYOF_UNICODE_ALL))
830 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
835 /* Can match anything (initialization) */
837 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
839 PERL_ARGS_ASSERT_CL_INIT;
841 Zero(cl, 1, struct regnode_charclass_class);
843 cl_anything(pRExC_state, cl);
844 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
847 /* These two functions currently do the exact same thing */
848 #define cl_init_zero S_cl_init
850 /* 'AND' a given class with another one. Can create false positives. 'cl'
851 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
852 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
854 S_cl_and(struct regnode_charclass_class *cl,
855 const struct regnode_charclass_class *and_with)
857 PERL_ARGS_ASSERT_CL_AND;
859 assert(PL_regkind[and_with->type] == ANYOF);
861 /* I (khw) am not sure all these restrictions are necessary XXX */
862 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
863 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
864 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
865 && !(and_with->flags & ANYOF_LOC_FOLD)
866 && !(cl->flags & ANYOF_LOC_FOLD)) {
869 if (and_with->flags & ANYOF_INVERT)
870 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
871 cl->bitmap[i] &= ~and_with->bitmap[i];
873 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
874 cl->bitmap[i] &= and_with->bitmap[i];
875 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
877 if (and_with->flags & ANYOF_INVERT) {
879 /* Here, the and'ed node is inverted. Get the AND of the flags that
880 * aren't affected by the inversion. Those that are affected are
881 * handled individually below */
882 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
883 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
884 cl->flags |= affected_flags;
886 /* We currently don't know how to deal with things that aren't in the
887 * bitmap, but we know that the intersection is no greater than what
888 * is already in cl, so let there be false positives that get sorted
889 * out after the synthetic start class succeeds, and the node is
890 * matched for real. */
892 /* The inversion of these two flags indicate that the resulting
893 * intersection doesn't have them */
894 if (and_with->flags & ANYOF_UNICODE_ALL) {
895 cl->flags &= ~ANYOF_UNICODE_ALL;
897 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
898 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
901 else { /* and'd node is not inverted */
902 U8 outside_bitmap_but_not_utf8; /* Temp variable */
904 if (! ANYOF_NONBITMAP(and_with)) {
906 /* Here 'and_with' doesn't match anything outside the bitmap
907 * (except possibly ANYOF_UNICODE_ALL), which means the
908 * intersection can't either, except for ANYOF_UNICODE_ALL, in
909 * which case we don't know what the intersection is, but it's no
910 * greater than what cl already has, so can just leave it alone,
911 * with possible false positives */
912 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
913 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
914 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
917 else if (! ANYOF_NONBITMAP(cl)) {
919 /* Here, 'and_with' does match something outside the bitmap, and cl
920 * doesn't have a list of things to match outside the bitmap. If
921 * cl can match all code points above 255, the intersection will
922 * be those above-255 code points that 'and_with' matches. If cl
923 * can't match all Unicode code points, it means that it can't
924 * match anything outside the bitmap (since the 'if' that got us
925 * into this block tested for that), so we leave the bitmap empty.
927 if (cl->flags & ANYOF_UNICODE_ALL) {
928 ARG_SET(cl, ARG(and_with));
930 /* and_with's ARG may match things that don't require UTF8.
931 * And now cl's will too, in spite of this being an 'and'. See
932 * the comments below about the kludge */
933 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
937 /* Here, both 'and_with' and cl match something outside the
938 * bitmap. Currently we do not do the intersection, so just match
939 * whatever cl had at the beginning. */
943 /* Take the intersection of the two sets of flags. However, the
944 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
945 * kludge around the fact that this flag is not treated like the others
946 * which are initialized in cl_anything(). The way the optimizer works
947 * is that the synthetic start class (SSC) is initialized to match
948 * anything, and then the first time a real node is encountered, its
949 * values are AND'd with the SSC's with the result being the values of
950 * the real node. However, there are paths through the optimizer where
951 * the AND never gets called, so those initialized bits are set
952 * inappropriately, which is not usually a big deal, as they just cause
953 * false positives in the SSC, which will just mean a probably
954 * imperceptible slow down in execution. However this bit has a
955 * higher false positive consequence in that it can cause utf8.pm,
956 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
957 * bigger slowdown and also causes significant extra memory to be used.
958 * In order to prevent this, the code now takes a different tack. The
959 * bit isn't set unless some part of the regular expression needs it,
960 * but once set it won't get cleared. This means that these extra
961 * modules won't get loaded unless there was some path through the
962 * pattern that would have required them anyway, and so any false
963 * positives that occur by not ANDing them out when they could be
964 * aren't as severe as they would be if we treated this bit like all
966 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
967 & ANYOF_NONBITMAP_NON_UTF8;
968 cl->flags &= and_with->flags;
969 cl->flags |= outside_bitmap_but_not_utf8;
973 /* 'OR' a given class with another one. Can create false positives. 'cl'
974 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
975 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
977 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
979 PERL_ARGS_ASSERT_CL_OR;
981 if (or_with->flags & ANYOF_INVERT) {
983 /* Here, the or'd node is to be inverted. This means we take the
984 * complement of everything not in the bitmap, but currently we don't
985 * know what that is, so give up and match anything */
986 if (ANYOF_NONBITMAP(or_with)) {
987 cl_anything(pRExC_state, cl);
990 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
991 * <= (B1 | !B2) | (CL1 | !CL2)
992 * which is wasteful if CL2 is small, but we ignore CL2:
993 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
994 * XXXX Can we handle case-fold? Unclear:
995 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
996 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
998 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
999 && !(or_with->flags & ANYOF_LOC_FOLD)
1000 && !(cl->flags & ANYOF_LOC_FOLD) ) {
1003 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1004 cl->bitmap[i] |= ~or_with->bitmap[i];
1005 } /* XXXX: logic is complicated otherwise */
1007 cl_anything(pRExC_state, cl);
1010 /* And, we can just take the union of the flags that aren't affected
1011 * by the inversion */
1012 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1014 /* For the remaining flags:
1015 ANYOF_UNICODE_ALL and inverted means to not match anything above
1016 255, which means that the union with cl should just be
1017 what cl has in it, so can ignore this flag
1018 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1019 is 127-255 to match them, but then invert that, so the
1020 union with cl should just be what cl has in it, so can
1023 } else { /* 'or_with' is not inverted */
1024 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1025 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1026 && (!(or_with->flags & ANYOF_LOC_FOLD)
1027 || (cl->flags & ANYOF_LOC_FOLD)) ) {
1030 /* OR char bitmap and class bitmap separately */
1031 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1032 cl->bitmap[i] |= or_with->bitmap[i];
1033 if (or_with->flags & ANYOF_CLASS) {
1034 ANYOF_CLASS_OR(or_with, cl);
1037 else { /* XXXX: logic is complicated, leave it along for a moment. */
1038 cl_anything(pRExC_state, cl);
1041 if (ANYOF_NONBITMAP(or_with)) {
1043 /* Use the added node's outside-the-bit-map match if there isn't a
1044 * conflict. If there is a conflict (both nodes match something
1045 * outside the bitmap, but what they match outside is not the same
1046 * pointer, and hence not easily compared until XXX we extend
1047 * inversion lists this far), give up and allow the start class to
1048 * match everything outside the bitmap. If that stuff is all above
1049 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1050 if (! ANYOF_NONBITMAP(cl)) {
1051 ARG_SET(cl, ARG(or_with));
1053 else if (ARG(cl) != ARG(or_with)) {
1055 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1056 cl_anything(pRExC_state, cl);
1059 cl->flags |= ANYOF_UNICODE_ALL;
1064 /* Take the union */
1065 cl->flags |= or_with->flags;
1069 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1070 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1071 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1072 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1077 dump_trie(trie,widecharmap,revcharmap)
1078 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1079 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1081 These routines dump out a trie in a somewhat readable format.
1082 The _interim_ variants are used for debugging the interim
1083 tables that are used to generate the final compressed
1084 representation which is what dump_trie expects.
1086 Part of the reason for their existence is to provide a form
1087 of documentation as to how the different representations function.
1092 Dumps the final compressed table form of the trie to Perl_debug_log.
1093 Used for debugging make_trie().
1097 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1098 AV *revcharmap, U32 depth)
1101 SV *sv=sv_newmortal();
1102 int colwidth= widecharmap ? 6 : 4;
1104 GET_RE_DEBUG_FLAGS_DECL;
1106 PERL_ARGS_ASSERT_DUMP_TRIE;
1108 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1109 (int)depth * 2 + 2,"",
1110 "Match","Base","Ofs" );
1112 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1113 SV ** const tmp = av_fetch( revcharmap, state, 0);
1115 PerlIO_printf( Perl_debug_log, "%*s",
1117 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1118 PL_colors[0], PL_colors[1],
1119 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1120 PERL_PV_ESCAPE_FIRSTCHAR
1125 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1126 (int)depth * 2 + 2,"");
1128 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1129 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1130 PerlIO_printf( Perl_debug_log, "\n");
1132 for( state = 1 ; state < trie->statecount ; state++ ) {
1133 const U32 base = trie->states[ state ].trans.base;
1135 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1137 if ( trie->states[ state ].wordnum ) {
1138 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1140 PerlIO_printf( Perl_debug_log, "%6s", "" );
1143 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1148 while( ( base + ofs < trie->uniquecharcount ) ||
1149 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1150 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1153 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1155 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1156 if ( ( base + ofs >= trie->uniquecharcount ) &&
1157 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1158 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1160 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1162 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1164 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1168 PerlIO_printf( Perl_debug_log, "]");
1171 PerlIO_printf( Perl_debug_log, "\n" );
1173 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1174 for (word=1; word <= trie->wordcount; word++) {
1175 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1176 (int)word, (int)(trie->wordinfo[word].prev),
1177 (int)(trie->wordinfo[word].len));
1179 PerlIO_printf(Perl_debug_log, "\n" );
1182 Dumps a fully constructed but uncompressed trie in list form.
1183 List tries normally only are used for construction when the number of
1184 possible chars (trie->uniquecharcount) is very high.
1185 Used for debugging make_trie().
1188 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1189 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1193 SV *sv=sv_newmortal();
1194 int colwidth= widecharmap ? 6 : 4;
1195 GET_RE_DEBUG_FLAGS_DECL;
1197 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1199 /* print out the table precompression. */
1200 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1201 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1202 "------:-----+-----------------\n" );
1204 for( state=1 ; state < next_alloc ; state ++ ) {
1207 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1208 (int)depth * 2 + 2,"", (UV)state );
1209 if ( ! trie->states[ state ].wordnum ) {
1210 PerlIO_printf( Perl_debug_log, "%5s| ","");
1212 PerlIO_printf( Perl_debug_log, "W%4x| ",
1213 trie->states[ state ].wordnum
1216 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1217 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1219 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1221 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1222 PL_colors[0], PL_colors[1],
1223 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1224 PERL_PV_ESCAPE_FIRSTCHAR
1226 TRIE_LIST_ITEM(state,charid).forid,
1227 (UV)TRIE_LIST_ITEM(state,charid).newstate
1230 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1231 (int)((depth * 2) + 14), "");
1234 PerlIO_printf( Perl_debug_log, "\n");
1239 Dumps a fully constructed but uncompressed trie in table form.
1240 This is the normal DFA style state transition table, with a few
1241 twists to facilitate compression later.
1242 Used for debugging make_trie().
1245 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1246 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1251 SV *sv=sv_newmortal();
1252 int colwidth= widecharmap ? 6 : 4;
1253 GET_RE_DEBUG_FLAGS_DECL;
1255 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1258 print out the table precompression so that we can do a visual check
1259 that they are identical.
1262 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1264 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1265 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1267 PerlIO_printf( Perl_debug_log, "%*s",
1269 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1270 PL_colors[0], PL_colors[1],
1271 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1272 PERL_PV_ESCAPE_FIRSTCHAR
1278 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1280 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1281 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1284 PerlIO_printf( Perl_debug_log, "\n" );
1286 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1288 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1289 (int)depth * 2 + 2,"",
1290 (UV)TRIE_NODENUM( state ) );
1292 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1293 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1295 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1297 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1299 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1300 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1302 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1303 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1311 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1312 startbranch: the first branch in the whole branch sequence
1313 first : start branch of sequence of branch-exact nodes.
1314 May be the same as startbranch
1315 last : Thing following the last branch.
1316 May be the same as tail.
1317 tail : item following the branch sequence
1318 count : words in the sequence
1319 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1320 depth : indent depth
1322 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1324 A trie is an N'ary tree where the branches are determined by digital
1325 decomposition of the key. IE, at the root node you look up the 1st character and
1326 follow that branch repeat until you find the end of the branches. Nodes can be
1327 marked as "accepting" meaning they represent a complete word. Eg:
1331 would convert into the following structure. Numbers represent states, letters
1332 following numbers represent valid transitions on the letter from that state, if
1333 the number is in square brackets it represents an accepting state, otherwise it
1334 will be in parenthesis.
1336 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1340 (1) +-i->(6)-+-s->[7]
1342 +-s->(3)-+-h->(4)-+-e->[5]
1344 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1346 This shows that when matching against the string 'hers' we will begin at state 1
1347 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1348 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1349 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1350 single traverse. We store a mapping from accepting to state to which word was
1351 matched, and then when we have multiple possibilities we try to complete the
1352 rest of the regex in the order in which they occured in the alternation.
1354 The only prior NFA like behaviour that would be changed by the TRIE support is
1355 the silent ignoring of duplicate alternations which are of the form:
1357 / (DUPE|DUPE) X? (?{ ... }) Y /x
1359 Thus EVAL blocks following a trie may be called a different number of times with
1360 and without the optimisation. With the optimisations dupes will be silently
1361 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1362 the following demonstrates:
1364 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1366 which prints out 'word' three times, but
1368 'words'=~/(word|word|word)(?{ print $1 })S/
1370 which doesnt print it out at all. This is due to other optimisations kicking in.
1372 Example of what happens on a structural level:
1374 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1376 1: CURLYM[1] {1,32767}(18)
1387 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1388 and should turn into:
1390 1: CURLYM[1] {1,32767}(18)
1392 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1400 Cases where tail != last would be like /(?foo|bar)baz/:
1410 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1411 and would end up looking like:
1414 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1421 d = uvuni_to_utf8_flags(d, uv, 0);
1423 is the recommended Unicode-aware way of saying
1428 #define TRIE_STORE_REVCHAR(val) \
1431 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1432 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1433 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1434 SvCUR_set(zlopp, kapow - flrbbbbb); \
1437 av_push(revcharmap, zlopp); \
1439 char ooooff = (char)val; \
1440 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1444 #define TRIE_READ_CHAR STMT_START { \
1447 /* if it is UTF then it is either already folded, or does not need folding */ \
1448 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1450 else if (folder == PL_fold_latin1) { \
1451 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1452 if ( foldlen > 0 ) { \
1453 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1459 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1460 skiplen = UNISKIP(uvc); \
1461 foldlen -= skiplen; \
1462 scan = foldbuf + skiplen; \
1465 /* raw data, will be folded later if needed */ \
1473 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1474 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1475 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1476 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1478 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1479 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1480 TRIE_LIST_CUR( state )++; \
1483 #define TRIE_LIST_NEW(state) STMT_START { \
1484 Newxz( trie->states[ state ].trans.list, \
1485 4, reg_trie_trans_le ); \
1486 TRIE_LIST_CUR( state ) = 1; \
1487 TRIE_LIST_LEN( state ) = 4; \
1490 #define TRIE_HANDLE_WORD(state) STMT_START { \
1491 U16 dupe= trie->states[ state ].wordnum; \
1492 regnode * const noper_next = regnext( noper ); \
1495 /* store the word for dumping */ \
1497 if (OP(noper) != NOTHING) \
1498 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1500 tmp = newSVpvn_utf8( "", 0, UTF ); \
1501 av_push( trie_words, tmp ); \
1505 trie->wordinfo[curword].prev = 0; \
1506 trie->wordinfo[curword].len = wordlen; \
1507 trie->wordinfo[curword].accept = state; \
1509 if ( noper_next < tail ) { \
1511 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1512 trie->jump[curword] = (U16)(noper_next - convert); \
1514 jumper = noper_next; \
1516 nextbranch= regnext(cur); \
1520 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1521 /* chain, so that when the bits of chain are later */\
1522 /* linked together, the dups appear in the chain */\
1523 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1524 trie->wordinfo[dupe].prev = curword; \
1526 /* we haven't inserted this word yet. */ \
1527 trie->states[ state ].wordnum = curword; \
1532 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1533 ( ( base + charid >= ucharcount \
1534 && base + charid < ubound \
1535 && state == trie->trans[ base - ucharcount + charid ].check \
1536 && trie->trans[ base - ucharcount + charid ].next ) \
1537 ? trie->trans[ base - ucharcount + charid ].next \
1538 : ( state==1 ? special : 0 ) \
1542 #define MADE_JUMP_TRIE 2
1543 #define MADE_EXACT_TRIE 4
1546 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1549 /* first pass, loop through and scan words */
1550 reg_trie_data *trie;
1551 HV *widecharmap = NULL;
1552 AV *revcharmap = newAV();
1554 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1559 regnode *jumper = NULL;
1560 regnode *nextbranch = NULL;
1561 regnode *convert = NULL;
1562 U32 *prev_states; /* temp array mapping each state to previous one */
1563 /* we just use folder as a flag in utf8 */
1564 const U8 * folder = NULL;
1567 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1568 AV *trie_words = NULL;
1569 /* along with revcharmap, this only used during construction but both are
1570 * useful during debugging so we store them in the struct when debugging.
1573 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1574 STRLEN trie_charcount=0;
1576 SV *re_trie_maxbuff;
1577 GET_RE_DEBUG_FLAGS_DECL;
1579 PERL_ARGS_ASSERT_MAKE_TRIE;
1581 PERL_UNUSED_ARG(depth);
1588 case EXACTFU_TRICKYFOLD:
1589 case EXACTFU: folder = PL_fold_latin1; break;
1590 case EXACTF: folder = PL_fold; break;
1591 case EXACTFL: folder = PL_fold_locale; break;
1592 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1595 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1597 trie->startstate = 1;
1598 trie->wordcount = word_count;
1599 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1600 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1602 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1603 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1604 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1607 trie_words = newAV();
1610 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1611 if (!SvIOK(re_trie_maxbuff)) {
1612 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1614 DEBUG_TRIE_COMPILE_r({
1615 PerlIO_printf( Perl_debug_log,
1616 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1617 (int)depth * 2 + 2, "",
1618 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1619 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1623 /* Find the node we are going to overwrite */
1624 if ( first == startbranch && OP( last ) != BRANCH ) {
1625 /* whole branch chain */
1628 /* branch sub-chain */
1629 convert = NEXTOPER( first );
1632 /* -- First loop and Setup --
1634 We first traverse the branches and scan each word to determine if it
1635 contains widechars, and how many unique chars there are, this is
1636 important as we have to build a table with at least as many columns as we
1639 We use an array of integers to represent the character codes 0..255
1640 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1641 native representation of the character value as the key and IV's for the
1644 *TODO* If we keep track of how many times each character is used we can
1645 remap the columns so that the table compression later on is more
1646 efficient in terms of memory by ensuring the most common value is in the
1647 middle and the least common are on the outside. IMO this would be better
1648 than a most to least common mapping as theres a decent chance the most
1649 common letter will share a node with the least common, meaning the node
1650 will not be compressible. With a middle is most common approach the worst
1651 case is when we have the least common nodes twice.
1655 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1656 regnode *noper = NEXTOPER( cur );
1657 const U8 *uc = (U8*)STRING( noper );
1658 const U8 *e = uc + STR_LEN( noper );
1660 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1662 const U8 *scan = (U8*)NULL;
1663 U32 wordlen = 0; /* required init */
1665 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1667 if (OP(noper) == NOTHING) {
1668 regnode *noper_next= regnext(noper);
1669 if (noper_next != tail && OP(noper_next) == flags) {
1671 uc= (U8*)STRING(noper);
1672 e= uc + STR_LEN(noper);
1673 trie->minlen= STR_LEN(noper);
1680 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1681 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1682 regardless of encoding */
1683 if (OP( noper ) == EXACTFU_SS) {
1684 /* false positives are ok, so just set this */
1685 TRIE_BITMAP_SET(trie,0xDF);
1688 for ( ; uc < e ; uc += len ) {
1689 TRIE_CHARCOUNT(trie)++;
1694 U8 folded= folder[ (U8) uvc ];
1695 if ( !trie->charmap[ folded ] ) {
1696 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1697 TRIE_STORE_REVCHAR( folded );
1700 if ( !trie->charmap[ uvc ] ) {
1701 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1702 TRIE_STORE_REVCHAR( uvc );
1705 /* store the codepoint in the bitmap, and its folded
1707 TRIE_BITMAP_SET(trie, uvc);
1709 /* store the folded codepoint */
1710 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1713 /* store first byte of utf8 representation of
1714 variant codepoints */
1715 if (! UNI_IS_INVARIANT(uvc)) {
1716 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1719 set_bit = 0; /* We've done our bit :-) */
1724 widecharmap = newHV();
1726 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1729 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1731 if ( !SvTRUE( *svpp ) ) {
1732 sv_setiv( *svpp, ++trie->uniquecharcount );
1733 TRIE_STORE_REVCHAR(uvc);
1737 if( cur == first ) {
1738 trie->minlen = chars;
1739 trie->maxlen = chars;
1740 } else if (chars < trie->minlen) {
1741 trie->minlen = chars;
1742 } else if (chars > trie->maxlen) {
1743 trie->maxlen = chars;
1745 if (OP( noper ) == EXACTFU_SS) {
1746 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1747 if (trie->minlen > 1)
1750 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1751 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1752 * - We assume that any such sequence might match a 2 byte string */
1753 if (trie->minlen > 2 )
1757 } /* end first pass */
1758 DEBUG_TRIE_COMPILE_r(
1759 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1760 (int)depth * 2 + 2,"",
1761 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1762 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1763 (int)trie->minlen, (int)trie->maxlen )
1767 We now know what we are dealing with in terms of unique chars and
1768 string sizes so we can calculate how much memory a naive
1769 representation using a flat table will take. If it's over a reasonable
1770 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1771 conservative but potentially much slower representation using an array
1774 At the end we convert both representations into the same compressed
1775 form that will be used in regexec.c for matching with. The latter
1776 is a form that cannot be used to construct with but has memory
1777 properties similar to the list form and access properties similar
1778 to the table form making it both suitable for fast searches and
1779 small enough that its feasable to store for the duration of a program.
1781 See the comment in the code where the compressed table is produced
1782 inplace from the flat tabe representation for an explanation of how
1783 the compression works.
1788 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1791 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1793 Second Pass -- Array Of Lists Representation
1795 Each state will be represented by a list of charid:state records
1796 (reg_trie_trans_le) the first such element holds the CUR and LEN
1797 points of the allocated array. (See defines above).
1799 We build the initial structure using the lists, and then convert
1800 it into the compressed table form which allows faster lookups
1801 (but cant be modified once converted).
1804 STRLEN transcount = 1;
1806 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1807 "%*sCompiling trie using list compiler\n",
1808 (int)depth * 2 + 2, ""));
1810 trie->states = (reg_trie_state *)
1811 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1812 sizeof(reg_trie_state) );
1816 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1818 regnode *noper = NEXTOPER( cur );
1819 U8 *uc = (U8*)STRING( noper );
1820 const U8 *e = uc + STR_LEN( noper );
1821 U32 state = 1; /* required init */
1822 U16 charid = 0; /* sanity init */
1823 U8 *scan = (U8*)NULL; /* sanity init */
1824 STRLEN foldlen = 0; /* required init */
1825 U32 wordlen = 0; /* required init */
1826 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1829 if (OP(noper) == NOTHING) {
1830 regnode *noper_next= regnext(noper);
1831 if (noper_next != tail && OP(noper_next) == flags) {
1833 uc= (U8*)STRING(noper);
1834 e= uc + STR_LEN(noper);
1838 if (OP(noper) != NOTHING) {
1839 for ( ; uc < e ; uc += len ) {
1844 charid = trie->charmap[ uvc ];
1846 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1850 charid=(U16)SvIV( *svpp );
1853 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1860 if ( !trie->states[ state ].trans.list ) {
1861 TRIE_LIST_NEW( state );
1863 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1864 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1865 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1870 newstate = next_alloc++;
1871 prev_states[newstate] = state;
1872 TRIE_LIST_PUSH( state, charid, newstate );
1877 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1881 TRIE_HANDLE_WORD(state);
1883 } /* end second pass */
1885 /* next alloc is the NEXT state to be allocated */
1886 trie->statecount = next_alloc;
1887 trie->states = (reg_trie_state *)
1888 PerlMemShared_realloc( trie->states,
1890 * sizeof(reg_trie_state) );
1892 /* and now dump it out before we compress it */
1893 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1894 revcharmap, next_alloc,
1898 trie->trans = (reg_trie_trans *)
1899 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1906 for( state=1 ; state < next_alloc ; state ++ ) {
1910 DEBUG_TRIE_COMPILE_MORE_r(
1911 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1915 if (trie->states[state].trans.list) {
1916 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1920 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1921 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1922 if ( forid < minid ) {
1924 } else if ( forid > maxid ) {
1928 if ( transcount < tp + maxid - minid + 1) {
1930 trie->trans = (reg_trie_trans *)
1931 PerlMemShared_realloc( trie->trans,
1933 * sizeof(reg_trie_trans) );
1934 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1936 base = trie->uniquecharcount + tp - minid;
1937 if ( maxid == minid ) {
1939 for ( ; zp < tp ; zp++ ) {
1940 if ( ! trie->trans[ zp ].next ) {
1941 base = trie->uniquecharcount + zp - minid;
1942 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1943 trie->trans[ zp ].check = state;
1949 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1950 trie->trans[ tp ].check = state;
1955 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1956 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1957 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1958 trie->trans[ tid ].check = state;
1960 tp += ( maxid - minid + 1 );
1962 Safefree(trie->states[ state ].trans.list);
1965 DEBUG_TRIE_COMPILE_MORE_r(
1966 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1969 trie->states[ state ].trans.base=base;
1971 trie->lasttrans = tp + 1;
1975 Second Pass -- Flat Table Representation.
1977 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1978 We know that we will need Charcount+1 trans at most to store the data
1979 (one row per char at worst case) So we preallocate both structures
1980 assuming worst case.
1982 We then construct the trie using only the .next slots of the entry
1985 We use the .check field of the first entry of the node temporarily to
1986 make compression both faster and easier by keeping track of how many non
1987 zero fields are in the node.
1989 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1992 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1993 number representing the first entry of the node, and state as a
1994 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1995 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1996 are 2 entrys per node. eg:
2004 The table is internally in the right hand, idx form. However as we also
2005 have to deal with the states array which is indexed by nodenum we have to
2006 use TRIE_NODENUM() to convert.
2009 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2010 "%*sCompiling trie using table compiler\n",
2011 (int)depth * 2 + 2, ""));
2013 trie->trans = (reg_trie_trans *)
2014 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2015 * trie->uniquecharcount + 1,
2016 sizeof(reg_trie_trans) );
2017 trie->states = (reg_trie_state *)
2018 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2019 sizeof(reg_trie_state) );
2020 next_alloc = trie->uniquecharcount + 1;
2023 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2025 regnode *noper = NEXTOPER( cur );
2026 const U8 *uc = (U8*)STRING( noper );
2027 const U8 *e = uc + STR_LEN( noper );
2029 U32 state = 1; /* required init */
2031 U16 charid = 0; /* sanity init */
2032 U32 accept_state = 0; /* sanity init */
2033 U8 *scan = (U8*)NULL; /* sanity init */
2035 STRLEN foldlen = 0; /* required init */
2036 U32 wordlen = 0; /* required init */
2038 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2040 if (OP(noper) == NOTHING) {
2041 regnode *noper_next= regnext(noper);
2042 if (noper_next != tail && OP(noper_next) == flags) {
2044 uc= (U8*)STRING(noper);
2045 e= uc + STR_LEN(noper);
2049 if ( OP(noper) != NOTHING ) {
2050 for ( ; uc < e ; uc += len ) {
2055 charid = trie->charmap[ uvc ];
2057 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2058 charid = svpp ? (U16)SvIV(*svpp) : 0;
2062 if ( !trie->trans[ state + charid ].next ) {
2063 trie->trans[ state + charid ].next = next_alloc;
2064 trie->trans[ state ].check++;
2065 prev_states[TRIE_NODENUM(next_alloc)]
2066 = TRIE_NODENUM(state);
2067 next_alloc += trie->uniquecharcount;
2069 state = trie->trans[ state + charid ].next;
2071 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2073 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2076 accept_state = TRIE_NODENUM( state );
2077 TRIE_HANDLE_WORD(accept_state);
2079 } /* end second pass */
2081 /* and now dump it out before we compress it */
2082 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2084 next_alloc, depth+1));
2088 * Inplace compress the table.*
2090 For sparse data sets the table constructed by the trie algorithm will
2091 be mostly 0/FAIL transitions or to put it another way mostly empty.
2092 (Note that leaf nodes will not contain any transitions.)
2094 This algorithm compresses the tables by eliminating most such
2095 transitions, at the cost of a modest bit of extra work during lookup:
2097 - Each states[] entry contains a .base field which indicates the
2098 index in the state[] array wheres its transition data is stored.
2100 - If .base is 0 there are no valid transitions from that node.
2102 - If .base is nonzero then charid is added to it to find an entry in
2105 -If trans[states[state].base+charid].check!=state then the
2106 transition is taken to be a 0/Fail transition. Thus if there are fail
2107 transitions at the front of the node then the .base offset will point
2108 somewhere inside the previous nodes data (or maybe even into a node
2109 even earlier), but the .check field determines if the transition is
2113 The following process inplace converts the table to the compressed
2114 table: We first do not compress the root node 1,and mark all its
2115 .check pointers as 1 and set its .base pointer as 1 as well. This
2116 allows us to do a DFA construction from the compressed table later,
2117 and ensures that any .base pointers we calculate later are greater
2120 - We set 'pos' to indicate the first entry of the second node.
2122 - We then iterate over the columns of the node, finding the first and
2123 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2124 and set the .check pointers accordingly, and advance pos
2125 appropriately and repreat for the next node. Note that when we copy
2126 the next pointers we have to convert them from the original
2127 NODEIDX form to NODENUM form as the former is not valid post
2130 - If a node has no transitions used we mark its base as 0 and do not
2131 advance the pos pointer.
2133 - If a node only has one transition we use a second pointer into the
2134 structure to fill in allocated fail transitions from other states.
2135 This pointer is independent of the main pointer and scans forward
2136 looking for null transitions that are allocated to a state. When it
2137 finds one it writes the single transition into the "hole". If the
2138 pointer doesnt find one the single transition is appended as normal.
2140 - Once compressed we can Renew/realloc the structures to release the
2143 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2144 specifically Fig 3.47 and the associated pseudocode.
2148 const U32 laststate = TRIE_NODENUM( next_alloc );
2151 trie->statecount = laststate;
2153 for ( state = 1 ; state < laststate ; state++ ) {
2155 const U32 stateidx = TRIE_NODEIDX( state );
2156 const U32 o_used = trie->trans[ stateidx ].check;
2157 U32 used = trie->trans[ stateidx ].check;
2158 trie->trans[ stateidx ].check = 0;
2160 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2161 if ( flag || trie->trans[ stateidx + charid ].next ) {
2162 if ( trie->trans[ stateidx + charid ].next ) {
2164 for ( ; zp < pos ; zp++ ) {
2165 if ( ! trie->trans[ zp ].next ) {
2169 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2170 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2171 trie->trans[ zp ].check = state;
2172 if ( ++zp > pos ) pos = zp;
2179 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2181 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2182 trie->trans[ pos ].check = state;
2187 trie->lasttrans = pos + 1;
2188 trie->states = (reg_trie_state *)
2189 PerlMemShared_realloc( trie->states, laststate
2190 * sizeof(reg_trie_state) );
2191 DEBUG_TRIE_COMPILE_MORE_r(
2192 PerlIO_printf( Perl_debug_log,
2193 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2194 (int)depth * 2 + 2,"",
2195 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2198 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2201 } /* end table compress */
2203 DEBUG_TRIE_COMPILE_MORE_r(
2204 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2205 (int)depth * 2 + 2, "",
2206 (UV)trie->statecount,
2207 (UV)trie->lasttrans)
2209 /* resize the trans array to remove unused space */
2210 trie->trans = (reg_trie_trans *)
2211 PerlMemShared_realloc( trie->trans, trie->lasttrans
2212 * sizeof(reg_trie_trans) );
2214 { /* Modify the program and insert the new TRIE node */
2215 U8 nodetype =(U8)(flags & 0xFF);
2219 regnode *optimize = NULL;
2220 #ifdef RE_TRACK_PATTERN_OFFSETS
2223 U32 mjd_nodelen = 0;
2224 #endif /* RE_TRACK_PATTERN_OFFSETS */
2225 #endif /* DEBUGGING */
2227 This means we convert either the first branch or the first Exact,
2228 depending on whether the thing following (in 'last') is a branch
2229 or not and whther first is the startbranch (ie is it a sub part of
2230 the alternation or is it the whole thing.)
2231 Assuming its a sub part we convert the EXACT otherwise we convert
2232 the whole branch sequence, including the first.
2234 /* Find the node we are going to overwrite */
2235 if ( first != startbranch || OP( last ) == BRANCH ) {
2236 /* branch sub-chain */
2237 NEXT_OFF( first ) = (U16)(last - first);
2238 #ifdef RE_TRACK_PATTERN_OFFSETS
2240 mjd_offset= Node_Offset((convert));
2241 mjd_nodelen= Node_Length((convert));
2244 /* whole branch chain */
2246 #ifdef RE_TRACK_PATTERN_OFFSETS
2249 const regnode *nop = NEXTOPER( convert );
2250 mjd_offset= Node_Offset((nop));
2251 mjd_nodelen= Node_Length((nop));
2255 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2256 (int)depth * 2 + 2, "",
2257 (UV)mjd_offset, (UV)mjd_nodelen)
2260 /* But first we check to see if there is a common prefix we can
2261 split out as an EXACT and put in front of the TRIE node. */
2262 trie->startstate= 1;
2263 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2265 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2269 const U32 base = trie->states[ state ].trans.base;
2271 if ( trie->states[state].wordnum )
2274 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2275 if ( ( base + ofs >= trie->uniquecharcount ) &&
2276 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2277 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2279 if ( ++count > 1 ) {
2280 SV **tmp = av_fetch( revcharmap, ofs, 0);
2281 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2282 if ( state == 1 ) break;
2284 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2286 PerlIO_printf(Perl_debug_log,
2287 "%*sNew Start State=%"UVuf" Class: [",
2288 (int)depth * 2 + 2, "",
2291 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2292 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2294 TRIE_BITMAP_SET(trie,*ch);
2296 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2298 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2302 TRIE_BITMAP_SET(trie,*ch);
2304 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2305 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2311 SV **tmp = av_fetch( revcharmap, idx, 0);
2313 char *ch = SvPV( *tmp, len );
2315 SV *sv=sv_newmortal();
2316 PerlIO_printf( Perl_debug_log,
2317 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2318 (int)depth * 2 + 2, "",
2320 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2321 PL_colors[0], PL_colors[1],
2322 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2323 PERL_PV_ESCAPE_FIRSTCHAR
2328 OP( convert ) = nodetype;
2329 str=STRING(convert);
2332 STR_LEN(convert) += len;
2338 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2343 trie->prefixlen = (state-1);
2345 regnode *n = convert+NODE_SZ_STR(convert);
2346 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2347 trie->startstate = state;
2348 trie->minlen -= (state - 1);
2349 trie->maxlen -= (state - 1);
2351 /* At least the UNICOS C compiler choked on this
2352 * being argument to DEBUG_r(), so let's just have
2355 #ifdef PERL_EXT_RE_BUILD
2361 regnode *fix = convert;
2362 U32 word = trie->wordcount;
2364 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2365 while( ++fix < n ) {
2366 Set_Node_Offset_Length(fix, 0, 0);
2369 SV ** const tmp = av_fetch( trie_words, word, 0 );
2371 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2372 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2374 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2382 NEXT_OFF(convert) = (U16)(tail - convert);
2383 DEBUG_r(optimize= n);
2389 if ( trie->maxlen ) {
2390 NEXT_OFF( convert ) = (U16)(tail - convert);
2391 ARG_SET( convert, data_slot );
2392 /* Store the offset to the first unabsorbed branch in
2393 jump[0], which is otherwise unused by the jump logic.
2394 We use this when dumping a trie and during optimisation. */
2396 trie->jump[0] = (U16)(nextbranch - convert);
2398 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2399 * and there is a bitmap
2400 * and the first "jump target" node we found leaves enough room
2401 * then convert the TRIE node into a TRIEC node, with the bitmap
2402 * embedded inline in the opcode - this is hypothetically faster.
2404 if ( !trie->states[trie->startstate].wordnum
2406 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2408 OP( convert ) = TRIEC;
2409 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2410 PerlMemShared_free(trie->bitmap);
2413 OP( convert ) = TRIE;
2415 /* store the type in the flags */
2416 convert->flags = nodetype;
2420 + regarglen[ OP( convert ) ];
2422 /* XXX We really should free up the resource in trie now,
2423 as we won't use them - (which resources?) dmq */
2425 /* needed for dumping*/
2426 DEBUG_r(if (optimize) {
2427 regnode *opt = convert;
2429 while ( ++opt < optimize) {
2430 Set_Node_Offset_Length(opt,0,0);
2433 Try to clean up some of the debris left after the
2436 while( optimize < jumper ) {
2437 mjd_nodelen += Node_Length((optimize));
2438 OP( optimize ) = OPTIMIZED;
2439 Set_Node_Offset_Length(optimize,0,0);
2442 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2444 } /* end node insert */
2445 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, convert);
2447 /* Finish populating the prev field of the wordinfo array. Walk back
2448 * from each accept state until we find another accept state, and if
2449 * so, point the first word's .prev field at the second word. If the
2450 * second already has a .prev field set, stop now. This will be the
2451 * case either if we've already processed that word's accept state,
2452 * or that state had multiple words, and the overspill words were
2453 * already linked up earlier.
2460 for (word=1; word <= trie->wordcount; word++) {
2462 if (trie->wordinfo[word].prev)
2464 state = trie->wordinfo[word].accept;
2466 state = prev_states[state];
2469 prev = trie->states[state].wordnum;
2473 trie->wordinfo[word].prev = prev;
2475 Safefree(prev_states);
2479 /* and now dump out the compressed format */
2480 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2482 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2484 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2485 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2487 SvREFCNT_dec_NN(revcharmap);
2491 : trie->startstate>1
2497 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2499 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2501 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2502 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2505 We find the fail state for each state in the trie, this state is the longest proper
2506 suffix of the current state's 'word' that is also a proper prefix of another word in our
2507 trie. State 1 represents the word '' and is thus the default fail state. This allows
2508 the DFA not to have to restart after its tried and failed a word at a given point, it
2509 simply continues as though it had been matching the other word in the first place.
2511 'abcdgu'=~/abcdefg|cdgu/
2512 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2513 fail, which would bring us to the state representing 'd' in the second word where we would
2514 try 'g' and succeed, proceeding to match 'cdgu'.
2516 /* add a fail transition */
2517 const U32 trie_offset = ARG(source);
2518 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2520 const U32 ucharcount = trie->uniquecharcount;
2521 const U32 numstates = trie->statecount;
2522 const U32 ubound = trie->lasttrans + ucharcount;
2526 U32 base = trie->states[ 1 ].trans.base;
2529 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2530 GET_RE_DEBUG_FLAGS_DECL;
2532 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2534 PERL_UNUSED_ARG(depth);
2538 ARG_SET( stclass, data_slot );
2539 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2540 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2541 aho->trie=trie_offset;
2542 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2543 Copy( trie->states, aho->states, numstates, reg_trie_state );
2544 Newxz( q, numstates, U32);
2545 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2548 /* initialize fail[0..1] to be 1 so that we always have
2549 a valid final fail state */
2550 fail[ 0 ] = fail[ 1 ] = 1;
2552 for ( charid = 0; charid < ucharcount ; charid++ ) {
2553 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2555 q[ q_write ] = newstate;
2556 /* set to point at the root */
2557 fail[ q[ q_write++ ] ]=1;
2560 while ( q_read < q_write) {
2561 const U32 cur = q[ q_read++ % numstates ];
2562 base = trie->states[ cur ].trans.base;
2564 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2565 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2567 U32 fail_state = cur;
2570 fail_state = fail[ fail_state ];
2571 fail_base = aho->states[ fail_state ].trans.base;
2572 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2574 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2575 fail[ ch_state ] = fail_state;
2576 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2578 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2580 q[ q_write++ % numstates] = ch_state;
2584 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2585 when we fail in state 1, this allows us to use the
2586 charclass scan to find a valid start char. This is based on the principle
2587 that theres a good chance the string being searched contains lots of stuff
2588 that cant be a start char.
2590 fail[ 0 ] = fail[ 1 ] = 0;
2591 DEBUG_TRIE_COMPILE_r({
2592 PerlIO_printf(Perl_debug_log,
2593 "%*sStclass Failtable (%"UVuf" states): 0",
2594 (int)(depth * 2), "", (UV)numstates
2596 for( q_read=1; q_read<numstates; q_read++ ) {
2597 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2599 PerlIO_printf(Perl_debug_log, "\n");
2602 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2607 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2608 * These need to be revisited when a newer toolchain becomes available.
2610 #if defined(__sparc64__) && defined(__GNUC__)
2611 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2612 # undef SPARC64_GCC_WORKAROUND
2613 # define SPARC64_GCC_WORKAROUND 1
2617 #define DEBUG_PEEP(str,scan,depth) \
2618 DEBUG_OPTIMISE_r({if (scan){ \
2619 SV * const mysv=sv_newmortal(); \
2620 regnode *Next = regnext(scan); \
2621 regprop(RExC_rx, mysv, scan); \
2622 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2623 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2624 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2628 /* The below joins as many adjacent EXACTish nodes as possible into a single
2629 * one. The regop may be changed if the node(s) contain certain sequences that
2630 * require special handling. The joining is only done if:
2631 * 1) there is room in the current conglomerated node to entirely contain the
2633 * 2) they are the exact same node type
2635 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2636 * these get optimized out
2638 * If a node is to match under /i (folded), the number of characters it matches
2639 * can be different than its character length if it contains a multi-character
2640 * fold. *min_subtract is set to the total delta of the input nodes.
2642 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2643 * and contains LATIN SMALL LETTER SHARP S
2645 * This is as good a place as any to discuss the design of handling these
2646 * multi-character fold sequences. It's been wrong in Perl for a very long
2647 * time. There are three code points in Unicode whose multi-character folds
2648 * were long ago discovered to mess things up. The previous designs for
2649 * dealing with these involved assigning a special node for them. This
2650 * approach doesn't work, as evidenced by this example:
2651 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2652 * Both these fold to "sss", but if the pattern is parsed to create a node that
2653 * would match just the \xDF, it won't be able to handle the case where a
2654 * successful match would have to cross the node's boundary. The new approach
2655 * that hopefully generally solves the problem generates an EXACTFU_SS node
2658 * It turns out that there are problems with all multi-character folds, and not
2659 * just these three. Now the code is general, for all such cases, but the
2660 * three still have some special handling. The approach taken is:
2661 * 1) This routine examines each EXACTFish node that could contain multi-
2662 * character fold sequences. It returns in *min_subtract how much to
2663 * subtract from the the actual length of the string to get a real minimum
2664 * match length; it is 0 if there are no multi-char folds. This delta is
2665 * used by the caller to adjust the min length of the match, and the delta
2666 * between min and max, so that the optimizer doesn't reject these
2667 * possibilities based on size constraints.
2668 * 2) Certain of these sequences require special handling by the trie code,
2669 * so, if found, this code changes the joined node type to special ops:
2670 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2671 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2672 * is used for an EXACTFU node that contains at least one "ss" sequence in
2673 * it. For non-UTF-8 patterns and strings, this is the only case where
2674 * there is a possible fold length change. That means that a regular
2675 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2676 * with length changes, and so can be processed faster. regexec.c takes
2677 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2678 * pre-folded by regcomp.c. This saves effort in regex matching.
2679 * However, the pre-folding isn't done for non-UTF8 patterns because the
2680 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2681 * down by forcing the pattern into UTF8 unless necessary. Also what
2682 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2683 * possibilities for the non-UTF8 patterns are quite simple, except for
2684 * the sharp s. All the ones that don't involve a UTF-8 target string are
2685 * members of a fold-pair, and arrays are set up for all of them so that
2686 * the other member of the pair can be found quickly. Code elsewhere in
2687 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2688 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2689 * described in the next item.
2690 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2691 * 'ss' or not is not knowable at compile time. It will match iff the
2692 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2693 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2694 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
2695 * described in item 3). An assumption that the optimizer part of
2696 * regexec.c (probably unwittingly) makes is that a character in the
2697 * pattern corresponds to at most a single character in the target string.
2698 * (And I do mean character, and not byte here, unlike other parts of the
2699 * documentation that have never been updated to account for multibyte
2700 * Unicode.) This assumption is wrong only in this case, as all other
2701 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2702 * virtue of having this file pre-fold UTF-8 patterns. I'm
2703 * reluctant to try to change this assumption, so instead the code punts.
2704 * This routine examines EXACTF nodes for the sharp s, and returns a
2705 * boolean indicating whether or not the node is an EXACTF node that
2706 * contains a sharp s. When it is true, the caller sets a flag that later
2707 * causes the optimizer in this file to not set values for the floating
2708 * and fixed string lengths, and thus avoids the optimizer code in
2709 * regexec.c that makes the invalid assumption. Thus, there is no
2710 * optimization based on string lengths for EXACTF nodes that contain the
2711 * sharp s. This only happens for /id rules (which means the pattern
2715 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2716 if (PL_regkind[OP(scan)] == EXACT) \
2717 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2720 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) {
2721 /* Merge several consecutive EXACTish nodes into one. */
2722 regnode *n = regnext(scan);
2724 regnode *next = scan + NODE_SZ_STR(scan);
2728 regnode *stop = scan;
2729 GET_RE_DEBUG_FLAGS_DECL;
2731 PERL_UNUSED_ARG(depth);
2734 PERL_ARGS_ASSERT_JOIN_EXACT;
2735 #ifndef EXPERIMENTAL_INPLACESCAN
2736 PERL_UNUSED_ARG(flags);
2737 PERL_UNUSED_ARG(val);
2739 DEBUG_PEEP("join",scan,depth);
2741 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2742 * EXACT ones that are mergeable to the current one. */
2744 && (PL_regkind[OP(n)] == NOTHING
2745 || (stringok && OP(n) == OP(scan)))
2747 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2750 if (OP(n) == TAIL || n > next)
2752 if (PL_regkind[OP(n)] == NOTHING) {
2753 DEBUG_PEEP("skip:",n,depth);
2754 NEXT_OFF(scan) += NEXT_OFF(n);
2755 next = n + NODE_STEP_REGNODE;
2762 else if (stringok) {
2763 const unsigned int oldl = STR_LEN(scan);
2764 regnode * const nnext = regnext(n);
2766 /* XXX I (khw) kind of doubt that this works on platforms where
2767 * U8_MAX is above 255 because of lots of other assumptions */
2768 /* Don't join if the sum can't fit into a single node */
2769 if (oldl + STR_LEN(n) > U8_MAX)
2772 DEBUG_PEEP("merg",n,depth);
2775 NEXT_OFF(scan) += NEXT_OFF(n);
2776 STR_LEN(scan) += STR_LEN(n);
2777 next = n + NODE_SZ_STR(n);
2778 /* Now we can overwrite *n : */
2779 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2787 #ifdef EXPERIMENTAL_INPLACESCAN
2788 if (flags && !NEXT_OFF(n)) {
2789 DEBUG_PEEP("atch", val, depth);
2790 if (reg_off_by_arg[OP(n)]) {
2791 ARG_SET(n, val - n);
2794 NEXT_OFF(n) = val - n;
2802 *has_exactf_sharp_s = FALSE;
2804 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2805 * can now analyze for sequences of problematic code points. (Prior to
2806 * this final joining, sequences could have been split over boundaries, and
2807 * hence missed). The sequences only happen in folding, hence for any
2808 * non-EXACT EXACTish node */
2809 if (OP(scan) != EXACT) {
2810 const U8 * const s0 = (U8*) STRING(scan);
2812 const U8 * const s_end = s0 + STR_LEN(scan);
2814 /* One pass is made over the node's string looking for all the
2815 * possibilities. to avoid some tests in the loop, there are two main
2816 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2820 /* Examine the string for a multi-character fold sequence. UTF-8
2821 * patterns have all characters pre-folded by the time this code is
2823 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2824 length sequence we are looking for is 2 */
2827 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2828 if (! len) { /* Not a multi-char fold: get next char */
2833 /* Nodes with 'ss' require special handling, except for EXACTFL
2834 * and EXACTFA for which there is no multi-char fold to this */
2835 if (len == 2 && *s == 's' && *(s+1) == 's'
2836 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2839 OP(scan) = EXACTFU_SS;
2842 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2843 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2844 COMBINING_DIAERESIS_UTF8
2845 COMBINING_ACUTE_ACCENT_UTF8,
2847 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2848 COMBINING_DIAERESIS_UTF8
2849 COMBINING_ACUTE_ACCENT_UTF8,
2854 /* These two folds require special handling by trie's, so
2855 * change the node type to indicate this. If EXACTFA and
2856 * EXACTFL were ever to be handled by trie's, this would
2857 * have to be changed. If this node has already been
2858 * changed to EXACTFU_SS in this loop, leave it as is. (I
2859 * (khw) think it doesn't matter in regexec.c for UTF
2860 * patterns, but no need to change it */
2861 if (OP(scan) == EXACTFU) {
2862 OP(scan) = EXACTFU_TRICKYFOLD;
2866 else { /* Here is a generic multi-char fold. */
2867 const U8* multi_end = s + len;
2869 /* Count how many characters in it. In the case of /l and
2870 * /aa, no folds which contain ASCII code points are
2871 * allowed, so check for those, and skip if found. (In
2872 * EXACTFL, no folds are allowed to any Latin1 code point,
2873 * not just ASCII. But there aren't any of these
2874 * currently, nor ever likely, so don't take the time to
2875 * test for them. The code that generates the
2876 * is_MULTI_foo() macros croaks should one actually get put
2877 * into Unicode .) */
2878 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2879 count = utf8_length(s, multi_end);
2883 while (s < multi_end) {
2886 goto next_iteration;
2896 /* The delta is how long the sequence is minus 1 (1 is how long
2897 * the character that folds to the sequence is) */
2898 *min_subtract += count - 1;
2902 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2904 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2905 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2906 * nodes can't have multi-char folds to this range (and there are
2907 * no existing ones in the upper latin1 range). In the EXACTF
2908 * case we look also for the sharp s, which can be in the final
2909 * position. Otherwise we can stop looking 1 byte earlier because
2910 * have to find at least two characters for a multi-fold */
2911 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2913 /* The below is perhaps overboard, but this allows us to save a
2914 * test each time through the loop at the expense of a mask. This
2915 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2916 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2917 * are 64. This uses an exclusive 'or' to find that bit and then
2918 * inverts it to form a mask, with just a single 0, in the bit
2919 * position where 'S' and 's' differ. */
2920 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2921 const U8 s_masked = 's' & S_or_s_mask;
2924 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2925 if (! len) { /* Not a multi-char fold. */
2926 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2928 *has_exactf_sharp_s = TRUE;
2935 && ((*s & S_or_s_mask) == s_masked)
2936 && ((*(s+1) & S_or_s_mask) == s_masked))
2939 /* EXACTF nodes need to know that the minimum length
2940 * changed so that a sharp s in the string can match this
2941 * ss in the pattern, but they remain EXACTF nodes, as they
2942 * won't match this unless the target string is is UTF-8,
2943 * which we don't know until runtime */
2944 if (OP(scan) != EXACTF) {
2945 OP(scan) = EXACTFU_SS;
2949 *min_subtract += len - 1;
2956 /* Allow dumping but overwriting the collection of skipped
2957 * ops and/or strings with fake optimized ops */
2958 n = scan + NODE_SZ_STR(scan);
2966 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2970 /* REx optimizer. Converts nodes into quicker variants "in place".
2971 Finds fixed substrings. */
2973 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2974 to the position after last scanned or to NULL. */
2976 #define INIT_AND_WITHP \
2977 assert(!and_withp); \
2978 Newx(and_withp,1,struct regnode_charclass_class); \
2979 SAVEFREEPV(and_withp)
2981 /* this is a chain of data about sub patterns we are processing that
2982 need to be handled separately/specially in study_chunk. Its so
2983 we can simulate recursion without losing state. */
2985 typedef struct scan_frame {
2986 regnode *last; /* last node to process in this frame */
2987 regnode *next; /* next node to process when last is reached */
2988 struct scan_frame *prev; /*previous frame*/
2989 I32 stop; /* what stopparen do we use */
2993 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2996 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2997 I32 *minlenp, I32 *deltap,
3002 struct regnode_charclass_class *and_withp,
3003 U32 flags, U32 depth)
3004 /* scanp: Start here (read-write). */
3005 /* deltap: Write maxlen-minlen here. */
3006 /* last: Stop before this one. */
3007 /* data: string data about the pattern */
3008 /* stopparen: treat close N as END */
3009 /* recursed: which subroutines have we recursed into */
3010 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3013 I32 min = 0; /* There must be at least this number of characters to match */
3015 regnode *scan = *scanp, *next;
3017 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3018 int is_inf_internal = 0; /* The studied chunk is infinite */
3019 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3020 scan_data_t data_fake;
3021 SV *re_trie_maxbuff = NULL;
3022 regnode *first_non_open = scan;
3023 I32 stopmin = I32_MAX;
3024 scan_frame *frame = NULL;
3025 GET_RE_DEBUG_FLAGS_DECL;
3027 PERL_ARGS_ASSERT_STUDY_CHUNK;
3030 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3034 while (first_non_open && OP(first_non_open) == OPEN)
3035 first_non_open=regnext(first_non_open);
3040 while ( scan && OP(scan) != END && scan < last ){
3041 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3042 node length to get a real minimum (because
3043 the folded version may be shorter) */
3044 bool has_exactf_sharp_s = FALSE;
3045 /* Peephole optimizer: */
3046 DEBUG_STUDYDATA("Peep:", data,depth);
3047 DEBUG_PEEP("Peep",scan,depth);
3049 /* Its not clear to khw or hv why this is done here, and not in the
3050 * clauses that deal with EXACT nodes. khw's guess is that it's
3051 * because of a previous design */
3052 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3054 /* Follow the next-chain of the current node and optimize
3055 away all the NOTHINGs from it. */
3056 if (OP(scan) != CURLYX) {
3057 const int max = (reg_off_by_arg[OP(scan)]
3059 /* I32 may be smaller than U16 on CRAYs! */
3060 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3061 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3065 /* Skip NOTHING and LONGJMP. */
3066 while ((n = regnext(n))
3067 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3068 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3069 && off + noff < max)
3071 if (reg_off_by_arg[OP(scan)])
3074 NEXT_OFF(scan) = off;
3079 /* The principal pseudo-switch. Cannot be a switch, since we
3080 look into several different things. */
3081 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3082 || OP(scan) == IFTHEN) {
3083 next = regnext(scan);
3085 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3087 if (OP(next) == code || code == IFTHEN) {
3088 /* NOTE - There is similar code to this block below for handling
3089 TRIE nodes on a re-study. If you change stuff here check there
3091 I32 max1 = 0, min1 = I32_MAX, num = 0;
3092 struct regnode_charclass_class accum;
3093 regnode * const startbranch=scan;
3095 if (flags & SCF_DO_SUBSTR)
3096 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3097 if (flags & SCF_DO_STCLASS)
3098 cl_init_zero(pRExC_state, &accum);
3100 while (OP(scan) == code) {
3101 I32 deltanext, minnext, f = 0, fake;
3102 struct regnode_charclass_class this_class;
3105 data_fake.flags = 0;
3107 data_fake.whilem_c = data->whilem_c;
3108 data_fake.last_closep = data->last_closep;
3111 data_fake.last_closep = &fake;
3113 data_fake.pos_delta = delta;
3114 next = regnext(scan);
3115 scan = NEXTOPER(scan);
3117 scan = NEXTOPER(scan);
3118 if (flags & SCF_DO_STCLASS) {
3119 cl_init(pRExC_state, &this_class);
3120 data_fake.start_class = &this_class;
3121 f = SCF_DO_STCLASS_AND;
3123 if (flags & SCF_WHILEM_VISITED_POS)
3124 f |= SCF_WHILEM_VISITED_POS;
3126 /* we suppose the run is continuous, last=next...*/
3127 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3129 stopparen, recursed, NULL, f,depth+1);
3132 if (deltanext == I32_MAX) {
3133 is_inf = is_inf_internal = 1;
3135 } else if (max1 < minnext + deltanext)
3136 max1 = minnext + deltanext;
3138 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3140 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3141 if ( stopmin > minnext)
3142 stopmin = min + min1;
3143 flags &= ~SCF_DO_SUBSTR;
3145 data->flags |= SCF_SEEN_ACCEPT;
3148 if (data_fake.flags & SF_HAS_EVAL)
3149 data->flags |= SF_HAS_EVAL;
3150 data->whilem_c = data_fake.whilem_c;
3152 if (flags & SCF_DO_STCLASS)
3153 cl_or(pRExC_state, &accum, &this_class);
3155 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3157 if (flags & SCF_DO_SUBSTR) {
3158 data->pos_min += min1;
3159 if (data->pos_delta >= I32_MAX - (max1 - min1))
3160 data->pos_delta = I32_MAX;
3162 data->pos_delta += max1 - min1;
3163 if (max1 != min1 || is_inf)
3164 data->longest = &(data->longest_float);
3167 if (delta == I32_MAX || I32_MAX - delta - (max1 - min1) < 0)
3170 delta += max1 - min1;
3171 if (flags & SCF_DO_STCLASS_OR) {
3172 cl_or(pRExC_state, data->start_class, &accum);
3174 cl_and(data->start_class, and_withp);
3175 flags &= ~SCF_DO_STCLASS;
3178 else if (flags & SCF_DO_STCLASS_AND) {
3180 cl_and(data->start_class, &accum);
3181 flags &= ~SCF_DO_STCLASS;
3184 /* Switch to OR mode: cache the old value of
3185 * data->start_class */
3187 StructCopy(data->start_class, and_withp,
3188 struct regnode_charclass_class);
3189 flags &= ~SCF_DO_STCLASS_AND;
3190 StructCopy(&accum, data->start_class,
3191 struct regnode_charclass_class);
3192 flags |= SCF_DO_STCLASS_OR;
3193 SET_SSC_EOS(data->start_class);
3197 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3200 Assuming this was/is a branch we are dealing with: 'scan' now
3201 points at the item that follows the branch sequence, whatever
3202 it is. We now start at the beginning of the sequence and look
3209 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3211 If we can find such a subsequence we need to turn the first
3212 element into a trie and then add the subsequent branch exact
3213 strings to the trie.
3217 1. patterns where the whole set of branches can be converted.
3219 2. patterns where only a subset can be converted.
3221 In case 1 we can replace the whole set with a single regop
3222 for the trie. In case 2 we need to keep the start and end
3225 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3226 becomes BRANCH TRIE; BRANCH X;
3228 There is an additional case, that being where there is a
3229 common prefix, which gets split out into an EXACT like node
3230 preceding the TRIE node.
3232 If x(1..n)==tail then we can do a simple trie, if not we make
3233 a "jump" trie, such that when we match the appropriate word
3234 we "jump" to the appropriate tail node. Essentially we turn
3235 a nested if into a case structure of sorts.
3240 if (!re_trie_maxbuff) {
3241 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3242 if (!SvIOK(re_trie_maxbuff))
3243 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3245 if ( SvIV(re_trie_maxbuff)>=0 ) {
3247 regnode *first = (regnode *)NULL;
3248 regnode *last = (regnode *)NULL;
3249 regnode *tail = scan;
3254 SV * const mysv = sv_newmortal(); /* for dumping */
3256 /* var tail is used because there may be a TAIL
3257 regop in the way. Ie, the exacts will point to the
3258 thing following the TAIL, but the last branch will
3259 point at the TAIL. So we advance tail. If we
3260 have nested (?:) we may have to move through several
3264 while ( OP( tail ) == TAIL ) {
3265 /* this is the TAIL generated by (?:) */
3266 tail = regnext( tail );
3270 DEBUG_TRIE_COMPILE_r({
3271 regprop(RExC_rx, mysv, tail );
3272 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3273 (int)depth * 2 + 2, "",
3274 "Looking for TRIE'able sequences. Tail node is: ",
3275 SvPV_nolen_const( mysv )
3281 Step through the branches
3282 cur represents each branch,
3283 noper is the first thing to be matched as part of that branch
3284 noper_next is the regnext() of that node.
3286 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3287 via a "jump trie" but we also support building with NOJUMPTRIE,
3288 which restricts the trie logic to structures like /FOO|BAR/.
3290 If noper is a trieable nodetype then the branch is a possible optimization
3291 target. If we are building under NOJUMPTRIE then we require that noper_next
3292 is the same as scan (our current position in the regex program).
3294 Once we have two or more consecutive such branches we can create a
3295 trie of the EXACT's contents and stitch it in place into the program.
3297 If the sequence represents all of the branches in the alternation we
3298 replace the entire thing with a single TRIE node.
3300 Otherwise when it is a subsequence we need to stitch it in place and
3301 replace only the relevant branches. This means the first branch has
3302 to remain as it is used by the alternation logic, and its next pointer,
3303 and needs to be repointed at the item on the branch chain following
3304 the last branch we have optimized away.
3306 This could be either a BRANCH, in which case the subsequence is internal,
3307 or it could be the item following the branch sequence in which case the
3308 subsequence is at the end (which does not necessarily mean the first node
3309 is the start of the alternation).
3311 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3314 ----------------+-----------
3318 EXACTFU_SS | EXACTFU
3319 EXACTFU_TRICKYFOLD | EXACTFU
3324 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3325 ( EXACT == (X) ) ? EXACT : \
3326 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3329 /* dont use tail as the end marker for this traverse */
3330 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3331 regnode * const noper = NEXTOPER( cur );
3332 U8 noper_type = OP( noper );
3333 U8 noper_trietype = TRIE_TYPE( noper_type );
3334 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3335 regnode * const noper_next = regnext( noper );
3336 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3337 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3340 DEBUG_TRIE_COMPILE_r({
3341 regprop(RExC_rx, mysv, cur);
3342 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3343 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3345 regprop(RExC_rx, mysv, noper);
3346 PerlIO_printf( Perl_debug_log, " -> %s",
3347 SvPV_nolen_const(mysv));
3350 regprop(RExC_rx, mysv, noper_next );
3351 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3352 SvPV_nolen_const(mysv));
3354 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3355 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3356 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3360 /* Is noper a trieable nodetype that can be merged with the
3361 * current trie (if there is one)? */
3365 ( noper_trietype == NOTHING)
3366 || ( trietype == NOTHING )
3367 || ( trietype == noper_trietype )
3370 && noper_next == tail
3374 /* Handle mergable triable node
3375 * Either we are the first node in a new trieable sequence,
3376 * in which case we do some bookkeeping, otherwise we update
3377 * the end pointer. */
3380 if ( noper_trietype == NOTHING ) {
3381 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3382 regnode * const noper_next = regnext( noper );
3383 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3384 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3387 if ( noper_next_trietype ) {
3388 trietype = noper_next_trietype;
3389 } else if (noper_next_type) {
3390 /* a NOTHING regop is 1 regop wide. We need at least two
3391 * for a trie so we can't merge this in */
3395 trietype = noper_trietype;
3398 if ( trietype == NOTHING )
3399 trietype = noper_trietype;
3404 } /* end handle mergable triable node */
3406 /* handle unmergable node -
3407 * noper may either be a triable node which can not be tried
3408 * together with the current trie, or a non triable node */
3410 /* If last is set and trietype is not NOTHING then we have found
3411 * at least two triable branch sequences in a row of a similar
3412 * trietype so we can turn them into a trie. If/when we
3413 * allow NOTHING to start a trie sequence this condition will be
3414 * required, and it isn't expensive so we leave it in for now. */
3415 if ( trietype && trietype != NOTHING )
3416 make_trie( pRExC_state,
3417 startbranch, first, cur, tail, count,
3418 trietype, depth+1 );
3419 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3423 && noper_next == tail
3426 /* noper is triable, so we can start a new trie sequence */
3429 trietype = noper_trietype;
3431 /* if we already saw a first but the current node is not triable then we have
3432 * to reset the first information. */
3437 } /* end handle unmergable node */
3438 } /* loop over branches */
3439 DEBUG_TRIE_COMPILE_r({
3440 regprop(RExC_rx, mysv, cur);
3441 PerlIO_printf( Perl_debug_log,
3442 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3443 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3446 if ( last && trietype ) {
3447 if ( trietype != NOTHING ) {
3448 /* the last branch of the sequence was part of a trie,
3449 * so we have to construct it here outside of the loop
3451 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3452 #ifdef TRIE_STUDY_OPT
3453 if ( ((made == MADE_EXACT_TRIE &&
3454 startbranch == first)
3455 || ( first_non_open == first )) &&
3457 flags |= SCF_TRIE_RESTUDY;
3458 if ( startbranch == first
3461 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3466 /* at this point we know whatever we have is a NOTHING sequence/branch
3467 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3469 if ( startbranch == first ) {
3471 /* the entire thing is a NOTHING sequence, something like this:
3472 * (?:|) So we can turn it into a plain NOTHING op. */
3473 DEBUG_TRIE_COMPILE_r({
3474 regprop(RExC_rx, mysv, cur);
3475 PerlIO_printf( Perl_debug_log,
3476 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3477 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3480 OP(startbranch)= NOTHING;
3481 NEXT_OFF(startbranch)= tail - startbranch;
3482 for ( opt= startbranch + 1; opt < tail ; opt++ )
3486 } /* end if ( last) */
3487 } /* TRIE_MAXBUF is non zero */
3492 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3493 scan = NEXTOPER(NEXTOPER(scan));
3494 } else /* single branch is optimized. */
3495 scan = NEXTOPER(scan);
3497 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3498 scan_frame *newframe = NULL;
3503 if (OP(scan) != SUSPEND) {
3504 /* set the pointer */
3505 if (OP(scan) == GOSUB) {
3507 RExC_recurse[ARG2L(scan)] = scan;
3508 start = RExC_open_parens[paren-1];
3509 end = RExC_close_parens[paren-1];
3512 start = RExC_rxi->program + 1;
3516 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3517 SAVEFREEPV(recursed);
3519 if (!PAREN_TEST(recursed,paren+1)) {
3520 PAREN_SET(recursed,paren+1);
3521 Newx(newframe,1,scan_frame);
3523 if (flags & SCF_DO_SUBSTR) {
3524 SCAN_COMMIT(pRExC_state,data,minlenp);
3525 data->longest = &(data->longest_float);
3527 is_inf = is_inf_internal = 1;
3528 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3529 cl_anything(pRExC_state, data->start_class);
3530 flags &= ~SCF_DO_STCLASS;
3533 Newx(newframe,1,scan_frame);
3536 end = regnext(scan);
3541 SAVEFREEPV(newframe);
3542 newframe->next = regnext(scan);
3543 newframe->last = last;
3544 newframe->stop = stopparen;
3545 newframe->prev = frame;
3555 else if (OP(scan) == EXACT) {
3556 I32 l = STR_LEN(scan);
3559 const U8 * const s = (U8*)STRING(scan);
3560 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3561 l = utf8_length(s, s + l);
3563 uc = *((U8*)STRING(scan));
3566 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3567 /* The code below prefers earlier match for fixed
3568 offset, later match for variable offset. */
3569 if (data->last_end == -1) { /* Update the start info. */
3570 data->last_start_min = data->pos_min;
3571 data->last_start_max = is_inf
3572 ? I32_MAX : data->pos_min + data->pos_delta;
3574 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3576 SvUTF8_on(data->last_found);
3578 SV * const sv = data->last_found;
3579 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3580 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3581 if (mg && mg->mg_len >= 0)
3582 mg->mg_len += utf8_length((U8*)STRING(scan),
3583 (U8*)STRING(scan)+STR_LEN(scan));
3585 data->last_end = data->pos_min + l;
3586 data->pos_min += l; /* As in the first entry. */
3587 data->flags &= ~SF_BEFORE_EOL;
3589 if (flags & SCF_DO_STCLASS_AND) {
3590 /* Check whether it is compatible with what we know already! */
3594 /* If compatible, we or it in below. It is compatible if is
3595 * in the bitmp and either 1) its bit or its fold is set, or 2)
3596 * it's for a locale. Even if there isn't unicode semantics
3597 * here, at runtime there may be because of matching against a
3598 * utf8 string, so accept a possible false positive for
3599 * latin1-range folds */
3601 (!(data->start_class->flags & ANYOF_LOCALE)
3602 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3603 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3604 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3609 ANYOF_CLASS_ZERO(data->start_class);
3610 ANYOF_BITMAP_ZERO(data->start_class);
3612 ANYOF_BITMAP_SET(data->start_class, uc);
3613 else if (uc >= 0x100) {
3616 /* Some Unicode code points fold to the Latin1 range; as
3617 * XXX temporary code, instead of figuring out if this is
3618 * one, just assume it is and set all the start class bits
3619 * that could be some such above 255 code point's fold
3620 * which will generate fals positives. As the code
3621 * elsewhere that does compute the fold settles down, it
3622 * can be extracted out and re-used here */
3623 for (i = 0; i < 256; i++){
3624 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3625 ANYOF_BITMAP_SET(data->start_class, i);
3629 CLEAR_SSC_EOS(data->start_class);
3631 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3633 else if (flags & SCF_DO_STCLASS_OR) {
3634 /* false positive possible if the class is case-folded */
3636 ANYOF_BITMAP_SET(data->start_class, uc);
3638 data->start_class->flags |= ANYOF_UNICODE_ALL;
3639 CLEAR_SSC_EOS(data->start_class);
3640 cl_and(data->start_class, and_withp);
3642 flags &= ~SCF_DO_STCLASS;
3644 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3645 I32 l = STR_LEN(scan);
3646 UV uc = *((U8*)STRING(scan));
3648 /* Search for fixed substrings supports EXACT only. */
3649 if (flags & SCF_DO_SUBSTR) {
3651 SCAN_COMMIT(pRExC_state, data, minlenp);
3654 const U8 * const s = (U8 *)STRING(scan);
3655 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3656 l = utf8_length(s, s + l);
3658 if (has_exactf_sharp_s) {
3659 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3661 min += l - min_subtract;
3663 delta += min_subtract;
3664 if (flags & SCF_DO_SUBSTR) {
3665 data->pos_min += l - min_subtract;
3666 if (data->pos_min < 0) {
3669 data->pos_delta += min_subtract;
3671 data->longest = &(data->longest_float);
3674 if (flags & SCF_DO_STCLASS_AND) {
3675 /* Check whether it is compatible with what we know already! */
3678 (!(data->start_class->flags & ANYOF_LOCALE)
3679 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3680 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3684 ANYOF_CLASS_ZERO(data->start_class);
3685 ANYOF_BITMAP_ZERO(data->start_class);
3687 ANYOF_BITMAP_SET(data->start_class, uc);
3688 CLEAR_SSC_EOS(data->start_class);
3689 if (OP(scan) == EXACTFL) {
3690 /* XXX This set is probably no longer necessary, and
3691 * probably wrong as LOCALE now is on in the initial
3693 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3697 /* Also set the other member of the fold pair. In case
3698 * that unicode semantics is called for at runtime, use
3699 * the full latin1 fold. (Can't do this for locale,
3700 * because not known until runtime) */
3701 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3703 /* All other (EXACTFL handled above) folds except under
3704 * /iaa that include s, S, and sharp_s also may include
3706 if (OP(scan) != EXACTFA) {
3707 if (uc == 's' || uc == 'S') {
3708 ANYOF_BITMAP_SET(data->start_class,
3709 LATIN_SMALL_LETTER_SHARP_S);
3711 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3712 ANYOF_BITMAP_SET(data->start_class, 's');
3713 ANYOF_BITMAP_SET(data->start_class, 'S');
3718 else if (uc >= 0x100) {
3720 for (i = 0; i < 256; i++){
3721 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3722 ANYOF_BITMAP_SET(data->start_class, i);
3727 else if (flags & SCF_DO_STCLASS_OR) {
3728 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3729 /* false positive possible if the class is case-folded.
3730 Assume that the locale settings are the same... */
3732 ANYOF_BITMAP_SET(data->start_class, uc);
3733 if (OP(scan) != EXACTFL) {
3735 /* And set the other member of the fold pair, but
3736 * can't do that in locale because not known until
3738 ANYOF_BITMAP_SET(data->start_class,
3739 PL_fold_latin1[uc]);
3741 /* All folds except under /iaa that include s, S,
3742 * and sharp_s also may include the others */
3743 if (OP(scan) != EXACTFA) {
3744 if (uc == 's' || uc == 'S') {
3745 ANYOF_BITMAP_SET(data->start_class,
3746 LATIN_SMALL_LETTER_SHARP_S);
3748 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3749 ANYOF_BITMAP_SET(data->start_class, 's');
3750 ANYOF_BITMAP_SET(data->start_class, 'S');
3755 CLEAR_SSC_EOS(data->start_class);
3757 cl_and(data->start_class, and_withp);
3759 flags &= ~SCF_DO_STCLASS;
3761 else if (REGNODE_VARIES(OP(scan))) {
3762 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3763 I32 f = flags, pos_before = 0;
3764 regnode * const oscan = scan;
3765 struct regnode_charclass_class this_class;
3766 struct regnode_charclass_class *oclass = NULL;
3767 I32 next_is_eval = 0;
3769 switch (PL_regkind[OP(scan)]) {
3770 case WHILEM: /* End of (?:...)* . */
3771 scan = NEXTOPER(scan);
3774 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3775 next = NEXTOPER(scan);
3776 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3778 maxcount = REG_INFTY;
3779 next = regnext(scan);
3780 scan = NEXTOPER(scan);
3784 if (flags & SCF_DO_SUBSTR)
3789 if (flags & SCF_DO_STCLASS) {
3791 maxcount = REG_INFTY;
3792 next = regnext(scan);
3793 scan = NEXTOPER(scan);
3796 is_inf = is_inf_internal = 1;
3797 scan = regnext(scan);
3798 if (flags & SCF_DO_SUBSTR) {
3799 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3800 data->longest = &(data->longest_float);
3802 goto optimize_curly_tail;
3804 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3805 && (scan->flags == stopparen))
3810 mincount = ARG1(scan);
3811 maxcount = ARG2(scan);
3813 next = regnext(scan);
3814 if (OP(scan) == CURLYX) {
3815 I32 lp = (data ? *(data->last_closep) : 0);
3816 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3818 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3819 next_is_eval = (OP(scan) == EVAL);
3821 if (flags & SCF_DO_SUBSTR) {
3822 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3823 pos_before = data->pos_min;
3827 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3829 data->flags |= SF_IS_INF;
3831 if (flags & SCF_DO_STCLASS) {
3832 cl_init(pRExC_state, &this_class);
3833 oclass = data->start_class;
3834 data->start_class = &this_class;
3835 f |= SCF_DO_STCLASS_AND;
3836 f &= ~SCF_DO_STCLASS_OR;
3838 /* Exclude from super-linear cache processing any {n,m}
3839 regops for which the combination of input pos and regex
3840 pos is not enough information to determine if a match
3843 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3844 regex pos at the \s*, the prospects for a match depend not
3845 only on the input position but also on how many (bar\s*)
3846 repeats into the {4,8} we are. */
3847 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3848 f &= ~SCF_WHILEM_VISITED_POS;
3850 /* This will finish on WHILEM, setting scan, or on NULL: */
3851 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3852 last, data, stopparen, recursed, NULL,
3854 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3856 if (flags & SCF_DO_STCLASS)
3857 data->start_class = oclass;
3858 if (mincount == 0 || minnext == 0) {
3859 if (flags & SCF_DO_STCLASS_OR) {
3860 cl_or(pRExC_state, data->start_class, &this_class);
3862 else if (flags & SCF_DO_STCLASS_AND) {
3863 /* Switch to OR mode: cache the old value of
3864 * data->start_class */
3866 StructCopy(data->start_class, and_withp,
3867 struct regnode_charclass_class);
3868 flags &= ~SCF_DO_STCLASS_AND;
3869 StructCopy(&this_class, data->start_class,
3870 struct regnode_charclass_class);
3871 flags |= SCF_DO_STCLASS_OR;
3872 SET_SSC_EOS(data->start_class);
3874 } else { /* Non-zero len */
3875 if (flags & SCF_DO_STCLASS_OR) {
3876 cl_or(pRExC_state, data->start_class, &this_class);
3877 cl_and(data->start_class, and_withp);
3879 else if (flags & SCF_DO_STCLASS_AND)
3880 cl_and(data->start_class, &this_class);
3881 flags &= ~SCF_DO_STCLASS;
3883 if (!scan) /* It was not CURLYX, but CURLY. */
3885 if ( /* ? quantifier ok, except for (?{ ... }) */
3886 (next_is_eval || !(mincount == 0 && maxcount == 1))
3887 && (minnext == 0) && (deltanext == 0)
3888 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3889 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3891 /* Fatal warnings may leak the regexp without this: */
3892 SAVEFREESV(RExC_rx_sv);
3893 ckWARNreg(RExC_parse,
3894 "Quantifier unexpected on zero-length expression");
3895 (void)ReREFCNT_inc(RExC_rx_sv);
3898 min += minnext * mincount;
3899 is_inf_internal |= deltanext == I32_MAX
3900 || (maxcount == REG_INFTY && minnext + deltanext > 0);
3901 is_inf |= is_inf_internal;
3905 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3907 /* Try powerful optimization CURLYX => CURLYN. */
3908 if ( OP(oscan) == CURLYX && data
3909 && data->flags & SF_IN_PAR
3910 && !(data->flags & SF_HAS_EVAL)
3911 && !deltanext && minnext == 1 ) {
3912 /* Try to optimize to CURLYN. */
3913 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3914 regnode * const nxt1 = nxt;
3921 if (!REGNODE_SIMPLE(OP(nxt))
3922 && !(PL_regkind[OP(nxt)] == EXACT
3923 && STR_LEN(nxt) == 1))
3929 if (OP(nxt) != CLOSE)
3931 if (RExC_open_parens) {
3932 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3933 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3935 /* Now we know that nxt2 is the only contents: */
3936 oscan->flags = (U8)ARG(nxt);
3938 OP(nxt1) = NOTHING; /* was OPEN. */
3941 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3942 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3943 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3944 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3945 OP(nxt + 1) = OPTIMIZED; /* was count. */
3946 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3951 /* Try optimization CURLYX => CURLYM. */
3952 if ( OP(oscan) == CURLYX && data
3953 && !(data->flags & SF_HAS_PAR)
3954 && !(data->flags & SF_HAS_EVAL)
3955 && !deltanext /* atom is fixed width */
3956 && minnext != 0 /* CURLYM can't handle zero width */
3957 && ! (RExC_seen & REG_SEEN_EXACTF_SHARP_S) /* Nor \xDF */
3959 /* XXXX How to optimize if data == 0? */
3960 /* Optimize to a simpler form. */
3961 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3965 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3966 && (OP(nxt2) != WHILEM))
3968 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3969 /* Need to optimize away parenths. */
3970 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
3971 /* Set the parenth number. */
3972 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3974 oscan->flags = (U8)ARG(nxt);
3975 if (RExC_open_parens) {
3976 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3977 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3979 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3980 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3983 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3984 OP(nxt + 1) = OPTIMIZED; /* was count. */
3985 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
3986 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
3989 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3990 regnode *nnxt = regnext(nxt1);
3992 if (reg_off_by_arg[OP(nxt1)])
3993 ARG_SET(nxt1, nxt2 - nxt1);
3994 else if (nxt2 - nxt1 < U16_MAX)
3995 NEXT_OFF(nxt1) = nxt2 - nxt1;
3997 OP(nxt) = NOTHING; /* Cannot beautify */
4002 /* Optimize again: */
4003 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4004 NULL, stopparen, recursed, NULL, 0,depth+1);
4009 else if ((OP(oscan) == CURLYX)
4010 && (flags & SCF_WHILEM_VISITED_POS)
4011 /* See the comment on a similar expression above.
4012 However, this time it's not a subexpression
4013 we care about, but the expression itself. */
4014 && (maxcount == REG_INFTY)
4015 && data && ++data->whilem_c < 16) {
4016 /* This stays as CURLYX, we can put the count/of pair. */
4017 /* Find WHILEM (as in regexec.c) */
4018 regnode *nxt = oscan + NEXT_OFF(oscan);
4020 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4022 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4023 | (RExC_whilem_seen << 4)); /* On WHILEM */
4025 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4027 if (flags & SCF_DO_SUBSTR) {
4028 SV *last_str = NULL;
4029 int counted = mincount != 0;
4031 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
4032 #if defined(SPARC64_GCC_WORKAROUND)
4035 const char *s = NULL;
4038 if (pos_before >= data->last_start_min)
4041 b = data->last_start_min;
4044 s = SvPV_const(data->last_found, l);
4045 old = b - data->last_start_min;
4048 I32 b = pos_before >= data->last_start_min
4049 ? pos_before : data->last_start_min;
4051 const char * const s = SvPV_const(data->last_found, l);
4052 I32 old = b - data->last_start_min;
4056 old = utf8_hop((U8*)s, old) - (U8*)s;
4058 /* Get the added string: */
4059 last_str = newSVpvn_utf8(s + old, l, UTF);
4060 if (deltanext == 0 && pos_before == b) {
4061 /* What was added is a constant string */
4063 SvGROW(last_str, (mincount * l) + 1);
4064 repeatcpy(SvPVX(last_str) + l,
4065 SvPVX_const(last_str), l, mincount - 1);
4066 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4067 /* Add additional parts. */
4068 SvCUR_set(data->last_found,
4069 SvCUR(data->last_found) - l);
4070 sv_catsv(data->last_found, last_str);
4072 SV * sv = data->last_found;
4074 SvUTF8(sv) && SvMAGICAL(sv) ?
4075 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4076 if (mg && mg->mg_len >= 0)
4077 mg->mg_len += CHR_SVLEN(last_str) - l;
4079 data->last_end += l * (mincount - 1);
4082 /* start offset must point into the last copy */
4083 data->last_start_min += minnext * (mincount - 1);
4084 data->last_start_max += is_inf ? I32_MAX
4085 : (maxcount - 1) * (minnext + data->pos_delta);
4088 /* It is counted once already... */
4089 data->pos_min += minnext * (mincount - counted);
4091 PerlIO_printf(Perl_debug_log, "counted=%d deltanext=%d I32_MAX=%d minnext=%d maxcount=%d mincount=%d\n",
4092 counted, deltanext, I32_MAX, minnext, maxcount, mincount);
4093 if (deltanext != I32_MAX)
4094 PerlIO_printf(Perl_debug_log, "LHS=%d RHS=%d\n", -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount, I32_MAX - data->pos_delta);
4096 if (deltanext == I32_MAX || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= I32_MAX - data->pos_delta)
4097 data->pos_delta = I32_MAX;
4099 data->pos_delta += - counted * deltanext +
4100 (minnext + deltanext) * maxcount - minnext * mincount;
4101 if (mincount != maxcount) {
4102 /* Cannot extend fixed substrings found inside
4104 SCAN_COMMIT(pRExC_state,data,minlenp);
4105 if (mincount && last_str) {
4106 SV * const sv = data->last_found;
4107 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4108 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4112 sv_setsv(sv, last_str);
4113 data->last_end = data->pos_min;
4114 data->last_start_min =
4115 data->pos_min - CHR_SVLEN(last_str);
4116 data->last_start_max = is_inf
4118 : data->pos_min + data->pos_delta
4119 - CHR_SVLEN(last_str);
4121 data->longest = &(data->longest_float);
4123 SvREFCNT_dec(last_str);
4125 if (data && (fl & SF_HAS_EVAL))
4126 data->flags |= SF_HAS_EVAL;
4127 optimize_curly_tail:
4128 if (OP(oscan) != CURLYX) {
4129 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4131 NEXT_OFF(oscan) += NEXT_OFF(next);
4134 default: /* REF, and CLUMP only? */
4135 if (flags & SCF_DO_SUBSTR) {
4136 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4137 data->longest = &(data->longest_float);
4139 is_inf = is_inf_internal = 1;
4140 if (flags & SCF_DO_STCLASS_OR)
4141 cl_anything(pRExC_state, data->start_class);
4142 flags &= ~SCF_DO_STCLASS;
4146 else if (OP(scan) == LNBREAK) {
4147 if (flags & SCF_DO_STCLASS) {
4149 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4150 if (flags & SCF_DO_STCLASS_AND) {
4151 for (value = 0; value < 256; value++)
4152 if (!is_VERTWS_cp(value))
4153 ANYOF_BITMAP_CLEAR(data->start_class, value);
4156 for (value = 0; value < 256; value++)
4157 if (is_VERTWS_cp(value))
4158 ANYOF_BITMAP_SET(data->start_class, value);
4160 if (flags & SCF_DO_STCLASS_OR)
4161 cl_and(data->start_class, and_withp);
4162 flags &= ~SCF_DO_STCLASS;
4165 delta++; /* Because of the 2 char string cr-lf */
4166 if (flags & SCF_DO_SUBSTR) {
4167 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4169 data->pos_delta += 1;
4170 data->longest = &(data->longest_float);
4173 else if (REGNODE_SIMPLE(OP(scan))) {
4176 if (flags & SCF_DO_SUBSTR) {
4177 SCAN_COMMIT(pRExC_state,data,minlenp);
4181 if (flags & SCF_DO_STCLASS) {
4183 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4185 /* Some of the logic below assumes that switching
4186 locale on will only add false positives. */
4187 switch (PL_regkind[OP(scan)]) {
4193 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan));
4196 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4197 cl_anything(pRExC_state, data->start_class);
4200 if (OP(scan) == SANY)
4202 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
4203 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
4204 || ANYOF_CLASS_TEST_ANY_SET(data->start_class));
4205 cl_anything(pRExC_state, data->start_class);
4207 if (flags & SCF_DO_STCLASS_AND || !value)
4208 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
4211 if (flags & SCF_DO_STCLASS_AND)
4212 cl_and(data->start_class,
4213 (struct regnode_charclass_class*)scan);
4215 cl_or(pRExC_state, data->start_class,
4216 (struct regnode_charclass_class*)scan);
4224 classnum = FLAGS(scan);
4225 if (flags & SCF_DO_STCLASS_AND) {
4226 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4227 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum) + 1);
4228 for (value = 0; value < loop_max; value++) {
4229 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4230 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4236 if (data->start_class->flags & ANYOF_LOCALE) {
4237 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum));
4241 /* Even if under locale, set the bits for non-locale
4242 * in case it isn't a true locale-node. This will
4243 * create false positives if it truly is locale */
4244 for (value = 0; value < loop_max; value++) {
4245 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4246 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4258 classnum = FLAGS(scan);
4259 if (flags & SCF_DO_STCLASS_AND) {
4260 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4261 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum));
4262 for (value = 0; value < loop_max; value++) {
4263 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4264 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4270 if (data->start_class->flags & ANYOF_LOCALE) {
4271 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum) + 1);
4275 /* Even if under locale, set the bits for non-locale in
4276 * case it isn't a true locale-node. This will create
4277 * false positives if it truly is locale */
4278 for (value = 0; value < loop_max; value++) {
4279 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4280 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4283 if (PL_regkind[OP(scan)] == NPOSIXD) {
4284 data->start_class->flags |= ANYOF_NON_UTF8_LATIN1_ALL;
4290 if (flags & SCF_DO_STCLASS_OR)
4291 cl_and(data->start_class, and_withp);
4292 flags &= ~SCF_DO_STCLASS;
4295 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
4296 data->flags |= (OP(scan) == MEOL
4299 SCAN_COMMIT(pRExC_state, data, minlenp);
4302 else if ( PL_regkind[OP(scan)] == BRANCHJ
4303 /* Lookbehind, or need to calculate parens/evals/stclass: */
4304 && (scan->flags || data || (flags & SCF_DO_STCLASS))
4305 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
4306 if ( OP(scan) == UNLESSM &&
4308 OP(NEXTOPER(NEXTOPER(scan))) == NOTHING &&
4309 OP(regnext(NEXTOPER(NEXTOPER(scan)))) == SUCCEED
4312 regnode *upto= regnext(scan);
4314 SV * const mysv_val=sv_newmortal();
4315 DEBUG_STUDYDATA("OPFAIL",data,depth);
4317 /*DEBUG_PARSE_MSG("opfail");*/
4318 regprop(RExC_rx, mysv_val, upto);
4319 PerlIO_printf(Perl_debug_log, "~ replace with OPFAIL pointed at %s (%"IVdf") offset %"IVdf"\n",
4320 SvPV_nolen_const(mysv_val),
4321 (IV)REG_NODE_NUM(upto),
4326 NEXT_OFF(scan) = upto - scan;
4327 for (opt= scan + 1; opt < upto ; opt++)
4328 OP(opt) = OPTIMIZED;
4332 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4333 || OP(scan) == UNLESSM )
4335 /* Negative Lookahead/lookbehind
4336 In this case we can't do fixed string optimisation.
4339 I32 deltanext, minnext, fake = 0;
4341 struct regnode_charclass_class intrnl;
4344 data_fake.flags = 0;
4346 data_fake.whilem_c = data->whilem_c;
4347 data_fake.last_closep = data->last_closep;
4350 data_fake.last_closep = &fake;
4351 data_fake.pos_delta = delta;
4352 if ( flags & SCF_DO_STCLASS && !scan->flags
4353 && OP(scan) == IFMATCH ) { /* Lookahead */
4354 cl_init(pRExC_state, &intrnl);
4355 data_fake.start_class = &intrnl;
4356 f |= SCF_DO_STCLASS_AND;
4358 if (flags & SCF_WHILEM_VISITED_POS)
4359 f |= SCF_WHILEM_VISITED_POS;
4360 next = regnext(scan);
4361 nscan = NEXTOPER(NEXTOPER(scan));
4362 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
4363 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
4366 FAIL("Variable length lookbehind not implemented");
4368 else if (minnext > (I32)U8_MAX) {
4369 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4371 scan->flags = (U8)minnext;
4374 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4376 if (data_fake.flags & SF_HAS_EVAL)
4377 data->flags |= SF_HAS_EVAL;
4378 data->whilem_c = data_fake.whilem_c;
4380 if (f & SCF_DO_STCLASS_AND) {
4381 if (flags & SCF_DO_STCLASS_OR) {
4382 /* OR before, AND after: ideally we would recurse with
4383 * data_fake to get the AND applied by study of the
4384 * remainder of the pattern, and then derecurse;
4385 * *** HACK *** for now just treat as "no information".
4386 * See [perl #56690].
4388 cl_init(pRExC_state, data->start_class);
4390 /* AND before and after: combine and continue */
4391 const int was = TEST_SSC_EOS(data->start_class);
4393 cl_and(data->start_class, &intrnl);
4395 SET_SSC_EOS(data->start_class);
4399 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4401 /* Positive Lookahead/lookbehind
4402 In this case we can do fixed string optimisation,
4403 but we must be careful about it. Note in the case of
4404 lookbehind the positions will be offset by the minimum
4405 length of the pattern, something we won't know about
4406 until after the recurse.
4408 I32 deltanext, fake = 0;
4410 struct regnode_charclass_class intrnl;
4412 /* We use SAVEFREEPV so that when the full compile
4413 is finished perl will clean up the allocated
4414 minlens when it's all done. This way we don't
4415 have to worry about freeing them when we know
4416 they wont be used, which would be a pain.
4419 Newx( minnextp, 1, I32 );
4420 SAVEFREEPV(minnextp);
4423 StructCopy(data, &data_fake, scan_data_t);
4424 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
4427 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
4428 data_fake.last_found=newSVsv(data->last_found);
4432 data_fake.last_closep = &fake;
4433 data_fake.flags = 0;
4434 data_fake.pos_delta = delta;
4436 data_fake.flags |= SF_IS_INF;
4437 if ( flags & SCF_DO_STCLASS && !scan->flags
4438 && OP(scan) == IFMATCH ) { /* Lookahead */
4439 cl_init(pRExC_state, &intrnl);
4440 data_fake.start_class = &intrnl;
4441 f |= SCF_DO_STCLASS_AND;
4443 if (flags & SCF_WHILEM_VISITED_POS)
4444 f |= SCF_WHILEM_VISITED_POS;
4445 next = regnext(scan);
4446 nscan = NEXTOPER(NEXTOPER(scan));
4448 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
4449 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
4452 FAIL("Variable length lookbehind not implemented");
4454 else if (*minnextp > (I32)U8_MAX) {
4455 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4457 scan->flags = (U8)*minnextp;
4462 if (f & SCF_DO_STCLASS_AND) {
4463 const int was = TEST_SSC_EOS(data.start_class);
4465 cl_and(data->start_class, &intrnl);
4467 SET_SSC_EOS(data->start_class);
4470 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4472 if (data_fake.flags & SF_HAS_EVAL)
4473 data->flags |= SF_HAS_EVAL;
4474 data->whilem_c = data_fake.whilem_c;
4475 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
4476 if (RExC_rx->minlen<*minnextp)
4477 RExC_rx->minlen=*minnextp;
4478 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
4479 SvREFCNT_dec_NN(data_fake.last_found);
4481 if ( data_fake.minlen_fixed != minlenp )
4483 data->offset_fixed= data_fake.offset_fixed;
4484 data->minlen_fixed= data_fake.minlen_fixed;
4485 data->lookbehind_fixed+= scan->flags;
4487 if ( data_fake.minlen_float != minlenp )
4489 data->minlen_float= data_fake.minlen_float;
4490 data->offset_float_min=data_fake.offset_float_min;
4491 data->offset_float_max=data_fake.offset_float_max;
4492 data->lookbehind_float+= scan->flags;
4499 else if (OP(scan) == OPEN) {
4500 if (stopparen != (I32)ARG(scan))
4503 else if (OP(scan) == CLOSE) {
4504 if (stopparen == (I32)ARG(scan)) {
4507 if ((I32)ARG(scan) == is_par) {
4508 next = regnext(scan);
4510 if ( next && (OP(next) != WHILEM) && next < last)
4511 is_par = 0; /* Disable optimization */
4514 *(data->last_closep) = ARG(scan);
4516 else if (OP(scan) == EVAL) {
4518 data->flags |= SF_HAS_EVAL;
4520 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
4521 if (flags & SCF_DO_SUBSTR) {
4522 SCAN_COMMIT(pRExC_state,data,minlenp);
4523 flags &= ~SCF_DO_SUBSTR;
4525 if (data && OP(scan)==ACCEPT) {
4526 data->flags |= SCF_SEEN_ACCEPT;
4531 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
4533 if (flags & SCF_DO_SUBSTR) {
4534 SCAN_COMMIT(pRExC_state,data,minlenp);
4535 data->longest = &(data->longest_float);
4537 is_inf = is_inf_internal = 1;
4538 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4539 cl_anything(pRExC_state, data->start_class);
4540 flags &= ~SCF_DO_STCLASS;
4542 else if (OP(scan) == GPOS) {
4543 if (!(RExC_rx->extflags & RXf_GPOS_FLOAT) &&
4544 !(delta || is_inf || (data && data->pos_delta)))
4546 if (!(RExC_rx->extflags & RXf_ANCH) && (flags & SCF_DO_SUBSTR))
4547 RExC_rx->extflags |= RXf_ANCH_GPOS;
4548 if (RExC_rx->gofs < (U32)min)
4549 RExC_rx->gofs = min;
4551 RExC_rx->extflags |= RXf_GPOS_FLOAT;
4555 #ifdef TRIE_STUDY_OPT
4556 #ifdef FULL_TRIE_STUDY
4557 else if (PL_regkind[OP(scan)] == TRIE) {
4558 /* NOTE - There is similar code to this block above for handling
4559 BRANCH nodes on the initial study. If you change stuff here
4561 regnode *trie_node= scan;
4562 regnode *tail= regnext(scan);
4563 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4564 I32 max1 = 0, min1 = I32_MAX;
4565 struct regnode_charclass_class accum;
4567 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
4568 SCAN_COMMIT(pRExC_state, data,minlenp); /* Cannot merge strings after this. */
4569 if (flags & SCF_DO_STCLASS)
4570 cl_init_zero(pRExC_state, &accum);
4576 const regnode *nextbranch= NULL;
4579 for ( word=1 ; word <= trie->wordcount ; word++)
4581 I32 deltanext=0, minnext=0, f = 0, fake;
4582 struct regnode_charclass_class this_class;
4584 data_fake.flags = 0;
4586 data_fake.whilem_c = data->whilem_c;
4587 data_fake.last_closep = data->last_closep;
4590 data_fake.last_closep = &fake;
4591 data_fake.pos_delta = delta;
4592 if (flags & SCF_DO_STCLASS) {
4593 cl_init(pRExC_state, &this_class);
4594 data_fake.start_class = &this_class;
4595 f = SCF_DO_STCLASS_AND;
4597 if (flags & SCF_WHILEM_VISITED_POS)
4598 f |= SCF_WHILEM_VISITED_POS;
4600 if (trie->jump[word]) {
4602 nextbranch = trie_node + trie->jump[0];
4603 scan= trie_node + trie->jump[word];
4604 /* We go from the jump point to the branch that follows
4605 it. Note this means we need the vestigal unused branches
4606 even though they arent otherwise used.
4608 minnext = study_chunk(pRExC_state, &scan, minlenp,
4609 &deltanext, (regnode *)nextbranch, &data_fake,
4610 stopparen, recursed, NULL, f,depth+1);
4612 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
4613 nextbranch= regnext((regnode*)nextbranch);
4615 if (min1 > (I32)(minnext + trie->minlen))
4616 min1 = minnext + trie->minlen;
4617 if (deltanext == I32_MAX) {
4618 is_inf = is_inf_internal = 1;
4620 } else if (max1 < (I32)(minnext + deltanext + trie->maxlen))
4621 max1 = minnext + deltanext + trie->maxlen;
4623 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4625 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4626 if ( stopmin > min + min1)
4627 stopmin = min + min1;
4628 flags &= ~SCF_DO_SUBSTR;
4630 data->flags |= SCF_SEEN_ACCEPT;
4633 if (data_fake.flags & SF_HAS_EVAL)
4634 data->flags |= SF_HAS_EVAL;
4635 data->whilem_c = data_fake.whilem_c;
4637 if (flags & SCF_DO_STCLASS)
4638 cl_or(pRExC_state, &accum, &this_class);
4641 if (flags & SCF_DO_SUBSTR) {
4642 data->pos_min += min1;
4643 data->pos_delta += max1 - min1;
4644 if (max1 != min1 || is_inf)
4645 data->longest = &(data->longest_float);
4648 delta += max1 - min1;
4649 if (flags & SCF_DO_STCLASS_OR) {
4650 cl_or(pRExC_state, data->start_class, &accum);
4652 cl_and(data->start_class, and_withp);
4653 flags &= ~SCF_DO_STCLASS;
4656 else if (flags & SCF_DO_STCLASS_AND) {
4658 cl_and(data->start_class, &accum);
4659 flags &= ~SCF_DO_STCLASS;
4662 /* Switch to OR mode: cache the old value of
4663 * data->start_class */
4665 StructCopy(data->start_class, and_withp,
4666 struct regnode_charclass_class);
4667 flags &= ~SCF_DO_STCLASS_AND;
4668 StructCopy(&accum, data->start_class,
4669 struct regnode_charclass_class);
4670 flags |= SCF_DO_STCLASS_OR;
4671 SET_SSC_EOS(data->start_class);
4678 else if (PL_regkind[OP(scan)] == TRIE) {
4679 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4682 min += trie->minlen;
4683 delta += (trie->maxlen - trie->minlen);
4684 flags &= ~SCF_DO_STCLASS; /* xxx */
4685 if (flags & SCF_DO_SUBSTR) {
4686 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4687 data->pos_min += trie->minlen;
4688 data->pos_delta += (trie->maxlen - trie->minlen);
4689 if (trie->maxlen != trie->minlen)
4690 data->longest = &(data->longest_float);
4692 if (trie->jump) /* no more substrings -- for now /grr*/
4693 flags &= ~SCF_DO_SUBSTR;
4695 #endif /* old or new */
4696 #endif /* TRIE_STUDY_OPT */
4698 /* Else: zero-length, ignore. */
4699 scan = regnext(scan);
4704 stopparen = frame->stop;
4705 frame = frame->prev;
4706 goto fake_study_recurse;
4711 DEBUG_STUDYDATA("pre-fin:",data,depth);
4714 *deltap = is_inf_internal ? I32_MAX : delta;
4715 if (flags & SCF_DO_SUBSTR && is_inf)
4716 data->pos_delta = I32_MAX - data->pos_min;
4717 if (is_par > (I32)U8_MAX)
4719 if (is_par && pars==1 && data) {
4720 data->flags |= SF_IN_PAR;
4721 data->flags &= ~SF_HAS_PAR;
4723 else if (pars && data) {
4724 data->flags |= SF_HAS_PAR;
4725 data->flags &= ~SF_IN_PAR;
4727 if (flags & SCF_DO_STCLASS_OR)
4728 cl_and(data->start_class, and_withp);
4729 if (flags & SCF_TRIE_RESTUDY)
4730 data->flags |= SCF_TRIE_RESTUDY;
4732 DEBUG_STUDYDATA("post-fin:",data,depth);
4734 return min < stopmin ? min : stopmin;
4738 S_add_data(RExC_state_t *pRExC_state, U32 n, const char *s)
4740 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
4742 PERL_ARGS_ASSERT_ADD_DATA;
4744 Renewc(RExC_rxi->data,
4745 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
4746 char, struct reg_data);
4748 Renew(RExC_rxi->data->what, count + n, U8);
4750 Newx(RExC_rxi->data->what, n, U8);
4751 RExC_rxi->data->count = count + n;
4752 Copy(s, RExC_rxi->data->what + count, n, U8);
4756 /*XXX: todo make this not included in a non debugging perl */
4757 #ifndef PERL_IN_XSUB_RE
4759 Perl_reginitcolors(pTHX)
4762 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
4764 char *t = savepv(s);
4768 t = strchr(t, '\t');
4774 PL_colors[i] = t = (char *)"";
4779 PL_colors[i++] = (char *)"";
4786 #ifdef TRIE_STUDY_OPT
4787 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
4790 (data.flags & SCF_TRIE_RESTUDY) \
4798 #define CHECK_RESTUDY_GOTO_butfirst
4802 * pregcomp - compile a regular expression into internal code
4804 * Decides which engine's compiler to call based on the hint currently in
4808 #ifndef PERL_IN_XSUB_RE
4810 /* return the currently in-scope regex engine (or the default if none) */
4812 regexp_engine const *
4813 Perl_current_re_engine(pTHX)
4817 if (IN_PERL_COMPILETIME) {
4818 HV * const table = GvHV(PL_hintgv);
4822 return &reh_regexp_engine;
4823 ptr = hv_fetchs(table, "regcomp", FALSE);
4824 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
4825 return &reh_regexp_engine;
4826 return INT2PTR(regexp_engine*,SvIV(*ptr));
4830 if (!PL_curcop->cop_hints_hash)
4831 return &reh_regexp_engine;
4832 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
4833 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
4834 return &reh_regexp_engine;
4835 return INT2PTR(regexp_engine*,SvIV(ptr));
4841 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
4844 regexp_engine const *eng = current_re_engine();
4845 GET_RE_DEBUG_FLAGS_DECL;
4847 PERL_ARGS_ASSERT_PREGCOMP;
4849 /* Dispatch a request to compile a regexp to correct regexp engine. */
4851 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
4854 return CALLREGCOMP_ENG(eng, pattern, flags);
4858 /* public(ish) entry point for the perl core's own regex compiling code.
4859 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
4860 * pattern rather than a list of OPs, and uses the internal engine rather
4861 * than the current one */
4864 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
4866 SV *pat = pattern; /* defeat constness! */
4867 PERL_ARGS_ASSERT_RE_COMPILE;
4868 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
4869 #ifdef PERL_IN_XSUB_RE
4874 NULL, NULL, rx_flags, 0);
4878 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
4879 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
4880 * point to the realloced string and length.
4882 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
4886 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
4887 char **pat_p, STRLEN *plen_p, int num_code_blocks)
4889 U8 *const src = (U8*)*pat_p;
4892 STRLEN s = 0, d = 0;
4894 GET_RE_DEBUG_FLAGS_DECL;
4896 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
4897 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
4899 Newx(dst, *plen_p * 2 + 1, U8);
4901 while (s < *plen_p) {
4902 const UV uv = NATIVE_TO_ASCII(src[s]);
4903 if (UNI_IS_INVARIANT(uv))
4904 dst[d] = (U8)UTF_TO_NATIVE(uv);
4906 dst[d++] = (U8)UTF8_EIGHT_BIT_HI(uv);
4907 dst[d] = (U8)UTF8_EIGHT_BIT_LO(uv);
4909 if (n < num_code_blocks) {
4910 if (!do_end && pRExC_state->code_blocks[n].start == s) {
4911 pRExC_state->code_blocks[n].start = d;
4912 assert(dst[d] == '(');
4915 else if (do_end && pRExC_state->code_blocks[n].end == s) {
4916 pRExC_state->code_blocks[n].end = d;
4917 assert(dst[d] == ')');
4927 *pat_p = (char*) dst;
4929 RExC_orig_utf8 = RExC_utf8 = 1;
4934 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
4935 * while recording any code block indices, and handling overloading,
4936 * nested qr// objects etc. If pat is null, it will allocate a new
4937 * string, or just return the first arg, if there's only one.
4939 * Returns the malloced/updated pat.
4940 * patternp and pat_count is the array of SVs to be concatted;
4941 * oplist is the optional list of ops that generated the SVs;
4942 * recompile_p is a pointer to a boolean that will be set if
4943 * the regex will need to be recompiled.
4944 * delim, if non-null is an SV that will be inserted between each element
4948 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
4949 SV *pat, SV ** const patternp, int pat_count,
4950 OP *oplist, bool *recompile_p, SV *delim)
4954 bool use_delim = FALSE;
4955 bool alloced = FALSE;
4957 /* if we know we have at least two args, create an empty string,
4958 * then concatenate args to that. For no args, return an empty string */
4959 if (!pat && pat_count != 1) {
4960 pat = newSVpvn("", 0);
4965 for (svp = patternp; svp < patternp + pat_count; svp++) {
4968 STRLEN orig_patlen = 0;
4970 SV *msv = use_delim ? delim : *svp;
4972 /* if we've got a delimiter, we go round the loop twice for each
4973 * svp slot (except the last), using the delimiter the second
4982 if (SvTYPE(msv) == SVt_PVAV) {
4983 /* we've encountered an interpolated array within
4984 * the pattern, e.g. /...@a..../. Expand the list of elements,
4985 * then recursively append elements.
4986 * The code in this block is based on S_pushav() */
4988 AV *const av = (AV*)msv;
4989 const I32 maxarg = AvFILL(av) + 1;
4993 assert(oplist->op_type == OP_PADAV
4994 || oplist->op_type == OP_RV2AV);
4995 oplist = oplist->op_sibling;;
4998 if (SvRMAGICAL(av)) {
5001 Newx(array, maxarg, SV*);
5003 for (i=0; i < (U32)maxarg; i++) {
5004 SV ** const svp = av_fetch(av, i, FALSE);
5005 array[i] = svp ? *svp : &PL_sv_undef;
5009 array = AvARRAY(av);
5011 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5012 array, maxarg, NULL, recompile_p,
5014 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5020 /* we make the assumption here that each op in the list of
5021 * op_siblings maps to one SV pushed onto the stack,
5022 * except for code blocks, with have both an OP_NULL and
5024 * This allows us to match up the list of SVs against the
5025 * list of OPs to find the next code block.
5027 * Note that PUSHMARK PADSV PADSV ..
5029 * PADRANGE PADSV PADSV ..
5030 * so the alignment still works. */
5033 if (oplist->op_type == OP_NULL
5034 && (oplist->op_flags & OPf_SPECIAL))
5036 assert(n < pRExC_state->num_code_blocks);
5037 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5038 pRExC_state->code_blocks[n].block = oplist;
5039 pRExC_state->code_blocks[n].src_regex = NULL;
5042 oplist = oplist->op_sibling; /* skip CONST */
5045 oplist = oplist->op_sibling;;
5048 /* apply magic and QR overloading to arg */
5051 if (SvROK(msv) && SvAMAGIC(msv)) {
5052 SV *sv = AMG_CALLunary(msv, regexp_amg);
5056 if (SvTYPE(sv) != SVt_REGEXP)
5057 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5062 /* try concatenation overload ... */
5063 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5064 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5067 /* overloading involved: all bets are off over literal
5068 * code. Pretend we haven't seen it */
5069 pRExC_state->num_code_blocks -= n;
5073 /* ... or failing that, try "" overload */
5074 while (SvAMAGIC(msv)
5075 && (sv = AMG_CALLunary(msv, string_amg))
5079 && SvRV(msv) == SvRV(sv))
5084 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5088 /* this is a partially unrolled
5089 * sv_catsv_nomg(pat, msv);
5090 * that allows us to adjust code block indices if
5093 char *dst = SvPV_force_nomg(pat, dlen);
5095 if (SvUTF8(msv) && !SvUTF8(pat)) {
5096 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
5097 sv_setpvn(pat, dst, dlen);
5100 sv_catsv_nomg(pat, msv);
5107 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5110 /* extract any code blocks within any embedded qr//'s */
5111 if (rx && SvTYPE(rx) == SVt_REGEXP
5112 && RX_ENGINE((REGEXP*)rx)->op_comp)
5115 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5116 if (ri->num_code_blocks) {
5118 /* the presence of an embedded qr// with code means
5119 * we should always recompile: the text of the
5120 * qr// may not have changed, but it may be a
5121 * different closure than last time */
5123 Renew(pRExC_state->code_blocks,
5124 pRExC_state->num_code_blocks + ri->num_code_blocks,
5125 struct reg_code_block);
5126 pRExC_state->num_code_blocks += ri->num_code_blocks;
5128 for (i=0; i < ri->num_code_blocks; i++) {
5129 struct reg_code_block *src, *dst;
5130 STRLEN offset = orig_patlen
5131 + ReANY((REGEXP *)rx)->pre_prefix;
5132 assert(n < pRExC_state->num_code_blocks);
5133 src = &ri->code_blocks[i];
5134 dst = &pRExC_state->code_blocks[n];
5135 dst->start = src->start + offset;
5136 dst->end = src->end + offset;
5137 dst->block = src->block;
5138 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5147 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
5156 /* see if there are any run-time code blocks in the pattern.
5157 * False positives are allowed */
5160 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5161 char *pat, STRLEN plen)
5166 for (s = 0; s < plen; s++) {
5167 if (n < pRExC_state->num_code_blocks
5168 && s == pRExC_state->code_blocks[n].start)
5170 s = pRExC_state->code_blocks[n].end;
5174 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
5176 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
5178 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
5185 /* Handle run-time code blocks. We will already have compiled any direct
5186 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
5187 * copy of it, but with any literal code blocks blanked out and
5188 * appropriate chars escaped; then feed it into
5190 * eval "qr'modified_pattern'"
5194 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
5198 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
5200 * After eval_sv()-ing that, grab any new code blocks from the returned qr
5201 * and merge them with any code blocks of the original regexp.
5203 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
5204 * instead, just save the qr and return FALSE; this tells our caller that
5205 * the original pattern needs upgrading to utf8.
5209 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5210 char *pat, STRLEN plen)
5214 GET_RE_DEBUG_FLAGS_DECL;
5216 if (pRExC_state->runtime_code_qr) {
5217 /* this is the second time we've been called; this should
5218 * only happen if the main pattern got upgraded to utf8
5219 * during compilation; re-use the qr we compiled first time
5220 * round (which should be utf8 too)
5222 qr = pRExC_state->runtime_code_qr;
5223 pRExC_state->runtime_code_qr = NULL;
5224 assert(RExC_utf8 && SvUTF8(qr));
5230 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
5234 /* determine how many extra chars we need for ' and \ escaping */
5235 for (s = 0; s < plen; s++) {
5236 if (pat[s] == '\'' || pat[s] == '\\')
5240 Newx(newpat, newlen, char);
5242 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
5244 for (s = 0; s < plen; s++) {
5245 if (n < pRExC_state->num_code_blocks
5246 && s == pRExC_state->code_blocks[n].start)
5248 /* blank out literal code block */
5249 assert(pat[s] == '(');
5250 while (s <= pRExC_state->code_blocks[n].end) {
5258 if (pat[s] == '\'' || pat[s] == '\\')
5263 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
5267 PerlIO_printf(Perl_debug_log,
5268 "%sre-parsing pattern for runtime code:%s %s\n",
5269 PL_colors[4],PL_colors[5],newpat);
5272 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5278 PUSHSTACKi(PERLSI_REQUIRE);
5279 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
5280 * parsing qr''; normally only q'' does this. It also alters
5282 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
5283 SvREFCNT_dec_NN(sv);
5288 SV * const errsv = ERRSV;
5289 if (SvTRUE_NN(errsv))
5291 Safefree(pRExC_state->code_blocks);
5292 /* use croak_sv ? */
5293 Perl_croak_nocontext("%s", SvPV_nolen_const(errsv));
5296 assert(SvROK(qr_ref));
5298 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5299 /* the leaving below frees the tmp qr_ref.
5300 * Give qr a life of its own */
5308 if (!RExC_utf8 && SvUTF8(qr)) {
5309 /* first time through; the pattern got upgraded; save the
5310 * qr for the next time through */
5311 assert(!pRExC_state->runtime_code_qr);
5312 pRExC_state->runtime_code_qr = qr;
5317 /* extract any code blocks within the returned qr// */
5320 /* merge the main (r1) and run-time (r2) code blocks into one */
5322 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
5323 struct reg_code_block *new_block, *dst;
5324 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
5327 if (!r2->num_code_blocks) /* we guessed wrong */
5329 SvREFCNT_dec_NN(qr);
5334 r1->num_code_blocks + r2->num_code_blocks,
5335 struct reg_code_block);
5338 while ( i1 < r1->num_code_blocks
5339 || i2 < r2->num_code_blocks)
5341 struct reg_code_block *src;
5344 if (i1 == r1->num_code_blocks) {
5345 src = &r2->code_blocks[i2++];
5348 else if (i2 == r2->num_code_blocks)
5349 src = &r1->code_blocks[i1++];
5350 else if ( r1->code_blocks[i1].start
5351 < r2->code_blocks[i2].start)
5353 src = &r1->code_blocks[i1++];
5354 assert(src->end < r2->code_blocks[i2].start);
5357 assert( r1->code_blocks[i1].start
5358 > r2->code_blocks[i2].start);
5359 src = &r2->code_blocks[i2++];
5361 assert(src->end < r1->code_blocks[i1].start);
5364 assert(pat[src->start] == '(');
5365 assert(pat[src->end] == ')');
5366 dst->start = src->start;
5367 dst->end = src->end;
5368 dst->block = src->block;
5369 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
5373 r1->num_code_blocks += r2->num_code_blocks;
5374 Safefree(r1->code_blocks);
5375 r1->code_blocks = new_block;
5378 SvREFCNT_dec_NN(qr);
5384 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)
5386 /* This is the common code for setting up the floating and fixed length
5387 * string data extracted from Perl_re_op_compile() below. Returns a boolean
5388 * as to whether succeeded or not */
5392 if (! (longest_length
5393 || (eol /* Can't have SEOL and MULTI */
5394 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
5396 /* See comments for join_exact for why REG_SEEN_EXACTF_SHARP_S */
5397 || (RExC_seen & REG_SEEN_EXACTF_SHARP_S))
5402 /* copy the information about the longest from the reg_scan_data
5403 over to the program. */
5404 if (SvUTF8(sv_longest)) {
5405 *rx_utf8 = sv_longest;
5408 *rx_substr = sv_longest;
5411 /* end_shift is how many chars that must be matched that
5412 follow this item. We calculate it ahead of time as once the
5413 lookbehind offset is added in we lose the ability to correctly
5415 ml = minlen ? *(minlen) : (I32)longest_length;
5416 *rx_end_shift = ml - offset
5417 - longest_length + (SvTAIL(sv_longest) != 0)
5420 t = (eol/* Can't have SEOL and MULTI */
5421 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
5422 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
5428 * Perl_re_op_compile - the perl internal RE engine's function to compile a
5429 * regular expression into internal code.
5430 * The pattern may be passed either as:
5431 * a list of SVs (patternp plus pat_count)
5432 * a list of OPs (expr)
5433 * If both are passed, the SV list is used, but the OP list indicates
5434 * which SVs are actually pre-compiled code blocks
5436 * The SVs in the list have magic and qr overloading applied to them (and
5437 * the list may be modified in-place with replacement SVs in the latter
5440 * If the pattern hasn't changed from old_re, then old_re will be
5443 * eng is the current engine. If that engine has an op_comp method, then
5444 * handle directly (i.e. we assume that op_comp was us); otherwise, just
5445 * do the initial concatenation of arguments and pass on to the external
5448 * If is_bare_re is not null, set it to a boolean indicating whether the
5449 * arg list reduced (after overloading) to a single bare regex which has
5450 * been returned (i.e. /$qr/).
5452 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
5454 * pm_flags contains the PMf_* flags, typically based on those from the
5455 * pm_flags field of the related PMOP. Currently we're only interested in
5456 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
5458 * We can't allocate space until we know how big the compiled form will be,
5459 * but we can't compile it (and thus know how big it is) until we've got a
5460 * place to put the code. So we cheat: we compile it twice, once with code
5461 * generation turned off and size counting turned on, and once "for real".
5462 * This also means that we don't allocate space until we are sure that the
5463 * thing really will compile successfully, and we never have to move the
5464 * code and thus invalidate pointers into it. (Note that it has to be in
5465 * one piece because free() must be able to free it all.) [NB: not true in perl]
5467 * Beware that the optimization-preparation code in here knows about some
5468 * of the structure of the compiled regexp. [I'll say.]
5472 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
5473 OP *expr, const regexp_engine* eng, REGEXP *old_re,
5474 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
5479 regexp_internal *ri;
5487 SV *code_blocksv = NULL;
5488 SV** new_patternp = patternp;
5490 /* these are all flags - maybe they should be turned
5491 * into a single int with different bit masks */
5492 I32 sawlookahead = 0;
5495 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
5497 bool runtime_code = 0;
5499 RExC_state_t RExC_state;
5500 RExC_state_t * const pRExC_state = &RExC_state;
5501 #ifdef TRIE_STUDY_OPT
5503 RExC_state_t copyRExC_state;
5505 GET_RE_DEBUG_FLAGS_DECL;
5507 PERL_ARGS_ASSERT_RE_OP_COMPILE;
5509 DEBUG_r(if (!PL_colorset) reginitcolors());
5511 #ifndef PERL_IN_XSUB_RE
5512 /* Initialize these here instead of as-needed, as is quick and avoids
5513 * having to test them each time otherwise */
5514 if (! PL_AboveLatin1) {
5515 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
5516 PL_ASCII = _new_invlist_C_array(ASCII_invlist);
5517 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
5519 PL_L1Posix_ptrs[_CC_ALPHANUMERIC]
5520 = _new_invlist_C_array(L1PosixAlnum_invlist);
5521 PL_Posix_ptrs[_CC_ALPHANUMERIC]
5522 = _new_invlist_C_array(PosixAlnum_invlist);
5524 PL_L1Posix_ptrs[_CC_ALPHA]
5525 = _new_invlist_C_array(L1PosixAlpha_invlist);
5526 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(PosixAlpha_invlist);
5528 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(PosixBlank_invlist);
5529 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(XPosixBlank_invlist);
5531 /* Cased is the same as Alpha in the ASCII range */
5532 PL_L1Posix_ptrs[_CC_CASED] = _new_invlist_C_array(L1Cased_invlist);
5533 PL_Posix_ptrs[_CC_CASED] = _new_invlist_C_array(PosixAlpha_invlist);
5535 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(PosixCntrl_invlist);
5536 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(XPosixCntrl_invlist);
5538 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5539 PL_L1Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5541 PL_L1Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(L1PosixGraph_invlist);
5542 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(PosixGraph_invlist);
5544 PL_L1Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(L1PosixLower_invlist);
5545 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(PosixLower_invlist);
5547 PL_L1Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(L1PosixPrint_invlist);
5548 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(PosixPrint_invlist);
5550 PL_L1Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(L1PosixPunct_invlist);
5551 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(PosixPunct_invlist);
5553 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(PerlSpace_invlist);
5554 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(XPerlSpace_invlist);
5555 PL_Posix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(PosixSpace_invlist);
5556 PL_XPosix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(XPosixSpace_invlist);
5558 PL_L1Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(L1PosixUpper_invlist);
5559 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(PosixUpper_invlist);
5561 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(VertSpace_invlist);
5563 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(PosixWord_invlist);
5564 PL_L1Posix_ptrs[_CC_WORDCHAR]
5565 = _new_invlist_C_array(L1PosixWord_invlist);
5567 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(PosixXDigit_invlist);
5568 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(XPosixXDigit_invlist);
5570 PL_HasMultiCharFold = _new_invlist_C_array(_Perl_Multi_Char_Folds_invlist);
5574 pRExC_state->code_blocks = NULL;
5575 pRExC_state->num_code_blocks = 0;
5578 *is_bare_re = FALSE;
5580 if (expr && (expr->op_type == OP_LIST ||
5581 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
5582 /* allocate code_blocks if needed */
5586 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling)
5587 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
5588 ncode++; /* count of DO blocks */
5590 pRExC_state->num_code_blocks = ncode;
5591 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
5596 /* compile-time pattern with just OP_CONSTs and DO blocks */
5601 /* find how many CONSTs there are */
5604 if (expr->op_type == OP_CONST)
5607 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5608 if (o->op_type == OP_CONST)
5612 /* fake up an SV array */
5614 assert(!new_patternp);
5615 Newx(new_patternp, n, SV*);
5616 SAVEFREEPV(new_patternp);
5620 if (expr->op_type == OP_CONST)
5621 new_patternp[n] = cSVOPx_sv(expr);
5623 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5624 if (o->op_type == OP_CONST)
5625 new_patternp[n++] = cSVOPo_sv;
5630 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5631 "Assembling pattern from %d elements%s\n", pat_count,
5632 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5634 /* set expr to the first arg op */
5636 if (pRExC_state->num_code_blocks
5637 && expr->op_type != OP_CONST)
5639 expr = cLISTOPx(expr)->op_first;
5640 assert( expr->op_type == OP_PUSHMARK
5641 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
5642 || expr->op_type == OP_PADRANGE);
5643 expr = expr->op_sibling;
5646 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
5647 expr, &recompile, NULL);
5649 /* handle bare (possibly after overloading) regex: foo =~ $re */
5654 if (SvTYPE(re) == SVt_REGEXP) {
5658 Safefree(pRExC_state->code_blocks);
5659 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5660 "Precompiled pattern%s\n",
5661 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5667 exp = SvPV_nomg(pat, plen);
5669 if (!eng->op_comp) {
5670 if ((SvUTF8(pat) && IN_BYTES)
5671 || SvGMAGICAL(pat) || SvAMAGIC(pat))
5673 /* make a temporary copy; either to convert to bytes,
5674 * or to avoid repeating get-magic / overloaded stringify */
5675 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
5676 (IN_BYTES ? 0 : SvUTF8(pat)));
5678 Safefree(pRExC_state->code_blocks);
5679 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
5682 /* ignore the utf8ness if the pattern is 0 length */
5683 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
5684 RExC_uni_semantics = 0;
5685 RExC_contains_locale = 0;
5686 pRExC_state->runtime_code_qr = NULL;
5689 SV *dsv= sv_newmortal();
5690 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
5691 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
5692 PL_colors[4],PL_colors[5],s);
5696 /* we jump here if we upgrade the pattern to utf8 and have to
5699 if ((pm_flags & PMf_USE_RE_EVAL)
5700 /* this second condition covers the non-regex literal case,
5701 * i.e. $foo =~ '(?{})'. */
5702 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
5704 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
5706 /* return old regex if pattern hasn't changed */
5707 /* XXX: note in the below we have to check the flags as well as the pattern.
5709 * Things get a touch tricky as we have to compare the utf8 flag independently
5710 * from the compile flags.
5715 && !!RX_UTF8(old_re) == !!RExC_utf8
5716 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
5717 && RX_PRECOMP(old_re)
5718 && RX_PRELEN(old_re) == plen
5719 && memEQ(RX_PRECOMP(old_re), exp, plen)
5720 && !runtime_code /* with runtime code, always recompile */ )
5722 Safefree(pRExC_state->code_blocks);
5726 rx_flags = orig_rx_flags;
5728 if (initial_charset == REGEX_LOCALE_CHARSET) {
5729 RExC_contains_locale = 1;
5731 else if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
5733 /* Set to use unicode semantics if the pattern is in utf8 and has the
5734 * 'depends' charset specified, as it means unicode when utf8 */
5735 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5739 RExC_flags = rx_flags;
5740 RExC_pm_flags = pm_flags;
5743 if (TAINTING_get && TAINT_get)
5744 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
5746 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
5747 /* whoops, we have a non-utf8 pattern, whilst run-time code
5748 * got compiled as utf8. Try again with a utf8 pattern */
5749 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
5750 pRExC_state->num_code_blocks);
5751 goto redo_first_pass;
5754 assert(!pRExC_state->runtime_code_qr);
5759 RExC_in_lookbehind = 0;
5760 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
5762 RExC_override_recoding = 0;
5763 RExC_in_multi_char_class = 0;
5765 /* First pass: determine size, legality. */
5768 RExC_end = exp + plen;
5773 RExC_emit = &PL_regdummy;
5774 RExC_whilem_seen = 0;
5775 RExC_open_parens = NULL;
5776 RExC_close_parens = NULL;
5778 RExC_paren_names = NULL;
5780 RExC_paren_name_list = NULL;
5782 RExC_recurse = NULL;
5783 RExC_recurse_count = 0;
5784 pRExC_state->code_index = 0;
5786 #if 0 /* REGC() is (currently) a NOP at the first pass.
5787 * Clever compilers notice this and complain. --jhi */
5788 REGC((U8)REG_MAGIC, (char*)RExC_emit);
5791 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
5793 RExC_lastparse=NULL;
5795 /* reg may croak on us, not giving us a chance to free
5796 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
5797 need it to survive as long as the regexp (qr/(?{})/).
5798 We must check that code_blocksv is not already set, because we may
5799 have jumped back to restart the sizing pass. */
5800 if (pRExC_state->code_blocks && !code_blocksv) {
5801 code_blocksv = newSV_type(SVt_PV);
5802 SAVEFREESV(code_blocksv);
5803 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
5804 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
5806 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5807 /* It's possible to write a regexp in ascii that represents Unicode
5808 codepoints outside of the byte range, such as via \x{100}. If we
5809 detect such a sequence we have to convert the entire pattern to utf8
5810 and then recompile, as our sizing calculation will have been based
5811 on 1 byte == 1 character, but we will need to use utf8 to encode
5812 at least some part of the pattern, and therefore must convert the whole
5815 if (flags & RESTART_UTF8) {
5816 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
5817 pRExC_state->num_code_blocks);
5818 goto redo_first_pass;
5820 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#X", flags);
5823 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
5826 PerlIO_printf(Perl_debug_log,
5827 "Required size %"IVdf" nodes\n"
5828 "Starting second pass (creation)\n",
5831 RExC_lastparse=NULL;
5834 /* The first pass could have found things that force Unicode semantics */
5835 if ((RExC_utf8 || RExC_uni_semantics)
5836 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
5838 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5841 /* Small enough for pointer-storage convention?
5842 If extralen==0, this means that we will not need long jumps. */
5843 if (RExC_size >= 0x10000L && RExC_extralen)
5844 RExC_size += RExC_extralen;
5847 if (RExC_whilem_seen > 15)
5848 RExC_whilem_seen = 15;
5850 /* Allocate space and zero-initialize. Note, the two step process
5851 of zeroing when in debug mode, thus anything assigned has to
5852 happen after that */
5853 rx = (REGEXP*) newSV_type(SVt_REGEXP);
5855 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
5856 char, regexp_internal);
5857 if ( r == NULL || ri == NULL )
5858 FAIL("Regexp out of space");
5860 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
5861 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode), char);
5863 /* bulk initialize base fields with 0. */
5864 Zero(ri, sizeof(regexp_internal), char);
5867 /* non-zero initialization begins here */
5870 r->extflags = rx_flags;
5871 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
5873 if (pm_flags & PMf_IS_QR) {
5874 ri->code_blocks = pRExC_state->code_blocks;
5875 ri->num_code_blocks = pRExC_state->num_code_blocks;
5880 for (n = 0; n < pRExC_state->num_code_blocks; n++)
5881 if (pRExC_state->code_blocks[n].src_regex)
5882 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
5883 SAVEFREEPV(pRExC_state->code_blocks);
5887 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
5888 bool has_charset = (get_regex_charset(r->extflags) != REGEX_DEPENDS_CHARSET);
5890 /* The caret is output if there are any defaults: if not all the STD
5891 * flags are set, or if no character set specifier is needed */
5893 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
5895 bool has_runon = ((RExC_seen & REG_SEEN_RUN_ON_COMMENT)==REG_SEEN_RUN_ON_COMMENT);
5896 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
5897 >> RXf_PMf_STD_PMMOD_SHIFT);
5898 const char *fptr = STD_PAT_MODS; /*"msix"*/
5900 /* Allocate for the worst case, which is all the std flags are turned
5901 * on. If more precision is desired, we could do a population count of
5902 * the flags set. This could be done with a small lookup table, or by
5903 * shifting, masking and adding, or even, when available, assembly
5904 * language for a machine-language population count.
5905 * We never output a minus, as all those are defaults, so are
5906 * covered by the caret */
5907 const STRLEN wraplen = plen + has_p + has_runon
5908 + has_default /* If needs a caret */
5910 /* If needs a character set specifier */
5911 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
5912 + (sizeof(STD_PAT_MODS) - 1)
5913 + (sizeof("(?:)") - 1);
5915 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
5916 r->xpv_len_u.xpvlenu_pv = p;
5918 SvFLAGS(rx) |= SVf_UTF8;
5921 /* If a default, cover it using the caret */
5923 *p++= DEFAULT_PAT_MOD;
5927 const char* const name = get_regex_charset_name(r->extflags, &len);
5928 Copy(name, p, len, char);
5932 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
5935 while((ch = *fptr++)) {
5943 Copy(RExC_precomp, p, plen, char);
5944 assert ((RX_WRAPPED(rx) - p) < 16);
5945 r->pre_prefix = p - RX_WRAPPED(rx);
5951 SvCUR_set(rx, p - RX_WRAPPED(rx));
5955 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
5957 if (RExC_seen & REG_SEEN_RECURSE) {
5958 Newxz(RExC_open_parens, RExC_npar,regnode *);
5959 SAVEFREEPV(RExC_open_parens);
5960 Newxz(RExC_close_parens,RExC_npar,regnode *);
5961 SAVEFREEPV(RExC_close_parens);
5964 /* Useful during FAIL. */
5965 #ifdef RE_TRACK_PATTERN_OFFSETS
5966 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
5967 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
5968 "%s %"UVuf" bytes for offset annotations.\n",
5969 ri->u.offsets ? "Got" : "Couldn't get",
5970 (UV)((2*RExC_size+1) * sizeof(U32))));
5972 SetProgLen(ri,RExC_size);
5976 REH_CALL_COMP_BEGIN_HOOK(pRExC_state->rx);
5978 /* Second pass: emit code. */
5979 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
5980 RExC_pm_flags = pm_flags;
5982 RExC_end = exp + plen;
5985 RExC_emit_start = ri->program;
5986 RExC_emit = ri->program;
5987 RExC_emit_bound = ri->program + RExC_size + 1;
5988 pRExC_state->code_index = 0;
5990 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
5991 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5993 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#X", flags);
5995 /* XXXX To minimize changes to RE engine we always allocate
5996 3-units-long substrs field. */
5997 Newx(r->substrs, 1, struct reg_substr_data);
5998 if (RExC_recurse_count) {
5999 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6000 SAVEFREEPV(RExC_recurse);
6004 r->minlen = minlen = sawlookahead = sawplus = sawopen = 0;
6005 Zero(r->substrs, 1, struct reg_substr_data);
6007 #ifdef TRIE_STUDY_OPT
6009 StructCopy(&zero_scan_data, &data, scan_data_t);
6010 copyRExC_state = RExC_state;
6013 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6015 RExC_state = copyRExC_state;
6016 if (seen & REG_TOP_LEVEL_BRANCHES)
6017 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
6019 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES;
6020 StructCopy(&zero_scan_data, &data, scan_data_t);
6023 StructCopy(&zero_scan_data, &data, scan_data_t);
6026 /* Dig out information for optimizations. */
6027 r->extflags = RExC_flags; /* was pm_op */
6028 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6031 SvUTF8_on(rx); /* Unicode in it? */
6032 ri->regstclass = NULL;
6033 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
6034 r->intflags |= PREGf_NAUGHTY;
6035 scan = ri->program + 1; /* First BRANCH. */
6037 /* testing for BRANCH here tells us whether there is "must appear"
6038 data in the pattern. If there is then we can use it for optimisations */
6039 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES)) { /* Only one top-level choice. */
6041 STRLEN longest_float_length, longest_fixed_length;
6042 struct regnode_charclass_class ch_class; /* pointed to by data */
6044 I32 last_close = 0; /* pointed to by data */
6045 regnode *first= scan;
6046 regnode *first_next= regnext(first);
6048 * Skip introductions and multiplicators >= 1
6049 * so that we can extract the 'meat' of the pattern that must
6050 * match in the large if() sequence following.
6051 * NOTE that EXACT is NOT covered here, as it is normally
6052 * picked up by the optimiser separately.
6054 * This is unfortunate as the optimiser isnt handling lookahead
6055 * properly currently.
6058 while ((OP(first) == OPEN && (sawopen = 1)) ||
6059 /* An OR of *one* alternative - should not happen now. */
6060 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6061 /* for now we can't handle lookbehind IFMATCH*/
6062 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6063 (OP(first) == PLUS) ||
6064 (OP(first) == MINMOD) ||
6065 /* An {n,m} with n>0 */
6066 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6067 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6070 * the only op that could be a regnode is PLUS, all the rest
6071 * will be regnode_1 or regnode_2.
6074 if (OP(first) == PLUS)
6077 first += regarglen[OP(first)];
6079 first = NEXTOPER(first);
6080 first_next= regnext(first);
6083 /* Starting-point info. */
6085 DEBUG_PEEP("first:",first,0);
6086 /* Ignore EXACT as we deal with it later. */
6087 if (PL_regkind[OP(first)] == EXACT) {
6088 if (OP(first) == EXACT)
6089 NOOP; /* Empty, get anchored substr later. */
6091 ri->regstclass = first;
6094 else if (PL_regkind[OP(first)] == TRIE &&
6095 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6098 /* this can happen only on restudy */
6099 if ( OP(first) == TRIE ) {
6100 struct regnode_1 *trieop = (struct regnode_1 *)
6101 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6102 StructCopy(first,trieop,struct regnode_1);
6103 trie_op=(regnode *)trieop;
6105 struct regnode_charclass *trieop = (struct regnode_charclass *)
6106 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6107 StructCopy(first,trieop,struct regnode_charclass);
6108 trie_op=(regnode *)trieop;
6111 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6112 ri->regstclass = trie_op;
6115 else if (REGNODE_SIMPLE(OP(first)))
6116 ri->regstclass = first;
6117 else if (PL_regkind[OP(first)] == BOUND ||
6118 PL_regkind[OP(first)] == NBOUND)
6119 ri->regstclass = first;
6120 else if (PL_regkind[OP(first)] == BOL) {
6121 r->extflags |= (OP(first) == MBOL
6123 : (OP(first) == SBOL
6126 first = NEXTOPER(first);
6129 else if (OP(first) == GPOS) {
6130 r->extflags |= RXf_ANCH_GPOS;
6131 first = NEXTOPER(first);
6134 else if ((!sawopen || !RExC_sawback) &&
6135 (OP(first) == STAR &&
6136 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6137 !(r->extflags & RXf_ANCH) && !pRExC_state->num_code_blocks)
6139 /* turn .* into ^.* with an implied $*=1 */
6141 (OP(NEXTOPER(first)) == REG_ANY)
6144 r->extflags |= type;
6145 r->intflags |= PREGf_IMPLICIT;
6146 first = NEXTOPER(first);
6149 if (sawplus && !sawlookahead && (!sawopen || !RExC_sawback)
6150 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6151 /* x+ must match at the 1st pos of run of x's */
6152 r->intflags |= PREGf_SKIP;
6154 /* Scan is after the zeroth branch, first is atomic matcher. */
6155 #ifdef TRIE_STUDY_OPT
6158 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6159 (IV)(first - scan + 1))
6163 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6164 (IV)(first - scan + 1))
6170 * If there's something expensive in the r.e., find the
6171 * longest literal string that must appear and make it the
6172 * regmust. Resolve ties in favor of later strings, since
6173 * the regstart check works with the beginning of the r.e.
6174 * and avoiding duplication strengthens checking. Not a
6175 * strong reason, but sufficient in the absence of others.
6176 * [Now we resolve ties in favor of the earlier string if
6177 * it happens that c_offset_min has been invalidated, since the
6178 * earlier string may buy us something the later one won't.]
6181 data.longest_fixed = newSVpvs("");
6182 data.longest_float = newSVpvs("");
6183 data.last_found = newSVpvs("");
6184 data.longest = &(data.longest_fixed);
6185 ENTER_with_name("study_chunk");
6186 SAVEFREESV(data.longest_fixed);
6187 SAVEFREESV(data.longest_float);
6188 SAVEFREESV(data.last_found);
6190 if (!ri->regstclass) {
6191 cl_init(pRExC_state, &ch_class);
6192 data.start_class = &ch_class;
6193 stclass_flag = SCF_DO_STCLASS_AND;
6194 } else /* XXXX Check for BOUND? */
6196 data.last_closep = &last_close;
6198 minlen = study_chunk(pRExC_state, &first, &minlen, &fake, scan + RExC_size, /* Up to end */
6199 &data, -1, NULL, NULL,
6200 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag,0);
6203 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6206 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6207 && data.last_start_min == 0 && data.last_end > 0
6208 && !RExC_seen_zerolen
6209 && !(RExC_seen & REG_SEEN_VERBARG)
6210 && (!(RExC_seen & REG_SEEN_GPOS) || (r->extflags & RXf_ANCH_GPOS)))
6211 r->extflags |= RXf_CHECK_ALL;
6212 scan_commit(pRExC_state, &data,&minlen,0);
6214 longest_float_length = CHR_SVLEN(data.longest_float);
6216 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6217 && data.offset_fixed == data.offset_float_min
6218 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6219 && S_setup_longest (aTHX_ pRExC_state,
6223 &(r->float_end_shift),
6224 data.lookbehind_float,
6225 data.offset_float_min,
6227 longest_float_length,
6228 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6229 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6231 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6232 r->float_max_offset = data.offset_float_max;
6233 if (data.offset_float_max < I32_MAX) /* Don't offset infinity */
6234 r->float_max_offset -= data.lookbehind_float;
6235 SvREFCNT_inc_simple_void_NN(data.longest_float);
6238 r->float_substr = r->float_utf8 = NULL;
6239 longest_float_length = 0;
6242 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6244 if (S_setup_longest (aTHX_ pRExC_state,
6246 &(r->anchored_utf8),
6247 &(r->anchored_substr),
6248 &(r->anchored_end_shift),
6249 data.lookbehind_fixed,
6252 longest_fixed_length,
6253 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6254 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6256 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6257 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6260 r->anchored_substr = r->anchored_utf8 = NULL;
6261 longest_fixed_length = 0;
6263 LEAVE_with_name("study_chunk");
6266 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6267 ri->regstclass = NULL;
6269 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6271 && ! TEST_SSC_EOS(data.start_class)
6272 && !cl_is_anything(data.start_class))
6274 const U32 n = add_data(pRExC_state, 1, "f");
6275 OP(data.start_class) = ANYOF_SYNTHETIC;
6277 Newx(RExC_rxi->data->data[n], 1,
6278 struct regnode_charclass_class);
6279 StructCopy(data.start_class,
6280 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6281 struct regnode_charclass_class);
6282 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6283 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6284 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6285 regprop(r, sv, (regnode*)data.start_class);
6286 PerlIO_printf(Perl_debug_log,
6287 "synthetic stclass \"%s\".\n",
6288 SvPVX_const(sv));});
6291 /* A temporary algorithm prefers floated substr to fixed one to dig more info. */
6292 if (longest_fixed_length > longest_float_length) {
6293 r->check_end_shift = r->anchored_end_shift;
6294 r->check_substr = r->anchored_substr;
6295 r->check_utf8 = r->anchored_utf8;
6296 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6297 if (r->extflags & RXf_ANCH_SINGLE)
6298 r->extflags |= RXf_NOSCAN;
6301 r->check_end_shift = r->float_end_shift;
6302 r->check_substr = r->float_substr;
6303 r->check_utf8 = r->float_utf8;
6304 r->check_offset_min = r->float_min_offset;
6305 r->check_offset_max = r->float_max_offset;
6307 /* XXXX Currently intuiting is not compatible with ANCH_GPOS.
6308 This should be changed ASAP! */
6309 if ((r->check_substr || r->check_utf8) && !(r->extflags & RXf_ANCH_GPOS)) {
6310 r->extflags |= RXf_USE_INTUIT;
6311 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6312 r->extflags |= RXf_INTUIT_TAIL;
6314 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6315 if ( (STRLEN)minlen < longest_float_length )
6316 minlen= longest_float_length;
6317 if ( (STRLEN)minlen < longest_fixed_length )
6318 minlen= longest_fixed_length;
6322 /* Several toplevels. Best we can is to set minlen. */
6324 struct regnode_charclass_class ch_class;
6327 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6329 scan = ri->program + 1;
6330 cl_init(pRExC_state, &ch_class);
6331 data.start_class = &ch_class;
6332 data.last_closep = &last_close;
6335 minlen = study_chunk(pRExC_state, &scan, &minlen, &fake, scan + RExC_size,
6336 &data, -1, NULL, NULL, SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS,0);
6338 CHECK_RESTUDY_GOTO_butfirst(NOOP);
6340 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
6341 = r->float_substr = r->float_utf8 = NULL;
6343 if (! TEST_SSC_EOS(data.start_class)
6344 && !cl_is_anything(data.start_class))
6346 const U32 n = add_data(pRExC_state, 1, "f");
6347 OP(data.start_class) = ANYOF_SYNTHETIC;
6349 Newx(RExC_rxi->data->data[n], 1,
6350 struct regnode_charclass_class);
6351 StructCopy(data.start_class,
6352 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6353 struct regnode_charclass_class);
6354 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6355 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6356 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
6357 regprop(r, sv, (regnode*)data.start_class);
6358 PerlIO_printf(Perl_debug_log,
6359 "synthetic stclass \"%s\".\n",
6360 SvPVX_const(sv));});
6364 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
6365 the "real" pattern. */
6367 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf"\n",
6368 (IV)minlen, (IV)r->minlen);
6370 r->minlenret = minlen;
6371 if (r->minlen < minlen)
6374 if (RExC_seen & REG_SEEN_GPOS)
6375 r->extflags |= RXf_GPOS_SEEN;
6376 if (RExC_seen & REG_SEEN_LOOKBEHIND)
6377 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the lookbehind */
6378 if (pRExC_state->num_code_blocks)
6379 r->extflags |= RXf_EVAL_SEEN;
6380 if (RExC_seen & REG_SEEN_CANY)
6381 r->extflags |= RXf_CANY_SEEN;
6382 if (RExC_seen & REG_SEEN_VERBARG)
6384 r->intflags |= PREGf_VERBARG_SEEN;
6385 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
6387 if (RExC_seen & REG_SEEN_CUTGROUP)
6388 r->intflags |= PREGf_CUTGROUP_SEEN;
6389 if (pm_flags & PMf_USE_RE_EVAL)
6390 r->intflags |= PREGf_USE_RE_EVAL;
6391 if (RExC_paren_names)
6392 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
6394 RXp_PAREN_NAMES(r) = NULL;
6397 regnode *first = ri->program + 1;
6399 regnode *next = NEXTOPER(first);
6402 if (PL_regkind[fop] == NOTHING && nop == END)
6403 r->extflags |= RXf_NULL;
6404 else if (PL_regkind[fop] == BOL && nop == END)
6405 r->extflags |= RXf_START_ONLY;
6406 else if (fop == PLUS && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE && OP(regnext(first)) == END)
6407 r->extflags |= RXf_WHITE;
6408 else if ( r->extflags & RXf_SPLIT && fop == EXACT && STR_LEN(first) == 1 && *(STRING(first)) == ' ' && OP(regnext(first)) == END )
6409 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
6413 if (RExC_paren_names) {
6414 ri->name_list_idx = add_data( pRExC_state, 1, "a" );
6415 ri->data->data[ri->name_list_idx] = (void*)SvREFCNT_inc(RExC_paren_name_list);
6418 ri->name_list_idx = 0;
6420 if (RExC_recurse_count) {
6421 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
6422 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
6423 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
6426 Newxz(r->offs, RExC_npar, regexp_paren_pair);
6427 /* assume we don't need to swap parens around before we match */
6430 PerlIO_printf(Perl_debug_log,"Final program:\n");
6433 #ifdef RE_TRACK_PATTERN_OFFSETS
6434 DEBUG_OFFSETS_r(if (ri->u.offsets) {
6435 const U32 len = ri->u.offsets[0];
6437 GET_RE_DEBUG_FLAGS_DECL;
6438 PerlIO_printf(Perl_debug_log, "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
6439 for (i = 1; i <= len; i++) {
6440 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
6441 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
6442 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
6444 PerlIO_printf(Perl_debug_log, "\n");
6449 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
6450 * by setting the regexp SV to readonly-only instead. If the
6451 * pattern's been recompiled, the USEDness should remain. */
6452 if (old_re && SvREADONLY(old_re))
6460 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
6463 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
6465 PERL_UNUSED_ARG(value);
6467 if (flags & RXapif_FETCH) {
6468 return reg_named_buff_fetch(rx, key, flags);
6469 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
6470 Perl_croak_no_modify();
6472 } else if (flags & RXapif_EXISTS) {
6473 return reg_named_buff_exists(rx, key, flags)
6476 } else if (flags & RXapif_REGNAMES) {
6477 return reg_named_buff_all(rx, flags);
6478 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
6479 return reg_named_buff_scalar(rx, flags);
6481 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
6487 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
6490 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
6491 PERL_UNUSED_ARG(lastkey);
6493 if (flags & RXapif_FIRSTKEY)
6494 return reg_named_buff_firstkey(rx, flags);
6495 else if (flags & RXapif_NEXTKEY)
6496 return reg_named_buff_nextkey(rx, flags);
6498 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter", (int)flags);
6504 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
6507 AV *retarray = NULL;
6509 struct regexp *const rx = ReANY(r);
6511 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
6513 if (flags & RXapif_ALL)
6516 if (rx && RXp_PAREN_NAMES(rx)) {
6517 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
6520 SV* sv_dat=HeVAL(he_str);
6521 I32 *nums=(I32*)SvPVX(sv_dat);
6522 for ( i=0; i<SvIVX(sv_dat); i++ ) {
6523 if ((I32)(rx->nparens) >= nums[i]
6524 && rx->offs[nums[i]].start != -1
6525 && rx->offs[nums[i]].end != -1)
6528 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
6533 ret = newSVsv(&PL_sv_undef);
6536 av_push(retarray, ret);
6539 return newRV_noinc(MUTABLE_SV(retarray));
6546 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
6549 struct regexp *const rx = ReANY(r);
6551 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
6553 if (rx && RXp_PAREN_NAMES(rx)) {
6554 if (flags & RXapif_ALL) {
6555 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
6557 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
6559 SvREFCNT_dec_NN(sv);
6571 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
6573 struct regexp *const rx = ReANY(r);
6575 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
6577 if ( rx && RXp_PAREN_NAMES(rx) ) {
6578 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
6580 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
6587 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
6589 struct regexp *const rx = ReANY(r);
6590 GET_RE_DEBUG_FLAGS_DECL;
6592 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
6594 if (rx && RXp_PAREN_NAMES(rx)) {
6595 HV *hv = RXp_PAREN_NAMES(rx);
6597 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6600 SV* sv_dat = HeVAL(temphe);
6601 I32 *nums = (I32*)SvPVX(sv_dat);
6602 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6603 if ((I32)(rx->lastparen) >= nums[i] &&
6604 rx->offs[nums[i]].start != -1 &&
6605 rx->offs[nums[i]].end != -1)
6611 if (parno || flags & RXapif_ALL) {
6612 return newSVhek(HeKEY_hek(temphe));
6620 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
6625 struct regexp *const rx = ReANY(r);
6627 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
6629 if (rx && RXp_PAREN_NAMES(rx)) {
6630 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
6631 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
6632 } else if (flags & RXapif_ONE) {
6633 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
6634 av = MUTABLE_AV(SvRV(ret));
6635 length = av_len(av);
6636 SvREFCNT_dec_NN(ret);
6637 return newSViv(length + 1);
6639 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar", (int)flags);
6643 return &PL_sv_undef;
6647 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
6649 struct regexp *const rx = ReANY(r);
6652 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
6654 if (rx && RXp_PAREN_NAMES(rx)) {
6655 HV *hv= RXp_PAREN_NAMES(rx);
6657 (void)hv_iterinit(hv);
6658 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6661 SV* sv_dat = HeVAL(temphe);
6662 I32 *nums = (I32*)SvPVX(sv_dat);
6663 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6664 if ((I32)(rx->lastparen) >= nums[i] &&
6665 rx->offs[nums[i]].start != -1 &&
6666 rx->offs[nums[i]].end != -1)
6672 if (parno || flags & RXapif_ALL) {
6673 av_push(av, newSVhek(HeKEY_hek(temphe)));
6678 return newRV_noinc(MUTABLE_SV(av));
6682 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
6685 struct regexp *const rx = ReANY(r);
6691 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
6693 if ( n == RX_BUFF_IDX_CARET_PREMATCH
6694 || n == RX_BUFF_IDX_CARET_FULLMATCH
6695 || n == RX_BUFF_IDX_CARET_POSTMATCH
6698 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
6700 /* on something like
6703 * the KEEPCOPY is set on the PMOP rather than the regex */
6704 if (PL_curpm && r == PM_GETRE(PL_curpm))
6705 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
6714 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
6715 /* no need to distinguish between them any more */
6716 n = RX_BUFF_IDX_FULLMATCH;
6718 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
6719 && rx->offs[0].start != -1)
6721 /* $`, ${^PREMATCH} */
6722 i = rx->offs[0].start;
6726 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
6727 && rx->offs[0].end != -1)
6729 /* $', ${^POSTMATCH} */
6730 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
6731 i = rx->sublen + rx->suboffset - rx->offs[0].end;
6734 if ( 0 <= n && n <= (I32)rx->nparens &&
6735 (s1 = rx->offs[n].start) != -1 &&
6736 (t1 = rx->offs[n].end) != -1)
6738 /* $&, ${^MATCH}, $1 ... */
6740 s = rx->subbeg + s1 - rx->suboffset;
6745 assert(s >= rx->subbeg);
6746 assert(rx->sublen >= (s - rx->subbeg) + i );
6748 #if NO_TAINT_SUPPORT
6749 sv_setpvn(sv, s, i);
6751 const int oldtainted = TAINT_get;
6753 sv_setpvn(sv, s, i);
6754 TAINT_set(oldtainted);
6756 if ( (rx->extflags & RXf_CANY_SEEN)
6757 ? (RXp_MATCH_UTF8(rx)
6758 && (!i || is_utf8_string((U8*)s, i)))
6759 : (RXp_MATCH_UTF8(rx)) )
6766 if (RXp_MATCH_TAINTED(rx)) {
6767 if (SvTYPE(sv) >= SVt_PVMG) {
6768 MAGIC* const mg = SvMAGIC(sv);
6771 SvMAGIC_set(sv, mg->mg_moremagic);
6773 if ((mgt = SvMAGIC(sv))) {
6774 mg->mg_moremagic = mgt;
6775 SvMAGIC_set(sv, mg);
6786 sv_setsv(sv,&PL_sv_undef);
6792 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
6793 SV const * const value)
6795 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
6797 PERL_UNUSED_ARG(rx);
6798 PERL_UNUSED_ARG(paren);
6799 PERL_UNUSED_ARG(value);
6802 Perl_croak_no_modify();
6806 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
6809 struct regexp *const rx = ReANY(r);
6813 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
6815 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
6816 || paren == RX_BUFF_IDX_CARET_FULLMATCH
6817 || paren == RX_BUFF_IDX_CARET_POSTMATCH
6820 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
6822 /* on something like
6825 * the KEEPCOPY is set on the PMOP rather than the regex */
6826 if (PL_curpm && r == PM_GETRE(PL_curpm))
6827 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
6833 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
6835 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
6836 case RX_BUFF_IDX_PREMATCH: /* $` */
6837 if (rx->offs[0].start != -1) {
6838 i = rx->offs[0].start;
6847 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
6848 case RX_BUFF_IDX_POSTMATCH: /* $' */
6849 if (rx->offs[0].end != -1) {
6850 i = rx->sublen - rx->offs[0].end;
6852 s1 = rx->offs[0].end;
6859 default: /* $& / ${^MATCH}, $1, $2, ... */
6860 if (paren <= (I32)rx->nparens &&
6861 (s1 = rx->offs[paren].start) != -1 &&
6862 (t1 = rx->offs[paren].end) != -1)
6868 if (ckWARN(WARN_UNINITIALIZED))
6869 report_uninit((const SV *)sv);
6874 if (i > 0 && RXp_MATCH_UTF8(rx)) {
6875 const char * const s = rx->subbeg - rx->suboffset + s1;
6880 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
6887 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
6889 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
6890 PERL_UNUSED_ARG(rx);
6894 return newSVpvs("Regexp");
6897 /* Scans the name of a named buffer from the pattern.
6898 * If flags is REG_RSN_RETURN_NULL returns null.
6899 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
6900 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
6901 * to the parsed name as looked up in the RExC_paren_names hash.
6902 * If there is an error throws a vFAIL().. type exception.
6905 #define REG_RSN_RETURN_NULL 0
6906 #define REG_RSN_RETURN_NAME 1
6907 #define REG_RSN_RETURN_DATA 2
6910 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
6912 char *name_start = RExC_parse;
6914 PERL_ARGS_ASSERT_REG_SCAN_NAME;
6916 if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
6917 /* skip IDFIRST by using do...while */
6920 RExC_parse += UTF8SKIP(RExC_parse);
6921 } while (isWORDCHAR_utf8((U8*)RExC_parse));
6925 } while (isWORDCHAR(*RExC_parse));
6927 RExC_parse++; /* so the <- from the vFAIL is after the offending character */
6928 vFAIL("Group name must start with a non-digit word character");
6932 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
6933 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
6934 if ( flags == REG_RSN_RETURN_NAME)
6936 else if (flags==REG_RSN_RETURN_DATA) {
6939 if ( ! sv_name ) /* should not happen*/
6940 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
6941 if (RExC_paren_names)
6942 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
6944 sv_dat = HeVAL(he_str);
6946 vFAIL("Reference to nonexistent named group");
6950 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
6951 (unsigned long) flags);
6953 assert(0); /* NOT REACHED */
6958 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
6959 int rem=(int)(RExC_end - RExC_parse); \
6968 if (RExC_lastparse!=RExC_parse) \
6969 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
6972 iscut ? "..." : "<" \
6975 PerlIO_printf(Perl_debug_log,"%16s",""); \
6978 num = RExC_size + 1; \
6980 num=REG_NODE_NUM(RExC_emit); \
6981 if (RExC_lastnum!=num) \
6982 PerlIO_printf(Perl_debug_log,"|%4d",num); \
6984 PerlIO_printf(Perl_debug_log,"|%4s",""); \
6985 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
6986 (int)((depth*2)), "", \
6990 RExC_lastparse=RExC_parse; \
6995 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
6996 DEBUG_PARSE_MSG((funcname)); \
6997 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
6999 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7000 DEBUG_PARSE_MSG((funcname)); \
7001 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7004 /* This section of code defines the inversion list object and its methods. The
7005 * interfaces are highly subject to change, so as much as possible is static to
7006 * this file. An inversion list is here implemented as a malloc'd C UV array
7007 * with some added info that is placed as UVs at the beginning in a header
7008 * portion. An inversion list for Unicode is an array of code points, sorted
7009 * by ordinal number. The zeroth element is the first code point in the list.
7010 * The 1th element is the first element beyond that not in the list. In other
7011 * words, the first range is
7012 * invlist[0]..(invlist[1]-1)
7013 * The other ranges follow. Thus every element whose index is divisible by two
7014 * marks the beginning of a range that is in the list, and every element not
7015 * divisible by two marks the beginning of a range not in the list. A single
7016 * element inversion list that contains the single code point N generally
7017 * consists of two elements
7020 * (The exception is when N is the highest representable value on the
7021 * machine, in which case the list containing just it would be a single
7022 * element, itself. By extension, if the last range in the list extends to
7023 * infinity, then the first element of that range will be in the inversion list
7024 * at a position that is divisible by two, and is the final element in the
7026 * Taking the complement (inverting) an inversion list is quite simple, if the
7027 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7028 * This implementation reserves an element at the beginning of each inversion
7029 * list to contain 0 when the list contains 0, and contains 1 otherwise. The
7030 * actual beginning of the list is either that element if 0, or the next one if
7033 * More about inversion lists can be found in "Unicode Demystified"
7034 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7035 * More will be coming when functionality is added later.
7037 * The inversion list data structure is currently implemented as an SV pointing
7038 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7039 * array of UV whose memory management is automatically handled by the existing
7040 * facilities for SV's.
7042 * Some of the methods should always be private to the implementation, and some
7043 * should eventually be made public */
7045 /* The header definitions are in F<inline_invlist.c> */
7046 #define TO_INTERNAL_SIZE(x) (((x) + HEADER_LENGTH) * sizeof(UV))
7047 #define FROM_INTERNAL_SIZE(x) (((x)/ sizeof(UV)) - HEADER_LENGTH)
7049 #define INVLIST_INITIAL_LEN 10
7051 PERL_STATIC_INLINE UV*
7052 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
7054 /* Returns a pointer to the first element in the inversion list's array.
7055 * This is called upon initialization of an inversion list. Where the
7056 * array begins depends on whether the list has the code point U+0000
7057 * in it or not. The other parameter tells it whether the code that
7058 * follows this call is about to put a 0 in the inversion list or not.
7059 * The first element is either the element with 0, if 0, or the next one,
7062 UV* zero = get_invlist_zero_addr(invlist);
7064 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7067 assert(! *_get_invlist_len_addr(invlist));
7069 /* 1^1 = 0; 1^0 = 1 */
7070 *zero = 1 ^ will_have_0;
7071 return zero + *zero;
7074 PERL_STATIC_INLINE UV*
7075 S_invlist_array(pTHX_ SV* const invlist)
7077 /* Returns the pointer to the inversion list's array. Every time the
7078 * length changes, this needs to be called in case malloc or realloc moved
7081 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7083 /* Must not be empty. If these fail, you probably didn't check for <len>
7084 * being non-zero before trying to get the array */
7085 assert(*_get_invlist_len_addr(invlist));
7086 assert(*get_invlist_zero_addr(invlist) == 0
7087 || *get_invlist_zero_addr(invlist) == 1);
7089 /* The array begins either at the element reserved for zero if the
7090 * list contains 0 (that element will be set to 0), or otherwise the next
7091 * element (in which case the reserved element will be set to 1). */
7092 return (UV *) (get_invlist_zero_addr(invlist)
7093 + *get_invlist_zero_addr(invlist));
7096 PERL_STATIC_INLINE void
7097 S_invlist_set_len(pTHX_ SV* const invlist, const UV len)
7099 /* Sets the current number of elements stored in the inversion list */
7101 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7103 *_get_invlist_len_addr(invlist) = len;
7105 assert(len <= SvLEN(invlist));
7107 SvCUR_set(invlist, TO_INTERNAL_SIZE(len));
7108 /* If the list contains U+0000, that element is part of the header,
7109 * and should not be counted as part of the array. It will contain
7110 * 0 in that case, and 1 otherwise. So we could flop 0=>1, 1=>0 and
7112 * SvCUR_set(invlist,
7113 * TO_INTERNAL_SIZE(len
7114 * - (*get_invlist_zero_addr(inv_list) ^ 1)));
7115 * But, this is only valid if len is not 0. The consequences of not doing
7116 * this is that the memory allocation code may think that 1 more UV is
7117 * being used than actually is, and so might do an unnecessary grow. That
7118 * seems worth not bothering to make this the precise amount.
7120 * Note that when inverting, SvCUR shouldn't change */
7123 PERL_STATIC_INLINE IV*
7124 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7126 /* Return the address of the UV that is reserved to hold the cached index
7129 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7131 return (IV *) (SvPVX(invlist) + (INVLIST_PREVIOUS_INDEX_OFFSET * sizeof (UV)));
7134 PERL_STATIC_INLINE IV
7135 S_invlist_previous_index(pTHX_ SV* const invlist)
7137 /* Returns cached index of previous search */
7139 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7141 return *get_invlist_previous_index_addr(invlist);
7144 PERL_STATIC_INLINE void
7145 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7147 /* Caches <index> for later retrieval */
7149 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7151 assert(index == 0 || index < (int) _invlist_len(invlist));
7153 *get_invlist_previous_index_addr(invlist) = index;
7156 PERL_STATIC_INLINE UV
7157 S_invlist_max(pTHX_ SV* const invlist)
7159 /* Returns the maximum number of elements storable in the inversion list's
7160 * array, without having to realloc() */
7162 PERL_ARGS_ASSERT_INVLIST_MAX;
7164 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7165 ? _invlist_len(invlist)
7166 : FROM_INTERNAL_SIZE(SvLEN(invlist));
7169 PERL_STATIC_INLINE UV*
7170 S_get_invlist_zero_addr(pTHX_ SV* invlist)
7172 /* Return the address of the UV that is reserved to hold 0 if the inversion
7173 * list contains 0. This has to be the last element of the heading, as the
7174 * list proper starts with either it if 0, or the next element if not.
7175 * (But we force it to contain either 0 or 1) */
7177 PERL_ARGS_ASSERT_GET_INVLIST_ZERO_ADDR;
7179 return (UV *) (SvPVX(invlist) + (INVLIST_ZERO_OFFSET * sizeof (UV)));
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) {
7194 initial_size = INVLIST_INITIAL_LEN;
7197 /* Allocate the initial space */
7198 new_list = newSV(TO_INTERNAL_SIZE(initial_size));
7199 invlist_set_len(new_list, 0);
7201 /* Force iterinit() to be used to get iteration to work */
7202 *get_invlist_iter_addr(new_list) = UV_MAX;
7204 /* This should force a segfault if a method doesn't initialize this
7206 *get_invlist_zero_addr(new_list) = UV_MAX;
7208 *get_invlist_previous_index_addr(new_list) = 0;
7209 *get_invlist_version_id_addr(new_list) = INVLIST_VERSION_ID;
7210 #if HEADER_LENGTH != 5
7211 # error Need to regenerate INVLIST_VERSION_ID by running perl -E 'say int(rand 2**31-1)', and then changing the #if to the new length
7219 S__new_invlist_C_array(pTHX_ UV* list)
7221 /* Return a pointer to a newly constructed inversion list, initialized to
7222 * point to <list>, which has to be in the exact correct inversion list
7223 * form, including internal fields. Thus this is a dangerous routine that
7224 * should not be used in the wrong hands */
7226 SV* invlist = newSV_type(SVt_PV);
7228 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7230 SvPV_set(invlist, (char *) list);
7231 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7232 shouldn't touch it */
7233 SvCUR_set(invlist, TO_INTERNAL_SIZE(_invlist_len(invlist)));
7235 if (*get_invlist_version_id_addr(invlist) != INVLIST_VERSION_ID) {
7236 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7239 /* Initialize the iteration pointer.
7240 * XXX This could be done at compile time in charclass_invlists.h, but I
7241 * (khw) am not confident that the suffixes for specifying the C constant
7242 * UV_MAX are portable, e.g. 'ull' on a 32 bit machine that is configured
7243 * to use 64 bits; might need a Configure probe */
7244 invlist_iterfinish(invlist);
7250 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7252 /* Grow the maximum size of an inversion list */
7254 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7256 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max));
7259 PERL_STATIC_INLINE void
7260 S_invlist_trim(pTHX_ SV* const invlist)
7262 PERL_ARGS_ASSERT_INVLIST_TRIM;
7264 /* Change the length of the inversion list to how many entries it currently
7267 SvPV_shrink_to_cur((SV *) invlist);
7270 #define _invlist_union_complement_2nd(a, b, output) _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
7273 S__append_range_to_invlist(pTHX_ SV* const invlist, const UV start, const UV end)
7275 /* Subject to change or removal. Append the range from 'start' to 'end' at
7276 * the end of the inversion list. The range must be above any existing
7280 UV max = invlist_max(invlist);
7281 UV len = _invlist_len(invlist);
7283 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7285 if (len == 0) { /* Empty lists must be initialized */
7286 array = _invlist_array_init(invlist, start == 0);
7289 /* Here, the existing list is non-empty. The current max entry in the
7290 * list is generally the first value not in the set, except when the
7291 * set extends to the end of permissible values, in which case it is
7292 * the first entry in that final set, and so this call is an attempt to
7293 * append out-of-order */
7295 UV final_element = len - 1;
7296 array = invlist_array(invlist);
7297 if (array[final_element] > start
7298 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
7300 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",
7301 array[final_element], start,
7302 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
7305 /* Here, it is a legal append. If the new range begins with the first
7306 * value not in the set, it is extending the set, so the new first
7307 * value not in the set is one greater than the newly extended range.
7309 if (array[final_element] == start) {
7310 if (end != UV_MAX) {
7311 array[final_element] = end + 1;
7314 /* But if the end is the maximum representable on the machine,
7315 * just let the range that this would extend to have no end */
7316 invlist_set_len(invlist, len - 1);
7322 /* Here the new range doesn't extend any existing set. Add it */
7324 len += 2; /* Includes an element each for the start and end of range */
7326 /* If overflows the existing space, extend, which may cause the array to be
7329 invlist_extend(invlist, len);
7330 invlist_set_len(invlist, len); /* Have to set len here to avoid assert
7331 failure in invlist_array() */
7332 array = invlist_array(invlist);
7335 invlist_set_len(invlist, len);
7338 /* The next item on the list starts the range, the one after that is
7339 * one past the new range. */
7340 array[len - 2] = start;
7341 if (end != UV_MAX) {
7342 array[len - 1] = end + 1;
7345 /* But if the end is the maximum representable on the machine, just let
7346 * the range have no end */
7347 invlist_set_len(invlist, len - 1);
7351 #ifndef PERL_IN_XSUB_RE
7354 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
7356 /* Searches the inversion list for the entry that contains the input code
7357 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
7358 * return value is the index into the list's array of the range that
7363 IV high = _invlist_len(invlist);
7364 const IV highest_element = high - 1;
7367 PERL_ARGS_ASSERT__INVLIST_SEARCH;
7369 /* If list is empty, return failure. */
7374 /* (We can't get the array unless we know the list is non-empty) */
7375 array = invlist_array(invlist);
7377 mid = invlist_previous_index(invlist);
7378 assert(mid >=0 && mid <= highest_element);
7380 /* <mid> contains the cache of the result of the previous call to this
7381 * function (0 the first time). See if this call is for the same result,
7382 * or if it is for mid-1. This is under the theory that calls to this
7383 * function will often be for related code points that are near each other.
7384 * And benchmarks show that caching gives better results. We also test
7385 * here if the code point is within the bounds of the list. These tests
7386 * replace others that would have had to be made anyway to make sure that
7387 * the array bounds were not exceeded, and these give us extra information
7388 * at the same time */
7389 if (cp >= array[mid]) {
7390 if (cp >= array[highest_element]) {
7391 return highest_element;
7394 /* Here, array[mid] <= cp < array[highest_element]. This means that
7395 * the final element is not the answer, so can exclude it; it also
7396 * means that <mid> is not the final element, so can refer to 'mid + 1'
7398 if (cp < array[mid + 1]) {
7404 else { /* cp < aray[mid] */
7405 if (cp < array[0]) { /* Fail if outside the array */
7409 if (cp >= array[mid - 1]) {
7414 /* Binary search. What we are looking for is <i> such that
7415 * array[i] <= cp < array[i+1]
7416 * The loop below converges on the i+1. Note that there may not be an
7417 * (i+1)th element in the array, and things work nonetheless */
7418 while (low < high) {
7419 mid = (low + high) / 2;
7420 assert(mid <= highest_element);
7421 if (array[mid] <= cp) { /* cp >= array[mid] */
7424 /* We could do this extra test to exit the loop early.
7425 if (cp < array[low]) {
7430 else { /* cp < array[mid] */
7437 invlist_set_previous_index(invlist, high);
7442 Perl__invlist_populate_swatch(pTHX_ SV* const invlist, const UV start, const UV end, U8* swatch)
7444 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
7445 * but is used when the swash has an inversion list. This makes this much
7446 * faster, as it uses a binary search instead of a linear one. This is
7447 * intimately tied to that function, and perhaps should be in utf8.c,
7448 * except it is intimately tied to inversion lists as well. It assumes
7449 * that <swatch> is all 0's on input */
7452 const IV len = _invlist_len(invlist);
7456 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
7458 if (len == 0) { /* Empty inversion list */
7462 array = invlist_array(invlist);
7464 /* Find which element it is */
7465 i = _invlist_search(invlist, start);
7467 /* We populate from <start> to <end> */
7468 while (current < end) {
7471 /* The inversion list gives the results for every possible code point
7472 * after the first one in the list. Only those ranges whose index is
7473 * even are ones that the inversion list matches. For the odd ones,
7474 * and if the initial code point is not in the list, we have to skip
7475 * forward to the next element */
7476 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
7478 if (i >= len) { /* Finished if beyond the end of the array */
7482 if (current >= end) { /* Finished if beyond the end of what we
7484 if (LIKELY(end < UV_MAX)) {
7488 /* We get here when the upper bound is the maximum
7489 * representable on the machine, and we are looking for just
7490 * that code point. Have to special case it */
7492 goto join_end_of_list;
7495 assert(current >= start);
7497 /* The current range ends one below the next one, except don't go past
7500 upper = (i < len && array[i] < end) ? array[i] : end;
7502 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
7503 * for each code point in it */
7504 for (; current < upper; current++) {
7505 const STRLEN offset = (STRLEN)(current - start);
7506 swatch[offset >> 3] |= 1 << (offset & 7);
7511 /* Quit if at the end of the list */
7514 /* But first, have to deal with the highest possible code point on
7515 * the platform. The previous code assumes that <end> is one
7516 * beyond where we want to populate, but that is impossible at the
7517 * platform's infinity, so have to handle it specially */
7518 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
7520 const STRLEN offset = (STRLEN)(end - start);
7521 swatch[offset >> 3] |= 1 << (offset & 7);
7526 /* Advance to the next range, which will be for code points not in the
7535 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** output)
7537 /* Take the union of two inversion lists and point <output> to it. *output
7538 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7539 * the reference count to that list will be decremented. The first list,
7540 * <a>, may be NULL, in which case a copy of the second list is returned.
7541 * If <complement_b> is TRUE, the union is taken of the complement
7542 * (inversion) of <b> instead of b itself.
7544 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7545 * Richard Gillam, published by Addison-Wesley, and explained at some
7546 * length there. The preface says to incorporate its examples into your
7547 * code at your own risk.
7549 * The algorithm is like a merge sort.
7551 * XXX A potential performance improvement is to keep track as we go along
7552 * if only one of the inputs contributes to the result, meaning the other
7553 * is a subset of that one. In that case, we can skip the final copy and
7554 * return the larger of the input lists, but then outside code might need
7555 * to keep track of whether to free the input list or not */
7557 UV* array_a; /* a's array */
7559 UV len_a; /* length of a's array */
7562 SV* u; /* the resulting union */
7566 UV i_a = 0; /* current index into a's array */
7570 /* running count, as explained in the algorithm source book; items are
7571 * stopped accumulating and are output when the count changes to/from 0.
7572 * The count is incremented when we start a range that's in the set, and
7573 * decremented when we start a range that's not in the set. So its range
7574 * is 0 to 2. Only when the count is zero is something not in the set.
7578 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
7581 /* If either one is empty, the union is the other one */
7582 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
7589 *output = invlist_clone(b);
7591 _invlist_invert(*output);
7593 } /* else *output already = b; */
7596 else if ((len_b = _invlist_len(b)) == 0) {
7601 /* The complement of an empty list is a list that has everything in it,
7602 * so the union with <a> includes everything too */
7607 *output = _new_invlist(1);
7608 _append_range_to_invlist(*output, 0, UV_MAX);
7610 else if (*output != a) {
7611 *output = invlist_clone(a);
7613 /* else *output already = a; */
7617 /* Here both lists exist and are non-empty */
7618 array_a = invlist_array(a);
7619 array_b = invlist_array(b);
7621 /* If are to take the union of 'a' with the complement of b, set it
7622 * up so are looking at b's complement. */
7625 /* To complement, we invert: if the first element is 0, remove it. To
7626 * do this, we just pretend the array starts one later, and clear the
7627 * flag as we don't have to do anything else later */
7628 if (array_b[0] == 0) {
7631 complement_b = FALSE;
7635 /* But if the first element is not zero, we unshift a 0 before the
7636 * array. The data structure reserves a space for that 0 (which
7637 * should be a '1' right now), so physical shifting is unneeded,
7638 * but temporarily change that element to 0. Before exiting the
7639 * routine, we must restore the element to '1' */
7646 /* Size the union for the worst case: that the sets are completely
7648 u = _new_invlist(len_a + len_b);
7650 /* Will contain U+0000 if either component does */
7651 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
7652 || (len_b > 0 && array_b[0] == 0));
7654 /* Go through each list item by item, stopping when exhausted one of
7656 while (i_a < len_a && i_b < len_b) {
7657 UV cp; /* The element to potentially add to the union's array */
7658 bool cp_in_set; /* is it in the the input list's set or not */
7660 /* We need to take one or the other of the two inputs for the union.
7661 * Since we are merging two sorted lists, we take the smaller of the
7662 * next items. In case of a tie, we take the one that is in its set
7663 * first. If we took one not in the set first, it would decrement the
7664 * count, possibly to 0 which would cause it to be output as ending the
7665 * range, and the next time through we would take the same number, and
7666 * output it again as beginning the next range. By doing it the
7667 * opposite way, there is no possibility that the count will be
7668 * momentarily decremented to 0, and thus the two adjoining ranges will
7669 * be seamlessly merged. (In a tie and both are in the set or both not
7670 * in the set, it doesn't matter which we take first.) */
7671 if (array_a[i_a] < array_b[i_b]
7672 || (array_a[i_a] == array_b[i_b]
7673 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7675 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7679 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7680 cp = array_b[i_b++];
7683 /* Here, have chosen which of the two inputs to look at. Only output
7684 * if the running count changes to/from 0, which marks the
7685 * beginning/end of a range in that's in the set */
7688 array_u[i_u++] = cp;
7695 array_u[i_u++] = cp;
7700 /* Here, we are finished going through at least one of the lists, which
7701 * means there is something remaining in at most one. We check if the list
7702 * that hasn't been exhausted is positioned such that we are in the middle
7703 * of a range in its set or not. (i_a and i_b point to the element beyond
7704 * the one we care about.) If in the set, we decrement 'count'; if 0, there
7705 * is potentially more to output.
7706 * There are four cases:
7707 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
7708 * in the union is entirely from the non-exhausted set.
7709 * 2) Both were in their sets, count is 2. Nothing further should
7710 * be output, as everything that remains will be in the exhausted
7711 * list's set, hence in the union; decrementing to 1 but not 0 insures
7713 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
7714 * Nothing further should be output because the union includes
7715 * everything from the exhausted set. Not decrementing ensures that.
7716 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
7717 * decrementing to 0 insures that we look at the remainder of the
7718 * non-exhausted set */
7719 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7720 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7725 /* The final length is what we've output so far, plus what else is about to
7726 * be output. (If 'count' is non-zero, then the input list we exhausted
7727 * has everything remaining up to the machine's limit in its set, and hence
7728 * in the union, so there will be no further output. */
7731 /* At most one of the subexpressions will be non-zero */
7732 len_u += (len_a - i_a) + (len_b - i_b);
7735 /* Set result to final length, which can change the pointer to array_u, so
7737 if (len_u != _invlist_len(u)) {
7738 invlist_set_len(u, len_u);
7740 array_u = invlist_array(u);
7743 /* When 'count' is 0, the list that was exhausted (if one was shorter than
7744 * the other) ended with everything above it not in its set. That means
7745 * that the remaining part of the union is precisely the same as the
7746 * non-exhausted list, so can just copy it unchanged. (If both list were
7747 * exhausted at the same time, then the operations below will be both 0.)
7750 IV copy_count; /* At most one will have a non-zero copy count */
7751 if ((copy_count = len_a - i_a) > 0) {
7752 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
7754 else if ((copy_count = len_b - i_b) > 0) {
7755 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
7759 /* If we've changed b, restore it */
7764 /* We may be removing a reference to one of the inputs */
7765 if (a == *output || b == *output) {
7766 assert(! invlist_is_iterating(*output));
7767 SvREFCNT_dec_NN(*output);
7775 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** i)
7777 /* Take the intersection of two inversion lists and point <i> to it. *i
7778 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7779 * the reference count to that list will be decremented.
7780 * If <complement_b> is TRUE, the result will be the intersection of <a>
7781 * and the complement (or inversion) of <b> instead of <b> directly.
7783 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7784 * Richard Gillam, published by Addison-Wesley, and explained at some
7785 * length there. The preface says to incorporate its examples into your
7786 * code at your own risk. In fact, it had bugs
7788 * The algorithm is like a merge sort, and is essentially the same as the
7792 UV* array_a; /* a's array */
7794 UV len_a; /* length of a's array */
7797 SV* r; /* the resulting intersection */
7801 UV i_a = 0; /* current index into a's array */
7805 /* running count, as explained in the algorithm source book; items are
7806 * stopped accumulating and are output when the count changes to/from 2.
7807 * The count is incremented when we start a range that's in the set, and
7808 * decremented when we start a range that's not in the set. So its range
7809 * is 0 to 2. Only when the count is 2 is something in the intersection.
7813 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
7816 /* Special case if either one is empty */
7817 len_a = _invlist_len(a);
7818 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
7820 if (len_a != 0 && complement_b) {
7822 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
7823 * be empty. Here, also we are using 'b's complement, which hence
7824 * must be every possible code point. Thus the intersection is
7827 *i = invlist_clone(a);
7833 /* else *i is already 'a' */
7837 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
7838 * intersection must be empty */
7845 *i = _new_invlist(0);
7849 /* Here both lists exist and are non-empty */
7850 array_a = invlist_array(a);
7851 array_b = invlist_array(b);
7853 /* If are to take the intersection of 'a' with the complement of b, set it
7854 * up so are looking at b's complement. */
7857 /* To complement, we invert: if the first element is 0, remove it. To
7858 * do this, we just pretend the array starts one later, and clear the
7859 * flag as we don't have to do anything else later */
7860 if (array_b[0] == 0) {
7863 complement_b = FALSE;
7867 /* But if the first element is not zero, we unshift a 0 before the
7868 * array. The data structure reserves a space for that 0 (which
7869 * should be a '1' right now), so physical shifting is unneeded,
7870 * but temporarily change that element to 0. Before exiting the
7871 * routine, we must restore the element to '1' */
7878 /* Size the intersection for the worst case: that the intersection ends up
7879 * fragmenting everything to be completely disjoint */
7880 r= _new_invlist(len_a + len_b);
7882 /* Will contain U+0000 iff both components do */
7883 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
7884 && len_b > 0 && array_b[0] == 0);
7886 /* Go through each list item by item, stopping when exhausted one of
7888 while (i_a < len_a && i_b < len_b) {
7889 UV cp; /* The element to potentially add to the intersection's
7891 bool cp_in_set; /* Is it in the input list's set or not */
7893 /* We need to take one or the other of the two inputs for the
7894 * intersection. Since we are merging two sorted lists, we take the
7895 * smaller of the next items. In case of a tie, we take the one that
7896 * is not in its set first (a difference from the union algorithm). If
7897 * we took one in the set first, it would increment the count, possibly
7898 * to 2 which would cause it to be output as starting a range in the
7899 * intersection, and the next time through we would take that same
7900 * number, and output it again as ending the set. By doing it the
7901 * opposite of this, there is no possibility that the count will be
7902 * momentarily incremented to 2. (In a tie and both are in the set or
7903 * both not in the set, it doesn't matter which we take first.) */
7904 if (array_a[i_a] < array_b[i_b]
7905 || (array_a[i_a] == array_b[i_b]
7906 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7908 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7912 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7916 /* Here, have chosen which of the two inputs to look at. Only output
7917 * if the running count changes to/from 2, which marks the
7918 * beginning/end of a range that's in the intersection */
7922 array_r[i_r++] = cp;
7927 array_r[i_r++] = cp;
7933 /* Here, we are finished going through at least one of the lists, which
7934 * means there is something remaining in at most one. We check if the list
7935 * that has been exhausted is positioned such that we are in the middle
7936 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
7937 * the ones we care about.) There are four cases:
7938 * 1) Both weren't in their sets, count is 0, and remains 0. There's
7939 * nothing left in the intersection.
7940 * 2) Both were in their sets, count is 2 and perhaps is incremented to
7941 * above 2. What should be output is exactly that which is in the
7942 * non-exhausted set, as everything it has is also in the intersection
7943 * set, and everything it doesn't have can't be in the intersection
7944 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
7945 * gets incremented to 2. Like the previous case, the intersection is
7946 * everything that remains in the non-exhausted set.
7947 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
7948 * remains 1. And the intersection has nothing more. */
7949 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7950 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7955 /* The final length is what we've output so far plus what else is in the
7956 * intersection. At most one of the subexpressions below will be non-zero */
7959 len_r += (len_a - i_a) + (len_b - i_b);
7962 /* Set result to final length, which can change the pointer to array_r, so
7964 if (len_r != _invlist_len(r)) {
7965 invlist_set_len(r, len_r);
7967 array_r = invlist_array(r);
7970 /* Finish outputting any remaining */
7971 if (count >= 2) { /* At most one will have a non-zero copy count */
7973 if ((copy_count = len_a - i_a) > 0) {
7974 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
7976 else if ((copy_count = len_b - i_b) > 0) {
7977 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
7981 /* If we've changed b, restore it */
7986 /* We may be removing a reference to one of the inputs */
7987 if (a == *i || b == *i) {
7988 assert(! invlist_is_iterating(*i));
7989 SvREFCNT_dec_NN(*i);
7997 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
7999 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8000 * set. A pointer to the inversion list is returned. This may actually be
8001 * a new list, in which case the passed in one has been destroyed. The
8002 * passed in inversion list can be NULL, in which case a new one is created
8003 * with just the one range in it */
8008 if (invlist == NULL) {
8009 invlist = _new_invlist(2);
8013 len = _invlist_len(invlist);
8016 /* If comes after the final entry actually in the list, can just append it
8019 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
8020 && start >= invlist_array(invlist)[len - 1]))
8022 _append_range_to_invlist(invlist, start, end);
8026 /* Here, can't just append things, create and return a new inversion list
8027 * which is the union of this range and the existing inversion list */
8028 range_invlist = _new_invlist(2);
8029 _append_range_to_invlist(range_invlist, start, end);
8031 _invlist_union(invlist, range_invlist, &invlist);
8033 /* The temporary can be freed */
8034 SvREFCNT_dec_NN(range_invlist);
8041 PERL_STATIC_INLINE SV*
8042 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
8043 return _add_range_to_invlist(invlist, cp, cp);
8046 #ifndef PERL_IN_XSUB_RE
8048 Perl__invlist_invert(pTHX_ SV* const invlist)
8050 /* Complement the input inversion list. This adds a 0 if the list didn't
8051 * have a zero; removes it otherwise. As described above, the data
8052 * structure is set up so that this is very efficient */
8054 UV* len_pos = _get_invlist_len_addr(invlist);
8056 PERL_ARGS_ASSERT__INVLIST_INVERT;
8058 assert(! invlist_is_iterating(invlist));
8060 /* The inverse of matching nothing is matching everything */
8061 if (*len_pos == 0) {
8062 _append_range_to_invlist(invlist, 0, UV_MAX);
8066 /* The exclusive or complents 0 to 1; and 1 to 0. If the result is 1, the
8067 * zero element was a 0, so it is being removed, so the length decrements
8068 * by 1; and vice-versa. SvCUR is unaffected */
8069 if (*get_invlist_zero_addr(invlist) ^= 1) {
8078 Perl__invlist_invert_prop(pTHX_ SV* const invlist)
8080 /* Complement the input inversion list (which must be a Unicode property,
8081 * all of which don't match above the Unicode maximum code point.) And
8082 * Perl has chosen to not have the inversion match above that either. This
8083 * adds a 0x110000 if the list didn't end with it, and removes it if it did
8089 PERL_ARGS_ASSERT__INVLIST_INVERT_PROP;
8091 _invlist_invert(invlist);
8093 len = _invlist_len(invlist);
8095 if (len != 0) { /* If empty do nothing */
8096 array = invlist_array(invlist);
8097 if (array[len - 1] != PERL_UNICODE_MAX + 1) {
8098 /* Add 0x110000. First, grow if necessary */
8100 if (invlist_max(invlist) < len) {
8101 invlist_extend(invlist, len);
8102 array = invlist_array(invlist);
8104 invlist_set_len(invlist, len);
8105 array[len - 1] = PERL_UNICODE_MAX + 1;
8107 else { /* Remove the 0x110000 */
8108 invlist_set_len(invlist, len - 1);
8116 PERL_STATIC_INLINE SV*
8117 S_invlist_clone(pTHX_ SV* const invlist)
8120 /* Return a new inversion list that is a copy of the input one, which is
8123 /* Need to allocate extra space to accommodate Perl's addition of a
8124 * trailing NUL to SvPV's, since it thinks they are always strings */
8125 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8126 STRLEN length = SvCUR(invlist);
8128 PERL_ARGS_ASSERT_INVLIST_CLONE;
8130 SvCUR_set(new_invlist, length); /* This isn't done automatically */
8131 Copy(SvPVX(invlist), SvPVX(new_invlist), length, char);
8136 PERL_STATIC_INLINE UV*
8137 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8139 /* Return the address of the UV that contains the current iteration
8142 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8144 return (UV *) (SvPVX(invlist) + (INVLIST_ITER_OFFSET * sizeof (UV)));
8147 PERL_STATIC_INLINE UV*
8148 S_get_invlist_version_id_addr(pTHX_ SV* invlist)
8150 /* Return the address of the UV that contains the version id. */
8152 PERL_ARGS_ASSERT_GET_INVLIST_VERSION_ID_ADDR;
8154 return (UV *) (SvPVX(invlist) + (INVLIST_VERSION_ID_OFFSET * sizeof (UV)));
8157 PERL_STATIC_INLINE void
8158 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8160 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8162 *get_invlist_iter_addr(invlist) = 0;
8165 PERL_STATIC_INLINE void
8166 S_invlist_iterfinish(pTHX_ SV* invlist)
8168 /* Terminate iterator for invlist. This is to catch development errors.
8169 * Any iteration that is interrupted before completed should call this
8170 * function. Functions that add code points anywhere else but to the end
8171 * of an inversion list assert that they are not in the middle of an
8172 * iteration. If they were, the addition would make the iteration
8173 * problematical: if the iteration hadn't reached the place where things
8174 * were being added, it would be ok */
8176 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8178 *get_invlist_iter_addr(invlist) = UV_MAX;
8182 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8184 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8185 * This call sets in <*start> and <*end>, the next range in <invlist>.
8186 * Returns <TRUE> if successful and the next call will return the next
8187 * range; <FALSE> if was already at the end of the list. If the latter,
8188 * <*start> and <*end> are unchanged, and the next call to this function
8189 * will start over at the beginning of the list */
8191 UV* pos = get_invlist_iter_addr(invlist);
8192 UV len = _invlist_len(invlist);
8195 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8198 *pos = UV_MAX; /* Force iterinit() to be required next time */
8202 array = invlist_array(invlist);
8204 *start = array[(*pos)++];
8210 *end = array[(*pos)++] - 1;
8216 PERL_STATIC_INLINE bool
8217 S_invlist_is_iterating(pTHX_ SV* const invlist)
8219 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8221 return *(get_invlist_iter_addr(invlist)) < UV_MAX;
8224 PERL_STATIC_INLINE UV
8225 S_invlist_highest(pTHX_ SV* const invlist)
8227 /* Returns the highest code point that matches an inversion list. This API
8228 * has an ambiguity, as it returns 0 under either the highest is actually
8229 * 0, or if the list is empty. If this distinction matters to you, check
8230 * for emptiness before calling this function */
8232 UV len = _invlist_len(invlist);
8235 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8241 array = invlist_array(invlist);
8243 /* The last element in the array in the inversion list always starts a
8244 * range that goes to infinity. That range may be for code points that are
8245 * matched in the inversion list, or it may be for ones that aren't
8246 * matched. In the latter case, the highest code point in the set is one
8247 * less than the beginning of this range; otherwise it is the final element
8248 * of this range: infinity */
8249 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8251 : array[len - 1] - 1;
8254 #ifndef PERL_IN_XSUB_RE
8256 Perl__invlist_contents(pTHX_ SV* const invlist)
8258 /* Get the contents of an inversion list into a string SV so that they can
8259 * be printed out. It uses the format traditionally done for debug tracing
8263 SV* output = newSVpvs("\n");
8265 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8267 assert(! invlist_is_iterating(invlist));
8269 invlist_iterinit(invlist);
8270 while (invlist_iternext(invlist, &start, &end)) {
8271 if (end == UV_MAX) {
8272 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8274 else if (end != start) {
8275 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8279 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
8287 #ifdef PERL_ARGS_ASSERT__INVLIST_DUMP
8289 Perl__invlist_dump(pTHX_ SV* const invlist, const char * const header)
8291 /* Dumps out the ranges in an inversion list. The string 'header'
8292 * if present is output on a line before the first range */
8296 PERL_ARGS_ASSERT__INVLIST_DUMP;
8298 if (header && strlen(header)) {
8299 PerlIO_printf(Perl_debug_log, "%s\n", header);
8301 if (invlist_is_iterating(invlist)) {
8302 PerlIO_printf(Perl_debug_log, "Can't dump because is in middle of iterating\n");
8306 invlist_iterinit(invlist);
8307 while (invlist_iternext(invlist, &start, &end)) {
8308 if (end == UV_MAX) {
8309 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. INFINITY\n", start);
8311 else if (end != start) {
8312 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. 0x%04"UVXf"\n",
8316 PerlIO_printf(Perl_debug_log, "0x%04"UVXf"\n", start);
8324 S__invlistEQ(pTHX_ SV* const a, SV* const b, bool complement_b)
8326 /* Return a boolean as to if the two passed in inversion lists are
8327 * identical. The final argument, if TRUE, says to take the complement of
8328 * the second inversion list before doing the comparison */
8330 UV* array_a = invlist_array(a);
8331 UV* array_b = invlist_array(b);
8332 UV len_a = _invlist_len(a);
8333 UV len_b = _invlist_len(b);
8335 UV i = 0; /* current index into the arrays */
8336 bool retval = TRUE; /* Assume are identical until proven otherwise */
8338 PERL_ARGS_ASSERT__INVLISTEQ;
8340 /* If are to compare 'a' with the complement of b, set it
8341 * up so are looking at b's complement. */
8344 /* The complement of nothing is everything, so <a> would have to have
8345 * just one element, starting at zero (ending at infinity) */
8347 return (len_a == 1 && array_a[0] == 0);
8349 else if (array_b[0] == 0) {
8351 /* Otherwise, to complement, we invert. Here, the first element is
8352 * 0, just remove it. To do this, we just pretend the array starts
8353 * one later, and clear the flag as we don't have to do anything
8358 complement_b = FALSE;
8362 /* But if the first element is not zero, we unshift a 0 before the
8363 * array. The data structure reserves a space for that 0 (which
8364 * should be a '1' right now), so physical shifting is unneeded,
8365 * but temporarily change that element to 0. Before exiting the
8366 * routine, we must restore the element to '1' */
8373 /* Make sure that the lengths are the same, as well as the final element
8374 * before looping through the remainder. (Thus we test the length, final,
8375 * and first elements right off the bat) */
8376 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
8379 else for (i = 0; i < len_a - 1; i++) {
8380 if (array_a[i] != array_b[i]) {
8393 #undef HEADER_LENGTH
8394 #undef INVLIST_INITIAL_LENGTH
8395 #undef TO_INTERNAL_SIZE
8396 #undef FROM_INTERNAL_SIZE
8397 #undef INVLIST_LEN_OFFSET
8398 #undef INVLIST_ZERO_OFFSET
8399 #undef INVLIST_ITER_OFFSET
8400 #undef INVLIST_VERSION_ID
8401 #undef INVLIST_PREVIOUS_INDEX_OFFSET
8403 /* End of inversion list object */
8406 S_parse_lparen_question_flags(pTHX_ struct RExC_state_t *pRExC_state)
8408 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
8409 * constructs, and updates RExC_flags with them. On input, RExC_parse
8410 * should point to the first flag; it is updated on output to point to the
8411 * final ')' or ':'. There needs to be at least one flag, or this will
8414 /* for (?g), (?gc), and (?o) warnings; warning
8415 about (?c) will warn about (?g) -- japhy */
8417 #define WASTED_O 0x01
8418 #define WASTED_G 0x02
8419 #define WASTED_C 0x04
8420 #define WASTED_GC (0x02|0x04)
8421 I32 wastedflags = 0x00;
8422 U32 posflags = 0, negflags = 0;
8423 U32 *flagsp = &posflags;
8424 char has_charset_modifier = '\0';
8426 bool has_use_defaults = FALSE;
8427 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
8429 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
8431 /* '^' as an initial flag sets certain defaults */
8432 if (UCHARAT(RExC_parse) == '^') {
8434 has_use_defaults = TRUE;
8435 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
8436 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
8437 ? REGEX_UNICODE_CHARSET
8438 : REGEX_DEPENDS_CHARSET);
8441 cs = get_regex_charset(RExC_flags);
8442 if (cs == REGEX_DEPENDS_CHARSET
8443 && (RExC_utf8 || RExC_uni_semantics))
8445 cs = REGEX_UNICODE_CHARSET;
8448 while (*RExC_parse) {
8449 /* && strchr("iogcmsx", *RExC_parse) */
8450 /* (?g), (?gc) and (?o) are useless here
8451 and must be globally applied -- japhy */
8452 switch (*RExC_parse) {
8454 /* Code for the imsx flags */
8455 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
8457 case LOCALE_PAT_MOD:
8458 if (has_charset_modifier) {
8459 goto excess_modifier;
8461 else if (flagsp == &negflags) {
8464 cs = REGEX_LOCALE_CHARSET;
8465 has_charset_modifier = LOCALE_PAT_MOD;
8466 RExC_contains_locale = 1;
8468 case UNICODE_PAT_MOD:
8469 if (has_charset_modifier) {
8470 goto excess_modifier;
8472 else if (flagsp == &negflags) {
8475 cs = REGEX_UNICODE_CHARSET;
8476 has_charset_modifier = UNICODE_PAT_MOD;
8478 case ASCII_RESTRICT_PAT_MOD:
8479 if (flagsp == &negflags) {
8482 if (has_charset_modifier) {
8483 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
8484 goto excess_modifier;
8486 /* Doubled modifier implies more restricted */
8487 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
8490 cs = REGEX_ASCII_RESTRICTED_CHARSET;
8492 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
8494 case DEPENDS_PAT_MOD:
8495 if (has_use_defaults) {
8496 goto fail_modifiers;
8498 else if (flagsp == &negflags) {
8501 else if (has_charset_modifier) {
8502 goto excess_modifier;
8505 /* The dual charset means unicode semantics if the
8506 * pattern (or target, not known until runtime) are
8507 * utf8, or something in the pattern indicates unicode
8509 cs = (RExC_utf8 || RExC_uni_semantics)
8510 ? REGEX_UNICODE_CHARSET
8511 : REGEX_DEPENDS_CHARSET;
8512 has_charset_modifier = DEPENDS_PAT_MOD;
8516 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
8517 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
8519 else if (has_charset_modifier == *(RExC_parse - 1)) {
8520 vFAIL2("Regexp modifier \"%c\" may not appear twice", *(RExC_parse - 1));
8523 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
8528 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"", *(RExC_parse - 1));
8530 case ONCE_PAT_MOD: /* 'o' */
8531 case GLOBAL_PAT_MOD: /* 'g' */
8532 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8533 const I32 wflagbit = *RExC_parse == 'o' ? WASTED_O : WASTED_G;
8534 if (! (wastedflags & wflagbit) ) {
8535 wastedflags |= wflagbit;
8538 "Useless (%s%c) - %suse /%c modifier",
8539 flagsp == &negflags ? "?-" : "?",
8541 flagsp == &negflags ? "don't " : "",
8548 case CONTINUE_PAT_MOD: /* 'c' */
8549 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8550 if (! (wastedflags & WASTED_C) ) {
8551 wastedflags |= WASTED_GC;
8554 "Useless (%sc) - %suse /gc modifier",
8555 flagsp == &negflags ? "?-" : "?",
8556 flagsp == &negflags ? "don't " : ""
8561 case KEEPCOPY_PAT_MOD: /* 'p' */
8562 if (flagsp == &negflags) {
8564 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
8566 *flagsp |= RXf_PMf_KEEPCOPY;
8570 /* A flag is a default iff it is following a minus, so
8571 * if there is a minus, it means will be trying to
8572 * re-specify a default which is an error */
8573 if (has_use_defaults || flagsp == &negflags) {
8574 goto fail_modifiers;
8577 wastedflags = 0; /* reset so (?g-c) warns twice */
8581 RExC_flags |= posflags;
8582 RExC_flags &= ~negflags;
8583 set_regex_charset(&RExC_flags, cs);
8589 vFAIL3("Sequence (%.*s...) not recognized",
8590 RExC_parse-seqstart, seqstart);
8599 - reg - regular expression, i.e. main body or parenthesized thing
8601 * Caller must absorb opening parenthesis.
8603 * Combining parenthesis handling with the base level of regular expression
8604 * is a trifle forced, but the need to tie the tails of the branches to what
8605 * follows makes it hard to avoid.
8607 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
8609 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
8611 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
8614 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
8615 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
8616 needs to be restarted.
8617 Otherwise would only return NULL if regbranch() returns NULL, which
8620 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
8621 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
8622 * 2 is like 1, but indicates that nextchar() has been called to advance
8623 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
8624 * this flag alerts us to the need to check for that */
8627 regnode *ret; /* Will be the head of the group. */
8630 regnode *ender = NULL;
8633 U32 oregflags = RExC_flags;
8634 bool have_branch = 0;
8636 I32 freeze_paren = 0;
8637 I32 after_freeze = 0;
8639 char * parse_start = RExC_parse; /* MJD */
8640 char * const oregcomp_parse = RExC_parse;
8642 GET_RE_DEBUG_FLAGS_DECL;
8644 PERL_ARGS_ASSERT_REG;
8645 DEBUG_PARSE("reg ");
8647 *flagp = 0; /* Tentatively. */
8650 /* Make an OPEN node, if parenthesized. */
8653 /* Under /x, space and comments can be gobbled up between the '(' and
8654 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
8655 * intervening space, as the sequence is a token, and a token should be
8657 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
8659 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
8660 char *start_verb = RExC_parse;
8661 STRLEN verb_len = 0;
8662 char *start_arg = NULL;
8663 unsigned char op = 0;
8665 int internal_argval = 0; /* internal_argval is only useful if !argok */
8667 if (has_intervening_patws && SIZE_ONLY) {
8668 ckWARNregdep(RExC_parse + 1, "In '(*VERB...)', splitting the initial '(*' is deprecated");
8670 while ( *RExC_parse && *RExC_parse != ')' ) {
8671 if ( *RExC_parse == ':' ) {
8672 start_arg = RExC_parse + 1;
8678 verb_len = RExC_parse - start_verb;
8681 while ( *RExC_parse && *RExC_parse != ')' )
8683 if ( *RExC_parse != ')' )
8684 vFAIL("Unterminated verb pattern argument");
8685 if ( RExC_parse == start_arg )
8688 if ( *RExC_parse != ')' )
8689 vFAIL("Unterminated verb pattern");
8692 switch ( *start_verb ) {
8693 case 'A': /* (*ACCEPT) */
8694 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
8696 internal_argval = RExC_nestroot;
8699 case 'C': /* (*COMMIT) */
8700 if ( memEQs(start_verb,verb_len,"COMMIT") )
8703 case 'F': /* (*FAIL) */
8704 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
8709 case ':': /* (*:NAME) */
8710 case 'M': /* (*MARK:NAME) */
8711 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
8716 case 'P': /* (*PRUNE) */
8717 if ( memEQs(start_verb,verb_len,"PRUNE") )
8720 case 'S': /* (*SKIP) */
8721 if ( memEQs(start_verb,verb_len,"SKIP") )
8724 case 'T': /* (*THEN) */
8725 /* [19:06] <TimToady> :: is then */
8726 if ( memEQs(start_verb,verb_len,"THEN") ) {
8728 RExC_seen |= REG_SEEN_CUTGROUP;
8734 vFAIL3("Unknown verb pattern '%.*s'",
8735 verb_len, start_verb);
8738 if ( start_arg && internal_argval ) {
8739 vFAIL3("Verb pattern '%.*s' may not have an argument",
8740 verb_len, start_verb);
8741 } else if ( argok < 0 && !start_arg ) {
8742 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
8743 verb_len, start_verb);
8745 ret = reganode(pRExC_state, op, internal_argval);
8746 if ( ! internal_argval && ! SIZE_ONLY ) {
8748 SV *sv = newSVpvn( start_arg, RExC_parse - start_arg);
8749 ARG(ret) = add_data( pRExC_state, 1, "S" );
8750 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
8757 if (!internal_argval)
8758 RExC_seen |= REG_SEEN_VERBARG;
8759 } else if ( start_arg ) {
8760 vFAIL3("Verb pattern '%.*s' may not have an argument",
8761 verb_len, start_verb);
8763 ret = reg_node(pRExC_state, op);
8765 nextchar(pRExC_state);
8768 if (*RExC_parse == '?') { /* (?...) */
8769 bool is_logical = 0;
8770 const char * const seqstart = RExC_parse;
8771 if (has_intervening_patws && SIZE_ONLY) {
8772 ckWARNregdep(RExC_parse + 1, "In '(?...)', splitting the initial '(?' is deprecated");
8776 paren = *RExC_parse++;
8777 ret = NULL; /* For look-ahead/behind. */
8780 case 'P': /* (?P...) variants for those used to PCRE/Python */
8781 paren = *RExC_parse++;
8782 if ( paren == '<') /* (?P<...>) named capture */
8784 else if (paren == '>') { /* (?P>name) named recursion */
8785 goto named_recursion;
8787 else if (paren == '=') { /* (?P=...) named backref */
8788 /* this pretty much dupes the code for \k<NAME> in regatom(), if
8789 you change this make sure you change that */
8790 char* name_start = RExC_parse;
8792 SV *sv_dat = reg_scan_name(pRExC_state,
8793 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8794 if (RExC_parse == name_start || *RExC_parse != ')')
8795 vFAIL2("Sequence %.3s... not terminated",parse_start);
8798 num = add_data( pRExC_state, 1, "S" );
8799 RExC_rxi->data->data[num]=(void*)sv_dat;
8800 SvREFCNT_inc_simple_void(sv_dat);
8803 ret = reganode(pRExC_state,
8806 : (ASCII_FOLD_RESTRICTED)
8808 : (AT_LEAST_UNI_SEMANTICS)
8816 Set_Node_Offset(ret, parse_start+1);
8817 Set_Node_Cur_Length(ret); /* MJD */
8819 nextchar(pRExC_state);
8823 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8825 case '<': /* (?<...) */
8826 if (*RExC_parse == '!')
8828 else if (*RExC_parse != '=')
8834 case '\'': /* (?'...') */
8835 name_start= RExC_parse;
8836 svname = reg_scan_name(pRExC_state,
8837 SIZE_ONLY ? /* reverse test from the others */
8838 REG_RSN_RETURN_NAME :
8839 REG_RSN_RETURN_NULL);
8840 if (RExC_parse == name_start) {
8842 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8845 if (*RExC_parse != paren)
8846 vFAIL2("Sequence (?%c... not terminated",
8847 paren=='>' ? '<' : paren);
8851 if (!svname) /* shouldn't happen */
8853 "panic: reg_scan_name returned NULL");
8854 if (!RExC_paren_names) {
8855 RExC_paren_names= newHV();
8856 sv_2mortal(MUTABLE_SV(RExC_paren_names));
8858 RExC_paren_name_list= newAV();
8859 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
8862 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
8864 sv_dat = HeVAL(he_str);
8866 /* croak baby croak */
8868 "panic: paren_name hash element allocation failed");
8869 } else if ( SvPOK(sv_dat) ) {
8870 /* (?|...) can mean we have dupes so scan to check
8871 its already been stored. Maybe a flag indicating
8872 we are inside such a construct would be useful,
8873 but the arrays are likely to be quite small, so
8874 for now we punt -- dmq */
8875 IV count = SvIV(sv_dat);
8876 I32 *pv = (I32*)SvPVX(sv_dat);
8878 for ( i = 0 ; i < count ; i++ ) {
8879 if ( pv[i] == RExC_npar ) {
8885 pv = (I32*)SvGROW(sv_dat, SvCUR(sv_dat) + sizeof(I32)+1);
8886 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
8887 pv[count] = RExC_npar;
8888 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
8891 (void)SvUPGRADE(sv_dat,SVt_PVNV);
8892 sv_setpvn(sv_dat, (char *)&(RExC_npar), sizeof(I32));
8894 SvIV_set(sv_dat, 1);
8897 /* Yes this does cause a memory leak in debugging Perls */
8898 if (!av_store(RExC_paren_name_list, RExC_npar, SvREFCNT_inc(svname)))
8899 SvREFCNT_dec_NN(svname);
8902 /*sv_dump(sv_dat);*/
8904 nextchar(pRExC_state);
8906 goto capturing_parens;
8908 RExC_seen |= REG_SEEN_LOOKBEHIND;
8909 RExC_in_lookbehind++;
8911 case '=': /* (?=...) */
8912 RExC_seen_zerolen++;
8914 case '!': /* (?!...) */
8915 RExC_seen_zerolen++;
8916 if (*RExC_parse == ')') {
8917 ret=reg_node(pRExC_state, OPFAIL);
8918 nextchar(pRExC_state);
8922 case '|': /* (?|...) */
8923 /* branch reset, behave like a (?:...) except that
8924 buffers in alternations share the same numbers */
8926 after_freeze = freeze_paren = RExC_npar;
8928 case ':': /* (?:...) */
8929 case '>': /* (?>...) */
8931 case '$': /* (?$...) */
8932 case '@': /* (?@...) */
8933 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
8935 case '#': /* (?#...) */
8936 /* XXX As soon as we disallow separating the '?' and '*' (by
8937 * spaces or (?#...) comment), it is believed that this case
8938 * will be unreachable and can be removed. See
8940 while (*RExC_parse && *RExC_parse != ')')
8942 if (*RExC_parse != ')')
8943 FAIL("Sequence (?#... not terminated");
8944 nextchar(pRExC_state);
8947 case '0' : /* (?0) */
8948 case 'R' : /* (?R) */
8949 if (*RExC_parse != ')')
8950 FAIL("Sequence (?R) not terminated");
8951 ret = reg_node(pRExC_state, GOSTART);
8952 *flagp |= POSTPONED;
8953 nextchar(pRExC_state);
8956 { /* named and numeric backreferences */
8958 case '&': /* (?&NAME) */
8959 parse_start = RExC_parse - 1;
8962 SV *sv_dat = reg_scan_name(pRExC_state,
8963 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8964 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8966 goto gen_recurse_regop;
8967 assert(0); /* NOT REACHED */
8969 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8971 vFAIL("Illegal pattern");
8973 goto parse_recursion;
8975 case '-': /* (?-1) */
8976 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8977 RExC_parse--; /* rewind to let it be handled later */
8981 case '1': case '2': case '3': case '4': /* (?1) */
8982 case '5': case '6': case '7': case '8': case '9':
8985 num = atoi(RExC_parse);
8986 parse_start = RExC_parse - 1; /* MJD */
8987 if (*RExC_parse == '-')
8989 while (isDIGIT(*RExC_parse))
8991 if (*RExC_parse!=')')
8992 vFAIL("Expecting close bracket");
8995 if ( paren == '-' ) {
8997 Diagram of capture buffer numbering.
8998 Top line is the normal capture buffer numbers
8999 Bottom line is the negative indexing as from
9003 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
9007 num = RExC_npar + num;
9010 vFAIL("Reference to nonexistent group");
9012 } else if ( paren == '+' ) {
9013 num = RExC_npar + num - 1;
9016 ret = reganode(pRExC_state, GOSUB, num);
9018 if (num > (I32)RExC_rx->nparens) {
9020 vFAIL("Reference to nonexistent group");
9022 ARG2L_SET( ret, RExC_recurse_count++);
9024 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9025 "Recurse #%"UVuf" to %"IVdf"\n", (UV)ARG(ret), (IV)ARG2L(ret)));
9029 RExC_seen |= REG_SEEN_RECURSE;
9030 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
9031 Set_Node_Offset(ret, parse_start); /* MJD */
9033 *flagp |= POSTPONED;
9034 nextchar(pRExC_state);
9036 } /* named and numeric backreferences */
9037 assert(0); /* NOT REACHED */
9039 case '?': /* (??...) */
9041 if (*RExC_parse != '{') {
9043 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
9046 *flagp |= POSTPONED;
9047 paren = *RExC_parse++;
9049 case '{': /* (?{...}) */
9052 struct reg_code_block *cb;
9054 RExC_seen_zerolen++;
9056 if ( !pRExC_state->num_code_blocks
9057 || pRExC_state->code_index >= pRExC_state->num_code_blocks
9058 || pRExC_state->code_blocks[pRExC_state->code_index].start
9059 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
9062 if (RExC_pm_flags & PMf_USE_RE_EVAL)
9063 FAIL("panic: Sequence (?{...}): no code block found\n");
9064 FAIL("Eval-group not allowed at runtime, use re 'eval'");
9066 /* this is a pre-compiled code block (?{...}) */
9067 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
9068 RExC_parse = RExC_start + cb->end;
9071 if (cb->src_regex) {
9072 n = add_data(pRExC_state, 2, "rl");
9073 RExC_rxi->data->data[n] =
9074 (void*)SvREFCNT_inc((SV*)cb->src_regex);
9075 RExC_rxi->data->data[n+1] = (void*)o;
9078 n = add_data(pRExC_state, 1,
9079 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l");
9080 RExC_rxi->data->data[n] = (void*)o;
9083 pRExC_state->code_index++;
9084 nextchar(pRExC_state);
9088 ret = reg_node(pRExC_state, LOGICAL);
9089 eval = reganode(pRExC_state, EVAL, n);
9092 /* for later propagation into (??{}) return value */
9093 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
9095 REGTAIL(pRExC_state, ret, eval);
9096 /* deal with the length of this later - MJD */
9099 ret = reganode(pRExC_state, EVAL, n);
9100 Set_Node_Length(ret, RExC_parse - parse_start + 1);
9101 Set_Node_Offset(ret, parse_start);
9104 case '(': /* (?(?{...})...) and (?(?=...)...) */
9107 if (RExC_parse[0] == '?') { /* (?(?...)) */
9108 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
9109 || RExC_parse[1] == '<'
9110 || RExC_parse[1] == '{') { /* Lookahead or eval. */
9114 ret = reg_node(pRExC_state, LOGICAL);
9118 tail = reg(pRExC_state, 1, &flag, depth+1);
9119 if (flag & RESTART_UTF8) {
9120 *flagp = RESTART_UTF8;
9123 REGTAIL(pRExC_state, ret, tail);
9127 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9128 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9130 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9131 char *name_start= RExC_parse++;
9133 SV *sv_dat=reg_scan_name(pRExC_state,
9134 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9135 if (RExC_parse == name_start || *RExC_parse != ch)
9136 vFAIL2("Sequence (?(%c... not terminated",
9137 (ch == '>' ? '<' : ch));
9140 num = add_data( pRExC_state, 1, "S" );
9141 RExC_rxi->data->data[num]=(void*)sv_dat;
9142 SvREFCNT_inc_simple_void(sv_dat);
9144 ret = reganode(pRExC_state,NGROUPP,num);
9145 goto insert_if_check_paren;
9147 else if (RExC_parse[0] == 'D' &&
9148 RExC_parse[1] == 'E' &&
9149 RExC_parse[2] == 'F' &&
9150 RExC_parse[3] == 'I' &&
9151 RExC_parse[4] == 'N' &&
9152 RExC_parse[5] == 'E')
9154 ret = reganode(pRExC_state,DEFINEP,0);
9157 goto insert_if_check_paren;
9159 else if (RExC_parse[0] == 'R') {
9162 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9163 parno = atoi(RExC_parse++);
9164 while (isDIGIT(*RExC_parse))
9166 } else if (RExC_parse[0] == '&') {
9169 sv_dat = reg_scan_name(pRExC_state,
9170 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9171 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9173 ret = reganode(pRExC_state,INSUBP,parno);
9174 goto insert_if_check_paren;
9176 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9179 parno = atoi(RExC_parse++);
9181 while (isDIGIT(*RExC_parse))
9183 ret = reganode(pRExC_state, GROUPP, parno);
9185 insert_if_check_paren:
9186 if ((c = *nextchar(pRExC_state)) != ')')
9187 vFAIL("Switch condition not recognized");
9189 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9190 br = regbranch(pRExC_state, &flags, 1,depth+1);
9192 if (flags & RESTART_UTF8) {
9193 *flagp = RESTART_UTF8;
9196 FAIL2("panic: regbranch returned NULL, flags=%#X",
9199 REGTAIL(pRExC_state, br, reganode(pRExC_state, LONGJMP, 0));
9200 c = *nextchar(pRExC_state);
9205 vFAIL("(?(DEFINE)....) does not allow branches");
9206 lastbr = reganode(pRExC_state, IFTHEN, 0); /* Fake one for optimizer. */
9207 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
9208 if (flags & RESTART_UTF8) {
9209 *flagp = RESTART_UTF8;
9212 FAIL2("panic: regbranch returned NULL, flags=%#X",
9215 REGTAIL(pRExC_state, ret, lastbr);
9218 c = *nextchar(pRExC_state);
9223 vFAIL("Switch (?(condition)... contains too many branches");
9224 ender = reg_node(pRExC_state, TAIL);
9225 REGTAIL(pRExC_state, br, ender);
9227 REGTAIL(pRExC_state, lastbr, ender);
9228 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
9231 REGTAIL(pRExC_state, ret, ender);
9232 RExC_size++; /* XXX WHY do we need this?!!
9233 For large programs it seems to be required
9234 but I can't figure out why. -- dmq*/
9238 vFAIL2("Unknown switch condition (?(%.2s", RExC_parse);
9241 case '[': /* (?[ ... ]) */
9242 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
9245 RExC_parse--; /* for vFAIL to print correctly */
9246 vFAIL("Sequence (? incomplete");
9248 default: /* e.g., (?i) */
9251 parse_lparen_question_flags(pRExC_state);
9252 if (UCHARAT(RExC_parse) != ':') {
9253 nextchar(pRExC_state);
9258 nextchar(pRExC_state);
9268 ret = reganode(pRExC_state, OPEN, parno);
9271 RExC_nestroot = parno;
9272 if (RExC_seen & REG_SEEN_RECURSE
9273 && !RExC_open_parens[parno-1])
9275 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9276 "Setting open paren #%"IVdf" to %d\n",
9277 (IV)parno, REG_NODE_NUM(ret)));
9278 RExC_open_parens[parno-1]= ret;
9281 Set_Node_Length(ret, 1); /* MJD */
9282 Set_Node_Offset(ret, RExC_parse); /* MJD */
9290 /* Pick up the branches, linking them together. */
9291 parse_start = RExC_parse; /* MJD */
9292 br = regbranch(pRExC_state, &flags, 1,depth+1);
9294 /* branch_len = (paren != 0); */
9297 if (flags & RESTART_UTF8) {
9298 *flagp = RESTART_UTF8;
9301 FAIL2("panic: regbranch returned NULL, flags=%#X", flags);
9303 if (*RExC_parse == '|') {
9304 if (!SIZE_ONLY && RExC_extralen) {
9305 reginsert(pRExC_state, BRANCHJ, br, depth+1);
9308 reginsert(pRExC_state, BRANCH, br, depth+1);
9309 Set_Node_Length(br, paren != 0);
9310 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
9314 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
9316 else if (paren == ':') {
9317 *flagp |= flags&SIMPLE;
9319 if (is_open) { /* Starts with OPEN. */
9320 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
9322 else if (paren != '?') /* Not Conditional */
9324 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9326 while (*RExC_parse == '|') {
9327 if (!SIZE_ONLY && RExC_extralen) {
9328 ender = reganode(pRExC_state, LONGJMP,0);
9329 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender); /* Append to the previous. */
9332 RExC_extralen += 2; /* Account for LONGJMP. */
9333 nextchar(pRExC_state);
9335 if (RExC_npar > after_freeze)
9336 after_freeze = RExC_npar;
9337 RExC_npar = freeze_paren;
9339 br = regbranch(pRExC_state, &flags, 0, depth+1);
9342 if (flags & RESTART_UTF8) {
9343 *flagp = RESTART_UTF8;
9346 FAIL2("panic: regbranch returned NULL, flags=%#X", flags);
9348 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
9350 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9353 if (have_branch || paren != ':') {
9354 /* Make a closing node, and hook it on the end. */
9357 ender = reg_node(pRExC_state, TAIL);
9360 ender = reganode(pRExC_state, CLOSE, parno);
9361 if (!SIZE_ONLY && RExC_seen & REG_SEEN_RECURSE) {
9362 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9363 "Setting close paren #%"IVdf" to %d\n",
9364 (IV)parno, REG_NODE_NUM(ender)));
9365 RExC_close_parens[parno-1]= ender;
9366 if (RExC_nestroot == parno)
9369 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
9370 Set_Node_Length(ender,1); /* MJD */
9376 *flagp &= ~HASWIDTH;
9379 ender = reg_node(pRExC_state, SUCCEED);
9382 ender = reg_node(pRExC_state, END);
9384 assert(!RExC_opend); /* there can only be one! */
9389 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9390 SV * const mysv_val1=sv_newmortal();
9391 SV * const mysv_val2=sv_newmortal();
9392 DEBUG_PARSE_MSG("lsbr");
9393 regprop(RExC_rx, mysv_val1, lastbr);
9394 regprop(RExC_rx, mysv_val2, ender);
9395 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9396 SvPV_nolen_const(mysv_val1),
9397 (IV)REG_NODE_NUM(lastbr),
9398 SvPV_nolen_const(mysv_val2),
9399 (IV)REG_NODE_NUM(ender),
9400 (IV)(ender - lastbr)
9403 REGTAIL(pRExC_state, lastbr, ender);
9405 if (have_branch && !SIZE_ONLY) {
9408 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
9410 /* Hook the tails of the branches to the closing node. */
9411 for (br = ret; br; br = regnext(br)) {
9412 const U8 op = PL_regkind[OP(br)];
9414 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
9415 if (OP(NEXTOPER(br)) != NOTHING || regnext(NEXTOPER(br)) != ender)
9418 else if (op == BRANCHJ) {
9419 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
9420 /* for now we always disable this optimisation * /
9421 if (OP(NEXTOPER(NEXTOPER(br))) != NOTHING || regnext(NEXTOPER(NEXTOPER(br))) != ender)
9427 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
9428 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9429 SV * const mysv_val1=sv_newmortal();
9430 SV * const mysv_val2=sv_newmortal();
9431 DEBUG_PARSE_MSG("NADA");
9432 regprop(RExC_rx, mysv_val1, ret);
9433 regprop(RExC_rx, mysv_val2, ender);
9434 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9435 SvPV_nolen_const(mysv_val1),
9436 (IV)REG_NODE_NUM(ret),
9437 SvPV_nolen_const(mysv_val2),
9438 (IV)REG_NODE_NUM(ender),
9443 if (OP(ender) == TAIL) {
9448 for ( opt= br + 1; opt < ender ; opt++ )
9450 NEXT_OFF(br)= ender - br;
9458 static const char parens[] = "=!<,>";
9460 if (paren && (p = strchr(parens, paren))) {
9461 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
9462 int flag = (p - parens) > 1;
9465 node = SUSPEND, flag = 0;
9466 reginsert(pRExC_state, node,ret, depth+1);
9467 Set_Node_Cur_Length(ret);
9468 Set_Node_Offset(ret, parse_start + 1);
9470 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
9474 /* Check for proper termination. */
9476 /* restore original flags, but keep (?p) */
9477 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
9478 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
9479 RExC_parse = oregcomp_parse;
9480 vFAIL("Unmatched (");
9483 else if (!paren && RExC_parse < RExC_end) {
9484 if (*RExC_parse == ')') {
9486 vFAIL("Unmatched )");
9489 FAIL("Junk on end of regexp"); /* "Can't happen". */
9490 assert(0); /* NOTREACHED */
9493 if (RExC_in_lookbehind) {
9494 RExC_in_lookbehind--;
9496 if (after_freeze > RExC_npar)
9497 RExC_npar = after_freeze;
9502 - regbranch - one alternative of an | operator
9504 * Implements the concatenation operator.
9506 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9510 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
9514 regnode *chain = NULL;
9516 I32 flags = 0, c = 0;
9517 GET_RE_DEBUG_FLAGS_DECL;
9519 PERL_ARGS_ASSERT_REGBRANCH;
9521 DEBUG_PARSE("brnc");
9526 if (!SIZE_ONLY && RExC_extralen)
9527 ret = reganode(pRExC_state, BRANCHJ,0);
9529 ret = reg_node(pRExC_state, BRANCH);
9530 Set_Node_Length(ret, 1);
9534 if (!first && SIZE_ONLY)
9535 RExC_extralen += 1; /* BRANCHJ */
9537 *flagp = WORST; /* Tentatively. */
9540 nextchar(pRExC_state);
9541 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
9543 latest = regpiece(pRExC_state, &flags,depth+1);
9544 if (latest == NULL) {
9545 if (flags & TRYAGAIN)
9547 if (flags & RESTART_UTF8) {
9548 *flagp = RESTART_UTF8;
9551 FAIL2("panic: regpiece returned NULL, flags=%#X", flags);
9553 else if (ret == NULL)
9555 *flagp |= flags&(HASWIDTH|POSTPONED);
9556 if (chain == NULL) /* First piece. */
9557 *flagp |= flags&SPSTART;
9560 REGTAIL(pRExC_state, chain, latest);
9565 if (chain == NULL) { /* Loop ran zero times. */
9566 chain = reg_node(pRExC_state, NOTHING);
9571 *flagp |= flags&SIMPLE;
9578 - regpiece - something followed by possible [*+?]
9580 * Note that the branching code sequences used for ? and the general cases
9581 * of * and + are somewhat optimized: they use the same NOTHING node as
9582 * both the endmarker for their branch list and the body of the last branch.
9583 * It might seem that this node could be dispensed with entirely, but the
9584 * endmarker role is not redundant.
9586 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
9588 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9592 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
9599 const char * const origparse = RExC_parse;
9601 I32 max = REG_INFTY;
9602 #ifdef RE_TRACK_PATTERN_OFFSETS
9605 const char *maxpos = NULL;
9607 /* Save the original in case we change the emitted regop to a FAIL. */
9608 regnode * const orig_emit = RExC_emit;
9610 GET_RE_DEBUG_FLAGS_DECL;
9612 PERL_ARGS_ASSERT_REGPIECE;
9614 DEBUG_PARSE("piec");
9616 ret = regatom(pRExC_state, &flags,depth+1);
9618 if (flags & (TRYAGAIN|RESTART_UTF8))
9619 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
9621 FAIL2("panic: regatom returned NULL, flags=%#X", flags);
9627 if (op == '{' && regcurly(RExC_parse, FALSE)) {
9629 #ifdef RE_TRACK_PATTERN_OFFSETS
9630 parse_start = RExC_parse; /* MJD */
9632 next = RExC_parse + 1;
9633 while (isDIGIT(*next) || *next == ',') {
9642 if (*next == '}') { /* got one */
9646 min = atoi(RExC_parse);
9650 maxpos = RExC_parse;
9652 if (!max && *maxpos != '0')
9653 max = REG_INFTY; /* meaning "infinity" */
9654 else if (max >= REG_INFTY)
9655 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
9657 nextchar(pRExC_state);
9658 if (max < min) { /* If can't match, warn and optimize to fail
9661 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
9663 /* We can't back off the size because we have to reserve
9664 * enough space for all the things we are about to throw
9665 * away, but we can shrink it by the ammount we are about
9667 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
9670 RExC_emit = orig_emit;
9672 ret = reg_node(pRExC_state, OPFAIL);
9677 if ((flags&SIMPLE)) {
9678 RExC_naughty += 2 + RExC_naughty / 2;
9679 reginsert(pRExC_state, CURLY, ret, depth+1);
9680 Set_Node_Offset(ret, parse_start+1); /* MJD */
9681 Set_Node_Cur_Length(ret);
9684 regnode * const w = reg_node(pRExC_state, WHILEM);
9687 REGTAIL(pRExC_state, ret, w);
9688 if (!SIZE_ONLY && RExC_extralen) {
9689 reginsert(pRExC_state, LONGJMP,ret, depth+1);
9690 reginsert(pRExC_state, NOTHING,ret, depth+1);
9691 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
9693 reginsert(pRExC_state, CURLYX,ret, depth+1);
9695 Set_Node_Offset(ret, parse_start+1);
9696 Set_Node_Length(ret,
9697 op == '{' ? (RExC_parse - parse_start) : 1);
9699 if (!SIZE_ONLY && RExC_extralen)
9700 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
9701 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
9703 RExC_whilem_seen++, RExC_extralen += 3;
9704 RExC_naughty += 4 + RExC_naughty; /* compound interest */
9713 ARG1_SET(ret, (U16)min);
9714 ARG2_SET(ret, (U16)max);
9726 #if 0 /* Now runtime fix should be reliable. */
9728 /* if this is reinstated, don't forget to put this back into perldiag:
9730 =item Regexp *+ operand could be empty at {#} in regex m/%s/
9732 (F) The part of the regexp subject to either the * or + quantifier
9733 could match an empty string. The {#} shows in the regular
9734 expression about where the problem was discovered.
9738 if (!(flags&HASWIDTH) && op != '?')
9739 vFAIL("Regexp *+ operand could be empty");
9742 #ifdef RE_TRACK_PATTERN_OFFSETS
9743 parse_start = RExC_parse;
9745 nextchar(pRExC_state);
9747 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
9749 if (op == '*' && (flags&SIMPLE)) {
9750 reginsert(pRExC_state, STAR, ret, depth+1);
9754 else if (op == '*') {
9758 else if (op == '+' && (flags&SIMPLE)) {
9759 reginsert(pRExC_state, PLUS, ret, depth+1);
9763 else if (op == '+') {
9767 else if (op == '?') {
9772 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
9773 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
9774 ckWARN3reg(RExC_parse,
9775 "%.*s matches null string many times",
9776 (int)(RExC_parse >= origparse ? RExC_parse - origparse : 0),
9778 (void)ReREFCNT_inc(RExC_rx_sv);
9781 if (RExC_parse < RExC_end && *RExC_parse == '?') {
9782 nextchar(pRExC_state);
9783 reginsert(pRExC_state, MINMOD, ret, depth+1);
9784 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
9786 #ifndef REG_ALLOW_MINMOD_SUSPEND
9789 if (RExC_parse < RExC_end && *RExC_parse == '+') {
9791 nextchar(pRExC_state);
9792 ender = reg_node(pRExC_state, SUCCEED);
9793 REGTAIL(pRExC_state, ret, ender);
9794 reginsert(pRExC_state, SUSPEND, ret, depth+1);
9796 ender = reg_node(pRExC_state, TAIL);
9797 REGTAIL(pRExC_state, ret, ender);
9801 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
9803 vFAIL("Nested quantifiers");
9810 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p, UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
9811 const bool strict /* Apply stricter parsing rules? */
9815 /* This is expected to be called by a parser routine that has recognized '\N'
9816 and needs to handle the rest. RExC_parse is expected to point at the first
9817 char following the N at the time of the call. On successful return,
9818 RExC_parse has been updated to point to just after the sequence identified
9819 by this routine, and <*flagp> has been updated.
9821 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
9824 \N may begin either a named sequence, or if outside a character class, mean
9825 to match a non-newline. For non single-quoted regexes, the tokenizer has
9826 attempted to decide which, and in the case of a named sequence, converted it
9827 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
9828 where c1... are the characters in the sequence. For single-quoted regexes,
9829 the tokenizer passes the \N sequence through unchanged; this code will not
9830 attempt to determine this nor expand those, instead raising a syntax error.
9831 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
9832 or there is no '}', it signals that this \N occurrence means to match a
9835 Only the \N{U+...} form should occur in a character class, for the same
9836 reason that '.' inside a character class means to just match a period: it
9837 just doesn't make sense.
9839 The function raises an error (via vFAIL), and doesn't return for various
9840 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
9841 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
9842 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
9843 only possible if node_p is non-NULL.
9846 If <valuep> is non-null, it means the caller can accept an input sequence
9847 consisting of a just a single code point; <*valuep> is set to that value
9848 if the input is such.
9850 If <node_p> is non-null it signifies that the caller can accept any other
9851 legal sequence (i.e., one that isn't just a single code point). <*node_p>
9853 1) \N means not-a-NL: points to a newly created REG_ANY node;
9854 2) \N{}: points to a new NOTHING node;
9855 3) otherwise: points to a new EXACT node containing the resolved
9857 Note that FALSE is returned for single code point sequences if <valuep> is
9861 char * endbrace; /* '}' following the name */
9863 char *endchar; /* Points to '.' or '}' ending cur char in the input
9865 bool has_multiple_chars; /* true if the input stream contains a sequence of
9866 more than one character */
9868 GET_RE_DEBUG_FLAGS_DECL;
9870 PERL_ARGS_ASSERT_GROK_BSLASH_N;
9874 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
9876 /* The [^\n] meaning of \N ignores spaces and comments under the /x
9877 * modifier. The other meaning does not */
9878 p = (RExC_flags & RXf_PMf_EXTENDED)
9879 ? regwhite( pRExC_state, RExC_parse )
9882 /* Disambiguate between \N meaning a named character versus \N meaning
9883 * [^\n]. The former is assumed when it can't be the latter. */
9884 if (*p != '{' || regcurly(p, FALSE)) {
9887 /* no bare \N in a charclass */
9888 if (in_char_class) {
9889 vFAIL("\\N in a character class must be a named character: \\N{...}");
9893 nextchar(pRExC_state);
9894 *node_p = reg_node(pRExC_state, REG_ANY);
9895 *flagp |= HASWIDTH|SIMPLE;
9898 Set_Node_Length(*node_p, 1); /* MJD */
9902 /* Here, we have decided it should be a named character or sequence */
9904 /* The test above made sure that the next real character is a '{', but
9905 * under the /x modifier, it could be separated by space (or a comment and
9906 * \n) and this is not allowed (for consistency with \x{...} and the
9907 * tokenizer handling of \N{NAME}). */
9908 if (*RExC_parse != '{') {
9909 vFAIL("Missing braces on \\N{}");
9912 RExC_parse++; /* Skip past the '{' */
9914 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
9915 || ! (endbrace == RExC_parse /* nothing between the {} */
9916 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below */
9917 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg) */
9919 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
9920 vFAIL("\\N{NAME} must be resolved by the lexer");
9923 if (endbrace == RExC_parse) { /* empty: \N{} */
9926 *node_p = reg_node(pRExC_state,NOTHING);
9928 else if (in_char_class) {
9929 if (SIZE_ONLY && in_char_class) {
9931 RExC_parse++; /* Position after the "}" */
9932 vFAIL("Zero length \\N{}");
9935 ckWARNreg(RExC_parse,
9936 "Ignoring zero length \\N{} in character class");
9944 nextchar(pRExC_state);
9948 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
9949 RExC_parse += 2; /* Skip past the 'U+' */
9951 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9953 /* Code points are separated by dots. If none, there is only one code
9954 * point, and is terminated by the brace */
9955 has_multiple_chars = (endchar < endbrace);
9957 if (valuep && (! has_multiple_chars || in_char_class)) {
9958 /* We only pay attention to the first char of
9959 multichar strings being returned in char classes. I kinda wonder
9960 if this makes sense as it does change the behaviour
9961 from earlier versions, OTOH that behaviour was broken
9962 as well. XXX Solution is to recharacterize as
9963 [rest-of-class]|multi1|multi2... */
9965 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
9966 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
9967 | PERL_SCAN_DISALLOW_PREFIX
9968 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
9970 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
9972 /* The tokenizer should have guaranteed validity, but it's possible to
9973 * bypass it by using single quoting, so check */
9974 if (length_of_hex == 0
9975 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
9977 RExC_parse += length_of_hex; /* Includes all the valid */
9978 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
9979 ? UTF8SKIP(RExC_parse)
9981 /* Guard against malformed utf8 */
9982 if (RExC_parse >= endchar) {
9983 RExC_parse = endchar;
9985 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9988 if (in_char_class && has_multiple_chars) {
9990 RExC_parse = endbrace;
9991 vFAIL("\\N{} in character class restricted to one character");
9994 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
9998 RExC_parse = endbrace + 1;
10000 else if (! node_p || ! has_multiple_chars) {
10002 /* Here, the input is legal, but not according to the caller's
10003 * options. We fail without advancing the parse, so that the
10004 * caller can try again */
10010 /* What is done here is to convert this to a sub-pattern of the form
10011 * (?:\x{char1}\x{char2}...)
10012 * and then call reg recursively. That way, it retains its atomicness,
10013 * while not having to worry about special handling that some code
10014 * points may have. toke.c has converted the original Unicode values
10015 * to native, so that we can just pass on the hex values unchanged. We
10016 * do have to set a flag to keep recoding from happening in the
10019 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
10021 char *orig_end = RExC_end;
10024 while (RExC_parse < endbrace) {
10026 /* Convert to notation the rest of the code understands */
10027 sv_catpv(substitute_parse, "\\x{");
10028 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
10029 sv_catpv(substitute_parse, "}");
10031 /* Point to the beginning of the next character in the sequence. */
10032 RExC_parse = endchar + 1;
10033 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10035 sv_catpv(substitute_parse, ")");
10037 RExC_parse = SvPV(substitute_parse, len);
10039 /* Don't allow empty number */
10041 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10043 RExC_end = RExC_parse + len;
10045 /* The values are Unicode, and therefore not subject to recoding */
10046 RExC_override_recoding = 1;
10048 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
10049 if (flags & RESTART_UTF8) {
10050 *flagp = RESTART_UTF8;
10053 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#X",
10056 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10058 RExC_parse = endbrace;
10059 RExC_end = orig_end;
10060 RExC_override_recoding = 0;
10062 nextchar(pRExC_state);
10072 * It returns the code point in utf8 for the value in *encp.
10073 * value: a code value in the source encoding
10074 * encp: a pointer to an Encode object
10076 * If the result from Encode is not a single character,
10077 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
10080 S_reg_recode(pTHX_ const char value, SV **encp)
10083 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
10084 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
10085 const STRLEN newlen = SvCUR(sv);
10086 UV uv = UNICODE_REPLACEMENT;
10088 PERL_ARGS_ASSERT_REG_RECODE;
10092 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
10095 if (!newlen || numlen != newlen) {
10096 uv = UNICODE_REPLACEMENT;
10102 PERL_STATIC_INLINE U8
10103 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
10107 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
10113 op = get_regex_charset(RExC_flags);
10114 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
10115 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
10116 been, so there is no hole */
10119 return op + EXACTF;
10122 PERL_STATIC_INLINE void
10123 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state, regnode *node, I32* flagp, STRLEN len, UV code_point)
10125 /* This knows the details about sizing an EXACTish node, setting flags for
10126 * it (by setting <*flagp>, and potentially populating it with a single
10129 * If <len> (the length in bytes) is non-zero, this function assumes that
10130 * the node has already been populated, and just does the sizing. In this
10131 * case <code_point> should be the final code point that has already been
10132 * placed into the node. This value will be ignored except that under some
10133 * circumstances <*flagp> is set based on it.
10135 * If <len> is zero, the function assumes that the node is to contain only
10136 * the single character given by <code_point> and calculates what <len>
10137 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10138 * additionally will populate the node's STRING with <code_point>, if <len>
10139 * is 0. In both cases <*flagp> is appropriately set
10141 * It knows that under FOLD, the Latin Sharp S and UTF characters above
10142 * 255, must be folded (the former only when the rules indicate it can
10145 bool len_passed_in = cBOOL(len != 0);
10146 U8 character[UTF8_MAXBYTES_CASE+1];
10148 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10150 if (! len_passed_in) {
10152 if (FOLD && (! LOC || code_point > 255)) {
10153 _to_uni_fold_flags(NATIVE_TO_UNI(code_point),
10156 FOLD_FLAGS_FULL | ((LOC)
10157 ? FOLD_FLAGS_LOCALE
10158 : (ASCII_FOLD_RESTRICTED)
10159 ? FOLD_FLAGS_NOMIX_ASCII
10163 uvchr_to_utf8( character, code_point);
10164 len = UTF8SKIP(character);
10168 || code_point != LATIN_SMALL_LETTER_SHARP_S
10169 || ASCII_FOLD_RESTRICTED
10170 || ! AT_LEAST_UNI_SEMANTICS)
10172 *character = (U8) code_point;
10177 *(character + 1) = 's';
10183 RExC_size += STR_SZ(len);
10186 RExC_emit += STR_SZ(len);
10187 STR_LEN(node) = len;
10188 if (! len_passed_in) {
10189 Copy((char *) character, STRING(node), len, char);
10193 *flagp |= HASWIDTH;
10195 /* A single character node is SIMPLE, except for the special-cased SHARP S
10197 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
10198 && (code_point != LATIN_SMALL_LETTER_SHARP_S
10199 || ! FOLD || ! DEPENDS_SEMANTICS))
10206 - regatom - the lowest level
10208 Try to identify anything special at the start of the pattern. If there
10209 is, then handle it as required. This may involve generating a single regop,
10210 such as for an assertion; or it may involve recursing, such as to
10211 handle a () structure.
10213 If the string doesn't start with something special then we gobble up
10214 as much literal text as we can.
10216 Once we have been able to handle whatever type of thing started the
10217 sequence, we return.
10219 Note: we have to be careful with escapes, as they can be both literal
10220 and special, and in the case of \10 and friends, context determines which.
10222 A summary of the code structure is:
10224 switch (first_byte) {
10225 cases for each special:
10226 handle this special;
10229 switch (2nd byte) {
10230 cases for each unambiguous special:
10231 handle this special;
10233 cases for each ambigous special/literal:
10235 if (special) handle here
10237 default: // unambiguously literal:
10240 default: // is a literal char
10243 create EXACTish node for literal;
10244 while (more input and node isn't full) {
10245 switch (input_byte) {
10246 cases for each special;
10247 make sure parse pointer is set so that the next call to
10248 regatom will see this special first
10249 goto loopdone; // EXACTish node terminated by prev. char
10251 append char to EXACTISH node;
10253 get next input byte;
10257 return the generated node;
10259 Specifically there are two separate switches for handling
10260 escape sequences, with the one for handling literal escapes requiring
10261 a dummy entry for all of the special escapes that are actually handled
10264 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
10266 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10268 Otherwise does not return NULL.
10272 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10275 regnode *ret = NULL;
10277 char *parse_start = RExC_parse;
10281 GET_RE_DEBUG_FLAGS_DECL;
10283 *flagp = WORST; /* Tentatively. */
10285 DEBUG_PARSE("atom");
10287 PERL_ARGS_ASSERT_REGATOM;
10290 switch ((U8)*RExC_parse) {
10292 RExC_seen_zerolen++;
10293 nextchar(pRExC_state);
10294 if (RExC_flags & RXf_PMf_MULTILINE)
10295 ret = reg_node(pRExC_state, MBOL);
10296 else if (RExC_flags & RXf_PMf_SINGLELINE)
10297 ret = reg_node(pRExC_state, SBOL);
10299 ret = reg_node(pRExC_state, BOL);
10300 Set_Node_Length(ret, 1); /* MJD */
10303 nextchar(pRExC_state);
10305 RExC_seen_zerolen++;
10306 if (RExC_flags & RXf_PMf_MULTILINE)
10307 ret = reg_node(pRExC_state, MEOL);
10308 else if (RExC_flags & RXf_PMf_SINGLELINE)
10309 ret = reg_node(pRExC_state, SEOL);
10311 ret = reg_node(pRExC_state, EOL);
10312 Set_Node_Length(ret, 1); /* MJD */
10315 nextchar(pRExC_state);
10316 if (RExC_flags & RXf_PMf_SINGLELINE)
10317 ret = reg_node(pRExC_state, SANY);
10319 ret = reg_node(pRExC_state, REG_ANY);
10320 *flagp |= HASWIDTH|SIMPLE;
10322 Set_Node_Length(ret, 1); /* MJD */
10326 char * const oregcomp_parse = ++RExC_parse;
10327 ret = regclass(pRExC_state, flagp,depth+1,
10328 FALSE, /* means parse the whole char class */
10329 TRUE, /* allow multi-char folds */
10330 FALSE, /* don't silence non-portable warnings. */
10332 if (*RExC_parse != ']') {
10333 RExC_parse = oregcomp_parse;
10334 vFAIL("Unmatched [");
10337 if (*flagp & RESTART_UTF8)
10339 FAIL2("panic: regclass returned NULL to regatom, flags=%#X",
10342 nextchar(pRExC_state);
10343 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
10347 nextchar(pRExC_state);
10348 ret = reg(pRExC_state, 2, &flags,depth+1);
10350 if (flags & TRYAGAIN) {
10351 if (RExC_parse == RExC_end) {
10352 /* Make parent create an empty node if needed. */
10353 *flagp |= TRYAGAIN;
10358 if (flags & RESTART_UTF8) {
10359 *flagp = RESTART_UTF8;
10362 FAIL2("panic: reg returned NULL to regatom, flags=%#X", flags);
10364 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10368 if (flags & TRYAGAIN) {
10369 *flagp |= TRYAGAIN;
10372 vFAIL("Internal urp");
10373 /* Supposed to be caught earlier. */
10376 if (!regcurly(RExC_parse, FALSE)) {
10385 vFAIL("Quantifier follows nothing");
10390 This switch handles escape sequences that resolve to some kind
10391 of special regop and not to literal text. Escape sequnces that
10392 resolve to literal text are handled below in the switch marked
10395 Every entry in this switch *must* have a corresponding entry
10396 in the literal escape switch. However, the opposite is not
10397 required, as the default for this switch is to jump to the
10398 literal text handling code.
10400 switch ((U8)*++RExC_parse) {
10402 /* Special Escapes */
10404 RExC_seen_zerolen++;
10405 ret = reg_node(pRExC_state, SBOL);
10407 goto finish_meta_pat;
10409 ret = reg_node(pRExC_state, GPOS);
10410 RExC_seen |= REG_SEEN_GPOS;
10412 goto finish_meta_pat;
10414 RExC_seen_zerolen++;
10415 ret = reg_node(pRExC_state, KEEPS);
10417 /* XXX:dmq : disabling in-place substitution seems to
10418 * be necessary here to avoid cases of memory corruption, as
10419 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
10421 RExC_seen |= REG_SEEN_LOOKBEHIND;
10422 goto finish_meta_pat;
10424 ret = reg_node(pRExC_state, SEOL);
10426 RExC_seen_zerolen++; /* Do not optimize RE away */
10427 goto finish_meta_pat;
10429 ret = reg_node(pRExC_state, EOS);
10431 RExC_seen_zerolen++; /* Do not optimize RE away */
10432 goto finish_meta_pat;
10434 ret = reg_node(pRExC_state, CANY);
10435 RExC_seen |= REG_SEEN_CANY;
10436 *flagp |= HASWIDTH|SIMPLE;
10437 goto finish_meta_pat;
10439 ret = reg_node(pRExC_state, CLUMP);
10440 *flagp |= HASWIDTH;
10441 goto finish_meta_pat;
10447 arg = ANYOF_WORDCHAR;
10451 RExC_seen_zerolen++;
10452 RExC_seen |= REG_SEEN_LOOKBEHIND;
10453 op = BOUND + get_regex_charset(RExC_flags);
10454 if (op > BOUNDA) { /* /aa is same as /a */
10457 ret = reg_node(pRExC_state, op);
10458 FLAGS(ret) = get_regex_charset(RExC_flags);
10460 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10461 ckWARNdep(RExC_parse, "\"\\b{\" is deprecated; use \"\\b\\{\" or \"\\b[{]\" instead");
10463 goto finish_meta_pat;
10465 RExC_seen_zerolen++;
10466 RExC_seen |= REG_SEEN_LOOKBEHIND;
10467 op = NBOUND + get_regex_charset(RExC_flags);
10468 if (op > NBOUNDA) { /* /aa is same as /a */
10471 ret = reg_node(pRExC_state, op);
10472 FLAGS(ret) = get_regex_charset(RExC_flags);
10474 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10475 ckWARNdep(RExC_parse, "\"\\B{\" is deprecated; use \"\\B\\{\" or \"\\B[{]\" instead");
10477 goto finish_meta_pat;
10487 ret = reg_node(pRExC_state, LNBREAK);
10488 *flagp |= HASWIDTH|SIMPLE;
10489 goto finish_meta_pat;
10497 goto join_posix_op_known;
10503 arg = ANYOF_VERTWS;
10505 goto join_posix_op_known;
10515 op = POSIXD + get_regex_charset(RExC_flags);
10516 if (op > POSIXA) { /* /aa is same as /a */
10520 join_posix_op_known:
10523 op += NPOSIXD - POSIXD;
10526 ret = reg_node(pRExC_state, op);
10528 FLAGS(ret) = namedclass_to_classnum(arg);
10531 *flagp |= HASWIDTH|SIMPLE;
10535 nextchar(pRExC_state);
10536 Set_Node_Length(ret, 2); /* MJD */
10542 char* parse_start = RExC_parse - 2;
10547 ret = regclass(pRExC_state, flagp,depth+1,
10548 TRUE, /* means just parse this element */
10549 FALSE, /* don't allow multi-char folds */
10550 FALSE, /* don't silence non-portable warnings.
10551 It would be a bug if these returned
10554 /* regclass() can only return RESTART_UTF8 if multi-char folds
10557 FAIL2("panic: regclass returned NULL to regatom, flags=%#X",
10562 Set_Node_Offset(ret, parse_start + 2);
10563 Set_Node_Cur_Length(ret);
10564 nextchar(pRExC_state);
10568 /* Handle \N and \N{NAME} with multiple code points here and not
10569 * below because it can be multicharacter. join_exact() will join
10570 * them up later on. Also this makes sure that things like
10571 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
10572 * The options to the grok function call causes it to fail if the
10573 * sequence is just a single code point. We then go treat it as
10574 * just another character in the current EXACT node, and hence it
10575 * gets uniform treatment with all the other characters. The
10576 * special treatment for quantifiers is not needed for such single
10577 * character sequences */
10579 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
10580 FALSE /* not strict */ )) {
10581 if (*flagp & RESTART_UTF8)
10587 case 'k': /* Handle \k<NAME> and \k'NAME' */
10590 char ch= RExC_parse[1];
10591 if (ch != '<' && ch != '\'' && ch != '{') {
10593 vFAIL2("Sequence %.2s... not terminated",parse_start);
10595 /* this pretty much dupes the code for (?P=...) in reg(), if
10596 you change this make sure you change that */
10597 char* name_start = (RExC_parse += 2);
10599 SV *sv_dat = reg_scan_name(pRExC_state,
10600 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10601 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
10602 if (RExC_parse == name_start || *RExC_parse != ch)
10603 vFAIL2("Sequence %.3s... not terminated",parse_start);
10606 num = add_data( pRExC_state, 1, "S" );
10607 RExC_rxi->data->data[num]=(void*)sv_dat;
10608 SvREFCNT_inc_simple_void(sv_dat);
10612 ret = reganode(pRExC_state,
10615 : (ASCII_FOLD_RESTRICTED)
10617 : (AT_LEAST_UNI_SEMANTICS)
10623 *flagp |= HASWIDTH;
10625 /* override incorrect value set in reganode MJD */
10626 Set_Node_Offset(ret, parse_start+1);
10627 Set_Node_Cur_Length(ret); /* MJD */
10628 nextchar(pRExC_state);
10634 case '1': case '2': case '3': case '4':
10635 case '5': case '6': case '7': case '8': case '9':
10638 bool isg = *RExC_parse == 'g';
10643 if (*RExC_parse == '{') {
10647 if (*RExC_parse == '-') {
10651 if (hasbrace && !isDIGIT(*RExC_parse)) {
10652 if (isrel) RExC_parse--;
10654 goto parse_named_seq;
10656 num = atoi(RExC_parse);
10657 if (isg && num == 0)
10658 vFAIL("Reference to invalid group 0");
10660 num = RExC_npar - num;
10662 vFAIL("Reference to nonexistent or unclosed group");
10664 if (!isg && num > 9 && num >= RExC_npar)
10665 /* Probably a character specified in octal, e.g. \35 */
10668 char * const parse_start = RExC_parse - 1; /* MJD */
10669 while (isDIGIT(*RExC_parse))
10671 if (parse_start == RExC_parse - 1)
10672 vFAIL("Unterminated \\g... pattern");
10674 if (*RExC_parse != '}')
10675 vFAIL("Unterminated \\g{...} pattern");
10679 if (num > (I32)RExC_rx->nparens)
10680 vFAIL("Reference to nonexistent group");
10683 ret = reganode(pRExC_state,
10686 : (ASCII_FOLD_RESTRICTED)
10688 : (AT_LEAST_UNI_SEMANTICS)
10694 *flagp |= HASWIDTH;
10696 /* override incorrect value set in reganode MJD */
10697 Set_Node_Offset(ret, parse_start+1);
10698 Set_Node_Cur_Length(ret); /* MJD */
10700 nextchar(pRExC_state);
10705 if (RExC_parse >= RExC_end)
10706 FAIL("Trailing \\");
10709 /* Do not generate "unrecognized" warnings here, we fall
10710 back into the quick-grab loop below */
10717 if (RExC_flags & RXf_PMf_EXTENDED) {
10718 if ( reg_skipcomment( pRExC_state ) )
10725 parse_start = RExC_parse - 1;
10734 #define MAX_NODE_STRING_SIZE 127
10735 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
10737 U8 upper_parse = MAX_NODE_STRING_SIZE;
10740 bool next_is_quantifier;
10741 char * oldp = NULL;
10743 /* If a folding node contains only code points that don't
10744 * participate in folds, it can be changed into an EXACT node,
10745 * which allows the optimizer more things to look for */
10749 node_type = compute_EXACTish(pRExC_state);
10750 ret = reg_node(pRExC_state, node_type);
10752 /* In pass1, folded, we use a temporary buffer instead of the
10753 * actual node, as the node doesn't exist yet */
10754 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
10760 /* We do the EXACTFish to EXACT node only if folding, and not if in
10761 * locale, as whether a character folds or not isn't known until
10763 maybe_exact = FOLD && ! LOC;
10765 /* XXX The node can hold up to 255 bytes, yet this only goes to
10766 * 127. I (khw) do not know why. Keeping it somewhat less than
10767 * 255 allows us to not have to worry about overflow due to
10768 * converting to utf8 and fold expansion, but that value is
10769 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
10770 * split up by this limit into a single one using the real max of
10771 * 255. Even at 127, this breaks under rare circumstances. If
10772 * folding, we do not want to split a node at a character that is a
10773 * non-final in a multi-char fold, as an input string could just
10774 * happen to want to match across the node boundary. The join
10775 * would solve that problem if the join actually happens. But a
10776 * series of more than two nodes in a row each of 127 would cause
10777 * the first join to succeed to get to 254, but then there wouldn't
10778 * be room for the next one, which could at be one of those split
10779 * multi-char folds. I don't know of any fool-proof solution. One
10780 * could back off to end with only a code point that isn't such a
10781 * non-final, but it is possible for there not to be any in the
10783 for (p = RExC_parse - 1;
10784 len < upper_parse && p < RExC_end;
10789 if (RExC_flags & RXf_PMf_EXTENDED)
10790 p = regwhite( pRExC_state, p );
10801 /* Literal Escapes Switch
10803 This switch is meant to handle escape sequences that
10804 resolve to a literal character.
10806 Every escape sequence that represents something
10807 else, like an assertion or a char class, is handled
10808 in the switch marked 'Special Escapes' above in this
10809 routine, but also has an entry here as anything that
10810 isn't explicitly mentioned here will be treated as
10811 an unescaped equivalent literal.
10814 switch ((U8)*++p) {
10815 /* These are all the special escapes. */
10816 case 'A': /* Start assertion */
10817 case 'b': case 'B': /* Word-boundary assertion*/
10818 case 'C': /* Single char !DANGEROUS! */
10819 case 'd': case 'D': /* digit class */
10820 case 'g': case 'G': /* generic-backref, pos assertion */
10821 case 'h': case 'H': /* HORIZWS */
10822 case 'k': case 'K': /* named backref, keep marker */
10823 case 'p': case 'P': /* Unicode property */
10824 case 'R': /* LNBREAK */
10825 case 's': case 'S': /* space class */
10826 case 'v': case 'V': /* VERTWS */
10827 case 'w': case 'W': /* word class */
10828 case 'X': /* eXtended Unicode "combining character sequence" */
10829 case 'z': case 'Z': /* End of line/string assertion */
10833 /* Anything after here is an escape that resolves to a
10834 literal. (Except digits, which may or may not)
10840 case 'N': /* Handle a single-code point named character. */
10841 /* The options cause it to fail if a multiple code
10842 * point sequence. Handle those in the switch() above
10844 RExC_parse = p + 1;
10845 if (! grok_bslash_N(pRExC_state, NULL, &ender,
10846 flagp, depth, FALSE,
10847 FALSE /* not strict */ ))
10849 if (*flagp & RESTART_UTF8)
10850 FAIL("panic: grok_bslash_N set RESTART_UTF8");
10851 RExC_parse = p = oldp;
10855 if (ender > 0xff) {
10872 ender = ASCII_TO_NATIVE('\033');
10876 ender = ASCII_TO_NATIVE('\007');
10882 const char* error_msg;
10884 bool valid = grok_bslash_o(&p,
10887 TRUE, /* out warnings */
10888 FALSE, /* not strict */
10889 TRUE, /* Output warnings
10894 RExC_parse = p; /* going to die anyway; point
10895 to exact spot of failure */
10899 if (PL_encoding && ender < 0x100) {
10900 goto recode_encoding;
10902 if (ender > 0xff) {
10909 UV result = UV_MAX; /* initialize to erroneous
10911 const char* error_msg;
10913 bool valid = grok_bslash_x(&p,
10916 TRUE, /* out warnings */
10917 FALSE, /* not strict */
10918 TRUE, /* Output warnings
10923 RExC_parse = p; /* going to die anyway; point
10924 to exact spot of failure */
10929 if (PL_encoding && ender < 0x100) {
10930 goto recode_encoding;
10932 if (ender > 0xff) {
10939 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
10941 case '0': case '1': case '2': case '3':case '4':
10942 case '5': case '6': case '7':
10944 (isDIGIT(p[1]) && atoi(p) >= RExC_npar))
10946 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
10948 ender = grok_oct(p, &numlen, &flags, NULL);
10949 if (ender > 0xff) {
10953 if (SIZE_ONLY /* like \08, \178 */
10956 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
10958 reg_warn_non_literal_string(
10960 form_short_octal_warning(p, numlen));
10963 else { /* Not to be treated as an octal constant, go
10968 if (PL_encoding && ender < 0x100)
10969 goto recode_encoding;
10971 case '8': case '9': /* These are illegal unless backrefs */
10972 if (atoi(p) <= RExC_npar) {
10973 --p; /* backup to backslash; handle as backref */
10978 if (! RExC_override_recoding) {
10979 SV* enc = PL_encoding;
10980 ender = reg_recode((const char)(U8)ender, &enc);
10981 if (!enc && SIZE_ONLY)
10982 ckWARNreg(p, "Invalid escape in the specified encoding");
10988 FAIL("Trailing \\");
10992 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
10993 /* Include any { following the alpha to emphasize
10994 * that it could be part of an escape at some point
10996 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
10997 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
10999 goto normal_default;
11000 } /* End of switch on '\' */
11002 default: /* A literal character */
11005 && RExC_flags & RXf_PMf_EXTENDED
11006 && ckWARN(WARN_DEPRECATED)
11007 && is_PATWS_non_low(p, UTF))
11009 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
11010 "Escape literal pattern white space under /x");
11014 if (UTF8_IS_START(*p) && UTF) {
11016 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
11017 &numlen, UTF8_ALLOW_DEFAULT);
11023 } /* End of switch on the literal */
11025 /* Here, have looked at the literal character and <ender>
11026 * contains its ordinal, <p> points to the character after it
11029 if ( RExC_flags & RXf_PMf_EXTENDED)
11030 p = regwhite( pRExC_state, p );
11032 /* If the next thing is a quantifier, it applies to this
11033 * character only, which means that this character has to be in
11034 * its own node and can't just be appended to the string in an
11035 * existing node, so if there are already other characters in
11036 * the node, close the node with just them, and set up to do
11037 * this character again next time through, when it will be the
11038 * only thing in its new node */
11039 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
11047 /* See comments for join_exact() as to why we fold
11048 * this non-UTF at compile time */
11049 || (node_type == EXACTFU
11050 && ender == LATIN_SMALL_LETTER_SHARP_S))
11054 /* Prime the casefolded buffer. Locale rules, which
11055 * apply only to code points < 256, aren't known until
11056 * execution, so for them, just output the original
11057 * character using utf8. If we start to fold non-UTF
11058 * patterns, be sure to update join_exact() */
11059 if (LOC && ender < 256) {
11060 if (UNI_IS_INVARIANT(ender)) {
11064 *s = UTF8_TWO_BYTE_HI(ender);
11065 *(s + 1) = UTF8_TWO_BYTE_LO(ender);
11070 UV folded = _to_uni_fold_flags(
11075 | ((LOC) ? FOLD_FLAGS_LOCALE
11076 : (ASCII_FOLD_RESTRICTED)
11077 ? FOLD_FLAGS_NOMIX_ASCII
11081 /* If this node only contains non-folding code
11082 * points so far, see if this new one is also
11085 if (folded != ender) {
11086 maybe_exact = FALSE;
11089 /* Here the fold is the original; we have
11090 * to check further to see if anything
11092 if (! PL_utf8_foldable) {
11093 SV* swash = swash_init("utf8",
11095 &PL_sv_undef, 1, 0);
11097 _get_swash_invlist(swash);
11098 SvREFCNT_dec_NN(swash);
11100 if (_invlist_contains_cp(PL_utf8_foldable,
11103 maybe_exact = FALSE;
11111 /* The loop increments <len> each time, as all but this
11112 * path (and the one just below for UTF) through it add
11113 * a single byte to the EXACTish node. But this one
11114 * has changed len to be the correct final value, so
11115 * subtract one to cancel out the increment that
11117 len += foldlen - 1;
11120 *(s++) = (char) ender;
11121 maybe_exact &= ! IS_IN_SOME_FOLD_L1(ender);
11125 const STRLEN unilen = reguni(pRExC_state, ender, s);
11131 /* See comment just above for - 1 */
11135 REGC((char)ender, s++);
11138 if (next_is_quantifier) {
11140 /* Here, the next input is a quantifier, and to get here,
11141 * the current character is the only one in the node.
11142 * Also, here <len> doesn't include the final byte for this
11148 } /* End of loop through literal characters */
11150 /* Here we have either exhausted the input or ran out of room in
11151 * the node. (If we encountered a character that can't be in the
11152 * node, transfer is made directly to <loopdone>, and so we
11153 * wouldn't have fallen off the end of the loop.) In the latter
11154 * case, we artificially have to split the node into two, because
11155 * we just don't have enough space to hold everything. This
11156 * creates a problem if the final character participates in a
11157 * multi-character fold in the non-final position, as a match that
11158 * should have occurred won't, due to the way nodes are matched,
11159 * and our artificial boundary. So back off until we find a non-
11160 * problematic character -- one that isn't at the beginning or
11161 * middle of such a fold. (Either it doesn't participate in any
11162 * folds, or appears only in the final position of all the folds it
11163 * does participate in.) A better solution with far fewer false
11164 * positives, and that would fill the nodes more completely, would
11165 * be to actually have available all the multi-character folds to
11166 * test against, and to back-off only far enough to be sure that
11167 * this node isn't ending with a partial one. <upper_parse> is set
11168 * further below (if we need to reparse the node) to include just
11169 * up through that final non-problematic character that this code
11170 * identifies, so when it is set to less than the full node, we can
11171 * skip the rest of this */
11172 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
11174 const STRLEN full_len = len;
11176 assert(len >= MAX_NODE_STRING_SIZE);
11178 /* Here, <s> points to the final byte of the final character.
11179 * Look backwards through the string until find a non-
11180 * problematic character */
11184 /* These two have no multi-char folds to non-UTF characters
11186 if (ASCII_FOLD_RESTRICTED || LOC) {
11190 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
11194 if (! PL_NonL1NonFinalFold) {
11195 PL_NonL1NonFinalFold = _new_invlist_C_array(
11196 NonL1_Perl_Non_Final_Folds_invlist);
11199 /* Point to the first byte of the final character */
11200 s = (char *) utf8_hop((U8 *) s, -1);
11202 while (s >= s0) { /* Search backwards until find
11203 non-problematic char */
11204 if (UTF8_IS_INVARIANT(*s)) {
11206 /* There are no ascii characters that participate
11207 * in multi-char folds under /aa. In EBCDIC, the
11208 * non-ascii invariants are all control characters,
11209 * so don't ever participate in any folds. */
11210 if (ASCII_FOLD_RESTRICTED
11211 || ! IS_NON_FINAL_FOLD(*s))
11216 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
11218 /* No Latin1 characters participate in multi-char
11219 * folds under /l */
11221 || ! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_UNI(
11227 else if (! _invlist_contains_cp(
11228 PL_NonL1NonFinalFold,
11229 valid_utf8_to_uvchr((U8 *) s, NULL)))
11234 /* Here, the current character is problematic in that
11235 * it does occur in the non-final position of some
11236 * fold, so try the character before it, but have to
11237 * special case the very first byte in the string, so
11238 * we don't read outside the string */
11239 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
11240 } /* End of loop backwards through the string */
11242 /* If there were only problematic characters in the string,
11243 * <s> will point to before s0, in which case the length
11244 * should be 0, otherwise include the length of the
11245 * non-problematic character just found */
11246 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
11249 /* Here, have found the final character, if any, that is
11250 * non-problematic as far as ending the node without splitting
11251 * it across a potential multi-char fold. <len> contains the
11252 * number of bytes in the node up-to and including that
11253 * character, or is 0 if there is no such character, meaning
11254 * the whole node contains only problematic characters. In
11255 * this case, give up and just take the node as-is. We can't
11261 /* Here, the node does contain some characters that aren't
11262 * problematic. If one such is the final character in the
11263 * node, we are done */
11264 if (len == full_len) {
11267 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
11269 /* If the final character is problematic, but the
11270 * penultimate is not, back-off that last character to
11271 * later start a new node with it */
11276 /* Here, the final non-problematic character is earlier
11277 * in the input than the penultimate character. What we do
11278 * is reparse from the beginning, going up only as far as
11279 * this final ok one, thus guaranteeing that the node ends
11280 * in an acceptable character. The reason we reparse is
11281 * that we know how far in the character is, but we don't
11282 * know how to correlate its position with the input parse.
11283 * An alternate implementation would be to build that
11284 * correlation as we go along during the original parse,
11285 * but that would entail extra work for every node, whereas
11286 * this code gets executed only when the string is too
11287 * large for the node, and the final two characters are
11288 * problematic, an infrequent occurrence. Yet another
11289 * possible strategy would be to save the tail of the
11290 * string, and the next time regatom is called, initialize
11291 * with that. The problem with this is that unless you
11292 * back off one more character, you won't be guaranteed
11293 * regatom will get called again, unless regbranch,
11294 * regpiece ... are also changed. If you do back off that
11295 * extra character, so that there is input guaranteed to
11296 * force calling regatom, you can't handle the case where
11297 * just the first character in the node is acceptable. I
11298 * (khw) decided to try this method which doesn't have that
11299 * pitfall; if performance issues are found, we can do a
11300 * combination of the current approach plus that one */
11306 } /* End of verifying node ends with an appropriate char */
11308 loopdone: /* Jumped to when encounters something that shouldn't be in
11311 /* If 'maybe_exact' is still set here, means there are no
11312 * code points in the node that participate in folds */
11313 if (FOLD && maybe_exact) {
11317 /* I (khw) don't know if you can get here with zero length, but the
11318 * old code handled this situation by creating a zero-length EXACT
11319 * node. Might as well be NOTHING instead */
11324 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
11327 RExC_parse = p - 1;
11328 Set_Node_Cur_Length(ret); /* MJD */
11329 nextchar(pRExC_state);
11331 /* len is STRLEN which is unsigned, need to copy to signed */
11334 vFAIL("Internal disaster");
11337 } /* End of label 'defchar:' */
11339 } /* End of giant switch on input character */
11345 S_regwhite( RExC_state_t *pRExC_state, char *p )
11347 const char *e = RExC_end;
11349 PERL_ARGS_ASSERT_REGWHITE;
11354 else if (*p == '#') {
11357 if (*p++ == '\n') {
11363 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11372 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
11374 /* Returns the next non-pattern-white space, non-comment character (the
11375 * latter only if 'recognize_comment is true) in the string p, which is
11376 * ended by RExC_end. If there is no line break ending a comment,
11377 * RExC_seen has added the REG_SEEN_RUN_ON_COMMENT flag; */
11378 const char *e = RExC_end;
11380 PERL_ARGS_ASSERT_REGPATWS;
11384 if ((len = is_PATWS_safe(p, e, UTF))) {
11387 else if (recognize_comment && *p == '#') {
11391 if (is_LNBREAK_safe(p, e, UTF)) {
11397 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11405 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
11406 Character classes ([:foo:]) can also be negated ([:^foo:]).
11407 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
11408 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
11409 but trigger failures because they are currently unimplemented. */
11411 #define POSIXCC_DONE(c) ((c) == ':')
11412 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
11413 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
11415 PERL_STATIC_INLINE I32
11416 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
11419 I32 namedclass = OOB_NAMEDCLASS;
11421 PERL_ARGS_ASSERT_REGPPOSIXCC;
11423 if (value == '[' && RExC_parse + 1 < RExC_end &&
11424 /* I smell either [: or [= or [. -- POSIX has been here, right? */
11425 POSIXCC(UCHARAT(RExC_parse)))
11427 const char c = UCHARAT(RExC_parse);
11428 char* const s = RExC_parse++;
11430 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
11432 if (RExC_parse == RExC_end) {
11435 /* Try to give a better location for the error (than the end of
11436 * the string) by looking for the matching ']' */
11438 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
11441 vFAIL2("Unmatched '%c' in POSIX class", c);
11443 /* Grandfather lone [:, [=, [. */
11447 const char* const t = RExC_parse++; /* skip over the c */
11450 if (UCHARAT(RExC_parse) == ']') {
11451 const char *posixcc = s + 1;
11452 RExC_parse++; /* skip over the ending ] */
11455 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
11456 const I32 skip = t - posixcc;
11458 /* Initially switch on the length of the name. */
11461 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
11462 this is the Perl \w
11464 namedclass = ANYOF_WORDCHAR;
11467 /* Names all of length 5. */
11468 /* alnum alpha ascii blank cntrl digit graph lower
11469 print punct space upper */
11470 /* Offset 4 gives the best switch position. */
11471 switch (posixcc[4]) {
11473 if (memEQ(posixcc, "alph", 4)) /* alpha */
11474 namedclass = ANYOF_ALPHA;
11477 if (memEQ(posixcc, "spac", 4)) /* space */
11478 namedclass = ANYOF_PSXSPC;
11481 if (memEQ(posixcc, "grap", 4)) /* graph */
11482 namedclass = ANYOF_GRAPH;
11485 if (memEQ(posixcc, "asci", 4)) /* ascii */
11486 namedclass = ANYOF_ASCII;
11489 if (memEQ(posixcc, "blan", 4)) /* blank */
11490 namedclass = ANYOF_BLANK;
11493 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
11494 namedclass = ANYOF_CNTRL;
11497 if (memEQ(posixcc, "alnu", 4)) /* alnum */
11498 namedclass = ANYOF_ALPHANUMERIC;
11501 if (memEQ(posixcc, "lowe", 4)) /* lower */
11502 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
11503 else if (memEQ(posixcc, "uppe", 4)) /* upper */
11504 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
11507 if (memEQ(posixcc, "digi", 4)) /* digit */
11508 namedclass = ANYOF_DIGIT;
11509 else if (memEQ(posixcc, "prin", 4)) /* print */
11510 namedclass = ANYOF_PRINT;
11511 else if (memEQ(posixcc, "punc", 4)) /* punct */
11512 namedclass = ANYOF_PUNCT;
11517 if (memEQ(posixcc, "xdigit", 6))
11518 namedclass = ANYOF_XDIGIT;
11522 if (namedclass == OOB_NAMEDCLASS)
11523 Simple_vFAIL3("POSIX class [:%.*s:] unknown",
11526 /* The #defines are structured so each complement is +1 to
11527 * the normal one */
11531 assert (posixcc[skip] == ':');
11532 assert (posixcc[skip+1] == ']');
11533 } else if (!SIZE_ONLY) {
11534 /* [[=foo=]] and [[.foo.]] are still future. */
11536 /* adjust RExC_parse so the warning shows after
11537 the class closes */
11538 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
11540 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11543 /* Maternal grandfather:
11544 * "[:" ending in ":" but not in ":]" */
11546 vFAIL("Unmatched '[' in POSIX class");
11549 /* Grandfather lone [:, [=, [. */
11559 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
11561 /* This applies some heuristics at the current parse position (which should
11562 * be at a '[') to see if what follows might be intended to be a [:posix:]
11563 * class. It returns true if it really is a posix class, of course, but it
11564 * also can return true if it thinks that what was intended was a posix
11565 * class that didn't quite make it.
11567 * It will return true for
11569 * [:alphanumerics] (as long as the ] isn't followed immediately by a
11570 * ')' indicating the end of the (?[
11571 * [:any garbage including %^&$ punctuation:]
11573 * This is designed to be called only from S_handle_regex_sets; it could be
11574 * easily adapted to be called from the spot at the beginning of regclass()
11575 * that checks to see in a normal bracketed class if the surrounding []
11576 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
11577 * change long-standing behavior, so I (khw) didn't do that */
11578 char* p = RExC_parse + 1;
11579 char first_char = *p;
11581 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
11583 assert(*(p - 1) == '[');
11585 if (! POSIXCC(first_char)) {
11590 while (p < RExC_end && isWORDCHAR(*p)) p++;
11592 if (p >= RExC_end) {
11596 if (p - RExC_parse > 2 /* Got at least 1 word character */
11597 && (*p == first_char
11598 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
11603 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
11606 && p - RExC_parse > 2 /* [:] evaluates to colon;
11607 [::] is a bad posix class. */
11608 && first_char == *(p - 1));
11612 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist, I32 *flagp, U32 depth,
11613 char * const oregcomp_parse)
11615 /* Handle the (?[...]) construct to do set operations */
11618 UV start, end; /* End points of code point ranges */
11620 char *save_end, *save_parse;
11625 const bool save_fold = FOLD;
11627 GET_RE_DEBUG_FLAGS_DECL;
11629 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
11632 vFAIL("(?[...]) not valid in locale");
11634 RExC_uni_semantics = 1;
11636 /* This will return only an ANYOF regnode, or (unlikely) something smaller
11637 * (such as EXACT). Thus we can skip most everything if just sizing. We
11638 * call regclass to handle '[]' so as to not have to reinvent its parsing
11639 * rules here (throwing away the size it computes each time). And, we exit
11640 * upon an unescaped ']' that isn't one ending a regclass. To do both
11641 * these things, we need to realize that something preceded by a backslash
11642 * is escaped, so we have to keep track of backslashes */
11645 Perl_ck_warner_d(aTHX_
11646 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
11647 "The regex_sets feature is experimental" REPORT_LOCATION,
11648 (int) (RExC_parse - RExC_precomp) , RExC_precomp, RExC_parse);
11650 while (RExC_parse < RExC_end) {
11651 SV* current = NULL;
11652 RExC_parse = regpatws(pRExC_state, RExC_parse,
11653 TRUE); /* means recognize comments */
11654 switch (*RExC_parse) {
11658 /* Skip the next byte (which could cause us to end up in
11659 * the middle of a UTF-8 character, but since none of those
11660 * are confusable with anything we currently handle in this
11661 * switch (invariants all), it's safe. We'll just hit the
11662 * default: case next time and keep on incrementing until
11663 * we find one of the invariants we do handle. */
11668 /* If this looks like it is a [:posix:] class, leave the
11669 * parse pointer at the '[' to fool regclass() into
11670 * thinking it is part of a '[[:posix:]]'. That function
11671 * will use strict checking to force a syntax error if it
11672 * doesn't work out to a legitimate class */
11673 bool is_posix_class
11674 = could_it_be_a_POSIX_class(pRExC_state);
11675 if (! is_posix_class) {
11679 /* regclass() can only return RESTART_UTF8 if multi-char
11680 folds are allowed. */
11681 if (!regclass(pRExC_state, flagp,depth+1,
11682 is_posix_class, /* parse the whole char
11683 class only if not a
11685 FALSE, /* don't allow multi-char folds */
11686 TRUE, /* silence non-portable warnings. */
11688 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X",
11691 /* function call leaves parse pointing to the ']', except
11692 * if we faked it */
11693 if (is_posix_class) {
11697 SvREFCNT_dec(current); /* In case it returned something */
11703 if (RExC_parse < RExC_end
11704 && *RExC_parse == ')')
11706 node = reganode(pRExC_state, ANYOF, 0);
11707 RExC_size += ANYOF_SKIP;
11708 nextchar(pRExC_state);
11709 Set_Node_Length(node,
11710 RExC_parse - oregcomp_parse + 1); /* MJD */
11719 FAIL("Syntax error in (?[...])");
11722 /* Pass 2 only after this. Everything in this construct is a
11723 * metacharacter. Operands begin with either a '\' (for an escape
11724 * sequence), or a '[' for a bracketed character class. Any other
11725 * character should be an operator, or parenthesis for grouping. Both
11726 * types of operands are handled by calling regclass() to parse them. It
11727 * is called with a parameter to indicate to return the computed inversion
11728 * list. The parsing here is implemented via a stack. Each entry on the
11729 * stack is a single character representing one of the operators, or the
11730 * '('; or else a pointer to an operand inversion list. */
11732 #define IS_OPERAND(a) (! SvIOK(a))
11734 /* The stack starts empty. It is a syntax error if the first thing parsed
11735 * is a binary operator; everything else is pushed on the stack. When an
11736 * operand is parsed, the top of the stack is examined. If it is a binary
11737 * operator, the item before it should be an operand, and both are replaced
11738 * by the result of doing that operation on the new operand and the one on
11739 * the stack. Thus a sequence of binary operands is reduced to a single
11740 * one before the next one is parsed.
11742 * A unary operator may immediately follow a binary in the input, for
11745 * When an operand is parsed and the top of the stack is a unary operator,
11746 * the operation is performed, and then the stack is rechecked to see if
11747 * this new operand is part of a binary operation; if so, it is handled as
11750 * A '(' is simply pushed on the stack; it is valid only if the stack is
11751 * empty, or the top element of the stack is an operator or another '('
11752 * (for which the parenthesized expression will become an operand). By the
11753 * time the corresponding ')' is parsed everything in between should have
11754 * been parsed and evaluated to a single operand (or else is a syntax
11755 * error), and is handled as a regular operand */
11757 sv_2mortal((SV *)(stack = newAV()));
11759 while (RExC_parse < RExC_end) {
11760 I32 top_index = av_tindex(stack);
11762 SV* current = NULL;
11764 /* Skip white space */
11765 RExC_parse = regpatws(pRExC_state, RExC_parse,
11766 TRUE); /* means recognize comments */
11767 if (RExC_parse >= RExC_end) {
11768 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
11770 if ((curchar = UCHARAT(RExC_parse)) == ']') {
11777 if (av_tindex(stack) >= 0 /* This makes sure that we can
11778 safely subtract 1 from
11779 RExC_parse in the next clause.
11780 If we have something on the
11781 stack, we have parsed something
11783 && UCHARAT(RExC_parse - 1) == '('
11784 && RExC_parse < RExC_end)
11786 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
11787 * This happens when we have some thing like
11789 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
11791 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
11793 * Here we would be handling the interpolated
11794 * '$thai_or_lao'. We handle this by a recursive call to
11795 * ourselves which returns the inversion list the
11796 * interpolated expression evaluates to. We use the flags
11797 * from the interpolated pattern. */
11798 U32 save_flags = RExC_flags;
11799 const char * const save_parse = ++RExC_parse;
11801 parse_lparen_question_flags(pRExC_state);
11803 if (RExC_parse == save_parse /* Makes sure there was at
11804 least one flag (or this
11805 embedding wasn't compiled)
11807 || RExC_parse >= RExC_end - 4
11808 || UCHARAT(RExC_parse) != ':'
11809 || UCHARAT(++RExC_parse) != '('
11810 || UCHARAT(++RExC_parse) != '?'
11811 || UCHARAT(++RExC_parse) != '[')
11814 /* In combination with the above, this moves the
11815 * pointer to the point just after the first erroneous
11816 * character (or if there are no flags, to where they
11817 * should have been) */
11818 if (RExC_parse >= RExC_end - 4) {
11819 RExC_parse = RExC_end;
11821 else if (RExC_parse != save_parse) {
11822 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11824 vFAIL("Expecting '(?flags:(?[...'");
11827 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
11828 depth+1, oregcomp_parse);
11830 /* Here, 'current' contains the embedded expression's
11831 * inversion list, and RExC_parse points to the trailing
11832 * ']'; the next character should be the ')' which will be
11833 * paired with the '(' that has been put on the stack, so
11834 * the whole embedded expression reduces to '(operand)' */
11837 RExC_flags = save_flags;
11838 goto handle_operand;
11843 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11844 vFAIL("Unexpected character");
11847 /* regclass() can only return RESTART_UTF8 if multi-char
11848 folds are allowed. */
11849 if (!regclass(pRExC_state, flagp,depth+1,
11850 TRUE, /* means parse just the next thing */
11851 FALSE, /* don't allow multi-char folds */
11852 FALSE, /* don't silence non-portable warnings. */
11854 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X",
11856 /* regclass() will return with parsing just the \ sequence,
11857 * leaving the parse pointer at the next thing to parse */
11859 goto handle_operand;
11861 case '[': /* Is a bracketed character class */
11863 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
11865 if (! is_posix_class) {
11869 /* regclass() can only return RESTART_UTF8 if multi-char
11870 folds are allowed. */
11871 if(!regclass(pRExC_state, flagp,depth+1,
11872 is_posix_class, /* parse the whole char class
11873 only if not a posix class */
11874 FALSE, /* don't allow multi-char folds */
11875 FALSE, /* don't silence non-portable warnings. */
11877 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#X",
11879 /* function call leaves parse pointing to the ']', except if we
11881 if (is_posix_class) {
11885 goto handle_operand;
11894 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
11895 || ! IS_OPERAND(*top_ptr))
11898 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
11900 av_push(stack, newSVuv(curchar));
11904 av_push(stack, newSVuv(curchar));
11908 if (top_index >= 0) {
11909 top_ptr = av_fetch(stack, top_index, FALSE);
11911 if (IS_OPERAND(*top_ptr)) {
11913 vFAIL("Unexpected '(' with no preceding operator");
11916 av_push(stack, newSVuv(curchar));
11923 || ! (current = av_pop(stack))
11924 || ! IS_OPERAND(current)
11925 || ! (lparen = av_pop(stack))
11926 || IS_OPERAND(lparen)
11927 || SvUV(lparen) != '(')
11929 SvREFCNT_dec(current);
11931 vFAIL("Unexpected ')'");
11934 SvREFCNT_dec_NN(lparen);
11941 /* Here, we have an operand to process, in 'current' */
11943 if (top_index < 0) { /* Just push if stack is empty */
11944 av_push(stack, current);
11947 SV* top = av_pop(stack);
11949 char current_operator;
11951 if (IS_OPERAND(top)) {
11952 SvREFCNT_dec_NN(top);
11953 SvREFCNT_dec_NN(current);
11954 vFAIL("Operand with no preceding operator");
11956 current_operator = (char) SvUV(top);
11957 switch (current_operator) {
11958 case '(': /* Push the '(' back on followed by the new
11960 av_push(stack, top);
11961 av_push(stack, current);
11962 SvREFCNT_inc(top); /* Counters the '_dec' done
11963 just after the 'break', so
11964 it doesn't get wrongly freed
11969 _invlist_invert(current);
11971 /* Unlike binary operators, the top of the stack,
11972 * now that this unary one has been popped off, may
11973 * legally be an operator, and we now have operand
11976 SvREFCNT_dec_NN(top);
11977 goto handle_operand;
11980 prev = av_pop(stack);
11981 _invlist_intersection(prev,
11984 av_push(stack, current);
11989 prev = av_pop(stack);
11990 _invlist_union(prev, current, ¤t);
11991 av_push(stack, current);
11995 prev = av_pop(stack);;
11996 _invlist_subtract(prev, current, ¤t);
11997 av_push(stack, current);
12000 case '^': /* The union minus the intersection */
12006 prev = av_pop(stack);
12007 _invlist_union(prev, current, &u);
12008 _invlist_intersection(prev, current, &i);
12009 /* _invlist_subtract will overwrite current
12010 without freeing what it already contains */
12012 _invlist_subtract(u, i, ¤t);
12013 av_push(stack, current);
12014 SvREFCNT_dec_NN(i);
12015 SvREFCNT_dec_NN(u);
12016 SvREFCNT_dec_NN(element);
12021 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
12023 SvREFCNT_dec_NN(top);
12024 SvREFCNT_dec(prev);
12028 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12031 if (av_tindex(stack) < 0 /* Was empty */
12032 || ((final = av_pop(stack)) == NULL)
12033 || ! IS_OPERAND(final)
12034 || av_tindex(stack) >= 0) /* More left on stack */
12036 vFAIL("Incomplete expression within '(?[ ])'");
12039 /* Here, 'final' is the resultant inversion list from evaluating the
12040 * expression. Return it if so requested */
12041 if (return_invlist) {
12042 *return_invlist = final;
12046 /* Otherwise generate a resultant node, based on 'final'. regclass() is
12047 * expecting a string of ranges and individual code points */
12048 invlist_iterinit(final);
12049 result_string = newSVpvs("");
12050 while (invlist_iternext(final, &start, &end)) {
12051 if (start == end) {
12052 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
12055 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
12060 save_parse = RExC_parse;
12061 RExC_parse = SvPV(result_string, len);
12062 save_end = RExC_end;
12063 RExC_end = RExC_parse + len;
12065 /* We turn off folding around the call, as the class we have constructed
12066 * already has all folding taken into consideration, and we don't want
12067 * regclass() to add to that */
12068 RExC_flags &= ~RXf_PMf_FOLD;
12069 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
12071 node = regclass(pRExC_state, flagp,depth+1,
12072 FALSE, /* means parse the whole char class */
12073 FALSE, /* don't allow multi-char folds */
12074 TRUE, /* silence non-portable warnings. The above may very
12075 well have generated non-portable code points, but
12076 they're valid on this machine */
12079 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
12082 RExC_flags |= RXf_PMf_FOLD;
12084 RExC_parse = save_parse + 1;
12085 RExC_end = save_end;
12086 SvREFCNT_dec_NN(final);
12087 SvREFCNT_dec_NN(result_string);
12089 nextchar(pRExC_state);
12090 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
12095 /* The names of properties whose definitions are not known at compile time are
12096 * stored in this SV, after a constant heading. So if the length has been
12097 * changed since initialization, then there is a run-time definition. */
12098 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION (SvCUR(listsv) != initial_listsv_len)
12101 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
12102 const bool stop_at_1, /* Just parse the next thing, don't
12103 look for a full character class */
12104 bool allow_multi_folds,
12105 const bool silence_non_portable, /* Don't output warnings
12108 SV** ret_invlist) /* Return an inversion list, not a node */
12110 /* parse a bracketed class specification. Most of these will produce an
12111 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
12112 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
12113 * under /i with multi-character folds: it will be rewritten following the
12114 * paradigm of this example, where the <multi-fold>s are characters which
12115 * fold to multiple character sequences:
12116 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
12117 * gets effectively rewritten as:
12118 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
12119 * reg() gets called (recursively) on the rewritten version, and this
12120 * function will return what it constructs. (Actually the <multi-fold>s
12121 * aren't physically removed from the [abcdefghi], it's just that they are
12122 * ignored in the recursion by means of a flag:
12123 * <RExC_in_multi_char_class>.)
12125 * ANYOF nodes contain a bit map for the first 256 characters, with the
12126 * corresponding bit set if that character is in the list. For characters
12127 * above 255, a range list or swash is used. There are extra bits for \w,
12128 * etc. in locale ANYOFs, as what these match is not determinable at
12131 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
12132 * to be restarted. This can only happen if ret_invlist is non-NULL.
12136 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
12138 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
12141 IV namedclass = OOB_NAMEDCLASS;
12142 char *rangebegin = NULL;
12143 bool need_class = 0;
12145 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
12146 than just initialized. */
12147 SV* properties = NULL; /* Code points that match \p{} \P{} */
12148 SV* posixes = NULL; /* Code points that match classes like, [:word:],
12149 extended beyond the Latin1 range */
12150 UV element_count = 0; /* Number of distinct elements in the class.
12151 Optimizations may be possible if this is tiny */
12152 AV * multi_char_matches = NULL; /* Code points that fold to more than one
12153 character; used under /i */
12155 char * stop_ptr = RExC_end; /* where to stop parsing */
12156 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
12158 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
12160 /* Unicode properties are stored in a swash; this holds the current one
12161 * being parsed. If this swash is the only above-latin1 component of the
12162 * character class, an optimization is to pass it directly on to the
12163 * execution engine. Otherwise, it is set to NULL to indicate that there
12164 * are other things in the class that have to be dealt with at execution
12166 SV* swash = NULL; /* Code points that match \p{} \P{} */
12168 /* Set if a component of this character class is user-defined; just passed
12169 * on to the engine */
12170 bool has_user_defined_property = FALSE;
12172 /* inversion list of code points this node matches only when the target
12173 * string is in UTF-8. (Because is under /d) */
12174 SV* depends_list = NULL;
12176 /* inversion list of code points this node matches. For much of the
12177 * function, it includes only those that match regardless of the utf8ness
12178 * of the target string */
12179 SV* cp_list = NULL;
12182 /* In a range, counts how many 0-2 of the ends of it came from literals,
12183 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
12184 UV literal_endpoint = 0;
12186 bool invert = FALSE; /* Is this class to be complemented */
12188 /* Is there any thing like \W or [:^digit:] that matches above the legal
12189 * Unicode range? */
12190 bool runtime_posix_matches_above_Unicode = FALSE;
12192 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
12193 case we need to change the emitted regop to an EXACT. */
12194 const char * orig_parse = RExC_parse;
12195 const I32 orig_size = RExC_size;
12196 GET_RE_DEBUG_FLAGS_DECL;
12198 PERL_ARGS_ASSERT_REGCLASS;
12200 PERL_UNUSED_ARG(depth);
12203 DEBUG_PARSE("clas");
12205 /* Assume we are going to generate an ANYOF node. */
12206 ret = reganode(pRExC_state, ANYOF, 0);
12209 RExC_size += ANYOF_SKIP;
12210 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
12213 ANYOF_FLAGS(ret) = 0;
12215 RExC_emit += ANYOF_SKIP;
12217 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
12219 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
12220 initial_listsv_len = SvCUR(listsv);
12221 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
12225 RExC_parse = regpatws(pRExC_state, RExC_parse,
12226 FALSE /* means don't recognize comments */);
12229 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
12232 allow_multi_folds = FALSE;
12235 RExC_parse = regpatws(pRExC_state, RExC_parse,
12236 FALSE /* means don't recognize comments */);
12240 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
12241 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
12242 const char *s = RExC_parse;
12243 const char c = *s++;
12245 while (isWORDCHAR(*s))
12247 if (*s && c == *s && s[1] == ']') {
12248 SAVEFREESV(RExC_rx_sv);
12250 "POSIX syntax [%c %c] belongs inside character classes",
12252 (void)ReREFCNT_inc(RExC_rx_sv);
12256 /* If the caller wants us to just parse a single element, accomplish this
12257 * by faking the loop ending condition */
12258 if (stop_at_1 && RExC_end > RExC_parse) {
12259 stop_ptr = RExC_parse + 1;
12262 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
12263 if (UCHARAT(RExC_parse) == ']')
12264 goto charclassloop;
12268 if (RExC_parse >= stop_ptr) {
12273 RExC_parse = regpatws(pRExC_state, RExC_parse,
12274 FALSE /* means don't recognize comments */);
12277 if (UCHARAT(RExC_parse) == ']') {
12283 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
12284 save_value = value;
12285 save_prevvalue = prevvalue;
12288 rangebegin = RExC_parse;
12292 value = utf8n_to_uvchr((U8*)RExC_parse,
12293 RExC_end - RExC_parse,
12294 &numlen, UTF8_ALLOW_DEFAULT);
12295 RExC_parse += numlen;
12298 value = UCHARAT(RExC_parse++);
12301 && RExC_parse < RExC_end
12302 && POSIXCC(UCHARAT(RExC_parse)))
12304 namedclass = regpposixcc(pRExC_state, value, strict);
12306 else if (value == '\\') {
12308 value = utf8n_to_uvchr((U8*)RExC_parse,
12309 RExC_end - RExC_parse,
12310 &numlen, UTF8_ALLOW_DEFAULT);
12311 RExC_parse += numlen;
12314 value = UCHARAT(RExC_parse++);
12316 /* Some compilers cannot handle switching on 64-bit integer
12317 * values, therefore value cannot be an UV. Yes, this will
12318 * be a problem later if we want switch on Unicode.
12319 * A similar issue a little bit later when switching on
12320 * namedclass. --jhi */
12322 /* If the \ is escaping white space when white space is being
12323 * skipped, it means that that white space is wanted literally, and
12324 * is already in 'value'. Otherwise, need to translate the escape
12325 * into what it signifies. */
12326 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
12328 case 'w': namedclass = ANYOF_WORDCHAR; break;
12329 case 'W': namedclass = ANYOF_NWORDCHAR; break;
12330 case 's': namedclass = ANYOF_SPACE; break;
12331 case 'S': namedclass = ANYOF_NSPACE; break;
12332 case 'd': namedclass = ANYOF_DIGIT; break;
12333 case 'D': namedclass = ANYOF_NDIGIT; break;
12334 case 'v': namedclass = ANYOF_VERTWS; break;
12335 case 'V': namedclass = ANYOF_NVERTWS; break;
12336 case 'h': namedclass = ANYOF_HORIZWS; break;
12337 case 'H': namedclass = ANYOF_NHORIZWS; break;
12338 case 'N': /* Handle \N{NAME} in class */
12340 /* We only pay attention to the first char of
12341 multichar strings being returned. I kinda wonder
12342 if this makes sense as it does change the behaviour
12343 from earlier versions, OTOH that behaviour was broken
12345 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
12346 TRUE, /* => charclass */
12349 if (*flagp & RESTART_UTF8)
12350 FAIL("panic: grok_bslash_N set RESTART_UTF8");
12360 /* We will handle any undefined properties ourselves */
12361 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF;
12363 if (RExC_parse >= RExC_end)
12364 vFAIL2("Empty \\%c{}", (U8)value);
12365 if (*RExC_parse == '{') {
12366 const U8 c = (U8)value;
12367 e = strchr(RExC_parse++, '}');
12369 vFAIL2("Missing right brace on \\%c{}", c);
12370 while (isSPACE(UCHARAT(RExC_parse)))
12372 if (e == RExC_parse)
12373 vFAIL2("Empty \\%c{}", c);
12374 n = e - RExC_parse;
12375 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
12386 if (UCHARAT(RExC_parse) == '^') {
12389 /* toggle. (The rhs xor gets the single bit that
12390 * differs between P and p; the other xor inverts just
12392 value ^= 'P' ^ 'p';
12394 while (isSPACE(UCHARAT(RExC_parse))) {
12399 /* Try to get the definition of the property into
12400 * <invlist>. If /i is in effect, the effective property
12401 * will have its name be <__NAME_i>. The design is
12402 * discussed in commit
12403 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
12404 Newx(name, n + sizeof("_i__\n"), char);
12406 sprintf(name, "%s%.*s%s\n",
12407 (FOLD) ? "__" : "",
12413 /* Look up the property name, and get its swash and
12414 * inversion list, if the property is found */
12416 SvREFCNT_dec_NN(swash);
12418 swash = _core_swash_init("utf8", name, &PL_sv_undef,
12421 NULL, /* No inversion list */
12424 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
12426 SvREFCNT_dec_NN(swash);
12430 /* Here didn't find it. It could be a user-defined
12431 * property that will be available at run-time. If we
12432 * accept only compile-time properties, is an error;
12433 * otherwise add it to the list for run-time look up */
12435 RExC_parse = e + 1;
12436 vFAIL3("Property '%.*s' is unknown", (int) n, name);
12438 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s\n",
12439 (value == 'p' ? '+' : '!'),
12441 has_user_defined_property = TRUE;
12443 /* We don't know yet, so have to assume that the
12444 * property could match something in the Latin1 range,
12445 * hence something that isn't utf8. Note that this
12446 * would cause things in <depends_list> to match
12447 * inappropriately, except that any \p{}, including
12448 * this one forces Unicode semantics, which means there
12449 * is <no depends_list> */
12450 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
12454 /* Here, did get the swash and its inversion list. If
12455 * the swash is from a user-defined property, then this
12456 * whole character class should be regarded as such */
12457 has_user_defined_property =
12459 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY);
12461 /* Invert if asking for the complement */
12462 if (value == 'P') {
12463 _invlist_union_complement_2nd(properties,
12467 /* The swash can't be used as-is, because we've
12468 * inverted things; delay removing it to here after
12469 * have copied its invlist above */
12470 SvREFCNT_dec_NN(swash);
12474 _invlist_union(properties, invlist, &properties);
12479 RExC_parse = e + 1;
12480 namedclass = ANYOF_UNIPROP; /* no official name, but it's
12483 /* \p means they want Unicode semantics */
12484 RExC_uni_semantics = 1;
12487 case 'n': value = '\n'; break;
12488 case 'r': value = '\r'; break;
12489 case 't': value = '\t'; break;
12490 case 'f': value = '\f'; break;
12491 case 'b': value = '\b'; break;
12492 case 'e': value = ASCII_TO_NATIVE('\033');break;
12493 case 'a': value = ASCII_TO_NATIVE('\007');break;
12495 RExC_parse--; /* function expects to be pointed at the 'o' */
12497 const char* error_msg;
12498 bool valid = grok_bslash_o(&RExC_parse,
12501 SIZE_ONLY, /* warnings in pass
12504 silence_non_portable,
12510 if (PL_encoding && value < 0x100) {
12511 goto recode_encoding;
12515 RExC_parse--; /* function expects to be pointed at the 'x' */
12517 const char* error_msg;
12518 bool valid = grok_bslash_x(&RExC_parse,
12521 TRUE, /* Output warnings */
12523 silence_non_portable,
12529 if (PL_encoding && value < 0x100)
12530 goto recode_encoding;
12533 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
12535 case '0': case '1': case '2': case '3': case '4':
12536 case '5': case '6': case '7':
12538 /* Take 1-3 octal digits */
12539 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12540 numlen = (strict) ? 4 : 3;
12541 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
12542 RExC_parse += numlen;
12545 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12546 vFAIL("Need exactly 3 octal digits");
12548 else if (! SIZE_ONLY /* like \08, \178 */
12550 && RExC_parse < RExC_end
12551 && isDIGIT(*RExC_parse)
12552 && ckWARN(WARN_REGEXP))
12554 SAVEFREESV(RExC_rx_sv);
12555 reg_warn_non_literal_string(
12557 form_short_octal_warning(RExC_parse, numlen));
12558 (void)ReREFCNT_inc(RExC_rx_sv);
12561 if (PL_encoding && value < 0x100)
12562 goto recode_encoding;
12566 if (! RExC_override_recoding) {
12567 SV* enc = PL_encoding;
12568 value = reg_recode((const char)(U8)value, &enc);
12571 vFAIL("Invalid escape in the specified encoding");
12573 else if (SIZE_ONLY) {
12574 ckWARNreg(RExC_parse,
12575 "Invalid escape in the specified encoding");
12581 /* Allow \_ to not give an error */
12582 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
12584 vFAIL2("Unrecognized escape \\%c in character class",
12588 SAVEFREESV(RExC_rx_sv);
12589 ckWARN2reg(RExC_parse,
12590 "Unrecognized escape \\%c in character class passed through",
12592 (void)ReREFCNT_inc(RExC_rx_sv);
12596 } /* End of switch on char following backslash */
12597 } /* end of handling backslash escape sequences */
12600 literal_endpoint++;
12603 /* Here, we have the current token in 'value' */
12605 /* What matches in a locale is not known until runtime. This includes
12606 * what the Posix classes (like \w, [:space:]) match. Room must be
12607 * reserved (one time per class) to store such classes, either if Perl
12608 * is compiled so that locale nodes always should have this space, or
12609 * if there is such class info to be stored. The space will contain a
12610 * bit for each named class that is to be matched against. This isn't
12611 * needed for \p{} and pseudo-classes, as they are not affected by
12612 * locale, and hence are dealt with separately */
12615 && (ANYOF_LOCALE == ANYOF_CLASS
12616 || (namedclass > OOB_NAMEDCLASS && namedclass < ANYOF_MAX)))
12620 RExC_size += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12623 RExC_emit += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12624 ANYOF_CLASS_ZERO(ret);
12626 ANYOF_FLAGS(ret) |= ANYOF_CLASS;
12629 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
12631 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
12632 * literal, as is the character that began the false range, i.e.
12633 * the 'a' in the examples */
12636 const int w = (RExC_parse >= rangebegin)
12637 ? RExC_parse - rangebegin
12640 vFAIL4("False [] range \"%*.*s\"", w, w, rangebegin);
12643 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
12644 ckWARN4reg(RExC_parse,
12645 "False [] range \"%*.*s\"",
12647 (void)ReREFCNT_inc(RExC_rx_sv);
12648 cp_list = add_cp_to_invlist(cp_list, '-');
12649 cp_list = add_cp_to_invlist(cp_list, prevvalue);
12653 range = 0; /* this was not a true range */
12654 element_count += 2; /* So counts for three values */
12658 U8 classnum = namedclass_to_classnum(namedclass);
12659 if (namedclass >= ANYOF_MAX) { /* If a special class */
12660 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
12662 /* Here, should be \h, \H, \v, or \V. Neither /d nor
12663 * /l make a difference in what these match. There
12664 * would be problems if these characters had folds
12665 * other than themselves, as cp_list is subject to
12667 if (classnum != _CC_VERTSPACE) {
12668 assert( namedclass == ANYOF_HORIZWS
12669 || namedclass == ANYOF_NHORIZWS);
12671 /* It turns out that \h is just a synonym for
12673 classnum = _CC_BLANK;
12676 _invlist_union_maybe_complement_2nd(
12678 PL_XPosix_ptrs[classnum],
12679 cBOOL(namedclass % 2), /* Complement if odd
12680 (NHORIZWS, NVERTWS)
12685 else if (classnum == _CC_ASCII) {
12688 ANYOF_CLASS_SET(ret, namedclass);
12691 #endif /* Not isascii(); just use the hard-coded definition for it */
12693 _invlist_union_maybe_complement_2nd(
12696 cBOOL(namedclass % 2), /* Complement if odd
12700 /* The code points 128-255 added above will be
12701 * subtracted out below under /d, so the flag needs to
12703 if (namedclass == ANYOF_NASCII && DEPENDS_SEMANTICS) {
12704 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL;
12708 else { /* Garden variety class */
12710 /* The ascii range inversion list */
12711 SV* ascii_source = PL_Posix_ptrs[classnum];
12713 /* The full Latin1 range inversion list */
12714 SV* l1_source = PL_L1Posix_ptrs[classnum];
12716 /* This code is structured into two major clauses. The
12717 * first is for classes whose complete definitions may not
12718 * already be known. It not, the Latin1 definition
12719 * (guaranteed to already known) is used plus code is
12720 * generated to load the rest at run-time (only if needed).
12721 * If the complete definition is known, it drops down to
12722 * the second clause, where the complete definition is
12725 if (classnum < _FIRST_NON_SWASH_CC) {
12727 /* Here, the class has a swash, which may or not
12728 * already be loaded */
12730 /* The name of the property to use to match the full
12731 * eXtended Unicode range swash for this character
12733 const char *Xname = swash_property_names[classnum];
12735 /* If returning the inversion list, we can't defer
12736 * getting this until runtime */
12737 if (ret_invlist && ! PL_utf8_swash_ptrs[classnum]) {
12738 PL_utf8_swash_ptrs[classnum] =
12739 _core_swash_init("utf8", Xname, &PL_sv_undef,
12742 NULL, /* No inversion list */
12743 NULL /* No flags */
12745 assert(PL_utf8_swash_ptrs[classnum]);
12747 if ( ! PL_utf8_swash_ptrs[classnum]) {
12748 if (namedclass % 2 == 0) { /* A non-complemented
12750 /* If not /a matching, there are code points we
12751 * don't know at compile time. Arrange for the
12752 * unknown matches to be loaded at run-time, if
12754 if (! AT_LEAST_ASCII_RESTRICTED) {
12755 Perl_sv_catpvf(aTHX_ listsv, "+utf8::%s\n",
12758 if (LOC) { /* Under locale, set run-time
12760 ANYOF_CLASS_SET(ret, namedclass);
12763 /* Add the current class's code points to
12764 * the running total */
12765 _invlist_union(posixes,
12766 (AT_LEAST_ASCII_RESTRICTED)
12772 else { /* A complemented class */
12773 if (AT_LEAST_ASCII_RESTRICTED) {
12774 /* Under /a should match everything above
12775 * ASCII, plus the complement of the set's
12777 _invlist_union_complement_2nd(posixes,
12782 /* Arrange for the unknown matches to be
12783 * loaded at run-time, if needed */
12784 Perl_sv_catpvf(aTHX_ listsv, "!utf8::%s\n",
12786 runtime_posix_matches_above_Unicode = TRUE;
12788 ANYOF_CLASS_SET(ret, namedclass);
12792 /* We want to match everything in
12793 * Latin1, except those things that
12794 * l1_source matches */
12795 SV* scratch_list = NULL;
12796 _invlist_subtract(PL_Latin1, l1_source,
12799 /* Add the list from this class to the
12802 posixes = scratch_list;
12805 _invlist_union(posixes,
12808 SvREFCNT_dec_NN(scratch_list);
12810 if (DEPENDS_SEMANTICS) {
12812 |= ANYOF_NON_UTF8_LATIN1_ALL;
12817 goto namedclass_done;
12820 /* Here, there is a swash loaded for the class. If no
12821 * inversion list for it yet, get it */
12822 if (! PL_XPosix_ptrs[classnum]) {
12823 PL_XPosix_ptrs[classnum]
12824 = _swash_to_invlist(PL_utf8_swash_ptrs[classnum]);
12828 /* Here there is an inversion list already loaded for the
12831 if (namedclass % 2 == 0) { /* A non-complemented class,
12832 like ANYOF_PUNCT */
12834 /* For non-locale, just add it to any existing list
12836 _invlist_union(posixes,
12837 (AT_LEAST_ASCII_RESTRICTED)
12839 : PL_XPosix_ptrs[classnum],
12842 else { /* Locale */
12843 SV* scratch_list = NULL;
12845 /* For above Latin1 code points, we use the full
12847 _invlist_intersection(PL_AboveLatin1,
12848 PL_XPosix_ptrs[classnum],
12850 /* And set the output to it, adding instead if
12851 * there already is an output. Checking if
12852 * 'posixes' is NULL first saves an extra clone.
12853 * Its reference count will be decremented at the
12854 * next union, etc, or if this is the only
12855 * instance, at the end of the routine */
12857 posixes = scratch_list;
12860 _invlist_union(posixes, scratch_list, &posixes);
12861 SvREFCNT_dec_NN(scratch_list);
12864 #ifndef HAS_ISBLANK
12865 if (namedclass != ANYOF_BLANK) {
12867 /* Set this class in the node for runtime
12869 ANYOF_CLASS_SET(ret, namedclass);
12870 #ifndef HAS_ISBLANK
12873 /* No isblank(), use the hard-coded ASCII-range
12874 * blanks, adding them to the running total. */
12876 _invlist_union(posixes, ascii_source, &posixes);
12881 else { /* A complemented class, like ANYOF_NPUNCT */
12883 _invlist_union_complement_2nd(
12885 (AT_LEAST_ASCII_RESTRICTED)
12887 : PL_XPosix_ptrs[classnum],
12889 /* Under /d, everything in the upper half of the
12890 * Latin1 range matches this complement */
12891 if (DEPENDS_SEMANTICS) {
12892 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL;
12895 else { /* Locale */
12896 SV* scratch_list = NULL;
12897 _invlist_subtract(PL_AboveLatin1,
12898 PL_XPosix_ptrs[classnum],
12901 posixes = scratch_list;
12904 _invlist_union(posixes, scratch_list, &posixes);
12905 SvREFCNT_dec_NN(scratch_list);
12907 #ifndef HAS_ISBLANK
12908 if (namedclass != ANYOF_NBLANK) {
12910 ANYOF_CLASS_SET(ret, namedclass);
12911 #ifndef HAS_ISBLANK
12914 /* Get the list of all code points in Latin1
12915 * that are not ASCII blanks, and add them to
12916 * the running total */
12917 _invlist_subtract(PL_Latin1, ascii_source,
12919 _invlist_union(posixes, scratch_list, &posixes);
12920 SvREFCNT_dec_NN(scratch_list);
12927 continue; /* Go get next character */
12929 } /* end of namedclass \blah */
12931 /* Here, we have a single value. If 'range' is set, it is the ending
12932 * of a range--check its validity. Later, we will handle each
12933 * individual code point in the range. If 'range' isn't set, this
12934 * could be the beginning of a range, so check for that by looking
12935 * ahead to see if the next real character to be processed is the range
12936 * indicator--the minus sign */
12939 RExC_parse = regpatws(pRExC_state, RExC_parse,
12940 FALSE /* means don't recognize comments */);
12944 if (prevvalue > value) /* b-a */ {
12945 const int w = RExC_parse - rangebegin;
12946 Simple_vFAIL4("Invalid [] range \"%*.*s\"", w, w, rangebegin);
12947 range = 0; /* not a valid range */
12951 prevvalue = value; /* save the beginning of the potential range */
12952 if (! stop_at_1 /* Can't be a range if parsing just one thing */
12953 && *RExC_parse == '-')
12955 char* next_char_ptr = RExC_parse + 1;
12956 if (skip_white) { /* Get the next real char after the '-' */
12957 next_char_ptr = regpatws(pRExC_state,
12959 FALSE); /* means don't recognize
12963 /* If the '-' is at the end of the class (just before the ']',
12964 * it is a literal minus; otherwise it is a range */
12965 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
12966 RExC_parse = next_char_ptr;
12968 /* a bad range like \w-, [:word:]- ? */
12969 if (namedclass > OOB_NAMEDCLASS) {
12970 if (strict || ckWARN(WARN_REGEXP)) {
12972 RExC_parse >= rangebegin ?
12973 RExC_parse - rangebegin : 0;
12975 vFAIL4("False [] range \"%*.*s\"",
12980 "False [] range \"%*.*s\"",
12985 cp_list = add_cp_to_invlist(cp_list, '-');
12989 range = 1; /* yeah, it's a range! */
12990 continue; /* but do it the next time */
12995 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
12998 /* non-Latin1 code point implies unicode semantics. Must be set in
12999 * pass1 so is there for the whole of pass 2 */
13001 RExC_uni_semantics = 1;
13004 /* Ready to process either the single value, or the completed range.
13005 * For single-valued non-inverted ranges, we consider the possibility
13006 * of multi-char folds. (We made a conscious decision to not do this
13007 * for the other cases because it can often lead to non-intuitive
13008 * results. For example, you have the peculiar case that:
13009 * "s s" =~ /^[^\xDF]+$/i => Y
13010 * "ss" =~ /^[^\xDF]+$/i => N
13012 * See [perl #89750] */
13013 if (FOLD && allow_multi_folds && value == prevvalue) {
13014 if (value == LATIN_SMALL_LETTER_SHARP_S
13015 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
13018 /* Here <value> is indeed a multi-char fold. Get what it is */
13020 U8 foldbuf[UTF8_MAXBYTES_CASE];
13023 UV folded = _to_uni_fold_flags(
13028 | ((LOC) ? FOLD_FLAGS_LOCALE
13029 : (ASCII_FOLD_RESTRICTED)
13030 ? FOLD_FLAGS_NOMIX_ASCII
13034 /* Here, <folded> should be the first character of the
13035 * multi-char fold of <value>, with <foldbuf> containing the
13036 * whole thing. But, if this fold is not allowed (because of
13037 * the flags), <fold> will be the same as <value>, and should
13038 * be processed like any other character, so skip the special
13040 if (folded != value) {
13042 /* Skip if we are recursed, currently parsing the class
13043 * again. Otherwise add this character to the list of
13044 * multi-char folds. */
13045 if (! RExC_in_multi_char_class) {
13046 AV** this_array_ptr;
13048 STRLEN cp_count = utf8_length(foldbuf,
13049 foldbuf + foldlen);
13050 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
13052 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
13055 if (! multi_char_matches) {
13056 multi_char_matches = newAV();
13059 /* <multi_char_matches> is actually an array of arrays.
13060 * There will be one or two top-level elements: [2],
13061 * and/or [3]. The [2] element is an array, each
13062 * element thereof is a character which folds to two
13063 * characters; likewise for [3]. (Unicode guarantees a
13064 * maximum of 3 characters in any fold.) When we
13065 * rewrite the character class below, we will do so
13066 * such that the longest folds are written first, so
13067 * that it prefers the longest matching strings first.
13068 * This is done even if it turns out that any
13069 * quantifier is non-greedy, out of programmer
13070 * laziness. Tom Christiansen has agreed that this is
13071 * ok. This makes the test for the ligature 'ffi' come
13072 * before the test for 'ff' */
13073 if (av_exists(multi_char_matches, cp_count)) {
13074 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13076 this_array = *this_array_ptr;
13079 this_array = newAV();
13080 av_store(multi_char_matches, cp_count,
13083 av_push(this_array, multi_fold);
13086 /* This element should not be processed further in this
13089 value = save_value;
13090 prevvalue = save_prevvalue;
13096 /* Deal with this element of the class */
13099 cp_list = _add_range_to_invlist(cp_list, prevvalue, value);
13101 SV* this_range = _new_invlist(1);
13102 _append_range_to_invlist(this_range, prevvalue, value);
13104 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
13105 * If this range was specified using something like 'i-j', we want
13106 * to include only the 'i' and the 'j', and not anything in
13107 * between, so exclude non-ASCII, non-alphabetics from it.
13108 * However, if the range was specified with something like
13109 * [\x89-\x91] or [\x89-j], all code points within it should be
13110 * included. literal_endpoint==2 means both ends of the range used
13111 * a literal character, not \x{foo} */
13112 if (literal_endpoint == 2
13113 && (prevvalue >= 'a' && value <= 'z')
13114 || (prevvalue >= 'A' && value <= 'Z'))
13116 _invlist_intersection(this_range, PL_Posix_ptrs[_CC_ALPHA],
13119 _invlist_union(cp_list, this_range, &cp_list);
13120 literal_endpoint = 0;
13124 range = 0; /* this range (if it was one) is done now */
13125 } /* End of loop through all the text within the brackets */
13127 /* If anything in the class expands to more than one character, we have to
13128 * deal with them by building up a substitute parse string, and recursively
13129 * calling reg() on it, instead of proceeding */
13130 if (multi_char_matches) {
13131 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
13134 char *save_end = RExC_end;
13135 char *save_parse = RExC_parse;
13136 bool first_time = TRUE; /* First multi-char occurrence doesn't get
13141 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
13142 because too confusing */
13144 sv_catpv(substitute_parse, "(?:");
13148 /* Look at the longest folds first */
13149 for (cp_count = av_len(multi_char_matches); cp_count > 0; cp_count--) {
13151 if (av_exists(multi_char_matches, cp_count)) {
13152 AV** this_array_ptr;
13155 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13157 while ((this_sequence = av_pop(*this_array_ptr)) !=
13160 if (! first_time) {
13161 sv_catpv(substitute_parse, "|");
13163 first_time = FALSE;
13165 sv_catpv(substitute_parse, SvPVX(this_sequence));
13170 /* If the character class contains anything else besides these
13171 * multi-character folds, have to include it in recursive parsing */
13172 if (element_count) {
13173 sv_catpv(substitute_parse, "|[");
13174 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
13175 sv_catpv(substitute_parse, "]");
13178 sv_catpv(substitute_parse, ")");
13181 /* This is a way to get the parse to skip forward a whole named
13182 * sequence instead of matching the 2nd character when it fails the
13184 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
13188 RExC_parse = SvPV(substitute_parse, len);
13189 RExC_end = RExC_parse + len;
13190 RExC_in_multi_char_class = 1;
13191 RExC_emit = (regnode *)orig_emit;
13193 ret = reg(pRExC_state, 1, ®_flags, depth+1);
13195 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
13197 RExC_parse = save_parse;
13198 RExC_end = save_end;
13199 RExC_in_multi_char_class = 0;
13200 SvREFCNT_dec_NN(multi_char_matches);
13204 /* If the character class contains only a single element, it may be
13205 * optimizable into another node type which is smaller and runs faster.
13206 * Check if this is the case for this class */
13207 if (element_count == 1 && ! ret_invlist) {
13211 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like \w or
13212 [:digit:] or \p{foo} */
13214 /* All named classes are mapped into POSIXish nodes, with its FLAG
13215 * argument giving which class it is */
13216 switch ((I32)namedclass) {
13217 case ANYOF_UNIPROP:
13220 /* These don't depend on the charset modifiers. They always
13221 * match under /u rules */
13222 case ANYOF_NHORIZWS:
13223 case ANYOF_HORIZWS:
13224 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
13227 case ANYOF_NVERTWS:
13232 /* The actual POSIXish node for all the rest depends on the
13233 * charset modifier. The ones in the first set depend only on
13234 * ASCII or, if available on this platform, locale */
13238 op = (LOC) ? POSIXL : POSIXA;
13249 /* under /a could be alpha */
13251 if (ASCII_RESTRICTED) {
13252 namedclass = ANYOF_ALPHA + (namedclass % 2);
13260 /* The rest have more possibilities depending on the charset.
13261 * We take advantage of the enum ordering of the charset
13262 * modifiers to get the exact node type, */
13264 op = POSIXD + get_regex_charset(RExC_flags);
13265 if (op > POSIXA) { /* /aa is same as /a */
13268 #ifndef HAS_ISBLANK
13270 && (namedclass == ANYOF_BLANK
13271 || namedclass == ANYOF_NBLANK))
13278 /* The odd numbered ones are the complements of the
13279 * next-lower even number one */
13280 if (namedclass % 2 == 1) {
13284 arg = namedclass_to_classnum(namedclass);
13288 else if (value == prevvalue) {
13290 /* Here, the class consists of just a single code point */
13293 if (! LOC && value == '\n') {
13294 op = REG_ANY; /* Optimize [^\n] */
13295 *flagp |= HASWIDTH|SIMPLE;
13299 else if (value < 256 || UTF) {
13301 /* Optimize a single value into an EXACTish node, but not if it
13302 * would require converting the pattern to UTF-8. */
13303 op = compute_EXACTish(pRExC_state);
13305 } /* Otherwise is a range */
13306 else if (! LOC) { /* locale could vary these */
13307 if (prevvalue == '0') {
13308 if (value == '9') {
13315 /* Here, we have changed <op> away from its initial value iff we found
13316 * an optimization */
13319 /* Throw away this ANYOF regnode, and emit the calculated one,
13320 * which should correspond to the beginning, not current, state of
13322 const char * cur_parse = RExC_parse;
13323 RExC_parse = (char *)orig_parse;
13327 /* To get locale nodes to not use the full ANYOF size would
13328 * require moving the code above that writes the portions
13329 * of it that aren't in other nodes to after this point.
13330 * e.g. ANYOF_CLASS_SET */
13331 RExC_size = orig_size;
13335 RExC_emit = (regnode *)orig_emit;
13336 if (PL_regkind[op] == POSIXD) {
13338 op += NPOSIXD - POSIXD;
13343 ret = reg_node(pRExC_state, op);
13345 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
13349 *flagp |= HASWIDTH|SIMPLE;
13351 else if (PL_regkind[op] == EXACT) {
13352 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13355 RExC_parse = (char *) cur_parse;
13357 SvREFCNT_dec(posixes);
13358 SvREFCNT_dec(cp_list);
13365 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
13367 /* If folding, we calculate all characters that could fold to or from the
13368 * ones already on the list */
13369 if (FOLD && cp_list) {
13370 UV start, end; /* End points of code point ranges */
13372 SV* fold_intersection = NULL;
13374 /* If the highest code point is within Latin1, we can use the
13375 * compiled-in Alphas list, and not have to go out to disk. This
13376 * yields two false positives, the masculine and feminine ordinal
13377 * indicators, which are weeded out below using the
13378 * IS_IN_SOME_FOLD_L1() macro */
13379 if (invlist_highest(cp_list) < 256) {
13380 _invlist_intersection(PL_L1Posix_ptrs[_CC_ALPHA], cp_list,
13381 &fold_intersection);
13385 /* Here, there are non-Latin1 code points, so we will have to go
13386 * fetch the list of all the characters that participate in folds
13388 if (! PL_utf8_foldable) {
13389 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13390 &PL_sv_undef, 1, 0);
13391 PL_utf8_foldable = _get_swash_invlist(swash);
13392 SvREFCNT_dec_NN(swash);
13395 /* This is a hash that for a particular fold gives all characters
13396 * that are involved in it */
13397 if (! PL_utf8_foldclosures) {
13399 /* If we were unable to find any folds, then we likely won't be
13400 * able to find the closures. So just create an empty list.
13401 * Folding will effectively be restricted to the non-Unicode
13402 * rules hard-coded into Perl. (This case happens legitimately
13403 * during compilation of Perl itself before the Unicode tables
13404 * are generated) */
13405 if (_invlist_len(PL_utf8_foldable) == 0) {
13406 PL_utf8_foldclosures = newHV();
13409 /* If the folds haven't been read in, call a fold function
13411 if (! PL_utf8_tofold) {
13412 U8 dummy[UTF8_MAXBYTES+1];
13414 /* This string is just a short named one above \xff */
13415 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
13416 assert(PL_utf8_tofold); /* Verify that worked */
13418 PL_utf8_foldclosures =
13419 _swash_inversion_hash(PL_utf8_tofold);
13423 /* Only the characters in this class that participate in folds need
13424 * be checked. Get the intersection of this class and all the
13425 * possible characters that are foldable. This can quickly narrow
13426 * down a large class */
13427 _invlist_intersection(PL_utf8_foldable, cp_list,
13428 &fold_intersection);
13431 /* Now look at the foldable characters in this class individually */
13432 invlist_iterinit(fold_intersection);
13433 while (invlist_iternext(fold_intersection, &start, &end)) {
13436 /* Locale folding for Latin1 characters is deferred until runtime */
13437 if (LOC && start < 256) {
13441 /* Look at every character in the range */
13442 for (j = start; j <= end; j++) {
13444 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
13450 /* We have the latin1 folding rules hard-coded here so that
13451 * an innocent-looking character class, like /[ks]/i won't
13452 * have to go out to disk to find the possible matches.
13453 * XXX It would be better to generate these via regen, in
13454 * case a new version of the Unicode standard adds new
13455 * mappings, though that is not really likely, and may be
13456 * caught by the default: case of the switch below. */
13458 if (IS_IN_SOME_FOLD_L1(j)) {
13460 /* ASCII is always matched; non-ASCII is matched only
13461 * under Unicode rules */
13462 if (isASCII(j) || AT_LEAST_UNI_SEMANTICS) {
13464 add_cp_to_invlist(cp_list, PL_fold_latin1[j]);
13468 add_cp_to_invlist(depends_list, PL_fold_latin1[j]);
13472 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
13473 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
13475 /* Certain Latin1 characters have matches outside
13476 * Latin1. To get here, <j> is one of those
13477 * characters. None of these matches is valid for
13478 * ASCII characters under /aa, which is why the 'if'
13479 * just above excludes those. These matches only
13480 * happen when the target string is utf8. The code
13481 * below adds the single fold closures for <j> to the
13482 * inversion list. */
13487 add_cp_to_invlist(cp_list, KELVIN_SIGN);
13491 cp_list = add_cp_to_invlist(cp_list,
13492 LATIN_SMALL_LETTER_LONG_S);
13495 cp_list = add_cp_to_invlist(cp_list,
13496 GREEK_CAPITAL_LETTER_MU);
13497 cp_list = add_cp_to_invlist(cp_list,
13498 GREEK_SMALL_LETTER_MU);
13500 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
13501 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
13503 add_cp_to_invlist(cp_list, ANGSTROM_SIGN);
13505 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
13506 cp_list = add_cp_to_invlist(cp_list,
13507 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
13509 case LATIN_SMALL_LETTER_SHARP_S:
13510 cp_list = add_cp_to_invlist(cp_list,
13511 LATIN_CAPITAL_LETTER_SHARP_S);
13513 case 'F': case 'f':
13514 case 'I': case 'i':
13515 case 'L': case 'l':
13516 case 'T': case 't':
13517 case 'A': case 'a':
13518 case 'H': case 'h':
13519 case 'J': case 'j':
13520 case 'N': case 'n':
13521 case 'W': case 'w':
13522 case 'Y': case 'y':
13523 /* These all are targets of multi-character
13524 * folds from code points that require UTF8 to
13525 * express, so they can't match unless the
13526 * target string is in UTF-8, so no action here
13527 * is necessary, as regexec.c properly handles
13528 * the general case for UTF-8 matching and
13529 * multi-char folds */
13532 /* Use deprecated warning to increase the
13533 * chances of this being output */
13534 ckWARN2regdep(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
13541 /* Here is an above Latin1 character. We don't have the rules
13542 * hard-coded for it. First, get its fold. This is the simple
13543 * fold, as the multi-character folds have been handled earlier
13544 * and separated out */
13545 _to_uni_fold_flags(j, foldbuf, &foldlen,
13547 ? FOLD_FLAGS_LOCALE
13548 : (ASCII_FOLD_RESTRICTED)
13549 ? FOLD_FLAGS_NOMIX_ASCII
13552 /* Single character fold of above Latin1. Add everything in
13553 * its fold closure to the list that this node should match.
13554 * The fold closures data structure is a hash with the keys
13555 * being the UTF-8 of every character that is folded to, like
13556 * 'k', and the values each an array of all code points that
13557 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
13558 * Multi-character folds are not included */
13559 if ((listp = hv_fetch(PL_utf8_foldclosures,
13560 (char *) foldbuf, foldlen, FALSE)))
13562 AV* list = (AV*) *listp;
13564 for (k = 0; k <= av_len(list); k++) {
13565 SV** c_p = av_fetch(list, k, FALSE);
13568 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
13572 /* /aa doesn't allow folds between ASCII and non-; /l
13573 * doesn't allow them between above and below 256 */
13574 if ((ASCII_FOLD_RESTRICTED
13575 && (isASCII(c) != isASCII(j)))
13576 || (LOC && ((c < 256) != (j < 256))))
13581 /* Folds involving non-ascii Latin1 characters
13582 * under /d are added to a separate list */
13583 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
13585 cp_list = add_cp_to_invlist(cp_list, c);
13588 depends_list = add_cp_to_invlist(depends_list, c);
13594 SvREFCNT_dec_NN(fold_intersection);
13597 /* And combine the result (if any) with any inversion list from posix
13598 * classes. The lists are kept separate up to now because we don't want to
13599 * fold the classes (folding of those is automatically handled by the swash
13600 * fetching code) */
13602 if (! DEPENDS_SEMANTICS) {
13604 _invlist_union(cp_list, posixes, &cp_list);
13605 SvREFCNT_dec_NN(posixes);
13612 /* Under /d, we put into a separate list the Latin1 things that
13613 * match only when the target string is utf8 */
13614 SV* nonascii_but_latin1_properties = NULL;
13615 _invlist_intersection(posixes, PL_Latin1,
13616 &nonascii_but_latin1_properties);
13617 _invlist_subtract(nonascii_but_latin1_properties, PL_ASCII,
13618 &nonascii_but_latin1_properties);
13619 _invlist_subtract(posixes, nonascii_but_latin1_properties,
13622 _invlist_union(cp_list, posixes, &cp_list);
13623 SvREFCNT_dec_NN(posixes);
13629 if (depends_list) {
13630 _invlist_union(depends_list, nonascii_but_latin1_properties,
13632 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
13635 depends_list = nonascii_but_latin1_properties;
13640 /* And combine the result (if any) with any inversion list from properties.
13641 * The lists are kept separate up to now so that we can distinguish the two
13642 * in regards to matching above-Unicode. A run-time warning is generated
13643 * if a Unicode property is matched against a non-Unicode code point. But,
13644 * we allow user-defined properties to match anything, without any warning,
13645 * and we also suppress the warning if there is a portion of the character
13646 * class that isn't a Unicode property, and which matches above Unicode, \W
13647 * or [\x{110000}] for example.
13648 * (Note that in this case, unlike the Posix one above, there is no
13649 * <depends_list>, because having a Unicode property forces Unicode
13652 bool warn_super = ! has_user_defined_property;
13655 /* If it matters to the final outcome, see if a non-property
13656 * component of the class matches above Unicode. If so, the
13657 * warning gets suppressed. This is true even if just a single
13658 * such code point is specified, as though not strictly correct if
13659 * another such code point is matched against, the fact that they
13660 * are using above-Unicode code points indicates they should know
13661 * the issues involved */
13663 bool non_prop_matches_above_Unicode =
13664 runtime_posix_matches_above_Unicode
13665 | (invlist_highest(cp_list) > PERL_UNICODE_MAX);
13667 non_prop_matches_above_Unicode =
13668 ! non_prop_matches_above_Unicode;
13670 warn_super = ! non_prop_matches_above_Unicode;
13673 _invlist_union(properties, cp_list, &cp_list);
13674 SvREFCNT_dec_NN(properties);
13677 cp_list = properties;
13681 OP(ret) = ANYOF_WARN_SUPER;
13685 /* Here, we have calculated what code points should be in the character
13688 * Now we can see about various optimizations. Fold calculation (which we
13689 * did above) needs to take place before inversion. Otherwise /[^k]/i
13690 * would invert to include K, which under /i would match k, which it
13691 * shouldn't. Therefore we can't invert folded locale now, as it won't be
13692 * folded until runtime */
13694 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
13695 * at compile time. Besides not inverting folded locale now, we can't
13696 * invert if there are things such as \w, which aren't known until runtime
13699 && ! (LOC && (FOLD || (ANYOF_FLAGS(ret) & ANYOF_CLASS)))
13701 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13703 _invlist_invert(cp_list);
13705 /* Any swash can't be used as-is, because we've inverted things */
13707 SvREFCNT_dec_NN(swash);
13711 /* Clear the invert flag since have just done it here */
13716 *ret_invlist = cp_list;
13717 SvREFCNT_dec(swash);
13719 /* Discard the generated node */
13721 RExC_size = orig_size;
13724 RExC_emit = orig_emit;
13729 /* If we didn't do folding, it's because some information isn't available
13730 * until runtime; set the run-time fold flag for these. (We don't have to
13731 * worry about properties folding, as that is taken care of by the swash
13735 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
13738 /* Some character classes are equivalent to other nodes. Such nodes take
13739 * up less room and generally fewer operations to execute than ANYOF nodes.
13740 * Above, we checked for and optimized into some such equivalents for
13741 * certain common classes that are easy to test. Getting to this point in
13742 * the code means that the class didn't get optimized there. Since this
13743 * code is only executed in Pass 2, it is too late to save space--it has
13744 * been allocated in Pass 1, and currently isn't given back. But turning
13745 * things into an EXACTish node can allow the optimizer to join it to any
13746 * adjacent such nodes. And if the class is equivalent to things like /./,
13747 * expensive run-time swashes can be avoided. Now that we have more
13748 * complete information, we can find things necessarily missed by the
13749 * earlier code. I (khw) am not sure how much to look for here. It would
13750 * be easy, but perhaps too slow, to check any candidates against all the
13751 * node types they could possibly match using _invlistEQ(). */
13756 && ! (ANYOF_FLAGS(ret) & ANYOF_CLASS)
13757 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13760 U8 op = END; /* The optimzation node-type */
13761 const char * cur_parse= RExC_parse;
13763 invlist_iterinit(cp_list);
13764 if (! invlist_iternext(cp_list, &start, &end)) {
13766 /* Here, the list is empty. This happens, for example, when a
13767 * Unicode property is the only thing in the character class, and
13768 * it doesn't match anything. (perluniprops.pod notes such
13771 *flagp |= HASWIDTH|SIMPLE;
13773 else if (start == end) { /* The range is a single code point */
13774 if (! invlist_iternext(cp_list, &start, &end)
13776 /* Don't do this optimization if it would require changing
13777 * the pattern to UTF-8 */
13778 && (start < 256 || UTF))
13780 /* Here, the list contains a single code point. Can optimize
13781 * into an EXACT node */
13790 /* A locale node under folding with one code point can be
13791 * an EXACTFL, as its fold won't be calculated until
13797 /* Here, we are generally folding, but there is only one
13798 * code point to match. If we have to, we use an EXACT
13799 * node, but it would be better for joining with adjacent
13800 * nodes in the optimization pass if we used the same
13801 * EXACTFish node that any such are likely to be. We can
13802 * do this iff the code point doesn't participate in any
13803 * folds. For example, an EXACTF of a colon is the same as
13804 * an EXACT one, since nothing folds to or from a colon. */
13806 if (IS_IN_SOME_FOLD_L1(value)) {
13811 if (! PL_utf8_foldable) {
13812 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13813 &PL_sv_undef, 1, 0);
13814 PL_utf8_foldable = _get_swash_invlist(swash);
13815 SvREFCNT_dec_NN(swash);
13817 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
13822 /* If we haven't found the node type, above, it means we
13823 * can use the prevailing one */
13825 op = compute_EXACTish(pRExC_state);
13830 else if (start == 0) {
13831 if (end == UV_MAX) {
13833 *flagp |= HASWIDTH|SIMPLE;
13836 else if (end == '\n' - 1
13837 && invlist_iternext(cp_list, &start, &end)
13838 && start == '\n' + 1 && end == UV_MAX)
13841 *flagp |= HASWIDTH|SIMPLE;
13845 invlist_iterfinish(cp_list);
13848 RExC_parse = (char *)orig_parse;
13849 RExC_emit = (regnode *)orig_emit;
13851 ret = reg_node(pRExC_state, op);
13853 RExC_parse = (char *)cur_parse;
13855 if (PL_regkind[op] == EXACT) {
13856 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13859 SvREFCNT_dec_NN(cp_list);
13864 /* Here, <cp_list> contains all the code points we can determine at
13865 * compile time that match under all conditions. Go through it, and
13866 * for things that belong in the bitmap, put them there, and delete from
13867 * <cp_list>. While we are at it, see if everything above 255 is in the
13868 * list, and if so, set a flag to speed up execution */
13869 ANYOF_BITMAP_ZERO(ret);
13872 /* This gets set if we actually need to modify things */
13873 bool change_invlist = FALSE;
13877 /* Start looking through <cp_list> */
13878 invlist_iterinit(cp_list);
13879 while (invlist_iternext(cp_list, &start, &end)) {
13883 if (end == UV_MAX && start <= 256) {
13884 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
13887 /* Quit if are above what we should change */
13892 change_invlist = TRUE;
13894 /* Set all the bits in the range, up to the max that we are doing */
13895 high = (end < 255) ? end : 255;
13896 for (i = start; i <= (int) high; i++) {
13897 if (! ANYOF_BITMAP_TEST(ret, i)) {
13898 ANYOF_BITMAP_SET(ret, i);
13904 invlist_iterfinish(cp_list);
13906 /* Done with loop; remove any code points that are in the bitmap from
13908 if (change_invlist) {
13909 _invlist_subtract(cp_list, PL_Latin1, &cp_list);
13912 /* If have completely emptied it, remove it completely */
13913 if (_invlist_len(cp_list) == 0) {
13914 SvREFCNT_dec_NN(cp_list);
13920 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
13923 /* Here, the bitmap has been populated with all the Latin1 code points that
13924 * always match. Can now add to the overall list those that match only
13925 * when the target string is UTF-8 (<depends_list>). */
13926 if (depends_list) {
13928 _invlist_union(cp_list, depends_list, &cp_list);
13929 SvREFCNT_dec_NN(depends_list);
13932 cp_list = depends_list;
13936 /* If there is a swash and more than one element, we can't use the swash in
13937 * the optimization below. */
13938 if (swash && element_count > 1) {
13939 SvREFCNT_dec_NN(swash);
13944 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13946 ARG_SET(ret, ANYOF_NONBITMAP_EMPTY);
13949 /* av[0] stores the character class description in its textual form:
13950 * used later (regexec.c:Perl_regclass_swash()) to initialize the
13951 * appropriate swash, and is also useful for dumping the regnode.
13952 * av[1] if NULL, is a placeholder to later contain the swash computed
13953 * from av[0]. But if no further computation need be done, the
13954 * swash is stored there now.
13955 * av[2] stores the cp_list inversion list for use in addition or
13956 * instead of av[0]; used only if av[1] is NULL
13957 * av[3] is set if any component of the class is from a user-defined
13958 * property; used only if av[1] is NULL */
13959 AV * const av = newAV();
13962 av_store(av, 0, (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13963 ? SvREFCNT_inc(listsv) : &PL_sv_undef);
13965 av_store(av, 1, swash);
13966 SvREFCNT_dec_NN(cp_list);
13969 av_store(av, 1, NULL);
13971 av_store(av, 2, cp_list);
13972 av_store(av, 3, newSVuv(has_user_defined_property));
13976 rv = newRV_noinc(MUTABLE_SV(av));
13977 n = add_data(pRExC_state, 1, "s");
13978 RExC_rxi->data->data[n] = (void*)rv;
13982 *flagp |= HASWIDTH|SIMPLE;
13985 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
13988 /* reg_skipcomment()
13990 Absorbs an /x style # comments from the input stream.
13991 Returns true if there is more text remaining in the stream.
13992 Will set the REG_SEEN_RUN_ON_COMMENT flag if the comment
13993 terminates the pattern without including a newline.
13995 Note its the callers responsibility to ensure that we are
13996 actually in /x mode
14001 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
14005 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
14007 while (RExC_parse < RExC_end)
14008 if (*RExC_parse++ == '\n') {
14013 /* we ran off the end of the pattern without ending
14014 the comment, so we have to add an \n when wrapping */
14015 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
14023 Advances the parse position, and optionally absorbs
14024 "whitespace" from the inputstream.
14026 Without /x "whitespace" means (?#...) style comments only,
14027 with /x this means (?#...) and # comments and whitespace proper.
14029 Returns the RExC_parse point from BEFORE the scan occurs.
14031 This is the /x friendly way of saying RExC_parse++.
14035 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
14037 char* const retval = RExC_parse++;
14039 PERL_ARGS_ASSERT_NEXTCHAR;
14042 if (RExC_end - RExC_parse >= 3
14043 && *RExC_parse == '('
14044 && RExC_parse[1] == '?'
14045 && RExC_parse[2] == '#')
14047 while (*RExC_parse != ')') {
14048 if (RExC_parse == RExC_end)
14049 FAIL("Sequence (?#... not terminated");
14055 if (RExC_flags & RXf_PMf_EXTENDED) {
14056 if (isSPACE(*RExC_parse)) {
14060 else if (*RExC_parse == '#') {
14061 if ( reg_skipcomment( pRExC_state ) )
14070 - reg_node - emit a node
14072 STATIC regnode * /* Location. */
14073 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
14077 regnode * const ret = RExC_emit;
14078 GET_RE_DEBUG_FLAGS_DECL;
14080 PERL_ARGS_ASSERT_REG_NODE;
14083 SIZE_ALIGN(RExC_size);
14087 if (RExC_emit >= RExC_emit_bound)
14088 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14089 op, RExC_emit, RExC_emit_bound);
14091 NODE_ALIGN_FILL(ret);
14093 FILL_ADVANCE_NODE(ptr, op);
14094 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (ptr) - 1);
14095 #ifdef RE_TRACK_PATTERN_OFFSETS
14096 if (RExC_offsets) { /* MJD */
14097 MJD_OFFSET_DEBUG(("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
14098 "reg_node", __LINE__,
14100 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
14101 ? "Overwriting end of array!\n" : "OK",
14102 (UV)(RExC_emit - RExC_emit_start),
14103 (UV)(RExC_parse - RExC_start),
14104 (UV)RExC_offsets[0]));
14105 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
14113 - reganode - emit a node with an argument
14115 STATIC regnode * /* Location. */
14116 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
14120 regnode * const ret = RExC_emit;
14121 GET_RE_DEBUG_FLAGS_DECL;
14123 PERL_ARGS_ASSERT_REGANODE;
14126 SIZE_ALIGN(RExC_size);
14131 assert(2==regarglen[op]+1);
14133 Anything larger than this has to allocate the extra amount.
14134 If we changed this to be:
14136 RExC_size += (1 + regarglen[op]);
14138 then it wouldn't matter. Its not clear what side effect
14139 might come from that so its not done so far.
14144 if (RExC_emit >= RExC_emit_bound)
14145 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14146 op, RExC_emit, RExC_emit_bound);
14148 NODE_ALIGN_FILL(ret);
14150 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
14151 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (ptr) - 2);
14152 #ifdef RE_TRACK_PATTERN_OFFSETS
14153 if (RExC_offsets) { /* MJD */
14154 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14158 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
14159 "Overwriting end of array!\n" : "OK",
14160 (UV)(RExC_emit - RExC_emit_start),
14161 (UV)(RExC_parse - RExC_start),
14162 (UV)RExC_offsets[0]));
14163 Set_Cur_Node_Offset;
14171 - reguni - emit (if appropriate) a Unicode character
14174 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
14178 PERL_ARGS_ASSERT_REGUNI;
14180 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
14184 - reginsert - insert an operator in front of already-emitted operand
14186 * Means relocating the operand.
14189 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
14195 const int offset = regarglen[(U8)op];
14196 const int size = NODE_STEP_REGNODE + offset;
14197 GET_RE_DEBUG_FLAGS_DECL;
14199 PERL_ARGS_ASSERT_REGINSERT;
14200 PERL_UNUSED_ARG(depth);
14201 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
14202 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
14211 if (RExC_open_parens) {
14213 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
14214 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
14215 if ( RExC_open_parens[paren] >= opnd ) {
14216 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
14217 RExC_open_parens[paren] += size;
14219 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
14221 if ( RExC_close_parens[paren] >= opnd ) {
14222 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
14223 RExC_close_parens[paren] += size;
14225 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
14230 while (src > opnd) {
14231 StructCopy(--src, --dst, regnode);
14232 #ifdef RE_TRACK_PATTERN_OFFSETS
14233 if (RExC_offsets) { /* MJD 20010112 */
14234 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
14238 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
14239 ? "Overwriting end of array!\n" : "OK",
14240 (UV)(src - RExC_emit_start),
14241 (UV)(dst - RExC_emit_start),
14242 (UV)RExC_offsets[0]));
14243 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
14244 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
14250 place = opnd; /* Op node, where operand used to be. */
14251 #ifdef RE_TRACK_PATTERN_OFFSETS
14252 if (RExC_offsets) { /* MJD */
14253 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14257 (UV)(place - RExC_emit_start) > RExC_offsets[0]
14258 ? "Overwriting end of array!\n" : "OK",
14259 (UV)(place - RExC_emit_start),
14260 (UV)(RExC_parse - RExC_start),
14261 (UV)RExC_offsets[0]));
14262 Set_Node_Offset(place, RExC_parse);
14263 Set_Node_Length(place, 1);
14266 src = NEXTOPER(place);
14267 FILL_ADVANCE_NODE(place, op);
14268 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (place) - 1);
14269 Zero(src, offset, regnode);
14273 - regtail - set the next-pointer at the end of a node chain of p to val.
14274 - SEE ALSO: regtail_study
14276 /* TODO: All three parms should be const */
14278 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14282 GET_RE_DEBUG_FLAGS_DECL;
14284 PERL_ARGS_ASSERT_REGTAIL;
14286 PERL_UNUSED_ARG(depth);
14292 /* Find last node. */
14295 regnode * const temp = regnext(scan);
14297 SV * const mysv=sv_newmortal();
14298 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
14299 regprop(RExC_rx, mysv, scan);
14300 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
14301 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
14302 (temp == NULL ? "->" : ""),
14303 (temp == NULL ? PL_reg_name[OP(val)] : "")
14311 if (reg_off_by_arg[OP(scan)]) {
14312 ARG_SET(scan, val - scan);
14315 NEXT_OFF(scan) = val - scan;
14321 - regtail_study - set the next-pointer at the end of a node chain of p to val.
14322 - Look for optimizable sequences at the same time.
14323 - currently only looks for EXACT chains.
14325 This is experimental code. The idea is to use this routine to perform
14326 in place optimizations on branches and groups as they are constructed,
14327 with the long term intention of removing optimization from study_chunk so
14328 that it is purely analytical.
14330 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
14331 to control which is which.
14334 /* TODO: All four parms should be const */
14337 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14342 #ifdef EXPERIMENTAL_INPLACESCAN
14345 GET_RE_DEBUG_FLAGS_DECL;
14347 PERL_ARGS_ASSERT_REGTAIL_STUDY;
14353 /* Find last node. */
14357 regnode * const temp = regnext(scan);
14358 #ifdef EXPERIMENTAL_INPLACESCAN
14359 if (PL_regkind[OP(scan)] == EXACT) {
14360 bool has_exactf_sharp_s; /* Unexamined in this routine */
14361 if (join_exact(pRExC_state,scan,&min, &has_exactf_sharp_s, 1,val,depth+1))
14366 switch (OP(scan)) {
14372 case EXACTFU_TRICKYFOLD:
14374 if( exact == PSEUDO )
14376 else if ( exact != OP(scan) )
14385 SV * const mysv=sv_newmortal();
14386 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
14387 regprop(RExC_rx, mysv, scan);
14388 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
14389 SvPV_nolen_const(mysv),
14390 REG_NODE_NUM(scan),
14391 PL_reg_name[exact]);
14398 SV * const mysv_val=sv_newmortal();
14399 DEBUG_PARSE_MSG("");
14400 regprop(RExC_rx, mysv_val, val);
14401 PerlIO_printf(Perl_debug_log, "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
14402 SvPV_nolen_const(mysv_val),
14403 (IV)REG_NODE_NUM(val),
14407 if (reg_off_by_arg[OP(scan)]) {
14408 ARG_SET(scan, val - scan);
14411 NEXT_OFF(scan) = val - scan;
14419 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
14423 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
14429 for (bit=0; bit<32; bit++) {
14430 if (flags & (1<<bit)) {
14431 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
14434 if (!set++ && lead)
14435 PerlIO_printf(Perl_debug_log, "%s",lead);
14436 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
14439 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
14440 if (!set++ && lead) {
14441 PerlIO_printf(Perl_debug_log, "%s",lead);
14444 case REGEX_UNICODE_CHARSET:
14445 PerlIO_printf(Perl_debug_log, "UNICODE");
14447 case REGEX_LOCALE_CHARSET:
14448 PerlIO_printf(Perl_debug_log, "LOCALE");
14450 case REGEX_ASCII_RESTRICTED_CHARSET:
14451 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
14453 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
14454 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
14457 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
14463 PerlIO_printf(Perl_debug_log, "\n");
14465 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14471 Perl_regdump(pTHX_ const regexp *r)
14475 SV * const sv = sv_newmortal();
14476 SV *dsv= sv_newmortal();
14477 RXi_GET_DECL(r,ri);
14478 GET_RE_DEBUG_FLAGS_DECL;
14480 PERL_ARGS_ASSERT_REGDUMP;
14482 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
14484 /* Header fields of interest. */
14485 if (r->anchored_substr) {
14486 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
14487 RE_SV_DUMPLEN(r->anchored_substr), 30);
14488 PerlIO_printf(Perl_debug_log,
14489 "anchored %s%s at %"IVdf" ",
14490 s, RE_SV_TAIL(r->anchored_substr),
14491 (IV)r->anchored_offset);
14492 } else if (r->anchored_utf8) {
14493 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
14494 RE_SV_DUMPLEN(r->anchored_utf8), 30);
14495 PerlIO_printf(Perl_debug_log,
14496 "anchored utf8 %s%s at %"IVdf" ",
14497 s, RE_SV_TAIL(r->anchored_utf8),
14498 (IV)r->anchored_offset);
14500 if (r->float_substr) {
14501 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
14502 RE_SV_DUMPLEN(r->float_substr), 30);
14503 PerlIO_printf(Perl_debug_log,
14504 "floating %s%s at %"IVdf"..%"UVuf" ",
14505 s, RE_SV_TAIL(r->float_substr),
14506 (IV)r->float_min_offset, (UV)r->float_max_offset);
14507 } else if (r->float_utf8) {
14508 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
14509 RE_SV_DUMPLEN(r->float_utf8), 30);
14510 PerlIO_printf(Perl_debug_log,
14511 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
14512 s, RE_SV_TAIL(r->float_utf8),
14513 (IV)r->float_min_offset, (UV)r->float_max_offset);
14515 if (r->check_substr || r->check_utf8)
14516 PerlIO_printf(Perl_debug_log,
14518 (r->check_substr == r->float_substr
14519 && r->check_utf8 == r->float_utf8
14520 ? "(checking floating" : "(checking anchored"));
14521 if (r->extflags & RXf_NOSCAN)
14522 PerlIO_printf(Perl_debug_log, " noscan");
14523 if (r->extflags & RXf_CHECK_ALL)
14524 PerlIO_printf(Perl_debug_log, " isall");
14525 if (r->check_substr || r->check_utf8)
14526 PerlIO_printf(Perl_debug_log, ") ");
14528 if (ri->regstclass) {
14529 regprop(r, sv, ri->regstclass);
14530 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
14532 if (r->extflags & RXf_ANCH) {
14533 PerlIO_printf(Perl_debug_log, "anchored");
14534 if (r->extflags & RXf_ANCH_BOL)
14535 PerlIO_printf(Perl_debug_log, "(BOL)");
14536 if (r->extflags & RXf_ANCH_MBOL)
14537 PerlIO_printf(Perl_debug_log, "(MBOL)");
14538 if (r->extflags & RXf_ANCH_SBOL)
14539 PerlIO_printf(Perl_debug_log, "(SBOL)");
14540 if (r->extflags & RXf_ANCH_GPOS)
14541 PerlIO_printf(Perl_debug_log, "(GPOS)");
14542 PerlIO_putc(Perl_debug_log, ' ');
14544 if (r->extflags & RXf_GPOS_SEEN)
14545 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
14546 if (r->intflags & PREGf_SKIP)
14547 PerlIO_printf(Perl_debug_log, "plus ");
14548 if (r->intflags & PREGf_IMPLICIT)
14549 PerlIO_printf(Perl_debug_log, "implicit ");
14550 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
14551 if (r->extflags & RXf_EVAL_SEEN)
14552 PerlIO_printf(Perl_debug_log, "with eval ");
14553 PerlIO_printf(Perl_debug_log, "\n");
14554 DEBUG_FLAGS_r(regdump_extflags("r->extflags: ",r->extflags));
14556 PERL_ARGS_ASSERT_REGDUMP;
14557 PERL_UNUSED_CONTEXT;
14558 PERL_UNUSED_ARG(r);
14559 #endif /* DEBUGGING */
14563 - regprop - printable representation of opcode
14565 #define EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags) \
14568 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]); \
14569 if (flags & ANYOF_INVERT) \
14570 /*make sure the invert info is in each */ \
14571 sv_catpvs(sv, "^"); \
14577 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
14583 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
14584 static const char * const anyofs[] = {
14585 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
14586 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
14587 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
14588 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
14589 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
14590 || _CC_VERTSPACE != 16
14591 #error Need to adjust order of anyofs[]
14628 RXi_GET_DECL(prog,progi);
14629 GET_RE_DEBUG_FLAGS_DECL;
14631 PERL_ARGS_ASSERT_REGPROP;
14635 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
14636 /* It would be nice to FAIL() here, but this may be called from
14637 regexec.c, and it would be hard to supply pRExC_state. */
14638 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(o), (int)REGNODE_MAX);
14639 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
14641 k = PL_regkind[OP(o)];
14644 sv_catpvs(sv, " ");
14645 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
14646 * is a crude hack but it may be the best for now since
14647 * we have no flag "this EXACTish node was UTF-8"
14649 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
14650 PERL_PV_ESCAPE_UNI_DETECT |
14651 PERL_PV_ESCAPE_NONASCII |
14652 PERL_PV_PRETTY_ELLIPSES |
14653 PERL_PV_PRETTY_LTGT |
14654 PERL_PV_PRETTY_NOCLEAR
14656 } else if (k == TRIE) {
14657 /* print the details of the trie in dumpuntil instead, as
14658 * progi->data isn't available here */
14659 const char op = OP(o);
14660 const U32 n = ARG(o);
14661 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
14662 (reg_ac_data *)progi->data->data[n] :
14664 const reg_trie_data * const trie
14665 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
14667 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
14668 DEBUG_TRIE_COMPILE_r(
14669 Perl_sv_catpvf(aTHX_ sv,
14670 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
14671 (UV)trie->startstate,
14672 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
14673 (UV)trie->wordcount,
14676 (UV)TRIE_CHARCOUNT(trie),
14677 (UV)trie->uniquecharcount
14680 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
14682 int rangestart = -1;
14683 U8* bitmap = IS_ANYOF_TRIE(op) ? (U8*)ANYOF_BITMAP(o) : (U8*)TRIE_BITMAP(trie);
14684 sv_catpvs(sv, "[");
14685 for (i = 0; i <= 256; i++) {
14686 if (i < 256 && BITMAP_TEST(bitmap,i)) {
14687 if (rangestart == -1)
14689 } else if (rangestart != -1) {
14690 if (i <= rangestart + 3)
14691 for (; rangestart < i; rangestart++)
14692 put_byte(sv, rangestart);
14694 put_byte(sv, rangestart);
14695 sv_catpvs(sv, "-");
14696 put_byte(sv, i - 1);
14701 sv_catpvs(sv, "]");
14704 } else if (k == CURLY) {
14705 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
14706 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
14707 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
14709 else if (k == WHILEM && o->flags) /* Ordinal/of */
14710 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
14711 else if (k == REF || k == OPEN || k == CLOSE || k == GROUPP || OP(o)==ACCEPT) {
14712 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
14713 if ( RXp_PAREN_NAMES(prog) ) {
14714 if ( k != REF || (OP(o) < NREF)) {
14715 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
14716 SV **name= av_fetch(list, ARG(o), 0 );
14718 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14721 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
14722 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
14723 I32 *nums=(I32*)SvPVX(sv_dat);
14724 SV **name= av_fetch(list, nums[0], 0 );
14727 for ( n=0; n<SvIVX(sv_dat); n++ ) {
14728 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
14729 (n ? "," : ""), (IV)nums[n]);
14731 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14735 } else if (k == GOSUB)
14736 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o)); /* Paren and offset */
14737 else if (k == VERB) {
14739 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
14740 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
14741 } else if (k == LOGICAL)
14742 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* 2: embedded, otherwise 1 */
14743 else if (k == ANYOF) {
14744 int i, rangestart = -1;
14745 const U8 flags = ANYOF_FLAGS(o);
14749 if (flags & ANYOF_LOCALE)
14750 sv_catpvs(sv, "{loc}");
14751 if (flags & ANYOF_LOC_FOLD)
14752 sv_catpvs(sv, "{i}");
14753 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
14754 if (flags & ANYOF_INVERT)
14755 sv_catpvs(sv, "^");
14757 /* output what the standard cp 0-255 bitmap matches */
14758 for (i = 0; i <= 256; i++) {
14759 if (i < 256 && ANYOF_BITMAP_TEST(o,i)) {
14760 if (rangestart == -1)
14762 } else if (rangestart != -1) {
14763 if (i <= rangestart + 3)
14764 for (; rangestart < i; rangestart++)
14765 put_byte(sv, rangestart);
14767 put_byte(sv, rangestart);
14768 sv_catpvs(sv, "-");
14769 put_byte(sv, i - 1);
14776 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14777 /* output any special charclass tests (used entirely under use locale) */
14778 if (ANYOF_CLASS_TEST_ANY_SET(o))
14779 for (i = 0; i < (int)(sizeof(anyofs)/sizeof(char*)); i++)
14780 if (ANYOF_CLASS_TEST(o,i)) {
14781 sv_catpv(sv, anyofs[i]);
14785 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14787 if (flags & ANYOF_NON_UTF8_LATIN1_ALL) {
14788 sv_catpvs(sv, "{non-utf8-latin1-all}");
14791 /* output information about the unicode matching */
14792 if (flags & ANYOF_UNICODE_ALL)
14793 sv_catpvs(sv, "{unicode_all}");
14794 else if (ANYOF_NONBITMAP(o))
14795 sv_catpvs(sv, "{unicode}");
14796 if (flags & ANYOF_NONBITMAP_NON_UTF8)
14797 sv_catpvs(sv, "{outside bitmap}");
14799 if (ANYOF_NONBITMAP(o)) {
14800 SV *lv; /* Set if there is something outside the bit map */
14801 SV * const sw = regclass_swash(prog, o, FALSE, &lv, NULL);
14802 bool byte_output = FALSE; /* If something in the bitmap has been
14805 if (lv && lv != &PL_sv_undef) {
14807 U8 s[UTF8_MAXBYTES_CASE+1];
14809 for (i = 0; i <= 256; i++) { /* Look at chars in bitmap */
14810 uvchr_to_utf8(s, i);
14813 && ! ANYOF_BITMAP_TEST(o, i) /* Don't duplicate
14817 && swash_fetch(sw, s, TRUE))
14819 if (rangestart == -1)
14821 } else if (rangestart != -1) {
14822 byte_output = TRUE;
14823 if (i <= rangestart + 3)
14824 for (; rangestart < i; rangestart++) {
14825 put_byte(sv, rangestart);
14828 put_byte(sv, rangestart);
14829 sv_catpvs(sv, "-");
14838 char *s = savesvpv(lv);
14839 char * const origs = s;
14841 while (*s && *s != '\n')
14845 const char * const t = ++s;
14848 sv_catpvs(sv, " ");
14854 /* Truncate very long output */
14855 if (s - origs > 256) {
14856 Perl_sv_catpvf(aTHX_ sv,
14858 (int) (s - origs - 1),
14864 else if (*s == '\t') {
14879 SvREFCNT_dec_NN(lv);
14883 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
14885 else if (k == POSIXD || k == NPOSIXD) {
14886 U8 index = FLAGS(o) * 2;
14887 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
14888 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
14891 sv_catpv(sv, anyofs[index]);
14894 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
14895 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
14897 PERL_UNUSED_CONTEXT;
14898 PERL_UNUSED_ARG(sv);
14899 PERL_UNUSED_ARG(o);
14900 PERL_UNUSED_ARG(prog);
14901 #endif /* DEBUGGING */
14905 Perl_re_intuit_string(pTHX_ REGEXP * const r)
14906 { /* Assume that RE_INTUIT is set */
14908 struct regexp *const prog = ReANY(r);
14909 GET_RE_DEBUG_FLAGS_DECL;
14911 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
14912 PERL_UNUSED_CONTEXT;
14916 const char * const s = SvPV_nolen_const(prog->check_substr
14917 ? prog->check_substr : prog->check_utf8);
14919 if (!PL_colorset) reginitcolors();
14920 PerlIO_printf(Perl_debug_log,
14921 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
14923 prog->check_substr ? "" : "utf8 ",
14924 PL_colors[5],PL_colors[0],
14927 (strlen(s) > 60 ? "..." : ""));
14930 return prog->check_substr ? prog->check_substr : prog->check_utf8;
14936 handles refcounting and freeing the perl core regexp structure. When
14937 it is necessary to actually free the structure the first thing it
14938 does is call the 'free' method of the regexp_engine associated to
14939 the regexp, allowing the handling of the void *pprivate; member
14940 first. (This routine is not overridable by extensions, which is why
14941 the extensions free is called first.)
14943 See regdupe and regdupe_internal if you change anything here.
14945 #ifndef PERL_IN_XSUB_RE
14947 Perl_pregfree(pTHX_ REGEXP *r)
14953 Perl_pregfree2(pTHX_ REGEXP *rx)
14956 struct regexp *const r = ReANY(rx);
14957 GET_RE_DEBUG_FLAGS_DECL;
14959 PERL_ARGS_ASSERT_PREGFREE2;
14961 if (r->mother_re) {
14962 ReREFCNT_dec(r->mother_re);
14964 CALLREGFREE_PVT(rx); /* free the private data */
14965 SvREFCNT_dec(RXp_PAREN_NAMES(r));
14966 Safefree(r->xpv_len_u.xpvlenu_pv);
14969 SvREFCNT_dec(r->anchored_substr);
14970 SvREFCNT_dec(r->anchored_utf8);
14971 SvREFCNT_dec(r->float_substr);
14972 SvREFCNT_dec(r->float_utf8);
14973 Safefree(r->substrs);
14975 RX_MATCH_COPY_FREE(rx);
14976 #ifdef PERL_ANY_COW
14977 SvREFCNT_dec(r->saved_copy);
14980 SvREFCNT_dec(r->qr_anoncv);
14981 rx->sv_u.svu_rx = 0;
14986 This is a hacky workaround to the structural issue of match results
14987 being stored in the regexp structure which is in turn stored in
14988 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
14989 could be PL_curpm in multiple contexts, and could require multiple
14990 result sets being associated with the pattern simultaneously, such
14991 as when doing a recursive match with (??{$qr})
14993 The solution is to make a lightweight copy of the regexp structure
14994 when a qr// is returned from the code executed by (??{$qr}) this
14995 lightweight copy doesn't actually own any of its data except for
14996 the starp/end and the actual regexp structure itself.
15002 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
15004 struct regexp *ret;
15005 struct regexp *const r = ReANY(rx);
15006 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
15008 PERL_ARGS_ASSERT_REG_TEMP_COPY;
15011 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
15013 SvOK_off((SV *)ret_x);
15015 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
15016 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
15017 made both spots point to the same regexp body.) */
15018 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
15019 assert(!SvPVX(ret_x));
15020 ret_x->sv_u.svu_rx = temp->sv_any;
15021 temp->sv_any = NULL;
15022 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
15023 SvREFCNT_dec_NN(temp);
15024 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
15025 ing below will not set it. */
15026 SvCUR_set(ret_x, SvCUR(rx));
15029 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
15030 sv_force_normal(sv) is called. */
15032 ret = ReANY(ret_x);
15034 SvFLAGS(ret_x) |= SvUTF8(rx);
15035 /* We share the same string buffer as the original regexp, on which we
15036 hold a reference count, incremented when mother_re is set below.
15037 The string pointer is copied here, being part of the regexp struct.
15039 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
15040 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
15042 const I32 npar = r->nparens+1;
15043 Newx(ret->offs, npar, regexp_paren_pair);
15044 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15047 Newx(ret->substrs, 1, struct reg_substr_data);
15048 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15050 SvREFCNT_inc_void(ret->anchored_substr);
15051 SvREFCNT_inc_void(ret->anchored_utf8);
15052 SvREFCNT_inc_void(ret->float_substr);
15053 SvREFCNT_inc_void(ret->float_utf8);
15055 /* check_substr and check_utf8, if non-NULL, point to either their
15056 anchored or float namesakes, and don't hold a second reference. */
15058 RX_MATCH_COPIED_off(ret_x);
15059 #ifdef PERL_ANY_COW
15060 ret->saved_copy = NULL;
15062 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
15063 SvREFCNT_inc_void(ret->qr_anoncv);
15069 /* regfree_internal()
15071 Free the private data in a regexp. This is overloadable by
15072 extensions. Perl takes care of the regexp structure in pregfree(),
15073 this covers the *pprivate pointer which technically perl doesn't
15074 know about, however of course we have to handle the
15075 regexp_internal structure when no extension is in use.
15077 Note this is called before freeing anything in the regexp
15082 Perl_regfree_internal(pTHX_ REGEXP * const rx)
15085 struct regexp *const r = ReANY(rx);
15086 RXi_GET_DECL(r,ri);
15087 GET_RE_DEBUG_FLAGS_DECL;
15089 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
15095 SV *dsv= sv_newmortal();
15096 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
15097 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
15098 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
15099 PL_colors[4],PL_colors[5],s);
15102 #ifdef RE_TRACK_PATTERN_OFFSETS
15104 Safefree(ri->u.offsets); /* 20010421 MJD */
15106 if (ri->code_blocks) {
15108 for (n = 0; n < ri->num_code_blocks; n++)
15109 SvREFCNT_dec(ri->code_blocks[n].src_regex);
15110 Safefree(ri->code_blocks);
15114 int n = ri->data->count;
15117 /* If you add a ->what type here, update the comment in regcomp.h */
15118 switch (ri->data->what[n]) {
15124 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
15127 Safefree(ri->data->data[n]);
15133 { /* Aho Corasick add-on structure for a trie node.
15134 Used in stclass optimization only */
15136 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
15138 refcount = --aho->refcount;
15141 PerlMemShared_free(aho->states);
15142 PerlMemShared_free(aho->fail);
15143 /* do this last!!!! */
15144 PerlMemShared_free(ri->data->data[n]);
15145 PerlMemShared_free(ri->regstclass);
15151 /* trie structure. */
15153 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
15155 refcount = --trie->refcount;
15158 PerlMemShared_free(trie->charmap);
15159 PerlMemShared_free(trie->states);
15160 PerlMemShared_free(trie->trans);
15162 PerlMemShared_free(trie->bitmap);
15164 PerlMemShared_free(trie->jump);
15165 PerlMemShared_free(trie->wordinfo);
15166 /* do this last!!!! */
15167 PerlMemShared_free(ri->data->data[n]);
15172 Perl_croak(aTHX_ "panic: regfree data code '%c'", ri->data->what[n]);
15175 Safefree(ri->data->what);
15176 Safefree(ri->data);
15182 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
15183 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
15184 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
15187 re_dup - duplicate a regexp.
15189 This routine is expected to clone a given regexp structure. It is only
15190 compiled under USE_ITHREADS.
15192 After all of the core data stored in struct regexp is duplicated
15193 the regexp_engine.dupe method is used to copy any private data
15194 stored in the *pprivate pointer. This allows extensions to handle
15195 any duplication it needs to do.
15197 See pregfree() and regfree_internal() if you change anything here.
15199 #if defined(USE_ITHREADS)
15200 #ifndef PERL_IN_XSUB_RE
15202 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
15206 const struct regexp *r = ReANY(sstr);
15207 struct regexp *ret = ReANY(dstr);
15209 PERL_ARGS_ASSERT_RE_DUP_GUTS;
15211 npar = r->nparens+1;
15212 Newx(ret->offs, npar, regexp_paren_pair);
15213 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15215 if (ret->substrs) {
15216 /* Do it this way to avoid reading from *r after the StructCopy().
15217 That way, if any of the sv_dup_inc()s dislodge *r from the L1
15218 cache, it doesn't matter. */
15219 const bool anchored = r->check_substr
15220 ? r->check_substr == r->anchored_substr
15221 : r->check_utf8 == r->anchored_utf8;
15222 Newx(ret->substrs, 1, struct reg_substr_data);
15223 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15225 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
15226 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
15227 ret->float_substr = sv_dup_inc(ret->float_substr, param);
15228 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
15230 /* check_substr and check_utf8, if non-NULL, point to either their
15231 anchored or float namesakes, and don't hold a second reference. */
15233 if (ret->check_substr) {
15235 assert(r->check_utf8 == r->anchored_utf8);
15236 ret->check_substr = ret->anchored_substr;
15237 ret->check_utf8 = ret->anchored_utf8;
15239 assert(r->check_substr == r->float_substr);
15240 assert(r->check_utf8 == r->float_utf8);
15241 ret->check_substr = ret->float_substr;
15242 ret->check_utf8 = ret->float_utf8;
15244 } else if (ret->check_utf8) {
15246 ret->check_utf8 = ret->anchored_utf8;
15248 ret->check_utf8 = ret->float_utf8;
15253 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
15254 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
15257 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
15259 if (RX_MATCH_COPIED(dstr))
15260 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
15262 ret->subbeg = NULL;
15263 #ifdef PERL_ANY_COW
15264 ret->saved_copy = NULL;
15267 /* Whether mother_re be set or no, we need to copy the string. We
15268 cannot refrain from copying it when the storage points directly to
15269 our mother regexp, because that's
15270 1: a buffer in a different thread
15271 2: something we no longer hold a reference on
15272 so we need to copy it locally. */
15273 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
15274 ret->mother_re = NULL;
15277 #endif /* PERL_IN_XSUB_RE */
15282 This is the internal complement to regdupe() which is used to copy
15283 the structure pointed to by the *pprivate pointer in the regexp.
15284 This is the core version of the extension overridable cloning hook.
15285 The regexp structure being duplicated will be copied by perl prior
15286 to this and will be provided as the regexp *r argument, however
15287 with the /old/ structures pprivate pointer value. Thus this routine
15288 may override any copying normally done by perl.
15290 It returns a pointer to the new regexp_internal structure.
15294 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
15297 struct regexp *const r = ReANY(rx);
15298 regexp_internal *reti;
15300 RXi_GET_DECL(r,ri);
15302 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
15306 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode), char, regexp_internal);
15307 Copy(ri->program, reti->program, len+1, regnode);
15309 reti->num_code_blocks = ri->num_code_blocks;
15310 if (ri->code_blocks) {
15312 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
15313 struct reg_code_block);
15314 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
15315 struct reg_code_block);
15316 for (n = 0; n < ri->num_code_blocks; n++)
15317 reti->code_blocks[n].src_regex = (REGEXP*)
15318 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
15321 reti->code_blocks = NULL;
15323 reti->regstclass = NULL;
15326 struct reg_data *d;
15327 const int count = ri->data->count;
15330 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
15331 char, struct reg_data);
15332 Newx(d->what, count, U8);
15335 for (i = 0; i < count; i++) {
15336 d->what[i] = ri->data->what[i];
15337 switch (d->what[i]) {
15338 /* see also regcomp.h and regfree_internal() */
15339 case 'a': /* actually an AV, but the dup function is identical. */
15343 case 'u': /* actually an HV, but the dup function is identical. */
15344 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
15347 /* This is cheating. */
15348 Newx(d->data[i], 1, struct regnode_charclass_class);
15349 StructCopy(ri->data->data[i], d->data[i],
15350 struct regnode_charclass_class);
15351 reti->regstclass = (regnode*)d->data[i];
15354 /* Trie stclasses are readonly and can thus be shared
15355 * without duplication. We free the stclass in pregfree
15356 * when the corresponding reg_ac_data struct is freed.
15358 reti->regstclass= ri->regstclass;
15362 ((reg_trie_data*)ri->data->data[i])->refcount++;
15367 d->data[i] = ri->data->data[i];
15370 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'", ri->data->what[i]);
15379 reti->name_list_idx = ri->name_list_idx;
15381 #ifdef RE_TRACK_PATTERN_OFFSETS
15382 if (ri->u.offsets) {
15383 Newx(reti->u.offsets, 2*len+1, U32);
15384 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
15387 SetProgLen(reti,len);
15390 return (void*)reti;
15393 #endif /* USE_ITHREADS */
15395 #ifndef PERL_IN_XSUB_RE
15398 - regnext - dig the "next" pointer out of a node
15401 Perl_regnext(pTHX_ regnode *p)
15409 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
15410 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(p), (int)REGNODE_MAX);
15413 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
15422 S_re_croak2(pTHX_ const char* pat1,const char* pat2,...)
15425 STRLEN l1 = strlen(pat1);
15426 STRLEN l2 = strlen(pat2);
15429 const char *message;
15431 PERL_ARGS_ASSERT_RE_CROAK2;
15437 Copy(pat1, buf, l1 , char);
15438 Copy(pat2, buf + l1, l2 , char);
15439 buf[l1 + l2] = '\n';
15440 buf[l1 + l2 + 1] = '\0';
15442 /* ANSI variant takes additional second argument */
15443 va_start(args, pat2);
15447 msv = vmess(buf, &args);
15449 message = SvPV_const(msv,l1);
15452 Copy(message, buf, l1 , char);
15453 buf[l1-1] = '\0'; /* Overwrite \n */
15454 Perl_croak(aTHX_ "%s", buf);
15457 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
15459 #ifndef PERL_IN_XSUB_RE
15461 Perl_save_re_context(pTHX)
15465 struct re_save_state *state;
15467 SAVEVPTR(PL_curcop);
15468 SSGROW(SAVESTACK_ALLOC_FOR_RE_SAVE_STATE + 1);
15470 state = (struct re_save_state *)(PL_savestack + PL_savestack_ix);
15471 PL_savestack_ix += SAVESTACK_ALLOC_FOR_RE_SAVE_STATE;
15472 SSPUSHUV(SAVEt_RE_STATE);
15474 Copy(&PL_reg_state, state, 1, struct re_save_state);
15476 PL_reg_oldsaved = NULL;
15477 PL_reg_oldsavedlen = 0;
15478 PL_reg_oldsavedoffset = 0;
15479 PL_reg_oldsavedcoffset = 0;
15480 PL_reg_maxiter = 0;
15481 PL_reg_leftiter = 0;
15482 PL_reg_poscache = NULL;
15483 PL_reg_poscache_size = 0;
15484 #ifdef PERL_ANY_COW
15488 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
15490 const REGEXP * const rx = PM_GETRE(PL_curpm);
15493 for (i = 1; i <= RX_NPARENS(rx); i++) {
15494 char digits[TYPE_CHARS(long)];
15495 const STRLEN len = my_snprintf(digits, sizeof(digits), "%lu", (long)i);
15496 GV *const *const gvp
15497 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
15500 GV * const gv = *gvp;
15501 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
15513 S_put_byte(pTHX_ SV *sv, int c)
15515 PERL_ARGS_ASSERT_PUT_BYTE;
15517 /* Our definition of isPRINT() ignores locales, so only bytes that are
15518 not part of UTF-8 are considered printable. I assume that the same
15519 holds for UTF-EBCDIC.
15520 Also, code point 255 is not printable in either (it's E0 in EBCDIC,
15521 which Wikipedia says:
15523 EO, or Eight Ones, is an 8-bit EBCDIC character code represented as all
15524 ones (binary 1111 1111, hexadecimal FF). It is similar, but not
15525 identical, to the ASCII delete (DEL) or rubout control character. ...
15526 it is typically mapped to hexadecimal code 9F, in order to provide a
15527 unique character mapping in both directions)
15529 So the old condition can be simplified to !isPRINT(c) */
15532 Perl_sv_catpvf(aTHX_ sv, "\\x%02x", c);
15535 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
15539 const char string = c;
15540 if (c == '-' || c == ']' || c == '\\' || c == '^')
15541 sv_catpvs(sv, "\\");
15542 sv_catpvn(sv, &string, 1);
15547 #define CLEAR_OPTSTART \
15548 if (optstart) STMT_START { \
15549 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
15553 #define DUMPUNTIL(b,e) CLEAR_OPTSTART; node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
15555 STATIC const regnode *
15556 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
15557 const regnode *last, const regnode *plast,
15558 SV* sv, I32 indent, U32 depth)
15561 U8 op = PSEUDO; /* Arbitrary non-END op. */
15562 const regnode *next;
15563 const regnode *optstart= NULL;
15565 RXi_GET_DECL(r,ri);
15566 GET_RE_DEBUG_FLAGS_DECL;
15568 PERL_ARGS_ASSERT_DUMPUNTIL;
15570 #ifdef DEBUG_DUMPUNTIL
15571 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
15572 last ? last-start : 0,plast ? plast-start : 0);
15575 if (plast && plast < last)
15578 while (PL_regkind[op] != END && (!last || node < last)) {
15579 /* While that wasn't END last time... */
15582 if (op == CLOSE || op == WHILEM)
15584 next = regnext((regnode *)node);
15587 if (OP(node) == OPTIMIZED) {
15588 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
15595 regprop(r, sv, node);
15596 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
15597 (int)(2*indent + 1), "", SvPVX_const(sv));
15599 if (OP(node) != OPTIMIZED) {
15600 if (next == NULL) /* Next ptr. */
15601 PerlIO_printf(Perl_debug_log, " (0)");
15602 else if (PL_regkind[(U8)op] == BRANCH && PL_regkind[OP(next)] != BRANCH )
15603 PerlIO_printf(Perl_debug_log, " (FAIL)");
15605 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
15606 (void)PerlIO_putc(Perl_debug_log, '\n');
15610 if (PL_regkind[(U8)op] == BRANCHJ) {
15613 const regnode *nnode = (OP(next) == LONGJMP
15614 ? regnext((regnode *)next)
15616 if (last && nnode > last)
15618 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
15621 else if (PL_regkind[(U8)op] == BRANCH) {
15623 DUMPUNTIL(NEXTOPER(node), next);
15625 else if ( PL_regkind[(U8)op] == TRIE ) {
15626 const regnode *this_trie = node;
15627 const char op = OP(node);
15628 const U32 n = ARG(node);
15629 const reg_ac_data * const ac = op>=AHOCORASICK ?
15630 (reg_ac_data *)ri->data->data[n] :
15632 const reg_trie_data * const trie =
15633 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
15635 AV *const trie_words = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
15637 const regnode *nextbranch= NULL;
15640 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
15641 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
15643 PerlIO_printf(Perl_debug_log, "%*s%s ",
15644 (int)(2*(indent+3)), "",
15645 elem_ptr ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr), SvCUR(*elem_ptr), 60,
15646 PL_colors[0], PL_colors[1],
15647 (SvUTF8(*elem_ptr) ? PERL_PV_ESCAPE_UNI : 0) |
15648 PERL_PV_PRETTY_ELLIPSES |
15649 PERL_PV_PRETTY_LTGT
15654 U16 dist= trie->jump[word_idx+1];
15655 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
15656 (UV)((dist ? this_trie + dist : next) - start));
15659 nextbranch= this_trie + trie->jump[0];
15660 DUMPUNTIL(this_trie + dist, nextbranch);
15662 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
15663 nextbranch= regnext((regnode *)nextbranch);
15665 PerlIO_printf(Perl_debug_log, "\n");
15668 if (last && next > last)
15673 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
15674 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
15675 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
15677 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
15679 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
15681 else if ( op == PLUS || op == STAR) {
15682 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
15684 else if (PL_regkind[(U8)op] == ANYOF) {
15685 /* arglen 1 + class block */
15686 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_CLASS)
15687 ? ANYOF_CLASS_SKIP : ANYOF_SKIP);
15688 node = NEXTOPER(node);
15690 else if (PL_regkind[(U8)op] == EXACT) {
15691 /* Literal string, where present. */
15692 node += NODE_SZ_STR(node) - 1;
15693 node = NEXTOPER(node);
15696 node = NEXTOPER(node);
15697 node += regarglen[(U8)op];
15699 if (op == CURLYX || op == OPEN)
15703 #ifdef DEBUG_DUMPUNTIL
15704 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
15709 #endif /* DEBUGGING */
15713 * c-indentation-style: bsd
15714 * c-basic-offset: 4
15715 * indent-tabs-mode: nil
15718 * ex: set ts=8 sts=4 sw=4 et: