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_C const struct regexp_engine my_reg_engine;
89 #include "dquote_inline.h"
90 #include "invlist_inline.h"
91 #include "unicode_constants.h"
93 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
97 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
98 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
101 #define STATIC static
104 /* this is a chain of data about sub patterns we are processing that
105 need to be handled separately/specially in study_chunk. Its so
106 we can simulate recursion without losing state. */
108 typedef struct scan_frame {
109 regnode *last_regnode; /* last node to process in this frame */
110 regnode *next_regnode; /* next node to process when last is reached */
111 U32 prev_recursed_depth;
112 I32 stopparen; /* what stopparen do we use */
113 U32 is_top_frame; /* what flags do we use? */
115 struct scan_frame *this_prev_frame; /* this previous frame */
116 struct scan_frame *prev_frame; /* previous frame */
117 struct scan_frame *next_frame; /* next frame */
120 /* Certain characters are output as a sequence with the first being a
122 #define isBACKSLASHED_PUNCT(c) strchr("-[]\\^", c)
125 struct RExC_state_t {
126 U32 flags; /* RXf_* are we folding, multilining? */
127 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
128 char *precomp; /* uncompiled string. */
129 char *precomp_end; /* pointer to end of uncompiled string. */
130 REGEXP *rx_sv; /* The SV that is the regexp. */
131 regexp *rx; /* perl core regexp structure */
132 regexp_internal *rxi; /* internal data for regexp object
134 char *start; /* Start of input for compile */
135 char *end; /* End of input for compile */
136 char *parse; /* Input-scan pointer. */
137 char *adjusted_start; /* 'start', adjusted. See code use */
138 STRLEN precomp_adj; /* an offset beyond precomp. See code use */
139 SSize_t whilem_seen; /* number of WHILEM in this expr */
140 regnode *emit_start; /* Start of emitted-code area */
141 regnode *emit_bound; /* First regnode outside of the
143 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
144 implies compiling, so don't emit */
145 regnode_ssc emit_dummy; /* placeholder for emit to point to;
146 large enough for the largest
147 non-EXACTish node, so can use it as
149 I32 naughty; /* How bad is this pattern? */
150 I32 sawback; /* Did we see \1, ...? */
152 SSize_t size; /* Code size. */
153 I32 npar; /* Capture buffer count, (OPEN) plus
154 one. ("par" 0 is the whole
156 I32 nestroot; /* root parens we are in - used by
160 regnode **open_parens; /* pointers to open parens */
161 regnode **close_parens; /* pointers to close parens */
162 regnode *end_op; /* END node in program */
163 I32 utf8; /* whether the pattern is utf8 or not */
164 I32 orig_utf8; /* whether the pattern was originally in utf8 */
165 /* XXX use this for future optimisation of case
166 * where pattern must be upgraded to utf8. */
167 I32 uni_semantics; /* If a d charset modifier should use unicode
168 rules, even if the pattern is not in
170 HV *paren_names; /* Paren names */
172 regnode **recurse; /* Recurse regops */
173 I32 recurse_count; /* Number of recurse regops we have generated */
174 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
176 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
179 I32 override_recoding;
181 I32 recode_x_to_native;
183 I32 in_multi_char_class;
184 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
186 int code_index; /* next code_blocks[] slot */
187 SSize_t maxlen; /* mininum possible number of chars in string to match */
188 scan_frame *frame_head;
189 scan_frame *frame_last;
192 #ifdef ADD_TO_REGEXEC
193 char *starttry; /* -Dr: where regtry was called. */
194 #define RExC_starttry (pRExC_state->starttry)
196 SV *runtime_code_qr; /* qr with the runtime code blocks */
198 const char *lastparse;
200 AV *paren_name_list; /* idx -> name */
201 U32 study_chunk_recursed_count;
204 #define RExC_lastparse (pRExC_state->lastparse)
205 #define RExC_lastnum (pRExC_state->lastnum)
206 #define RExC_paren_name_list (pRExC_state->paren_name_list)
207 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
208 #define RExC_mysv (pRExC_state->mysv1)
209 #define RExC_mysv1 (pRExC_state->mysv1)
210 #define RExC_mysv2 (pRExC_state->mysv2)
213 bool seen_unfolded_sharp_s;
218 #define RExC_flags (pRExC_state->flags)
219 #define RExC_pm_flags (pRExC_state->pm_flags)
220 #define RExC_precomp (pRExC_state->precomp)
221 #define RExC_precomp_adj (pRExC_state->precomp_adj)
222 #define RExC_adjusted_start (pRExC_state->adjusted_start)
223 #define RExC_precomp_end (pRExC_state->precomp_end)
224 #define RExC_rx_sv (pRExC_state->rx_sv)
225 #define RExC_rx (pRExC_state->rx)
226 #define RExC_rxi (pRExC_state->rxi)
227 #define RExC_start (pRExC_state->start)
228 #define RExC_end (pRExC_state->end)
229 #define RExC_parse (pRExC_state->parse)
230 #define RExC_whilem_seen (pRExC_state->whilem_seen)
232 /* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
233 * EXACTF node, hence was parsed under /di rules. If later in the parse,
234 * something forces the pattern into using /ui rules, the sharp s should be
235 * folded into the sequence 'ss', which takes up more space than previously
236 * calculated. This means that the sizing pass needs to be restarted. (The
237 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
238 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
239 * so there is no need to resize [perl #125990]. */
240 #define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
242 #ifdef RE_TRACK_PATTERN_OFFSETS
243 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
246 #define RExC_emit (pRExC_state->emit)
247 #define RExC_emit_dummy (pRExC_state->emit_dummy)
248 #define RExC_emit_start (pRExC_state->emit_start)
249 #define RExC_emit_bound (pRExC_state->emit_bound)
250 #define RExC_sawback (pRExC_state->sawback)
251 #define RExC_seen (pRExC_state->seen)
252 #define RExC_size (pRExC_state->size)
253 #define RExC_maxlen (pRExC_state->maxlen)
254 #define RExC_npar (pRExC_state->npar)
255 #define RExC_nestroot (pRExC_state->nestroot)
256 #define RExC_extralen (pRExC_state->extralen)
257 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
258 #define RExC_utf8 (pRExC_state->utf8)
259 #define RExC_uni_semantics (pRExC_state->uni_semantics)
260 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
261 #define RExC_open_parens (pRExC_state->open_parens)
262 #define RExC_close_parens (pRExC_state->close_parens)
263 #define RExC_end_op (pRExC_state->end_op)
264 #define RExC_paren_names (pRExC_state->paren_names)
265 #define RExC_recurse (pRExC_state->recurse)
266 #define RExC_recurse_count (pRExC_state->recurse_count)
267 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
268 #define RExC_study_chunk_recursed_bytes \
269 (pRExC_state->study_chunk_recursed_bytes)
270 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
271 #define RExC_contains_locale (pRExC_state->contains_locale)
273 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
275 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
276 #define RExC_frame_head (pRExC_state->frame_head)
277 #define RExC_frame_last (pRExC_state->frame_last)
278 #define RExC_frame_count (pRExC_state->frame_count)
279 #define RExC_strict (pRExC_state->strict)
280 #define RExC_study_started (pRExC_state->study_started)
281 #define RExC_warn_text (pRExC_state->warn_text)
283 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
284 * a flag to disable back-off on the fixed/floating substrings - if it's
285 * a high complexity pattern we assume the benefit of avoiding a full match
286 * is worth the cost of checking for the substrings even if they rarely help.
288 #define RExC_naughty (pRExC_state->naughty)
289 #define TOO_NAUGHTY (10)
290 #define MARK_NAUGHTY(add) \
291 if (RExC_naughty < TOO_NAUGHTY) \
292 RExC_naughty += (add)
293 #define MARK_NAUGHTY_EXP(exp, add) \
294 if (RExC_naughty < TOO_NAUGHTY) \
295 RExC_naughty += RExC_naughty / (exp) + (add)
297 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
298 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
299 ((*s) == '{' && regcurly(s)))
302 * Flags to be passed up and down.
304 #define WORST 0 /* Worst case. */
305 #define HASWIDTH 0x01 /* Known to match non-null strings. */
307 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
308 * character. (There needs to be a case: in the switch statement in regexec.c
309 * for any node marked SIMPLE.) Note that this is not the same thing as
312 #define SPSTART 0x04 /* Starts with * or + */
313 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
314 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
315 #define RESTART_PASS1 0x20 /* Need to restart sizing pass */
316 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
317 calcuate sizes as UTF-8 */
319 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
321 /* whether trie related optimizations are enabled */
322 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
323 #define TRIE_STUDY_OPT
324 #define FULL_TRIE_STUDY
330 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
331 #define PBITVAL(paren) (1 << ((paren) & 7))
332 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
333 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
334 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
336 #define REQUIRE_UTF8(flagp) STMT_START { \
339 *flagp = RESTART_PASS1|NEED_UTF8; \
344 /* Change from /d into /u rules, and restart the parse if we've already seen
345 * something whose size would increase as a result, by setting *flagp and
346 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
347 * we've change to /u during the parse. */
348 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
350 if (DEPENDS_SEMANTICS) { \
352 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
353 RExC_uni_semantics = 1; \
354 if (RExC_seen_unfolded_sharp_s) { \
355 *flagp |= RESTART_PASS1; \
356 return restart_retval; \
361 /* This converts the named class defined in regcomp.h to its equivalent class
362 * number defined in handy.h. */
363 #define namedclass_to_classnum(class) ((int) ((class) / 2))
364 #define classnum_to_namedclass(classnum) ((classnum) * 2)
366 #define _invlist_union_complement_2nd(a, b, output) \
367 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
368 #define _invlist_intersection_complement_2nd(a, b, output) \
369 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
371 /* About scan_data_t.
373 During optimisation we recurse through the regexp program performing
374 various inplace (keyhole style) optimisations. In addition study_chunk
375 and scan_commit populate this data structure with information about
376 what strings MUST appear in the pattern. We look for the longest
377 string that must appear at a fixed location, and we look for the
378 longest string that may appear at a floating location. So for instance
383 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
384 strings (because they follow a .* construct). study_chunk will identify
385 both FOO and BAR as being the longest fixed and floating strings respectively.
387 The strings can be composites, for instance
391 will result in a composite fixed substring 'foo'.
393 For each string some basic information is maintained:
395 - offset or min_offset
396 This is the position the string must appear at, or not before.
397 It also implicitly (when combined with minlenp) tells us how many
398 characters must match before the string we are searching for.
399 Likewise when combined with minlenp and the length of the string it
400 tells us how many characters must appear after the string we have
404 Only used for floating strings. This is the rightmost point that
405 the string can appear at. If set to SSize_t_MAX it indicates that the
406 string can occur infinitely far to the right.
409 A pointer to the minimum number of characters of the pattern that the
410 string was found inside. This is important as in the case of positive
411 lookahead or positive lookbehind we can have multiple patterns
416 The minimum length of the pattern overall is 3, the minimum length
417 of the lookahead part is 3, but the minimum length of the part that
418 will actually match is 1. So 'FOO's minimum length is 3, but the
419 minimum length for the F is 1. This is important as the minimum length
420 is used to determine offsets in front of and behind the string being
421 looked for. Since strings can be composites this is the length of the
422 pattern at the time it was committed with a scan_commit. Note that
423 the length is calculated by study_chunk, so that the minimum lengths
424 are not known until the full pattern has been compiled, thus the
425 pointer to the value.
429 In the case of lookbehind the string being searched for can be
430 offset past the start point of the final matching string.
431 If this value was just blithely removed from the min_offset it would
432 invalidate some of the calculations for how many chars must match
433 before or after (as they are derived from min_offset and minlen and
434 the length of the string being searched for).
435 When the final pattern is compiled and the data is moved from the
436 scan_data_t structure into the regexp structure the information
437 about lookbehind is factored in, with the information that would
438 have been lost precalculated in the end_shift field for the
441 The fields pos_min and pos_delta are used to store the minimum offset
442 and the delta to the maximum offset at the current point in the pattern.
446 typedef struct scan_data_t {
447 /*I32 len_min; unused */
448 /*I32 len_delta; unused */
452 SSize_t last_end; /* min value, <0 unless valid. */
453 SSize_t last_start_min;
454 SSize_t last_start_max;
455 SV **longest; /* Either &l_fixed, or &l_float. */
456 SV *longest_fixed; /* longest fixed string found in pattern */
457 SSize_t offset_fixed; /* offset where it starts */
458 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
459 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
460 SV *longest_float; /* longest floating string found in pattern */
461 SSize_t offset_float_min; /* earliest point in string it can appear */
462 SSize_t offset_float_max; /* latest point in string it can appear */
463 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
464 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
467 SSize_t *last_closep;
468 regnode_ssc *start_class;
472 * Forward declarations for pregcomp()'s friends.
475 static const scan_data_t zero_scan_data =
476 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
478 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
479 #define SF_BEFORE_SEOL 0x0001
480 #define SF_BEFORE_MEOL 0x0002
481 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
482 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
484 #define SF_FIX_SHIFT_EOL (+2)
485 #define SF_FL_SHIFT_EOL (+4)
487 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
488 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
490 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
491 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
492 #define SF_IS_INF 0x0040
493 #define SF_HAS_PAR 0x0080
494 #define SF_IN_PAR 0x0100
495 #define SF_HAS_EVAL 0x0200
498 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
499 * longest substring in the pattern. When it is not set the optimiser keeps
500 * track of position, but does not keep track of the actual strings seen,
502 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
505 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
506 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
507 * turned off because of the alternation (BRANCH). */
508 #define SCF_DO_SUBSTR 0x0400
510 #define SCF_DO_STCLASS_AND 0x0800
511 #define SCF_DO_STCLASS_OR 0x1000
512 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
513 #define SCF_WHILEM_VISITED_POS 0x2000
515 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
516 #define SCF_SEEN_ACCEPT 0x8000
517 #define SCF_TRIE_DOING_RESTUDY 0x10000
518 #define SCF_IN_DEFINE 0x20000
523 #define UTF cBOOL(RExC_utf8)
525 /* The enums for all these are ordered so things work out correctly */
526 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
527 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
528 == REGEX_DEPENDS_CHARSET)
529 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
530 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
531 >= REGEX_UNICODE_CHARSET)
532 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
533 == REGEX_ASCII_RESTRICTED_CHARSET)
534 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
535 >= REGEX_ASCII_RESTRICTED_CHARSET)
536 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
537 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
539 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
541 /* For programs that want to be strictly Unicode compatible by dying if any
542 * attempt is made to match a non-Unicode code point against a Unicode
544 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
546 #define OOB_NAMEDCLASS -1
548 /* There is no code point that is out-of-bounds, so this is problematic. But
549 * its only current use is to initialize a variable that is always set before
551 #define OOB_UNICODE 0xDEADBEEF
553 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
556 /* length of regex to show in messages that don't mark a position within */
557 #define RegexLengthToShowInErrorMessages 127
560 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
561 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
562 * op/pragma/warn/regcomp.
564 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
565 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
567 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
568 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
570 /* The code in this file in places uses one level of recursion with parsing
571 * rebased to an alternate string constructed by us in memory. This can take
572 * the form of something that is completely different from the input, or
573 * something that uses the input as part of the alternate. In the first case,
574 * there should be no possibility of an error, as we are in complete control of
575 * the alternate string. But in the second case we don't control the input
576 * portion, so there may be errors in that. Here's an example:
578 * is handled specially because \x{df} folds to a sequence of more than one
579 * character, 'ss'. What is done is to create and parse an alternate string,
580 * which looks like this:
581 * /(?:\x{DF}|[abc\x{DF}def])/ui
582 * where it uses the input unchanged in the middle of something it constructs,
583 * which is a branch for the DF outside the character class, and clustering
584 * parens around the whole thing. (It knows enough to skip the DF inside the
585 * class while in this substitute parse.) 'abc' and 'def' may have errors that
586 * need to be reported. The general situation looks like this:
589 * Input: ----------------------------------------------------
590 * Constructed: ---------------------------------------------------
593 * The input string sI..eI is the input pattern. The string sC..EC is the
594 * constructed substitute parse string. The portions sC..tC and eC..EC are
595 * constructed by us. The portion tC..eC is an exact duplicate of the input
596 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
597 * while parsing, we find an error at xC. We want to display a message showing
598 * the real input string. Thus we need to find the point xI in it which
599 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
600 * been constructed by us, and so shouldn't have errors. We get:
602 * xI = sI + (tI - sI) + (xC - tC)
604 * and, the offset into sI is:
606 * (xI - sI) = (tI - sI) + (xC - tC)
608 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
609 * and we save tC as RExC_adjusted_start.
611 * During normal processing of the input pattern, everything points to that,
612 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
615 #define tI_sI RExC_precomp_adj
616 #define tC RExC_adjusted_start
617 #define sC RExC_precomp
618 #define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
619 #define xI(xC) (sC + xI_offset(xC))
620 #define eC RExC_precomp_end
622 #define REPORT_LOCATION_ARGS(xC) \
624 (xI(xC) > eC) /* Don't run off end */ \
625 ? eC - sC /* Length before the <--HERE */ \
627 sC), /* The input pattern printed up to the <--HERE */ \
629 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
630 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
632 /* Used to point after bad bytes for an error message, but avoid skipping
633 * past a nul byte. */
634 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
637 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
638 * arg. Show regex, up to a maximum length. If it's too long, chop and add
641 #define _FAIL(code) STMT_START { \
642 const char *ellipses = ""; \
643 IV len = RExC_precomp_end - RExC_precomp; \
646 SAVEFREESV(RExC_rx_sv); \
647 if (len > RegexLengthToShowInErrorMessages) { \
648 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
649 len = RegexLengthToShowInErrorMessages - 10; \
655 #define FAIL(msg) _FAIL( \
656 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
657 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
659 #define FAIL2(msg,arg) _FAIL( \
660 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
661 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
664 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
666 #define Simple_vFAIL(m) STMT_START { \
667 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
668 m, REPORT_LOCATION_ARGS(RExC_parse)); \
672 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
674 #define vFAIL(m) STMT_START { \
676 SAVEFREESV(RExC_rx_sv); \
681 * Like Simple_vFAIL(), but accepts two arguments.
683 #define Simple_vFAIL2(m,a1) STMT_START { \
684 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
685 REPORT_LOCATION_ARGS(RExC_parse)); \
689 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
691 #define vFAIL2(m,a1) STMT_START { \
693 SAVEFREESV(RExC_rx_sv); \
694 Simple_vFAIL2(m, a1); \
699 * Like Simple_vFAIL(), but accepts three arguments.
701 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
702 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
703 REPORT_LOCATION_ARGS(RExC_parse)); \
707 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
709 #define vFAIL3(m,a1,a2) STMT_START { \
711 SAVEFREESV(RExC_rx_sv); \
712 Simple_vFAIL3(m, a1, a2); \
716 * Like Simple_vFAIL(), but accepts four arguments.
718 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
719 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
720 REPORT_LOCATION_ARGS(RExC_parse)); \
723 #define vFAIL4(m,a1,a2,a3) STMT_START { \
725 SAVEFREESV(RExC_rx_sv); \
726 Simple_vFAIL4(m, a1, a2, a3); \
729 /* A specialized version of vFAIL2 that works with UTF8f */
730 #define vFAIL2utf8f(m, a1) STMT_START { \
732 SAVEFREESV(RExC_rx_sv); \
733 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
734 REPORT_LOCATION_ARGS(RExC_parse)); \
737 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
739 SAVEFREESV(RExC_rx_sv); \
740 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
741 REPORT_LOCATION_ARGS(RExC_parse)); \
744 /* These have asserts in them because of [perl #122671] Many warnings in
745 * regcomp.c can occur twice. If they get output in pass1 and later in that
746 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
747 * would get output again. So they should be output in pass2, and these
748 * asserts make sure new warnings follow that paradigm. */
750 /* m is not necessarily a "literal string", in this macro */
751 #define reg_warn_non_literal_string(loc, m) STMT_START { \
752 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
753 "%s" REPORT_LOCATION, \
754 m, REPORT_LOCATION_ARGS(loc)); \
757 #define ckWARNreg(loc,m) STMT_START { \
758 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
760 REPORT_LOCATION_ARGS(loc)); \
763 #define vWARN(loc, m) STMT_START { \
764 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
766 REPORT_LOCATION_ARGS(loc)); \
769 #define vWARN_dep(loc, m) STMT_START { \
770 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
772 REPORT_LOCATION_ARGS(loc)); \
775 #define ckWARNdep(loc,m) STMT_START { \
776 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
778 REPORT_LOCATION_ARGS(loc)); \
781 #define ckWARNregdep(loc,m) STMT_START { \
782 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
785 REPORT_LOCATION_ARGS(loc)); \
788 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
789 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
791 a1, REPORT_LOCATION_ARGS(loc)); \
794 #define ckWARN2reg(loc, m, a1) STMT_START { \
795 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
797 a1, REPORT_LOCATION_ARGS(loc)); \
800 #define vWARN3(loc, m, a1, a2) STMT_START { \
801 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
803 a1, a2, REPORT_LOCATION_ARGS(loc)); \
806 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
807 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
810 REPORT_LOCATION_ARGS(loc)); \
813 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
814 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
817 REPORT_LOCATION_ARGS(loc)); \
820 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
821 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
824 REPORT_LOCATION_ARGS(loc)); \
827 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
828 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
831 REPORT_LOCATION_ARGS(loc)); \
834 /* Macros for recording node offsets. 20001227 mjd@plover.com
835 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
836 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
837 * Element 0 holds the number n.
838 * Position is 1 indexed.
840 #ifndef RE_TRACK_PATTERN_OFFSETS
841 #define Set_Node_Offset_To_R(node,byte)
842 #define Set_Node_Offset(node,byte)
843 #define Set_Cur_Node_Offset
844 #define Set_Node_Length_To_R(node,len)
845 #define Set_Node_Length(node,len)
846 #define Set_Node_Cur_Length(node,start)
847 #define Node_Offset(n)
848 #define Node_Length(n)
849 #define Set_Node_Offset_Length(node,offset,len)
850 #define ProgLen(ri) ri->u.proglen
851 #define SetProgLen(ri,x) ri->u.proglen = x
853 #define ProgLen(ri) ri->u.offsets[0]
854 #define SetProgLen(ri,x) ri->u.offsets[0] = x
855 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
857 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
858 __LINE__, (int)(node), (int)(byte))); \
860 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
863 RExC_offsets[2*(node)-1] = (byte); \
868 #define Set_Node_Offset(node,byte) \
869 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
870 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
872 #define Set_Node_Length_To_R(node,len) STMT_START { \
874 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
875 __LINE__, (int)(node), (int)(len))); \
877 Perl_croak(aTHX_ "value of node is %d in Length macro", \
880 RExC_offsets[2*(node)] = (len); \
885 #define Set_Node_Length(node,len) \
886 Set_Node_Length_To_R((node)-RExC_emit_start, len)
887 #define Set_Node_Cur_Length(node, start) \
888 Set_Node_Length(node, RExC_parse - start)
890 /* Get offsets and lengths */
891 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
892 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
894 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
895 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
896 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
900 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
901 #define EXPERIMENTAL_INPLACESCAN
902 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
906 Perl_re_printf(pTHX_ const char *fmt, ...)
910 PerlIO *f= Perl_debug_log;
911 PERL_ARGS_ASSERT_RE_PRINTF;
913 result = PerlIO_vprintf(f, fmt, ap);
919 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
923 PerlIO *f= Perl_debug_log;
924 PERL_ARGS_ASSERT_RE_INDENTF;
926 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
927 result = PerlIO_vprintf(f, fmt, ap);
931 #endif /* DEBUGGING */
933 #define DEBUG_RExC_seen() \
934 DEBUG_OPTIMISE_MORE_r({ \
935 Perl_re_printf( aTHX_ "RExC_seen: "); \
937 if (RExC_seen & REG_ZERO_LEN_SEEN) \
938 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
940 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
941 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
943 if (RExC_seen & REG_GPOS_SEEN) \
944 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
946 if (RExC_seen & REG_RECURSE_SEEN) \
947 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
949 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
950 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
952 if (RExC_seen & REG_VERBARG_SEEN) \
953 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
955 if (RExC_seen & REG_CUTGROUP_SEEN) \
956 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
958 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
959 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
961 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
962 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
964 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
965 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
967 Perl_re_printf( aTHX_ "\n"); \
970 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
971 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
973 #define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
975 Perl_re_printf( aTHX_ "%s", open_str); \
976 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
977 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
978 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
979 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
980 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
981 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
982 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
983 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
984 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
985 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
986 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
987 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
988 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
989 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
990 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
991 Perl_re_printf( aTHX_ "%s", close_str); \
995 #define DEBUG_STUDYDATA(str,data,depth) \
996 DEBUG_OPTIMISE_MORE_r(if(data){ \
997 Perl_re_indentf( aTHX_ "" str "Pos:%" IVdf "/%" IVdf \
998 " Flags: 0x%" UVXf, \
1000 (IV)((data)->pos_min), \
1001 (IV)((data)->pos_delta), \
1002 (UV)((data)->flags) \
1004 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
1005 Perl_re_printf( aTHX_ \
1006 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s", \
1007 (IV)((data)->whilem_c), \
1008 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
1009 is_inf ? "INF " : "" \
1011 if ((data)->last_found) \
1012 Perl_re_printf( aTHX_ \
1013 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf \
1014 " %sFixed:'%s' @ %" IVdf \
1015 " %sFloat: '%s' @ %" IVdf "/%" IVdf, \
1016 SvPVX_const((data)->last_found), \
1017 (IV)((data)->last_end), \
1018 (IV)((data)->last_start_min), \
1019 (IV)((data)->last_start_max), \
1020 ((data)->longest && \
1021 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
1022 SvPVX_const((data)->longest_fixed), \
1023 (IV)((data)->offset_fixed), \
1024 ((data)->longest && \
1025 (data)->longest==&((data)->longest_float)) ? "*" : "", \
1026 SvPVX_const((data)->longest_float), \
1027 (IV)((data)->offset_float_min), \
1028 (IV)((data)->offset_float_max) \
1030 Perl_re_printf( aTHX_ "\n"); \
1034 /* =========================================================
1035 * BEGIN edit_distance stuff.
1037 * This calculates how many single character changes of any type are needed to
1038 * transform a string into another one. It is taken from version 3.1 of
1040 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1043 /* Our unsorted dictionary linked list. */
1044 /* Note we use UVs, not chars. */
1049 struct dictionary* next;
1051 typedef struct dictionary item;
1054 PERL_STATIC_INLINE item*
1055 push(UV key,item* curr)
1058 Newxz(head, 1, item);
1066 PERL_STATIC_INLINE item*
1067 find(item* head, UV key)
1069 item* iterator = head;
1071 if (iterator->key == key){
1074 iterator = iterator->next;
1080 PERL_STATIC_INLINE item*
1081 uniquePush(item* head,UV key)
1083 item* iterator = head;
1086 if (iterator->key == key) {
1089 iterator = iterator->next;
1092 return push(key,head);
1095 PERL_STATIC_INLINE void
1096 dict_free(item* head)
1098 item* iterator = head;
1101 item* temp = iterator;
1102 iterator = iterator->next;
1109 /* End of Dictionary Stuff */
1111 /* All calculations/work are done here */
1113 S_edit_distance(const UV* src,
1115 const STRLEN x, /* length of src[] */
1116 const STRLEN y, /* length of tgt[] */
1117 const SSize_t maxDistance
1121 UV swapCount,swapScore,targetCharCount,i,j;
1123 UV score_ceil = x + y;
1125 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1127 /* intialize matrix start values */
1128 Newxz(scores, ( (x + 2) * (y + 2)), UV);
1129 scores[0] = score_ceil;
1130 scores[1 * (y + 2) + 0] = score_ceil;
1131 scores[0 * (y + 2) + 1] = score_ceil;
1132 scores[1 * (y + 2) + 1] = 0;
1133 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1138 for (i=1;i<=x;i++) {
1140 head = uniquePush(head,src[i]);
1141 scores[(i+1) * (y + 2) + 1] = i;
1142 scores[(i+1) * (y + 2) + 0] = score_ceil;
1145 for (j=1;j<=y;j++) {
1148 head = uniquePush(head,tgt[j]);
1149 scores[1 * (y + 2) + (j + 1)] = j;
1150 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1153 targetCharCount = find(head,tgt[j-1])->value;
1154 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1156 if (src[i-1] != tgt[j-1]){
1157 scores[(i+1) * (y + 2) + (j + 1)] = MIN(swapScore,(MIN(scores[i * (y + 2) + j], MIN(scores[(i+1) * (y + 2) + j], scores[i * (y + 2) + (j + 1)])) + 1));
1161 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1165 find(head,src[i-1])->value = i;
1169 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1172 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1176 /* END of edit_distance() stuff
1177 * ========================================================= */
1179 /* is c a control character for which we have a mnemonic? */
1180 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1183 S_cntrl_to_mnemonic(const U8 c)
1185 /* Returns the mnemonic string that represents character 'c', if one
1186 * exists; NULL otherwise. The only ones that exist for the purposes of
1187 * this routine are a few control characters */
1190 case '\a': return "\\a";
1191 case '\b': return "\\b";
1192 case ESC_NATIVE: return "\\e";
1193 case '\f': return "\\f";
1194 case '\n': return "\\n";
1195 case '\r': return "\\r";
1196 case '\t': return "\\t";
1202 /* Mark that we cannot extend a found fixed substring at this point.
1203 Update the longest found anchored substring and the longest found
1204 floating substrings if needed. */
1207 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1208 SSize_t *minlenp, int is_inf)
1210 const STRLEN l = CHR_SVLEN(data->last_found);
1211 const STRLEN old_l = CHR_SVLEN(*data->longest);
1212 GET_RE_DEBUG_FLAGS_DECL;
1214 PERL_ARGS_ASSERT_SCAN_COMMIT;
1216 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1217 SvSetMagicSV(*data->longest, data->last_found);
1218 if (*data->longest == data->longest_fixed) {
1219 data->offset_fixed = l ? data->last_start_min : data->pos_min;
1220 if (data->flags & SF_BEFORE_EOL)
1222 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
1224 data->flags &= ~SF_FIX_BEFORE_EOL;
1225 data->minlen_fixed=minlenp;
1226 data->lookbehind_fixed=0;
1228 else { /* *data->longest == data->longest_float */
1229 data->offset_float_min = l ? data->last_start_min : data->pos_min;
1230 data->offset_float_max = (l
1231 ? data->last_start_max
1232 : (data->pos_delta > SSize_t_MAX - data->pos_min
1234 : data->pos_min + data->pos_delta));
1236 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
1237 data->offset_float_max = SSize_t_MAX;
1238 if (data->flags & SF_BEFORE_EOL)
1240 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
1242 data->flags &= ~SF_FL_BEFORE_EOL;
1243 data->minlen_float=minlenp;
1244 data->lookbehind_float=0;
1247 SvCUR_set(data->last_found, 0);
1249 SV * const sv = data->last_found;
1250 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1251 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1256 data->last_end = -1;
1257 data->flags &= ~SF_BEFORE_EOL;
1258 DEBUG_STUDYDATA("commit: ",data,0);
1261 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1262 * list that describes which code points it matches */
1265 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1267 /* Set the SSC 'ssc' to match an empty string or any code point */
1269 PERL_ARGS_ASSERT_SSC_ANYTHING;
1271 assert(is_ANYOF_SYNTHETIC(ssc));
1273 /* mortalize so won't leak */
1274 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1275 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1279 S_ssc_is_anything(const regnode_ssc *ssc)
1281 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1282 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1283 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1284 * in any way, so there's no point in using it */
1289 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1291 assert(is_ANYOF_SYNTHETIC(ssc));
1293 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1297 /* See if the list consists solely of the range 0 - Infinity */
1298 invlist_iterinit(ssc->invlist);
1299 ret = invlist_iternext(ssc->invlist, &start, &end)
1303 invlist_iterfinish(ssc->invlist);
1309 /* If e.g., both \w and \W are set, matches everything */
1310 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1312 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1313 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1323 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1325 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1326 * string, any code point, or any posix class under locale */
1328 PERL_ARGS_ASSERT_SSC_INIT;
1330 Zero(ssc, 1, regnode_ssc);
1331 set_ANYOF_SYNTHETIC(ssc);
1332 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1335 /* If any portion of the regex is to operate under locale rules that aren't
1336 * fully known at compile time, initialization includes it. The reason
1337 * this isn't done for all regexes is that the optimizer was written under
1338 * the assumption that locale was all-or-nothing. Given the complexity and
1339 * lack of documentation in the optimizer, and that there are inadequate
1340 * test cases for locale, many parts of it may not work properly, it is
1341 * safest to avoid locale unless necessary. */
1342 if (RExC_contains_locale) {
1343 ANYOF_POSIXL_SETALL(ssc);
1346 ANYOF_POSIXL_ZERO(ssc);
1351 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1352 const regnode_ssc *ssc)
1354 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1355 * to the list of code points matched, and locale posix classes; hence does
1356 * not check its flags) */
1361 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1363 assert(is_ANYOF_SYNTHETIC(ssc));
1365 invlist_iterinit(ssc->invlist);
1366 ret = invlist_iternext(ssc->invlist, &start, &end)
1370 invlist_iterfinish(ssc->invlist);
1376 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1384 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1385 const regnode_charclass* const node)
1387 /* Returns a mortal inversion list defining which code points are matched
1388 * by 'node', which is of type ANYOF. Handles complementing the result if
1389 * appropriate. If some code points aren't knowable at this time, the
1390 * returned list must, and will, contain every code point that is a
1394 SV* only_utf8_locale_invlist = NULL;
1396 const U32 n = ARG(node);
1397 bool new_node_has_latin1 = FALSE;
1399 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1401 /* Look at the data structure created by S_set_ANYOF_arg() */
1402 if (n != ANYOF_ONLY_HAS_BITMAP) {
1403 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1404 AV * const av = MUTABLE_AV(SvRV(rv));
1405 SV **const ary = AvARRAY(av);
1406 assert(RExC_rxi->data->what[n] == 's');
1408 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1409 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1411 else if (ary[0] && ary[0] != &PL_sv_undef) {
1413 /* Here, no compile-time swash, and there are things that won't be
1414 * known until runtime -- we have to assume it could be anything */
1415 invlist = sv_2mortal(_new_invlist(1));
1416 return _add_range_to_invlist(invlist, 0, UV_MAX);
1418 else if (ary[3] && ary[3] != &PL_sv_undef) {
1420 /* Here no compile-time swash, and no run-time only data. Use the
1421 * node's inversion list */
1422 invlist = sv_2mortal(invlist_clone(ary[3]));
1425 /* Get the code points valid only under UTF-8 locales */
1426 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1427 && ary[2] && ary[2] != &PL_sv_undef)
1429 only_utf8_locale_invlist = ary[2];
1434 invlist = sv_2mortal(_new_invlist(0));
1437 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1438 * code points, and an inversion list for the others, but if there are code
1439 * points that should match only conditionally on the target string being
1440 * UTF-8, those are placed in the inversion list, and not the bitmap.
1441 * Since there are circumstances under which they could match, they are
1442 * included in the SSC. But if the ANYOF node is to be inverted, we have
1443 * to exclude them here, so that when we invert below, the end result
1444 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1445 * have to do this here before we add the unconditionally matched code
1447 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1448 _invlist_intersection_complement_2nd(invlist,
1453 /* Add in the points from the bit map */
1454 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1455 if (ANYOF_BITMAP_TEST(node, i)) {
1456 unsigned int start = i++;
1458 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1461 invlist = _add_range_to_invlist(invlist, start, i-1);
1462 new_node_has_latin1 = TRUE;
1466 /* If this can match all upper Latin1 code points, have to add them
1467 * as well. But don't add them if inverting, as when that gets done below,
1468 * it would exclude all these characters, including the ones it shouldn't
1469 * that were added just above */
1470 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1471 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1473 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1476 /* Similarly for these */
1477 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1478 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1481 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1482 _invlist_invert(invlist);
1484 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1486 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1487 * locale. We can skip this if there are no 0-255 at all. */
1488 _invlist_union(invlist, PL_Latin1, &invlist);
1491 /* Similarly add the UTF-8 locale possible matches. These have to be
1492 * deferred until after the non-UTF-8 locale ones are taken care of just
1493 * above, or it leads to wrong results under ANYOF_INVERT */
1494 if (only_utf8_locale_invlist) {
1495 _invlist_union_maybe_complement_2nd(invlist,
1496 only_utf8_locale_invlist,
1497 ANYOF_FLAGS(node) & ANYOF_INVERT,
1504 /* These two functions currently do the exact same thing */
1505 #define ssc_init_zero ssc_init
1507 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1508 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1510 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1511 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1512 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1515 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1516 const regnode_charclass *and_with)
1518 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1519 * another SSC or a regular ANYOF class. Can create false positives. */
1524 PERL_ARGS_ASSERT_SSC_AND;
1526 assert(is_ANYOF_SYNTHETIC(ssc));
1528 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1529 * the code point inversion list and just the relevant flags */
1530 if (is_ANYOF_SYNTHETIC(and_with)) {
1531 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1532 anded_flags = ANYOF_FLAGS(and_with);
1534 /* XXX This is a kludge around what appears to be deficiencies in the
1535 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1536 * there are paths through the optimizer where it doesn't get weeded
1537 * out when it should. And if we don't make some extra provision for
1538 * it like the code just below, it doesn't get added when it should.
1539 * This solution is to add it only when AND'ing, which is here, and
1540 * only when what is being AND'ed is the pristine, original node
1541 * matching anything. Thus it is like adding it to ssc_anything() but
1542 * only when the result is to be AND'ed. Probably the same solution
1543 * could be adopted for the same problem we have with /l matching,
1544 * which is solved differently in S_ssc_init(), and that would lead to
1545 * fewer false positives than that solution has. But if this solution
1546 * creates bugs, the consequences are only that a warning isn't raised
1547 * that should be; while the consequences for having /l bugs is
1548 * incorrect matches */
1549 if (ssc_is_anything((regnode_ssc *)and_with)) {
1550 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1554 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1555 if (OP(and_with) == ANYOFD) {
1556 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1559 anded_flags = ANYOF_FLAGS(and_with)
1560 &( ANYOF_COMMON_FLAGS
1561 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1562 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1563 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1565 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1570 ANYOF_FLAGS(ssc) &= anded_flags;
1572 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1573 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1574 * 'and_with' may be inverted. When not inverted, we have the situation of
1576 * (C1 | P1) & (C2 | P2)
1577 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1578 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1579 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1580 * <= ((C1 & C2) | P1 | P2)
1581 * Alternatively, the last few steps could be:
1582 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1583 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1584 * <= (C1 | C2 | (P1 & P2))
1585 * We favor the second approach if either P1 or P2 is non-empty. This is
1586 * because these components are a barrier to doing optimizations, as what
1587 * they match cannot be known until the moment of matching as they are
1588 * dependent on the current locale, 'AND"ing them likely will reduce or
1590 * But we can do better if we know that C1,P1 are in their initial state (a
1591 * frequent occurrence), each matching everything:
1592 * (<everything>) & (C2 | P2) = C2 | P2
1593 * Similarly, if C2,P2 are in their initial state (again a frequent
1594 * occurrence), the result is a no-op
1595 * (C1 | P1) & (<everything>) = C1 | P1
1598 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1599 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1600 * <= (C1 & ~C2) | (P1 & ~P2)
1603 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1604 && ! is_ANYOF_SYNTHETIC(and_with))
1608 ssc_intersection(ssc,
1610 FALSE /* Has already been inverted */
1613 /* If either P1 or P2 is empty, the intersection will be also; can skip
1615 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1616 ANYOF_POSIXL_ZERO(ssc);
1618 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1620 /* Note that the Posix class component P from 'and_with' actually
1622 * P = Pa | Pb | ... | Pn
1623 * where each component is one posix class, such as in [\w\s].
1625 * ~P = ~(Pa | Pb | ... | Pn)
1626 * = ~Pa & ~Pb & ... & ~Pn
1627 * <= ~Pa | ~Pb | ... | ~Pn
1628 * The last is something we can easily calculate, but unfortunately
1629 * is likely to have many false positives. We could do better
1630 * in some (but certainly not all) instances if two classes in
1631 * P have known relationships. For example
1632 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1634 * :lower: & :print: = :lower:
1635 * And similarly for classes that must be disjoint. For example,
1636 * since \s and \w can have no elements in common based on rules in
1637 * the POSIX standard,
1638 * \w & ^\S = nothing
1639 * Unfortunately, some vendor locales do not meet the Posix
1640 * standard, in particular almost everything by Microsoft.
1641 * The loop below just changes e.g., \w into \W and vice versa */
1643 regnode_charclass_posixl temp;
1644 int add = 1; /* To calculate the index of the complement */
1646 ANYOF_POSIXL_ZERO(&temp);
1647 for (i = 0; i < ANYOF_MAX; i++) {
1649 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1650 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1652 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1653 ANYOF_POSIXL_SET(&temp, i + add);
1655 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1657 ANYOF_POSIXL_AND(&temp, ssc);
1659 } /* else ssc already has no posixes */
1660 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1661 in its initial state */
1662 else if (! is_ANYOF_SYNTHETIC(and_with)
1663 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1665 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1666 * copy it over 'ssc' */
1667 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1668 if (is_ANYOF_SYNTHETIC(and_with)) {
1669 StructCopy(and_with, ssc, regnode_ssc);
1672 ssc->invlist = anded_cp_list;
1673 ANYOF_POSIXL_ZERO(ssc);
1674 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1675 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1679 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1680 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1682 /* One or the other of P1, P2 is non-empty. */
1683 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1684 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1686 ssc_union(ssc, anded_cp_list, FALSE);
1688 else { /* P1 = P2 = empty */
1689 ssc_intersection(ssc, anded_cp_list, FALSE);
1695 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1696 const regnode_charclass *or_with)
1698 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1699 * another SSC or a regular ANYOF class. Can create false positives if
1700 * 'or_with' is to be inverted. */
1705 PERL_ARGS_ASSERT_SSC_OR;
1707 assert(is_ANYOF_SYNTHETIC(ssc));
1709 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1710 * the code point inversion list and just the relevant flags */
1711 if (is_ANYOF_SYNTHETIC(or_with)) {
1712 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1713 ored_flags = ANYOF_FLAGS(or_with);
1716 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1717 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1718 if (OP(or_with) != ANYOFD) {
1720 |= ANYOF_FLAGS(or_with)
1721 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1722 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1723 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1725 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1730 ANYOF_FLAGS(ssc) |= ored_flags;
1732 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1733 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1734 * 'or_with' may be inverted. When not inverted, we have the simple
1735 * situation of computing:
1736 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1737 * If P1|P2 yields a situation with both a class and its complement are
1738 * set, like having both \w and \W, this matches all code points, and we
1739 * can delete these from the P component of the ssc going forward. XXX We
1740 * might be able to delete all the P components, but I (khw) am not certain
1741 * about this, and it is better to be safe.
1744 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1745 * <= (C1 | P1) | ~C2
1746 * <= (C1 | ~C2) | P1
1747 * (which results in actually simpler code than the non-inverted case)
1750 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1751 && ! is_ANYOF_SYNTHETIC(or_with))
1753 /* We ignore P2, leaving P1 going forward */
1754 } /* else Not inverted */
1755 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1756 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1757 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1759 for (i = 0; i < ANYOF_MAX; i += 2) {
1760 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1762 ssc_match_all_cp(ssc);
1763 ANYOF_POSIXL_CLEAR(ssc, i);
1764 ANYOF_POSIXL_CLEAR(ssc, i+1);
1772 FALSE /* Already has been inverted */
1776 PERL_STATIC_INLINE void
1777 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1779 PERL_ARGS_ASSERT_SSC_UNION;
1781 assert(is_ANYOF_SYNTHETIC(ssc));
1783 _invlist_union_maybe_complement_2nd(ssc->invlist,
1789 PERL_STATIC_INLINE void
1790 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1792 const bool invert2nd)
1794 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1796 assert(is_ANYOF_SYNTHETIC(ssc));
1798 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1804 PERL_STATIC_INLINE void
1805 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1807 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1809 assert(is_ANYOF_SYNTHETIC(ssc));
1811 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1814 PERL_STATIC_INLINE void
1815 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1817 /* AND just the single code point 'cp' into the SSC 'ssc' */
1819 SV* cp_list = _new_invlist(2);
1821 PERL_ARGS_ASSERT_SSC_CP_AND;
1823 assert(is_ANYOF_SYNTHETIC(ssc));
1825 cp_list = add_cp_to_invlist(cp_list, cp);
1826 ssc_intersection(ssc, cp_list,
1827 FALSE /* Not inverted */
1829 SvREFCNT_dec_NN(cp_list);
1832 PERL_STATIC_INLINE void
1833 S_ssc_clear_locale(regnode_ssc *ssc)
1835 /* Set the SSC 'ssc' to not match any locale things */
1836 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1838 assert(is_ANYOF_SYNTHETIC(ssc));
1840 ANYOF_POSIXL_ZERO(ssc);
1841 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1844 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1847 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1849 /* The synthetic start class is used to hopefully quickly winnow down
1850 * places where a pattern could start a match in the target string. If it
1851 * doesn't really narrow things down that much, there isn't much point to
1852 * having the overhead of using it. This function uses some very crude
1853 * heuristics to decide if to use the ssc or not.
1855 * It returns TRUE if 'ssc' rules out more than half what it considers to
1856 * be the "likely" possible matches, but of course it doesn't know what the
1857 * actual things being matched are going to be; these are only guesses
1859 * For /l matches, it assumes that the only likely matches are going to be
1860 * in the 0-255 range, uniformly distributed, so half of that is 127
1861 * For /a and /d matches, it assumes that the likely matches will be just
1862 * the ASCII range, so half of that is 63
1863 * For /u and there isn't anything matching above the Latin1 range, it
1864 * assumes that that is the only range likely to be matched, and uses
1865 * half that as the cut-off: 127. If anything matches above Latin1,
1866 * it assumes that all of Unicode could match (uniformly), except for
1867 * non-Unicode code points and things in the General Category "Other"
1868 * (unassigned, private use, surrogates, controls and formats). This
1869 * is a much large number. */
1871 U32 count = 0; /* Running total of number of code points matched by
1873 UV start, end; /* Start and end points of current range in inversion
1875 const U32 max_code_points = (LOC)
1877 : (( ! UNI_SEMANTICS
1878 || invlist_highest(ssc->invlist) < 256)
1881 const U32 max_match = max_code_points / 2;
1883 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1885 invlist_iterinit(ssc->invlist);
1886 while (invlist_iternext(ssc->invlist, &start, &end)) {
1887 if (start >= max_code_points) {
1890 end = MIN(end, max_code_points - 1);
1891 count += end - start + 1;
1892 if (count >= max_match) {
1893 invlist_iterfinish(ssc->invlist);
1903 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1905 /* The inversion list in the SSC is marked mortal; now we need a more
1906 * permanent copy, which is stored the same way that is done in a regular
1907 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1910 SV* invlist = invlist_clone(ssc->invlist);
1912 PERL_ARGS_ASSERT_SSC_FINALIZE;
1914 assert(is_ANYOF_SYNTHETIC(ssc));
1916 /* The code in this file assumes that all but these flags aren't relevant
1917 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1918 * by the time we reach here */
1919 assert(! (ANYOF_FLAGS(ssc)
1920 & ~( ANYOF_COMMON_FLAGS
1921 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1922 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
1924 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1926 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1927 NULL, NULL, NULL, FALSE);
1929 /* Make sure is clone-safe */
1930 ssc->invlist = NULL;
1932 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1933 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1936 if (RExC_contains_locale) {
1940 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1943 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1944 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1945 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1946 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1947 ? (TRIE_LIST_CUR( idx ) - 1) \
1953 dump_trie(trie,widecharmap,revcharmap)
1954 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1955 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1957 These routines dump out a trie in a somewhat readable format.
1958 The _interim_ variants are used for debugging the interim
1959 tables that are used to generate the final compressed
1960 representation which is what dump_trie expects.
1962 Part of the reason for their existence is to provide a form
1963 of documentation as to how the different representations function.
1968 Dumps the final compressed table form of the trie to Perl_debug_log.
1969 Used for debugging make_trie().
1973 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1974 AV *revcharmap, U32 depth)
1977 SV *sv=sv_newmortal();
1978 int colwidth= widecharmap ? 6 : 4;
1980 GET_RE_DEBUG_FLAGS_DECL;
1982 PERL_ARGS_ASSERT_DUMP_TRIE;
1984 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
1985 depth+1, "Match","Base","Ofs" );
1987 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1988 SV ** const tmp = av_fetch( revcharmap, state, 0);
1990 Perl_re_printf( aTHX_ "%*s",
1992 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1993 PL_colors[0], PL_colors[1],
1994 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1995 PERL_PV_ESCAPE_FIRSTCHAR
2000 Perl_re_printf( aTHX_ "\n");
2001 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2003 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2004 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2005 Perl_re_printf( aTHX_ "\n");
2007 for( state = 1 ; state < trie->statecount ; state++ ) {
2008 const U32 base = trie->states[ state ].trans.base;
2010 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2012 if ( trie->states[ state ].wordnum ) {
2013 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2015 Perl_re_printf( aTHX_ "%6s", "" );
2018 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2023 while( ( base + ofs < trie->uniquecharcount ) ||
2024 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2025 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2029 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2031 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2032 if ( ( base + ofs >= trie->uniquecharcount )
2033 && ( base + ofs - trie->uniquecharcount
2035 && trie->trans[ base + ofs
2036 - trie->uniquecharcount ].check == state )
2038 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2039 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2042 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
2046 Perl_re_printf( aTHX_ "]");
2049 Perl_re_printf( aTHX_ "\n" );
2051 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2053 for (word=1; word <= trie->wordcount; word++) {
2054 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2055 (int)word, (int)(trie->wordinfo[word].prev),
2056 (int)(trie->wordinfo[word].len));
2058 Perl_re_printf( aTHX_ "\n" );
2061 Dumps a fully constructed but uncompressed trie in list form.
2062 List tries normally only are used for construction when the number of
2063 possible chars (trie->uniquecharcount) is very high.
2064 Used for debugging make_trie().
2067 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2068 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2072 SV *sv=sv_newmortal();
2073 int colwidth= widecharmap ? 6 : 4;
2074 GET_RE_DEBUG_FLAGS_DECL;
2076 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2078 /* print out the table precompression. */
2079 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2081 Perl_re_indentf( aTHX_ "%s",
2082 depth+1, "------:-----+-----------------\n" );
2084 for( state=1 ; state < next_alloc ; state ++ ) {
2087 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2088 depth+1, (UV)state );
2089 if ( ! trie->states[ state ].wordnum ) {
2090 Perl_re_printf( aTHX_ "%5s| ","");
2092 Perl_re_printf( aTHX_ "W%4x| ",
2093 trie->states[ state ].wordnum
2096 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2097 SV ** const tmp = av_fetch( revcharmap,
2098 TRIE_LIST_ITEM(state,charid).forid, 0);
2100 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2102 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2104 PL_colors[0], PL_colors[1],
2105 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2106 | PERL_PV_ESCAPE_FIRSTCHAR
2108 TRIE_LIST_ITEM(state,charid).forid,
2109 (UV)TRIE_LIST_ITEM(state,charid).newstate
2112 Perl_re_printf( aTHX_ "\n%*s| ",
2113 (int)((depth * 2) + 14), "");
2116 Perl_re_printf( aTHX_ "\n");
2121 Dumps a fully constructed but uncompressed trie in table form.
2122 This is the normal DFA style state transition table, with a few
2123 twists to facilitate compression later.
2124 Used for debugging make_trie().
2127 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2128 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2133 SV *sv=sv_newmortal();
2134 int colwidth= widecharmap ? 6 : 4;
2135 GET_RE_DEBUG_FLAGS_DECL;
2137 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2140 print out the table precompression so that we can do a visual check
2141 that they are identical.
2144 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2146 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2147 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2149 Perl_re_printf( aTHX_ "%*s",
2151 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2152 PL_colors[0], PL_colors[1],
2153 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2154 PERL_PV_ESCAPE_FIRSTCHAR
2160 Perl_re_printf( aTHX_ "\n");
2161 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2163 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2164 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2167 Perl_re_printf( aTHX_ "\n" );
2169 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2171 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2173 (UV)TRIE_NODENUM( state ) );
2175 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2176 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2178 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2180 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2182 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2183 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2184 (UV)trie->trans[ state ].check );
2186 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2187 (UV)trie->trans[ state ].check,
2188 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2196 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2197 startbranch: the first branch in the whole branch sequence
2198 first : start branch of sequence of branch-exact nodes.
2199 May be the same as startbranch
2200 last : Thing following the last branch.
2201 May be the same as tail.
2202 tail : item following the branch sequence
2203 count : words in the sequence
2204 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2205 depth : indent depth
2207 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2209 A trie is an N'ary tree where the branches are determined by digital
2210 decomposition of the key. IE, at the root node you look up the 1st character and
2211 follow that branch repeat until you find the end of the branches. Nodes can be
2212 marked as "accepting" meaning they represent a complete word. Eg:
2216 would convert into the following structure. Numbers represent states, letters
2217 following numbers represent valid transitions on the letter from that state, if
2218 the number is in square brackets it represents an accepting state, otherwise it
2219 will be in parenthesis.
2221 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2225 (1) +-i->(6)-+-s->[7]
2227 +-s->(3)-+-h->(4)-+-e->[5]
2229 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2231 This shows that when matching against the string 'hers' we will begin at state 1
2232 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2233 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2234 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2235 single traverse. We store a mapping from accepting to state to which word was
2236 matched, and then when we have multiple possibilities we try to complete the
2237 rest of the regex in the order in which they occurred in the alternation.
2239 The only prior NFA like behaviour that would be changed by the TRIE support is
2240 the silent ignoring of duplicate alternations which are of the form:
2242 / (DUPE|DUPE) X? (?{ ... }) Y /x
2244 Thus EVAL blocks following a trie may be called a different number of times with
2245 and without the optimisation. With the optimisations dupes will be silently
2246 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2247 the following demonstrates:
2249 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2251 which prints out 'word' three times, but
2253 'words'=~/(word|word|word)(?{ print $1 })S/
2255 which doesnt print it out at all. This is due to other optimisations kicking in.
2257 Example of what happens on a structural level:
2259 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2261 1: CURLYM[1] {1,32767}(18)
2272 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2273 and should turn into:
2275 1: CURLYM[1] {1,32767}(18)
2277 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2285 Cases where tail != last would be like /(?foo|bar)baz/:
2295 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2296 and would end up looking like:
2299 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2306 d = uvchr_to_utf8_flags(d, uv, 0);
2308 is the recommended Unicode-aware way of saying
2313 #define TRIE_STORE_REVCHAR(val) \
2316 SV *zlopp = newSV(UTF8_MAXBYTES); \
2317 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2318 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2319 SvCUR_set(zlopp, kapow - flrbbbbb); \
2322 av_push(revcharmap, zlopp); \
2324 char ooooff = (char)val; \
2325 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2329 /* This gets the next character from the input, folding it if not already
2331 #define TRIE_READ_CHAR STMT_START { \
2334 /* if it is UTF then it is either already folded, or does not need \
2336 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2338 else if (folder == PL_fold_latin1) { \
2339 /* This folder implies Unicode rules, which in the range expressible \
2340 * by not UTF is the lower case, with the two exceptions, one of \
2341 * which should have been taken care of before calling this */ \
2342 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2343 uvc = toLOWER_L1(*uc); \
2344 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2347 /* raw data, will be folded later if needed */ \
2355 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2356 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2357 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2358 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2359 TRIE_LIST_LEN( state ) = ging; \
2361 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2362 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2363 TRIE_LIST_CUR( state )++; \
2366 #define TRIE_LIST_NEW(state) STMT_START { \
2367 Newxz( trie->states[ state ].trans.list, \
2368 4, reg_trie_trans_le ); \
2369 TRIE_LIST_CUR( state ) = 1; \
2370 TRIE_LIST_LEN( state ) = 4; \
2373 #define TRIE_HANDLE_WORD(state) STMT_START { \
2374 U16 dupe= trie->states[ state ].wordnum; \
2375 regnode * const noper_next = regnext( noper ); \
2378 /* store the word for dumping */ \
2380 if (OP(noper) != NOTHING) \
2381 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2383 tmp = newSVpvn_utf8( "", 0, UTF ); \
2384 av_push( trie_words, tmp ); \
2388 trie->wordinfo[curword].prev = 0; \
2389 trie->wordinfo[curword].len = wordlen; \
2390 trie->wordinfo[curword].accept = state; \
2392 if ( noper_next < tail ) { \
2394 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2396 trie->jump[curword] = (U16)(noper_next - convert); \
2398 jumper = noper_next; \
2400 nextbranch= regnext(cur); \
2404 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2405 /* chain, so that when the bits of chain are later */\
2406 /* linked together, the dups appear in the chain */\
2407 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2408 trie->wordinfo[dupe].prev = curword; \
2410 /* we haven't inserted this word yet. */ \
2411 trie->states[ state ].wordnum = curword; \
2416 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2417 ( ( base + charid >= ucharcount \
2418 && base + charid < ubound \
2419 && state == trie->trans[ base - ucharcount + charid ].check \
2420 && trie->trans[ base - ucharcount + charid ].next ) \
2421 ? trie->trans[ base - ucharcount + charid ].next \
2422 : ( state==1 ? special : 0 ) \
2425 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2427 TRIE_BITMAP_SET(trie, uvc); \
2428 /* store the folded codepoint */ \
2430 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2433 /* store first byte of utf8 representation of */ \
2434 /* variant codepoints */ \
2435 if (! UVCHR_IS_INVARIANT(uvc)) { \
2436 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2441 #define MADE_JUMP_TRIE 2
2442 #define MADE_EXACT_TRIE 4
2445 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2446 regnode *first, regnode *last, regnode *tail,
2447 U32 word_count, U32 flags, U32 depth)
2449 /* first pass, loop through and scan words */
2450 reg_trie_data *trie;
2451 HV *widecharmap = NULL;
2452 AV *revcharmap = newAV();
2458 regnode *jumper = NULL;
2459 regnode *nextbranch = NULL;
2460 regnode *convert = NULL;
2461 U32 *prev_states; /* temp array mapping each state to previous one */
2462 /* we just use folder as a flag in utf8 */
2463 const U8 * folder = NULL;
2466 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2467 AV *trie_words = NULL;
2468 /* along with revcharmap, this only used during construction but both are
2469 * useful during debugging so we store them in the struct when debugging.
2472 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2473 STRLEN trie_charcount=0;
2475 SV *re_trie_maxbuff;
2476 GET_RE_DEBUG_FLAGS_DECL;
2478 PERL_ARGS_ASSERT_MAKE_TRIE;
2480 PERL_UNUSED_ARG(depth);
2484 case EXACT: case EXACTL: break;
2488 case EXACTFLU8: folder = PL_fold_latin1; break;
2489 case EXACTF: folder = PL_fold; break;
2490 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2493 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2495 trie->startstate = 1;
2496 trie->wordcount = word_count;
2497 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2498 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2499 if (flags == EXACT || flags == EXACTL)
2500 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2501 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2502 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2505 trie_words = newAV();
2508 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2509 assert(re_trie_maxbuff);
2510 if (!SvIOK(re_trie_maxbuff)) {
2511 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2513 DEBUG_TRIE_COMPILE_r({
2514 Perl_re_indentf( aTHX_
2515 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2517 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2518 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2521 /* Find the node we are going to overwrite */
2522 if ( first == startbranch && OP( last ) != BRANCH ) {
2523 /* whole branch chain */
2526 /* branch sub-chain */
2527 convert = NEXTOPER( first );
2530 /* -- First loop and Setup --
2532 We first traverse the branches and scan each word to determine if it
2533 contains widechars, and how many unique chars there are, this is
2534 important as we have to build a table with at least as many columns as we
2537 We use an array of integers to represent the character codes 0..255
2538 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2539 the native representation of the character value as the key and IV's for
2542 *TODO* If we keep track of how many times each character is used we can
2543 remap the columns so that the table compression later on is more
2544 efficient in terms of memory by ensuring the most common value is in the
2545 middle and the least common are on the outside. IMO this would be better
2546 than a most to least common mapping as theres a decent chance the most
2547 common letter will share a node with the least common, meaning the node
2548 will not be compressible. With a middle is most common approach the worst
2549 case is when we have the least common nodes twice.
2553 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2554 regnode *noper = NEXTOPER( cur );
2558 U32 wordlen = 0; /* required init */
2559 STRLEN minchars = 0;
2560 STRLEN maxchars = 0;
2561 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2564 if (OP(noper) == NOTHING) {
2565 /* skip past a NOTHING at the start of an alternation
2566 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2568 regnode *noper_next= regnext(noper);
2569 if (noper_next < tail)
2573 if ( noper < tail &&
2575 OP(noper) == flags ||
2578 OP(noper) == EXACTFU_SS
2582 uc= (U8*)STRING(noper);
2583 e= uc + STR_LEN(noper);
2590 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2591 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2592 regardless of encoding */
2593 if (OP( noper ) == EXACTFU_SS) {
2594 /* false positives are ok, so just set this */
2595 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2599 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2601 TRIE_CHARCOUNT(trie)++;
2604 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2605 * is in effect. Under /i, this character can match itself, or
2606 * anything that folds to it. If not under /i, it can match just
2607 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2608 * all fold to k, and all are single characters. But some folds
2609 * expand to more than one character, so for example LATIN SMALL
2610 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2611 * the string beginning at 'uc' is 'ffi', it could be matched by
2612 * three characters, or just by the one ligature character. (It
2613 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2614 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2615 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2616 * match.) The trie needs to know the minimum and maximum number
2617 * of characters that could match so that it can use size alone to
2618 * quickly reject many match attempts. The max is simple: it is
2619 * the number of folded characters in this branch (since a fold is
2620 * never shorter than what folds to it. */
2624 /* And the min is equal to the max if not under /i (indicated by
2625 * 'folder' being NULL), or there are no multi-character folds. If
2626 * there is a multi-character fold, the min is incremented just
2627 * once, for the character that folds to the sequence. Each
2628 * character in the sequence needs to be added to the list below of
2629 * characters in the trie, but we count only the first towards the
2630 * min number of characters needed. This is done through the
2631 * variable 'foldlen', which is returned by the macros that look
2632 * for these sequences as the number of bytes the sequence
2633 * occupies. Each time through the loop, we decrement 'foldlen' by
2634 * how many bytes the current char occupies. Only when it reaches
2635 * 0 do we increment 'minchars' or look for another multi-character
2637 if (folder == NULL) {
2640 else if (foldlen > 0) {
2641 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2646 /* See if *uc is the beginning of a multi-character fold. If
2647 * so, we decrement the length remaining to look at, to account
2648 * for the current character this iteration. (We can use 'uc'
2649 * instead of the fold returned by TRIE_READ_CHAR because for
2650 * non-UTF, the latin1_safe macro is smart enough to account
2651 * for all the unfolded characters, and because for UTF, the
2652 * string will already have been folded earlier in the
2653 * compilation process */
2655 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2656 foldlen -= UTF8SKIP(uc);
2659 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2664 /* The current character (and any potential folds) should be added
2665 * to the possible matching characters for this position in this
2669 U8 folded= folder[ (U8) uvc ];
2670 if ( !trie->charmap[ folded ] ) {
2671 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2672 TRIE_STORE_REVCHAR( folded );
2675 if ( !trie->charmap[ uvc ] ) {
2676 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2677 TRIE_STORE_REVCHAR( uvc );
2680 /* store the codepoint in the bitmap, and its folded
2682 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2683 set_bit = 0; /* We've done our bit :-) */
2687 /* XXX We could come up with the list of code points that fold
2688 * to this using PL_utf8_foldclosures, except not for
2689 * multi-char folds, as there may be multiple combinations
2690 * there that could work, which needs to wait until runtime to
2691 * resolve (The comment about LIGATURE FFI above is such an
2696 widecharmap = newHV();
2698 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2701 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2703 if ( !SvTRUE( *svpp ) ) {
2704 sv_setiv( *svpp, ++trie->uniquecharcount );
2705 TRIE_STORE_REVCHAR(uvc);
2708 } /* end loop through characters in this branch of the trie */
2710 /* We take the min and max for this branch and combine to find the min
2711 * and max for all branches processed so far */
2712 if( cur == first ) {
2713 trie->minlen = minchars;
2714 trie->maxlen = maxchars;
2715 } else if (minchars < trie->minlen) {
2716 trie->minlen = minchars;
2717 } else if (maxchars > trie->maxlen) {
2718 trie->maxlen = maxchars;
2720 } /* end first pass */
2721 DEBUG_TRIE_COMPILE_r(
2722 Perl_re_indentf( aTHX_
2723 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2725 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2726 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2727 (int)trie->minlen, (int)trie->maxlen )
2731 We now know what we are dealing with in terms of unique chars and
2732 string sizes so we can calculate how much memory a naive
2733 representation using a flat table will take. If it's over a reasonable
2734 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2735 conservative but potentially much slower representation using an array
2738 At the end we convert both representations into the same compressed
2739 form that will be used in regexec.c for matching with. The latter
2740 is a form that cannot be used to construct with but has memory
2741 properties similar to the list form and access properties similar
2742 to the table form making it both suitable for fast searches and
2743 small enough that its feasable to store for the duration of a program.
2745 See the comment in the code where the compressed table is produced
2746 inplace from the flat tabe representation for an explanation of how
2747 the compression works.
2752 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2755 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2756 > SvIV(re_trie_maxbuff) )
2759 Second Pass -- Array Of Lists Representation
2761 Each state will be represented by a list of charid:state records
2762 (reg_trie_trans_le) the first such element holds the CUR and LEN
2763 points of the allocated array. (See defines above).
2765 We build the initial structure using the lists, and then convert
2766 it into the compressed table form which allows faster lookups
2767 (but cant be modified once converted).
2770 STRLEN transcount = 1;
2772 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2775 trie->states = (reg_trie_state *)
2776 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2777 sizeof(reg_trie_state) );
2781 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2783 regnode *noper = NEXTOPER( cur );
2784 U32 state = 1; /* required init */
2785 U16 charid = 0; /* sanity init */
2786 U32 wordlen = 0; /* required init */
2788 if (OP(noper) == NOTHING) {
2789 regnode *noper_next= regnext(noper);
2790 if (noper_next < tail)
2794 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2795 const U8 *uc= (U8*)STRING(noper);
2796 const U8 *e= uc + STR_LEN(noper);
2798 for ( ; uc < e ; uc += len ) {
2803 charid = trie->charmap[ uvc ];
2805 SV** const svpp = hv_fetch( widecharmap,
2812 charid=(U16)SvIV( *svpp );
2815 /* charid is now 0 if we dont know the char read, or
2816 * nonzero if we do */
2823 if ( !trie->states[ state ].trans.list ) {
2824 TRIE_LIST_NEW( state );
2827 check <= TRIE_LIST_USED( state );
2830 if ( TRIE_LIST_ITEM( state, check ).forid
2833 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2838 newstate = next_alloc++;
2839 prev_states[newstate] = state;
2840 TRIE_LIST_PUSH( state, charid, newstate );
2845 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
2849 TRIE_HANDLE_WORD(state);
2851 } /* end second pass */
2853 /* next alloc is the NEXT state to be allocated */
2854 trie->statecount = next_alloc;
2855 trie->states = (reg_trie_state *)
2856 PerlMemShared_realloc( trie->states,
2858 * sizeof(reg_trie_state) );
2860 /* and now dump it out before we compress it */
2861 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2862 revcharmap, next_alloc,
2866 trie->trans = (reg_trie_trans *)
2867 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2874 for( state=1 ; state < next_alloc ; state ++ ) {
2878 DEBUG_TRIE_COMPILE_MORE_r(
2879 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
2883 if (trie->states[state].trans.list) {
2884 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2888 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2889 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2890 if ( forid < minid ) {
2892 } else if ( forid > maxid ) {
2896 if ( transcount < tp + maxid - minid + 1) {
2898 trie->trans = (reg_trie_trans *)
2899 PerlMemShared_realloc( trie->trans,
2901 * sizeof(reg_trie_trans) );
2902 Zero( trie->trans + (transcount / 2),
2906 base = trie->uniquecharcount + tp - minid;
2907 if ( maxid == minid ) {
2909 for ( ; zp < tp ; zp++ ) {
2910 if ( ! trie->trans[ zp ].next ) {
2911 base = trie->uniquecharcount + zp - minid;
2912 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2914 trie->trans[ zp ].check = state;
2920 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2922 trie->trans[ tp ].check = state;
2927 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2928 const U32 tid = base
2929 - trie->uniquecharcount
2930 + TRIE_LIST_ITEM( state, idx ).forid;
2931 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2933 trie->trans[ tid ].check = state;
2935 tp += ( maxid - minid + 1 );
2937 Safefree(trie->states[ state ].trans.list);
2940 DEBUG_TRIE_COMPILE_MORE_r(
2941 Perl_re_printf( aTHX_ " base: %d\n",base);
2944 trie->states[ state ].trans.base=base;
2946 trie->lasttrans = tp + 1;
2950 Second Pass -- Flat Table Representation.
2952 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2953 each. We know that we will need Charcount+1 trans at most to store
2954 the data (one row per char at worst case) So we preallocate both
2955 structures assuming worst case.
2957 We then construct the trie using only the .next slots of the entry
2960 We use the .check field of the first entry of the node temporarily
2961 to make compression both faster and easier by keeping track of how
2962 many non zero fields are in the node.
2964 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2967 There are two terms at use here: state as a TRIE_NODEIDX() which is
2968 a number representing the first entry of the node, and state as a
2969 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2970 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2971 if there are 2 entrys per node. eg:
2979 The table is internally in the right hand, idx form. However as we
2980 also have to deal with the states array which is indexed by nodenum
2981 we have to use TRIE_NODENUM() to convert.
2984 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
2987 trie->trans = (reg_trie_trans *)
2988 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2989 * trie->uniquecharcount + 1,
2990 sizeof(reg_trie_trans) );
2991 trie->states = (reg_trie_state *)
2992 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2993 sizeof(reg_trie_state) );
2994 next_alloc = trie->uniquecharcount + 1;
2997 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2999 regnode *noper = NEXTOPER( cur );
3001 U32 state = 1; /* required init */
3003 U16 charid = 0; /* sanity init */
3004 U32 accept_state = 0; /* sanity init */
3006 U32 wordlen = 0; /* required init */
3008 if (OP(noper) == NOTHING) {
3009 regnode *noper_next= regnext(noper);
3010 if (noper_next < tail)
3014 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3015 const U8 *uc= (U8*)STRING(noper);
3016 const U8 *e= uc + STR_LEN(noper);
3018 for ( ; uc < e ; uc += len ) {
3023 charid = trie->charmap[ uvc ];
3025 SV* const * const svpp = hv_fetch( widecharmap,
3029 charid = svpp ? (U16)SvIV(*svpp) : 0;
3033 if ( !trie->trans[ state + charid ].next ) {
3034 trie->trans[ state + charid ].next = next_alloc;
3035 trie->trans[ state ].check++;
3036 prev_states[TRIE_NODENUM(next_alloc)]
3037 = TRIE_NODENUM(state);
3038 next_alloc += trie->uniquecharcount;
3040 state = trie->trans[ state + charid ].next;
3042 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3044 /* charid is now 0 if we dont know the char read, or
3045 * nonzero if we do */
3048 accept_state = TRIE_NODENUM( state );
3049 TRIE_HANDLE_WORD(accept_state);
3051 } /* end second pass */
3053 /* and now dump it out before we compress it */
3054 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3056 next_alloc, depth+1));
3060 * Inplace compress the table.*
3062 For sparse data sets the table constructed by the trie algorithm will
3063 be mostly 0/FAIL transitions or to put it another way mostly empty.
3064 (Note that leaf nodes will not contain any transitions.)
3066 This algorithm compresses the tables by eliminating most such
3067 transitions, at the cost of a modest bit of extra work during lookup:
3069 - Each states[] entry contains a .base field which indicates the
3070 index in the state[] array wheres its transition data is stored.
3072 - If .base is 0 there are no valid transitions from that node.
3074 - If .base is nonzero then charid is added to it to find an entry in
3077 -If trans[states[state].base+charid].check!=state then the
3078 transition is taken to be a 0/Fail transition. Thus if there are fail
3079 transitions at the front of the node then the .base offset will point
3080 somewhere inside the previous nodes data (or maybe even into a node
3081 even earlier), but the .check field determines if the transition is
3085 The following process inplace converts the table to the compressed
3086 table: We first do not compress the root node 1,and mark all its
3087 .check pointers as 1 and set its .base pointer as 1 as well. This
3088 allows us to do a DFA construction from the compressed table later,
3089 and ensures that any .base pointers we calculate later are greater
3092 - We set 'pos' to indicate the first entry of the second node.
3094 - We then iterate over the columns of the node, finding the first and
3095 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3096 and set the .check pointers accordingly, and advance pos
3097 appropriately and repreat for the next node. Note that when we copy
3098 the next pointers we have to convert them from the original
3099 NODEIDX form to NODENUM form as the former is not valid post
3102 - If a node has no transitions used we mark its base as 0 and do not
3103 advance the pos pointer.
3105 - If a node only has one transition we use a second pointer into the
3106 structure to fill in allocated fail transitions from other states.
3107 This pointer is independent of the main pointer and scans forward
3108 looking for null transitions that are allocated to a state. When it
3109 finds one it writes the single transition into the "hole". If the
3110 pointer doesnt find one the single transition is appended as normal.
3112 - Once compressed we can Renew/realloc the structures to release the
3115 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3116 specifically Fig 3.47 and the associated pseudocode.
3120 const U32 laststate = TRIE_NODENUM( next_alloc );
3123 trie->statecount = laststate;
3125 for ( state = 1 ; state < laststate ; state++ ) {
3127 const U32 stateidx = TRIE_NODEIDX( state );
3128 const U32 o_used = trie->trans[ stateidx ].check;
3129 U32 used = trie->trans[ stateidx ].check;
3130 trie->trans[ stateidx ].check = 0;
3133 used && charid < trie->uniquecharcount;
3136 if ( flag || trie->trans[ stateidx + charid ].next ) {
3137 if ( trie->trans[ stateidx + charid ].next ) {
3139 for ( ; zp < pos ; zp++ ) {
3140 if ( ! trie->trans[ zp ].next ) {
3144 trie->states[ state ].trans.base
3146 + trie->uniquecharcount
3148 trie->trans[ zp ].next
3149 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3151 trie->trans[ zp ].check = state;
3152 if ( ++zp > pos ) pos = zp;
3159 trie->states[ state ].trans.base
3160 = pos + trie->uniquecharcount - charid ;
3162 trie->trans[ pos ].next
3163 = SAFE_TRIE_NODENUM(
3164 trie->trans[ stateidx + charid ].next );
3165 trie->trans[ pos ].check = state;
3170 trie->lasttrans = pos + 1;
3171 trie->states = (reg_trie_state *)
3172 PerlMemShared_realloc( trie->states, laststate
3173 * sizeof(reg_trie_state) );
3174 DEBUG_TRIE_COMPILE_MORE_r(
3175 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3177 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3181 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3184 } /* end table compress */
3186 DEBUG_TRIE_COMPILE_MORE_r(
3187 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3189 (UV)trie->statecount,
3190 (UV)trie->lasttrans)
3192 /* resize the trans array to remove unused space */
3193 trie->trans = (reg_trie_trans *)
3194 PerlMemShared_realloc( trie->trans, trie->lasttrans
3195 * sizeof(reg_trie_trans) );
3197 { /* Modify the program and insert the new TRIE node */
3198 U8 nodetype =(U8)(flags & 0xFF);
3202 regnode *optimize = NULL;
3203 #ifdef RE_TRACK_PATTERN_OFFSETS
3206 U32 mjd_nodelen = 0;
3207 #endif /* RE_TRACK_PATTERN_OFFSETS */
3208 #endif /* DEBUGGING */
3210 This means we convert either the first branch or the first Exact,
3211 depending on whether the thing following (in 'last') is a branch
3212 or not and whther first is the startbranch (ie is it a sub part of
3213 the alternation or is it the whole thing.)
3214 Assuming its a sub part we convert the EXACT otherwise we convert
3215 the whole branch sequence, including the first.
3217 /* Find the node we are going to overwrite */
3218 if ( first != startbranch || OP( last ) == BRANCH ) {
3219 /* branch sub-chain */
3220 NEXT_OFF( first ) = (U16)(last - first);
3221 #ifdef RE_TRACK_PATTERN_OFFSETS
3223 mjd_offset= Node_Offset((convert));
3224 mjd_nodelen= Node_Length((convert));
3227 /* whole branch chain */
3229 #ifdef RE_TRACK_PATTERN_OFFSETS
3232 const regnode *nop = NEXTOPER( convert );
3233 mjd_offset= Node_Offset((nop));
3234 mjd_nodelen= Node_Length((nop));
3238 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3240 (UV)mjd_offset, (UV)mjd_nodelen)
3243 /* But first we check to see if there is a common prefix we can
3244 split out as an EXACT and put in front of the TRIE node. */
3245 trie->startstate= 1;
3246 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3247 /* we want to find the first state that has more than
3248 * one transition, if that state is not the first state
3249 * then we have a common prefix which we can remove.
3252 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3254 I32 first_ofs = -1; /* keeps track of the ofs of the first
3255 transition, -1 means none */
3257 const U32 base = trie->states[ state ].trans.base;
3259 /* does this state terminate an alternation? */
3260 if ( trie->states[state].wordnum )
3263 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3264 if ( ( base + ofs >= trie->uniquecharcount ) &&
3265 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3266 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3268 if ( ++count > 1 ) {
3269 /* we have more than one transition */
3272 /* if this is the first state there is no common prefix
3273 * to extract, so we can exit */
3274 if ( state == 1 ) break;
3275 tmp = av_fetch( revcharmap, ofs, 0);
3276 ch = (U8*)SvPV_nolen_const( *tmp );
3278 /* if we are on count 2 then we need to initialize the
3279 * bitmap, and store the previous char if there was one
3282 /* clear the bitmap */
3283 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3285 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3288 if (first_ofs >= 0) {
3289 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3290 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3292 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3294 Perl_re_printf( aTHX_ "%s", (char*)ch)
3298 /* store the current firstchar in the bitmap */
3299 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3300 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3306 /* This state has only one transition, its transition is part
3307 * of a common prefix - we need to concatenate the char it
3308 * represents to what we have so far. */
3309 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3311 char *ch = SvPV( *tmp, len );
3313 SV *sv=sv_newmortal();
3314 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3316 (UV)state, (UV)first_ofs,
3317 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3318 PL_colors[0], PL_colors[1],
3319 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3320 PERL_PV_ESCAPE_FIRSTCHAR
3325 OP( convert ) = nodetype;
3326 str=STRING(convert);
3329 STR_LEN(convert) += len;
3335 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3340 trie->prefixlen = (state-1);
3342 regnode *n = convert+NODE_SZ_STR(convert);
3343 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3344 trie->startstate = state;
3345 trie->minlen -= (state - 1);
3346 trie->maxlen -= (state - 1);
3348 /* At least the UNICOS C compiler choked on this
3349 * being argument to DEBUG_r(), so let's just have
3352 #ifdef PERL_EXT_RE_BUILD
3358 regnode *fix = convert;
3359 U32 word = trie->wordcount;
3361 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3362 while( ++fix < n ) {
3363 Set_Node_Offset_Length(fix, 0, 0);
3366 SV ** const tmp = av_fetch( trie_words, word, 0 );
3368 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3369 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3371 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3379 NEXT_OFF(convert) = (U16)(tail - convert);
3380 DEBUG_r(optimize= n);
3386 if ( trie->maxlen ) {
3387 NEXT_OFF( convert ) = (U16)(tail - convert);
3388 ARG_SET( convert, data_slot );
3389 /* Store the offset to the first unabsorbed branch in
3390 jump[0], which is otherwise unused by the jump logic.
3391 We use this when dumping a trie and during optimisation. */
3393 trie->jump[0] = (U16)(nextbranch - convert);
3395 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3396 * and there is a bitmap
3397 * and the first "jump target" node we found leaves enough room
3398 * then convert the TRIE node into a TRIEC node, with the bitmap
3399 * embedded inline in the opcode - this is hypothetically faster.
3401 if ( !trie->states[trie->startstate].wordnum
3403 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3405 OP( convert ) = TRIEC;
3406 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3407 PerlMemShared_free(trie->bitmap);
3410 OP( convert ) = TRIE;
3412 /* store the type in the flags */
3413 convert->flags = nodetype;
3417 + regarglen[ OP( convert ) ];
3419 /* XXX We really should free up the resource in trie now,
3420 as we won't use them - (which resources?) dmq */
3422 /* needed for dumping*/
3423 DEBUG_r(if (optimize) {
3424 regnode *opt = convert;
3426 while ( ++opt < optimize) {
3427 Set_Node_Offset_Length(opt,0,0);
3430 Try to clean up some of the debris left after the
3433 while( optimize < jumper ) {
3434 mjd_nodelen += Node_Length((optimize));
3435 OP( optimize ) = OPTIMIZED;
3436 Set_Node_Offset_Length(optimize,0,0);
3439 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3441 } /* end node insert */
3443 /* Finish populating the prev field of the wordinfo array. Walk back
3444 * from each accept state until we find another accept state, and if
3445 * so, point the first word's .prev field at the second word. If the
3446 * second already has a .prev field set, stop now. This will be the
3447 * case either if we've already processed that word's accept state,
3448 * or that state had multiple words, and the overspill words were
3449 * already linked up earlier.
3456 for (word=1; word <= trie->wordcount; word++) {
3458 if (trie->wordinfo[word].prev)
3460 state = trie->wordinfo[word].accept;
3462 state = prev_states[state];
3465 prev = trie->states[state].wordnum;
3469 trie->wordinfo[word].prev = prev;
3471 Safefree(prev_states);
3475 /* and now dump out the compressed format */
3476 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3478 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3480 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3481 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3483 SvREFCNT_dec_NN(revcharmap);
3487 : trie->startstate>1
3493 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3495 /* The Trie is constructed and compressed now so we can build a fail array if
3498 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3500 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3504 We find the fail state for each state in the trie, this state is the longest
3505 proper suffix of the current state's 'word' that is also a proper prefix of
3506 another word in our trie. State 1 represents the word '' and is thus the
3507 default fail state. This allows the DFA not to have to restart after its
3508 tried and failed a word at a given point, it simply continues as though it
3509 had been matching the other word in the first place.
3511 'abcdgu'=~/abcdefg|cdgu/
3512 When we get to 'd' we are still matching the first word, we would encounter
3513 'g' which would fail, which would bring us to the state representing 'd' in
3514 the second word where we would try 'g' and succeed, proceeding to match
3517 /* add a fail transition */
3518 const U32 trie_offset = ARG(source);
3519 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3521 const U32 ucharcount = trie->uniquecharcount;
3522 const U32 numstates = trie->statecount;
3523 const U32 ubound = trie->lasttrans + ucharcount;
3527 U32 base = trie->states[ 1 ].trans.base;
3530 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3532 GET_RE_DEBUG_FLAGS_DECL;
3534 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3535 PERL_UNUSED_CONTEXT;
3537 PERL_UNUSED_ARG(depth);
3540 if ( OP(source) == TRIE ) {
3541 struct regnode_1 *op = (struct regnode_1 *)
3542 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3543 StructCopy(source,op,struct regnode_1);
3544 stclass = (regnode *)op;
3546 struct regnode_charclass *op = (struct regnode_charclass *)
3547 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3548 StructCopy(source,op,struct regnode_charclass);
3549 stclass = (regnode *)op;
3551 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3553 ARG_SET( stclass, data_slot );
3554 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3555 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3556 aho->trie=trie_offset;
3557 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3558 Copy( trie->states, aho->states, numstates, reg_trie_state );
3559 Newxz( q, numstates, U32);
3560 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3563 /* initialize fail[0..1] to be 1 so that we always have
3564 a valid final fail state */
3565 fail[ 0 ] = fail[ 1 ] = 1;
3567 for ( charid = 0; charid < ucharcount ; charid++ ) {
3568 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3570 q[ q_write ] = newstate;
3571 /* set to point at the root */
3572 fail[ q[ q_write++ ] ]=1;
3575 while ( q_read < q_write) {
3576 const U32 cur = q[ q_read++ % numstates ];
3577 base = trie->states[ cur ].trans.base;
3579 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3580 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3582 U32 fail_state = cur;
3585 fail_state = fail[ fail_state ];
3586 fail_base = aho->states[ fail_state ].trans.base;
3587 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3589 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3590 fail[ ch_state ] = fail_state;
3591 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3593 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3595 q[ q_write++ % numstates] = ch_state;
3599 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3600 when we fail in state 1, this allows us to use the
3601 charclass scan to find a valid start char. This is based on the principle
3602 that theres a good chance the string being searched contains lots of stuff
3603 that cant be a start char.
3605 fail[ 0 ] = fail[ 1 ] = 0;
3606 DEBUG_TRIE_COMPILE_r({
3607 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3608 depth, (UV)numstates
3610 for( q_read=1; q_read<numstates; q_read++ ) {
3611 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3613 Perl_re_printf( aTHX_ "\n");
3616 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3621 #define DEBUG_PEEP(str,scan,depth) \
3622 DEBUG_OPTIMISE_r({if (scan){ \
3623 regnode *Next = regnext(scan); \
3624 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);\
3625 Perl_re_indentf( aTHX_ "" str ">%3d: %s (%d)", \
3626 depth, REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3627 Next ? (REG_NODE_NUM(Next)) : 0 );\
3628 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3629 Perl_re_printf( aTHX_ "\n"); \
3632 /* The below joins as many adjacent EXACTish nodes as possible into a single
3633 * one. The regop may be changed if the node(s) contain certain sequences that
3634 * require special handling. The joining is only done if:
3635 * 1) there is room in the current conglomerated node to entirely contain the
3637 * 2) they are the exact same node type
3639 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3640 * these get optimized out
3642 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3643 * as possible, even if that means splitting an existing node so that its first
3644 * part is moved to the preceeding node. This would maximise the efficiency of
3645 * memEQ during matching. Elsewhere in this file, khw proposes splitting
3646 * EXACTFish nodes into portions that don't change under folding vs those that
3647 * do. Those portions that don't change may be the only things in the pattern that
3648 * could be used to find fixed and floating strings.
3650 * If a node is to match under /i (folded), the number of characters it matches
3651 * can be different than its character length if it contains a multi-character
3652 * fold. *min_subtract is set to the total delta number of characters of the
3655 * And *unfolded_multi_char is set to indicate whether or not the node contains
3656 * an unfolded multi-char fold. This happens when whether the fold is valid or
3657 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3658 * SMALL LETTER SHARP S, as only if the target string being matched against
3659 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3660 * folding rules depend on the locale in force at runtime. (Multi-char folds
3661 * whose components are all above the Latin1 range are not run-time locale
3662 * dependent, and have already been folded by the time this function is
3665 * This is as good a place as any to discuss the design of handling these
3666 * multi-character fold sequences. It's been wrong in Perl for a very long
3667 * time. There are three code points in Unicode whose multi-character folds
3668 * were long ago discovered to mess things up. The previous designs for
3669 * dealing with these involved assigning a special node for them. This
3670 * approach doesn't always work, as evidenced by this example:
3671 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3672 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3673 * would match just the \xDF, it won't be able to handle the case where a
3674 * successful match would have to cross the node's boundary. The new approach
3675 * that hopefully generally solves the problem generates an EXACTFU_SS node
3676 * that is "sss" in this case.
3678 * It turns out that there are problems with all multi-character folds, and not
3679 * just these three. Now the code is general, for all such cases. The
3680 * approach taken is:
3681 * 1) This routine examines each EXACTFish node that could contain multi-
3682 * character folded sequences. Since a single character can fold into
3683 * such a sequence, the minimum match length for this node is less than
3684 * the number of characters in the node. This routine returns in
3685 * *min_subtract how many characters to subtract from the the actual
3686 * length of the string to get a real minimum match length; it is 0 if
3687 * there are no multi-char foldeds. This delta is used by the caller to
3688 * adjust the min length of the match, and the delta between min and max,
3689 * so that the optimizer doesn't reject these possibilities based on size
3691 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3692 * is used for an EXACTFU node that contains at least one "ss" sequence in
3693 * it. For non-UTF-8 patterns and strings, this is the only case where
3694 * there is a possible fold length change. That means that a regular
3695 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3696 * with length changes, and so can be processed faster. regexec.c takes
3697 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3698 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3699 * known until runtime). This saves effort in regex matching. However,
3700 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3701 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3702 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3703 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3704 * possibilities for the non-UTF8 patterns are quite simple, except for
3705 * the sharp s. All the ones that don't involve a UTF-8 target string are
3706 * members of a fold-pair, and arrays are set up for all of them so that
3707 * the other member of the pair can be found quickly. Code elsewhere in
3708 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3709 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3710 * described in the next item.
3711 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3712 * validity of the fold won't be known until runtime, and so must remain
3713 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3714 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3715 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3716 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3717 * The reason this is a problem is that the optimizer part of regexec.c
3718 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3719 * that a character in the pattern corresponds to at most a single
3720 * character in the target string. (And I do mean character, and not byte
3721 * here, unlike other parts of the documentation that have never been
3722 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3723 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3724 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3725 * nodes, violate the assumption, and they are the only instances where it
3726 * is violated. I'm reluctant to try to change the assumption, as the
3727 * code involved is impenetrable to me (khw), so instead the code here
3728 * punts. This routine examines EXACTFL nodes, and (when the pattern
3729 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3730 * boolean indicating whether or not the node contains such a fold. When
3731 * it is true, the caller sets a flag that later causes the optimizer in
3732 * this file to not set values for the floating and fixed string lengths,
3733 * and thus avoids the optimizer code in regexec.c that makes the invalid
3734 * assumption. Thus, there is no optimization based on string lengths for
3735 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3736 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3737 * assumption is wrong only in these cases is that all other non-UTF-8
3738 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3739 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3740 * EXACTF nodes because we don't know at compile time if it actually
3741 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3742 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3743 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3744 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3745 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3746 * string would require the pattern to be forced into UTF-8, the overhead
3747 * of which we want to avoid. Similarly the unfolded multi-char folds in
3748 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3751 * Similarly, the code that generates tries doesn't currently handle
3752 * not-already-folded multi-char folds, and it looks like a pain to change
3753 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3754 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3755 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3756 * using /iaa matching will be doing so almost entirely with ASCII
3757 * strings, so this should rarely be encountered in practice */
3759 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3760 if (PL_regkind[OP(scan)] == EXACT) \
3761 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3764 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3765 UV *min_subtract, bool *unfolded_multi_char,
3766 U32 flags,regnode *val, U32 depth)
3768 /* Merge several consecutive EXACTish nodes into one. */
3769 regnode *n = regnext(scan);
3771 regnode *next = scan + NODE_SZ_STR(scan);
3775 regnode *stop = scan;
3776 GET_RE_DEBUG_FLAGS_DECL;
3778 PERL_UNUSED_ARG(depth);
3781 PERL_ARGS_ASSERT_JOIN_EXACT;
3782 #ifndef EXPERIMENTAL_INPLACESCAN
3783 PERL_UNUSED_ARG(flags);
3784 PERL_UNUSED_ARG(val);
3786 DEBUG_PEEP("join",scan,depth);
3788 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3789 * EXACT ones that are mergeable to the current one. */
3791 && (PL_regkind[OP(n)] == NOTHING
3792 || (stringok && OP(n) == OP(scan)))
3794 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3797 if (OP(n) == TAIL || n > next)
3799 if (PL_regkind[OP(n)] == NOTHING) {
3800 DEBUG_PEEP("skip:",n,depth);
3801 NEXT_OFF(scan) += NEXT_OFF(n);
3802 next = n + NODE_STEP_REGNODE;
3809 else if (stringok) {
3810 const unsigned int oldl = STR_LEN(scan);
3811 regnode * const nnext = regnext(n);
3813 /* XXX I (khw) kind of doubt that this works on platforms (should
3814 * Perl ever run on one) where U8_MAX is above 255 because of lots
3815 * of other assumptions */
3816 /* Don't join if the sum can't fit into a single node */
3817 if (oldl + STR_LEN(n) > U8_MAX)
3820 DEBUG_PEEP("merg",n,depth);
3823 NEXT_OFF(scan) += NEXT_OFF(n);
3824 STR_LEN(scan) += STR_LEN(n);
3825 next = n + NODE_SZ_STR(n);
3826 /* Now we can overwrite *n : */
3827 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3835 #ifdef EXPERIMENTAL_INPLACESCAN
3836 if (flags && !NEXT_OFF(n)) {
3837 DEBUG_PEEP("atch", val, depth);
3838 if (reg_off_by_arg[OP(n)]) {
3839 ARG_SET(n, val - n);
3842 NEXT_OFF(n) = val - n;
3850 *unfolded_multi_char = FALSE;
3852 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3853 * can now analyze for sequences of problematic code points. (Prior to
3854 * this final joining, sequences could have been split over boundaries, and
3855 * hence missed). The sequences only happen in folding, hence for any
3856 * non-EXACT EXACTish node */
3857 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3858 U8* s0 = (U8*) STRING(scan);
3860 U8* s_end = s0 + STR_LEN(scan);
3862 int total_count_delta = 0; /* Total delta number of characters that
3863 multi-char folds expand to */
3865 /* One pass is made over the node's string looking for all the
3866 * possibilities. To avoid some tests in the loop, there are two main
3867 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3872 if (OP(scan) == EXACTFL) {
3875 /* An EXACTFL node would already have been changed to another
3876 * node type unless there is at least one character in it that
3877 * is problematic; likely a character whose fold definition
3878 * won't be known until runtime, and so has yet to be folded.
3879 * For all but the UTF-8 locale, folds are 1-1 in length, but
3880 * to handle the UTF-8 case, we need to create a temporary
3881 * folded copy using UTF-8 locale rules in order to analyze it.
3882 * This is because our macros that look to see if a sequence is
3883 * a multi-char fold assume everything is folded (otherwise the
3884 * tests in those macros would be too complicated and slow).
3885 * Note that here, the non-problematic folds will have already
3886 * been done, so we can just copy such characters. We actually
3887 * don't completely fold the EXACTFL string. We skip the
3888 * unfolded multi-char folds, as that would just create work
3889 * below to figure out the size they already are */
3891 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3894 STRLEN s_len = UTF8SKIP(s);
3895 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3896 Copy(s, d, s_len, U8);
3899 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3900 *unfolded_multi_char = TRUE;
3901 Copy(s, d, s_len, U8);
3904 else if (isASCII(*s)) {
3905 *(d++) = toFOLD(*s);
3909 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
3915 /* Point the remainder of the routine to look at our temporary
3919 } /* End of creating folded copy of EXACTFL string */
3921 /* Examine the string for a multi-character fold sequence. UTF-8
3922 * patterns have all characters pre-folded by the time this code is
3924 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3925 length sequence we are looking for is 2 */
3927 int count = 0; /* How many characters in a multi-char fold */
3928 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3929 if (! len) { /* Not a multi-char fold: get next char */
3934 /* Nodes with 'ss' require special handling, except for
3935 * EXACTFA-ish for which there is no multi-char fold to this */
3936 if (len == 2 && *s == 's' && *(s+1) == 's'
3937 && OP(scan) != EXACTFA
3938 && OP(scan) != EXACTFA_NO_TRIE)
3941 if (OP(scan) != EXACTFL) {
3942 OP(scan) = EXACTFU_SS;
3946 else { /* Here is a generic multi-char fold. */
3947 U8* multi_end = s + len;
3949 /* Count how many characters are in it. In the case of
3950 * /aa, no folds which contain ASCII code points are
3951 * allowed, so check for those, and skip if found. */
3952 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3953 count = utf8_length(s, multi_end);
3957 while (s < multi_end) {
3960 goto next_iteration;
3970 /* The delta is how long the sequence is minus 1 (1 is how long
3971 * the character that folds to the sequence is) */
3972 total_count_delta += count - 1;
3976 /* We created a temporary folded copy of the string in EXACTFL
3977 * nodes. Therefore we need to be sure it doesn't go below zero,
3978 * as the real string could be shorter */
3979 if (OP(scan) == EXACTFL) {
3980 int total_chars = utf8_length((U8*) STRING(scan),
3981 (U8*) STRING(scan) + STR_LEN(scan));
3982 if (total_count_delta > total_chars) {
3983 total_count_delta = total_chars;
3987 *min_subtract += total_count_delta;
3990 else if (OP(scan) == EXACTFA) {
3992 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3993 * fold to the ASCII range (and there are no existing ones in the
3994 * upper latin1 range). But, as outlined in the comments preceding
3995 * this function, we need to flag any occurrences of the sharp s.
3996 * This character forbids trie formation (because of added
3998 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
3999 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4000 || UNICODE_DOT_DOT_VERSION > 0)
4002 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4003 OP(scan) = EXACTFA_NO_TRIE;
4004 *unfolded_multi_char = TRUE;
4012 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
4013 * folds that are all Latin1. As explained in the comments
4014 * preceding this function, we look also for the sharp s in EXACTF
4015 * and EXACTFL nodes; it can be in the final position. Otherwise
4016 * we can stop looking 1 byte earlier because have to find at least
4017 * two characters for a multi-fold */
4018 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4023 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4024 if (! len) { /* Not a multi-char fold. */
4025 if (*s == LATIN_SMALL_LETTER_SHARP_S
4026 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4028 *unfolded_multi_char = TRUE;
4035 && isALPHA_FOLD_EQ(*s, 's')
4036 && isALPHA_FOLD_EQ(*(s+1), 's'))
4039 /* EXACTF nodes need to know that the minimum length
4040 * changed so that a sharp s in the string can match this
4041 * ss in the pattern, but they remain EXACTF nodes, as they
4042 * won't match this unless the target string is is UTF-8,
4043 * which we don't know until runtime. EXACTFL nodes can't
4044 * transform into EXACTFU nodes */
4045 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4046 OP(scan) = EXACTFU_SS;
4050 *min_subtract += len - 1;
4058 /* Allow dumping but overwriting the collection of skipped
4059 * ops and/or strings with fake optimized ops */
4060 n = scan + NODE_SZ_STR(scan);
4068 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
4072 /* REx optimizer. Converts nodes into quicker variants "in place".
4073 Finds fixed substrings. */
4075 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4076 to the position after last scanned or to NULL. */
4078 #define INIT_AND_WITHP \
4079 assert(!and_withp); \
4080 Newx(and_withp,1, regnode_ssc); \
4081 SAVEFREEPV(and_withp)
4085 S_unwind_scan_frames(pTHX_ const void *p)
4087 scan_frame *f= (scan_frame *)p;
4089 scan_frame *n= f->next_frame;
4097 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4098 SSize_t *minlenp, SSize_t *deltap,
4103 regnode_ssc *and_withp,
4104 U32 flags, U32 depth)
4105 /* scanp: Start here (read-write). */
4106 /* deltap: Write maxlen-minlen here. */
4107 /* last: Stop before this one. */
4108 /* data: string data about the pattern */
4109 /* stopparen: treat close N as END */
4110 /* recursed: which subroutines have we recursed into */
4111 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4113 /* There must be at least this number of characters to match */
4116 regnode *scan = *scanp, *next;
4118 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4119 int is_inf_internal = 0; /* The studied chunk is infinite */
4120 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4121 scan_data_t data_fake;
4122 SV *re_trie_maxbuff = NULL;
4123 regnode *first_non_open = scan;
4124 SSize_t stopmin = SSize_t_MAX;
4125 scan_frame *frame = NULL;
4126 GET_RE_DEBUG_FLAGS_DECL;
4128 PERL_ARGS_ASSERT_STUDY_CHUNK;
4129 RExC_study_started= 1;
4133 while (first_non_open && OP(first_non_open) == OPEN)
4134 first_non_open=regnext(first_non_open);
4140 RExC_study_chunk_recursed_count++;
4142 DEBUG_OPTIMISE_MORE_r(
4144 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4145 depth, (long)stopparen,
4146 (unsigned long)RExC_study_chunk_recursed_count,
4147 (unsigned long)depth, (unsigned long)recursed_depth,
4150 if (recursed_depth) {
4153 for ( j = 0 ; j < recursed_depth ; j++ ) {
4154 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
4156 PAREN_TEST(RExC_study_chunk_recursed +
4157 ( j * RExC_study_chunk_recursed_bytes), i )
4160 !PAREN_TEST(RExC_study_chunk_recursed +
4161 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
4164 Perl_re_printf( aTHX_ " %d",(int)i);
4168 if ( j + 1 < recursed_depth ) {
4169 Perl_re_printf( aTHX_ ",");
4173 Perl_re_printf( aTHX_ "\n");
4176 while ( scan && OP(scan) != END && scan < last ){
4177 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4178 node length to get a real minimum (because
4179 the folded version may be shorter) */
4180 bool unfolded_multi_char = FALSE;
4181 /* Peephole optimizer: */
4182 DEBUG_STUDYDATA("Peep:", data, depth);
4183 DEBUG_PEEP("Peep", scan, depth);
4186 /* The reason we do this here is that we need to deal with things like
4187 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4188 * parsing code, as each (?:..) is handled by a different invocation of
4191 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
4193 /* Follow the next-chain of the current node and optimize
4194 away all the NOTHINGs from it. */
4195 if (OP(scan) != CURLYX) {
4196 const int max = (reg_off_by_arg[OP(scan)]
4198 /* I32 may be smaller than U16 on CRAYs! */
4199 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4200 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
4204 /* Skip NOTHING and LONGJMP. */
4205 while ((n = regnext(n))
4206 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4207 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
4208 && off + noff < max)
4210 if (reg_off_by_arg[OP(scan)])
4213 NEXT_OFF(scan) = off;
4216 /* The principal pseudo-switch. Cannot be a switch, since we
4217 look into several different things. */
4218 if ( OP(scan) == DEFINEP ) {
4220 SSize_t deltanext = 0;
4221 SSize_t fake_last_close = 0;
4222 I32 f = SCF_IN_DEFINE;
4224 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4225 scan = regnext(scan);
4226 assert( OP(scan) == IFTHEN );
4227 DEBUG_PEEP("expect IFTHEN", scan, depth);
4229 data_fake.last_closep= &fake_last_close;
4231 next = regnext(scan);
4232 scan = NEXTOPER(NEXTOPER(scan));
4233 DEBUG_PEEP("scan", scan, depth);
4234 DEBUG_PEEP("next", next, depth);
4236 /* we suppose the run is continuous, last=next...
4237 * NOTE we dont use the return here! */
4238 (void)study_chunk(pRExC_state, &scan, &minlen,
4239 &deltanext, next, &data_fake, stopparen,
4240 recursed_depth, NULL, f, depth+1);
4245 OP(scan) == BRANCH ||
4246 OP(scan) == BRANCHJ ||
4249 next = regnext(scan);
4252 /* The op(next)==code check below is to see if we
4253 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4254 * IFTHEN is special as it might not appear in pairs.
4255 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4256 * we dont handle it cleanly. */
4257 if (OP(next) == code || code == IFTHEN) {
4258 /* NOTE - There is similar code to this block below for
4259 * handling TRIE nodes on a re-study. If you change stuff here
4260 * check there too. */
4261 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
4263 regnode * const startbranch=scan;
4265 if (flags & SCF_DO_SUBSTR) {
4266 /* Cannot merge strings after this. */
4267 scan_commit(pRExC_state, data, minlenp, is_inf);
4270 if (flags & SCF_DO_STCLASS)
4271 ssc_init_zero(pRExC_state, &accum);
4273 while (OP(scan) == code) {
4274 SSize_t deltanext, minnext, fake;
4276 regnode_ssc this_class;
4278 DEBUG_PEEP("Branch", scan, depth);
4281 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4283 data_fake.whilem_c = data->whilem_c;
4284 data_fake.last_closep = data->last_closep;
4287 data_fake.last_closep = &fake;
4289 data_fake.pos_delta = delta;
4290 next = regnext(scan);
4292 scan = NEXTOPER(scan); /* everything */
4293 if (code != BRANCH) /* everything but BRANCH */
4294 scan = NEXTOPER(scan);
4296 if (flags & SCF_DO_STCLASS) {
4297 ssc_init(pRExC_state, &this_class);
4298 data_fake.start_class = &this_class;
4299 f = SCF_DO_STCLASS_AND;
4301 if (flags & SCF_WHILEM_VISITED_POS)
4302 f |= SCF_WHILEM_VISITED_POS;
4304 /* we suppose the run is continuous, last=next...*/
4305 minnext = study_chunk(pRExC_state, &scan, minlenp,
4306 &deltanext, next, &data_fake, stopparen,
4307 recursed_depth, NULL, f,depth+1);
4311 if (deltanext == SSize_t_MAX) {
4312 is_inf = is_inf_internal = 1;
4314 } else if (max1 < minnext + deltanext)
4315 max1 = minnext + deltanext;
4317 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4319 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4320 if ( stopmin > minnext)
4321 stopmin = min + min1;
4322 flags &= ~SCF_DO_SUBSTR;
4324 data->flags |= SCF_SEEN_ACCEPT;
4327 if (data_fake.flags & SF_HAS_EVAL)
4328 data->flags |= SF_HAS_EVAL;
4329 data->whilem_c = data_fake.whilem_c;
4331 if (flags & SCF_DO_STCLASS)
4332 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4334 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4336 if (flags & SCF_DO_SUBSTR) {
4337 data->pos_min += min1;
4338 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
4339 data->pos_delta = SSize_t_MAX;
4341 data->pos_delta += max1 - min1;
4342 if (max1 != min1 || is_inf)
4343 data->longest = &(data->longest_float);
4346 if (delta == SSize_t_MAX
4347 || SSize_t_MAX - delta - (max1 - min1) < 0)
4348 delta = SSize_t_MAX;
4350 delta += max1 - min1;
4351 if (flags & SCF_DO_STCLASS_OR) {
4352 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4354 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4355 flags &= ~SCF_DO_STCLASS;
4358 else if (flags & SCF_DO_STCLASS_AND) {
4360 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4361 flags &= ~SCF_DO_STCLASS;
4364 /* Switch to OR mode: cache the old value of
4365 * data->start_class */
4367 StructCopy(data->start_class, and_withp, regnode_ssc);
4368 flags &= ~SCF_DO_STCLASS_AND;
4369 StructCopy(&accum, data->start_class, regnode_ssc);
4370 flags |= SCF_DO_STCLASS_OR;
4374 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4375 OP( startbranch ) == BRANCH )
4379 Assuming this was/is a branch we are dealing with: 'scan'
4380 now points at the item that follows the branch sequence,
4381 whatever it is. We now start at the beginning of the
4382 sequence and look for subsequences of
4388 which would be constructed from a pattern like
4391 If we can find such a subsequence we need to turn the first
4392 element into a trie and then add the subsequent branch exact
4393 strings to the trie.
4397 1. patterns where the whole set of branches can be
4400 2. patterns where only a subset can be converted.
4402 In case 1 we can replace the whole set with a single regop
4403 for the trie. In case 2 we need to keep the start and end
4406 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4407 becomes BRANCH TRIE; BRANCH X;
4409 There is an additional case, that being where there is a
4410 common prefix, which gets split out into an EXACT like node
4411 preceding the TRIE node.
4413 If x(1..n)==tail then we can do a simple trie, if not we make
4414 a "jump" trie, such that when we match the appropriate word
4415 we "jump" to the appropriate tail node. Essentially we turn
4416 a nested if into a case structure of sorts.
4421 if (!re_trie_maxbuff) {
4422 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4423 if (!SvIOK(re_trie_maxbuff))
4424 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4426 if ( SvIV(re_trie_maxbuff)>=0 ) {
4428 regnode *first = (regnode *)NULL;
4429 regnode *last = (regnode *)NULL;
4430 regnode *tail = scan;
4434 /* var tail is used because there may be a TAIL
4435 regop in the way. Ie, the exacts will point to the
4436 thing following the TAIL, but the last branch will
4437 point at the TAIL. So we advance tail. If we
4438 have nested (?:) we may have to move through several
4442 while ( OP( tail ) == TAIL ) {
4443 /* this is the TAIL generated by (?:) */
4444 tail = regnext( tail );
4448 DEBUG_TRIE_COMPILE_r({
4449 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4450 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4452 "Looking for TRIE'able sequences. Tail node is ",
4453 (UV)(tail - RExC_emit_start),
4454 SvPV_nolen_const( RExC_mysv )
4460 Step through the branches
4461 cur represents each branch,
4462 noper is the first thing to be matched as part
4464 noper_next is the regnext() of that node.
4466 We normally handle a case like this
4467 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4468 support building with NOJUMPTRIE, which restricts
4469 the trie logic to structures like /FOO|BAR/.
4471 If noper is a trieable nodetype then the branch is
4472 a possible optimization target. If we are building
4473 under NOJUMPTRIE then we require that noper_next is
4474 the same as scan (our current position in the regex
4477 Once we have two or more consecutive such branches
4478 we can create a trie of the EXACT's contents and
4479 stitch it in place into the program.
4481 If the sequence represents all of the branches in
4482 the alternation we replace the entire thing with a
4485 Otherwise when it is a subsequence we need to
4486 stitch it in place and replace only the relevant
4487 branches. This means the first branch has to remain
4488 as it is used by the alternation logic, and its
4489 next pointer, and needs to be repointed at the item
4490 on the branch chain following the last branch we
4491 have optimized away.
4493 This could be either a BRANCH, in which case the
4494 subsequence is internal, or it could be the item
4495 following the branch sequence in which case the
4496 subsequence is at the end (which does not
4497 necessarily mean the first node is the start of the
4500 TRIE_TYPE(X) is a define which maps the optype to a
4504 ----------------+-----------
4508 EXACTFU_SS | EXACTFU
4511 EXACTFLU8 | EXACTFLU8
4515 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4517 : ( EXACT == (X) ) \
4519 : ( EXACTFU == (X) || EXACTFU_SS == (X) ) \
4521 : ( EXACTFA == (X) ) \
4523 : ( EXACTL == (X) ) \
4525 : ( EXACTFLU8 == (X) ) \
4529 /* dont use tail as the end marker for this traverse */
4530 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4531 regnode * const noper = NEXTOPER( cur );
4532 U8 noper_type = OP( noper );
4533 U8 noper_trietype = TRIE_TYPE( noper_type );
4534 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4535 regnode * const noper_next = regnext( noper );
4536 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4537 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4540 DEBUG_TRIE_COMPILE_r({
4541 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4542 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4544 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4546 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4547 Perl_re_printf( aTHX_ " -> %d:%s",
4548 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
4551 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4552 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
4553 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
4555 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
4556 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4557 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4561 /* Is noper a trieable nodetype that can be merged
4562 * with the current trie (if there is one)? */
4566 ( noper_trietype == NOTHING )
4567 || ( trietype == NOTHING )
4568 || ( trietype == noper_trietype )
4571 && noper_next >= tail
4575 /* Handle mergable triable node Either we are
4576 * the first node in a new trieable sequence,
4577 * in which case we do some bookkeeping,
4578 * otherwise we update the end pointer. */
4581 if ( noper_trietype == NOTHING ) {
4582 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4583 regnode * const noper_next = regnext( noper );
4584 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4585 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4588 if ( noper_next_trietype ) {
4589 trietype = noper_next_trietype;
4590 } else if (noper_next_type) {
4591 /* a NOTHING regop is 1 regop wide.
4592 * We need at least two for a trie
4593 * so we can't merge this in */
4597 trietype = noper_trietype;
4600 if ( trietype == NOTHING )
4601 trietype = noper_trietype;
4606 } /* end handle mergable triable node */
4608 /* handle unmergable node -
4609 * noper may either be a triable node which can
4610 * not be tried together with the current trie,
4611 * or a non triable node */
4613 /* If last is set and trietype is not
4614 * NOTHING then we have found at least two
4615 * triable branch sequences in a row of a
4616 * similar trietype so we can turn them
4617 * into a trie. If/when we allow NOTHING to
4618 * start a trie sequence this condition
4619 * will be required, and it isn't expensive
4620 * so we leave it in for now. */
4621 if ( trietype && trietype != NOTHING )
4622 make_trie( pRExC_state,
4623 startbranch, first, cur, tail,
4624 count, trietype, depth+1 );
4625 last = NULL; /* note: we clear/update
4626 first, trietype etc below,
4627 so we dont do it here */
4631 && noper_next >= tail
4634 /* noper is triable, so we can start a new
4638 trietype = noper_trietype;
4640 /* if we already saw a first but the
4641 * current node is not triable then we have
4642 * to reset the first information. */
4647 } /* end handle unmergable node */
4648 } /* loop over branches */
4649 DEBUG_TRIE_COMPILE_r({
4650 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4651 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
4652 depth+1, SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4653 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
4654 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4655 PL_reg_name[trietype]
4659 if ( last && trietype ) {
4660 if ( trietype != NOTHING ) {
4661 /* the last branch of the sequence was part of
4662 * a trie, so we have to construct it here
4663 * outside of the loop */
4664 made= make_trie( pRExC_state, startbranch,
4665 first, scan, tail, count,
4666 trietype, depth+1 );
4667 #ifdef TRIE_STUDY_OPT
4668 if ( ((made == MADE_EXACT_TRIE &&
4669 startbranch == first)
4670 || ( first_non_open == first )) &&
4672 flags |= SCF_TRIE_RESTUDY;
4673 if ( startbranch == first
4676 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4681 /* at this point we know whatever we have is a
4682 * NOTHING sequence/branch AND if 'startbranch'
4683 * is 'first' then we can turn the whole thing
4686 if ( startbranch == first ) {
4688 /* the entire thing is a NOTHING sequence,
4689 * something like this: (?:|) So we can
4690 * turn it into a plain NOTHING op. */
4691 DEBUG_TRIE_COMPILE_r({
4692 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4693 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
4695 SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4698 OP(startbranch)= NOTHING;
4699 NEXT_OFF(startbranch)= tail - startbranch;
4700 for ( opt= startbranch + 1; opt < tail ; opt++ )
4704 } /* end if ( last) */
4705 } /* TRIE_MAXBUF is non zero */
4710 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4711 scan = NEXTOPER(NEXTOPER(scan));
4712 } else /* single branch is optimized. */
4713 scan = NEXTOPER(scan);
4715 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
4717 regnode *start = NULL;
4718 regnode *end = NULL;
4719 U32 my_recursed_depth= recursed_depth;
4721 if (OP(scan) != SUSPEND) { /* GOSUB */
4722 /* Do setup, note this code has side effects beyond
4723 * the rest of this block. Specifically setting
4724 * RExC_recurse[] must happen at least once during
4727 RExC_recurse[ARG2L(scan)] = scan;
4728 start = RExC_open_parens[paren];
4729 end = RExC_close_parens[paren];
4731 /* NOTE we MUST always execute the above code, even
4732 * if we do nothing with a GOSUB */
4734 ( flags & SCF_IN_DEFINE )
4737 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
4739 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
4742 /* no need to do anything here if we are in a define. */
4743 /* or we are after some kind of infinite construct
4744 * so we can skip recursing into this item.
4745 * Since it is infinite we will not change the maxlen
4746 * or delta, and if we miss something that might raise
4747 * the minlen it will merely pessimise a little.
4749 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
4750 * might result in a minlen of 1 and not of 4,
4751 * but this doesn't make us mismatch, just try a bit
4752 * harder than we should.
4754 scan= regnext(scan);
4761 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4763 /* it is quite possible that there are more efficient ways
4764 * to do this. We maintain a bitmap per level of recursion
4765 * of which patterns we have entered so we can detect if a
4766 * pattern creates a possible infinite loop. When we
4767 * recurse down a level we copy the previous levels bitmap
4768 * down. When we are at recursion level 0 we zero the top
4769 * level bitmap. It would be nice to implement a different
4770 * more efficient way of doing this. In particular the top
4771 * level bitmap may be unnecessary.
4773 if (!recursed_depth) {
4774 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4776 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4777 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4778 RExC_study_chunk_recursed_bytes, U8);
4780 /* we havent recursed into this paren yet, so recurse into it */
4781 DEBUG_STUDYDATA("gosub-set:", data,depth);
4782 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4783 my_recursed_depth= recursed_depth + 1;
4785 DEBUG_STUDYDATA("gosub-inf:", data,depth);
4786 /* some form of infinite recursion, assume infinite length
4788 if (flags & SCF_DO_SUBSTR) {
4789 scan_commit(pRExC_state, data, minlenp, is_inf);
4790 data->longest = &(data->longest_float);
4792 is_inf = is_inf_internal = 1;
4793 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4794 ssc_anything(data->start_class);
4795 flags &= ~SCF_DO_STCLASS;
4797 start= NULL; /* reset start so we dont recurse later on. */
4802 end = regnext(scan);
4805 scan_frame *newframe;
4807 if (!RExC_frame_last) {
4808 Newxz(newframe, 1, scan_frame);
4809 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
4810 RExC_frame_head= newframe;
4812 } else if (!RExC_frame_last->next_frame) {
4813 Newxz(newframe,1,scan_frame);
4814 RExC_frame_last->next_frame= newframe;
4815 newframe->prev_frame= RExC_frame_last;
4818 newframe= RExC_frame_last->next_frame;
4820 RExC_frame_last= newframe;
4822 newframe->next_regnode = regnext(scan);
4823 newframe->last_regnode = last;
4824 newframe->stopparen = stopparen;
4825 newframe->prev_recursed_depth = recursed_depth;
4826 newframe->this_prev_frame= frame;
4828 DEBUG_STUDYDATA("frame-new:",data,depth);
4829 DEBUG_PEEP("fnew", scan, depth);
4836 recursed_depth= my_recursed_depth;
4841 else if (OP(scan) == EXACT || OP(scan) == EXACTL) {
4842 SSize_t l = STR_LEN(scan);
4845 const U8 * const s = (U8*)STRING(scan);
4846 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4847 l = utf8_length(s, s + l);
4849 uc = *((U8*)STRING(scan));
4852 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4853 /* The code below prefers earlier match for fixed
4854 offset, later match for variable offset. */
4855 if (data->last_end == -1) { /* Update the start info. */
4856 data->last_start_min = data->pos_min;
4857 data->last_start_max = is_inf
4858 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4860 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4862 SvUTF8_on(data->last_found);
4864 SV * const sv = data->last_found;
4865 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4866 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4867 if (mg && mg->mg_len >= 0)
4868 mg->mg_len += utf8_length((U8*)STRING(scan),
4869 (U8*)STRING(scan)+STR_LEN(scan));
4871 data->last_end = data->pos_min + l;
4872 data->pos_min += l; /* As in the first entry. */
4873 data->flags &= ~SF_BEFORE_EOL;
4876 /* ANDing the code point leaves at most it, and not in locale, and
4877 * can't match null string */
4878 if (flags & SCF_DO_STCLASS_AND) {
4879 ssc_cp_and(data->start_class, uc);
4880 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4881 ssc_clear_locale(data->start_class);
4883 else if (flags & SCF_DO_STCLASS_OR) {
4884 ssc_add_cp(data->start_class, uc);
4885 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4887 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4888 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4890 flags &= ~SCF_DO_STCLASS;
4892 else if (PL_regkind[OP(scan)] == EXACT) {
4893 /* But OP != EXACT!, so is EXACTFish */
4894 SSize_t l = STR_LEN(scan);
4895 const U8 * s = (U8*)STRING(scan);
4897 /* Search for fixed substrings supports EXACT only. */
4898 if (flags & SCF_DO_SUBSTR) {
4900 scan_commit(pRExC_state, data, minlenp, is_inf);
4903 l = utf8_length(s, s + l);
4905 if (unfolded_multi_char) {
4906 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4908 min += l - min_subtract;
4910 delta += min_subtract;
4911 if (flags & SCF_DO_SUBSTR) {
4912 data->pos_min += l - min_subtract;
4913 if (data->pos_min < 0) {
4916 data->pos_delta += min_subtract;
4918 data->longest = &(data->longest_float);
4922 if (flags & SCF_DO_STCLASS) {
4923 SV* EXACTF_invlist = _make_exactf_invlist(pRExC_state, scan);
4925 assert(EXACTF_invlist);
4926 if (flags & SCF_DO_STCLASS_AND) {
4927 if (OP(scan) != EXACTFL)
4928 ssc_clear_locale(data->start_class);
4929 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4930 ANYOF_POSIXL_ZERO(data->start_class);
4931 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4933 else { /* SCF_DO_STCLASS_OR */
4934 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4935 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4937 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4938 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4940 flags &= ~SCF_DO_STCLASS;
4941 SvREFCNT_dec(EXACTF_invlist);
4944 else if (REGNODE_VARIES(OP(scan))) {
4945 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4946 I32 fl = 0, f = flags;
4947 regnode * const oscan = scan;
4948 regnode_ssc this_class;
4949 regnode_ssc *oclass = NULL;
4950 I32 next_is_eval = 0;
4952 switch (PL_regkind[OP(scan)]) {
4953 case WHILEM: /* End of (?:...)* . */
4954 scan = NEXTOPER(scan);
4957 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4958 next = NEXTOPER(scan);
4959 if (OP(next) == EXACT
4960 || OP(next) == EXACTL
4961 || (flags & SCF_DO_STCLASS))
4964 maxcount = REG_INFTY;
4965 next = regnext(scan);
4966 scan = NEXTOPER(scan);
4970 if (flags & SCF_DO_SUBSTR)
4975 if (flags & SCF_DO_STCLASS) {
4977 maxcount = REG_INFTY;
4978 next = regnext(scan);
4979 scan = NEXTOPER(scan);
4982 if (flags & SCF_DO_SUBSTR) {
4983 scan_commit(pRExC_state, data, minlenp, is_inf);
4984 /* Cannot extend fixed substrings */
4985 data->longest = &(data->longest_float);
4987 is_inf = is_inf_internal = 1;
4988 scan = regnext(scan);
4989 goto optimize_curly_tail;
4991 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
4992 && (scan->flags == stopparen))
4997 mincount = ARG1(scan);
4998 maxcount = ARG2(scan);
5000 next = regnext(scan);
5001 if (OP(scan) == CURLYX) {
5002 I32 lp = (data ? *(data->last_closep) : 0);
5003 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5005 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5006 next_is_eval = (OP(scan) == EVAL);
5008 if (flags & SCF_DO_SUBSTR) {
5010 scan_commit(pRExC_state, data, minlenp, is_inf);
5011 /* Cannot extend fixed substrings */
5012 pos_before = data->pos_min;
5016 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5018 data->flags |= SF_IS_INF;
5020 if (flags & SCF_DO_STCLASS) {
5021 ssc_init(pRExC_state, &this_class);
5022 oclass = data->start_class;
5023 data->start_class = &this_class;
5024 f |= SCF_DO_STCLASS_AND;
5025 f &= ~SCF_DO_STCLASS_OR;
5027 /* Exclude from super-linear cache processing any {n,m}
5028 regops for which the combination of input pos and regex
5029 pos is not enough information to determine if a match
5032 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5033 regex pos at the \s*, the prospects for a match depend not
5034 only on the input position but also on how many (bar\s*)
5035 repeats into the {4,8} we are. */
5036 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5037 f &= ~SCF_WHILEM_VISITED_POS;
5039 /* This will finish on WHILEM, setting scan, or on NULL: */
5040 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5041 last, data, stopparen, recursed_depth, NULL,
5043 ? (f & ~SCF_DO_SUBSTR)
5047 if (flags & SCF_DO_STCLASS)
5048 data->start_class = oclass;
5049 if (mincount == 0 || minnext == 0) {
5050 if (flags & SCF_DO_STCLASS_OR) {
5051 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5053 else if (flags & SCF_DO_STCLASS_AND) {
5054 /* Switch to OR mode: cache the old value of
5055 * data->start_class */
5057 StructCopy(data->start_class, and_withp, regnode_ssc);
5058 flags &= ~SCF_DO_STCLASS_AND;
5059 StructCopy(&this_class, data->start_class, regnode_ssc);
5060 flags |= SCF_DO_STCLASS_OR;
5061 ANYOF_FLAGS(data->start_class)
5062 |= SSC_MATCHES_EMPTY_STRING;
5064 } else { /* Non-zero len */
5065 if (flags & SCF_DO_STCLASS_OR) {
5066 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5067 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5069 else if (flags & SCF_DO_STCLASS_AND)
5070 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5071 flags &= ~SCF_DO_STCLASS;
5073 if (!scan) /* It was not CURLYX, but CURLY. */
5075 if (!(flags & SCF_TRIE_DOING_RESTUDY)
5076 /* ? quantifier ok, except for (?{ ... }) */
5077 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5078 && (minnext == 0) && (deltanext == 0)
5079 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5080 && maxcount <= REG_INFTY/3) /* Complement check for big
5083 /* Fatal warnings may leak the regexp without this: */
5084 SAVEFREESV(RExC_rx_sv);
5085 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5086 "Quantifier unexpected on zero-length expression "
5087 "in regex m/%" UTF8f "/",
5088 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5090 (void)ReREFCNT_inc(RExC_rx_sv);
5093 min += minnext * mincount;
5094 is_inf_internal |= deltanext == SSize_t_MAX
5095 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5096 is_inf |= is_inf_internal;
5098 delta = SSize_t_MAX;
5100 delta += (minnext + deltanext) * maxcount
5101 - minnext * mincount;
5103 /* Try powerful optimization CURLYX => CURLYN. */
5104 if ( OP(oscan) == CURLYX && data
5105 && data->flags & SF_IN_PAR
5106 && !(data->flags & SF_HAS_EVAL)
5107 && !deltanext && minnext == 1 ) {
5108 /* Try to optimize to CURLYN. */
5109 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5110 regnode * const nxt1 = nxt;
5117 if (!REGNODE_SIMPLE(OP(nxt))
5118 && !(PL_regkind[OP(nxt)] == EXACT
5119 && STR_LEN(nxt) == 1))
5125 if (OP(nxt) != CLOSE)
5127 if (RExC_open_parens) {
5128 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5129 RExC_close_parens[ARG(nxt1)]=nxt+2; /*close->while*/
5131 /* Now we know that nxt2 is the only contents: */
5132 oscan->flags = (U8)ARG(nxt);
5134 OP(nxt1) = NOTHING; /* was OPEN. */
5137 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5138 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5139 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5140 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5141 OP(nxt + 1) = OPTIMIZED; /* was count. */
5142 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5147 /* Try optimization CURLYX => CURLYM. */
5148 if ( OP(oscan) == CURLYX && data
5149 && !(data->flags & SF_HAS_PAR)
5150 && !(data->flags & SF_HAS_EVAL)
5151 && !deltanext /* atom is fixed width */
5152 && minnext != 0 /* CURLYM can't handle zero width */
5154 /* Nor characters whose fold at run-time may be
5155 * multi-character */
5156 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5158 /* XXXX How to optimize if data == 0? */
5159 /* Optimize to a simpler form. */
5160 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5164 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5165 && (OP(nxt2) != WHILEM))
5167 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5168 /* Need to optimize away parenths. */
5169 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5170 /* Set the parenth number. */
5171 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5173 oscan->flags = (U8)ARG(nxt);
5174 if (RExC_open_parens) {
5175 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5176 RExC_close_parens[ARG(nxt1)]=nxt2+1; /*close->NOTHING*/
5178 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5179 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5182 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5183 OP(nxt + 1) = OPTIMIZED; /* was count. */
5184 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5185 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5188 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5189 regnode *nnxt = regnext(nxt1);
5191 if (reg_off_by_arg[OP(nxt1)])
5192 ARG_SET(nxt1, nxt2 - nxt1);
5193 else if (nxt2 - nxt1 < U16_MAX)
5194 NEXT_OFF(nxt1) = nxt2 - nxt1;
5196 OP(nxt) = NOTHING; /* Cannot beautify */
5201 /* Optimize again: */
5202 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5203 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
5208 else if ((OP(oscan) == CURLYX)
5209 && (flags & SCF_WHILEM_VISITED_POS)
5210 /* See the comment on a similar expression above.
5211 However, this time it's not a subexpression
5212 we care about, but the expression itself. */
5213 && (maxcount == REG_INFTY)
5215 /* This stays as CURLYX, we can put the count/of pair. */
5216 /* Find WHILEM (as in regexec.c) */
5217 regnode *nxt = oscan + NEXT_OFF(oscan);
5219 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5221 nxt = PREVOPER(nxt);
5222 if (nxt->flags & 0xf) {
5223 /* we've already set whilem count on this node */
5224 } else if (++data->whilem_c < 16) {
5225 assert(data->whilem_c <= RExC_whilem_seen);
5226 nxt->flags = (U8)(data->whilem_c
5227 | (RExC_whilem_seen << 4)); /* On WHILEM */
5230 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5232 if (flags & SCF_DO_SUBSTR) {
5233 SV *last_str = NULL;
5234 STRLEN last_chrs = 0;
5235 int counted = mincount != 0;
5237 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5239 SSize_t b = pos_before >= data->last_start_min
5240 ? pos_before : data->last_start_min;
5242 const char * const s = SvPV_const(data->last_found, l);
5243 SSize_t old = b - data->last_start_min;
5246 old = utf8_hop((U8*)s, old) - (U8*)s;
5248 /* Get the added string: */
5249 last_str = newSVpvn_utf8(s + old, l, UTF);
5250 last_chrs = UTF ? utf8_length((U8*)(s + old),
5251 (U8*)(s + old + l)) : l;
5252 if (deltanext == 0 && pos_before == b) {
5253 /* What was added is a constant string */
5256 SvGROW(last_str, (mincount * l) + 1);
5257 repeatcpy(SvPVX(last_str) + l,
5258 SvPVX_const(last_str), l,
5260 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5261 /* Add additional parts. */
5262 SvCUR_set(data->last_found,
5263 SvCUR(data->last_found) - l);
5264 sv_catsv(data->last_found, last_str);
5266 SV * sv = data->last_found;
5268 SvUTF8(sv) && SvMAGICAL(sv) ?
5269 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5270 if (mg && mg->mg_len >= 0)
5271 mg->mg_len += last_chrs * (mincount-1);
5273 last_chrs *= mincount;
5274 data->last_end += l * (mincount - 1);
5277 /* start offset must point into the last copy */
5278 data->last_start_min += minnext * (mincount - 1);
5279 data->last_start_max =
5282 : data->last_start_max +
5283 (maxcount - 1) * (minnext + data->pos_delta);
5286 /* It is counted once already... */
5287 data->pos_min += minnext * (mincount - counted);
5289 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5290 " SSize_t_MAX=%" UVuf " minnext=%" UVuf
5291 " maxcount=%" UVuf " mincount=%" UVuf "\n",
5292 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
5294 if (deltanext != SSize_t_MAX)
5295 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5296 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5297 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
5299 if (deltanext == SSize_t_MAX
5300 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
5301 data->pos_delta = SSize_t_MAX;
5303 data->pos_delta += - counted * deltanext +
5304 (minnext + deltanext) * maxcount - minnext * mincount;
5305 if (mincount != maxcount) {
5306 /* Cannot extend fixed substrings found inside
5308 scan_commit(pRExC_state, data, minlenp, is_inf);
5309 if (mincount && last_str) {
5310 SV * const sv = data->last_found;
5311 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5312 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5316 sv_setsv(sv, last_str);
5317 data->last_end = data->pos_min;
5318 data->last_start_min = data->pos_min - last_chrs;
5319 data->last_start_max = is_inf
5321 : data->pos_min + data->pos_delta - last_chrs;
5323 data->longest = &(data->longest_float);
5325 SvREFCNT_dec(last_str);
5327 if (data && (fl & SF_HAS_EVAL))
5328 data->flags |= SF_HAS_EVAL;
5329 optimize_curly_tail:
5330 if (OP(oscan) != CURLYX) {
5331 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
5333 NEXT_OFF(oscan) += NEXT_OFF(next);
5339 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5344 if (flags & SCF_DO_SUBSTR) {
5345 /* Cannot expect anything... */
5346 scan_commit(pRExC_state, data, minlenp, is_inf);
5347 data->longest = &(data->longest_float);
5349 is_inf = is_inf_internal = 1;
5350 if (flags & SCF_DO_STCLASS_OR) {
5351 if (OP(scan) == CLUMP) {
5352 /* Actually is any start char, but very few code points
5353 * aren't start characters */
5354 ssc_match_all_cp(data->start_class);
5357 ssc_anything(data->start_class);
5360 flags &= ~SCF_DO_STCLASS;
5364 else if (OP(scan) == LNBREAK) {
5365 if (flags & SCF_DO_STCLASS) {
5366 if (flags & SCF_DO_STCLASS_AND) {
5367 ssc_intersection(data->start_class,
5368 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5369 ssc_clear_locale(data->start_class);
5370 ANYOF_FLAGS(data->start_class)
5371 &= ~SSC_MATCHES_EMPTY_STRING;
5373 else if (flags & SCF_DO_STCLASS_OR) {
5374 ssc_union(data->start_class,
5375 PL_XPosix_ptrs[_CC_VERTSPACE],
5377 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5379 /* See commit msg for
5380 * 749e076fceedeb708a624933726e7989f2302f6a */
5381 ANYOF_FLAGS(data->start_class)
5382 &= ~SSC_MATCHES_EMPTY_STRING;
5384 flags &= ~SCF_DO_STCLASS;
5387 if (delta != SSize_t_MAX)
5388 delta++; /* Because of the 2 char string cr-lf */
5389 if (flags & SCF_DO_SUBSTR) {
5390 /* Cannot expect anything... */
5391 scan_commit(pRExC_state, data, minlenp, is_inf);
5393 data->pos_delta += 1;
5394 data->longest = &(data->longest_float);
5397 else if (REGNODE_SIMPLE(OP(scan))) {
5399 if (flags & SCF_DO_SUBSTR) {
5400 scan_commit(pRExC_state, data, minlenp, is_inf);
5404 if (flags & SCF_DO_STCLASS) {
5406 SV* my_invlist = NULL;
5409 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5410 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5412 /* Some of the logic below assumes that switching
5413 locale on will only add false positives. */
5418 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5422 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5423 ssc_match_all_cp(data->start_class);
5428 SV* REG_ANY_invlist = _new_invlist(2);
5429 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5431 if (flags & SCF_DO_STCLASS_OR) {
5432 ssc_union(data->start_class,
5434 TRUE /* TRUE => invert, hence all but \n
5438 else if (flags & SCF_DO_STCLASS_AND) {
5439 ssc_intersection(data->start_class,
5441 TRUE /* TRUE => invert */
5443 ssc_clear_locale(data->start_class);
5445 SvREFCNT_dec_NN(REG_ANY_invlist);
5452 if (flags & SCF_DO_STCLASS_AND)
5453 ssc_and(pRExC_state, data->start_class,
5454 (regnode_charclass *) scan);
5456 ssc_or(pRExC_state, data->start_class,
5457 (regnode_charclass *) scan);
5465 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5466 if (flags & SCF_DO_STCLASS_AND) {
5467 bool was_there = cBOOL(
5468 ANYOF_POSIXL_TEST(data->start_class,
5470 ANYOF_POSIXL_ZERO(data->start_class);
5471 if (was_there) { /* Do an AND */
5472 ANYOF_POSIXL_SET(data->start_class, namedclass);
5474 /* No individual code points can now match */
5475 data->start_class->invlist
5476 = sv_2mortal(_new_invlist(0));
5479 int complement = namedclass + ((invert) ? -1 : 1);
5481 assert(flags & SCF_DO_STCLASS_OR);
5483 /* If the complement of this class was already there,
5484 * the result is that they match all code points,
5485 * (\d + \D == everything). Remove the classes from
5486 * future consideration. Locale is not relevant in
5488 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5489 ssc_match_all_cp(data->start_class);
5490 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5491 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5493 else { /* The usual case; just add this class to the
5495 ANYOF_POSIXL_SET(data->start_class, namedclass);
5500 case NPOSIXA: /* For these, we always know the exact set of
5505 if (FLAGS(scan) == _CC_ASCII) {
5506 my_invlist = invlist_clone(PL_XPosix_ptrs[_CC_ASCII]);
5509 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
5510 PL_XPosix_ptrs[_CC_ASCII],
5521 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5523 /* NPOSIXD matches all upper Latin1 code points unless the
5524 * target string being matched is UTF-8, which is
5525 * unknowable until match time. Since we are going to
5526 * invert, we want to get rid of all of them so that the
5527 * inversion will match all */
5528 if (OP(scan) == NPOSIXD) {
5529 _invlist_subtract(my_invlist, PL_UpperLatin1,
5535 if (flags & SCF_DO_STCLASS_AND) {
5536 ssc_intersection(data->start_class, my_invlist, invert);
5537 ssc_clear_locale(data->start_class);
5540 assert(flags & SCF_DO_STCLASS_OR);
5541 ssc_union(data->start_class, my_invlist, invert);
5543 SvREFCNT_dec(my_invlist);
5545 if (flags & SCF_DO_STCLASS_OR)
5546 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5547 flags &= ~SCF_DO_STCLASS;
5550 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5551 data->flags |= (OP(scan) == MEOL
5554 scan_commit(pRExC_state, data, minlenp, is_inf);
5557 else if ( PL_regkind[OP(scan)] == BRANCHJ
5558 /* Lookbehind, or need to calculate parens/evals/stclass: */
5559 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5560 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5562 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5563 || OP(scan) == UNLESSM )
5565 /* Negative Lookahead/lookbehind
5566 In this case we can't do fixed string optimisation.
5569 SSize_t deltanext, minnext, fake = 0;
5574 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5576 data_fake.whilem_c = data->whilem_c;
5577 data_fake.last_closep = data->last_closep;
5580 data_fake.last_closep = &fake;
5581 data_fake.pos_delta = delta;
5582 if ( flags & SCF_DO_STCLASS && !scan->flags
5583 && OP(scan) == IFMATCH ) { /* Lookahead */
5584 ssc_init(pRExC_state, &intrnl);
5585 data_fake.start_class = &intrnl;
5586 f |= SCF_DO_STCLASS_AND;
5588 if (flags & SCF_WHILEM_VISITED_POS)
5589 f |= SCF_WHILEM_VISITED_POS;
5590 next = regnext(scan);
5591 nscan = NEXTOPER(NEXTOPER(scan));
5592 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5593 last, &data_fake, stopparen,
5594 recursed_depth, NULL, f, depth+1);
5597 FAIL("Variable length lookbehind not implemented");
5599 else if (minnext > (I32)U8_MAX) {
5600 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5603 scan->flags = (U8)minnext;
5606 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5608 if (data_fake.flags & SF_HAS_EVAL)
5609 data->flags |= SF_HAS_EVAL;
5610 data->whilem_c = data_fake.whilem_c;
5612 if (f & SCF_DO_STCLASS_AND) {
5613 if (flags & SCF_DO_STCLASS_OR) {
5614 /* OR before, AND after: ideally we would recurse with
5615 * data_fake to get the AND applied by study of the
5616 * remainder of the pattern, and then derecurse;
5617 * *** HACK *** for now just treat as "no information".
5618 * See [perl #56690].
5620 ssc_init(pRExC_state, data->start_class);
5622 /* AND before and after: combine and continue. These
5623 * assertions are zero-length, so can match an EMPTY
5625 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5626 ANYOF_FLAGS(data->start_class)
5627 |= SSC_MATCHES_EMPTY_STRING;
5631 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5633 /* Positive Lookahead/lookbehind
5634 In this case we can do fixed string optimisation,
5635 but we must be careful about it. Note in the case of
5636 lookbehind the positions will be offset by the minimum
5637 length of the pattern, something we won't know about
5638 until after the recurse.
5640 SSize_t deltanext, fake = 0;
5644 /* We use SAVEFREEPV so that when the full compile
5645 is finished perl will clean up the allocated
5646 minlens when it's all done. This way we don't
5647 have to worry about freeing them when we know
5648 they wont be used, which would be a pain.
5651 Newx( minnextp, 1, SSize_t );
5652 SAVEFREEPV(minnextp);
5655 StructCopy(data, &data_fake, scan_data_t);
5656 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5659 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5660 data_fake.last_found=newSVsv(data->last_found);
5664 data_fake.last_closep = &fake;
5665 data_fake.flags = 0;
5666 data_fake.pos_delta = delta;
5668 data_fake.flags |= SF_IS_INF;
5669 if ( flags & SCF_DO_STCLASS && !scan->flags
5670 && OP(scan) == IFMATCH ) { /* Lookahead */
5671 ssc_init(pRExC_state, &intrnl);
5672 data_fake.start_class = &intrnl;
5673 f |= SCF_DO_STCLASS_AND;
5675 if (flags & SCF_WHILEM_VISITED_POS)
5676 f |= SCF_WHILEM_VISITED_POS;
5677 next = regnext(scan);
5678 nscan = NEXTOPER(NEXTOPER(scan));
5680 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5681 &deltanext, last, &data_fake,
5682 stopparen, recursed_depth, NULL,
5686 FAIL("Variable length lookbehind not implemented");
5688 else if (*minnextp > (I32)U8_MAX) {
5689 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5692 scan->flags = (U8)*minnextp;
5697 if (f & SCF_DO_STCLASS_AND) {
5698 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5699 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
5702 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5704 if (data_fake.flags & SF_HAS_EVAL)
5705 data->flags |= SF_HAS_EVAL;
5706 data->whilem_c = data_fake.whilem_c;
5707 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5708 if (RExC_rx->minlen<*minnextp)
5709 RExC_rx->minlen=*minnextp;
5710 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5711 SvREFCNT_dec_NN(data_fake.last_found);
5713 if ( data_fake.minlen_fixed != minlenp )
5715 data->offset_fixed= data_fake.offset_fixed;
5716 data->minlen_fixed= data_fake.minlen_fixed;
5717 data->lookbehind_fixed+= scan->flags;
5719 if ( data_fake.minlen_float != minlenp )
5721 data->minlen_float= data_fake.minlen_float;
5722 data->offset_float_min=data_fake.offset_float_min;
5723 data->offset_float_max=data_fake.offset_float_max;
5724 data->lookbehind_float+= scan->flags;
5731 else if (OP(scan) == OPEN) {
5732 if (stopparen != (I32)ARG(scan))
5735 else if (OP(scan) == CLOSE) {
5736 if (stopparen == (I32)ARG(scan)) {
5739 if ((I32)ARG(scan) == is_par) {
5740 next = regnext(scan);
5742 if ( next && (OP(next) != WHILEM) && next < last)
5743 is_par = 0; /* Disable optimization */
5746 *(data->last_closep) = ARG(scan);
5748 else if (OP(scan) == EVAL) {
5750 data->flags |= SF_HAS_EVAL;
5752 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5753 if (flags & SCF_DO_SUBSTR) {
5754 scan_commit(pRExC_state, data, minlenp, is_inf);
5755 flags &= ~SCF_DO_SUBSTR;
5757 if (data && OP(scan)==ACCEPT) {
5758 data->flags |= SCF_SEEN_ACCEPT;
5763 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5765 if (flags & SCF_DO_SUBSTR) {
5766 scan_commit(pRExC_state, data, minlenp, is_inf);
5767 data->longest = &(data->longest_float);
5769 is_inf = is_inf_internal = 1;
5770 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5771 ssc_anything(data->start_class);
5772 flags &= ~SCF_DO_STCLASS;
5774 else if (OP(scan) == GPOS) {
5775 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5776 !(delta || is_inf || (data && data->pos_delta)))
5778 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5779 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5780 if (RExC_rx->gofs < (STRLEN)min)
5781 RExC_rx->gofs = min;
5783 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5787 #ifdef TRIE_STUDY_OPT
5788 #ifdef FULL_TRIE_STUDY
5789 else if (PL_regkind[OP(scan)] == TRIE) {
5790 /* NOTE - There is similar code to this block above for handling
5791 BRANCH nodes on the initial study. If you change stuff here
5793 regnode *trie_node= scan;
5794 regnode *tail= regnext(scan);
5795 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5796 SSize_t max1 = 0, min1 = SSize_t_MAX;
5799 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5800 /* Cannot merge strings after this. */
5801 scan_commit(pRExC_state, data, minlenp, is_inf);
5803 if (flags & SCF_DO_STCLASS)
5804 ssc_init_zero(pRExC_state, &accum);
5810 const regnode *nextbranch= NULL;
5813 for ( word=1 ; word <= trie->wordcount ; word++)
5815 SSize_t deltanext=0, minnext=0, f = 0, fake;
5816 regnode_ssc this_class;
5818 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5820 data_fake.whilem_c = data->whilem_c;
5821 data_fake.last_closep = data->last_closep;
5824 data_fake.last_closep = &fake;
5825 data_fake.pos_delta = delta;
5826 if (flags & SCF_DO_STCLASS) {
5827 ssc_init(pRExC_state, &this_class);
5828 data_fake.start_class = &this_class;
5829 f = SCF_DO_STCLASS_AND;
5831 if (flags & SCF_WHILEM_VISITED_POS)
5832 f |= SCF_WHILEM_VISITED_POS;
5834 if (trie->jump[word]) {
5836 nextbranch = trie_node + trie->jump[0];
5837 scan= trie_node + trie->jump[word];
5838 /* We go from the jump point to the branch that follows
5839 it. Note this means we need the vestigal unused
5840 branches even though they arent otherwise used. */
5841 minnext = study_chunk(pRExC_state, &scan, minlenp,
5842 &deltanext, (regnode *)nextbranch, &data_fake,
5843 stopparen, recursed_depth, NULL, f,depth+1);
5845 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5846 nextbranch= regnext((regnode*)nextbranch);
5848 if (min1 > (SSize_t)(minnext + trie->minlen))
5849 min1 = minnext + trie->minlen;
5850 if (deltanext == SSize_t_MAX) {
5851 is_inf = is_inf_internal = 1;
5853 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5854 max1 = minnext + deltanext + trie->maxlen;
5856 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5858 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5859 if ( stopmin > min + min1)
5860 stopmin = min + min1;
5861 flags &= ~SCF_DO_SUBSTR;
5863 data->flags |= SCF_SEEN_ACCEPT;
5866 if (data_fake.flags & SF_HAS_EVAL)
5867 data->flags |= SF_HAS_EVAL;
5868 data->whilem_c = data_fake.whilem_c;
5870 if (flags & SCF_DO_STCLASS)
5871 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5874 if (flags & SCF_DO_SUBSTR) {
5875 data->pos_min += min1;
5876 data->pos_delta += max1 - min1;
5877 if (max1 != min1 || is_inf)
5878 data->longest = &(data->longest_float);
5881 if (delta != SSize_t_MAX)
5882 delta += max1 - min1;
5883 if (flags & SCF_DO_STCLASS_OR) {
5884 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5886 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5887 flags &= ~SCF_DO_STCLASS;
5890 else if (flags & SCF_DO_STCLASS_AND) {
5892 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5893 flags &= ~SCF_DO_STCLASS;
5896 /* Switch to OR mode: cache the old value of
5897 * data->start_class */
5899 StructCopy(data->start_class, and_withp, regnode_ssc);
5900 flags &= ~SCF_DO_STCLASS_AND;
5901 StructCopy(&accum, data->start_class, regnode_ssc);
5902 flags |= SCF_DO_STCLASS_OR;
5909 else if (PL_regkind[OP(scan)] == TRIE) {
5910 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5913 min += trie->minlen;
5914 delta += (trie->maxlen - trie->minlen);
5915 flags &= ~SCF_DO_STCLASS; /* xxx */
5916 if (flags & SCF_DO_SUBSTR) {
5917 /* Cannot expect anything... */
5918 scan_commit(pRExC_state, data, minlenp, is_inf);
5919 data->pos_min += trie->minlen;
5920 data->pos_delta += (trie->maxlen - trie->minlen);
5921 if (trie->maxlen != trie->minlen)
5922 data->longest = &(data->longest_float);
5924 if (trie->jump) /* no more substrings -- for now /grr*/
5925 flags &= ~SCF_DO_SUBSTR;
5927 #endif /* old or new */
5928 #endif /* TRIE_STUDY_OPT */
5930 /* Else: zero-length, ignore. */
5931 scan = regnext(scan);
5936 /* we need to unwind recursion. */
5939 DEBUG_STUDYDATA("frame-end:",data,depth);
5940 DEBUG_PEEP("fend", scan, depth);
5942 /* restore previous context */
5943 last = frame->last_regnode;
5944 scan = frame->next_regnode;
5945 stopparen = frame->stopparen;
5946 recursed_depth = frame->prev_recursed_depth;
5948 RExC_frame_last = frame->prev_frame;
5949 frame = frame->this_prev_frame;
5950 goto fake_study_recurse;
5954 DEBUG_STUDYDATA("pre-fin:",data,depth);
5957 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5959 if (flags & SCF_DO_SUBSTR && is_inf)
5960 data->pos_delta = SSize_t_MAX - data->pos_min;
5961 if (is_par > (I32)U8_MAX)
5963 if (is_par && pars==1 && data) {
5964 data->flags |= SF_IN_PAR;
5965 data->flags &= ~SF_HAS_PAR;
5967 else if (pars && data) {
5968 data->flags |= SF_HAS_PAR;
5969 data->flags &= ~SF_IN_PAR;
5971 if (flags & SCF_DO_STCLASS_OR)
5972 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5973 if (flags & SCF_TRIE_RESTUDY)
5974 data->flags |= SCF_TRIE_RESTUDY;
5976 DEBUG_STUDYDATA("post-fin:",data,depth);
5979 SSize_t final_minlen= min < stopmin ? min : stopmin;
5981 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
5982 if (final_minlen > SSize_t_MAX - delta)
5983 RExC_maxlen = SSize_t_MAX;
5984 else if (RExC_maxlen < final_minlen + delta)
5985 RExC_maxlen = final_minlen + delta;
5987 return final_minlen;
5989 NOT_REACHED; /* NOTREACHED */
5993 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5995 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5997 PERL_ARGS_ASSERT_ADD_DATA;
5999 Renewc(RExC_rxi->data,
6000 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6001 char, struct reg_data);
6003 Renew(RExC_rxi->data->what, count + n, U8);
6005 Newx(RExC_rxi->data->what, n, U8);
6006 RExC_rxi->data->count = count + n;
6007 Copy(s, RExC_rxi->data->what + count, n, U8);
6011 /*XXX: todo make this not included in a non debugging perl, but appears to be
6012 * used anyway there, in 'use re' */
6013 #ifndef PERL_IN_XSUB_RE
6015 Perl_reginitcolors(pTHX)
6017 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6019 char *t = savepv(s);
6023 t = strchr(t, '\t');
6029 PL_colors[i] = t = (char *)"";
6034 PL_colors[i++] = (char *)"";
6041 #ifdef TRIE_STUDY_OPT
6042 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6045 (data.flags & SCF_TRIE_RESTUDY) \
6053 #define CHECK_RESTUDY_GOTO_butfirst
6057 * pregcomp - compile a regular expression into internal code
6059 * Decides which engine's compiler to call based on the hint currently in
6063 #ifndef PERL_IN_XSUB_RE
6065 /* return the currently in-scope regex engine (or the default if none) */
6067 regexp_engine const *
6068 Perl_current_re_engine(pTHX)
6070 if (IN_PERL_COMPILETIME) {
6071 HV * const table = GvHV(PL_hintgv);
6074 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6075 return &PL_core_reg_engine;
6076 ptr = hv_fetchs(table, "regcomp", FALSE);
6077 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6078 return &PL_core_reg_engine;
6079 return INT2PTR(regexp_engine*,SvIV(*ptr));
6083 if (!PL_curcop->cop_hints_hash)
6084 return &PL_core_reg_engine;
6085 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6086 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6087 return &PL_core_reg_engine;
6088 return INT2PTR(regexp_engine*,SvIV(ptr));
6094 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6096 regexp_engine const *eng = current_re_engine();
6097 GET_RE_DEBUG_FLAGS_DECL;
6099 PERL_ARGS_ASSERT_PREGCOMP;
6101 /* Dispatch a request to compile a regexp to correct regexp engine. */
6103 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6106 return CALLREGCOMP_ENG(eng, pattern, flags);
6110 /* public(ish) entry point for the perl core's own regex compiling code.
6111 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6112 * pattern rather than a list of OPs, and uses the internal engine rather
6113 * than the current one */
6116 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6118 SV *pat = pattern; /* defeat constness! */
6119 PERL_ARGS_ASSERT_RE_COMPILE;
6120 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6121 #ifdef PERL_IN_XSUB_RE
6124 &PL_core_reg_engine,
6126 NULL, NULL, rx_flags, 0);
6131 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6135 if (--cbs->refcnt > 0)
6137 for (n = 0; n < cbs->count; n++) {
6138 REGEXP *rx = cbs->cb[n].src_regex;
6139 cbs->cb[n].src_regex = NULL;
6147 static struct reg_code_blocks *
6148 S_alloc_code_blocks(pTHX_ int ncode)
6150 struct reg_code_blocks *cbs;
6151 Newx(cbs, 1, struct reg_code_blocks);
6154 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6156 Newx(cbs->cb, ncode, struct reg_code_block);
6163 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6164 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6165 * point to the realloced string and length.
6167 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6171 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6172 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6174 U8 *const src = (U8*)*pat_p;
6179 GET_RE_DEBUG_FLAGS_DECL;
6181 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6182 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6184 Newx(dst, *plen_p * 2 + 1, U8);
6187 while (s < *plen_p) {
6188 append_utf8_from_native_byte(src[s], &d);
6190 if (n < num_code_blocks) {
6191 assert(pRExC_state->code_blocks);
6192 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6193 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6194 assert(*(d - 1) == '(');
6197 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6198 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6199 assert(*(d - 1) == ')');
6208 *pat_p = (char*) dst;
6210 RExC_orig_utf8 = RExC_utf8 = 1;
6215 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6216 * while recording any code block indices, and handling overloading,
6217 * nested qr// objects etc. If pat is null, it will allocate a new
6218 * string, or just return the first arg, if there's only one.
6220 * Returns the malloced/updated pat.
6221 * patternp and pat_count is the array of SVs to be concatted;
6222 * oplist is the optional list of ops that generated the SVs;
6223 * recompile_p is a pointer to a boolean that will be set if
6224 * the regex will need to be recompiled.
6225 * delim, if non-null is an SV that will be inserted between each element
6229 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6230 SV *pat, SV ** const patternp, int pat_count,
6231 OP *oplist, bool *recompile_p, SV *delim)
6235 bool use_delim = FALSE;
6236 bool alloced = FALSE;
6238 /* if we know we have at least two args, create an empty string,
6239 * then concatenate args to that. For no args, return an empty string */
6240 if (!pat && pat_count != 1) {
6246 for (svp = patternp; svp < patternp + pat_count; svp++) {
6249 STRLEN orig_patlen = 0;
6251 SV *msv = use_delim ? delim : *svp;
6252 if (!msv) msv = &PL_sv_undef;
6254 /* if we've got a delimiter, we go round the loop twice for each
6255 * svp slot (except the last), using the delimiter the second
6264 if (SvTYPE(msv) == SVt_PVAV) {
6265 /* we've encountered an interpolated array within
6266 * the pattern, e.g. /...@a..../. Expand the list of elements,
6267 * then recursively append elements.
6268 * The code in this block is based on S_pushav() */
6270 AV *const av = (AV*)msv;
6271 const SSize_t maxarg = AvFILL(av) + 1;
6275 assert(oplist->op_type == OP_PADAV
6276 || oplist->op_type == OP_RV2AV);
6277 oplist = OpSIBLING(oplist);
6280 if (SvRMAGICAL(av)) {
6283 Newx(array, maxarg, SV*);
6285 for (i=0; i < maxarg; i++) {
6286 SV ** const svp = av_fetch(av, i, FALSE);
6287 array[i] = svp ? *svp : &PL_sv_undef;
6291 array = AvARRAY(av);
6293 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6294 array, maxarg, NULL, recompile_p,
6296 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6302 /* we make the assumption here that each op in the list of
6303 * op_siblings maps to one SV pushed onto the stack,
6304 * except for code blocks, with have both an OP_NULL and
6306 * This allows us to match up the list of SVs against the
6307 * list of OPs to find the next code block.
6309 * Note that PUSHMARK PADSV PADSV ..
6311 * PADRANGE PADSV PADSV ..
6312 * so the alignment still works. */
6315 if (oplist->op_type == OP_NULL
6316 && (oplist->op_flags & OPf_SPECIAL))
6318 assert(n < pRExC_state->code_blocks->count);
6319 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6320 pRExC_state->code_blocks->cb[n].block = oplist;
6321 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6324 oplist = OpSIBLING(oplist); /* skip CONST */
6327 oplist = OpSIBLING(oplist);;
6330 /* apply magic and QR overloading to arg */
6333 if (SvROK(msv) && SvAMAGIC(msv)) {
6334 SV *sv = AMG_CALLunary(msv, regexp_amg);
6338 if (SvTYPE(sv) != SVt_REGEXP)
6339 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6344 /* try concatenation overload ... */
6345 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6346 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6349 /* overloading involved: all bets are off over literal
6350 * code. Pretend we haven't seen it */
6352 pRExC_state->code_blocks->count -= n;
6356 /* ... or failing that, try "" overload */
6357 while (SvAMAGIC(msv)
6358 && (sv = AMG_CALLunary(msv, string_amg))
6362 && SvRV(msv) == SvRV(sv))
6367 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6371 /* this is a partially unrolled
6372 * sv_catsv_nomg(pat, msv);
6373 * that allows us to adjust code block indices if
6376 char *dst = SvPV_force_nomg(pat, dlen);
6378 if (SvUTF8(msv) && !SvUTF8(pat)) {
6379 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6380 sv_setpvn(pat, dst, dlen);
6383 sv_catsv_nomg(pat, msv);
6387 /* We have only one SV to process, but we need to verify
6388 * it is properly null terminated or we will fail asserts
6389 * later. In theory we probably shouldn't get such SV's,
6390 * but if we do we should handle it gracefully. */
6391 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) ) {
6392 /* not a string, or a string with a trailing null */
6395 /* a string with no trailing null, we need to copy it
6396 * so it we have a trailing null */
6402 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
6405 /* extract any code blocks within any embedded qr//'s */
6406 if (rx && SvTYPE(rx) == SVt_REGEXP
6407 && RX_ENGINE((REGEXP*)rx)->op_comp)
6410 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6411 if (ri->code_blocks && ri->code_blocks->count) {
6413 /* the presence of an embedded qr// with code means
6414 * we should always recompile: the text of the
6415 * qr// may not have changed, but it may be a
6416 * different closure than last time */
6418 if (pRExC_state->code_blocks) {
6419 int new_count = pRExC_state->code_blocks->count
6420 + ri->code_blocks->count;
6421 Renew(pRExC_state->code_blocks->cb,
6422 new_count, struct reg_code_block);
6423 pRExC_state->code_blocks->count = new_count;
6426 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
6427 ri->code_blocks->count);
6429 for (i=0; i < ri->code_blocks->count; i++) {
6430 struct reg_code_block *src, *dst;
6431 STRLEN offset = orig_patlen
6432 + ReANY((REGEXP *)rx)->pre_prefix;
6433 assert(n < pRExC_state->code_blocks->count);
6434 src = &ri->code_blocks->cb[i];
6435 dst = &pRExC_state->code_blocks->cb[n];
6436 dst->start = src->start + offset;
6437 dst->end = src->end + offset;
6438 dst->block = src->block;
6439 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6448 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6457 /* see if there are any run-time code blocks in the pattern.
6458 * False positives are allowed */
6461 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6462 char *pat, STRLEN plen)
6467 PERL_UNUSED_CONTEXT;
6469 for (s = 0; s < plen; s++) {
6470 if ( pRExC_state->code_blocks
6471 && n < pRExC_state->code_blocks->count
6472 && s == pRExC_state->code_blocks->cb[n].start)
6474 s = pRExC_state->code_blocks->cb[n].end;
6478 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6480 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6482 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6489 /* Handle run-time code blocks. We will already have compiled any direct
6490 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6491 * copy of it, but with any literal code blocks blanked out and
6492 * appropriate chars escaped; then feed it into
6494 * eval "qr'modified_pattern'"
6498 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6502 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6504 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6505 * and merge them with any code blocks of the original regexp.
6507 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6508 * instead, just save the qr and return FALSE; this tells our caller that
6509 * the original pattern needs upgrading to utf8.
6513 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6514 char *pat, STRLEN plen)
6518 GET_RE_DEBUG_FLAGS_DECL;
6520 if (pRExC_state->runtime_code_qr) {
6521 /* this is the second time we've been called; this should
6522 * only happen if the main pattern got upgraded to utf8
6523 * during compilation; re-use the qr we compiled first time
6524 * round (which should be utf8 too)
6526 qr = pRExC_state->runtime_code_qr;
6527 pRExC_state->runtime_code_qr = NULL;
6528 assert(RExC_utf8 && SvUTF8(qr));
6534 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
6538 /* determine how many extra chars we need for ' and \ escaping */
6539 for (s = 0; s < plen; s++) {
6540 if (pat[s] == '\'' || pat[s] == '\\')
6544 Newx(newpat, newlen, char);
6546 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6548 for (s = 0; s < plen; s++) {
6549 if ( pRExC_state->code_blocks
6550 && n < pRExC_state->code_blocks->count
6551 && s == pRExC_state->code_blocks->cb[n].start)
6553 /* blank out literal code block */
6554 assert(pat[s] == '(');
6555 while (s <= pRExC_state->code_blocks->cb[n].end) {
6563 if (pat[s] == '\'' || pat[s] == '\\')
6568 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
6570 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
6576 Perl_re_printf( aTHX_
6577 "%sre-parsing pattern for runtime code:%s %s\n",
6578 PL_colors[4],PL_colors[5],newpat);
6581 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6587 PUSHSTACKi(PERLSI_REQUIRE);
6588 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6589 * parsing qr''; normally only q'' does this. It also alters
6591 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6592 SvREFCNT_dec_NN(sv);
6597 SV * const errsv = ERRSV;
6598 if (SvTRUE_NN(errsv))
6599 /* use croak_sv ? */
6600 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
6602 assert(SvROK(qr_ref));
6604 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6605 /* the leaving below frees the tmp qr_ref.
6606 * Give qr a life of its own */
6614 if (!RExC_utf8 && SvUTF8(qr)) {
6615 /* first time through; the pattern got upgraded; save the
6616 * qr for the next time through */
6617 assert(!pRExC_state->runtime_code_qr);
6618 pRExC_state->runtime_code_qr = qr;
6623 /* extract any code blocks within the returned qr// */
6626 /* merge the main (r1) and run-time (r2) code blocks into one */
6628 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6629 struct reg_code_block *new_block, *dst;
6630 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6634 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
6636 SvREFCNT_dec_NN(qr);
6640 if (!r1->code_blocks)
6641 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
6643 r1c = r1->code_blocks->count;
6644 r2c = r2->code_blocks->count;
6646 Newx(new_block, r1c + r2c, struct reg_code_block);
6650 while (i1 < r1c || i2 < r2c) {
6651 struct reg_code_block *src;
6655 src = &r2->code_blocks->cb[i2++];
6659 src = &r1->code_blocks->cb[i1++];
6660 else if ( r1->code_blocks->cb[i1].start
6661 < r2->code_blocks->cb[i2].start)
6663 src = &r1->code_blocks->cb[i1++];
6664 assert(src->end < r2->code_blocks->cb[i2].start);
6667 assert( r1->code_blocks->cb[i1].start
6668 > r2->code_blocks->cb[i2].start);
6669 src = &r2->code_blocks->cb[i2++];
6671 assert(src->end < r1->code_blocks->cb[i1].start);
6674 assert(pat[src->start] == '(');
6675 assert(pat[src->end] == ')');
6676 dst->start = src->start;
6677 dst->end = src->end;
6678 dst->block = src->block;
6679 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6683 r1->code_blocks->count += r2c;
6684 Safefree(r1->code_blocks->cb);
6685 r1->code_blocks->cb = new_block;
6688 SvREFCNT_dec_NN(qr);
6694 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6695 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6696 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6697 STRLEN longest_length, bool eol, bool meol)
6699 /* This is the common code for setting up the floating and fixed length
6700 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6701 * as to whether succeeded or not */
6706 if (! (longest_length
6707 || (eol /* Can't have SEOL and MULTI */
6708 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6710 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6711 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6716 /* copy the information about the longest from the reg_scan_data
6717 over to the program. */
6718 if (SvUTF8(sv_longest)) {
6719 *rx_utf8 = sv_longest;
6722 *rx_substr = sv_longest;
6725 /* end_shift is how many chars that must be matched that
6726 follow this item. We calculate it ahead of time as once the
6727 lookbehind offset is added in we lose the ability to correctly
6729 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6730 *rx_end_shift = ml - offset
6732 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
6734 + (SvTAIL(sv_longest) != 0)
6738 t = (eol/* Can't have SEOL and MULTI */
6739 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6740 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6746 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6747 * regular expression into internal code.
6748 * The pattern may be passed either as:
6749 * a list of SVs (patternp plus pat_count)
6750 * a list of OPs (expr)
6751 * If both are passed, the SV list is used, but the OP list indicates
6752 * which SVs are actually pre-compiled code blocks
6754 * The SVs in the list have magic and qr overloading applied to them (and
6755 * the list may be modified in-place with replacement SVs in the latter
6758 * If the pattern hasn't changed from old_re, then old_re will be
6761 * eng is the current engine. If that engine has an op_comp method, then
6762 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6763 * do the initial concatenation of arguments and pass on to the external
6766 * If is_bare_re is not null, set it to a boolean indicating whether the
6767 * arg list reduced (after overloading) to a single bare regex which has
6768 * been returned (i.e. /$qr/).
6770 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6772 * pm_flags contains the PMf_* flags, typically based on those from the
6773 * pm_flags field of the related PMOP. Currently we're only interested in
6774 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6776 * We can't allocate space until we know how big the compiled form will be,
6777 * but we can't compile it (and thus know how big it is) until we've got a
6778 * place to put the code. So we cheat: we compile it twice, once with code
6779 * generation turned off and size counting turned on, and once "for real".
6780 * This also means that we don't allocate space until we are sure that the
6781 * thing really will compile successfully, and we never have to move the
6782 * code and thus invalidate pointers into it. (Note that it has to be in
6783 * one piece because free() must be able to free it all.) [NB: not true in perl]
6785 * Beware that the optimization-preparation code in here knows about some
6786 * of the structure of the compiled regexp. [I'll say.]
6790 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6791 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6792 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6796 regexp_internal *ri;
6804 SV** new_patternp = patternp;
6806 /* these are all flags - maybe they should be turned
6807 * into a single int with different bit masks */
6808 I32 sawlookahead = 0;
6813 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6815 bool runtime_code = 0;
6817 RExC_state_t RExC_state;
6818 RExC_state_t * const pRExC_state = &RExC_state;
6819 #ifdef TRIE_STUDY_OPT
6821 RExC_state_t copyRExC_state;
6823 GET_RE_DEBUG_FLAGS_DECL;
6825 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6827 DEBUG_r(if (!PL_colorset) reginitcolors());
6829 /* Initialize these here instead of as-needed, as is quick and avoids
6830 * having to test them each time otherwise */
6831 if (! PL_AboveLatin1) {
6833 char * dump_len_string;
6836 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6837 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6838 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6839 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6840 PL_HasMultiCharFold =
6841 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6843 /* This is calculated here, because the Perl program that generates the
6844 * static global ones doesn't currently have access to
6845 * NUM_ANYOF_CODE_POINTS */
6846 PL_InBitmap = _new_invlist(2);
6847 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
6848 NUM_ANYOF_CODE_POINTS - 1);
6850 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
6851 if ( ! dump_len_string
6852 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
6854 PL_dump_re_max_len = 0;
6859 pRExC_state->warn_text = NULL;
6860 pRExC_state->code_blocks = NULL;
6863 *is_bare_re = FALSE;
6865 if (expr && (expr->op_type == OP_LIST ||
6866 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6867 /* allocate code_blocks if needed */
6871 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
6872 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6873 ncode++; /* count of DO blocks */
6876 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
6880 /* compile-time pattern with just OP_CONSTs and DO blocks */
6885 /* find how many CONSTs there are */
6888 if (expr->op_type == OP_CONST)
6891 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6892 if (o->op_type == OP_CONST)
6896 /* fake up an SV array */
6898 assert(!new_patternp);
6899 Newx(new_patternp, n, SV*);
6900 SAVEFREEPV(new_patternp);
6904 if (expr->op_type == OP_CONST)
6905 new_patternp[n] = cSVOPx_sv(expr);
6907 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6908 if (o->op_type == OP_CONST)
6909 new_patternp[n++] = cSVOPo_sv;
6914 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6915 "Assembling pattern from %d elements%s\n", pat_count,
6916 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6918 /* set expr to the first arg op */
6920 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
6921 && expr->op_type != OP_CONST)
6923 expr = cLISTOPx(expr)->op_first;
6924 assert( expr->op_type == OP_PUSHMARK
6925 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6926 || expr->op_type == OP_PADRANGE);
6927 expr = OpSIBLING(expr);
6930 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6931 expr, &recompile, NULL);
6933 /* handle bare (possibly after overloading) regex: foo =~ $re */
6938 if (SvTYPE(re) == SVt_REGEXP) {
6942 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6943 "Precompiled pattern%s\n",
6944 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6950 exp = SvPV_nomg(pat, plen);
6952 if (!eng->op_comp) {
6953 if ((SvUTF8(pat) && IN_BYTES)
6954 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6956 /* make a temporary copy; either to convert to bytes,
6957 * or to avoid repeating get-magic / overloaded stringify */
6958 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6959 (IN_BYTES ? 0 : SvUTF8(pat)));
6961 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6964 /* ignore the utf8ness if the pattern is 0 length */
6965 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6967 RExC_uni_semantics = 0;
6968 RExC_seen_unfolded_sharp_s = 0;
6969 RExC_contains_locale = 0;
6970 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
6971 RExC_study_started = 0;
6972 pRExC_state->runtime_code_qr = NULL;
6973 RExC_frame_head= NULL;
6974 RExC_frame_last= NULL;
6975 RExC_frame_count= 0;
6978 RExC_mysv1= sv_newmortal();
6979 RExC_mysv2= sv_newmortal();
6982 SV *dsv= sv_newmortal();
6983 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6984 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
6985 PL_colors[4],PL_colors[5],s);
6989 /* we jump here if we have to recompile, e.g., from upgrading the pattern
6992 if ((pm_flags & PMf_USE_RE_EVAL)
6993 /* this second condition covers the non-regex literal case,
6994 * i.e. $foo =~ '(?{})'. */
6995 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6997 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6999 /* return old regex if pattern hasn't changed */
7000 /* XXX: note in the below we have to check the flags as well as the
7003 * Things get a touch tricky as we have to compare the utf8 flag
7004 * independently from the compile flags. */
7008 && !!RX_UTF8(old_re) == !!RExC_utf8
7009 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7010 && RX_PRECOMP(old_re)
7011 && RX_PRELEN(old_re) == plen
7012 && memEQ(RX_PRECOMP(old_re), exp, plen)
7013 && !runtime_code /* with runtime code, always recompile */ )
7018 rx_flags = orig_rx_flags;
7020 if ( initial_charset == REGEX_DEPENDS_CHARSET
7021 && (RExC_utf8 ||RExC_uni_semantics))
7024 /* Set to use unicode semantics if the pattern is in utf8 and has the
7025 * 'depends' charset specified, as it means unicode when utf8 */
7026 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7030 RExC_precomp_adj = 0;
7031 RExC_flags = rx_flags;
7032 RExC_pm_flags = pm_flags;
7035 assert(TAINTING_get || !TAINT_get);
7037 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7039 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7040 /* whoops, we have a non-utf8 pattern, whilst run-time code
7041 * got compiled as utf8. Try again with a utf8 pattern */
7042 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7043 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7044 goto redo_first_pass;
7047 assert(!pRExC_state->runtime_code_qr);
7053 RExC_in_lookbehind = 0;
7054 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7057 RExC_recode_x_to_native = 0;
7059 RExC_in_multi_char_class = 0;
7061 /* First pass: determine size, legality. */
7063 RExC_start = RExC_adjusted_start = exp;
7064 RExC_end = exp + plen;
7065 RExC_precomp_end = RExC_end;
7070 RExC_emit = (regnode *) &RExC_emit_dummy;
7071 RExC_whilem_seen = 0;
7072 RExC_open_parens = NULL;
7073 RExC_close_parens = NULL;
7075 RExC_paren_names = NULL;
7077 RExC_paren_name_list = NULL;
7079 RExC_recurse = NULL;
7080 RExC_study_chunk_recursed = NULL;
7081 RExC_study_chunk_recursed_bytes= 0;
7082 RExC_recurse_count = 0;
7083 pRExC_state->code_index = 0;
7085 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7086 * code makes sure the final byte is an uncounted NUL. But should this
7087 * ever not be the case, lots of things could read beyond the end of the
7088 * buffer: loops like
7089 * while(isFOO(*RExC_parse)) RExC_parse++;
7090 * strchr(RExC_parse, "foo");
7091 * etc. So it is worth noting. */
7092 assert(*RExC_end == '\0');
7095 Perl_re_printf( aTHX_ "Starting first pass (sizing)\n");
7097 RExC_lastparse=NULL;
7100 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7101 /* It's possible to write a regexp in ascii that represents Unicode
7102 codepoints outside of the byte range, such as via \x{100}. If we
7103 detect such a sequence we have to convert the entire pattern to utf8
7104 and then recompile, as our sizing calculation will have been based
7105 on 1 byte == 1 character, but we will need to use utf8 to encode
7106 at least some part of the pattern, and therefore must convert the whole
7109 if (flags & RESTART_PASS1) {
7110 if (flags & NEED_UTF8) {
7111 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7112 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7115 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7116 "Need to redo pass 1\n"));
7119 goto redo_first_pass;
7121 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#" UVxf, (UV) flags);
7125 Perl_re_printf( aTHX_
7126 "Required size %" IVdf " nodes\n"
7127 "Starting second pass (creation)\n",
7130 RExC_lastparse=NULL;
7133 /* The first pass could have found things that force Unicode semantics */
7134 if ((RExC_utf8 || RExC_uni_semantics)
7135 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
7137 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7140 /* Small enough for pointer-storage convention?
7141 If extralen==0, this means that we will not need long jumps. */
7142 if (RExC_size >= 0x10000L && RExC_extralen)
7143 RExC_size += RExC_extralen;
7146 if (RExC_whilem_seen > 15)
7147 RExC_whilem_seen = 15;
7149 /* Allocate space and zero-initialize. Note, the two step process
7150 of zeroing when in debug mode, thus anything assigned has to
7151 happen after that */
7152 rx = (REGEXP*) newSV_type(SVt_REGEXP);
7154 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7155 char, regexp_internal);
7156 if ( r == NULL || ri == NULL )
7157 FAIL("Regexp out of space");
7159 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
7160 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7163 /* bulk initialize base fields with 0. */
7164 Zero(ri, sizeof(regexp_internal), char);
7167 /* non-zero initialization begins here */
7170 r->extflags = rx_flags;
7171 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7173 if (pm_flags & PMf_IS_QR) {
7174 ri->code_blocks = pRExC_state->code_blocks;
7175 if (ri->code_blocks)
7176 ri->code_blocks->refcnt++;
7180 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7181 bool has_charset = (get_regex_charset(r->extflags)
7182 != REGEX_DEPENDS_CHARSET);
7184 /* The caret is output if there are any defaults: if not all the STD
7185 * flags are set, or if no character set specifier is needed */
7187 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7189 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7190 == REG_RUN_ON_COMMENT_SEEN);
7191 U8 reganch = (U8)((r->extflags & RXf_PMf_STD_PMMOD)
7192 >> RXf_PMf_STD_PMMOD_SHIFT);
7193 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7196 /* We output all the necessary flags; we never output a minus, as all
7197 * those are defaults, so are
7198 * covered by the caret */
7199 const STRLEN wraplen = plen + has_p + has_runon
7200 + has_default /* If needs a caret */
7201 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7203 /* If needs a character set specifier */
7204 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7205 + (sizeof("(?:)") - 1);
7207 /* make sure PL_bitcount bounds not exceeded */
7208 assert(sizeof(STD_PAT_MODS) <= 8);
7210 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
7211 r->xpv_len_u.xpvlenu_pv = p;
7213 SvFLAGS(rx) |= SVf_UTF8;
7216 /* If a default, cover it using the caret */
7218 *p++= DEFAULT_PAT_MOD;
7222 const char* const name = get_regex_charset_name(r->extflags, &len);
7223 Copy(name, p, len, char);
7227 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7230 while((ch = *fptr++)) {
7238 Copy(RExC_precomp, p, plen, char);
7239 assert ((RX_WRAPPED(rx) - p) < 16);
7240 r->pre_prefix = p - RX_WRAPPED(rx);
7246 SvCUR_set(rx, p - RX_WRAPPED(rx));
7250 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
7252 /* Useful during FAIL. */
7253 #ifdef RE_TRACK_PATTERN_OFFSETS
7254 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
7255 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7256 "%s %" UVuf " bytes for offset annotations.\n",
7257 ri->u.offsets ? "Got" : "Couldn't get",
7258 (UV)((2*RExC_size+1) * sizeof(U32))));
7260 SetProgLen(ri,RExC_size);
7265 /* Second pass: emit code. */
7266 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7267 RExC_pm_flags = pm_flags;
7269 RExC_end = exp + plen;
7271 RExC_emit_start = ri->program;
7272 RExC_emit = ri->program;
7273 RExC_emit_bound = ri->program + RExC_size + 1;
7274 pRExC_state->code_index = 0;
7276 *((char*) RExC_emit++) = (char) REG_MAGIC;
7277 /* setup various meta data about recursion, this all requires
7278 * RExC_npar to be correctly set, and a bit later on we clear it */
7279 if (RExC_seen & REG_RECURSE_SEEN) {
7280 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
7281 "%*s%*s Setting up open/close parens\n",
7282 22, "| |", (int)(0 * 2 + 1), ""));
7284 /* setup RExC_open_parens, which holds the address of each
7285 * OPEN tag, and to make things simpler for the 0 index
7286 * the start of the program - this is used later for offsets */
7287 Newxz(RExC_open_parens, RExC_npar,regnode *);
7288 SAVEFREEPV(RExC_open_parens);
7289 RExC_open_parens[0] = RExC_emit;
7291 /* setup RExC_close_parens, which holds the address of each
7292 * CLOSE tag, and to make things simpler for the 0 index
7293 * the end of the program - this is used later for offsets */
7294 Newxz(RExC_close_parens, RExC_npar,regnode *);
7295 SAVEFREEPV(RExC_close_parens);
7296 /* we dont know where end op starts yet, so we dont
7297 * need to set RExC_close_parens[0] like we do RExC_open_parens[0] above */
7299 /* Note, RExC_npar is 1 + the number of parens in a pattern.
7300 * So its 1 if there are no parens. */
7301 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
7302 ((RExC_npar & 0x07) != 0);
7303 Newx(RExC_study_chunk_recursed,
7304 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7305 SAVEFREEPV(RExC_study_chunk_recursed);
7308 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7310 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#" UVxf, (UV) flags);
7313 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7316 /* XXXX To minimize changes to RE engine we always allocate
7317 3-units-long substrs field. */
7318 Newx(r->substrs, 1, struct reg_substr_data);
7319 if (RExC_recurse_count) {
7320 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
7321 SAVEFREEPV(RExC_recurse);
7325 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7327 RExC_study_chunk_recursed_count= 0;
7329 Zero(r->substrs, 1, struct reg_substr_data);
7330 if (RExC_study_chunk_recursed) {
7331 Zero(RExC_study_chunk_recursed,
7332 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7336 #ifdef TRIE_STUDY_OPT
7338 StructCopy(&zero_scan_data, &data, scan_data_t);
7339 copyRExC_state = RExC_state;
7342 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
7344 RExC_state = copyRExC_state;
7345 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
7346 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
7348 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
7349 StructCopy(&zero_scan_data, &data, scan_data_t);
7352 StructCopy(&zero_scan_data, &data, scan_data_t);
7355 /* Dig out information for optimizations. */
7356 r->extflags = RExC_flags; /* was pm_op */
7357 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
7360 SvUTF8_on(rx); /* Unicode in it? */
7361 ri->regstclass = NULL;
7362 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
7363 r->intflags |= PREGf_NAUGHTY;
7364 scan = ri->program + 1; /* First BRANCH. */
7366 /* testing for BRANCH here tells us whether there is "must appear"
7367 data in the pattern. If there is then we can use it for optimisations */
7368 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7371 STRLEN longest_float_length, longest_fixed_length;
7372 regnode_ssc ch_class; /* pointed to by data */
7374 SSize_t last_close = 0; /* pointed to by data */
7375 regnode *first= scan;
7376 regnode *first_next= regnext(first);
7378 * Skip introductions and multiplicators >= 1
7379 * so that we can extract the 'meat' of the pattern that must
7380 * match in the large if() sequence following.
7381 * NOTE that EXACT is NOT covered here, as it is normally
7382 * picked up by the optimiser separately.
7384 * This is unfortunate as the optimiser isnt handling lookahead
7385 * properly currently.
7388 while ((OP(first) == OPEN && (sawopen = 1)) ||
7389 /* An OR of *one* alternative - should not happen now. */
7390 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7391 /* for now we can't handle lookbehind IFMATCH*/
7392 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7393 (OP(first) == PLUS) ||
7394 (OP(first) == MINMOD) ||
7395 /* An {n,m} with n>0 */
7396 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7397 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7400 * the only op that could be a regnode is PLUS, all the rest
7401 * will be regnode_1 or regnode_2.
7403 * (yves doesn't think this is true)
7405 if (OP(first) == PLUS)
7408 if (OP(first) == MINMOD)
7410 first += regarglen[OP(first)];
7412 first = NEXTOPER(first);
7413 first_next= regnext(first);
7416 /* Starting-point info. */
7418 DEBUG_PEEP("first:",first,0);
7419 /* Ignore EXACT as we deal with it later. */
7420 if (PL_regkind[OP(first)] == EXACT) {
7421 if (OP(first) == EXACT || OP(first) == EXACTL)
7422 NOOP; /* Empty, get anchored substr later. */
7424 ri->regstclass = first;
7427 else if (PL_regkind[OP(first)] == TRIE &&
7428 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
7430 /* this can happen only on restudy */
7431 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
7434 else if (REGNODE_SIMPLE(OP(first)))
7435 ri->regstclass = first;
7436 else if (PL_regkind[OP(first)] == BOUND ||
7437 PL_regkind[OP(first)] == NBOUND)
7438 ri->regstclass = first;
7439 else if (PL_regkind[OP(first)] == BOL) {
7440 r->intflags |= (OP(first) == MBOL
7443 first = NEXTOPER(first);
7446 else if (OP(first) == GPOS) {
7447 r->intflags |= PREGf_ANCH_GPOS;
7448 first = NEXTOPER(first);
7451 else if ((!sawopen || !RExC_sawback) &&
7453 (OP(first) == STAR &&
7454 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7455 !(r->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
7457 /* turn .* into ^.* with an implied $*=1 */
7459 (OP(NEXTOPER(first)) == REG_ANY)
7462 r->intflags |= (type | PREGf_IMPLICIT);
7463 first = NEXTOPER(first);
7466 if (sawplus && !sawminmod && !sawlookahead
7467 && (!sawopen || !RExC_sawback)
7468 && !pRExC_state->code_blocks) /* May examine pos and $& */
7469 /* x+ must match at the 1st pos of run of x's */
7470 r->intflags |= PREGf_SKIP;
7472 /* Scan is after the zeroth branch, first is atomic matcher. */
7473 #ifdef TRIE_STUDY_OPT
7476 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7477 (IV)(first - scan + 1))
7481 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7482 (IV)(first - scan + 1))
7488 * If there's something expensive in the r.e., find the
7489 * longest literal string that must appear and make it the
7490 * regmust. Resolve ties in favor of later strings, since
7491 * the regstart check works with the beginning of the r.e.
7492 * and avoiding duplication strengthens checking. Not a
7493 * strong reason, but sufficient in the absence of others.
7494 * [Now we resolve ties in favor of the earlier string if
7495 * it happens that c_offset_min has been invalidated, since the
7496 * earlier string may buy us something the later one won't.]
7499 data.longest_fixed = newSVpvs("");
7500 data.longest_float = newSVpvs("");
7501 data.last_found = newSVpvs("");
7502 data.longest = &(data.longest_fixed);
7503 ENTER_with_name("study_chunk");
7504 SAVEFREESV(data.longest_fixed);
7505 SAVEFREESV(data.longest_float);
7506 SAVEFREESV(data.last_found);
7508 if (!ri->regstclass) {
7509 ssc_init(pRExC_state, &ch_class);
7510 data.start_class = &ch_class;
7511 stclass_flag = SCF_DO_STCLASS_AND;
7512 } else /* XXXX Check for BOUND? */
7514 data.last_closep = &last_close;
7517 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
7518 scan + RExC_size, /* Up to end */
7520 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
7521 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
7525 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
7528 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
7529 && data.last_start_min == 0 && data.last_end > 0
7530 && !RExC_seen_zerolen
7531 && !(RExC_seen & REG_VERBARG_SEEN)
7532 && !(RExC_seen & REG_GPOS_SEEN)
7534 r->extflags |= RXf_CHECK_ALL;
7536 scan_commit(pRExC_state, &data,&minlen,0);
7538 longest_float_length = CHR_SVLEN(data.longest_float);
7540 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
7541 && data.offset_fixed == data.offset_float_min
7542 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
7543 && S_setup_longest (aTHX_ pRExC_state,
7547 &(r->float_end_shift),
7548 data.lookbehind_float,
7549 data.offset_float_min,
7551 longest_float_length,
7552 cBOOL(data.flags & SF_FL_BEFORE_EOL),
7553 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
7555 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
7556 r->float_max_offset = data.offset_float_max;
7557 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
7558 r->float_max_offset -= data.lookbehind_float;
7559 SvREFCNT_inc_simple_void_NN(data.longest_float);
7562 r->float_substr = r->float_utf8 = NULL;
7563 longest_float_length = 0;
7566 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
7568 if (S_setup_longest (aTHX_ pRExC_state,
7570 &(r->anchored_utf8),
7571 &(r->anchored_substr),
7572 &(r->anchored_end_shift),
7573 data.lookbehind_fixed,
7576 longest_fixed_length,
7577 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
7578 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
7580 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
7581 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
7584 r->anchored_substr = r->anchored_utf8 = NULL;
7585 longest_fixed_length = 0;
7587 LEAVE_with_name("study_chunk");
7590 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
7591 ri->regstclass = NULL;
7593 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
7595 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7596 && is_ssc_worth_it(pRExC_state, data.start_class))
7598 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7600 ssc_finalize(pRExC_state, data.start_class);
7602 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7603 StructCopy(data.start_class,
7604 (regnode_ssc*)RExC_rxi->data->data[n],
7606 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7607 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7608 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7609 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7610 Perl_re_printf( aTHX_
7611 "synthetic stclass \"%s\".\n",
7612 SvPVX_const(sv));});
7613 data.start_class = NULL;
7616 /* A temporary algorithm prefers floated substr to fixed one to dig
7618 if (longest_fixed_length > longest_float_length) {
7619 r->substrs->check_ix = 0;
7620 r->check_end_shift = r->anchored_end_shift;
7621 r->check_substr = r->anchored_substr;
7622 r->check_utf8 = r->anchored_utf8;
7623 r->check_offset_min = r->check_offset_max = r->anchored_offset;
7624 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
7625 r->intflags |= PREGf_NOSCAN;
7628 r->substrs->check_ix = 1;
7629 r->check_end_shift = r->float_end_shift;
7630 r->check_substr = r->float_substr;
7631 r->check_utf8 = r->float_utf8;
7632 r->check_offset_min = r->float_min_offset;
7633 r->check_offset_max = r->float_max_offset;
7635 if ((r->check_substr || r->check_utf8) ) {
7636 r->extflags |= RXf_USE_INTUIT;
7637 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7638 r->extflags |= RXf_INTUIT_TAIL;
7640 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
7642 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7643 if ( (STRLEN)minlen < longest_float_length )
7644 minlen= longest_float_length;
7645 if ( (STRLEN)minlen < longest_fixed_length )
7646 minlen= longest_fixed_length;
7650 /* Several toplevels. Best we can is to set minlen. */
7652 regnode_ssc ch_class;
7653 SSize_t last_close = 0;
7655 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
7657 scan = ri->program + 1;
7658 ssc_init(pRExC_state, &ch_class);
7659 data.start_class = &ch_class;
7660 data.last_closep = &last_close;
7663 minlen = study_chunk(pRExC_state,
7664 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7665 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7666 ? SCF_TRIE_DOING_RESTUDY
7670 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7672 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7673 = r->float_substr = r->float_utf8 = NULL;
7675 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7676 && is_ssc_worth_it(pRExC_state, data.start_class))
7678 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7680 ssc_finalize(pRExC_state, data.start_class);
7682 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7683 StructCopy(data.start_class,
7684 (regnode_ssc*)RExC_rxi->data->data[n],
7686 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7687 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7688 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7689 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7690 Perl_re_printf( aTHX_
7691 "synthetic stclass \"%s\".\n",
7692 SvPVX_const(sv));});
7693 data.start_class = NULL;
7697 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7698 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7699 r->maxlen = REG_INFTY;
7702 r->maxlen = RExC_maxlen;
7705 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7706 the "real" pattern. */
7708 Perl_re_printf( aTHX_ "minlen: %" IVdf " r->minlen:%" IVdf " maxlen:%" IVdf "\n",
7709 (IV)minlen, (IV)r->minlen, (IV)RExC_maxlen);
7711 r->minlenret = minlen;
7712 if (r->minlen < minlen)
7715 if (RExC_seen & REG_RECURSE_SEEN ) {
7716 r->intflags |= PREGf_RECURSE_SEEN;
7717 Newxz(r->recurse_locinput, r->nparens + 1, char *);
7719 if (RExC_seen & REG_GPOS_SEEN)
7720 r->intflags |= PREGf_GPOS_SEEN;
7721 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7722 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7724 if (pRExC_state->code_blocks)
7725 r->extflags |= RXf_EVAL_SEEN;
7726 if (RExC_seen & REG_VERBARG_SEEN)
7728 r->intflags |= PREGf_VERBARG_SEEN;
7729 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7731 if (RExC_seen & REG_CUTGROUP_SEEN)
7732 r->intflags |= PREGf_CUTGROUP_SEEN;
7733 if (pm_flags & PMf_USE_RE_EVAL)
7734 r->intflags |= PREGf_USE_RE_EVAL;
7735 if (RExC_paren_names)
7736 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7738 RXp_PAREN_NAMES(r) = NULL;
7740 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7741 * so it can be used in pp.c */
7742 if (r->intflags & PREGf_ANCH)
7743 r->extflags |= RXf_IS_ANCHORED;
7747 /* this is used to identify "special" patterns that might result
7748 * in Perl NOT calling the regex engine and instead doing the match "itself",
7749 * particularly special cases in split//. By having the regex compiler
7750 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7751 * we avoid weird issues with equivalent patterns resulting in different behavior,
7752 * AND we allow non Perl engines to get the same optimizations by the setting the
7753 * flags appropriately - Yves */
7754 regnode *first = ri->program + 1;
7756 regnode *next = regnext(first);
7759 if (PL_regkind[fop] == NOTHING && nop == END)
7760 r->extflags |= RXf_NULL;
7761 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
7762 /* when fop is SBOL first->flags will be true only when it was
7763 * produced by parsing /\A/, and not when parsing /^/. This is
7764 * very important for the split code as there we want to
7765 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
7766 * See rt #122761 for more details. -- Yves */
7767 r->extflags |= RXf_START_ONLY;
7768 else if (fop == PLUS
7769 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7771 r->extflags |= RXf_WHITE;
7772 else if ( r->extflags & RXf_SPLIT
7773 && (fop == EXACT || fop == EXACTL)
7774 && STR_LEN(first) == 1
7775 && *(STRING(first)) == ' '
7777 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7781 if (RExC_contains_locale) {
7782 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7786 if (RExC_paren_names) {
7787 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7788 ri->data->data[ri->name_list_idx]
7789 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7792 ri->name_list_idx = 0;
7794 while ( RExC_recurse_count > 0 ) {
7795 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
7797 * This data structure is set up in study_chunk() and is used
7798 * to calculate the distance between a GOSUB regopcode and
7799 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
7802 * If for some reason someone writes code that optimises
7803 * away a GOSUB opcode then the assert should be changed to
7804 * an if(scan) to guard the ARG2L_SET() - Yves
7807 assert(scan && OP(scan) == GOSUB);
7808 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - scan );
7811 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7812 /* assume we don't need to swap parens around before we match */
7814 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
7815 (unsigned long)RExC_study_chunk_recursed_count);
7819 Perl_re_printf( aTHX_ "Final program:\n");
7822 #ifdef RE_TRACK_PATTERN_OFFSETS
7823 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7824 const STRLEN len = ri->u.offsets[0];
7826 GET_RE_DEBUG_FLAGS_DECL;
7827 Perl_re_printf( aTHX_
7828 "Offsets: [%" UVuf "]\n\t", (UV)ri->u.offsets[0]);
7829 for (i = 1; i <= len; i++) {
7830 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7831 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7832 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7834 Perl_re_printf( aTHX_ "\n");
7839 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7840 * by setting the regexp SV to readonly-only instead. If the
7841 * pattern's been recompiled, the USEDness should remain. */
7842 if (old_re && SvREADONLY(old_re))
7850 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7853 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7855 PERL_UNUSED_ARG(value);
7857 if (flags & RXapif_FETCH) {
7858 return reg_named_buff_fetch(rx, key, flags);
7859 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7860 Perl_croak_no_modify();
7862 } else if (flags & RXapif_EXISTS) {
7863 return reg_named_buff_exists(rx, key, flags)
7866 } else if (flags & RXapif_REGNAMES) {
7867 return reg_named_buff_all(rx, flags);
7868 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7869 return reg_named_buff_scalar(rx, flags);
7871 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7877 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7880 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7881 PERL_UNUSED_ARG(lastkey);
7883 if (flags & RXapif_FIRSTKEY)
7884 return reg_named_buff_firstkey(rx, flags);
7885 else if (flags & RXapif_NEXTKEY)
7886 return reg_named_buff_nextkey(rx, flags);
7888 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7895 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7899 struct regexp *const rx = ReANY(r);
7901 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7903 if (rx && RXp_PAREN_NAMES(rx)) {
7904 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7907 SV* sv_dat=HeVAL(he_str);
7908 I32 *nums=(I32*)SvPVX(sv_dat);
7909 AV * const retarray = (flags & RXapif_ALL) ? newAV() : NULL;
7910 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7911 if ((I32)(rx->nparens) >= nums[i]
7912 && rx->offs[nums[i]].start != -1
7913 && rx->offs[nums[i]].end != -1)
7916 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7921 ret = newSVsv(&PL_sv_undef);
7924 av_push(retarray, ret);
7927 return newRV_noinc(MUTABLE_SV(retarray));
7934 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7937 struct regexp *const rx = ReANY(r);
7939 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7941 if (rx && RXp_PAREN_NAMES(rx)) {
7942 if (flags & RXapif_ALL) {
7943 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7945 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7947 SvREFCNT_dec_NN(sv);
7959 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7961 struct regexp *const rx = ReANY(r);
7963 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7965 if ( rx && RXp_PAREN_NAMES(rx) ) {
7966 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7968 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7975 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7977 struct regexp *const rx = ReANY(r);
7978 GET_RE_DEBUG_FLAGS_DECL;
7980 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7982 if (rx && RXp_PAREN_NAMES(rx)) {
7983 HV *hv = RXp_PAREN_NAMES(rx);
7985 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7988 SV* sv_dat = HeVAL(temphe);
7989 I32 *nums = (I32*)SvPVX(sv_dat);
7990 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7991 if ((I32)(rx->lastparen) >= nums[i] &&
7992 rx->offs[nums[i]].start != -1 &&
7993 rx->offs[nums[i]].end != -1)
7999 if (parno || flags & RXapif_ALL) {
8000 return newSVhek(HeKEY_hek(temphe));
8008 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8013 struct regexp *const rx = ReANY(r);
8015 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8017 if (rx && RXp_PAREN_NAMES(rx)) {
8018 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8019 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8020 } else if (flags & RXapif_ONE) {
8021 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8022 av = MUTABLE_AV(SvRV(ret));
8023 length = av_tindex(av);
8024 SvREFCNT_dec_NN(ret);
8025 return newSViv(length + 1);
8027 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8032 return &PL_sv_undef;
8036 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8038 struct regexp *const rx = ReANY(r);
8041 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8043 if (rx && RXp_PAREN_NAMES(rx)) {
8044 HV *hv= RXp_PAREN_NAMES(rx);
8046 (void)hv_iterinit(hv);
8047 while ( (temphe = hv_iternext_flags(hv,0)) ) {
8050 SV* sv_dat = HeVAL(temphe);
8051 I32 *nums = (I32*)SvPVX(sv_dat);
8052 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8053 if ((I32)(rx->lastparen) >= nums[i] &&
8054 rx->offs[nums[i]].start != -1 &&
8055 rx->offs[nums[i]].end != -1)
8061 if (parno || flags & RXapif_ALL) {
8062 av_push(av, newSVhek(HeKEY_hek(temphe)));
8067 return newRV_noinc(MUTABLE_SV(av));
8071 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8074 struct regexp *const rx = ReANY(r);
8080 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8082 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8083 || n == RX_BUFF_IDX_CARET_FULLMATCH
8084 || n == RX_BUFF_IDX_CARET_POSTMATCH
8087 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8089 /* on something like
8092 * the KEEPCOPY is set on the PMOP rather than the regex */
8093 if (PL_curpm && r == PM_GETRE(PL_curpm))
8094 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8103 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8104 /* no need to distinguish between them any more */
8105 n = RX_BUFF_IDX_FULLMATCH;
8107 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8108 && rx->offs[0].start != -1)
8110 /* $`, ${^PREMATCH} */
8111 i = rx->offs[0].start;
8115 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8116 && rx->offs[0].end != -1)
8118 /* $', ${^POSTMATCH} */
8119 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8120 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8123 if ( 0 <= n && n <= (I32)rx->nparens &&
8124 (s1 = rx->offs[n].start) != -1 &&
8125 (t1 = rx->offs[n].end) != -1)
8127 /* $&, ${^MATCH}, $1 ... */
8129 s = rx->subbeg + s1 - rx->suboffset;
8134 assert(s >= rx->subbeg);
8135 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8137 #ifdef NO_TAINT_SUPPORT
8138 sv_setpvn(sv, s, i);
8140 const int oldtainted = TAINT_get;
8142 sv_setpvn(sv, s, i);
8143 TAINT_set(oldtainted);
8145 if (RXp_MATCH_UTF8(rx))
8150 if (RXp_MATCH_TAINTED(rx)) {
8151 if (SvTYPE(sv) >= SVt_PVMG) {
8152 MAGIC* const mg = SvMAGIC(sv);
8155 SvMAGIC_set(sv, mg->mg_moremagic);
8157 if ((mgt = SvMAGIC(sv))) {
8158 mg->mg_moremagic = mgt;
8159 SvMAGIC_set(sv, mg);
8176 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8177 SV const * const value)
8179 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8181 PERL_UNUSED_ARG(rx);
8182 PERL_UNUSED_ARG(paren);
8183 PERL_UNUSED_ARG(value);
8186 Perl_croak_no_modify();
8190 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8193 struct regexp *const rx = ReANY(r);
8197 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8199 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8200 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8201 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8204 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8206 /* on something like
8209 * the KEEPCOPY is set on the PMOP rather than the regex */
8210 if (PL_curpm && r == PM_GETRE(PL_curpm))
8211 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8217 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8219 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8220 case RX_BUFF_IDX_PREMATCH: /* $` */
8221 if (rx->offs[0].start != -1) {
8222 i = rx->offs[0].start;
8231 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8232 case RX_BUFF_IDX_POSTMATCH: /* $' */
8233 if (rx->offs[0].end != -1) {
8234 i = rx->sublen - rx->offs[0].end;
8236 s1 = rx->offs[0].end;
8243 default: /* $& / ${^MATCH}, $1, $2, ... */
8244 if (paren <= (I32)rx->nparens &&
8245 (s1 = rx->offs[paren].start) != -1 &&
8246 (t1 = rx->offs[paren].end) != -1)
8252 if (ckWARN(WARN_UNINITIALIZED))
8253 report_uninit((const SV *)sv);
8258 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8259 const char * const s = rx->subbeg - rx->suboffset + s1;
8264 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8271 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8273 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8274 PERL_UNUSED_ARG(rx);
8278 return newSVpvs("Regexp");
8281 /* Scans the name of a named buffer from the pattern.
8282 * If flags is REG_RSN_RETURN_NULL returns null.
8283 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8284 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8285 * to the parsed name as looked up in the RExC_paren_names hash.
8286 * If there is an error throws a vFAIL().. type exception.
8289 #define REG_RSN_RETURN_NULL 0
8290 #define REG_RSN_RETURN_NAME 1
8291 #define REG_RSN_RETURN_DATA 2
8294 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8296 char *name_start = RExC_parse;
8298 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8300 assert (RExC_parse <= RExC_end);
8301 if (RExC_parse == RExC_end) NOOP;
8302 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8303 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8304 * using do...while */
8307 RExC_parse += UTF8SKIP(RExC_parse);
8308 } while ( RExC_parse < RExC_end
8309 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8313 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8315 RExC_parse++; /* so the <- from the vFAIL is after the offending
8317 vFAIL("Group name must start with a non-digit word character");
8321 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8322 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8323 if ( flags == REG_RSN_RETURN_NAME)
8325 else if (flags==REG_RSN_RETURN_DATA) {
8328 if ( ! sv_name ) /* should not happen*/
8329 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8330 if (RExC_paren_names)
8331 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8333 sv_dat = HeVAL(he_str);
8335 vFAIL("Reference to nonexistent named group");
8339 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8340 (unsigned long) flags);
8342 NOT_REACHED; /* NOTREACHED */
8347 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8349 if (RExC_lastparse!=RExC_parse) { \
8350 Perl_re_printf( aTHX_ "%s", \
8351 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8352 RExC_end - RExC_parse, 16, \
8354 PERL_PV_ESCAPE_UNI_DETECT | \
8355 PERL_PV_PRETTY_ELLIPSES | \
8356 PERL_PV_PRETTY_LTGT | \
8357 PERL_PV_ESCAPE_RE | \
8358 PERL_PV_PRETTY_EXACTSIZE \
8362 Perl_re_printf( aTHX_ "%16s",""); \
8365 num = RExC_size + 1; \
8367 num=REG_NODE_NUM(RExC_emit); \
8368 if (RExC_lastnum!=num) \
8369 Perl_re_printf( aTHX_ "|%4d",num); \
8371 Perl_re_printf( aTHX_ "|%4s",""); \
8372 Perl_re_printf( aTHX_ "|%*s%-4s", \
8373 (int)((depth*2)), "", \
8377 RExC_lastparse=RExC_parse; \
8382 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8383 DEBUG_PARSE_MSG((funcname)); \
8384 Perl_re_printf( aTHX_ "%4s","\n"); \
8386 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8387 DEBUG_PARSE_MSG((funcname)); \
8388 Perl_re_printf( aTHX_ fmt "\n",args); \
8391 /* This section of code defines the inversion list object and its methods. The
8392 * interfaces are highly subject to change, so as much as possible is static to
8393 * this file. An inversion list is here implemented as a malloc'd C UV array
8394 * as an SVt_INVLIST scalar.
8396 * An inversion list for Unicode is an array of code points, sorted by ordinal
8397 * number. Each element gives the code point that begins a range that extends
8398 * up-to but not including the code point given by the next element. The final
8399 * element gives the first code point of a range that extends to the platform's
8400 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8401 * ...) give ranges whose code points are all in the inversion list. We say
8402 * that those ranges are in the set. The odd-numbered elements give ranges
8403 * whose code points are not in the inversion list, and hence not in the set.
8404 * Thus, element [0] is the first code point in the list. Element [1]
8405 * is the first code point beyond that not in the list; and element [2] is the
8406 * first code point beyond that that is in the list. In other words, the first
8407 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8408 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8409 * all code points in that range are not in the inversion list. The third
8410 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8411 * list, and so forth. Thus every element whose index is divisible by two
8412 * gives the beginning of a range that is in the list, and every element whose
8413 * index is not divisible by two gives the beginning of a range not in the
8414 * list. If the final element's index is divisible by two, the inversion list
8415 * extends to the platform's infinity; otherwise the highest code point in the
8416 * inversion list is the contents of that element minus 1.
8418 * A range that contains just a single code point N will look like
8420 * invlist[i+1] == N+1
8422 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8423 * impossible to represent, so element [i+1] is omitted. The single element
8425 * invlist[0] == UV_MAX
8426 * contains just UV_MAX, but is interpreted as matching to infinity.
8428 * Taking the complement (inverting) an inversion list is quite simple, if the
8429 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8430 * This implementation reserves an element at the beginning of each inversion
8431 * list to always contain 0; there is an additional flag in the header which
8432 * indicates if the list begins at the 0, or is offset to begin at the next
8433 * element. This means that the inversion list can be inverted without any
8434 * copying; just flip the flag.
8436 * More about inversion lists can be found in "Unicode Demystified"
8437 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8439 * The inversion list data structure is currently implemented as an SV pointing
8440 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8441 * array of UV whose memory management is automatically handled by the existing
8442 * facilities for SV's.
8444 * Some of the methods should always be private to the implementation, and some
8445 * should eventually be made public */
8447 /* The header definitions are in F<invlist_inline.h> */
8449 #ifndef PERL_IN_XSUB_RE
8451 PERL_STATIC_INLINE UV*
8452 S__invlist_array_init(SV* const invlist, const bool will_have_0)
8454 /* Returns a pointer to the first element in the inversion list's array.
8455 * This is called upon initialization of an inversion list. Where the
8456 * array begins depends on whether the list has the code point U+0000 in it
8457 * or not. The other parameter tells it whether the code that follows this
8458 * call is about to put a 0 in the inversion list or not. The first
8459 * element is either the element reserved for 0, if TRUE, or the element
8460 * after it, if FALSE */
8462 bool* offset = get_invlist_offset_addr(invlist);
8463 UV* zero_addr = (UV *) SvPVX(invlist);
8465 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8468 assert(! _invlist_len(invlist));
8472 /* 1^1 = 0; 1^0 = 1 */
8473 *offset = 1 ^ will_have_0;
8474 return zero_addr + *offset;
8479 PERL_STATIC_INLINE void
8480 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8482 /* Sets the current number of elements stored in the inversion list.
8483 * Updates SvCUR correspondingly */
8484 PERL_UNUSED_CONTEXT;
8485 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8487 assert(SvTYPE(invlist) == SVt_INVLIST);
8492 : TO_INTERNAL_SIZE(len + offset));
8493 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8496 #ifndef PERL_IN_XSUB_RE
8499 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
8501 /* Replaces the inversion list in 'dest' with the one from 'src'. It
8502 * steals the list from 'src', so 'src' is made to have a NULL list. This
8503 * is similar to what SvSetMagicSV() would do, if it were implemented on
8504 * inversion lists, though this routine avoids a copy */
8506 const UV src_len = _invlist_len(src);
8507 const bool src_offset = *get_invlist_offset_addr(src);
8508 const STRLEN src_byte_len = SvLEN(src);
8509 char * array = SvPVX(src);
8511 const int oldtainted = TAINT_get;
8513 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
8515 assert(SvTYPE(src) == SVt_INVLIST);
8516 assert(SvTYPE(dest) == SVt_INVLIST);
8517 assert(! invlist_is_iterating(src));
8518 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
8520 /* Make sure it ends in the right place with a NUL, as our inversion list
8521 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
8523 array[src_byte_len - 1] = '\0';
8525 TAINT_NOT; /* Otherwise it breaks */
8526 sv_usepvn_flags(dest,
8530 /* This flag is documented to cause a copy to be avoided */
8531 SV_HAS_TRAILING_NUL);
8532 TAINT_set(oldtainted);
8537 /* Finish up copying over the other fields in an inversion list */
8538 *get_invlist_offset_addr(dest) = src_offset;
8539 invlist_set_len(dest, src_len, src_offset);
8540 *get_invlist_previous_index_addr(dest) = 0;
8541 invlist_iterfinish(dest);
8544 PERL_STATIC_INLINE IV*
8545 S_get_invlist_previous_index_addr(SV* invlist)
8547 /* Return the address of the IV that is reserved to hold the cached index
8549 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8551 assert(SvTYPE(invlist) == SVt_INVLIST);
8553 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8556 PERL_STATIC_INLINE IV
8557 S_invlist_previous_index(SV* const invlist)
8559 /* Returns cached index of previous search */
8561 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8563 return *get_invlist_previous_index_addr(invlist);
8566 PERL_STATIC_INLINE void
8567 S_invlist_set_previous_index(SV* const invlist, const IV index)
8569 /* Caches <index> for later retrieval */
8571 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8573 assert(index == 0 || index < (int) _invlist_len(invlist));
8575 *get_invlist_previous_index_addr(invlist) = index;
8578 PERL_STATIC_INLINE void
8579 S_invlist_trim(SV* invlist)
8581 /* Free the not currently-being-used space in an inversion list */
8583 /* But don't free up the space needed for the 0 UV that is always at the
8584 * beginning of the list, nor the trailing NUL */
8585 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
8587 PERL_ARGS_ASSERT_INVLIST_TRIM;
8589 assert(SvTYPE(invlist) == SVt_INVLIST);
8591 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
8594 PERL_STATIC_INLINE void
8595 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
8597 PERL_ARGS_ASSERT_INVLIST_CLEAR;
8599 assert(SvTYPE(invlist) == SVt_INVLIST);
8601 invlist_set_len(invlist, 0, 0);
8602 invlist_trim(invlist);
8605 #endif /* ifndef PERL_IN_XSUB_RE */
8607 PERL_STATIC_INLINE bool
8608 S_invlist_is_iterating(SV* const invlist)
8610 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8612 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8615 #ifndef PERL_IN_XSUB_RE
8617 PERL_STATIC_INLINE UV
8618 S_invlist_max(SV* const invlist)
8620 /* Returns the maximum number of elements storable in the inversion list's
8621 * array, without having to realloc() */
8623 PERL_ARGS_ASSERT_INVLIST_MAX;
8625 assert(SvTYPE(invlist) == SVt_INVLIST);
8627 /* Assumes worst case, in which the 0 element is not counted in the
8628 * inversion list, so subtracts 1 for that */
8629 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8630 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8631 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8634 Perl__new_invlist(pTHX_ IV initial_size)
8637 /* Return a pointer to a newly constructed inversion list, with enough
8638 * space to store 'initial_size' elements. If that number is negative, a
8639 * system default is used instead */
8643 if (initial_size < 0) {
8647 /* Allocate the initial space */
8648 new_list = newSV_type(SVt_INVLIST);
8650 /* First 1 is in case the zero element isn't in the list; second 1 is for
8652 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8653 invlist_set_len(new_list, 0, 0);
8655 /* Force iterinit() to be used to get iteration to work */
8656 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8658 *get_invlist_previous_index_addr(new_list) = 0;
8664 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8666 /* Return a pointer to a newly constructed inversion list, initialized to
8667 * point to <list>, which has to be in the exact correct inversion list
8668 * form, including internal fields. Thus this is a dangerous routine that
8669 * should not be used in the wrong hands. The passed in 'list' contains
8670 * several header fields at the beginning that are not part of the
8671 * inversion list body proper */
8673 const STRLEN length = (STRLEN) list[0];
8674 const UV version_id = list[1];
8675 const bool offset = cBOOL(list[2]);
8676 #define HEADER_LENGTH 3
8677 /* If any of the above changes in any way, you must change HEADER_LENGTH
8678 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8679 * perl -E 'say int(rand 2**31-1)'
8681 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8682 data structure type, so that one being
8683 passed in can be validated to be an
8684 inversion list of the correct vintage.
8687 SV* invlist = newSV_type(SVt_INVLIST);
8689 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8691 if (version_id != INVLIST_VERSION_ID) {
8692 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8695 /* The generated array passed in includes header elements that aren't part
8696 * of the list proper, so start it just after them */
8697 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8699 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8700 shouldn't touch it */
8702 *(get_invlist_offset_addr(invlist)) = offset;
8704 /* The 'length' passed to us is the physical number of elements in the
8705 * inversion list. But if there is an offset the logical number is one
8707 invlist_set_len(invlist, length - offset, offset);
8709 invlist_set_previous_index(invlist, 0);
8711 /* Initialize the iteration pointer. */
8712 invlist_iterfinish(invlist);
8714 SvREADONLY_on(invlist);
8720 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8722 /* Grow the maximum size of an inversion list */
8724 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8726 assert(SvTYPE(invlist) == SVt_INVLIST);
8728 /* Add one to account for the zero element at the beginning which may not
8729 * be counted by the calling parameters */
8730 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8734 S__append_range_to_invlist(pTHX_ SV* const invlist,
8735 const UV start, const UV end)
8737 /* Subject to change or removal. Append the range from 'start' to 'end' at
8738 * the end of the inversion list. The range must be above any existing
8742 UV max = invlist_max(invlist);
8743 UV len = _invlist_len(invlist);
8746 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8748 if (len == 0) { /* Empty lists must be initialized */
8749 offset = start != 0;
8750 array = _invlist_array_init(invlist, ! offset);
8753 /* Here, the existing list is non-empty. The current max entry in the
8754 * list is generally the first value not in the set, except when the
8755 * set extends to the end of permissible values, in which case it is
8756 * the first entry in that final set, and so this call is an attempt to
8757 * append out-of-order */
8759 UV final_element = len - 1;
8760 array = invlist_array(invlist);
8761 if ( array[final_element] > start
8762 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8764 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",
8765 array[final_element], start,
8766 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8769 /* Here, it is a legal append. If the new range begins 1 above the end
8770 * of the range below it, it is extending the range below it, so the
8771 * new first value not in the set is one greater than the newly
8772 * extended range. */
8773 offset = *get_invlist_offset_addr(invlist);
8774 if (array[final_element] == start) {
8775 if (end != UV_MAX) {
8776 array[final_element] = end + 1;
8779 /* But if the end is the maximum representable on the machine,
8780 * assume that infinity was actually what was meant. Just let
8781 * the range that this would extend to have no end */
8782 invlist_set_len(invlist, len - 1, offset);
8788 /* Here the new range doesn't extend any existing set. Add it */
8790 len += 2; /* Includes an element each for the start and end of range */
8792 /* If wll overflow the existing space, extend, which may cause the array to
8795 invlist_extend(invlist, len);
8797 /* Have to set len here to avoid assert failure in invlist_array() */
8798 invlist_set_len(invlist, len, offset);
8800 array = invlist_array(invlist);
8803 invlist_set_len(invlist, len, offset);
8806 /* The next item on the list starts the range, the one after that is
8807 * one past the new range. */
8808 array[len - 2] = start;
8809 if (end != UV_MAX) {
8810 array[len - 1] = end + 1;
8813 /* But if the end is the maximum representable on the machine, just let
8814 * the range have no end */
8815 invlist_set_len(invlist, len - 1, offset);
8820 Perl__invlist_search(SV* const invlist, const UV cp)
8822 /* Searches the inversion list for the entry that contains the input code
8823 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8824 * return value is the index into the list's array of the range that
8825 * contains <cp>, that is, 'i' such that
8826 * array[i] <= cp < array[i+1]
8831 IV high = _invlist_len(invlist);
8832 const IV highest_element = high - 1;
8835 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8837 /* If list is empty, return failure. */
8842 /* (We can't get the array unless we know the list is non-empty) */
8843 array = invlist_array(invlist);
8845 mid = invlist_previous_index(invlist);
8847 if (mid > highest_element) {
8848 mid = highest_element;
8851 /* <mid> contains the cache of the result of the previous call to this
8852 * function (0 the first time). See if this call is for the same result,
8853 * or if it is for mid-1. This is under the theory that calls to this
8854 * function will often be for related code points that are near each other.
8855 * And benchmarks show that caching gives better results. We also test
8856 * here if the code point is within the bounds of the list. These tests
8857 * replace others that would have had to be made anyway to make sure that
8858 * the array bounds were not exceeded, and these give us extra information
8859 * at the same time */
8860 if (cp >= array[mid]) {
8861 if (cp >= array[highest_element]) {
8862 return highest_element;
8865 /* Here, array[mid] <= cp < array[highest_element]. This means that
8866 * the final element is not the answer, so can exclude it; it also
8867 * means that <mid> is not the final element, so can refer to 'mid + 1'
8869 if (cp < array[mid + 1]) {
8875 else { /* cp < aray[mid] */
8876 if (cp < array[0]) { /* Fail if outside the array */
8880 if (cp >= array[mid - 1]) {
8885 /* Binary search. What we are looking for is <i> such that
8886 * array[i] <= cp < array[i+1]
8887 * The loop below converges on the i+1. Note that there may not be an
8888 * (i+1)th element in the array, and things work nonetheless */
8889 while (low < high) {
8890 mid = (low + high) / 2;
8891 assert(mid <= highest_element);
8892 if (array[mid] <= cp) { /* cp >= array[mid] */
8895 /* We could do this extra test to exit the loop early.
8896 if (cp < array[low]) {
8901 else { /* cp < array[mid] */
8908 invlist_set_previous_index(invlist, high);
8913 Perl__invlist_populate_swatch(SV* const invlist,
8914 const UV start, const UV end, U8* swatch)
8916 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8917 * but is used when the swash has an inversion list. This makes this much
8918 * faster, as it uses a binary search instead of a linear one. This is
8919 * intimately tied to that function, and perhaps should be in utf8.c,
8920 * except it is intimately tied to inversion lists as well. It assumes
8921 * that <swatch> is all 0's on input */
8924 const IV len = _invlist_len(invlist);
8928 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8930 if (len == 0) { /* Empty inversion list */
8934 array = invlist_array(invlist);
8936 /* Find which element it is */
8937 i = _invlist_search(invlist, start);
8939 /* We populate from <start> to <end> */
8940 while (current < end) {
8943 /* The inversion list gives the results for every possible code point
8944 * after the first one in the list. Only those ranges whose index is
8945 * even are ones that the inversion list matches. For the odd ones,
8946 * and if the initial code point is not in the list, we have to skip
8947 * forward to the next element */
8948 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8950 if (i >= len) { /* Finished if beyond the end of the array */
8954 if (current >= end) { /* Finished if beyond the end of what we
8956 if (LIKELY(end < UV_MAX)) {
8960 /* We get here when the upper bound is the maximum
8961 * representable on the machine, and we are looking for just
8962 * that code point. Have to special case it */
8964 goto join_end_of_list;
8967 assert(current >= start);
8969 /* The current range ends one below the next one, except don't go past
8972 upper = (i < len && array[i] < end) ? array[i] : end;
8974 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8975 * for each code point in it */
8976 for (; current < upper; current++) {
8977 const STRLEN offset = (STRLEN)(current - start);
8978 swatch[offset >> 3] |= 1 << (offset & 7);
8983 /* Quit if at the end of the list */
8986 /* But first, have to deal with the highest possible code point on
8987 * the platform. The previous code assumes that <end> is one
8988 * beyond where we want to populate, but that is impossible at the
8989 * platform's infinity, so have to handle it specially */
8990 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8992 const STRLEN offset = (STRLEN)(end - start);
8993 swatch[offset >> 3] |= 1 << (offset & 7);
8998 /* Advance to the next range, which will be for code points not in the
9007 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9008 const bool complement_b, SV** output)
9010 /* Take the union of two inversion lists and point '*output' to it. On
9011 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9012 * even 'a' or 'b'). If to an inversion list, the contents of the original
9013 * list will be replaced by the union. The first list, 'a', may be
9014 * NULL, in which case a copy of the second list is placed in '*output'.
9015 * If 'complement_b' is TRUE, the union is taken of the complement
9016 * (inversion) of 'b' instead of b itself.
9018 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9019 * Richard Gillam, published by Addison-Wesley, and explained at some
9020 * length there. The preface says to incorporate its examples into your
9021 * code at your own risk.
9023 * The algorithm is like a merge sort. */
9025 const UV* array_a; /* a's array */
9027 UV len_a; /* length of a's array */
9030 SV* u; /* the resulting union */
9034 UV i_a = 0; /* current index into a's array */
9038 /* running count, as explained in the algorithm source book; items are
9039 * stopped accumulating and are output when the count changes to/from 0.
9040 * The count is incremented when we start a range that's in an input's set,
9041 * and decremented when we start a range that's not in a set. So this
9042 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9043 * and hence nothing goes into the union; 1, just one of the inputs is in
9044 * its set (and its current range gets added to the union); and 2 when both
9045 * inputs are in their sets. */
9048 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9050 assert(*output == NULL || SvTYPE(*output) == SVt_INVLIST);
9052 len_b = _invlist_len(b);
9055 /* Here, 'b' is empty, hence it's complement is all possible code
9056 * points. So if the union includes the complement of 'b', it includes
9057 * everything, and we need not even look at 'a'. It's easiest to
9058 * create a new inversion list that matches everything. */
9060 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9062 if (*output == NULL) { /* If the output didn't exist, just point it
9064 *output = everything;
9066 else { /* Otherwise, replace its contents with the new list */
9067 invlist_replace_list_destroys_src(*output, everything);
9068 SvREFCNT_dec_NN(everything);
9074 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9075 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9076 * output will be empty */
9078 if (a == NULL || _invlist_len(a) == 0) {
9079 if (*output == NULL) {
9080 *output = _new_invlist(0);
9083 invlist_clear(*output);
9088 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9089 * union. We can just return a copy of 'a' if '*output' doesn't point
9090 * to an existing list */
9091 if (*output == NULL) {
9092 *output = invlist_clone(a);
9096 /* If the output is to overwrite 'a', we have a no-op, as it's
9102 /* Here, '*output' is to be overwritten by 'a' */
9103 u = invlist_clone(a);
9104 invlist_replace_list_destroys_src(*output, u);
9110 /* Here 'b' is not empty. See about 'a' */
9112 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9114 /* Here, 'a' is empty (and b is not). That means the union will come
9115 * entirely from 'b'. If '*output' is NULL, we can directly return a
9116 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9119 SV ** dest = (*output == NULL) ? output : &u;
9120 *dest = invlist_clone(b);
9122 _invlist_invert(*dest);
9126 invlist_replace_list_destroys_src(*output, u);
9133 /* Here both lists exist and are non-empty */
9134 array_a = invlist_array(a);
9135 array_b = invlist_array(b);
9137 /* If are to take the union of 'a' with the complement of b, set it
9138 * up so are looking at b's complement. */
9141 /* To complement, we invert: if the first element is 0, remove it. To
9142 * do this, we just pretend the array starts one later */
9143 if (array_b[0] == 0) {
9149 /* But if the first element is not zero, we pretend the list starts
9150 * at the 0 that is always stored immediately before the array. */
9156 /* Size the union for the worst case: that the sets are completely
9158 u = _new_invlist(len_a + len_b);
9160 /* Will contain U+0000 if either component does */
9161 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9162 || (len_b > 0 && array_b[0] == 0));
9164 /* Go through each input list item by item, stopping when have exhausted
9166 while (i_a < len_a && i_b < len_b) {
9167 UV cp; /* The element to potentially add to the union's array */
9168 bool cp_in_set; /* is it in the the input list's set or not */
9170 /* We need to take one or the other of the two inputs for the union.
9171 * Since we are merging two sorted lists, we take the smaller of the
9172 * next items. In case of a tie, we take first the one that is in its
9173 * set. If we first took the one not in its set, it would decrement
9174 * the count, possibly to 0 which would cause it to be output as ending
9175 * the range, and the next time through we would take the same number,
9176 * and output it again as beginning the next range. By doing it the
9177 * opposite way, there is no possibility that the count will be
9178 * momentarily decremented to 0, and thus the two adjoining ranges will
9179 * be seamlessly merged. (In a tie and both are in the set or both not
9180 * in the set, it doesn't matter which we take first.) */
9181 if ( array_a[i_a] < array_b[i_b]
9182 || ( array_a[i_a] == array_b[i_b]
9183 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9185 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9186 cp = array_a[i_a++];
9189 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9190 cp = array_b[i_b++];
9193 /* Here, have chosen which of the two inputs to look at. Only output
9194 * if the running count changes to/from 0, which marks the
9195 * beginning/end of a range that's in the set */
9198 array_u[i_u++] = cp;
9205 array_u[i_u++] = cp;
9211 /* The loop above increments the index into exactly one of the input lists
9212 * each iteration, and ends when either index gets to its list end. That
9213 * means the other index is lower than its end, and so something is
9214 * remaining in that one. We decrement 'count', as explained below, if
9215 * that list is in its set. (i_a and i_b each currently index the element
9216 * beyond the one we care about.) */
9217 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9218 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9223 /* Above we decremented 'count' if the list that had unexamined elements in
9224 * it was in its set. This has made it so that 'count' being non-zero
9225 * means there isn't anything left to output; and 'count' equal to 0 means
9226 * that what is left to output is precisely that which is left in the
9227 * non-exhausted input list.
9229 * To see why, note first that the exhausted input obviously has nothing
9230 * left to add to the union. If it was in its set at its end, that means
9231 * the set extends from here to the platform's infinity, and hence so does
9232 * the union and the non-exhausted set is irrelevant. The exhausted set
9233 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9234 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9235 * 'count' remains at 1. This is consistent with the decremented 'count'
9236 * != 0 meaning there's nothing left to add to the union.
9238 * But if the exhausted input wasn't in its set, it contributed 0 to
9239 * 'count', and the rest of the union will be whatever the other input is.
9240 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9241 * otherwise it gets decremented to 0. This is consistent with 'count'
9242 * == 0 meaning the remainder of the union is whatever is left in the
9243 * non-exhausted list. */
9248 IV copy_count = len_a - i_a;
9249 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9250 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9252 else { /* The non-exhausted input is b */
9253 copy_count = len_b - i_b;
9254 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9256 len_u = i_u + copy_count;
9259 /* Set the result to the final length, which can change the pointer to
9260 * array_u, so re-find it. (Note that it is unlikely that this will
9261 * change, as we are shrinking the space, not enlarging it) */
9262 if (len_u != _invlist_len(u)) {
9263 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9265 array_u = invlist_array(u);
9268 if (*output == NULL) { /* Simply return the new inversion list */
9272 /* Otherwise, overwrite the inversion list that was in '*output'. We
9273 * could instead free '*output', and then set it to 'u', but experience
9274 * has shown [perl #127392] that if the input is a mortal, we can get a
9275 * huge build-up of these during regex compilation before they get
9277 invlist_replace_list_destroys_src(*output, u);
9285 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9286 const bool complement_b, SV** i)
9288 /* Take the intersection of two inversion lists and point '*i' to it. On
9289 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9290 * even 'a' or 'b'). If to an inversion list, the contents of the original
9291 * list will be replaced by the intersection. The first list, 'a', may be
9292 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9293 * TRUE, the result will be the intersection of 'a' and the complement (or
9294 * inversion) of 'b' instead of 'b' directly.
9296 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9297 * Richard Gillam, published by Addison-Wesley, and explained at some
9298 * length there. The preface says to incorporate its examples into your
9299 * code at your own risk. In fact, it had bugs
9301 * The algorithm is like a merge sort, and is essentially the same as the
9305 const UV* array_a; /* a's array */
9307 UV len_a; /* length of a's array */
9310 SV* r; /* the resulting intersection */
9314 UV i_a = 0; /* current index into a's array */
9318 /* running count of how many of the two inputs are postitioned at ranges
9319 * that are in their sets. As explained in the algorithm source book,
9320 * items are stopped accumulating and are output when the count changes
9321 * to/from 2. The count is incremented when we start a range that's in an
9322 * input's set, and decremented when we start a range that's not in a set.
9323 * Only when it is 2 are we in the intersection. */
9326 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9328 assert(*i == NULL || SvTYPE(*i) == SVt_INVLIST);
9330 /* Special case if either one is empty */
9331 len_a = (a == NULL) ? 0 : _invlist_len(a);
9332 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9333 if (len_a != 0 && complement_b) {
9335 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9336 * must be empty. Here, also we are using 'b's complement, which
9337 * hence must be every possible code point. Thus the intersection
9340 if (*i == a) { /* No-op */
9345 *i = invlist_clone(a);
9349 r = invlist_clone(a);
9350 invlist_replace_list_destroys_src(*i, r);
9355 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9356 * intersection must be empty */
9358 *i = _new_invlist(0);
9366 /* Here both lists exist and are non-empty */
9367 array_a = invlist_array(a);
9368 array_b = invlist_array(b);
9370 /* If are to take the intersection of 'a' with the complement of b, set it
9371 * up so are looking at b's complement. */
9374 /* To complement, we invert: if the first element is 0, remove it. To
9375 * do this, we just pretend the array starts one later */
9376 if (array_b[0] == 0) {
9382 /* But if the first element is not zero, we pretend the list starts
9383 * at the 0 that is always stored immediately before the array. */
9389 /* Size the intersection for the worst case: that the intersection ends up
9390 * fragmenting everything to be completely disjoint */
9391 r= _new_invlist(len_a + len_b);
9393 /* Will contain U+0000 iff both components do */
9394 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9395 && len_b > 0 && array_b[0] == 0);
9397 /* Go through each list item by item, stopping when have exhausted one of
9399 while (i_a < len_a && i_b < len_b) {
9400 UV cp; /* The element to potentially add to the intersection's
9402 bool cp_in_set; /* Is it in the input list's set or not */
9404 /* We need to take one or the other of the two inputs for the
9405 * intersection. Since we are merging two sorted lists, we take the
9406 * smaller of the next items. In case of a tie, we take first the one
9407 * that is not in its set (a difference from the union algorithm). If
9408 * we first took the one in its set, it would increment the count,
9409 * possibly to 2 which would cause it to be output as starting a range
9410 * in the intersection, and the next time through we would take that
9411 * same number, and output it again as ending the set. By doing the
9412 * opposite of this, there is no possibility that the count will be
9413 * momentarily incremented to 2. (In a tie and both are in the set or
9414 * both not in the set, it doesn't matter which we take first.) */
9415 if ( array_a[i_a] < array_b[i_b]
9416 || ( array_a[i_a] == array_b[i_b]
9417 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9419 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9420 cp = array_a[i_a++];
9423 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9427 /* Here, have chosen which of the two inputs to look at. Only output
9428 * if the running count changes to/from 2, which marks the
9429 * beginning/end of a range that's in the intersection */
9433 array_r[i_r++] = cp;
9438 array_r[i_r++] = cp;
9445 /* The loop above increments the index into exactly one of the input lists
9446 * each iteration, and ends when either index gets to its list end. That
9447 * means the other index is lower than its end, and so something is
9448 * remaining in that one. We increment 'count', as explained below, if the
9449 * exhausted list was in its set. (i_a and i_b each currently index the
9450 * element beyond the one we care about.) */
9451 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9452 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9457 /* Above we incremented 'count' if the exhausted list was in its set. This
9458 * has made it so that 'count' being below 2 means there is nothing left to
9459 * output; otheriwse what's left to add to the intersection is precisely
9460 * that which is left in the non-exhausted input list.
9462 * To see why, note first that the exhausted input obviously has nothing
9463 * left to affect the intersection. If it was in its set at its end, that
9464 * means the set extends from here to the platform's infinity, and hence
9465 * anything in the non-exhausted's list will be in the intersection, and
9466 * anything not in it won't be. Hence, the rest of the intersection is
9467 * precisely what's in the non-exhausted list The exhausted set also
9468 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9469 * it means 'count' is now at least 2. This is consistent with the
9470 * incremented 'count' being >= 2 means to add the non-exhausted list to
9473 * But if the exhausted input wasn't in its set, it contributed 0 to
9474 * 'count', and the intersection can't include anything further; the
9475 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9476 * incremented. This is consistent with 'count' being < 2 meaning nothing
9477 * further to add to the intersection. */
9478 if (count < 2) { /* Nothing left to put in the intersection. */
9481 else { /* copy the non-exhausted list, unchanged. */
9482 IV copy_count = len_a - i_a;
9483 if (copy_count > 0) { /* a is the one with stuff left */
9484 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9486 else { /* b is the one with stuff left */
9487 copy_count = len_b - i_b;
9488 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9490 len_r = i_r + copy_count;
9493 /* Set the result to the final length, which can change the pointer to
9494 * array_r, so re-find it. (Note that it is unlikely that this will
9495 * change, as we are shrinking the space, not enlarging it) */
9496 if (len_r != _invlist_len(r)) {
9497 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9499 array_r = invlist_array(r);
9502 if (*i == NULL) { /* Simply return the calculated intersection */
9505 else { /* Otherwise, replace the existing inversion list in '*i'. We could
9506 instead free '*i', and then set it to 'r', but experience has
9507 shown [perl #127392] that if the input is a mortal, we can get a
9508 huge build-up of these during regex compilation before they get
9511 invlist_replace_list_destroys_src(*i, r);
9523 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9525 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9526 * set. A pointer to the inversion list is returned. This may actually be
9527 * a new list, in which case the passed in one has been destroyed. The
9528 * passed-in inversion list can be NULL, in which case a new one is created
9529 * with just the one range in it. The new list is not necessarily
9530 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9531 * result of this function. The gain would not be large, and in many
9532 * cases, this is called multiple times on a single inversion list, so
9533 * anything freed may almost immediately be needed again.
9535 * This used to mostly call the 'union' routine, but that is much more
9536 * heavyweight than really needed for a single range addition */
9538 UV* array; /* The array implementing the inversion list */
9539 UV len; /* How many elements in 'array' */
9540 SSize_t i_s; /* index into the invlist array where 'start'
9542 SSize_t i_e = 0; /* And the index where 'end' should go */
9543 UV cur_highest; /* The highest code point in the inversion list
9544 upon entry to this function */
9546 /* This range becomes the whole inversion list if none already existed */
9547 if (invlist == NULL) {
9548 invlist = _new_invlist(2);
9549 _append_range_to_invlist(invlist, start, end);
9553 /* Likewise, if the inversion list is currently empty */
9554 len = _invlist_len(invlist);
9556 _append_range_to_invlist(invlist, start, end);
9560 /* Starting here, we have to know the internals of the list */
9561 array = invlist_array(invlist);
9563 /* If the new range ends higher than the current highest ... */
9564 cur_highest = invlist_highest(invlist);
9565 if (end > cur_highest) {
9567 /* If the whole range is higher, we can just append it */
9568 if (start > cur_highest) {
9569 _append_range_to_invlist(invlist, start, end);
9573 /* Otherwise, add the portion that is higher ... */
9574 _append_range_to_invlist(invlist, cur_highest + 1, end);
9576 /* ... and continue on below to handle the rest. As a result of the
9577 * above append, we know that the index of the end of the range is the
9578 * final even numbered one of the array. Recall that the final element
9579 * always starts a range that extends to infinity. If that range is in
9580 * the set (meaning the set goes from here to infinity), it will be an
9581 * even index, but if it isn't in the set, it's odd, and the final
9582 * range in the set is one less, which is even. */
9583 if (end == UV_MAX) {
9591 /* We have dealt with appending, now see about prepending. If the new
9592 * range starts lower than the current lowest ... */
9593 if (start < array[0]) {
9595 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
9596 * Let the union code handle it, rather than having to know the
9597 * trickiness in two code places. */
9598 if (UNLIKELY(start == 0)) {
9601 range_invlist = _new_invlist(2);
9602 _append_range_to_invlist(range_invlist, start, end);
9604 _invlist_union(invlist, range_invlist, &invlist);
9606 SvREFCNT_dec_NN(range_invlist);
9611 /* If the whole new range comes before the first entry, and doesn't
9612 * extend it, we have to insert it as an additional range */
9613 if (end < array[0] - 1) {
9615 goto splice_in_new_range;
9618 /* Here the new range adjoins the existing first range, extending it
9622 /* And continue on below to handle the rest. We know that the index of
9623 * the beginning of the range is the first one of the array */
9626 else { /* Not prepending any part of the new range to the existing list.
9627 * Find where in the list it should go. This finds i_s, such that:
9628 * invlist[i_s] <= start < array[i_s+1]
9630 i_s = _invlist_search(invlist, start);
9633 /* At this point, any extending before the beginning of the inversion list
9634 * and/or after the end has been done. This has made it so that, in the
9635 * code below, each endpoint of the new range is either in a range that is
9636 * in the set, or is in a gap between two ranges that are. This means we
9637 * don't have to worry about exceeding the array bounds.
9639 * Find where in the list the new range ends (but we can skip this if we
9640 * have already determined what it is, or if it will be the same as i_s,
9641 * which we already have computed) */
9643 i_e = (start == end)
9645 : _invlist_search(invlist, end);
9648 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
9649 * is a range that goes to infinity there is no element at invlist[i_e+1],
9650 * so only the first relation holds. */
9652 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9654 /* Here, the ranges on either side of the beginning of the new range
9655 * are in the set, and this range starts in the gap between them.
9657 * The new range extends the range above it downwards if the new range
9658 * ends at or above that range's start */
9659 const bool extends_the_range_above = ( end == UV_MAX
9660 || end + 1 >= array[i_s+1]);
9662 /* The new range extends the range below it upwards if it begins just
9663 * after where that range ends */
9664 if (start == array[i_s]) {
9666 /* If the new range fills the entire gap between the other ranges,
9667 * they will get merged together. Other ranges may also get
9668 * merged, depending on how many of them the new range spans. In
9669 * the general case, we do the merge later, just once, after we
9670 * figure out how many to merge. But in the case where the new
9671 * range exactly spans just this one gap (possibly extending into
9672 * the one above), we do the merge here, and an early exit. This
9673 * is done here to avoid having to special case later. */
9674 if (i_e - i_s <= 1) {
9676 /* If i_e - i_s == 1, it means that the new range terminates
9677 * within the range above, and hence 'extends_the_range_above'
9678 * must be true. (If the range above it extends to infinity,
9679 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
9680 * will be 0, so no harm done.) */
9681 if (extends_the_range_above) {
9682 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
9683 invlist_set_len(invlist,
9685 *(get_invlist_offset_addr(invlist)));
9689 /* Here, i_e must == i_s. We keep them in sync, as they apply
9690 * to the same range, and below we are about to decrement i_s
9695 /* Here, the new range is adjacent to the one below. (It may also
9696 * span beyond the range above, but that will get resolved later.)
9697 * Extend the range below to include this one. */
9698 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
9702 else if (extends_the_range_above) {
9704 /* Here the new range only extends the range above it, but not the
9705 * one below. It merges with the one above. Again, we keep i_e
9706 * and i_s in sync if they point to the same range */
9715 /* Here, we've dealt with the new range start extending any adjoining
9718 * If the new range extends to infinity, it is now the final one,
9719 * regardless of what was there before */
9720 if (UNLIKELY(end == UV_MAX)) {
9721 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
9725 /* If i_e started as == i_s, it has also been dealt with,
9726 * and been updated to the new i_s, which will fail the following if */
9727 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
9729 /* Here, the ranges on either side of the end of the new range are in
9730 * the set, and this range ends in the gap between them.
9732 * If this range is adjacent to (hence extends) the range above it, it
9733 * becomes part of that range; likewise if it extends the range below,
9734 * it becomes part of that range */
9735 if (end + 1 == array[i_e+1]) {
9739 else if (start <= array[i_e]) {
9740 array[i_e] = end + 1;
9747 /* If the range fits entirely in an existing range (as possibly already
9748 * extended above), it doesn't add anything new */
9749 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9753 /* Here, no part of the range is in the list. Must add it. It will
9754 * occupy 2 more slots */
9755 splice_in_new_range:
9757 invlist_extend(invlist, len + 2);
9758 array = invlist_array(invlist);
9759 /* Move the rest of the array down two slots. Don't include any
9761 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
9763 /* Do the actual splice */
9764 array[i_e+1] = start;
9765 array[i_e+2] = end + 1;
9766 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
9770 /* Here the new range crossed the boundaries of a pre-existing range. The
9771 * code above has adjusted things so that both ends are in ranges that are
9772 * in the set. This means everything in between must also be in the set.
9773 * Just squash things together */
9774 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
9775 invlist_set_len(invlist,
9777 *(get_invlist_offset_addr(invlist)));
9783 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9784 UV** other_elements_ptr)
9786 /* Create and return an inversion list whose contents are to be populated
9787 * by the caller. The caller gives the number of elements (in 'size') and
9788 * the very first element ('element0'). This function will set
9789 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9792 * Obviously there is some trust involved that the caller will properly
9793 * fill in the other elements of the array.
9795 * (The first element needs to be passed in, as the underlying code does
9796 * things differently depending on whether it is zero or non-zero) */
9798 SV* invlist = _new_invlist(size);
9801 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9803 invlist = add_cp_to_invlist(invlist, element0);
9804 offset = *get_invlist_offset_addr(invlist);
9806 invlist_set_len(invlist, size, offset);
9807 *other_elements_ptr = invlist_array(invlist) + 1;
9813 PERL_STATIC_INLINE SV*
9814 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9815 return _add_range_to_invlist(invlist, cp, cp);
9818 #ifndef PERL_IN_XSUB_RE
9820 Perl__invlist_invert(pTHX_ SV* const invlist)
9822 /* Complement the input inversion list. This adds a 0 if the list didn't
9823 * have a zero; removes it otherwise. As described above, the data
9824 * structure is set up so that this is very efficient */
9826 PERL_ARGS_ASSERT__INVLIST_INVERT;
9828 assert(! invlist_is_iterating(invlist));
9830 /* The inverse of matching nothing is matching everything */
9831 if (_invlist_len(invlist) == 0) {
9832 _append_range_to_invlist(invlist, 0, UV_MAX);
9836 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9841 PERL_STATIC_INLINE SV*
9842 S_invlist_clone(pTHX_ SV* const invlist)
9845 /* Return a new inversion list that is a copy of the input one, which is
9846 * unchanged. The new list will not be mortal even if the old one was. */
9848 /* Need to allocate extra space to accommodate Perl's addition of a
9849 * trailing NUL to SvPV's, since it thinks they are always strings */
9850 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9851 STRLEN physical_length = SvCUR(invlist);
9852 bool offset = *(get_invlist_offset_addr(invlist));
9854 PERL_ARGS_ASSERT_INVLIST_CLONE;
9856 *(get_invlist_offset_addr(new_invlist)) = offset;
9857 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9858 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9863 PERL_STATIC_INLINE STRLEN*
9864 S_get_invlist_iter_addr(SV* invlist)
9866 /* Return the address of the UV that contains the current iteration
9869 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9871 assert(SvTYPE(invlist) == SVt_INVLIST);
9873 return &(((XINVLIST*) SvANY(invlist))->iterator);
9876 PERL_STATIC_INLINE void
9877 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9879 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9881 *get_invlist_iter_addr(invlist) = 0;
9884 PERL_STATIC_INLINE void
9885 S_invlist_iterfinish(SV* invlist)
9887 /* Terminate iterator for invlist. This is to catch development errors.
9888 * Any iteration that is interrupted before completed should call this
9889 * function. Functions that add code points anywhere else but to the end
9890 * of an inversion list assert that they are not in the middle of an
9891 * iteration. If they were, the addition would make the iteration
9892 * problematical: if the iteration hadn't reached the place where things
9893 * were being added, it would be ok */
9895 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9897 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9901 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9903 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9904 * This call sets in <*start> and <*end>, the next range in <invlist>.
9905 * Returns <TRUE> if successful and the next call will return the next
9906 * range; <FALSE> if was already at the end of the list. If the latter,
9907 * <*start> and <*end> are unchanged, and the next call to this function
9908 * will start over at the beginning of the list */
9910 STRLEN* pos = get_invlist_iter_addr(invlist);
9911 UV len = _invlist_len(invlist);
9914 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9917 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9921 array = invlist_array(invlist);
9923 *start = array[(*pos)++];
9929 *end = array[(*pos)++] - 1;
9935 PERL_STATIC_INLINE UV
9936 S_invlist_highest(SV* const invlist)
9938 /* Returns the highest code point that matches an inversion list. This API
9939 * has an ambiguity, as it returns 0 under either the highest is actually
9940 * 0, or if the list is empty. If this distinction matters to you, check
9941 * for emptiness before calling this function */
9943 UV len = _invlist_len(invlist);
9946 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9952 array = invlist_array(invlist);
9954 /* The last element in the array in the inversion list always starts a
9955 * range that goes to infinity. That range may be for code points that are
9956 * matched in the inversion list, or it may be for ones that aren't
9957 * matched. In the latter case, the highest code point in the set is one
9958 * less than the beginning of this range; otherwise it is the final element
9959 * of this range: infinity */
9960 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
9962 : array[len - 1] - 1;
9966 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
9968 /* Get the contents of an inversion list into a string SV so that they can
9969 * be printed out. If 'traditional_style' is TRUE, it uses the format
9970 * traditionally done for debug tracing; otherwise it uses a format
9971 * suitable for just copying to the output, with blanks between ranges and
9972 * a dash between range components */
9976 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
9977 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
9979 if (traditional_style) {
9980 output = newSVpvs("\n");
9983 output = newSVpvs("");
9986 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
9988 assert(! invlist_is_iterating(invlist));
9990 invlist_iterinit(invlist);
9991 while (invlist_iternext(invlist, &start, &end)) {
9992 if (end == UV_MAX) {
9993 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFINITY%c",
9994 start, intra_range_delimiter,
9995 inter_range_delimiter);
9997 else if (end != start) {
9998 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10000 intra_range_delimiter,
10001 end, inter_range_delimiter);
10004 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10005 start, inter_range_delimiter);
10009 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10010 SvCUR_set(output, SvCUR(output) - 1);
10016 #ifndef PERL_IN_XSUB_RE
10018 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10019 const char * const indent, SV* const invlist)
10021 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10022 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10023 * the string 'indent'. The output looks like this:
10024 [0] 0x000A .. 0x000D
10026 [4] 0x2028 .. 0x2029
10027 [6] 0x3104 .. INFINITY
10028 * This means that the first range of code points matched by the list are
10029 * 0xA through 0xD; the second range contains only the single code point
10030 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10031 * are used to define each range (except if the final range extends to
10032 * infinity, only a single element is needed). The array index of the
10033 * first element for the corresponding range is given in brackets. */
10038 PERL_ARGS_ASSERT__INVLIST_DUMP;
10040 if (invlist_is_iterating(invlist)) {
10041 Perl_dump_indent(aTHX_ level, file,
10042 "%sCan't dump inversion list because is in middle of iterating\n",
10047 invlist_iterinit(invlist);
10048 while (invlist_iternext(invlist, &start, &end)) {
10049 if (end == UV_MAX) {
10050 Perl_dump_indent(aTHX_ level, file,
10051 "%s[%" UVuf "] 0x%04" UVXf " .. INFINITY\n",
10052 indent, (UV)count, start);
10054 else if (end != start) {
10055 Perl_dump_indent(aTHX_ level, file,
10056 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10057 indent, (UV)count, start, end);
10060 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10061 indent, (UV)count, start);
10068 Perl__load_PL_utf8_foldclosures (pTHX)
10070 assert(! PL_utf8_foldclosures);
10072 /* If the folds haven't been read in, call a fold function
10074 if (! PL_utf8_tofold) {
10075 U8 dummy[UTF8_MAXBYTES_CASE+1];
10076 const U8 hyphen[] = HYPHEN_UTF8;
10078 /* This string is just a short named one above \xff */
10079 toFOLD_utf8_safe(hyphen, hyphen + sizeof(hyphen) - 1, dummy, NULL);
10080 assert(PL_utf8_tofold); /* Verify that worked */
10082 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
10086 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10088 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10090 /* Return a boolean as to if the two passed in inversion lists are
10091 * identical. The final argument, if TRUE, says to take the complement of
10092 * the second inversion list before doing the comparison */
10094 const UV* array_a = invlist_array(a);
10095 const UV* array_b = invlist_array(b);
10096 UV len_a = _invlist_len(a);
10097 UV len_b = _invlist_len(b);
10099 PERL_ARGS_ASSERT__INVLISTEQ;
10101 /* If are to compare 'a' with the complement of b, set it
10102 * up so are looking at b's complement. */
10103 if (complement_b) {
10105 /* The complement of nothing is everything, so <a> would have to have
10106 * just one element, starting at zero (ending at infinity) */
10108 return (len_a == 1 && array_a[0] == 0);
10110 else if (array_b[0] == 0) {
10112 /* Otherwise, to complement, we invert. Here, the first element is
10113 * 0, just remove it. To do this, we just pretend the array starts
10121 /* But if the first element is not zero, we pretend the list starts
10122 * at the 0 that is always stored immediately before the array. */
10128 return len_a == len_b
10129 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10135 * As best we can, determine the characters that can match the start of
10136 * the given EXACTF-ish node.
10138 * Returns the invlist as a new SV*; it is the caller's responsibility to
10139 * call SvREFCNT_dec() when done with it.
10142 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10144 const U8 * s = (U8*)STRING(node);
10145 SSize_t bytelen = STR_LEN(node);
10147 /* Start out big enough for 2 separate code points */
10148 SV* invlist = _new_invlist(4);
10150 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
10155 /* We punt and assume can match anything if the node begins
10156 * with a multi-character fold. Things are complicated. For
10157 * example, /ffi/i could match any of:
10158 * "\N{LATIN SMALL LIGATURE FFI}"
10159 * "\N{LATIN SMALL LIGATURE FF}I"
10160 * "F\N{LATIN SMALL LIGATURE FI}"
10161 * plus several other things; and making sure we have all the
10162 * possibilities is hard. */
10163 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10164 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10167 /* Any Latin1 range character can potentially match any
10168 * other depending on the locale */
10169 if (OP(node) == EXACTFL) {
10170 _invlist_union(invlist, PL_Latin1, &invlist);
10173 /* But otherwise, it matches at least itself. We can
10174 * quickly tell if it has a distinct fold, and if so,
10175 * it matches that as well */
10176 invlist = add_cp_to_invlist(invlist, uc);
10177 if (IS_IN_SOME_FOLD_L1(uc))
10178 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10181 /* Some characters match above-Latin1 ones under /i. This
10182 * is true of EXACTFL ones when the locale is UTF-8 */
10183 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10184 && (! isASCII(uc) || (OP(node) != EXACTFA
10185 && OP(node) != EXACTFA_NO_TRIE)))
10187 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10191 else { /* Pattern is UTF-8 */
10192 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10193 STRLEN foldlen = UTF8SKIP(s);
10194 const U8* e = s + bytelen;
10197 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10199 /* The only code points that aren't folded in a UTF EXACTFish
10200 * node are are the problematic ones in EXACTFL nodes */
10201 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10202 /* We need to check for the possibility that this EXACTFL
10203 * node begins with a multi-char fold. Therefore we fold
10204 * the first few characters of it so that we can make that
10209 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10211 *(d++) = (U8) toFOLD(*s);
10216 toFOLD_utf8_safe(s, e, d, &len);
10222 /* And set up so the code below that looks in this folded
10223 * buffer instead of the node's string */
10225 foldlen = UTF8SKIP(folded);
10229 /* When we reach here 's' points to the fold of the first
10230 * character(s) of the node; and 'e' points to far enough along
10231 * the folded string to be just past any possible multi-char
10232 * fold. 'foldlen' is the length in bytes of the first
10235 * Unlike the non-UTF-8 case, the macro for determining if a
10236 * string is a multi-char fold requires all the characters to
10237 * already be folded. This is because of all the complications
10238 * if not. Note that they are folded anyway, except in EXACTFL
10239 * nodes. Like the non-UTF case above, we punt if the node
10240 * begins with a multi-char fold */
10242 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10243 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10245 else { /* Single char fold */
10247 /* It matches all the things that fold to it, which are
10248 * found in PL_utf8_foldclosures (including itself) */
10249 invlist = add_cp_to_invlist(invlist, uc);
10250 if (! PL_utf8_foldclosures)
10251 _load_PL_utf8_foldclosures();
10252 if ((listp = hv_fetch(PL_utf8_foldclosures,
10253 (char *) s, foldlen, FALSE)))
10255 AV* list = (AV*) *listp;
10257 for (k = 0; k <= av_tindex_skip_len_mg(list); k++) {
10258 SV** c_p = av_fetch(list, k, FALSE);
10264 /* /aa doesn't allow folds between ASCII and non- */
10265 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
10266 && isASCII(c) != isASCII(uc))
10271 invlist = add_cp_to_invlist(invlist, c);
10280 #undef HEADER_LENGTH
10281 #undef TO_INTERNAL_SIZE
10282 #undef FROM_INTERNAL_SIZE
10283 #undef INVLIST_VERSION_ID
10285 /* End of inversion list object */
10288 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10290 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10291 * constructs, and updates RExC_flags with them. On input, RExC_parse
10292 * should point to the first flag; it is updated on output to point to the
10293 * final ')' or ':'. There needs to be at least one flag, or this will
10296 /* for (?g), (?gc), and (?o) warnings; warning
10297 about (?c) will warn about (?g) -- japhy */
10299 #define WASTED_O 0x01
10300 #define WASTED_G 0x02
10301 #define WASTED_C 0x04
10302 #define WASTED_GC (WASTED_G|WASTED_C)
10303 I32 wastedflags = 0x00;
10304 U32 posflags = 0, negflags = 0;
10305 U32 *flagsp = &posflags;
10306 char has_charset_modifier = '\0';
10308 bool has_use_defaults = FALSE;
10309 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10310 int x_mod_count = 0;
10312 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10314 /* '^' as an initial flag sets certain defaults */
10315 if (UCHARAT(RExC_parse) == '^') {
10317 has_use_defaults = TRUE;
10318 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10319 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
10320 ? REGEX_UNICODE_CHARSET
10321 : REGEX_DEPENDS_CHARSET);
10324 cs = get_regex_charset(RExC_flags);
10325 if (cs == REGEX_DEPENDS_CHARSET
10326 && (RExC_utf8 || RExC_uni_semantics))
10328 cs = REGEX_UNICODE_CHARSET;
10331 while (RExC_parse < RExC_end) {
10332 /* && strchr("iogcmsx", *RExC_parse) */
10333 /* (?g), (?gc) and (?o) are useless here
10334 and must be globally applied -- japhy */
10335 switch (*RExC_parse) {
10337 /* Code for the imsxn flags */
10338 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10340 case LOCALE_PAT_MOD:
10341 if (has_charset_modifier) {
10342 goto excess_modifier;
10344 else if (flagsp == &negflags) {
10347 cs = REGEX_LOCALE_CHARSET;
10348 has_charset_modifier = LOCALE_PAT_MOD;
10350 case UNICODE_PAT_MOD:
10351 if (has_charset_modifier) {
10352 goto excess_modifier;
10354 else if (flagsp == &negflags) {
10357 cs = REGEX_UNICODE_CHARSET;
10358 has_charset_modifier = UNICODE_PAT_MOD;
10360 case ASCII_RESTRICT_PAT_MOD:
10361 if (flagsp == &negflags) {
10364 if (has_charset_modifier) {
10365 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10366 goto excess_modifier;
10368 /* Doubled modifier implies more restricted */
10369 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10372 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10374 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10376 case DEPENDS_PAT_MOD:
10377 if (has_use_defaults) {
10378 goto fail_modifiers;
10380 else if (flagsp == &negflags) {
10383 else if (has_charset_modifier) {
10384 goto excess_modifier;
10387 /* The dual charset means unicode semantics if the
10388 * pattern (or target, not known until runtime) are
10389 * utf8, or something in the pattern indicates unicode
10391 cs = (RExC_utf8 || RExC_uni_semantics)
10392 ? REGEX_UNICODE_CHARSET
10393 : REGEX_DEPENDS_CHARSET;
10394 has_charset_modifier = DEPENDS_PAT_MOD;
10398 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10399 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10401 else if (has_charset_modifier == *(RExC_parse - 1)) {
10402 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10403 *(RExC_parse - 1));
10406 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10408 NOT_REACHED; /*NOTREACHED*/
10411 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10412 *(RExC_parse - 1));
10413 NOT_REACHED; /*NOTREACHED*/
10414 case ONCE_PAT_MOD: /* 'o' */
10415 case GLOBAL_PAT_MOD: /* 'g' */
10416 if (PASS2 && ckWARN(WARN_REGEXP)) {
10417 const I32 wflagbit = *RExC_parse == 'o'
10420 if (! (wastedflags & wflagbit) ) {
10421 wastedflags |= wflagbit;
10422 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10425 "Useless (%s%c) - %suse /%c modifier",
10426 flagsp == &negflags ? "?-" : "?",
10428 flagsp == &negflags ? "don't " : "",
10435 case CONTINUE_PAT_MOD: /* 'c' */
10436 if (PASS2 && ckWARN(WARN_REGEXP)) {
10437 if (! (wastedflags & WASTED_C) ) {
10438 wastedflags |= WASTED_GC;
10439 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10442 "Useless (%sc) - %suse /gc modifier",
10443 flagsp == &negflags ? "?-" : "?",
10444 flagsp == &negflags ? "don't " : ""
10449 case KEEPCOPY_PAT_MOD: /* 'p' */
10450 if (flagsp == &negflags) {
10452 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10454 *flagsp |= RXf_PMf_KEEPCOPY;
10458 /* A flag is a default iff it is following a minus, so
10459 * if there is a minus, it means will be trying to
10460 * re-specify a default which is an error */
10461 if (has_use_defaults || flagsp == &negflags) {
10462 goto fail_modifiers;
10464 flagsp = &negflags;
10465 wastedflags = 0; /* reset so (?g-c) warns twice */
10471 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
10472 negflags |= RXf_PMf_EXTENDED_MORE;
10474 RExC_flags |= posflags;
10476 if (negflags & RXf_PMf_EXTENDED) {
10477 negflags |= RXf_PMf_EXTENDED_MORE;
10479 RExC_flags &= ~negflags;
10480 set_regex_charset(&RExC_flags, cs);
10485 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10486 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10487 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
10488 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10489 NOT_REACHED; /*NOTREACHED*/
10492 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10495 vFAIL("Sequence (?... not terminated");
10499 - reg - regular expression, i.e. main body or parenthesized thing
10501 * Caller must absorb opening parenthesis.
10503 * Combining parenthesis handling with the base level of regular expression
10504 * is a trifle forced, but the need to tie the tails of the branches to what
10505 * follows makes it hard to avoid.
10507 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10509 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10511 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10514 PERL_STATIC_INLINE regnode *
10515 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10517 char * parse_start,
10522 char* name_start = RExC_parse;
10524 SV *sv_dat = reg_scan_name(pRExC_state, SIZE_ONLY
10525 ? REG_RSN_RETURN_NULL
10526 : REG_RSN_RETURN_DATA);
10527 GET_RE_DEBUG_FLAGS_DECL;
10529 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10531 if (RExC_parse == name_start || *RExC_parse != ch) {
10532 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10533 vFAIL2("Sequence %.3s... not terminated",parse_start);
10537 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10538 RExC_rxi->data->data[num]=(void*)sv_dat;
10539 SvREFCNT_inc_simple_void(sv_dat);
10542 ret = reganode(pRExC_state,
10545 : (ASCII_FOLD_RESTRICTED)
10547 : (AT_LEAST_UNI_SEMANTICS)
10553 *flagp |= HASWIDTH;
10555 Set_Node_Offset(ret, parse_start+1);
10556 Set_Node_Cur_Length(ret, parse_start);
10558 nextchar(pRExC_state);
10562 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
10563 flags. Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan
10564 needs to be restarted, or'd with NEED_UTF8 if the pattern needs to be
10565 upgraded to UTF-8. Otherwise would only return NULL if regbranch() returns
10566 NULL, which cannot happen. */
10568 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
10569 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10570 * 2 is like 1, but indicates that nextchar() has been called to advance
10571 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10572 * this flag alerts us to the need to check for that */
10574 regnode *ret; /* Will be the head of the group. */
10577 regnode *ender = NULL;
10580 U32 oregflags = RExC_flags;
10581 bool have_branch = 0;
10583 I32 freeze_paren = 0;
10584 I32 after_freeze = 0;
10585 I32 num; /* numeric backreferences */
10587 char * parse_start = RExC_parse; /* MJD */
10588 char * const oregcomp_parse = RExC_parse;
10590 GET_RE_DEBUG_FLAGS_DECL;
10592 PERL_ARGS_ASSERT_REG;
10593 DEBUG_PARSE("reg ");
10595 *flagp = 0; /* Tentatively. */
10597 /* Having this true makes it feasible to have a lot fewer tests for the
10598 * parse pointer being in scope. For example, we can write
10599 * while(isFOO(*RExC_parse)) RExC_parse++;
10601 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
10603 assert(*RExC_end == '\0');
10605 /* Make an OPEN node, if parenthesized. */
10608 /* Under /x, space and comments can be gobbled up between the '(' and
10609 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
10610 * intervening space, as the sequence is a token, and a token should be
10612 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
10614 if (RExC_parse >= RExC_end) {
10615 vFAIL("Unmatched (");
10618 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
10619 char *start_verb = RExC_parse + 1;
10621 char *start_arg = NULL;
10622 unsigned char op = 0;
10623 int arg_required = 0;
10624 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
10626 if (has_intervening_patws) {
10627 RExC_parse++; /* past the '*' */
10628 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
10630 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
10631 if ( *RExC_parse == ':' ) {
10632 start_arg = RExC_parse + 1;
10635 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10637 verb_len = RExC_parse - start_verb;
10639 if (RExC_parse >= RExC_end) {
10640 goto unterminated_verb_pattern;
10642 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10643 while ( RExC_parse < RExC_end && *RExC_parse != ')' )
10644 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10645 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10646 unterminated_verb_pattern:
10647 vFAIL("Unterminated verb pattern argument");
10648 if ( RExC_parse == start_arg )
10651 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10652 vFAIL("Unterminated verb pattern");
10655 /* Here, we know that RExC_parse < RExC_end */
10657 switch ( *start_verb ) {
10658 case 'A': /* (*ACCEPT) */
10659 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
10661 internal_argval = RExC_nestroot;
10664 case 'C': /* (*COMMIT) */
10665 if ( memEQs(start_verb,verb_len,"COMMIT") )
10668 case 'F': /* (*FAIL) */
10669 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
10673 case ':': /* (*:NAME) */
10674 case 'M': /* (*MARK:NAME) */
10675 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
10680 case 'P': /* (*PRUNE) */
10681 if ( memEQs(start_verb,verb_len,"PRUNE") )
10684 case 'S': /* (*SKIP) */
10685 if ( memEQs(start_verb,verb_len,"SKIP") )
10688 case 'T': /* (*THEN) */
10689 /* [19:06] <TimToady> :: is then */
10690 if ( memEQs(start_verb,verb_len,"THEN") ) {
10692 RExC_seen |= REG_CUTGROUP_SEEN;
10697 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10699 "Unknown verb pattern '%" UTF8f "'",
10700 UTF8fARG(UTF, verb_len, start_verb));
10702 if ( arg_required && !start_arg ) {
10703 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
10704 verb_len, start_verb);
10706 if (internal_argval == -1) {
10707 ret = reganode(pRExC_state, op, 0);
10709 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
10711 RExC_seen |= REG_VERBARG_SEEN;
10712 if ( ! SIZE_ONLY ) {
10714 SV *sv = newSVpvn( start_arg,
10715 RExC_parse - start_arg);
10716 ARG(ret) = add_data( pRExC_state,
10717 STR_WITH_LEN("S"));
10718 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
10723 if ( internal_argval != -1 )
10724 ARG2L_SET(ret, internal_argval);
10726 nextchar(pRExC_state);
10729 else if (*RExC_parse == '?') { /* (?...) */
10730 bool is_logical = 0;
10731 const char * const seqstart = RExC_parse;
10732 const char * endptr;
10733 if (has_intervening_patws) {
10735 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
10738 RExC_parse++; /* past the '?' */
10739 paren = *RExC_parse; /* might be a trailing NUL, if not
10741 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10742 if (RExC_parse > RExC_end) {
10745 ret = NULL; /* For look-ahead/behind. */
10748 case 'P': /* (?P...) variants for those used to PCRE/Python */
10749 paren = *RExC_parse;
10750 if ( paren == '<') { /* (?P<...>) named capture */
10752 if (RExC_parse >= RExC_end) {
10753 vFAIL("Sequence (?P<... not terminated");
10755 goto named_capture;
10757 else if (paren == '>') { /* (?P>name) named recursion */
10759 if (RExC_parse >= RExC_end) {
10760 vFAIL("Sequence (?P>... not terminated");
10762 goto named_recursion;
10764 else if (paren == '=') { /* (?P=...) named backref */
10766 return handle_named_backref(pRExC_state, flagp,
10769 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10770 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10771 vFAIL3("Sequence (%.*s...) not recognized",
10772 RExC_parse-seqstart, seqstart);
10773 NOT_REACHED; /*NOTREACHED*/
10774 case '<': /* (?<...) */
10775 if (*RExC_parse == '!')
10777 else if (*RExC_parse != '=')
10784 case '\'': /* (?'...') */
10785 name_start = RExC_parse;
10786 svname = reg_scan_name(pRExC_state,
10787 SIZE_ONLY /* reverse test from the others */
10788 ? REG_RSN_RETURN_NAME
10789 : REG_RSN_RETURN_NULL);
10790 if ( RExC_parse == name_start
10791 || RExC_parse >= RExC_end
10792 || *RExC_parse != paren)
10794 vFAIL2("Sequence (?%c... not terminated",
10795 paren=='>' ? '<' : paren);
10800 if (!svname) /* shouldn't happen */
10802 "panic: reg_scan_name returned NULL");
10803 if (!RExC_paren_names) {
10804 RExC_paren_names= newHV();
10805 sv_2mortal(MUTABLE_SV(RExC_paren_names));
10807 RExC_paren_name_list= newAV();
10808 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10811 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10813 sv_dat = HeVAL(he_str);
10815 /* croak baby croak */
10817 "panic: paren_name hash element allocation failed");
10818 } else if ( SvPOK(sv_dat) ) {
10819 /* (?|...) can mean we have dupes so scan to check
10820 its already been stored. Maybe a flag indicating
10821 we are inside such a construct would be useful,
10822 but the arrays are likely to be quite small, so
10823 for now we punt -- dmq */
10824 IV count = SvIV(sv_dat);
10825 I32 *pv = (I32*)SvPVX(sv_dat);
10827 for ( i = 0 ; i < count ; i++ ) {
10828 if ( pv[i] == RExC_npar ) {
10834 pv = (I32*)SvGROW(sv_dat,
10835 SvCUR(sv_dat) + sizeof(I32)+1);
10836 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10837 pv[count] = RExC_npar;
10838 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10841 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10842 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10845 SvIV_set(sv_dat, 1);
10848 /* Yes this does cause a memory leak in debugging Perls
10850 if (!av_store(RExC_paren_name_list,
10851 RExC_npar, SvREFCNT_inc(svname)))
10852 SvREFCNT_dec_NN(svname);
10855 /*sv_dump(sv_dat);*/
10857 nextchar(pRExC_state);
10859 goto capturing_parens;
10861 RExC_seen |= REG_LOOKBEHIND_SEEN;
10862 RExC_in_lookbehind++;
10864 if (RExC_parse >= RExC_end) {
10865 vFAIL("Sequence (?... not terminated");
10869 case '=': /* (?=...) */
10870 RExC_seen_zerolen++;
10872 case '!': /* (?!...) */
10873 RExC_seen_zerolen++;
10874 /* check if we're really just a "FAIL" assertion */
10875 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
10876 FALSE /* Don't force to /x */ );
10877 if (*RExC_parse == ')') {
10878 ret=reganode(pRExC_state, OPFAIL, 0);
10879 nextchar(pRExC_state);
10883 case '|': /* (?|...) */
10884 /* branch reset, behave like a (?:...) except that
10885 buffers in alternations share the same numbers */
10887 after_freeze = freeze_paren = RExC_npar;
10889 case ':': /* (?:...) */
10890 case '>': /* (?>...) */
10892 case '$': /* (?$...) */
10893 case '@': /* (?@...) */
10894 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10896 case '0' : /* (?0) */
10897 case 'R' : /* (?R) */
10898 if (RExC_parse == RExC_end || *RExC_parse != ')')
10899 FAIL("Sequence (?R) not terminated");
10901 RExC_seen |= REG_RECURSE_SEEN;
10902 *flagp |= POSTPONED;
10903 goto gen_recurse_regop;
10905 /* named and numeric backreferences */
10906 case '&': /* (?&NAME) */
10907 parse_start = RExC_parse - 1;
10910 SV *sv_dat = reg_scan_name(pRExC_state,
10911 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10912 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10914 if (RExC_parse >= RExC_end || *RExC_parse != ')')
10915 vFAIL("Sequence (?&... not terminated");
10916 goto gen_recurse_regop;
10919 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10921 vFAIL("Illegal pattern");
10923 goto parse_recursion;
10925 case '-': /* (?-1) */
10926 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10927 RExC_parse--; /* rewind to let it be handled later */
10931 case '1': case '2': case '3': case '4': /* (?1) */
10932 case '5': case '6': case '7': case '8': case '9':
10933 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
10936 bool is_neg = FALSE;
10938 parse_start = RExC_parse - 1; /* MJD */
10939 if (*RExC_parse == '-') {
10943 if (grok_atoUV(RExC_parse, &unum, &endptr)
10947 RExC_parse = (char*)endptr;
10951 /* Some limit for num? */
10955 if (*RExC_parse!=')')
10956 vFAIL("Expecting close bracket");
10959 if ( paren == '-' ) {
10961 Diagram of capture buffer numbering.
10962 Top line is the normal capture buffer numbers
10963 Bottom line is the negative indexing as from
10967 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
10971 num = RExC_npar + num;
10974 vFAIL("Reference to nonexistent group");
10976 } else if ( paren == '+' ) {
10977 num = RExC_npar + num - 1;
10979 /* We keep track how many GOSUB items we have produced.
10980 To start off the ARG2L() of the GOSUB holds its "id",
10981 which is used later in conjunction with RExC_recurse
10982 to calculate the offset we need to jump for the GOSUB,
10983 which it will store in the final representation.
10984 We have to defer the actual calculation until much later
10985 as the regop may move.
10988 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
10990 if (num > (I32)RExC_rx->nparens) {
10992 vFAIL("Reference to nonexistent group");
10994 RExC_recurse_count++;
10995 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
10996 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
10997 22, "| |", (int)(depth * 2 + 1), "",
10998 (UV)ARG(ret), (IV)ARG2L(ret)));
11000 RExC_seen |= REG_RECURSE_SEEN;
11002 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
11003 Set_Node_Offset(ret, parse_start); /* MJD */
11005 *flagp |= POSTPONED;
11006 assert(*RExC_parse == ')');
11007 nextchar(pRExC_state);
11012 case '?': /* (??...) */
11014 if (*RExC_parse != '{') {
11015 RExC_parse += SKIP_IF_CHAR(RExC_parse);
11016 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11018 "Sequence (%" UTF8f "...) not recognized",
11019 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11020 NOT_REACHED; /*NOTREACHED*/
11022 *flagp |= POSTPONED;
11026 case '{': /* (?{...}) */
11029 struct reg_code_block *cb;
11031 RExC_seen_zerolen++;
11033 if ( !pRExC_state->code_blocks
11034 || pRExC_state->code_index
11035 >= pRExC_state->code_blocks->count
11036 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11037 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11040 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11041 FAIL("panic: Sequence (?{...}): no code block found\n");
11042 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11044 /* this is a pre-compiled code block (?{...}) */
11045 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11046 RExC_parse = RExC_start + cb->end;
11049 if (cb->src_regex) {
11050 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11051 RExC_rxi->data->data[n] =
11052 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11053 RExC_rxi->data->data[n+1] = (void*)o;
11056 n = add_data(pRExC_state,
11057 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11058 RExC_rxi->data->data[n] = (void*)o;
11061 pRExC_state->code_index++;
11062 nextchar(pRExC_state);
11066 ret = reg_node(pRExC_state, LOGICAL);
11068 eval = reg2Lanode(pRExC_state, EVAL,
11071 /* for later propagation into (??{})
11073 RExC_flags & RXf_PMf_COMPILETIME
11078 REGTAIL(pRExC_state, ret, eval);
11079 /* deal with the length of this later - MJD */
11082 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11083 Set_Node_Length(ret, RExC_parse - parse_start + 1);
11084 Set_Node_Offset(ret, parse_start);
11087 case '(': /* (?(?{...})...) and (?(?=...)...) */
11090 const int DEFINE_len = sizeof("DEFINE") - 1;
11091 if (RExC_parse[0] == '?') { /* (?(?...)) */
11092 if ( RExC_parse < RExC_end - 1
11093 && ( RExC_parse[1] == '='
11094 || RExC_parse[1] == '!'
11095 || RExC_parse[1] == '<'
11096 || RExC_parse[1] == '{')
11097 ) { /* Lookahead or eval. */
11101 ret = reg_node(pRExC_state, LOGICAL);
11105 tail = reg(pRExC_state, 1, &flag, depth+1);
11106 if (flag & (RESTART_PASS1|NEED_UTF8)) {
11107 *flagp = flag & (RESTART_PASS1|NEED_UTF8);
11110 REGTAIL(pRExC_state, ret, tail);
11113 /* Fall through to ‘Unknown switch condition’ at the
11114 end of the if/else chain. */
11116 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11117 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11119 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11120 char *name_start= RExC_parse++;
11122 SV *sv_dat=reg_scan_name(pRExC_state,
11123 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11124 if ( RExC_parse == name_start
11125 || RExC_parse >= RExC_end
11126 || *RExC_parse != ch)
11128 vFAIL2("Sequence (?(%c... not terminated",
11129 (ch == '>' ? '<' : ch));
11133 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11134 RExC_rxi->data->data[num]=(void*)sv_dat;
11135 SvREFCNT_inc_simple_void(sv_dat);
11137 ret = reganode(pRExC_state,NGROUPP,num);
11138 goto insert_if_check_paren;
11140 else if (RExC_end - RExC_parse >= DEFINE_len
11141 && strnEQ(RExC_parse, "DEFINE", DEFINE_len))
11143 ret = reganode(pRExC_state,DEFINEP,0);
11144 RExC_parse += DEFINE_len;
11146 goto insert_if_check_paren;
11148 else if (RExC_parse[0] == 'R') {
11150 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11151 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11152 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11155 if (RExC_parse[0] == '0') {
11159 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11161 if (grok_atoUV(RExC_parse, &uv, &endptr)
11164 parno = (I32)uv + 1;
11165 RExC_parse = (char*)endptr;
11167 /* else "Switch condition not recognized" below */
11168 } else if (RExC_parse[0] == '&') {
11171 sv_dat = reg_scan_name(pRExC_state,
11173 ? REG_RSN_RETURN_NULL
11174 : REG_RSN_RETURN_DATA);
11176 /* we should only have a false sv_dat when
11177 * SIZE_ONLY is true, and we always have false
11178 * sv_dat when SIZE_ONLY is true.
11179 * reg_scan_name() will VFAIL() if the name is
11180 * unknown when SIZE_ONLY is false, and otherwise
11181 * will return something, and when SIZE_ONLY is
11182 * true, reg_scan_name() just parses the string,
11183 * and doesnt return anything. (in theory) */
11184 assert(SIZE_ONLY ? !sv_dat : !!sv_dat);
11187 parno = 1 + *((I32 *)SvPVX(sv_dat));
11189 ret = reganode(pRExC_state,INSUBP,parno);
11190 goto insert_if_check_paren;
11192 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11196 if (grok_atoUV(RExC_parse, &uv, &endptr)
11200 RExC_parse = (char*)endptr;
11203 vFAIL("panic: grok_atoUV returned FALSE");
11205 ret = reganode(pRExC_state, GROUPP, parno);
11207 insert_if_check_paren:
11208 if (UCHARAT(RExC_parse) != ')') {
11209 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11210 vFAIL("Switch condition not recognized");
11212 nextchar(pRExC_state);
11214 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
11215 br = regbranch(pRExC_state, &flags, 1,depth+1);
11217 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11218 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11221 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11224 REGTAIL(pRExC_state, br, reganode(pRExC_state,
11226 c = UCHARAT(RExC_parse);
11227 nextchar(pRExC_state);
11228 if (flags&HASWIDTH)
11229 *flagp |= HASWIDTH;
11232 vFAIL("(?(DEFINE)....) does not allow branches");
11234 /* Fake one for optimizer. */
11235 lastbr = reganode(pRExC_state, IFTHEN, 0);
11237 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
11238 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11239 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11242 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11245 REGTAIL(pRExC_state, ret, lastbr);
11246 if (flags&HASWIDTH)
11247 *flagp |= HASWIDTH;
11248 c = UCHARAT(RExC_parse);
11249 nextchar(pRExC_state);
11254 if (RExC_parse >= RExC_end)
11255 vFAIL("Switch (?(condition)... not terminated");
11257 vFAIL("Switch (?(condition)... contains too many branches");
11259 ender = reg_node(pRExC_state, TAIL);
11260 REGTAIL(pRExC_state, br, ender);
11262 REGTAIL(pRExC_state, lastbr, ender);
11263 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11266 REGTAIL(pRExC_state, ret, ender);
11267 RExC_size++; /* XXX WHY do we need this?!!
11268 For large programs it seems to be required
11269 but I can't figure out why. -- dmq*/
11272 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11273 vFAIL("Unknown switch condition (?(...))");
11275 case '[': /* (?[ ... ]) */
11276 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
11278 case 0: /* A NUL */
11279 RExC_parse--; /* for vFAIL to print correctly */
11280 vFAIL("Sequence (? incomplete");
11282 default: /* e.g., (?i) */
11283 RExC_parse = (char *) seqstart + 1;
11285 parse_lparen_question_flags(pRExC_state);
11286 if (UCHARAT(RExC_parse) != ':') {
11287 if (RExC_parse < RExC_end)
11288 nextchar(pRExC_state);
11293 nextchar(pRExC_state);
11298 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11303 ret = reganode(pRExC_state, OPEN, parno);
11305 if (!RExC_nestroot)
11306 RExC_nestroot = parno;
11307 if (RExC_open_parens && !RExC_open_parens[parno])
11309 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11310 "%*s%*s Setting open paren #%" IVdf " to %d\n",
11311 22, "| |", (int)(depth * 2 + 1), "",
11312 (IV)parno, REG_NODE_NUM(ret)));
11313 RExC_open_parens[parno]= ret;
11316 Set_Node_Length(ret, 1); /* MJD */
11317 Set_Node_Offset(ret, RExC_parse); /* MJD */
11320 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
11329 /* Pick up the branches, linking them together. */
11330 parse_start = RExC_parse; /* MJD */
11331 br = regbranch(pRExC_state, &flags, 1,depth+1);
11333 /* branch_len = (paren != 0); */
11336 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11337 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11340 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11342 if (*RExC_parse == '|') {
11343 if (!SIZE_ONLY && RExC_extralen) {
11344 reginsert(pRExC_state, BRANCHJ, br, depth+1);
11347 reginsert(pRExC_state, BRANCH, br, depth+1);
11348 Set_Node_Length(br, paren != 0);
11349 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
11353 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
11355 else if (paren == ':') {
11356 *flagp |= flags&SIMPLE;
11358 if (is_open) { /* Starts with OPEN. */
11359 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
11361 else if (paren != '?') /* Not Conditional */
11363 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11365 while (*RExC_parse == '|') {
11366 if (!SIZE_ONLY && RExC_extralen) {
11367 ender = reganode(pRExC_state, LONGJMP,0);
11369 /* Append to the previous. */
11370 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11373 RExC_extralen += 2; /* Account for LONGJMP. */
11374 nextchar(pRExC_state);
11375 if (freeze_paren) {
11376 if (RExC_npar > after_freeze)
11377 after_freeze = RExC_npar;
11378 RExC_npar = freeze_paren;
11380 br = regbranch(pRExC_state, &flags, 0, depth+1);
11383 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11384 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11387 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11389 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
11391 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11394 if (have_branch || paren != ':') {
11395 /* Make a closing node, and hook it on the end. */
11398 ender = reg_node(pRExC_state, TAIL);
11401 ender = reganode(pRExC_state, CLOSE, parno);
11402 if ( RExC_close_parens ) {
11403 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11404 "%*s%*s Setting close paren #%" IVdf " to %d\n",
11405 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
11406 RExC_close_parens[parno]= ender;
11407 if (RExC_nestroot == parno)
11410 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
11411 Set_Node_Length(ender,1); /* MJD */
11417 *flagp &= ~HASWIDTH;
11420 ender = reg_node(pRExC_state, SUCCEED);
11423 ender = reg_node(pRExC_state, END);
11425 assert(!RExC_end_op); /* there can only be one! */
11426 RExC_end_op = ender;
11427 if (RExC_close_parens) {
11428 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11429 "%*s%*s Setting close paren #0 (END) to %d\n",
11430 22, "| |", (int)(depth * 2 + 1), "", REG_NODE_NUM(ender)));
11432 RExC_close_parens[0]= ender;
11437 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11438 DEBUG_PARSE_MSG("lsbr");
11439 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
11440 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11441 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11442 SvPV_nolen_const(RExC_mysv1),
11443 (IV)REG_NODE_NUM(lastbr),
11444 SvPV_nolen_const(RExC_mysv2),
11445 (IV)REG_NODE_NUM(ender),
11446 (IV)(ender - lastbr)
11449 REGTAIL(pRExC_state, lastbr, ender);
11451 if (have_branch && !SIZE_ONLY) {
11452 char is_nothing= 1;
11454 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
11456 /* Hook the tails of the branches to the closing node. */
11457 for (br = ret; br; br = regnext(br)) {
11458 const U8 op = PL_regkind[OP(br)];
11459 if (op == BRANCH) {
11460 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
11461 if ( OP(NEXTOPER(br)) != NOTHING
11462 || regnext(NEXTOPER(br)) != ender)
11465 else if (op == BRANCHJ) {
11466 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
11467 /* for now we always disable this optimisation * /
11468 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
11469 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
11475 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
11476 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11477 DEBUG_PARSE_MSG("NADA");
11478 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
11479 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11480 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11481 SvPV_nolen_const(RExC_mysv1),
11482 (IV)REG_NODE_NUM(ret),
11483 SvPV_nolen_const(RExC_mysv2),
11484 (IV)REG_NODE_NUM(ender),
11489 if (OP(ender) == TAIL) {
11494 for ( opt= br + 1; opt < ender ; opt++ )
11495 OP(opt)= OPTIMIZED;
11496 NEXT_OFF(br)= ender - br;
11504 static const char parens[] = "=!<,>";
11506 if (paren && (p = strchr(parens, paren))) {
11507 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
11508 int flag = (p - parens) > 1;
11511 node = SUSPEND, flag = 0;
11512 reginsert(pRExC_state, node,ret, depth+1);
11513 Set_Node_Cur_Length(ret, parse_start);
11514 Set_Node_Offset(ret, parse_start + 1);
11516 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
11520 /* Check for proper termination. */
11522 /* restore original flags, but keep (?p) and, if we've changed from /d
11523 * rules to /u, keep the /u */
11524 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
11525 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
11526 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
11528 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
11529 RExC_parse = oregcomp_parse;
11530 vFAIL("Unmatched (");
11532 nextchar(pRExC_state);
11534 else if (!paren && RExC_parse < RExC_end) {
11535 if (*RExC_parse == ')') {
11537 vFAIL("Unmatched )");
11540 FAIL("Junk on end of regexp"); /* "Can't happen". */
11541 NOT_REACHED; /* NOTREACHED */
11544 if (RExC_in_lookbehind) {
11545 RExC_in_lookbehind--;
11547 if (after_freeze > RExC_npar)
11548 RExC_npar = after_freeze;
11553 - regbranch - one alternative of an | operator
11555 * Implements the concatenation operator.
11557 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11558 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11561 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
11564 regnode *chain = NULL;
11566 I32 flags = 0, c = 0;
11567 GET_RE_DEBUG_FLAGS_DECL;
11569 PERL_ARGS_ASSERT_REGBRANCH;
11571 DEBUG_PARSE("brnc");
11576 if (!SIZE_ONLY && RExC_extralen)
11577 ret = reganode(pRExC_state, BRANCHJ,0);
11579 ret = reg_node(pRExC_state, BRANCH);
11580 Set_Node_Length(ret, 1);
11584 if (!first && SIZE_ONLY)
11585 RExC_extralen += 1; /* BRANCHJ */
11587 *flagp = WORST; /* Tentatively. */
11589 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11590 FALSE /* Don't force to /x */ );
11591 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
11592 flags &= ~TRYAGAIN;
11593 latest = regpiece(pRExC_state, &flags,depth+1);
11594 if (latest == NULL) {
11595 if (flags & TRYAGAIN)
11597 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11598 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11601 FAIL2("panic: regpiece returned NULL, flags=%#" UVxf, (UV) flags);
11603 else if (ret == NULL)
11605 *flagp |= flags&(HASWIDTH|POSTPONED);
11606 if (chain == NULL) /* First piece. */
11607 *flagp |= flags&SPSTART;
11609 /* FIXME adding one for every branch after the first is probably
11610 * excessive now we have TRIE support. (hv) */
11612 REGTAIL(pRExC_state, chain, latest);
11617 if (chain == NULL) { /* Loop ran zero times. */
11618 chain = reg_node(pRExC_state, NOTHING);
11623 *flagp |= flags&SIMPLE;
11630 - regpiece - something followed by possible quantifier * + ? {n,m}
11632 * Note that the branching code sequences used for ? and the general cases
11633 * of * and + are somewhat optimized: they use the same NOTHING node as
11634 * both the endmarker for their branch list and the body of the last branch.
11635 * It might seem that this node could be dispensed with entirely, but the
11636 * endmarker role is not redundant.
11638 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
11640 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11641 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11644 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11650 const char * const origparse = RExC_parse;
11652 I32 max = REG_INFTY;
11653 #ifdef RE_TRACK_PATTERN_OFFSETS
11656 const char *maxpos = NULL;
11659 /* Save the original in case we change the emitted regop to a FAIL. */
11660 regnode * const orig_emit = RExC_emit;
11662 GET_RE_DEBUG_FLAGS_DECL;
11664 PERL_ARGS_ASSERT_REGPIECE;
11666 DEBUG_PARSE("piec");
11668 ret = regatom(pRExC_state, &flags,depth+1);
11670 if (flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8))
11671 *flagp |= flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8);
11673 FAIL2("panic: regatom returned NULL, flags=%#" UVxf, (UV) flags);
11679 if (op == '{' && regcurly(RExC_parse)) {
11681 #ifdef RE_TRACK_PATTERN_OFFSETS
11682 parse_start = RExC_parse; /* MJD */
11684 next = RExC_parse + 1;
11685 while (isDIGIT(*next) || *next == ',') {
11686 if (*next == ',') {
11694 if (*next == '}') { /* got one */
11695 const char* endptr;
11699 if (isDIGIT(*RExC_parse)) {
11700 if (!grok_atoUV(RExC_parse, &uv, &endptr))
11701 vFAIL("Invalid quantifier in {,}");
11702 if (uv >= REG_INFTY)
11703 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11708 if (*maxpos == ',')
11711 maxpos = RExC_parse;
11712 if (isDIGIT(*maxpos)) {
11713 if (!grok_atoUV(maxpos, &uv, &endptr))
11714 vFAIL("Invalid quantifier in {,}");
11715 if (uv >= REG_INFTY)
11716 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11719 max = REG_INFTY; /* meaning "infinity" */
11722 nextchar(pRExC_state);
11723 if (max < min) { /* If can't match, warn and optimize to fail
11725 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
11727 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
11728 NEXT_OFF(orig_emit)= regarglen[OPFAIL] + NODE_STEP_REGNODE;
11732 else if (min == max && *RExC_parse == '?')
11735 ckWARN2reg(RExC_parse + 1,
11736 "Useless use of greediness modifier '%c'",
11742 if ((flags&SIMPLE)) {
11743 if (min == 0 && max == REG_INFTY) {
11744 reginsert(pRExC_state, STAR, ret, depth+1);
11747 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11750 if (min == 1 && max == REG_INFTY) {
11751 reginsert(pRExC_state, PLUS, ret, depth+1);
11754 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11757 MARK_NAUGHTY_EXP(2, 2);
11758 reginsert(pRExC_state, CURLY, ret, depth+1);
11759 Set_Node_Offset(ret, parse_start+1); /* MJD */
11760 Set_Node_Cur_Length(ret, parse_start);
11763 regnode * const w = reg_node(pRExC_state, WHILEM);
11766 REGTAIL(pRExC_state, ret, w);
11767 if (!SIZE_ONLY && RExC_extralen) {
11768 reginsert(pRExC_state, LONGJMP,ret, depth+1);
11769 reginsert(pRExC_state, NOTHING,ret, depth+1);
11770 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
11772 reginsert(pRExC_state, CURLYX,ret, depth+1);
11774 Set_Node_Offset(ret, parse_start+1);
11775 Set_Node_Length(ret,
11776 op == '{' ? (RExC_parse - parse_start) : 1);
11778 if (!SIZE_ONLY && RExC_extralen)
11779 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
11780 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
11782 RExC_whilem_seen++, RExC_extralen += 3;
11783 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
11790 *flagp |= HASWIDTH;
11792 ARG1_SET(ret, (U16)min);
11793 ARG2_SET(ret, (U16)max);
11795 if (max == REG_INFTY)
11796 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11802 if (!ISMULT1(op)) {
11807 #if 0 /* Now runtime fix should be reliable. */
11809 /* if this is reinstated, don't forget to put this back into perldiag:
11811 =item Regexp *+ operand could be empty at {#} in regex m/%s/
11813 (F) The part of the regexp subject to either the * or + quantifier
11814 could match an empty string. The {#} shows in the regular
11815 expression about where the problem was discovered.
11819 if (!(flags&HASWIDTH) && op != '?')
11820 vFAIL("Regexp *+ operand could be empty");
11823 #ifdef RE_TRACK_PATTERN_OFFSETS
11824 parse_start = RExC_parse;
11826 nextchar(pRExC_state);
11828 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
11834 else if (op == '+') {
11838 else if (op == '?') {
11843 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
11844 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
11845 ckWARN2reg(RExC_parse,
11846 "%" UTF8f " matches null string many times",
11847 UTF8fARG(UTF, (RExC_parse >= origparse
11848 ? RExC_parse - origparse
11851 (void)ReREFCNT_inc(RExC_rx_sv);
11854 if (*RExC_parse == '?') {
11855 nextchar(pRExC_state);
11856 reginsert(pRExC_state, MINMOD, ret, depth+1);
11857 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
11859 else if (*RExC_parse == '+') {
11861 nextchar(pRExC_state);
11862 ender = reg_node(pRExC_state, SUCCEED);
11863 REGTAIL(pRExC_state, ret, ender);
11864 reginsert(pRExC_state, SUSPEND, ret, depth+1);
11866 ender = reg_node(pRExC_state, TAIL);
11867 REGTAIL(pRExC_state, ret, ender);
11870 if (ISMULT2(RExC_parse)) {
11872 vFAIL("Nested quantifiers");
11879 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
11888 /* This routine teases apart the various meanings of \N and returns
11889 * accordingly. The input parameters constrain which meaning(s) is/are valid
11890 * in the current context.
11892 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
11894 * If <code_point_p> is not NULL, the context is expecting the result to be a
11895 * single code point. If this \N instance turns out to a single code point,
11896 * the function returns TRUE and sets *code_point_p to that code point.
11898 * If <node_p> is not NULL, the context is expecting the result to be one of
11899 * the things representable by a regnode. If this \N instance turns out to be
11900 * one such, the function generates the regnode, returns TRUE and sets *node_p
11901 * to point to that regnode.
11903 * If this instance of \N isn't legal in any context, this function will
11904 * generate a fatal error and not return.
11906 * On input, RExC_parse should point to the first char following the \N at the
11907 * time of the call. On successful return, RExC_parse will have been updated
11908 * to point to just after the sequence identified by this routine. Also
11909 * *flagp has been updated as needed.
11911 * When there is some problem with the current context and this \N instance,
11912 * the function returns FALSE, without advancing RExC_parse, nor setting
11913 * *node_p, nor *code_point_p, nor *flagp.
11915 * If <cp_count> is not NULL, the caller wants to know the length (in code
11916 * points) that this \N sequence matches. This is set even if the function
11917 * returns FALSE, as detailed below.
11919 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
11921 * Probably the most common case is for the \N to specify a single code point.
11922 * *cp_count will be set to 1, and *code_point_p will be set to that code
11925 * Another possibility is for the input to be an empty \N{}, which for
11926 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
11927 * will be set to a generated NOTHING node.
11929 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
11930 * set to 0. *node_p will be set to a generated REG_ANY node.
11932 * The fourth possibility is that \N resolves to a sequence of more than one
11933 * code points. *cp_count will be set to the number of code points in the
11934 * sequence. *node_p * will be set to a generated node returned by this
11935 * function calling S_reg().
11937 * The final possibility is that it is premature to be calling this function;
11938 * that pass1 needs to be restarted. This can happen when this changes from
11939 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
11940 * latter occurs only when the fourth possibility would otherwise be in
11941 * effect, and is because one of those code points requires the pattern to be
11942 * recompiled as UTF-8. The function returns FALSE, and sets the
11943 * RESTART_PASS1 and NEED_UTF8 flags in *flagp, as appropriate. When this
11944 * happens, the caller needs to desist from continuing parsing, and return
11945 * this information to its caller. This is not set for when there is only one
11946 * code point, as this can be called as part of an ANYOF node, and they can
11947 * store above-Latin1 code points without the pattern having to be in UTF-8.
11949 * For non-single-quoted regexes, the tokenizer has resolved character and
11950 * sequence names inside \N{...} into their Unicode values, normalizing the
11951 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
11952 * hex-represented code points in the sequence. This is done there because
11953 * the names can vary based on what charnames pragma is in scope at the time,
11954 * so we need a way to take a snapshot of what they resolve to at the time of
11955 * the original parse. [perl #56444].
11957 * That parsing is skipped for single-quoted regexes, so we may here get
11958 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
11959 * parser. But if the single-quoted regex is something like '\N{U+41}', that
11960 * is legal and handled here. The code point is Unicode, and has to be
11961 * translated into the native character set for non-ASCII platforms.
11964 char * endbrace; /* points to '}' following the name */
11965 char *endchar; /* Points to '.' or '}' ending cur char in the input
11967 char* p = RExC_parse; /* Temporary */
11969 GET_RE_DEBUG_FLAGS_DECL;
11971 PERL_ARGS_ASSERT_GROK_BSLASH_N;
11973 GET_RE_DEBUG_FLAGS;
11975 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
11976 assert(! (node_p && cp_count)); /* At most 1 should be set */
11978 if (cp_count) { /* Initialize return for the most common case */
11982 /* The [^\n] meaning of \N ignores spaces and comments under the /x
11983 * modifier. The other meanings do not, so use a temporary until we find
11984 * out which we are being called with */
11985 skip_to_be_ignored_text(pRExC_state, &p,
11986 FALSE /* Don't force to /x */ );
11988 /* Disambiguate between \N meaning a named character versus \N meaning
11989 * [^\n]. The latter is assumed when the {...} following the \N is a legal
11990 * quantifier, or there is no '{' at all */
11991 if (*p != '{' || regcurly(p)) {
12001 *node_p = reg_node(pRExC_state, REG_ANY);
12002 *flagp |= HASWIDTH|SIMPLE;
12004 Set_Node_Length(*node_p, 1); /* MJD */
12008 /* Here, we have decided it should be a named character or sequence */
12010 /* The test above made sure that the next real character is a '{', but
12011 * under the /x modifier, it could be separated by space (or a comment and
12012 * \n) and this is not allowed (for consistency with \x{...} and the
12013 * tokenizer handling of \N{NAME}). */
12014 if (*RExC_parse != '{') {
12015 vFAIL("Missing braces on \\N{}");
12018 RExC_parse++; /* Skip past the '{' */
12020 endbrace = strchr(RExC_parse, '}');
12021 if (! endbrace) { /* no trailing brace */
12022 vFAIL2("Missing right brace on \\%c{}", 'N');
12024 else if(!(endbrace == RExC_parse /* nothing between the {} */
12025 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked... */
12026 && strnEQ(RExC_parse, "U+", 2)))) /* ... below for a better
12029 RExC_parse = endbrace; /* position msg's '<--HERE' */
12030 vFAIL("\\N{NAME} must be resolved by the lexer");
12033 REQUIRE_UNI_RULES(flagp, FALSE); /* Unicode named chars imply Unicode
12036 if (endbrace == RExC_parse) { /* empty: \N{} */
12038 RExC_parse++; /* Position after the "}" */
12039 vFAIL("Zero length \\N{}");
12044 nextchar(pRExC_state);
12049 *node_p = reg_node(pRExC_state,NOTHING);
12053 RExC_parse += 2; /* Skip past the 'U+' */
12055 /* Because toke.c has generated a special construct for us guaranteed not
12056 * to have NULs, we can use a str function */
12057 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12059 /* Code points are separated by dots. If none, there is only one code
12060 * point, and is terminated by the brace */
12062 if (endchar >= endbrace) {
12063 STRLEN length_of_hex;
12064 I32 grok_hex_flags;
12066 /* Here, exactly one code point. If that isn't what is wanted, fail */
12067 if (! code_point_p) {
12072 /* Convert code point from hex */
12073 length_of_hex = (STRLEN)(endchar - RExC_parse);
12074 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
12075 | PERL_SCAN_DISALLOW_PREFIX
12077 /* No errors in the first pass (See [perl
12078 * #122671].) We let the code below find the
12079 * errors when there are multiple chars. */
12081 ? PERL_SCAN_SILENT_ILLDIGIT
12084 /* This routine is the one place where both single- and double-quotish
12085 * \N{U+xxxx} are evaluated. The value is a Unicode code point which
12086 * must be converted to native. */
12087 *code_point_p = UNI_TO_NATIVE(grok_hex(RExC_parse,
12092 /* The tokenizer should have guaranteed validity, but it's possible to
12093 * bypass it by using single quoting, so check. Don't do the check
12094 * here when there are multiple chars; we do it below anyway. */
12095 if (length_of_hex == 0
12096 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
12098 RExC_parse += length_of_hex; /* Includes all the valid */
12099 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12100 ? UTF8SKIP(RExC_parse)
12102 /* Guard against malformed utf8 */
12103 if (RExC_parse >= endchar) {
12104 RExC_parse = endchar;
12106 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12109 RExC_parse = endbrace + 1;
12112 else { /* Is a multiple character sequence */
12113 SV * substitute_parse;
12115 char *orig_end = RExC_end;
12116 char *save_start = RExC_start;
12119 /* Count the code points, if desired, in the sequence */
12122 while (RExC_parse < endbrace) {
12123 /* Point to the beginning of the next character in the sequence. */
12124 RExC_parse = endchar + 1;
12125 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12130 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12131 * But don't backup up the pointer if the caller want to know how many
12132 * code points there are (they can then handle things) */
12140 /* What is done here is to convert this to a sub-pattern of the form
12141 * \x{char1}\x{char2}... and then call reg recursively to parse it
12142 * (enclosing in "(?: ... )" ). That way, it retains its atomicness,
12143 * while not having to worry about special handling that some code
12144 * points may have. */
12146 substitute_parse = newSVpvs("?:");
12148 while (RExC_parse < endbrace) {
12150 /* Convert to notation the rest of the code understands */
12151 sv_catpv(substitute_parse, "\\x{");
12152 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
12153 sv_catpv(substitute_parse, "}");
12155 /* Point to the beginning of the next character in the sequence. */
12156 RExC_parse = endchar + 1;
12157 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12160 sv_catpv(substitute_parse, ")");
12162 RExC_parse = RExC_start = RExC_adjusted_start = SvPV(substitute_parse,
12165 /* Don't allow empty number */
12166 if (len < (STRLEN) 8) {
12167 RExC_parse = endbrace;
12168 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12170 RExC_end = RExC_parse + len;
12172 /* The values are Unicode, and therefore not subject to recoding, but
12173 * have to be converted to native on a non-Unicode (meaning non-ASCII)
12176 RExC_recode_x_to_native = 1;
12180 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
12181 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12182 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12185 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#" UVxf,
12188 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12191 /* Restore the saved values */
12192 RExC_start = RExC_adjusted_start = save_start;
12193 RExC_parse = endbrace;
12194 RExC_end = orig_end;
12196 RExC_recode_x_to_native = 0;
12199 SvREFCNT_dec_NN(substitute_parse);
12200 nextchar(pRExC_state);
12207 PERL_STATIC_INLINE U8
12208 S_compute_EXACTish(RExC_state_t *pRExC_state)
12212 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
12220 op = get_regex_charset(RExC_flags);
12221 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
12222 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
12223 been, so there is no hole */
12226 return op + EXACTF;
12229 PERL_STATIC_INLINE void
12230 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
12231 regnode *node, I32* flagp, STRLEN len, UV code_point,
12234 /* This knows the details about sizing an EXACTish node, setting flags for
12235 * it (by setting <*flagp>, and potentially populating it with a single
12238 * If <len> (the length in bytes) is non-zero, this function assumes that
12239 * the node has already been populated, and just does the sizing. In this
12240 * case <code_point> should be the final code point that has already been
12241 * placed into the node. This value will be ignored except that under some
12242 * circumstances <*flagp> is set based on it.
12244 * If <len> is zero, the function assumes that the node is to contain only
12245 * the single character given by <code_point> and calculates what <len>
12246 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
12247 * additionally will populate the node's STRING with <code_point> or its
12250 * In both cases <*flagp> is appropriately set
12252 * It knows that under FOLD, the Latin Sharp S and UTF characters above
12253 * 255, must be folded (the former only when the rules indicate it can
12256 * When it does the populating, it looks at the flag 'downgradable'. If
12257 * true with a node that folds, it checks if the single code point
12258 * participates in a fold, and if not downgrades the node to an EXACT.
12259 * This helps the optimizer */
12261 bool len_passed_in = cBOOL(len != 0);
12262 U8 character[UTF8_MAXBYTES_CASE+1];
12264 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
12266 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
12267 * sizing difference, and is extra work that is thrown away */
12268 if (downgradable && ! PASS2) {
12269 downgradable = FALSE;
12272 if (! len_passed_in) {
12274 if (UVCHR_IS_INVARIANT(code_point)) {
12275 if (LOC || ! FOLD) { /* /l defers folding until runtime */
12276 *character = (U8) code_point;
12278 else { /* Here is /i and not /l. (toFOLD() is defined on just
12279 ASCII, which isn't the same thing as INVARIANT on
12280 EBCDIC, but it works there, as the extra invariants
12281 fold to themselves) */
12282 *character = toFOLD((U8) code_point);
12284 /* We can downgrade to an EXACT node if this character
12285 * isn't a folding one. Note that this assumes that
12286 * nothing above Latin1 folds to some other invariant than
12287 * one of these alphabetics; otherwise we would also have
12289 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12290 * || ASCII_FOLD_RESTRICTED))
12292 if (downgradable && PL_fold[code_point] == code_point) {
12298 else if (FOLD && (! LOC
12299 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
12300 { /* Folding, and ok to do so now */
12301 UV folded = _to_uni_fold_flags(
12305 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12306 ? FOLD_FLAGS_NOMIX_ASCII
12309 && folded == code_point /* This quickly rules out many
12310 cases, avoiding the
12311 _invlist_contains_cp() overhead
12313 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
12320 else if (code_point <= MAX_UTF8_TWO_BYTE) {
12322 /* Not folding this cp, and can output it directly */
12323 *character = UTF8_TWO_BYTE_HI(code_point);
12324 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
12328 uvchr_to_utf8( character, code_point);
12329 len = UTF8SKIP(character);
12331 } /* Else pattern isn't UTF8. */
12333 *character = (U8) code_point;
12335 } /* Else is folded non-UTF8 */
12336 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12337 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12338 || UNICODE_DOT_DOT_VERSION > 0)
12339 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
12343 /* We don't fold any non-UTF8 except possibly the Sharp s (see
12344 * comments at join_exact()); */
12345 *character = (U8) code_point;
12348 /* Can turn into an EXACT node if we know the fold at compile time,
12349 * and it folds to itself and doesn't particpate in other folds */
12352 && PL_fold_latin1[code_point] == code_point
12353 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12354 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
12358 } /* else is Sharp s. May need to fold it */
12359 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
12361 *(character + 1) = 's';
12365 *character = LATIN_SMALL_LETTER_SHARP_S;
12371 RExC_size += STR_SZ(len);
12374 RExC_emit += STR_SZ(len);
12375 STR_LEN(node) = len;
12376 if (! len_passed_in) {
12377 Copy((char *) character, STRING(node), len, char);
12381 *flagp |= HASWIDTH;
12383 /* A single character node is SIMPLE, except for the special-cased SHARP S
12385 if ((len == 1 || (UTF && len == UVCHR_SKIP(code_point)))
12386 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12387 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12388 || UNICODE_DOT_DOT_VERSION > 0)
12389 && ( code_point != LATIN_SMALL_LETTER_SHARP_S
12390 || ! FOLD || ! DEPENDS_SEMANTICS)
12396 /* The OP may not be well defined in PASS1 */
12397 if (PASS2 && OP(node) == EXACTFL) {
12398 RExC_contains_locale = 1;
12403 S_new_regcurly(const char *s, const char *e)
12405 /* This is a temporary function designed to match the most lenient form of
12406 * a {m,n} quantifier we ever envision, with either number omitted, and
12407 * spaces anywhere between/before/after them.
12409 * If this function fails, then the string it matches is very unlikely to
12410 * ever be considered a valid quantifier, so we can allow the '{' that
12411 * begins it to be considered as a literal */
12413 bool has_min = FALSE;
12414 bool has_max = FALSE;
12416 PERL_ARGS_ASSERT_NEW_REGCURLY;
12418 if (s >= e || *s++ != '{')
12421 while (s < e && isSPACE(*s)) {
12424 while (s < e && isDIGIT(*s)) {
12428 while (s < e && isSPACE(*s)) {
12434 while (s < e && isSPACE(*s)) {
12437 while (s < e && isDIGIT(*s)) {
12441 while (s < e && isSPACE(*s)) {
12446 return s < e && *s == '}' && (has_min || has_max);
12449 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
12450 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
12453 S_backref_value(char *p)
12455 const char* endptr;
12457 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
12464 - regatom - the lowest level
12466 Try to identify anything special at the start of the current parse position.
12467 If there is, then handle it as required. This may involve generating a
12468 single regop, such as for an assertion; or it may involve recursing, such as
12469 to handle a () structure.
12471 If the string doesn't start with something special then we gobble up
12472 as much literal text as we can. If we encounter a quantifier, we have to
12473 back off the final literal character, as that quantifier applies to just it
12474 and not to the whole string of literals.
12476 Once we have been able to handle whatever type of thing started the
12477 sequence, we return.
12479 Note: we have to be careful with escapes, as they can be both literal
12480 and special, and in the case of \10 and friends, context determines which.
12482 A summary of the code structure is:
12484 switch (first_byte) {
12485 cases for each special:
12486 handle this special;
12489 switch (2nd byte) {
12490 cases for each unambiguous special:
12491 handle this special;
12493 cases for each ambigous special/literal:
12495 if (special) handle here
12497 default: // unambiguously literal:
12500 default: // is a literal char
12503 create EXACTish node for literal;
12504 while (more input and node isn't full) {
12505 switch (input_byte) {
12506 cases for each special;
12507 make sure parse pointer is set so that the next call to
12508 regatom will see this special first
12509 goto loopdone; // EXACTish node terminated by prev. char
12511 append char to EXACTISH node;
12513 get next input byte;
12517 return the generated node;
12519 Specifically there are two separate switches for handling
12520 escape sequences, with the one for handling literal escapes requiring
12521 a dummy entry for all of the special escapes that are actually handled
12524 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
12526 Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
12527 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
12528 Otherwise does not return NULL.
12532 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12534 regnode *ret = NULL;
12541 GET_RE_DEBUG_FLAGS_DECL;
12543 *flagp = WORST; /* Tentatively. */
12545 DEBUG_PARSE("atom");
12547 PERL_ARGS_ASSERT_REGATOM;
12550 parse_start = RExC_parse;
12551 assert(RExC_parse < RExC_end);
12552 switch ((U8)*RExC_parse) {
12554 RExC_seen_zerolen++;
12555 nextchar(pRExC_state);
12556 if (RExC_flags & RXf_PMf_MULTILINE)
12557 ret = reg_node(pRExC_state, MBOL);
12559 ret = reg_node(pRExC_state, SBOL);
12560 Set_Node_Length(ret, 1); /* MJD */
12563 nextchar(pRExC_state);
12565 RExC_seen_zerolen++;
12566 if (RExC_flags & RXf_PMf_MULTILINE)
12567 ret = reg_node(pRExC_state, MEOL);
12569 ret = reg_node(pRExC_state, SEOL);
12570 Set_Node_Length(ret, 1); /* MJD */
12573 nextchar(pRExC_state);
12574 if (RExC_flags & RXf_PMf_SINGLELINE)
12575 ret = reg_node(pRExC_state, SANY);
12577 ret = reg_node(pRExC_state, REG_ANY);
12578 *flagp |= HASWIDTH|SIMPLE;
12580 Set_Node_Length(ret, 1); /* MJD */
12584 char * const oregcomp_parse = ++RExC_parse;
12585 ret = regclass(pRExC_state, flagp,depth+1,
12586 FALSE, /* means parse the whole char class */
12587 TRUE, /* allow multi-char folds */
12588 FALSE, /* don't silence non-portable warnings. */
12589 (bool) RExC_strict,
12590 TRUE, /* Allow an optimized regnode result */
12594 if (*flagp & (RESTART_PASS1|NEED_UTF8))
12596 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12599 if (*RExC_parse != ']') {
12600 RExC_parse = oregcomp_parse;
12601 vFAIL("Unmatched [");
12603 nextchar(pRExC_state);
12604 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
12608 nextchar(pRExC_state);
12609 ret = reg(pRExC_state, 2, &flags,depth+1);
12611 if (flags & TRYAGAIN) {
12612 if (RExC_parse >= RExC_end) {
12613 /* Make parent create an empty node if needed. */
12614 *flagp |= TRYAGAIN;
12619 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12620 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12623 FAIL2("panic: reg returned NULL to regatom, flags=%#" UVxf,
12626 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12630 if (flags & TRYAGAIN) {
12631 *flagp |= TRYAGAIN;
12634 vFAIL("Internal urp");
12635 /* Supposed to be caught earlier. */
12641 vFAIL("Quantifier follows nothing");
12646 This switch handles escape sequences that resolve to some kind
12647 of special regop and not to literal text. Escape sequnces that
12648 resolve to literal text are handled below in the switch marked
12651 Every entry in this switch *must* have a corresponding entry
12652 in the literal escape switch. However, the opposite is not
12653 required, as the default for this switch is to jump to the
12654 literal text handling code.
12657 switch ((U8)*RExC_parse) {
12658 /* Special Escapes */
12660 RExC_seen_zerolen++;
12661 ret = reg_node(pRExC_state, SBOL);
12662 /* SBOL is shared with /^/ so we set the flags so we can tell
12663 * /\A/ from /^/ in split. We check ret because first pass we
12664 * have no regop struct to set the flags on. */
12668 goto finish_meta_pat;
12670 ret = reg_node(pRExC_state, GPOS);
12671 RExC_seen |= REG_GPOS_SEEN;
12673 goto finish_meta_pat;
12675 RExC_seen_zerolen++;
12676 ret = reg_node(pRExC_state, KEEPS);
12678 /* XXX:dmq : disabling in-place substitution seems to
12679 * be necessary here to avoid cases of memory corruption, as
12680 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
12682 RExC_seen |= REG_LOOKBEHIND_SEEN;
12683 goto finish_meta_pat;
12685 ret = reg_node(pRExC_state, SEOL);
12687 RExC_seen_zerolen++; /* Do not optimize RE away */
12688 goto finish_meta_pat;
12690 ret = reg_node(pRExC_state, EOS);
12692 RExC_seen_zerolen++; /* Do not optimize RE away */
12693 goto finish_meta_pat;
12695 vFAIL("\\C no longer supported");
12697 ret = reg_node(pRExC_state, CLUMP);
12698 *flagp |= HASWIDTH;
12699 goto finish_meta_pat;
12705 arg = ANYOF_WORDCHAR;
12713 regex_charset charset = get_regex_charset(RExC_flags);
12715 RExC_seen_zerolen++;
12716 RExC_seen |= REG_LOOKBEHIND_SEEN;
12717 op = BOUND + charset;
12719 if (op == BOUNDL) {
12720 RExC_contains_locale = 1;
12723 ret = reg_node(pRExC_state, op);
12725 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
12726 FLAGS(ret) = TRADITIONAL_BOUND;
12727 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
12733 char name = *RExC_parse;
12736 endbrace = strchr(RExC_parse, '}');
12739 vFAIL2("Missing right brace on \\%c{}", name);
12741 /* XXX Need to decide whether to take spaces or not. Should be
12742 * consistent with \p{}, but that currently is SPACE, which
12743 * means vertical too, which seems wrong
12744 * while (isBLANK(*RExC_parse)) {
12747 if (endbrace == RExC_parse) {
12748 RExC_parse++; /* After the '}' */
12749 vFAIL2("Empty \\%c{}", name);
12751 length = endbrace - RExC_parse;
12752 /*while (isBLANK(*(RExC_parse + length - 1))) {
12755 switch (*RExC_parse) {
12758 && (length != 3 || strnNE(RExC_parse + 1, "cb", 2)))
12760 goto bad_bound_type;
12762 FLAGS(ret) = GCB_BOUND;
12765 if (length != 2 || *(RExC_parse + 1) != 'b') {
12766 goto bad_bound_type;
12768 FLAGS(ret) = LB_BOUND;
12771 if (length != 2 || *(RExC_parse + 1) != 'b') {
12772 goto bad_bound_type;
12774 FLAGS(ret) = SB_BOUND;
12777 if (length != 2 || *(RExC_parse + 1) != 'b') {
12778 goto bad_bound_type;
12780 FLAGS(ret) = WB_BOUND;
12784 RExC_parse = endbrace;
12786 "'%" UTF8f "' is an unknown bound type",
12787 UTF8fARG(UTF, length, endbrace - length));
12788 NOT_REACHED; /*NOTREACHED*/
12790 RExC_parse = endbrace;
12791 REQUIRE_UNI_RULES(flagp, NULL);
12793 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
12797 /* Don't have to worry about UTF-8, in this message because
12798 * to get here the contents of the \b must be ASCII */
12799 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
12800 "Using /u for '%.*s' instead of /%s",
12802 endbrace - length + 1,
12803 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
12804 ? ASCII_RESTRICT_PAT_MODS
12805 : ASCII_MORE_RESTRICT_PAT_MODS);
12809 if (PASS2 && invert) {
12810 OP(ret) += NBOUND - BOUND;
12812 goto finish_meta_pat;
12820 if (! DEPENDS_SEMANTICS) {
12824 /* \d doesn't have any matches in the upper Latin1 range, hence /d
12825 * is equivalent to /u. Changing to /u saves some branches at
12828 goto join_posix_op_known;
12831 ret = reg_node(pRExC_state, LNBREAK);
12832 *flagp |= HASWIDTH|SIMPLE;
12833 goto finish_meta_pat;
12841 goto join_posix_op_known;
12847 arg = ANYOF_VERTWS;
12849 goto join_posix_op_known;
12859 op = POSIXD + get_regex_charset(RExC_flags);
12860 if (op > POSIXA) { /* /aa is same as /a */
12863 else if (op == POSIXL) {
12864 RExC_contains_locale = 1;
12867 join_posix_op_known:
12870 op += NPOSIXD - POSIXD;
12873 ret = reg_node(pRExC_state, op);
12875 FLAGS(ret) = namedclass_to_classnum(arg);
12878 *flagp |= HASWIDTH|SIMPLE;
12882 if ( UCHARAT(RExC_parse + 1) == '{'
12883 && UNLIKELY(! new_regcurly(RExC_parse + 1, RExC_end)))
12886 vFAIL("Unescaped left brace in regex is illegal here");
12888 nextchar(pRExC_state);
12889 Set_Node_Length(ret, 2); /* MJD */
12895 ret = regclass(pRExC_state, flagp,depth+1,
12896 TRUE, /* means just parse this element */
12897 FALSE, /* don't allow multi-char folds */
12898 FALSE, /* don't silence non-portable warnings. It
12899 would be a bug if these returned
12901 (bool) RExC_strict,
12902 TRUE, /* Allow an optimized regnode result */
12905 if (*flagp & RESTART_PASS1)
12907 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
12908 * multi-char folds are allowed. */
12910 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12915 Set_Node_Offset(ret, parse_start);
12916 Set_Node_Cur_Length(ret, parse_start - 2);
12917 nextchar(pRExC_state);
12920 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
12921 * \N{...} evaluates to a sequence of more than one code points).
12922 * The function call below returns a regnode, which is our result.
12923 * The parameters cause it to fail if the \N{} evaluates to a
12924 * single code point; we handle those like any other literal. The
12925 * reason that the multicharacter case is handled here and not as
12926 * part of the EXACtish code is because of quantifiers. In
12927 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
12928 * this way makes that Just Happen. dmq.
12929 * join_exact() will join this up with adjacent EXACTish nodes
12930 * later on, if appropriate. */
12932 if (grok_bslash_N(pRExC_state,
12933 &ret, /* Want a regnode returned */
12934 NULL, /* Fail if evaluates to a single code
12936 NULL, /* Don't need a count of how many code
12945 if (*flagp & RESTART_PASS1)
12948 /* Here, evaluates to a single code point. Go get that */
12949 RExC_parse = parse_start;
12952 case 'k': /* Handle \k<NAME> and \k'NAME' */
12956 if ( RExC_parse >= RExC_end - 1
12957 || (( ch = RExC_parse[1]) != '<'
12962 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12963 vFAIL2("Sequence %.2s... not terminated",parse_start);
12966 ret = handle_named_backref(pRExC_state,
12978 case '1': case '2': case '3': case '4':
12979 case '5': case '6': case '7': case '8': case '9':
12984 if (*RExC_parse == 'g') {
12988 if (*RExC_parse == '{') {
12992 if (*RExC_parse == '-') {
12996 if (hasbrace && !isDIGIT(*RExC_parse)) {
12997 if (isrel) RExC_parse--;
12999 goto parse_named_seq;
13002 if (RExC_parse >= RExC_end) {
13003 goto unterminated_g;
13005 num = S_backref_value(RExC_parse);
13007 vFAIL("Reference to invalid group 0");
13008 else if (num == I32_MAX) {
13009 if (isDIGIT(*RExC_parse))
13010 vFAIL("Reference to nonexistent group");
13013 vFAIL("Unterminated \\g... pattern");
13017 num = RExC_npar - num;
13019 vFAIL("Reference to nonexistent or unclosed group");
13023 num = S_backref_value(RExC_parse);
13024 /* bare \NNN might be backref or octal - if it is larger
13025 * than or equal RExC_npar then it is assumed to be an
13026 * octal escape. Note RExC_npar is +1 from the actual
13027 * number of parens. */
13028 /* Note we do NOT check if num == I32_MAX here, as that is
13029 * handled by the RExC_npar check */
13032 /* any numeric escape < 10 is always a backref */
13034 /* any numeric escape < RExC_npar is a backref */
13035 && num >= RExC_npar
13036 /* cannot be an octal escape if it starts with 8 */
13037 && *RExC_parse != '8'
13038 /* cannot be an octal escape it it starts with 9 */
13039 && *RExC_parse != '9'
13042 /* Probably not a backref, instead likely to be an
13043 * octal character escape, e.g. \35 or \777.
13044 * The above logic should make it obvious why using
13045 * octal escapes in patterns is problematic. - Yves */
13046 RExC_parse = parse_start;
13051 /* At this point RExC_parse points at a numeric escape like
13052 * \12 or \88 or something similar, which we should NOT treat
13053 * as an octal escape. It may or may not be a valid backref
13054 * escape. For instance \88888888 is unlikely to be a valid
13056 while (isDIGIT(*RExC_parse))
13059 if (*RExC_parse != '}')
13060 vFAIL("Unterminated \\g{...} pattern");
13064 if (num > (I32)RExC_rx->nparens)
13065 vFAIL("Reference to nonexistent group");
13068 ret = reganode(pRExC_state,
13071 : (ASCII_FOLD_RESTRICTED)
13073 : (AT_LEAST_UNI_SEMANTICS)
13079 *flagp |= HASWIDTH;
13081 /* override incorrect value set in reganode MJD */
13082 Set_Node_Offset(ret, parse_start);
13083 Set_Node_Cur_Length(ret, parse_start-1);
13084 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13085 FALSE /* Don't force to /x */ );
13089 if (RExC_parse >= RExC_end)
13090 FAIL("Trailing \\");
13093 /* Do not generate "unrecognized" warnings here, we fall
13094 back into the quick-grab loop below */
13095 RExC_parse = parse_start;
13097 } /* end of switch on a \foo sequence */
13102 /* '#' comments should have been spaced over before this function was
13104 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13106 if (RExC_flags & RXf_PMf_EXTENDED) {
13107 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13108 if (RExC_parse < RExC_end)
13118 /* Here, we have determined that the next thing is probably a
13119 * literal character. RExC_parse points to the first byte of its
13120 * definition. (It still may be an escape sequence that evaluates
13121 * to a single character) */
13127 #define MAX_NODE_STRING_SIZE 127
13128 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
13130 U8 upper_parse = MAX_NODE_STRING_SIZE;
13131 U8 node_type = compute_EXACTish(pRExC_state);
13132 bool next_is_quantifier;
13133 char * oldp = NULL;
13135 /* We can convert EXACTF nodes to EXACTFU if they contain only
13136 * characters that match identically regardless of the target
13137 * string's UTF8ness. The reason to do this is that EXACTF is not
13138 * trie-able, EXACTFU is.
13140 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13141 * contain only above-Latin1 characters (hence must be in UTF8),
13142 * which don't participate in folds with Latin1-range characters,
13143 * as the latter's folds aren't known until runtime. (We don't
13144 * need to figure this out until pass 2) */
13145 bool maybe_exactfu = PASS2
13146 && (node_type == EXACTF || node_type == EXACTFL);
13148 /* If a folding node contains only code points that don't
13149 * participate in folds, it can be changed into an EXACT node,
13150 * which allows the optimizer more things to look for */
13153 ret = reg_node(pRExC_state, node_type);
13155 /* In pass1, folded, we use a temporary buffer instead of the
13156 * actual node, as the node doesn't exist yet */
13157 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
13163 /* We look for the EXACTFish to EXACT node optimizaton only if
13164 * folding. (And we don't need to figure this out until pass 2).
13165 * XXX It might actually make sense to split the node into portions
13166 * that are exact and ones that aren't, so that we could later use
13167 * the exact ones to find the longest fixed and floating strings.
13168 * One would want to join them back into a larger node. One could
13169 * use a pseudo regnode like 'EXACT_ORIG_FOLD' */
13170 maybe_exact = FOLD && PASS2;
13172 /* XXX The node can hold up to 255 bytes, yet this only goes to
13173 * 127. I (khw) do not know why. Keeping it somewhat less than
13174 * 255 allows us to not have to worry about overflow due to
13175 * converting to utf8 and fold expansion, but that value is
13176 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
13177 * split up by this limit into a single one using the real max of
13178 * 255. Even at 127, this breaks under rare circumstances. If
13179 * folding, we do not want to split a node at a character that is a
13180 * non-final in a multi-char fold, as an input string could just
13181 * happen to want to match across the node boundary. The join
13182 * would solve that problem if the join actually happens. But a
13183 * series of more than two nodes in a row each of 127 would cause
13184 * the first join to succeed to get to 254, but then there wouldn't
13185 * be room for the next one, which could at be one of those split
13186 * multi-char folds. I don't know of any fool-proof solution. One
13187 * could back off to end with only a code point that isn't such a
13188 * non-final, but it is possible for there not to be any in the
13191 assert( ! UTF /* Is at the beginning of a character */
13192 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13193 || UTF8_IS_START(UCHARAT(RExC_parse)));
13195 /* Here, we have a literal character. Find the maximal string of
13196 * them in the input that we can fit into a single EXACTish node.
13197 * We quit at the first non-literal or when the node gets full */
13198 for (p = RExC_parse;
13199 len < upper_parse && p < RExC_end;
13204 /* White space has already been ignored */
13205 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13206 || ! is_PATWS_safe((p), RExC_end, UTF));
13218 /* Literal Escapes Switch
13220 This switch is meant to handle escape sequences that
13221 resolve to a literal character.
13223 Every escape sequence that represents something
13224 else, like an assertion or a char class, is handled
13225 in the switch marked 'Special Escapes' above in this
13226 routine, but also has an entry here as anything that
13227 isn't explicitly mentioned here will be treated as
13228 an unescaped equivalent literal.
13231 switch ((U8)*++p) {
13232 /* These are all the special escapes. */
13233 case 'A': /* Start assertion */
13234 case 'b': case 'B': /* Word-boundary assertion*/
13235 case 'C': /* Single char !DANGEROUS! */
13236 case 'd': case 'D': /* digit class */
13237 case 'g': case 'G': /* generic-backref, pos assertion */
13238 case 'h': case 'H': /* HORIZWS */
13239 case 'k': case 'K': /* named backref, keep marker */
13240 case 'p': case 'P': /* Unicode property */
13241 case 'R': /* LNBREAK */
13242 case 's': case 'S': /* space class */
13243 case 'v': case 'V': /* VERTWS */
13244 case 'w': case 'W': /* word class */
13245 case 'X': /* eXtended Unicode "combining
13246 character sequence" */
13247 case 'z': case 'Z': /* End of line/string assertion */
13251 /* Anything after here is an escape that resolves to a
13252 literal. (Except digits, which may or may not)
13258 case 'N': /* Handle a single-code point named character. */
13259 RExC_parse = p + 1;
13260 if (! grok_bslash_N(pRExC_state,
13261 NULL, /* Fail if evaluates to
13262 anything other than a
13263 single code point */
13264 &ender, /* The returned single code
13266 NULL, /* Don't need a count of
13267 how many code points */
13272 if (*flagp & NEED_UTF8)
13273 FAIL("panic: grok_bslash_N set NEED_UTF8");
13274 if (*flagp & RESTART_PASS1)
13277 /* Here, it wasn't a single code point. Go close
13278 * up this EXACTish node. The switch() prior to
13279 * this switch handles the other cases */
13280 RExC_parse = p = oldp;
13284 if (ender > 0xff) {
13285 REQUIRE_UTF8(flagp);
13301 ender = ESC_NATIVE;
13311 const char* error_msg;
13313 bool valid = grok_bslash_o(&p,
13316 PASS2, /* out warnings */
13317 (bool) RExC_strict,
13318 TRUE, /* Output warnings
13323 RExC_parse = p; /* going to die anyway; point
13324 to exact spot of failure */
13328 if (ender > 0xff) {
13329 REQUIRE_UTF8(flagp);
13335 UV result = UV_MAX; /* initialize to erroneous
13337 const char* error_msg;
13339 bool valid = grok_bslash_x(&p,
13342 PASS2, /* out warnings */
13343 (bool) RExC_strict,
13344 TRUE, /* Silence warnings
13349 RExC_parse = p; /* going to die anyway; point
13350 to exact spot of failure */
13355 if (ender < 0x100) {
13357 if (RExC_recode_x_to_native) {
13358 ender = LATIN1_TO_NATIVE(ender);
13363 REQUIRE_UTF8(flagp);
13369 ender = grok_bslash_c(*p++, PASS2);
13371 case '8': case '9': /* must be a backreference */
13373 /* we have an escape like \8 which cannot be an octal escape
13374 * so we exit the loop, and let the outer loop handle this
13375 * escape which may or may not be a legitimate backref. */
13377 case '1': case '2': case '3':case '4':
13378 case '5': case '6': case '7':
13379 /* When we parse backslash escapes there is ambiguity
13380 * between backreferences and octal escapes. Any escape
13381 * from \1 - \9 is a backreference, any multi-digit
13382 * escape which does not start with 0 and which when
13383 * evaluated as decimal could refer to an already
13384 * parsed capture buffer is a back reference. Anything
13387 * Note this implies that \118 could be interpreted as
13388 * 118 OR as "\11" . "8" depending on whether there
13389 * were 118 capture buffers defined already in the
13392 /* NOTE, RExC_npar is 1 more than the actual number of
13393 * parens we have seen so far, hence the < RExC_npar below. */
13395 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
13396 { /* Not to be treated as an octal constant, go
13404 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13406 ender = grok_oct(p, &numlen, &flags, NULL);
13407 if (ender > 0xff) {
13408 REQUIRE_UTF8(flagp);
13411 if (PASS2 /* like \08, \178 */
13413 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
13415 reg_warn_non_literal_string(
13417 form_short_octal_warning(p, numlen));
13423 FAIL("Trailing \\");
13426 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
13427 /* Include any left brace following the alpha to emphasize
13428 * that it could be part of an escape at some point
13430 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
13431 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
13433 goto normal_default;
13434 } /* End of switch on '\' */
13437 /* Currently we allow an lbrace at the start of a construct
13438 * without raising a warning. This is because we think we
13439 * will never want such a brace to be meant to be other
13440 * than taken literally. */
13441 if (len || (p > RExC_start && isALPHA_A(*(p - 1)))) {
13443 /* But, we raise a fatal warning otherwise, as the
13444 * deprecation cycle has come and gone. Except that it
13445 * turns out that some heavily-relied on upstream
13446 * software, notably GNU Autoconf, have failed to fix
13447 * their uses. For these, don't make it fatal unless
13448 * we anticipate using the '{' for something else.
13449 * This happens after any alpha, and for a looser {m,n}
13450 * quantifier specification */
13452 || ( p > parse_start + 1
13453 && isALPHA_A(*(p - 1))
13454 && *(p - 2) == '\\')
13455 || new_regcurly(p, RExC_end))
13457 RExC_parse = p + 1;
13458 vFAIL("Unescaped left brace in regex is "
13462 ckWARNregdep(p + 1,
13463 "Unescaped left brace in regex is "
13464 "deprecated here (and will be fatal "
13465 "in Perl 5.30), passed through");
13468 goto normal_default;
13471 if (PASS2 && p > RExC_parse && RExC_strict) {
13472 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
13475 default: /* A literal character */
13477 if (! UTF8_IS_INVARIANT(*p) && UTF) {
13479 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
13480 &numlen, UTF8_ALLOW_DEFAULT);
13486 } /* End of switch on the literal */
13488 /* Here, have looked at the literal character and <ender>
13489 * contains its ordinal, <p> points to the character after it.
13490 * We need to check if the next non-ignored thing is a
13491 * quantifier. Move <p> to after anything that should be
13492 * ignored, which, as a side effect, positions <p> for the next
13493 * loop iteration */
13494 skip_to_be_ignored_text(pRExC_state, &p,
13495 FALSE /* Don't force to /x */ );
13497 /* If the next thing is a quantifier, it applies to this
13498 * character only, which means that this character has to be in
13499 * its own node and can't just be appended to the string in an
13500 * existing node, so if there are already other characters in
13501 * the node, close the node with just them, and set up to do
13502 * this character again next time through, when it will be the
13503 * only thing in its new node */
13505 next_is_quantifier = LIKELY(p < RExC_end)
13506 && UNLIKELY(ISMULT2(p));
13508 if (next_is_quantifier && LIKELY(len)) {
13513 /* Ready to add 'ender' to the node */
13515 if (! FOLD) { /* The simple case, just append the literal */
13517 /* In the sizing pass, we need only the size of the
13518 * character we are appending, hence we can delay getting
13519 * its representation until PASS2. */
13522 const STRLEN unilen = UVCHR_SKIP(ender);
13525 /* We have to subtract 1 just below (and again in
13526 * the corresponding PASS2 code) because the loop
13527 * increments <len> each time, as all but this path
13528 * (and one other) through it add a single byte to
13529 * the EXACTish node. But these paths would change
13530 * len to be the correct final value, so cancel out
13531 * the increment that follows */
13537 } else { /* PASS2 */
13540 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
13541 len += (char *) new_s - s - 1;
13542 s = (char *) new_s;
13545 *(s++) = (char) ender;
13549 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
13551 /* Here are folding under /l, and the code point is
13552 * problematic. First, we know we can't simplify things */
13553 maybe_exact = FALSE;
13554 maybe_exactfu = FALSE;
13556 /* A problematic code point in this context means that its
13557 * fold isn't known until runtime, so we can't fold it now.
13558 * (The non-problematic code points are the above-Latin1
13559 * ones that fold to also all above-Latin1. Their folds
13560 * don't vary no matter what the locale is.) But here we
13561 * have characters whose fold depends on the locale.
13562 * Unlike the non-folding case above, we have to keep track
13563 * of these in the sizing pass, so that we can make sure we
13564 * don't split too-long nodes in the middle of a potential
13565 * multi-char fold. And unlike the regular fold case
13566 * handled in the else clauses below, we don't actually
13567 * fold and don't have special cases to consider. What we
13568 * do for both passes is the PASS2 code for non-folding */
13569 goto not_fold_common;
13571 else /* A regular FOLD code point */
13573 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13574 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13575 || UNICODE_DOT_DOT_VERSION > 0)
13576 /* See comments for join_exact() as to why we fold
13577 * this non-UTF at compile time */
13578 || ( node_type == EXACTFU
13579 && ender == LATIN_SMALL_LETTER_SHARP_S)
13582 /* Here, are folding and are not UTF-8 encoded; therefore
13583 * the character must be in the range 0-255, and is not /l
13584 * (Not /l because we already handled these under /l in
13585 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
13586 if (IS_IN_SOME_FOLD_L1(ender)) {
13587 maybe_exact = FALSE;
13589 /* See if the character's fold differs between /d and
13590 * /u. This includes the multi-char fold SHARP S to
13592 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
13593 RExC_seen_unfolded_sharp_s = 1;
13594 maybe_exactfu = FALSE;
13596 else if (maybe_exactfu
13597 && (PL_fold[ender] != PL_fold_latin1[ender]
13598 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13599 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13600 || UNICODE_DOT_DOT_VERSION > 0)
13602 && isALPHA_FOLD_EQ(ender, 's')
13603 && isALPHA_FOLD_EQ(*(s-1), 's'))
13606 maybe_exactfu = FALSE;
13610 /* Even when folding, we store just the input character, as
13611 * we have an array that finds its fold quickly */
13612 *(s++) = (char) ender;
13614 else { /* FOLD, and UTF (or sharp s) */
13615 /* Unlike the non-fold case, we do actually have to
13616 * calculate the results here in pass 1. This is for two
13617 * reasons, the folded length may be longer than the
13618 * unfolded, and we have to calculate how many EXACTish
13619 * nodes it will take; and we may run out of room in a node
13620 * in the middle of a potential multi-char fold, and have
13621 * to back off accordingly. */
13624 if (isASCII_uni(ender)) {
13625 folded = toFOLD(ender);
13626 *(s)++ = (U8) folded;
13631 folded = _to_uni_fold_flags(
13635 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
13636 ? FOLD_FLAGS_NOMIX_ASCII
13640 /* The loop increments <len> each time, as all but this
13641 * path (and one other) through it add a single byte to
13642 * the EXACTish node. But this one has changed len to
13643 * be the correct final value, so subtract one to
13644 * cancel out the increment that follows */
13645 len += foldlen - 1;
13647 /* If this node only contains non-folding code points so
13648 * far, see if this new one is also non-folding */
13650 if (folded != ender) {
13651 maybe_exact = FALSE;
13654 /* Here the fold is the original; we have to check
13655 * further to see if anything folds to it */
13656 if (_invlist_contains_cp(PL_utf8_foldable,
13659 maybe_exact = FALSE;
13666 if (next_is_quantifier) {
13668 /* Here, the next input is a quantifier, and to get here,
13669 * the current character is the only one in the node.
13670 * Also, here <len> doesn't include the final byte for this
13676 } /* End of loop through literal characters */
13678 /* Here we have either exhausted the input or ran out of room in
13679 * the node. (If we encountered a character that can't be in the
13680 * node, transfer is made directly to <loopdone>, and so we
13681 * wouldn't have fallen off the end of the loop.) In the latter
13682 * case, we artificially have to split the node into two, because
13683 * we just don't have enough space to hold everything. This
13684 * creates a problem if the final character participates in a
13685 * multi-character fold in the non-final position, as a match that
13686 * should have occurred won't, due to the way nodes are matched,
13687 * and our artificial boundary. So back off until we find a non-
13688 * problematic character -- one that isn't at the beginning or
13689 * middle of such a fold. (Either it doesn't participate in any
13690 * folds, or appears only in the final position of all the folds it
13691 * does participate in.) A better solution with far fewer false
13692 * positives, and that would fill the nodes more completely, would
13693 * be to actually have available all the multi-character folds to
13694 * test against, and to back-off only far enough to be sure that
13695 * this node isn't ending with a partial one. <upper_parse> is set
13696 * further below (if we need to reparse the node) to include just
13697 * up through that final non-problematic character that this code
13698 * identifies, so when it is set to less than the full node, we can
13699 * skip the rest of this */
13700 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
13702 const STRLEN full_len = len;
13704 assert(len >= MAX_NODE_STRING_SIZE);
13706 /* Here, <s> points to the final byte of the final character.
13707 * Look backwards through the string until find a non-
13708 * problematic character */
13712 /* This has no multi-char folds to non-UTF characters */
13713 if (ASCII_FOLD_RESTRICTED) {
13717 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
13721 if (! PL_NonL1NonFinalFold) {
13722 PL_NonL1NonFinalFold = _new_invlist_C_array(
13723 NonL1_Perl_Non_Final_Folds_invlist);
13726 /* Point to the first byte of the final character */
13727 s = (char *) utf8_hop((U8 *) s, -1);
13729 while (s >= s0) { /* Search backwards until find
13730 non-problematic char */
13731 if (UTF8_IS_INVARIANT(*s)) {
13733 /* There are no ascii characters that participate
13734 * in multi-char folds under /aa. In EBCDIC, the
13735 * non-ascii invariants are all control characters,
13736 * so don't ever participate in any folds. */
13737 if (ASCII_FOLD_RESTRICTED
13738 || ! IS_NON_FINAL_FOLD(*s))
13743 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
13744 if (! IS_NON_FINAL_FOLD(EIGHT_BIT_UTF8_TO_NATIVE(
13750 else if (! _invlist_contains_cp(
13751 PL_NonL1NonFinalFold,
13752 valid_utf8_to_uvchr((U8 *) s, NULL)))
13757 /* Here, the current character is problematic in that
13758 * it does occur in the non-final position of some
13759 * fold, so try the character before it, but have to
13760 * special case the very first byte in the string, so
13761 * we don't read outside the string */
13762 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
13763 } /* End of loop backwards through the string */
13765 /* If there were only problematic characters in the string,
13766 * <s> will point to before s0, in which case the length
13767 * should be 0, otherwise include the length of the
13768 * non-problematic character just found */
13769 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
13772 /* Here, have found the final character, if any, that is
13773 * non-problematic as far as ending the node without splitting
13774 * it across a potential multi-char fold. <len> contains the
13775 * number of bytes in the node up-to and including that
13776 * character, or is 0 if there is no such character, meaning
13777 * the whole node contains only problematic characters. In
13778 * this case, give up and just take the node as-is. We can't
13783 /* If the node ends in an 's' we make sure it stays EXACTF,
13784 * as if it turns into an EXACTFU, it could later get
13785 * joined with another 's' that would then wrongly match
13787 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
13789 maybe_exactfu = FALSE;
13793 /* Here, the node does contain some characters that aren't
13794 * problematic. If one such is the final character in the
13795 * node, we are done */
13796 if (len == full_len) {
13799 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
13801 /* If the final character is problematic, but the
13802 * penultimate is not, back-off that last character to
13803 * later start a new node with it */
13808 /* Here, the final non-problematic character is earlier
13809 * in the input than the penultimate character. What we do
13810 * is reparse from the beginning, going up only as far as
13811 * this final ok one, thus guaranteeing that the node ends
13812 * in an acceptable character. The reason we reparse is
13813 * that we know how far in the character is, but we don't
13814 * know how to correlate its position with the input parse.
13815 * An alternate implementation would be to build that
13816 * correlation as we go along during the original parse,
13817 * but that would entail extra work for every node, whereas
13818 * this code gets executed only when the string is too
13819 * large for the node, and the final two characters are
13820 * problematic, an infrequent occurrence. Yet another
13821 * possible strategy would be to save the tail of the
13822 * string, and the next time regatom is called, initialize
13823 * with that. The problem with this is that unless you
13824 * back off one more character, you won't be guaranteed
13825 * regatom will get called again, unless regbranch,
13826 * regpiece ... are also changed. If you do back off that
13827 * extra character, so that there is input guaranteed to
13828 * force calling regatom, you can't handle the case where
13829 * just the first character in the node is acceptable. I
13830 * (khw) decided to try this method which doesn't have that
13831 * pitfall; if performance issues are found, we can do a
13832 * combination of the current approach plus that one */
13838 } /* End of verifying node ends with an appropriate char */
13840 loopdone: /* Jumped to when encounters something that shouldn't be
13843 /* I (khw) don't know if you can get here with zero length, but the
13844 * old code handled this situation by creating a zero-length EXACT
13845 * node. Might as well be NOTHING instead */
13851 /* If 'maybe_exact' is still set here, means there are no
13852 * code points in the node that participate in folds;
13853 * similarly for 'maybe_exactfu' and code points that match
13854 * differently depending on UTF8ness of the target string
13855 * (for /u), or depending on locale for /l */
13861 else if (maybe_exactfu) {
13867 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
13868 FALSE /* Don't look to see if could
13869 be turned into an EXACT
13870 node, as we have already
13875 RExC_parse = p - 1;
13876 Set_Node_Cur_Length(ret, parse_start);
13879 /* len is STRLEN which is unsigned, need to copy to signed */
13882 vFAIL("Internal disaster");
13885 } /* End of label 'defchar:' */
13887 } /* End of giant switch on input character */
13889 /* Position parse to next real character */
13890 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13891 FALSE /* Don't force to /x */ );
13892 if (PASS2 && *RExC_parse == '{' && OP(ret) != SBOL && ! regcurly(RExC_parse)) {
13893 ckWARNregdep(RExC_parse + 1, "Unescaped left brace in regex is deprecated here (and will be fatal in Perl 5.30), passed through");
13901 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
13903 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
13904 * sets up the bitmap and any flags, removing those code points from the
13905 * inversion list, setting it to NULL should it become completely empty */
13907 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
13908 assert(PL_regkind[OP(node)] == ANYOF);
13910 ANYOF_BITMAP_ZERO(node);
13911 if (*invlist_ptr) {
13913 /* This gets set if we actually need to modify things */
13914 bool change_invlist = FALSE;
13918 /* Start looking through *invlist_ptr */
13919 invlist_iterinit(*invlist_ptr);
13920 while (invlist_iternext(*invlist_ptr, &start, &end)) {
13924 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
13925 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
13928 /* Quit if are above what we should change */
13929 if (start >= NUM_ANYOF_CODE_POINTS) {
13933 change_invlist = TRUE;
13935 /* Set all the bits in the range, up to the max that we are doing */
13936 high = (end < NUM_ANYOF_CODE_POINTS - 1)
13938 : NUM_ANYOF_CODE_POINTS - 1;
13939 for (i = start; i <= (int) high; i++) {
13940 if (! ANYOF_BITMAP_TEST(node, i)) {
13941 ANYOF_BITMAP_SET(node, i);
13945 invlist_iterfinish(*invlist_ptr);
13947 /* Done with loop; remove any code points that are in the bitmap from
13948 * *invlist_ptr; similarly for code points above the bitmap if we have
13949 * a flag to match all of them anyways */
13950 if (change_invlist) {
13951 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
13953 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
13954 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
13957 /* If have completely emptied it, remove it completely */
13958 if (_invlist_len(*invlist_ptr) == 0) {
13959 SvREFCNT_dec_NN(*invlist_ptr);
13960 *invlist_ptr = NULL;
13965 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
13966 Character classes ([:foo:]) can also be negated ([:^foo:]).
13967 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
13968 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
13969 but trigger failures because they are currently unimplemented. */
13971 #define POSIXCC_DONE(c) ((c) == ':')
13972 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
13973 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
13974 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
13976 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
13977 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
13978 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
13980 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
13982 /* 'posix_warnings' and 'warn_text' are names of variables in the following
13984 #define ADD_POSIX_WARNING(p, text) STMT_START { \
13985 if (posix_warnings) { \
13986 if (! RExC_warn_text ) RExC_warn_text = (AV *) sv_2mortal((SV *) newAV()); \
13987 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
13991 REPORT_LOCATION_ARGS(p))); \
13996 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
13998 const char * const s, /* Where the putative posix class begins.
13999 Normally, this is one past the '['. This
14000 parameter exists so it can be somewhere
14001 besides RExC_parse. */
14002 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
14004 AV ** posix_warnings, /* Where to place any generated warnings, or
14006 const bool check_only /* Don't die if error */
14009 /* This parses what the caller thinks may be one of the three POSIX
14011 * 1) a character class, like [:blank:]
14012 * 2) a collating symbol, like [. .]
14013 * 3) an equivalence class, like [= =]
14014 * In the latter two cases, it croaks if it finds a syntactically legal
14015 * one, as these are not handled by Perl.
14017 * The main purpose is to look for a POSIX character class. It returns:
14018 * a) the class number
14019 * if it is a completely syntactically and semantically legal class.
14020 * 'updated_parse_ptr', if not NULL, is set to point to just after the
14021 * closing ']' of the class
14022 * b) OOB_NAMEDCLASS
14023 * if it appears that one of the three POSIX constructs was meant, but
14024 * its specification was somehow defective. 'updated_parse_ptr', if
14025 * not NULL, is set to point to the character just after the end
14026 * character of the class. See below for handling of warnings.
14027 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
14028 * if it doesn't appear that a POSIX construct was intended.
14029 * 'updated_parse_ptr' is not changed. No warnings nor errors are
14032 * In b) there may be errors or warnings generated. If 'check_only' is
14033 * TRUE, then any errors are discarded. Warnings are returned to the
14034 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
14035 * instead it is NULL, warnings are suppressed. This is done in all
14036 * passes. The reason for this is that the rest of the parsing is heavily
14037 * dependent on whether this routine found a valid posix class or not. If
14038 * it did, the closing ']' is absorbed as part of the class. If no class,
14039 * or an invalid one is found, any ']' will be considered the terminator of
14040 * the outer bracketed character class, leading to very different results.
14041 * In particular, a '(?[ ])' construct will likely have a syntax error if
14042 * the class is parsed other than intended, and this will happen in pass1,
14043 * before the warnings would normally be output. This mechanism allows the
14044 * caller to output those warnings in pass1 just before dieing, giving a
14045 * much better clue as to what is wrong.
14047 * The reason for this function, and its complexity is that a bracketed
14048 * character class can contain just about anything. But it's easy to
14049 * mistype the very specific posix class syntax but yielding a valid
14050 * regular bracketed class, so it silently gets compiled into something
14051 * quite unintended.
14053 * The solution adopted here maintains backward compatibility except that
14054 * it adds a warning if it looks like a posix class was intended but
14055 * improperly specified. The warning is not raised unless what is input
14056 * very closely resembles one of the 14 legal posix classes. To do this,
14057 * it uses fuzzy parsing. It calculates how many single-character edits it
14058 * would take to transform what was input into a legal posix class. Only
14059 * if that number is quite small does it think that the intention was a
14060 * posix class. Obviously these are heuristics, and there will be cases
14061 * where it errs on one side or another, and they can be tweaked as
14062 * experience informs.
14064 * The syntax for a legal posix class is:
14066 * qr/(?xa: \[ : \^? [:lower:]{4,6} : \] )/
14068 * What this routine considers syntactically to be an intended posix class
14069 * is this (the comments indicate some restrictions that the pattern
14072 * qr/(?x: \[? # The left bracket, possibly
14074 * \h* # possibly followed by blanks
14075 * (?: \^ \h* )? # possibly a misplaced caret
14076 * [:;]? # The opening class character,
14077 * # possibly omitted. A typo
14078 * # semi-colon can also be used.
14080 * \^? # possibly a correctly placed
14081 * # caret, but not if there was also
14082 * # a misplaced one
14084 * .{3,15} # The class name. If there are
14085 * # deviations from the legal syntax,
14086 * # its edit distance must be close
14087 * # to a real class name in order
14088 * # for it to be considered to be
14089 * # an intended posix class.
14091 * [:punct:]? # The closing class character,
14092 * # possibly omitted. If not a colon
14093 * # nor semi colon, the class name
14094 * # must be even closer to a valid
14097 * \]? # The right bracket, possibly
14101 * In the above, \h must be ASCII-only.
14103 * These are heuristics, and can be tweaked as field experience dictates.
14104 * There will be cases when someone didn't intend to specify a posix class
14105 * that this warns as being so. The goal is to minimize these, while
14106 * maximizing the catching of things intended to be a posix class that
14107 * aren't parsed as such.
14111 const char * const e = RExC_end;
14112 unsigned complement = 0; /* If to complement the class */
14113 bool found_problem = FALSE; /* Assume OK until proven otherwise */
14114 bool has_opening_bracket = FALSE;
14115 bool has_opening_colon = FALSE;
14116 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
14118 const char * possible_end = NULL; /* used for a 2nd parse pass */
14119 const char* name_start; /* ptr to class name first char */
14121 /* If the number of single-character typos the input name is away from a
14122 * legal name is no more than this number, it is considered to have meant
14123 * the legal name */
14124 int max_distance = 2;
14126 /* to store the name. The size determines the maximum length before we
14127 * decide that no posix class was intended. Should be at least
14128 * sizeof("alphanumeric") */
14130 STATIC_ASSERT_DECL(C_ARRAY_LENGTH(input_text) >= sizeof "alphanumeric");
14132 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
14134 if (posix_warnings && RExC_warn_text)
14135 av_clear(RExC_warn_text);
14138 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14141 if (*(p - 1) != '[') {
14142 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
14143 found_problem = TRUE;
14146 has_opening_bracket = TRUE;
14149 /* They could be confused and think you can put spaces between the
14152 found_problem = TRUE;
14156 } while (p < e && isBLANK(*p));
14158 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14161 /* For [. .] and [= =]. These are quite different internally from [: :],
14162 * so they are handled separately. */
14163 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
14164 and 1 for at least one char in it
14167 const char open_char = *p;
14168 const char * temp_ptr = p + 1;
14170 /* These two constructs are not handled by perl, and if we find a
14171 * syntactically valid one, we croak. khw, who wrote this code, finds
14172 * this explanation of them very unclear:
14173 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
14174 * And searching the rest of the internet wasn't very helpful either.
14175 * It looks like just about any byte can be in these constructs,
14176 * depending on the locale. But unless the pattern is being compiled
14177 * under /l, which is very rare, Perl runs under the C or POSIX locale.
14178 * In that case, it looks like [= =] isn't allowed at all, and that
14179 * [. .] could be any single code point, but for longer strings the
14180 * constituent characters would have to be the ASCII alphabetics plus
14181 * the minus-hyphen. Any sensible locale definition would limit itself
14182 * to these. And any portable one definitely should. Trying to parse
14183 * the general case is a nightmare (see [perl #127604]). So, this code
14184 * looks only for interiors of these constructs that match:
14186 * Using \w relaxes the apparent rules a little, without adding much
14187 * danger of mistaking something else for one of these constructs.
14189 * [. .] in some implementations described on the internet is usable to
14190 * escape a character that otherwise is special in bracketed character
14191 * classes. For example [.].] means a literal right bracket instead of
14192 * the ending of the class
14194 * [= =] can legitimately contain a [. .] construct, but we don't
14195 * handle this case, as that [. .] construct will later get parsed
14196 * itself and croak then. And [= =] is checked for even when not under
14197 * /l, as Perl has long done so.
14199 * The code below relies on there being a trailing NUL, so it doesn't
14200 * have to keep checking if the parse ptr < e.
14202 if (temp_ptr[1] == open_char) {
14205 else while ( temp_ptr < e
14206 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
14211 if (*temp_ptr == open_char) {
14213 if (*temp_ptr == ']') {
14215 if (! found_problem && ! check_only) {
14216 RExC_parse = (char *) temp_ptr;
14217 vFAIL3("POSIX syntax [%c %c] is reserved for future "
14218 "extensions", open_char, open_char);
14221 /* Here, the syntax wasn't completely valid, or else the call
14222 * is to check-only */
14223 if (updated_parse_ptr) {
14224 *updated_parse_ptr = (char *) temp_ptr;
14227 return OOB_NAMEDCLASS;
14231 /* If we find something that started out to look like one of these
14232 * constructs, but isn't, we continue below so that it can be checked
14233 * for being a class name with a typo of '.' or '=' instead of a colon.
14237 /* Here, we think there is a possibility that a [: :] class was meant, and
14238 * we have the first real character. It could be they think the '^' comes
14241 found_problem = TRUE;
14242 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
14247 found_problem = TRUE;
14251 } while (p < e && isBLANK(*p));
14253 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14257 /* But the first character should be a colon, which they could have easily
14258 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
14259 * distinguish from a colon, so treat that as a colon). */
14262 has_opening_colon = TRUE;
14264 else if (*p == ';') {
14265 found_problem = TRUE;
14267 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14268 has_opening_colon = TRUE;
14271 found_problem = TRUE;
14272 ADD_POSIX_WARNING(p, "there must be a starting ':'");
14274 /* Consider an initial punctuation (not one of the recognized ones) to
14275 * be a left terminator */
14276 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
14281 /* They may think that you can put spaces between the components */
14283 found_problem = TRUE;
14287 } while (p < e && isBLANK(*p));
14289 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14294 /* We consider something like [^:^alnum:]] to not have been intended to
14295 * be a posix class, but XXX maybe we should */
14297 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14304 /* Again, they may think that you can put spaces between the components */
14306 found_problem = TRUE;
14310 } while (p < e && isBLANK(*p));
14312 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14317 /* XXX This ']' may be a typo, and something else was meant. But
14318 * treating it as such creates enough complications, that that
14319 * possibility isn't currently considered here. So we assume that the
14320 * ']' is what is intended, and if we've already found an initial '[',
14321 * this leaves this construct looking like [:] or [:^], which almost
14322 * certainly weren't intended to be posix classes */
14323 if (has_opening_bracket) {
14324 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14327 /* But this function can be called when we parse the colon for
14328 * something like qr/[alpha:]]/, so we back up to look for the
14333 found_problem = TRUE;
14334 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14336 else if (*p != ':') {
14338 /* XXX We are currently very restrictive here, so this code doesn't
14339 * consider the possibility that, say, /[alpha.]]/ was intended to
14340 * be a posix class. */
14341 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14344 /* Here we have something like 'foo:]'. There was no initial colon,
14345 * and we back up over 'foo. XXX Unlike the going forward case, we
14346 * don't handle typos of non-word chars in the middle */
14347 has_opening_colon = FALSE;
14350 while (p > RExC_start && isWORDCHAR(*p)) {
14355 /* Here, we have positioned ourselves to where we think the first
14356 * character in the potential class is */
14359 /* Now the interior really starts. There are certain key characters that
14360 * can end the interior, or these could just be typos. To catch both
14361 * cases, we may have to do two passes. In the first pass, we keep on
14362 * going unless we come to a sequence that matches
14363 * qr/ [[:punct:]] [[:blank:]]* \] /xa
14364 * This means it takes a sequence to end the pass, so two typos in a row if
14365 * that wasn't what was intended. If the class is perfectly formed, just
14366 * this one pass is needed. We also stop if there are too many characters
14367 * being accumulated, but this number is deliberately set higher than any
14368 * real class. It is set high enough so that someone who thinks that
14369 * 'alphanumeric' is a correct name would get warned that it wasn't.
14370 * While doing the pass, we keep track of where the key characters were in
14371 * it. If we don't find an end to the class, and one of the key characters
14372 * was found, we redo the pass, but stop when we get to that character.
14373 * Thus the key character was considered a typo in the first pass, but a
14374 * terminator in the second. If two key characters are found, we stop at
14375 * the second one in the first pass. Again this can miss two typos, but
14376 * catches a single one
14378 * In the first pass, 'possible_end' starts as NULL, and then gets set to
14379 * point to the first key character. For the second pass, it starts as -1.
14385 bool has_blank = FALSE;
14386 bool has_upper = FALSE;
14387 bool has_terminating_colon = FALSE;
14388 bool has_terminating_bracket = FALSE;
14389 bool has_semi_colon = FALSE;
14390 unsigned int name_len = 0;
14391 int punct_count = 0;
14395 /* Squeeze out blanks when looking up the class name below */
14396 if (isBLANK(*p) ) {
14398 found_problem = TRUE;
14403 /* The name will end with a punctuation */
14405 const char * peek = p + 1;
14407 /* Treat any non-']' punctuation followed by a ']' (possibly
14408 * with intervening blanks) as trying to terminate the class.
14409 * ']]' is very likely to mean a class was intended (but
14410 * missing the colon), but the warning message that gets
14411 * generated shows the error position better if we exit the
14412 * loop at the bottom (eventually), so skip it here. */
14414 if (peek < e && isBLANK(*peek)) {
14416 found_problem = TRUE;
14419 } while (peek < e && isBLANK(*peek));
14422 if (peek < e && *peek == ']') {
14423 has_terminating_bracket = TRUE;
14425 has_terminating_colon = TRUE;
14427 else if (*p == ';') {
14428 has_semi_colon = TRUE;
14429 has_terminating_colon = TRUE;
14432 found_problem = TRUE;
14439 /* Here we have punctuation we thought didn't end the class.
14440 * Keep track of the position of the key characters that are
14441 * more likely to have been class-enders */
14442 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
14444 /* Allow just one such possible class-ender not actually
14445 * ending the class. */
14446 if (possible_end) {
14452 /* If we have too many punctuation characters, no use in
14454 if (++punct_count > max_distance) {
14458 /* Treat the punctuation as a typo. */
14459 input_text[name_len++] = *p;
14462 else if (isUPPER(*p)) { /* Use lowercase for lookup */
14463 input_text[name_len++] = toLOWER(*p);
14465 found_problem = TRUE;
14467 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
14468 input_text[name_len++] = *p;
14472 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
14476 /* The declaration of 'input_text' is how long we allow a potential
14477 * class name to be, before saying they didn't mean a class name at
14479 if (name_len >= C_ARRAY_LENGTH(input_text)) {
14484 /* We get to here when the possible class name hasn't been properly
14485 * terminated before:
14486 * 1) we ran off the end of the pattern; or
14487 * 2) found two characters, each of which might have been intended to
14488 * be the name's terminator
14489 * 3) found so many punctuation characters in the purported name,
14490 * that the edit distance to a valid one is exceeded
14491 * 4) we decided it was more characters than anyone could have
14492 * intended to be one. */
14494 found_problem = TRUE;
14496 /* In the final two cases, we know that looking up what we've
14497 * accumulated won't lead to a match, even a fuzzy one. */
14498 if ( name_len >= C_ARRAY_LENGTH(input_text)
14499 || punct_count > max_distance)
14501 /* If there was an intermediate key character that could have been
14502 * an intended end, redo the parse, but stop there */
14503 if (possible_end && possible_end != (char *) -1) {
14504 possible_end = (char *) -1; /* Special signal value to say
14505 we've done a first pass */
14510 /* Otherwise, it can't have meant to have been a class */
14511 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14514 /* If we ran off the end, and the final character was a punctuation
14515 * one, back up one, to look at that final one just below. Later, we
14516 * will restore the parse pointer if appropriate */
14517 if (name_len && p == e && isPUNCT(*(p-1))) {
14522 if (p < e && isPUNCT(*p)) {
14524 has_terminating_bracket = TRUE;
14526 /* If this is a 2nd ']', and the first one is just below this
14527 * one, consider that to be the real terminator. This gives a
14528 * uniform and better positioning for the warning message */
14530 && possible_end != (char *) -1
14531 && *possible_end == ']'
14532 && name_len && input_text[name_len - 1] == ']')
14537 /* And this is actually equivalent to having done the 2nd
14538 * pass now, so set it to not try again */
14539 possible_end = (char *) -1;
14544 has_terminating_colon = TRUE;
14546 else if (*p == ';') {
14547 has_semi_colon = TRUE;
14548 has_terminating_colon = TRUE;
14556 /* Here, we have a class name to look up. We can short circuit the
14557 * stuff below for short names that can't possibly be meant to be a
14558 * class name. (We can do this on the first pass, as any second pass
14559 * will yield an even shorter name) */
14560 if (name_len < 3) {
14561 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14564 /* Find which class it is. Initially switch on the length of the name.
14566 switch (name_len) {
14568 if (memEQ(name_start, "word", 4)) {
14569 /* this is not POSIX, this is the Perl \w */
14570 class_number = ANYOF_WORDCHAR;
14574 /* Names all of length 5: alnum alpha ascii blank cntrl digit
14575 * graph lower print punct space upper
14576 * Offset 4 gives the best switch position. */
14577 switch (name_start[4]) {
14579 if (memEQ(name_start, "alph", 4)) /* alpha */
14580 class_number = ANYOF_ALPHA;
14583 if (memEQ(name_start, "spac", 4)) /* space */
14584 class_number = ANYOF_SPACE;
14587 if (memEQ(name_start, "grap", 4)) /* graph */
14588 class_number = ANYOF_GRAPH;
14591 if (memEQ(name_start, "asci", 4)) /* ascii */
14592 class_number = ANYOF_ASCII;
14595 if (memEQ(name_start, "blan", 4)) /* blank */
14596 class_number = ANYOF_BLANK;
14599 if (memEQ(name_start, "cntr", 4)) /* cntrl */
14600 class_number = ANYOF_CNTRL;
14603 if (memEQ(name_start, "alnu", 4)) /* alnum */
14604 class_number = ANYOF_ALPHANUMERIC;
14607 if (memEQ(name_start, "lowe", 4)) /* lower */
14608 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
14609 else if (memEQ(name_start, "uppe", 4)) /* upper */
14610 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
14613 if (memEQ(name_start, "digi", 4)) /* digit */
14614 class_number = ANYOF_DIGIT;
14615 else if (memEQ(name_start, "prin", 4)) /* print */
14616 class_number = ANYOF_PRINT;
14617 else if (memEQ(name_start, "punc", 4)) /* punct */
14618 class_number = ANYOF_PUNCT;
14623 if (memEQ(name_start, "xdigit", 6))
14624 class_number = ANYOF_XDIGIT;
14628 /* If the name exactly matches a posix class name the class number will
14629 * here be set to it, and the input almost certainly was meant to be a
14630 * posix class, so we can skip further checking. If instead the syntax
14631 * is exactly correct, but the name isn't one of the legal ones, we
14632 * will return that as an error below. But if neither of these apply,
14633 * it could be that no posix class was intended at all, or that one
14634 * was, but there was a typo. We tease these apart by doing fuzzy
14635 * matching on the name */
14636 if (class_number == OOB_NAMEDCLASS && found_problem) {
14637 const UV posix_names[][6] = {
14638 { 'a', 'l', 'n', 'u', 'm' },
14639 { 'a', 'l', 'p', 'h', 'a' },
14640 { 'a', 's', 'c', 'i', 'i' },
14641 { 'b', 'l', 'a', 'n', 'k' },
14642 { 'c', 'n', 't', 'r', 'l' },
14643 { 'd', 'i', 'g', 'i', 't' },
14644 { 'g', 'r', 'a', 'p', 'h' },
14645 { 'l', 'o', 'w', 'e', 'r' },
14646 { 'p', 'r', 'i', 'n', 't' },
14647 { 'p', 'u', 'n', 'c', 't' },
14648 { 's', 'p', 'a', 'c', 'e' },
14649 { 'u', 'p', 'p', 'e', 'r' },
14650 { 'w', 'o', 'r', 'd' },
14651 { 'x', 'd', 'i', 'g', 'i', 't' }
14653 /* The names of the above all have added NULs to make them the same
14654 * size, so we need to also have the real lengths */
14655 const UV posix_name_lengths[] = {
14656 sizeof("alnum") - 1,
14657 sizeof("alpha") - 1,
14658 sizeof("ascii") - 1,
14659 sizeof("blank") - 1,
14660 sizeof("cntrl") - 1,
14661 sizeof("digit") - 1,
14662 sizeof("graph") - 1,
14663 sizeof("lower") - 1,
14664 sizeof("print") - 1,
14665 sizeof("punct") - 1,
14666 sizeof("space") - 1,
14667 sizeof("upper") - 1,
14668 sizeof("word") - 1,
14669 sizeof("xdigit")- 1
14672 int temp_max = max_distance; /* Use a temporary, so if we
14673 reparse, we haven't changed the
14676 /* Use a smaller max edit distance if we are missing one of the
14678 if ( has_opening_bracket + has_opening_colon < 2
14679 || has_terminating_bracket + has_terminating_colon < 2)
14684 /* See if the input name is close to a legal one */
14685 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
14687 /* Short circuit call if the lengths are too far apart to be
14689 if (abs( (int) (name_len - posix_name_lengths[i]))
14695 if (edit_distance(input_text,
14698 posix_name_lengths[i],
14702 { /* If it is close, it probably was intended to be a class */
14703 goto probably_meant_to_be;
14707 /* Here the input name is not close enough to a valid class name
14708 * for us to consider it to be intended to be a posix class. If
14709 * we haven't already done so, and the parse found a character that
14710 * could have been terminators for the name, but which we absorbed
14711 * as typos during the first pass, repeat the parse, signalling it
14712 * to stop at that character */
14713 if (possible_end && possible_end != (char *) -1) {
14714 possible_end = (char *) -1;
14719 /* Here neither pass found a close-enough class name */
14720 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14723 probably_meant_to_be:
14725 /* Here we think that a posix specification was intended. Update any
14727 if (updated_parse_ptr) {
14728 *updated_parse_ptr = (char *) p;
14731 /* If a posix class name was intended but incorrectly specified, we
14732 * output or return the warnings */
14733 if (found_problem) {
14735 /* We set flags for these issues in the parse loop above instead of
14736 * adding them to the list of warnings, because we can parse it
14737 * twice, and we only want one warning instance */
14739 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
14742 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14744 if (has_semi_colon) {
14745 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14747 else if (! has_terminating_colon) {
14748 ADD_POSIX_WARNING(p, "there is no terminating ':'");
14750 if (! has_terminating_bracket) {
14751 ADD_POSIX_WARNING(p, "there is no terminating ']'");
14754 if (posix_warnings && RExC_warn_text && av_top_index(RExC_warn_text) > -1) {
14755 *posix_warnings = RExC_warn_text;
14758 else if (class_number != OOB_NAMEDCLASS) {
14759 /* If it is a known class, return the class. The class number
14760 * #defines are structured so each complement is +1 to the normal
14762 return class_number + complement;
14764 else if (! check_only) {
14766 /* Here, it is an unrecognized class. This is an error (unless the
14767 * call is to check only, which we've already handled above) */
14768 const char * const complement_string = (complement)
14771 RExC_parse = (char *) p;
14772 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
14774 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
14778 return OOB_NAMEDCLASS;
14780 #undef ADD_POSIX_WARNING
14782 STATIC unsigned int
14783 S_regex_set_precedence(const U8 my_operator) {
14785 /* Returns the precedence in the (?[...]) construct of the input operator,
14786 * specified by its character representation. The precedence follows
14787 * general Perl rules, but it extends this so that ')' and ']' have (low)
14788 * precedence even though they aren't really operators */
14790 switch (my_operator) {
14806 NOT_REACHED; /* NOTREACHED */
14807 return 0; /* Silence compiler warning */
14811 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
14812 I32 *flagp, U32 depth,
14813 char * const oregcomp_parse)
14815 /* Handle the (?[...]) construct to do set operations */
14817 U8 curchar; /* Current character being parsed */
14818 UV start, end; /* End points of code point ranges */
14819 SV* final = NULL; /* The end result inversion list */
14820 SV* result_string; /* 'final' stringified */
14821 AV* stack; /* stack of operators and operands not yet
14823 AV* fence_stack = NULL; /* A stack containing the positions in
14824 'stack' of where the undealt-with left
14825 parens would be if they were actually
14827 /* The 'VOL' (expanding to 'volatile') is a workaround for an optimiser bug
14828 * in Solaris Studio 12.3. See RT #127455 */
14829 VOL IV fence = 0; /* Position of where most recent undealt-
14830 with left paren in stack is; -1 if none.
14832 STRLEN len; /* Temporary */
14833 regnode* node; /* Temporary, and final regnode returned by
14835 const bool save_fold = FOLD; /* Temporary */
14836 char *save_end, *save_parse; /* Temporaries */
14837 const bool in_locale = LOC; /* we turn off /l during processing */
14838 AV* posix_warnings = NULL;
14840 GET_RE_DEBUG_FLAGS_DECL;
14842 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
14845 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
14848 REQUIRE_UNI_RULES(flagp, NULL); /* The use of this operator implies /u.
14849 This is required so that the compile
14850 time values are valid in all runtime
14853 /* This will return only an ANYOF regnode, or (unlikely) something smaller
14854 * (such as EXACT). Thus we can skip most everything if just sizing. We
14855 * call regclass to handle '[]' so as to not have to reinvent its parsing
14856 * rules here (throwing away the size it computes each time). And, we exit
14857 * upon an unescaped ']' that isn't one ending a regclass. To do both
14858 * these things, we need to realize that something preceded by a backslash
14859 * is escaped, so we have to keep track of backslashes */
14861 UV depth = 0; /* how many nested (?[...]) constructs */
14863 while (RExC_parse < RExC_end) {
14864 SV* current = NULL;
14866 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14867 TRUE /* Force /x */ );
14869 switch (*RExC_parse) {
14871 if (RExC_parse[1] == '[') depth++, RExC_parse++;
14876 /* Skip past this, so the next character gets skipped, after
14879 if (*RExC_parse == 'c') {
14880 /* Skip the \cX notation for control characters */
14881 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14887 /* See if this is a [:posix:] class. */
14888 bool is_posix_class = (OOB_NAMEDCLASS
14889 < handle_possible_posix(pRExC_state,
14893 TRUE /* checking only */));
14894 /* If it is a posix class, leave the parse pointer at the
14895 * '[' to fool regclass() into thinking it is part of a
14896 * '[[:posix:]]'. */
14897 if (! is_posix_class) {
14901 /* regclass() can only return RESTART_PASS1 and NEED_UTF8
14902 * if multi-char folds are allowed. */
14903 if (!regclass(pRExC_state, flagp,depth+1,
14904 is_posix_class, /* parse the whole char
14905 class only if not a
14907 FALSE, /* don't allow multi-char folds */
14908 TRUE, /* silence non-portable warnings. */
14910 FALSE, /* Require return to be an ANYOF */
14914 FAIL2("panic: regclass returned NULL to handle_sets, "
14915 "flags=%#" UVxf, (UV) *flagp);
14917 /* function call leaves parse pointing to the ']', except
14918 * if we faked it */
14919 if (is_posix_class) {
14923 SvREFCNT_dec(current); /* In case it returned something */
14928 if (depth--) break;
14930 if (*RExC_parse == ')') {
14931 node = reganode(pRExC_state, ANYOF, 0);
14932 RExC_size += ANYOF_SKIP;
14933 nextchar(pRExC_state);
14934 Set_Node_Length(node,
14935 RExC_parse - oregcomp_parse + 1); /* MJD */
14937 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
14945 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14949 /* We output the messages even if warnings are off, because we'll fail
14950 * the very next thing, and these give a likely diagnosis for that */
14951 if (posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
14952 output_or_return_posix_warnings(pRExC_state, posix_warnings, NULL);
14955 FAIL("Syntax error in (?[...])");
14958 /* Pass 2 only after this. */
14959 Perl_ck_warner_d(aTHX_
14960 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
14961 "The regex_sets feature is experimental" REPORT_LOCATION,
14962 REPORT_LOCATION_ARGS(RExC_parse));
14964 /* Everything in this construct is a metacharacter. Operands begin with
14965 * either a '\' (for an escape sequence), or a '[' for a bracketed
14966 * character class. Any other character should be an operator, or
14967 * parenthesis for grouping. Both types of operands are handled by calling
14968 * regclass() to parse them. It is called with a parameter to indicate to
14969 * return the computed inversion list. The parsing here is implemented via
14970 * a stack. Each entry on the stack is a single character representing one
14971 * of the operators; or else a pointer to an operand inversion list. */
14973 #define IS_OPERATOR(a) SvIOK(a)
14974 #define IS_OPERAND(a) (! IS_OPERATOR(a))
14976 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
14977 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
14978 * with pronouncing it called it Reverse Polish instead, but now that YOU
14979 * know how to pronounce it you can use the correct term, thus giving due
14980 * credit to the person who invented it, and impressing your geek friends.
14981 * Wikipedia says that the pronounciation of "Ł" has been changing so that
14982 * it is now more like an English initial W (as in wonk) than an L.)
14984 * This means that, for example, 'a | b & c' is stored on the stack as
14992 * where the numbers in brackets give the stack [array] element number.
14993 * In this implementation, parentheses are not stored on the stack.
14994 * Instead a '(' creates a "fence" so that the part of the stack below the
14995 * fence is invisible except to the corresponding ')' (this allows us to
14996 * replace testing for parens, by using instead subtraction of the fence
14997 * position). As new operands are processed they are pushed onto the stack
14998 * (except as noted in the next paragraph). New operators of higher
14999 * precedence than the current final one are inserted on the stack before
15000 * the lhs operand (so that when the rhs is pushed next, everything will be
15001 * in the correct positions shown above. When an operator of equal or
15002 * lower precedence is encountered in parsing, all the stacked operations
15003 * of equal or higher precedence are evaluated, leaving the result as the
15004 * top entry on the stack. This makes higher precedence operations
15005 * evaluate before lower precedence ones, and causes operations of equal
15006 * precedence to left associate.
15008 * The only unary operator '!' is immediately pushed onto the stack when
15009 * encountered. When an operand is encountered, if the top of the stack is
15010 * a '!", the complement is immediately performed, and the '!' popped. The
15011 * resulting value is treated as a new operand, and the logic in the
15012 * previous paragraph is executed. Thus in the expression
15014 * the stack looks like
15020 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
15027 * A ')' is treated as an operator with lower precedence than all the
15028 * aforementioned ones, which causes all operations on the stack above the
15029 * corresponding '(' to be evaluated down to a single resultant operand.
15030 * Then the fence for the '(' is removed, and the operand goes through the
15031 * algorithm above, without the fence.
15033 * A separate stack is kept of the fence positions, so that the position of
15034 * the latest so-far unbalanced '(' is at the top of it.
15036 * The ']' ending the construct is treated as the lowest operator of all,
15037 * so that everything gets evaluated down to a single operand, which is the
15040 sv_2mortal((SV *)(stack = newAV()));
15041 sv_2mortal((SV *)(fence_stack = newAV()));
15043 while (RExC_parse < RExC_end) {
15044 I32 top_index; /* Index of top-most element in 'stack' */
15045 SV** top_ptr; /* Pointer to top 'stack' element */
15046 SV* current = NULL; /* To contain the current inversion list
15048 SV* only_to_avoid_leaks;
15050 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15051 TRUE /* Force /x */ );
15052 if (RExC_parse >= RExC_end) {
15053 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
15056 curchar = UCHARAT(RExC_parse);
15060 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15061 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
15062 DEBUG_U(dump_regex_sets_structures(pRExC_state,
15063 stack, fence, fence_stack));
15066 top_index = av_tindex_skip_len_mg(stack);
15069 SV** stacked_ptr; /* Ptr to something already on 'stack' */
15070 char stacked_operator; /* The topmost operator on the 'stack'. */
15071 SV* lhs; /* Operand to the left of the operator */
15072 SV* rhs; /* Operand to the right of the operator */
15073 SV* fence_ptr; /* Pointer to top element of the fence
15078 if ( RExC_parse < RExC_end - 1
15079 && (UCHARAT(RExC_parse + 1) == '?'))
15081 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
15082 * This happens when we have some thing like
15084 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
15086 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
15088 * Here we would be handling the interpolated
15089 * '$thai_or_lao'. We handle this by a recursive call to
15090 * ourselves which returns the inversion list the
15091 * interpolated expression evaluates to. We use the flags
15092 * from the interpolated pattern. */
15093 U32 save_flags = RExC_flags;
15094 const char * save_parse;
15096 RExC_parse += 2; /* Skip past the '(?' */
15097 save_parse = RExC_parse;
15099 /* Parse any flags for the '(?' */
15100 parse_lparen_question_flags(pRExC_state);
15102 if (RExC_parse == save_parse /* Makes sure there was at
15103 least one flag (or else
15104 this embedding wasn't
15106 || RExC_parse >= RExC_end - 4
15107 || UCHARAT(RExC_parse) != ':'
15108 || UCHARAT(++RExC_parse) != '('
15109 || UCHARAT(++RExC_parse) != '?'
15110 || UCHARAT(++RExC_parse) != '[')
15113 /* In combination with the above, this moves the
15114 * pointer to the point just after the first erroneous
15115 * character (or if there are no flags, to where they
15116 * should have been) */
15117 if (RExC_parse >= RExC_end - 4) {
15118 RExC_parse = RExC_end;
15120 else if (RExC_parse != save_parse) {
15121 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15123 vFAIL("Expecting '(?flags:(?[...'");
15126 /* Recurse, with the meat of the embedded expression */
15128 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
15129 depth+1, oregcomp_parse);
15131 /* Here, 'current' contains the embedded expression's
15132 * inversion list, and RExC_parse points to the trailing
15133 * ']'; the next character should be the ')' */
15135 assert(UCHARAT(RExC_parse) == ')');
15137 /* Then the ')' matching the original '(' handled by this
15138 * case: statement */
15140 assert(UCHARAT(RExC_parse) == ')');
15143 RExC_flags = save_flags;
15144 goto handle_operand;
15147 /* A regular '('. Look behind for illegal syntax */
15148 if (top_index - fence >= 0) {
15149 /* If the top entry on the stack is an operator, it had
15150 * better be a '!', otherwise the entry below the top
15151 * operand should be an operator */
15152 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
15153 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
15154 || ( IS_OPERAND(*top_ptr)
15155 && ( top_index - fence < 1
15156 || ! (stacked_ptr = av_fetch(stack,
15159 || ! IS_OPERATOR(*stacked_ptr))))
15162 vFAIL("Unexpected '(' with no preceding operator");
15166 /* Stack the position of this undealt-with left paren */
15167 av_push(fence_stack, newSViv(fence));
15168 fence = top_index + 1;
15172 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15173 * multi-char folds are allowed. */
15174 if (!regclass(pRExC_state, flagp,depth+1,
15175 TRUE, /* means parse just the next thing */
15176 FALSE, /* don't allow multi-char folds */
15177 FALSE, /* don't silence non-portable warnings. */
15179 FALSE, /* Require return to be an ANYOF */
15183 FAIL2("panic: regclass returned NULL to handle_sets, "
15184 "flags=%#" UVxf, (UV) *flagp);
15187 /* regclass() will return with parsing just the \ sequence,
15188 * leaving the parse pointer at the next thing to parse */
15190 goto handle_operand;
15192 case '[': /* Is a bracketed character class */
15194 /* See if this is a [:posix:] class. */
15195 bool is_posix_class = (OOB_NAMEDCLASS
15196 < handle_possible_posix(pRExC_state,
15200 TRUE /* checking only */));
15201 /* If it is a posix class, leave the parse pointer at the '['
15202 * to fool regclass() into thinking it is part of a
15203 * '[[:posix:]]'. */
15204 if (! is_posix_class) {
15208 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15209 * multi-char folds are allowed. */
15210 if (!regclass(pRExC_state, flagp,depth+1,
15211 is_posix_class, /* parse the whole char
15212 class only if not a
15214 FALSE, /* don't allow multi-char folds */
15215 TRUE, /* silence non-portable warnings. */
15217 FALSE, /* Require return to be an ANYOF */
15222 FAIL2("panic: regclass returned NULL to handle_sets, "
15223 "flags=%#" UVxf, (UV) *flagp);
15226 /* function call leaves parse pointing to the ']', except if we
15228 if (is_posix_class) {
15232 goto handle_operand;
15236 if (top_index >= 1) {
15237 goto join_operators;
15240 /* Only a single operand on the stack: are done */
15244 if (av_tindex_skip_len_mg(fence_stack) < 0) {
15246 vFAIL("Unexpected ')'");
15249 /* If nothing after the fence, is missing an operand */
15250 if (top_index - fence < 0) {
15254 /* If at least two things on the stack, treat this as an
15256 if (top_index - fence >= 1) {
15257 goto join_operators;
15260 /* Here only a single thing on the fenced stack, and there is a
15261 * fence. Get rid of it */
15262 fence_ptr = av_pop(fence_stack);
15264 fence = SvIV(fence_ptr) - 1;
15265 SvREFCNT_dec_NN(fence_ptr);
15272 /* Having gotten rid of the fence, we pop the operand at the
15273 * stack top and process it as a newly encountered operand */
15274 current = av_pop(stack);
15275 if (IS_OPERAND(current)) {
15276 goto handle_operand;
15288 /* These binary operators should have a left operand already
15290 if ( top_index - fence < 0
15291 || top_index - fence == 1
15292 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
15293 || ! IS_OPERAND(*top_ptr))
15295 goto unexpected_binary;
15298 /* If only the one operand is on the part of the stack visible
15299 * to us, we just place this operator in the proper position */
15300 if (top_index - fence < 2) {
15302 /* Place the operator before the operand */
15304 SV* lhs = av_pop(stack);
15305 av_push(stack, newSVuv(curchar));
15306 av_push(stack, lhs);
15310 /* But if there is something else on the stack, we need to
15311 * process it before this new operator if and only if the
15312 * stacked operation has equal or higher precedence than the
15317 /* The operator on the stack is supposed to be below both its
15319 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
15320 || IS_OPERAND(*stacked_ptr))
15322 /* But if not, it's legal and indicates we are completely
15323 * done if and only if we're currently processing a ']',
15324 * which should be the final thing in the expression */
15325 if (curchar == ']') {
15331 vFAIL2("Unexpected binary operator '%c' with no "
15332 "preceding operand", curchar);
15334 stacked_operator = (char) SvUV(*stacked_ptr);
15336 if (regex_set_precedence(curchar)
15337 > regex_set_precedence(stacked_operator))
15339 /* Here, the new operator has higher precedence than the
15340 * stacked one. This means we need to add the new one to
15341 * the stack to await its rhs operand (and maybe more
15342 * stuff). We put it before the lhs operand, leaving
15343 * untouched the stacked operator and everything below it
15345 lhs = av_pop(stack);
15346 assert(IS_OPERAND(lhs));
15348 av_push(stack, newSVuv(curchar));
15349 av_push(stack, lhs);
15353 /* Here, the new operator has equal or lower precedence than
15354 * what's already there. This means the operation already
15355 * there should be performed now, before the new one. */
15357 rhs = av_pop(stack);
15358 if (! IS_OPERAND(rhs)) {
15360 /* This can happen when a ! is not followed by an operand,
15361 * like in /(?[\t &!])/ */
15365 lhs = av_pop(stack);
15367 if (! IS_OPERAND(lhs)) {
15369 /* This can happen when there is an empty (), like in
15370 * /(?[[0]+()+])/ */
15374 switch (stacked_operator) {
15376 _invlist_intersection(lhs, rhs, &rhs);
15381 _invlist_union(lhs, rhs, &rhs);
15385 _invlist_subtract(lhs, rhs, &rhs);
15388 case '^': /* The union minus the intersection */
15393 _invlist_union(lhs, rhs, &u);
15394 _invlist_intersection(lhs, rhs, &i);
15395 _invlist_subtract(u, i, &rhs);
15396 SvREFCNT_dec_NN(i);
15397 SvREFCNT_dec_NN(u);
15403 /* Here, the higher precedence operation has been done, and the
15404 * result is in 'rhs'. We overwrite the stacked operator with
15405 * the result. Then we redo this code to either push the new
15406 * operator onto the stack or perform any higher precedence
15407 * stacked operation */
15408 only_to_avoid_leaks = av_pop(stack);
15409 SvREFCNT_dec(only_to_avoid_leaks);
15410 av_push(stack, rhs);
15413 case '!': /* Highest priority, right associative */
15415 /* If what's already at the top of the stack is another '!",
15416 * they just cancel each other out */
15417 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
15418 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
15420 only_to_avoid_leaks = av_pop(stack);
15421 SvREFCNT_dec(only_to_avoid_leaks);
15423 else { /* Otherwise, since it's right associative, just push
15425 av_push(stack, newSVuv(curchar));
15430 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15431 vFAIL("Unexpected character");
15435 /* Here 'current' is the operand. If something is already on the
15436 * stack, we have to check if it is a !. But first, the code above
15437 * may have altered the stack in the time since we earlier set
15440 top_index = av_tindex_skip_len_mg(stack);
15441 if (top_index - fence >= 0) {
15442 /* If the top entry on the stack is an operator, it had better
15443 * be a '!', otherwise the entry below the top operand should
15444 * be an operator */
15445 top_ptr = av_fetch(stack, top_index, FALSE);
15447 if (IS_OPERATOR(*top_ptr)) {
15449 /* The only permissible operator at the top of the stack is
15450 * '!', which is applied immediately to this operand. */
15451 curchar = (char) SvUV(*top_ptr);
15452 if (curchar != '!') {
15453 SvREFCNT_dec(current);
15454 vFAIL2("Unexpected binary operator '%c' with no "
15455 "preceding operand", curchar);
15458 _invlist_invert(current);
15460 only_to_avoid_leaks = av_pop(stack);
15461 SvREFCNT_dec(only_to_avoid_leaks);
15463 /* And we redo with the inverted operand. This allows
15464 * handling multiple ! in a row */
15465 goto handle_operand;
15467 /* Single operand is ok only for the non-binary ')'
15469 else if ((top_index - fence == 0 && curchar != ')')
15470 || (top_index - fence > 0
15471 && (! (stacked_ptr = av_fetch(stack,
15474 || IS_OPERAND(*stacked_ptr))))
15476 SvREFCNT_dec(current);
15477 vFAIL("Operand with no preceding operator");
15481 /* Here there was nothing on the stack or the top element was
15482 * another operand. Just add this new one */
15483 av_push(stack, current);
15485 } /* End of switch on next parse token */
15487 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15488 } /* End of loop parsing through the construct */
15491 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
15492 vFAIL("Unmatched (");
15495 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
15496 || ((final = av_pop(stack)) == NULL)
15497 || ! IS_OPERAND(final)
15498 || SvTYPE(final) != SVt_INVLIST
15499 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
15502 SvREFCNT_dec(final);
15503 vFAIL("Incomplete expression within '(?[ ])'");
15506 /* Here, 'final' is the resultant inversion list from evaluating the
15507 * expression. Return it if so requested */
15508 if (return_invlist) {
15509 *return_invlist = final;
15513 /* Otherwise generate a resultant node, based on 'final'. regclass() is
15514 * expecting a string of ranges and individual code points */
15515 invlist_iterinit(final);
15516 result_string = newSVpvs("");
15517 while (invlist_iternext(final, &start, &end)) {
15518 if (start == end) {
15519 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
15522 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%" UVXf "}",
15527 /* About to generate an ANYOF (or similar) node from the inversion list we
15528 * have calculated */
15529 save_parse = RExC_parse;
15530 RExC_parse = SvPV(result_string, len);
15531 save_end = RExC_end;
15532 RExC_end = RExC_parse + len;
15534 /* We turn off folding around the call, as the class we have constructed
15535 * already has all folding taken into consideration, and we don't want
15536 * regclass() to add to that */
15537 RExC_flags &= ~RXf_PMf_FOLD;
15538 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if multi-char
15539 * folds are allowed. */
15540 node = regclass(pRExC_state, flagp,depth+1,
15541 FALSE, /* means parse the whole char class */
15542 FALSE, /* don't allow multi-char folds */
15543 TRUE, /* silence non-portable warnings. The above may very
15544 well have generated non-portable code points, but
15545 they're valid on this machine */
15546 FALSE, /* similarly, no need for strict */
15547 FALSE, /* Require return to be an ANYOF */
15552 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#" UVxf,
15555 /* Fix up the node type if we are in locale. (We have pretended we are
15556 * under /u for the purposes of regclass(), as this construct will only
15557 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
15558 * as to cause any warnings about bad locales to be output in regexec.c),
15559 * and add the flag that indicates to check if not in a UTF-8 locale. The
15560 * reason we above forbid optimization into something other than an ANYOF
15561 * node is simply to minimize the number of code changes in regexec.c.
15562 * Otherwise we would have to create new EXACTish node types and deal with
15563 * them. This decision could be revisited should this construct become
15566 * (One might think we could look at the resulting ANYOF node and suppress
15567 * the flag if everything is above 255, as those would be UTF-8 only,
15568 * but this isn't true, as the components that led to that result could
15569 * have been locale-affected, and just happen to cancel each other out
15570 * under UTF-8 locales.) */
15572 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
15574 assert(OP(node) == ANYOF);
15578 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
15582 RExC_flags |= RXf_PMf_FOLD;
15585 RExC_parse = save_parse + 1;
15586 RExC_end = save_end;
15587 SvREFCNT_dec_NN(final);
15588 SvREFCNT_dec_NN(result_string);
15590 nextchar(pRExC_state);
15591 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
15595 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15598 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
15599 AV * stack, const IV fence, AV * fence_stack)
15600 { /* Dumps the stacks in handle_regex_sets() */
15602 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
15603 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
15606 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
15608 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
15610 if (stack_top < 0) {
15611 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
15614 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
15615 for (i = stack_top; i >= 0; i--) {
15616 SV ** element_ptr = av_fetch(stack, i, FALSE);
15617 if (! element_ptr) {
15620 if (IS_OPERATOR(*element_ptr)) {
15621 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
15622 (int) i, (int) SvIV(*element_ptr));
15625 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
15626 sv_dump(*element_ptr);
15631 if (fence_stack_top < 0) {
15632 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
15635 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
15636 for (i = fence_stack_top; i >= 0; i--) {
15637 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
15638 if (! element_ptr) {
15641 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
15642 (int) i, (int) SvIV(*element_ptr));
15653 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
15655 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
15656 * innocent-looking character class, like /[ks]/i won't have to go out to
15657 * disk to find the possible matches.
15659 * This should be called only for a Latin1-range code points, cp, which is
15660 * known to be involved in a simple fold with other code points above
15661 * Latin1. It would give false results if /aa has been specified.
15662 * Multi-char folds are outside the scope of this, and must be handled
15665 * XXX It would be better to generate these via regen, in case a new
15666 * version of the Unicode standard adds new mappings, though that is not
15667 * really likely, and may be caught by the default: case of the switch
15670 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
15672 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
15678 add_cp_to_invlist(*invlist, KELVIN_SIGN);
15682 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
15685 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
15686 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
15688 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
15689 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
15690 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
15692 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
15693 *invlist = add_cp_to_invlist(*invlist,
15694 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
15697 #ifdef LATIN_CAPITAL_LETTER_SHARP_S /* not defined in early Unicode releases */
15699 case LATIN_SMALL_LETTER_SHARP_S:
15700 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
15705 #if UNICODE_MAJOR_VERSION < 3 \
15706 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0)
15708 /* In 3.0 and earlier, U+0130 folded simply to 'i'; and in 3.0.1 so did
15713 add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
15714 # if UNICODE_DOT_DOT_VERSION == 1
15715 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_DOTLESS_I);
15721 /* Use deprecated warning to increase the chances of this being
15724 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
15731 S_output_or_return_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings, AV** return_posix_warnings)
15733 /* If the final parameter is NULL, output the elements of the array given
15734 * by '*posix_warnings' as REGEXP warnings. Otherwise, the elements are
15735 * pushed onto it, (creating if necessary) */
15738 const bool first_is_fatal = ! return_posix_warnings
15739 && ckDEAD(packWARN(WARN_REGEXP));
15741 PERL_ARGS_ASSERT_OUTPUT_OR_RETURN_POSIX_WARNINGS;
15743 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
15744 if (return_posix_warnings) {
15745 if (! *return_posix_warnings) { /* mortalize to not leak if
15746 warnings are fatal */
15747 *return_posix_warnings = (AV *) sv_2mortal((SV *) newAV());
15749 av_push(*return_posix_warnings, msg);
15752 if (first_is_fatal) { /* Avoid leaking this */
15753 av_undef(posix_warnings); /* This isn't necessary if the
15754 array is mortal, but is a
15756 (void) sv_2mortal(msg);
15758 SAVEFREESV(RExC_rx_sv);
15761 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
15762 SvREFCNT_dec_NN(msg);
15768 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
15770 /* This adds the string scalar <multi_string> to the array
15771 * <multi_char_matches>. <multi_string> is known to have exactly
15772 * <cp_count> code points in it. This is used when constructing a
15773 * bracketed character class and we find something that needs to match more
15774 * than a single character.
15776 * <multi_char_matches> is actually an array of arrays. Each top-level
15777 * element is an array that contains all the strings known so far that are
15778 * the same length. And that length (in number of code points) is the same
15779 * as the index of the top-level array. Hence, the [2] element is an
15780 * array, each element thereof is a string containing TWO code points;
15781 * while element [3] is for strings of THREE characters, and so on. Since
15782 * this is for multi-char strings there can never be a [0] nor [1] element.
15784 * When we rewrite the character class below, we will do so such that the
15785 * longest strings are written first, so that it prefers the longest
15786 * matching strings first. This is done even if it turns out that any
15787 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
15788 * Christiansen has agreed that this is ok. This makes the test for the
15789 * ligature 'ffi' come before the test for 'ff', for example */
15792 AV** this_array_ptr;
15794 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
15796 if (! multi_char_matches) {
15797 multi_char_matches = newAV();
15800 if (av_exists(multi_char_matches, cp_count)) {
15801 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
15802 this_array = *this_array_ptr;
15805 this_array = newAV();
15806 av_store(multi_char_matches, cp_count,
15809 av_push(this_array, multi_string);
15811 return multi_char_matches;
15814 /* The names of properties whose definitions are not known at compile time are
15815 * stored in this SV, after a constant heading. So if the length has been
15816 * changed since initialization, then there is a run-time definition. */
15817 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
15818 (SvCUR(listsv) != initial_listsv_len)
15820 /* There is a restricted set of white space characters that are legal when
15821 * ignoring white space in a bracketed character class. This generates the
15822 * code to skip them.
15824 * There is a line below that uses the same white space criteria but is outside
15825 * this macro. Both here and there must use the same definition */
15826 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
15829 while (isBLANK_A(UCHARAT(p))) \
15837 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
15838 const bool stop_at_1, /* Just parse the next thing, don't
15839 look for a full character class */
15840 bool allow_multi_folds,
15841 const bool silence_non_portable, /* Don't output warnings
15845 bool optimizable, /* ? Allow a non-ANYOF return
15847 SV** ret_invlist, /* Return an inversion list, not a node */
15848 AV** return_posix_warnings
15851 /* parse a bracketed class specification. Most of these will produce an
15852 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
15853 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
15854 * under /i with multi-character folds: it will be rewritten following the
15855 * paradigm of this example, where the <multi-fold>s are characters which
15856 * fold to multiple character sequences:
15857 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
15858 * gets effectively rewritten as:
15859 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
15860 * reg() gets called (recursively) on the rewritten version, and this
15861 * function will return what it constructs. (Actually the <multi-fold>s
15862 * aren't physically removed from the [abcdefghi], it's just that they are
15863 * ignored in the recursion by means of a flag:
15864 * <RExC_in_multi_char_class>.)
15866 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
15867 * characters, with the corresponding bit set if that character is in the
15868 * list. For characters above this, a range list or swash is used. There
15869 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
15870 * determinable at compile time
15872 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs
15873 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded
15874 * to UTF-8. This can only happen if ret_invlist is non-NULL.
15877 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
15879 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
15882 int namedclass = OOB_NAMEDCLASS;
15883 char *rangebegin = NULL;
15884 bool need_class = 0;
15886 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
15887 than just initialized. */
15888 SV* properties = NULL; /* Code points that match \p{} \P{} */
15889 SV* posixes = NULL; /* Code points that match classes like [:word:],
15890 extended beyond the Latin1 range. These have to
15891 be kept separate from other code points for much
15892 of this function because their handling is
15893 different under /i, and for most classes under
15895 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
15896 separate for a while from the non-complemented
15897 versions because of complications with /d
15899 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
15900 treated more simply than the general case,
15901 leading to less compilation and execution
15903 UV element_count = 0; /* Number of distinct elements in the class.
15904 Optimizations may be possible if this is tiny */
15905 AV * multi_char_matches = NULL; /* Code points that fold to more than one
15906 character; used under /i */
15908 char * stop_ptr = RExC_end; /* where to stop parsing */
15910 /* ignore unescaped whitespace? */
15911 const bool skip_white = cBOOL( ret_invlist
15912 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
15914 /* Unicode properties are stored in a swash; this holds the current one
15915 * being parsed. If this swash is the only above-latin1 component of the
15916 * character class, an optimization is to pass it directly on to the
15917 * execution engine. Otherwise, it is set to NULL to indicate that there
15918 * are other things in the class that have to be dealt with at execution
15920 SV* swash = NULL; /* Code points that match \p{} \P{} */
15922 /* Set if a component of this character class is user-defined; just passed
15923 * on to the engine */
15924 bool has_user_defined_property = FALSE;
15926 /* inversion list of code points this node matches only when the target
15927 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
15929 SV* has_upper_latin1_only_utf8_matches = NULL;
15931 /* Inversion list of code points this node matches regardless of things
15932 * like locale, folding, utf8ness of the target string */
15933 SV* cp_list = NULL;
15935 /* Like cp_list, but code points on this list need to be checked for things
15936 * that fold to/from them under /i */
15937 SV* cp_foldable_list = NULL;
15939 /* Like cp_list, but code points on this list are valid only when the
15940 * runtime locale is UTF-8 */
15941 SV* only_utf8_locale_list = NULL;
15943 /* In a range, if one of the endpoints is non-character-set portable,
15944 * meaning that it hard-codes a code point that may mean a different
15945 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
15946 * mnemonic '\t' which each mean the same character no matter which
15947 * character set the platform is on. */
15948 unsigned int non_portable_endpoint = 0;
15950 /* Is the range unicode? which means on a platform that isn't 1-1 native
15951 * to Unicode (i.e. non-ASCII), each code point in it should be considered
15952 * to be a Unicode value. */
15953 bool unicode_range = FALSE;
15954 bool invert = FALSE; /* Is this class to be complemented */
15956 bool warn_super = ALWAYS_WARN_SUPER;
15958 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
15959 case we need to change the emitted regop to an EXACT. */
15960 const char * orig_parse = RExC_parse;
15961 const SSize_t orig_size = RExC_size;
15962 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
15964 /* This variable is used to mark where the end in the input is of something
15965 * that looks like a POSIX construct but isn't. During the parse, when
15966 * something looks like it could be such a construct is encountered, it is
15967 * checked for being one, but not if we've already checked this area of the
15968 * input. Only after this position is reached do we check again */
15969 char *not_posix_region_end = RExC_parse - 1;
15971 AV* posix_warnings = NULL;
15972 const bool do_posix_warnings = return_posix_warnings
15973 || (PASS2 && ckWARN(WARN_REGEXP));
15975 GET_RE_DEBUG_FLAGS_DECL;
15977 PERL_ARGS_ASSERT_REGCLASS;
15979 PERL_UNUSED_ARG(depth);
15982 DEBUG_PARSE("clas");
15984 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
15985 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
15986 && UNICODE_DOT_DOT_VERSION == 0)
15987 allow_multi_folds = FALSE;
15990 /* Assume we are going to generate an ANYOF node. */
15991 ret = reganode(pRExC_state,
15998 RExC_size += ANYOF_SKIP;
15999 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
16002 ANYOF_FLAGS(ret) = 0;
16004 RExC_emit += ANYOF_SKIP;
16005 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
16006 initial_listsv_len = SvCUR(listsv);
16007 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
16010 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16012 assert(RExC_parse <= RExC_end);
16014 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
16017 allow_multi_folds = FALSE;
16019 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16022 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
16023 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
16024 int maybe_class = handle_possible_posix(pRExC_state,
16026 ¬_posix_region_end,
16028 TRUE /* checking only */);
16029 if (PASS2 && maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
16030 SAVEFREESV(RExC_rx_sv);
16031 ckWARN4reg(not_posix_region_end,
16032 "POSIX syntax [%c %c] belongs inside character classes%s",
16033 *RExC_parse, *RExC_parse,
16034 (maybe_class == OOB_NAMEDCLASS)
16035 ? ((POSIXCC_NOTYET(*RExC_parse))
16036 ? " (but this one isn't implemented)"
16037 : " (but this one isn't fully valid)")
16040 (void)ReREFCNT_inc(RExC_rx_sv);
16044 /* If the caller wants us to just parse a single element, accomplish this
16045 * by faking the loop ending condition */
16046 if (stop_at_1 && RExC_end > RExC_parse) {
16047 stop_ptr = RExC_parse + 1;
16050 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
16051 if (UCHARAT(RExC_parse) == ']')
16052 goto charclassloop;
16056 if ( posix_warnings
16057 && av_tindex_skip_len_mg(posix_warnings) >= 0
16058 && RExC_parse > not_posix_region_end)
16060 /* Warnings about posix class issues are considered tentative until
16061 * we are far enough along in the parse that we can no longer
16062 * change our mind, at which point we either output them or add
16063 * them, if it has so specified, to what gets returned to the
16064 * caller. This is done each time through the loop so that a later
16065 * class won't zap them before they have been dealt with. */
16066 output_or_return_posix_warnings(pRExC_state, posix_warnings,
16067 return_posix_warnings);
16070 if (RExC_parse >= stop_ptr) {
16074 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16076 if (UCHARAT(RExC_parse) == ']') {
16082 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
16083 save_value = value;
16084 save_prevvalue = prevvalue;
16087 rangebegin = RExC_parse;
16089 non_portable_endpoint = 0;
16091 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
16092 value = utf8n_to_uvchr((U8*)RExC_parse,
16093 RExC_end - RExC_parse,
16094 &numlen, UTF8_ALLOW_DEFAULT);
16095 RExC_parse += numlen;
16098 value = UCHARAT(RExC_parse++);
16100 if (value == '[') {
16101 char * posix_class_end;
16102 namedclass = handle_possible_posix(pRExC_state,
16105 do_posix_warnings ? &posix_warnings : NULL,
16106 FALSE /* die if error */);
16107 if (namedclass > OOB_NAMEDCLASS) {
16109 /* If there was an earlier attempt to parse this particular
16110 * posix class, and it failed, it was a false alarm, as this
16111 * successful one proves */
16112 if ( posix_warnings
16113 && av_tindex_skip_len_mg(posix_warnings) >= 0
16114 && not_posix_region_end >= RExC_parse
16115 && not_posix_region_end <= posix_class_end)
16117 av_undef(posix_warnings);
16120 RExC_parse = posix_class_end;
16122 else if (namedclass == OOB_NAMEDCLASS) {
16123 not_posix_region_end = posix_class_end;
16126 namedclass = OOB_NAMEDCLASS;
16129 else if ( RExC_parse - 1 > not_posix_region_end
16130 && MAYBE_POSIXCC(value))
16132 (void) handle_possible_posix(
16134 RExC_parse - 1, /* -1 because parse has already been
16136 ¬_posix_region_end,
16137 do_posix_warnings ? &posix_warnings : NULL,
16138 TRUE /* checking only */);
16140 else if (value == '\\') {
16141 /* Is a backslash; get the code point of the char after it */
16143 if (RExC_parse >= RExC_end) {
16144 vFAIL("Unmatched [");
16147 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
16148 value = utf8n_to_uvchr((U8*)RExC_parse,
16149 RExC_end - RExC_parse,
16150 &numlen, UTF8_ALLOW_DEFAULT);
16151 RExC_parse += numlen;
16154 value = UCHARAT(RExC_parse++);
16156 /* Some compilers cannot handle switching on 64-bit integer
16157 * values, therefore value cannot be an UV. Yes, this will
16158 * be a problem later if we want switch on Unicode.
16159 * A similar issue a little bit later when switching on
16160 * namedclass. --jhi */
16162 /* If the \ is escaping white space when white space is being
16163 * skipped, it means that that white space is wanted literally, and
16164 * is already in 'value'. Otherwise, need to translate the escape
16165 * into what it signifies. */
16166 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
16168 case 'w': namedclass = ANYOF_WORDCHAR; break;
16169 case 'W': namedclass = ANYOF_NWORDCHAR; break;
16170 case 's': namedclass = ANYOF_SPACE; break;
16171 case 'S': namedclass = ANYOF_NSPACE; break;
16172 case 'd': namedclass = ANYOF_DIGIT; break;
16173 case 'D': namedclass = ANYOF_NDIGIT; break;
16174 case 'v': namedclass = ANYOF_VERTWS; break;
16175 case 'V': namedclass = ANYOF_NVERTWS; break;
16176 case 'h': namedclass = ANYOF_HORIZWS; break;
16177 case 'H': namedclass = ANYOF_NHORIZWS; break;
16178 case 'N': /* Handle \N{NAME} in class */
16180 const char * const backslash_N_beg = RExC_parse - 2;
16183 if (! grok_bslash_N(pRExC_state,
16184 NULL, /* No regnode */
16185 &value, /* Yes single value */
16186 &cp_count, /* Multiple code pt count */
16192 if (*flagp & NEED_UTF8)
16193 FAIL("panic: grok_bslash_N set NEED_UTF8");
16194 if (*flagp & RESTART_PASS1)
16197 if (cp_count < 0) {
16198 vFAIL("\\N in a character class must be a named character: \\N{...}");
16200 else if (cp_count == 0) {
16202 ckWARNreg(RExC_parse,
16203 "Ignoring zero length \\N{} in character class");
16206 else { /* cp_count > 1 */
16207 if (! RExC_in_multi_char_class) {
16208 if (invert || range || *RExC_parse == '-') {
16211 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
16214 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
16216 break; /* <value> contains the first code
16217 point. Drop out of the switch to
16221 SV * multi_char_N = newSVpvn(backslash_N_beg,
16222 RExC_parse - backslash_N_beg);
16224 = add_multi_match(multi_char_matches,
16229 } /* End of cp_count != 1 */
16231 /* This element should not be processed further in this
16234 value = save_value;
16235 prevvalue = save_prevvalue;
16236 continue; /* Back to top of loop to get next char */
16239 /* Here, is a single code point, and <value> contains it */
16240 unicode_range = TRUE; /* \N{} are Unicode */
16248 /* We will handle any undefined properties ourselves */
16249 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
16250 /* And we actually would prefer to get
16251 * the straight inversion list of the
16252 * swash, since we will be accessing it
16253 * anyway, to save a little time */
16254 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
16256 if (RExC_parse >= RExC_end)
16257 vFAIL2("Empty \\%c", (U8)value);
16258 if (*RExC_parse == '{') {
16259 const U8 c = (U8)value;
16260 e = strchr(RExC_parse, '}');
16263 vFAIL2("Missing right brace on \\%c{}", c);
16267 while (isSPACE(*RExC_parse)) {
16271 if (UCHARAT(RExC_parse) == '^') {
16273 /* toggle. (The rhs xor gets the single bit that
16274 * differs between P and p; the other xor inverts just
16276 value ^= 'P' ^ 'p';
16279 while (isSPACE(*RExC_parse)) {
16284 if (e == RExC_parse)
16285 vFAIL2("Empty \\%c{}", c);
16287 n = e - RExC_parse;
16288 while (isSPACE(*(RExC_parse + n - 1)))
16290 } /* The \p isn't immediately followed by a '{' */
16291 else if (! isALPHA(*RExC_parse)) {
16292 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16293 vFAIL2("Character following \\%c must be '{' or a "
16294 "single-character Unicode property name",
16304 char* base_name; /* name after any packages are stripped */
16305 char* lookup_name = NULL;
16306 const char * const colon_colon = "::";
16308 /* Try to get the definition of the property into
16309 * <invlist>. If /i is in effect, the effective property
16310 * will have its name be <__NAME_i>. The design is
16311 * discussed in commit
16312 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
16313 name = savepv(Perl_form(aTHX_ "%.*s", (int)n, RExC_parse));
16316 lookup_name = savepv(Perl_form(aTHX_ "__%s_i", name));
16318 /* The function call just below that uses this can fail
16319 * to return, leaking memory if we don't do this */
16320 SAVEFREEPV(lookup_name);
16323 /* Look up the property name, and get its swash and
16324 * inversion list, if the property is found */
16325 SvREFCNT_dec(swash); /* Free any left-overs */
16326 swash = _core_swash_init("utf8",
16333 NULL, /* No inversion list */
16336 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
16337 HV* curpkg = (IN_PERL_COMPILETIME)
16339 : CopSTASH(PL_curcop);
16343 if (swash) { /* Got a swash but no inversion list.
16344 Something is likely wrong that will
16345 be sorted-out later */
16346 SvREFCNT_dec_NN(swash);
16350 /* Here didn't find it. It could be a an error (like a
16351 * typo) in specifying a Unicode property, or it could
16352 * be a user-defined property that will be available at
16353 * run-time. The names of these must begin with 'In'
16354 * or 'Is' (after any packages are stripped off). So
16355 * if not one of those, or if we accept only
16356 * compile-time properties, is an error; otherwise add
16357 * it to the list for run-time look up. */
16358 if ((base_name = rninstr(name, name + n,
16359 colon_colon, colon_colon + 2)))
16360 { /* Has ::. We know this must be a user-defined
16363 final_n -= base_name - name;
16372 || base_name[0] != 'I'
16373 || (base_name[1] != 's' && base_name[1] != 'n')
16376 const char * const msg
16378 ? "Illegal user-defined property name"
16379 : "Can't find Unicode property definition";
16380 RExC_parse = e + 1;
16382 /* diag_listed_as: Can't find Unicode property definition "%s" */
16383 vFAIL3utf8f("%s \"%" UTF8f "\"",
16384 msg, UTF8fARG(UTF, n, name));
16387 /* If the property name doesn't already have a package
16388 * name, add the current one to it so that it can be
16389 * referred to outside it. [perl #121777] */
16390 if (! has_pkg && curpkg) {
16391 char* pkgname = HvNAME(curpkg);
16392 if (strNE(pkgname, "main")) {
16393 char* full_name = Perl_form(aTHX_
16397 n = strlen(full_name);
16398 name = savepvn(full_name, n);
16402 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s%" UTF8f "%s\n",
16403 (value == 'p' ? '+' : '!'),
16404 (FOLD) ? "__" : "",
16405 UTF8fARG(UTF, n, name),
16406 (FOLD) ? "_i" : "");
16407 has_user_defined_property = TRUE;
16408 optimizable = FALSE; /* Will have to leave this an
16411 /* We don't know yet what this matches, so have to flag
16413 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
16417 /* Here, did get the swash and its inversion list. If
16418 * the swash is from a user-defined property, then this
16419 * whole character class should be regarded as such */
16420 if (swash_init_flags
16421 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
16423 has_user_defined_property = TRUE;
16426 /* We warn on matching an above-Unicode code point
16427 * if the match would return true, except don't
16428 * warn for \p{All}, which has exactly one element
16430 (_invlist_contains_cp(invlist, 0x110000)
16431 && (! (_invlist_len(invlist) == 1
16432 && *invlist_array(invlist) == 0)))
16438 /* Invert if asking for the complement */
16439 if (value == 'P') {
16440 _invlist_union_complement_2nd(properties,
16444 /* The swash can't be used as-is, because we've
16445 * inverted things; delay removing it to here after
16446 * have copied its invlist above */
16447 SvREFCNT_dec_NN(swash);
16451 _invlist_union(properties, invlist, &properties);
16455 RExC_parse = e + 1;
16456 namedclass = ANYOF_UNIPROP; /* no official name, but it's
16459 /* \p means they want Unicode semantics */
16460 REQUIRE_UNI_RULES(flagp, NULL);
16463 case 'n': value = '\n'; break;
16464 case 'r': value = '\r'; break;
16465 case 't': value = '\t'; break;
16466 case 'f': value = '\f'; break;
16467 case 'b': value = '\b'; break;
16468 case 'e': value = ESC_NATIVE; break;
16469 case 'a': value = '\a'; break;
16471 RExC_parse--; /* function expects to be pointed at the 'o' */
16473 const char* error_msg;
16474 bool valid = grok_bslash_o(&RExC_parse,
16477 PASS2, /* warnings only in
16480 silence_non_portable,
16486 non_portable_endpoint++;
16489 RExC_parse--; /* function expects to be pointed at the 'x' */
16491 const char* error_msg;
16492 bool valid = grok_bslash_x(&RExC_parse,
16495 PASS2, /* Output warnings */
16497 silence_non_portable,
16503 non_portable_endpoint++;
16506 value = grok_bslash_c(*RExC_parse++, PASS2);
16507 non_portable_endpoint++;
16509 case '0': case '1': case '2': case '3': case '4':
16510 case '5': case '6': case '7':
16512 /* Take 1-3 octal digits */
16513 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
16514 numlen = (strict) ? 4 : 3;
16515 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
16516 RExC_parse += numlen;
16519 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16520 vFAIL("Need exactly 3 octal digits");
16522 else if (! SIZE_ONLY /* like \08, \178 */
16524 && RExC_parse < RExC_end
16525 && isDIGIT(*RExC_parse)
16526 && ckWARN(WARN_REGEXP))
16528 SAVEFREESV(RExC_rx_sv);
16529 reg_warn_non_literal_string(
16531 form_short_octal_warning(RExC_parse, numlen));
16532 (void)ReREFCNT_inc(RExC_rx_sv);
16535 non_portable_endpoint++;
16539 /* Allow \_ to not give an error */
16540 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
16542 vFAIL2("Unrecognized escape \\%c in character class",
16546 SAVEFREESV(RExC_rx_sv);
16547 ckWARN2reg(RExC_parse,
16548 "Unrecognized escape \\%c in character class passed through",
16550 (void)ReREFCNT_inc(RExC_rx_sv);
16554 } /* End of switch on char following backslash */
16555 } /* end of handling backslash escape sequences */
16557 /* Here, we have the current token in 'value' */
16559 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
16562 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
16563 * literal, as is the character that began the false range, i.e.
16564 * the 'a' in the examples */
16567 const int w = (RExC_parse >= rangebegin)
16568 ? RExC_parse - rangebegin
16572 "False [] range \"%" UTF8f "\"",
16573 UTF8fARG(UTF, w, rangebegin));
16576 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
16577 ckWARN2reg(RExC_parse,
16578 "False [] range \"%" UTF8f "\"",
16579 UTF8fARG(UTF, w, rangebegin));
16580 (void)ReREFCNT_inc(RExC_rx_sv);
16581 cp_list = add_cp_to_invlist(cp_list, '-');
16582 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
16587 range = 0; /* this was not a true range */
16588 element_count += 2; /* So counts for three values */
16591 classnum = namedclass_to_classnum(namedclass);
16593 if (LOC && namedclass < ANYOF_POSIXL_MAX
16594 #ifndef HAS_ISASCII
16595 && classnum != _CC_ASCII
16598 /* What the Posix classes (like \w, [:space:]) match in locale
16599 * isn't knowable under locale until actual match time. Room
16600 * must be reserved (one time per outer bracketed class) to
16601 * store such classes. The space will contain a bit for each
16602 * named class that is to be matched against. This isn't
16603 * needed for \p{} and pseudo-classes, as they are not affected
16604 * by locale, and hence are dealt with separately */
16605 if (! need_class) {
16608 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16611 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16613 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
16614 ANYOF_POSIXL_ZERO(ret);
16616 /* We can't change this into some other type of node
16617 * (unless this is the only element, in which case there
16618 * are nodes that mean exactly this) as has runtime
16620 optimizable = FALSE;
16623 /* Coverity thinks it is possible for this to be negative; both
16624 * jhi and khw think it's not, but be safer */
16625 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16626 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
16628 /* See if it already matches the complement of this POSIX
16630 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16631 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
16635 posixl_matches_all = TRUE;
16636 break; /* No need to continue. Since it matches both
16637 e.g., \w and \W, it matches everything, and the
16638 bracketed class can be optimized into qr/./s */
16641 /* Add this class to those that should be checked at runtime */
16642 ANYOF_POSIXL_SET(ret, namedclass);
16644 /* The above-Latin1 characters are not subject to locale rules.
16645 * Just add them, in the second pass, to the
16646 * unconditionally-matched list */
16648 SV* scratch_list = NULL;
16650 /* Get the list of the above-Latin1 code points this
16652 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
16653 PL_XPosix_ptrs[classnum],
16655 /* Odd numbers are complements, like
16656 * NDIGIT, NASCII, ... */
16657 namedclass % 2 != 0,
16659 /* Checking if 'cp_list' is NULL first saves an extra
16660 * clone. Its reference count will be decremented at the
16661 * next union, etc, or if this is the only instance, at the
16662 * end of the routine */
16664 cp_list = scratch_list;
16667 _invlist_union(cp_list, scratch_list, &cp_list);
16668 SvREFCNT_dec_NN(scratch_list);
16670 continue; /* Go get next character */
16673 else if (! SIZE_ONLY) {
16675 /* Here, not in pass1 (in that pass we skip calculating the
16676 * contents of this class), and is not /l, or is a POSIX class
16677 * for which /l doesn't matter (or is a Unicode property, which
16678 * is skipped here). */
16679 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
16680 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
16682 /* Here, should be \h, \H, \v, or \V. None of /d, /i
16683 * nor /l make a difference in what these match,
16684 * therefore we just add what they match to cp_list. */
16685 if (classnum != _CC_VERTSPACE) {
16686 assert( namedclass == ANYOF_HORIZWS
16687 || namedclass == ANYOF_NHORIZWS);
16689 /* It turns out that \h is just a synonym for
16691 classnum = _CC_BLANK;
16694 _invlist_union_maybe_complement_2nd(
16696 PL_XPosix_ptrs[classnum],
16697 namedclass % 2 != 0, /* Complement if odd
16698 (NHORIZWS, NVERTWS)
16703 else if ( UNI_SEMANTICS
16704 || classnum == _CC_ASCII
16705 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
16706 || classnum == _CC_XDIGIT)))
16708 /* We usually have to worry about /d and /a affecting what
16709 * POSIX classes match, with special code needed for /d
16710 * because we won't know until runtime what all matches.
16711 * But there is no extra work needed under /u, and
16712 * [:ascii:] is unaffected by /a and /d; and :digit: and
16713 * :xdigit: don't have runtime differences under /d. So we
16714 * can special case these, and avoid some extra work below,
16715 * and at runtime. */
16716 _invlist_union_maybe_complement_2nd(
16718 PL_XPosix_ptrs[classnum],
16719 namedclass % 2 != 0,
16722 else { /* Garden variety class. If is NUPPER, NALPHA, ...
16723 complement and use nposixes */
16724 SV** posixes_ptr = namedclass % 2 == 0
16727 _invlist_union_maybe_complement_2nd(
16729 PL_XPosix_ptrs[classnum],
16730 namedclass % 2 != 0,
16734 } /* end of namedclass \blah */
16736 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16738 /* If 'range' is set, 'value' is the ending of a range--check its
16739 * validity. (If value isn't a single code point in the case of a
16740 * range, we should have figured that out above in the code that
16741 * catches false ranges). Later, we will handle each individual code
16742 * point in the range. If 'range' isn't set, this could be the
16743 * beginning of a range, so check for that by looking ahead to see if
16744 * the next real character to be processed is the range indicator--the
16749 /* For unicode ranges, we have to test that the Unicode as opposed
16750 * to the native values are not decreasing. (Above 255, there is
16751 * no difference between native and Unicode) */
16752 if (unicode_range && prevvalue < 255 && value < 255) {
16753 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
16754 goto backwards_range;
16759 if (prevvalue > value) /* b-a */ {
16764 w = RExC_parse - rangebegin;
16766 "Invalid [] range \"%" UTF8f "\"",
16767 UTF8fARG(UTF, w, rangebegin));
16768 NOT_REACHED; /* NOTREACHED */
16772 prevvalue = value; /* save the beginning of the potential range */
16773 if (! stop_at_1 /* Can't be a range if parsing just one thing */
16774 && *RExC_parse == '-')
16776 char* next_char_ptr = RExC_parse + 1;
16778 /* Get the next real char after the '-' */
16779 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
16781 /* If the '-' is at the end of the class (just before the ']',
16782 * it is a literal minus; otherwise it is a range */
16783 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
16784 RExC_parse = next_char_ptr;
16786 /* a bad range like \w-, [:word:]- ? */
16787 if (namedclass > OOB_NAMEDCLASS) {
16788 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
16789 const int w = RExC_parse >= rangebegin
16790 ? RExC_parse - rangebegin
16793 vFAIL4("False [] range \"%*.*s\"",
16798 "False [] range \"%*.*s\"",
16803 cp_list = add_cp_to_invlist(cp_list, '-');
16807 range = 1; /* yeah, it's a range! */
16808 continue; /* but do it the next time */
16813 if (namedclass > OOB_NAMEDCLASS) {
16817 /* Here, we have a single value this time through the loop, and
16818 * <prevvalue> is the beginning of the range, if any; or <value> if
16821 /* non-Latin1 code point implies unicode semantics. Must be set in
16822 * pass1 so is there for the whole of pass 2 */
16824 REQUIRE_UNI_RULES(flagp, NULL);
16827 /* Ready to process either the single value, or the completed range.
16828 * For single-valued non-inverted ranges, we consider the possibility
16829 * of multi-char folds. (We made a conscious decision to not do this
16830 * for the other cases because it can often lead to non-intuitive
16831 * results. For example, you have the peculiar case that:
16832 * "s s" =~ /^[^\xDF]+$/i => Y
16833 * "ss" =~ /^[^\xDF]+$/i => N
16835 * See [perl #89750] */
16836 if (FOLD && allow_multi_folds && value == prevvalue) {
16837 if (value == LATIN_SMALL_LETTER_SHARP_S
16838 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
16841 /* Here <value> is indeed a multi-char fold. Get what it is */
16843 U8 foldbuf[UTF8_MAXBYTES_CASE];
16846 UV folded = _to_uni_fold_flags(
16850 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
16851 ? FOLD_FLAGS_NOMIX_ASCII
16855 /* Here, <folded> should be the first character of the
16856 * multi-char fold of <value>, with <foldbuf> containing the
16857 * whole thing. But, if this fold is not allowed (because of
16858 * the flags), <fold> will be the same as <value>, and should
16859 * be processed like any other character, so skip the special
16861 if (folded != value) {
16863 /* Skip if we are recursed, currently parsing the class
16864 * again. Otherwise add this character to the list of
16865 * multi-char folds. */
16866 if (! RExC_in_multi_char_class) {
16867 STRLEN cp_count = utf8_length(foldbuf,
16868 foldbuf + foldlen);
16869 SV* multi_fold = sv_2mortal(newSVpvs(""));
16871 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
16874 = add_multi_match(multi_char_matches,
16880 /* This element should not be processed further in this
16883 value = save_value;
16884 prevvalue = save_prevvalue;
16890 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
16893 /* If the range starts above 255, everything is portable and
16894 * likely to be so for any forseeable character set, so don't
16896 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
16897 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
16899 else if (prevvalue != value) {
16901 /* Under strict, ranges that stop and/or end in an ASCII
16902 * printable should have each end point be a portable value
16903 * for it (preferably like 'A', but we don't warn if it is
16904 * a (portable) Unicode name or code point), and the range
16905 * must be be all digits or all letters of the same case.
16906 * Otherwise, the range is non-portable and unclear as to
16907 * what it contains */
16908 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
16909 && ( non_portable_endpoint
16910 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
16911 || (isLOWER_A(prevvalue) && isLOWER_A(value))
16912 || (isUPPER_A(prevvalue) && isUPPER_A(value))
16914 vWARN(RExC_parse, "Ranges of ASCII printables should"
16915 " be some subset of \"0-9\","
16916 " \"A-Z\", or \"a-z\"");
16918 else if (prevvalue >= 0x660) { /* ARABIC_INDIC_DIGIT_ZERO */
16919 SSize_t index_start;
16920 SSize_t index_final;
16922 /* But the nature of Unicode and languages mean we
16923 * can't do the same checks for above-ASCII ranges,
16924 * except in the case of digit ones. These should
16925 * contain only digits from the same group of 10. The
16926 * ASCII case is handled just above. 0x660 is the
16927 * first digit character beyond ASCII. Hence here, the
16928 * range could be a range of digits. First some
16929 * unlikely special cases. Grandfather in that a range
16930 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
16931 * if its starting value is one of the 10 digits prior
16932 * to it. This is because it is an alternate way of
16933 * writing 19D1, and some people may expect it to be in
16934 * that group. But it is bad, because it won't give
16935 * the expected results. In Unicode 5.2 it was
16936 * considered to be in that group (of 11, hence), but
16937 * this was fixed in the next version */
16939 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
16940 goto warn_bad_digit_range;
16942 else if (UNLIKELY( prevvalue >= 0x1D7CE
16943 && value <= 0x1D7FF))
16945 /* This is the only other case currently in Unicode
16946 * where the algorithm below fails. The code
16947 * points just above are the end points of a single
16948 * range containing only decimal digits. It is 5
16949 * different series of 0-9. All other ranges of
16950 * digits currently in Unicode are just a single
16951 * series. (And mktables will notify us if a later
16952 * Unicode version breaks this.)
16954 * If the range being checked is at most 9 long,
16955 * and the digit values represented are in
16956 * numerical order, they are from the same series.
16958 if ( value - prevvalue > 9
16959 || ((( value - 0x1D7CE) % 10)
16960 <= (prevvalue - 0x1D7CE) % 10))
16962 goto warn_bad_digit_range;
16967 /* For all other ranges of digits in Unicode, the
16968 * algorithm is just to check if both end points
16969 * are in the same series, which is the same range.
16971 index_start = _invlist_search(
16972 PL_XPosix_ptrs[_CC_DIGIT],
16975 /* Warn if the range starts and ends with a digit,
16976 * and they are not in the same group of 10. */
16977 if ( index_start >= 0
16978 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
16980 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
16981 value)) != index_start
16982 && index_final >= 0
16983 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
16985 warn_bad_digit_range:
16986 vWARN(RExC_parse, "Ranges of digits should be"
16987 " from the same group of"
16994 if ((! range || prevvalue == value) && non_portable_endpoint) {
16995 if (isPRINT_A(value)) {
16998 if (isBACKSLASHED_PUNCT(value)) {
16999 literal[d++] = '\\';
17001 literal[d++] = (char) value;
17002 literal[d++] = '\0';
17005 "\"%.*s\" is more clearly written simply as \"%s\"",
17006 (int) (RExC_parse - rangebegin),
17011 else if isMNEMONIC_CNTRL(value) {
17013 "\"%.*s\" is more clearly written simply as \"%s\"",
17014 (int) (RExC_parse - rangebegin),
17016 cntrl_to_mnemonic((U8) value)
17022 /* Deal with this element of the class */
17026 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17029 /* On non-ASCII platforms, for ranges that span all of 0..255, and
17030 * ones that don't require special handling, we can just add the
17031 * range like we do for ASCII platforms */
17032 if ((UNLIKELY(prevvalue == 0) && value >= 255)
17033 || ! (prevvalue < 256
17035 || (! non_portable_endpoint
17036 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
17037 || (isUPPER_A(prevvalue)
17038 && isUPPER_A(value)))))))
17040 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17044 /* Here, requires special handling. This can be because it is
17045 * a range whose code points are considered to be Unicode, and
17046 * so must be individually translated into native, or because
17047 * its a subrange of 'A-Z' or 'a-z' which each aren't
17048 * contiguous in EBCDIC, but we have defined them to include
17049 * only the "expected" upper or lower case ASCII alphabetics.
17050 * Subranges above 255 are the same in native and Unicode, so
17051 * can be added as a range */
17052 U8 start = NATIVE_TO_LATIN1(prevvalue);
17054 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
17055 for (j = start; j <= end; j++) {
17056 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
17059 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17066 range = 0; /* this range (if it was one) is done now */
17067 } /* End of loop through all the text within the brackets */
17070 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
17071 output_or_return_posix_warnings(pRExC_state, posix_warnings,
17072 return_posix_warnings);
17075 /* If anything in the class expands to more than one character, we have to
17076 * deal with them by building up a substitute parse string, and recursively
17077 * calling reg() on it, instead of proceeding */
17078 if (multi_char_matches) {
17079 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
17082 char *save_end = RExC_end;
17083 char *save_parse = RExC_parse;
17084 char *save_start = RExC_start;
17085 STRLEN prefix_end = 0; /* We copy the character class after a
17086 prefix supplied here. This is the size
17087 + 1 of that prefix */
17088 bool first_time = TRUE; /* First multi-char occurrence doesn't get
17093 assert(RExC_precomp_adj == 0); /* Only one level of recursion allowed */
17095 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
17096 because too confusing */
17098 sv_catpv(substitute_parse, "(?:");
17102 /* Look at the longest folds first */
17103 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
17108 if (av_exists(multi_char_matches, cp_count)) {
17109 AV** this_array_ptr;
17112 this_array_ptr = (AV**) av_fetch(multi_char_matches,
17114 while ((this_sequence = av_pop(*this_array_ptr)) !=
17117 if (! first_time) {
17118 sv_catpv(substitute_parse, "|");
17120 first_time = FALSE;
17122 sv_catpv(substitute_parse, SvPVX(this_sequence));
17127 /* If the character class contains anything else besides these
17128 * multi-character folds, have to include it in recursive parsing */
17129 if (element_count) {
17130 sv_catpv(substitute_parse, "|[");
17131 prefix_end = SvCUR(substitute_parse);
17132 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
17134 /* Put in a closing ']' only if not going off the end, as otherwise
17135 * we are adding something that really isn't there */
17136 if (RExC_parse < RExC_end) {
17137 sv_catpv(substitute_parse, "]");
17141 sv_catpv(substitute_parse, ")");
17144 /* This is a way to get the parse to skip forward a whole named
17145 * sequence instead of matching the 2nd character when it fails the
17147 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
17151 /* Set up the data structure so that any errors will be properly
17152 * reported. See the comments at the definition of
17153 * REPORT_LOCATION_ARGS for details */
17154 RExC_precomp_adj = orig_parse - RExC_precomp;
17155 RExC_start = RExC_parse = SvPV(substitute_parse, len);
17156 RExC_adjusted_start = RExC_start + prefix_end;
17157 RExC_end = RExC_parse + len;
17158 RExC_in_multi_char_class = 1;
17159 RExC_emit = (regnode *)orig_emit;
17161 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17163 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PASS1|NEED_UTF8);
17165 /* And restore so can parse the rest of the pattern */
17166 RExC_parse = save_parse;
17167 RExC_start = RExC_adjusted_start = save_start;
17168 RExC_precomp_adj = 0;
17169 RExC_end = save_end;
17170 RExC_in_multi_char_class = 0;
17171 SvREFCNT_dec_NN(multi_char_matches);
17175 /* Here, we've gone through the entire class and dealt with multi-char
17176 * folds. We are now in a position that we can do some checks to see if we
17177 * can optimize this ANYOF node into a simpler one, even in Pass 1.
17178 * Currently we only do two checks:
17179 * 1) is in the unlikely event that the user has specified both, eg. \w and
17180 * \W under /l, then the class matches everything. (This optimization
17181 * is done only to make the optimizer code run later work.)
17182 * 2) if the character class contains only a single element (including a
17183 * single range), we see if there is an equivalent node for it.
17184 * Other checks are possible */
17186 && ! ret_invlist /* Can't optimize if returning the constructed
17188 && (UNLIKELY(posixl_matches_all) || element_count == 1))
17193 if (UNLIKELY(posixl_matches_all)) {
17196 else if (namedclass > OOB_NAMEDCLASS) { /* this is a single named
17197 class, like \w or [:digit:]
17200 /* All named classes are mapped into POSIXish nodes, with its FLAG
17201 * argument giving which class it is */
17202 switch ((I32)namedclass) {
17203 case ANYOF_UNIPROP:
17206 /* These don't depend on the charset modifiers. They always
17207 * match under /u rules */
17208 case ANYOF_NHORIZWS:
17209 case ANYOF_HORIZWS:
17210 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
17213 case ANYOF_NVERTWS:
17218 /* The actual POSIXish node for all the rest depends on the
17219 * charset modifier. The ones in the first set depend only on
17220 * ASCII or, if available on this platform, also locale */
17224 op = (LOC) ? POSIXL : POSIXA;
17230 /* The following don't have any matches in the upper Latin1
17231 * range, hence /d is equivalent to /u for them. Making it /u
17232 * saves some branches at runtime */
17236 case ANYOF_NXDIGIT:
17237 if (! DEPENDS_SEMANTICS) {
17238 goto treat_as_default;
17244 /* The following change to CASED under /i */
17250 namedclass = ANYOF_CASED + (namedclass % 2);
17254 /* The rest have more possibilities depending on the charset.
17255 * We take advantage of the enum ordering of the charset
17256 * modifiers to get the exact node type, */
17259 op = POSIXD + get_regex_charset(RExC_flags);
17260 if (op > POSIXA) { /* /aa is same as /a */
17265 /* The odd numbered ones are the complements of the
17266 * next-lower even number one */
17267 if (namedclass % 2 == 1) {
17271 arg = namedclass_to_classnum(namedclass);
17275 else if (value == prevvalue) {
17277 /* Here, the class consists of just a single code point */
17280 if (! LOC && value == '\n') {
17281 op = REG_ANY; /* Optimize [^\n] */
17282 *flagp |= HASWIDTH|SIMPLE;
17286 else if (value < 256 || UTF) {
17288 /* Optimize a single value into an EXACTish node, but not if it
17289 * would require converting the pattern to UTF-8. */
17290 op = compute_EXACTish(pRExC_state);
17292 } /* Otherwise is a range */
17293 else if (! LOC) { /* locale could vary these */
17294 if (prevvalue == '0') {
17295 if (value == '9') {
17300 else if (! FOLD || ASCII_FOLD_RESTRICTED) {
17301 /* We can optimize A-Z or a-z, but not if they could match
17302 * something like the KELVIN SIGN under /i. */
17303 if (prevvalue == 'A') {
17306 && ! non_portable_endpoint
17309 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
17313 else if (prevvalue == 'a') {
17316 && ! non_portable_endpoint
17319 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
17326 /* Here, we have changed <op> away from its initial value iff we found
17327 * an optimization */
17330 /* Throw away this ANYOF regnode, and emit the calculated one,
17331 * which should correspond to the beginning, not current, state of
17333 const char * cur_parse = RExC_parse;
17334 RExC_parse = (char *)orig_parse;
17338 /* To get locale nodes to not use the full ANYOF size would
17339 * require moving the code above that writes the portions
17340 * of it that aren't in other nodes to after this point.
17341 * e.g. ANYOF_POSIXL_SET */
17342 RExC_size = orig_size;
17346 RExC_emit = (regnode *)orig_emit;
17347 if (PL_regkind[op] == POSIXD) {
17348 if (op == POSIXL) {
17349 RExC_contains_locale = 1;
17352 op += NPOSIXD - POSIXD;
17357 ret = reg_node(pRExC_state, op);
17359 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17363 *flagp |= HASWIDTH|SIMPLE;
17365 else if (PL_regkind[op] == EXACT) {
17366 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17367 TRUE /* downgradable to EXACT */
17371 RExC_parse = (char *) cur_parse;
17373 SvREFCNT_dec(posixes);
17374 SvREFCNT_dec(nposixes);
17375 SvREFCNT_dec(simple_posixes);
17376 SvREFCNT_dec(cp_list);
17377 SvREFCNT_dec(cp_foldable_list);
17384 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
17386 /* If folding, we calculate all characters that could fold to or from the
17387 * ones already on the list */
17388 if (cp_foldable_list) {
17390 UV start, end; /* End points of code point ranges */
17392 SV* fold_intersection = NULL;
17395 /* Our calculated list will be for Unicode rules. For locale
17396 * matching, we have to keep a separate list that is consulted at
17397 * runtime only when the locale indicates Unicode rules. For
17398 * non-locale, we just use the general list */
17400 use_list = &only_utf8_locale_list;
17403 use_list = &cp_list;
17406 /* Only the characters in this class that participate in folds need
17407 * be checked. Get the intersection of this class and all the
17408 * possible characters that are foldable. This can quickly narrow
17409 * down a large class */
17410 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
17411 &fold_intersection);
17413 /* The folds for all the Latin1 characters are hard-coded into this
17414 * program, but we have to go out to disk to get the others. */
17415 if (invlist_highest(cp_foldable_list) >= 256) {
17417 /* This is a hash that for a particular fold gives all
17418 * characters that are involved in it */
17419 if (! PL_utf8_foldclosures) {
17420 _load_PL_utf8_foldclosures();
17424 /* Now look at the foldable characters in this class individually */
17425 invlist_iterinit(fold_intersection);
17426 while (invlist_iternext(fold_intersection, &start, &end)) {
17429 /* Look at every character in the range */
17430 for (j = start; j <= end; j++) {
17431 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17437 if (IS_IN_SOME_FOLD_L1(j)) {
17439 /* ASCII is always matched; non-ASCII is matched
17440 * only under Unicode rules (which could happen
17441 * under /l if the locale is a UTF-8 one */
17442 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
17443 *use_list = add_cp_to_invlist(*use_list,
17444 PL_fold_latin1[j]);
17447 has_upper_latin1_only_utf8_matches
17448 = add_cp_to_invlist(
17449 has_upper_latin1_only_utf8_matches,
17450 PL_fold_latin1[j]);
17454 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
17455 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
17457 add_above_Latin1_folds(pRExC_state,
17464 /* Here is an above Latin1 character. We don't have the
17465 * rules hard-coded for it. First, get its fold. This is
17466 * the simple fold, as the multi-character folds have been
17467 * handled earlier and separated out */
17468 _to_uni_fold_flags(j, foldbuf, &foldlen,
17469 (ASCII_FOLD_RESTRICTED)
17470 ? FOLD_FLAGS_NOMIX_ASCII
17473 /* Single character fold of above Latin1. Add everything in
17474 * its fold closure to the list that this node should match.
17475 * The fold closures data structure is a hash with the keys
17476 * being the UTF-8 of every character that is folded to, like
17477 * 'k', and the values each an array of all code points that
17478 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
17479 * Multi-character folds are not included */
17480 if ((listp = hv_fetch(PL_utf8_foldclosures,
17481 (char *) foldbuf, foldlen, FALSE)))
17483 AV* list = (AV*) *listp;
17485 for (k = 0; k <= av_tindex_skip_len_mg(list); k++) {
17486 SV** c_p = av_fetch(list, k, FALSE);
17492 /* /aa doesn't allow folds between ASCII and non- */
17493 if ((ASCII_FOLD_RESTRICTED
17494 && (isASCII(c) != isASCII(j))))
17499 /* Folds under /l which cross the 255/256 boundary
17500 * are added to a separate list. (These are valid
17501 * only when the locale is UTF-8.) */
17502 if (c < 256 && LOC) {
17503 *use_list = add_cp_to_invlist(*use_list, c);
17507 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
17509 cp_list = add_cp_to_invlist(cp_list, c);
17512 /* Similarly folds involving non-ascii Latin1
17513 * characters under /d are added to their list */
17514 has_upper_latin1_only_utf8_matches
17515 = add_cp_to_invlist(
17516 has_upper_latin1_only_utf8_matches,
17523 SvREFCNT_dec_NN(fold_intersection);
17526 /* Now that we have finished adding all the folds, there is no reason
17527 * to keep the foldable list separate */
17528 _invlist_union(cp_list, cp_foldable_list, &cp_list);
17529 SvREFCNT_dec_NN(cp_foldable_list);
17532 /* And combine the result (if any) with any inversion lists from posix
17533 * classes. The lists are kept separate up to now because we don't want to
17534 * fold the classes (folding of those is automatically handled by the swash
17535 * fetching code) */
17536 if (simple_posixes) { /* These are the classes known to be unaffected by
17539 _invlist_union(cp_list, simple_posixes, &cp_list);
17540 SvREFCNT_dec_NN(simple_posixes);
17543 cp_list = simple_posixes;
17546 if (posixes || nposixes) {
17548 /* We have to adjust /a and /aa */
17549 if (AT_LEAST_ASCII_RESTRICTED) {
17551 /* Under /a and /aa, nothing above ASCII matches these */
17553 _invlist_intersection(posixes,
17554 PL_XPosix_ptrs[_CC_ASCII],
17558 /* Under /a and /aa, everything above ASCII matches these
17561 _invlist_union_complement_2nd(nposixes,
17562 PL_XPosix_ptrs[_CC_ASCII],
17567 if (! DEPENDS_SEMANTICS) {
17569 /* For everything but /d, we can just add the current 'posixes' and
17570 * 'nposixes' to the main list */
17573 _invlist_union(cp_list, posixes, &cp_list);
17574 SvREFCNT_dec_NN(posixes);
17582 _invlist_union(cp_list, nposixes, &cp_list);
17583 SvREFCNT_dec_NN(nposixes);
17586 cp_list = nposixes;
17591 /* Under /d, things like \w match upper Latin1 characters only if
17592 * the target string is in UTF-8. But things like \W match all the
17593 * upper Latin1 characters if the target string is not in UTF-8.
17595 * Handle the case where there something like \W separately */
17597 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1);
17599 /* A complemented posix class matches all upper Latin1
17600 * characters if not in UTF-8. And it matches just certain
17601 * ones when in UTF-8. That means those certain ones are
17602 * matched regardless, so can just be added to the
17603 * unconditional list */
17605 _invlist_union(cp_list, nposixes, &cp_list);
17606 SvREFCNT_dec_NN(nposixes);
17610 cp_list = nposixes;
17613 /* Likewise for 'posixes' */
17614 _invlist_union(posixes, cp_list, &cp_list);
17616 /* Likewise for anything else in the range that matched only
17618 if (has_upper_latin1_only_utf8_matches) {
17619 _invlist_union(cp_list,
17620 has_upper_latin1_only_utf8_matches,
17622 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17623 has_upper_latin1_only_utf8_matches = NULL;
17626 /* If we don't match all the upper Latin1 characters regardless
17627 * of UTF-8ness, we have to set a flag to match the rest when
17629 _invlist_subtract(only_non_utf8_list, cp_list,
17630 &only_non_utf8_list);
17631 if (_invlist_len(only_non_utf8_list) != 0) {
17632 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17636 /* Here there were no complemented posix classes. That means
17637 * the upper Latin1 characters in 'posixes' match only when the
17638 * target string is in UTF-8. So we have to add them to the
17639 * list of those types of code points, while adding the
17640 * remainder to the unconditional list.
17642 * First calculate what they are */
17643 SV* nonascii_but_latin1_properties = NULL;
17644 _invlist_intersection(posixes, PL_UpperLatin1,
17645 &nonascii_but_latin1_properties);
17647 /* And add them to the final list of such characters. */
17648 _invlist_union(has_upper_latin1_only_utf8_matches,
17649 nonascii_but_latin1_properties,
17650 &has_upper_latin1_only_utf8_matches);
17652 /* Remove them from what now becomes the unconditional list */
17653 _invlist_subtract(posixes, nonascii_but_latin1_properties,
17656 /* And add those unconditional ones to the final list */
17658 _invlist_union(cp_list, posixes, &cp_list);
17659 SvREFCNT_dec_NN(posixes);
17666 SvREFCNT_dec(nonascii_but_latin1_properties);
17668 /* Get rid of any characters that we now know are matched
17669 * unconditionally from the conditional list, which may make
17670 * that list empty */
17671 _invlist_subtract(has_upper_latin1_only_utf8_matches,
17673 &has_upper_latin1_only_utf8_matches);
17674 if (_invlist_len(has_upper_latin1_only_utf8_matches) == 0) {
17675 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17676 has_upper_latin1_only_utf8_matches = NULL;
17682 /* And combine the result (if any) with any inversion list from properties.
17683 * The lists are kept separate up to now so that we can distinguish the two
17684 * in regards to matching above-Unicode. A run-time warning is generated
17685 * if a Unicode property is matched against a non-Unicode code point. But,
17686 * we allow user-defined properties to match anything, without any warning,
17687 * and we also suppress the warning if there is a portion of the character
17688 * class that isn't a Unicode property, and which matches above Unicode, \W
17689 * or [\x{110000}] for example.
17690 * (Note that in this case, unlike the Posix one above, there is no
17691 * <has_upper_latin1_only_utf8_matches>, because having a Unicode property
17692 * forces Unicode semantics */
17696 /* If it matters to the final outcome, see if a non-property
17697 * component of the class matches above Unicode. If so, the
17698 * warning gets suppressed. This is true even if just a single
17699 * such code point is specified, as, though not strictly correct if
17700 * another such code point is matched against, the fact that they
17701 * are using above-Unicode code points indicates they should know
17702 * the issues involved */
17704 warn_super = ! (invert
17705 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
17708 _invlist_union(properties, cp_list, &cp_list);
17709 SvREFCNT_dec_NN(properties);
17712 cp_list = properties;
17717 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17719 /* Because an ANYOF node is the only one that warns, this node
17720 * can't be optimized into something else */
17721 optimizable = FALSE;
17725 /* Here, we have calculated what code points should be in the character
17728 * Now we can see about various optimizations. Fold calculation (which we
17729 * did above) needs to take place before inversion. Otherwise /[^k]/i
17730 * would invert to include K, which under /i would match k, which it
17731 * shouldn't. Therefore we can't invert folded locale now, as it won't be
17732 * folded until runtime */
17734 /* If we didn't do folding, it's because some information isn't available
17735 * until runtime; set the run-time fold flag for these. (We don't have to
17736 * worry about properties folding, as that is taken care of by the swash
17737 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
17738 * locales, or the class matches at least one 0-255 range code point */
17741 /* Some things on the list might be unconditionally included because of
17742 * other components. Remove them, and clean up the list if it goes to
17744 if (only_utf8_locale_list && cp_list) {
17745 _invlist_subtract(only_utf8_locale_list, cp_list,
17746 &only_utf8_locale_list);
17748 if (_invlist_len(only_utf8_locale_list) == 0) {
17749 SvREFCNT_dec_NN(only_utf8_locale_list);
17750 only_utf8_locale_list = NULL;
17753 if (only_utf8_locale_list) {
17756 |ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17758 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
17760 invlist_iterinit(cp_list);
17761 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
17762 ANYOF_FLAGS(ret) |= ANYOFL_FOLD;
17764 invlist_iterfinish(cp_list);
17767 else if ( DEPENDS_SEMANTICS
17768 && ( has_upper_latin1_only_utf8_matches
17769 || (ANYOF_FLAGS(ret) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
17772 optimizable = FALSE;
17776 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
17777 * at compile time. Besides not inverting folded locale now, we can't
17778 * invert if there are things such as \w, which aren't known until runtime
17782 && OP(ret) != ANYOFD
17783 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
17784 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17786 _invlist_invert(cp_list);
17788 /* Any swash can't be used as-is, because we've inverted things */
17790 SvREFCNT_dec_NN(swash);
17794 /* Clear the invert flag since have just done it here */
17801 *ret_invlist = cp_list;
17802 SvREFCNT_dec(swash);
17804 /* Discard the generated node */
17806 RExC_size = orig_size;
17809 RExC_emit = orig_emit;
17814 /* Some character classes are equivalent to other nodes. Such nodes take
17815 * up less room and generally fewer operations to execute than ANYOF nodes.
17816 * Above, we checked for and optimized into some such equivalents for
17817 * certain common classes that are easy to test. Getting to this point in
17818 * the code means that the class didn't get optimized there. Since this
17819 * code is only executed in Pass 2, it is too late to save space--it has
17820 * been allocated in Pass 1, and currently isn't given back. But turning
17821 * things into an EXACTish node can allow the optimizer to join it to any
17822 * adjacent such nodes. And if the class is equivalent to things like /./,
17823 * expensive run-time swashes can be avoided. Now that we have more
17824 * complete information, we can find things necessarily missed by the
17825 * earlier code. Another possible "optimization" that isn't done is that
17826 * something like [Ee] could be changed into an EXACTFU. khw tried this
17827 * and found that the ANYOF is faster, including for code points not in the
17828 * bitmap. This still might make sense to do, provided it got joined with
17829 * an adjacent node(s) to create a longer EXACTFU one. This could be
17830 * accomplished by creating a pseudo ANYOF_EXACTFU node type that the join
17831 * routine would know is joinable. If that didn't happen, the node type
17832 * could then be made a straight ANYOF */
17834 if (optimizable && cp_list && ! invert) {
17836 U8 op = END; /* The optimzation node-type */
17837 int posix_class = -1; /* Illegal value */
17838 const char * cur_parse= RExC_parse;
17840 invlist_iterinit(cp_list);
17841 if (! invlist_iternext(cp_list, &start, &end)) {
17843 /* Here, the list is empty. This happens, for example, when a
17844 * Unicode property that doesn't match anything is the only element
17845 * in the character class (perluniprops.pod notes such properties).
17848 *flagp |= HASWIDTH|SIMPLE;
17850 else if (start == end) { /* The range is a single code point */
17851 if (! invlist_iternext(cp_list, &start, &end)
17853 /* Don't do this optimization if it would require changing
17854 * the pattern to UTF-8 */
17855 && (start < 256 || UTF))
17857 /* Here, the list contains a single code point. Can optimize
17858 * into an EXACTish node */
17869 /* A locale node under folding with one code point can be
17870 * an EXACTFL, as its fold won't be calculated until
17876 /* Here, we are generally folding, but there is only one
17877 * code point to match. If we have to, we use an EXACT
17878 * node, but it would be better for joining with adjacent
17879 * nodes in the optimization pass if we used the same
17880 * EXACTFish node that any such are likely to be. We can
17881 * do this iff the code point doesn't participate in any
17882 * folds. For example, an EXACTF of a colon is the same as
17883 * an EXACT one, since nothing folds to or from a colon. */
17885 if (IS_IN_SOME_FOLD_L1(value)) {
17890 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
17895 /* If we haven't found the node type, above, it means we
17896 * can use the prevailing one */
17898 op = compute_EXACTish(pRExC_state);
17902 } /* End of first range contains just a single code point */
17903 else if (start == 0) {
17904 if (end == UV_MAX) {
17906 *flagp |= HASWIDTH|SIMPLE;
17909 else if (end == '\n' - 1
17910 && invlist_iternext(cp_list, &start, &end)
17911 && start == '\n' + 1 && end == UV_MAX)
17914 *flagp |= HASWIDTH|SIMPLE;
17918 invlist_iterfinish(cp_list);
17921 const UV cp_list_len = _invlist_len(cp_list);
17922 const UV* cp_list_array = invlist_array(cp_list);
17924 /* Here, didn't find an optimization. See if this matches any of
17925 * the POSIX classes. These run slightly faster for above-Unicode
17926 * code points, so don't bother with POSIXA ones nor the 2 that
17927 * have no above-Unicode matches. We can avoid these checks unless
17928 * the ANYOF matches at least as high as the lowest POSIX one
17929 * (which was manually found to be \v. The actual code point may
17930 * increase in later Unicode releases, if a higher code point is
17931 * assigned to be \v, but this code will never break. It would
17932 * just mean we could execute the checks for posix optimizations
17933 * unnecessarily) */
17935 if (cp_list_array[cp_list_len-1] > 0x2029) {
17936 for (posix_class = 0;
17937 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
17941 if (posix_class == _CC_ASCII || posix_class == _CC_CNTRL) {
17944 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
17946 /* Check if matches normal or inverted */
17947 if (_invlistEQ(cp_list,
17948 PL_XPosix_ptrs[posix_class],
17951 op = (try_inverted)
17954 *flagp |= HASWIDTH|SIMPLE;
17964 RExC_parse = (char *)orig_parse;
17965 RExC_emit = (regnode *)orig_emit;
17967 if (regarglen[op]) {
17968 ret = reganode(pRExC_state, op, 0);
17970 ret = reg_node(pRExC_state, op);
17973 RExC_parse = (char *)cur_parse;
17975 if (PL_regkind[op] == EXACT) {
17976 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17977 TRUE /* downgradable to EXACT */
17980 else if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17981 FLAGS(ret) = posix_class;
17984 SvREFCNT_dec_NN(cp_list);
17989 /* Here, <cp_list> contains all the code points we can determine at
17990 * compile time that match under all conditions. Go through it, and
17991 * for things that belong in the bitmap, put them there, and delete from
17992 * <cp_list>. While we are at it, see if everything above 255 is in the
17993 * list, and if so, set a flag to speed up execution */
17995 populate_ANYOF_from_invlist(ret, &cp_list);
17998 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
18001 /* Here, the bitmap has been populated with all the Latin1 code points that
18002 * always match. Can now add to the overall list those that match only
18003 * when the target string is UTF-8 (<has_upper_latin1_only_utf8_matches>).
18005 if (has_upper_latin1_only_utf8_matches) {
18007 _invlist_union(cp_list,
18008 has_upper_latin1_only_utf8_matches,
18010 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
18013 cp_list = has_upper_latin1_only_utf8_matches;
18015 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
18018 /* If there is a swash and more than one element, we can't use the swash in
18019 * the optimization below. */
18020 if (swash && element_count > 1) {
18021 SvREFCNT_dec_NN(swash);
18025 /* Note that the optimization of using 'swash' if it is the only thing in
18026 * the class doesn't have us change swash at all, so it can include things
18027 * that are also in the bitmap; otherwise we have purposely deleted that
18028 * duplicate information */
18029 set_ANYOF_arg(pRExC_state, ret, cp_list,
18030 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
18032 only_utf8_locale_list,
18033 swash, has_user_defined_property);
18035 *flagp |= HASWIDTH|SIMPLE;
18037 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
18038 RExC_contains_locale = 1;
18044 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
18047 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
18048 regnode* const node,
18050 SV* const runtime_defns,
18051 SV* const only_utf8_locale_list,
18053 const bool has_user_defined_property)
18055 /* Sets the arg field of an ANYOF-type node 'node', using information about
18056 * the node passed-in. If there is nothing outside the node's bitmap, the
18057 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
18058 * the count returned by add_data(), having allocated and stored an array,
18059 * av, that that count references, as follows:
18060 * av[0] stores the character class description in its textual form.
18061 * This is used later (regexec.c:Perl_regclass_swash()) to
18062 * initialize the appropriate swash, and is also useful for dumping
18063 * the regnode. This is set to &PL_sv_undef if the textual
18064 * description is not needed at run-time (as happens if the other
18065 * elements completely define the class)
18066 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
18067 * computed from av[0]. But if no further computation need be done,
18068 * the swash is stored here now (and av[0] is &PL_sv_undef).
18069 * av[2] stores the inversion list of code points that match only if the
18070 * current locale is UTF-8
18071 * av[3] stores the cp_list inversion list for use in addition or instead
18072 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
18073 * (Otherwise everything needed is already in av[0] and av[1])
18074 * av[4] is set if any component of the class is from a user-defined
18075 * property; used only if av[3] exists */
18079 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
18081 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
18082 assert(! (ANYOF_FLAGS(node)
18083 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
18084 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
18087 AV * const av = newAV();
18090 av_store(av, 0, (runtime_defns)
18091 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
18094 av_store(av, 1, swash);
18095 SvREFCNT_dec_NN(cp_list);
18098 av_store(av, 1, &PL_sv_undef);
18100 av_store(av, 3, cp_list);
18101 av_store(av, 4, newSVuv(has_user_defined_property));
18105 if (only_utf8_locale_list) {
18106 av_store(av, 2, only_utf8_locale_list);
18109 av_store(av, 2, &PL_sv_undef);
18112 rv = newRV_noinc(MUTABLE_SV(av));
18113 n = add_data(pRExC_state, STR_WITH_LEN("s"));
18114 RExC_rxi->data->data[n] = (void*)rv;
18119 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
18121 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
18122 const regnode* node,
18125 SV** only_utf8_locale_ptr,
18126 SV** output_invlist)
18129 /* For internal core use only.
18130 * Returns the swash for the input 'node' in the regex 'prog'.
18131 * If <doinit> is 'true', will attempt to create the swash if not already
18133 * If <listsvp> is non-null, will return the printable contents of the
18134 * swash. This can be used to get debugging information even before the
18135 * swash exists, by calling this function with 'doinit' set to false, in
18136 * which case the components that will be used to eventually create the
18137 * swash are returned (in a printable form).
18138 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
18139 * store an inversion list of code points that should match only if the
18140 * execution-time locale is a UTF-8 one.
18141 * If <output_invlist> is not NULL, it is where this routine is to store an
18142 * inversion list of the code points that would be instead returned in
18143 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
18144 * when this parameter is used, is just the non-code point data that
18145 * will go into creating the swash. This currently should be just
18146 * user-defined properties whose definitions were not known at compile
18147 * time. Using this parameter allows for easier manipulation of the
18148 * swash's data by the caller. It is illegal to call this function with
18149 * this parameter set, but not <listsvp>
18151 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
18152 * that, in spite of this function's name, the swash it returns may include
18153 * the bitmap data as well */
18156 SV *si = NULL; /* Input swash initialization string */
18157 SV* invlist = NULL;
18159 RXi_GET_DECL(prog,progi);
18160 const struct reg_data * const data = prog ? progi->data : NULL;
18162 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
18163 assert(! output_invlist || listsvp);
18165 if (data && data->count) {
18166 const U32 n = ARG(node);
18168 if (data->what[n] == 's') {
18169 SV * const rv = MUTABLE_SV(data->data[n]);
18170 AV * const av = MUTABLE_AV(SvRV(rv));
18171 SV **const ary = AvARRAY(av);
18172 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
18174 si = *ary; /* ary[0] = the string to initialize the swash with */
18176 if (av_tindex_skip_len_mg(av) >= 2) {
18177 if (only_utf8_locale_ptr
18179 && ary[2] != &PL_sv_undef)
18181 *only_utf8_locale_ptr = ary[2];
18184 assert(only_utf8_locale_ptr);
18185 *only_utf8_locale_ptr = NULL;
18188 /* Elements 3 and 4 are either both present or both absent. [3]
18189 * is any inversion list generated at compile time; [4]
18190 * indicates if that inversion list has any user-defined
18191 * properties in it. */
18192 if (av_tindex_skip_len_mg(av) >= 3) {
18194 if (SvUV(ary[4])) {
18195 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
18203 /* Element [1] is reserved for the set-up swash. If already there,
18204 * return it; if not, create it and store it there */
18205 if (ary[1] && SvROK(ary[1])) {
18208 else if (doinit && ((si && si != &PL_sv_undef)
18209 || (invlist && invlist != &PL_sv_undef))) {
18211 sw = _core_swash_init("utf8", /* the utf8 package */
18215 0, /* not from tr/// */
18217 &swash_init_flags);
18218 (void)av_store(av, 1, sw);
18223 /* If requested, return a printable version of what this swash matches */
18225 SV* matches_string = NULL;
18227 /* The swash should be used, if possible, to get the data, as it
18228 * contains the resolved data. But this function can be called at
18229 * compile-time, before everything gets resolved, in which case we
18230 * return the currently best available information, which is the string
18231 * that will eventually be used to do that resolving, 'si' */
18232 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
18233 && (si && si != &PL_sv_undef))
18235 /* Here, we only have 'si' (and possibly some passed-in data in
18236 * 'invlist', which is handled below) If the caller only wants
18237 * 'si', use that. */
18238 if (! output_invlist) {
18239 matches_string = newSVsv(si);
18242 /* But if the caller wants an inversion list of the node, we
18243 * need to parse 'si' and place as much as possible in the
18244 * desired output inversion list, making 'matches_string' only
18245 * contain the currently unresolvable things */
18246 const char *si_string = SvPVX(si);
18247 STRLEN remaining = SvCUR(si);
18251 /* Ignore everything before the first new-line */
18252 while (*si_string != '\n' && remaining > 0) {
18256 assert(remaining > 0);
18261 while (remaining > 0) {
18263 /* The data consists of just strings defining user-defined
18264 * property names, but in prior incarnations, and perhaps
18265 * somehow from pluggable regex engines, it could still
18266 * hold hex code point definitions. Each component of a
18267 * range would be separated by a tab, and each range by a
18268 * new-line. If these are found, instead add them to the
18269 * inversion list */
18270 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
18271 |PERL_SCAN_SILENT_NON_PORTABLE;
18272 STRLEN len = remaining;
18273 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
18275 /* If the hex decode routine found something, it should go
18276 * up to the next \n */
18277 if ( *(si_string + len) == '\n') {
18278 if (count) { /* 2nd code point on line */
18279 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
18282 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
18285 goto prepare_for_next_iteration;
18288 /* If the hex decode was instead for the lower range limit,
18289 * save it, and go parse the upper range limit */
18290 if (*(si_string + len) == '\t') {
18291 assert(count == 0);
18295 prepare_for_next_iteration:
18296 si_string += len + 1;
18297 remaining -= len + 1;
18301 /* Here, didn't find a legal hex number. Just add it from
18302 * here to the next \n */
18305 while (*(si_string + len) != '\n' && remaining > 0) {
18309 if (*(si_string + len) == '\n') {
18313 if (matches_string) {
18314 sv_catpvn(matches_string, si_string, len - 1);
18317 matches_string = newSVpvn(si_string, len - 1);
18320 sv_catpvs(matches_string, " ");
18321 } /* end of loop through the text */
18323 assert(matches_string);
18324 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
18325 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
18327 } /* end of has an 'si' but no swash */
18330 /* If we have a swash in place, its equivalent inversion list was above
18331 * placed into 'invlist'. If not, this variable may contain a stored
18332 * inversion list which is information beyond what is in 'si' */
18335 /* Again, if the caller doesn't want the output inversion list, put
18336 * everything in 'matches-string' */
18337 if (! output_invlist) {
18338 if ( ! matches_string) {
18339 matches_string = newSVpvs("\n");
18341 sv_catsv(matches_string, invlist_contents(invlist,
18342 TRUE /* traditional style */
18345 else if (! *output_invlist) {
18346 *output_invlist = invlist_clone(invlist);
18349 _invlist_union(*output_invlist, invlist, output_invlist);
18353 *listsvp = matches_string;
18358 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
18360 /* reg_skipcomment()
18362 Absorbs an /x style # comment from the input stream,
18363 returning a pointer to the first character beyond the comment, or if the
18364 comment terminates the pattern without anything following it, this returns
18365 one past the final character of the pattern (in other words, RExC_end) and
18366 sets the REG_RUN_ON_COMMENT_SEEN flag.
18368 Note it's the callers responsibility to ensure that we are
18369 actually in /x mode
18373 PERL_STATIC_INLINE char*
18374 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
18376 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
18380 while (p < RExC_end) {
18381 if (*(++p) == '\n') {
18386 /* we ran off the end of the pattern without ending the comment, so we have
18387 * to add an \n when wrapping */
18388 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
18393 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
18395 const bool force_to_xmod
18398 /* If the text at the current parse position '*p' is a '(?#...)' comment,
18399 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
18400 * is /x whitespace, advance '*p' so that on exit it points to the first
18401 * byte past all such white space and comments */
18403 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
18405 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
18407 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
18410 if (RExC_end - (*p) >= 3
18412 && *(*p + 1) == '?'
18413 && *(*p + 2) == '#')
18415 while (*(*p) != ')') {
18416 if ((*p) == RExC_end)
18417 FAIL("Sequence (?#... not terminated");
18425 const char * save_p = *p;
18426 while ((*p) < RExC_end) {
18428 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
18431 else if (*(*p) == '#') {
18432 (*p) = reg_skipcomment(pRExC_state, (*p));
18438 if (*p != save_p) {
18451 Advances the parse position by one byte, unless that byte is the beginning
18452 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
18453 those two cases, the parse position is advanced beyond all such comments and
18456 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
18460 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
18462 PERL_ARGS_ASSERT_NEXTCHAR;
18464 if (RExC_parse < RExC_end) {
18466 || UTF8_IS_INVARIANT(*RExC_parse)
18467 || UTF8_IS_START(*RExC_parse));
18469 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
18471 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
18472 FALSE /* Don't force /x */ );
18477 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
18479 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
18480 * space. In pass1, it aligns and increments RExC_size; in pass2,
18483 regnode * const ret = RExC_emit;
18484 GET_RE_DEBUG_FLAGS_DECL;
18486 PERL_ARGS_ASSERT_REGNODE_GUTS;
18488 assert(extra_size >= regarglen[op]);
18491 SIZE_ALIGN(RExC_size);
18492 RExC_size += 1 + extra_size;
18495 if (RExC_emit >= RExC_emit_bound)
18496 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
18497 op, (void*)RExC_emit, (void*)RExC_emit_bound);
18499 NODE_ALIGN_FILL(ret);
18500 #ifndef RE_TRACK_PATTERN_OFFSETS
18501 PERL_UNUSED_ARG(name);
18503 if (RExC_offsets) { /* MJD */
18505 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
18508 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
18509 ? "Overwriting end of array!\n" : "OK",
18510 (UV)(RExC_emit - RExC_emit_start),
18511 (UV)(RExC_parse - RExC_start),
18512 (UV)RExC_offsets[0]));
18513 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
18520 - reg_node - emit a node
18522 STATIC regnode * /* Location. */
18523 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
18525 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
18527 PERL_ARGS_ASSERT_REG_NODE;
18529 assert(regarglen[op] == 0);
18532 regnode *ptr = ret;
18533 FILL_ADVANCE_NODE(ptr, op);
18540 - reganode - emit a node with an argument
18542 STATIC regnode * /* Location. */
18543 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
18545 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
18547 PERL_ARGS_ASSERT_REGANODE;
18549 assert(regarglen[op] == 1);
18552 regnode *ptr = ret;
18553 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
18560 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
18562 /* emit a node with U32 and I32 arguments */
18564 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
18566 PERL_ARGS_ASSERT_REG2LANODE;
18568 assert(regarglen[op] == 2);
18571 regnode *ptr = ret;
18572 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
18579 - reginsert - insert an operator in front of already-emitted operand
18581 * Means relocating the operand.
18583 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
18584 * set up NEXT_OFF() of the inserted node if needed. Something like this:
18586 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
18588 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
18592 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *operand, U32 depth)
18597 const int offset = regarglen[(U8)op];
18598 const int size = NODE_STEP_REGNODE + offset;
18599 GET_RE_DEBUG_FLAGS_DECL;
18601 PERL_ARGS_ASSERT_REGINSERT;
18602 PERL_UNUSED_CONTEXT;
18603 PERL_UNUSED_ARG(depth);
18604 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
18605 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
18610 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
18611 studying. If this is wrong then we need to adjust RExC_recurse
18612 below like we do with RExC_open_parens/RExC_close_parens. */
18616 if (RExC_open_parens) {
18618 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
18619 /* remember that RExC_npar is rex->nparens + 1,
18620 * iow it is 1 more than the number of parens seen in
18621 * the pattern so far. */
18622 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
18623 /* note, RExC_open_parens[0] is the start of the
18624 * regex, it can't move. RExC_close_parens[0] is the end
18625 * of the regex, it *can* move. */
18626 if ( paren && RExC_open_parens[paren] >= operand ) {
18627 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
18628 RExC_open_parens[paren] += size;
18630 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
18632 if ( RExC_close_parens[paren] >= operand ) {
18633 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
18634 RExC_close_parens[paren] += size;
18636 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
18641 RExC_end_op += size;
18643 while (src > operand) {
18644 StructCopy(--src, --dst, regnode);
18645 #ifdef RE_TRACK_PATTERN_OFFSETS
18646 if (RExC_offsets) { /* MJD 20010112 */
18648 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
18652 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
18653 ? "Overwriting end of array!\n" : "OK",
18654 (UV)(src - RExC_emit_start),
18655 (UV)(dst - RExC_emit_start),
18656 (UV)RExC_offsets[0]));
18657 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
18658 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
18664 place = operand; /* Op node, where operand used to be. */
18665 #ifdef RE_TRACK_PATTERN_OFFSETS
18666 if (RExC_offsets) { /* MJD */
18668 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
18672 (UV)(place - RExC_emit_start) > RExC_offsets[0]
18673 ? "Overwriting end of array!\n" : "OK",
18674 (UV)(place - RExC_emit_start),
18675 (UV)(RExC_parse - RExC_start),
18676 (UV)RExC_offsets[0]));
18677 Set_Node_Offset(place, RExC_parse);
18678 Set_Node_Length(place, 1);
18681 src = NEXTOPER(place);
18682 FILL_ADVANCE_NODE(place, op);
18683 Zero(src, offset, regnode);
18687 - regtail - set the next-pointer at the end of a node chain of p to val.
18688 - SEE ALSO: regtail_study
18691 S_regtail(pTHX_ RExC_state_t * pRExC_state,
18692 const regnode * const p,
18693 const regnode * const val,
18697 GET_RE_DEBUG_FLAGS_DECL;
18699 PERL_ARGS_ASSERT_REGTAIL;
18701 PERL_UNUSED_ARG(depth);
18707 /* Find last node. */
18708 scan = (regnode *) p;
18710 regnode * const temp = regnext(scan);
18712 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
18713 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18714 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
18715 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
18716 (temp == NULL ? "->" : ""),
18717 (temp == NULL ? PL_reg_name[OP(val)] : "")
18725 if (reg_off_by_arg[OP(scan)]) {
18726 ARG_SET(scan, val - scan);
18729 NEXT_OFF(scan) = val - scan;
18735 - regtail_study - set the next-pointer at the end of a node chain of p to val.
18736 - Look for optimizable sequences at the same time.
18737 - currently only looks for EXACT chains.
18739 This is experimental code. The idea is to use this routine to perform
18740 in place optimizations on branches and groups as they are constructed,
18741 with the long term intention of removing optimization from study_chunk so
18742 that it is purely analytical.
18744 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
18745 to control which is which.
18748 /* TODO: All four parms should be const */
18751 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
18752 const regnode *val,U32 depth)
18756 #ifdef EXPERIMENTAL_INPLACESCAN
18759 GET_RE_DEBUG_FLAGS_DECL;
18761 PERL_ARGS_ASSERT_REGTAIL_STUDY;
18767 /* Find last node. */
18771 regnode * const temp = regnext(scan);
18772 #ifdef EXPERIMENTAL_INPLACESCAN
18773 if (PL_regkind[OP(scan)] == EXACT) {
18774 bool unfolded_multi_char; /* Unexamined in this routine */
18775 if (join_exact(pRExC_state, scan, &min,
18776 &unfolded_multi_char, 1, val, depth+1))
18781 switch (OP(scan)) {
18785 case EXACTFA_NO_TRIE:
18791 if( exact == PSEUDO )
18793 else if ( exact != OP(scan) )
18802 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
18803 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18804 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
18805 SvPV_nolen_const(RExC_mysv),
18806 REG_NODE_NUM(scan),
18807 PL_reg_name[exact]);
18814 DEBUG_PARSE_MSG("");
18815 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
18816 Perl_re_printf( aTHX_
18817 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
18818 SvPV_nolen_const(RExC_mysv),
18819 (IV)REG_NODE_NUM(val),
18823 if (reg_off_by_arg[OP(scan)]) {
18824 ARG_SET(scan, val - scan);
18827 NEXT_OFF(scan) = val - scan;
18835 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
18840 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
18845 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18847 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
18848 if (flags & (1<<bit)) {
18849 if (!set++ && lead)
18850 Perl_re_printf( aTHX_ "%s",lead);
18851 Perl_re_printf( aTHX_ "%s ",PL_reg_intflags_name[bit]);
18856 Perl_re_printf( aTHX_ "\n");
18858 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18863 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
18869 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18871 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
18872 if (flags & (1<<bit)) {
18873 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
18876 if (!set++ && lead)
18877 Perl_re_printf( aTHX_ "%s",lead);
18878 Perl_re_printf( aTHX_ "%s ",PL_reg_extflags_name[bit]);
18881 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
18882 if (!set++ && lead) {
18883 Perl_re_printf( aTHX_ "%s",lead);
18886 case REGEX_UNICODE_CHARSET:
18887 Perl_re_printf( aTHX_ "UNICODE");
18889 case REGEX_LOCALE_CHARSET:
18890 Perl_re_printf( aTHX_ "LOCALE");
18892 case REGEX_ASCII_RESTRICTED_CHARSET:
18893 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
18895 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
18896 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
18899 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
18905 Perl_re_printf( aTHX_ "\n");
18907 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18913 Perl_regdump(pTHX_ const regexp *r)
18916 SV * const sv = sv_newmortal();
18917 SV *dsv= sv_newmortal();
18918 RXi_GET_DECL(r,ri);
18919 GET_RE_DEBUG_FLAGS_DECL;
18921 PERL_ARGS_ASSERT_REGDUMP;
18923 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
18925 /* Header fields of interest. */
18926 if (r->anchored_substr) {
18927 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
18928 RE_SV_DUMPLEN(r->anchored_substr), 30);
18929 Perl_re_printf( aTHX_
18930 "anchored %s%s at %" IVdf " ",
18931 s, RE_SV_TAIL(r->anchored_substr),
18932 (IV)r->anchored_offset);
18933 } else if (r->anchored_utf8) {
18934 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
18935 RE_SV_DUMPLEN(r->anchored_utf8), 30);
18936 Perl_re_printf( aTHX_
18937 "anchored utf8 %s%s at %" IVdf " ",
18938 s, RE_SV_TAIL(r->anchored_utf8),
18939 (IV)r->anchored_offset);
18941 if (r->float_substr) {
18942 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
18943 RE_SV_DUMPLEN(r->float_substr), 30);
18944 Perl_re_printf( aTHX_
18945 "floating %s%s at %" IVdf "..%" UVuf " ",
18946 s, RE_SV_TAIL(r->float_substr),
18947 (IV)r->float_min_offset, (UV)r->float_max_offset);
18948 } else if (r->float_utf8) {
18949 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
18950 RE_SV_DUMPLEN(r->float_utf8), 30);
18951 Perl_re_printf( aTHX_
18952 "floating utf8 %s%s at %" IVdf "..%" UVuf " ",
18953 s, RE_SV_TAIL(r->float_utf8),
18954 (IV)r->float_min_offset, (UV)r->float_max_offset);
18956 if (r->check_substr || r->check_utf8)
18957 Perl_re_printf( aTHX_
18959 (r->check_substr == r->float_substr
18960 && r->check_utf8 == r->float_utf8
18961 ? "(checking floating" : "(checking anchored"));
18962 if (r->intflags & PREGf_NOSCAN)
18963 Perl_re_printf( aTHX_ " noscan");
18964 if (r->extflags & RXf_CHECK_ALL)
18965 Perl_re_printf( aTHX_ " isall");
18966 if (r->check_substr || r->check_utf8)
18967 Perl_re_printf( aTHX_ ") ");
18969 if (ri->regstclass) {
18970 regprop(r, sv, ri->regstclass, NULL, NULL);
18971 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
18973 if (r->intflags & PREGf_ANCH) {
18974 Perl_re_printf( aTHX_ "anchored");
18975 if (r->intflags & PREGf_ANCH_MBOL)
18976 Perl_re_printf( aTHX_ "(MBOL)");
18977 if (r->intflags & PREGf_ANCH_SBOL)
18978 Perl_re_printf( aTHX_ "(SBOL)");
18979 if (r->intflags & PREGf_ANCH_GPOS)
18980 Perl_re_printf( aTHX_ "(GPOS)");
18981 Perl_re_printf( aTHX_ " ");
18983 if (r->intflags & PREGf_GPOS_SEEN)
18984 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
18985 if (r->intflags & PREGf_SKIP)
18986 Perl_re_printf( aTHX_ "plus ");
18987 if (r->intflags & PREGf_IMPLICIT)
18988 Perl_re_printf( aTHX_ "implicit ");
18989 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
18990 if (r->extflags & RXf_EVAL_SEEN)
18991 Perl_re_printf( aTHX_ "with eval ");
18992 Perl_re_printf( aTHX_ "\n");
18994 regdump_extflags("r->extflags: ",r->extflags);
18995 regdump_intflags("r->intflags: ",r->intflags);
18998 PERL_ARGS_ASSERT_REGDUMP;
18999 PERL_UNUSED_CONTEXT;
19000 PERL_UNUSED_ARG(r);
19001 #endif /* DEBUGGING */
19004 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
19007 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
19008 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
19009 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
19010 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
19011 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
19012 || _CC_VERTSPACE != 15
19013 # error Need to adjust order of anyofs[]
19015 static const char * const anyofs[] = {
19052 - regprop - printable representation of opcode, with run time support
19056 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
19060 RXi_GET_DECL(prog,progi);
19061 GET_RE_DEBUG_FLAGS_DECL;
19063 PERL_ARGS_ASSERT_REGPROP;
19067 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
19068 /* It would be nice to FAIL() here, but this may be called from
19069 regexec.c, and it would be hard to supply pRExC_state. */
19070 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
19071 (int)OP(o), (int)REGNODE_MAX);
19072 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
19074 k = PL_regkind[OP(o)];
19077 sv_catpvs(sv, " ");
19078 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
19079 * is a crude hack but it may be the best for now since
19080 * we have no flag "this EXACTish node was UTF-8"
19082 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
19083 PERL_PV_ESCAPE_UNI_DETECT |
19084 PERL_PV_ESCAPE_NONASCII |
19085 PERL_PV_PRETTY_ELLIPSES |
19086 PERL_PV_PRETTY_LTGT |
19087 PERL_PV_PRETTY_NOCLEAR
19089 } else if (k == TRIE) {
19090 /* print the details of the trie in dumpuntil instead, as
19091 * progi->data isn't available here */
19092 const char op = OP(o);
19093 const U32 n = ARG(o);
19094 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
19095 (reg_ac_data *)progi->data->data[n] :
19097 const reg_trie_data * const trie
19098 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
19100 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
19101 DEBUG_TRIE_COMPILE_r({
19103 sv_catpvs(sv, "(JUMP)");
19104 Perl_sv_catpvf(aTHX_ sv,
19105 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
19106 (UV)trie->startstate,
19107 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
19108 (UV)trie->wordcount,
19111 (UV)TRIE_CHARCOUNT(trie),
19112 (UV)trie->uniquecharcount
19115 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
19116 sv_catpvs(sv, "[");
19117 (void) put_charclass_bitmap_innards(sv,
19118 ((IS_ANYOF_TRIE(op))
19120 : TRIE_BITMAP(trie)),
19126 sv_catpvs(sv, "]");
19128 } else if (k == CURLY) {
19129 U32 lo = ARG1(o), hi = ARG2(o);
19130 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
19131 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
19132 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
19133 if (hi == REG_INFTY)
19134 sv_catpvs(sv, "INFTY");
19136 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
19137 sv_catpvs(sv, "}");
19139 else if (k == WHILEM && o->flags) /* Ordinal/of */
19140 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
19141 else if (k == REF || k == OPEN || k == CLOSE
19142 || k == GROUPP || OP(o)==ACCEPT)
19144 AV *name_list= NULL;
19145 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
19146 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
19147 if ( RXp_PAREN_NAMES(prog) ) {
19148 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19149 } else if ( pRExC_state ) {
19150 name_list= RExC_paren_name_list;
19153 if ( k != REF || (OP(o) < NREF)) {
19154 SV **name= av_fetch(name_list, parno, 0 );
19156 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19159 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
19160 I32 *nums=(I32*)SvPVX(sv_dat);
19161 SV **name= av_fetch(name_list, nums[0], 0 );
19164 for ( n=0; n<SvIVX(sv_dat); n++ ) {
19165 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
19166 (n ? "," : ""), (IV)nums[n]);
19168 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19172 if ( k == REF && reginfo) {
19173 U32 n = ARG(o); /* which paren pair */
19174 I32 ln = prog->offs[n].start;
19175 if (prog->lastparen < n || ln == -1)
19176 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
19177 else if (ln == prog->offs[n].end)
19178 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
19180 const char *s = reginfo->strbeg + ln;
19181 Perl_sv_catpvf(aTHX_ sv, ": ");
19182 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
19183 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
19186 } else if (k == GOSUB) {
19187 AV *name_list= NULL;
19188 if ( RXp_PAREN_NAMES(prog) ) {
19189 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19190 } else if ( pRExC_state ) {
19191 name_list= RExC_paren_name_list;
19194 /* Paren and offset */
19195 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
19196 (int)((o + (int)ARG2L(o)) - progi->program) );
19198 SV **name= av_fetch(name_list, ARG(o), 0 );
19200 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19203 else if (k == LOGICAL)
19204 /* 2: embedded, otherwise 1 */
19205 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
19206 else if (k == ANYOF) {
19207 const U8 flags = ANYOF_FLAGS(o);
19208 bool do_sep = FALSE; /* Do we need to separate various components of
19210 /* Set if there is still an unresolved user-defined property */
19211 SV *unresolved = NULL;
19213 /* Things that are ignored except when the runtime locale is UTF-8 */
19214 SV *only_utf8_locale_invlist = NULL;
19216 /* Code points that don't fit in the bitmap */
19217 SV *nonbitmap_invlist = NULL;
19219 /* And things that aren't in the bitmap, but are small enough to be */
19220 SV* bitmap_range_not_in_bitmap = NULL;
19222 const bool inverted = flags & ANYOF_INVERT;
19224 if (OP(o) == ANYOFL) {
19225 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
19226 sv_catpvs(sv, "{utf8-locale-reqd}");
19228 if (flags & ANYOFL_FOLD) {
19229 sv_catpvs(sv, "{i}");
19233 /* If there is stuff outside the bitmap, get it */
19234 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
19235 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
19237 &only_utf8_locale_invlist,
19238 &nonbitmap_invlist);
19239 /* The non-bitmap data may contain stuff that could fit in the
19240 * bitmap. This could come from a user-defined property being
19241 * finally resolved when this call was done; or much more likely
19242 * because there are matches that require UTF-8 to be valid, and so
19243 * aren't in the bitmap. This is teased apart later */
19244 _invlist_intersection(nonbitmap_invlist,
19246 &bitmap_range_not_in_bitmap);
19247 /* Leave just the things that don't fit into the bitmap */
19248 _invlist_subtract(nonbitmap_invlist,
19250 &nonbitmap_invlist);
19253 /* Obey this flag to add all above-the-bitmap code points */
19254 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
19255 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
19256 NUM_ANYOF_CODE_POINTS,
19260 /* Ready to start outputting. First, the initial left bracket */
19261 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
19263 /* Then all the things that could fit in the bitmap */
19264 do_sep = put_charclass_bitmap_innards(sv,
19266 bitmap_range_not_in_bitmap,
19267 only_utf8_locale_invlist,
19270 /* Can't try inverting for a
19271 * better display if there are
19272 * things that haven't been
19274 unresolved != NULL);
19275 SvREFCNT_dec(bitmap_range_not_in_bitmap);
19277 /* If there are user-defined properties which haven't been defined yet,
19278 * output them. If the result is not to be inverted, it is clearest to
19279 * output them in a separate [] from the bitmap range stuff. If the
19280 * result is to be complemented, we have to show everything in one [],
19281 * as the inversion applies to the whole thing. Use {braces} to
19282 * separate them from anything in the bitmap and anything above the
19286 if (! do_sep) { /* If didn't output anything in the bitmap */
19287 sv_catpvs(sv, "^");
19289 sv_catpvs(sv, "{");
19292 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19294 sv_catsv(sv, unresolved);
19296 sv_catpvs(sv, "}");
19298 do_sep = ! inverted;
19301 /* And, finally, add the above-the-bitmap stuff */
19302 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
19305 /* See if truncation size is overridden */
19306 const STRLEN dump_len = (PL_dump_re_max_len)
19307 ? PL_dump_re_max_len
19310 /* This is output in a separate [] */
19312 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19315 /* And, for easy of understanding, it is shown in the
19316 * uncomplemented form if possible. The one exception being if
19317 * there are unresolved items, where the inversion has to be
19318 * delayed until runtime */
19319 if (inverted && ! unresolved) {
19320 _invlist_invert(nonbitmap_invlist);
19321 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
19324 contents = invlist_contents(nonbitmap_invlist,
19325 FALSE /* output suitable for catsv */
19328 /* If the output is shorter than the permissible maximum, just do it. */
19329 if (SvCUR(contents) <= dump_len) {
19330 sv_catsv(sv, contents);
19333 const char * contents_string = SvPVX(contents);
19334 STRLEN i = dump_len;
19336 /* Otherwise, start at the permissible max and work back to the
19337 * first break possibility */
19338 while (i > 0 && contents_string[i] != ' ') {
19341 if (i == 0) { /* Fail-safe. Use the max if we couldn't
19342 find a legal break */
19346 sv_catpvn(sv, contents_string, i);
19347 sv_catpvs(sv, "...");
19350 SvREFCNT_dec_NN(contents);
19351 SvREFCNT_dec_NN(nonbitmap_invlist);
19354 /* And finally the matching, closing ']' */
19355 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
19357 SvREFCNT_dec(unresolved);
19359 else if (k == POSIXD || k == NPOSIXD) {
19360 U8 index = FLAGS(o) * 2;
19361 if (index < C_ARRAY_LENGTH(anyofs)) {
19362 if (*anyofs[index] != '[') {
19365 sv_catpv(sv, anyofs[index]);
19366 if (*anyofs[index] != '[') {
19371 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
19374 else if (k == BOUND || k == NBOUND) {
19375 /* Must be synced with order of 'bound_type' in regcomp.h */
19376 const char * const bounds[] = {
19377 "", /* Traditional */
19383 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
19384 sv_catpv(sv, bounds[FLAGS(o)]);
19386 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
19387 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
19388 else if (OP(o) == SBOL)
19389 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
19391 /* add on the verb argument if there is one */
19392 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
19393 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
19394 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
19397 PERL_UNUSED_CONTEXT;
19398 PERL_UNUSED_ARG(sv);
19399 PERL_UNUSED_ARG(o);
19400 PERL_UNUSED_ARG(prog);
19401 PERL_UNUSED_ARG(reginfo);
19402 PERL_UNUSED_ARG(pRExC_state);
19403 #endif /* DEBUGGING */
19409 Perl_re_intuit_string(pTHX_ REGEXP * const r)
19410 { /* Assume that RE_INTUIT is set */
19411 struct regexp *const prog = ReANY(r);
19412 GET_RE_DEBUG_FLAGS_DECL;
19414 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
19415 PERL_UNUSED_CONTEXT;
19419 const char * const s = SvPV_nolen_const(RX_UTF8(r)
19420 ? prog->check_utf8 : prog->check_substr);
19422 if (!PL_colorset) reginitcolors();
19423 Perl_re_printf( aTHX_
19424 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
19426 RX_UTF8(r) ? "utf8 " : "",
19427 PL_colors[5],PL_colors[0],
19430 (strlen(s) > 60 ? "..." : ""));
19433 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
19434 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
19440 handles refcounting and freeing the perl core regexp structure. When
19441 it is necessary to actually free the structure the first thing it
19442 does is call the 'free' method of the regexp_engine associated to
19443 the regexp, allowing the handling of the void *pprivate; member
19444 first. (This routine is not overridable by extensions, which is why
19445 the extensions free is called first.)
19447 See regdupe and regdupe_internal if you change anything here.
19449 #ifndef PERL_IN_XSUB_RE
19451 Perl_pregfree(pTHX_ REGEXP *r)
19457 Perl_pregfree2(pTHX_ REGEXP *rx)
19459 struct regexp *const r = ReANY(rx);
19460 GET_RE_DEBUG_FLAGS_DECL;
19462 PERL_ARGS_ASSERT_PREGFREE2;
19464 if (r->mother_re) {
19465 ReREFCNT_dec(r->mother_re);
19467 CALLREGFREE_PVT(rx); /* free the private data */
19468 SvREFCNT_dec(RXp_PAREN_NAMES(r));
19469 Safefree(r->xpv_len_u.xpvlenu_pv);
19472 SvREFCNT_dec(r->anchored_substr);
19473 SvREFCNT_dec(r->anchored_utf8);
19474 SvREFCNT_dec(r->float_substr);
19475 SvREFCNT_dec(r->float_utf8);
19476 Safefree(r->substrs);
19478 RX_MATCH_COPY_FREE(rx);
19479 #ifdef PERL_ANY_COW
19480 SvREFCNT_dec(r->saved_copy);
19483 SvREFCNT_dec(r->qr_anoncv);
19484 if (r->recurse_locinput)
19485 Safefree(r->recurse_locinput);
19486 rx->sv_u.svu_rx = 0;
19491 This is a hacky workaround to the structural issue of match results
19492 being stored in the regexp structure which is in turn stored in
19493 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
19494 could be PL_curpm in multiple contexts, and could require multiple
19495 result sets being associated with the pattern simultaneously, such
19496 as when doing a recursive match with (??{$qr})
19498 The solution is to make a lightweight copy of the regexp structure
19499 when a qr// is returned from the code executed by (??{$qr}) this
19500 lightweight copy doesn't actually own any of its data except for
19501 the starp/end and the actual regexp structure itself.
19507 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
19509 struct regexp *ret;
19510 struct regexp *const r = ReANY(rx);
19511 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
19513 PERL_ARGS_ASSERT_REG_TEMP_COPY;
19516 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
19518 SvOK_off((SV *)ret_x);
19520 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
19521 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
19522 made both spots point to the same regexp body.) */
19523 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
19524 assert(!SvPVX(ret_x));
19525 ret_x->sv_u.svu_rx = temp->sv_any;
19526 temp->sv_any = NULL;
19527 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
19528 SvREFCNT_dec_NN(temp);
19529 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
19530 ing below will not set it. */
19531 SvCUR_set(ret_x, SvCUR(rx));
19534 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
19535 sv_force_normal(sv) is called. */
19537 ret = ReANY(ret_x);
19539 SvFLAGS(ret_x) |= SvUTF8(rx);
19540 /* We share the same string buffer as the original regexp, on which we
19541 hold a reference count, incremented when mother_re is set below.
19542 The string pointer is copied here, being part of the regexp struct.
19544 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
19545 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
19547 const I32 npar = r->nparens+1;
19548 Newx(ret->offs, npar, regexp_paren_pair);
19549 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19552 Newx(ret->substrs, 1, struct reg_substr_data);
19553 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19555 SvREFCNT_inc_void(ret->anchored_substr);
19556 SvREFCNT_inc_void(ret->anchored_utf8);
19557 SvREFCNT_inc_void(ret->float_substr);
19558 SvREFCNT_inc_void(ret->float_utf8);
19560 /* check_substr and check_utf8, if non-NULL, point to either their
19561 anchored or float namesakes, and don't hold a second reference. */
19563 RX_MATCH_COPIED_off(ret_x);
19564 #ifdef PERL_ANY_COW
19565 ret->saved_copy = NULL;
19567 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
19568 SvREFCNT_inc_void(ret->qr_anoncv);
19569 if (r->recurse_locinput)
19570 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19576 /* regfree_internal()
19578 Free the private data in a regexp. This is overloadable by
19579 extensions. Perl takes care of the regexp structure in pregfree(),
19580 this covers the *pprivate pointer which technically perl doesn't
19581 know about, however of course we have to handle the
19582 regexp_internal structure when no extension is in use.
19584 Note this is called before freeing anything in the regexp
19589 Perl_regfree_internal(pTHX_ REGEXP * const rx)
19591 struct regexp *const r = ReANY(rx);
19592 RXi_GET_DECL(r,ri);
19593 GET_RE_DEBUG_FLAGS_DECL;
19595 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
19601 SV *dsv= sv_newmortal();
19602 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
19603 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
19604 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
19605 PL_colors[4],PL_colors[5],s);
19608 #ifdef RE_TRACK_PATTERN_OFFSETS
19610 Safefree(ri->u.offsets); /* 20010421 MJD */
19612 if (ri->code_blocks)
19613 S_free_codeblocks(aTHX_ ri->code_blocks);
19616 int n = ri->data->count;
19619 /* If you add a ->what type here, update the comment in regcomp.h */
19620 switch (ri->data->what[n]) {
19626 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
19629 Safefree(ri->data->data[n]);
19635 { /* Aho Corasick add-on structure for a trie node.
19636 Used in stclass optimization only */
19638 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
19639 #ifdef USE_ITHREADS
19643 refcount = --aho->refcount;
19646 PerlMemShared_free(aho->states);
19647 PerlMemShared_free(aho->fail);
19648 /* do this last!!!! */
19649 PerlMemShared_free(ri->data->data[n]);
19650 /* we should only ever get called once, so
19651 * assert as much, and also guard the free
19652 * which /might/ happen twice. At the least
19653 * it will make code anlyzers happy and it
19654 * doesn't cost much. - Yves */
19655 assert(ri->regstclass);
19656 if (ri->regstclass) {
19657 PerlMemShared_free(ri->regstclass);
19658 ri->regstclass = 0;
19665 /* trie structure. */
19667 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
19668 #ifdef USE_ITHREADS
19672 refcount = --trie->refcount;
19675 PerlMemShared_free(trie->charmap);
19676 PerlMemShared_free(trie->states);
19677 PerlMemShared_free(trie->trans);
19679 PerlMemShared_free(trie->bitmap);
19681 PerlMemShared_free(trie->jump);
19682 PerlMemShared_free(trie->wordinfo);
19683 /* do this last!!!! */
19684 PerlMemShared_free(ri->data->data[n]);
19689 Perl_croak(aTHX_ "panic: regfree data code '%c'",
19690 ri->data->what[n]);
19693 Safefree(ri->data->what);
19694 Safefree(ri->data);
19700 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
19701 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
19702 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
19705 re_dup_guts - duplicate a regexp.
19707 This routine is expected to clone a given regexp structure. It is only
19708 compiled under USE_ITHREADS.
19710 After all of the core data stored in struct regexp is duplicated
19711 the regexp_engine.dupe method is used to copy any private data
19712 stored in the *pprivate pointer. This allows extensions to handle
19713 any duplication it needs to do.
19715 See pregfree() and regfree_internal() if you change anything here.
19717 #if defined(USE_ITHREADS)
19718 #ifndef PERL_IN_XSUB_RE
19720 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
19724 const struct regexp *r = ReANY(sstr);
19725 struct regexp *ret = ReANY(dstr);
19727 PERL_ARGS_ASSERT_RE_DUP_GUTS;
19729 npar = r->nparens+1;
19730 Newx(ret->offs, npar, regexp_paren_pair);
19731 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19733 if (ret->substrs) {
19734 /* Do it this way to avoid reading from *r after the StructCopy().
19735 That way, if any of the sv_dup_inc()s dislodge *r from the L1
19736 cache, it doesn't matter. */
19737 const bool anchored = r->check_substr
19738 ? r->check_substr == r->anchored_substr
19739 : r->check_utf8 == r->anchored_utf8;
19740 Newx(ret->substrs, 1, struct reg_substr_data);
19741 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19743 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
19744 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
19745 ret->float_substr = sv_dup_inc(ret->float_substr, param);
19746 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
19748 /* check_substr and check_utf8, if non-NULL, point to either their
19749 anchored or float namesakes, and don't hold a second reference. */
19751 if (ret->check_substr) {
19753 assert(r->check_utf8 == r->anchored_utf8);
19754 ret->check_substr = ret->anchored_substr;
19755 ret->check_utf8 = ret->anchored_utf8;
19757 assert(r->check_substr == r->float_substr);
19758 assert(r->check_utf8 == r->float_utf8);
19759 ret->check_substr = ret->float_substr;
19760 ret->check_utf8 = ret->float_utf8;
19762 } else if (ret->check_utf8) {
19764 ret->check_utf8 = ret->anchored_utf8;
19766 ret->check_utf8 = ret->float_utf8;
19771 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
19772 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
19773 if (r->recurse_locinput)
19774 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19777 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
19779 if (RX_MATCH_COPIED(dstr))
19780 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
19782 ret->subbeg = NULL;
19783 #ifdef PERL_ANY_COW
19784 ret->saved_copy = NULL;
19787 /* Whether mother_re be set or no, we need to copy the string. We
19788 cannot refrain from copying it when the storage points directly to
19789 our mother regexp, because that's
19790 1: a buffer in a different thread
19791 2: something we no longer hold a reference on
19792 so we need to copy it locally. */
19793 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
19794 ret->mother_re = NULL;
19796 #endif /* PERL_IN_XSUB_RE */
19801 This is the internal complement to regdupe() which is used to copy
19802 the structure pointed to by the *pprivate pointer in the regexp.
19803 This is the core version of the extension overridable cloning hook.
19804 The regexp structure being duplicated will be copied by perl prior
19805 to this and will be provided as the regexp *r argument, however
19806 with the /old/ structures pprivate pointer value. Thus this routine
19807 may override any copying normally done by perl.
19809 It returns a pointer to the new regexp_internal structure.
19813 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
19816 struct regexp *const r = ReANY(rx);
19817 regexp_internal *reti;
19819 RXi_GET_DECL(r,ri);
19821 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
19825 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
19826 char, regexp_internal);
19827 Copy(ri->program, reti->program, len+1, regnode);
19830 if (ri->code_blocks) {
19832 Newx(reti->code_blocks, 1, struct reg_code_blocks);
19833 Newx(reti->code_blocks->cb, ri->code_blocks->count,
19834 struct reg_code_block);
19835 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
19836 ri->code_blocks->count, struct reg_code_block);
19837 for (n = 0; n < ri->code_blocks->count; n++)
19838 reti->code_blocks->cb[n].src_regex = (REGEXP*)
19839 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
19840 reti->code_blocks->count = ri->code_blocks->count;
19841 reti->code_blocks->refcnt = 1;
19844 reti->code_blocks = NULL;
19846 reti->regstclass = NULL;
19849 struct reg_data *d;
19850 const int count = ri->data->count;
19853 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
19854 char, struct reg_data);
19855 Newx(d->what, count, U8);
19858 for (i = 0; i < count; i++) {
19859 d->what[i] = ri->data->what[i];
19860 switch (d->what[i]) {
19861 /* see also regcomp.h and regfree_internal() */
19862 case 'a': /* actually an AV, but the dup function is identical. */
19866 case 'u': /* actually an HV, but the dup function is identical. */
19867 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
19870 /* This is cheating. */
19871 Newx(d->data[i], 1, regnode_ssc);
19872 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
19873 reti->regstclass = (regnode*)d->data[i];
19876 /* Trie stclasses are readonly and can thus be shared
19877 * without duplication. We free the stclass in pregfree
19878 * when the corresponding reg_ac_data struct is freed.
19880 reti->regstclass= ri->regstclass;
19884 ((reg_trie_data*)ri->data->data[i])->refcount++;
19889 d->data[i] = ri->data->data[i];
19892 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
19893 ri->data->what[i]);
19902 reti->name_list_idx = ri->name_list_idx;
19904 #ifdef RE_TRACK_PATTERN_OFFSETS
19905 if (ri->u.offsets) {
19906 Newx(reti->u.offsets, 2*len+1, U32);
19907 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
19910 SetProgLen(reti,len);
19913 return (void*)reti;
19916 #endif /* USE_ITHREADS */
19918 #ifndef PERL_IN_XSUB_RE
19921 - regnext - dig the "next" pointer out of a node
19924 Perl_regnext(pTHX_ regnode *p)
19931 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
19932 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
19933 (int)OP(p), (int)REGNODE_MAX);
19936 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
19945 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
19948 STRLEN l1 = strlen(pat1);
19949 STRLEN l2 = strlen(pat2);
19952 const char *message;
19954 PERL_ARGS_ASSERT_RE_CROAK2;
19960 Copy(pat1, buf, l1 , char);
19961 Copy(pat2, buf + l1, l2 , char);
19962 buf[l1 + l2] = '\n';
19963 buf[l1 + l2 + 1] = '\0';
19964 va_start(args, pat2);
19965 msv = vmess(buf, &args);
19967 message = SvPV_const(msv,l1);
19970 Copy(message, buf, l1 , char);
19971 /* l1-1 to avoid \n */
19972 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, l1-1, buf));
19975 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
19977 #ifndef PERL_IN_XSUB_RE
19979 Perl_save_re_context(pTHX)
19984 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
19987 const REGEXP * const rx = PM_GETRE(PL_curpm);
19989 nparens = RX_NPARENS(rx);
19992 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
19993 * that PL_curpm will be null, but that utf8.pm and the modules it
19994 * loads will only use $1..$3.
19995 * The t/porting/re_context.t test file checks this assumption.
20000 for (i = 1; i <= nparens; i++) {
20001 char digits[TYPE_CHARS(long)];
20002 const STRLEN len = my_snprintf(digits, sizeof(digits),
20004 GV *const *const gvp
20005 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
20008 GV * const gv = *gvp;
20009 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
20019 S_put_code_point(pTHX_ SV *sv, UV c)
20021 PERL_ARGS_ASSERT_PUT_CODE_POINT;
20024 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
20026 else if (isPRINT(c)) {
20027 const char string = (char) c;
20029 /* We use {phrase} as metanotation in the class, so also escape literal
20031 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
20032 sv_catpvs(sv, "\\");
20033 sv_catpvn(sv, &string, 1);
20035 else if (isMNEMONIC_CNTRL(c)) {
20036 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
20039 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
20043 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
20046 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
20048 /* Appends to 'sv' a displayable version of the range of code points from
20049 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
20050 * that have them, when they occur at the beginning or end of the range.
20051 * It uses hex to output the remaining code points, unless 'allow_literals'
20052 * is true, in which case the printable ASCII ones are output as-is (though
20053 * some of these will be escaped by put_code_point()).
20055 * NOTE: This is designed only for printing ranges of code points that fit
20056 * inside an ANYOF bitmap. Higher code points are simply suppressed
20059 const unsigned int min_range_count = 3;
20061 assert(start <= end);
20063 PERL_ARGS_ASSERT_PUT_RANGE;
20065 while (start <= end) {
20067 const char * format;
20069 if (end - start < min_range_count) {
20071 /* Output chars individually when they occur in short ranges */
20072 for (; start <= end; start++) {
20073 put_code_point(sv, start);
20078 /* If permitted by the input options, and there is a possibility that
20079 * this range contains a printable literal, look to see if there is
20081 if (allow_literals && start <= MAX_PRINT_A) {
20083 /* If the character at the beginning of the range isn't an ASCII
20084 * printable, effectively split the range into two parts:
20085 * 1) the portion before the first such printable,
20087 * and output them separately. */
20088 if (! isPRINT_A(start)) {
20089 UV temp_end = start + 1;
20091 /* There is no point looking beyond the final possible
20092 * printable, in MAX_PRINT_A */
20093 UV max = MIN(end, MAX_PRINT_A);
20095 while (temp_end <= max && ! isPRINT_A(temp_end)) {
20099 /* Here, temp_end points to one beyond the first printable if
20100 * found, or to one beyond 'max' if not. If none found, make
20101 * sure that we use the entire range */
20102 if (temp_end > MAX_PRINT_A) {
20103 temp_end = end + 1;
20106 /* Output the first part of the split range: the part that
20107 * doesn't have printables, with the parameter set to not look
20108 * for literals (otherwise we would infinitely recurse) */
20109 put_range(sv, start, temp_end - 1, FALSE);
20111 /* The 2nd part of the range (if any) starts here. */
20114 /* We do a continue, instead of dropping down, because even if
20115 * the 2nd part is non-empty, it could be so short that we want
20116 * to output it as individual characters, as tested for at the
20117 * top of this loop. */
20121 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
20122 * output a sub-range of just the digits or letters, then process
20123 * the remaining portion as usual. */
20124 if (isALPHANUMERIC_A(start)) {
20125 UV mask = (isDIGIT_A(start))
20130 UV temp_end = start + 1;
20132 /* Find the end of the sub-range that includes just the
20133 * characters in the same class as the first character in it */
20134 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
20139 /* For short ranges, don't duplicate the code above to output
20140 * them; just call recursively */
20141 if (temp_end - start < min_range_count) {
20142 put_range(sv, start, temp_end, FALSE);
20144 else { /* Output as a range */
20145 put_code_point(sv, start);
20146 sv_catpvs(sv, "-");
20147 put_code_point(sv, temp_end);
20149 start = temp_end + 1;
20153 /* We output any other printables as individual characters */
20154 if (isPUNCT_A(start) || isSPACE_A(start)) {
20155 while (start <= end && (isPUNCT_A(start)
20156 || isSPACE_A(start)))
20158 put_code_point(sv, start);
20163 } /* End of looking for literals */
20165 /* Here is not to output as a literal. Some control characters have
20166 * mnemonic names. Split off any of those at the beginning and end of
20167 * the range to print mnemonically. It isn't possible for many of
20168 * these to be in a row, so this won't overwhelm with output */
20170 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
20172 while (isMNEMONIC_CNTRL(start) && start <= end) {
20173 put_code_point(sv, start);
20177 /* If this didn't take care of the whole range ... */
20178 if (start <= end) {
20180 /* Look backwards from the end to find the final non-mnemonic
20183 while (isMNEMONIC_CNTRL(temp_end)) {
20187 /* And separately output the interior range that doesn't start
20188 * or end with mnemonics */
20189 put_range(sv, start, temp_end, FALSE);
20191 /* Then output the mnemonic trailing controls */
20192 start = temp_end + 1;
20193 while (start <= end) {
20194 put_code_point(sv, start);
20201 /* As a final resort, output the range or subrange as hex. */
20203 this_end = (end < NUM_ANYOF_CODE_POINTS)
20205 : NUM_ANYOF_CODE_POINTS - 1;
20206 #if NUM_ANYOF_CODE_POINTS > 256
20207 format = (this_end < 256)
20208 ? "\\x%02" UVXf "-\\x%02" UVXf
20209 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
20211 format = "\\x%02" UVXf "-\\x%02" UVXf;
20213 GCC_DIAG_IGNORE(-Wformat-nonliteral);
20214 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
20221 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
20223 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
20227 bool allow_literals = TRUE;
20229 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
20231 /* Generally, it is more readable if printable characters are output as
20232 * literals, but if a range (nearly) spans all of them, it's best to output
20233 * it as a single range. This code will use a single range if all but 2
20234 * ASCII printables are in it */
20235 invlist_iterinit(invlist);
20236 while (invlist_iternext(invlist, &start, &end)) {
20238 /* If the range starts beyond the final printable, it doesn't have any
20240 if (start > MAX_PRINT_A) {
20244 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
20245 * all but two, the range must start and end no later than 2 from
20247 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
20248 if (end > MAX_PRINT_A) {
20254 if (end - start >= MAX_PRINT_A - ' ' - 2) {
20255 allow_literals = FALSE;
20260 invlist_iterfinish(invlist);
20262 /* Here we have figured things out. Output each range */
20263 invlist_iterinit(invlist);
20264 while (invlist_iternext(invlist, &start, &end)) {
20265 if (start >= NUM_ANYOF_CODE_POINTS) {
20268 put_range(sv, start, end, allow_literals);
20270 invlist_iterfinish(invlist);
20276 S_put_charclass_bitmap_innards_common(pTHX_
20277 SV* invlist, /* The bitmap */
20278 SV* posixes, /* Under /l, things like [:word:], \S */
20279 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
20280 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
20281 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
20282 const bool invert /* Is the result to be inverted? */
20285 /* Create and return an SV containing a displayable version of the bitmap
20286 * and associated information determined by the input parameters. If the
20287 * output would have been only the inversion indicator '^', NULL is instead
20292 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
20295 output = newSVpvs("^");
20298 output = newSVpvs("");
20301 /* First, the code points in the bitmap that are unconditionally there */
20302 put_charclass_bitmap_innards_invlist(output, invlist);
20304 /* Traditionally, these have been placed after the main code points */
20306 sv_catsv(output, posixes);
20309 if (only_utf8 && _invlist_len(only_utf8)) {
20310 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
20311 put_charclass_bitmap_innards_invlist(output, only_utf8);
20314 if (not_utf8 && _invlist_len(not_utf8)) {
20315 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
20316 put_charclass_bitmap_innards_invlist(output, not_utf8);
20319 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
20320 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
20321 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
20323 /* This is the only list in this routine that can legally contain code
20324 * points outside the bitmap range. The call just above to
20325 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
20326 * output them here. There's about a half-dozen possible, and none in
20327 * contiguous ranges longer than 2 */
20328 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20330 SV* above_bitmap = NULL;
20332 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
20334 invlist_iterinit(above_bitmap);
20335 while (invlist_iternext(above_bitmap, &start, &end)) {
20338 for (i = start; i <= end; i++) {
20339 put_code_point(output, i);
20342 invlist_iterfinish(above_bitmap);
20343 SvREFCNT_dec_NN(above_bitmap);
20347 if (invert && SvCUR(output) == 1) {
20355 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
20357 SV *nonbitmap_invlist,
20358 SV *only_utf8_locale_invlist,
20359 const regnode * const node,
20360 const bool force_as_is_display)
20362 /* Appends to 'sv' a displayable version of the innards of the bracketed
20363 * character class defined by the other arguments:
20364 * 'bitmap' points to the bitmap.
20365 * 'nonbitmap_invlist' is an inversion list of the code points that are in
20366 * the bitmap range, but for some reason aren't in the bitmap; NULL if
20367 * none. The reasons for this could be that they require some
20368 * condition such as the target string being or not being in UTF-8
20369 * (under /d), or because they came from a user-defined property that
20370 * was not resolved at the time of the regex compilation (under /u)
20371 * 'only_utf8_locale_invlist' is an inversion list of the code points that
20372 * are valid only if the runtime locale is a UTF-8 one; NULL if none
20373 * 'node' is the regex pattern node. It is needed only when the above two
20374 * parameters are not null, and is passed so that this routine can
20375 * tease apart the various reasons for them.
20376 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
20377 * to invert things to see if that leads to a cleaner display. If
20378 * FALSE, this routine is free to use its judgment about doing this.
20380 * It returns TRUE if there was actually something output. (It may be that
20381 * the bitmap, etc is empty.)
20383 * When called for outputting the bitmap of a non-ANYOF node, just pass the
20384 * bitmap, with the succeeding parameters set to NULL, and the final one to
20388 /* In general, it tries to display the 'cleanest' representation of the
20389 * innards, choosing whether to display them inverted or not, regardless of
20390 * whether the class itself is to be inverted. However, there are some
20391 * cases where it can't try inverting, as what actually matches isn't known
20392 * until runtime, and hence the inversion isn't either. */
20393 bool inverting_allowed = ! force_as_is_display;
20396 STRLEN orig_sv_cur = SvCUR(sv);
20398 SV* invlist; /* Inversion list we accumulate of code points that
20399 are unconditionally matched */
20400 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
20402 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
20404 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
20405 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
20408 SV* as_is_display; /* The output string when we take the inputs
20410 SV* inverted_display; /* The output string when we invert the inputs */
20412 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
20414 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
20416 /* We are biased in favor of displaying things without them being inverted,
20417 * as that is generally easier to understand */
20418 const int bias = 5;
20420 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
20422 /* Start off with whatever code points are passed in. (We clone, so we
20423 * don't change the caller's list) */
20424 if (nonbitmap_invlist) {
20425 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
20426 invlist = invlist_clone(nonbitmap_invlist);
20428 else { /* Worst case size is every other code point is matched */
20429 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
20433 if (OP(node) == ANYOFD) {
20435 /* This flag indicates that the code points below 0x100 in the
20436 * nonbitmap list are precisely the ones that match only when the
20437 * target is UTF-8 (they should all be non-ASCII). */
20438 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
20440 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
20441 _invlist_subtract(invlist, only_utf8, &invlist);
20444 /* And this flag for matching all non-ASCII 0xFF and below */
20445 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
20447 not_utf8 = invlist_clone(PL_UpperLatin1);
20450 else if (OP(node) == ANYOFL) {
20452 /* If either of these flags are set, what matches isn't
20453 * determinable except during execution, so don't know enough here
20455 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
20456 inverting_allowed = FALSE;
20459 /* What the posix classes match also varies at runtime, so these
20460 * will be output symbolically. */
20461 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
20464 posixes = newSVpvs("");
20465 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
20466 if (ANYOF_POSIXL_TEST(node,i)) {
20467 sv_catpv(posixes, anyofs[i]);
20474 /* Accumulate the bit map into the unconditional match list */
20475 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
20476 if (BITMAP_TEST(bitmap, i)) {
20478 for (; i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i); i++) {
20481 invlist = _add_range_to_invlist(invlist, start, i-1);
20485 /* Make sure that the conditional match lists don't have anything in them
20486 * that match unconditionally; otherwise the output is quite confusing.
20487 * This could happen if the code that populates these misses some
20490 _invlist_subtract(only_utf8, invlist, &only_utf8);
20493 _invlist_subtract(not_utf8, invlist, ¬_utf8);
20496 if (only_utf8_locale_invlist) {
20498 /* Since this list is passed in, we have to make a copy before
20500 only_utf8_locale = invlist_clone(only_utf8_locale_invlist);
20502 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
20504 /* And, it can get really weird for us to try outputting an inverted
20505 * form of this list when it has things above the bitmap, so don't even
20507 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20508 inverting_allowed = FALSE;
20512 /* Calculate what the output would be if we take the input as-is */
20513 as_is_display = put_charclass_bitmap_innards_common(invlist,
20520 /* If have to take the output as-is, just do that */
20521 if (! inverting_allowed) {
20522 if (as_is_display) {
20523 sv_catsv(sv, as_is_display);
20524 SvREFCNT_dec_NN(as_is_display);
20527 else { /* But otherwise, create the output again on the inverted input, and
20528 use whichever version is shorter */
20530 int inverted_bias, as_is_bias;
20532 /* We will apply our bias to whichever of the the results doesn't have
20542 inverted_bias = bias;
20545 /* Now invert each of the lists that contribute to the output,
20546 * excluding from the result things outside the possible range */
20548 /* For the unconditional inversion list, we have to add in all the
20549 * conditional code points, so that when inverted, they will be gone
20551 _invlist_union(only_utf8, invlist, &invlist);
20552 _invlist_union(not_utf8, invlist, &invlist);
20553 _invlist_union(only_utf8_locale, invlist, &invlist);
20554 _invlist_invert(invlist);
20555 _invlist_intersection(invlist, PL_InBitmap, &invlist);
20558 _invlist_invert(only_utf8);
20559 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
20561 else if (not_utf8) {
20563 /* If a code point matches iff the target string is not in UTF-8,
20564 * then complementing the result has it not match iff not in UTF-8,
20565 * which is the same thing as matching iff it is UTF-8. */
20566 only_utf8 = not_utf8;
20570 if (only_utf8_locale) {
20571 _invlist_invert(only_utf8_locale);
20572 _invlist_intersection(only_utf8_locale,
20574 &only_utf8_locale);
20577 inverted_display = put_charclass_bitmap_innards_common(
20582 only_utf8_locale, invert);
20584 /* Use the shortest representation, taking into account our bias
20585 * against showing it inverted */
20586 if ( inverted_display
20587 && ( ! as_is_display
20588 || ( SvCUR(inverted_display) + inverted_bias
20589 < SvCUR(as_is_display) + as_is_bias)))
20591 sv_catsv(sv, inverted_display);
20593 else if (as_is_display) {
20594 sv_catsv(sv, as_is_display);
20597 SvREFCNT_dec(as_is_display);
20598 SvREFCNT_dec(inverted_display);
20601 SvREFCNT_dec_NN(invlist);
20602 SvREFCNT_dec(only_utf8);
20603 SvREFCNT_dec(not_utf8);
20604 SvREFCNT_dec(posixes);
20605 SvREFCNT_dec(only_utf8_locale);
20607 return SvCUR(sv) > orig_sv_cur;
20610 #define CLEAR_OPTSTART \
20611 if (optstart) STMT_START { \
20612 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
20613 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
20617 #define DUMPUNTIL(b,e) \
20619 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
20621 STATIC const regnode *
20622 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
20623 const regnode *last, const regnode *plast,
20624 SV* sv, I32 indent, U32 depth)
20626 U8 op = PSEUDO; /* Arbitrary non-END op. */
20627 const regnode *next;
20628 const regnode *optstart= NULL;
20630 RXi_GET_DECL(r,ri);
20631 GET_RE_DEBUG_FLAGS_DECL;
20633 PERL_ARGS_ASSERT_DUMPUNTIL;
20635 #ifdef DEBUG_DUMPUNTIL
20636 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n",indent,node-start,
20637 last ? last-start : 0,plast ? plast-start : 0);
20640 if (plast && plast < last)
20643 while (PL_regkind[op] != END && (!last || node < last)) {
20645 /* While that wasn't END last time... */
20648 if (op == CLOSE || op == WHILEM)
20650 next = regnext((regnode *)node);
20653 if (OP(node) == OPTIMIZED) {
20654 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
20661 regprop(r, sv, node, NULL, NULL);
20662 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
20663 (int)(2*indent + 1), "", SvPVX_const(sv));
20665 if (OP(node) != OPTIMIZED) {
20666 if (next == NULL) /* Next ptr. */
20667 Perl_re_printf( aTHX_ " (0)");
20668 else if (PL_regkind[(U8)op] == BRANCH
20669 && PL_regkind[OP(next)] != BRANCH )
20670 Perl_re_printf( aTHX_ " (FAIL)");
20672 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
20673 Perl_re_printf( aTHX_ "\n");
20677 if (PL_regkind[(U8)op] == BRANCHJ) {
20680 const regnode *nnode = (OP(next) == LONGJMP
20681 ? regnext((regnode *)next)
20683 if (last && nnode > last)
20685 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
20688 else if (PL_regkind[(U8)op] == BRANCH) {
20690 DUMPUNTIL(NEXTOPER(node), next);
20692 else if ( PL_regkind[(U8)op] == TRIE ) {
20693 const regnode *this_trie = node;
20694 const char op = OP(node);
20695 const U32 n = ARG(node);
20696 const reg_ac_data * const ac = op>=AHOCORASICK ?
20697 (reg_ac_data *)ri->data->data[n] :
20699 const reg_trie_data * const trie =
20700 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
20702 AV *const trie_words
20703 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
20705 const regnode *nextbranch= NULL;
20708 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
20709 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
20711 Perl_re_indentf( aTHX_ "%s ",
20714 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
20715 SvCUR(*elem_ptr), 60,
20716 PL_colors[0], PL_colors[1],
20718 ? PERL_PV_ESCAPE_UNI
20720 | PERL_PV_PRETTY_ELLIPSES
20721 | PERL_PV_PRETTY_LTGT
20726 U16 dist= trie->jump[word_idx+1];
20727 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
20728 (UV)((dist ? this_trie + dist : next) - start));
20731 nextbranch= this_trie + trie->jump[0];
20732 DUMPUNTIL(this_trie + dist, nextbranch);
20734 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
20735 nextbranch= regnext((regnode *)nextbranch);
20737 Perl_re_printf( aTHX_ "\n");
20740 if (last && next > last)
20745 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
20746 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
20747 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
20749 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
20751 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
20753 else if ( op == PLUS || op == STAR) {
20754 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
20756 else if (PL_regkind[(U8)op] == ANYOF) {
20757 /* arglen 1 + class block */
20758 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
20759 ? ANYOF_POSIXL_SKIP
20761 node = NEXTOPER(node);
20763 else if (PL_regkind[(U8)op] == EXACT) {
20764 /* Literal string, where present. */
20765 node += NODE_SZ_STR(node) - 1;
20766 node = NEXTOPER(node);
20769 node = NEXTOPER(node);
20770 node += regarglen[(U8)op];
20772 if (op == CURLYX || op == OPEN)
20776 #ifdef DEBUG_DUMPUNTIL
20777 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
20782 #endif /* DEBUGGING */
20785 * ex: set ts=8 sts=4 sw=4 et: