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) \
123 ((c) == '-' || (c) == ']' || (c) == '\\' || (c) == '^')
126 struct RExC_state_t {
127 U32 flags; /* RXf_* are we folding, multilining? */
128 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
129 char *precomp; /* uncompiled string. */
130 char *precomp_end; /* pointer to end of uncompiled string. */
131 REGEXP *rx_sv; /* The SV that is the regexp. */
132 regexp *rx; /* perl core regexp structure */
133 regexp_internal *rxi; /* internal data for regexp object
135 char *start; /* Start of input for compile */
136 char *end; /* End of input for compile */
137 char *parse; /* Input-scan pointer. */
138 char *adjusted_start; /* 'start', adjusted. See code use */
139 STRLEN precomp_adj; /* an offset beyond precomp. See code use */
140 SSize_t whilem_seen; /* number of WHILEM in this expr */
141 regnode *emit_start; /* Start of emitted-code area */
142 regnode *emit_bound; /* First regnode outside of the
144 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
145 implies compiling, so don't emit */
146 regnode_ssc emit_dummy; /* placeholder for emit to point to;
147 large enough for the largest
148 non-EXACTish node, so can use it as
150 I32 naughty; /* How bad is this pattern? */
151 I32 sawback; /* Did we see \1, ...? */
153 SSize_t size; /* Code size. */
154 I32 npar; /* Capture buffer count, (OPEN) plus
155 one. ("par" 0 is the whole
157 I32 nestroot; /* root parens we are in - used by
161 regnode **open_parens; /* pointers to open parens */
162 regnode **close_parens; /* pointers to close parens */
163 regnode *end_op; /* END node in program */
164 I32 utf8; /* whether the pattern is utf8 or not */
165 I32 orig_utf8; /* whether the pattern was originally in utf8 */
166 /* XXX use this for future optimisation of case
167 * where pattern must be upgraded to utf8. */
168 I32 uni_semantics; /* If a d charset modifier should use unicode
169 rules, even if the pattern is not in
171 HV *paren_names; /* Paren names */
173 regnode **recurse; /* Recurse regops */
174 I32 recurse_count; /* Number of recurse regops we have generated */
175 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
177 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
180 I32 override_recoding;
182 I32 recode_x_to_native;
184 I32 in_multi_char_class;
185 struct reg_code_block *code_blocks; /* positions of literal (?{})
187 int num_code_blocks; /* size of code_blocks[] */
188 int code_index; /* next code_blocks[] slot */
189 SSize_t maxlen; /* mininum possible number of chars in string to match */
190 scan_frame *frame_head;
191 scan_frame *frame_last;
194 #ifdef ADD_TO_REGEXEC
195 char *starttry; /* -Dr: where regtry was called. */
196 #define RExC_starttry (pRExC_state->starttry)
198 SV *runtime_code_qr; /* qr with the runtime code blocks */
200 const char *lastparse;
202 AV *paren_name_list; /* idx -> name */
203 U32 study_chunk_recursed_count;
206 #define RExC_lastparse (pRExC_state->lastparse)
207 #define RExC_lastnum (pRExC_state->lastnum)
208 #define RExC_paren_name_list (pRExC_state->paren_name_list)
209 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
210 #define RExC_mysv (pRExC_state->mysv1)
211 #define RExC_mysv1 (pRExC_state->mysv1)
212 #define RExC_mysv2 (pRExC_state->mysv2)
215 bool seen_unfolded_sharp_s;
220 #define RExC_flags (pRExC_state->flags)
221 #define RExC_pm_flags (pRExC_state->pm_flags)
222 #define RExC_precomp (pRExC_state->precomp)
223 #define RExC_precomp_adj (pRExC_state->precomp_adj)
224 #define RExC_adjusted_start (pRExC_state->adjusted_start)
225 #define RExC_precomp_end (pRExC_state->precomp_end)
226 #define RExC_rx_sv (pRExC_state->rx_sv)
227 #define RExC_rx (pRExC_state->rx)
228 #define RExC_rxi (pRExC_state->rxi)
229 #define RExC_start (pRExC_state->start)
230 #define RExC_end (pRExC_state->end)
231 #define RExC_parse (pRExC_state->parse)
232 #define RExC_whilem_seen (pRExC_state->whilem_seen)
234 /* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
235 * EXACTF node, hence was parsed under /di rules. If later in the parse,
236 * something forces the pattern into using /ui rules, the sharp s should be
237 * folded into the sequence 'ss', which takes up more space than previously
238 * calculated. This means that the sizing pass needs to be restarted. (The
239 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
240 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
241 * so there is no need to resize [perl #125990]. */
242 #define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
244 #ifdef RE_TRACK_PATTERN_OFFSETS
245 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
248 #define RExC_emit (pRExC_state->emit)
249 #define RExC_emit_dummy (pRExC_state->emit_dummy)
250 #define RExC_emit_start (pRExC_state->emit_start)
251 #define RExC_emit_bound (pRExC_state->emit_bound)
252 #define RExC_sawback (pRExC_state->sawback)
253 #define RExC_seen (pRExC_state->seen)
254 #define RExC_size (pRExC_state->size)
255 #define RExC_maxlen (pRExC_state->maxlen)
256 #define RExC_npar (pRExC_state->npar)
257 #define RExC_nestroot (pRExC_state->nestroot)
258 #define RExC_extralen (pRExC_state->extralen)
259 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
260 #define RExC_utf8 (pRExC_state->utf8)
261 #define RExC_uni_semantics (pRExC_state->uni_semantics)
262 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
263 #define RExC_open_parens (pRExC_state->open_parens)
264 #define RExC_close_parens (pRExC_state->close_parens)
265 #define RExC_end_op (pRExC_state->end_op)
266 #define RExC_paren_names (pRExC_state->paren_names)
267 #define RExC_recurse (pRExC_state->recurse)
268 #define RExC_recurse_count (pRExC_state->recurse_count)
269 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
270 #define RExC_study_chunk_recursed_bytes \
271 (pRExC_state->study_chunk_recursed_bytes)
272 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
273 #define RExC_contains_locale (pRExC_state->contains_locale)
275 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
277 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
278 #define RExC_frame_head (pRExC_state->frame_head)
279 #define RExC_frame_last (pRExC_state->frame_last)
280 #define RExC_frame_count (pRExC_state->frame_count)
281 #define RExC_strict (pRExC_state->strict)
282 #define RExC_study_started (pRExC_state->study_started)
283 #define RExC_warn_text (pRExC_state->warn_text)
285 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
286 * a flag to disable back-off on the fixed/floating substrings - if it's
287 * a high complexity pattern we assume the benefit of avoiding a full match
288 * is worth the cost of checking for the substrings even if they rarely help.
290 #define RExC_naughty (pRExC_state->naughty)
291 #define TOO_NAUGHTY (10)
292 #define MARK_NAUGHTY(add) \
293 if (RExC_naughty < TOO_NAUGHTY) \
294 RExC_naughty += (add)
295 #define MARK_NAUGHTY_EXP(exp, add) \
296 if (RExC_naughty < TOO_NAUGHTY) \
297 RExC_naughty += RExC_naughty / (exp) + (add)
299 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
300 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
301 ((*s) == '{' && regcurly(s)))
304 * Flags to be passed up and down.
306 #define WORST 0 /* Worst case. */
307 #define HASWIDTH 0x01 /* Known to match non-null strings. */
309 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
310 * character. (There needs to be a case: in the switch statement in regexec.c
311 * for any node marked SIMPLE.) Note that this is not the same thing as
314 #define SPSTART 0x04 /* Starts with * or + */
315 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
316 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
317 #define RESTART_PASS1 0x20 /* Need to restart sizing pass */
318 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
319 calcuate sizes as UTF-8 */
321 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
323 /* whether trie related optimizations are enabled */
324 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
325 #define TRIE_STUDY_OPT
326 #define FULL_TRIE_STUDY
332 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
333 #define PBITVAL(paren) (1 << ((paren) & 7))
334 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
335 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
336 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
338 #define REQUIRE_UTF8(flagp) STMT_START { \
341 *flagp = RESTART_PASS1|NEED_UTF8; \
346 /* Change from /d into /u rules, and restart the parse if we've already seen
347 * something whose size would increase as a result, by setting *flagp and
348 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
349 * we've change to /u during the parse. */
350 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
352 if (DEPENDS_SEMANTICS) { \
354 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
355 RExC_uni_semantics = 1; \
356 if (RExC_seen_unfolded_sharp_s) { \
357 *flagp |= RESTART_PASS1; \
358 return restart_retval; \
363 /* This converts the named class defined in regcomp.h to its equivalent class
364 * number defined in handy.h. */
365 #define namedclass_to_classnum(class) ((int) ((class) / 2))
366 #define classnum_to_namedclass(classnum) ((classnum) * 2)
368 #define _invlist_union_complement_2nd(a, b, output) \
369 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
370 #define _invlist_intersection_complement_2nd(a, b, output) \
371 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
373 /* About scan_data_t.
375 During optimisation we recurse through the regexp program performing
376 various inplace (keyhole style) optimisations. In addition study_chunk
377 and scan_commit populate this data structure with information about
378 what strings MUST appear in the pattern. We look for the longest
379 string that must appear at a fixed location, and we look for the
380 longest string that may appear at a floating location. So for instance
385 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
386 strings (because they follow a .* construct). study_chunk will identify
387 both FOO and BAR as being the longest fixed and floating strings respectively.
389 The strings can be composites, for instance
393 will result in a composite fixed substring 'foo'.
395 For each string some basic information is maintained:
397 - offset or min_offset
398 This is the position the string must appear at, or not before.
399 It also implicitly (when combined with minlenp) tells us how many
400 characters must match before the string we are searching for.
401 Likewise when combined with minlenp and the length of the string it
402 tells us how many characters must appear after the string we have
406 Only used for floating strings. This is the rightmost point that
407 the string can appear at. If set to SSize_t_MAX it indicates that the
408 string can occur infinitely far to the right.
411 A pointer to the minimum number of characters of the pattern that the
412 string was found inside. This is important as in the case of positive
413 lookahead or positive lookbehind we can have multiple patterns
418 The minimum length of the pattern overall is 3, the minimum length
419 of the lookahead part is 3, but the minimum length of the part that
420 will actually match is 1. So 'FOO's minimum length is 3, but the
421 minimum length for the F is 1. This is important as the minimum length
422 is used to determine offsets in front of and behind the string being
423 looked for. Since strings can be composites this is the length of the
424 pattern at the time it was committed with a scan_commit. Note that
425 the length is calculated by study_chunk, so that the minimum lengths
426 are not known until the full pattern has been compiled, thus the
427 pointer to the value.
431 In the case of lookbehind the string being searched for can be
432 offset past the start point of the final matching string.
433 If this value was just blithely removed from the min_offset it would
434 invalidate some of the calculations for how many chars must match
435 before or after (as they are derived from min_offset and minlen and
436 the length of the string being searched for).
437 When the final pattern is compiled and the data is moved from the
438 scan_data_t structure into the regexp structure the information
439 about lookbehind is factored in, with the information that would
440 have been lost precalculated in the end_shift field for the
443 The fields pos_min and pos_delta are used to store the minimum offset
444 and the delta to the maximum offset at the current point in the pattern.
448 typedef struct scan_data_t {
449 /*I32 len_min; unused */
450 /*I32 len_delta; unused */
454 SSize_t last_end; /* min value, <0 unless valid. */
455 SSize_t last_start_min;
456 SSize_t last_start_max;
457 SV **longest; /* Either &l_fixed, or &l_float. */
458 SV *longest_fixed; /* longest fixed string found in pattern */
459 SSize_t offset_fixed; /* offset where it starts */
460 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
461 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
462 SV *longest_float; /* longest floating string found in pattern */
463 SSize_t offset_float_min; /* earliest point in string it can appear */
464 SSize_t offset_float_max; /* latest point in string it can appear */
465 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
466 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
469 SSize_t *last_closep;
470 regnode_ssc *start_class;
474 * Forward declarations for pregcomp()'s friends.
477 static const scan_data_t zero_scan_data =
478 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
480 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
481 #define SF_BEFORE_SEOL 0x0001
482 #define SF_BEFORE_MEOL 0x0002
483 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
484 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
486 #define SF_FIX_SHIFT_EOL (+2)
487 #define SF_FL_SHIFT_EOL (+4)
489 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
490 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
492 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
493 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
494 #define SF_IS_INF 0x0040
495 #define SF_HAS_PAR 0x0080
496 #define SF_IN_PAR 0x0100
497 #define SF_HAS_EVAL 0x0200
500 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
501 * longest substring in the pattern. When it is not set the optimiser keeps
502 * track of position, but does not keep track of the actual strings seen,
504 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
507 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
508 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
509 * turned off because of the alternation (BRANCH). */
510 #define SCF_DO_SUBSTR 0x0400
512 #define SCF_DO_STCLASS_AND 0x0800
513 #define SCF_DO_STCLASS_OR 0x1000
514 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
515 #define SCF_WHILEM_VISITED_POS 0x2000
517 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
518 #define SCF_SEEN_ACCEPT 0x8000
519 #define SCF_TRIE_DOING_RESTUDY 0x10000
520 #define SCF_IN_DEFINE 0x20000
525 #define UTF cBOOL(RExC_utf8)
527 /* The enums for all these are ordered so things work out correctly */
528 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
529 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
530 == REGEX_DEPENDS_CHARSET)
531 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
532 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
533 >= REGEX_UNICODE_CHARSET)
534 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
535 == REGEX_ASCII_RESTRICTED_CHARSET)
536 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
537 >= REGEX_ASCII_RESTRICTED_CHARSET)
538 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
539 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
541 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
543 /* For programs that want to be strictly Unicode compatible by dying if any
544 * attempt is made to match a non-Unicode code point against a Unicode
546 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
548 #define OOB_NAMEDCLASS -1
550 /* There is no code point that is out-of-bounds, so this is problematic. But
551 * its only current use is to initialize a variable that is always set before
553 #define OOB_UNICODE 0xDEADBEEF
555 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
558 /* length of regex to show in messages that don't mark a position within */
559 #define RegexLengthToShowInErrorMessages 127
562 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
563 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
564 * op/pragma/warn/regcomp.
566 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
567 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
569 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
570 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
572 /* The code in this file in places uses one level of recursion with parsing
573 * rebased to an alternate string constructed by us in memory. This can take
574 * the form of something that is completely different from the input, or
575 * something that uses the input as part of the alternate. In the first case,
576 * there should be no possibility of an error, as we are in complete control of
577 * the alternate string. But in the second case we don't control the input
578 * portion, so there may be errors in that. Here's an example:
580 * is handled specially because \x{df} folds to a sequence of more than one
581 * character, 'ss'. What is done is to create and parse an alternate string,
582 * which looks like this:
583 * /(?:\x{DF}|[abc\x{DF}def])/ui
584 * where it uses the input unchanged in the middle of something it constructs,
585 * which is a branch for the DF outside the character class, and clustering
586 * parens around the whole thing. (It knows enough to skip the DF inside the
587 * class while in this substitute parse.) 'abc' and 'def' may have errors that
588 * need to be reported. The general situation looks like this:
591 * Input: ----------------------------------------------------
592 * Constructed: ---------------------------------------------------
595 * The input string sI..eI is the input pattern. The string sC..EC is the
596 * constructed substitute parse string. The portions sC..tC and eC..EC are
597 * constructed by us. The portion tC..eC is an exact duplicate of the input
598 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
599 * while parsing, we find an error at xC. We want to display a message showing
600 * the real input string. Thus we need to find the point xI in it which
601 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
602 * been constructed by us, and so shouldn't have errors. We get:
604 * xI = sI + (tI - sI) + (xC - tC)
606 * and, the offset into sI is:
608 * (xI - sI) = (tI - sI) + (xC - tC)
610 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
611 * and we save tC as RExC_adjusted_start.
613 * During normal processing of the input pattern, everything points to that,
614 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
617 #define tI_sI RExC_precomp_adj
618 #define tC RExC_adjusted_start
619 #define sC RExC_precomp
620 #define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
621 #define xI(xC) (sC + xI_offset(xC))
622 #define eC RExC_precomp_end
624 #define REPORT_LOCATION_ARGS(xC) \
626 (xI(xC) > eC) /* Don't run off end */ \
627 ? eC - sC /* Length before the <--HERE */ \
629 sC), /* The input pattern printed up to the <--HERE */ \
631 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
632 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
634 /* Used to point after bad bytes for an error message, but avoid skipping
635 * past a nul byte. */
636 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
639 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
640 * arg. Show regex, up to a maximum length. If it's too long, chop and add
643 #define _FAIL(code) STMT_START { \
644 const char *ellipses = ""; \
645 IV len = RExC_precomp_end - RExC_precomp; \
648 SAVEFREESV(RExC_rx_sv); \
649 if (len > RegexLengthToShowInErrorMessages) { \
650 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
651 len = RegexLengthToShowInErrorMessages - 10; \
657 #define FAIL(msg) _FAIL( \
658 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
659 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
661 #define FAIL2(msg,arg) _FAIL( \
662 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
663 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
666 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
668 #define Simple_vFAIL(m) STMT_START { \
669 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
670 m, REPORT_LOCATION_ARGS(RExC_parse)); \
674 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
676 #define vFAIL(m) STMT_START { \
678 SAVEFREESV(RExC_rx_sv); \
683 * Like Simple_vFAIL(), but accepts two arguments.
685 #define Simple_vFAIL2(m,a1) STMT_START { \
686 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
687 REPORT_LOCATION_ARGS(RExC_parse)); \
691 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
693 #define vFAIL2(m,a1) STMT_START { \
695 SAVEFREESV(RExC_rx_sv); \
696 Simple_vFAIL2(m, a1); \
701 * Like Simple_vFAIL(), but accepts three arguments.
703 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
704 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
705 REPORT_LOCATION_ARGS(RExC_parse)); \
709 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
711 #define vFAIL3(m,a1,a2) STMT_START { \
713 SAVEFREESV(RExC_rx_sv); \
714 Simple_vFAIL3(m, a1, a2); \
718 * Like Simple_vFAIL(), but accepts four arguments.
720 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
721 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
722 REPORT_LOCATION_ARGS(RExC_parse)); \
725 #define vFAIL4(m,a1,a2,a3) STMT_START { \
727 SAVEFREESV(RExC_rx_sv); \
728 Simple_vFAIL4(m, a1, a2, a3); \
731 /* A specialized version of vFAIL2 that works with UTF8f */
732 #define vFAIL2utf8f(m, a1) STMT_START { \
734 SAVEFREESV(RExC_rx_sv); \
735 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
736 REPORT_LOCATION_ARGS(RExC_parse)); \
739 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
741 SAVEFREESV(RExC_rx_sv); \
742 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
743 REPORT_LOCATION_ARGS(RExC_parse)); \
746 /* These have asserts in them because of [perl #122671] Many warnings in
747 * regcomp.c can occur twice. If they get output in pass1 and later in that
748 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
749 * would get output again. So they should be output in pass2, and these
750 * asserts make sure new warnings follow that paradigm. */
752 /* m is not necessarily a "literal string", in this macro */
753 #define reg_warn_non_literal_string(loc, m) STMT_START { \
754 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
755 "%s" REPORT_LOCATION, \
756 m, REPORT_LOCATION_ARGS(loc)); \
759 #define ckWARNreg(loc,m) STMT_START { \
760 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
762 REPORT_LOCATION_ARGS(loc)); \
765 #define vWARN(loc, m) STMT_START { \
766 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
768 REPORT_LOCATION_ARGS(loc)); \
771 #define vWARN_dep(loc, m) STMT_START { \
772 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
774 REPORT_LOCATION_ARGS(loc)); \
777 #define ckWARNdep(loc,m) STMT_START { \
778 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
780 REPORT_LOCATION_ARGS(loc)); \
783 #define ckWARNregdep(loc,m) STMT_START { \
784 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
787 REPORT_LOCATION_ARGS(loc)); \
790 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
791 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
793 a1, REPORT_LOCATION_ARGS(loc)); \
796 #define ckWARN2reg(loc, m, a1) STMT_START { \
797 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
799 a1, REPORT_LOCATION_ARGS(loc)); \
802 #define vWARN3(loc, m, a1, a2) STMT_START { \
803 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
805 a1, a2, REPORT_LOCATION_ARGS(loc)); \
808 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
809 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
812 REPORT_LOCATION_ARGS(loc)); \
815 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
816 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
819 REPORT_LOCATION_ARGS(loc)); \
822 #define vWARN4dep(loc, m, a1, a2, a3) STMT_START { \
823 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN2(WARN_REGEXP,WARN_DEPRECATED), \
826 REPORT_LOCATION_ARGS(loc)); \
829 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
830 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
833 REPORT_LOCATION_ARGS(loc)); \
836 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
837 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
840 REPORT_LOCATION_ARGS(loc)); \
843 /* Macros for recording node offsets. 20001227 mjd@plover.com
844 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
845 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
846 * Element 0 holds the number n.
847 * Position is 1 indexed.
849 #ifndef RE_TRACK_PATTERN_OFFSETS
850 #define Set_Node_Offset_To_R(node,byte)
851 #define Set_Node_Offset(node,byte)
852 #define Set_Cur_Node_Offset
853 #define Set_Node_Length_To_R(node,len)
854 #define Set_Node_Length(node,len)
855 #define Set_Node_Cur_Length(node,start)
856 #define Node_Offset(n)
857 #define Node_Length(n)
858 #define Set_Node_Offset_Length(node,offset,len)
859 #define ProgLen(ri) ri->u.proglen
860 #define SetProgLen(ri,x) ri->u.proglen = x
862 #define ProgLen(ri) ri->u.offsets[0]
863 #define SetProgLen(ri,x) ri->u.offsets[0] = x
864 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
866 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
867 __LINE__, (int)(node), (int)(byte))); \
869 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
872 RExC_offsets[2*(node)-1] = (byte); \
877 #define Set_Node_Offset(node,byte) \
878 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
879 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
881 #define Set_Node_Length_To_R(node,len) STMT_START { \
883 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
884 __LINE__, (int)(node), (int)(len))); \
886 Perl_croak(aTHX_ "value of node is %d in Length macro", \
889 RExC_offsets[2*(node)] = (len); \
894 #define Set_Node_Length(node,len) \
895 Set_Node_Length_To_R((node)-RExC_emit_start, len)
896 #define Set_Node_Cur_Length(node, start) \
897 Set_Node_Length(node, RExC_parse - start)
899 /* Get offsets and lengths */
900 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
901 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
903 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
904 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
905 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
909 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
910 #define EXPERIMENTAL_INPLACESCAN
911 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
915 Perl_re_printf(pTHX_ const char *fmt, ...)
919 PerlIO *f= Perl_debug_log;
920 PERL_ARGS_ASSERT_RE_PRINTF;
922 result = PerlIO_vprintf(f, fmt, ap);
928 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
932 PerlIO *f= Perl_debug_log;
933 PERL_ARGS_ASSERT_RE_INDENTF;
935 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
936 result = PerlIO_vprintf(f, fmt, ap);
940 #endif /* DEBUGGING */
942 #define DEBUG_RExC_seen() \
943 DEBUG_OPTIMISE_MORE_r({ \
944 Perl_re_printf( aTHX_ "RExC_seen: "); \
946 if (RExC_seen & REG_ZERO_LEN_SEEN) \
947 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
949 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
950 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
952 if (RExC_seen & REG_GPOS_SEEN) \
953 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
955 if (RExC_seen & REG_RECURSE_SEEN) \
956 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
958 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
959 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
961 if (RExC_seen & REG_VERBARG_SEEN) \
962 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
964 if (RExC_seen & REG_CUTGROUP_SEEN) \
965 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
967 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
968 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
970 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
971 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
973 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
974 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
976 Perl_re_printf( aTHX_ "\n"); \
979 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
980 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
982 #define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
984 Perl_re_printf( aTHX_ "%s", open_str); \
985 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
986 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
987 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
988 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
989 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
990 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
991 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
992 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
993 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
994 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
995 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
996 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
997 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
998 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
999 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
1000 Perl_re_printf( aTHX_ "%s", close_str); \
1004 #define DEBUG_STUDYDATA(str,data,depth) \
1005 DEBUG_OPTIMISE_MORE_r(if(data){ \
1006 Perl_re_indentf( aTHX_ "" str "Pos:%" IVdf "/%" IVdf \
1007 " Flags: 0x%" UVXf, \
1009 (IV)((data)->pos_min), \
1010 (IV)((data)->pos_delta), \
1011 (UV)((data)->flags) \
1013 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
1014 Perl_re_printf( aTHX_ \
1015 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s", \
1016 (IV)((data)->whilem_c), \
1017 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
1018 is_inf ? "INF " : "" \
1020 if ((data)->last_found) \
1021 Perl_re_printf( aTHX_ \
1022 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf \
1023 " %sFixed:'%s' @ %" IVdf \
1024 " %sFloat: '%s' @ %" IVdf "/%" IVdf, \
1025 SvPVX_const((data)->last_found), \
1026 (IV)((data)->last_end), \
1027 (IV)((data)->last_start_min), \
1028 (IV)((data)->last_start_max), \
1029 ((data)->longest && \
1030 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
1031 SvPVX_const((data)->longest_fixed), \
1032 (IV)((data)->offset_fixed), \
1033 ((data)->longest && \
1034 (data)->longest==&((data)->longest_float)) ? "*" : "", \
1035 SvPVX_const((data)->longest_float), \
1036 (IV)((data)->offset_float_min), \
1037 (IV)((data)->offset_float_max) \
1039 Perl_re_printf( aTHX_ "\n"); \
1043 /* =========================================================
1044 * BEGIN edit_distance stuff.
1046 * This calculates how many single character changes of any type are needed to
1047 * transform a string into another one. It is taken from version 3.1 of
1049 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1052 /* Our unsorted dictionary linked list. */
1053 /* Note we use UVs, not chars. */
1058 struct dictionary* next;
1060 typedef struct dictionary item;
1063 PERL_STATIC_INLINE item*
1064 push(UV key,item* curr)
1067 Newxz(head, 1, item);
1075 PERL_STATIC_INLINE item*
1076 find(item* head, UV key)
1078 item* iterator = head;
1080 if (iterator->key == key){
1083 iterator = iterator->next;
1089 PERL_STATIC_INLINE item*
1090 uniquePush(item* head,UV key)
1092 item* iterator = head;
1095 if (iterator->key == key) {
1098 iterator = iterator->next;
1101 return push(key,head);
1104 PERL_STATIC_INLINE void
1105 dict_free(item* head)
1107 item* iterator = head;
1110 item* temp = iterator;
1111 iterator = iterator->next;
1118 /* End of Dictionary Stuff */
1120 /* All calculations/work are done here */
1122 S_edit_distance(const UV* src,
1124 const STRLEN x, /* length of src[] */
1125 const STRLEN y, /* length of tgt[] */
1126 const SSize_t maxDistance
1130 UV swapCount,swapScore,targetCharCount,i,j;
1132 UV score_ceil = x + y;
1134 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1136 /* intialize matrix start values */
1137 Newxz(scores, ( (x + 2) * (y + 2)), UV);
1138 scores[0] = score_ceil;
1139 scores[1 * (y + 2) + 0] = score_ceil;
1140 scores[0 * (y + 2) + 1] = score_ceil;
1141 scores[1 * (y + 2) + 1] = 0;
1142 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1147 for (i=1;i<=x;i++) {
1149 head = uniquePush(head,src[i]);
1150 scores[(i+1) * (y + 2) + 1] = i;
1151 scores[(i+1) * (y + 2) + 0] = score_ceil;
1154 for (j=1;j<=y;j++) {
1157 head = uniquePush(head,tgt[j]);
1158 scores[1 * (y + 2) + (j + 1)] = j;
1159 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1162 targetCharCount = find(head,tgt[j-1])->value;
1163 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1165 if (src[i-1] != tgt[j-1]){
1166 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));
1170 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1174 find(head,src[i-1])->value = i;
1178 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1181 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1185 /* END of edit_distance() stuff
1186 * ========================================================= */
1188 /* is c a control character for which we have a mnemonic? */
1189 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1192 S_cntrl_to_mnemonic(const U8 c)
1194 /* Returns the mnemonic string that represents character 'c', if one
1195 * exists; NULL otherwise. The only ones that exist for the purposes of
1196 * this routine are a few control characters */
1199 case '\a': return "\\a";
1200 case '\b': return "\\b";
1201 case ESC_NATIVE: return "\\e";
1202 case '\f': return "\\f";
1203 case '\n': return "\\n";
1204 case '\r': return "\\r";
1205 case '\t': return "\\t";
1211 /* Mark that we cannot extend a found fixed substring at this point.
1212 Update the longest found anchored substring and the longest found
1213 floating substrings if needed. */
1216 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1217 SSize_t *minlenp, int is_inf)
1219 const STRLEN l = CHR_SVLEN(data->last_found);
1220 const STRLEN old_l = CHR_SVLEN(*data->longest);
1221 GET_RE_DEBUG_FLAGS_DECL;
1223 PERL_ARGS_ASSERT_SCAN_COMMIT;
1225 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1226 SvSetMagicSV(*data->longest, data->last_found);
1227 if (*data->longest == data->longest_fixed) {
1228 data->offset_fixed = l ? data->last_start_min : data->pos_min;
1229 if (data->flags & SF_BEFORE_EOL)
1231 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
1233 data->flags &= ~SF_FIX_BEFORE_EOL;
1234 data->minlen_fixed=minlenp;
1235 data->lookbehind_fixed=0;
1237 else { /* *data->longest == data->longest_float */
1238 data->offset_float_min = l ? data->last_start_min : data->pos_min;
1239 data->offset_float_max = (l
1240 ? data->last_start_max
1241 : (data->pos_delta > SSize_t_MAX - data->pos_min
1243 : data->pos_min + data->pos_delta));
1245 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
1246 data->offset_float_max = SSize_t_MAX;
1247 if (data->flags & SF_BEFORE_EOL)
1249 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
1251 data->flags &= ~SF_FL_BEFORE_EOL;
1252 data->minlen_float=minlenp;
1253 data->lookbehind_float=0;
1256 SvCUR_set(data->last_found, 0);
1258 SV * const sv = data->last_found;
1259 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1260 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1265 data->last_end = -1;
1266 data->flags &= ~SF_BEFORE_EOL;
1267 DEBUG_STUDYDATA("commit: ",data,0);
1270 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1271 * list that describes which code points it matches */
1274 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1276 /* Set the SSC 'ssc' to match an empty string or any code point */
1278 PERL_ARGS_ASSERT_SSC_ANYTHING;
1280 assert(is_ANYOF_SYNTHETIC(ssc));
1282 /* mortalize so won't leak */
1283 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1284 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1288 S_ssc_is_anything(const regnode_ssc *ssc)
1290 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1291 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1292 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1293 * in any way, so there's no point in using it */
1298 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1300 assert(is_ANYOF_SYNTHETIC(ssc));
1302 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1306 /* See if the list consists solely of the range 0 - Infinity */
1307 invlist_iterinit(ssc->invlist);
1308 ret = invlist_iternext(ssc->invlist, &start, &end)
1312 invlist_iterfinish(ssc->invlist);
1318 /* If e.g., both \w and \W are set, matches everything */
1319 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1321 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1322 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1332 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1334 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1335 * string, any code point, or any posix class under locale */
1337 PERL_ARGS_ASSERT_SSC_INIT;
1339 Zero(ssc, 1, regnode_ssc);
1340 set_ANYOF_SYNTHETIC(ssc);
1341 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1344 /* If any portion of the regex is to operate under locale rules that aren't
1345 * fully known at compile time, initialization includes it. The reason
1346 * this isn't done for all regexes is that the optimizer was written under
1347 * the assumption that locale was all-or-nothing. Given the complexity and
1348 * lack of documentation in the optimizer, and that there are inadequate
1349 * test cases for locale, many parts of it may not work properly, it is
1350 * safest to avoid locale unless necessary. */
1351 if (RExC_contains_locale) {
1352 ANYOF_POSIXL_SETALL(ssc);
1355 ANYOF_POSIXL_ZERO(ssc);
1360 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1361 const regnode_ssc *ssc)
1363 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1364 * to the list of code points matched, and locale posix classes; hence does
1365 * not check its flags) */
1370 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1372 assert(is_ANYOF_SYNTHETIC(ssc));
1374 invlist_iterinit(ssc->invlist);
1375 ret = invlist_iternext(ssc->invlist, &start, &end)
1379 invlist_iterfinish(ssc->invlist);
1385 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1393 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1394 const regnode_charclass* const node)
1396 /* Returns a mortal inversion list defining which code points are matched
1397 * by 'node', which is of type ANYOF. Handles complementing the result if
1398 * appropriate. If some code points aren't knowable at this time, the
1399 * returned list must, and will, contain every code point that is a
1403 SV* only_utf8_locale_invlist = NULL;
1405 const U32 n = ARG(node);
1406 bool new_node_has_latin1 = FALSE;
1408 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1410 /* Look at the data structure created by S_set_ANYOF_arg() */
1411 if (n != ANYOF_ONLY_HAS_BITMAP) {
1412 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1413 AV * const av = MUTABLE_AV(SvRV(rv));
1414 SV **const ary = AvARRAY(av);
1415 assert(RExC_rxi->data->what[n] == 's');
1417 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1418 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1420 else if (ary[0] && ary[0] != &PL_sv_undef) {
1422 /* Here, no compile-time swash, and there are things that won't be
1423 * known until runtime -- we have to assume it could be anything */
1424 invlist = sv_2mortal(_new_invlist(1));
1425 return _add_range_to_invlist(invlist, 0, UV_MAX);
1427 else if (ary[3] && ary[3] != &PL_sv_undef) {
1429 /* Here no compile-time swash, and no run-time only data. Use the
1430 * node's inversion list */
1431 invlist = sv_2mortal(invlist_clone(ary[3]));
1434 /* Get the code points valid only under UTF-8 locales */
1435 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1436 && ary[2] && ary[2] != &PL_sv_undef)
1438 only_utf8_locale_invlist = ary[2];
1443 invlist = sv_2mortal(_new_invlist(0));
1446 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1447 * code points, and an inversion list for the others, but if there are code
1448 * points that should match only conditionally on the target string being
1449 * UTF-8, those are placed in the inversion list, and not the bitmap.
1450 * Since there are circumstances under which they could match, they are
1451 * included in the SSC. But if the ANYOF node is to be inverted, we have
1452 * to exclude them here, so that when we invert below, the end result
1453 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1454 * have to do this here before we add the unconditionally matched code
1456 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1457 _invlist_intersection_complement_2nd(invlist,
1462 /* Add in the points from the bit map */
1463 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1464 if (ANYOF_BITMAP_TEST(node, i)) {
1465 unsigned int start = i++;
1467 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1470 invlist = _add_range_to_invlist(invlist, start, i-1);
1471 new_node_has_latin1 = TRUE;
1475 /* If this can match all upper Latin1 code points, have to add them
1476 * as well. But don't add them if inverting, as when that gets done below,
1477 * it would exclude all these characters, including the ones it shouldn't
1478 * that were added just above */
1479 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1480 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1482 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1485 /* Similarly for these */
1486 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1487 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1490 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1491 _invlist_invert(invlist);
1493 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1495 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1496 * locale. We can skip this if there are no 0-255 at all. */
1497 _invlist_union(invlist, PL_Latin1, &invlist);
1500 /* Similarly add the UTF-8 locale possible matches. These have to be
1501 * deferred until after the non-UTF-8 locale ones are taken care of just
1502 * above, or it leads to wrong results under ANYOF_INVERT */
1503 if (only_utf8_locale_invlist) {
1504 _invlist_union_maybe_complement_2nd(invlist,
1505 only_utf8_locale_invlist,
1506 ANYOF_FLAGS(node) & ANYOF_INVERT,
1513 /* These two functions currently do the exact same thing */
1514 #define ssc_init_zero ssc_init
1516 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1517 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1519 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1520 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1521 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1524 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1525 const regnode_charclass *and_with)
1527 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1528 * another SSC or a regular ANYOF class. Can create false positives. */
1533 PERL_ARGS_ASSERT_SSC_AND;
1535 assert(is_ANYOF_SYNTHETIC(ssc));
1537 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1538 * the code point inversion list and just the relevant flags */
1539 if (is_ANYOF_SYNTHETIC(and_with)) {
1540 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1541 anded_flags = ANYOF_FLAGS(and_with);
1543 /* XXX This is a kludge around what appears to be deficiencies in the
1544 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1545 * there are paths through the optimizer where it doesn't get weeded
1546 * out when it should. And if we don't make some extra provision for
1547 * it like the code just below, it doesn't get added when it should.
1548 * This solution is to add it only when AND'ing, which is here, and
1549 * only when what is being AND'ed is the pristine, original node
1550 * matching anything. Thus it is like adding it to ssc_anything() but
1551 * only when the result is to be AND'ed. Probably the same solution
1552 * could be adopted for the same problem we have with /l matching,
1553 * which is solved differently in S_ssc_init(), and that would lead to
1554 * fewer false positives than that solution has. But if this solution
1555 * creates bugs, the consequences are only that a warning isn't raised
1556 * that should be; while the consequences for having /l bugs is
1557 * incorrect matches */
1558 if (ssc_is_anything((regnode_ssc *)and_with)) {
1559 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1563 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1564 if (OP(and_with) == ANYOFD) {
1565 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1568 anded_flags = ANYOF_FLAGS(and_with)
1569 &( ANYOF_COMMON_FLAGS
1570 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1571 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1572 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1574 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1579 ANYOF_FLAGS(ssc) &= anded_flags;
1581 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1582 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1583 * 'and_with' may be inverted. When not inverted, we have the situation of
1585 * (C1 | P1) & (C2 | P2)
1586 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1587 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1588 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1589 * <= ((C1 & C2) | P1 | P2)
1590 * Alternatively, the last few steps could be:
1591 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1592 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1593 * <= (C1 | C2 | (P1 & P2))
1594 * We favor the second approach if either P1 or P2 is non-empty. This is
1595 * because these components are a barrier to doing optimizations, as what
1596 * they match cannot be known until the moment of matching as they are
1597 * dependent on the current locale, 'AND"ing them likely will reduce or
1599 * But we can do better if we know that C1,P1 are in their initial state (a
1600 * frequent occurrence), each matching everything:
1601 * (<everything>) & (C2 | P2) = C2 | P2
1602 * Similarly, if C2,P2 are in their initial state (again a frequent
1603 * occurrence), the result is a no-op
1604 * (C1 | P1) & (<everything>) = C1 | P1
1607 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1608 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1609 * <= (C1 & ~C2) | (P1 & ~P2)
1612 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1613 && ! is_ANYOF_SYNTHETIC(and_with))
1617 ssc_intersection(ssc,
1619 FALSE /* Has already been inverted */
1622 /* If either P1 or P2 is empty, the intersection will be also; can skip
1624 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1625 ANYOF_POSIXL_ZERO(ssc);
1627 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1629 /* Note that the Posix class component P from 'and_with' actually
1631 * P = Pa | Pb | ... | Pn
1632 * where each component is one posix class, such as in [\w\s].
1634 * ~P = ~(Pa | Pb | ... | Pn)
1635 * = ~Pa & ~Pb & ... & ~Pn
1636 * <= ~Pa | ~Pb | ... | ~Pn
1637 * The last is something we can easily calculate, but unfortunately
1638 * is likely to have many false positives. We could do better
1639 * in some (but certainly not all) instances if two classes in
1640 * P have known relationships. For example
1641 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1643 * :lower: & :print: = :lower:
1644 * And similarly for classes that must be disjoint. For example,
1645 * since \s and \w can have no elements in common based on rules in
1646 * the POSIX standard,
1647 * \w & ^\S = nothing
1648 * Unfortunately, some vendor locales do not meet the Posix
1649 * standard, in particular almost everything by Microsoft.
1650 * The loop below just changes e.g., \w into \W and vice versa */
1652 regnode_charclass_posixl temp;
1653 int add = 1; /* To calculate the index of the complement */
1655 ANYOF_POSIXL_ZERO(&temp);
1656 for (i = 0; i < ANYOF_MAX; i++) {
1658 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1659 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1661 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1662 ANYOF_POSIXL_SET(&temp, i + add);
1664 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1666 ANYOF_POSIXL_AND(&temp, ssc);
1668 } /* else ssc already has no posixes */
1669 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1670 in its initial state */
1671 else if (! is_ANYOF_SYNTHETIC(and_with)
1672 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1674 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1675 * copy it over 'ssc' */
1676 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1677 if (is_ANYOF_SYNTHETIC(and_with)) {
1678 StructCopy(and_with, ssc, regnode_ssc);
1681 ssc->invlist = anded_cp_list;
1682 ANYOF_POSIXL_ZERO(ssc);
1683 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1684 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1688 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1689 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1691 /* One or the other of P1, P2 is non-empty. */
1692 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1693 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1695 ssc_union(ssc, anded_cp_list, FALSE);
1697 else { /* P1 = P2 = empty */
1698 ssc_intersection(ssc, anded_cp_list, FALSE);
1704 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1705 const regnode_charclass *or_with)
1707 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1708 * another SSC or a regular ANYOF class. Can create false positives if
1709 * 'or_with' is to be inverted. */
1714 PERL_ARGS_ASSERT_SSC_OR;
1716 assert(is_ANYOF_SYNTHETIC(ssc));
1718 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1719 * the code point inversion list and just the relevant flags */
1720 if (is_ANYOF_SYNTHETIC(or_with)) {
1721 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1722 ored_flags = ANYOF_FLAGS(or_with);
1725 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1726 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1727 if (OP(or_with) != ANYOFD) {
1729 |= ANYOF_FLAGS(or_with)
1730 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1731 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1732 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1734 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1739 ANYOF_FLAGS(ssc) |= ored_flags;
1741 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1742 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1743 * 'or_with' may be inverted. When not inverted, we have the simple
1744 * situation of computing:
1745 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1746 * If P1|P2 yields a situation with both a class and its complement are
1747 * set, like having both \w and \W, this matches all code points, and we
1748 * can delete these from the P component of the ssc going forward. XXX We
1749 * might be able to delete all the P components, but I (khw) am not certain
1750 * about this, and it is better to be safe.
1753 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1754 * <= (C1 | P1) | ~C2
1755 * <= (C1 | ~C2) | P1
1756 * (which results in actually simpler code than the non-inverted case)
1759 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1760 && ! is_ANYOF_SYNTHETIC(or_with))
1762 /* We ignore P2, leaving P1 going forward */
1763 } /* else Not inverted */
1764 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1765 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1766 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1768 for (i = 0; i < ANYOF_MAX; i += 2) {
1769 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1771 ssc_match_all_cp(ssc);
1772 ANYOF_POSIXL_CLEAR(ssc, i);
1773 ANYOF_POSIXL_CLEAR(ssc, i+1);
1781 FALSE /* Already has been inverted */
1785 PERL_STATIC_INLINE void
1786 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1788 PERL_ARGS_ASSERT_SSC_UNION;
1790 assert(is_ANYOF_SYNTHETIC(ssc));
1792 _invlist_union_maybe_complement_2nd(ssc->invlist,
1798 PERL_STATIC_INLINE void
1799 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1801 const bool invert2nd)
1803 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1805 assert(is_ANYOF_SYNTHETIC(ssc));
1807 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1813 PERL_STATIC_INLINE void
1814 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1816 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1818 assert(is_ANYOF_SYNTHETIC(ssc));
1820 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1823 PERL_STATIC_INLINE void
1824 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1826 /* AND just the single code point 'cp' into the SSC 'ssc' */
1828 SV* cp_list = _new_invlist(2);
1830 PERL_ARGS_ASSERT_SSC_CP_AND;
1832 assert(is_ANYOF_SYNTHETIC(ssc));
1834 cp_list = add_cp_to_invlist(cp_list, cp);
1835 ssc_intersection(ssc, cp_list,
1836 FALSE /* Not inverted */
1838 SvREFCNT_dec_NN(cp_list);
1841 PERL_STATIC_INLINE void
1842 S_ssc_clear_locale(regnode_ssc *ssc)
1844 /* Set the SSC 'ssc' to not match any locale things */
1845 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1847 assert(is_ANYOF_SYNTHETIC(ssc));
1849 ANYOF_POSIXL_ZERO(ssc);
1850 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1853 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1856 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1858 /* The synthetic start class is used to hopefully quickly winnow down
1859 * places where a pattern could start a match in the target string. If it
1860 * doesn't really narrow things down that much, there isn't much point to
1861 * having the overhead of using it. This function uses some very crude
1862 * heuristics to decide if to use the ssc or not.
1864 * It returns TRUE if 'ssc' rules out more than half what it considers to
1865 * be the "likely" possible matches, but of course it doesn't know what the
1866 * actual things being matched are going to be; these are only guesses
1868 * For /l matches, it assumes that the only likely matches are going to be
1869 * in the 0-255 range, uniformly distributed, so half of that is 127
1870 * For /a and /d matches, it assumes that the likely matches will be just
1871 * the ASCII range, so half of that is 63
1872 * For /u and there isn't anything matching above the Latin1 range, it
1873 * assumes that that is the only range likely to be matched, and uses
1874 * half that as the cut-off: 127. If anything matches above Latin1,
1875 * it assumes that all of Unicode could match (uniformly), except for
1876 * non-Unicode code points and things in the General Category "Other"
1877 * (unassigned, private use, surrogates, controls and formats). This
1878 * is a much large number. */
1880 U32 count = 0; /* Running total of number of code points matched by
1882 UV start, end; /* Start and end points of current range in inversion
1884 const U32 max_code_points = (LOC)
1886 : (( ! UNI_SEMANTICS
1887 || invlist_highest(ssc->invlist) < 256)
1890 const U32 max_match = max_code_points / 2;
1892 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1894 invlist_iterinit(ssc->invlist);
1895 while (invlist_iternext(ssc->invlist, &start, &end)) {
1896 if (start >= max_code_points) {
1899 end = MIN(end, max_code_points - 1);
1900 count += end - start + 1;
1901 if (count >= max_match) {
1902 invlist_iterfinish(ssc->invlist);
1912 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1914 /* The inversion list in the SSC is marked mortal; now we need a more
1915 * permanent copy, which is stored the same way that is done in a regular
1916 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1919 SV* invlist = invlist_clone(ssc->invlist);
1921 PERL_ARGS_ASSERT_SSC_FINALIZE;
1923 assert(is_ANYOF_SYNTHETIC(ssc));
1925 /* The code in this file assumes that all but these flags aren't relevant
1926 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1927 * by the time we reach here */
1928 assert(! (ANYOF_FLAGS(ssc)
1929 & ~( ANYOF_COMMON_FLAGS
1930 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1931 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
1933 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1935 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1936 NULL, NULL, NULL, FALSE);
1938 /* Make sure is clone-safe */
1939 ssc->invlist = NULL;
1941 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1942 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1945 if (RExC_contains_locale) {
1949 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1952 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1953 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1954 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1955 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1956 ? (TRIE_LIST_CUR( idx ) - 1) \
1962 dump_trie(trie,widecharmap,revcharmap)
1963 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1964 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1966 These routines dump out a trie in a somewhat readable format.
1967 The _interim_ variants are used for debugging the interim
1968 tables that are used to generate the final compressed
1969 representation which is what dump_trie expects.
1971 Part of the reason for their existence is to provide a form
1972 of documentation as to how the different representations function.
1977 Dumps the final compressed table form of the trie to Perl_debug_log.
1978 Used for debugging make_trie().
1982 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1983 AV *revcharmap, U32 depth)
1986 SV *sv=sv_newmortal();
1987 int colwidth= widecharmap ? 6 : 4;
1989 GET_RE_DEBUG_FLAGS_DECL;
1991 PERL_ARGS_ASSERT_DUMP_TRIE;
1993 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
1994 depth+1, "Match","Base","Ofs" );
1996 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1997 SV ** const tmp = av_fetch( revcharmap, state, 0);
1999 Perl_re_printf( aTHX_ "%*s",
2001 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2002 PL_colors[0], PL_colors[1],
2003 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2004 PERL_PV_ESCAPE_FIRSTCHAR
2009 Perl_re_printf( aTHX_ "\n");
2010 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2012 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2013 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2014 Perl_re_printf( aTHX_ "\n");
2016 for( state = 1 ; state < trie->statecount ; state++ ) {
2017 const U32 base = trie->states[ state ].trans.base;
2019 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2021 if ( trie->states[ state ].wordnum ) {
2022 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2024 Perl_re_printf( aTHX_ "%6s", "" );
2027 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2032 while( ( base + ofs < trie->uniquecharcount ) ||
2033 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2034 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2038 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2040 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2041 if ( ( base + ofs >= trie->uniquecharcount )
2042 && ( base + ofs - trie->uniquecharcount
2044 && trie->trans[ base + ofs
2045 - trie->uniquecharcount ].check == state )
2047 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2048 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2051 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
2055 Perl_re_printf( aTHX_ "]");
2058 Perl_re_printf( aTHX_ "\n" );
2060 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2062 for (word=1; word <= trie->wordcount; word++) {
2063 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2064 (int)word, (int)(trie->wordinfo[word].prev),
2065 (int)(trie->wordinfo[word].len));
2067 Perl_re_printf( aTHX_ "\n" );
2070 Dumps a fully constructed but uncompressed trie in list form.
2071 List tries normally only are used for construction when the number of
2072 possible chars (trie->uniquecharcount) is very high.
2073 Used for debugging make_trie().
2076 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2077 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2081 SV *sv=sv_newmortal();
2082 int colwidth= widecharmap ? 6 : 4;
2083 GET_RE_DEBUG_FLAGS_DECL;
2085 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2087 /* print out the table precompression. */
2088 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2090 Perl_re_indentf( aTHX_ "%s",
2091 depth+1, "------:-----+-----------------\n" );
2093 for( state=1 ; state < next_alloc ; state ++ ) {
2096 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2097 depth+1, (UV)state );
2098 if ( ! trie->states[ state ].wordnum ) {
2099 Perl_re_printf( aTHX_ "%5s| ","");
2101 Perl_re_printf( aTHX_ "W%4x| ",
2102 trie->states[ state ].wordnum
2105 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2106 SV ** const tmp = av_fetch( revcharmap,
2107 TRIE_LIST_ITEM(state,charid).forid, 0);
2109 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2111 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2113 PL_colors[0], PL_colors[1],
2114 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2115 | PERL_PV_ESCAPE_FIRSTCHAR
2117 TRIE_LIST_ITEM(state,charid).forid,
2118 (UV)TRIE_LIST_ITEM(state,charid).newstate
2121 Perl_re_printf( aTHX_ "\n%*s| ",
2122 (int)((depth * 2) + 14), "");
2125 Perl_re_printf( aTHX_ "\n");
2130 Dumps a fully constructed but uncompressed trie in table form.
2131 This is the normal DFA style state transition table, with a few
2132 twists to facilitate compression later.
2133 Used for debugging make_trie().
2136 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2137 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2142 SV *sv=sv_newmortal();
2143 int colwidth= widecharmap ? 6 : 4;
2144 GET_RE_DEBUG_FLAGS_DECL;
2146 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2149 print out the table precompression so that we can do a visual check
2150 that they are identical.
2153 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2155 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2156 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2158 Perl_re_printf( aTHX_ "%*s",
2160 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2161 PL_colors[0], PL_colors[1],
2162 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2163 PERL_PV_ESCAPE_FIRSTCHAR
2169 Perl_re_printf( aTHX_ "\n");
2170 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2172 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2173 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2176 Perl_re_printf( aTHX_ "\n" );
2178 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2180 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2182 (UV)TRIE_NODENUM( state ) );
2184 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2185 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2187 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2189 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2191 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2192 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2193 (UV)trie->trans[ state ].check );
2195 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2196 (UV)trie->trans[ state ].check,
2197 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2205 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2206 startbranch: the first branch in the whole branch sequence
2207 first : start branch of sequence of branch-exact nodes.
2208 May be the same as startbranch
2209 last : Thing following the last branch.
2210 May be the same as tail.
2211 tail : item following the branch sequence
2212 count : words in the sequence
2213 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2214 depth : indent depth
2216 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2218 A trie is an N'ary tree where the branches are determined by digital
2219 decomposition of the key. IE, at the root node you look up the 1st character and
2220 follow that branch repeat until you find the end of the branches. Nodes can be
2221 marked as "accepting" meaning they represent a complete word. Eg:
2225 would convert into the following structure. Numbers represent states, letters
2226 following numbers represent valid transitions on the letter from that state, if
2227 the number is in square brackets it represents an accepting state, otherwise it
2228 will be in parenthesis.
2230 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2234 (1) +-i->(6)-+-s->[7]
2236 +-s->(3)-+-h->(4)-+-e->[5]
2238 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2240 This shows that when matching against the string 'hers' we will begin at state 1
2241 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2242 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2243 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2244 single traverse. We store a mapping from accepting to state to which word was
2245 matched, and then when we have multiple possibilities we try to complete the
2246 rest of the regex in the order in which they occurred in the alternation.
2248 The only prior NFA like behaviour that would be changed by the TRIE support is
2249 the silent ignoring of duplicate alternations which are of the form:
2251 / (DUPE|DUPE) X? (?{ ... }) Y /x
2253 Thus EVAL blocks following a trie may be called a different number of times with
2254 and without the optimisation. With the optimisations dupes will be silently
2255 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2256 the following demonstrates:
2258 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2260 which prints out 'word' three times, but
2262 'words'=~/(word|word|word)(?{ print $1 })S/
2264 which doesnt print it out at all. This is due to other optimisations kicking in.
2266 Example of what happens on a structural level:
2268 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2270 1: CURLYM[1] {1,32767}(18)
2281 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2282 and should turn into:
2284 1: CURLYM[1] {1,32767}(18)
2286 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2294 Cases where tail != last would be like /(?foo|bar)baz/:
2304 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2305 and would end up looking like:
2308 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2315 d = uvchr_to_utf8_flags(d, uv, 0);
2317 is the recommended Unicode-aware way of saying
2322 #define TRIE_STORE_REVCHAR(val) \
2325 SV *zlopp = newSV(UTF8_MAXBYTES); \
2326 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2327 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2328 SvCUR_set(zlopp, kapow - flrbbbbb); \
2331 av_push(revcharmap, zlopp); \
2333 char ooooff = (char)val; \
2334 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2338 /* This gets the next character from the input, folding it if not already
2340 #define TRIE_READ_CHAR STMT_START { \
2343 /* if it is UTF then it is either already folded, or does not need \
2345 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2347 else if (folder == PL_fold_latin1) { \
2348 /* This folder implies Unicode rules, which in the range expressible \
2349 * by not UTF is the lower case, with the two exceptions, one of \
2350 * which should have been taken care of before calling this */ \
2351 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2352 uvc = toLOWER_L1(*uc); \
2353 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2356 /* raw data, will be folded later if needed */ \
2364 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2365 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2366 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2367 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2369 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2370 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2371 TRIE_LIST_CUR( state )++; \
2374 #define TRIE_LIST_NEW(state) STMT_START { \
2375 Newxz( trie->states[ state ].trans.list, \
2376 4, reg_trie_trans_le ); \
2377 TRIE_LIST_CUR( state ) = 1; \
2378 TRIE_LIST_LEN( state ) = 4; \
2381 #define TRIE_HANDLE_WORD(state) STMT_START { \
2382 U16 dupe= trie->states[ state ].wordnum; \
2383 regnode * const noper_next = regnext( noper ); \
2386 /* store the word for dumping */ \
2388 if (OP(noper) != NOTHING) \
2389 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2391 tmp = newSVpvn_utf8( "", 0, UTF ); \
2392 av_push( trie_words, tmp ); \
2396 trie->wordinfo[curword].prev = 0; \
2397 trie->wordinfo[curword].len = wordlen; \
2398 trie->wordinfo[curword].accept = state; \
2400 if ( noper_next < tail ) { \
2402 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2404 trie->jump[curword] = (U16)(noper_next - convert); \
2406 jumper = noper_next; \
2408 nextbranch= regnext(cur); \
2412 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2413 /* chain, so that when the bits of chain are later */\
2414 /* linked together, the dups appear in the chain */\
2415 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2416 trie->wordinfo[dupe].prev = curword; \
2418 /* we haven't inserted this word yet. */ \
2419 trie->states[ state ].wordnum = curword; \
2424 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2425 ( ( base + charid >= ucharcount \
2426 && base + charid < ubound \
2427 && state == trie->trans[ base - ucharcount + charid ].check \
2428 && trie->trans[ base - ucharcount + charid ].next ) \
2429 ? trie->trans[ base - ucharcount + charid ].next \
2430 : ( state==1 ? special : 0 ) \
2433 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2435 TRIE_BITMAP_SET(trie, uvc); \
2436 /* store the folded codepoint */ \
2438 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2441 /* store first byte of utf8 representation of */ \
2442 /* variant codepoints */ \
2443 if (! UVCHR_IS_INVARIANT(uvc)) { \
2444 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2449 #define MADE_JUMP_TRIE 2
2450 #define MADE_EXACT_TRIE 4
2453 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2454 regnode *first, regnode *last, regnode *tail,
2455 U32 word_count, U32 flags, U32 depth)
2457 /* first pass, loop through and scan words */
2458 reg_trie_data *trie;
2459 HV *widecharmap = NULL;
2460 AV *revcharmap = newAV();
2466 regnode *jumper = NULL;
2467 regnode *nextbranch = NULL;
2468 regnode *convert = NULL;
2469 U32 *prev_states; /* temp array mapping each state to previous one */
2470 /* we just use folder as a flag in utf8 */
2471 const U8 * folder = NULL;
2474 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2475 AV *trie_words = NULL;
2476 /* along with revcharmap, this only used during construction but both are
2477 * useful during debugging so we store them in the struct when debugging.
2480 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2481 STRLEN trie_charcount=0;
2483 SV *re_trie_maxbuff;
2484 GET_RE_DEBUG_FLAGS_DECL;
2486 PERL_ARGS_ASSERT_MAKE_TRIE;
2488 PERL_UNUSED_ARG(depth);
2492 case EXACT: case EXACTL: break;
2496 case EXACTFLU8: folder = PL_fold_latin1; break;
2497 case EXACTF: folder = PL_fold; break;
2498 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2501 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2503 trie->startstate = 1;
2504 trie->wordcount = word_count;
2505 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2506 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2507 if (flags == EXACT || flags == EXACTL)
2508 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2509 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2510 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2513 trie_words = newAV();
2516 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2517 assert(re_trie_maxbuff);
2518 if (!SvIOK(re_trie_maxbuff)) {
2519 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2521 DEBUG_TRIE_COMPILE_r({
2522 Perl_re_indentf( aTHX_
2523 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2525 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2526 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2529 /* Find the node we are going to overwrite */
2530 if ( first == startbranch && OP( last ) != BRANCH ) {
2531 /* whole branch chain */
2534 /* branch sub-chain */
2535 convert = NEXTOPER( first );
2538 /* -- First loop and Setup --
2540 We first traverse the branches and scan each word to determine if it
2541 contains widechars, and how many unique chars there are, this is
2542 important as we have to build a table with at least as many columns as we
2545 We use an array of integers to represent the character codes 0..255
2546 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2547 the native representation of the character value as the key and IV's for
2550 *TODO* If we keep track of how many times each character is used we can
2551 remap the columns so that the table compression later on is more
2552 efficient in terms of memory by ensuring the most common value is in the
2553 middle and the least common are on the outside. IMO this would be better
2554 than a most to least common mapping as theres a decent chance the most
2555 common letter will share a node with the least common, meaning the node
2556 will not be compressible. With a middle is most common approach the worst
2557 case is when we have the least common nodes twice.
2561 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2562 regnode *noper = NEXTOPER( cur );
2566 U32 wordlen = 0; /* required init */
2567 STRLEN minchars = 0;
2568 STRLEN maxchars = 0;
2569 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2572 if (OP(noper) == NOTHING) {
2573 /* skip past a NOTHING at the start of an alternation
2574 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2576 regnode *noper_next= regnext(noper);
2577 if (noper_next < tail)
2581 if ( noper < tail &&
2583 OP(noper) == flags ||
2586 OP(noper) == EXACTFU_SS
2590 uc= (U8*)STRING(noper);
2591 e= uc + STR_LEN(noper);
2598 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2599 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2600 regardless of encoding */
2601 if (OP( noper ) == EXACTFU_SS) {
2602 /* false positives are ok, so just set this */
2603 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2607 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2609 TRIE_CHARCOUNT(trie)++;
2612 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2613 * is in effect. Under /i, this character can match itself, or
2614 * anything that folds to it. If not under /i, it can match just
2615 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2616 * all fold to k, and all are single characters. But some folds
2617 * expand to more than one character, so for example LATIN SMALL
2618 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2619 * the string beginning at 'uc' is 'ffi', it could be matched by
2620 * three characters, or just by the one ligature character. (It
2621 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2622 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2623 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2624 * match.) The trie needs to know the minimum and maximum number
2625 * of characters that could match so that it can use size alone to
2626 * quickly reject many match attempts. The max is simple: it is
2627 * the number of folded characters in this branch (since a fold is
2628 * never shorter than what folds to it. */
2632 /* And the min is equal to the max if not under /i (indicated by
2633 * 'folder' being NULL), or there are no multi-character folds. If
2634 * there is a multi-character fold, the min is incremented just
2635 * once, for the character that folds to the sequence. Each
2636 * character in the sequence needs to be added to the list below of
2637 * characters in the trie, but we count only the first towards the
2638 * min number of characters needed. This is done through the
2639 * variable 'foldlen', which is returned by the macros that look
2640 * for these sequences as the number of bytes the sequence
2641 * occupies. Each time through the loop, we decrement 'foldlen' by
2642 * how many bytes the current char occupies. Only when it reaches
2643 * 0 do we increment 'minchars' or look for another multi-character
2645 if (folder == NULL) {
2648 else if (foldlen > 0) {
2649 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2654 /* See if *uc is the beginning of a multi-character fold. If
2655 * so, we decrement the length remaining to look at, to account
2656 * for the current character this iteration. (We can use 'uc'
2657 * instead of the fold returned by TRIE_READ_CHAR because for
2658 * non-UTF, the latin1_safe macro is smart enough to account
2659 * for all the unfolded characters, and because for UTF, the
2660 * string will already have been folded earlier in the
2661 * compilation process */
2663 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2664 foldlen -= UTF8SKIP(uc);
2667 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2672 /* The current character (and any potential folds) should be added
2673 * to the possible matching characters for this position in this
2677 U8 folded= folder[ (U8) uvc ];
2678 if ( !trie->charmap[ folded ] ) {
2679 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2680 TRIE_STORE_REVCHAR( folded );
2683 if ( !trie->charmap[ uvc ] ) {
2684 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2685 TRIE_STORE_REVCHAR( uvc );
2688 /* store the codepoint in the bitmap, and its folded
2690 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2691 set_bit = 0; /* We've done our bit :-) */
2695 /* XXX We could come up with the list of code points that fold
2696 * to this using PL_utf8_foldclosures, except not for
2697 * multi-char folds, as there may be multiple combinations
2698 * there that could work, which needs to wait until runtime to
2699 * resolve (The comment about LIGATURE FFI above is such an
2704 widecharmap = newHV();
2706 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2709 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2711 if ( !SvTRUE( *svpp ) ) {
2712 sv_setiv( *svpp, ++trie->uniquecharcount );
2713 TRIE_STORE_REVCHAR(uvc);
2716 } /* end loop through characters in this branch of the trie */
2718 /* We take the min and max for this branch and combine to find the min
2719 * and max for all branches processed so far */
2720 if( cur == first ) {
2721 trie->minlen = minchars;
2722 trie->maxlen = maxchars;
2723 } else if (minchars < trie->minlen) {
2724 trie->minlen = minchars;
2725 } else if (maxchars > trie->maxlen) {
2726 trie->maxlen = maxchars;
2728 } /* end first pass */
2729 DEBUG_TRIE_COMPILE_r(
2730 Perl_re_indentf( aTHX_
2731 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2733 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2734 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2735 (int)trie->minlen, (int)trie->maxlen )
2739 We now know what we are dealing with in terms of unique chars and
2740 string sizes so we can calculate how much memory a naive
2741 representation using a flat table will take. If it's over a reasonable
2742 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2743 conservative but potentially much slower representation using an array
2746 At the end we convert both representations into the same compressed
2747 form that will be used in regexec.c for matching with. The latter
2748 is a form that cannot be used to construct with but has memory
2749 properties similar to the list form and access properties similar
2750 to the table form making it both suitable for fast searches and
2751 small enough that its feasable to store for the duration of a program.
2753 See the comment in the code where the compressed table is produced
2754 inplace from the flat tabe representation for an explanation of how
2755 the compression works.
2760 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2763 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2764 > SvIV(re_trie_maxbuff) )
2767 Second Pass -- Array Of Lists Representation
2769 Each state will be represented by a list of charid:state records
2770 (reg_trie_trans_le) the first such element holds the CUR and LEN
2771 points of the allocated array. (See defines above).
2773 We build the initial structure using the lists, and then convert
2774 it into the compressed table form which allows faster lookups
2775 (but cant be modified once converted).
2778 STRLEN transcount = 1;
2780 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2783 trie->states = (reg_trie_state *)
2784 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2785 sizeof(reg_trie_state) );
2789 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2791 regnode *noper = NEXTOPER( cur );
2792 U32 state = 1; /* required init */
2793 U16 charid = 0; /* sanity init */
2794 U32 wordlen = 0; /* required init */
2796 if (OP(noper) == NOTHING) {
2797 regnode *noper_next= regnext(noper);
2798 if (noper_next < tail)
2802 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2803 const U8 *uc= (U8*)STRING(noper);
2804 const U8 *e= uc + STR_LEN(noper);
2806 for ( ; uc < e ; uc += len ) {
2811 charid = trie->charmap[ uvc ];
2813 SV** const svpp = hv_fetch( widecharmap,
2820 charid=(U16)SvIV( *svpp );
2823 /* charid is now 0 if we dont know the char read, or
2824 * nonzero if we do */
2831 if ( !trie->states[ state ].trans.list ) {
2832 TRIE_LIST_NEW( state );
2835 check <= TRIE_LIST_USED( state );
2838 if ( TRIE_LIST_ITEM( state, check ).forid
2841 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2846 newstate = next_alloc++;
2847 prev_states[newstate] = state;
2848 TRIE_LIST_PUSH( state, charid, newstate );
2853 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
2857 TRIE_HANDLE_WORD(state);
2859 } /* end second pass */
2861 /* next alloc is the NEXT state to be allocated */
2862 trie->statecount = next_alloc;
2863 trie->states = (reg_trie_state *)
2864 PerlMemShared_realloc( trie->states,
2866 * sizeof(reg_trie_state) );
2868 /* and now dump it out before we compress it */
2869 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2870 revcharmap, next_alloc,
2874 trie->trans = (reg_trie_trans *)
2875 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2882 for( state=1 ; state < next_alloc ; state ++ ) {
2886 DEBUG_TRIE_COMPILE_MORE_r(
2887 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
2891 if (trie->states[state].trans.list) {
2892 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2896 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2897 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2898 if ( forid < minid ) {
2900 } else if ( forid > maxid ) {
2904 if ( transcount < tp + maxid - minid + 1) {
2906 trie->trans = (reg_trie_trans *)
2907 PerlMemShared_realloc( trie->trans,
2909 * sizeof(reg_trie_trans) );
2910 Zero( trie->trans + (transcount / 2),
2914 base = trie->uniquecharcount + tp - minid;
2915 if ( maxid == minid ) {
2917 for ( ; zp < tp ; zp++ ) {
2918 if ( ! trie->trans[ zp ].next ) {
2919 base = trie->uniquecharcount + zp - minid;
2920 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2922 trie->trans[ zp ].check = state;
2928 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2930 trie->trans[ tp ].check = state;
2935 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2936 const U32 tid = base
2937 - trie->uniquecharcount
2938 + TRIE_LIST_ITEM( state, idx ).forid;
2939 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2941 trie->trans[ tid ].check = state;
2943 tp += ( maxid - minid + 1 );
2945 Safefree(trie->states[ state ].trans.list);
2948 DEBUG_TRIE_COMPILE_MORE_r(
2949 Perl_re_printf( aTHX_ " base: %d\n",base);
2952 trie->states[ state ].trans.base=base;
2954 trie->lasttrans = tp + 1;
2958 Second Pass -- Flat Table Representation.
2960 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2961 each. We know that we will need Charcount+1 trans at most to store
2962 the data (one row per char at worst case) So we preallocate both
2963 structures assuming worst case.
2965 We then construct the trie using only the .next slots of the entry
2968 We use the .check field of the first entry of the node temporarily
2969 to make compression both faster and easier by keeping track of how
2970 many non zero fields are in the node.
2972 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2975 There are two terms at use here: state as a TRIE_NODEIDX() which is
2976 a number representing the first entry of the node, and state as a
2977 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2978 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2979 if there are 2 entrys per node. eg:
2987 The table is internally in the right hand, idx form. However as we
2988 also have to deal with the states array which is indexed by nodenum
2989 we have to use TRIE_NODENUM() to convert.
2992 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
2995 trie->trans = (reg_trie_trans *)
2996 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2997 * trie->uniquecharcount + 1,
2998 sizeof(reg_trie_trans) );
2999 trie->states = (reg_trie_state *)
3000 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3001 sizeof(reg_trie_state) );
3002 next_alloc = trie->uniquecharcount + 1;
3005 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3007 regnode *noper = NEXTOPER( cur );
3009 U32 state = 1; /* required init */
3011 U16 charid = 0; /* sanity init */
3012 U32 accept_state = 0; /* sanity init */
3014 U32 wordlen = 0; /* required init */
3016 if (OP(noper) == NOTHING) {
3017 regnode *noper_next= regnext(noper);
3018 if (noper_next < tail)
3022 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3023 const U8 *uc= (U8*)STRING(noper);
3024 const U8 *e= uc + STR_LEN(noper);
3026 for ( ; uc < e ; uc += len ) {
3031 charid = trie->charmap[ uvc ];
3033 SV* const * const svpp = hv_fetch( widecharmap,
3037 charid = svpp ? (U16)SvIV(*svpp) : 0;
3041 if ( !trie->trans[ state + charid ].next ) {
3042 trie->trans[ state + charid ].next = next_alloc;
3043 trie->trans[ state ].check++;
3044 prev_states[TRIE_NODENUM(next_alloc)]
3045 = TRIE_NODENUM(state);
3046 next_alloc += trie->uniquecharcount;
3048 state = trie->trans[ state + charid ].next;
3050 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3052 /* charid is now 0 if we dont know the char read, or
3053 * nonzero if we do */
3056 accept_state = TRIE_NODENUM( state );
3057 TRIE_HANDLE_WORD(accept_state);
3059 } /* end second pass */
3061 /* and now dump it out before we compress it */
3062 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3064 next_alloc, depth+1));
3068 * Inplace compress the table.*
3070 For sparse data sets the table constructed by the trie algorithm will
3071 be mostly 0/FAIL transitions or to put it another way mostly empty.
3072 (Note that leaf nodes will not contain any transitions.)
3074 This algorithm compresses the tables by eliminating most such
3075 transitions, at the cost of a modest bit of extra work during lookup:
3077 - Each states[] entry contains a .base field which indicates the
3078 index in the state[] array wheres its transition data is stored.
3080 - If .base is 0 there are no valid transitions from that node.
3082 - If .base is nonzero then charid is added to it to find an entry in
3085 -If trans[states[state].base+charid].check!=state then the
3086 transition is taken to be a 0/Fail transition. Thus if there are fail
3087 transitions at the front of the node then the .base offset will point
3088 somewhere inside the previous nodes data (or maybe even into a node
3089 even earlier), but the .check field determines if the transition is
3093 The following process inplace converts the table to the compressed
3094 table: We first do not compress the root node 1,and mark all its
3095 .check pointers as 1 and set its .base pointer as 1 as well. This
3096 allows us to do a DFA construction from the compressed table later,
3097 and ensures that any .base pointers we calculate later are greater
3100 - We set 'pos' to indicate the first entry of the second node.
3102 - We then iterate over the columns of the node, finding the first and
3103 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3104 and set the .check pointers accordingly, and advance pos
3105 appropriately and repreat for the next node. Note that when we copy
3106 the next pointers we have to convert them from the original
3107 NODEIDX form to NODENUM form as the former is not valid post
3110 - If a node has no transitions used we mark its base as 0 and do not
3111 advance the pos pointer.
3113 - If a node only has one transition we use a second pointer into the
3114 structure to fill in allocated fail transitions from other states.
3115 This pointer is independent of the main pointer and scans forward
3116 looking for null transitions that are allocated to a state. When it
3117 finds one it writes the single transition into the "hole". If the
3118 pointer doesnt find one the single transition is appended as normal.
3120 - Once compressed we can Renew/realloc the structures to release the
3123 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3124 specifically Fig 3.47 and the associated pseudocode.
3128 const U32 laststate = TRIE_NODENUM( next_alloc );
3131 trie->statecount = laststate;
3133 for ( state = 1 ; state < laststate ; state++ ) {
3135 const U32 stateidx = TRIE_NODEIDX( state );
3136 const U32 o_used = trie->trans[ stateidx ].check;
3137 U32 used = trie->trans[ stateidx ].check;
3138 trie->trans[ stateidx ].check = 0;
3141 used && charid < trie->uniquecharcount;
3144 if ( flag || trie->trans[ stateidx + charid ].next ) {
3145 if ( trie->trans[ stateidx + charid ].next ) {
3147 for ( ; zp < pos ; zp++ ) {
3148 if ( ! trie->trans[ zp ].next ) {
3152 trie->states[ state ].trans.base
3154 + trie->uniquecharcount
3156 trie->trans[ zp ].next
3157 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3159 trie->trans[ zp ].check = state;
3160 if ( ++zp > pos ) pos = zp;
3167 trie->states[ state ].trans.base
3168 = pos + trie->uniquecharcount - charid ;
3170 trie->trans[ pos ].next
3171 = SAFE_TRIE_NODENUM(
3172 trie->trans[ stateidx + charid ].next );
3173 trie->trans[ pos ].check = state;
3178 trie->lasttrans = pos + 1;
3179 trie->states = (reg_trie_state *)
3180 PerlMemShared_realloc( trie->states, laststate
3181 * sizeof(reg_trie_state) );
3182 DEBUG_TRIE_COMPILE_MORE_r(
3183 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3185 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3189 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3192 } /* end table compress */
3194 DEBUG_TRIE_COMPILE_MORE_r(
3195 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3197 (UV)trie->statecount,
3198 (UV)trie->lasttrans)
3200 /* resize the trans array to remove unused space */
3201 trie->trans = (reg_trie_trans *)
3202 PerlMemShared_realloc( trie->trans, trie->lasttrans
3203 * sizeof(reg_trie_trans) );
3205 { /* Modify the program and insert the new TRIE node */
3206 U8 nodetype =(U8)(flags & 0xFF);
3210 regnode *optimize = NULL;
3211 #ifdef RE_TRACK_PATTERN_OFFSETS
3214 U32 mjd_nodelen = 0;
3215 #endif /* RE_TRACK_PATTERN_OFFSETS */
3216 #endif /* DEBUGGING */
3218 This means we convert either the first branch or the first Exact,
3219 depending on whether the thing following (in 'last') is a branch
3220 or not and whther first is the startbranch (ie is it a sub part of
3221 the alternation or is it the whole thing.)
3222 Assuming its a sub part we convert the EXACT otherwise we convert
3223 the whole branch sequence, including the first.
3225 /* Find the node we are going to overwrite */
3226 if ( first != startbranch || OP( last ) == BRANCH ) {
3227 /* branch sub-chain */
3228 NEXT_OFF( first ) = (U16)(last - first);
3229 #ifdef RE_TRACK_PATTERN_OFFSETS
3231 mjd_offset= Node_Offset((convert));
3232 mjd_nodelen= Node_Length((convert));
3235 /* whole branch chain */
3237 #ifdef RE_TRACK_PATTERN_OFFSETS
3240 const regnode *nop = NEXTOPER( convert );
3241 mjd_offset= Node_Offset((nop));
3242 mjd_nodelen= Node_Length((nop));
3246 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3248 (UV)mjd_offset, (UV)mjd_nodelen)
3251 /* But first we check to see if there is a common prefix we can
3252 split out as an EXACT and put in front of the TRIE node. */
3253 trie->startstate= 1;
3254 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3255 /* we want to find the first state that has more than
3256 * one transition, if that state is not the first state
3257 * then we have a common prefix which we can remove.
3260 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3262 I32 first_ofs = -1; /* keeps track of the ofs of the first
3263 transition, -1 means none */
3265 const U32 base = trie->states[ state ].trans.base;
3267 /* does this state terminate an alternation? */
3268 if ( trie->states[state].wordnum )
3271 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3272 if ( ( base + ofs >= trie->uniquecharcount ) &&
3273 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3274 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3276 if ( ++count > 1 ) {
3277 /* we have more than one transition */
3280 /* if this is the first state there is no common prefix
3281 * to extract, so we can exit */
3282 if ( state == 1 ) break;
3283 tmp = av_fetch( revcharmap, ofs, 0);
3284 ch = (U8*)SvPV_nolen_const( *tmp );
3286 /* if we are on count 2 then we need to initialize the
3287 * bitmap, and store the previous char if there was one
3290 /* clear the bitmap */
3291 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3293 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3296 if (first_ofs >= 0) {
3297 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3298 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3300 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3302 Perl_re_printf( aTHX_ "%s", (char*)ch)
3306 /* store the current firstchar in the bitmap */
3307 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3308 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3314 /* This state has only one transition, its transition is part
3315 * of a common prefix - we need to concatenate the char it
3316 * represents to what we have so far. */
3317 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3319 char *ch = SvPV( *tmp, len );
3321 SV *sv=sv_newmortal();
3322 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3324 (UV)state, (UV)first_ofs,
3325 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3326 PL_colors[0], PL_colors[1],
3327 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3328 PERL_PV_ESCAPE_FIRSTCHAR
3333 OP( convert ) = nodetype;
3334 str=STRING(convert);
3337 STR_LEN(convert) += len;
3343 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3348 trie->prefixlen = (state-1);
3350 regnode *n = convert+NODE_SZ_STR(convert);
3351 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3352 trie->startstate = state;
3353 trie->minlen -= (state - 1);
3354 trie->maxlen -= (state - 1);
3356 /* At least the UNICOS C compiler choked on this
3357 * being argument to DEBUG_r(), so let's just have
3360 #ifdef PERL_EXT_RE_BUILD
3366 regnode *fix = convert;
3367 U32 word = trie->wordcount;
3369 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3370 while( ++fix < n ) {
3371 Set_Node_Offset_Length(fix, 0, 0);
3374 SV ** const tmp = av_fetch( trie_words, word, 0 );
3376 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3377 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3379 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3387 NEXT_OFF(convert) = (U16)(tail - convert);
3388 DEBUG_r(optimize= n);
3394 if ( trie->maxlen ) {
3395 NEXT_OFF( convert ) = (U16)(tail - convert);
3396 ARG_SET( convert, data_slot );
3397 /* Store the offset to the first unabsorbed branch in
3398 jump[0], which is otherwise unused by the jump logic.
3399 We use this when dumping a trie and during optimisation. */
3401 trie->jump[0] = (U16)(nextbranch - convert);
3403 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3404 * and there is a bitmap
3405 * and the first "jump target" node we found leaves enough room
3406 * then convert the TRIE node into a TRIEC node, with the bitmap
3407 * embedded inline in the opcode - this is hypothetically faster.
3409 if ( !trie->states[trie->startstate].wordnum
3411 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3413 OP( convert ) = TRIEC;
3414 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3415 PerlMemShared_free(trie->bitmap);
3418 OP( convert ) = TRIE;
3420 /* store the type in the flags */
3421 convert->flags = nodetype;
3425 + regarglen[ OP( convert ) ];
3427 /* XXX We really should free up the resource in trie now,
3428 as we won't use them - (which resources?) dmq */
3430 /* needed for dumping*/
3431 DEBUG_r(if (optimize) {
3432 regnode *opt = convert;
3434 while ( ++opt < optimize) {
3435 Set_Node_Offset_Length(opt,0,0);
3438 Try to clean up some of the debris left after the
3441 while( optimize < jumper ) {
3442 mjd_nodelen += Node_Length((optimize));
3443 OP( optimize ) = OPTIMIZED;
3444 Set_Node_Offset_Length(optimize,0,0);
3447 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3449 } /* end node insert */
3451 /* Finish populating the prev field of the wordinfo array. Walk back
3452 * from each accept state until we find another accept state, and if
3453 * so, point the first word's .prev field at the second word. If the
3454 * second already has a .prev field set, stop now. This will be the
3455 * case either if we've already processed that word's accept state,
3456 * or that state had multiple words, and the overspill words were
3457 * already linked up earlier.
3464 for (word=1; word <= trie->wordcount; word++) {
3466 if (trie->wordinfo[word].prev)
3468 state = trie->wordinfo[word].accept;
3470 state = prev_states[state];
3473 prev = trie->states[state].wordnum;
3477 trie->wordinfo[word].prev = prev;
3479 Safefree(prev_states);
3483 /* and now dump out the compressed format */
3484 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3486 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3488 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3489 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3491 SvREFCNT_dec_NN(revcharmap);
3495 : trie->startstate>1
3501 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3503 /* The Trie is constructed and compressed now so we can build a fail array if
3506 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3508 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3512 We find the fail state for each state in the trie, this state is the longest
3513 proper suffix of the current state's 'word' that is also a proper prefix of
3514 another word in our trie. State 1 represents the word '' and is thus the
3515 default fail state. This allows the DFA not to have to restart after its
3516 tried and failed a word at a given point, it simply continues as though it
3517 had been matching the other word in the first place.
3519 'abcdgu'=~/abcdefg|cdgu/
3520 When we get to 'd' we are still matching the first word, we would encounter
3521 'g' which would fail, which would bring us to the state representing 'd' in
3522 the second word where we would try 'g' and succeed, proceeding to match
3525 /* add a fail transition */
3526 const U32 trie_offset = ARG(source);
3527 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3529 const U32 ucharcount = trie->uniquecharcount;
3530 const U32 numstates = trie->statecount;
3531 const U32 ubound = trie->lasttrans + ucharcount;
3535 U32 base = trie->states[ 1 ].trans.base;
3538 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3540 GET_RE_DEBUG_FLAGS_DECL;
3542 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3543 PERL_UNUSED_CONTEXT;
3545 PERL_UNUSED_ARG(depth);
3548 if ( OP(source) == TRIE ) {
3549 struct regnode_1 *op = (struct regnode_1 *)
3550 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3551 StructCopy(source,op,struct regnode_1);
3552 stclass = (regnode *)op;
3554 struct regnode_charclass *op = (struct regnode_charclass *)
3555 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3556 StructCopy(source,op,struct regnode_charclass);
3557 stclass = (regnode *)op;
3559 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3561 ARG_SET( stclass, data_slot );
3562 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3563 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3564 aho->trie=trie_offset;
3565 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3566 Copy( trie->states, aho->states, numstates, reg_trie_state );
3567 Newxz( q, numstates, U32);
3568 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3571 /* initialize fail[0..1] to be 1 so that we always have
3572 a valid final fail state */
3573 fail[ 0 ] = fail[ 1 ] = 1;
3575 for ( charid = 0; charid < ucharcount ; charid++ ) {
3576 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3578 q[ q_write ] = newstate;
3579 /* set to point at the root */
3580 fail[ q[ q_write++ ] ]=1;
3583 while ( q_read < q_write) {
3584 const U32 cur = q[ q_read++ % numstates ];
3585 base = trie->states[ cur ].trans.base;
3587 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3588 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3590 U32 fail_state = cur;
3593 fail_state = fail[ fail_state ];
3594 fail_base = aho->states[ fail_state ].trans.base;
3595 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3597 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3598 fail[ ch_state ] = fail_state;
3599 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3601 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3603 q[ q_write++ % numstates] = ch_state;
3607 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3608 when we fail in state 1, this allows us to use the
3609 charclass scan to find a valid start char. This is based on the principle
3610 that theres a good chance the string being searched contains lots of stuff
3611 that cant be a start char.
3613 fail[ 0 ] = fail[ 1 ] = 0;
3614 DEBUG_TRIE_COMPILE_r({
3615 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3616 depth, (UV)numstates
3618 for( q_read=1; q_read<numstates; q_read++ ) {
3619 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3621 Perl_re_printf( aTHX_ "\n");
3624 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3629 #define DEBUG_PEEP(str,scan,depth) \
3630 DEBUG_OPTIMISE_r({if (scan){ \
3631 regnode *Next = regnext(scan); \
3632 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);\
3633 Perl_re_indentf( aTHX_ "" str ">%3d: %s (%d)", \
3634 depth, REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3635 Next ? (REG_NODE_NUM(Next)) : 0 );\
3636 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3637 Perl_re_printf( aTHX_ "\n"); \
3640 /* The below joins as many adjacent EXACTish nodes as possible into a single
3641 * one. The regop may be changed if the node(s) contain certain sequences that
3642 * require special handling. The joining is only done if:
3643 * 1) there is room in the current conglomerated node to entirely contain the
3645 * 2) they are the exact same node type
3647 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3648 * these get optimized out
3650 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3651 * as possible, even if that means splitting an existing node so that its first
3652 * part is moved to the preceeding node. This would maximise the efficiency of
3653 * memEQ during matching. Elsewhere in this file, khw proposes splitting
3654 * EXACTFish nodes into portions that don't change under folding vs those that
3655 * do. Those portions that don't change may be the only things in the pattern that
3656 * could be used to find fixed and floating strings.
3658 * If a node is to match under /i (folded), the number of characters it matches
3659 * can be different than its character length if it contains a multi-character
3660 * fold. *min_subtract is set to the total delta number of characters of the
3663 * And *unfolded_multi_char is set to indicate whether or not the node contains
3664 * an unfolded multi-char fold. This happens when whether the fold is valid or
3665 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3666 * SMALL LETTER SHARP S, as only if the target string being matched against
3667 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3668 * folding rules depend on the locale in force at runtime. (Multi-char folds
3669 * whose components are all above the Latin1 range are not run-time locale
3670 * dependent, and have already been folded by the time this function is
3673 * This is as good a place as any to discuss the design of handling these
3674 * multi-character fold sequences. It's been wrong in Perl for a very long
3675 * time. There are three code points in Unicode whose multi-character folds
3676 * were long ago discovered to mess things up. The previous designs for
3677 * dealing with these involved assigning a special node for them. This
3678 * approach doesn't always work, as evidenced by this example:
3679 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3680 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3681 * would match just the \xDF, it won't be able to handle the case where a
3682 * successful match would have to cross the node's boundary. The new approach
3683 * that hopefully generally solves the problem generates an EXACTFU_SS node
3684 * that is "sss" in this case.
3686 * It turns out that there are problems with all multi-character folds, and not
3687 * just these three. Now the code is general, for all such cases. The
3688 * approach taken is:
3689 * 1) This routine examines each EXACTFish node that could contain multi-
3690 * character folded sequences. Since a single character can fold into
3691 * such a sequence, the minimum match length for this node is less than
3692 * the number of characters in the node. This routine returns in
3693 * *min_subtract how many characters to subtract from the the actual
3694 * length of the string to get a real minimum match length; it is 0 if
3695 * there are no multi-char foldeds. This delta is used by the caller to
3696 * adjust the min length of the match, and the delta between min and max,
3697 * so that the optimizer doesn't reject these possibilities based on size
3699 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3700 * is used for an EXACTFU node that contains at least one "ss" sequence in
3701 * it. For non-UTF-8 patterns and strings, this is the only case where
3702 * there is a possible fold length change. That means that a regular
3703 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3704 * with length changes, and so can be processed faster. regexec.c takes
3705 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3706 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3707 * known until runtime). This saves effort in regex matching. However,
3708 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3709 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3710 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3711 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3712 * possibilities for the non-UTF8 patterns are quite simple, except for
3713 * the sharp s. All the ones that don't involve a UTF-8 target string are
3714 * members of a fold-pair, and arrays are set up for all of them so that
3715 * the other member of the pair can be found quickly. Code elsewhere in
3716 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3717 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3718 * described in the next item.
3719 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3720 * validity of the fold won't be known until runtime, and so must remain
3721 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3722 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3723 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3724 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3725 * The reason this is a problem is that the optimizer part of regexec.c
3726 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3727 * that a character in the pattern corresponds to at most a single
3728 * character in the target string. (And I do mean character, and not byte
3729 * here, unlike other parts of the documentation that have never been
3730 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3731 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3732 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3733 * nodes, violate the assumption, and they are the only instances where it
3734 * is violated. I'm reluctant to try to change the assumption, as the
3735 * code involved is impenetrable to me (khw), so instead the code here
3736 * punts. This routine examines EXACTFL nodes, and (when the pattern
3737 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3738 * boolean indicating whether or not the node contains such a fold. When
3739 * it is true, the caller sets a flag that later causes the optimizer in
3740 * this file to not set values for the floating and fixed string lengths,
3741 * and thus avoids the optimizer code in regexec.c that makes the invalid
3742 * assumption. Thus, there is no optimization based on string lengths for
3743 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3744 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3745 * assumption is wrong only in these cases is that all other non-UTF-8
3746 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3747 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3748 * EXACTF nodes because we don't know at compile time if it actually
3749 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3750 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3751 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3752 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3753 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3754 * string would require the pattern to be forced into UTF-8, the overhead
3755 * of which we want to avoid. Similarly the unfolded multi-char folds in
3756 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3759 * Similarly, the code that generates tries doesn't currently handle
3760 * not-already-folded multi-char folds, and it looks like a pain to change
3761 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3762 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3763 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3764 * using /iaa matching will be doing so almost entirely with ASCII
3765 * strings, so this should rarely be encountered in practice */
3767 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3768 if (PL_regkind[OP(scan)] == EXACT) \
3769 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3772 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3773 UV *min_subtract, bool *unfolded_multi_char,
3774 U32 flags,regnode *val, U32 depth)
3776 /* Merge several consecutive EXACTish nodes into one. */
3777 regnode *n = regnext(scan);
3779 regnode *next = scan + NODE_SZ_STR(scan);
3783 regnode *stop = scan;
3784 GET_RE_DEBUG_FLAGS_DECL;
3786 PERL_UNUSED_ARG(depth);
3789 PERL_ARGS_ASSERT_JOIN_EXACT;
3790 #ifndef EXPERIMENTAL_INPLACESCAN
3791 PERL_UNUSED_ARG(flags);
3792 PERL_UNUSED_ARG(val);
3794 DEBUG_PEEP("join",scan,depth);
3796 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3797 * EXACT ones that are mergeable to the current one. */
3799 && (PL_regkind[OP(n)] == NOTHING
3800 || (stringok && OP(n) == OP(scan)))
3802 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3805 if (OP(n) == TAIL || n > next)
3807 if (PL_regkind[OP(n)] == NOTHING) {
3808 DEBUG_PEEP("skip:",n,depth);
3809 NEXT_OFF(scan) += NEXT_OFF(n);
3810 next = n + NODE_STEP_REGNODE;
3817 else if (stringok) {
3818 const unsigned int oldl = STR_LEN(scan);
3819 regnode * const nnext = regnext(n);
3821 /* XXX I (khw) kind of doubt that this works on platforms (should
3822 * Perl ever run on one) where U8_MAX is above 255 because of lots
3823 * of other assumptions */
3824 /* Don't join if the sum can't fit into a single node */
3825 if (oldl + STR_LEN(n) > U8_MAX)
3828 DEBUG_PEEP("merg",n,depth);
3831 NEXT_OFF(scan) += NEXT_OFF(n);
3832 STR_LEN(scan) += STR_LEN(n);
3833 next = n + NODE_SZ_STR(n);
3834 /* Now we can overwrite *n : */
3835 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3843 #ifdef EXPERIMENTAL_INPLACESCAN
3844 if (flags && !NEXT_OFF(n)) {
3845 DEBUG_PEEP("atch", val, depth);
3846 if (reg_off_by_arg[OP(n)]) {
3847 ARG_SET(n, val - n);
3850 NEXT_OFF(n) = val - n;
3858 *unfolded_multi_char = FALSE;
3860 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3861 * can now analyze for sequences of problematic code points. (Prior to
3862 * this final joining, sequences could have been split over boundaries, and
3863 * hence missed). The sequences only happen in folding, hence for any
3864 * non-EXACT EXACTish node */
3865 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3866 U8* s0 = (U8*) STRING(scan);
3868 U8* s_end = s0 + STR_LEN(scan);
3870 int total_count_delta = 0; /* Total delta number of characters that
3871 multi-char folds expand to */
3873 /* One pass is made over the node's string looking for all the
3874 * possibilities. To avoid some tests in the loop, there are two main
3875 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3880 if (OP(scan) == EXACTFL) {
3883 /* An EXACTFL node would already have been changed to another
3884 * node type unless there is at least one character in it that
3885 * is problematic; likely a character whose fold definition
3886 * won't be known until runtime, and so has yet to be folded.
3887 * For all but the UTF-8 locale, folds are 1-1 in length, but
3888 * to handle the UTF-8 case, we need to create a temporary
3889 * folded copy using UTF-8 locale rules in order to analyze it.
3890 * This is because our macros that look to see if a sequence is
3891 * a multi-char fold assume everything is folded (otherwise the
3892 * tests in those macros would be too complicated and slow).
3893 * Note that here, the non-problematic folds will have already
3894 * been done, so we can just copy such characters. We actually
3895 * don't completely fold the EXACTFL string. We skip the
3896 * unfolded multi-char folds, as that would just create work
3897 * below to figure out the size they already are */
3899 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3902 STRLEN s_len = UTF8SKIP(s);
3903 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3904 Copy(s, d, s_len, U8);
3907 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3908 *unfolded_multi_char = TRUE;
3909 Copy(s, d, s_len, U8);
3912 else if (isASCII(*s)) {
3913 *(d++) = toFOLD(*s);
3917 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
3923 /* Point the remainder of the routine to look at our temporary
3927 } /* End of creating folded copy of EXACTFL string */
3929 /* Examine the string for a multi-character fold sequence. UTF-8
3930 * patterns have all characters pre-folded by the time this code is
3932 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3933 length sequence we are looking for is 2 */
3935 int count = 0; /* How many characters in a multi-char fold */
3936 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3937 if (! len) { /* Not a multi-char fold: get next char */
3942 /* Nodes with 'ss' require special handling, except for
3943 * EXACTFA-ish for which there is no multi-char fold to this */
3944 if (len == 2 && *s == 's' && *(s+1) == 's'
3945 && OP(scan) != EXACTFA
3946 && OP(scan) != EXACTFA_NO_TRIE)
3949 if (OP(scan) != EXACTFL) {
3950 OP(scan) = EXACTFU_SS;
3954 else { /* Here is a generic multi-char fold. */
3955 U8* multi_end = s + len;
3957 /* Count how many characters are in it. In the case of
3958 * /aa, no folds which contain ASCII code points are
3959 * allowed, so check for those, and skip if found. */
3960 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3961 count = utf8_length(s, multi_end);
3965 while (s < multi_end) {
3968 goto next_iteration;
3978 /* The delta is how long the sequence is minus 1 (1 is how long
3979 * the character that folds to the sequence is) */
3980 total_count_delta += count - 1;
3984 /* We created a temporary folded copy of the string in EXACTFL
3985 * nodes. Therefore we need to be sure it doesn't go below zero,
3986 * as the real string could be shorter */
3987 if (OP(scan) == EXACTFL) {
3988 int total_chars = utf8_length((U8*) STRING(scan),
3989 (U8*) STRING(scan) + STR_LEN(scan));
3990 if (total_count_delta > total_chars) {
3991 total_count_delta = total_chars;
3995 *min_subtract += total_count_delta;
3998 else if (OP(scan) == EXACTFA) {
4000 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
4001 * fold to the ASCII range (and there are no existing ones in the
4002 * upper latin1 range). But, as outlined in the comments preceding
4003 * this function, we need to flag any occurrences of the sharp s.
4004 * This character forbids trie formation (because of added
4006 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4007 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4008 || UNICODE_DOT_DOT_VERSION > 0)
4010 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4011 OP(scan) = EXACTFA_NO_TRIE;
4012 *unfolded_multi_char = TRUE;
4020 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
4021 * folds that are all Latin1. As explained in the comments
4022 * preceding this function, we look also for the sharp s in EXACTF
4023 * and EXACTFL nodes; it can be in the final position. Otherwise
4024 * we can stop looking 1 byte earlier because have to find at least
4025 * two characters for a multi-fold */
4026 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4031 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4032 if (! len) { /* Not a multi-char fold. */
4033 if (*s == LATIN_SMALL_LETTER_SHARP_S
4034 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4036 *unfolded_multi_char = TRUE;
4043 && isALPHA_FOLD_EQ(*s, 's')
4044 && isALPHA_FOLD_EQ(*(s+1), 's'))
4047 /* EXACTF nodes need to know that the minimum length
4048 * changed so that a sharp s in the string can match this
4049 * ss in the pattern, but they remain EXACTF nodes, as they
4050 * won't match this unless the target string is is UTF-8,
4051 * which we don't know until runtime. EXACTFL nodes can't
4052 * transform into EXACTFU nodes */
4053 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4054 OP(scan) = EXACTFU_SS;
4058 *min_subtract += len - 1;
4066 /* Allow dumping but overwriting the collection of skipped
4067 * ops and/or strings with fake optimized ops */
4068 n = scan + NODE_SZ_STR(scan);
4076 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
4080 /* REx optimizer. Converts nodes into quicker variants "in place".
4081 Finds fixed substrings. */
4083 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4084 to the position after last scanned or to NULL. */
4086 #define INIT_AND_WITHP \
4087 assert(!and_withp); \
4088 Newx(and_withp,1, regnode_ssc); \
4089 SAVEFREEPV(and_withp)
4093 S_unwind_scan_frames(pTHX_ const void *p)
4095 scan_frame *f= (scan_frame *)p;
4097 scan_frame *n= f->next_frame;
4105 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4106 SSize_t *minlenp, SSize_t *deltap,
4111 regnode_ssc *and_withp,
4112 U32 flags, U32 depth)
4113 /* scanp: Start here (read-write). */
4114 /* deltap: Write maxlen-minlen here. */
4115 /* last: Stop before this one. */
4116 /* data: string data about the pattern */
4117 /* stopparen: treat close N as END */
4118 /* recursed: which subroutines have we recursed into */
4119 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4121 /* There must be at least this number of characters to match */
4124 regnode *scan = *scanp, *next;
4126 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4127 int is_inf_internal = 0; /* The studied chunk is infinite */
4128 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4129 scan_data_t data_fake;
4130 SV *re_trie_maxbuff = NULL;
4131 regnode *first_non_open = scan;
4132 SSize_t stopmin = SSize_t_MAX;
4133 scan_frame *frame = NULL;
4134 GET_RE_DEBUG_FLAGS_DECL;
4136 PERL_ARGS_ASSERT_STUDY_CHUNK;
4137 RExC_study_started= 1;
4141 while (first_non_open && OP(first_non_open) == OPEN)
4142 first_non_open=regnext(first_non_open);
4148 RExC_study_chunk_recursed_count++;
4150 DEBUG_OPTIMISE_MORE_r(
4152 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4153 depth, (long)stopparen,
4154 (unsigned long)RExC_study_chunk_recursed_count,
4155 (unsigned long)depth, (unsigned long)recursed_depth,
4158 if (recursed_depth) {
4161 for ( j = 0 ; j < recursed_depth ; j++ ) {
4162 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
4164 PAREN_TEST(RExC_study_chunk_recursed +
4165 ( j * RExC_study_chunk_recursed_bytes), i )
4168 !PAREN_TEST(RExC_study_chunk_recursed +
4169 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
4172 Perl_re_printf( aTHX_ " %d",(int)i);
4176 if ( j + 1 < recursed_depth ) {
4177 Perl_re_printf( aTHX_ ",");
4181 Perl_re_printf( aTHX_ "\n");
4184 while ( scan && OP(scan) != END && scan < last ){
4185 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4186 node length to get a real minimum (because
4187 the folded version may be shorter) */
4188 bool unfolded_multi_char = FALSE;
4189 /* Peephole optimizer: */
4190 DEBUG_STUDYDATA("Peep:", data, depth);
4191 DEBUG_PEEP("Peep", scan, depth);
4194 /* The reason we do this here is that we need to deal with things like
4195 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4196 * parsing code, as each (?:..) is handled by a different invocation of
4199 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
4201 /* Follow the next-chain of the current node and optimize
4202 away all the NOTHINGs from it. */
4203 if (OP(scan) != CURLYX) {
4204 const int max = (reg_off_by_arg[OP(scan)]
4206 /* I32 may be smaller than U16 on CRAYs! */
4207 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4208 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
4212 /* Skip NOTHING and LONGJMP. */
4213 while ((n = regnext(n))
4214 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4215 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
4216 && off + noff < max)
4218 if (reg_off_by_arg[OP(scan)])
4221 NEXT_OFF(scan) = off;
4224 /* The principal pseudo-switch. Cannot be a switch, since we
4225 look into several different things. */
4226 if ( OP(scan) == DEFINEP ) {
4228 SSize_t deltanext = 0;
4229 SSize_t fake_last_close = 0;
4230 I32 f = SCF_IN_DEFINE;
4232 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4233 scan = regnext(scan);
4234 assert( OP(scan) == IFTHEN );
4235 DEBUG_PEEP("expect IFTHEN", scan, depth);
4237 data_fake.last_closep= &fake_last_close;
4239 next = regnext(scan);
4240 scan = NEXTOPER(NEXTOPER(scan));
4241 DEBUG_PEEP("scan", scan, depth);
4242 DEBUG_PEEP("next", next, depth);
4244 /* we suppose the run is continuous, last=next...
4245 * NOTE we dont use the return here! */
4246 (void)study_chunk(pRExC_state, &scan, &minlen,
4247 &deltanext, next, &data_fake, stopparen,
4248 recursed_depth, NULL, f, depth+1);
4253 OP(scan) == BRANCH ||
4254 OP(scan) == BRANCHJ ||
4257 next = regnext(scan);
4260 /* The op(next)==code check below is to see if we
4261 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4262 * IFTHEN is special as it might not appear in pairs.
4263 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4264 * we dont handle it cleanly. */
4265 if (OP(next) == code || code == IFTHEN) {
4266 /* NOTE - There is similar code to this block below for
4267 * handling TRIE nodes on a re-study. If you change stuff here
4268 * check there too. */
4269 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
4271 regnode * const startbranch=scan;
4273 if (flags & SCF_DO_SUBSTR) {
4274 /* Cannot merge strings after this. */
4275 scan_commit(pRExC_state, data, minlenp, is_inf);
4278 if (flags & SCF_DO_STCLASS)
4279 ssc_init_zero(pRExC_state, &accum);
4281 while (OP(scan) == code) {
4282 SSize_t deltanext, minnext, fake;
4284 regnode_ssc this_class;
4286 DEBUG_PEEP("Branch", scan, depth);
4289 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4291 data_fake.whilem_c = data->whilem_c;
4292 data_fake.last_closep = data->last_closep;
4295 data_fake.last_closep = &fake;
4297 data_fake.pos_delta = delta;
4298 next = regnext(scan);
4300 scan = NEXTOPER(scan); /* everything */
4301 if (code != BRANCH) /* everything but BRANCH */
4302 scan = NEXTOPER(scan);
4304 if (flags & SCF_DO_STCLASS) {
4305 ssc_init(pRExC_state, &this_class);
4306 data_fake.start_class = &this_class;
4307 f = SCF_DO_STCLASS_AND;
4309 if (flags & SCF_WHILEM_VISITED_POS)
4310 f |= SCF_WHILEM_VISITED_POS;
4312 /* we suppose the run is continuous, last=next...*/
4313 minnext = study_chunk(pRExC_state, &scan, minlenp,
4314 &deltanext, next, &data_fake, stopparen,
4315 recursed_depth, NULL, f,depth+1);
4319 if (deltanext == SSize_t_MAX) {
4320 is_inf = is_inf_internal = 1;
4322 } else if (max1 < minnext + deltanext)
4323 max1 = minnext + deltanext;
4325 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4327 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4328 if ( stopmin > minnext)
4329 stopmin = min + min1;
4330 flags &= ~SCF_DO_SUBSTR;
4332 data->flags |= SCF_SEEN_ACCEPT;
4335 if (data_fake.flags & SF_HAS_EVAL)
4336 data->flags |= SF_HAS_EVAL;
4337 data->whilem_c = data_fake.whilem_c;
4339 if (flags & SCF_DO_STCLASS)
4340 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4342 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4344 if (flags & SCF_DO_SUBSTR) {
4345 data->pos_min += min1;
4346 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
4347 data->pos_delta = SSize_t_MAX;
4349 data->pos_delta += max1 - min1;
4350 if (max1 != min1 || is_inf)
4351 data->longest = &(data->longest_float);
4354 if (delta == SSize_t_MAX
4355 || SSize_t_MAX - delta - (max1 - min1) < 0)
4356 delta = SSize_t_MAX;
4358 delta += max1 - min1;
4359 if (flags & SCF_DO_STCLASS_OR) {
4360 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4362 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4363 flags &= ~SCF_DO_STCLASS;
4366 else if (flags & SCF_DO_STCLASS_AND) {
4368 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4369 flags &= ~SCF_DO_STCLASS;
4372 /* Switch to OR mode: cache the old value of
4373 * data->start_class */
4375 StructCopy(data->start_class, and_withp, regnode_ssc);
4376 flags &= ~SCF_DO_STCLASS_AND;
4377 StructCopy(&accum, data->start_class, regnode_ssc);
4378 flags |= SCF_DO_STCLASS_OR;
4382 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4383 OP( startbranch ) == BRANCH )
4387 Assuming this was/is a branch we are dealing with: 'scan'
4388 now points at the item that follows the branch sequence,
4389 whatever it is. We now start at the beginning of the
4390 sequence and look for subsequences of
4396 which would be constructed from a pattern like
4399 If we can find such a subsequence we need to turn the first
4400 element into a trie and then add the subsequent branch exact
4401 strings to the trie.
4405 1. patterns where the whole set of branches can be
4408 2. patterns where only a subset can be converted.
4410 In case 1 we can replace the whole set with a single regop
4411 for the trie. In case 2 we need to keep the start and end
4414 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4415 becomes BRANCH TRIE; BRANCH X;
4417 There is an additional case, that being where there is a
4418 common prefix, which gets split out into an EXACT like node
4419 preceding the TRIE node.
4421 If x(1..n)==tail then we can do a simple trie, if not we make
4422 a "jump" trie, such that when we match the appropriate word
4423 we "jump" to the appropriate tail node. Essentially we turn
4424 a nested if into a case structure of sorts.
4429 if (!re_trie_maxbuff) {
4430 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4431 if (!SvIOK(re_trie_maxbuff))
4432 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4434 if ( SvIV(re_trie_maxbuff)>=0 ) {
4436 regnode *first = (regnode *)NULL;
4437 regnode *last = (regnode *)NULL;
4438 regnode *tail = scan;
4442 /* var tail is used because there may be a TAIL
4443 regop in the way. Ie, the exacts will point to the
4444 thing following the TAIL, but the last branch will
4445 point at the TAIL. So we advance tail. If we
4446 have nested (?:) we may have to move through several
4450 while ( OP( tail ) == TAIL ) {
4451 /* this is the TAIL generated by (?:) */
4452 tail = regnext( tail );
4456 DEBUG_TRIE_COMPILE_r({
4457 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4458 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4460 "Looking for TRIE'able sequences. Tail node is ",
4461 (UV)(tail - RExC_emit_start),
4462 SvPV_nolen_const( RExC_mysv )
4468 Step through the branches
4469 cur represents each branch,
4470 noper is the first thing to be matched as part
4472 noper_next is the regnext() of that node.
4474 We normally handle a case like this
4475 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4476 support building with NOJUMPTRIE, which restricts
4477 the trie logic to structures like /FOO|BAR/.
4479 If noper is a trieable nodetype then the branch is
4480 a possible optimization target. If we are building
4481 under NOJUMPTRIE then we require that noper_next is
4482 the same as scan (our current position in the regex
4485 Once we have two or more consecutive such branches
4486 we can create a trie of the EXACT's contents and
4487 stitch it in place into the program.
4489 If the sequence represents all of the branches in
4490 the alternation we replace the entire thing with a
4493 Otherwise when it is a subsequence we need to
4494 stitch it in place and replace only the relevant
4495 branches. This means the first branch has to remain
4496 as it is used by the alternation logic, and its
4497 next pointer, and needs to be repointed at the item
4498 on the branch chain following the last branch we
4499 have optimized away.
4501 This could be either a BRANCH, in which case the
4502 subsequence is internal, or it could be the item
4503 following the branch sequence in which case the
4504 subsequence is at the end (which does not
4505 necessarily mean the first node is the start of the
4508 TRIE_TYPE(X) is a define which maps the optype to a
4512 ----------------+-----------
4516 EXACTFU_SS | EXACTFU
4519 EXACTFLU8 | EXACTFLU8
4523 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4525 : ( EXACT == (X) ) \
4527 : ( EXACTFU == (X) || EXACTFU_SS == (X) ) \
4529 : ( EXACTFA == (X) ) \
4531 : ( EXACTL == (X) ) \
4533 : ( EXACTFLU8 == (X) ) \
4537 /* dont use tail as the end marker for this traverse */
4538 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4539 regnode * const noper = NEXTOPER( cur );
4540 U8 noper_type = OP( noper );
4541 U8 noper_trietype = TRIE_TYPE( noper_type );
4542 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4543 regnode * const noper_next = regnext( noper );
4544 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4545 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4548 DEBUG_TRIE_COMPILE_r({
4549 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4550 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4552 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4554 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4555 Perl_re_printf( aTHX_ " -> %d:%s",
4556 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
4559 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4560 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
4561 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
4563 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
4564 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4565 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4569 /* Is noper a trieable nodetype that can be merged
4570 * with the current trie (if there is one)? */
4574 ( noper_trietype == NOTHING )
4575 || ( trietype == NOTHING )
4576 || ( trietype == noper_trietype )
4579 && noper_next >= tail
4583 /* Handle mergable triable node Either we are
4584 * the first node in a new trieable sequence,
4585 * in which case we do some bookkeeping,
4586 * otherwise we update the end pointer. */
4589 if ( noper_trietype == NOTHING ) {
4590 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4591 regnode * const noper_next = regnext( noper );
4592 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4593 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4596 if ( noper_next_trietype ) {
4597 trietype = noper_next_trietype;
4598 } else if (noper_next_type) {
4599 /* a NOTHING regop is 1 regop wide.
4600 * We need at least two for a trie
4601 * so we can't merge this in */
4605 trietype = noper_trietype;
4608 if ( trietype == NOTHING )
4609 trietype = noper_trietype;
4614 } /* end handle mergable triable node */
4616 /* handle unmergable node -
4617 * noper may either be a triable node which can
4618 * not be tried together with the current trie,
4619 * or a non triable node */
4621 /* If last is set and trietype is not
4622 * NOTHING then we have found at least two
4623 * triable branch sequences in a row of a
4624 * similar trietype so we can turn them
4625 * into a trie. If/when we allow NOTHING to
4626 * start a trie sequence this condition
4627 * will be required, and it isn't expensive
4628 * so we leave it in for now. */
4629 if ( trietype && trietype != NOTHING )
4630 make_trie( pRExC_state,
4631 startbranch, first, cur, tail,
4632 count, trietype, depth+1 );
4633 last = NULL; /* note: we clear/update
4634 first, trietype etc below,
4635 so we dont do it here */
4639 && noper_next >= tail
4642 /* noper is triable, so we can start a new
4646 trietype = noper_trietype;
4648 /* if we already saw a first but the
4649 * current node is not triable then we have
4650 * to reset the first information. */
4655 } /* end handle unmergable node */
4656 } /* loop over branches */
4657 DEBUG_TRIE_COMPILE_r({
4658 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4659 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
4660 depth+1, SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4661 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
4662 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4663 PL_reg_name[trietype]
4667 if ( last && trietype ) {
4668 if ( trietype != NOTHING ) {
4669 /* the last branch of the sequence was part of
4670 * a trie, so we have to construct it here
4671 * outside of the loop */
4672 made= make_trie( pRExC_state, startbranch,
4673 first, scan, tail, count,
4674 trietype, depth+1 );
4675 #ifdef TRIE_STUDY_OPT
4676 if ( ((made == MADE_EXACT_TRIE &&
4677 startbranch == first)
4678 || ( first_non_open == first )) &&
4680 flags |= SCF_TRIE_RESTUDY;
4681 if ( startbranch == first
4684 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4689 /* at this point we know whatever we have is a
4690 * NOTHING sequence/branch AND if 'startbranch'
4691 * is 'first' then we can turn the whole thing
4694 if ( startbranch == first ) {
4696 /* the entire thing is a NOTHING sequence,
4697 * something like this: (?:|) So we can
4698 * turn it into a plain NOTHING op. */
4699 DEBUG_TRIE_COMPILE_r({
4700 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4701 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
4703 SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4706 OP(startbranch)= NOTHING;
4707 NEXT_OFF(startbranch)= tail - startbranch;
4708 for ( opt= startbranch + 1; opt < tail ; opt++ )
4712 } /* end if ( last) */
4713 } /* TRIE_MAXBUF is non zero */
4718 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4719 scan = NEXTOPER(NEXTOPER(scan));
4720 } else /* single branch is optimized. */
4721 scan = NEXTOPER(scan);
4723 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
4725 regnode *start = NULL;
4726 regnode *end = NULL;
4727 U32 my_recursed_depth= recursed_depth;
4729 if (OP(scan) != SUSPEND) { /* GOSUB */
4730 /* Do setup, note this code has side effects beyond
4731 * the rest of this block. Specifically setting
4732 * RExC_recurse[] must happen at least once during
4735 RExC_recurse[ARG2L(scan)] = scan;
4736 start = RExC_open_parens[paren];
4737 end = RExC_close_parens[paren];
4739 /* NOTE we MUST always execute the above code, even
4740 * if we do nothing with a GOSUB */
4742 ( flags & SCF_IN_DEFINE )
4745 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
4747 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
4750 /* no need to do anything here if we are in a define. */
4751 /* or we are after some kind of infinite construct
4752 * so we can skip recursing into this item.
4753 * Since it is infinite we will not change the maxlen
4754 * or delta, and if we miss something that might raise
4755 * the minlen it will merely pessimise a little.
4757 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
4758 * might result in a minlen of 1 and not of 4,
4759 * but this doesn't make us mismatch, just try a bit
4760 * harder than we should.
4762 scan= regnext(scan);
4769 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4771 /* it is quite possible that there are more efficient ways
4772 * to do this. We maintain a bitmap per level of recursion
4773 * of which patterns we have entered so we can detect if a
4774 * pattern creates a possible infinite loop. When we
4775 * recurse down a level we copy the previous levels bitmap
4776 * down. When we are at recursion level 0 we zero the top
4777 * level bitmap. It would be nice to implement a different
4778 * more efficient way of doing this. In particular the top
4779 * level bitmap may be unnecessary.
4781 if (!recursed_depth) {
4782 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4784 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4785 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4786 RExC_study_chunk_recursed_bytes, U8);
4788 /* we havent recursed into this paren yet, so recurse into it */
4789 DEBUG_STUDYDATA("gosub-set:", data,depth);
4790 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4791 my_recursed_depth= recursed_depth + 1;
4793 DEBUG_STUDYDATA("gosub-inf:", data,depth);
4794 /* some form of infinite recursion, assume infinite length
4796 if (flags & SCF_DO_SUBSTR) {
4797 scan_commit(pRExC_state, data, minlenp, is_inf);
4798 data->longest = &(data->longest_float);
4800 is_inf = is_inf_internal = 1;
4801 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4802 ssc_anything(data->start_class);
4803 flags &= ~SCF_DO_STCLASS;
4805 start= NULL; /* reset start so we dont recurse later on. */
4810 end = regnext(scan);
4813 scan_frame *newframe;
4815 if (!RExC_frame_last) {
4816 Newxz(newframe, 1, scan_frame);
4817 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
4818 RExC_frame_head= newframe;
4820 } else if (!RExC_frame_last->next_frame) {
4821 Newxz(newframe,1,scan_frame);
4822 RExC_frame_last->next_frame= newframe;
4823 newframe->prev_frame= RExC_frame_last;
4826 newframe= RExC_frame_last->next_frame;
4828 RExC_frame_last= newframe;
4830 newframe->next_regnode = regnext(scan);
4831 newframe->last_regnode = last;
4832 newframe->stopparen = stopparen;
4833 newframe->prev_recursed_depth = recursed_depth;
4834 newframe->this_prev_frame= frame;
4836 DEBUG_STUDYDATA("frame-new:",data,depth);
4837 DEBUG_PEEP("fnew", scan, depth);
4844 recursed_depth= my_recursed_depth;
4849 else if (OP(scan) == EXACT || OP(scan) == EXACTL) {
4850 SSize_t l = STR_LEN(scan);
4853 const U8 * const s = (U8*)STRING(scan);
4854 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4855 l = utf8_length(s, s + l);
4857 uc = *((U8*)STRING(scan));
4860 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4861 /* The code below prefers earlier match for fixed
4862 offset, later match for variable offset. */
4863 if (data->last_end == -1) { /* Update the start info. */
4864 data->last_start_min = data->pos_min;
4865 data->last_start_max = is_inf
4866 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4868 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4870 SvUTF8_on(data->last_found);
4872 SV * const sv = data->last_found;
4873 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4874 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4875 if (mg && mg->mg_len >= 0)
4876 mg->mg_len += utf8_length((U8*)STRING(scan),
4877 (U8*)STRING(scan)+STR_LEN(scan));
4879 data->last_end = data->pos_min + l;
4880 data->pos_min += l; /* As in the first entry. */
4881 data->flags &= ~SF_BEFORE_EOL;
4884 /* ANDing the code point leaves at most it, and not in locale, and
4885 * can't match null string */
4886 if (flags & SCF_DO_STCLASS_AND) {
4887 ssc_cp_and(data->start_class, uc);
4888 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4889 ssc_clear_locale(data->start_class);
4891 else if (flags & SCF_DO_STCLASS_OR) {
4892 ssc_add_cp(data->start_class, uc);
4893 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4895 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4896 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4898 flags &= ~SCF_DO_STCLASS;
4900 else if (PL_regkind[OP(scan)] == EXACT) {
4901 /* But OP != EXACT!, so is EXACTFish */
4902 SSize_t l = STR_LEN(scan);
4903 const U8 * s = (U8*)STRING(scan);
4905 /* Search for fixed substrings supports EXACT only. */
4906 if (flags & SCF_DO_SUBSTR) {
4908 scan_commit(pRExC_state, data, minlenp, is_inf);
4911 l = utf8_length(s, s + l);
4913 if (unfolded_multi_char) {
4914 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4916 min += l - min_subtract;
4918 delta += min_subtract;
4919 if (flags & SCF_DO_SUBSTR) {
4920 data->pos_min += l - min_subtract;
4921 if (data->pos_min < 0) {
4924 data->pos_delta += min_subtract;
4926 data->longest = &(data->longest_float);
4930 if (flags & SCF_DO_STCLASS) {
4931 SV* EXACTF_invlist = _make_exactf_invlist(pRExC_state, scan);
4933 assert(EXACTF_invlist);
4934 if (flags & SCF_DO_STCLASS_AND) {
4935 if (OP(scan) != EXACTFL)
4936 ssc_clear_locale(data->start_class);
4937 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4938 ANYOF_POSIXL_ZERO(data->start_class);
4939 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4941 else { /* SCF_DO_STCLASS_OR */
4942 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4943 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4945 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4946 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4948 flags &= ~SCF_DO_STCLASS;
4949 SvREFCNT_dec(EXACTF_invlist);
4952 else if (REGNODE_VARIES(OP(scan))) {
4953 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4954 I32 fl = 0, f = flags;
4955 regnode * const oscan = scan;
4956 regnode_ssc this_class;
4957 regnode_ssc *oclass = NULL;
4958 I32 next_is_eval = 0;
4960 switch (PL_regkind[OP(scan)]) {
4961 case WHILEM: /* End of (?:...)* . */
4962 scan = NEXTOPER(scan);
4965 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4966 next = NEXTOPER(scan);
4967 if (OP(next) == EXACT
4968 || OP(next) == EXACTL
4969 || (flags & SCF_DO_STCLASS))
4972 maxcount = REG_INFTY;
4973 next = regnext(scan);
4974 scan = NEXTOPER(scan);
4978 if (flags & SCF_DO_SUBSTR)
4983 if (flags & SCF_DO_STCLASS) {
4985 maxcount = REG_INFTY;
4986 next = regnext(scan);
4987 scan = NEXTOPER(scan);
4990 if (flags & SCF_DO_SUBSTR) {
4991 scan_commit(pRExC_state, data, minlenp, is_inf);
4992 /* Cannot extend fixed substrings */
4993 data->longest = &(data->longest_float);
4995 is_inf = is_inf_internal = 1;
4996 scan = regnext(scan);
4997 goto optimize_curly_tail;
4999 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
5000 && (scan->flags == stopparen))
5005 mincount = ARG1(scan);
5006 maxcount = ARG2(scan);
5008 next = regnext(scan);
5009 if (OP(scan) == CURLYX) {
5010 I32 lp = (data ? *(data->last_closep) : 0);
5011 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5013 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5014 next_is_eval = (OP(scan) == EVAL);
5016 if (flags & SCF_DO_SUBSTR) {
5018 scan_commit(pRExC_state, data, minlenp, is_inf);
5019 /* Cannot extend fixed substrings */
5020 pos_before = data->pos_min;
5024 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5026 data->flags |= SF_IS_INF;
5028 if (flags & SCF_DO_STCLASS) {
5029 ssc_init(pRExC_state, &this_class);
5030 oclass = data->start_class;
5031 data->start_class = &this_class;
5032 f |= SCF_DO_STCLASS_AND;
5033 f &= ~SCF_DO_STCLASS_OR;
5035 /* Exclude from super-linear cache processing any {n,m}
5036 regops for which the combination of input pos and regex
5037 pos is not enough information to determine if a match
5040 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5041 regex pos at the \s*, the prospects for a match depend not
5042 only on the input position but also on how many (bar\s*)
5043 repeats into the {4,8} we are. */
5044 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5045 f &= ~SCF_WHILEM_VISITED_POS;
5047 /* This will finish on WHILEM, setting scan, or on NULL: */
5048 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5049 last, data, stopparen, recursed_depth, NULL,
5051 ? (f & ~SCF_DO_SUBSTR)
5055 if (flags & SCF_DO_STCLASS)
5056 data->start_class = oclass;
5057 if (mincount == 0 || minnext == 0) {
5058 if (flags & SCF_DO_STCLASS_OR) {
5059 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5061 else if (flags & SCF_DO_STCLASS_AND) {
5062 /* Switch to OR mode: cache the old value of
5063 * data->start_class */
5065 StructCopy(data->start_class, and_withp, regnode_ssc);
5066 flags &= ~SCF_DO_STCLASS_AND;
5067 StructCopy(&this_class, data->start_class, regnode_ssc);
5068 flags |= SCF_DO_STCLASS_OR;
5069 ANYOF_FLAGS(data->start_class)
5070 |= SSC_MATCHES_EMPTY_STRING;
5072 } else { /* Non-zero len */
5073 if (flags & SCF_DO_STCLASS_OR) {
5074 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5075 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5077 else if (flags & SCF_DO_STCLASS_AND)
5078 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5079 flags &= ~SCF_DO_STCLASS;
5081 if (!scan) /* It was not CURLYX, but CURLY. */
5083 if (!(flags & SCF_TRIE_DOING_RESTUDY)
5084 /* ? quantifier ok, except for (?{ ... }) */
5085 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5086 && (minnext == 0) && (deltanext == 0)
5087 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5088 && maxcount <= REG_INFTY/3) /* Complement check for big
5091 /* Fatal warnings may leak the regexp without this: */
5092 SAVEFREESV(RExC_rx_sv);
5093 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5094 "Quantifier unexpected on zero-length expression "
5095 "in regex m/%" UTF8f "/",
5096 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5098 (void)ReREFCNT_inc(RExC_rx_sv);
5101 min += minnext * mincount;
5102 is_inf_internal |= deltanext == SSize_t_MAX
5103 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5104 is_inf |= is_inf_internal;
5106 delta = SSize_t_MAX;
5108 delta += (minnext + deltanext) * maxcount
5109 - minnext * mincount;
5111 /* Try powerful optimization CURLYX => CURLYN. */
5112 if ( OP(oscan) == CURLYX && data
5113 && data->flags & SF_IN_PAR
5114 && !(data->flags & SF_HAS_EVAL)
5115 && !deltanext && minnext == 1 ) {
5116 /* Try to optimize to CURLYN. */
5117 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5118 regnode * const nxt1 = nxt;
5125 if (!REGNODE_SIMPLE(OP(nxt))
5126 && !(PL_regkind[OP(nxt)] == EXACT
5127 && STR_LEN(nxt) == 1))
5133 if (OP(nxt) != CLOSE)
5135 if (RExC_open_parens) {
5136 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5137 RExC_close_parens[ARG(nxt1)]=nxt+2; /*close->while*/
5139 /* Now we know that nxt2 is the only contents: */
5140 oscan->flags = (U8)ARG(nxt);
5142 OP(nxt1) = NOTHING; /* was OPEN. */
5145 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5146 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5147 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5148 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5149 OP(nxt + 1) = OPTIMIZED; /* was count. */
5150 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5155 /* Try optimization CURLYX => CURLYM. */
5156 if ( OP(oscan) == CURLYX && data
5157 && !(data->flags & SF_HAS_PAR)
5158 && !(data->flags & SF_HAS_EVAL)
5159 && !deltanext /* atom is fixed width */
5160 && minnext != 0 /* CURLYM can't handle zero width */
5162 /* Nor characters whose fold at run-time may be
5163 * multi-character */
5164 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5166 /* XXXX How to optimize if data == 0? */
5167 /* Optimize to a simpler form. */
5168 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5172 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5173 && (OP(nxt2) != WHILEM))
5175 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5176 /* Need to optimize away parenths. */
5177 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5178 /* Set the parenth number. */
5179 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5181 oscan->flags = (U8)ARG(nxt);
5182 if (RExC_open_parens) {
5183 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5184 RExC_close_parens[ARG(nxt1)]=nxt2+1; /*close->NOTHING*/
5186 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5187 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5190 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5191 OP(nxt + 1) = OPTIMIZED; /* was count. */
5192 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5193 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5196 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5197 regnode *nnxt = regnext(nxt1);
5199 if (reg_off_by_arg[OP(nxt1)])
5200 ARG_SET(nxt1, nxt2 - nxt1);
5201 else if (nxt2 - nxt1 < U16_MAX)
5202 NEXT_OFF(nxt1) = nxt2 - nxt1;
5204 OP(nxt) = NOTHING; /* Cannot beautify */
5209 /* Optimize again: */
5210 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5211 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
5216 else if ((OP(oscan) == CURLYX)
5217 && (flags & SCF_WHILEM_VISITED_POS)
5218 /* See the comment on a similar expression above.
5219 However, this time it's not a subexpression
5220 we care about, but the expression itself. */
5221 && (maxcount == REG_INFTY)
5222 && data && ++data->whilem_c < 16) {
5223 /* This stays as CURLYX, we can put the count/of pair. */
5224 /* Find WHILEM (as in regexec.c) */
5225 regnode *nxt = oscan + NEXT_OFF(oscan);
5227 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5229 PREVOPER(nxt)->flags = (U8)(data->whilem_c
5230 | (RExC_whilem_seen << 4)); /* On WHILEM */
5232 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5234 if (flags & SCF_DO_SUBSTR) {
5235 SV *last_str = NULL;
5236 STRLEN last_chrs = 0;
5237 int counted = mincount != 0;
5239 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5241 SSize_t b = pos_before >= data->last_start_min
5242 ? pos_before : data->last_start_min;
5244 const char * const s = SvPV_const(data->last_found, l);
5245 SSize_t old = b - data->last_start_min;
5248 old = utf8_hop((U8*)s, old) - (U8*)s;
5250 /* Get the added string: */
5251 last_str = newSVpvn_utf8(s + old, l, UTF);
5252 last_chrs = UTF ? utf8_length((U8*)(s + old),
5253 (U8*)(s + old + l)) : l;
5254 if (deltanext == 0 && pos_before == b) {
5255 /* What was added is a constant string */
5258 SvGROW(last_str, (mincount * l) + 1);
5259 repeatcpy(SvPVX(last_str) + l,
5260 SvPVX_const(last_str), l,
5262 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5263 /* Add additional parts. */
5264 SvCUR_set(data->last_found,
5265 SvCUR(data->last_found) - l);
5266 sv_catsv(data->last_found, last_str);
5268 SV * sv = data->last_found;
5270 SvUTF8(sv) && SvMAGICAL(sv) ?
5271 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5272 if (mg && mg->mg_len >= 0)
5273 mg->mg_len += last_chrs * (mincount-1);
5275 last_chrs *= mincount;
5276 data->last_end += l * (mincount - 1);
5279 /* start offset must point into the last copy */
5280 data->last_start_min += minnext * (mincount - 1);
5281 data->last_start_max =
5284 : data->last_start_max +
5285 (maxcount - 1) * (minnext + data->pos_delta);
5288 /* It is counted once already... */
5289 data->pos_min += minnext * (mincount - counted);
5291 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5292 " SSize_t_MAX=%" UVuf " minnext=%" UVuf
5293 " maxcount=%" UVuf " mincount=%" UVuf "\n",
5294 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
5296 if (deltanext != SSize_t_MAX)
5297 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5298 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5299 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
5301 if (deltanext == SSize_t_MAX
5302 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
5303 data->pos_delta = SSize_t_MAX;
5305 data->pos_delta += - counted * deltanext +
5306 (minnext + deltanext) * maxcount - minnext * mincount;
5307 if (mincount != maxcount) {
5308 /* Cannot extend fixed substrings found inside
5310 scan_commit(pRExC_state, data, minlenp, is_inf);
5311 if (mincount && last_str) {
5312 SV * const sv = data->last_found;
5313 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5314 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5318 sv_setsv(sv, last_str);
5319 data->last_end = data->pos_min;
5320 data->last_start_min = data->pos_min - last_chrs;
5321 data->last_start_max = is_inf
5323 : data->pos_min + data->pos_delta - last_chrs;
5325 data->longest = &(data->longest_float);
5327 SvREFCNT_dec(last_str);
5329 if (data && (fl & SF_HAS_EVAL))
5330 data->flags |= SF_HAS_EVAL;
5331 optimize_curly_tail:
5332 if (OP(oscan) != CURLYX) {
5333 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
5335 NEXT_OFF(oscan) += NEXT_OFF(next);
5341 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5346 if (flags & SCF_DO_SUBSTR) {
5347 /* Cannot expect anything... */
5348 scan_commit(pRExC_state, data, minlenp, is_inf);
5349 data->longest = &(data->longest_float);
5351 is_inf = is_inf_internal = 1;
5352 if (flags & SCF_DO_STCLASS_OR) {
5353 if (OP(scan) == CLUMP) {
5354 /* Actually is any start char, but very few code points
5355 * aren't start characters */
5356 ssc_match_all_cp(data->start_class);
5359 ssc_anything(data->start_class);
5362 flags &= ~SCF_DO_STCLASS;
5366 else if (OP(scan) == LNBREAK) {
5367 if (flags & SCF_DO_STCLASS) {
5368 if (flags & SCF_DO_STCLASS_AND) {
5369 ssc_intersection(data->start_class,
5370 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5371 ssc_clear_locale(data->start_class);
5372 ANYOF_FLAGS(data->start_class)
5373 &= ~SSC_MATCHES_EMPTY_STRING;
5375 else if (flags & SCF_DO_STCLASS_OR) {
5376 ssc_union(data->start_class,
5377 PL_XPosix_ptrs[_CC_VERTSPACE],
5379 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5381 /* See commit msg for
5382 * 749e076fceedeb708a624933726e7989f2302f6a */
5383 ANYOF_FLAGS(data->start_class)
5384 &= ~SSC_MATCHES_EMPTY_STRING;
5386 flags &= ~SCF_DO_STCLASS;
5389 if (delta != SSize_t_MAX)
5390 delta++; /* Because of the 2 char string cr-lf */
5391 if (flags & SCF_DO_SUBSTR) {
5392 /* Cannot expect anything... */
5393 scan_commit(pRExC_state, data, minlenp, is_inf);
5395 data->pos_delta += 1;
5396 data->longest = &(data->longest_float);
5399 else if (REGNODE_SIMPLE(OP(scan))) {
5401 if (flags & SCF_DO_SUBSTR) {
5402 scan_commit(pRExC_state, data, minlenp, is_inf);
5406 if (flags & SCF_DO_STCLASS) {
5408 SV* my_invlist = NULL;
5411 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5412 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5414 /* Some of the logic below assumes that switching
5415 locale on will only add false positives. */
5420 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5424 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5425 ssc_match_all_cp(data->start_class);
5430 SV* REG_ANY_invlist = _new_invlist(2);
5431 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5433 if (flags & SCF_DO_STCLASS_OR) {
5434 ssc_union(data->start_class,
5436 TRUE /* TRUE => invert, hence all but \n
5440 else if (flags & SCF_DO_STCLASS_AND) {
5441 ssc_intersection(data->start_class,
5443 TRUE /* TRUE => invert */
5445 ssc_clear_locale(data->start_class);
5447 SvREFCNT_dec_NN(REG_ANY_invlist);
5454 if (flags & SCF_DO_STCLASS_AND)
5455 ssc_and(pRExC_state, data->start_class,
5456 (regnode_charclass *) scan);
5458 ssc_or(pRExC_state, data->start_class,
5459 (regnode_charclass *) scan);
5467 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5468 if (flags & SCF_DO_STCLASS_AND) {
5469 bool was_there = cBOOL(
5470 ANYOF_POSIXL_TEST(data->start_class,
5472 ANYOF_POSIXL_ZERO(data->start_class);
5473 if (was_there) { /* Do an AND */
5474 ANYOF_POSIXL_SET(data->start_class, namedclass);
5476 /* No individual code points can now match */
5477 data->start_class->invlist
5478 = sv_2mortal(_new_invlist(0));
5481 int complement = namedclass + ((invert) ? -1 : 1);
5483 assert(flags & SCF_DO_STCLASS_OR);
5485 /* If the complement of this class was already there,
5486 * the result is that they match all code points,
5487 * (\d + \D == everything). Remove the classes from
5488 * future consideration. Locale is not relevant in
5490 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5491 ssc_match_all_cp(data->start_class);
5492 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5493 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5495 else { /* The usual case; just add this class to the
5497 ANYOF_POSIXL_SET(data->start_class, namedclass);
5502 case NPOSIXA: /* For these, we always know the exact set of
5507 if (FLAGS(scan) == _CC_ASCII) {
5508 my_invlist = invlist_clone(PL_XPosix_ptrs[_CC_ASCII]);
5511 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
5512 PL_XPosix_ptrs[_CC_ASCII],
5523 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5525 /* NPOSIXD matches all upper Latin1 code points unless the
5526 * target string being matched is UTF-8, which is
5527 * unknowable until match time. Since we are going to
5528 * invert, we want to get rid of all of them so that the
5529 * inversion will match all */
5530 if (OP(scan) == NPOSIXD) {
5531 _invlist_subtract(my_invlist, PL_UpperLatin1,
5537 if (flags & SCF_DO_STCLASS_AND) {
5538 ssc_intersection(data->start_class, my_invlist, invert);
5539 ssc_clear_locale(data->start_class);
5542 assert(flags & SCF_DO_STCLASS_OR);
5543 ssc_union(data->start_class, my_invlist, invert);
5545 SvREFCNT_dec(my_invlist);
5547 if (flags & SCF_DO_STCLASS_OR)
5548 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5549 flags &= ~SCF_DO_STCLASS;
5552 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5553 data->flags |= (OP(scan) == MEOL
5556 scan_commit(pRExC_state, data, minlenp, is_inf);
5559 else if ( PL_regkind[OP(scan)] == BRANCHJ
5560 /* Lookbehind, or need to calculate parens/evals/stclass: */
5561 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5562 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5564 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5565 || OP(scan) == UNLESSM )
5567 /* Negative Lookahead/lookbehind
5568 In this case we can't do fixed string optimisation.
5571 SSize_t deltanext, minnext, fake = 0;
5576 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5578 data_fake.whilem_c = data->whilem_c;
5579 data_fake.last_closep = data->last_closep;
5582 data_fake.last_closep = &fake;
5583 data_fake.pos_delta = delta;
5584 if ( flags & SCF_DO_STCLASS && !scan->flags
5585 && OP(scan) == IFMATCH ) { /* Lookahead */
5586 ssc_init(pRExC_state, &intrnl);
5587 data_fake.start_class = &intrnl;
5588 f |= SCF_DO_STCLASS_AND;
5590 if (flags & SCF_WHILEM_VISITED_POS)
5591 f |= SCF_WHILEM_VISITED_POS;
5592 next = regnext(scan);
5593 nscan = NEXTOPER(NEXTOPER(scan));
5594 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5595 last, &data_fake, stopparen,
5596 recursed_depth, NULL, f, depth+1);
5599 FAIL("Variable length lookbehind not implemented");
5601 else if (minnext > (I32)U8_MAX) {
5602 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5605 scan->flags = (U8)minnext;
5608 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5610 if (data_fake.flags & SF_HAS_EVAL)
5611 data->flags |= SF_HAS_EVAL;
5612 data->whilem_c = data_fake.whilem_c;
5614 if (f & SCF_DO_STCLASS_AND) {
5615 if (flags & SCF_DO_STCLASS_OR) {
5616 /* OR before, AND after: ideally we would recurse with
5617 * data_fake to get the AND applied by study of the
5618 * remainder of the pattern, and then derecurse;
5619 * *** HACK *** for now just treat as "no information".
5620 * See [perl #56690].
5622 ssc_init(pRExC_state, data->start_class);
5624 /* AND before and after: combine and continue. These
5625 * assertions are zero-length, so can match an EMPTY
5627 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5628 ANYOF_FLAGS(data->start_class)
5629 |= SSC_MATCHES_EMPTY_STRING;
5633 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5635 /* Positive Lookahead/lookbehind
5636 In this case we can do fixed string optimisation,
5637 but we must be careful about it. Note in the case of
5638 lookbehind the positions will be offset by the minimum
5639 length of the pattern, something we won't know about
5640 until after the recurse.
5642 SSize_t deltanext, fake = 0;
5646 /* We use SAVEFREEPV so that when the full compile
5647 is finished perl will clean up the allocated
5648 minlens when it's all done. This way we don't
5649 have to worry about freeing them when we know
5650 they wont be used, which would be a pain.
5653 Newx( minnextp, 1, SSize_t );
5654 SAVEFREEPV(minnextp);
5657 StructCopy(data, &data_fake, scan_data_t);
5658 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5661 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5662 data_fake.last_found=newSVsv(data->last_found);
5666 data_fake.last_closep = &fake;
5667 data_fake.flags = 0;
5668 data_fake.pos_delta = delta;
5670 data_fake.flags |= SF_IS_INF;
5671 if ( flags & SCF_DO_STCLASS && !scan->flags
5672 && OP(scan) == IFMATCH ) { /* Lookahead */
5673 ssc_init(pRExC_state, &intrnl);
5674 data_fake.start_class = &intrnl;
5675 f |= SCF_DO_STCLASS_AND;
5677 if (flags & SCF_WHILEM_VISITED_POS)
5678 f |= SCF_WHILEM_VISITED_POS;
5679 next = regnext(scan);
5680 nscan = NEXTOPER(NEXTOPER(scan));
5682 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5683 &deltanext, last, &data_fake,
5684 stopparen, recursed_depth, NULL,
5688 FAIL("Variable length lookbehind not implemented");
5690 else if (*minnextp > (I32)U8_MAX) {
5691 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5694 scan->flags = (U8)*minnextp;
5699 if (f & SCF_DO_STCLASS_AND) {
5700 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5701 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
5704 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5706 if (data_fake.flags & SF_HAS_EVAL)
5707 data->flags |= SF_HAS_EVAL;
5708 data->whilem_c = data_fake.whilem_c;
5709 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5710 if (RExC_rx->minlen<*minnextp)
5711 RExC_rx->minlen=*minnextp;
5712 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5713 SvREFCNT_dec_NN(data_fake.last_found);
5715 if ( data_fake.minlen_fixed != minlenp )
5717 data->offset_fixed= data_fake.offset_fixed;
5718 data->minlen_fixed= data_fake.minlen_fixed;
5719 data->lookbehind_fixed+= scan->flags;
5721 if ( data_fake.minlen_float != minlenp )
5723 data->minlen_float= data_fake.minlen_float;
5724 data->offset_float_min=data_fake.offset_float_min;
5725 data->offset_float_max=data_fake.offset_float_max;
5726 data->lookbehind_float+= scan->flags;
5733 else if (OP(scan) == OPEN) {
5734 if (stopparen != (I32)ARG(scan))
5737 else if (OP(scan) == CLOSE) {
5738 if (stopparen == (I32)ARG(scan)) {
5741 if ((I32)ARG(scan) == is_par) {
5742 next = regnext(scan);
5744 if ( next && (OP(next) != WHILEM) && next < last)
5745 is_par = 0; /* Disable optimization */
5748 *(data->last_closep) = ARG(scan);
5750 else if (OP(scan) == EVAL) {
5752 data->flags |= SF_HAS_EVAL;
5754 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5755 if (flags & SCF_DO_SUBSTR) {
5756 scan_commit(pRExC_state, data, minlenp, is_inf);
5757 flags &= ~SCF_DO_SUBSTR;
5759 if (data && OP(scan)==ACCEPT) {
5760 data->flags |= SCF_SEEN_ACCEPT;
5765 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5767 if (flags & SCF_DO_SUBSTR) {
5768 scan_commit(pRExC_state, data, minlenp, is_inf);
5769 data->longest = &(data->longest_float);
5771 is_inf = is_inf_internal = 1;
5772 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5773 ssc_anything(data->start_class);
5774 flags &= ~SCF_DO_STCLASS;
5776 else if (OP(scan) == GPOS) {
5777 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5778 !(delta || is_inf || (data && data->pos_delta)))
5780 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5781 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5782 if (RExC_rx->gofs < (STRLEN)min)
5783 RExC_rx->gofs = min;
5785 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5789 #ifdef TRIE_STUDY_OPT
5790 #ifdef FULL_TRIE_STUDY
5791 else if (PL_regkind[OP(scan)] == TRIE) {
5792 /* NOTE - There is similar code to this block above for handling
5793 BRANCH nodes on the initial study. If you change stuff here
5795 regnode *trie_node= scan;
5796 regnode *tail= regnext(scan);
5797 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5798 SSize_t max1 = 0, min1 = SSize_t_MAX;
5801 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5802 /* Cannot merge strings after this. */
5803 scan_commit(pRExC_state, data, minlenp, is_inf);
5805 if (flags & SCF_DO_STCLASS)
5806 ssc_init_zero(pRExC_state, &accum);
5812 const regnode *nextbranch= NULL;
5815 for ( word=1 ; word <= trie->wordcount ; word++)
5817 SSize_t deltanext=0, minnext=0, f = 0, fake;
5818 regnode_ssc this_class;
5820 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5822 data_fake.whilem_c = data->whilem_c;
5823 data_fake.last_closep = data->last_closep;
5826 data_fake.last_closep = &fake;
5827 data_fake.pos_delta = delta;
5828 if (flags & SCF_DO_STCLASS) {
5829 ssc_init(pRExC_state, &this_class);
5830 data_fake.start_class = &this_class;
5831 f = SCF_DO_STCLASS_AND;
5833 if (flags & SCF_WHILEM_VISITED_POS)
5834 f |= SCF_WHILEM_VISITED_POS;
5836 if (trie->jump[word]) {
5838 nextbranch = trie_node + trie->jump[0];
5839 scan= trie_node + trie->jump[word];
5840 /* We go from the jump point to the branch that follows
5841 it. Note this means we need the vestigal unused
5842 branches even though they arent otherwise used. */
5843 minnext = study_chunk(pRExC_state, &scan, minlenp,
5844 &deltanext, (regnode *)nextbranch, &data_fake,
5845 stopparen, recursed_depth, NULL, f,depth+1);
5847 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5848 nextbranch= regnext((regnode*)nextbranch);
5850 if (min1 > (SSize_t)(minnext + trie->minlen))
5851 min1 = minnext + trie->minlen;
5852 if (deltanext == SSize_t_MAX) {
5853 is_inf = is_inf_internal = 1;
5855 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5856 max1 = minnext + deltanext + trie->maxlen;
5858 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5860 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5861 if ( stopmin > min + min1)
5862 stopmin = min + min1;
5863 flags &= ~SCF_DO_SUBSTR;
5865 data->flags |= SCF_SEEN_ACCEPT;
5868 if (data_fake.flags & SF_HAS_EVAL)
5869 data->flags |= SF_HAS_EVAL;
5870 data->whilem_c = data_fake.whilem_c;
5872 if (flags & SCF_DO_STCLASS)
5873 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5876 if (flags & SCF_DO_SUBSTR) {
5877 data->pos_min += min1;
5878 data->pos_delta += max1 - min1;
5879 if (max1 != min1 || is_inf)
5880 data->longest = &(data->longest_float);
5883 if (delta != SSize_t_MAX)
5884 delta += max1 - min1;
5885 if (flags & SCF_DO_STCLASS_OR) {
5886 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5888 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5889 flags &= ~SCF_DO_STCLASS;
5892 else if (flags & SCF_DO_STCLASS_AND) {
5894 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5895 flags &= ~SCF_DO_STCLASS;
5898 /* Switch to OR mode: cache the old value of
5899 * data->start_class */
5901 StructCopy(data->start_class, and_withp, regnode_ssc);
5902 flags &= ~SCF_DO_STCLASS_AND;
5903 StructCopy(&accum, data->start_class, regnode_ssc);
5904 flags |= SCF_DO_STCLASS_OR;
5911 else if (PL_regkind[OP(scan)] == TRIE) {
5912 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5915 min += trie->minlen;
5916 delta += (trie->maxlen - trie->minlen);
5917 flags &= ~SCF_DO_STCLASS; /* xxx */
5918 if (flags & SCF_DO_SUBSTR) {
5919 /* Cannot expect anything... */
5920 scan_commit(pRExC_state, data, minlenp, is_inf);
5921 data->pos_min += trie->minlen;
5922 data->pos_delta += (trie->maxlen - trie->minlen);
5923 if (trie->maxlen != trie->minlen)
5924 data->longest = &(data->longest_float);
5926 if (trie->jump) /* no more substrings -- for now /grr*/
5927 flags &= ~SCF_DO_SUBSTR;
5929 #endif /* old or new */
5930 #endif /* TRIE_STUDY_OPT */
5932 /* Else: zero-length, ignore. */
5933 scan = regnext(scan);
5938 /* we need to unwind recursion. */
5941 DEBUG_STUDYDATA("frame-end:",data,depth);
5942 DEBUG_PEEP("fend", scan, depth);
5944 /* restore previous context */
5945 last = frame->last_regnode;
5946 scan = frame->next_regnode;
5947 stopparen = frame->stopparen;
5948 recursed_depth = frame->prev_recursed_depth;
5950 RExC_frame_last = frame->prev_frame;
5951 frame = frame->this_prev_frame;
5952 goto fake_study_recurse;
5956 DEBUG_STUDYDATA("pre-fin:",data,depth);
5959 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5961 if (flags & SCF_DO_SUBSTR && is_inf)
5962 data->pos_delta = SSize_t_MAX - data->pos_min;
5963 if (is_par > (I32)U8_MAX)
5965 if (is_par && pars==1 && data) {
5966 data->flags |= SF_IN_PAR;
5967 data->flags &= ~SF_HAS_PAR;
5969 else if (pars && data) {
5970 data->flags |= SF_HAS_PAR;
5971 data->flags &= ~SF_IN_PAR;
5973 if (flags & SCF_DO_STCLASS_OR)
5974 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5975 if (flags & SCF_TRIE_RESTUDY)
5976 data->flags |= SCF_TRIE_RESTUDY;
5978 DEBUG_STUDYDATA("post-fin:",data,depth);
5981 SSize_t final_minlen= min < stopmin ? min : stopmin;
5983 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
5984 if (final_minlen > SSize_t_MAX - delta)
5985 RExC_maxlen = SSize_t_MAX;
5986 else if (RExC_maxlen < final_minlen + delta)
5987 RExC_maxlen = final_minlen + delta;
5989 return final_minlen;
5991 NOT_REACHED; /* NOTREACHED */
5995 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5997 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5999 PERL_ARGS_ASSERT_ADD_DATA;
6001 Renewc(RExC_rxi->data,
6002 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6003 char, struct reg_data);
6005 Renew(RExC_rxi->data->what, count + n, U8);
6007 Newx(RExC_rxi->data->what, n, U8);
6008 RExC_rxi->data->count = count + n;
6009 Copy(s, RExC_rxi->data->what + count, n, U8);
6013 /*XXX: todo make this not included in a non debugging perl, but appears to be
6014 * used anyway there, in 'use re' */
6015 #ifndef PERL_IN_XSUB_RE
6017 Perl_reginitcolors(pTHX)
6019 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6021 char *t = savepv(s);
6025 t = strchr(t, '\t');
6031 PL_colors[i] = t = (char *)"";
6036 PL_colors[i++] = (char *)"";
6043 #ifdef TRIE_STUDY_OPT
6044 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6047 (data.flags & SCF_TRIE_RESTUDY) \
6055 #define CHECK_RESTUDY_GOTO_butfirst
6059 * pregcomp - compile a regular expression into internal code
6061 * Decides which engine's compiler to call based on the hint currently in
6065 #ifndef PERL_IN_XSUB_RE
6067 /* return the currently in-scope regex engine (or the default if none) */
6069 regexp_engine const *
6070 Perl_current_re_engine(pTHX)
6072 if (IN_PERL_COMPILETIME) {
6073 HV * const table = GvHV(PL_hintgv);
6076 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6077 return &PL_core_reg_engine;
6078 ptr = hv_fetchs(table, "regcomp", FALSE);
6079 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6080 return &PL_core_reg_engine;
6081 return INT2PTR(regexp_engine*,SvIV(*ptr));
6085 if (!PL_curcop->cop_hints_hash)
6086 return &PL_core_reg_engine;
6087 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6088 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6089 return &PL_core_reg_engine;
6090 return INT2PTR(regexp_engine*,SvIV(ptr));
6096 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6098 regexp_engine const *eng = current_re_engine();
6099 GET_RE_DEBUG_FLAGS_DECL;
6101 PERL_ARGS_ASSERT_PREGCOMP;
6103 /* Dispatch a request to compile a regexp to correct regexp engine. */
6105 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6108 return CALLREGCOMP_ENG(eng, pattern, flags);
6112 /* public(ish) entry point for the perl core's own regex compiling code.
6113 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6114 * pattern rather than a list of OPs, and uses the internal engine rather
6115 * than the current one */
6118 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6120 SV *pat = pattern; /* defeat constness! */
6121 PERL_ARGS_ASSERT_RE_COMPILE;
6122 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6123 #ifdef PERL_IN_XSUB_RE
6126 &PL_core_reg_engine,
6128 NULL, NULL, rx_flags, 0);
6132 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6133 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6134 * point to the realloced string and length.
6136 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6140 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6141 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6143 U8 *const src = (U8*)*pat_p;
6148 GET_RE_DEBUG_FLAGS_DECL;
6150 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6151 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6153 Newx(dst, *plen_p * 2 + 1, U8);
6156 while (s < *plen_p) {
6157 append_utf8_from_native_byte(src[s], &d);
6158 if (n < num_code_blocks) {
6159 if (!do_end && pRExC_state->code_blocks[n].start == s) {
6160 pRExC_state->code_blocks[n].start = d - dst - 1;
6161 assert(*(d - 1) == '(');
6164 else if (do_end && pRExC_state->code_blocks[n].end == s) {
6165 pRExC_state->code_blocks[n].end = d - dst - 1;
6166 assert(*(d - 1) == ')');
6175 *pat_p = (char*) dst;
6177 RExC_orig_utf8 = RExC_utf8 = 1;
6182 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6183 * while recording any code block indices, and handling overloading,
6184 * nested qr// objects etc. If pat is null, it will allocate a new
6185 * string, or just return the first arg, if there's only one.
6187 * Returns the malloced/updated pat.
6188 * patternp and pat_count is the array of SVs to be concatted;
6189 * oplist is the optional list of ops that generated the SVs;
6190 * recompile_p is a pointer to a boolean that will be set if
6191 * the regex will need to be recompiled.
6192 * delim, if non-null is an SV that will be inserted between each element
6196 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6197 SV *pat, SV ** const patternp, int pat_count,
6198 OP *oplist, bool *recompile_p, SV *delim)
6202 bool use_delim = FALSE;
6203 bool alloced = FALSE;
6205 /* if we know we have at least two args, create an empty string,
6206 * then concatenate args to that. For no args, return an empty string */
6207 if (!pat && pat_count != 1) {
6213 for (svp = patternp; svp < patternp + pat_count; svp++) {
6216 STRLEN orig_patlen = 0;
6218 SV *msv = use_delim ? delim : *svp;
6219 if (!msv) msv = &PL_sv_undef;
6221 /* if we've got a delimiter, we go round the loop twice for each
6222 * svp slot (except the last), using the delimiter the second
6231 if (SvTYPE(msv) == SVt_PVAV) {
6232 /* we've encountered an interpolated array within
6233 * the pattern, e.g. /...@a..../. Expand the list of elements,
6234 * then recursively append elements.
6235 * The code in this block is based on S_pushav() */
6237 AV *const av = (AV*)msv;
6238 const SSize_t maxarg = AvFILL(av) + 1;
6242 assert(oplist->op_type == OP_PADAV
6243 || oplist->op_type == OP_RV2AV);
6244 oplist = OpSIBLING(oplist);
6247 if (SvRMAGICAL(av)) {
6250 Newx(array, maxarg, SV*);
6252 for (i=0; i < maxarg; i++) {
6253 SV ** const svp = av_fetch(av, i, FALSE);
6254 array[i] = svp ? *svp : &PL_sv_undef;
6258 array = AvARRAY(av);
6260 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6261 array, maxarg, NULL, recompile_p,
6263 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6269 /* we make the assumption here that each op in the list of
6270 * op_siblings maps to one SV pushed onto the stack,
6271 * except for code blocks, with have both an OP_NULL and
6273 * This allows us to match up the list of SVs against the
6274 * list of OPs to find the next code block.
6276 * Note that PUSHMARK PADSV PADSV ..
6278 * PADRANGE PADSV PADSV ..
6279 * so the alignment still works. */
6282 if (oplist->op_type == OP_NULL
6283 && (oplist->op_flags & OPf_SPECIAL))
6285 assert(n < pRExC_state->num_code_blocks);
6286 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
6287 pRExC_state->code_blocks[n].block = oplist;
6288 pRExC_state->code_blocks[n].src_regex = NULL;
6291 oplist = OpSIBLING(oplist); /* skip CONST */
6294 oplist = OpSIBLING(oplist);;
6297 /* apply magic and QR overloading to arg */
6300 if (SvROK(msv) && SvAMAGIC(msv)) {
6301 SV *sv = AMG_CALLunary(msv, regexp_amg);
6305 if (SvTYPE(sv) != SVt_REGEXP)
6306 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6311 /* try concatenation overload ... */
6312 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6313 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6316 /* overloading involved: all bets are off over literal
6317 * code. Pretend we haven't seen it */
6318 pRExC_state->num_code_blocks -= n;
6322 /* ... or failing that, try "" overload */
6323 while (SvAMAGIC(msv)
6324 && (sv = AMG_CALLunary(msv, string_amg))
6328 && SvRV(msv) == SvRV(sv))
6333 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6337 /* this is a partially unrolled
6338 * sv_catsv_nomg(pat, msv);
6339 * that allows us to adjust code block indices if
6342 char *dst = SvPV_force_nomg(pat, dlen);
6344 if (SvUTF8(msv) && !SvUTF8(pat)) {
6345 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6346 sv_setpvn(pat, dst, dlen);
6349 sv_catsv_nomg(pat, msv);
6353 /* We have only one SV to process, but we need to verify
6354 * it is properly null terminated or we will fail asserts
6355 * later. In theory we probably shouldn't get such SV's,
6356 * but if we do we should handle it gracefully. */
6357 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) ) {
6358 /* not a string, or a string with a trailing null */
6361 /* a string with no trailing null, we need to copy it
6362 * so it we have a trailing null */
6368 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
6371 /* extract any code blocks within any embedded qr//'s */
6372 if (rx && SvTYPE(rx) == SVt_REGEXP
6373 && RX_ENGINE((REGEXP*)rx)->op_comp)
6376 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6377 if (ri->num_code_blocks) {
6379 /* the presence of an embedded qr// with code means
6380 * we should always recompile: the text of the
6381 * qr// may not have changed, but it may be a
6382 * different closure than last time */
6384 Renew(pRExC_state->code_blocks,
6385 pRExC_state->num_code_blocks + ri->num_code_blocks,
6386 struct reg_code_block);
6387 pRExC_state->num_code_blocks += ri->num_code_blocks;
6389 for (i=0; i < ri->num_code_blocks; i++) {
6390 struct reg_code_block *src, *dst;
6391 STRLEN offset = orig_patlen
6392 + ReANY((REGEXP *)rx)->pre_prefix;
6393 assert(n < pRExC_state->num_code_blocks);
6394 src = &ri->code_blocks[i];
6395 dst = &pRExC_state->code_blocks[n];
6396 dst->start = src->start + offset;
6397 dst->end = src->end + offset;
6398 dst->block = src->block;
6399 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6408 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6417 /* see if there are any run-time code blocks in the pattern.
6418 * False positives are allowed */
6421 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6422 char *pat, STRLEN plen)
6427 PERL_UNUSED_CONTEXT;
6429 for (s = 0; s < plen; s++) {
6430 if (n < pRExC_state->num_code_blocks
6431 && s == pRExC_state->code_blocks[n].start)
6433 s = pRExC_state->code_blocks[n].end;
6437 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6439 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6441 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6448 /* Handle run-time code blocks. We will already have compiled any direct
6449 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6450 * copy of it, but with any literal code blocks blanked out and
6451 * appropriate chars escaped; then feed it into
6453 * eval "qr'modified_pattern'"
6457 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6461 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6463 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6464 * and merge them with any code blocks of the original regexp.
6466 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6467 * instead, just save the qr and return FALSE; this tells our caller that
6468 * the original pattern needs upgrading to utf8.
6472 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6473 char *pat, STRLEN plen)
6477 GET_RE_DEBUG_FLAGS_DECL;
6479 if (pRExC_state->runtime_code_qr) {
6480 /* this is the second time we've been called; this should
6481 * only happen if the main pattern got upgraded to utf8
6482 * during compilation; re-use the qr we compiled first time
6483 * round (which should be utf8 too)
6485 qr = pRExC_state->runtime_code_qr;
6486 pRExC_state->runtime_code_qr = NULL;
6487 assert(RExC_utf8 && SvUTF8(qr));
6493 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
6497 /* determine how many extra chars we need for ' and \ escaping */
6498 for (s = 0; s < plen; s++) {
6499 if (pat[s] == '\'' || pat[s] == '\\')
6503 Newx(newpat, newlen, char);
6505 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6507 for (s = 0; s < plen; s++) {
6508 if (n < pRExC_state->num_code_blocks
6509 && s == pRExC_state->code_blocks[n].start)
6511 /* blank out literal code block */
6512 assert(pat[s] == '(');
6513 while (s <= pRExC_state->code_blocks[n].end) {
6521 if (pat[s] == '\'' || pat[s] == '\\')
6526 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
6528 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
6534 Perl_re_printf( aTHX_
6535 "%sre-parsing pattern for runtime code:%s %s\n",
6536 PL_colors[4],PL_colors[5],newpat);
6539 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6545 PUSHSTACKi(PERLSI_REQUIRE);
6546 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6547 * parsing qr''; normally only q'' does this. It also alters
6549 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6550 SvREFCNT_dec_NN(sv);
6555 SV * const errsv = ERRSV;
6556 if (SvTRUE_NN(errsv))
6558 Safefree(pRExC_state->code_blocks);
6559 /* use croak_sv ? */
6560 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
6563 assert(SvROK(qr_ref));
6565 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6566 /* the leaving below frees the tmp qr_ref.
6567 * Give qr a life of its own */
6575 if (!RExC_utf8 && SvUTF8(qr)) {
6576 /* first time through; the pattern got upgraded; save the
6577 * qr for the next time through */
6578 assert(!pRExC_state->runtime_code_qr);
6579 pRExC_state->runtime_code_qr = qr;
6584 /* extract any code blocks within the returned qr// */
6587 /* merge the main (r1) and run-time (r2) code blocks into one */
6589 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6590 struct reg_code_block *new_block, *dst;
6591 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6594 if (!r2->num_code_blocks) /* we guessed wrong */
6596 SvREFCNT_dec_NN(qr);
6601 r1->num_code_blocks + r2->num_code_blocks,
6602 struct reg_code_block);
6605 while ( i1 < r1->num_code_blocks
6606 || i2 < r2->num_code_blocks)
6608 struct reg_code_block *src;
6611 if (i1 == r1->num_code_blocks) {
6612 src = &r2->code_blocks[i2++];
6615 else if (i2 == r2->num_code_blocks)
6616 src = &r1->code_blocks[i1++];
6617 else if ( r1->code_blocks[i1].start
6618 < r2->code_blocks[i2].start)
6620 src = &r1->code_blocks[i1++];
6621 assert(src->end < r2->code_blocks[i2].start);
6624 assert( r1->code_blocks[i1].start
6625 > r2->code_blocks[i2].start);
6626 src = &r2->code_blocks[i2++];
6628 assert(src->end < r1->code_blocks[i1].start);
6631 assert(pat[src->start] == '(');
6632 assert(pat[src->end] == ')');
6633 dst->start = src->start;
6634 dst->end = src->end;
6635 dst->block = src->block;
6636 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6640 r1->num_code_blocks += r2->num_code_blocks;
6641 Safefree(r1->code_blocks);
6642 r1->code_blocks = new_block;
6645 SvREFCNT_dec_NN(qr);
6651 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6652 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6653 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6654 STRLEN longest_length, bool eol, bool meol)
6656 /* This is the common code for setting up the floating and fixed length
6657 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6658 * as to whether succeeded or not */
6663 if (! (longest_length
6664 || (eol /* Can't have SEOL and MULTI */
6665 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6667 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6668 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6673 /* copy the information about the longest from the reg_scan_data
6674 over to the program. */
6675 if (SvUTF8(sv_longest)) {
6676 *rx_utf8 = sv_longest;
6679 *rx_substr = sv_longest;
6682 /* end_shift is how many chars that must be matched that
6683 follow this item. We calculate it ahead of time as once the
6684 lookbehind offset is added in we lose the ability to correctly
6686 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6687 *rx_end_shift = ml - offset
6689 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
6691 + (SvTAIL(sv_longest) != 0)
6695 t = (eol/* Can't have SEOL and MULTI */
6696 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6697 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6703 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6704 * regular expression into internal code.
6705 * The pattern may be passed either as:
6706 * a list of SVs (patternp plus pat_count)
6707 * a list of OPs (expr)
6708 * If both are passed, the SV list is used, but the OP list indicates
6709 * which SVs are actually pre-compiled code blocks
6711 * The SVs in the list have magic and qr overloading applied to them (and
6712 * the list may be modified in-place with replacement SVs in the latter
6715 * If the pattern hasn't changed from old_re, then old_re will be
6718 * eng is the current engine. If that engine has an op_comp method, then
6719 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6720 * do the initial concatenation of arguments and pass on to the external
6723 * If is_bare_re is not null, set it to a boolean indicating whether the
6724 * arg list reduced (after overloading) to a single bare regex which has
6725 * been returned (i.e. /$qr/).
6727 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6729 * pm_flags contains the PMf_* flags, typically based on those from the
6730 * pm_flags field of the related PMOP. Currently we're only interested in
6731 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6733 * We can't allocate space until we know how big the compiled form will be,
6734 * but we can't compile it (and thus know how big it is) until we've got a
6735 * place to put the code. So we cheat: we compile it twice, once with code
6736 * generation turned off and size counting turned on, and once "for real".
6737 * This also means that we don't allocate space until we are sure that the
6738 * thing really will compile successfully, and we never have to move the
6739 * code and thus invalidate pointers into it. (Note that it has to be in
6740 * one piece because free() must be able to free it all.) [NB: not true in perl]
6742 * Beware that the optimization-preparation code in here knows about some
6743 * of the structure of the compiled regexp. [I'll say.]
6747 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6748 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6749 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6753 regexp_internal *ri;
6761 SV *code_blocksv = NULL;
6762 SV** new_patternp = patternp;
6764 /* these are all flags - maybe they should be turned
6765 * into a single int with different bit masks */
6766 I32 sawlookahead = 0;
6771 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6773 bool runtime_code = 0;
6775 RExC_state_t RExC_state;
6776 RExC_state_t * const pRExC_state = &RExC_state;
6777 #ifdef TRIE_STUDY_OPT
6779 RExC_state_t copyRExC_state;
6781 GET_RE_DEBUG_FLAGS_DECL;
6783 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6785 DEBUG_r(if (!PL_colorset) reginitcolors());
6787 /* Initialize these here instead of as-needed, as is quick and avoids
6788 * having to test them each time otherwise */
6789 if (! PL_AboveLatin1) {
6791 char * dump_len_string;
6794 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6795 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6796 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6797 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6798 PL_HasMultiCharFold =
6799 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6801 /* This is calculated here, because the Perl program that generates the
6802 * static global ones doesn't currently have access to
6803 * NUM_ANYOF_CODE_POINTS */
6804 PL_InBitmap = _new_invlist(2);
6805 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
6806 NUM_ANYOF_CODE_POINTS - 1);
6808 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
6809 if ( ! dump_len_string
6810 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
6812 PL_dump_re_max_len = 0;
6817 pRExC_state->warn_text = NULL;
6818 pRExC_state->code_blocks = NULL;
6819 pRExC_state->num_code_blocks = 0;
6822 *is_bare_re = FALSE;
6824 if (expr && (expr->op_type == OP_LIST ||
6825 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6826 /* allocate code_blocks if needed */
6830 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
6831 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6832 ncode++; /* count of DO blocks */
6834 pRExC_state->num_code_blocks = ncode;
6835 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
6840 /* compile-time pattern with just OP_CONSTs and DO blocks */
6845 /* find how many CONSTs there are */
6848 if (expr->op_type == OP_CONST)
6851 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6852 if (o->op_type == OP_CONST)
6856 /* fake up an SV array */
6858 assert(!new_patternp);
6859 Newx(new_patternp, n, SV*);
6860 SAVEFREEPV(new_patternp);
6864 if (expr->op_type == OP_CONST)
6865 new_patternp[n] = cSVOPx_sv(expr);
6867 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6868 if (o->op_type == OP_CONST)
6869 new_patternp[n++] = cSVOPo_sv;
6874 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6875 "Assembling pattern from %d elements%s\n", pat_count,
6876 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6878 /* set expr to the first arg op */
6880 if (pRExC_state->num_code_blocks
6881 && expr->op_type != OP_CONST)
6883 expr = cLISTOPx(expr)->op_first;
6884 assert( expr->op_type == OP_PUSHMARK
6885 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6886 || expr->op_type == OP_PADRANGE);
6887 expr = OpSIBLING(expr);
6890 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6891 expr, &recompile, NULL);
6893 /* handle bare (possibly after overloading) regex: foo =~ $re */
6898 if (SvTYPE(re) == SVt_REGEXP) {
6902 Safefree(pRExC_state->code_blocks);
6903 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6904 "Precompiled pattern%s\n",
6905 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6911 exp = SvPV_nomg(pat, plen);
6913 if (!eng->op_comp) {
6914 if ((SvUTF8(pat) && IN_BYTES)
6915 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6917 /* make a temporary copy; either to convert to bytes,
6918 * or to avoid repeating get-magic / overloaded stringify */
6919 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6920 (IN_BYTES ? 0 : SvUTF8(pat)));
6922 Safefree(pRExC_state->code_blocks);
6923 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6926 /* ignore the utf8ness if the pattern is 0 length */
6927 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6929 RExC_uni_semantics = 0;
6930 RExC_seen_unfolded_sharp_s = 0;
6931 RExC_contains_locale = 0;
6932 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
6933 RExC_study_started = 0;
6934 pRExC_state->runtime_code_qr = NULL;
6935 RExC_frame_head= NULL;
6936 RExC_frame_last= NULL;
6937 RExC_frame_count= 0;
6940 RExC_mysv1= sv_newmortal();
6941 RExC_mysv2= sv_newmortal();
6944 SV *dsv= sv_newmortal();
6945 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6946 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
6947 PL_colors[4],PL_colors[5],s);
6951 /* we jump here if we have to recompile, e.g., from upgrading the pattern
6954 if ((pm_flags & PMf_USE_RE_EVAL)
6955 /* this second condition covers the non-regex literal case,
6956 * i.e. $foo =~ '(?{})'. */
6957 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6959 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6961 /* return old regex if pattern hasn't changed */
6962 /* XXX: note in the below we have to check the flags as well as the
6965 * Things get a touch tricky as we have to compare the utf8 flag
6966 * independently from the compile flags. */
6970 && !!RX_UTF8(old_re) == !!RExC_utf8
6971 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
6972 && RX_PRECOMP(old_re)
6973 && RX_PRELEN(old_re) == plen
6974 && memEQ(RX_PRECOMP(old_re), exp, plen)
6975 && !runtime_code /* with runtime code, always recompile */ )
6977 Safefree(pRExC_state->code_blocks);
6981 rx_flags = orig_rx_flags;
6983 if ( initial_charset == REGEX_DEPENDS_CHARSET
6984 && (RExC_utf8 ||RExC_uni_semantics))
6987 /* Set to use unicode semantics if the pattern is in utf8 and has the
6988 * 'depends' charset specified, as it means unicode when utf8 */
6989 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6993 RExC_precomp_adj = 0;
6994 RExC_flags = rx_flags;
6995 RExC_pm_flags = pm_flags;
6998 assert(TAINTING_get || !TAINT_get);
7000 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7002 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7003 /* whoops, we have a non-utf8 pattern, whilst run-time code
7004 * got compiled as utf8. Try again with a utf8 pattern */
7005 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7006 pRExC_state->num_code_blocks);
7007 goto redo_first_pass;
7010 assert(!pRExC_state->runtime_code_qr);
7016 RExC_in_lookbehind = 0;
7017 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7020 RExC_recode_x_to_native = 0;
7022 RExC_in_multi_char_class = 0;
7024 /* First pass: determine size, legality. */
7026 RExC_start = RExC_adjusted_start = exp;
7027 RExC_end = exp + plen;
7028 RExC_precomp_end = RExC_end;
7033 RExC_emit = (regnode *) &RExC_emit_dummy;
7034 RExC_whilem_seen = 0;
7035 RExC_open_parens = NULL;
7036 RExC_close_parens = NULL;
7038 RExC_paren_names = NULL;
7040 RExC_paren_name_list = NULL;
7042 RExC_recurse = NULL;
7043 RExC_study_chunk_recursed = NULL;
7044 RExC_study_chunk_recursed_bytes= 0;
7045 RExC_recurse_count = 0;
7046 pRExC_state->code_index = 0;
7048 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7049 * code makes sure the final byte is an uncounted NUL. But should this
7050 * ever not be the case, lots of things could read beyond the end of the
7051 * buffer: loops like
7052 * while(isFOO(*RExC_parse)) RExC_parse++;
7053 * strchr(RExC_parse, "foo");
7054 * etc. So it is worth noting. */
7055 assert(*RExC_end == '\0');
7058 Perl_re_printf( aTHX_ "Starting first pass (sizing)\n");
7060 RExC_lastparse=NULL;
7062 /* reg may croak on us, not giving us a chance to free
7063 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
7064 need it to survive as long as the regexp (qr/(?{})/).
7065 We must check that code_blocksv is not already set, because we may
7066 have jumped back to restart the sizing pass. */
7067 if (pRExC_state->code_blocks && !code_blocksv) {
7068 code_blocksv = newSV_type(SVt_PV);
7069 SAVEFREESV(code_blocksv);
7070 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
7071 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
7073 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7074 /* It's possible to write a regexp in ascii that represents Unicode
7075 codepoints outside of the byte range, such as via \x{100}. If we
7076 detect such a sequence we have to convert the entire pattern to utf8
7077 and then recompile, as our sizing calculation will have been based
7078 on 1 byte == 1 character, but we will need to use utf8 to encode
7079 at least some part of the pattern, and therefore must convert the whole
7082 if (flags & RESTART_PASS1) {
7083 if (flags & NEED_UTF8) {
7084 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7085 pRExC_state->num_code_blocks);
7088 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7089 "Need to redo pass 1\n"));
7092 goto redo_first_pass;
7094 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#" UVxf, (UV) flags);
7097 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
7100 Perl_re_printf( aTHX_
7101 "Required size %" IVdf " nodes\n"
7102 "Starting second pass (creation)\n",
7105 RExC_lastparse=NULL;
7108 /* The first pass could have found things that force Unicode semantics */
7109 if ((RExC_utf8 || RExC_uni_semantics)
7110 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
7112 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7115 /* Small enough for pointer-storage convention?
7116 If extralen==0, this means that we will not need long jumps. */
7117 if (RExC_size >= 0x10000L && RExC_extralen)
7118 RExC_size += RExC_extralen;
7121 if (RExC_whilem_seen > 15)
7122 RExC_whilem_seen = 15;
7124 /* Allocate space and zero-initialize. Note, the two step process
7125 of zeroing when in debug mode, thus anything assigned has to
7126 happen after that */
7127 rx = (REGEXP*) newSV_type(SVt_REGEXP);
7129 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7130 char, regexp_internal);
7131 if ( r == NULL || ri == NULL )
7132 FAIL("Regexp out of space");
7134 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
7135 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7138 /* bulk initialize base fields with 0. */
7139 Zero(ri, sizeof(regexp_internal), char);
7142 /* non-zero initialization begins here */
7145 r->extflags = rx_flags;
7146 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7148 if (pm_flags & PMf_IS_QR) {
7149 ri->code_blocks = pRExC_state->code_blocks;
7150 ri->num_code_blocks = pRExC_state->num_code_blocks;
7155 for (n = 0; n < pRExC_state->num_code_blocks; n++)
7156 if (pRExC_state->code_blocks[n].src_regex)
7157 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
7158 if(pRExC_state->code_blocks)
7159 SAVEFREEPV(pRExC_state->code_blocks); /* often null */
7163 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7164 bool has_charset = (get_regex_charset(r->extflags)
7165 != REGEX_DEPENDS_CHARSET);
7167 /* The caret is output if there are any defaults: if not all the STD
7168 * flags are set, or if no character set specifier is needed */
7170 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7172 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7173 == REG_RUN_ON_COMMENT_SEEN);
7174 U8 reganch = (U8)((r->extflags & RXf_PMf_STD_PMMOD)
7175 >> RXf_PMf_STD_PMMOD_SHIFT);
7176 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7179 /* We output all the necessary flags; we never output a minus, as all
7180 * those are defaults, so are
7181 * covered by the caret */
7182 const STRLEN wraplen = plen + has_p + has_runon
7183 + has_default /* If needs a caret */
7184 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7186 /* If needs a character set specifier */
7187 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7188 + (sizeof("(?:)") - 1);
7190 /* make sure PL_bitcount bounds not exceeded */
7191 assert(sizeof(STD_PAT_MODS) <= 8);
7193 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
7194 r->xpv_len_u.xpvlenu_pv = p;
7196 SvFLAGS(rx) |= SVf_UTF8;
7199 /* If a default, cover it using the caret */
7201 *p++= DEFAULT_PAT_MOD;
7205 const char* const name = get_regex_charset_name(r->extflags, &len);
7206 Copy(name, p, len, char);
7210 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7213 while((ch = *fptr++)) {
7221 Copy(RExC_precomp, p, plen, char);
7222 assert ((RX_WRAPPED(rx) - p) < 16);
7223 r->pre_prefix = p - RX_WRAPPED(rx);
7229 SvCUR_set(rx, p - RX_WRAPPED(rx));
7233 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
7235 /* Useful during FAIL. */
7236 #ifdef RE_TRACK_PATTERN_OFFSETS
7237 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
7238 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7239 "%s %" UVuf " bytes for offset annotations.\n",
7240 ri->u.offsets ? "Got" : "Couldn't get",
7241 (UV)((2*RExC_size+1) * sizeof(U32))));
7243 SetProgLen(ri,RExC_size);
7248 /* Second pass: emit code. */
7249 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7250 RExC_pm_flags = pm_flags;
7252 RExC_end = exp + plen;
7254 RExC_emit_start = ri->program;
7255 RExC_emit = ri->program;
7256 RExC_emit_bound = ri->program + RExC_size + 1;
7257 pRExC_state->code_index = 0;
7259 *((char*) RExC_emit++) = (char) REG_MAGIC;
7260 /* setup various meta data about recursion, this all requires
7261 * RExC_npar to be correctly set, and a bit later on we clear it */
7262 if (RExC_seen & REG_RECURSE_SEEN) {
7263 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
7264 "%*s%*s Setting up open/close parens\n",
7265 22, "| |", (int)(0 * 2 + 1), ""));
7267 /* setup RExC_open_parens, which holds the address of each
7268 * OPEN tag, and to make things simpler for the 0 index
7269 * the start of the program - this is used later for offsets */
7270 Newxz(RExC_open_parens, RExC_npar,regnode *);
7271 SAVEFREEPV(RExC_open_parens);
7272 RExC_open_parens[0] = RExC_emit;
7274 /* setup RExC_close_parens, which holds the address of each
7275 * CLOSE tag, and to make things simpler for the 0 index
7276 * the end of the program - this is used later for offsets */
7277 Newxz(RExC_close_parens, RExC_npar,regnode *);
7278 SAVEFREEPV(RExC_close_parens);
7279 /* we dont know where end op starts yet, so we dont
7280 * need to set RExC_close_parens[0] like we do RExC_open_parens[0] above */
7282 /* Note, RExC_npar is 1 + the number of parens in a pattern.
7283 * So its 1 if there are no parens. */
7284 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
7285 ((RExC_npar & 0x07) != 0);
7286 Newx(RExC_study_chunk_recursed,
7287 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7288 SAVEFREEPV(RExC_study_chunk_recursed);
7291 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7293 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#" UVxf, (UV) flags);
7296 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7299 /* XXXX To minimize changes to RE engine we always allocate
7300 3-units-long substrs field. */
7301 Newx(r->substrs, 1, struct reg_substr_data);
7302 if (RExC_recurse_count) {
7303 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
7304 SAVEFREEPV(RExC_recurse);
7308 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7310 RExC_study_chunk_recursed_count= 0;
7312 Zero(r->substrs, 1, struct reg_substr_data);
7313 if (RExC_study_chunk_recursed) {
7314 Zero(RExC_study_chunk_recursed,
7315 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7319 #ifdef TRIE_STUDY_OPT
7321 StructCopy(&zero_scan_data, &data, scan_data_t);
7322 copyRExC_state = RExC_state;
7325 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
7327 RExC_state = copyRExC_state;
7328 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
7329 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
7331 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
7332 StructCopy(&zero_scan_data, &data, scan_data_t);
7335 StructCopy(&zero_scan_data, &data, scan_data_t);
7338 /* Dig out information for optimizations. */
7339 r->extflags = RExC_flags; /* was pm_op */
7340 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
7343 SvUTF8_on(rx); /* Unicode in it? */
7344 ri->regstclass = NULL;
7345 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
7346 r->intflags |= PREGf_NAUGHTY;
7347 scan = ri->program + 1; /* First BRANCH. */
7349 /* testing for BRANCH here tells us whether there is "must appear"
7350 data in the pattern. If there is then we can use it for optimisations */
7351 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7354 STRLEN longest_float_length, longest_fixed_length;
7355 regnode_ssc ch_class; /* pointed to by data */
7357 SSize_t last_close = 0; /* pointed to by data */
7358 regnode *first= scan;
7359 regnode *first_next= regnext(first);
7361 * Skip introductions and multiplicators >= 1
7362 * so that we can extract the 'meat' of the pattern that must
7363 * match in the large if() sequence following.
7364 * NOTE that EXACT is NOT covered here, as it is normally
7365 * picked up by the optimiser separately.
7367 * This is unfortunate as the optimiser isnt handling lookahead
7368 * properly currently.
7371 while ((OP(first) == OPEN && (sawopen = 1)) ||
7372 /* An OR of *one* alternative - should not happen now. */
7373 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7374 /* for now we can't handle lookbehind IFMATCH*/
7375 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7376 (OP(first) == PLUS) ||
7377 (OP(first) == MINMOD) ||
7378 /* An {n,m} with n>0 */
7379 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7380 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7383 * the only op that could be a regnode is PLUS, all the rest
7384 * will be regnode_1 or regnode_2.
7386 * (yves doesn't think this is true)
7388 if (OP(first) == PLUS)
7391 if (OP(first) == MINMOD)
7393 first += regarglen[OP(first)];
7395 first = NEXTOPER(first);
7396 first_next= regnext(first);
7399 /* Starting-point info. */
7401 DEBUG_PEEP("first:",first,0);
7402 /* Ignore EXACT as we deal with it later. */
7403 if (PL_regkind[OP(first)] == EXACT) {
7404 if (OP(first) == EXACT || OP(first) == EXACTL)
7405 NOOP; /* Empty, get anchored substr later. */
7407 ri->regstclass = first;
7410 else if (PL_regkind[OP(first)] == TRIE &&
7411 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
7413 /* this can happen only on restudy */
7414 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
7417 else if (REGNODE_SIMPLE(OP(first)))
7418 ri->regstclass = first;
7419 else if (PL_regkind[OP(first)] == BOUND ||
7420 PL_regkind[OP(first)] == NBOUND)
7421 ri->regstclass = first;
7422 else if (PL_regkind[OP(first)] == BOL) {
7423 r->intflags |= (OP(first) == MBOL
7426 first = NEXTOPER(first);
7429 else if (OP(first) == GPOS) {
7430 r->intflags |= PREGf_ANCH_GPOS;
7431 first = NEXTOPER(first);
7434 else if ((!sawopen || !RExC_sawback) &&
7436 (OP(first) == STAR &&
7437 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7438 !(r->intflags & PREGf_ANCH) && !pRExC_state->num_code_blocks)
7440 /* turn .* into ^.* with an implied $*=1 */
7442 (OP(NEXTOPER(first)) == REG_ANY)
7445 r->intflags |= (type | PREGf_IMPLICIT);
7446 first = NEXTOPER(first);
7449 if (sawplus && !sawminmod && !sawlookahead
7450 && (!sawopen || !RExC_sawback)
7451 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
7452 /* x+ must match at the 1st pos of run of x's */
7453 r->intflags |= PREGf_SKIP;
7455 /* Scan is after the zeroth branch, first is atomic matcher. */
7456 #ifdef TRIE_STUDY_OPT
7459 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7460 (IV)(first - scan + 1))
7464 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7465 (IV)(first - scan + 1))
7471 * If there's something expensive in the r.e., find the
7472 * longest literal string that must appear and make it the
7473 * regmust. Resolve ties in favor of later strings, since
7474 * the regstart check works with the beginning of the r.e.
7475 * and avoiding duplication strengthens checking. Not a
7476 * strong reason, but sufficient in the absence of others.
7477 * [Now we resolve ties in favor of the earlier string if
7478 * it happens that c_offset_min has been invalidated, since the
7479 * earlier string may buy us something the later one won't.]
7482 data.longest_fixed = newSVpvs("");
7483 data.longest_float = newSVpvs("");
7484 data.last_found = newSVpvs("");
7485 data.longest = &(data.longest_fixed);
7486 ENTER_with_name("study_chunk");
7487 SAVEFREESV(data.longest_fixed);
7488 SAVEFREESV(data.longest_float);
7489 SAVEFREESV(data.last_found);
7491 if (!ri->regstclass) {
7492 ssc_init(pRExC_state, &ch_class);
7493 data.start_class = &ch_class;
7494 stclass_flag = SCF_DO_STCLASS_AND;
7495 } else /* XXXX Check for BOUND? */
7497 data.last_closep = &last_close;
7500 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
7501 scan + RExC_size, /* Up to end */
7503 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
7504 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
7508 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
7511 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
7512 && data.last_start_min == 0 && data.last_end > 0
7513 && !RExC_seen_zerolen
7514 && !(RExC_seen & REG_VERBARG_SEEN)
7515 && !(RExC_seen & REG_GPOS_SEEN)
7517 r->extflags |= RXf_CHECK_ALL;
7519 scan_commit(pRExC_state, &data,&minlen,0);
7521 longest_float_length = CHR_SVLEN(data.longest_float);
7523 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
7524 && data.offset_fixed == data.offset_float_min
7525 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
7526 && S_setup_longest (aTHX_ pRExC_state,
7530 &(r->float_end_shift),
7531 data.lookbehind_float,
7532 data.offset_float_min,
7534 longest_float_length,
7535 cBOOL(data.flags & SF_FL_BEFORE_EOL),
7536 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
7538 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
7539 r->float_max_offset = data.offset_float_max;
7540 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
7541 r->float_max_offset -= data.lookbehind_float;
7542 SvREFCNT_inc_simple_void_NN(data.longest_float);
7545 r->float_substr = r->float_utf8 = NULL;
7546 longest_float_length = 0;
7549 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
7551 if (S_setup_longest (aTHX_ pRExC_state,
7553 &(r->anchored_utf8),
7554 &(r->anchored_substr),
7555 &(r->anchored_end_shift),
7556 data.lookbehind_fixed,
7559 longest_fixed_length,
7560 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
7561 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
7563 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
7564 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
7567 r->anchored_substr = r->anchored_utf8 = NULL;
7568 longest_fixed_length = 0;
7570 LEAVE_with_name("study_chunk");
7573 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
7574 ri->regstclass = NULL;
7576 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
7578 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7579 && is_ssc_worth_it(pRExC_state, data.start_class))
7581 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7583 ssc_finalize(pRExC_state, data.start_class);
7585 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7586 StructCopy(data.start_class,
7587 (regnode_ssc*)RExC_rxi->data->data[n],
7589 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7590 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7591 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7592 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7593 Perl_re_printf( aTHX_
7594 "synthetic stclass \"%s\".\n",
7595 SvPVX_const(sv));});
7596 data.start_class = NULL;
7599 /* A temporary algorithm prefers floated substr to fixed one to dig
7601 if (longest_fixed_length > longest_float_length) {
7602 r->substrs->check_ix = 0;
7603 r->check_end_shift = r->anchored_end_shift;
7604 r->check_substr = r->anchored_substr;
7605 r->check_utf8 = r->anchored_utf8;
7606 r->check_offset_min = r->check_offset_max = r->anchored_offset;
7607 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
7608 r->intflags |= PREGf_NOSCAN;
7611 r->substrs->check_ix = 1;
7612 r->check_end_shift = r->float_end_shift;
7613 r->check_substr = r->float_substr;
7614 r->check_utf8 = r->float_utf8;
7615 r->check_offset_min = r->float_min_offset;
7616 r->check_offset_max = r->float_max_offset;
7618 if ((r->check_substr || r->check_utf8) ) {
7619 r->extflags |= RXf_USE_INTUIT;
7620 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7621 r->extflags |= RXf_INTUIT_TAIL;
7623 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
7625 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7626 if ( (STRLEN)minlen < longest_float_length )
7627 minlen= longest_float_length;
7628 if ( (STRLEN)minlen < longest_fixed_length )
7629 minlen= longest_fixed_length;
7633 /* Several toplevels. Best we can is to set minlen. */
7635 regnode_ssc ch_class;
7636 SSize_t last_close = 0;
7638 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
7640 scan = ri->program + 1;
7641 ssc_init(pRExC_state, &ch_class);
7642 data.start_class = &ch_class;
7643 data.last_closep = &last_close;
7646 minlen = study_chunk(pRExC_state,
7647 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7648 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7649 ? SCF_TRIE_DOING_RESTUDY
7653 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7655 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7656 = r->float_substr = r->float_utf8 = NULL;
7658 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7659 && is_ssc_worth_it(pRExC_state, data.start_class))
7661 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7663 ssc_finalize(pRExC_state, data.start_class);
7665 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7666 StructCopy(data.start_class,
7667 (regnode_ssc*)RExC_rxi->data->data[n],
7669 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7670 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7671 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7672 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7673 Perl_re_printf( aTHX_
7674 "synthetic stclass \"%s\".\n",
7675 SvPVX_const(sv));});
7676 data.start_class = NULL;
7680 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7681 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7682 r->maxlen = REG_INFTY;
7685 r->maxlen = RExC_maxlen;
7688 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7689 the "real" pattern. */
7691 Perl_re_printf( aTHX_ "minlen: %" IVdf " r->minlen:%" IVdf " maxlen:%" IVdf "\n",
7692 (IV)minlen, (IV)r->minlen, (IV)RExC_maxlen);
7694 r->minlenret = minlen;
7695 if (r->minlen < minlen)
7698 if (RExC_seen & REG_RECURSE_SEEN ) {
7699 r->intflags |= PREGf_RECURSE_SEEN;
7700 Newxz(r->recurse_locinput, r->nparens + 1, char *);
7702 if (RExC_seen & REG_GPOS_SEEN)
7703 r->intflags |= PREGf_GPOS_SEEN;
7704 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7705 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7707 if (pRExC_state->num_code_blocks)
7708 r->extflags |= RXf_EVAL_SEEN;
7709 if (RExC_seen & REG_VERBARG_SEEN)
7711 r->intflags |= PREGf_VERBARG_SEEN;
7712 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7714 if (RExC_seen & REG_CUTGROUP_SEEN)
7715 r->intflags |= PREGf_CUTGROUP_SEEN;
7716 if (pm_flags & PMf_USE_RE_EVAL)
7717 r->intflags |= PREGf_USE_RE_EVAL;
7718 if (RExC_paren_names)
7719 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7721 RXp_PAREN_NAMES(r) = NULL;
7723 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7724 * so it can be used in pp.c */
7725 if (r->intflags & PREGf_ANCH)
7726 r->extflags |= RXf_IS_ANCHORED;
7730 /* this is used to identify "special" patterns that might result
7731 * in Perl NOT calling the regex engine and instead doing the match "itself",
7732 * particularly special cases in split//. By having the regex compiler
7733 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7734 * we avoid weird issues with equivalent patterns resulting in different behavior,
7735 * AND we allow non Perl engines to get the same optimizations by the setting the
7736 * flags appropriately - Yves */
7737 regnode *first = ri->program + 1;
7739 regnode *next = regnext(first);
7742 if (PL_regkind[fop] == NOTHING && nop == END)
7743 r->extflags |= RXf_NULL;
7744 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
7745 /* when fop is SBOL first->flags will be true only when it was
7746 * produced by parsing /\A/, and not when parsing /^/. This is
7747 * very important for the split code as there we want to
7748 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
7749 * See rt #122761 for more details. -- Yves */
7750 r->extflags |= RXf_START_ONLY;
7751 else if (fop == PLUS
7752 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7754 r->extflags |= RXf_WHITE;
7755 else if ( r->extflags & RXf_SPLIT
7756 && (fop == EXACT || fop == EXACTL)
7757 && STR_LEN(first) == 1
7758 && *(STRING(first)) == ' '
7760 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7764 if (RExC_contains_locale) {
7765 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7769 if (RExC_paren_names) {
7770 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7771 ri->data->data[ri->name_list_idx]
7772 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7775 ri->name_list_idx = 0;
7777 while ( RExC_recurse_count > 0 ) {
7778 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
7779 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - scan );
7782 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7783 /* assume we don't need to swap parens around before we match */
7785 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
7786 (unsigned long)RExC_study_chunk_recursed_count);
7790 Perl_re_printf( aTHX_ "Final program:\n");
7793 #ifdef RE_TRACK_PATTERN_OFFSETS
7794 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7795 const STRLEN len = ri->u.offsets[0];
7797 GET_RE_DEBUG_FLAGS_DECL;
7798 Perl_re_printf( aTHX_
7799 "Offsets: [%" UVuf "]\n\t", (UV)ri->u.offsets[0]);
7800 for (i = 1; i <= len; i++) {
7801 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7802 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7803 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7805 Perl_re_printf( aTHX_ "\n");
7810 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7811 * by setting the regexp SV to readonly-only instead. If the
7812 * pattern's been recompiled, the USEDness should remain. */
7813 if (old_re && SvREADONLY(old_re))
7821 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7824 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7826 PERL_UNUSED_ARG(value);
7828 if (flags & RXapif_FETCH) {
7829 return reg_named_buff_fetch(rx, key, flags);
7830 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7831 Perl_croak_no_modify();
7833 } else if (flags & RXapif_EXISTS) {
7834 return reg_named_buff_exists(rx, key, flags)
7837 } else if (flags & RXapif_REGNAMES) {
7838 return reg_named_buff_all(rx, flags);
7839 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7840 return reg_named_buff_scalar(rx, flags);
7842 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7848 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7851 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7852 PERL_UNUSED_ARG(lastkey);
7854 if (flags & RXapif_FIRSTKEY)
7855 return reg_named_buff_firstkey(rx, flags);
7856 else if (flags & RXapif_NEXTKEY)
7857 return reg_named_buff_nextkey(rx, flags);
7859 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7866 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7869 AV *retarray = NULL;
7871 struct regexp *const rx = ReANY(r);
7873 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7875 if (flags & RXapif_ALL)
7878 if (rx && RXp_PAREN_NAMES(rx)) {
7879 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7882 SV* sv_dat=HeVAL(he_str);
7883 I32 *nums=(I32*)SvPVX(sv_dat);
7884 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7885 if ((I32)(rx->nparens) >= nums[i]
7886 && rx->offs[nums[i]].start != -1
7887 && rx->offs[nums[i]].end != -1)
7890 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7895 ret = newSVsv(&PL_sv_undef);
7898 av_push(retarray, ret);
7901 return newRV_noinc(MUTABLE_SV(retarray));
7908 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7911 struct regexp *const rx = ReANY(r);
7913 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7915 if (rx && RXp_PAREN_NAMES(rx)) {
7916 if (flags & RXapif_ALL) {
7917 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7919 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7921 SvREFCNT_dec_NN(sv);
7933 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7935 struct regexp *const rx = ReANY(r);
7937 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7939 if ( rx && RXp_PAREN_NAMES(rx) ) {
7940 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7942 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7949 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7951 struct regexp *const rx = ReANY(r);
7952 GET_RE_DEBUG_FLAGS_DECL;
7954 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7956 if (rx && RXp_PAREN_NAMES(rx)) {
7957 HV *hv = RXp_PAREN_NAMES(rx);
7959 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7962 SV* sv_dat = HeVAL(temphe);
7963 I32 *nums = (I32*)SvPVX(sv_dat);
7964 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7965 if ((I32)(rx->lastparen) >= nums[i] &&
7966 rx->offs[nums[i]].start != -1 &&
7967 rx->offs[nums[i]].end != -1)
7973 if (parno || flags & RXapif_ALL) {
7974 return newSVhek(HeKEY_hek(temphe));
7982 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
7987 struct regexp *const rx = ReANY(r);
7989 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
7991 if (rx && RXp_PAREN_NAMES(rx)) {
7992 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
7993 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
7994 } else if (flags & RXapif_ONE) {
7995 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
7996 av = MUTABLE_AV(SvRV(ret));
7997 length = av_tindex(av);
7998 SvREFCNT_dec_NN(ret);
7999 return newSViv(length + 1);
8001 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8006 return &PL_sv_undef;
8010 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8012 struct regexp *const rx = ReANY(r);
8015 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8017 if (rx && RXp_PAREN_NAMES(rx)) {
8018 HV *hv= RXp_PAREN_NAMES(rx);
8020 (void)hv_iterinit(hv);
8021 while ( (temphe = hv_iternext_flags(hv,0)) ) {
8024 SV* sv_dat = HeVAL(temphe);
8025 I32 *nums = (I32*)SvPVX(sv_dat);
8026 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8027 if ((I32)(rx->lastparen) >= nums[i] &&
8028 rx->offs[nums[i]].start != -1 &&
8029 rx->offs[nums[i]].end != -1)
8035 if (parno || flags & RXapif_ALL) {
8036 av_push(av, newSVhek(HeKEY_hek(temphe)));
8041 return newRV_noinc(MUTABLE_SV(av));
8045 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8048 struct regexp *const rx = ReANY(r);
8054 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8056 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8057 || n == RX_BUFF_IDX_CARET_FULLMATCH
8058 || n == RX_BUFF_IDX_CARET_POSTMATCH
8061 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8063 /* on something like
8066 * the KEEPCOPY is set on the PMOP rather than the regex */
8067 if (PL_curpm && r == PM_GETRE(PL_curpm))
8068 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8077 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8078 /* no need to distinguish between them any more */
8079 n = RX_BUFF_IDX_FULLMATCH;
8081 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8082 && rx->offs[0].start != -1)
8084 /* $`, ${^PREMATCH} */
8085 i = rx->offs[0].start;
8089 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8090 && rx->offs[0].end != -1)
8092 /* $', ${^POSTMATCH} */
8093 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8094 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8097 if ( 0 <= n && n <= (I32)rx->nparens &&
8098 (s1 = rx->offs[n].start) != -1 &&
8099 (t1 = rx->offs[n].end) != -1)
8101 /* $&, ${^MATCH}, $1 ... */
8103 s = rx->subbeg + s1 - rx->suboffset;
8108 assert(s >= rx->subbeg);
8109 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8111 #ifdef NO_TAINT_SUPPORT
8112 sv_setpvn(sv, s, i);
8114 const int oldtainted = TAINT_get;
8116 sv_setpvn(sv, s, i);
8117 TAINT_set(oldtainted);
8119 if (RXp_MATCH_UTF8(rx))
8124 if (RXp_MATCH_TAINTED(rx)) {
8125 if (SvTYPE(sv) >= SVt_PVMG) {
8126 MAGIC* const mg = SvMAGIC(sv);
8129 SvMAGIC_set(sv, mg->mg_moremagic);
8131 if ((mgt = SvMAGIC(sv))) {
8132 mg->mg_moremagic = mgt;
8133 SvMAGIC_set(sv, mg);
8150 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8151 SV const * const value)
8153 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8155 PERL_UNUSED_ARG(rx);
8156 PERL_UNUSED_ARG(paren);
8157 PERL_UNUSED_ARG(value);
8160 Perl_croak_no_modify();
8164 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8167 struct regexp *const rx = ReANY(r);
8171 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8173 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8174 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8175 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8178 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8180 /* on something like
8183 * the KEEPCOPY is set on the PMOP rather than the regex */
8184 if (PL_curpm && r == PM_GETRE(PL_curpm))
8185 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8191 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8193 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8194 case RX_BUFF_IDX_PREMATCH: /* $` */
8195 if (rx->offs[0].start != -1) {
8196 i = rx->offs[0].start;
8205 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8206 case RX_BUFF_IDX_POSTMATCH: /* $' */
8207 if (rx->offs[0].end != -1) {
8208 i = rx->sublen - rx->offs[0].end;
8210 s1 = rx->offs[0].end;
8217 default: /* $& / ${^MATCH}, $1, $2, ... */
8218 if (paren <= (I32)rx->nparens &&
8219 (s1 = rx->offs[paren].start) != -1 &&
8220 (t1 = rx->offs[paren].end) != -1)
8226 if (ckWARN(WARN_UNINITIALIZED))
8227 report_uninit((const SV *)sv);
8232 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8233 const char * const s = rx->subbeg - rx->suboffset + s1;
8238 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8245 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8247 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8248 PERL_UNUSED_ARG(rx);
8252 return newSVpvs("Regexp");
8255 /* Scans the name of a named buffer from the pattern.
8256 * If flags is REG_RSN_RETURN_NULL returns null.
8257 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8258 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8259 * to the parsed name as looked up in the RExC_paren_names hash.
8260 * If there is an error throws a vFAIL().. type exception.
8263 #define REG_RSN_RETURN_NULL 0
8264 #define REG_RSN_RETURN_NAME 1
8265 #define REG_RSN_RETURN_DATA 2
8268 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8270 char *name_start = RExC_parse;
8272 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8274 assert (RExC_parse <= RExC_end);
8275 if (RExC_parse == RExC_end) NOOP;
8276 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8277 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8278 * using do...while */
8281 RExC_parse += UTF8SKIP(RExC_parse);
8282 } while ( RExC_parse < RExC_end
8283 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8287 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8289 RExC_parse++; /* so the <- from the vFAIL is after the offending
8291 vFAIL("Group name must start with a non-digit word character");
8295 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8296 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8297 if ( flags == REG_RSN_RETURN_NAME)
8299 else if (flags==REG_RSN_RETURN_DATA) {
8302 if ( ! sv_name ) /* should not happen*/
8303 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8304 if (RExC_paren_names)
8305 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8307 sv_dat = HeVAL(he_str);
8309 vFAIL("Reference to nonexistent named group");
8313 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8314 (unsigned long) flags);
8316 NOT_REACHED; /* NOTREACHED */
8321 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8323 if (RExC_lastparse!=RExC_parse) { \
8324 Perl_re_printf( aTHX_ "%s", \
8325 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8326 RExC_end - RExC_parse, 16, \
8328 PERL_PV_ESCAPE_UNI_DETECT | \
8329 PERL_PV_PRETTY_ELLIPSES | \
8330 PERL_PV_PRETTY_LTGT | \
8331 PERL_PV_ESCAPE_RE | \
8332 PERL_PV_PRETTY_EXACTSIZE \
8336 Perl_re_printf( aTHX_ "%16s",""); \
8339 num = RExC_size + 1; \
8341 num=REG_NODE_NUM(RExC_emit); \
8342 if (RExC_lastnum!=num) \
8343 Perl_re_printf( aTHX_ "|%4d",num); \
8345 Perl_re_printf( aTHX_ "|%4s",""); \
8346 Perl_re_printf( aTHX_ "|%*s%-4s", \
8347 (int)((depth*2)), "", \
8351 RExC_lastparse=RExC_parse; \
8356 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8357 DEBUG_PARSE_MSG((funcname)); \
8358 Perl_re_printf( aTHX_ "%4s","\n"); \
8360 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8361 DEBUG_PARSE_MSG((funcname)); \
8362 Perl_re_printf( aTHX_ fmt "\n",args); \
8365 /* This section of code defines the inversion list object and its methods. The
8366 * interfaces are highly subject to change, so as much as possible is static to
8367 * this file. An inversion list is here implemented as a malloc'd C UV array
8368 * as an SVt_INVLIST scalar.
8370 * An inversion list for Unicode is an array of code points, sorted by ordinal
8371 * number. Each element gives the code point that begins a range that extends
8372 * up-to but not including the code point given by the next element. The final
8373 * element gives the first code point of a range that extends to the platform's
8374 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8375 * ...) give ranges whose code points are all in the inversion list. We say
8376 * that those ranges are in the set. The odd-numbered elements give ranges
8377 * whose code points are not in the inversion list, and hence not in the set.
8378 * Thus, element [0] is the first code point in the list. Element [1]
8379 * is the first code point beyond that not in the list; and element [2] is the
8380 * first code point beyond that that is in the list. In other words, the first
8381 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8382 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8383 * all code points in that range are not in the inversion list. The third
8384 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8385 * list, and so forth. Thus every element whose index is divisible by two
8386 * gives the beginning of a range that is in the list, and every element whose
8387 * index is not divisible by two gives the beginning of a range not in the
8388 * list. If the final element's index is divisible by two, the inversion list
8389 * extends to the platform's infinity; otherwise the highest code point in the
8390 * inversion list is the contents of that element minus 1.
8392 * A range that contains just a single code point N will look like
8394 * invlist[i+1] == N+1
8396 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8397 * impossible to represent, so element [i+1] is omitted. The single element
8399 * invlist[0] == UV_MAX
8400 * contains just UV_MAX, but is interpreted as matching to infinity.
8402 * Taking the complement (inverting) an inversion list is quite simple, if the
8403 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8404 * This implementation reserves an element at the beginning of each inversion
8405 * list to always contain 0; there is an additional flag in the header which
8406 * indicates if the list begins at the 0, or is offset to begin at the next
8407 * element. This means that the inversion list can be inverted without any
8408 * copying; just flip the flag.
8410 * More about inversion lists can be found in "Unicode Demystified"
8411 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8413 * The inversion list data structure is currently implemented as an SV pointing
8414 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8415 * array of UV whose memory management is automatically handled by the existing
8416 * facilities for SV's.
8418 * Some of the methods should always be private to the implementation, and some
8419 * should eventually be made public */
8421 /* The header definitions are in F<invlist_inline.h> */
8423 #ifndef PERL_IN_XSUB_RE
8425 PERL_STATIC_INLINE UV*
8426 S__invlist_array_init(SV* const invlist, const bool will_have_0)
8428 /* Returns a pointer to the first element in the inversion list's array.
8429 * This is called upon initialization of an inversion list. Where the
8430 * array begins depends on whether the list has the code point U+0000 in it
8431 * or not. The other parameter tells it whether the code that follows this
8432 * call is about to put a 0 in the inversion list or not. The first
8433 * element is either the element reserved for 0, if TRUE, or the element
8434 * after it, if FALSE */
8436 bool* offset = get_invlist_offset_addr(invlist);
8437 UV* zero_addr = (UV *) SvPVX(invlist);
8439 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8442 assert(! _invlist_len(invlist));
8446 /* 1^1 = 0; 1^0 = 1 */
8447 *offset = 1 ^ will_have_0;
8448 return zero_addr + *offset;
8453 PERL_STATIC_INLINE void
8454 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8456 /* Sets the current number of elements stored in the inversion list.
8457 * Updates SvCUR correspondingly */
8458 PERL_UNUSED_CONTEXT;
8459 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8461 assert(SvTYPE(invlist) == SVt_INVLIST);
8466 : TO_INTERNAL_SIZE(len + offset));
8467 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8470 #ifndef PERL_IN_XSUB_RE
8473 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
8475 /* Replaces the inversion list in 'dest' with the one from 'src'. It
8476 * steals the list from 'src', so 'src' is made to have a NULL list. This
8477 * is similar to what SvSetMagicSV() would do, if it were implemented on
8478 * inversion lists, though this routine avoids a copy */
8480 const UV src_len = _invlist_len(src);
8481 const bool src_offset = *get_invlist_offset_addr(src);
8482 const STRLEN src_byte_len = SvLEN(src);
8483 char * array = SvPVX(src);
8485 const int oldtainted = TAINT_get;
8487 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
8489 assert(SvTYPE(src) == SVt_INVLIST);
8490 assert(SvTYPE(dest) == SVt_INVLIST);
8491 assert(! invlist_is_iterating(src));
8492 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
8494 /* Make sure it ends in the right place with a NUL, as our inversion list
8495 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
8497 array[src_byte_len - 1] = '\0';
8499 TAINT_NOT; /* Otherwise it breaks */
8500 sv_usepvn_flags(dest,
8504 /* This flag is documented to cause a copy to be avoided */
8505 SV_HAS_TRAILING_NUL);
8506 TAINT_set(oldtainted);
8511 /* Finish up copying over the other fields in an inversion list */
8512 *get_invlist_offset_addr(dest) = src_offset;
8513 invlist_set_len(dest, src_len, src_offset);
8514 *get_invlist_previous_index_addr(dest) = 0;
8515 invlist_iterfinish(dest);
8518 PERL_STATIC_INLINE IV*
8519 S_get_invlist_previous_index_addr(SV* invlist)
8521 /* Return the address of the IV that is reserved to hold the cached index
8523 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8525 assert(SvTYPE(invlist) == SVt_INVLIST);
8527 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8530 PERL_STATIC_INLINE IV
8531 S_invlist_previous_index(SV* const invlist)
8533 /* Returns cached index of previous search */
8535 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8537 return *get_invlist_previous_index_addr(invlist);
8540 PERL_STATIC_INLINE void
8541 S_invlist_set_previous_index(SV* const invlist, const IV index)
8543 /* Caches <index> for later retrieval */
8545 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8547 assert(index == 0 || index < (int) _invlist_len(invlist));
8549 *get_invlist_previous_index_addr(invlist) = index;
8552 PERL_STATIC_INLINE void
8553 S_invlist_trim(SV* invlist)
8555 /* Free the not currently-being-used space in an inversion list */
8557 /* But don't free up the space needed for the 0 UV that is always at the
8558 * beginning of the list, nor the trailing NUL */
8559 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
8561 PERL_ARGS_ASSERT_INVLIST_TRIM;
8563 assert(SvTYPE(invlist) == SVt_INVLIST);
8565 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
8568 PERL_STATIC_INLINE void
8569 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
8571 PERL_ARGS_ASSERT_INVLIST_CLEAR;
8573 assert(SvTYPE(invlist) == SVt_INVLIST);
8575 invlist_set_len(invlist, 0, 0);
8576 invlist_trim(invlist);
8579 #endif /* ifndef PERL_IN_XSUB_RE */
8581 PERL_STATIC_INLINE bool
8582 S_invlist_is_iterating(SV* const invlist)
8584 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8586 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8589 #ifndef PERL_IN_XSUB_RE
8591 PERL_STATIC_INLINE UV
8592 S_invlist_max(SV* const invlist)
8594 /* Returns the maximum number of elements storable in the inversion list's
8595 * array, without having to realloc() */
8597 PERL_ARGS_ASSERT_INVLIST_MAX;
8599 assert(SvTYPE(invlist) == SVt_INVLIST);
8601 /* Assumes worst case, in which the 0 element is not counted in the
8602 * inversion list, so subtracts 1 for that */
8603 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8604 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8605 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8608 Perl__new_invlist(pTHX_ IV initial_size)
8611 /* Return a pointer to a newly constructed inversion list, with enough
8612 * space to store 'initial_size' elements. If that number is negative, a
8613 * system default is used instead */
8617 if (initial_size < 0) {
8621 /* Allocate the initial space */
8622 new_list = newSV_type(SVt_INVLIST);
8624 /* First 1 is in case the zero element isn't in the list; second 1 is for
8626 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8627 invlist_set_len(new_list, 0, 0);
8629 /* Force iterinit() to be used to get iteration to work */
8630 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8632 *get_invlist_previous_index_addr(new_list) = 0;
8638 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8640 /* Return a pointer to a newly constructed inversion list, initialized to
8641 * point to <list>, which has to be in the exact correct inversion list
8642 * form, including internal fields. Thus this is a dangerous routine that
8643 * should not be used in the wrong hands. The passed in 'list' contains
8644 * several header fields at the beginning that are not part of the
8645 * inversion list body proper */
8647 const STRLEN length = (STRLEN) list[0];
8648 const UV version_id = list[1];
8649 const bool offset = cBOOL(list[2]);
8650 #define HEADER_LENGTH 3
8651 /* If any of the above changes in any way, you must change HEADER_LENGTH
8652 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8653 * perl -E 'say int(rand 2**31-1)'
8655 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8656 data structure type, so that one being
8657 passed in can be validated to be an
8658 inversion list of the correct vintage.
8661 SV* invlist = newSV_type(SVt_INVLIST);
8663 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8665 if (version_id != INVLIST_VERSION_ID) {
8666 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8669 /* The generated array passed in includes header elements that aren't part
8670 * of the list proper, so start it just after them */
8671 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8673 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8674 shouldn't touch it */
8676 *(get_invlist_offset_addr(invlist)) = offset;
8678 /* The 'length' passed to us is the physical number of elements in the
8679 * inversion list. But if there is an offset the logical number is one
8681 invlist_set_len(invlist, length - offset, offset);
8683 invlist_set_previous_index(invlist, 0);
8685 /* Initialize the iteration pointer. */
8686 invlist_iterfinish(invlist);
8688 SvREADONLY_on(invlist);
8694 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8696 /* Grow the maximum size of an inversion list */
8698 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8700 assert(SvTYPE(invlist) == SVt_INVLIST);
8702 /* Add one to account for the zero element at the beginning which may not
8703 * be counted by the calling parameters */
8704 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8708 S__append_range_to_invlist(pTHX_ SV* const invlist,
8709 const UV start, const UV end)
8711 /* Subject to change or removal. Append the range from 'start' to 'end' at
8712 * the end of the inversion list. The range must be above any existing
8716 UV max = invlist_max(invlist);
8717 UV len = _invlist_len(invlist);
8720 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8722 if (len == 0) { /* Empty lists must be initialized */
8723 offset = start != 0;
8724 array = _invlist_array_init(invlist, ! offset);
8727 /* Here, the existing list is non-empty. The current max entry in the
8728 * list is generally the first value not in the set, except when the
8729 * set extends to the end of permissible values, in which case it is
8730 * the first entry in that final set, and so this call is an attempt to
8731 * append out-of-order */
8733 UV final_element = len - 1;
8734 array = invlist_array(invlist);
8735 if ( array[final_element] > start
8736 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8738 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",
8739 array[final_element], start,
8740 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8743 /* Here, it is a legal append. If the new range begins 1 above the end
8744 * of the range below it, it is extending the range below it, so the
8745 * new first value not in the set is one greater than the newly
8746 * extended range. */
8747 offset = *get_invlist_offset_addr(invlist);
8748 if (array[final_element] == start) {
8749 if (end != UV_MAX) {
8750 array[final_element] = end + 1;
8753 /* But if the end is the maximum representable on the machine,
8754 * assume that infinity was actually what was meant. Just let
8755 * the range that this would extend to have no end */
8756 invlist_set_len(invlist, len - 1, offset);
8762 /* Here the new range doesn't extend any existing set. Add it */
8764 len += 2; /* Includes an element each for the start and end of range */
8766 /* If wll overflow the existing space, extend, which may cause the array to
8769 invlist_extend(invlist, len);
8771 /* Have to set len here to avoid assert failure in invlist_array() */
8772 invlist_set_len(invlist, len, offset);
8774 array = invlist_array(invlist);
8777 invlist_set_len(invlist, len, offset);
8780 /* The next item on the list starts the range, the one after that is
8781 * one past the new range. */
8782 array[len - 2] = start;
8783 if (end != UV_MAX) {
8784 array[len - 1] = end + 1;
8787 /* But if the end is the maximum representable on the machine, just let
8788 * the range have no end */
8789 invlist_set_len(invlist, len - 1, offset);
8794 Perl__invlist_search(SV* const invlist, const UV cp)
8796 /* Searches the inversion list for the entry that contains the input code
8797 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8798 * return value is the index into the list's array of the range that
8799 * contains <cp>, that is, 'i' such that
8800 * array[i] <= cp < array[i+1]
8805 IV high = _invlist_len(invlist);
8806 const IV highest_element = high - 1;
8809 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8811 /* If list is empty, return failure. */
8816 /* (We can't get the array unless we know the list is non-empty) */
8817 array = invlist_array(invlist);
8819 mid = invlist_previous_index(invlist);
8821 if (mid > highest_element) {
8822 mid = highest_element;
8825 /* <mid> contains the cache of the result of the previous call to this
8826 * function (0 the first time). See if this call is for the same result,
8827 * or if it is for mid-1. This is under the theory that calls to this
8828 * function will often be for related code points that are near each other.
8829 * And benchmarks show that caching gives better results. We also test
8830 * here if the code point is within the bounds of the list. These tests
8831 * replace others that would have had to be made anyway to make sure that
8832 * the array bounds were not exceeded, and these give us extra information
8833 * at the same time */
8834 if (cp >= array[mid]) {
8835 if (cp >= array[highest_element]) {
8836 return highest_element;
8839 /* Here, array[mid] <= cp < array[highest_element]. This means that
8840 * the final element is not the answer, so can exclude it; it also
8841 * means that <mid> is not the final element, so can refer to 'mid + 1'
8843 if (cp < array[mid + 1]) {
8849 else { /* cp < aray[mid] */
8850 if (cp < array[0]) { /* Fail if outside the array */
8854 if (cp >= array[mid - 1]) {
8859 /* Binary search. What we are looking for is <i> such that
8860 * array[i] <= cp < array[i+1]
8861 * The loop below converges on the i+1. Note that there may not be an
8862 * (i+1)th element in the array, and things work nonetheless */
8863 while (low < high) {
8864 mid = (low + high) / 2;
8865 assert(mid <= highest_element);
8866 if (array[mid] <= cp) { /* cp >= array[mid] */
8869 /* We could do this extra test to exit the loop early.
8870 if (cp < array[low]) {
8875 else { /* cp < array[mid] */
8882 invlist_set_previous_index(invlist, high);
8887 Perl__invlist_populate_swatch(SV* const invlist,
8888 const UV start, const UV end, U8* swatch)
8890 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8891 * but is used when the swash has an inversion list. This makes this much
8892 * faster, as it uses a binary search instead of a linear one. This is
8893 * intimately tied to that function, and perhaps should be in utf8.c,
8894 * except it is intimately tied to inversion lists as well. It assumes
8895 * that <swatch> is all 0's on input */
8898 const IV len = _invlist_len(invlist);
8902 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8904 if (len == 0) { /* Empty inversion list */
8908 array = invlist_array(invlist);
8910 /* Find which element it is */
8911 i = _invlist_search(invlist, start);
8913 /* We populate from <start> to <end> */
8914 while (current < end) {
8917 /* The inversion list gives the results for every possible code point
8918 * after the first one in the list. Only those ranges whose index is
8919 * even are ones that the inversion list matches. For the odd ones,
8920 * and if the initial code point is not in the list, we have to skip
8921 * forward to the next element */
8922 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8924 if (i >= len) { /* Finished if beyond the end of the array */
8928 if (current >= end) { /* Finished if beyond the end of what we
8930 if (LIKELY(end < UV_MAX)) {
8934 /* We get here when the upper bound is the maximum
8935 * representable on the machine, and we are looking for just
8936 * that code point. Have to special case it */
8938 goto join_end_of_list;
8941 assert(current >= start);
8943 /* The current range ends one below the next one, except don't go past
8946 upper = (i < len && array[i] < end) ? array[i] : end;
8948 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8949 * for each code point in it */
8950 for (; current < upper; current++) {
8951 const STRLEN offset = (STRLEN)(current - start);
8952 swatch[offset >> 3] |= 1 << (offset & 7);
8957 /* Quit if at the end of the list */
8960 /* But first, have to deal with the highest possible code point on
8961 * the platform. The previous code assumes that <end> is one
8962 * beyond where we want to populate, but that is impossible at the
8963 * platform's infinity, so have to handle it specially */
8964 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8966 const STRLEN offset = (STRLEN)(end - start);
8967 swatch[offset >> 3] |= 1 << (offset & 7);
8972 /* Advance to the next range, which will be for code points not in the
8981 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8982 const bool complement_b, SV** output)
8984 /* Take the union of two inversion lists and point '*output' to it. On
8985 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
8986 * even 'a' or 'b'). If to an inversion list, the contents of the original
8987 * list will be replaced by the union. The first list, 'a', may be
8988 * NULL, in which case a copy of the second list is placed in '*output'.
8989 * If 'complement_b' is TRUE, the union is taken of the complement
8990 * (inversion) of 'b' instead of b itself.
8992 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8993 * Richard Gillam, published by Addison-Wesley, and explained at some
8994 * length there. The preface says to incorporate its examples into your
8995 * code at your own risk.
8997 * The algorithm is like a merge sort. */
8999 const UV* array_a; /* a's array */
9001 UV len_a; /* length of a's array */
9004 SV* u; /* the resulting union */
9008 UV i_a = 0; /* current index into a's array */
9012 /* running count, as explained in the algorithm source book; items are
9013 * stopped accumulating and are output when the count changes to/from 0.
9014 * The count is incremented when we start a range that's in an input's set,
9015 * and decremented when we start a range that's not in a set. So this
9016 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9017 * and hence nothing goes into the union; 1, just one of the inputs is in
9018 * its set (and its current range gets added to the union); and 2 when both
9019 * inputs are in their sets. */
9022 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9024 assert(*output == NULL || SvTYPE(*output) == SVt_INVLIST);
9026 len_b = _invlist_len(b);
9029 /* Here, 'b' is empty, hence it's complement is all possible code
9030 * points. So if the union includes the complement of 'b', it includes
9031 * everything, and we need not even look at 'a'. It's easiest to
9032 * create a new inversion list that matches everything. */
9034 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9036 if (*output == NULL) { /* If the output didn't exist, just point it
9038 *output = everything;
9040 else { /* Otherwise, replace its contents with the new list */
9041 invlist_replace_list_destroys_src(*output, everything);
9042 SvREFCNT_dec_NN(everything);
9048 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9049 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9050 * output will be empty */
9052 if (a == NULL || _invlist_len(a) == 0) {
9053 if (*output == NULL) {
9054 *output = _new_invlist(0);
9057 invlist_clear(*output);
9062 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9063 * union. We can just return a copy of 'a' if '*output' doesn't point
9064 * to an existing list */
9065 if (*output == NULL) {
9066 *output = invlist_clone(a);
9070 /* If the output is to overwrite 'a', we have a no-op, as it's
9076 /* Here, '*output' is to be overwritten by 'a' */
9077 u = invlist_clone(a);
9078 invlist_replace_list_destroys_src(*output, u);
9084 /* Here 'b' is not empty. See about 'a' */
9086 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9088 /* Here, 'a' is empty (and b is not). That means the union will come
9089 * entirely from 'b'. If '*output' is NULL, we can directly return a
9090 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9093 SV ** dest = (*output == NULL) ? output : &u;
9094 *dest = invlist_clone(b);
9096 _invlist_invert(*dest);
9100 invlist_replace_list_destroys_src(*output, u);
9107 /* Here both lists exist and are non-empty */
9108 array_a = invlist_array(a);
9109 array_b = invlist_array(b);
9111 /* If are to take the union of 'a' with the complement of b, set it
9112 * up so are looking at b's complement. */
9115 /* To complement, we invert: if the first element is 0, remove it. To
9116 * do this, we just pretend the array starts one later */
9117 if (array_b[0] == 0) {
9123 /* But if the first element is not zero, we pretend the list starts
9124 * at the 0 that is always stored immediately before the array. */
9130 /* Size the union for the worst case: that the sets are completely
9132 u = _new_invlist(len_a + len_b);
9134 /* Will contain U+0000 if either component does */
9135 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9136 || (len_b > 0 && array_b[0] == 0));
9138 /* Go through each input list item by item, stopping when have exhausted
9140 while (i_a < len_a && i_b < len_b) {
9141 UV cp; /* The element to potentially add to the union's array */
9142 bool cp_in_set; /* is it in the the input list's set or not */
9144 /* We need to take one or the other of the two inputs for the union.
9145 * Since we are merging two sorted lists, we take the smaller of the
9146 * next items. In case of a tie, we take first the one that is in its
9147 * set. If we first took the one not in its set, it would decrement
9148 * the count, possibly to 0 which would cause it to be output as ending
9149 * the range, and the next time through we would take the same number,
9150 * and output it again as beginning the next range. By doing it the
9151 * opposite way, there is no possibility that the count will be
9152 * momentarily decremented to 0, and thus the two adjoining ranges will
9153 * be seamlessly merged. (In a tie and both are in the set or both not
9154 * in the set, it doesn't matter which we take first.) */
9155 if ( array_a[i_a] < array_b[i_b]
9156 || ( array_a[i_a] == array_b[i_b]
9157 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9159 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9160 cp = array_a[i_a++];
9163 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9164 cp = array_b[i_b++];
9167 /* Here, have chosen which of the two inputs to look at. Only output
9168 * if the running count changes to/from 0, which marks the
9169 * beginning/end of a range that's in the set */
9172 array_u[i_u++] = cp;
9179 array_u[i_u++] = cp;
9185 /* The loop above increments the index into exactly one of the input lists
9186 * each iteration, and ends when either index gets to its list end. That
9187 * means the other index is lower than its end, and so something is
9188 * remaining in that one. We decrement 'count', as explained below, if
9189 * that list is in its set. (i_a and i_b each currently index the element
9190 * beyond the one we care about.) */
9191 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9192 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9197 /* Above we decremented 'count' if the list that had unexamined elements in
9198 * it was in its set. This has made it so that 'count' being non-zero
9199 * means there isn't anything left to output; and 'count' equal to 0 means
9200 * that what is left to output is precisely that which is left in the
9201 * non-exhausted input list.
9203 * To see why, note first that the exhausted input obviously has nothing
9204 * left to add to the union. If it was in its set at its end, that means
9205 * the set extends from here to the platform's infinity, and hence so does
9206 * the union and the non-exhausted set is irrelevant. The exhausted set
9207 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9208 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9209 * 'count' remains at 1. This is consistent with the decremented 'count'
9210 * != 0 meaning there's nothing left to add to the union.
9212 * But if the exhausted input wasn't in its set, it contributed 0 to
9213 * 'count', and the rest of the union will be whatever the other input is.
9214 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9215 * otherwise it gets decremented to 0. This is consistent with 'count'
9216 * == 0 meaning the remainder of the union is whatever is left in the
9217 * non-exhausted list. */
9222 IV copy_count = len_a - i_a;
9223 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9224 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9226 else { /* The non-exhausted input is b */
9227 copy_count = len_b - i_b;
9228 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9230 len_u = i_u + copy_count;
9233 /* Set the result to the final length, which can change the pointer to
9234 * array_u, so re-find it. (Note that it is unlikely that this will
9235 * change, as we are shrinking the space, not enlarging it) */
9236 if (len_u != _invlist_len(u)) {
9237 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9239 array_u = invlist_array(u);
9242 if (*output == NULL) { /* Simply return the new inversion list */
9246 /* Otherwise, overwrite the inversion list that was in '*output'. We
9247 * could instead free '*output', and then set it to 'u', but experience
9248 * has shown [perl #127392] that if the input is a mortal, we can get a
9249 * huge build-up of these during regex compilation before they get
9251 invlist_replace_list_destroys_src(*output, u);
9259 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9260 const bool complement_b, SV** i)
9262 /* Take the intersection of two inversion lists and point '*i' to it. On
9263 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9264 * even 'a' or 'b'). If to an inversion list, the contents of the original
9265 * list will be replaced by the intersection. The first list, 'a', may be
9266 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9267 * TRUE, the result will be the intersection of 'a' and the complement (or
9268 * inversion) of 'b' instead of 'b' directly.
9270 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9271 * Richard Gillam, published by Addison-Wesley, and explained at some
9272 * length there. The preface says to incorporate its examples into your
9273 * code at your own risk. In fact, it had bugs
9275 * The algorithm is like a merge sort, and is essentially the same as the
9279 const UV* array_a; /* a's array */
9281 UV len_a; /* length of a's array */
9284 SV* r; /* the resulting intersection */
9288 UV i_a = 0; /* current index into a's array */
9292 /* running count of how many of the two inputs are postitioned at ranges
9293 * that are in their sets. As explained in the algorithm source book,
9294 * items are stopped accumulating and are output when the count changes
9295 * to/from 2. The count is incremented when we start a range that's in an
9296 * input's set, and decremented when we start a range that's not in a set.
9297 * Only when it is 2 are we in the intersection. */
9300 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9302 assert(*i == NULL || SvTYPE(*i) == SVt_INVLIST);
9304 /* Special case if either one is empty */
9305 len_a = (a == NULL) ? 0 : _invlist_len(a);
9306 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9307 if (len_a != 0 && complement_b) {
9309 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9310 * must be empty. Here, also we are using 'b's complement, which
9311 * hence must be every possible code point. Thus the intersection
9314 if (*i == a) { /* No-op */
9319 *i = invlist_clone(a);
9323 r = invlist_clone(a);
9324 invlist_replace_list_destroys_src(*i, r);
9329 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9330 * intersection must be empty */
9332 *i = _new_invlist(0);
9340 /* Here both lists exist and are non-empty */
9341 array_a = invlist_array(a);
9342 array_b = invlist_array(b);
9344 /* If are to take the intersection of 'a' with the complement of b, set it
9345 * up so are looking at b's complement. */
9348 /* To complement, we invert: if the first element is 0, remove it. To
9349 * do this, we just pretend the array starts one later */
9350 if (array_b[0] == 0) {
9356 /* But if the first element is not zero, we pretend the list starts
9357 * at the 0 that is always stored immediately before the array. */
9363 /* Size the intersection for the worst case: that the intersection ends up
9364 * fragmenting everything to be completely disjoint */
9365 r= _new_invlist(len_a + len_b);
9367 /* Will contain U+0000 iff both components do */
9368 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9369 && len_b > 0 && array_b[0] == 0);
9371 /* Go through each list item by item, stopping when have exhausted one of
9373 while (i_a < len_a && i_b < len_b) {
9374 UV cp; /* The element to potentially add to the intersection's
9376 bool cp_in_set; /* Is it in the input list's set or not */
9378 /* We need to take one or the other of the two inputs for the
9379 * intersection. Since we are merging two sorted lists, we take the
9380 * smaller of the next items. In case of a tie, we take first the one
9381 * that is not in its set (a difference from the union algorithm). If
9382 * we first took the one in its set, it would increment the count,
9383 * possibly to 2 which would cause it to be output as starting a range
9384 * in the intersection, and the next time through we would take that
9385 * same number, and output it again as ending the set. By doing the
9386 * opposite of this, there is no possibility that the count will be
9387 * momentarily incremented to 2. (In a tie and both are in the set or
9388 * both not in the set, it doesn't matter which we take first.) */
9389 if ( array_a[i_a] < array_b[i_b]
9390 || ( array_a[i_a] == array_b[i_b]
9391 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9393 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9394 cp = array_a[i_a++];
9397 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9401 /* Here, have chosen which of the two inputs to look at. Only output
9402 * if the running count changes to/from 2, which marks the
9403 * beginning/end of a range that's in the intersection */
9407 array_r[i_r++] = cp;
9412 array_r[i_r++] = cp;
9419 /* The loop above increments the index into exactly one of the input lists
9420 * each iteration, and ends when either index gets to its list end. That
9421 * means the other index is lower than its end, and so something is
9422 * remaining in that one. We increment 'count', as explained below, if the
9423 * exhausted list was in its set. (i_a and i_b each currently index the
9424 * element beyond the one we care about.) */
9425 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9426 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9431 /* Above we incremented 'count' if the exhausted list was in its set. This
9432 * has made it so that 'count' being below 2 means there is nothing left to
9433 * output; otheriwse what's left to add to the intersection is precisely
9434 * that which is left in the non-exhausted input list.
9436 * To see why, note first that the exhausted input obviously has nothing
9437 * left to affect the intersection. If it was in its set at its end, that
9438 * means the set extends from here to the platform's infinity, and hence
9439 * anything in the non-exhausted's list will be in the intersection, and
9440 * anything not in it won't be. Hence, the rest of the intersection is
9441 * precisely what's in the non-exhausted list The exhausted set also
9442 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9443 * it means 'count' is now at least 2. This is consistent with the
9444 * incremented 'count' being >= 2 means to add the non-exhausted list to
9447 * But if the exhausted input wasn't in its set, it contributed 0 to
9448 * 'count', and the intersection can't include anything further; the
9449 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9450 * incremented. This is consistent with 'count' being < 2 meaning nothing
9451 * further to add to the intersection. */
9452 if (count < 2) { /* Nothing left to put in the intersection. */
9455 else { /* copy the non-exhausted list, unchanged. */
9456 IV copy_count = len_a - i_a;
9457 if (copy_count > 0) { /* a is the one with stuff left */
9458 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9460 else { /* b is the one with stuff left */
9461 copy_count = len_b - i_b;
9462 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9464 len_r = i_r + copy_count;
9467 /* Set the result to the final length, which can change the pointer to
9468 * array_r, so re-find it. (Note that it is unlikely that this will
9469 * change, as we are shrinking the space, not enlarging it) */
9470 if (len_r != _invlist_len(r)) {
9471 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9473 array_r = invlist_array(r);
9476 if (*i == NULL) { /* Simply return the calculated intersection */
9479 else { /* Otherwise, replace the existing inversion list in '*i'. We could
9480 instead free '*i', and then set it to 'r', but experience has
9481 shown [perl #127392] that if the input is a mortal, we can get a
9482 huge build-up of these during regex compilation before they get
9485 invlist_replace_list_destroys_src(*i, r);
9497 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9499 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9500 * set. A pointer to the inversion list is returned. This may actually be
9501 * a new list, in which case the passed in one has been destroyed. The
9502 * passed-in inversion list can be NULL, in which case a new one is created
9503 * with just the one range in it. The new list is not necessarily
9504 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9505 * result of this function. The gain would not be large, and in many
9506 * cases, this is called multiple times on a single inversion list, so
9507 * anything freed may almost immediately be needed again.
9509 * This used to mostly call the 'union' routine, but that is much more
9510 * heavyweight than really needed for a single range addition */
9512 UV* array; /* The array implementing the inversion list */
9513 UV len; /* How many elements in 'array' */
9514 SSize_t i_s; /* index into the invlist array where 'start'
9516 SSize_t i_e = 0; /* And the index where 'end' should go */
9517 UV cur_highest; /* The highest code point in the inversion list
9518 upon entry to this function */
9520 /* This range becomes the whole inversion list if none already existed */
9521 if (invlist == NULL) {
9522 invlist = _new_invlist(2);
9523 _append_range_to_invlist(invlist, start, end);
9527 /* Likewise, if the inversion list is currently empty */
9528 len = _invlist_len(invlist);
9530 _append_range_to_invlist(invlist, start, end);
9534 /* Starting here, we have to know the internals of the list */
9535 array = invlist_array(invlist);
9537 /* If the new range ends higher than the current highest ... */
9538 cur_highest = invlist_highest(invlist);
9539 if (end > cur_highest) {
9541 /* If the whole range is higher, we can just append it */
9542 if (start > cur_highest) {
9543 _append_range_to_invlist(invlist, start, end);
9547 /* Otherwise, add the portion that is higher ... */
9548 _append_range_to_invlist(invlist, cur_highest + 1, end);
9550 /* ... and continue on below to handle the rest. As a result of the
9551 * above append, we know that the index of the end of the range is the
9552 * final even numbered one of the array. Recall that the final element
9553 * always starts a range that extends to infinity. If that range is in
9554 * the set (meaning the set goes from here to infinity), it will be an
9555 * even index, but if it isn't in the set, it's odd, and the final
9556 * range in the set is one less, which is even. */
9557 if (end == UV_MAX) {
9565 /* We have dealt with appending, now see about prepending. If the new
9566 * range starts lower than the current lowest ... */
9567 if (start < array[0]) {
9569 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
9570 * Let the union code handle it, rather than having to know the
9571 * trickiness in two code places. */
9572 if (UNLIKELY(start == 0)) {
9575 range_invlist = _new_invlist(2);
9576 _append_range_to_invlist(range_invlist, start, end);
9578 _invlist_union(invlist, range_invlist, &invlist);
9580 SvREFCNT_dec_NN(range_invlist);
9585 /* If the whole new range comes before the first entry, and doesn't
9586 * extend it, we have to insert it as an additional range */
9587 if (end < array[0] - 1) {
9589 goto splice_in_new_range;
9592 /* Here the new range adjoins the existing first range, extending it
9596 /* And continue on below to handle the rest. We know that the index of
9597 * the beginning of the range is the first one of the array */
9600 else { /* Not prepending any part of the new range to the existing list.
9601 * Find where in the list it should go. This finds i_s, such that:
9602 * invlist[i_s] <= start < array[i_s+1]
9604 i_s = _invlist_search(invlist, start);
9607 /* At this point, any extending before the beginning of the inversion list
9608 * and/or after the end has been done. This has made it so that, in the
9609 * code below, each endpoint of the new range is either in a range that is
9610 * in the set, or is in a gap between two ranges that are. This means we
9611 * don't have to worry about exceeding the array bounds.
9613 * Find where in the list the new range ends (but we can skip this if we
9614 * have already determined what it is, or if it will be the same as i_s,
9615 * which we already have computed) */
9617 i_e = (start == end)
9619 : _invlist_search(invlist, end);
9622 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
9623 * is a range that goes to infinity there is no element at invlist[i_e+1],
9624 * so only the first relation holds. */
9626 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9628 /* Here, the ranges on either side of the beginning of the new range
9629 * are in the set, and this range starts in the gap between them.
9631 * The new range extends the range above it downwards if the new range
9632 * ends at or above that range's start */
9633 const bool extends_the_range_above = ( end == UV_MAX
9634 || end + 1 >= array[i_s+1]);
9636 /* The new range extends the range below it upwards if it begins just
9637 * after where that range ends */
9638 if (start == array[i_s]) {
9640 /* If the new range fills the entire gap between the other ranges,
9641 * they will get merged together. Other ranges may also get
9642 * merged, depending on how many of them the new range spans. In
9643 * the general case, we do the merge later, just once, after we
9644 * figure out how many to merge. But in the case where the new
9645 * range exactly spans just this one gap (possibly extending into
9646 * the one above), we do the merge here, and an early exit. This
9647 * is done here to avoid having to special case later. */
9648 if (i_e - i_s <= 1) {
9650 /* If i_e - i_s == 1, it means that the new range terminates
9651 * within the range above, and hence 'extends_the_range_above'
9652 * must be true. (If the range above it extends to infinity,
9653 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
9654 * will be 0, so no harm done.) */
9655 if (extends_the_range_above) {
9656 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
9657 invlist_set_len(invlist,
9659 *(get_invlist_offset_addr(invlist)));
9663 /* Here, i_e must == i_s. We keep them in sync, as they apply
9664 * to the same range, and below we are about to decrement i_s
9669 /* Here, the new range is adjacent to the one below. (It may also
9670 * span beyond the range above, but that will get resolved later.)
9671 * Extend the range below to include this one. */
9672 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
9676 else if (extends_the_range_above) {
9678 /* Here the new range only extends the range above it, but not the
9679 * one below. It merges with the one above. Again, we keep i_e
9680 * and i_s in sync if they point to the same range */
9689 /* Here, we've dealt with the new range start extending any adjoining
9692 * If the new range extends to infinity, it is now the final one,
9693 * regardless of what was there before */
9694 if (UNLIKELY(end == UV_MAX)) {
9695 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
9699 /* If i_e started as == i_s, it has also been dealt with,
9700 * and been updated to the new i_s, which will fail the following if */
9701 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
9703 /* Here, the ranges on either side of the end of the new range are in
9704 * the set, and this range ends in the gap between them.
9706 * If this range is adjacent to (hence extends) the range above it, it
9707 * becomes part of that range; likewise if it extends the range below,
9708 * it becomes part of that range */
9709 if (end + 1 == array[i_e+1]) {
9713 else if (start <= array[i_e]) {
9714 array[i_e] = end + 1;
9721 /* If the range fits entirely in an existing range (as possibly already
9722 * extended above), it doesn't add anything new */
9723 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9727 /* Here, no part of the range is in the list. Must add it. It will
9728 * occupy 2 more slots */
9729 splice_in_new_range:
9731 invlist_extend(invlist, len + 2);
9732 array = invlist_array(invlist);
9733 /* Move the rest of the array down two slots. Don't include any
9735 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
9737 /* Do the actual splice */
9738 array[i_e+1] = start;
9739 array[i_e+2] = end + 1;
9740 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
9744 /* Here the new range crossed the boundaries of a pre-existing range. The
9745 * code above has adjusted things so that both ends are in ranges that are
9746 * in the set. This means everything in between must also be in the set.
9747 * Just squash things together */
9748 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
9749 invlist_set_len(invlist,
9751 *(get_invlist_offset_addr(invlist)));
9757 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9758 UV** other_elements_ptr)
9760 /* Create and return an inversion list whose contents are to be populated
9761 * by the caller. The caller gives the number of elements (in 'size') and
9762 * the very first element ('element0'). This function will set
9763 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9766 * Obviously there is some trust involved that the caller will properly
9767 * fill in the other elements of the array.
9769 * (The first element needs to be passed in, as the underlying code does
9770 * things differently depending on whether it is zero or non-zero) */
9772 SV* invlist = _new_invlist(size);
9775 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9777 invlist = add_cp_to_invlist(invlist, element0);
9778 offset = *get_invlist_offset_addr(invlist);
9780 invlist_set_len(invlist, size, offset);
9781 *other_elements_ptr = invlist_array(invlist) + 1;
9787 PERL_STATIC_INLINE SV*
9788 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9789 return _add_range_to_invlist(invlist, cp, cp);
9792 #ifndef PERL_IN_XSUB_RE
9794 Perl__invlist_invert(pTHX_ SV* const invlist)
9796 /* Complement the input inversion list. This adds a 0 if the list didn't
9797 * have a zero; removes it otherwise. As described above, the data
9798 * structure is set up so that this is very efficient */
9800 PERL_ARGS_ASSERT__INVLIST_INVERT;
9802 assert(! invlist_is_iterating(invlist));
9804 /* The inverse of matching nothing is matching everything */
9805 if (_invlist_len(invlist) == 0) {
9806 _append_range_to_invlist(invlist, 0, UV_MAX);
9810 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9815 PERL_STATIC_INLINE SV*
9816 S_invlist_clone(pTHX_ SV* const invlist)
9819 /* Return a new inversion list that is a copy of the input one, which is
9820 * unchanged. The new list will not be mortal even if the old one was. */
9822 /* Need to allocate extra space to accommodate Perl's addition of a
9823 * trailing NUL to SvPV's, since it thinks they are always strings */
9824 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9825 STRLEN physical_length = SvCUR(invlist);
9826 bool offset = *(get_invlist_offset_addr(invlist));
9828 PERL_ARGS_ASSERT_INVLIST_CLONE;
9830 *(get_invlist_offset_addr(new_invlist)) = offset;
9831 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9832 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9837 PERL_STATIC_INLINE STRLEN*
9838 S_get_invlist_iter_addr(SV* invlist)
9840 /* Return the address of the UV that contains the current iteration
9843 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9845 assert(SvTYPE(invlist) == SVt_INVLIST);
9847 return &(((XINVLIST*) SvANY(invlist))->iterator);
9850 PERL_STATIC_INLINE void
9851 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9853 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9855 *get_invlist_iter_addr(invlist) = 0;
9858 PERL_STATIC_INLINE void
9859 S_invlist_iterfinish(SV* invlist)
9861 /* Terminate iterator for invlist. This is to catch development errors.
9862 * Any iteration that is interrupted before completed should call this
9863 * function. Functions that add code points anywhere else but to the end
9864 * of an inversion list assert that they are not in the middle of an
9865 * iteration. If they were, the addition would make the iteration
9866 * problematical: if the iteration hadn't reached the place where things
9867 * were being added, it would be ok */
9869 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9871 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9875 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9877 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9878 * This call sets in <*start> and <*end>, the next range in <invlist>.
9879 * Returns <TRUE> if successful and the next call will return the next
9880 * range; <FALSE> if was already at the end of the list. If the latter,
9881 * <*start> and <*end> are unchanged, and the next call to this function
9882 * will start over at the beginning of the list */
9884 STRLEN* pos = get_invlist_iter_addr(invlist);
9885 UV len = _invlist_len(invlist);
9888 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9891 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9895 array = invlist_array(invlist);
9897 *start = array[(*pos)++];
9903 *end = array[(*pos)++] - 1;
9909 PERL_STATIC_INLINE UV
9910 S_invlist_highest(SV* const invlist)
9912 /* Returns the highest code point that matches an inversion list. This API
9913 * has an ambiguity, as it returns 0 under either the highest is actually
9914 * 0, or if the list is empty. If this distinction matters to you, check
9915 * for emptiness before calling this function */
9917 UV len = _invlist_len(invlist);
9920 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9926 array = invlist_array(invlist);
9928 /* The last element in the array in the inversion list always starts a
9929 * range that goes to infinity. That range may be for code points that are
9930 * matched in the inversion list, or it may be for ones that aren't
9931 * matched. In the latter case, the highest code point in the set is one
9932 * less than the beginning of this range; otherwise it is the final element
9933 * of this range: infinity */
9934 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
9936 : array[len - 1] - 1;
9940 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
9942 /* Get the contents of an inversion list into a string SV so that they can
9943 * be printed out. If 'traditional_style' is TRUE, it uses the format
9944 * traditionally done for debug tracing; otherwise it uses a format
9945 * suitable for just copying to the output, with blanks between ranges and
9946 * a dash between range components */
9950 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
9951 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
9953 if (traditional_style) {
9954 output = newSVpvs("\n");
9957 output = newSVpvs("");
9960 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
9962 assert(! invlist_is_iterating(invlist));
9964 invlist_iterinit(invlist);
9965 while (invlist_iternext(invlist, &start, &end)) {
9966 if (end == UV_MAX) {
9967 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFINITY%c",
9968 start, intra_range_delimiter,
9969 inter_range_delimiter);
9971 else if (end != start) {
9972 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
9974 intra_range_delimiter,
9975 end, inter_range_delimiter);
9978 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
9979 start, inter_range_delimiter);
9983 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
9984 SvCUR_set(output, SvCUR(output) - 1);
9990 #ifndef PERL_IN_XSUB_RE
9992 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
9993 const char * const indent, SV* const invlist)
9995 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
9996 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
9997 * the string 'indent'. The output looks like this:
9998 [0] 0x000A .. 0x000D
10000 [4] 0x2028 .. 0x2029
10001 [6] 0x3104 .. INFINITY
10002 * This means that the first range of code points matched by the list are
10003 * 0xA through 0xD; the second range contains only the single code point
10004 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10005 * are used to define each range (except if the final range extends to
10006 * infinity, only a single element is needed). The array index of the
10007 * first element for the corresponding range is given in brackets. */
10012 PERL_ARGS_ASSERT__INVLIST_DUMP;
10014 if (invlist_is_iterating(invlist)) {
10015 Perl_dump_indent(aTHX_ level, file,
10016 "%sCan't dump inversion list because is in middle of iterating\n",
10021 invlist_iterinit(invlist);
10022 while (invlist_iternext(invlist, &start, &end)) {
10023 if (end == UV_MAX) {
10024 Perl_dump_indent(aTHX_ level, file,
10025 "%s[%" UVuf "] 0x%04" UVXf " .. INFINITY\n",
10026 indent, (UV)count, start);
10028 else if (end != start) {
10029 Perl_dump_indent(aTHX_ level, file,
10030 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10031 indent, (UV)count, start, end);
10034 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10035 indent, (UV)count, start);
10042 Perl__load_PL_utf8_foldclosures (pTHX)
10044 assert(! PL_utf8_foldclosures);
10046 /* If the folds haven't been read in, call a fold function
10048 if (! PL_utf8_tofold) {
10049 U8 dummy[UTF8_MAXBYTES_CASE+1];
10050 const U8 hyphen[] = HYPHEN_UTF8;
10052 /* This string is just a short named one above \xff */
10053 toFOLD_utf8_safe(hyphen, hyphen + sizeof(hyphen) - 1, dummy, NULL);
10054 assert(PL_utf8_tofold); /* Verify that worked */
10056 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
10060 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10062 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10064 /* Return a boolean as to if the two passed in inversion lists are
10065 * identical. The final argument, if TRUE, says to take the complement of
10066 * the second inversion list before doing the comparison */
10068 const UV* array_a = invlist_array(a);
10069 const UV* array_b = invlist_array(b);
10070 UV len_a = _invlist_len(a);
10071 UV len_b = _invlist_len(b);
10073 PERL_ARGS_ASSERT__INVLISTEQ;
10075 /* If are to compare 'a' with the complement of b, set it
10076 * up so are looking at b's complement. */
10077 if (complement_b) {
10079 /* The complement of nothing is everything, so <a> would have to have
10080 * just one element, starting at zero (ending at infinity) */
10082 return (len_a == 1 && array_a[0] == 0);
10084 else if (array_b[0] == 0) {
10086 /* Otherwise, to complement, we invert. Here, the first element is
10087 * 0, just remove it. To do this, we just pretend the array starts
10095 /* But if the first element is not zero, we pretend the list starts
10096 * at the 0 that is always stored immediately before the array. */
10102 return len_a == len_b
10103 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10109 * As best we can, determine the characters that can match the start of
10110 * the given EXACTF-ish node.
10112 * Returns the invlist as a new SV*; it is the caller's responsibility to
10113 * call SvREFCNT_dec() when done with it.
10116 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10118 const U8 * s = (U8*)STRING(node);
10119 SSize_t bytelen = STR_LEN(node);
10121 /* Start out big enough for 2 separate code points */
10122 SV* invlist = _new_invlist(4);
10124 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
10129 /* We punt and assume can match anything if the node begins
10130 * with a multi-character fold. Things are complicated. For
10131 * example, /ffi/i could match any of:
10132 * "\N{LATIN SMALL LIGATURE FFI}"
10133 * "\N{LATIN SMALL LIGATURE FF}I"
10134 * "F\N{LATIN SMALL LIGATURE FI}"
10135 * plus several other things; and making sure we have all the
10136 * possibilities is hard. */
10137 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10138 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10141 /* Any Latin1 range character can potentially match any
10142 * other depending on the locale */
10143 if (OP(node) == EXACTFL) {
10144 _invlist_union(invlist, PL_Latin1, &invlist);
10147 /* But otherwise, it matches at least itself. We can
10148 * quickly tell if it has a distinct fold, and if so,
10149 * it matches that as well */
10150 invlist = add_cp_to_invlist(invlist, uc);
10151 if (IS_IN_SOME_FOLD_L1(uc))
10152 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10155 /* Some characters match above-Latin1 ones under /i. This
10156 * is true of EXACTFL ones when the locale is UTF-8 */
10157 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10158 && (! isASCII(uc) || (OP(node) != EXACTFA
10159 && OP(node) != EXACTFA_NO_TRIE)))
10161 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10165 else { /* Pattern is UTF-8 */
10166 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10167 STRLEN foldlen = UTF8SKIP(s);
10168 const U8* e = s + bytelen;
10171 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10173 /* The only code points that aren't folded in a UTF EXACTFish
10174 * node are are the problematic ones in EXACTFL nodes */
10175 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10176 /* We need to check for the possibility that this EXACTFL
10177 * node begins with a multi-char fold. Therefore we fold
10178 * the first few characters of it so that we can make that
10183 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10185 *(d++) = (U8) toFOLD(*s);
10190 toFOLD_utf8_safe(s, e, d, &len);
10196 /* And set up so the code below that looks in this folded
10197 * buffer instead of the node's string */
10199 foldlen = UTF8SKIP(folded);
10203 /* When we reach here 's' points to the fold of the first
10204 * character(s) of the node; and 'e' points to far enough along
10205 * the folded string to be just past any possible multi-char
10206 * fold. 'foldlen' is the length in bytes of the first
10209 * Unlike the non-UTF-8 case, the macro for determining if a
10210 * string is a multi-char fold requires all the characters to
10211 * already be folded. This is because of all the complications
10212 * if not. Note that they are folded anyway, except in EXACTFL
10213 * nodes. Like the non-UTF case above, we punt if the node
10214 * begins with a multi-char fold */
10216 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10217 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10219 else { /* Single char fold */
10221 /* It matches all the things that fold to it, which are
10222 * found in PL_utf8_foldclosures (including itself) */
10223 invlist = add_cp_to_invlist(invlist, uc);
10224 if (! PL_utf8_foldclosures)
10225 _load_PL_utf8_foldclosures();
10226 if ((listp = hv_fetch(PL_utf8_foldclosures,
10227 (char *) s, foldlen, FALSE)))
10229 AV* list = (AV*) *listp;
10231 for (k = 0; k <= av_tindex_nomg(list); k++) {
10232 SV** c_p = av_fetch(list, k, FALSE);
10238 /* /aa doesn't allow folds between ASCII and non- */
10239 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
10240 && isASCII(c) != isASCII(uc))
10245 invlist = add_cp_to_invlist(invlist, c);
10254 #undef HEADER_LENGTH
10255 #undef TO_INTERNAL_SIZE
10256 #undef FROM_INTERNAL_SIZE
10257 #undef INVLIST_VERSION_ID
10259 /* End of inversion list object */
10262 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10264 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10265 * constructs, and updates RExC_flags with them. On input, RExC_parse
10266 * should point to the first flag; it is updated on output to point to the
10267 * final ')' or ':'. There needs to be at least one flag, or this will
10270 /* for (?g), (?gc), and (?o) warnings; warning
10271 about (?c) will warn about (?g) -- japhy */
10273 #define WASTED_O 0x01
10274 #define WASTED_G 0x02
10275 #define WASTED_C 0x04
10276 #define WASTED_GC (WASTED_G|WASTED_C)
10277 I32 wastedflags = 0x00;
10278 U32 posflags = 0, negflags = 0;
10279 U32 *flagsp = &posflags;
10280 char has_charset_modifier = '\0';
10282 bool has_use_defaults = FALSE;
10283 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10284 int x_mod_count = 0;
10286 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10288 /* '^' as an initial flag sets certain defaults */
10289 if (UCHARAT(RExC_parse) == '^') {
10291 has_use_defaults = TRUE;
10292 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10293 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
10294 ? REGEX_UNICODE_CHARSET
10295 : REGEX_DEPENDS_CHARSET);
10298 cs = get_regex_charset(RExC_flags);
10299 if (cs == REGEX_DEPENDS_CHARSET
10300 && (RExC_utf8 || RExC_uni_semantics))
10302 cs = REGEX_UNICODE_CHARSET;
10305 while (RExC_parse < RExC_end) {
10306 /* && strchr("iogcmsx", *RExC_parse) */
10307 /* (?g), (?gc) and (?o) are useless here
10308 and must be globally applied -- japhy */
10309 switch (*RExC_parse) {
10311 /* Code for the imsxn flags */
10312 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10314 case LOCALE_PAT_MOD:
10315 if (has_charset_modifier) {
10316 goto excess_modifier;
10318 else if (flagsp == &negflags) {
10321 cs = REGEX_LOCALE_CHARSET;
10322 has_charset_modifier = LOCALE_PAT_MOD;
10324 case UNICODE_PAT_MOD:
10325 if (has_charset_modifier) {
10326 goto excess_modifier;
10328 else if (flagsp == &negflags) {
10331 cs = REGEX_UNICODE_CHARSET;
10332 has_charset_modifier = UNICODE_PAT_MOD;
10334 case ASCII_RESTRICT_PAT_MOD:
10335 if (flagsp == &negflags) {
10338 if (has_charset_modifier) {
10339 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10340 goto excess_modifier;
10342 /* Doubled modifier implies more restricted */
10343 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10346 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10348 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10350 case DEPENDS_PAT_MOD:
10351 if (has_use_defaults) {
10352 goto fail_modifiers;
10354 else if (flagsp == &negflags) {
10357 else if (has_charset_modifier) {
10358 goto excess_modifier;
10361 /* The dual charset means unicode semantics if the
10362 * pattern (or target, not known until runtime) are
10363 * utf8, or something in the pattern indicates unicode
10365 cs = (RExC_utf8 || RExC_uni_semantics)
10366 ? REGEX_UNICODE_CHARSET
10367 : REGEX_DEPENDS_CHARSET;
10368 has_charset_modifier = DEPENDS_PAT_MOD;
10372 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10373 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10375 else if (has_charset_modifier == *(RExC_parse - 1)) {
10376 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10377 *(RExC_parse - 1));
10380 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10382 NOT_REACHED; /*NOTREACHED*/
10385 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10386 *(RExC_parse - 1));
10387 NOT_REACHED; /*NOTREACHED*/
10388 case ONCE_PAT_MOD: /* 'o' */
10389 case GLOBAL_PAT_MOD: /* 'g' */
10390 if (PASS2 && ckWARN(WARN_REGEXP)) {
10391 const I32 wflagbit = *RExC_parse == 'o'
10394 if (! (wastedflags & wflagbit) ) {
10395 wastedflags |= wflagbit;
10396 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10399 "Useless (%s%c) - %suse /%c modifier",
10400 flagsp == &negflags ? "?-" : "?",
10402 flagsp == &negflags ? "don't " : "",
10409 case CONTINUE_PAT_MOD: /* 'c' */
10410 if (PASS2 && ckWARN(WARN_REGEXP)) {
10411 if (! (wastedflags & WASTED_C) ) {
10412 wastedflags |= WASTED_GC;
10413 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10416 "Useless (%sc) - %suse /gc modifier",
10417 flagsp == &negflags ? "?-" : "?",
10418 flagsp == &negflags ? "don't " : ""
10423 case KEEPCOPY_PAT_MOD: /* 'p' */
10424 if (flagsp == &negflags) {
10426 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10428 *flagsp |= RXf_PMf_KEEPCOPY;
10432 /* A flag is a default iff it is following a minus, so
10433 * if there is a minus, it means will be trying to
10434 * re-specify a default which is an error */
10435 if (has_use_defaults || flagsp == &negflags) {
10436 goto fail_modifiers;
10438 flagsp = &negflags;
10439 wastedflags = 0; /* reset so (?g-c) warns twice */
10445 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
10446 negflags |= RXf_PMf_EXTENDED_MORE;
10448 RExC_flags |= posflags;
10450 if (negflags & RXf_PMf_EXTENDED) {
10451 negflags |= RXf_PMf_EXTENDED_MORE;
10453 RExC_flags &= ~negflags;
10454 set_regex_charset(&RExC_flags, cs);
10459 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10460 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10461 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
10462 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10463 NOT_REACHED; /*NOTREACHED*/
10466 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10469 vFAIL("Sequence (?... not terminated");
10473 - reg - regular expression, i.e. main body or parenthesized thing
10475 * Caller must absorb opening parenthesis.
10477 * Combining parenthesis handling with the base level of regular expression
10478 * is a trifle forced, but the need to tie the tails of the branches to what
10479 * follows makes it hard to avoid.
10481 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10483 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10485 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10488 PERL_STATIC_INLINE regnode *
10489 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10491 char * parse_start,
10496 char* name_start = RExC_parse;
10498 SV *sv_dat = reg_scan_name(pRExC_state, SIZE_ONLY
10499 ? REG_RSN_RETURN_NULL
10500 : REG_RSN_RETURN_DATA);
10501 GET_RE_DEBUG_FLAGS_DECL;
10503 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10505 if (RExC_parse == name_start || *RExC_parse != ch) {
10506 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10507 vFAIL2("Sequence %.3s... not terminated",parse_start);
10511 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10512 RExC_rxi->data->data[num]=(void*)sv_dat;
10513 SvREFCNT_inc_simple_void(sv_dat);
10516 ret = reganode(pRExC_state,
10519 : (ASCII_FOLD_RESTRICTED)
10521 : (AT_LEAST_UNI_SEMANTICS)
10527 *flagp |= HASWIDTH;
10529 Set_Node_Offset(ret, parse_start+1);
10530 Set_Node_Cur_Length(ret, parse_start);
10532 nextchar(pRExC_state);
10536 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
10537 flags. Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan
10538 needs to be restarted, or'd with NEED_UTF8 if the pattern needs to be
10539 upgraded to UTF-8. Otherwise would only return NULL if regbranch() returns
10540 NULL, which cannot happen. */
10542 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
10543 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10544 * 2 is like 1, but indicates that nextchar() has been called to advance
10545 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10546 * this flag alerts us to the need to check for that */
10548 regnode *ret; /* Will be the head of the group. */
10551 regnode *ender = NULL;
10554 U32 oregflags = RExC_flags;
10555 bool have_branch = 0;
10557 I32 freeze_paren = 0;
10558 I32 after_freeze = 0;
10559 I32 num; /* numeric backreferences */
10561 char * parse_start = RExC_parse; /* MJD */
10562 char * const oregcomp_parse = RExC_parse;
10564 GET_RE_DEBUG_FLAGS_DECL;
10566 PERL_ARGS_ASSERT_REG;
10567 DEBUG_PARSE("reg ");
10569 *flagp = 0; /* Tentatively. */
10571 /* Having this true makes it feasible to have a lot fewer tests for the
10572 * parse pointer being in scope. For example, we can write
10573 * while(isFOO(*RExC_parse)) RExC_parse++;
10575 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
10577 assert(*RExC_end == '\0');
10579 /* Make an OPEN node, if parenthesized. */
10582 /* Under /x, space and comments can be gobbled up between the '(' and
10583 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
10584 * intervening space, as the sequence is a token, and a token should be
10586 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
10588 if (RExC_parse >= RExC_end) {
10589 vFAIL("Unmatched (");
10592 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
10593 char *start_verb = RExC_parse + 1;
10595 char *start_arg = NULL;
10596 unsigned char op = 0;
10597 int arg_required = 0;
10598 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
10600 if (has_intervening_patws) {
10601 RExC_parse++; /* past the '*' */
10602 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
10604 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
10605 if ( *RExC_parse == ':' ) {
10606 start_arg = RExC_parse + 1;
10609 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10611 verb_len = RExC_parse - start_verb;
10613 if (RExC_parse >= RExC_end) {
10614 goto unterminated_verb_pattern;
10616 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10617 while ( RExC_parse < RExC_end && *RExC_parse != ')' )
10618 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10619 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10620 unterminated_verb_pattern:
10621 vFAIL("Unterminated verb pattern argument");
10622 if ( RExC_parse == start_arg )
10625 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10626 vFAIL("Unterminated verb pattern");
10629 /* Here, we know that RExC_parse < RExC_end */
10631 switch ( *start_verb ) {
10632 case 'A': /* (*ACCEPT) */
10633 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
10635 internal_argval = RExC_nestroot;
10638 case 'C': /* (*COMMIT) */
10639 if ( memEQs(start_verb,verb_len,"COMMIT") )
10642 case 'F': /* (*FAIL) */
10643 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
10647 case ':': /* (*:NAME) */
10648 case 'M': /* (*MARK:NAME) */
10649 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
10654 case 'P': /* (*PRUNE) */
10655 if ( memEQs(start_verb,verb_len,"PRUNE") )
10658 case 'S': /* (*SKIP) */
10659 if ( memEQs(start_verb,verb_len,"SKIP") )
10662 case 'T': /* (*THEN) */
10663 /* [19:06] <TimToady> :: is then */
10664 if ( memEQs(start_verb,verb_len,"THEN") ) {
10666 RExC_seen |= REG_CUTGROUP_SEEN;
10671 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10673 "Unknown verb pattern '%" UTF8f "'",
10674 UTF8fARG(UTF, verb_len, start_verb));
10676 if ( arg_required && !start_arg ) {
10677 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
10678 verb_len, start_verb);
10680 if (internal_argval == -1) {
10681 ret = reganode(pRExC_state, op, 0);
10683 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
10685 RExC_seen |= REG_VERBARG_SEEN;
10686 if ( ! SIZE_ONLY ) {
10688 SV *sv = newSVpvn( start_arg,
10689 RExC_parse - start_arg);
10690 ARG(ret) = add_data( pRExC_state,
10691 STR_WITH_LEN("S"));
10692 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
10697 if ( internal_argval != -1 )
10698 ARG2L_SET(ret, internal_argval);
10700 nextchar(pRExC_state);
10703 else if (*RExC_parse == '?') { /* (?...) */
10704 bool is_logical = 0;
10705 const char * const seqstart = RExC_parse;
10706 const char * endptr;
10707 if (has_intervening_patws) {
10709 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
10712 RExC_parse++; /* past the '?' */
10713 paren = *RExC_parse; /* might be a trailing NUL, if not
10715 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10716 if (RExC_parse > RExC_end) {
10719 ret = NULL; /* For look-ahead/behind. */
10722 case 'P': /* (?P...) variants for those used to PCRE/Python */
10723 paren = *RExC_parse;
10724 if ( paren == '<') { /* (?P<...>) named capture */
10726 if (RExC_parse >= RExC_end) {
10727 vFAIL("Sequence (?P<... not terminated");
10729 goto named_capture;
10731 else if (paren == '>') { /* (?P>name) named recursion */
10733 if (RExC_parse >= RExC_end) {
10734 vFAIL("Sequence (?P>... not terminated");
10736 goto named_recursion;
10738 else if (paren == '=') { /* (?P=...) named backref */
10740 return handle_named_backref(pRExC_state, flagp,
10743 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10744 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10745 vFAIL3("Sequence (%.*s...) not recognized",
10746 RExC_parse-seqstart, seqstart);
10747 NOT_REACHED; /*NOTREACHED*/
10748 case '<': /* (?<...) */
10749 if (*RExC_parse == '!')
10751 else if (*RExC_parse != '=')
10758 case '\'': /* (?'...') */
10759 name_start = RExC_parse;
10760 svname = reg_scan_name(pRExC_state,
10761 SIZE_ONLY /* reverse test from the others */
10762 ? REG_RSN_RETURN_NAME
10763 : REG_RSN_RETURN_NULL);
10764 if ( RExC_parse == name_start
10765 || RExC_parse >= RExC_end
10766 || *RExC_parse != paren)
10768 vFAIL2("Sequence (?%c... not terminated",
10769 paren=='>' ? '<' : paren);
10774 if (!svname) /* shouldn't happen */
10776 "panic: reg_scan_name returned NULL");
10777 if (!RExC_paren_names) {
10778 RExC_paren_names= newHV();
10779 sv_2mortal(MUTABLE_SV(RExC_paren_names));
10781 RExC_paren_name_list= newAV();
10782 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10785 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10787 sv_dat = HeVAL(he_str);
10789 /* croak baby croak */
10791 "panic: paren_name hash element allocation failed");
10792 } else if ( SvPOK(sv_dat) ) {
10793 /* (?|...) can mean we have dupes so scan to check
10794 its already been stored. Maybe a flag indicating
10795 we are inside such a construct would be useful,
10796 but the arrays are likely to be quite small, so
10797 for now we punt -- dmq */
10798 IV count = SvIV(sv_dat);
10799 I32 *pv = (I32*)SvPVX(sv_dat);
10801 for ( i = 0 ; i < count ; i++ ) {
10802 if ( pv[i] == RExC_npar ) {
10808 pv = (I32*)SvGROW(sv_dat,
10809 SvCUR(sv_dat) + sizeof(I32)+1);
10810 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10811 pv[count] = RExC_npar;
10812 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10815 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10816 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10819 SvIV_set(sv_dat, 1);
10822 /* Yes this does cause a memory leak in debugging Perls
10824 if (!av_store(RExC_paren_name_list,
10825 RExC_npar, SvREFCNT_inc(svname)))
10826 SvREFCNT_dec_NN(svname);
10829 /*sv_dump(sv_dat);*/
10831 nextchar(pRExC_state);
10833 goto capturing_parens;
10835 RExC_seen |= REG_LOOKBEHIND_SEEN;
10836 RExC_in_lookbehind++;
10838 if (RExC_parse >= RExC_end) {
10839 vFAIL("Sequence (?... not terminated");
10843 case '=': /* (?=...) */
10844 RExC_seen_zerolen++;
10846 case '!': /* (?!...) */
10847 RExC_seen_zerolen++;
10848 /* check if we're really just a "FAIL" assertion */
10849 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
10850 FALSE /* Don't force to /x */ );
10851 if (*RExC_parse == ')') {
10852 ret=reganode(pRExC_state, OPFAIL, 0);
10853 nextchar(pRExC_state);
10857 case '|': /* (?|...) */
10858 /* branch reset, behave like a (?:...) except that
10859 buffers in alternations share the same numbers */
10861 after_freeze = freeze_paren = RExC_npar;
10863 case ':': /* (?:...) */
10864 case '>': /* (?>...) */
10866 case '$': /* (?$...) */
10867 case '@': /* (?@...) */
10868 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10870 case '0' : /* (?0) */
10871 case 'R' : /* (?R) */
10872 if (RExC_parse == RExC_end || *RExC_parse != ')')
10873 FAIL("Sequence (?R) not terminated");
10875 RExC_seen |= REG_RECURSE_SEEN;
10876 *flagp |= POSTPONED;
10877 goto gen_recurse_regop;
10879 /* named and numeric backreferences */
10880 case '&': /* (?&NAME) */
10881 parse_start = RExC_parse - 1;
10884 SV *sv_dat = reg_scan_name(pRExC_state,
10885 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10886 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10888 if (RExC_parse >= RExC_end || *RExC_parse != ')')
10889 vFAIL("Sequence (?&... not terminated");
10890 goto gen_recurse_regop;
10893 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10895 vFAIL("Illegal pattern");
10897 goto parse_recursion;
10899 case '-': /* (?-1) */
10900 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10901 RExC_parse--; /* rewind to let it be handled later */
10905 case '1': case '2': case '3': case '4': /* (?1) */
10906 case '5': case '6': case '7': case '8': case '9':
10907 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
10910 bool is_neg = FALSE;
10912 parse_start = RExC_parse - 1; /* MJD */
10913 if (*RExC_parse == '-') {
10917 if (grok_atoUV(RExC_parse, &unum, &endptr)
10921 RExC_parse = (char*)endptr;
10925 /* Some limit for num? */
10929 if (*RExC_parse!=')')
10930 vFAIL("Expecting close bracket");
10933 if ( paren == '-' ) {
10935 Diagram of capture buffer numbering.
10936 Top line is the normal capture buffer numbers
10937 Bottom line is the negative indexing as from
10941 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
10945 num = RExC_npar + num;
10948 vFAIL("Reference to nonexistent group");
10950 } else if ( paren == '+' ) {
10951 num = RExC_npar + num - 1;
10953 /* We keep track how many GOSUB items we have produced.
10954 To start off the ARG2L() of the GOSUB holds its "id",
10955 which is used later in conjunction with RExC_recurse
10956 to calculate the offset we need to jump for the GOSUB,
10957 which it will store in the final representation.
10958 We have to defer the actual calculation until much later
10959 as the regop may move.
10962 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
10964 if (num > (I32)RExC_rx->nparens) {
10966 vFAIL("Reference to nonexistent group");
10968 RExC_recurse_count++;
10969 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
10970 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
10971 22, "| |", (int)(depth * 2 + 1), "",
10972 (UV)ARG(ret), (IV)ARG2L(ret)));
10974 RExC_seen |= REG_RECURSE_SEEN;
10976 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
10977 Set_Node_Offset(ret, parse_start); /* MJD */
10979 *flagp |= POSTPONED;
10980 assert(*RExC_parse == ')');
10981 nextchar(pRExC_state);
10986 case '?': /* (??...) */
10988 if (*RExC_parse != '{') {
10989 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10990 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10992 "Sequence (%" UTF8f "...) not recognized",
10993 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10994 NOT_REACHED; /*NOTREACHED*/
10996 *flagp |= POSTPONED;
11000 case '{': /* (?{...}) */
11003 struct reg_code_block *cb;
11005 RExC_seen_zerolen++;
11007 if ( !pRExC_state->num_code_blocks
11008 || pRExC_state->code_index >= pRExC_state->num_code_blocks
11009 || pRExC_state->code_blocks[pRExC_state->code_index].start
11010 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11013 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11014 FAIL("panic: Sequence (?{...}): no code block found\n");
11015 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11017 /* this is a pre-compiled code block (?{...}) */
11018 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
11019 RExC_parse = RExC_start + cb->end;
11022 if (cb->src_regex) {
11023 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11024 RExC_rxi->data->data[n] =
11025 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11026 RExC_rxi->data->data[n+1] = (void*)o;
11029 n = add_data(pRExC_state,
11030 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11031 RExC_rxi->data->data[n] = (void*)o;
11034 pRExC_state->code_index++;
11035 nextchar(pRExC_state);
11039 ret = reg_node(pRExC_state, LOGICAL);
11041 eval = reg2Lanode(pRExC_state, EVAL,
11044 /* for later propagation into (??{})
11046 RExC_flags & RXf_PMf_COMPILETIME
11051 REGTAIL(pRExC_state, ret, eval);
11052 /* deal with the length of this later - MJD */
11055 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11056 Set_Node_Length(ret, RExC_parse - parse_start + 1);
11057 Set_Node_Offset(ret, parse_start);
11060 case '(': /* (?(?{...})...) and (?(?=...)...) */
11063 const int DEFINE_len = sizeof("DEFINE") - 1;
11064 if (RExC_parse[0] == '?') { /* (?(?...)) */
11065 if ( RExC_parse < RExC_end - 1
11066 && ( RExC_parse[1] == '='
11067 || RExC_parse[1] == '!'
11068 || RExC_parse[1] == '<'
11069 || RExC_parse[1] == '{')
11070 ) { /* Lookahead or eval. */
11074 ret = reg_node(pRExC_state, LOGICAL);
11078 tail = reg(pRExC_state, 1, &flag, depth+1);
11079 if (flag & (RESTART_PASS1|NEED_UTF8)) {
11080 *flagp = flag & (RESTART_PASS1|NEED_UTF8);
11083 REGTAIL(pRExC_state, ret, tail);
11086 /* Fall through to ‘Unknown switch condition’ at the
11087 end of the if/else chain. */
11089 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11090 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11092 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11093 char *name_start= RExC_parse++;
11095 SV *sv_dat=reg_scan_name(pRExC_state,
11096 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11097 if ( RExC_parse == name_start
11098 || RExC_parse >= RExC_end
11099 || *RExC_parse != ch)
11101 vFAIL2("Sequence (?(%c... not terminated",
11102 (ch == '>' ? '<' : ch));
11106 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11107 RExC_rxi->data->data[num]=(void*)sv_dat;
11108 SvREFCNT_inc_simple_void(sv_dat);
11110 ret = reganode(pRExC_state,NGROUPP,num);
11111 goto insert_if_check_paren;
11113 else if (RExC_end - RExC_parse >= DEFINE_len
11114 && strnEQ(RExC_parse, "DEFINE", DEFINE_len))
11116 ret = reganode(pRExC_state,DEFINEP,0);
11117 RExC_parse += DEFINE_len;
11119 goto insert_if_check_paren;
11121 else if (RExC_parse[0] == 'R') {
11123 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11124 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11125 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11128 if (RExC_parse[0] == '0') {
11132 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11134 if (grok_atoUV(RExC_parse, &uv, &endptr)
11137 parno = (I32)uv + 1;
11138 RExC_parse = (char*)endptr;
11140 /* else "Switch condition not recognized" below */
11141 } else if (RExC_parse[0] == '&') {
11144 sv_dat = reg_scan_name(pRExC_state,
11146 ? REG_RSN_RETURN_NULL
11147 : REG_RSN_RETURN_DATA);
11149 /* we should only have a false sv_dat when
11150 * SIZE_ONLY is true, and we always have false
11151 * sv_dat when SIZE_ONLY is true.
11152 * reg_scan_name() will VFAIL() if the name is
11153 * unknown when SIZE_ONLY is false, and otherwise
11154 * will return something, and when SIZE_ONLY is
11155 * true, reg_scan_name() just parses the string,
11156 * and doesnt return anything. (in theory) */
11157 assert(SIZE_ONLY ? !sv_dat : !!sv_dat);
11160 parno = 1 + *((I32 *)SvPVX(sv_dat));
11162 ret = reganode(pRExC_state,INSUBP,parno);
11163 goto insert_if_check_paren;
11165 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11169 if (grok_atoUV(RExC_parse, &uv, &endptr)
11173 RExC_parse = (char*)endptr;
11176 vFAIL("panic: grok_atoUV returned FALSE");
11178 ret = reganode(pRExC_state, GROUPP, parno);
11180 insert_if_check_paren:
11181 if (UCHARAT(RExC_parse) != ')') {
11182 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11183 vFAIL("Switch condition not recognized");
11185 nextchar(pRExC_state);
11187 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
11188 br = regbranch(pRExC_state, &flags, 1,depth+1);
11190 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11191 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11194 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11197 REGTAIL(pRExC_state, br, reganode(pRExC_state,
11199 c = UCHARAT(RExC_parse);
11200 nextchar(pRExC_state);
11201 if (flags&HASWIDTH)
11202 *flagp |= HASWIDTH;
11205 vFAIL("(?(DEFINE)....) does not allow branches");
11207 /* Fake one for optimizer. */
11208 lastbr = reganode(pRExC_state, IFTHEN, 0);
11210 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
11211 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11212 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11215 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11218 REGTAIL(pRExC_state, ret, lastbr);
11219 if (flags&HASWIDTH)
11220 *flagp |= HASWIDTH;
11221 c = UCHARAT(RExC_parse);
11222 nextchar(pRExC_state);
11227 if (RExC_parse >= RExC_end)
11228 vFAIL("Switch (?(condition)... not terminated");
11230 vFAIL("Switch (?(condition)... contains too many branches");
11232 ender = reg_node(pRExC_state, TAIL);
11233 REGTAIL(pRExC_state, br, ender);
11235 REGTAIL(pRExC_state, lastbr, ender);
11236 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11239 REGTAIL(pRExC_state, ret, ender);
11240 RExC_size++; /* XXX WHY do we need this?!!
11241 For large programs it seems to be required
11242 but I can't figure out why. -- dmq*/
11245 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11246 vFAIL("Unknown switch condition (?(...))");
11248 case '[': /* (?[ ... ]) */
11249 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
11251 case 0: /* A NUL */
11252 RExC_parse--; /* for vFAIL to print correctly */
11253 vFAIL("Sequence (? incomplete");
11255 default: /* e.g., (?i) */
11256 RExC_parse = (char *) seqstart + 1;
11258 parse_lparen_question_flags(pRExC_state);
11259 if (UCHARAT(RExC_parse) != ':') {
11260 if (RExC_parse < RExC_end)
11261 nextchar(pRExC_state);
11266 nextchar(pRExC_state);
11271 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11276 ret = reganode(pRExC_state, OPEN, parno);
11278 if (!RExC_nestroot)
11279 RExC_nestroot = parno;
11280 if (RExC_open_parens && !RExC_open_parens[parno])
11282 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11283 "%*s%*s Setting open paren #%" IVdf " to %d\n",
11284 22, "| |", (int)(depth * 2 + 1), "",
11285 (IV)parno, REG_NODE_NUM(ret)));
11286 RExC_open_parens[parno]= ret;
11289 Set_Node_Length(ret, 1); /* MJD */
11290 Set_Node_Offset(ret, RExC_parse); /* MJD */
11293 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
11302 /* Pick up the branches, linking them together. */
11303 parse_start = RExC_parse; /* MJD */
11304 br = regbranch(pRExC_state, &flags, 1,depth+1);
11306 /* branch_len = (paren != 0); */
11309 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11310 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11313 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11315 if (*RExC_parse == '|') {
11316 if (!SIZE_ONLY && RExC_extralen) {
11317 reginsert(pRExC_state, BRANCHJ, br, depth+1);
11320 reginsert(pRExC_state, BRANCH, br, depth+1);
11321 Set_Node_Length(br, paren != 0);
11322 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
11326 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
11328 else if (paren == ':') {
11329 *flagp |= flags&SIMPLE;
11331 if (is_open) { /* Starts with OPEN. */
11332 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
11334 else if (paren != '?') /* Not Conditional */
11336 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11338 while (*RExC_parse == '|') {
11339 if (!SIZE_ONLY && RExC_extralen) {
11340 ender = reganode(pRExC_state, LONGJMP,0);
11342 /* Append to the previous. */
11343 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11346 RExC_extralen += 2; /* Account for LONGJMP. */
11347 nextchar(pRExC_state);
11348 if (freeze_paren) {
11349 if (RExC_npar > after_freeze)
11350 after_freeze = RExC_npar;
11351 RExC_npar = freeze_paren;
11353 br = regbranch(pRExC_state, &flags, 0, depth+1);
11356 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11357 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11360 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11362 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
11364 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11367 if (have_branch || paren != ':') {
11368 /* Make a closing node, and hook it on the end. */
11371 ender = reg_node(pRExC_state, TAIL);
11374 ender = reganode(pRExC_state, CLOSE, parno);
11375 if ( RExC_close_parens ) {
11376 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11377 "%*s%*s Setting close paren #%" IVdf " to %d\n",
11378 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
11379 RExC_close_parens[parno]= ender;
11380 if (RExC_nestroot == parno)
11383 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
11384 Set_Node_Length(ender,1); /* MJD */
11390 *flagp &= ~HASWIDTH;
11393 ender = reg_node(pRExC_state, SUCCEED);
11396 ender = reg_node(pRExC_state, END);
11398 assert(!RExC_end_op); /* there can only be one! */
11399 RExC_end_op = ender;
11400 if (RExC_close_parens) {
11401 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11402 "%*s%*s Setting close paren #0 (END) to %d\n",
11403 22, "| |", (int)(depth * 2 + 1), "", REG_NODE_NUM(ender)));
11405 RExC_close_parens[0]= ender;
11410 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11411 DEBUG_PARSE_MSG("lsbr");
11412 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
11413 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11414 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11415 SvPV_nolen_const(RExC_mysv1),
11416 (IV)REG_NODE_NUM(lastbr),
11417 SvPV_nolen_const(RExC_mysv2),
11418 (IV)REG_NODE_NUM(ender),
11419 (IV)(ender - lastbr)
11422 REGTAIL(pRExC_state, lastbr, ender);
11424 if (have_branch && !SIZE_ONLY) {
11425 char is_nothing= 1;
11427 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
11429 /* Hook the tails of the branches to the closing node. */
11430 for (br = ret; br; br = regnext(br)) {
11431 const U8 op = PL_regkind[OP(br)];
11432 if (op == BRANCH) {
11433 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
11434 if ( OP(NEXTOPER(br)) != NOTHING
11435 || regnext(NEXTOPER(br)) != ender)
11438 else if (op == BRANCHJ) {
11439 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
11440 /* for now we always disable this optimisation * /
11441 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
11442 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
11448 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
11449 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11450 DEBUG_PARSE_MSG("NADA");
11451 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
11452 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11453 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11454 SvPV_nolen_const(RExC_mysv1),
11455 (IV)REG_NODE_NUM(ret),
11456 SvPV_nolen_const(RExC_mysv2),
11457 (IV)REG_NODE_NUM(ender),
11462 if (OP(ender) == TAIL) {
11467 for ( opt= br + 1; opt < ender ; opt++ )
11468 OP(opt)= OPTIMIZED;
11469 NEXT_OFF(br)= ender - br;
11477 static const char parens[] = "=!<,>";
11479 if (paren && (p = strchr(parens, paren))) {
11480 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
11481 int flag = (p - parens) > 1;
11484 node = SUSPEND, flag = 0;
11485 reginsert(pRExC_state, node,ret, depth+1);
11486 Set_Node_Cur_Length(ret, parse_start);
11487 Set_Node_Offset(ret, parse_start + 1);
11489 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
11493 /* Check for proper termination. */
11495 /* restore original flags, but keep (?p) and, if we've changed from /d
11496 * rules to /u, keep the /u */
11497 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
11498 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
11499 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
11501 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
11502 RExC_parse = oregcomp_parse;
11503 vFAIL("Unmatched (");
11505 nextchar(pRExC_state);
11507 else if (!paren && RExC_parse < RExC_end) {
11508 if (*RExC_parse == ')') {
11510 vFAIL("Unmatched )");
11513 FAIL("Junk on end of regexp"); /* "Can't happen". */
11514 NOT_REACHED; /* NOTREACHED */
11517 if (RExC_in_lookbehind) {
11518 RExC_in_lookbehind--;
11520 if (after_freeze > RExC_npar)
11521 RExC_npar = after_freeze;
11526 - regbranch - one alternative of an | operator
11528 * Implements the concatenation operator.
11530 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11531 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11534 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
11537 regnode *chain = NULL;
11539 I32 flags = 0, c = 0;
11540 GET_RE_DEBUG_FLAGS_DECL;
11542 PERL_ARGS_ASSERT_REGBRANCH;
11544 DEBUG_PARSE("brnc");
11549 if (!SIZE_ONLY && RExC_extralen)
11550 ret = reganode(pRExC_state, BRANCHJ,0);
11552 ret = reg_node(pRExC_state, BRANCH);
11553 Set_Node_Length(ret, 1);
11557 if (!first && SIZE_ONLY)
11558 RExC_extralen += 1; /* BRANCHJ */
11560 *flagp = WORST; /* Tentatively. */
11562 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11563 FALSE /* Don't force to /x */ );
11564 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
11565 flags &= ~TRYAGAIN;
11566 latest = regpiece(pRExC_state, &flags,depth+1);
11567 if (latest == NULL) {
11568 if (flags & TRYAGAIN)
11570 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11571 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11574 FAIL2("panic: regpiece returned NULL, flags=%#" UVxf, (UV) flags);
11576 else if (ret == NULL)
11578 *flagp |= flags&(HASWIDTH|POSTPONED);
11579 if (chain == NULL) /* First piece. */
11580 *flagp |= flags&SPSTART;
11582 /* FIXME adding one for every branch after the first is probably
11583 * excessive now we have TRIE support. (hv) */
11585 REGTAIL(pRExC_state, chain, latest);
11590 if (chain == NULL) { /* Loop ran zero times. */
11591 chain = reg_node(pRExC_state, NOTHING);
11596 *flagp |= flags&SIMPLE;
11603 - regpiece - something followed by possible quantifier * + ? {n,m}
11605 * Note that the branching code sequences used for ? and the general cases
11606 * of * and + are somewhat optimized: they use the same NOTHING node as
11607 * both the endmarker for their branch list and the body of the last branch.
11608 * It might seem that this node could be dispensed with entirely, but the
11609 * endmarker role is not redundant.
11611 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
11613 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11614 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11617 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11623 const char * const origparse = RExC_parse;
11625 I32 max = REG_INFTY;
11626 #ifdef RE_TRACK_PATTERN_OFFSETS
11629 const char *maxpos = NULL;
11632 /* Save the original in case we change the emitted regop to a FAIL. */
11633 regnode * const orig_emit = RExC_emit;
11635 GET_RE_DEBUG_FLAGS_DECL;
11637 PERL_ARGS_ASSERT_REGPIECE;
11639 DEBUG_PARSE("piec");
11641 ret = regatom(pRExC_state, &flags,depth+1);
11643 if (flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8))
11644 *flagp |= flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8);
11646 FAIL2("panic: regatom returned NULL, flags=%#" UVxf, (UV) flags);
11652 if (op == '{' && regcurly(RExC_parse)) {
11654 #ifdef RE_TRACK_PATTERN_OFFSETS
11655 parse_start = RExC_parse; /* MJD */
11657 next = RExC_parse + 1;
11658 while (isDIGIT(*next) || *next == ',') {
11659 if (*next == ',') {
11667 if (*next == '}') { /* got one */
11668 const char* endptr;
11672 if (isDIGIT(*RExC_parse)) {
11673 if (!grok_atoUV(RExC_parse, &uv, &endptr))
11674 vFAIL("Invalid quantifier in {,}");
11675 if (uv >= REG_INFTY)
11676 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11681 if (*maxpos == ',')
11684 maxpos = RExC_parse;
11685 if (isDIGIT(*maxpos)) {
11686 if (!grok_atoUV(maxpos, &uv, &endptr))
11687 vFAIL("Invalid quantifier in {,}");
11688 if (uv >= REG_INFTY)
11689 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11692 max = REG_INFTY; /* meaning "infinity" */
11695 nextchar(pRExC_state);
11696 if (max < min) { /* If can't match, warn and optimize to fail
11700 /* We can't back off the size because we have to reserve
11701 * enough space for all the things we are about to throw
11702 * away, but we can shrink it by the amount we are about
11703 * to re-use here */
11704 RExC_size += PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
11707 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
11708 RExC_emit = orig_emit;
11710 ret = reganode(pRExC_state, OPFAIL, 0);
11713 else if (min == max && *RExC_parse == '?')
11716 ckWARN2reg(RExC_parse + 1,
11717 "Useless use of greediness modifier '%c'",
11723 if ((flags&SIMPLE)) {
11724 if (min == 0 && max == REG_INFTY) {
11725 reginsert(pRExC_state, STAR, ret, depth+1);
11728 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11731 if (min == 1 && max == REG_INFTY) {
11732 reginsert(pRExC_state, PLUS, ret, depth+1);
11735 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11738 MARK_NAUGHTY_EXP(2, 2);
11739 reginsert(pRExC_state, CURLY, ret, depth+1);
11740 Set_Node_Offset(ret, parse_start+1); /* MJD */
11741 Set_Node_Cur_Length(ret, parse_start);
11744 regnode * const w = reg_node(pRExC_state, WHILEM);
11747 REGTAIL(pRExC_state, ret, w);
11748 if (!SIZE_ONLY && RExC_extralen) {
11749 reginsert(pRExC_state, LONGJMP,ret, depth+1);
11750 reginsert(pRExC_state, NOTHING,ret, depth+1);
11751 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
11753 reginsert(pRExC_state, CURLYX,ret, depth+1);
11755 Set_Node_Offset(ret, parse_start+1);
11756 Set_Node_Length(ret,
11757 op == '{' ? (RExC_parse - parse_start) : 1);
11759 if (!SIZE_ONLY && RExC_extralen)
11760 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
11761 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
11763 RExC_whilem_seen++, RExC_extralen += 3;
11764 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
11771 *flagp |= HASWIDTH;
11773 ARG1_SET(ret, (U16)min);
11774 ARG2_SET(ret, (U16)max);
11776 if (max == REG_INFTY)
11777 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11783 if (!ISMULT1(op)) {
11788 #if 0 /* Now runtime fix should be reliable. */
11790 /* if this is reinstated, don't forget to put this back into perldiag:
11792 =item Regexp *+ operand could be empty at {#} in regex m/%s/
11794 (F) The part of the regexp subject to either the * or + quantifier
11795 could match an empty string. The {#} shows in the regular
11796 expression about where the problem was discovered.
11800 if (!(flags&HASWIDTH) && op != '?')
11801 vFAIL("Regexp *+ operand could be empty");
11804 #ifdef RE_TRACK_PATTERN_OFFSETS
11805 parse_start = RExC_parse;
11807 nextchar(pRExC_state);
11809 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
11815 else if (op == '+') {
11819 else if (op == '?') {
11824 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
11825 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
11826 ckWARN2reg(RExC_parse,
11827 "%" UTF8f " matches null string many times",
11828 UTF8fARG(UTF, (RExC_parse >= origparse
11829 ? RExC_parse - origparse
11832 (void)ReREFCNT_inc(RExC_rx_sv);
11835 if (*RExC_parse == '?') {
11836 nextchar(pRExC_state);
11837 reginsert(pRExC_state, MINMOD, ret, depth+1);
11838 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
11840 else if (*RExC_parse == '+') {
11842 nextchar(pRExC_state);
11843 ender = reg_node(pRExC_state, SUCCEED);
11844 REGTAIL(pRExC_state, ret, ender);
11845 reginsert(pRExC_state, SUSPEND, ret, depth+1);
11847 ender = reg_node(pRExC_state, TAIL);
11848 REGTAIL(pRExC_state, ret, ender);
11851 if (ISMULT2(RExC_parse)) {
11853 vFAIL("Nested quantifiers");
11860 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
11869 /* This routine teases apart the various meanings of \N and returns
11870 * accordingly. The input parameters constrain which meaning(s) is/are valid
11871 * in the current context.
11873 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
11875 * If <code_point_p> is not NULL, the context is expecting the result to be a
11876 * single code point. If this \N instance turns out to a single code point,
11877 * the function returns TRUE and sets *code_point_p to that code point.
11879 * If <node_p> is not NULL, the context is expecting the result to be one of
11880 * the things representable by a regnode. If this \N instance turns out to be
11881 * one such, the function generates the regnode, returns TRUE and sets *node_p
11882 * to point to that regnode.
11884 * If this instance of \N isn't legal in any context, this function will
11885 * generate a fatal error and not return.
11887 * On input, RExC_parse should point to the first char following the \N at the
11888 * time of the call. On successful return, RExC_parse will have been updated
11889 * to point to just after the sequence identified by this routine. Also
11890 * *flagp has been updated as needed.
11892 * When there is some problem with the current context and this \N instance,
11893 * the function returns FALSE, without advancing RExC_parse, nor setting
11894 * *node_p, nor *code_point_p, nor *flagp.
11896 * If <cp_count> is not NULL, the caller wants to know the length (in code
11897 * points) that this \N sequence matches. This is set even if the function
11898 * returns FALSE, as detailed below.
11900 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
11902 * Probably the most common case is for the \N to specify a single code point.
11903 * *cp_count will be set to 1, and *code_point_p will be set to that code
11906 * Another possibility is for the input to be an empty \N{}, which for
11907 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
11908 * will be set to a generated NOTHING node.
11910 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
11911 * set to 0. *node_p will be set to a generated REG_ANY node.
11913 * The fourth possibility is that \N resolves to a sequence of more than one
11914 * code points. *cp_count will be set to the number of code points in the
11915 * sequence. *node_p * will be set to a generated node returned by this
11916 * function calling S_reg().
11918 * The final possibility is that it is premature to be calling this function;
11919 * that pass1 needs to be restarted. This can happen when this changes from
11920 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
11921 * latter occurs only when the fourth possibility would otherwise be in
11922 * effect, and is because one of those code points requires the pattern to be
11923 * recompiled as UTF-8. The function returns FALSE, and sets the
11924 * RESTART_PASS1 and NEED_UTF8 flags in *flagp, as appropriate. When this
11925 * happens, the caller needs to desist from continuing parsing, and return
11926 * this information to its caller. This is not set for when there is only one
11927 * code point, as this can be called as part of an ANYOF node, and they can
11928 * store above-Latin1 code points without the pattern having to be in UTF-8.
11930 * For non-single-quoted regexes, the tokenizer has resolved character and
11931 * sequence names inside \N{...} into their Unicode values, normalizing the
11932 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
11933 * hex-represented code points in the sequence. This is done there because
11934 * the names can vary based on what charnames pragma is in scope at the time,
11935 * so we need a way to take a snapshot of what they resolve to at the time of
11936 * the original parse. [perl #56444].
11938 * That parsing is skipped for single-quoted regexes, so we may here get
11939 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
11940 * parser. But if the single-quoted regex is something like '\N{U+41}', that
11941 * is legal and handled here. The code point is Unicode, and has to be
11942 * translated into the native character set for non-ASCII platforms.
11945 char * endbrace; /* points to '}' following the name */
11946 char *endchar; /* Points to '.' or '}' ending cur char in the input
11948 char* p = RExC_parse; /* Temporary */
11950 GET_RE_DEBUG_FLAGS_DECL;
11952 PERL_ARGS_ASSERT_GROK_BSLASH_N;
11954 GET_RE_DEBUG_FLAGS;
11956 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
11957 assert(! (node_p && cp_count)); /* At most 1 should be set */
11959 if (cp_count) { /* Initialize return for the most common case */
11963 /* The [^\n] meaning of \N ignores spaces and comments under the /x
11964 * modifier. The other meanings do not, so use a temporary until we find
11965 * out which we are being called with */
11966 skip_to_be_ignored_text(pRExC_state, &p,
11967 FALSE /* Don't force to /x */ );
11969 /* Disambiguate between \N meaning a named character versus \N meaning
11970 * [^\n]. The latter is assumed when the {...} following the \N is a legal
11971 * quantifier, or there is no '{' at all */
11972 if (*p != '{' || regcurly(p)) {
11982 *node_p = reg_node(pRExC_state, REG_ANY);
11983 *flagp |= HASWIDTH|SIMPLE;
11985 Set_Node_Length(*node_p, 1); /* MJD */
11989 /* Here, we have decided it should be a named character or sequence */
11991 /* The test above made sure that the next real character is a '{', but
11992 * under the /x modifier, it could be separated by space (or a comment and
11993 * \n) and this is not allowed (for consistency with \x{...} and the
11994 * tokenizer handling of \N{NAME}). */
11995 if (*RExC_parse != '{') {
11996 vFAIL("Missing braces on \\N{}");
11999 RExC_parse++; /* Skip past the '{' */
12001 if (! (endbrace = strchr(RExC_parse, '}'))) { /* no trailing brace */
12002 vFAIL2("Missing right brace on \\%c{}", 'N');
12004 else if(!(endbrace == RExC_parse /* nothing between the {} */
12005 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked... */
12006 && strnEQ(RExC_parse, "U+", 2)))) /* ... below for a better
12009 RExC_parse = endbrace; /* position msg's '<--HERE' */
12010 vFAIL("\\N{NAME} must be resolved by the lexer");
12013 REQUIRE_UNI_RULES(flagp, FALSE); /* Unicode named chars imply Unicode
12016 if (endbrace == RExC_parse) { /* empty: \N{} */
12018 RExC_parse++; /* Position after the "}" */
12019 vFAIL("Zero length \\N{}");
12024 nextchar(pRExC_state);
12029 *node_p = reg_node(pRExC_state,NOTHING);
12033 RExC_parse += 2; /* Skip past the 'U+' */
12035 /* Because toke.c has generated a special construct for us guaranteed not
12036 * to have NULs, we can use a str function */
12037 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12039 /* Code points are separated by dots. If none, there is only one code
12040 * point, and is terminated by the brace */
12042 if (endchar >= endbrace) {
12043 STRLEN length_of_hex;
12044 I32 grok_hex_flags;
12046 /* Here, exactly one code point. If that isn't what is wanted, fail */
12047 if (! code_point_p) {
12052 /* Convert code point from hex */
12053 length_of_hex = (STRLEN)(endchar - RExC_parse);
12054 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
12055 | PERL_SCAN_DISALLOW_PREFIX
12057 /* No errors in the first pass (See [perl
12058 * #122671].) We let the code below find the
12059 * errors when there are multiple chars. */
12061 ? PERL_SCAN_SILENT_ILLDIGIT
12064 /* This routine is the one place where both single- and double-quotish
12065 * \N{U+xxxx} are evaluated. The value is a Unicode code point which
12066 * must be converted to native. */
12067 *code_point_p = UNI_TO_NATIVE(grok_hex(RExC_parse,
12072 /* The tokenizer should have guaranteed validity, but it's possible to
12073 * bypass it by using single quoting, so check. Don't do the check
12074 * here when there are multiple chars; we do it below anyway. */
12075 if (length_of_hex == 0
12076 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
12078 RExC_parse += length_of_hex; /* Includes all the valid */
12079 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12080 ? UTF8SKIP(RExC_parse)
12082 /* Guard against malformed utf8 */
12083 if (RExC_parse >= endchar) {
12084 RExC_parse = endchar;
12086 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12089 RExC_parse = endbrace + 1;
12092 else { /* Is a multiple character sequence */
12093 SV * substitute_parse;
12095 char *orig_end = RExC_end;
12096 char *save_start = RExC_start;
12099 /* Count the code points, if desired, in the sequence */
12102 while (RExC_parse < endbrace) {
12103 /* Point to the beginning of the next character in the sequence. */
12104 RExC_parse = endchar + 1;
12105 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12110 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12111 * But don't backup up the pointer if the caller want to know how many
12112 * code points there are (they can then handle things) */
12120 /* What is done here is to convert this to a sub-pattern of the form
12121 * \x{char1}\x{char2}... and then call reg recursively to parse it
12122 * (enclosing in "(?: ... )" ). That way, it retains its atomicness,
12123 * while not having to worry about special handling that some code
12124 * points may have. */
12126 substitute_parse = newSVpvs("?:");
12128 while (RExC_parse < endbrace) {
12130 /* Convert to notation the rest of the code understands */
12131 sv_catpv(substitute_parse, "\\x{");
12132 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
12133 sv_catpv(substitute_parse, "}");
12135 /* Point to the beginning of the next character in the sequence. */
12136 RExC_parse = endchar + 1;
12137 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12140 sv_catpv(substitute_parse, ")");
12142 RExC_parse = RExC_start = RExC_adjusted_start = SvPV(substitute_parse,
12145 /* Don't allow empty number */
12146 if (len < (STRLEN) 8) {
12147 RExC_parse = endbrace;
12148 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12150 RExC_end = RExC_parse + len;
12152 /* The values are Unicode, and therefore not subject to recoding, but
12153 * have to be converted to native on a non-Unicode (meaning non-ASCII)
12156 RExC_recode_x_to_native = 1;
12160 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
12161 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12162 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12165 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#" UVxf,
12168 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12171 /* Restore the saved values */
12172 RExC_start = RExC_adjusted_start = save_start;
12173 RExC_parse = endbrace;
12174 RExC_end = orig_end;
12176 RExC_recode_x_to_native = 0;
12179 SvREFCNT_dec_NN(substitute_parse);
12180 nextchar(pRExC_state);
12187 PERL_STATIC_INLINE U8
12188 S_compute_EXACTish(RExC_state_t *pRExC_state)
12192 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
12200 op = get_regex_charset(RExC_flags);
12201 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
12202 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
12203 been, so there is no hole */
12206 return op + EXACTF;
12209 PERL_STATIC_INLINE void
12210 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
12211 regnode *node, I32* flagp, STRLEN len, UV code_point,
12214 /* This knows the details about sizing an EXACTish node, setting flags for
12215 * it (by setting <*flagp>, and potentially populating it with a single
12218 * If <len> (the length in bytes) is non-zero, this function assumes that
12219 * the node has already been populated, and just does the sizing. In this
12220 * case <code_point> should be the final code point that has already been
12221 * placed into the node. This value will be ignored except that under some
12222 * circumstances <*flagp> is set based on it.
12224 * If <len> is zero, the function assumes that the node is to contain only
12225 * the single character given by <code_point> and calculates what <len>
12226 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
12227 * additionally will populate the node's STRING with <code_point> or its
12230 * In both cases <*flagp> is appropriately set
12232 * It knows that under FOLD, the Latin Sharp S and UTF characters above
12233 * 255, must be folded (the former only when the rules indicate it can
12236 * When it does the populating, it looks at the flag 'downgradable'. If
12237 * true with a node that folds, it checks if the single code point
12238 * participates in a fold, and if not downgrades the node to an EXACT.
12239 * This helps the optimizer */
12241 bool len_passed_in = cBOOL(len != 0);
12242 U8 character[UTF8_MAXBYTES_CASE+1];
12244 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
12246 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
12247 * sizing difference, and is extra work that is thrown away */
12248 if (downgradable && ! PASS2) {
12249 downgradable = FALSE;
12252 if (! len_passed_in) {
12254 if (UVCHR_IS_INVARIANT(code_point)) {
12255 if (LOC || ! FOLD) { /* /l defers folding until runtime */
12256 *character = (U8) code_point;
12258 else { /* Here is /i and not /l. (toFOLD() is defined on just
12259 ASCII, which isn't the same thing as INVARIANT on
12260 EBCDIC, but it works there, as the extra invariants
12261 fold to themselves) */
12262 *character = toFOLD((U8) code_point);
12264 /* We can downgrade to an EXACT node if this character
12265 * isn't a folding one. Note that this assumes that
12266 * nothing above Latin1 folds to some other invariant than
12267 * one of these alphabetics; otherwise we would also have
12269 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12270 * || ASCII_FOLD_RESTRICTED))
12272 if (downgradable && PL_fold[code_point] == code_point) {
12278 else if (FOLD && (! LOC
12279 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
12280 { /* Folding, and ok to do so now */
12281 UV folded = _to_uni_fold_flags(
12285 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12286 ? FOLD_FLAGS_NOMIX_ASCII
12289 && folded == code_point /* This quickly rules out many
12290 cases, avoiding the
12291 _invlist_contains_cp() overhead
12293 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
12300 else if (code_point <= MAX_UTF8_TWO_BYTE) {
12302 /* Not folding this cp, and can output it directly */
12303 *character = UTF8_TWO_BYTE_HI(code_point);
12304 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
12308 uvchr_to_utf8( character, code_point);
12309 len = UTF8SKIP(character);
12311 } /* Else pattern isn't UTF8. */
12313 *character = (U8) code_point;
12315 } /* Else is folded non-UTF8 */
12316 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12317 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12318 || UNICODE_DOT_DOT_VERSION > 0)
12319 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
12323 /* We don't fold any non-UTF8 except possibly the Sharp s (see
12324 * comments at join_exact()); */
12325 *character = (U8) code_point;
12328 /* Can turn into an EXACT node if we know the fold at compile time,
12329 * and it folds to itself and doesn't particpate in other folds */
12332 && PL_fold_latin1[code_point] == code_point
12333 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12334 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
12338 } /* else is Sharp s. May need to fold it */
12339 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
12341 *(character + 1) = 's';
12345 *character = LATIN_SMALL_LETTER_SHARP_S;
12351 RExC_size += STR_SZ(len);
12354 RExC_emit += STR_SZ(len);
12355 STR_LEN(node) = len;
12356 if (! len_passed_in) {
12357 Copy((char *) character, STRING(node), len, char);
12361 *flagp |= HASWIDTH;
12363 /* A single character node is SIMPLE, except for the special-cased SHARP S
12365 if ((len == 1 || (UTF && len == UVCHR_SKIP(code_point)))
12366 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12367 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12368 || UNICODE_DOT_DOT_VERSION > 0)
12369 && ( code_point != LATIN_SMALL_LETTER_SHARP_S
12370 || ! FOLD || ! DEPENDS_SEMANTICS)
12376 /* The OP may not be well defined in PASS1 */
12377 if (PASS2 && OP(node) == EXACTFL) {
12378 RExC_contains_locale = 1;
12383 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
12384 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
12387 S_backref_value(char *p)
12389 const char* endptr;
12391 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
12398 - regatom - the lowest level
12400 Try to identify anything special at the start of the current parse position.
12401 If there is, then handle it as required. This may involve generating a
12402 single regop, such as for an assertion; or it may involve recursing, such as
12403 to handle a () structure.
12405 If the string doesn't start with something special then we gobble up
12406 as much literal text as we can. If we encounter a quantifier, we have to
12407 back off the final literal character, as that quantifier applies to just it
12408 and not to the whole string of literals.
12410 Once we have been able to handle whatever type of thing started the
12411 sequence, we return.
12413 Note: we have to be careful with escapes, as they can be both literal
12414 and special, and in the case of \10 and friends, context determines which.
12416 A summary of the code structure is:
12418 switch (first_byte) {
12419 cases for each special:
12420 handle this special;
12423 switch (2nd byte) {
12424 cases for each unambiguous special:
12425 handle this special;
12427 cases for each ambigous special/literal:
12429 if (special) handle here
12431 default: // unambiguously literal:
12434 default: // is a literal char
12437 create EXACTish node for literal;
12438 while (more input and node isn't full) {
12439 switch (input_byte) {
12440 cases for each special;
12441 make sure parse pointer is set so that the next call to
12442 regatom will see this special first
12443 goto loopdone; // EXACTish node terminated by prev. char
12445 append char to EXACTISH node;
12447 get next input byte;
12451 return the generated node;
12453 Specifically there are two separate switches for handling
12454 escape sequences, with the one for handling literal escapes requiring
12455 a dummy entry for all of the special escapes that are actually handled
12458 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
12460 Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
12461 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
12462 Otherwise does not return NULL.
12466 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12468 regnode *ret = NULL;
12475 GET_RE_DEBUG_FLAGS_DECL;
12477 *flagp = WORST; /* Tentatively. */
12479 DEBUG_PARSE("atom");
12481 PERL_ARGS_ASSERT_REGATOM;
12484 parse_start = RExC_parse;
12485 assert(RExC_parse < RExC_end);
12486 switch ((U8)*RExC_parse) {
12488 RExC_seen_zerolen++;
12489 nextchar(pRExC_state);
12490 if (RExC_flags & RXf_PMf_MULTILINE)
12491 ret = reg_node(pRExC_state, MBOL);
12493 ret = reg_node(pRExC_state, SBOL);
12494 Set_Node_Length(ret, 1); /* MJD */
12497 nextchar(pRExC_state);
12499 RExC_seen_zerolen++;
12500 if (RExC_flags & RXf_PMf_MULTILINE)
12501 ret = reg_node(pRExC_state, MEOL);
12503 ret = reg_node(pRExC_state, SEOL);
12504 Set_Node_Length(ret, 1); /* MJD */
12507 nextchar(pRExC_state);
12508 if (RExC_flags & RXf_PMf_SINGLELINE)
12509 ret = reg_node(pRExC_state, SANY);
12511 ret = reg_node(pRExC_state, REG_ANY);
12512 *flagp |= HASWIDTH|SIMPLE;
12514 Set_Node_Length(ret, 1); /* MJD */
12518 char * const oregcomp_parse = ++RExC_parse;
12519 ret = regclass(pRExC_state, flagp,depth+1,
12520 FALSE, /* means parse the whole char class */
12521 TRUE, /* allow multi-char folds */
12522 FALSE, /* don't silence non-portable warnings. */
12523 (bool) RExC_strict,
12524 TRUE, /* Allow an optimized regnode result */
12528 if (*flagp & (RESTART_PASS1|NEED_UTF8))
12530 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12533 if (*RExC_parse != ']') {
12534 RExC_parse = oregcomp_parse;
12535 vFAIL("Unmatched [");
12537 nextchar(pRExC_state);
12538 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
12542 nextchar(pRExC_state);
12543 ret = reg(pRExC_state, 2, &flags,depth+1);
12545 if (flags & TRYAGAIN) {
12546 if (RExC_parse >= RExC_end) {
12547 /* Make parent create an empty node if needed. */
12548 *flagp |= TRYAGAIN;
12553 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12554 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12557 FAIL2("panic: reg returned NULL to regatom, flags=%#" UVxf,
12560 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12564 if (flags & TRYAGAIN) {
12565 *flagp |= TRYAGAIN;
12568 vFAIL("Internal urp");
12569 /* Supposed to be caught earlier. */
12575 vFAIL("Quantifier follows nothing");
12580 This switch handles escape sequences that resolve to some kind
12581 of special regop and not to literal text. Escape sequnces that
12582 resolve to literal text are handled below in the switch marked
12585 Every entry in this switch *must* have a corresponding entry
12586 in the literal escape switch. However, the opposite is not
12587 required, as the default for this switch is to jump to the
12588 literal text handling code.
12591 switch ((U8)*RExC_parse) {
12592 /* Special Escapes */
12594 RExC_seen_zerolen++;
12595 ret = reg_node(pRExC_state, SBOL);
12596 /* SBOL is shared with /^/ so we set the flags so we can tell
12597 * /\A/ from /^/ in split. We check ret because first pass we
12598 * have no regop struct to set the flags on. */
12602 goto finish_meta_pat;
12604 ret = reg_node(pRExC_state, GPOS);
12605 RExC_seen |= REG_GPOS_SEEN;
12607 goto finish_meta_pat;
12609 RExC_seen_zerolen++;
12610 ret = reg_node(pRExC_state, KEEPS);
12612 /* XXX:dmq : disabling in-place substitution seems to
12613 * be necessary here to avoid cases of memory corruption, as
12614 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
12616 RExC_seen |= REG_LOOKBEHIND_SEEN;
12617 goto finish_meta_pat;
12619 ret = reg_node(pRExC_state, SEOL);
12621 RExC_seen_zerolen++; /* Do not optimize RE away */
12622 goto finish_meta_pat;
12624 ret = reg_node(pRExC_state, EOS);
12626 RExC_seen_zerolen++; /* Do not optimize RE away */
12627 goto finish_meta_pat;
12629 vFAIL("\\C no longer supported");
12631 ret = reg_node(pRExC_state, CLUMP);
12632 *flagp |= HASWIDTH;
12633 goto finish_meta_pat;
12639 arg = ANYOF_WORDCHAR;
12647 regex_charset charset = get_regex_charset(RExC_flags);
12649 RExC_seen_zerolen++;
12650 RExC_seen |= REG_LOOKBEHIND_SEEN;
12651 op = BOUND + charset;
12653 if (op == BOUNDL) {
12654 RExC_contains_locale = 1;
12657 ret = reg_node(pRExC_state, op);
12659 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
12660 FLAGS(ret) = TRADITIONAL_BOUND;
12661 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
12667 char name = *RExC_parse;
12670 endbrace = strchr(RExC_parse, '}');
12673 vFAIL2("Missing right brace on \\%c{}", name);
12675 /* XXX Need to decide whether to take spaces or not. Should be
12676 * consistent with \p{}, but that currently is SPACE, which
12677 * means vertical too, which seems wrong
12678 * while (isBLANK(*RExC_parse)) {
12681 if (endbrace == RExC_parse) {
12682 RExC_parse++; /* After the '}' */
12683 vFAIL2("Empty \\%c{}", name);
12685 length = endbrace - RExC_parse;
12686 /*while (isBLANK(*(RExC_parse + length - 1))) {
12689 switch (*RExC_parse) {
12692 && (length != 3 || strnNE(RExC_parse + 1, "cb", 2)))
12694 goto bad_bound_type;
12696 FLAGS(ret) = GCB_BOUND;
12699 if (length != 2 || *(RExC_parse + 1) != 'b') {
12700 goto bad_bound_type;
12702 FLAGS(ret) = LB_BOUND;
12705 if (length != 2 || *(RExC_parse + 1) != 'b') {
12706 goto bad_bound_type;
12708 FLAGS(ret) = SB_BOUND;
12711 if (length != 2 || *(RExC_parse + 1) != 'b') {
12712 goto bad_bound_type;
12714 FLAGS(ret) = WB_BOUND;
12718 RExC_parse = endbrace;
12720 "'%" UTF8f "' is an unknown bound type",
12721 UTF8fARG(UTF, length, endbrace - length));
12722 NOT_REACHED; /*NOTREACHED*/
12724 RExC_parse = endbrace;
12725 REQUIRE_UNI_RULES(flagp, NULL);
12727 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
12731 /* Don't have to worry about UTF-8, in this message because
12732 * to get here the contents of the \b must be ASCII */
12733 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
12734 "Using /u for '%.*s' instead of /%s",
12736 endbrace - length + 1,
12737 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
12738 ? ASCII_RESTRICT_PAT_MODS
12739 : ASCII_MORE_RESTRICT_PAT_MODS);
12743 if (PASS2 && invert) {
12744 OP(ret) += NBOUND - BOUND;
12746 goto finish_meta_pat;
12754 if (! DEPENDS_SEMANTICS) {
12758 /* \d doesn't have any matches in the upper Latin1 range, hence /d
12759 * is equivalent to /u. Changing to /u saves some branches at
12762 goto join_posix_op_known;
12765 ret = reg_node(pRExC_state, LNBREAK);
12766 *flagp |= HASWIDTH|SIMPLE;
12767 goto finish_meta_pat;
12775 goto join_posix_op_known;
12781 arg = ANYOF_VERTWS;
12783 goto join_posix_op_known;
12793 op = POSIXD + get_regex_charset(RExC_flags);
12794 if (op > POSIXA) { /* /aa is same as /a */
12797 else if (op == POSIXL) {
12798 RExC_contains_locale = 1;
12801 join_posix_op_known:
12804 op += NPOSIXD - POSIXD;
12807 ret = reg_node(pRExC_state, op);
12809 FLAGS(ret) = namedclass_to_classnum(arg);
12812 *flagp |= HASWIDTH|SIMPLE;
12816 nextchar(pRExC_state);
12817 Set_Node_Length(ret, 2); /* MJD */
12823 ret = regclass(pRExC_state, flagp,depth+1,
12824 TRUE, /* means just parse this element */
12825 FALSE, /* don't allow multi-char folds */
12826 FALSE, /* don't silence non-portable warnings. It
12827 would be a bug if these returned
12829 (bool) RExC_strict,
12830 TRUE, /* Allow an optimized regnode result */
12833 if (*flagp & RESTART_PASS1)
12835 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
12836 * multi-char folds are allowed. */
12838 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12843 Set_Node_Offset(ret, parse_start);
12844 Set_Node_Cur_Length(ret, parse_start - 2);
12845 nextchar(pRExC_state);
12848 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
12849 * \N{...} evaluates to a sequence of more than one code points).
12850 * The function call below returns a regnode, which is our result.
12851 * The parameters cause it to fail if the \N{} evaluates to a
12852 * single code point; we handle those like any other literal. The
12853 * reason that the multicharacter case is handled here and not as
12854 * part of the EXACtish code is because of quantifiers. In
12855 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
12856 * this way makes that Just Happen. dmq.
12857 * join_exact() will join this up with adjacent EXACTish nodes
12858 * later on, if appropriate. */
12860 if (grok_bslash_N(pRExC_state,
12861 &ret, /* Want a regnode returned */
12862 NULL, /* Fail if evaluates to a single code
12864 NULL, /* Don't need a count of how many code
12873 if (*flagp & RESTART_PASS1)
12876 /* Here, evaluates to a single code point. Go get that */
12877 RExC_parse = parse_start;
12880 case 'k': /* Handle \k<NAME> and \k'NAME' */
12884 if ( RExC_parse >= RExC_end - 1
12885 || (( ch = RExC_parse[1]) != '<'
12890 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12891 vFAIL2("Sequence %.2s... not terminated",parse_start);
12894 ret = handle_named_backref(pRExC_state,
12906 case '1': case '2': case '3': case '4':
12907 case '5': case '6': case '7': case '8': case '9':
12912 if (*RExC_parse == 'g') {
12916 if (*RExC_parse == '{') {
12920 if (*RExC_parse == '-') {
12924 if (hasbrace && !isDIGIT(*RExC_parse)) {
12925 if (isrel) RExC_parse--;
12927 goto parse_named_seq;
12930 if (RExC_parse >= RExC_end) {
12931 goto unterminated_g;
12933 num = S_backref_value(RExC_parse);
12935 vFAIL("Reference to invalid group 0");
12936 else if (num == I32_MAX) {
12937 if (isDIGIT(*RExC_parse))
12938 vFAIL("Reference to nonexistent group");
12941 vFAIL("Unterminated \\g... pattern");
12945 num = RExC_npar - num;
12947 vFAIL("Reference to nonexistent or unclosed group");
12951 num = S_backref_value(RExC_parse);
12952 /* bare \NNN might be backref or octal - if it is larger
12953 * than or equal RExC_npar then it is assumed to be an
12954 * octal escape. Note RExC_npar is +1 from the actual
12955 * number of parens. */
12956 /* Note we do NOT check if num == I32_MAX here, as that is
12957 * handled by the RExC_npar check */
12960 /* any numeric escape < 10 is always a backref */
12962 /* any numeric escape < RExC_npar is a backref */
12963 && num >= RExC_npar
12964 /* cannot be an octal escape if it starts with 8 */
12965 && *RExC_parse != '8'
12966 /* cannot be an octal escape it it starts with 9 */
12967 && *RExC_parse != '9'
12970 /* Probably not a backref, instead likely to be an
12971 * octal character escape, e.g. \35 or \777.
12972 * The above logic should make it obvious why using
12973 * octal escapes in patterns is problematic. - Yves */
12974 RExC_parse = parse_start;
12979 /* At this point RExC_parse points at a numeric escape like
12980 * \12 or \88 or something similar, which we should NOT treat
12981 * as an octal escape. It may or may not be a valid backref
12982 * escape. For instance \88888888 is unlikely to be a valid
12984 while (isDIGIT(*RExC_parse))
12987 if (*RExC_parse != '}')
12988 vFAIL("Unterminated \\g{...} pattern");
12992 if (num > (I32)RExC_rx->nparens)
12993 vFAIL("Reference to nonexistent group");
12996 ret = reganode(pRExC_state,
12999 : (ASCII_FOLD_RESTRICTED)
13001 : (AT_LEAST_UNI_SEMANTICS)
13007 *flagp |= HASWIDTH;
13009 /* override incorrect value set in reganode MJD */
13010 Set_Node_Offset(ret, parse_start);
13011 Set_Node_Cur_Length(ret, parse_start-1);
13012 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13013 FALSE /* Don't force to /x */ );
13017 if (RExC_parse >= RExC_end)
13018 FAIL("Trailing \\");
13021 /* Do not generate "unrecognized" warnings here, we fall
13022 back into the quick-grab loop below */
13023 RExC_parse = parse_start;
13025 } /* end of switch on a \foo sequence */
13030 /* '#' comments should have been spaced over before this function was
13032 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13034 if (RExC_flags & RXf_PMf_EXTENDED) {
13035 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13036 if (RExC_parse < RExC_end)
13046 /* Here, we have determined that the next thing is probably a
13047 * literal character. RExC_parse points to the first byte of its
13048 * definition. (It still may be an escape sequence that evaluates
13049 * to a single character) */
13055 #define MAX_NODE_STRING_SIZE 127
13056 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
13058 U8 upper_parse = MAX_NODE_STRING_SIZE;
13059 U8 node_type = compute_EXACTish(pRExC_state);
13060 bool next_is_quantifier;
13061 char * oldp = NULL;
13063 /* We can convert EXACTF nodes to EXACTFU if they contain only
13064 * characters that match identically regardless of the target
13065 * string's UTF8ness. The reason to do this is that EXACTF is not
13066 * trie-able, EXACTFU is.
13068 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13069 * contain only above-Latin1 characters (hence must be in UTF8),
13070 * which don't participate in folds with Latin1-range characters,
13071 * as the latter's folds aren't known until runtime. (We don't
13072 * need to figure this out until pass 2) */
13073 bool maybe_exactfu = PASS2
13074 && (node_type == EXACTF || node_type == EXACTFL);
13076 /* If a folding node contains only code points that don't
13077 * participate in folds, it can be changed into an EXACT node,
13078 * which allows the optimizer more things to look for */
13081 ret = reg_node(pRExC_state, node_type);
13083 /* In pass1, folded, we use a temporary buffer instead of the
13084 * actual node, as the node doesn't exist yet */
13085 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
13091 /* We look for the EXACTFish to EXACT node optimizaton only if
13092 * folding. (And we don't need to figure this out until pass 2).
13093 * XXX It might actually make sense to split the node into portions
13094 * that are exact and ones that aren't, so that we could later use
13095 * the exact ones to find the longest fixed and floating strings.
13096 * One would want to join them back into a larger node. One could
13097 * use a pseudo regnode like 'EXACT_ORIG_FOLD' */
13098 maybe_exact = FOLD && PASS2;
13100 /* XXX The node can hold up to 255 bytes, yet this only goes to
13101 * 127. I (khw) do not know why. Keeping it somewhat less than
13102 * 255 allows us to not have to worry about overflow due to
13103 * converting to utf8 and fold expansion, but that value is
13104 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
13105 * split up by this limit into a single one using the real max of
13106 * 255. Even at 127, this breaks under rare circumstances. If
13107 * folding, we do not want to split a node at a character that is a
13108 * non-final in a multi-char fold, as an input string could just
13109 * happen to want to match across the node boundary. The join
13110 * would solve that problem if the join actually happens. But a
13111 * series of more than two nodes in a row each of 127 would cause
13112 * the first join to succeed to get to 254, but then there wouldn't
13113 * be room for the next one, which could at be one of those split
13114 * multi-char folds. I don't know of any fool-proof solution. One
13115 * could back off to end with only a code point that isn't such a
13116 * non-final, but it is possible for there not to be any in the
13119 assert( ! UTF /* Is at the beginning of a character */
13120 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13121 || UTF8_IS_START(UCHARAT(RExC_parse)));
13123 /* Here, we have a literal character. Find the maximal string of
13124 * them in the input that we can fit into a single EXACTish node.
13125 * We quit at the first non-literal or when the node gets full */
13126 for (p = RExC_parse;
13127 len < upper_parse && p < RExC_end;
13132 /* White space has already been ignored */
13133 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13134 || ! is_PATWS_safe((p), RExC_end, UTF));
13146 /* Literal Escapes Switch
13148 This switch is meant to handle escape sequences that
13149 resolve to a literal character.
13151 Every escape sequence that represents something
13152 else, like an assertion or a char class, is handled
13153 in the switch marked 'Special Escapes' above in this
13154 routine, but also has an entry here as anything that
13155 isn't explicitly mentioned here will be treated as
13156 an unescaped equivalent literal.
13159 switch ((U8)*++p) {
13160 /* These are all the special escapes. */
13161 case 'A': /* Start assertion */
13162 case 'b': case 'B': /* Word-boundary assertion*/
13163 case 'C': /* Single char !DANGEROUS! */
13164 case 'd': case 'D': /* digit class */
13165 case 'g': case 'G': /* generic-backref, pos assertion */
13166 case 'h': case 'H': /* HORIZWS */
13167 case 'k': case 'K': /* named backref, keep marker */
13168 case 'p': case 'P': /* Unicode property */
13169 case 'R': /* LNBREAK */
13170 case 's': case 'S': /* space class */
13171 case 'v': case 'V': /* VERTWS */
13172 case 'w': case 'W': /* word class */
13173 case 'X': /* eXtended Unicode "combining
13174 character sequence" */
13175 case 'z': case 'Z': /* End of line/string assertion */
13179 /* Anything after here is an escape that resolves to a
13180 literal. (Except digits, which may or may not)
13186 case 'N': /* Handle a single-code point named character. */
13187 RExC_parse = p + 1;
13188 if (! grok_bslash_N(pRExC_state,
13189 NULL, /* Fail if evaluates to
13190 anything other than a
13191 single code point */
13192 &ender, /* The returned single code
13194 NULL, /* Don't need a count of
13195 how many code points */
13200 if (*flagp & NEED_UTF8)
13201 FAIL("panic: grok_bslash_N set NEED_UTF8");
13202 if (*flagp & RESTART_PASS1)
13205 /* Here, it wasn't a single code point. Go close
13206 * up this EXACTish node. The switch() prior to
13207 * this switch handles the other cases */
13208 RExC_parse = p = oldp;
13212 if (ender > 0xff) {
13213 REQUIRE_UTF8(flagp);
13229 ender = ESC_NATIVE;
13239 const char* error_msg;
13241 bool valid = grok_bslash_o(&p,
13244 PASS2, /* out warnings */
13245 (bool) RExC_strict,
13246 TRUE, /* Output warnings
13251 RExC_parse = p; /* going to die anyway; point
13252 to exact spot of failure */
13256 if (ender > 0xff) {
13257 REQUIRE_UTF8(flagp);
13263 UV result = UV_MAX; /* initialize to erroneous
13265 const char* error_msg;
13267 bool valid = grok_bslash_x(&p,
13270 PASS2, /* out warnings */
13271 (bool) RExC_strict,
13272 TRUE, /* Silence warnings
13277 RExC_parse = p; /* going to die anyway; point
13278 to exact spot of failure */
13283 if (ender < 0x100) {
13285 if (RExC_recode_x_to_native) {
13286 ender = LATIN1_TO_NATIVE(ender);
13291 REQUIRE_UTF8(flagp);
13297 ender = grok_bslash_c(*p++, PASS2);
13299 case '8': case '9': /* must be a backreference */
13301 /* we have an escape like \8 which cannot be an octal escape
13302 * so we exit the loop, and let the outer loop handle this
13303 * escape which may or may not be a legitimate backref. */
13305 case '1': case '2': case '3':case '4':
13306 case '5': case '6': case '7':
13307 /* When we parse backslash escapes there is ambiguity
13308 * between backreferences and octal escapes. Any escape
13309 * from \1 - \9 is a backreference, any multi-digit
13310 * escape which does not start with 0 and which when
13311 * evaluated as decimal could refer to an already
13312 * parsed capture buffer is a back reference. Anything
13315 * Note this implies that \118 could be interpreted as
13316 * 118 OR as "\11" . "8" depending on whether there
13317 * were 118 capture buffers defined already in the
13320 /* NOTE, RExC_npar is 1 more than the actual number of
13321 * parens we have seen so far, hence the < RExC_npar below. */
13323 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
13324 { /* Not to be treated as an octal constant, go
13332 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13334 ender = grok_oct(p, &numlen, &flags, NULL);
13335 if (ender > 0xff) {
13336 REQUIRE_UTF8(flagp);
13339 if (PASS2 /* like \08, \178 */
13341 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
13343 reg_warn_non_literal_string(
13345 form_short_octal_warning(p, numlen));
13351 FAIL("Trailing \\");
13354 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
13355 /* Include any left brace following the alpha to emphasize
13356 * that it could be part of an escape at some point
13358 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
13359 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
13361 goto normal_default;
13362 } /* End of switch on '\' */
13365 /* Currently we don't care if the lbrace is at the start
13366 * of a construct. This catches it in the middle of a
13367 * literal string, or when it's the first thing after
13368 * something like "\b" */
13369 if (len || (p > RExC_start && isALPHA_A(*(p -1)))) {
13370 RExC_parse = p + 1;
13371 vFAIL("Unescaped left brace in regex is illegal here");
13373 goto normal_default;
13376 if (PASS2 && p > RExC_parse && RExC_strict) {
13377 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
13380 default: /* A literal character */
13382 if (! UTF8_IS_INVARIANT(*p) && UTF) {
13384 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
13385 &numlen, UTF8_ALLOW_DEFAULT);
13391 } /* End of switch on the literal */
13393 /* Here, have looked at the literal character and <ender>
13394 * contains its ordinal, <p> points to the character after it.
13395 * We need to check if the next non-ignored thing is a
13396 * quantifier. Move <p> to after anything that should be
13397 * ignored, which, as a side effect, positions <p> for the next
13398 * loop iteration */
13399 skip_to_be_ignored_text(pRExC_state, &p,
13400 FALSE /* Don't force to /x */ );
13402 /* If the next thing is a quantifier, it applies to this
13403 * character only, which means that this character has to be in
13404 * its own node and can't just be appended to the string in an
13405 * existing node, so if there are already other characters in
13406 * the node, close the node with just them, and set up to do
13407 * this character again next time through, when it will be the
13408 * only thing in its new node */
13410 if ((next_is_quantifier = ( LIKELY(p < RExC_end)
13411 && UNLIKELY(ISMULT2(p))))
13418 /* Ready to add 'ender' to the node */
13420 if (! FOLD) { /* The simple case, just append the literal */
13422 /* In the sizing pass, we need only the size of the
13423 * character we are appending, hence we can delay getting
13424 * its representation until PASS2. */
13427 const STRLEN unilen = UVCHR_SKIP(ender);
13430 /* We have to subtract 1 just below (and again in
13431 * the corresponding PASS2 code) because the loop
13432 * increments <len> each time, as all but this path
13433 * (and one other) through it add a single byte to
13434 * the EXACTish node. But these paths would change
13435 * len to be the correct final value, so cancel out
13436 * the increment that follows */
13442 } else { /* PASS2 */
13445 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
13446 len += (char *) new_s - s - 1;
13447 s = (char *) new_s;
13450 *(s++) = (char) ender;
13454 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
13456 /* Here are folding under /l, and the code point is
13457 * problematic. First, we know we can't simplify things */
13458 maybe_exact = FALSE;
13459 maybe_exactfu = FALSE;
13461 /* A problematic code point in this context means that its
13462 * fold isn't known until runtime, so we can't fold it now.
13463 * (The non-problematic code points are the above-Latin1
13464 * ones that fold to also all above-Latin1. Their folds
13465 * don't vary no matter what the locale is.) But here we
13466 * have characters whose fold depends on the locale.
13467 * Unlike the non-folding case above, we have to keep track
13468 * of these in the sizing pass, so that we can make sure we
13469 * don't split too-long nodes in the middle of a potential
13470 * multi-char fold. And unlike the regular fold case
13471 * handled in the else clauses below, we don't actually
13472 * fold and don't have special cases to consider. What we
13473 * do for both passes is the PASS2 code for non-folding */
13474 goto not_fold_common;
13476 else /* A regular FOLD code point */
13478 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13479 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13480 || UNICODE_DOT_DOT_VERSION > 0)
13481 /* See comments for join_exact() as to why we fold
13482 * this non-UTF at compile time */
13483 || ( node_type == EXACTFU
13484 && ender == LATIN_SMALL_LETTER_SHARP_S)
13487 /* Here, are folding and are not UTF-8 encoded; therefore
13488 * the character must be in the range 0-255, and is not /l
13489 * (Not /l because we already handled these under /l in
13490 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
13491 if (IS_IN_SOME_FOLD_L1(ender)) {
13492 maybe_exact = FALSE;
13494 /* See if the character's fold differs between /d and
13495 * /u. This includes the multi-char fold SHARP S to
13497 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
13498 RExC_seen_unfolded_sharp_s = 1;
13499 maybe_exactfu = FALSE;
13501 else if (maybe_exactfu
13502 && (PL_fold[ender] != PL_fold_latin1[ender]
13503 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13504 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13505 || UNICODE_DOT_DOT_VERSION > 0)
13507 && isALPHA_FOLD_EQ(ender, 's')
13508 && isALPHA_FOLD_EQ(*(s-1), 's'))
13511 maybe_exactfu = FALSE;
13515 /* Even when folding, we store just the input character, as
13516 * we have an array that finds its fold quickly */
13517 *(s++) = (char) ender;
13519 else { /* FOLD, and UTF (or sharp s) */
13520 /* Unlike the non-fold case, we do actually have to
13521 * calculate the results here in pass 1. This is for two
13522 * reasons, the folded length may be longer than the
13523 * unfolded, and we have to calculate how many EXACTish
13524 * nodes it will take; and we may run out of room in a node
13525 * in the middle of a potential multi-char fold, and have
13526 * to back off accordingly. */
13529 if (isASCII_uni(ender)) {
13530 folded = toFOLD(ender);
13531 *(s)++ = (U8) folded;
13536 folded = _to_uni_fold_flags(
13540 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
13541 ? FOLD_FLAGS_NOMIX_ASCII
13545 /* The loop increments <len> each time, as all but this
13546 * path (and one other) through it add a single byte to
13547 * the EXACTish node. But this one has changed len to
13548 * be the correct final value, so subtract one to
13549 * cancel out the increment that follows */
13550 len += foldlen - 1;
13552 /* If this node only contains non-folding code points so
13553 * far, see if this new one is also non-folding */
13555 if (folded != ender) {
13556 maybe_exact = FALSE;
13559 /* Here the fold is the original; we have to check
13560 * further to see if anything folds to it */
13561 if (_invlist_contains_cp(PL_utf8_foldable,
13564 maybe_exact = FALSE;
13571 if (next_is_quantifier) {
13573 /* Here, the next input is a quantifier, and to get here,
13574 * the current character is the only one in the node.
13575 * Also, here <len> doesn't include the final byte for this
13581 } /* End of loop through literal characters */
13583 /* Here we have either exhausted the input or ran out of room in
13584 * the node. (If we encountered a character that can't be in the
13585 * node, transfer is made directly to <loopdone>, and so we
13586 * wouldn't have fallen off the end of the loop.) In the latter
13587 * case, we artificially have to split the node into two, because
13588 * we just don't have enough space to hold everything. This
13589 * creates a problem if the final character participates in a
13590 * multi-character fold in the non-final position, as a match that
13591 * should have occurred won't, due to the way nodes are matched,
13592 * and our artificial boundary. So back off until we find a non-
13593 * problematic character -- one that isn't at the beginning or
13594 * middle of such a fold. (Either it doesn't participate in any
13595 * folds, or appears only in the final position of all the folds it
13596 * does participate in.) A better solution with far fewer false
13597 * positives, and that would fill the nodes more completely, would
13598 * be to actually have available all the multi-character folds to
13599 * test against, and to back-off only far enough to be sure that
13600 * this node isn't ending with a partial one. <upper_parse> is set
13601 * further below (if we need to reparse the node) to include just
13602 * up through that final non-problematic character that this code
13603 * identifies, so when it is set to less than the full node, we can
13604 * skip the rest of this */
13605 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
13607 const STRLEN full_len = len;
13609 assert(len >= MAX_NODE_STRING_SIZE);
13611 /* Here, <s> points to the final byte of the final character.
13612 * Look backwards through the string until find a non-
13613 * problematic character */
13617 /* This has no multi-char folds to non-UTF characters */
13618 if (ASCII_FOLD_RESTRICTED) {
13622 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
13626 if (! PL_NonL1NonFinalFold) {
13627 PL_NonL1NonFinalFold = _new_invlist_C_array(
13628 NonL1_Perl_Non_Final_Folds_invlist);
13631 /* Point to the first byte of the final character */
13632 s = (char *) utf8_hop((U8 *) s, -1);
13634 while (s >= s0) { /* Search backwards until find
13635 non-problematic char */
13636 if (UTF8_IS_INVARIANT(*s)) {
13638 /* There are no ascii characters that participate
13639 * in multi-char folds under /aa. In EBCDIC, the
13640 * non-ascii invariants are all control characters,
13641 * so don't ever participate in any folds. */
13642 if (ASCII_FOLD_RESTRICTED
13643 || ! IS_NON_FINAL_FOLD(*s))
13648 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
13649 if (! IS_NON_FINAL_FOLD(EIGHT_BIT_UTF8_TO_NATIVE(
13655 else if (! _invlist_contains_cp(
13656 PL_NonL1NonFinalFold,
13657 valid_utf8_to_uvchr((U8 *) s, NULL)))
13662 /* Here, the current character is problematic in that
13663 * it does occur in the non-final position of some
13664 * fold, so try the character before it, but have to
13665 * special case the very first byte in the string, so
13666 * we don't read outside the string */
13667 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
13668 } /* End of loop backwards through the string */
13670 /* If there were only problematic characters in the string,
13671 * <s> will point to before s0, in which case the length
13672 * should be 0, otherwise include the length of the
13673 * non-problematic character just found */
13674 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
13677 /* Here, have found the final character, if any, that is
13678 * non-problematic as far as ending the node without splitting
13679 * it across a potential multi-char fold. <len> contains the
13680 * number of bytes in the node up-to and including that
13681 * character, or is 0 if there is no such character, meaning
13682 * the whole node contains only problematic characters. In
13683 * this case, give up and just take the node as-is. We can't
13688 /* If the node ends in an 's' we make sure it stays EXACTF,
13689 * as if it turns into an EXACTFU, it could later get
13690 * joined with another 's' that would then wrongly match
13692 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
13694 maybe_exactfu = FALSE;
13698 /* Here, the node does contain some characters that aren't
13699 * problematic. If one such is the final character in the
13700 * node, we are done */
13701 if (len == full_len) {
13704 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
13706 /* If the final character is problematic, but the
13707 * penultimate is not, back-off that last character to
13708 * later start a new node with it */
13713 /* Here, the final non-problematic character is earlier
13714 * in the input than the penultimate character. What we do
13715 * is reparse from the beginning, going up only as far as
13716 * this final ok one, thus guaranteeing that the node ends
13717 * in an acceptable character. The reason we reparse is
13718 * that we know how far in the character is, but we don't
13719 * know how to correlate its position with the input parse.
13720 * An alternate implementation would be to build that
13721 * correlation as we go along during the original parse,
13722 * but that would entail extra work for every node, whereas
13723 * this code gets executed only when the string is too
13724 * large for the node, and the final two characters are
13725 * problematic, an infrequent occurrence. Yet another
13726 * possible strategy would be to save the tail of the
13727 * string, and the next time regatom is called, initialize
13728 * with that. The problem with this is that unless you
13729 * back off one more character, you won't be guaranteed
13730 * regatom will get called again, unless regbranch,
13731 * regpiece ... are also changed. If you do back off that
13732 * extra character, so that there is input guaranteed to
13733 * force calling regatom, you can't handle the case where
13734 * just the first character in the node is acceptable. I
13735 * (khw) decided to try this method which doesn't have that
13736 * pitfall; if performance issues are found, we can do a
13737 * combination of the current approach plus that one */
13743 } /* End of verifying node ends with an appropriate char */
13745 loopdone: /* Jumped to when encounters something that shouldn't be
13748 /* I (khw) don't know if you can get here with zero length, but the
13749 * old code handled this situation by creating a zero-length EXACT
13750 * node. Might as well be NOTHING instead */
13756 /* If 'maybe_exact' is still set here, means there are no
13757 * code points in the node that participate in folds;
13758 * similarly for 'maybe_exactfu' and code points that match
13759 * differently depending on UTF8ness of the target string
13760 * (for /u), or depending on locale for /l */
13766 else if (maybe_exactfu) {
13772 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
13773 FALSE /* Don't look to see if could
13774 be turned into an EXACT
13775 node, as we have already
13780 RExC_parse = p - 1;
13781 Set_Node_Cur_Length(ret, parse_start);
13784 /* len is STRLEN which is unsigned, need to copy to signed */
13787 vFAIL("Internal disaster");
13790 } /* End of label 'defchar:' */
13792 } /* End of giant switch on input character */
13794 /* Position parse to next real character */
13795 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13796 FALSE /* Don't force to /x */ );
13797 if (PASS2 && *RExC_parse == '{' && OP(ret) != SBOL && ! regcurly(RExC_parse)) {
13798 ckWARNregdep(RExC_parse + 1, "Unescaped left brace in regex is deprecated here (and will be fatal in Perl 5.30), passed through");
13806 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
13808 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
13809 * sets up the bitmap and any flags, removing those code points from the
13810 * inversion list, setting it to NULL should it become completely empty */
13812 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
13813 assert(PL_regkind[OP(node)] == ANYOF);
13815 ANYOF_BITMAP_ZERO(node);
13816 if (*invlist_ptr) {
13818 /* This gets set if we actually need to modify things */
13819 bool change_invlist = FALSE;
13823 /* Start looking through *invlist_ptr */
13824 invlist_iterinit(*invlist_ptr);
13825 while (invlist_iternext(*invlist_ptr, &start, &end)) {
13829 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
13830 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
13833 /* Quit if are above what we should change */
13834 if (start >= NUM_ANYOF_CODE_POINTS) {
13838 change_invlist = TRUE;
13840 /* Set all the bits in the range, up to the max that we are doing */
13841 high = (end < NUM_ANYOF_CODE_POINTS - 1)
13843 : NUM_ANYOF_CODE_POINTS - 1;
13844 for (i = start; i <= (int) high; i++) {
13845 if (! ANYOF_BITMAP_TEST(node, i)) {
13846 ANYOF_BITMAP_SET(node, i);
13850 invlist_iterfinish(*invlist_ptr);
13852 /* Done with loop; remove any code points that are in the bitmap from
13853 * *invlist_ptr; similarly for code points above the bitmap if we have
13854 * a flag to match all of them anyways */
13855 if (change_invlist) {
13856 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
13858 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
13859 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
13862 /* If have completely emptied it, remove it completely */
13863 if (_invlist_len(*invlist_ptr) == 0) {
13864 SvREFCNT_dec_NN(*invlist_ptr);
13865 *invlist_ptr = NULL;
13870 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
13871 Character classes ([:foo:]) can also be negated ([:^foo:]).
13872 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
13873 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
13874 but trigger failures because they are currently unimplemented. */
13876 #define POSIXCC_DONE(c) ((c) == ':')
13877 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
13878 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
13879 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
13881 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
13882 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
13883 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
13885 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
13887 /* 'posix_warnings' and 'warn_text' are names of variables in the following
13889 #define ADD_POSIX_WARNING(p, text) STMT_START { \
13890 if (posix_warnings) { \
13891 if (! RExC_warn_text ) RExC_warn_text = (AV *) sv_2mortal((SV *) newAV()); \
13892 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
13896 REPORT_LOCATION_ARGS(p))); \
13901 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
13903 const char * const s, /* Where the putative posix class begins.
13904 Normally, this is one past the '['. This
13905 parameter exists so it can be somewhere
13906 besides RExC_parse. */
13907 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
13909 AV ** posix_warnings, /* Where to place any generated warnings, or
13911 const bool check_only /* Don't die if error */
13914 /* This parses what the caller thinks may be one of the three POSIX
13916 * 1) a character class, like [:blank:]
13917 * 2) a collating symbol, like [. .]
13918 * 3) an equivalence class, like [= =]
13919 * In the latter two cases, it croaks if it finds a syntactically legal
13920 * one, as these are not handled by Perl.
13922 * The main purpose is to look for a POSIX character class. It returns:
13923 * a) the class number
13924 * if it is a completely syntactically and semantically legal class.
13925 * 'updated_parse_ptr', if not NULL, is set to point to just after the
13926 * closing ']' of the class
13927 * b) OOB_NAMEDCLASS
13928 * if it appears that one of the three POSIX constructs was meant, but
13929 * its specification was somehow defective. 'updated_parse_ptr', if
13930 * not NULL, is set to point to the character just after the end
13931 * character of the class. See below for handling of warnings.
13932 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
13933 * if it doesn't appear that a POSIX construct was intended.
13934 * 'updated_parse_ptr' is not changed. No warnings nor errors are
13937 * In b) there may be errors or warnings generated. If 'check_only' is
13938 * TRUE, then any errors are discarded. Warnings are returned to the
13939 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
13940 * instead it is NULL, warnings are suppressed. This is done in all
13941 * passes. The reason for this is that the rest of the parsing is heavily
13942 * dependent on whether this routine found a valid posix class or not. If
13943 * it did, the closing ']' is absorbed as part of the class. If no class,
13944 * or an invalid one is found, any ']' will be considered the terminator of
13945 * the outer bracketed character class, leading to very different results.
13946 * In particular, a '(?[ ])' construct will likely have a syntax error if
13947 * the class is parsed other than intended, and this will happen in pass1,
13948 * before the warnings would normally be output. This mechanism allows the
13949 * caller to output those warnings in pass1 just before dieing, giving a
13950 * much better clue as to what is wrong.
13952 * The reason for this function, and its complexity is that a bracketed
13953 * character class can contain just about anything. But it's easy to
13954 * mistype the very specific posix class syntax but yielding a valid
13955 * regular bracketed class, so it silently gets compiled into something
13956 * quite unintended.
13958 * The solution adopted here maintains backward compatibility except that
13959 * it adds a warning if it looks like a posix class was intended but
13960 * improperly specified. The warning is not raised unless what is input
13961 * very closely resembles one of the 14 legal posix classes. To do this,
13962 * it uses fuzzy parsing. It calculates how many single-character edits it
13963 * would take to transform what was input into a legal posix class. Only
13964 * if that number is quite small does it think that the intention was a
13965 * posix class. Obviously these are heuristics, and there will be cases
13966 * where it errs on one side or another, and they can be tweaked as
13967 * experience informs.
13969 * The syntax for a legal posix class is:
13971 * qr/(?xa: \[ : \^? [:lower:]{4,6} : \] )/
13973 * What this routine considers syntactically to be an intended posix class
13974 * is this (the comments indicate some restrictions that the pattern
13977 * qr/(?x: \[? # The left bracket, possibly
13979 * \h* # possibly followed by blanks
13980 * (?: \^ \h* )? # possibly a misplaced caret
13981 * [:;]? # The opening class character,
13982 * # possibly omitted. A typo
13983 * # semi-colon can also be used.
13985 * \^? # possibly a correctly placed
13986 * # caret, but not if there was also
13987 * # a misplaced one
13989 * .{3,15} # The class name. If there are
13990 * # deviations from the legal syntax,
13991 * # its edit distance must be close
13992 * # to a real class name in order
13993 * # for it to be considered to be
13994 * # an intended posix class.
13996 * [:punct:]? # The closing class character,
13997 * # possibly omitted. If not a colon
13998 * # nor semi colon, the class name
13999 * # must be even closer to a valid
14002 * \]? # The right bracket, possibly
14006 * In the above, \h must be ASCII-only.
14008 * These are heuristics, and can be tweaked as field experience dictates.
14009 * There will be cases when someone didn't intend to specify a posix class
14010 * that this warns as being so. The goal is to minimize these, while
14011 * maximizing the catching of things intended to be a posix class that
14012 * aren't parsed as such.
14016 const char * const e = RExC_end;
14017 unsigned complement = 0; /* If to complement the class */
14018 bool found_problem = FALSE; /* Assume OK until proven otherwise */
14019 bool has_opening_bracket = FALSE;
14020 bool has_opening_colon = FALSE;
14021 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
14023 const char * possible_end = NULL; /* used for a 2nd parse pass */
14024 const char* name_start; /* ptr to class name first char */
14026 /* If the number of single-character typos the input name is away from a
14027 * legal name is no more than this number, it is considered to have meant
14028 * the legal name */
14029 int max_distance = 2;
14031 /* to store the name. The size determines the maximum length before we
14032 * decide that no posix class was intended. Should be at least
14033 * sizeof("alphanumeric") */
14036 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
14038 if (posix_warnings && RExC_warn_text)
14039 av_clear(RExC_warn_text);
14042 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14045 if (*(p - 1) != '[') {
14046 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
14047 found_problem = TRUE;
14050 has_opening_bracket = TRUE;
14053 /* They could be confused and think you can put spaces between the
14056 found_problem = TRUE;
14060 } while (p < e && isBLANK(*p));
14062 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14065 /* For [. .] and [= =]. These are quite different internally from [: :],
14066 * so they are handled separately. */
14067 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
14068 and 1 for at least one char in it
14071 const char open_char = *p;
14072 const char * temp_ptr = p + 1;
14074 /* These two constructs are not handled by perl, and if we find a
14075 * syntactically valid one, we croak. khw, who wrote this code, finds
14076 * this explanation of them very unclear:
14077 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
14078 * And searching the rest of the internet wasn't very helpful either.
14079 * It looks like just about any byte can be in these constructs,
14080 * depending on the locale. But unless the pattern is being compiled
14081 * under /l, which is very rare, Perl runs under the C or POSIX locale.
14082 * In that case, it looks like [= =] isn't allowed at all, and that
14083 * [. .] could be any single code point, but for longer strings the
14084 * constituent characters would have to be the ASCII alphabetics plus
14085 * the minus-hyphen. Any sensible locale definition would limit itself
14086 * to these. And any portable one definitely should. Trying to parse
14087 * the general case is a nightmare (see [perl #127604]). So, this code
14088 * looks only for interiors of these constructs that match:
14090 * Using \w relaxes the apparent rules a little, without adding much
14091 * danger of mistaking something else for one of these constructs.
14093 * [. .] in some implementations described on the internet is usable to
14094 * escape a character that otherwise is special in bracketed character
14095 * classes. For example [.].] means a literal right bracket instead of
14096 * the ending of the class
14098 * [= =] can legitimately contain a [. .] construct, but we don't
14099 * handle this case, as that [. .] construct will later get parsed
14100 * itself and croak then. And [= =] is checked for even when not under
14101 * /l, as Perl has long done so.
14103 * The code below relies on there being a trailing NUL, so it doesn't
14104 * have to keep checking if the parse ptr < e.
14106 if (temp_ptr[1] == open_char) {
14109 else while ( temp_ptr < e
14110 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
14115 if (*temp_ptr == open_char) {
14117 if (*temp_ptr == ']') {
14119 if (! found_problem && ! check_only) {
14120 RExC_parse = (char *) temp_ptr;
14121 vFAIL3("POSIX syntax [%c %c] is reserved for future "
14122 "extensions", open_char, open_char);
14125 /* Here, the syntax wasn't completely valid, or else the call
14126 * is to check-only */
14127 if (updated_parse_ptr) {
14128 *updated_parse_ptr = (char *) temp_ptr;
14131 return OOB_NAMEDCLASS;
14135 /* If we find something that started out to look like one of these
14136 * constructs, but isn't, we continue below so that it can be checked
14137 * for being a class name with a typo of '.' or '=' instead of a colon.
14141 /* Here, we think there is a possibility that a [: :] class was meant, and
14142 * we have the first real character. It could be they think the '^' comes
14145 found_problem = TRUE;
14146 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
14151 found_problem = TRUE;
14155 } while (p < e && isBLANK(*p));
14157 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14161 /* But the first character should be a colon, which they could have easily
14162 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
14163 * distinguish from a colon, so treat that as a colon). */
14166 has_opening_colon = TRUE;
14168 else if (*p == ';') {
14169 found_problem = TRUE;
14171 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14172 has_opening_colon = TRUE;
14175 found_problem = TRUE;
14176 ADD_POSIX_WARNING(p, "there must be a starting ':'");
14178 /* Consider an initial punctuation (not one of the recognized ones) to
14179 * be a left terminator */
14180 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
14185 /* They may think that you can put spaces between the components */
14187 found_problem = TRUE;
14191 } while (p < e && isBLANK(*p));
14193 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14198 /* We consider something like [^:^alnum:]] to not have been intended to
14199 * be a posix class, but XXX maybe we should */
14201 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14208 /* Again, they may think that you can put spaces between the components */
14210 found_problem = TRUE;
14214 } while (p < e && isBLANK(*p));
14216 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14221 /* XXX This ']' may be a typo, and something else was meant. But
14222 * treating it as such creates enough complications, that that
14223 * possibility isn't currently considered here. So we assume that the
14224 * ']' is what is intended, and if we've already found an initial '[',
14225 * this leaves this construct looking like [:] or [:^], which almost
14226 * certainly weren't intended to be posix classes */
14227 if (has_opening_bracket) {
14228 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14231 /* But this function can be called when we parse the colon for
14232 * something like qr/[alpha:]]/, so we back up to look for the
14237 found_problem = TRUE;
14238 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14240 else if (*p != ':') {
14242 /* XXX We are currently very restrictive here, so this code doesn't
14243 * consider the possibility that, say, /[alpha.]]/ was intended to
14244 * be a posix class. */
14245 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14248 /* Here we have something like 'foo:]'. There was no initial colon,
14249 * and we back up over 'foo. XXX Unlike the going forward case, we
14250 * don't handle typos of non-word chars in the middle */
14251 has_opening_colon = FALSE;
14254 while (p > RExC_start && isWORDCHAR(*p)) {
14259 /* Here, we have positioned ourselves to where we think the first
14260 * character in the potential class is */
14263 /* Now the interior really starts. There are certain key characters that
14264 * can end the interior, or these could just be typos. To catch both
14265 * cases, we may have to do two passes. In the first pass, we keep on
14266 * going unless we come to a sequence that matches
14267 * qr/ [[:punct:]] [[:blank:]]* \] /xa
14268 * This means it takes a sequence to end the pass, so two typos in a row if
14269 * that wasn't what was intended. If the class is perfectly formed, just
14270 * this one pass is needed. We also stop if there are too many characters
14271 * being accumulated, but this number is deliberately set higher than any
14272 * real class. It is set high enough so that someone who thinks that
14273 * 'alphanumeric' is a correct name would get warned that it wasn't.
14274 * While doing the pass, we keep track of where the key characters were in
14275 * it. If we don't find an end to the class, and one of the key characters
14276 * was found, we redo the pass, but stop when we get to that character.
14277 * Thus the key character was considered a typo in the first pass, but a
14278 * terminator in the second. If two key characters are found, we stop at
14279 * the second one in the first pass. Again this can miss two typos, but
14280 * catches a single one
14282 * In the first pass, 'possible_end' starts as NULL, and then gets set to
14283 * point to the first key character. For the second pass, it starts as -1.
14289 bool has_blank = FALSE;
14290 bool has_upper = FALSE;
14291 bool has_terminating_colon = FALSE;
14292 bool has_terminating_bracket = FALSE;
14293 bool has_semi_colon = FALSE;
14294 unsigned int name_len = 0;
14295 int punct_count = 0;
14299 /* Squeeze out blanks when looking up the class name below */
14300 if (isBLANK(*p) ) {
14302 found_problem = TRUE;
14307 /* The name will end with a punctuation */
14309 const char * peek = p + 1;
14311 /* Treat any non-']' punctuation followed by a ']' (possibly
14312 * with intervening blanks) as trying to terminate the class.
14313 * ']]' is very likely to mean a class was intended (but
14314 * missing the colon), but the warning message that gets
14315 * generated shows the error position better if we exit the
14316 * loop at the bottom (eventually), so skip it here. */
14318 if (peek < e && isBLANK(*peek)) {
14320 found_problem = TRUE;
14323 } while (peek < e && isBLANK(*peek));
14326 if (peek < e && *peek == ']') {
14327 has_terminating_bracket = TRUE;
14329 has_terminating_colon = TRUE;
14331 else if (*p == ';') {
14332 has_semi_colon = TRUE;
14333 has_terminating_colon = TRUE;
14336 found_problem = TRUE;
14343 /* Here we have punctuation we thought didn't end the class.
14344 * Keep track of the position of the key characters that are
14345 * more likely to have been class-enders */
14346 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
14348 /* Allow just one such possible class-ender not actually
14349 * ending the class. */
14350 if (possible_end) {
14356 /* If we have too many punctuation characters, no use in
14358 if (++punct_count > max_distance) {
14362 /* Treat the punctuation as a typo. */
14363 input_text[name_len++] = *p;
14366 else if (isUPPER(*p)) { /* Use lowercase for lookup */
14367 input_text[name_len++] = toLOWER(*p);
14369 found_problem = TRUE;
14371 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
14372 input_text[name_len++] = *p;
14376 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
14380 /* The declaration of 'input_text' is how long we allow a potential
14381 * class name to be, before saying they didn't mean a class name at
14383 if (name_len >= C_ARRAY_LENGTH(input_text)) {
14388 /* We get to here when the possible class name hasn't been properly
14389 * terminated before:
14390 * 1) we ran off the end of the pattern; or
14391 * 2) found two characters, each of which might have been intended to
14392 * be the name's terminator
14393 * 3) found so many punctuation characters in the purported name,
14394 * that the edit distance to a valid one is exceeded
14395 * 4) we decided it was more characters than anyone could have
14396 * intended to be one. */
14398 found_problem = TRUE;
14400 /* In the final two cases, we know that looking up what we've
14401 * accumulated won't lead to a match, even a fuzzy one. */
14402 if ( name_len >= C_ARRAY_LENGTH(input_text)
14403 || punct_count > max_distance)
14405 /* If there was an intermediate key character that could have been
14406 * an intended end, redo the parse, but stop there */
14407 if (possible_end && possible_end != (char *) -1) {
14408 possible_end = (char *) -1; /* Special signal value to say
14409 we've done a first pass */
14414 /* Otherwise, it can't have meant to have been a class */
14415 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14418 /* If we ran off the end, and the final character was a punctuation
14419 * one, back up one, to look at that final one just below. Later, we
14420 * will restore the parse pointer if appropriate */
14421 if (name_len && p == e && isPUNCT(*(p-1))) {
14426 if (p < e && isPUNCT(*p)) {
14428 has_terminating_bracket = TRUE;
14430 /* If this is a 2nd ']', and the first one is just below this
14431 * one, consider that to be the real terminator. This gives a
14432 * uniform and better positioning for the warning message */
14434 && possible_end != (char *) -1
14435 && *possible_end == ']'
14436 && name_len && input_text[name_len - 1] == ']')
14441 /* And this is actually equivalent to having done the 2nd
14442 * pass now, so set it to not try again */
14443 possible_end = (char *) -1;
14448 has_terminating_colon = TRUE;
14450 else if (*p == ';') {
14451 has_semi_colon = TRUE;
14452 has_terminating_colon = TRUE;
14460 /* Here, we have a class name to look up. We can short circuit the
14461 * stuff below for short names that can't possibly be meant to be a
14462 * class name. (We can do this on the first pass, as any second pass
14463 * will yield an even shorter name) */
14464 if (name_len < 3) {
14465 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14468 /* Find which class it is. Initially switch on the length of the name.
14470 switch (name_len) {
14472 if (memEQ(name_start, "word", 4)) {
14473 /* this is not POSIX, this is the Perl \w */
14474 class_number = ANYOF_WORDCHAR;
14478 /* Names all of length 5: alnum alpha ascii blank cntrl digit
14479 * graph lower print punct space upper
14480 * Offset 4 gives the best switch position. */
14481 switch (name_start[4]) {
14483 if (memEQ(name_start, "alph", 4)) /* alpha */
14484 class_number = ANYOF_ALPHA;
14487 if (memEQ(name_start, "spac", 4)) /* space */
14488 class_number = ANYOF_SPACE;
14491 if (memEQ(name_start, "grap", 4)) /* graph */
14492 class_number = ANYOF_GRAPH;
14495 if (memEQ(name_start, "asci", 4)) /* ascii */
14496 class_number = ANYOF_ASCII;
14499 if (memEQ(name_start, "blan", 4)) /* blank */
14500 class_number = ANYOF_BLANK;
14503 if (memEQ(name_start, "cntr", 4)) /* cntrl */
14504 class_number = ANYOF_CNTRL;
14507 if (memEQ(name_start, "alnu", 4)) /* alnum */
14508 class_number = ANYOF_ALPHANUMERIC;
14511 if (memEQ(name_start, "lowe", 4)) /* lower */
14512 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
14513 else if (memEQ(name_start, "uppe", 4)) /* upper */
14514 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
14517 if (memEQ(name_start, "digi", 4)) /* digit */
14518 class_number = ANYOF_DIGIT;
14519 else if (memEQ(name_start, "prin", 4)) /* print */
14520 class_number = ANYOF_PRINT;
14521 else if (memEQ(name_start, "punc", 4)) /* punct */
14522 class_number = ANYOF_PUNCT;
14527 if (memEQ(name_start, "xdigit", 6))
14528 class_number = ANYOF_XDIGIT;
14532 /* If the name exactly matches a posix class name the class number will
14533 * here be set to it, and the input almost certainly was meant to be a
14534 * posix class, so we can skip further checking. If instead the syntax
14535 * is exactly correct, but the name isn't one of the legal ones, we
14536 * will return that as an error below. But if neither of these apply,
14537 * it could be that no posix class was intended at all, or that one
14538 * was, but there was a typo. We tease these apart by doing fuzzy
14539 * matching on the name */
14540 if (class_number == OOB_NAMEDCLASS && found_problem) {
14541 const UV posix_names[][6] = {
14542 { 'a', 'l', 'n', 'u', 'm' },
14543 { 'a', 'l', 'p', 'h', 'a' },
14544 { 'a', 's', 'c', 'i', 'i' },
14545 { 'b', 'l', 'a', 'n', 'k' },
14546 { 'c', 'n', 't', 'r', 'l' },
14547 { 'd', 'i', 'g', 'i', 't' },
14548 { 'g', 'r', 'a', 'p', 'h' },
14549 { 'l', 'o', 'w', 'e', 'r' },
14550 { 'p', 'r', 'i', 'n', 't' },
14551 { 'p', 'u', 'n', 'c', 't' },
14552 { 's', 'p', 'a', 'c', 'e' },
14553 { 'u', 'p', 'p', 'e', 'r' },
14554 { 'w', 'o', 'r', 'd' },
14555 { 'x', 'd', 'i', 'g', 'i', 't' }
14557 /* The names of the above all have added NULs to make them the same
14558 * size, so we need to also have the real lengths */
14559 const UV posix_name_lengths[] = {
14560 sizeof("alnum") - 1,
14561 sizeof("alpha") - 1,
14562 sizeof("ascii") - 1,
14563 sizeof("blank") - 1,
14564 sizeof("cntrl") - 1,
14565 sizeof("digit") - 1,
14566 sizeof("graph") - 1,
14567 sizeof("lower") - 1,
14568 sizeof("print") - 1,
14569 sizeof("punct") - 1,
14570 sizeof("space") - 1,
14571 sizeof("upper") - 1,
14572 sizeof("word") - 1,
14573 sizeof("xdigit")- 1
14576 int temp_max = max_distance; /* Use a temporary, so if we
14577 reparse, we haven't changed the
14580 /* Use a smaller max edit distance if we are missing one of the
14582 if ( has_opening_bracket + has_opening_colon < 2
14583 || has_terminating_bracket + has_terminating_colon < 2)
14588 /* See if the input name is close to a legal one */
14589 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
14591 /* Short circuit call if the lengths are too far apart to be
14593 if (abs( (int) (name_len - posix_name_lengths[i]))
14599 if (edit_distance(input_text,
14602 posix_name_lengths[i],
14606 { /* If it is close, it probably was intended to be a class */
14607 goto probably_meant_to_be;
14611 /* Here the input name is not close enough to a valid class name
14612 * for us to consider it to be intended to be a posix class. If
14613 * we haven't already done so, and the parse found a character that
14614 * could have been terminators for the name, but which we absorbed
14615 * as typos during the first pass, repeat the parse, signalling it
14616 * to stop at that character */
14617 if (possible_end && possible_end != (char *) -1) {
14618 possible_end = (char *) -1;
14623 /* Here neither pass found a close-enough class name */
14624 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14627 probably_meant_to_be:
14629 /* Here we think that a posix specification was intended. Update any
14631 if (updated_parse_ptr) {
14632 *updated_parse_ptr = (char *) p;
14635 /* If a posix class name was intended but incorrectly specified, we
14636 * output or return the warnings */
14637 if (found_problem) {
14639 /* We set flags for these issues in the parse loop above instead of
14640 * adding them to the list of warnings, because we can parse it
14641 * twice, and we only want one warning instance */
14643 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
14646 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14648 if (has_semi_colon) {
14649 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14651 else if (! has_terminating_colon) {
14652 ADD_POSIX_WARNING(p, "there is no terminating ':'");
14654 if (! has_terminating_bracket) {
14655 ADD_POSIX_WARNING(p, "there is no terminating ']'");
14658 if (posix_warnings && RExC_warn_text && av_top_index(RExC_warn_text) > -1) {
14659 *posix_warnings = RExC_warn_text;
14662 else if (class_number != OOB_NAMEDCLASS) {
14663 /* If it is a known class, return the class. The class number
14664 * #defines are structured so each complement is +1 to the normal
14666 return class_number + complement;
14668 else if (! check_only) {
14670 /* Here, it is an unrecognized class. This is an error (unless the
14671 * call is to check only, which we've already handled above) */
14672 const char * const complement_string = (complement)
14675 RExC_parse = (char *) p;
14676 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
14678 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
14682 return OOB_NAMEDCLASS;
14684 #undef ADD_POSIX_WARNING
14686 STATIC unsigned int
14687 S_regex_set_precedence(const U8 my_operator) {
14689 /* Returns the precedence in the (?[...]) construct of the input operator,
14690 * specified by its character representation. The precedence follows
14691 * general Perl rules, but it extends this so that ')' and ']' have (low)
14692 * precedence even though they aren't really operators */
14694 switch (my_operator) {
14710 NOT_REACHED; /* NOTREACHED */
14711 return 0; /* Silence compiler warning */
14715 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
14716 I32 *flagp, U32 depth,
14717 char * const oregcomp_parse)
14719 /* Handle the (?[...]) construct to do set operations */
14721 U8 curchar; /* Current character being parsed */
14722 UV start, end; /* End points of code point ranges */
14723 SV* final = NULL; /* The end result inversion list */
14724 SV* result_string; /* 'final' stringified */
14725 AV* stack; /* stack of operators and operands not yet
14727 AV* fence_stack = NULL; /* A stack containing the positions in
14728 'stack' of where the undealt-with left
14729 parens would be if they were actually
14731 /* The 'VOL' (expanding to 'volatile') is a workaround for an optimiser bug
14732 * in Solaris Studio 12.3. See RT #127455 */
14733 VOL IV fence = 0; /* Position of where most recent undealt-
14734 with left paren in stack is; -1 if none.
14736 STRLEN len; /* Temporary */
14737 regnode* node; /* Temporary, and final regnode returned by
14739 const bool save_fold = FOLD; /* Temporary */
14740 char *save_end, *save_parse; /* Temporaries */
14741 const bool in_locale = LOC; /* we turn off /l during processing */
14742 AV* posix_warnings = NULL;
14744 GET_RE_DEBUG_FLAGS_DECL;
14746 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
14749 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
14752 REQUIRE_UNI_RULES(flagp, NULL); /* The use of this operator implies /u.
14753 This is required so that the compile
14754 time values are valid in all runtime
14757 /* This will return only an ANYOF regnode, or (unlikely) something smaller
14758 * (such as EXACT). Thus we can skip most everything if just sizing. We
14759 * call regclass to handle '[]' so as to not have to reinvent its parsing
14760 * rules here (throwing away the size it computes each time). And, we exit
14761 * upon an unescaped ']' that isn't one ending a regclass. To do both
14762 * these things, we need to realize that something preceded by a backslash
14763 * is escaped, so we have to keep track of backslashes */
14765 UV depth = 0; /* how many nested (?[...]) constructs */
14767 while (RExC_parse < RExC_end) {
14768 SV* current = NULL;
14770 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14771 TRUE /* Force /x */ );
14773 switch (*RExC_parse) {
14775 if (RExC_parse[1] == '[') depth++, RExC_parse++;
14780 /* Skip past this, so the next character gets skipped, after
14783 if (*RExC_parse == 'c') {
14784 /* Skip the \cX notation for control characters */
14785 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14791 /* See if this is a [:posix:] class. */
14792 bool is_posix_class = (OOB_NAMEDCLASS
14793 < handle_possible_posix(pRExC_state,
14797 TRUE /* checking only */));
14798 /* If it is a posix class, leave the parse pointer at the
14799 * '[' to fool regclass() into thinking it is part of a
14800 * '[[:posix:]]'. */
14801 if (! is_posix_class) {
14805 /* regclass() can only return RESTART_PASS1 and NEED_UTF8
14806 * if multi-char folds are allowed. */
14807 if (!regclass(pRExC_state, flagp,depth+1,
14808 is_posix_class, /* parse the whole char
14809 class only if not a
14811 FALSE, /* don't allow multi-char folds */
14812 TRUE, /* silence non-portable warnings. */
14814 FALSE, /* Require return to be an ANYOF */
14818 FAIL2("panic: regclass returned NULL to handle_sets, "
14819 "flags=%#" UVxf, (UV) *flagp);
14821 /* function call leaves parse pointing to the ']', except
14822 * if we faked it */
14823 if (is_posix_class) {
14827 SvREFCNT_dec(current); /* In case it returned something */
14832 if (depth--) break;
14834 if (*RExC_parse == ')') {
14835 node = reganode(pRExC_state, ANYOF, 0);
14836 RExC_size += ANYOF_SKIP;
14837 nextchar(pRExC_state);
14838 Set_Node_Length(node,
14839 RExC_parse - oregcomp_parse + 1); /* MJD */
14841 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
14849 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14853 /* We output the messages even if warnings are off, because we'll fail
14854 * the very next thing, and these give a likely diagnosis for that */
14855 if (posix_warnings && av_tindex_nomg(posix_warnings) >= 0) {
14856 output_or_return_posix_warnings(pRExC_state, posix_warnings, NULL);
14859 FAIL("Syntax error in (?[...])");
14862 /* Pass 2 only after this. */
14863 Perl_ck_warner_d(aTHX_
14864 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
14865 "The regex_sets feature is experimental" REPORT_LOCATION,
14866 REPORT_LOCATION_ARGS(RExC_parse));
14868 /* Everything in this construct is a metacharacter. Operands begin with
14869 * either a '\' (for an escape sequence), or a '[' for a bracketed
14870 * character class. Any other character should be an operator, or
14871 * parenthesis for grouping. Both types of operands are handled by calling
14872 * regclass() to parse them. It is called with a parameter to indicate to
14873 * return the computed inversion list. The parsing here is implemented via
14874 * a stack. Each entry on the stack is a single character representing one
14875 * of the operators; or else a pointer to an operand inversion list. */
14877 #define IS_OPERATOR(a) SvIOK(a)
14878 #define IS_OPERAND(a) (! IS_OPERATOR(a))
14880 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
14881 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
14882 * with pronouncing it called it Reverse Polish instead, but now that YOU
14883 * know how to pronounce it you can use the correct term, thus giving due
14884 * credit to the person who invented it, and impressing your geek friends.
14885 * Wikipedia says that the pronounciation of "Ł" has been changing so that
14886 * it is now more like an English initial W (as in wonk) than an L.)
14888 * This means that, for example, 'a | b & c' is stored on the stack as
14896 * where the numbers in brackets give the stack [array] element number.
14897 * In this implementation, parentheses are not stored on the stack.
14898 * Instead a '(' creates a "fence" so that the part of the stack below the
14899 * fence is invisible except to the corresponding ')' (this allows us to
14900 * replace testing for parens, by using instead subtraction of the fence
14901 * position). As new operands are processed they are pushed onto the stack
14902 * (except as noted in the next paragraph). New operators of higher
14903 * precedence than the current final one are inserted on the stack before
14904 * the lhs operand (so that when the rhs is pushed next, everything will be
14905 * in the correct positions shown above. When an operator of equal or
14906 * lower precedence is encountered in parsing, all the stacked operations
14907 * of equal or higher precedence are evaluated, leaving the result as the
14908 * top entry on the stack. This makes higher precedence operations
14909 * evaluate before lower precedence ones, and causes operations of equal
14910 * precedence to left associate.
14912 * The only unary operator '!' is immediately pushed onto the stack when
14913 * encountered. When an operand is encountered, if the top of the stack is
14914 * a '!", the complement is immediately performed, and the '!' popped. The
14915 * resulting value is treated as a new operand, and the logic in the
14916 * previous paragraph is executed. Thus in the expression
14918 * the stack looks like
14924 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
14931 * A ')' is treated as an operator with lower precedence than all the
14932 * aforementioned ones, which causes all operations on the stack above the
14933 * corresponding '(' to be evaluated down to a single resultant operand.
14934 * Then the fence for the '(' is removed, and the operand goes through the
14935 * algorithm above, without the fence.
14937 * A separate stack is kept of the fence positions, so that the position of
14938 * the latest so-far unbalanced '(' is at the top of it.
14940 * The ']' ending the construct is treated as the lowest operator of all,
14941 * so that everything gets evaluated down to a single operand, which is the
14944 sv_2mortal((SV *)(stack = newAV()));
14945 sv_2mortal((SV *)(fence_stack = newAV()));
14947 while (RExC_parse < RExC_end) {
14948 I32 top_index; /* Index of top-most element in 'stack' */
14949 SV** top_ptr; /* Pointer to top 'stack' element */
14950 SV* current = NULL; /* To contain the current inversion list
14952 SV* only_to_avoid_leaks;
14954 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14955 TRUE /* Force /x */ );
14956 if (RExC_parse >= RExC_end) {
14957 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
14960 curchar = UCHARAT(RExC_parse);
14964 #ifdef ENABLE_REGEX_SETS_DEBUGGING
14965 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
14966 DEBUG_U(dump_regex_sets_structures(pRExC_state,
14967 stack, fence, fence_stack));
14970 top_index = av_tindex_nomg(stack);
14973 SV** stacked_ptr; /* Ptr to something already on 'stack' */
14974 char stacked_operator; /* The topmost operator on the 'stack'. */
14975 SV* lhs; /* Operand to the left of the operator */
14976 SV* rhs; /* Operand to the right of the operator */
14977 SV* fence_ptr; /* Pointer to top element of the fence
14982 if ( RExC_parse < RExC_end - 1
14983 && (UCHARAT(RExC_parse + 1) == '?'))
14985 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
14986 * This happens when we have some thing like
14988 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
14990 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
14992 * Here we would be handling the interpolated
14993 * '$thai_or_lao'. We handle this by a recursive call to
14994 * ourselves which returns the inversion list the
14995 * interpolated expression evaluates to. We use the flags
14996 * from the interpolated pattern. */
14997 U32 save_flags = RExC_flags;
14998 const char * save_parse;
15000 RExC_parse += 2; /* Skip past the '(?' */
15001 save_parse = RExC_parse;
15003 /* Parse any flags for the '(?' */
15004 parse_lparen_question_flags(pRExC_state);
15006 if (RExC_parse == save_parse /* Makes sure there was at
15007 least one flag (or else
15008 this embedding wasn't
15010 || RExC_parse >= RExC_end - 4
15011 || UCHARAT(RExC_parse) != ':'
15012 || UCHARAT(++RExC_parse) != '('
15013 || UCHARAT(++RExC_parse) != '?'
15014 || UCHARAT(++RExC_parse) != '[')
15017 /* In combination with the above, this moves the
15018 * pointer to the point just after the first erroneous
15019 * character (or if there are no flags, to where they
15020 * should have been) */
15021 if (RExC_parse >= RExC_end - 4) {
15022 RExC_parse = RExC_end;
15024 else if (RExC_parse != save_parse) {
15025 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15027 vFAIL("Expecting '(?flags:(?[...'");
15030 /* Recurse, with the meat of the embedded expression */
15032 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
15033 depth+1, oregcomp_parse);
15035 /* Here, 'current' contains the embedded expression's
15036 * inversion list, and RExC_parse points to the trailing
15037 * ']'; the next character should be the ')' */
15039 assert(UCHARAT(RExC_parse) == ')');
15041 /* Then the ')' matching the original '(' handled by this
15042 * case: statement */
15044 assert(UCHARAT(RExC_parse) == ')');
15047 RExC_flags = save_flags;
15048 goto handle_operand;
15051 /* A regular '('. Look behind for illegal syntax */
15052 if (top_index - fence >= 0) {
15053 /* If the top entry on the stack is an operator, it had
15054 * better be a '!', otherwise the entry below the top
15055 * operand should be an operator */
15056 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
15057 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
15058 || ( IS_OPERAND(*top_ptr)
15059 && ( top_index - fence < 1
15060 || ! (stacked_ptr = av_fetch(stack,
15063 || ! IS_OPERATOR(*stacked_ptr))))
15066 vFAIL("Unexpected '(' with no preceding operator");
15070 /* Stack the position of this undealt-with left paren */
15071 av_push(fence_stack, newSViv(fence));
15072 fence = top_index + 1;
15076 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15077 * multi-char folds are allowed. */
15078 if (!regclass(pRExC_state, flagp,depth+1,
15079 TRUE, /* means parse just the next thing */
15080 FALSE, /* don't allow multi-char folds */
15081 FALSE, /* don't silence non-portable warnings. */
15083 FALSE, /* Require return to be an ANYOF */
15087 FAIL2("panic: regclass returned NULL to handle_sets, "
15088 "flags=%#" UVxf, (UV) *flagp);
15091 /* regclass() will return with parsing just the \ sequence,
15092 * leaving the parse pointer at the next thing to parse */
15094 goto handle_operand;
15096 case '[': /* Is a bracketed character class */
15098 /* See if this is a [:posix:] class. */
15099 bool is_posix_class = (OOB_NAMEDCLASS
15100 < handle_possible_posix(pRExC_state,
15104 TRUE /* checking only */));
15105 /* If it is a posix class, leave the parse pointer at the '['
15106 * to fool regclass() into thinking it is part of a
15107 * '[[:posix:]]'. */
15108 if (! is_posix_class) {
15112 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15113 * multi-char folds are allowed. */
15114 if (!regclass(pRExC_state, flagp,depth+1,
15115 is_posix_class, /* parse the whole char
15116 class only if not a
15118 FALSE, /* don't allow multi-char folds */
15119 TRUE, /* silence non-portable warnings. */
15121 FALSE, /* Require return to be an ANYOF */
15126 FAIL2("panic: regclass returned NULL to handle_sets, "
15127 "flags=%#" UVxf, (UV) *flagp);
15130 /* function call leaves parse pointing to the ']', except if we
15132 if (is_posix_class) {
15136 goto handle_operand;
15140 if (top_index >= 1) {
15141 goto join_operators;
15144 /* Only a single operand on the stack: are done */
15148 if (av_tindex_nomg(fence_stack) < 0) {
15150 vFAIL("Unexpected ')'");
15153 /* If nothing after the fence, is missing an operand */
15154 if (top_index - fence < 0) {
15158 /* If at least two things on the stack, treat this as an
15160 if (top_index - fence >= 1) {
15161 goto join_operators;
15164 /* Here only a single thing on the fenced stack, and there is a
15165 * fence. Get rid of it */
15166 fence_ptr = av_pop(fence_stack);
15168 fence = SvIV(fence_ptr) - 1;
15169 SvREFCNT_dec_NN(fence_ptr);
15176 /* Having gotten rid of the fence, we pop the operand at the
15177 * stack top and process it as a newly encountered operand */
15178 current = av_pop(stack);
15179 if (IS_OPERAND(current)) {
15180 goto handle_operand;
15192 /* These binary operators should have a left operand already
15194 if ( top_index - fence < 0
15195 || top_index - fence == 1
15196 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
15197 || ! IS_OPERAND(*top_ptr))
15199 goto unexpected_binary;
15202 /* If only the one operand is on the part of the stack visible
15203 * to us, we just place this operator in the proper position */
15204 if (top_index - fence < 2) {
15206 /* Place the operator before the operand */
15208 SV* lhs = av_pop(stack);
15209 av_push(stack, newSVuv(curchar));
15210 av_push(stack, lhs);
15214 /* But if there is something else on the stack, we need to
15215 * process it before this new operator if and only if the
15216 * stacked operation has equal or higher precedence than the
15221 /* The operator on the stack is supposed to be below both its
15223 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
15224 || IS_OPERAND(*stacked_ptr))
15226 /* But if not, it's legal and indicates we are completely
15227 * done if and only if we're currently processing a ']',
15228 * which should be the final thing in the expression */
15229 if (curchar == ']') {
15235 vFAIL2("Unexpected binary operator '%c' with no "
15236 "preceding operand", curchar);
15238 stacked_operator = (char) SvUV(*stacked_ptr);
15240 if (regex_set_precedence(curchar)
15241 > regex_set_precedence(stacked_operator))
15243 /* Here, the new operator has higher precedence than the
15244 * stacked one. This means we need to add the new one to
15245 * the stack to await its rhs operand (and maybe more
15246 * stuff). We put it before the lhs operand, leaving
15247 * untouched the stacked operator and everything below it
15249 lhs = av_pop(stack);
15250 assert(IS_OPERAND(lhs));
15252 av_push(stack, newSVuv(curchar));
15253 av_push(stack, lhs);
15257 /* Here, the new operator has equal or lower precedence than
15258 * what's already there. This means the operation already
15259 * there should be performed now, before the new one. */
15261 rhs = av_pop(stack);
15262 if (! IS_OPERAND(rhs)) {
15264 /* This can happen when a ! is not followed by an operand,
15265 * like in /(?[\t &!])/ */
15269 lhs = av_pop(stack);
15271 if (! IS_OPERAND(lhs)) {
15273 /* This can happen when there is an empty (), like in
15274 * /(?[[0]+()+])/ */
15278 switch (stacked_operator) {
15280 _invlist_intersection(lhs, rhs, &rhs);
15285 _invlist_union(lhs, rhs, &rhs);
15289 _invlist_subtract(lhs, rhs, &rhs);
15292 case '^': /* The union minus the intersection */
15297 _invlist_union(lhs, rhs, &u);
15298 _invlist_intersection(lhs, rhs, &i);
15299 _invlist_subtract(u, i, &rhs);
15300 SvREFCNT_dec_NN(i);
15301 SvREFCNT_dec_NN(u);
15307 /* Here, the higher precedence operation has been done, and the
15308 * result is in 'rhs'. We overwrite the stacked operator with
15309 * the result. Then we redo this code to either push the new
15310 * operator onto the stack or perform any higher precedence
15311 * stacked operation */
15312 only_to_avoid_leaks = av_pop(stack);
15313 SvREFCNT_dec(only_to_avoid_leaks);
15314 av_push(stack, rhs);
15317 case '!': /* Highest priority, right associative */
15319 /* If what's already at the top of the stack is another '!",
15320 * they just cancel each other out */
15321 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
15322 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
15324 only_to_avoid_leaks = av_pop(stack);
15325 SvREFCNT_dec(only_to_avoid_leaks);
15327 else { /* Otherwise, since it's right associative, just push
15329 av_push(stack, newSVuv(curchar));
15334 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15335 vFAIL("Unexpected character");
15339 /* Here 'current' is the operand. If something is already on the
15340 * stack, we have to check if it is a !. But first, the code above
15341 * may have altered the stack in the time since we earlier set
15344 top_index = av_tindex_nomg(stack);
15345 if (top_index - fence >= 0) {
15346 /* If the top entry on the stack is an operator, it had better
15347 * be a '!', otherwise the entry below the top operand should
15348 * be an operator */
15349 top_ptr = av_fetch(stack, top_index, FALSE);
15351 if (IS_OPERATOR(*top_ptr)) {
15353 /* The only permissible operator at the top of the stack is
15354 * '!', which is applied immediately to this operand. */
15355 curchar = (char) SvUV(*top_ptr);
15356 if (curchar != '!') {
15357 SvREFCNT_dec(current);
15358 vFAIL2("Unexpected binary operator '%c' with no "
15359 "preceding operand", curchar);
15362 _invlist_invert(current);
15364 only_to_avoid_leaks = av_pop(stack);
15365 SvREFCNT_dec(only_to_avoid_leaks);
15367 /* And we redo with the inverted operand. This allows
15368 * handling multiple ! in a row */
15369 goto handle_operand;
15371 /* Single operand is ok only for the non-binary ')'
15373 else if ((top_index - fence == 0 && curchar != ')')
15374 || (top_index - fence > 0
15375 && (! (stacked_ptr = av_fetch(stack,
15378 || IS_OPERAND(*stacked_ptr))))
15380 SvREFCNT_dec(current);
15381 vFAIL("Operand with no preceding operator");
15385 /* Here there was nothing on the stack or the top element was
15386 * another operand. Just add this new one */
15387 av_push(stack, current);
15389 } /* End of switch on next parse token */
15391 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15392 } /* End of loop parsing through the construct */
15395 if (av_tindex_nomg(fence_stack) >= 0) {
15396 vFAIL("Unmatched (");
15399 if (av_tindex_nomg(stack) < 0 /* Was empty */
15400 || ((final = av_pop(stack)) == NULL)
15401 || ! IS_OPERAND(final)
15402 || SvTYPE(final) != SVt_INVLIST
15403 || av_tindex_nomg(stack) >= 0) /* More left on stack */
15406 SvREFCNT_dec(final);
15407 vFAIL("Incomplete expression within '(?[ ])'");
15410 /* Here, 'final' is the resultant inversion list from evaluating the
15411 * expression. Return it if so requested */
15412 if (return_invlist) {
15413 *return_invlist = final;
15417 /* Otherwise generate a resultant node, based on 'final'. regclass() is
15418 * expecting a string of ranges and individual code points */
15419 invlist_iterinit(final);
15420 result_string = newSVpvs("");
15421 while (invlist_iternext(final, &start, &end)) {
15422 if (start == end) {
15423 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
15426 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%" UVXf "}",
15431 /* About to generate an ANYOF (or similar) node from the inversion list we
15432 * have calculated */
15433 save_parse = RExC_parse;
15434 RExC_parse = SvPV(result_string, len);
15435 save_end = RExC_end;
15436 RExC_end = RExC_parse + len;
15438 /* We turn off folding around the call, as the class we have constructed
15439 * already has all folding taken into consideration, and we don't want
15440 * regclass() to add to that */
15441 RExC_flags &= ~RXf_PMf_FOLD;
15442 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if multi-char
15443 * folds are allowed. */
15444 node = regclass(pRExC_state, flagp,depth+1,
15445 FALSE, /* means parse the whole char class */
15446 FALSE, /* don't allow multi-char folds */
15447 TRUE, /* silence non-portable warnings. The above may very
15448 well have generated non-portable code points, but
15449 they're valid on this machine */
15450 FALSE, /* similarly, no need for strict */
15451 FALSE, /* Require return to be an ANYOF */
15456 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#" UVxf,
15459 /* Fix up the node type if we are in locale. (We have pretended we are
15460 * under /u for the purposes of regclass(), as this construct will only
15461 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
15462 * as to cause any warnings about bad locales to be output in regexec.c),
15463 * and add the flag that indicates to check if not in a UTF-8 locale. The
15464 * reason we above forbid optimization into something other than an ANYOF
15465 * node is simply to minimize the number of code changes in regexec.c.
15466 * Otherwise we would have to create new EXACTish node types and deal with
15467 * them. This decision could be revisited should this construct become
15470 * (One might think we could look at the resulting ANYOF node and suppress
15471 * the flag if everything is above 255, as those would be UTF-8 only,
15472 * but this isn't true, as the components that led to that result could
15473 * have been locale-affected, and just happen to cancel each other out
15474 * under UTF-8 locales.) */
15476 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
15478 assert(OP(node) == ANYOF);
15482 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
15486 RExC_flags |= RXf_PMf_FOLD;
15489 RExC_parse = save_parse + 1;
15490 RExC_end = save_end;
15491 SvREFCNT_dec_NN(final);
15492 SvREFCNT_dec_NN(result_string);
15494 nextchar(pRExC_state);
15495 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
15499 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15502 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
15503 AV * stack, const IV fence, AV * fence_stack)
15504 { /* Dumps the stacks in handle_regex_sets() */
15506 const SSize_t stack_top = av_tindex_nomg(stack);
15507 const SSize_t fence_stack_top = av_tindex_nomg(fence_stack);
15510 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
15512 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
15514 if (stack_top < 0) {
15515 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
15518 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
15519 for (i = stack_top; i >= 0; i--) {
15520 SV ** element_ptr = av_fetch(stack, i, FALSE);
15521 if (! element_ptr) {
15524 if (IS_OPERATOR(*element_ptr)) {
15525 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
15526 (int) i, (int) SvIV(*element_ptr));
15529 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
15530 sv_dump(*element_ptr);
15535 if (fence_stack_top < 0) {
15536 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
15539 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
15540 for (i = fence_stack_top; i >= 0; i--) {
15541 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
15542 if (! element_ptr) {
15545 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
15546 (int) i, (int) SvIV(*element_ptr));
15557 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
15559 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
15560 * innocent-looking character class, like /[ks]/i won't have to go out to
15561 * disk to find the possible matches.
15563 * This should be called only for a Latin1-range code points, cp, which is
15564 * known to be involved in a simple fold with other code points above
15565 * Latin1. It would give false results if /aa has been specified.
15566 * Multi-char folds are outside the scope of this, and must be handled
15569 * XXX It would be better to generate these via regen, in case a new
15570 * version of the Unicode standard adds new mappings, though that is not
15571 * really likely, and may be caught by the default: case of the switch
15574 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
15576 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
15582 add_cp_to_invlist(*invlist, KELVIN_SIGN);
15586 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
15589 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
15590 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
15592 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
15593 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
15594 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
15596 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
15597 *invlist = add_cp_to_invlist(*invlist,
15598 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
15601 #ifdef LATIN_CAPITAL_LETTER_SHARP_S /* not defined in early Unicode releases */
15603 case LATIN_SMALL_LETTER_SHARP_S:
15604 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
15609 #if UNICODE_MAJOR_VERSION < 3 \
15610 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0)
15612 /* In 3.0 and earlier, U+0130 folded simply to 'i'; and in 3.0.1 so did
15617 add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
15618 # if UNICODE_DOT_DOT_VERSION == 1
15619 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_DOTLESS_I);
15625 /* Use deprecated warning to increase the chances of this being
15628 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
15635 S_output_or_return_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings, AV** return_posix_warnings)
15637 /* If the final parameter is NULL, output the elements of the array given
15638 * by '*posix_warnings' as REGEXP warnings. Otherwise, the elements are
15639 * pushed onto it, (creating if necessary) */
15642 const bool first_is_fatal = ! return_posix_warnings
15643 && ckDEAD(packWARN(WARN_REGEXP));
15645 PERL_ARGS_ASSERT_OUTPUT_OR_RETURN_POSIX_WARNINGS;
15647 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
15648 if (return_posix_warnings) {
15649 if (! *return_posix_warnings) { /* mortalize to not leak if
15650 warnings are fatal */
15651 *return_posix_warnings = (AV *) sv_2mortal((SV *) newAV());
15653 av_push(*return_posix_warnings, msg);
15656 if (first_is_fatal) { /* Avoid leaking this */
15657 av_undef(posix_warnings); /* This isn't necessary if the
15658 array is mortal, but is a
15660 (void) sv_2mortal(msg);
15662 SAVEFREESV(RExC_rx_sv);
15665 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
15666 SvREFCNT_dec_NN(msg);
15672 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
15674 /* This adds the string scalar <multi_string> to the array
15675 * <multi_char_matches>. <multi_string> is known to have exactly
15676 * <cp_count> code points in it. This is used when constructing a
15677 * bracketed character class and we find something that needs to match more
15678 * than a single character.
15680 * <multi_char_matches> is actually an array of arrays. Each top-level
15681 * element is an array that contains all the strings known so far that are
15682 * the same length. And that length (in number of code points) is the same
15683 * as the index of the top-level array. Hence, the [2] element is an
15684 * array, each element thereof is a string containing TWO code points;
15685 * while element [3] is for strings of THREE characters, and so on. Since
15686 * this is for multi-char strings there can never be a [0] nor [1] element.
15688 * When we rewrite the character class below, we will do so such that the
15689 * longest strings are written first, so that it prefers the longest
15690 * matching strings first. This is done even if it turns out that any
15691 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
15692 * Christiansen has agreed that this is ok. This makes the test for the
15693 * ligature 'ffi' come before the test for 'ff', for example */
15696 AV** this_array_ptr;
15698 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
15700 if (! multi_char_matches) {
15701 multi_char_matches = newAV();
15704 if (av_exists(multi_char_matches, cp_count)) {
15705 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
15706 this_array = *this_array_ptr;
15709 this_array = newAV();
15710 av_store(multi_char_matches, cp_count,
15713 av_push(this_array, multi_string);
15715 return multi_char_matches;
15718 /* The names of properties whose definitions are not known at compile time are
15719 * stored in this SV, after a constant heading. So if the length has been
15720 * changed since initialization, then there is a run-time definition. */
15721 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
15722 (SvCUR(listsv) != initial_listsv_len)
15724 /* There is a restricted set of white space characters that are legal when
15725 * ignoring white space in a bracketed character class. This generates the
15726 * code to skip them.
15728 * There is a line below that uses the same white space criteria but is outside
15729 * this macro. Both here and there must use the same definition */
15730 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
15733 while (isBLANK_A(UCHARAT(p))) \
15741 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
15742 const bool stop_at_1, /* Just parse the next thing, don't
15743 look for a full character class */
15744 bool allow_multi_folds,
15745 const bool silence_non_portable, /* Don't output warnings
15749 bool optimizable, /* ? Allow a non-ANYOF return
15751 SV** ret_invlist, /* Return an inversion list, not a node */
15752 AV** return_posix_warnings
15755 /* parse a bracketed class specification. Most of these will produce an
15756 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
15757 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
15758 * under /i with multi-character folds: it will be rewritten following the
15759 * paradigm of this example, where the <multi-fold>s are characters which
15760 * fold to multiple character sequences:
15761 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
15762 * gets effectively rewritten as:
15763 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
15764 * reg() gets called (recursively) on the rewritten version, and this
15765 * function will return what it constructs. (Actually the <multi-fold>s
15766 * aren't physically removed from the [abcdefghi], it's just that they are
15767 * ignored in the recursion by means of a flag:
15768 * <RExC_in_multi_char_class>.)
15770 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
15771 * characters, with the corresponding bit set if that character is in the
15772 * list. For characters above this, a range list or swash is used. There
15773 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
15774 * determinable at compile time
15776 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs
15777 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded
15778 * to UTF-8. This can only happen if ret_invlist is non-NULL.
15781 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
15783 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
15786 int namedclass = OOB_NAMEDCLASS;
15787 char *rangebegin = NULL;
15788 bool need_class = 0;
15790 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
15791 than just initialized. */
15792 SV* properties = NULL; /* Code points that match \p{} \P{} */
15793 SV* posixes = NULL; /* Code points that match classes like [:word:],
15794 extended beyond the Latin1 range. These have to
15795 be kept separate from other code points for much
15796 of this function because their handling is
15797 different under /i, and for most classes under
15799 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
15800 separate for a while from the non-complemented
15801 versions because of complications with /d
15803 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
15804 treated more simply than the general case,
15805 leading to less compilation and execution
15807 UV element_count = 0; /* Number of distinct elements in the class.
15808 Optimizations may be possible if this is tiny */
15809 AV * multi_char_matches = NULL; /* Code points that fold to more than one
15810 character; used under /i */
15812 char * stop_ptr = RExC_end; /* where to stop parsing */
15814 /* ignore unescaped whitespace? */
15815 const bool skip_white = cBOOL( ret_invlist
15816 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
15818 /* Unicode properties are stored in a swash; this holds the current one
15819 * being parsed. If this swash is the only above-latin1 component of the
15820 * character class, an optimization is to pass it directly on to the
15821 * execution engine. Otherwise, it is set to NULL to indicate that there
15822 * are other things in the class that have to be dealt with at execution
15824 SV* swash = NULL; /* Code points that match \p{} \P{} */
15826 /* Set if a component of this character class is user-defined; just passed
15827 * on to the engine */
15828 bool has_user_defined_property = FALSE;
15830 /* inversion list of code points this node matches only when the target
15831 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
15833 SV* has_upper_latin1_only_utf8_matches = NULL;
15835 /* Inversion list of code points this node matches regardless of things
15836 * like locale, folding, utf8ness of the target string */
15837 SV* cp_list = NULL;
15839 /* Like cp_list, but code points on this list need to be checked for things
15840 * that fold to/from them under /i */
15841 SV* cp_foldable_list = NULL;
15843 /* Like cp_list, but code points on this list are valid only when the
15844 * runtime locale is UTF-8 */
15845 SV* only_utf8_locale_list = NULL;
15847 /* In a range, if one of the endpoints is non-character-set portable,
15848 * meaning that it hard-codes a code point that may mean a different
15849 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
15850 * mnemonic '\t' which each mean the same character no matter which
15851 * character set the platform is on. */
15852 unsigned int non_portable_endpoint = 0;
15854 /* Is the range unicode? which means on a platform that isn't 1-1 native
15855 * to Unicode (i.e. non-ASCII), each code point in it should be considered
15856 * to be a Unicode value. */
15857 bool unicode_range = FALSE;
15858 bool invert = FALSE; /* Is this class to be complemented */
15860 bool warn_super = ALWAYS_WARN_SUPER;
15862 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
15863 case we need to change the emitted regop to an EXACT. */
15864 const char * orig_parse = RExC_parse;
15865 const SSize_t orig_size = RExC_size;
15866 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
15868 /* This variable is used to mark where the end in the input is of something
15869 * that looks like a POSIX construct but isn't. During the parse, when
15870 * something looks like it could be such a construct is encountered, it is
15871 * checked for being one, but not if we've already checked this area of the
15872 * input. Only after this position is reached do we check again */
15873 char *not_posix_region_end = RExC_parse - 1;
15875 AV* posix_warnings = NULL;
15876 const bool do_posix_warnings = return_posix_warnings
15877 || (PASS2 && ckWARN(WARN_REGEXP));
15879 GET_RE_DEBUG_FLAGS_DECL;
15881 PERL_ARGS_ASSERT_REGCLASS;
15883 PERL_UNUSED_ARG(depth);
15886 DEBUG_PARSE("clas");
15888 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
15889 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
15890 && UNICODE_DOT_DOT_VERSION == 0)
15891 allow_multi_folds = FALSE;
15894 /* Assume we are going to generate an ANYOF node. */
15895 ret = reganode(pRExC_state,
15902 RExC_size += ANYOF_SKIP;
15903 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
15906 ANYOF_FLAGS(ret) = 0;
15908 RExC_emit += ANYOF_SKIP;
15909 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
15910 initial_listsv_len = SvCUR(listsv);
15911 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
15914 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15916 assert(RExC_parse <= RExC_end);
15918 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
15921 allow_multi_folds = FALSE;
15923 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15926 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
15927 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
15928 int maybe_class = handle_possible_posix(pRExC_state,
15930 ¬_posix_region_end,
15932 TRUE /* checking only */);
15933 if (PASS2 && maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
15934 SAVEFREESV(RExC_rx_sv);
15935 ckWARN4reg(not_posix_region_end,
15936 "POSIX syntax [%c %c] belongs inside character classes%s",
15937 *RExC_parse, *RExC_parse,
15938 (maybe_class == OOB_NAMEDCLASS)
15939 ? ((POSIXCC_NOTYET(*RExC_parse))
15940 ? " (but this one isn't implemented)"
15941 : " (but this one isn't fully valid)")
15944 (void)ReREFCNT_inc(RExC_rx_sv);
15948 /* If the caller wants us to just parse a single element, accomplish this
15949 * by faking the loop ending condition */
15950 if (stop_at_1 && RExC_end > RExC_parse) {
15951 stop_ptr = RExC_parse + 1;
15954 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
15955 if (UCHARAT(RExC_parse) == ']')
15956 goto charclassloop;
15960 if ( posix_warnings
15961 && av_tindex_nomg(posix_warnings) >= 0
15962 && RExC_parse > not_posix_region_end)
15964 /* Warnings about posix class issues are considered tentative until
15965 * we are far enough along in the parse that we can no longer
15966 * change our mind, at which point we either output them or add
15967 * them, if it has so specified, to what gets returned to the
15968 * caller. This is done each time through the loop so that a later
15969 * class won't zap them before they have been dealt with. */
15970 output_or_return_posix_warnings(pRExC_state, posix_warnings,
15971 return_posix_warnings);
15974 if (RExC_parse >= stop_ptr) {
15978 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15980 if (UCHARAT(RExC_parse) == ']') {
15986 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
15987 save_value = value;
15988 save_prevvalue = prevvalue;
15991 rangebegin = RExC_parse;
15993 non_portable_endpoint = 0;
15995 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
15996 value = utf8n_to_uvchr((U8*)RExC_parse,
15997 RExC_end - RExC_parse,
15998 &numlen, UTF8_ALLOW_DEFAULT);
15999 RExC_parse += numlen;
16002 value = UCHARAT(RExC_parse++);
16004 if (value == '[') {
16005 char * posix_class_end;
16006 namedclass = handle_possible_posix(pRExC_state,
16009 do_posix_warnings ? &posix_warnings : NULL,
16010 FALSE /* die if error */);
16011 if (namedclass > OOB_NAMEDCLASS) {
16013 /* If there was an earlier attempt to parse this particular
16014 * posix class, and it failed, it was a false alarm, as this
16015 * successful one proves */
16016 if ( posix_warnings
16017 && av_tindex_nomg(posix_warnings) >= 0
16018 && not_posix_region_end >= RExC_parse
16019 && not_posix_region_end <= posix_class_end)
16021 av_undef(posix_warnings);
16024 RExC_parse = posix_class_end;
16026 else if (namedclass == OOB_NAMEDCLASS) {
16027 not_posix_region_end = posix_class_end;
16030 namedclass = OOB_NAMEDCLASS;
16033 else if ( RExC_parse - 1 > not_posix_region_end
16034 && MAYBE_POSIXCC(value))
16036 (void) handle_possible_posix(
16038 RExC_parse - 1, /* -1 because parse has already been
16040 ¬_posix_region_end,
16041 do_posix_warnings ? &posix_warnings : NULL,
16042 TRUE /* checking only */);
16044 else if (value == '\\') {
16045 /* Is a backslash; get the code point of the char after it */
16047 if (RExC_parse >= RExC_end) {
16048 vFAIL("Unmatched [");
16051 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
16052 value = utf8n_to_uvchr((U8*)RExC_parse,
16053 RExC_end - RExC_parse,
16054 &numlen, UTF8_ALLOW_DEFAULT);
16055 RExC_parse += numlen;
16058 value = UCHARAT(RExC_parse++);
16060 /* Some compilers cannot handle switching on 64-bit integer
16061 * values, therefore value cannot be an UV. Yes, this will
16062 * be a problem later if we want switch on Unicode.
16063 * A similar issue a little bit later when switching on
16064 * namedclass. --jhi */
16066 /* If the \ is escaping white space when white space is being
16067 * skipped, it means that that white space is wanted literally, and
16068 * is already in 'value'. Otherwise, need to translate the escape
16069 * into what it signifies. */
16070 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
16072 case 'w': namedclass = ANYOF_WORDCHAR; break;
16073 case 'W': namedclass = ANYOF_NWORDCHAR; break;
16074 case 's': namedclass = ANYOF_SPACE; break;
16075 case 'S': namedclass = ANYOF_NSPACE; break;
16076 case 'd': namedclass = ANYOF_DIGIT; break;
16077 case 'D': namedclass = ANYOF_NDIGIT; break;
16078 case 'v': namedclass = ANYOF_VERTWS; break;
16079 case 'V': namedclass = ANYOF_NVERTWS; break;
16080 case 'h': namedclass = ANYOF_HORIZWS; break;
16081 case 'H': namedclass = ANYOF_NHORIZWS; break;
16082 case 'N': /* Handle \N{NAME} in class */
16084 const char * const backslash_N_beg = RExC_parse - 2;
16087 if (! grok_bslash_N(pRExC_state,
16088 NULL, /* No regnode */
16089 &value, /* Yes single value */
16090 &cp_count, /* Multiple code pt count */
16096 if (*flagp & NEED_UTF8)
16097 FAIL("panic: grok_bslash_N set NEED_UTF8");
16098 if (*flagp & RESTART_PASS1)
16101 if (cp_count < 0) {
16102 vFAIL("\\N in a character class must be a named character: \\N{...}");
16104 else if (cp_count == 0) {
16106 ckWARNreg(RExC_parse,
16107 "Ignoring zero length \\N{} in character class");
16110 else { /* cp_count > 1 */
16111 if (! RExC_in_multi_char_class) {
16112 if (invert || range || *RExC_parse == '-') {
16115 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
16118 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
16120 break; /* <value> contains the first code
16121 point. Drop out of the switch to
16125 SV * multi_char_N = newSVpvn(backslash_N_beg,
16126 RExC_parse - backslash_N_beg);
16128 = add_multi_match(multi_char_matches,
16133 } /* End of cp_count != 1 */
16135 /* This element should not be processed further in this
16138 value = save_value;
16139 prevvalue = save_prevvalue;
16140 continue; /* Back to top of loop to get next char */
16143 /* Here, is a single code point, and <value> contains it */
16144 unicode_range = TRUE; /* \N{} are Unicode */
16152 /* We will handle any undefined properties ourselves */
16153 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
16154 /* And we actually would prefer to get
16155 * the straight inversion list of the
16156 * swash, since we will be accessing it
16157 * anyway, to save a little time */
16158 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
16160 if (RExC_parse >= RExC_end)
16161 vFAIL2("Empty \\%c", (U8)value);
16162 if (*RExC_parse == '{') {
16163 const U8 c = (U8)value;
16164 e = strchr(RExC_parse, '}');
16167 vFAIL2("Missing right brace on \\%c{}", c);
16171 while (isSPACE(*RExC_parse)) {
16175 if (UCHARAT(RExC_parse) == '^') {
16177 /* toggle. (The rhs xor gets the single bit that
16178 * differs between P and p; the other xor inverts just
16180 value ^= 'P' ^ 'p';
16183 while (isSPACE(*RExC_parse)) {
16188 if (e == RExC_parse)
16189 vFAIL2("Empty \\%c{}", c);
16191 n = e - RExC_parse;
16192 while (isSPACE(*(RExC_parse + n - 1)))
16194 } /* The \p isn't immediately followed by a '{' */
16195 else if (! isALPHA(*RExC_parse)) {
16196 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16197 vFAIL2("Character following \\%c must be '{' or a "
16198 "single-character Unicode property name",
16208 char* base_name; /* name after any packages are stripped */
16209 char* lookup_name = NULL;
16210 const char * const colon_colon = "::";
16212 /* Try to get the definition of the property into
16213 * <invlist>. If /i is in effect, the effective property
16214 * will have its name be <__NAME_i>. The design is
16215 * discussed in commit
16216 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
16217 name = savepv(Perl_form(aTHX_ "%.*s", (int)n, RExC_parse));
16220 lookup_name = savepv(Perl_form(aTHX_ "__%s_i", name));
16222 /* The function call just below that uses this can fail
16223 * to return, leaking memory if we don't do this */
16224 SAVEFREEPV(lookup_name);
16227 /* Look up the property name, and get its swash and
16228 * inversion list, if the property is found */
16229 SvREFCNT_dec(swash); /* Free any left-overs */
16230 swash = _core_swash_init("utf8",
16237 NULL, /* No inversion list */
16240 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
16241 HV* curpkg = (IN_PERL_COMPILETIME)
16243 : CopSTASH(PL_curcop);
16247 if (swash) { /* Got a swash but no inversion list.
16248 Something is likely wrong that will
16249 be sorted-out later */
16250 SvREFCNT_dec_NN(swash);
16254 /* Here didn't find it. It could be a an error (like a
16255 * typo) in specifying a Unicode property, or it could
16256 * be a user-defined property that will be available at
16257 * run-time. The names of these must begin with 'In'
16258 * or 'Is' (after any packages are stripped off). So
16259 * if not one of those, or if we accept only
16260 * compile-time properties, is an error; otherwise add
16261 * it to the list for run-time look up. */
16262 if ((base_name = rninstr(name, name + n,
16263 colon_colon, colon_colon + 2)))
16264 { /* Has ::. We know this must be a user-defined
16267 final_n -= base_name - name;
16276 || base_name[0] != 'I'
16277 || (base_name[1] != 's' && base_name[1] != 'n')
16280 const char * const msg
16282 ? "Illegal user-defined property name"
16283 : "Can't find Unicode property definition";
16284 RExC_parse = e + 1;
16286 /* diag_listed_as: Can't find Unicode property definition "%s" */
16287 vFAIL3utf8f("%s \"%" UTF8f "\"",
16288 msg, UTF8fARG(UTF, n, name));
16291 /* If the property name doesn't already have a package
16292 * name, add the current one to it so that it can be
16293 * referred to outside it. [perl #121777] */
16294 if (! has_pkg && curpkg) {
16295 char* pkgname = HvNAME(curpkg);
16296 if (strNE(pkgname, "main")) {
16297 char* full_name = Perl_form(aTHX_
16301 n = strlen(full_name);
16302 name = savepvn(full_name, n);
16306 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s%" UTF8f "%s\n",
16307 (value == 'p' ? '+' : '!'),
16308 (FOLD) ? "__" : "",
16309 UTF8fARG(UTF, n, name),
16310 (FOLD) ? "_i" : "");
16311 has_user_defined_property = TRUE;
16312 optimizable = FALSE; /* Will have to leave this an
16315 /* We don't know yet what this matches, so have to flag
16317 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
16321 /* Here, did get the swash and its inversion list. If
16322 * the swash is from a user-defined property, then this
16323 * whole character class should be regarded as such */
16324 if (swash_init_flags
16325 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
16327 has_user_defined_property = TRUE;
16330 /* We warn on matching an above-Unicode code point
16331 * if the match would return true, except don't
16332 * warn for \p{All}, which has exactly one element
16334 (_invlist_contains_cp(invlist, 0x110000)
16335 && (! (_invlist_len(invlist) == 1
16336 && *invlist_array(invlist) == 0)))
16342 /* Invert if asking for the complement */
16343 if (value == 'P') {
16344 _invlist_union_complement_2nd(properties,
16348 /* The swash can't be used as-is, because we've
16349 * inverted things; delay removing it to here after
16350 * have copied its invlist above */
16351 SvREFCNT_dec_NN(swash);
16355 _invlist_union(properties, invlist, &properties);
16359 RExC_parse = e + 1;
16360 namedclass = ANYOF_UNIPROP; /* no official name, but it's
16363 /* \p means they want Unicode semantics */
16364 REQUIRE_UNI_RULES(flagp, NULL);
16367 case 'n': value = '\n'; break;
16368 case 'r': value = '\r'; break;
16369 case 't': value = '\t'; break;
16370 case 'f': value = '\f'; break;
16371 case 'b': value = '\b'; break;
16372 case 'e': value = ESC_NATIVE; break;
16373 case 'a': value = '\a'; break;
16375 RExC_parse--; /* function expects to be pointed at the 'o' */
16377 const char* error_msg;
16378 bool valid = grok_bslash_o(&RExC_parse,
16381 PASS2, /* warnings only in
16384 silence_non_portable,
16390 non_portable_endpoint++;
16393 RExC_parse--; /* function expects to be pointed at the 'x' */
16395 const char* error_msg;
16396 bool valid = grok_bslash_x(&RExC_parse,
16399 PASS2, /* Output warnings */
16401 silence_non_portable,
16407 non_portable_endpoint++;
16410 value = grok_bslash_c(*RExC_parse++, PASS2);
16411 non_portable_endpoint++;
16413 case '0': case '1': case '2': case '3': case '4':
16414 case '5': case '6': case '7':
16416 /* Take 1-3 octal digits */
16417 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
16418 numlen = (strict) ? 4 : 3;
16419 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
16420 RExC_parse += numlen;
16423 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16424 vFAIL("Need exactly 3 octal digits");
16426 else if (! SIZE_ONLY /* like \08, \178 */
16428 && RExC_parse < RExC_end
16429 && isDIGIT(*RExC_parse)
16430 && ckWARN(WARN_REGEXP))
16432 SAVEFREESV(RExC_rx_sv);
16433 reg_warn_non_literal_string(
16435 form_short_octal_warning(RExC_parse, numlen));
16436 (void)ReREFCNT_inc(RExC_rx_sv);
16439 non_portable_endpoint++;
16443 /* Allow \_ to not give an error */
16444 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
16446 vFAIL2("Unrecognized escape \\%c in character class",
16450 SAVEFREESV(RExC_rx_sv);
16451 ckWARN2reg(RExC_parse,
16452 "Unrecognized escape \\%c in character class passed through",
16454 (void)ReREFCNT_inc(RExC_rx_sv);
16458 } /* End of switch on char following backslash */
16459 } /* end of handling backslash escape sequences */
16461 /* Here, we have the current token in 'value' */
16463 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
16466 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
16467 * literal, as is the character that began the false range, i.e.
16468 * the 'a' in the examples */
16471 const int w = (RExC_parse >= rangebegin)
16472 ? RExC_parse - rangebegin
16476 "False [] range \"%" UTF8f "\"",
16477 UTF8fARG(UTF, w, rangebegin));
16480 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
16481 ckWARN2reg(RExC_parse,
16482 "False [] range \"%" UTF8f "\"",
16483 UTF8fARG(UTF, w, rangebegin));
16484 (void)ReREFCNT_inc(RExC_rx_sv);
16485 cp_list = add_cp_to_invlist(cp_list, '-');
16486 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
16491 range = 0; /* this was not a true range */
16492 element_count += 2; /* So counts for three values */
16495 classnum = namedclass_to_classnum(namedclass);
16497 if (LOC && namedclass < ANYOF_POSIXL_MAX
16498 #ifndef HAS_ISASCII
16499 && classnum != _CC_ASCII
16502 /* What the Posix classes (like \w, [:space:]) match in locale
16503 * isn't knowable under locale until actual match time. Room
16504 * must be reserved (one time per outer bracketed class) to
16505 * store such classes. The space will contain a bit for each
16506 * named class that is to be matched against. This isn't
16507 * needed for \p{} and pseudo-classes, as they are not affected
16508 * by locale, and hence are dealt with separately */
16509 if (! need_class) {
16512 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16515 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16517 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
16518 ANYOF_POSIXL_ZERO(ret);
16520 /* We can't change this into some other type of node
16521 * (unless this is the only element, in which case there
16522 * are nodes that mean exactly this) as has runtime
16524 optimizable = FALSE;
16527 /* Coverity thinks it is possible for this to be negative; both
16528 * jhi and khw think it's not, but be safer */
16529 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16530 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
16532 /* See if it already matches the complement of this POSIX
16534 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16535 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
16539 posixl_matches_all = TRUE;
16540 break; /* No need to continue. Since it matches both
16541 e.g., \w and \W, it matches everything, and the
16542 bracketed class can be optimized into qr/./s */
16545 /* Add this class to those that should be checked at runtime */
16546 ANYOF_POSIXL_SET(ret, namedclass);
16548 /* The above-Latin1 characters are not subject to locale rules.
16549 * Just add them, in the second pass, to the
16550 * unconditionally-matched list */
16552 SV* scratch_list = NULL;
16554 /* Get the list of the above-Latin1 code points this
16556 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
16557 PL_XPosix_ptrs[classnum],
16559 /* Odd numbers are complements, like
16560 * NDIGIT, NASCII, ... */
16561 namedclass % 2 != 0,
16563 /* Checking if 'cp_list' is NULL first saves an extra
16564 * clone. Its reference count will be decremented at the
16565 * next union, etc, or if this is the only instance, at the
16566 * end of the routine */
16568 cp_list = scratch_list;
16571 _invlist_union(cp_list, scratch_list, &cp_list);
16572 SvREFCNT_dec_NN(scratch_list);
16574 continue; /* Go get next character */
16577 else if (! SIZE_ONLY) {
16579 /* Here, not in pass1 (in that pass we skip calculating the
16580 * contents of this class), and is not /l, or is a POSIX class
16581 * for which /l doesn't matter (or is a Unicode property, which
16582 * is skipped here). */
16583 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
16584 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
16586 /* Here, should be \h, \H, \v, or \V. None of /d, /i
16587 * nor /l make a difference in what these match,
16588 * therefore we just add what they match to cp_list. */
16589 if (classnum != _CC_VERTSPACE) {
16590 assert( namedclass == ANYOF_HORIZWS
16591 || namedclass == ANYOF_NHORIZWS);
16593 /* It turns out that \h is just a synonym for
16595 classnum = _CC_BLANK;
16598 _invlist_union_maybe_complement_2nd(
16600 PL_XPosix_ptrs[classnum],
16601 namedclass % 2 != 0, /* Complement if odd
16602 (NHORIZWS, NVERTWS)
16607 else if ( UNI_SEMANTICS
16608 || classnum == _CC_ASCII
16609 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
16610 || classnum == _CC_XDIGIT)))
16612 /* We usually have to worry about /d and /a affecting what
16613 * POSIX classes match, with special code needed for /d
16614 * because we won't know until runtime what all matches.
16615 * But there is no extra work needed under /u, and
16616 * [:ascii:] is unaffected by /a and /d; and :digit: and
16617 * :xdigit: don't have runtime differences under /d. So we
16618 * can special case these, and avoid some extra work below,
16619 * and at runtime. */
16620 _invlist_union_maybe_complement_2nd(
16622 PL_XPosix_ptrs[classnum],
16623 namedclass % 2 != 0,
16626 else { /* Garden variety class. If is NUPPER, NALPHA, ...
16627 complement and use nposixes */
16628 SV** posixes_ptr = namedclass % 2 == 0
16631 _invlist_union_maybe_complement_2nd(
16633 PL_XPosix_ptrs[classnum],
16634 namedclass % 2 != 0,
16638 } /* end of namedclass \blah */
16640 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16642 /* If 'range' is set, 'value' is the ending of a range--check its
16643 * validity. (If value isn't a single code point in the case of a
16644 * range, we should have figured that out above in the code that
16645 * catches false ranges). Later, we will handle each individual code
16646 * point in the range. If 'range' isn't set, this could be the
16647 * beginning of a range, so check for that by looking ahead to see if
16648 * the next real character to be processed is the range indicator--the
16653 /* For unicode ranges, we have to test that the Unicode as opposed
16654 * to the native values are not decreasing. (Above 255, there is
16655 * no difference between native and Unicode) */
16656 if (unicode_range && prevvalue < 255 && value < 255) {
16657 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
16658 goto backwards_range;
16663 if (prevvalue > value) /* b-a */ {
16668 w = RExC_parse - rangebegin;
16670 "Invalid [] range \"%" UTF8f "\"",
16671 UTF8fARG(UTF, w, rangebegin));
16672 NOT_REACHED; /* NOTREACHED */
16676 prevvalue = value; /* save the beginning of the potential range */
16677 if (! stop_at_1 /* Can't be a range if parsing just one thing */
16678 && *RExC_parse == '-')
16680 char* next_char_ptr = RExC_parse + 1;
16682 /* Get the next real char after the '-' */
16683 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
16685 /* If the '-' is at the end of the class (just before the ']',
16686 * it is a literal minus; otherwise it is a range */
16687 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
16688 RExC_parse = next_char_ptr;
16690 /* a bad range like \w-, [:word:]- ? */
16691 if (namedclass > OOB_NAMEDCLASS) {
16692 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
16693 const int w = RExC_parse >= rangebegin
16694 ? RExC_parse - rangebegin
16697 vFAIL4("False [] range \"%*.*s\"",
16702 "False [] range \"%*.*s\"",
16707 cp_list = add_cp_to_invlist(cp_list, '-');
16711 range = 1; /* yeah, it's a range! */
16712 continue; /* but do it the next time */
16717 if (namedclass > OOB_NAMEDCLASS) {
16721 /* Here, we have a single value this time through the loop, and
16722 * <prevvalue> is the beginning of the range, if any; or <value> if
16725 /* non-Latin1 code point implies unicode semantics. Must be set in
16726 * pass1 so is there for the whole of pass 2 */
16728 REQUIRE_UNI_RULES(flagp, NULL);
16731 /* Ready to process either the single value, or the completed range.
16732 * For single-valued non-inverted ranges, we consider the possibility
16733 * of multi-char folds. (We made a conscious decision to not do this
16734 * for the other cases because it can often lead to non-intuitive
16735 * results. For example, you have the peculiar case that:
16736 * "s s" =~ /^[^\xDF]+$/i => Y
16737 * "ss" =~ /^[^\xDF]+$/i => N
16739 * See [perl #89750] */
16740 if (FOLD && allow_multi_folds && value == prevvalue) {
16741 if (value == LATIN_SMALL_LETTER_SHARP_S
16742 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
16745 /* Here <value> is indeed a multi-char fold. Get what it is */
16747 U8 foldbuf[UTF8_MAXBYTES_CASE];
16750 UV folded = _to_uni_fold_flags(
16754 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
16755 ? FOLD_FLAGS_NOMIX_ASCII
16759 /* Here, <folded> should be the first character of the
16760 * multi-char fold of <value>, with <foldbuf> containing the
16761 * whole thing. But, if this fold is not allowed (because of
16762 * the flags), <fold> will be the same as <value>, and should
16763 * be processed like any other character, so skip the special
16765 if (folded != value) {
16767 /* Skip if we are recursed, currently parsing the class
16768 * again. Otherwise add this character to the list of
16769 * multi-char folds. */
16770 if (! RExC_in_multi_char_class) {
16771 STRLEN cp_count = utf8_length(foldbuf,
16772 foldbuf + foldlen);
16773 SV* multi_fold = sv_2mortal(newSVpvs(""));
16775 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
16778 = add_multi_match(multi_char_matches,
16784 /* This element should not be processed further in this
16787 value = save_value;
16788 prevvalue = save_prevvalue;
16794 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
16797 /* If the range starts above 255, everything is portable and
16798 * likely to be so for any forseeable character set, so don't
16800 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
16801 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
16803 else if (prevvalue != value) {
16805 /* Under strict, ranges that stop and/or end in an ASCII
16806 * printable should have each end point be a portable value
16807 * for it (preferably like 'A', but we don't warn if it is
16808 * a (portable) Unicode name or code point), and the range
16809 * must be be all digits or all letters of the same case.
16810 * Otherwise, the range is non-portable and unclear as to
16811 * what it contains */
16812 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
16813 && ( non_portable_endpoint
16814 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
16815 || (isLOWER_A(prevvalue) && isLOWER_A(value))
16816 || (isUPPER_A(prevvalue) && isUPPER_A(value))
16818 vWARN(RExC_parse, "Ranges of ASCII printables should"
16819 " be some subset of \"0-9\","
16820 " \"A-Z\", or \"a-z\"");
16822 else if (prevvalue >= 0x660) { /* ARABIC_INDIC_DIGIT_ZERO */
16823 SSize_t index_start;
16824 SSize_t index_final;
16826 /* But the nature of Unicode and languages mean we
16827 * can't do the same checks for above-ASCII ranges,
16828 * except in the case of digit ones. These should
16829 * contain only digits from the same group of 10. The
16830 * ASCII case is handled just above. 0x660 is the
16831 * first digit character beyond ASCII. Hence here, the
16832 * range could be a range of digits. First some
16833 * unlikely special cases. Grandfather in that a range
16834 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
16835 * if its starting value is one of the 10 digits prior
16836 * to it. This is because it is an alternate way of
16837 * writing 19D1, and some people may expect it to be in
16838 * that group. But it is bad, because it won't give
16839 * the expected results. In Unicode 5.2 it was
16840 * considered to be in that group (of 11, hence), but
16841 * this was fixed in the next version */
16843 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
16844 goto warn_bad_digit_range;
16846 else if (UNLIKELY( prevvalue >= 0x1D7CE
16847 && value <= 0x1D7FF))
16849 /* This is the only other case currently in Unicode
16850 * where the algorithm below fails. The code
16851 * points just above are the end points of a single
16852 * range containing only decimal digits. It is 5
16853 * different series of 0-9. All other ranges of
16854 * digits currently in Unicode are just a single
16855 * series. (And mktables will notify us if a later
16856 * Unicode version breaks this.)
16858 * If the range being checked is at most 9 long,
16859 * and the digit values represented are in
16860 * numerical order, they are from the same series.
16862 if ( value - prevvalue > 9
16863 || ((( value - 0x1D7CE) % 10)
16864 <= (prevvalue - 0x1D7CE) % 10))
16866 goto warn_bad_digit_range;
16871 /* For all other ranges of digits in Unicode, the
16872 * algorithm is just to check if both end points
16873 * are in the same series, which is the same range.
16875 index_start = _invlist_search(
16876 PL_XPosix_ptrs[_CC_DIGIT],
16879 /* Warn if the range starts and ends with a digit,
16880 * and they are not in the same group of 10. */
16881 if ( index_start >= 0
16882 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
16884 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
16885 value)) != index_start
16886 && index_final >= 0
16887 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
16889 warn_bad_digit_range:
16890 vWARN(RExC_parse, "Ranges of digits should be"
16891 " from the same group of"
16898 if ((! range || prevvalue == value) && non_portable_endpoint) {
16899 if (isPRINT_A(value)) {
16902 if (isBACKSLASHED_PUNCT(value)) {
16903 literal[d++] = '\\';
16905 literal[d++] = (char) value;
16906 literal[d++] = '\0';
16908 vWARN4dep(RExC_parse,
16909 "\"%.*s\" is more clearly written simply as \"%s\". "
16910 "This will be a fatal error in Perl 5.28",
16911 (int) (RExC_parse - rangebegin),
16916 else if isMNEMONIC_CNTRL(value) {
16917 vWARN4dep(RExC_parse,
16918 "\"%.*s\" is more clearly written simply as \"%s\". "
16919 "This will be a fatal error in Perl 5.28",
16920 (int) (RExC_parse - rangebegin),
16922 cntrl_to_mnemonic((U8) value)
16928 /* Deal with this element of the class */
16932 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16935 /* On non-ASCII platforms, for ranges that span all of 0..255, and
16936 * ones that don't require special handling, we can just add the
16937 * range like we do for ASCII platforms */
16938 if ((UNLIKELY(prevvalue == 0) && value >= 255)
16939 || ! (prevvalue < 256
16941 || (! non_portable_endpoint
16942 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
16943 || (isUPPER_A(prevvalue)
16944 && isUPPER_A(value)))))))
16946 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16950 /* Here, requires special handling. This can be because it is
16951 * a range whose code points are considered to be Unicode, and
16952 * so must be individually translated into native, or because
16953 * its a subrange of 'A-Z' or 'a-z' which each aren't
16954 * contiguous in EBCDIC, but we have defined them to include
16955 * only the "expected" upper or lower case ASCII alphabetics.
16956 * Subranges above 255 are the same in native and Unicode, so
16957 * can be added as a range */
16958 U8 start = NATIVE_TO_LATIN1(prevvalue);
16960 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
16961 for (j = start; j <= end; j++) {
16962 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
16965 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16972 range = 0; /* this range (if it was one) is done now */
16973 } /* End of loop through all the text within the brackets */
16976 if ( posix_warnings && av_tindex_nomg(posix_warnings) >= 0) {
16977 output_or_return_posix_warnings(pRExC_state, posix_warnings,
16978 return_posix_warnings);
16981 /* If anything in the class expands to more than one character, we have to
16982 * deal with them by building up a substitute parse string, and recursively
16983 * calling reg() on it, instead of proceeding */
16984 if (multi_char_matches) {
16985 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
16988 char *save_end = RExC_end;
16989 char *save_parse = RExC_parse;
16990 char *save_start = RExC_start;
16991 STRLEN prefix_end = 0; /* We copy the character class after a
16992 prefix supplied here. This is the size
16993 + 1 of that prefix */
16994 bool first_time = TRUE; /* First multi-char occurrence doesn't get
16999 assert(RExC_precomp_adj == 0); /* Only one level of recursion allowed */
17001 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
17002 because too confusing */
17004 sv_catpv(substitute_parse, "(?:");
17008 /* Look at the longest folds first */
17009 for (cp_count = av_tindex_nomg(multi_char_matches);
17014 if (av_exists(multi_char_matches, cp_count)) {
17015 AV** this_array_ptr;
17018 this_array_ptr = (AV**) av_fetch(multi_char_matches,
17020 while ((this_sequence = av_pop(*this_array_ptr)) !=
17023 if (! first_time) {
17024 sv_catpv(substitute_parse, "|");
17026 first_time = FALSE;
17028 sv_catpv(substitute_parse, SvPVX(this_sequence));
17033 /* If the character class contains anything else besides these
17034 * multi-character folds, have to include it in recursive parsing */
17035 if (element_count) {
17036 sv_catpv(substitute_parse, "|[");
17037 prefix_end = SvCUR(substitute_parse);
17038 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
17040 /* Put in a closing ']' only if not going off the end, as otherwise
17041 * we are adding something that really isn't there */
17042 if (RExC_parse < RExC_end) {
17043 sv_catpv(substitute_parse, "]");
17047 sv_catpv(substitute_parse, ")");
17050 /* This is a way to get the parse to skip forward a whole named
17051 * sequence instead of matching the 2nd character when it fails the
17053 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
17057 /* Set up the data structure so that any errors will be properly
17058 * reported. See the comments at the definition of
17059 * REPORT_LOCATION_ARGS for details */
17060 RExC_precomp_adj = orig_parse - RExC_precomp;
17061 RExC_start = RExC_parse = SvPV(substitute_parse, len);
17062 RExC_adjusted_start = RExC_start + prefix_end;
17063 RExC_end = RExC_parse + len;
17064 RExC_in_multi_char_class = 1;
17065 RExC_emit = (regnode *)orig_emit;
17067 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17069 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PASS1|NEED_UTF8);
17071 /* And restore so can parse the rest of the pattern */
17072 RExC_parse = save_parse;
17073 RExC_start = RExC_adjusted_start = save_start;
17074 RExC_precomp_adj = 0;
17075 RExC_end = save_end;
17076 RExC_in_multi_char_class = 0;
17077 SvREFCNT_dec_NN(multi_char_matches);
17081 /* Here, we've gone through the entire class and dealt with multi-char
17082 * folds. We are now in a position that we can do some checks to see if we
17083 * can optimize this ANYOF node into a simpler one, even in Pass 1.
17084 * Currently we only do two checks:
17085 * 1) is in the unlikely event that the user has specified both, eg. \w and
17086 * \W under /l, then the class matches everything. (This optimization
17087 * is done only to make the optimizer code run later work.)
17088 * 2) if the character class contains only a single element (including a
17089 * single range), we see if there is an equivalent node for it.
17090 * Other checks are possible */
17092 && ! ret_invlist /* Can't optimize if returning the constructed
17094 && (UNLIKELY(posixl_matches_all) || element_count == 1))
17099 if (UNLIKELY(posixl_matches_all)) {
17102 else if (namedclass > OOB_NAMEDCLASS) { /* this is a single named
17103 class, like \w or [:digit:]
17106 /* All named classes are mapped into POSIXish nodes, with its FLAG
17107 * argument giving which class it is */
17108 switch ((I32)namedclass) {
17109 case ANYOF_UNIPROP:
17112 /* These don't depend on the charset modifiers. They always
17113 * match under /u rules */
17114 case ANYOF_NHORIZWS:
17115 case ANYOF_HORIZWS:
17116 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
17119 case ANYOF_NVERTWS:
17124 /* The actual POSIXish node for all the rest depends on the
17125 * charset modifier. The ones in the first set depend only on
17126 * ASCII or, if available on this platform, also locale */
17130 op = (LOC) ? POSIXL : POSIXA;
17136 /* The following don't have any matches in the upper Latin1
17137 * range, hence /d is equivalent to /u for them. Making it /u
17138 * saves some branches at runtime */
17142 case ANYOF_NXDIGIT:
17143 if (! DEPENDS_SEMANTICS) {
17144 goto treat_as_default;
17150 /* The following change to CASED under /i */
17156 namedclass = ANYOF_CASED + (namedclass % 2);
17160 /* The rest have more possibilities depending on the charset.
17161 * We take advantage of the enum ordering of the charset
17162 * modifiers to get the exact node type, */
17165 op = POSIXD + get_regex_charset(RExC_flags);
17166 if (op > POSIXA) { /* /aa is same as /a */
17171 /* The odd numbered ones are the complements of the
17172 * next-lower even number one */
17173 if (namedclass % 2 == 1) {
17177 arg = namedclass_to_classnum(namedclass);
17181 else if (value == prevvalue) {
17183 /* Here, the class consists of just a single code point */
17186 if (! LOC && value == '\n') {
17187 op = REG_ANY; /* Optimize [^\n] */
17188 *flagp |= HASWIDTH|SIMPLE;
17192 else if (value < 256 || UTF) {
17194 /* Optimize a single value into an EXACTish node, but not if it
17195 * would require converting the pattern to UTF-8. */
17196 op = compute_EXACTish(pRExC_state);
17198 } /* Otherwise is a range */
17199 else if (! LOC) { /* locale could vary these */
17200 if (prevvalue == '0') {
17201 if (value == '9') {
17206 else if (! FOLD || ASCII_FOLD_RESTRICTED) {
17207 /* We can optimize A-Z or a-z, but not if they could match
17208 * something like the KELVIN SIGN under /i. */
17209 if (prevvalue == 'A') {
17212 && ! non_portable_endpoint
17215 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
17219 else if (prevvalue == 'a') {
17222 && ! non_portable_endpoint
17225 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
17232 /* Here, we have changed <op> away from its initial value iff we found
17233 * an optimization */
17236 /* Throw away this ANYOF regnode, and emit the calculated one,
17237 * which should correspond to the beginning, not current, state of
17239 const char * cur_parse = RExC_parse;
17240 RExC_parse = (char *)orig_parse;
17244 /* To get locale nodes to not use the full ANYOF size would
17245 * require moving the code above that writes the portions
17246 * of it that aren't in other nodes to after this point.
17247 * e.g. ANYOF_POSIXL_SET */
17248 RExC_size = orig_size;
17252 RExC_emit = (regnode *)orig_emit;
17253 if (PL_regkind[op] == POSIXD) {
17254 if (op == POSIXL) {
17255 RExC_contains_locale = 1;
17258 op += NPOSIXD - POSIXD;
17263 ret = reg_node(pRExC_state, op);
17265 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17269 *flagp |= HASWIDTH|SIMPLE;
17271 else if (PL_regkind[op] == EXACT) {
17272 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17273 TRUE /* downgradable to EXACT */
17277 RExC_parse = (char *) cur_parse;
17279 SvREFCNT_dec(posixes);
17280 SvREFCNT_dec(nposixes);
17281 SvREFCNT_dec(simple_posixes);
17282 SvREFCNT_dec(cp_list);
17283 SvREFCNT_dec(cp_foldable_list);
17290 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
17292 /* If folding, we calculate all characters that could fold to or from the
17293 * ones already on the list */
17294 if (cp_foldable_list) {
17296 UV start, end; /* End points of code point ranges */
17298 SV* fold_intersection = NULL;
17301 /* Our calculated list will be for Unicode rules. For locale
17302 * matching, we have to keep a separate list that is consulted at
17303 * runtime only when the locale indicates Unicode rules. For
17304 * non-locale, we just use the general list */
17306 use_list = &only_utf8_locale_list;
17309 use_list = &cp_list;
17312 /* Only the characters in this class that participate in folds need
17313 * be checked. Get the intersection of this class and all the
17314 * possible characters that are foldable. This can quickly narrow
17315 * down a large class */
17316 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
17317 &fold_intersection);
17319 /* The folds for all the Latin1 characters are hard-coded into this
17320 * program, but we have to go out to disk to get the others. */
17321 if (invlist_highest(cp_foldable_list) >= 256) {
17323 /* This is a hash that for a particular fold gives all
17324 * characters that are involved in it */
17325 if (! PL_utf8_foldclosures) {
17326 _load_PL_utf8_foldclosures();
17330 /* Now look at the foldable characters in this class individually */
17331 invlist_iterinit(fold_intersection);
17332 while (invlist_iternext(fold_intersection, &start, &end)) {
17335 /* Look at every character in the range */
17336 for (j = start; j <= end; j++) {
17337 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17343 if (IS_IN_SOME_FOLD_L1(j)) {
17345 /* ASCII is always matched; non-ASCII is matched
17346 * only under Unicode rules (which could happen
17347 * under /l if the locale is a UTF-8 one */
17348 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
17349 *use_list = add_cp_to_invlist(*use_list,
17350 PL_fold_latin1[j]);
17353 has_upper_latin1_only_utf8_matches
17354 = add_cp_to_invlist(
17355 has_upper_latin1_only_utf8_matches,
17356 PL_fold_latin1[j]);
17360 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
17361 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
17363 add_above_Latin1_folds(pRExC_state,
17370 /* Here is an above Latin1 character. We don't have the
17371 * rules hard-coded for it. First, get its fold. This is
17372 * the simple fold, as the multi-character folds have been
17373 * handled earlier and separated out */
17374 _to_uni_fold_flags(j, foldbuf, &foldlen,
17375 (ASCII_FOLD_RESTRICTED)
17376 ? FOLD_FLAGS_NOMIX_ASCII
17379 /* Single character fold of above Latin1. Add everything in
17380 * its fold closure to the list that this node should match.
17381 * The fold closures data structure is a hash with the keys
17382 * being the UTF-8 of every character that is folded to, like
17383 * 'k', and the values each an array of all code points that
17384 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
17385 * Multi-character folds are not included */
17386 if ((listp = hv_fetch(PL_utf8_foldclosures,
17387 (char *) foldbuf, foldlen, FALSE)))
17389 AV* list = (AV*) *listp;
17391 for (k = 0; k <= av_tindex_nomg(list); k++) {
17392 SV** c_p = av_fetch(list, k, FALSE);
17398 /* /aa doesn't allow folds between ASCII and non- */
17399 if ((ASCII_FOLD_RESTRICTED
17400 && (isASCII(c) != isASCII(j))))
17405 /* Folds under /l which cross the 255/256 boundary
17406 * are added to a separate list. (These are valid
17407 * only when the locale is UTF-8.) */
17408 if (c < 256 && LOC) {
17409 *use_list = add_cp_to_invlist(*use_list, c);
17413 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
17415 cp_list = add_cp_to_invlist(cp_list, c);
17418 /* Similarly folds involving non-ascii Latin1
17419 * characters under /d are added to their list */
17420 has_upper_latin1_only_utf8_matches
17421 = add_cp_to_invlist(
17422 has_upper_latin1_only_utf8_matches,
17429 SvREFCNT_dec_NN(fold_intersection);
17432 /* Now that we have finished adding all the folds, there is no reason
17433 * to keep the foldable list separate */
17434 _invlist_union(cp_list, cp_foldable_list, &cp_list);
17435 SvREFCNT_dec_NN(cp_foldable_list);
17438 /* And combine the result (if any) with any inversion lists from posix
17439 * classes. The lists are kept separate up to now because we don't want to
17440 * fold the classes (folding of those is automatically handled by the swash
17441 * fetching code) */
17442 if (simple_posixes) { /* These are the classes known to be unaffected by
17445 _invlist_union(cp_list, simple_posixes, &cp_list);
17446 SvREFCNT_dec_NN(simple_posixes);
17449 cp_list = simple_posixes;
17452 if (posixes || nposixes) {
17454 /* We have to adjust /a and /aa */
17455 if (AT_LEAST_ASCII_RESTRICTED) {
17457 /* Under /a and /aa, nothing above ASCII matches these */
17459 _invlist_intersection(posixes,
17460 PL_XPosix_ptrs[_CC_ASCII],
17464 /* Under /a and /aa, everything above ASCII matches these
17467 _invlist_union_complement_2nd(nposixes,
17468 PL_XPosix_ptrs[_CC_ASCII],
17473 if (! DEPENDS_SEMANTICS) {
17475 /* For everything but /d, we can just add the current 'posixes' and
17476 * 'nposixes' to the main list */
17479 _invlist_union(cp_list, posixes, &cp_list);
17480 SvREFCNT_dec_NN(posixes);
17488 _invlist_union(cp_list, nposixes, &cp_list);
17489 SvREFCNT_dec_NN(nposixes);
17492 cp_list = nposixes;
17497 /* Under /d, things like \w match upper Latin1 characters only if
17498 * the target string is in UTF-8. But things like \W match all the
17499 * upper Latin1 characters if the target string is not in UTF-8.
17501 * Handle the case where there something like \W separately */
17503 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1);
17505 /* A complemented posix class matches all upper Latin1
17506 * characters if not in UTF-8. And it matches just certain
17507 * ones when in UTF-8. That means those certain ones are
17508 * matched regardless, so can just be added to the
17509 * unconditional list */
17511 _invlist_union(cp_list, nposixes, &cp_list);
17512 SvREFCNT_dec_NN(nposixes);
17516 cp_list = nposixes;
17519 /* Likewise for 'posixes' */
17520 _invlist_union(posixes, cp_list, &cp_list);
17522 /* Likewise for anything else in the range that matched only
17524 if (has_upper_latin1_only_utf8_matches) {
17525 _invlist_union(cp_list,
17526 has_upper_latin1_only_utf8_matches,
17528 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17529 has_upper_latin1_only_utf8_matches = NULL;
17532 /* If we don't match all the upper Latin1 characters regardless
17533 * of UTF-8ness, we have to set a flag to match the rest when
17535 _invlist_subtract(only_non_utf8_list, cp_list,
17536 &only_non_utf8_list);
17537 if (_invlist_len(only_non_utf8_list) != 0) {
17538 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17542 /* Here there were no complemented posix classes. That means
17543 * the upper Latin1 characters in 'posixes' match only when the
17544 * target string is in UTF-8. So we have to add them to the
17545 * list of those types of code points, while adding the
17546 * remainder to the unconditional list.
17548 * First calculate what they are */
17549 SV* nonascii_but_latin1_properties = NULL;
17550 _invlist_intersection(posixes, PL_UpperLatin1,
17551 &nonascii_but_latin1_properties);
17553 /* And add them to the final list of such characters. */
17554 _invlist_union(has_upper_latin1_only_utf8_matches,
17555 nonascii_but_latin1_properties,
17556 &has_upper_latin1_only_utf8_matches);
17558 /* Remove them from what now becomes the unconditional list */
17559 _invlist_subtract(posixes, nonascii_but_latin1_properties,
17562 /* And add those unconditional ones to the final list */
17564 _invlist_union(cp_list, posixes, &cp_list);
17565 SvREFCNT_dec_NN(posixes);
17572 SvREFCNT_dec(nonascii_but_latin1_properties);
17574 /* Get rid of any characters that we now know are matched
17575 * unconditionally from the conditional list, which may make
17576 * that list empty */
17577 _invlist_subtract(has_upper_latin1_only_utf8_matches,
17579 &has_upper_latin1_only_utf8_matches);
17580 if (_invlist_len(has_upper_latin1_only_utf8_matches) == 0) {
17581 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17582 has_upper_latin1_only_utf8_matches = NULL;
17588 /* And combine the result (if any) with any inversion list from properties.
17589 * The lists are kept separate up to now so that we can distinguish the two
17590 * in regards to matching above-Unicode. A run-time warning is generated
17591 * if a Unicode property is matched against a non-Unicode code point. But,
17592 * we allow user-defined properties to match anything, without any warning,
17593 * and we also suppress the warning if there is a portion of the character
17594 * class that isn't a Unicode property, and which matches above Unicode, \W
17595 * or [\x{110000}] for example.
17596 * (Note that in this case, unlike the Posix one above, there is no
17597 * <has_upper_latin1_only_utf8_matches>, because having a Unicode property
17598 * forces Unicode semantics */
17602 /* If it matters to the final outcome, see if a non-property
17603 * component of the class matches above Unicode. If so, the
17604 * warning gets suppressed. This is true even if just a single
17605 * such code point is specified, as, though not strictly correct if
17606 * another such code point is matched against, the fact that they
17607 * are using above-Unicode code points indicates they should know
17608 * the issues involved */
17610 warn_super = ! (invert
17611 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
17614 _invlist_union(properties, cp_list, &cp_list);
17615 SvREFCNT_dec_NN(properties);
17618 cp_list = properties;
17623 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17625 /* Because an ANYOF node is the only one that warns, this node
17626 * can't be optimized into something else */
17627 optimizable = FALSE;
17631 /* Here, we have calculated what code points should be in the character
17634 * Now we can see about various optimizations. Fold calculation (which we
17635 * did above) needs to take place before inversion. Otherwise /[^k]/i
17636 * would invert to include K, which under /i would match k, which it
17637 * shouldn't. Therefore we can't invert folded locale now, as it won't be
17638 * folded until runtime */
17640 /* If we didn't do folding, it's because some information isn't available
17641 * until runtime; set the run-time fold flag for these. (We don't have to
17642 * worry about properties folding, as that is taken care of by the swash
17643 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
17644 * locales, or the class matches at least one 0-255 range code point */
17647 /* Some things on the list might be unconditionally included because of
17648 * other components. Remove them, and clean up the list if it goes to
17650 if (only_utf8_locale_list && cp_list) {
17651 _invlist_subtract(only_utf8_locale_list, cp_list,
17652 &only_utf8_locale_list);
17654 if (_invlist_len(only_utf8_locale_list) == 0) {
17655 SvREFCNT_dec_NN(only_utf8_locale_list);
17656 only_utf8_locale_list = NULL;
17659 if (only_utf8_locale_list) {
17662 |ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17664 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
17666 invlist_iterinit(cp_list);
17667 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
17668 ANYOF_FLAGS(ret) |= ANYOFL_FOLD;
17670 invlist_iterfinish(cp_list);
17673 else if ( DEPENDS_SEMANTICS
17674 && ( has_upper_latin1_only_utf8_matches
17675 || (ANYOF_FLAGS(ret) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
17678 optimizable = FALSE;
17682 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
17683 * at compile time. Besides not inverting folded locale now, we can't
17684 * invert if there are things such as \w, which aren't known until runtime
17688 && OP(ret) != ANYOFD
17689 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
17690 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17692 _invlist_invert(cp_list);
17694 /* Any swash can't be used as-is, because we've inverted things */
17696 SvREFCNT_dec_NN(swash);
17700 /* Clear the invert flag since have just done it here */
17707 *ret_invlist = cp_list;
17708 SvREFCNT_dec(swash);
17710 /* Discard the generated node */
17712 RExC_size = orig_size;
17715 RExC_emit = orig_emit;
17720 /* Some character classes are equivalent to other nodes. Such nodes take
17721 * up less room and generally fewer operations to execute than ANYOF nodes.
17722 * Above, we checked for and optimized into some such equivalents for
17723 * certain common classes that are easy to test. Getting to this point in
17724 * the code means that the class didn't get optimized there. Since this
17725 * code is only executed in Pass 2, it is too late to save space--it has
17726 * been allocated in Pass 1, and currently isn't given back. But turning
17727 * things into an EXACTish node can allow the optimizer to join it to any
17728 * adjacent such nodes. And if the class is equivalent to things like /./,
17729 * expensive run-time swashes can be avoided. Now that we have more
17730 * complete information, we can find things necessarily missed by the
17731 * earlier code. Another possible "optimization" that isn't done is that
17732 * something like [Ee] could be changed into an EXACTFU. khw tried this
17733 * and found that the ANYOF is faster, including for code points not in the
17734 * bitmap. This still might make sense to do, provided it got joined with
17735 * an adjacent node(s) to create a longer EXACTFU one. This could be
17736 * accomplished by creating a pseudo ANYOF_EXACTFU node type that the join
17737 * routine would know is joinable. If that didn't happen, the node type
17738 * could then be made a straight ANYOF */
17740 if (optimizable && cp_list && ! invert) {
17742 U8 op = END; /* The optimzation node-type */
17743 int posix_class = -1; /* Illegal value */
17744 const char * cur_parse= RExC_parse;
17746 invlist_iterinit(cp_list);
17747 if (! invlist_iternext(cp_list, &start, &end)) {
17749 /* Here, the list is empty. This happens, for example, when a
17750 * Unicode property that doesn't match anything is the only element
17751 * in the character class (perluniprops.pod notes such properties).
17754 *flagp |= HASWIDTH|SIMPLE;
17756 else if (start == end) { /* The range is a single code point */
17757 if (! invlist_iternext(cp_list, &start, &end)
17759 /* Don't do this optimization if it would require changing
17760 * the pattern to UTF-8 */
17761 && (start < 256 || UTF))
17763 /* Here, the list contains a single code point. Can optimize
17764 * into an EXACTish node */
17775 /* A locale node under folding with one code point can be
17776 * an EXACTFL, as its fold won't be calculated until
17782 /* Here, we are generally folding, but there is only one
17783 * code point to match. If we have to, we use an EXACT
17784 * node, but it would be better for joining with adjacent
17785 * nodes in the optimization pass if we used the same
17786 * EXACTFish node that any such are likely to be. We can
17787 * do this iff the code point doesn't participate in any
17788 * folds. For example, an EXACTF of a colon is the same as
17789 * an EXACT one, since nothing folds to or from a colon. */
17791 if (IS_IN_SOME_FOLD_L1(value)) {
17796 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
17801 /* If we haven't found the node type, above, it means we
17802 * can use the prevailing one */
17804 op = compute_EXACTish(pRExC_state);
17808 } /* End of first range contains just a single code point */
17809 else if (start == 0) {
17810 if (end == UV_MAX) {
17812 *flagp |= HASWIDTH|SIMPLE;
17815 else if (end == '\n' - 1
17816 && invlist_iternext(cp_list, &start, &end)
17817 && start == '\n' + 1 && end == UV_MAX)
17820 *flagp |= HASWIDTH|SIMPLE;
17824 invlist_iterfinish(cp_list);
17827 const UV cp_list_len = _invlist_len(cp_list);
17828 const UV* cp_list_array = invlist_array(cp_list);
17830 /* Here, didn't find an optimization. See if this matches any of
17831 * the POSIX classes. These run slightly faster for above-Unicode
17832 * code points, so don't bother with POSIXA ones nor the 2 that
17833 * have no above-Unicode matches. We can avoid these checks unless
17834 * the ANYOF matches at least as high as the lowest POSIX one
17835 * (which was manually found to be \v. The actual code point may
17836 * increase in later Unicode releases, if a higher code point is
17837 * assigned to be \v, but this code will never break. It would
17838 * just mean we could execute the checks for posix optimizations
17839 * unnecessarily) */
17841 if (cp_list_array[cp_list_len-1] > 0x2029) {
17842 for (posix_class = 0;
17843 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
17847 if (posix_class == _CC_ASCII || posix_class == _CC_CNTRL) {
17850 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
17852 /* Check if matches normal or inverted */
17853 if (_invlistEQ(cp_list,
17854 PL_XPosix_ptrs[posix_class],
17857 op = (try_inverted)
17860 *flagp |= HASWIDTH|SIMPLE;
17870 RExC_parse = (char *)orig_parse;
17871 RExC_emit = (regnode *)orig_emit;
17873 if (regarglen[op]) {
17874 ret = reganode(pRExC_state, op, 0);
17876 ret = reg_node(pRExC_state, op);
17879 RExC_parse = (char *)cur_parse;
17881 if (PL_regkind[op] == EXACT) {
17882 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17883 TRUE /* downgradable to EXACT */
17886 else if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17887 FLAGS(ret) = posix_class;
17890 SvREFCNT_dec_NN(cp_list);
17895 /* Here, <cp_list> contains all the code points we can determine at
17896 * compile time that match under all conditions. Go through it, and
17897 * for things that belong in the bitmap, put them there, and delete from
17898 * <cp_list>. While we are at it, see if everything above 255 is in the
17899 * list, and if so, set a flag to speed up execution */
17901 populate_ANYOF_from_invlist(ret, &cp_list);
17904 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
17907 /* Here, the bitmap has been populated with all the Latin1 code points that
17908 * always match. Can now add to the overall list those that match only
17909 * when the target string is UTF-8 (<has_upper_latin1_only_utf8_matches>).
17911 if (has_upper_latin1_only_utf8_matches) {
17913 _invlist_union(cp_list,
17914 has_upper_latin1_only_utf8_matches,
17916 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17919 cp_list = has_upper_latin1_only_utf8_matches;
17921 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
17924 /* If there is a swash and more than one element, we can't use the swash in
17925 * the optimization below. */
17926 if (swash && element_count > 1) {
17927 SvREFCNT_dec_NN(swash);
17931 /* Note that the optimization of using 'swash' if it is the only thing in
17932 * the class doesn't have us change swash at all, so it can include things
17933 * that are also in the bitmap; otherwise we have purposely deleted that
17934 * duplicate information */
17935 set_ANYOF_arg(pRExC_state, ret, cp_list,
17936 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17938 only_utf8_locale_list,
17939 swash, has_user_defined_property);
17941 *flagp |= HASWIDTH|SIMPLE;
17943 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
17944 RExC_contains_locale = 1;
17950 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
17953 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
17954 regnode* const node,
17956 SV* const runtime_defns,
17957 SV* const only_utf8_locale_list,
17959 const bool has_user_defined_property)
17961 /* Sets the arg field of an ANYOF-type node 'node', using information about
17962 * the node passed-in. If there is nothing outside the node's bitmap, the
17963 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
17964 * the count returned by add_data(), having allocated and stored an array,
17965 * av, that that count references, as follows:
17966 * av[0] stores the character class description in its textual form.
17967 * This is used later (regexec.c:Perl_regclass_swash()) to
17968 * initialize the appropriate swash, and is also useful for dumping
17969 * the regnode. This is set to &PL_sv_undef if the textual
17970 * description is not needed at run-time (as happens if the other
17971 * elements completely define the class)
17972 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
17973 * computed from av[0]. But if no further computation need be done,
17974 * the swash is stored here now (and av[0] is &PL_sv_undef).
17975 * av[2] stores the inversion list of code points that match only if the
17976 * current locale is UTF-8
17977 * av[3] stores the cp_list inversion list for use in addition or instead
17978 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
17979 * (Otherwise everything needed is already in av[0] and av[1])
17980 * av[4] is set if any component of the class is from a user-defined
17981 * property; used only if av[3] exists */
17985 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
17987 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
17988 assert(! (ANYOF_FLAGS(node)
17989 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
17990 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
17993 AV * const av = newAV();
17996 av_store(av, 0, (runtime_defns)
17997 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
18000 av_store(av, 1, swash);
18001 SvREFCNT_dec_NN(cp_list);
18004 av_store(av, 1, &PL_sv_undef);
18006 av_store(av, 3, cp_list);
18007 av_store(av, 4, newSVuv(has_user_defined_property));
18011 if (only_utf8_locale_list) {
18012 av_store(av, 2, only_utf8_locale_list);
18015 av_store(av, 2, &PL_sv_undef);
18018 rv = newRV_noinc(MUTABLE_SV(av));
18019 n = add_data(pRExC_state, STR_WITH_LEN("s"));
18020 RExC_rxi->data->data[n] = (void*)rv;
18025 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
18027 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
18028 const regnode* node,
18031 SV** only_utf8_locale_ptr,
18032 SV** output_invlist)
18035 /* For internal core use only.
18036 * Returns the swash for the input 'node' in the regex 'prog'.
18037 * If <doinit> is 'true', will attempt to create the swash if not already
18039 * If <listsvp> is non-null, will return the printable contents of the
18040 * swash. This can be used to get debugging information even before the
18041 * swash exists, by calling this function with 'doinit' set to false, in
18042 * which case the components that will be used to eventually create the
18043 * swash are returned (in a printable form).
18044 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
18045 * store an inversion list of code points that should match only if the
18046 * execution-time locale is a UTF-8 one.
18047 * If <output_invlist> is not NULL, it is where this routine is to store an
18048 * inversion list of the code points that would be instead returned in
18049 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
18050 * when this parameter is used, is just the non-code point data that
18051 * will go into creating the swash. This currently should be just
18052 * user-defined properties whose definitions were not known at compile
18053 * time. Using this parameter allows for easier manipulation of the
18054 * swash's data by the caller. It is illegal to call this function with
18055 * this parameter set, but not <listsvp>
18057 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
18058 * that, in spite of this function's name, the swash it returns may include
18059 * the bitmap data as well */
18062 SV *si = NULL; /* Input swash initialization string */
18063 SV* invlist = NULL;
18065 RXi_GET_DECL(prog,progi);
18066 const struct reg_data * const data = prog ? progi->data : NULL;
18068 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
18069 assert(! output_invlist || listsvp);
18071 if (data && data->count) {
18072 const U32 n = ARG(node);
18074 if (data->what[n] == 's') {
18075 SV * const rv = MUTABLE_SV(data->data[n]);
18076 AV * const av = MUTABLE_AV(SvRV(rv));
18077 SV **const ary = AvARRAY(av);
18078 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
18080 si = *ary; /* ary[0] = the string to initialize the swash with */
18082 if (av_tindex_nomg(av) >= 2) {
18083 if (only_utf8_locale_ptr
18085 && ary[2] != &PL_sv_undef)
18087 *only_utf8_locale_ptr = ary[2];
18090 assert(only_utf8_locale_ptr);
18091 *only_utf8_locale_ptr = NULL;
18094 /* Elements 3 and 4 are either both present or both absent. [3]
18095 * is any inversion list generated at compile time; [4]
18096 * indicates if that inversion list has any user-defined
18097 * properties in it. */
18098 if (av_tindex_nomg(av) >= 3) {
18100 if (SvUV(ary[4])) {
18101 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
18109 /* Element [1] is reserved for the set-up swash. If already there,
18110 * return it; if not, create it and store it there */
18111 if (ary[1] && SvROK(ary[1])) {
18114 else if (doinit && ((si && si != &PL_sv_undef)
18115 || (invlist && invlist != &PL_sv_undef))) {
18117 sw = _core_swash_init("utf8", /* the utf8 package */
18121 0, /* not from tr/// */
18123 &swash_init_flags);
18124 (void)av_store(av, 1, sw);
18129 /* If requested, return a printable version of what this swash matches */
18131 SV* matches_string = NULL;
18133 /* The swash should be used, if possible, to get the data, as it
18134 * contains the resolved data. But this function can be called at
18135 * compile-time, before everything gets resolved, in which case we
18136 * return the currently best available information, which is the string
18137 * that will eventually be used to do that resolving, 'si' */
18138 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
18139 && (si && si != &PL_sv_undef))
18141 /* Here, we only have 'si' (and possibly some passed-in data in
18142 * 'invlist', which is handled below) If the caller only wants
18143 * 'si', use that. */
18144 if (! output_invlist) {
18145 matches_string = newSVsv(si);
18148 /* But if the caller wants an inversion list of the node, we
18149 * need to parse 'si' and place as much as possible in the
18150 * desired output inversion list, making 'matches_string' only
18151 * contain the currently unresolvable things */
18152 const char *si_string = SvPVX(si);
18153 STRLEN remaining = SvCUR(si);
18157 /* Ignore everything before the first new-line */
18158 while (*si_string != '\n' && remaining > 0) {
18162 assert(remaining > 0);
18167 while (remaining > 0) {
18169 /* The data consists of just strings defining user-defined
18170 * property names, but in prior incarnations, and perhaps
18171 * somehow from pluggable regex engines, it could still
18172 * hold hex code point definitions. Each component of a
18173 * range would be separated by a tab, and each range by a
18174 * new-line. If these are found, instead add them to the
18175 * inversion list */
18176 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
18177 |PERL_SCAN_SILENT_NON_PORTABLE;
18178 STRLEN len = remaining;
18179 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
18181 /* If the hex decode routine found something, it should go
18182 * up to the next \n */
18183 if ( *(si_string + len) == '\n') {
18184 if (count) { /* 2nd code point on line */
18185 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
18188 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
18191 goto prepare_for_next_iteration;
18194 /* If the hex decode was instead for the lower range limit,
18195 * save it, and go parse the upper range limit */
18196 if (*(si_string + len) == '\t') {
18197 assert(count == 0);
18201 prepare_for_next_iteration:
18202 si_string += len + 1;
18203 remaining -= len + 1;
18207 /* Here, didn't find a legal hex number. Just add it from
18208 * here to the next \n */
18211 while (*(si_string + len) != '\n' && remaining > 0) {
18215 if (*(si_string + len) == '\n') {
18219 if (matches_string) {
18220 sv_catpvn(matches_string, si_string, len - 1);
18223 matches_string = newSVpvn(si_string, len - 1);
18226 sv_catpvs(matches_string, " ");
18227 } /* end of loop through the text */
18229 assert(matches_string);
18230 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
18231 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
18233 } /* end of has an 'si' but no swash */
18236 /* If we have a swash in place, its equivalent inversion list was above
18237 * placed into 'invlist'. If not, this variable may contain a stored
18238 * inversion list which is information beyond what is in 'si' */
18241 /* Again, if the caller doesn't want the output inversion list, put
18242 * everything in 'matches-string' */
18243 if (! output_invlist) {
18244 if ( ! matches_string) {
18245 matches_string = newSVpvs("\n");
18247 sv_catsv(matches_string, invlist_contents(invlist,
18248 TRUE /* traditional style */
18251 else if (! *output_invlist) {
18252 *output_invlist = invlist_clone(invlist);
18255 _invlist_union(*output_invlist, invlist, output_invlist);
18259 *listsvp = matches_string;
18264 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
18266 /* reg_skipcomment()
18268 Absorbs an /x style # comment from the input stream,
18269 returning a pointer to the first character beyond the comment, or if the
18270 comment terminates the pattern without anything following it, this returns
18271 one past the final character of the pattern (in other words, RExC_end) and
18272 sets the REG_RUN_ON_COMMENT_SEEN flag.
18274 Note it's the callers responsibility to ensure that we are
18275 actually in /x mode
18279 PERL_STATIC_INLINE char*
18280 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
18282 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
18286 while (p < RExC_end) {
18287 if (*(++p) == '\n') {
18292 /* we ran off the end of the pattern without ending the comment, so we have
18293 * to add an \n when wrapping */
18294 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
18299 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
18301 const bool force_to_xmod
18304 /* If the text at the current parse position '*p' is a '(?#...)' comment,
18305 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
18306 * is /x whitespace, advance '*p' so that on exit it points to the first
18307 * byte past all such white space and comments */
18309 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
18311 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
18313 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
18316 if (RExC_end - (*p) >= 3
18318 && *(*p + 1) == '?'
18319 && *(*p + 2) == '#')
18321 while (*(*p) != ')') {
18322 if ((*p) == RExC_end)
18323 FAIL("Sequence (?#... not terminated");
18331 const char * save_p = *p;
18332 while ((*p) < RExC_end) {
18334 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
18337 else if (*(*p) == '#') {
18338 (*p) = reg_skipcomment(pRExC_state, (*p));
18344 if (*p != save_p) {
18357 Advances the parse position by one byte, unless that byte is the beginning
18358 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
18359 those two cases, the parse position is advanced beyond all such comments and
18362 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
18366 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
18368 PERL_ARGS_ASSERT_NEXTCHAR;
18370 if (RExC_parse < RExC_end) {
18372 || UTF8_IS_INVARIANT(*RExC_parse)
18373 || UTF8_IS_START(*RExC_parse));
18375 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
18377 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
18378 FALSE /* Don't force /x */ );
18383 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
18385 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
18386 * space. In pass1, it aligns and increments RExC_size; in pass2,
18389 regnode * const ret = RExC_emit;
18390 GET_RE_DEBUG_FLAGS_DECL;
18392 PERL_ARGS_ASSERT_REGNODE_GUTS;
18394 assert(extra_size >= regarglen[op]);
18397 SIZE_ALIGN(RExC_size);
18398 RExC_size += 1 + extra_size;
18401 if (RExC_emit >= RExC_emit_bound)
18402 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
18403 op, (void*)RExC_emit, (void*)RExC_emit_bound);
18405 NODE_ALIGN_FILL(ret);
18406 #ifndef RE_TRACK_PATTERN_OFFSETS
18407 PERL_UNUSED_ARG(name);
18409 if (RExC_offsets) { /* MJD */
18411 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
18414 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
18415 ? "Overwriting end of array!\n" : "OK",
18416 (UV)(RExC_emit - RExC_emit_start),
18417 (UV)(RExC_parse - RExC_start),
18418 (UV)RExC_offsets[0]));
18419 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
18426 - reg_node - emit a node
18428 STATIC regnode * /* Location. */
18429 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
18431 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
18433 PERL_ARGS_ASSERT_REG_NODE;
18435 assert(regarglen[op] == 0);
18438 regnode *ptr = ret;
18439 FILL_ADVANCE_NODE(ptr, op);
18446 - reganode - emit a node with an argument
18448 STATIC regnode * /* Location. */
18449 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
18451 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
18453 PERL_ARGS_ASSERT_REGANODE;
18455 assert(regarglen[op] == 1);
18458 regnode *ptr = ret;
18459 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
18466 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
18468 /* emit a node with U32 and I32 arguments */
18470 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
18472 PERL_ARGS_ASSERT_REG2LANODE;
18474 assert(regarglen[op] == 2);
18477 regnode *ptr = ret;
18478 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
18485 - reginsert - insert an operator in front of already-emitted operand
18487 * Means relocating the operand.
18490 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
18495 const int offset = regarglen[(U8)op];
18496 const int size = NODE_STEP_REGNODE + offset;
18497 GET_RE_DEBUG_FLAGS_DECL;
18499 PERL_ARGS_ASSERT_REGINSERT;
18500 PERL_UNUSED_CONTEXT;
18501 PERL_UNUSED_ARG(depth);
18502 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
18503 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
18508 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
18509 studying. If this is wrong then we need to adjust RExC_recurse
18510 below like we do with RExC_open_parens/RExC_close_parens. */
18514 if (RExC_open_parens) {
18516 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
18517 /* remember that RExC_npar is rex->nparens + 1,
18518 * iow it is 1 more than the number of parens seen in
18519 * the pattern so far. */
18520 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
18521 /* note, RExC_open_parens[0] is the start of the
18522 * regex, it can't move. RExC_close_parens[0] is the end
18523 * of the regex, it *can* move. */
18524 if ( paren && RExC_open_parens[paren] >= opnd ) {
18525 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
18526 RExC_open_parens[paren] += size;
18528 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
18530 if ( RExC_close_parens[paren] >= opnd ) {
18531 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
18532 RExC_close_parens[paren] += size;
18534 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
18539 RExC_end_op += size;
18541 while (src > opnd) {
18542 StructCopy(--src, --dst, regnode);
18543 #ifdef RE_TRACK_PATTERN_OFFSETS
18544 if (RExC_offsets) { /* MJD 20010112 */
18546 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
18550 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
18551 ? "Overwriting end of array!\n" : "OK",
18552 (UV)(src - RExC_emit_start),
18553 (UV)(dst - RExC_emit_start),
18554 (UV)RExC_offsets[0]));
18555 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
18556 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
18562 place = opnd; /* Op node, where operand used to be. */
18563 #ifdef RE_TRACK_PATTERN_OFFSETS
18564 if (RExC_offsets) { /* MJD */
18566 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
18570 (UV)(place - RExC_emit_start) > RExC_offsets[0]
18571 ? "Overwriting end of array!\n" : "OK",
18572 (UV)(place - RExC_emit_start),
18573 (UV)(RExC_parse - RExC_start),
18574 (UV)RExC_offsets[0]));
18575 Set_Node_Offset(place, RExC_parse);
18576 Set_Node_Length(place, 1);
18579 src = NEXTOPER(place);
18580 FILL_ADVANCE_NODE(place, op);
18581 Zero(src, offset, regnode);
18585 - regtail - set the next-pointer at the end of a node chain of p to val.
18586 - SEE ALSO: regtail_study
18589 S_regtail(pTHX_ RExC_state_t * pRExC_state,
18590 const regnode * const p,
18591 const regnode * const val,
18595 GET_RE_DEBUG_FLAGS_DECL;
18597 PERL_ARGS_ASSERT_REGTAIL;
18599 PERL_UNUSED_ARG(depth);
18605 /* Find last node. */
18606 scan = (regnode *) p;
18608 regnode * const temp = regnext(scan);
18610 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
18611 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18612 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
18613 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
18614 (temp == NULL ? "->" : ""),
18615 (temp == NULL ? PL_reg_name[OP(val)] : "")
18623 if (reg_off_by_arg[OP(scan)]) {
18624 ARG_SET(scan, val - scan);
18627 NEXT_OFF(scan) = val - scan;
18633 - regtail_study - set the next-pointer at the end of a node chain of p to val.
18634 - Look for optimizable sequences at the same time.
18635 - currently only looks for EXACT chains.
18637 This is experimental code. The idea is to use this routine to perform
18638 in place optimizations on branches and groups as they are constructed,
18639 with the long term intention of removing optimization from study_chunk so
18640 that it is purely analytical.
18642 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
18643 to control which is which.
18646 /* TODO: All four parms should be const */
18649 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
18650 const regnode *val,U32 depth)
18654 #ifdef EXPERIMENTAL_INPLACESCAN
18657 GET_RE_DEBUG_FLAGS_DECL;
18659 PERL_ARGS_ASSERT_REGTAIL_STUDY;
18665 /* Find last node. */
18669 regnode * const temp = regnext(scan);
18670 #ifdef EXPERIMENTAL_INPLACESCAN
18671 if (PL_regkind[OP(scan)] == EXACT) {
18672 bool unfolded_multi_char; /* Unexamined in this routine */
18673 if (join_exact(pRExC_state, scan, &min,
18674 &unfolded_multi_char, 1, val, depth+1))
18679 switch (OP(scan)) {
18683 case EXACTFA_NO_TRIE:
18689 if( exact == PSEUDO )
18691 else if ( exact != OP(scan) )
18700 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
18701 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18702 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
18703 SvPV_nolen_const(RExC_mysv),
18704 REG_NODE_NUM(scan),
18705 PL_reg_name[exact]);
18712 DEBUG_PARSE_MSG("");
18713 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
18714 Perl_re_printf( aTHX_
18715 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
18716 SvPV_nolen_const(RExC_mysv),
18717 (IV)REG_NODE_NUM(val),
18721 if (reg_off_by_arg[OP(scan)]) {
18722 ARG_SET(scan, val - scan);
18725 NEXT_OFF(scan) = val - scan;
18733 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
18738 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
18743 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18745 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
18746 if (flags & (1<<bit)) {
18747 if (!set++ && lead)
18748 Perl_re_printf( aTHX_ "%s",lead);
18749 Perl_re_printf( aTHX_ "%s ",PL_reg_intflags_name[bit]);
18754 Perl_re_printf( aTHX_ "\n");
18756 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18761 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
18767 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18769 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
18770 if (flags & (1<<bit)) {
18771 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
18774 if (!set++ && lead)
18775 Perl_re_printf( aTHX_ "%s",lead);
18776 Perl_re_printf( aTHX_ "%s ",PL_reg_extflags_name[bit]);
18779 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
18780 if (!set++ && lead) {
18781 Perl_re_printf( aTHX_ "%s",lead);
18784 case REGEX_UNICODE_CHARSET:
18785 Perl_re_printf( aTHX_ "UNICODE");
18787 case REGEX_LOCALE_CHARSET:
18788 Perl_re_printf( aTHX_ "LOCALE");
18790 case REGEX_ASCII_RESTRICTED_CHARSET:
18791 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
18793 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
18794 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
18797 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
18803 Perl_re_printf( aTHX_ "\n");
18805 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18811 Perl_regdump(pTHX_ const regexp *r)
18814 SV * const sv = sv_newmortal();
18815 SV *dsv= sv_newmortal();
18816 RXi_GET_DECL(r,ri);
18817 GET_RE_DEBUG_FLAGS_DECL;
18819 PERL_ARGS_ASSERT_REGDUMP;
18821 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
18823 /* Header fields of interest. */
18824 if (r->anchored_substr) {
18825 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
18826 RE_SV_DUMPLEN(r->anchored_substr), 30);
18827 Perl_re_printf( aTHX_
18828 "anchored %s%s at %" IVdf " ",
18829 s, RE_SV_TAIL(r->anchored_substr),
18830 (IV)r->anchored_offset);
18831 } else if (r->anchored_utf8) {
18832 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
18833 RE_SV_DUMPLEN(r->anchored_utf8), 30);
18834 Perl_re_printf( aTHX_
18835 "anchored utf8 %s%s at %" IVdf " ",
18836 s, RE_SV_TAIL(r->anchored_utf8),
18837 (IV)r->anchored_offset);
18839 if (r->float_substr) {
18840 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
18841 RE_SV_DUMPLEN(r->float_substr), 30);
18842 Perl_re_printf( aTHX_
18843 "floating %s%s at %" IVdf "..%" UVuf " ",
18844 s, RE_SV_TAIL(r->float_substr),
18845 (IV)r->float_min_offset, (UV)r->float_max_offset);
18846 } else if (r->float_utf8) {
18847 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
18848 RE_SV_DUMPLEN(r->float_utf8), 30);
18849 Perl_re_printf( aTHX_
18850 "floating utf8 %s%s at %" IVdf "..%" UVuf " ",
18851 s, RE_SV_TAIL(r->float_utf8),
18852 (IV)r->float_min_offset, (UV)r->float_max_offset);
18854 if (r->check_substr || r->check_utf8)
18855 Perl_re_printf( aTHX_
18857 (r->check_substr == r->float_substr
18858 && r->check_utf8 == r->float_utf8
18859 ? "(checking floating" : "(checking anchored"));
18860 if (r->intflags & PREGf_NOSCAN)
18861 Perl_re_printf( aTHX_ " noscan");
18862 if (r->extflags & RXf_CHECK_ALL)
18863 Perl_re_printf( aTHX_ " isall");
18864 if (r->check_substr || r->check_utf8)
18865 Perl_re_printf( aTHX_ ") ");
18867 if (ri->regstclass) {
18868 regprop(r, sv, ri->regstclass, NULL, NULL);
18869 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
18871 if (r->intflags & PREGf_ANCH) {
18872 Perl_re_printf( aTHX_ "anchored");
18873 if (r->intflags & PREGf_ANCH_MBOL)
18874 Perl_re_printf( aTHX_ "(MBOL)");
18875 if (r->intflags & PREGf_ANCH_SBOL)
18876 Perl_re_printf( aTHX_ "(SBOL)");
18877 if (r->intflags & PREGf_ANCH_GPOS)
18878 Perl_re_printf( aTHX_ "(GPOS)");
18879 Perl_re_printf( aTHX_ " ");
18881 if (r->intflags & PREGf_GPOS_SEEN)
18882 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
18883 if (r->intflags & PREGf_SKIP)
18884 Perl_re_printf( aTHX_ "plus ");
18885 if (r->intflags & PREGf_IMPLICIT)
18886 Perl_re_printf( aTHX_ "implicit ");
18887 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
18888 if (r->extflags & RXf_EVAL_SEEN)
18889 Perl_re_printf( aTHX_ "with eval ");
18890 Perl_re_printf( aTHX_ "\n");
18892 regdump_extflags("r->extflags: ",r->extflags);
18893 regdump_intflags("r->intflags: ",r->intflags);
18896 PERL_ARGS_ASSERT_REGDUMP;
18897 PERL_UNUSED_CONTEXT;
18898 PERL_UNUSED_ARG(r);
18899 #endif /* DEBUGGING */
18902 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
18905 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
18906 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
18907 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
18908 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
18909 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
18910 || _CC_VERTSPACE != 15
18911 # error Need to adjust order of anyofs[]
18913 static const char * const anyofs[] = {
18950 - regprop - printable representation of opcode, with run time support
18954 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
18958 RXi_GET_DECL(prog,progi);
18959 GET_RE_DEBUG_FLAGS_DECL;
18961 PERL_ARGS_ASSERT_REGPROP;
18965 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
18966 /* It would be nice to FAIL() here, but this may be called from
18967 regexec.c, and it would be hard to supply pRExC_state. */
18968 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
18969 (int)OP(o), (int)REGNODE_MAX);
18970 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
18972 k = PL_regkind[OP(o)];
18975 sv_catpvs(sv, " ");
18976 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
18977 * is a crude hack but it may be the best for now since
18978 * we have no flag "this EXACTish node was UTF-8"
18980 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
18981 PERL_PV_ESCAPE_UNI_DETECT |
18982 PERL_PV_ESCAPE_NONASCII |
18983 PERL_PV_PRETTY_ELLIPSES |
18984 PERL_PV_PRETTY_LTGT |
18985 PERL_PV_PRETTY_NOCLEAR
18987 } else if (k == TRIE) {
18988 /* print the details of the trie in dumpuntil instead, as
18989 * progi->data isn't available here */
18990 const char op = OP(o);
18991 const U32 n = ARG(o);
18992 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
18993 (reg_ac_data *)progi->data->data[n] :
18995 const reg_trie_data * const trie
18996 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
18998 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
18999 DEBUG_TRIE_COMPILE_r({
19001 sv_catpvs(sv, "(JUMP)");
19002 Perl_sv_catpvf(aTHX_ sv,
19003 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
19004 (UV)trie->startstate,
19005 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
19006 (UV)trie->wordcount,
19009 (UV)TRIE_CHARCOUNT(trie),
19010 (UV)trie->uniquecharcount
19013 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
19014 sv_catpvs(sv, "[");
19015 (void) put_charclass_bitmap_innards(sv,
19016 ((IS_ANYOF_TRIE(op))
19018 : TRIE_BITMAP(trie)),
19024 sv_catpvs(sv, "]");
19026 } else if (k == CURLY) {
19027 U32 lo = ARG1(o), hi = ARG2(o);
19028 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
19029 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
19030 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
19031 if (hi == REG_INFTY)
19032 sv_catpvs(sv, "INFTY");
19034 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
19035 sv_catpvs(sv, "}");
19037 else if (k == WHILEM && o->flags) /* Ordinal/of */
19038 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
19039 else if (k == REF || k == OPEN || k == CLOSE
19040 || k == GROUPP || OP(o)==ACCEPT)
19042 AV *name_list= NULL;
19043 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
19044 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
19045 if ( RXp_PAREN_NAMES(prog) ) {
19046 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19047 } else if ( pRExC_state ) {
19048 name_list= RExC_paren_name_list;
19051 if ( k != REF || (OP(o) < NREF)) {
19052 SV **name= av_fetch(name_list, parno, 0 );
19054 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19057 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
19058 I32 *nums=(I32*)SvPVX(sv_dat);
19059 SV **name= av_fetch(name_list, nums[0], 0 );
19062 for ( n=0; n<SvIVX(sv_dat); n++ ) {
19063 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
19064 (n ? "," : ""), (IV)nums[n]);
19066 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19070 if ( k == REF && reginfo) {
19071 U32 n = ARG(o); /* which paren pair */
19072 I32 ln = prog->offs[n].start;
19073 if (prog->lastparen < n || ln == -1)
19074 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
19075 else if (ln == prog->offs[n].end)
19076 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
19078 const char *s = reginfo->strbeg + ln;
19079 Perl_sv_catpvf(aTHX_ sv, ": ");
19080 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
19081 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
19084 } else if (k == GOSUB) {
19085 AV *name_list= NULL;
19086 if ( RXp_PAREN_NAMES(prog) ) {
19087 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19088 } else if ( pRExC_state ) {
19089 name_list= RExC_paren_name_list;
19092 /* Paren and offset */
19093 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
19094 (int)((o + (int)ARG2L(o)) - progi->program) );
19096 SV **name= av_fetch(name_list, ARG(o), 0 );
19098 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19101 else if (k == LOGICAL)
19102 /* 2: embedded, otherwise 1 */
19103 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
19104 else if (k == ANYOF) {
19105 const U8 flags = ANYOF_FLAGS(o);
19106 bool do_sep = FALSE; /* Do we need to separate various components of
19108 /* Set if there is still an unresolved user-defined property */
19109 SV *unresolved = NULL;
19111 /* Things that are ignored except when the runtime locale is UTF-8 */
19112 SV *only_utf8_locale_invlist = NULL;
19114 /* Code points that don't fit in the bitmap */
19115 SV *nonbitmap_invlist = NULL;
19117 /* And things that aren't in the bitmap, but are small enough to be */
19118 SV* bitmap_range_not_in_bitmap = NULL;
19120 const bool inverted = flags & ANYOF_INVERT;
19122 if (OP(o) == ANYOFL) {
19123 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
19124 sv_catpvs(sv, "{utf8-locale-reqd}");
19126 if (flags & ANYOFL_FOLD) {
19127 sv_catpvs(sv, "{i}");
19131 /* If there is stuff outside the bitmap, get it */
19132 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
19133 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
19135 &only_utf8_locale_invlist,
19136 &nonbitmap_invlist);
19137 /* The non-bitmap data may contain stuff that could fit in the
19138 * bitmap. This could come from a user-defined property being
19139 * finally resolved when this call was done; or much more likely
19140 * because there are matches that require UTF-8 to be valid, and so
19141 * aren't in the bitmap. This is teased apart later */
19142 _invlist_intersection(nonbitmap_invlist,
19144 &bitmap_range_not_in_bitmap);
19145 /* Leave just the things that don't fit into the bitmap */
19146 _invlist_subtract(nonbitmap_invlist,
19148 &nonbitmap_invlist);
19151 /* Obey this flag to add all above-the-bitmap code points */
19152 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
19153 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
19154 NUM_ANYOF_CODE_POINTS,
19158 /* Ready to start outputting. First, the initial left bracket */
19159 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
19161 /* Then all the things that could fit in the bitmap */
19162 do_sep = put_charclass_bitmap_innards(sv,
19164 bitmap_range_not_in_bitmap,
19165 only_utf8_locale_invlist,
19168 /* Can't try inverting for a
19169 * better display if there are
19170 * things that haven't been
19172 unresolved != NULL);
19173 SvREFCNT_dec(bitmap_range_not_in_bitmap);
19175 /* If there are user-defined properties which haven't been defined yet,
19176 * output them. If the result is not to be inverted, it is clearest to
19177 * output them in a separate [] from the bitmap range stuff. If the
19178 * result is to be complemented, we have to show everything in one [],
19179 * as the inversion applies to the whole thing. Use {braces} to
19180 * separate them from anything in the bitmap and anything above the
19184 if (! do_sep) { /* If didn't output anything in the bitmap */
19185 sv_catpvs(sv, "^");
19187 sv_catpvs(sv, "{");
19190 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19192 sv_catsv(sv, unresolved);
19194 sv_catpvs(sv, "}");
19196 do_sep = ! inverted;
19199 /* And, finally, add the above-the-bitmap stuff */
19200 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
19203 /* See if truncation size is overridden */
19204 const STRLEN dump_len = (PL_dump_re_max_len)
19205 ? PL_dump_re_max_len
19208 /* This is output in a separate [] */
19210 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19213 /* And, for easy of understanding, it is shown in the
19214 * uncomplemented form if possible. The one exception being if
19215 * there are unresolved items, where the inversion has to be
19216 * delayed until runtime */
19217 if (inverted && ! unresolved) {
19218 _invlist_invert(nonbitmap_invlist);
19219 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
19222 contents = invlist_contents(nonbitmap_invlist,
19223 FALSE /* output suitable for catsv */
19226 /* If the output is shorter than the permissible maximum, just do it. */
19227 if (SvCUR(contents) <= dump_len) {
19228 sv_catsv(sv, contents);
19231 const char * contents_string = SvPVX(contents);
19232 STRLEN i = dump_len;
19234 /* Otherwise, start at the permissible max and work back to the
19235 * first break possibility */
19236 while (i > 0 && contents_string[i] != ' ') {
19239 if (i == 0) { /* Fail-safe. Use the max if we couldn't
19240 find a legal break */
19244 sv_catpvn(sv, contents_string, i);
19245 sv_catpvs(sv, "...");
19248 SvREFCNT_dec_NN(contents);
19249 SvREFCNT_dec_NN(nonbitmap_invlist);
19252 /* And finally the matching, closing ']' */
19253 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
19255 SvREFCNT_dec(unresolved);
19257 else if (k == POSIXD || k == NPOSIXD) {
19258 U8 index = FLAGS(o) * 2;
19259 if (index < C_ARRAY_LENGTH(anyofs)) {
19260 if (*anyofs[index] != '[') {
19263 sv_catpv(sv, anyofs[index]);
19264 if (*anyofs[index] != '[') {
19269 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
19272 else if (k == BOUND || k == NBOUND) {
19273 /* Must be synced with order of 'bound_type' in regcomp.h */
19274 const char * const bounds[] = {
19275 "", /* Traditional */
19281 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
19282 sv_catpv(sv, bounds[FLAGS(o)]);
19284 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
19285 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
19286 else if (OP(o) == SBOL)
19287 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
19289 /* add on the verb argument if there is one */
19290 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
19291 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
19292 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
19295 PERL_UNUSED_CONTEXT;
19296 PERL_UNUSED_ARG(sv);
19297 PERL_UNUSED_ARG(o);
19298 PERL_UNUSED_ARG(prog);
19299 PERL_UNUSED_ARG(reginfo);
19300 PERL_UNUSED_ARG(pRExC_state);
19301 #endif /* DEBUGGING */
19307 Perl_re_intuit_string(pTHX_ REGEXP * const r)
19308 { /* Assume that RE_INTUIT is set */
19309 struct regexp *const prog = ReANY(r);
19310 GET_RE_DEBUG_FLAGS_DECL;
19312 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
19313 PERL_UNUSED_CONTEXT;
19317 const char * const s = SvPV_nolen_const(RX_UTF8(r)
19318 ? prog->check_utf8 : prog->check_substr);
19320 if (!PL_colorset) reginitcolors();
19321 Perl_re_printf( aTHX_
19322 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
19324 RX_UTF8(r) ? "utf8 " : "",
19325 PL_colors[5],PL_colors[0],
19328 (strlen(s) > 60 ? "..." : ""));
19331 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
19332 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
19338 handles refcounting and freeing the perl core regexp structure. When
19339 it is necessary to actually free the structure the first thing it
19340 does is call the 'free' method of the regexp_engine associated to
19341 the regexp, allowing the handling of the void *pprivate; member
19342 first. (This routine is not overridable by extensions, which is why
19343 the extensions free is called first.)
19345 See regdupe and regdupe_internal if you change anything here.
19347 #ifndef PERL_IN_XSUB_RE
19349 Perl_pregfree(pTHX_ REGEXP *r)
19355 Perl_pregfree2(pTHX_ REGEXP *rx)
19357 struct regexp *const r = ReANY(rx);
19358 GET_RE_DEBUG_FLAGS_DECL;
19360 PERL_ARGS_ASSERT_PREGFREE2;
19362 if (r->mother_re) {
19363 ReREFCNT_dec(r->mother_re);
19365 CALLREGFREE_PVT(rx); /* free the private data */
19366 SvREFCNT_dec(RXp_PAREN_NAMES(r));
19367 Safefree(r->xpv_len_u.xpvlenu_pv);
19370 SvREFCNT_dec(r->anchored_substr);
19371 SvREFCNT_dec(r->anchored_utf8);
19372 SvREFCNT_dec(r->float_substr);
19373 SvREFCNT_dec(r->float_utf8);
19374 Safefree(r->substrs);
19376 RX_MATCH_COPY_FREE(rx);
19377 #ifdef PERL_ANY_COW
19378 SvREFCNT_dec(r->saved_copy);
19381 SvREFCNT_dec(r->qr_anoncv);
19382 if (r->recurse_locinput)
19383 Safefree(r->recurse_locinput);
19384 rx->sv_u.svu_rx = 0;
19389 This is a hacky workaround to the structural issue of match results
19390 being stored in the regexp structure which is in turn stored in
19391 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
19392 could be PL_curpm in multiple contexts, and could require multiple
19393 result sets being associated with the pattern simultaneously, such
19394 as when doing a recursive match with (??{$qr})
19396 The solution is to make a lightweight copy of the regexp structure
19397 when a qr// is returned from the code executed by (??{$qr}) this
19398 lightweight copy doesn't actually own any of its data except for
19399 the starp/end and the actual regexp structure itself.
19405 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
19407 struct regexp *ret;
19408 struct regexp *const r = ReANY(rx);
19409 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
19411 PERL_ARGS_ASSERT_REG_TEMP_COPY;
19414 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
19416 SvOK_off((SV *)ret_x);
19418 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
19419 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
19420 made both spots point to the same regexp body.) */
19421 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
19422 assert(!SvPVX(ret_x));
19423 ret_x->sv_u.svu_rx = temp->sv_any;
19424 temp->sv_any = NULL;
19425 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
19426 SvREFCNT_dec_NN(temp);
19427 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
19428 ing below will not set it. */
19429 SvCUR_set(ret_x, SvCUR(rx));
19432 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
19433 sv_force_normal(sv) is called. */
19435 ret = ReANY(ret_x);
19437 SvFLAGS(ret_x) |= SvUTF8(rx);
19438 /* We share the same string buffer as the original regexp, on which we
19439 hold a reference count, incremented when mother_re is set below.
19440 The string pointer is copied here, being part of the regexp struct.
19442 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
19443 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
19445 const I32 npar = r->nparens+1;
19446 Newx(ret->offs, npar, regexp_paren_pair);
19447 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19450 Newx(ret->substrs, 1, struct reg_substr_data);
19451 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19453 SvREFCNT_inc_void(ret->anchored_substr);
19454 SvREFCNT_inc_void(ret->anchored_utf8);
19455 SvREFCNT_inc_void(ret->float_substr);
19456 SvREFCNT_inc_void(ret->float_utf8);
19458 /* check_substr and check_utf8, if non-NULL, point to either their
19459 anchored or float namesakes, and don't hold a second reference. */
19461 RX_MATCH_COPIED_off(ret_x);
19462 #ifdef PERL_ANY_COW
19463 ret->saved_copy = NULL;
19465 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
19466 SvREFCNT_inc_void(ret->qr_anoncv);
19467 if (r->recurse_locinput)
19468 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19474 /* regfree_internal()
19476 Free the private data in a regexp. This is overloadable by
19477 extensions. Perl takes care of the regexp structure in pregfree(),
19478 this covers the *pprivate pointer which technically perl doesn't
19479 know about, however of course we have to handle the
19480 regexp_internal structure when no extension is in use.
19482 Note this is called before freeing anything in the regexp
19487 Perl_regfree_internal(pTHX_ REGEXP * const rx)
19489 struct regexp *const r = ReANY(rx);
19490 RXi_GET_DECL(r,ri);
19491 GET_RE_DEBUG_FLAGS_DECL;
19493 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
19499 SV *dsv= sv_newmortal();
19500 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
19501 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
19502 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
19503 PL_colors[4],PL_colors[5],s);
19506 #ifdef RE_TRACK_PATTERN_OFFSETS
19508 Safefree(ri->u.offsets); /* 20010421 MJD */
19510 if (ri->code_blocks) {
19512 for (n = 0; n < ri->num_code_blocks; n++)
19513 SvREFCNT_dec(ri->code_blocks[n].src_regex);
19514 Safefree(ri->code_blocks);
19518 int n = ri->data->count;
19521 /* If you add a ->what type here, update the comment in regcomp.h */
19522 switch (ri->data->what[n]) {
19528 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
19531 Safefree(ri->data->data[n]);
19537 { /* Aho Corasick add-on structure for a trie node.
19538 Used in stclass optimization only */
19540 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
19541 #ifdef USE_ITHREADS
19545 refcount = --aho->refcount;
19548 PerlMemShared_free(aho->states);
19549 PerlMemShared_free(aho->fail);
19550 /* do this last!!!! */
19551 PerlMemShared_free(ri->data->data[n]);
19552 /* we should only ever get called once, so
19553 * assert as much, and also guard the free
19554 * which /might/ happen twice. At the least
19555 * it will make code anlyzers happy and it
19556 * doesn't cost much. - Yves */
19557 assert(ri->regstclass);
19558 if (ri->regstclass) {
19559 PerlMemShared_free(ri->regstclass);
19560 ri->regstclass = 0;
19567 /* trie structure. */
19569 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
19570 #ifdef USE_ITHREADS
19574 refcount = --trie->refcount;
19577 PerlMemShared_free(trie->charmap);
19578 PerlMemShared_free(trie->states);
19579 PerlMemShared_free(trie->trans);
19581 PerlMemShared_free(trie->bitmap);
19583 PerlMemShared_free(trie->jump);
19584 PerlMemShared_free(trie->wordinfo);
19585 /* do this last!!!! */
19586 PerlMemShared_free(ri->data->data[n]);
19591 Perl_croak(aTHX_ "panic: regfree data code '%c'",
19592 ri->data->what[n]);
19595 Safefree(ri->data->what);
19596 Safefree(ri->data);
19602 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
19603 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
19604 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
19607 re_dup_guts - duplicate a regexp.
19609 This routine is expected to clone a given regexp structure. It is only
19610 compiled under USE_ITHREADS.
19612 After all of the core data stored in struct regexp is duplicated
19613 the regexp_engine.dupe method is used to copy any private data
19614 stored in the *pprivate pointer. This allows extensions to handle
19615 any duplication it needs to do.
19617 See pregfree() and regfree_internal() if you change anything here.
19619 #if defined(USE_ITHREADS)
19620 #ifndef PERL_IN_XSUB_RE
19622 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
19626 const struct regexp *r = ReANY(sstr);
19627 struct regexp *ret = ReANY(dstr);
19629 PERL_ARGS_ASSERT_RE_DUP_GUTS;
19631 npar = r->nparens+1;
19632 Newx(ret->offs, npar, regexp_paren_pair);
19633 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19635 if (ret->substrs) {
19636 /* Do it this way to avoid reading from *r after the StructCopy().
19637 That way, if any of the sv_dup_inc()s dislodge *r from the L1
19638 cache, it doesn't matter. */
19639 const bool anchored = r->check_substr
19640 ? r->check_substr == r->anchored_substr
19641 : r->check_utf8 == r->anchored_utf8;
19642 Newx(ret->substrs, 1, struct reg_substr_data);
19643 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19645 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
19646 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
19647 ret->float_substr = sv_dup_inc(ret->float_substr, param);
19648 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
19650 /* check_substr and check_utf8, if non-NULL, point to either their
19651 anchored or float namesakes, and don't hold a second reference. */
19653 if (ret->check_substr) {
19655 assert(r->check_utf8 == r->anchored_utf8);
19656 ret->check_substr = ret->anchored_substr;
19657 ret->check_utf8 = ret->anchored_utf8;
19659 assert(r->check_substr == r->float_substr);
19660 assert(r->check_utf8 == r->float_utf8);
19661 ret->check_substr = ret->float_substr;
19662 ret->check_utf8 = ret->float_utf8;
19664 } else if (ret->check_utf8) {
19666 ret->check_utf8 = ret->anchored_utf8;
19668 ret->check_utf8 = ret->float_utf8;
19673 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
19674 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
19675 if (r->recurse_locinput)
19676 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19679 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
19681 if (RX_MATCH_COPIED(dstr))
19682 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
19684 ret->subbeg = NULL;
19685 #ifdef PERL_ANY_COW
19686 ret->saved_copy = NULL;
19689 /* Whether mother_re be set or no, we need to copy the string. We
19690 cannot refrain from copying it when the storage points directly to
19691 our mother regexp, because that's
19692 1: a buffer in a different thread
19693 2: something we no longer hold a reference on
19694 so we need to copy it locally. */
19695 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
19696 ret->mother_re = NULL;
19698 #endif /* PERL_IN_XSUB_RE */
19703 This is the internal complement to regdupe() which is used to copy
19704 the structure pointed to by the *pprivate pointer in the regexp.
19705 This is the core version of the extension overridable cloning hook.
19706 The regexp structure being duplicated will be copied by perl prior
19707 to this and will be provided as the regexp *r argument, however
19708 with the /old/ structures pprivate pointer value. Thus this routine
19709 may override any copying normally done by perl.
19711 It returns a pointer to the new regexp_internal structure.
19715 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
19718 struct regexp *const r = ReANY(rx);
19719 regexp_internal *reti;
19721 RXi_GET_DECL(r,ri);
19723 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
19727 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
19728 char, regexp_internal);
19729 Copy(ri->program, reti->program, len+1, regnode);
19732 reti->num_code_blocks = ri->num_code_blocks;
19733 if (ri->code_blocks) {
19735 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
19736 struct reg_code_block);
19737 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
19738 struct reg_code_block);
19739 for (n = 0; n < ri->num_code_blocks; n++)
19740 reti->code_blocks[n].src_regex = (REGEXP*)
19741 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
19744 reti->code_blocks = NULL;
19746 reti->regstclass = NULL;
19749 struct reg_data *d;
19750 const int count = ri->data->count;
19753 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
19754 char, struct reg_data);
19755 Newx(d->what, count, U8);
19758 for (i = 0; i < count; i++) {
19759 d->what[i] = ri->data->what[i];
19760 switch (d->what[i]) {
19761 /* see also regcomp.h and regfree_internal() */
19762 case 'a': /* actually an AV, but the dup function is identical. */
19766 case 'u': /* actually an HV, but the dup function is identical. */
19767 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
19770 /* This is cheating. */
19771 Newx(d->data[i], 1, regnode_ssc);
19772 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
19773 reti->regstclass = (regnode*)d->data[i];
19776 /* Trie stclasses are readonly and can thus be shared
19777 * without duplication. We free the stclass in pregfree
19778 * when the corresponding reg_ac_data struct is freed.
19780 reti->regstclass= ri->regstclass;
19784 ((reg_trie_data*)ri->data->data[i])->refcount++;
19789 d->data[i] = ri->data->data[i];
19792 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
19793 ri->data->what[i]);
19802 reti->name_list_idx = ri->name_list_idx;
19804 #ifdef RE_TRACK_PATTERN_OFFSETS
19805 if (ri->u.offsets) {
19806 Newx(reti->u.offsets, 2*len+1, U32);
19807 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
19810 SetProgLen(reti,len);
19813 return (void*)reti;
19816 #endif /* USE_ITHREADS */
19818 #ifndef PERL_IN_XSUB_RE
19821 - regnext - dig the "next" pointer out of a node
19824 Perl_regnext(pTHX_ regnode *p)
19831 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
19832 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
19833 (int)OP(p), (int)REGNODE_MAX);
19836 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
19845 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
19848 STRLEN l1 = strlen(pat1);
19849 STRLEN l2 = strlen(pat2);
19852 const char *message;
19854 PERL_ARGS_ASSERT_RE_CROAK2;
19860 Copy(pat1, buf, l1 , char);
19861 Copy(pat2, buf + l1, l2 , char);
19862 buf[l1 + l2] = '\n';
19863 buf[l1 + l2 + 1] = '\0';
19864 va_start(args, pat2);
19865 msv = vmess(buf, &args);
19867 message = SvPV_const(msv,l1);
19870 Copy(message, buf, l1 , char);
19871 /* l1-1 to avoid \n */
19872 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, l1-1, buf));
19875 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
19877 #ifndef PERL_IN_XSUB_RE
19879 Perl_save_re_context(pTHX)
19884 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
19887 const REGEXP * const rx = PM_GETRE(PL_curpm);
19889 nparens = RX_NPARENS(rx);
19892 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
19893 * that PL_curpm will be null, but that utf8.pm and the modules it
19894 * loads will only use $1..$3.
19895 * The t/porting/re_context.t test file checks this assumption.
19900 for (i = 1; i <= nparens; i++) {
19901 char digits[TYPE_CHARS(long)];
19902 const STRLEN len = my_snprintf(digits, sizeof(digits),
19904 GV *const *const gvp
19905 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
19908 GV * const gv = *gvp;
19909 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
19919 S_put_code_point(pTHX_ SV *sv, UV c)
19921 PERL_ARGS_ASSERT_PUT_CODE_POINT;
19924 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
19926 else if (isPRINT(c)) {
19927 const char string = (char) c;
19929 /* We use {phrase} as metanotation in the class, so also escape literal
19931 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
19932 sv_catpvs(sv, "\\");
19933 sv_catpvn(sv, &string, 1);
19935 else if (isMNEMONIC_CNTRL(c)) {
19936 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
19939 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
19943 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
19946 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
19948 /* Appends to 'sv' a displayable version of the range of code points from
19949 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
19950 * that have them, when they occur at the beginning or end of the range.
19951 * It uses hex to output the remaining code points, unless 'allow_literals'
19952 * is true, in which case the printable ASCII ones are output as-is (though
19953 * some of these will be escaped by put_code_point()).
19955 * NOTE: This is designed only for printing ranges of code points that fit
19956 * inside an ANYOF bitmap. Higher code points are simply suppressed
19959 const unsigned int min_range_count = 3;
19961 assert(start <= end);
19963 PERL_ARGS_ASSERT_PUT_RANGE;
19965 while (start <= end) {
19967 const char * format;
19969 if (end - start < min_range_count) {
19971 /* Output chars individually when they occur in short ranges */
19972 for (; start <= end; start++) {
19973 put_code_point(sv, start);
19978 /* If permitted by the input options, and there is a possibility that
19979 * this range contains a printable literal, look to see if there is
19981 if (allow_literals && start <= MAX_PRINT_A) {
19983 /* If the character at the beginning of the range isn't an ASCII
19984 * printable, effectively split the range into two parts:
19985 * 1) the portion before the first such printable,
19987 * and output them separately. */
19988 if (! isPRINT_A(start)) {
19989 UV temp_end = start + 1;
19991 /* There is no point looking beyond the final possible
19992 * printable, in MAX_PRINT_A */
19993 UV max = MIN(end, MAX_PRINT_A);
19995 while (temp_end <= max && ! isPRINT_A(temp_end)) {
19999 /* Here, temp_end points to one beyond the first printable if
20000 * found, or to one beyond 'max' if not. If none found, make
20001 * sure that we use the entire range */
20002 if (temp_end > MAX_PRINT_A) {
20003 temp_end = end + 1;
20006 /* Output the first part of the split range: the part that
20007 * doesn't have printables, with the parameter set to not look
20008 * for literals (otherwise we would infinitely recurse) */
20009 put_range(sv, start, temp_end - 1, FALSE);
20011 /* The 2nd part of the range (if any) starts here. */
20014 /* We do a continue, instead of dropping down, because even if
20015 * the 2nd part is non-empty, it could be so short that we want
20016 * to output it as individual characters, as tested for at the
20017 * top of this loop. */
20021 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
20022 * output a sub-range of just the digits or letters, then process
20023 * the remaining portion as usual. */
20024 if (isALPHANUMERIC_A(start)) {
20025 UV mask = (isDIGIT_A(start))
20030 UV temp_end = start + 1;
20032 /* Find the end of the sub-range that includes just the
20033 * characters in the same class as the first character in it */
20034 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
20039 /* For short ranges, don't duplicate the code above to output
20040 * them; just call recursively */
20041 if (temp_end - start < min_range_count) {
20042 put_range(sv, start, temp_end, FALSE);
20044 else { /* Output as a range */
20045 put_code_point(sv, start);
20046 sv_catpvs(sv, "-");
20047 put_code_point(sv, temp_end);
20049 start = temp_end + 1;
20053 /* We output any other printables as individual characters */
20054 if (isPUNCT_A(start) || isSPACE_A(start)) {
20055 while (start <= end && (isPUNCT_A(start)
20056 || isSPACE_A(start)))
20058 put_code_point(sv, start);
20063 } /* End of looking for literals */
20065 /* Here is not to output as a literal. Some control characters have
20066 * mnemonic names. Split off any of those at the beginning and end of
20067 * the range to print mnemonically. It isn't possible for many of
20068 * these to be in a row, so this won't overwhelm with output */
20070 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
20072 while (isMNEMONIC_CNTRL(start) && start <= end) {
20073 put_code_point(sv, start);
20077 /* If this didn't take care of the whole range ... */
20078 if (start <= end) {
20080 /* Look backwards from the end to find the final non-mnemonic
20083 while (isMNEMONIC_CNTRL(temp_end)) {
20087 /* And separately output the interior range that doesn't start
20088 * or end with mnemonics */
20089 put_range(sv, start, temp_end, FALSE);
20091 /* Then output the mnemonic trailing controls */
20092 start = temp_end + 1;
20093 while (start <= end) {
20094 put_code_point(sv, start);
20101 /* As a final resort, output the range or subrange as hex. */
20103 this_end = (end < NUM_ANYOF_CODE_POINTS)
20105 : NUM_ANYOF_CODE_POINTS - 1;
20106 #if NUM_ANYOF_CODE_POINTS > 256
20107 format = (this_end < 256)
20108 ? "\\x%02" UVXf "-\\x%02" UVXf
20109 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
20111 format = "\\x%02" UVXf "-\\x%02" UVXf;
20113 GCC_DIAG_IGNORE(-Wformat-nonliteral);
20114 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
20121 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
20123 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
20127 bool allow_literals = TRUE;
20129 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
20131 /* Generally, it is more readable if printable characters are output as
20132 * literals, but if a range (nearly) spans all of them, it's best to output
20133 * it as a single range. This code will use a single range if all but 2
20134 * ASCII printables are in it */
20135 invlist_iterinit(invlist);
20136 while (invlist_iternext(invlist, &start, &end)) {
20138 /* If the range starts beyond the final printable, it doesn't have any
20140 if (start > MAX_PRINT_A) {
20144 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
20145 * all but two, the range must start and end no later than 2 from
20147 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
20148 if (end > MAX_PRINT_A) {
20154 if (end - start >= MAX_PRINT_A - ' ' - 2) {
20155 allow_literals = FALSE;
20160 invlist_iterfinish(invlist);
20162 /* Here we have figured things out. Output each range */
20163 invlist_iterinit(invlist);
20164 while (invlist_iternext(invlist, &start, &end)) {
20165 if (start >= NUM_ANYOF_CODE_POINTS) {
20168 put_range(sv, start, end, allow_literals);
20170 invlist_iterfinish(invlist);
20176 S_put_charclass_bitmap_innards_common(pTHX_
20177 SV* invlist, /* The bitmap */
20178 SV* posixes, /* Under /l, things like [:word:], \S */
20179 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
20180 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
20181 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
20182 const bool invert /* Is the result to be inverted? */
20185 /* Create and return an SV containing a displayable version of the bitmap
20186 * and associated information determined by the input parameters. If the
20187 * output would have been only the inversion indicator '^', NULL is instead
20192 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
20195 output = newSVpvs("^");
20198 output = newSVpvs("");
20201 /* First, the code points in the bitmap that are unconditionally there */
20202 put_charclass_bitmap_innards_invlist(output, invlist);
20204 /* Traditionally, these have been placed after the main code points */
20206 sv_catsv(output, posixes);
20209 if (only_utf8 && _invlist_len(only_utf8)) {
20210 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
20211 put_charclass_bitmap_innards_invlist(output, only_utf8);
20214 if (not_utf8 && _invlist_len(not_utf8)) {
20215 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
20216 put_charclass_bitmap_innards_invlist(output, not_utf8);
20219 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
20220 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
20221 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
20223 /* This is the only list in this routine that can legally contain code
20224 * points outside the bitmap range. The call just above to
20225 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
20226 * output them here. There's about a half-dozen possible, and none in
20227 * contiguous ranges longer than 2 */
20228 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20230 SV* above_bitmap = NULL;
20232 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
20234 invlist_iterinit(above_bitmap);
20235 while (invlist_iternext(above_bitmap, &start, &end)) {
20238 for (i = start; i <= end; i++) {
20239 put_code_point(output, i);
20242 invlist_iterfinish(above_bitmap);
20243 SvREFCNT_dec_NN(above_bitmap);
20247 if (invert && SvCUR(output) == 1) {
20255 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
20257 SV *nonbitmap_invlist,
20258 SV *only_utf8_locale_invlist,
20259 const regnode * const node,
20260 const bool force_as_is_display)
20262 /* Appends to 'sv' a displayable version of the innards of the bracketed
20263 * character class defined by the other arguments:
20264 * 'bitmap' points to the bitmap.
20265 * 'nonbitmap_invlist' is an inversion list of the code points that are in
20266 * the bitmap range, but for some reason aren't in the bitmap; NULL if
20267 * none. The reasons for this could be that they require some
20268 * condition such as the target string being or not being in UTF-8
20269 * (under /d), or because they came from a user-defined property that
20270 * was not resolved at the time of the regex compilation (under /u)
20271 * 'only_utf8_locale_invlist' is an inversion list of the code points that
20272 * are valid only if the runtime locale is a UTF-8 one; NULL if none
20273 * 'node' is the regex pattern node. It is needed only when the above two
20274 * parameters are not null, and is passed so that this routine can
20275 * tease apart the various reasons for them.
20276 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
20277 * to invert things to see if that leads to a cleaner display. If
20278 * FALSE, this routine is free to use its judgment about doing this.
20280 * It returns TRUE if there was actually something output. (It may be that
20281 * the bitmap, etc is empty.)
20283 * When called for outputting the bitmap of a non-ANYOF node, just pass the
20284 * bitmap, with the succeeding parameters set to NULL, and the final one to
20288 /* In general, it tries to display the 'cleanest' representation of the
20289 * innards, choosing whether to display them inverted or not, regardless of
20290 * whether the class itself is to be inverted. However, there are some
20291 * cases where it can't try inverting, as what actually matches isn't known
20292 * until runtime, and hence the inversion isn't either. */
20293 bool inverting_allowed = ! force_as_is_display;
20296 STRLEN orig_sv_cur = SvCUR(sv);
20298 SV* invlist; /* Inversion list we accumulate of code points that
20299 are unconditionally matched */
20300 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
20302 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
20304 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
20305 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
20308 SV* as_is_display; /* The output string when we take the inputs
20310 SV* inverted_display; /* The output string when we invert the inputs */
20312 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
20314 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
20316 /* We are biased in favor of displaying things without them being inverted,
20317 * as that is generally easier to understand */
20318 const int bias = 5;
20320 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
20322 /* Start off with whatever code points are passed in. (We clone, so we
20323 * don't change the caller's list) */
20324 if (nonbitmap_invlist) {
20325 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
20326 invlist = invlist_clone(nonbitmap_invlist);
20328 else { /* Worst case size is every other code point is matched */
20329 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
20333 if (OP(node) == ANYOFD) {
20335 /* This flag indicates that the code points below 0x100 in the
20336 * nonbitmap list are precisely the ones that match only when the
20337 * target is UTF-8 (they should all be non-ASCII). */
20338 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
20340 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
20341 _invlist_subtract(invlist, only_utf8, &invlist);
20344 /* And this flag for matching all non-ASCII 0xFF and below */
20345 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
20347 not_utf8 = invlist_clone(PL_UpperLatin1);
20350 else if (OP(node) == ANYOFL) {
20352 /* If either of these flags are set, what matches isn't
20353 * determinable except during execution, so don't know enough here
20355 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
20356 inverting_allowed = FALSE;
20359 /* What the posix classes match also varies at runtime, so these
20360 * will be output symbolically. */
20361 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
20364 posixes = newSVpvs("");
20365 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
20366 if (ANYOF_POSIXL_TEST(node,i)) {
20367 sv_catpv(posixes, anyofs[i]);
20374 /* Accumulate the bit map into the unconditional match list */
20375 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
20376 if (BITMAP_TEST(bitmap, i)) {
20378 for (; i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i); i++) {
20381 invlist = _add_range_to_invlist(invlist, start, i-1);
20385 /* Make sure that the conditional match lists don't have anything in them
20386 * that match unconditionally; otherwise the output is quite confusing.
20387 * This could happen if the code that populates these misses some
20390 _invlist_subtract(only_utf8, invlist, &only_utf8);
20393 _invlist_subtract(not_utf8, invlist, ¬_utf8);
20396 if (only_utf8_locale_invlist) {
20398 /* Since this list is passed in, we have to make a copy before
20400 only_utf8_locale = invlist_clone(only_utf8_locale_invlist);
20402 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
20404 /* And, it can get really weird for us to try outputting an inverted
20405 * form of this list when it has things above the bitmap, so don't even
20407 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20408 inverting_allowed = FALSE;
20412 /* Calculate what the output would be if we take the input as-is */
20413 as_is_display = put_charclass_bitmap_innards_common(invlist,
20420 /* If have to take the output as-is, just do that */
20421 if (! inverting_allowed) {
20422 if (as_is_display) {
20423 sv_catsv(sv, as_is_display);
20424 SvREFCNT_dec_NN(as_is_display);
20427 else { /* But otherwise, create the output again on the inverted input, and
20428 use whichever version is shorter */
20430 int inverted_bias, as_is_bias;
20432 /* We will apply our bias to whichever of the the results doesn't have
20442 inverted_bias = bias;
20445 /* Now invert each of the lists that contribute to the output,
20446 * excluding from the result things outside the possible range */
20448 /* For the unconditional inversion list, we have to add in all the
20449 * conditional code points, so that when inverted, they will be gone
20451 _invlist_union(only_utf8, invlist, &invlist);
20452 _invlist_union(not_utf8, invlist, &invlist);
20453 _invlist_union(only_utf8_locale, invlist, &invlist);
20454 _invlist_invert(invlist);
20455 _invlist_intersection(invlist, PL_InBitmap, &invlist);
20458 _invlist_invert(only_utf8);
20459 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
20461 else if (not_utf8) {
20463 /* If a code point matches iff the target string is not in UTF-8,
20464 * then complementing the result has it not match iff not in UTF-8,
20465 * which is the same thing as matching iff it is UTF-8. */
20466 only_utf8 = not_utf8;
20470 if (only_utf8_locale) {
20471 _invlist_invert(only_utf8_locale);
20472 _invlist_intersection(only_utf8_locale,
20474 &only_utf8_locale);
20477 inverted_display = put_charclass_bitmap_innards_common(
20482 only_utf8_locale, invert);
20484 /* Use the shortest representation, taking into account our bias
20485 * against showing it inverted */
20486 if ( inverted_display
20487 && ( ! as_is_display
20488 || ( SvCUR(inverted_display) + inverted_bias
20489 < SvCUR(as_is_display) + as_is_bias)))
20491 sv_catsv(sv, inverted_display);
20493 else if (as_is_display) {
20494 sv_catsv(sv, as_is_display);
20497 SvREFCNT_dec(as_is_display);
20498 SvREFCNT_dec(inverted_display);
20501 SvREFCNT_dec_NN(invlist);
20502 SvREFCNT_dec(only_utf8);
20503 SvREFCNT_dec(not_utf8);
20504 SvREFCNT_dec(posixes);
20505 SvREFCNT_dec(only_utf8_locale);
20507 return SvCUR(sv) > orig_sv_cur;
20510 #define CLEAR_OPTSTART \
20511 if (optstart) STMT_START { \
20512 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
20513 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
20517 #define DUMPUNTIL(b,e) \
20519 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
20521 STATIC const regnode *
20522 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
20523 const regnode *last, const regnode *plast,
20524 SV* sv, I32 indent, U32 depth)
20526 U8 op = PSEUDO; /* Arbitrary non-END op. */
20527 const regnode *next;
20528 const regnode *optstart= NULL;
20530 RXi_GET_DECL(r,ri);
20531 GET_RE_DEBUG_FLAGS_DECL;
20533 PERL_ARGS_ASSERT_DUMPUNTIL;
20535 #ifdef DEBUG_DUMPUNTIL
20536 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n",indent,node-start,
20537 last ? last-start : 0,plast ? plast-start : 0);
20540 if (plast && plast < last)
20543 while (PL_regkind[op] != END && (!last || node < last)) {
20545 /* While that wasn't END last time... */
20548 if (op == CLOSE || op == WHILEM)
20550 next = regnext((regnode *)node);
20553 if (OP(node) == OPTIMIZED) {
20554 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
20561 regprop(r, sv, node, NULL, NULL);
20562 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
20563 (int)(2*indent + 1), "", SvPVX_const(sv));
20565 if (OP(node) != OPTIMIZED) {
20566 if (next == NULL) /* Next ptr. */
20567 Perl_re_printf( aTHX_ " (0)");
20568 else if (PL_regkind[(U8)op] == BRANCH
20569 && PL_regkind[OP(next)] != BRANCH )
20570 Perl_re_printf( aTHX_ " (FAIL)");
20572 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
20573 Perl_re_printf( aTHX_ "\n");
20577 if (PL_regkind[(U8)op] == BRANCHJ) {
20580 const regnode *nnode = (OP(next) == LONGJMP
20581 ? regnext((regnode *)next)
20583 if (last && nnode > last)
20585 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
20588 else if (PL_regkind[(U8)op] == BRANCH) {
20590 DUMPUNTIL(NEXTOPER(node), next);
20592 else if ( PL_regkind[(U8)op] == TRIE ) {
20593 const regnode *this_trie = node;
20594 const char op = OP(node);
20595 const U32 n = ARG(node);
20596 const reg_ac_data * const ac = op>=AHOCORASICK ?
20597 (reg_ac_data *)ri->data->data[n] :
20599 const reg_trie_data * const trie =
20600 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
20602 AV *const trie_words
20603 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
20605 const regnode *nextbranch= NULL;
20608 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
20609 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
20611 Perl_re_indentf( aTHX_ "%s ",
20614 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
20615 SvCUR(*elem_ptr), 60,
20616 PL_colors[0], PL_colors[1],
20618 ? PERL_PV_ESCAPE_UNI
20620 | PERL_PV_PRETTY_ELLIPSES
20621 | PERL_PV_PRETTY_LTGT
20626 U16 dist= trie->jump[word_idx+1];
20627 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
20628 (UV)((dist ? this_trie + dist : next) - start));
20631 nextbranch= this_trie + trie->jump[0];
20632 DUMPUNTIL(this_trie + dist, nextbranch);
20634 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
20635 nextbranch= regnext((regnode *)nextbranch);
20637 Perl_re_printf( aTHX_ "\n");
20640 if (last && next > last)
20645 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
20646 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
20647 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
20649 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
20651 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
20653 else if ( op == PLUS || op == STAR) {
20654 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
20656 else if (PL_regkind[(U8)op] == ANYOF) {
20657 /* arglen 1 + class block */
20658 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
20659 ? ANYOF_POSIXL_SKIP
20661 node = NEXTOPER(node);
20663 else if (PL_regkind[(U8)op] == EXACT) {
20664 /* Literal string, where present. */
20665 node += NODE_SZ_STR(node) - 1;
20666 node = NEXTOPER(node);
20669 node = NEXTOPER(node);
20670 node += regarglen[(U8)op];
20672 if (op == CURLYX || op == OPEN)
20676 #ifdef DEBUG_DUMPUNTIL
20677 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
20682 #endif /* DEBUGGING */
20685 * ex: set ts=8 sts=4 sw=4 et: