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_blocks *code_blocks;/* positions of literal (?{})
187 int code_index; /* next code_blocks[] slot */
188 SSize_t maxlen; /* mininum possible number of chars in string to match */
189 scan_frame *frame_head;
190 scan_frame *frame_last;
193 #ifdef ADD_TO_REGEXEC
194 char *starttry; /* -Dr: where regtry was called. */
195 #define RExC_starttry (pRExC_state->starttry)
197 SV *runtime_code_qr; /* qr with the runtime code blocks */
199 const char *lastparse;
201 AV *paren_name_list; /* idx -> name */
202 U32 study_chunk_recursed_count;
205 #define RExC_lastparse (pRExC_state->lastparse)
206 #define RExC_lastnum (pRExC_state->lastnum)
207 #define RExC_paren_name_list (pRExC_state->paren_name_list)
208 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
209 #define RExC_mysv (pRExC_state->mysv1)
210 #define RExC_mysv1 (pRExC_state->mysv1)
211 #define RExC_mysv2 (pRExC_state->mysv2)
214 bool seen_unfolded_sharp_s;
219 #define RExC_flags (pRExC_state->flags)
220 #define RExC_pm_flags (pRExC_state->pm_flags)
221 #define RExC_precomp (pRExC_state->precomp)
222 #define RExC_precomp_adj (pRExC_state->precomp_adj)
223 #define RExC_adjusted_start (pRExC_state->adjusted_start)
224 #define RExC_precomp_end (pRExC_state->precomp_end)
225 #define RExC_rx_sv (pRExC_state->rx_sv)
226 #define RExC_rx (pRExC_state->rx)
227 #define RExC_rxi (pRExC_state->rxi)
228 #define RExC_start (pRExC_state->start)
229 #define RExC_end (pRExC_state->end)
230 #define RExC_parse (pRExC_state->parse)
231 #define RExC_whilem_seen (pRExC_state->whilem_seen)
233 /* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
234 * EXACTF node, hence was parsed under /di rules. If later in the parse,
235 * something forces the pattern into using /ui rules, the sharp s should be
236 * folded into the sequence 'ss', which takes up more space than previously
237 * calculated. This means that the sizing pass needs to be restarted. (The
238 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
239 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
240 * so there is no need to resize [perl #125990]. */
241 #define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
243 #ifdef RE_TRACK_PATTERN_OFFSETS
244 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
247 #define RExC_emit (pRExC_state->emit)
248 #define RExC_emit_dummy (pRExC_state->emit_dummy)
249 #define RExC_emit_start (pRExC_state->emit_start)
250 #define RExC_emit_bound (pRExC_state->emit_bound)
251 #define RExC_sawback (pRExC_state->sawback)
252 #define RExC_seen (pRExC_state->seen)
253 #define RExC_size (pRExC_state->size)
254 #define RExC_maxlen (pRExC_state->maxlen)
255 #define RExC_npar (pRExC_state->npar)
256 #define RExC_nestroot (pRExC_state->nestroot)
257 #define RExC_extralen (pRExC_state->extralen)
258 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
259 #define RExC_utf8 (pRExC_state->utf8)
260 #define RExC_uni_semantics (pRExC_state->uni_semantics)
261 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
262 #define RExC_open_parens (pRExC_state->open_parens)
263 #define RExC_close_parens (pRExC_state->close_parens)
264 #define RExC_end_op (pRExC_state->end_op)
265 #define RExC_paren_names (pRExC_state->paren_names)
266 #define RExC_recurse (pRExC_state->recurse)
267 #define RExC_recurse_count (pRExC_state->recurse_count)
268 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
269 #define RExC_study_chunk_recursed_bytes \
270 (pRExC_state->study_chunk_recursed_bytes)
271 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
272 #define RExC_contains_locale (pRExC_state->contains_locale)
274 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
276 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
277 #define RExC_frame_head (pRExC_state->frame_head)
278 #define RExC_frame_last (pRExC_state->frame_last)
279 #define RExC_frame_count (pRExC_state->frame_count)
280 #define RExC_strict (pRExC_state->strict)
281 #define RExC_study_started (pRExC_state->study_started)
282 #define RExC_warn_text (pRExC_state->warn_text)
284 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
285 * a flag to disable back-off on the fixed/floating substrings - if it's
286 * a high complexity pattern we assume the benefit of avoiding a full match
287 * is worth the cost of checking for the substrings even if they rarely help.
289 #define RExC_naughty (pRExC_state->naughty)
290 #define TOO_NAUGHTY (10)
291 #define MARK_NAUGHTY(add) \
292 if (RExC_naughty < TOO_NAUGHTY) \
293 RExC_naughty += (add)
294 #define MARK_NAUGHTY_EXP(exp, add) \
295 if (RExC_naughty < TOO_NAUGHTY) \
296 RExC_naughty += RExC_naughty / (exp) + (add)
298 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
299 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
300 ((*s) == '{' && regcurly(s)))
303 * Flags to be passed up and down.
305 #define WORST 0 /* Worst case. */
306 #define HASWIDTH 0x01 /* Known to match non-null strings. */
308 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
309 * character. (There needs to be a case: in the switch statement in regexec.c
310 * for any node marked SIMPLE.) Note that this is not the same thing as
313 #define SPSTART 0x04 /* Starts with * or + */
314 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
315 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
316 #define RESTART_PASS1 0x20 /* Need to restart sizing pass */
317 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
318 calcuate sizes as UTF-8 */
320 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
322 /* whether trie related optimizations are enabled */
323 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
324 #define TRIE_STUDY_OPT
325 #define FULL_TRIE_STUDY
331 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
332 #define PBITVAL(paren) (1 << ((paren) & 7))
333 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
334 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
335 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
337 #define REQUIRE_UTF8(flagp) STMT_START { \
340 *flagp = RESTART_PASS1|NEED_UTF8; \
345 /* Change from /d into /u rules, and restart the parse if we've already seen
346 * something whose size would increase as a result, by setting *flagp and
347 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
348 * we've change to /u during the parse. */
349 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
351 if (DEPENDS_SEMANTICS) { \
353 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
354 RExC_uni_semantics = 1; \
355 if (RExC_seen_unfolded_sharp_s) { \
356 *flagp |= RESTART_PASS1; \
357 return restart_retval; \
362 /* This converts the named class defined in regcomp.h to its equivalent class
363 * number defined in handy.h. */
364 #define namedclass_to_classnum(class) ((int) ((class) / 2))
365 #define classnum_to_namedclass(classnum) ((classnum) * 2)
367 #define _invlist_union_complement_2nd(a, b, output) \
368 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
369 #define _invlist_intersection_complement_2nd(a, b, output) \
370 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
372 /* About scan_data_t.
374 During optimisation we recurse through the regexp program performing
375 various inplace (keyhole style) optimisations. In addition study_chunk
376 and scan_commit populate this data structure with information about
377 what strings MUST appear in the pattern. We look for the longest
378 string that must appear at a fixed location, and we look for the
379 longest string that may appear at a floating location. So for instance
384 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
385 strings (because they follow a .* construct). study_chunk will identify
386 both FOO and BAR as being the longest fixed and floating strings respectively.
388 The strings can be composites, for instance
392 will result in a composite fixed substring 'foo'.
394 For each string some basic information is maintained:
396 - offset or min_offset
397 This is the position the string must appear at, or not before.
398 It also implicitly (when combined with minlenp) tells us how many
399 characters must match before the string we are searching for.
400 Likewise when combined with minlenp and the length of the string it
401 tells us how many characters must appear after the string we have
405 Only used for floating strings. This is the rightmost point that
406 the string can appear at. If set to SSize_t_MAX it indicates that the
407 string can occur infinitely far to the right.
410 A pointer to the minimum number of characters of the pattern that the
411 string was found inside. This is important as in the case of positive
412 lookahead or positive lookbehind we can have multiple patterns
417 The minimum length of the pattern overall is 3, the minimum length
418 of the lookahead part is 3, but the minimum length of the part that
419 will actually match is 1. So 'FOO's minimum length is 3, but the
420 minimum length for the F is 1. This is important as the minimum length
421 is used to determine offsets in front of and behind the string being
422 looked for. Since strings can be composites this is the length of the
423 pattern at the time it was committed with a scan_commit. Note that
424 the length is calculated by study_chunk, so that the minimum lengths
425 are not known until the full pattern has been compiled, thus the
426 pointer to the value.
430 In the case of lookbehind the string being searched for can be
431 offset past the start point of the final matching string.
432 If this value was just blithely removed from the min_offset it would
433 invalidate some of the calculations for how many chars must match
434 before or after (as they are derived from min_offset and minlen and
435 the length of the string being searched for).
436 When the final pattern is compiled and the data is moved from the
437 scan_data_t structure into the regexp structure the information
438 about lookbehind is factored in, with the information that would
439 have been lost precalculated in the end_shift field for the
442 The fields pos_min and pos_delta are used to store the minimum offset
443 and the delta to the maximum offset at the current point in the pattern.
447 typedef struct scan_data_t {
448 /*I32 len_min; unused */
449 /*I32 len_delta; unused */
453 SSize_t last_end; /* min value, <0 unless valid. */
454 SSize_t last_start_min;
455 SSize_t last_start_max;
456 SV **longest; /* Either &l_fixed, or &l_float. */
457 SV *longest_fixed; /* longest fixed string found in pattern */
458 SSize_t offset_fixed; /* offset where it starts */
459 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
460 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
461 SV *longest_float; /* longest floating string found in pattern */
462 SSize_t offset_float_min; /* earliest point in string it can appear */
463 SSize_t offset_float_max; /* latest point in string it can appear */
464 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
465 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
468 SSize_t *last_closep;
469 regnode_ssc *start_class;
473 * Forward declarations for pregcomp()'s friends.
476 static const scan_data_t zero_scan_data =
477 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
479 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
480 #define SF_BEFORE_SEOL 0x0001
481 #define SF_BEFORE_MEOL 0x0002
482 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
483 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
485 #define SF_FIX_SHIFT_EOL (+2)
486 #define SF_FL_SHIFT_EOL (+4)
488 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
489 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
491 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
492 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
493 #define SF_IS_INF 0x0040
494 #define SF_HAS_PAR 0x0080
495 #define SF_IN_PAR 0x0100
496 #define SF_HAS_EVAL 0x0200
499 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
500 * longest substring in the pattern. When it is not set the optimiser keeps
501 * track of position, but does not keep track of the actual strings seen,
503 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
506 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
507 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
508 * turned off because of the alternation (BRANCH). */
509 #define SCF_DO_SUBSTR 0x0400
511 #define SCF_DO_STCLASS_AND 0x0800
512 #define SCF_DO_STCLASS_OR 0x1000
513 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
514 #define SCF_WHILEM_VISITED_POS 0x2000
516 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
517 #define SCF_SEEN_ACCEPT 0x8000
518 #define SCF_TRIE_DOING_RESTUDY 0x10000
519 #define SCF_IN_DEFINE 0x20000
524 #define UTF cBOOL(RExC_utf8)
526 /* The enums for all these are ordered so things work out correctly */
527 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
528 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
529 == REGEX_DEPENDS_CHARSET)
530 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
531 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
532 >= REGEX_UNICODE_CHARSET)
533 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
534 == REGEX_ASCII_RESTRICTED_CHARSET)
535 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
536 >= REGEX_ASCII_RESTRICTED_CHARSET)
537 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
538 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
540 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
542 /* For programs that want to be strictly Unicode compatible by dying if any
543 * attempt is made to match a non-Unicode code point against a Unicode
545 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
547 #define OOB_NAMEDCLASS -1
549 /* There is no code point that is out-of-bounds, so this is problematic. But
550 * its only current use is to initialize a variable that is always set before
552 #define OOB_UNICODE 0xDEADBEEF
554 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
557 /* length of regex to show in messages that don't mark a position within */
558 #define RegexLengthToShowInErrorMessages 127
561 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
562 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
563 * op/pragma/warn/regcomp.
565 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
566 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
568 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
569 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
571 /* The code in this file in places uses one level of recursion with parsing
572 * rebased to an alternate string constructed by us in memory. This can take
573 * the form of something that is completely different from the input, or
574 * something that uses the input as part of the alternate. In the first case,
575 * there should be no possibility of an error, as we are in complete control of
576 * the alternate string. But in the second case we don't control the input
577 * portion, so there may be errors in that. Here's an example:
579 * is handled specially because \x{df} folds to a sequence of more than one
580 * character, 'ss'. What is done is to create and parse an alternate string,
581 * which looks like this:
582 * /(?:\x{DF}|[abc\x{DF}def])/ui
583 * where it uses the input unchanged in the middle of something it constructs,
584 * which is a branch for the DF outside the character class, and clustering
585 * parens around the whole thing. (It knows enough to skip the DF inside the
586 * class while in this substitute parse.) 'abc' and 'def' may have errors that
587 * need to be reported. The general situation looks like this:
590 * Input: ----------------------------------------------------
591 * Constructed: ---------------------------------------------------
594 * The input string sI..eI is the input pattern. The string sC..EC is the
595 * constructed substitute parse string. The portions sC..tC and eC..EC are
596 * constructed by us. The portion tC..eC is an exact duplicate of the input
597 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
598 * while parsing, we find an error at xC. We want to display a message showing
599 * the real input string. Thus we need to find the point xI in it which
600 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
601 * been constructed by us, and so shouldn't have errors. We get:
603 * xI = sI + (tI - sI) + (xC - tC)
605 * and, the offset into sI is:
607 * (xI - sI) = (tI - sI) + (xC - tC)
609 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
610 * and we save tC as RExC_adjusted_start.
612 * During normal processing of the input pattern, everything points to that,
613 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
616 #define tI_sI RExC_precomp_adj
617 #define tC RExC_adjusted_start
618 #define sC RExC_precomp
619 #define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
620 #define xI(xC) (sC + xI_offset(xC))
621 #define eC RExC_precomp_end
623 #define REPORT_LOCATION_ARGS(xC) \
625 (xI(xC) > eC) /* Don't run off end */ \
626 ? eC - sC /* Length before the <--HERE */ \
628 sC), /* The input pattern printed up to the <--HERE */ \
630 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
631 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
633 /* Used to point after bad bytes for an error message, but avoid skipping
634 * past a nul byte. */
635 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
638 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
639 * arg. Show regex, up to a maximum length. If it's too long, chop and add
642 #define _FAIL(code) STMT_START { \
643 const char *ellipses = ""; \
644 IV len = RExC_precomp_end - RExC_precomp; \
647 SAVEFREESV(RExC_rx_sv); \
648 if (len > RegexLengthToShowInErrorMessages) { \
649 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
650 len = RegexLengthToShowInErrorMessages - 10; \
656 #define FAIL(msg) _FAIL( \
657 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
658 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
660 #define FAIL2(msg,arg) _FAIL( \
661 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
662 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
665 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
667 #define Simple_vFAIL(m) STMT_START { \
668 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
669 m, REPORT_LOCATION_ARGS(RExC_parse)); \
673 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
675 #define vFAIL(m) STMT_START { \
677 SAVEFREESV(RExC_rx_sv); \
682 * Like Simple_vFAIL(), but accepts two arguments.
684 #define Simple_vFAIL2(m,a1) STMT_START { \
685 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
686 REPORT_LOCATION_ARGS(RExC_parse)); \
690 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
692 #define vFAIL2(m,a1) STMT_START { \
694 SAVEFREESV(RExC_rx_sv); \
695 Simple_vFAIL2(m, a1); \
700 * Like Simple_vFAIL(), but accepts three arguments.
702 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
703 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
704 REPORT_LOCATION_ARGS(RExC_parse)); \
708 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
710 #define vFAIL3(m,a1,a2) STMT_START { \
712 SAVEFREESV(RExC_rx_sv); \
713 Simple_vFAIL3(m, a1, a2); \
717 * Like Simple_vFAIL(), but accepts four arguments.
719 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
720 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
721 REPORT_LOCATION_ARGS(RExC_parse)); \
724 #define vFAIL4(m,a1,a2,a3) STMT_START { \
726 SAVEFREESV(RExC_rx_sv); \
727 Simple_vFAIL4(m, a1, a2, a3); \
730 /* A specialized version of vFAIL2 that works with UTF8f */
731 #define vFAIL2utf8f(m, a1) STMT_START { \
733 SAVEFREESV(RExC_rx_sv); \
734 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
735 REPORT_LOCATION_ARGS(RExC_parse)); \
738 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
740 SAVEFREESV(RExC_rx_sv); \
741 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
742 REPORT_LOCATION_ARGS(RExC_parse)); \
745 /* These have asserts in them because of [perl #122671] Many warnings in
746 * regcomp.c can occur twice. If they get output in pass1 and later in that
747 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
748 * would get output again. So they should be output in pass2, and these
749 * asserts make sure new warnings follow that paradigm. */
751 /* m is not necessarily a "literal string", in this macro */
752 #define reg_warn_non_literal_string(loc, m) STMT_START { \
753 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
754 "%s" REPORT_LOCATION, \
755 m, REPORT_LOCATION_ARGS(loc)); \
758 #define ckWARNreg(loc,m) STMT_START { \
759 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
761 REPORT_LOCATION_ARGS(loc)); \
764 #define vWARN(loc, m) STMT_START { \
765 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
767 REPORT_LOCATION_ARGS(loc)); \
770 #define vWARN_dep(loc, m) STMT_START { \
771 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
773 REPORT_LOCATION_ARGS(loc)); \
776 #define ckWARNdep(loc,m) STMT_START { \
777 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
779 REPORT_LOCATION_ARGS(loc)); \
782 #define ckWARNregdep(loc,m) STMT_START { \
783 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
786 REPORT_LOCATION_ARGS(loc)); \
789 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
790 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
792 a1, REPORT_LOCATION_ARGS(loc)); \
795 #define ckWARN2reg(loc, m, a1) STMT_START { \
796 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
798 a1, REPORT_LOCATION_ARGS(loc)); \
801 #define vWARN3(loc, m, a1, a2) STMT_START { \
802 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
804 a1, a2, REPORT_LOCATION_ARGS(loc)); \
807 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
808 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
811 REPORT_LOCATION_ARGS(loc)); \
814 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
815 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
818 REPORT_LOCATION_ARGS(loc)); \
821 #define vWARN4dep(loc, m, a1, a2, a3) STMT_START { \
822 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN2(WARN_REGEXP,WARN_DEPRECATED), \
825 REPORT_LOCATION_ARGS(loc)); \
828 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
829 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
832 REPORT_LOCATION_ARGS(loc)); \
835 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
836 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
839 REPORT_LOCATION_ARGS(loc)); \
842 /* Macros for recording node offsets. 20001227 mjd@plover.com
843 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
844 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
845 * Element 0 holds the number n.
846 * Position is 1 indexed.
848 #ifndef RE_TRACK_PATTERN_OFFSETS
849 #define Set_Node_Offset_To_R(node,byte)
850 #define Set_Node_Offset(node,byte)
851 #define Set_Cur_Node_Offset
852 #define Set_Node_Length_To_R(node,len)
853 #define Set_Node_Length(node,len)
854 #define Set_Node_Cur_Length(node,start)
855 #define Node_Offset(n)
856 #define Node_Length(n)
857 #define Set_Node_Offset_Length(node,offset,len)
858 #define ProgLen(ri) ri->u.proglen
859 #define SetProgLen(ri,x) ri->u.proglen = x
861 #define ProgLen(ri) ri->u.offsets[0]
862 #define SetProgLen(ri,x) ri->u.offsets[0] = x
863 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
865 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
866 __LINE__, (int)(node), (int)(byte))); \
868 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
871 RExC_offsets[2*(node)-1] = (byte); \
876 #define Set_Node_Offset(node,byte) \
877 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
878 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
880 #define Set_Node_Length_To_R(node,len) STMT_START { \
882 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
883 __LINE__, (int)(node), (int)(len))); \
885 Perl_croak(aTHX_ "value of node is %d in Length macro", \
888 RExC_offsets[2*(node)] = (len); \
893 #define Set_Node_Length(node,len) \
894 Set_Node_Length_To_R((node)-RExC_emit_start, len)
895 #define Set_Node_Cur_Length(node, start) \
896 Set_Node_Length(node, RExC_parse - start)
898 /* Get offsets and lengths */
899 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
900 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
902 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
903 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
904 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
908 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
909 #define EXPERIMENTAL_INPLACESCAN
910 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
914 Perl_re_printf(pTHX_ const char *fmt, ...)
918 PerlIO *f= Perl_debug_log;
919 PERL_ARGS_ASSERT_RE_PRINTF;
921 result = PerlIO_vprintf(f, fmt, ap);
927 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
931 PerlIO *f= Perl_debug_log;
932 PERL_ARGS_ASSERT_RE_INDENTF;
934 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
935 result = PerlIO_vprintf(f, fmt, ap);
939 #endif /* DEBUGGING */
941 #define DEBUG_RExC_seen() \
942 DEBUG_OPTIMISE_MORE_r({ \
943 Perl_re_printf( aTHX_ "RExC_seen: "); \
945 if (RExC_seen & REG_ZERO_LEN_SEEN) \
946 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
948 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
949 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
951 if (RExC_seen & REG_GPOS_SEEN) \
952 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
954 if (RExC_seen & REG_RECURSE_SEEN) \
955 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
957 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
958 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
960 if (RExC_seen & REG_VERBARG_SEEN) \
961 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
963 if (RExC_seen & REG_CUTGROUP_SEEN) \
964 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
966 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
967 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
969 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
970 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
972 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
973 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
975 Perl_re_printf( aTHX_ "\n"); \
978 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
979 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
981 #define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
983 Perl_re_printf( aTHX_ "%s", open_str); \
984 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
985 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
986 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
987 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
988 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
989 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
990 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
991 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
992 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
993 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
994 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
995 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
996 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
997 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
998 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
999 Perl_re_printf( aTHX_ "%s", close_str); \
1003 #define DEBUG_STUDYDATA(str,data,depth) \
1004 DEBUG_OPTIMISE_MORE_r(if(data){ \
1005 Perl_re_indentf( aTHX_ "" str "Pos:%" IVdf "/%" IVdf \
1006 " Flags: 0x%" UVXf, \
1008 (IV)((data)->pos_min), \
1009 (IV)((data)->pos_delta), \
1010 (UV)((data)->flags) \
1012 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
1013 Perl_re_printf( aTHX_ \
1014 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s", \
1015 (IV)((data)->whilem_c), \
1016 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
1017 is_inf ? "INF " : "" \
1019 if ((data)->last_found) \
1020 Perl_re_printf( aTHX_ \
1021 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf \
1022 " %sFixed:'%s' @ %" IVdf \
1023 " %sFloat: '%s' @ %" IVdf "/%" IVdf, \
1024 SvPVX_const((data)->last_found), \
1025 (IV)((data)->last_end), \
1026 (IV)((data)->last_start_min), \
1027 (IV)((data)->last_start_max), \
1028 ((data)->longest && \
1029 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
1030 SvPVX_const((data)->longest_fixed), \
1031 (IV)((data)->offset_fixed), \
1032 ((data)->longest && \
1033 (data)->longest==&((data)->longest_float)) ? "*" : "", \
1034 SvPVX_const((data)->longest_float), \
1035 (IV)((data)->offset_float_min), \
1036 (IV)((data)->offset_float_max) \
1038 Perl_re_printf( aTHX_ "\n"); \
1042 /* =========================================================
1043 * BEGIN edit_distance stuff.
1045 * This calculates how many single character changes of any type are needed to
1046 * transform a string into another one. It is taken from version 3.1 of
1048 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1051 /* Our unsorted dictionary linked list. */
1052 /* Note we use UVs, not chars. */
1057 struct dictionary* next;
1059 typedef struct dictionary item;
1062 PERL_STATIC_INLINE item*
1063 push(UV key,item* curr)
1066 Newxz(head, 1, item);
1074 PERL_STATIC_INLINE item*
1075 find(item* head, UV key)
1077 item* iterator = head;
1079 if (iterator->key == key){
1082 iterator = iterator->next;
1088 PERL_STATIC_INLINE item*
1089 uniquePush(item* head,UV key)
1091 item* iterator = head;
1094 if (iterator->key == key) {
1097 iterator = iterator->next;
1100 return push(key,head);
1103 PERL_STATIC_INLINE void
1104 dict_free(item* head)
1106 item* iterator = head;
1109 item* temp = iterator;
1110 iterator = iterator->next;
1117 /* End of Dictionary Stuff */
1119 /* All calculations/work are done here */
1121 S_edit_distance(const UV* src,
1123 const STRLEN x, /* length of src[] */
1124 const STRLEN y, /* length of tgt[] */
1125 const SSize_t maxDistance
1129 UV swapCount,swapScore,targetCharCount,i,j;
1131 UV score_ceil = x + y;
1133 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1135 /* intialize matrix start values */
1136 Newxz(scores, ( (x + 2) * (y + 2)), UV);
1137 scores[0] = score_ceil;
1138 scores[1 * (y + 2) + 0] = score_ceil;
1139 scores[0 * (y + 2) + 1] = score_ceil;
1140 scores[1 * (y + 2) + 1] = 0;
1141 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1146 for (i=1;i<=x;i++) {
1148 head = uniquePush(head,src[i]);
1149 scores[(i+1) * (y + 2) + 1] = i;
1150 scores[(i+1) * (y + 2) + 0] = score_ceil;
1153 for (j=1;j<=y;j++) {
1156 head = uniquePush(head,tgt[j]);
1157 scores[1 * (y + 2) + (j + 1)] = j;
1158 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1161 targetCharCount = find(head,tgt[j-1])->value;
1162 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1164 if (src[i-1] != tgt[j-1]){
1165 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));
1169 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1173 find(head,src[i-1])->value = i;
1177 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1180 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1184 /* END of edit_distance() stuff
1185 * ========================================================= */
1187 /* is c a control character for which we have a mnemonic? */
1188 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1191 S_cntrl_to_mnemonic(const U8 c)
1193 /* Returns the mnemonic string that represents character 'c', if one
1194 * exists; NULL otherwise. The only ones that exist for the purposes of
1195 * this routine are a few control characters */
1198 case '\a': return "\\a";
1199 case '\b': return "\\b";
1200 case ESC_NATIVE: return "\\e";
1201 case '\f': return "\\f";
1202 case '\n': return "\\n";
1203 case '\r': return "\\r";
1204 case '\t': return "\\t";
1210 /* Mark that we cannot extend a found fixed substring at this point.
1211 Update the longest found anchored substring and the longest found
1212 floating substrings if needed. */
1215 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1216 SSize_t *minlenp, int is_inf)
1218 const STRLEN l = CHR_SVLEN(data->last_found);
1219 const STRLEN old_l = CHR_SVLEN(*data->longest);
1220 GET_RE_DEBUG_FLAGS_DECL;
1222 PERL_ARGS_ASSERT_SCAN_COMMIT;
1224 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1225 SvSetMagicSV(*data->longest, data->last_found);
1226 if (*data->longest == data->longest_fixed) {
1227 data->offset_fixed = l ? data->last_start_min : data->pos_min;
1228 if (data->flags & SF_BEFORE_EOL)
1230 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
1232 data->flags &= ~SF_FIX_BEFORE_EOL;
1233 data->minlen_fixed=minlenp;
1234 data->lookbehind_fixed=0;
1236 else { /* *data->longest == data->longest_float */
1237 data->offset_float_min = l ? data->last_start_min : data->pos_min;
1238 data->offset_float_max = (l
1239 ? data->last_start_max
1240 : (data->pos_delta > SSize_t_MAX - data->pos_min
1242 : data->pos_min + data->pos_delta));
1244 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
1245 data->offset_float_max = SSize_t_MAX;
1246 if (data->flags & SF_BEFORE_EOL)
1248 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
1250 data->flags &= ~SF_FL_BEFORE_EOL;
1251 data->minlen_float=minlenp;
1252 data->lookbehind_float=0;
1255 SvCUR_set(data->last_found, 0);
1257 SV * const sv = data->last_found;
1258 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1259 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1264 data->last_end = -1;
1265 data->flags &= ~SF_BEFORE_EOL;
1266 DEBUG_STUDYDATA("commit: ",data,0);
1269 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1270 * list that describes which code points it matches */
1273 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1275 /* Set the SSC 'ssc' to match an empty string or any code point */
1277 PERL_ARGS_ASSERT_SSC_ANYTHING;
1279 assert(is_ANYOF_SYNTHETIC(ssc));
1281 /* mortalize so won't leak */
1282 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1283 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1287 S_ssc_is_anything(const regnode_ssc *ssc)
1289 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1290 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1291 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1292 * in any way, so there's no point in using it */
1297 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1299 assert(is_ANYOF_SYNTHETIC(ssc));
1301 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1305 /* See if the list consists solely of the range 0 - Infinity */
1306 invlist_iterinit(ssc->invlist);
1307 ret = invlist_iternext(ssc->invlist, &start, &end)
1311 invlist_iterfinish(ssc->invlist);
1317 /* If e.g., both \w and \W are set, matches everything */
1318 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1320 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1321 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1331 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1333 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1334 * string, any code point, or any posix class under locale */
1336 PERL_ARGS_ASSERT_SSC_INIT;
1338 Zero(ssc, 1, regnode_ssc);
1339 set_ANYOF_SYNTHETIC(ssc);
1340 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1343 /* If any portion of the regex is to operate under locale rules that aren't
1344 * fully known at compile time, initialization includes it. The reason
1345 * this isn't done for all regexes is that the optimizer was written under
1346 * the assumption that locale was all-or-nothing. Given the complexity and
1347 * lack of documentation in the optimizer, and that there are inadequate
1348 * test cases for locale, many parts of it may not work properly, it is
1349 * safest to avoid locale unless necessary. */
1350 if (RExC_contains_locale) {
1351 ANYOF_POSIXL_SETALL(ssc);
1354 ANYOF_POSIXL_ZERO(ssc);
1359 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1360 const regnode_ssc *ssc)
1362 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1363 * to the list of code points matched, and locale posix classes; hence does
1364 * not check its flags) */
1369 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1371 assert(is_ANYOF_SYNTHETIC(ssc));
1373 invlist_iterinit(ssc->invlist);
1374 ret = invlist_iternext(ssc->invlist, &start, &end)
1378 invlist_iterfinish(ssc->invlist);
1384 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1392 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1393 const regnode_charclass* const node)
1395 /* Returns a mortal inversion list defining which code points are matched
1396 * by 'node', which is of type ANYOF. Handles complementing the result if
1397 * appropriate. If some code points aren't knowable at this time, the
1398 * returned list must, and will, contain every code point that is a
1402 SV* only_utf8_locale_invlist = NULL;
1404 const U32 n = ARG(node);
1405 bool new_node_has_latin1 = FALSE;
1407 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1409 /* Look at the data structure created by S_set_ANYOF_arg() */
1410 if (n != ANYOF_ONLY_HAS_BITMAP) {
1411 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1412 AV * const av = MUTABLE_AV(SvRV(rv));
1413 SV **const ary = AvARRAY(av);
1414 assert(RExC_rxi->data->what[n] == 's');
1416 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1417 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1419 else if (ary[0] && ary[0] != &PL_sv_undef) {
1421 /* Here, no compile-time swash, and there are things that won't be
1422 * known until runtime -- we have to assume it could be anything */
1423 invlist = sv_2mortal(_new_invlist(1));
1424 return _add_range_to_invlist(invlist, 0, UV_MAX);
1426 else if (ary[3] && ary[3] != &PL_sv_undef) {
1428 /* Here no compile-time swash, and no run-time only data. Use the
1429 * node's inversion list */
1430 invlist = sv_2mortal(invlist_clone(ary[3]));
1433 /* Get the code points valid only under UTF-8 locales */
1434 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1435 && ary[2] && ary[2] != &PL_sv_undef)
1437 only_utf8_locale_invlist = ary[2];
1442 invlist = sv_2mortal(_new_invlist(0));
1445 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1446 * code points, and an inversion list for the others, but if there are code
1447 * points that should match only conditionally on the target string being
1448 * UTF-8, those are placed in the inversion list, and not the bitmap.
1449 * Since there are circumstances under which they could match, they are
1450 * included in the SSC. But if the ANYOF node is to be inverted, we have
1451 * to exclude them here, so that when we invert below, the end result
1452 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1453 * have to do this here before we add the unconditionally matched code
1455 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1456 _invlist_intersection_complement_2nd(invlist,
1461 /* Add in the points from the bit map */
1462 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1463 if (ANYOF_BITMAP_TEST(node, i)) {
1464 unsigned int start = i++;
1466 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1469 invlist = _add_range_to_invlist(invlist, start, i-1);
1470 new_node_has_latin1 = TRUE;
1474 /* If this can match all upper Latin1 code points, have to add them
1475 * as well. But don't add them if inverting, as when that gets done below,
1476 * it would exclude all these characters, including the ones it shouldn't
1477 * that were added just above */
1478 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1479 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1481 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1484 /* Similarly for these */
1485 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1486 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1489 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1490 _invlist_invert(invlist);
1492 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1494 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1495 * locale. We can skip this if there are no 0-255 at all. */
1496 _invlist_union(invlist, PL_Latin1, &invlist);
1499 /* Similarly add the UTF-8 locale possible matches. These have to be
1500 * deferred until after the non-UTF-8 locale ones are taken care of just
1501 * above, or it leads to wrong results under ANYOF_INVERT */
1502 if (only_utf8_locale_invlist) {
1503 _invlist_union_maybe_complement_2nd(invlist,
1504 only_utf8_locale_invlist,
1505 ANYOF_FLAGS(node) & ANYOF_INVERT,
1512 /* These two functions currently do the exact same thing */
1513 #define ssc_init_zero ssc_init
1515 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1516 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1518 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1519 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1520 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1523 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1524 const regnode_charclass *and_with)
1526 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1527 * another SSC or a regular ANYOF class. Can create false positives. */
1532 PERL_ARGS_ASSERT_SSC_AND;
1534 assert(is_ANYOF_SYNTHETIC(ssc));
1536 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1537 * the code point inversion list and just the relevant flags */
1538 if (is_ANYOF_SYNTHETIC(and_with)) {
1539 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1540 anded_flags = ANYOF_FLAGS(and_with);
1542 /* XXX This is a kludge around what appears to be deficiencies in the
1543 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1544 * there are paths through the optimizer where it doesn't get weeded
1545 * out when it should. And if we don't make some extra provision for
1546 * it like the code just below, it doesn't get added when it should.
1547 * This solution is to add it only when AND'ing, which is here, and
1548 * only when what is being AND'ed is the pristine, original node
1549 * matching anything. Thus it is like adding it to ssc_anything() but
1550 * only when the result is to be AND'ed. Probably the same solution
1551 * could be adopted for the same problem we have with /l matching,
1552 * which is solved differently in S_ssc_init(), and that would lead to
1553 * fewer false positives than that solution has. But if this solution
1554 * creates bugs, the consequences are only that a warning isn't raised
1555 * that should be; while the consequences for having /l bugs is
1556 * incorrect matches */
1557 if (ssc_is_anything((regnode_ssc *)and_with)) {
1558 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1562 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1563 if (OP(and_with) == ANYOFD) {
1564 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1567 anded_flags = ANYOF_FLAGS(and_with)
1568 &( ANYOF_COMMON_FLAGS
1569 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1570 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1571 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1573 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1578 ANYOF_FLAGS(ssc) &= anded_flags;
1580 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1581 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1582 * 'and_with' may be inverted. When not inverted, we have the situation of
1584 * (C1 | P1) & (C2 | P2)
1585 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1586 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1587 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1588 * <= ((C1 & C2) | P1 | P2)
1589 * Alternatively, the last few steps could be:
1590 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1591 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1592 * <= (C1 | C2 | (P1 & P2))
1593 * We favor the second approach if either P1 or P2 is non-empty. This is
1594 * because these components are a barrier to doing optimizations, as what
1595 * they match cannot be known until the moment of matching as they are
1596 * dependent on the current locale, 'AND"ing them likely will reduce or
1598 * But we can do better if we know that C1,P1 are in their initial state (a
1599 * frequent occurrence), each matching everything:
1600 * (<everything>) & (C2 | P2) = C2 | P2
1601 * Similarly, if C2,P2 are in their initial state (again a frequent
1602 * occurrence), the result is a no-op
1603 * (C1 | P1) & (<everything>) = C1 | P1
1606 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1607 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1608 * <= (C1 & ~C2) | (P1 & ~P2)
1611 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1612 && ! is_ANYOF_SYNTHETIC(and_with))
1616 ssc_intersection(ssc,
1618 FALSE /* Has already been inverted */
1621 /* If either P1 or P2 is empty, the intersection will be also; can skip
1623 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1624 ANYOF_POSIXL_ZERO(ssc);
1626 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1628 /* Note that the Posix class component P from 'and_with' actually
1630 * P = Pa | Pb | ... | Pn
1631 * where each component is one posix class, such as in [\w\s].
1633 * ~P = ~(Pa | Pb | ... | Pn)
1634 * = ~Pa & ~Pb & ... & ~Pn
1635 * <= ~Pa | ~Pb | ... | ~Pn
1636 * The last is something we can easily calculate, but unfortunately
1637 * is likely to have many false positives. We could do better
1638 * in some (but certainly not all) instances if two classes in
1639 * P have known relationships. For example
1640 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1642 * :lower: & :print: = :lower:
1643 * And similarly for classes that must be disjoint. For example,
1644 * since \s and \w can have no elements in common based on rules in
1645 * the POSIX standard,
1646 * \w & ^\S = nothing
1647 * Unfortunately, some vendor locales do not meet the Posix
1648 * standard, in particular almost everything by Microsoft.
1649 * The loop below just changes e.g., \w into \W and vice versa */
1651 regnode_charclass_posixl temp;
1652 int add = 1; /* To calculate the index of the complement */
1654 ANYOF_POSIXL_ZERO(&temp);
1655 for (i = 0; i < ANYOF_MAX; i++) {
1657 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1658 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1660 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1661 ANYOF_POSIXL_SET(&temp, i + add);
1663 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1665 ANYOF_POSIXL_AND(&temp, ssc);
1667 } /* else ssc already has no posixes */
1668 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1669 in its initial state */
1670 else if (! is_ANYOF_SYNTHETIC(and_with)
1671 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1673 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1674 * copy it over 'ssc' */
1675 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1676 if (is_ANYOF_SYNTHETIC(and_with)) {
1677 StructCopy(and_with, ssc, regnode_ssc);
1680 ssc->invlist = anded_cp_list;
1681 ANYOF_POSIXL_ZERO(ssc);
1682 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1683 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1687 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1688 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1690 /* One or the other of P1, P2 is non-empty. */
1691 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1692 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1694 ssc_union(ssc, anded_cp_list, FALSE);
1696 else { /* P1 = P2 = empty */
1697 ssc_intersection(ssc, anded_cp_list, FALSE);
1703 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1704 const regnode_charclass *or_with)
1706 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1707 * another SSC or a regular ANYOF class. Can create false positives if
1708 * 'or_with' is to be inverted. */
1713 PERL_ARGS_ASSERT_SSC_OR;
1715 assert(is_ANYOF_SYNTHETIC(ssc));
1717 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1718 * the code point inversion list and just the relevant flags */
1719 if (is_ANYOF_SYNTHETIC(or_with)) {
1720 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1721 ored_flags = ANYOF_FLAGS(or_with);
1724 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1725 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1726 if (OP(or_with) != ANYOFD) {
1728 |= ANYOF_FLAGS(or_with)
1729 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1730 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1731 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1733 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1738 ANYOF_FLAGS(ssc) |= ored_flags;
1740 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1741 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1742 * 'or_with' may be inverted. When not inverted, we have the simple
1743 * situation of computing:
1744 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1745 * If P1|P2 yields a situation with both a class and its complement are
1746 * set, like having both \w and \W, this matches all code points, and we
1747 * can delete these from the P component of the ssc going forward. XXX We
1748 * might be able to delete all the P components, but I (khw) am not certain
1749 * about this, and it is better to be safe.
1752 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1753 * <= (C1 | P1) | ~C2
1754 * <= (C1 | ~C2) | P1
1755 * (which results in actually simpler code than the non-inverted case)
1758 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1759 && ! is_ANYOF_SYNTHETIC(or_with))
1761 /* We ignore P2, leaving P1 going forward */
1762 } /* else Not inverted */
1763 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1764 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1765 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1767 for (i = 0; i < ANYOF_MAX; i += 2) {
1768 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1770 ssc_match_all_cp(ssc);
1771 ANYOF_POSIXL_CLEAR(ssc, i);
1772 ANYOF_POSIXL_CLEAR(ssc, i+1);
1780 FALSE /* Already has been inverted */
1784 PERL_STATIC_INLINE void
1785 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1787 PERL_ARGS_ASSERT_SSC_UNION;
1789 assert(is_ANYOF_SYNTHETIC(ssc));
1791 _invlist_union_maybe_complement_2nd(ssc->invlist,
1797 PERL_STATIC_INLINE void
1798 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1800 const bool invert2nd)
1802 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1804 assert(is_ANYOF_SYNTHETIC(ssc));
1806 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1812 PERL_STATIC_INLINE void
1813 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1815 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1817 assert(is_ANYOF_SYNTHETIC(ssc));
1819 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1822 PERL_STATIC_INLINE void
1823 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1825 /* AND just the single code point 'cp' into the SSC 'ssc' */
1827 SV* cp_list = _new_invlist(2);
1829 PERL_ARGS_ASSERT_SSC_CP_AND;
1831 assert(is_ANYOF_SYNTHETIC(ssc));
1833 cp_list = add_cp_to_invlist(cp_list, cp);
1834 ssc_intersection(ssc, cp_list,
1835 FALSE /* Not inverted */
1837 SvREFCNT_dec_NN(cp_list);
1840 PERL_STATIC_INLINE void
1841 S_ssc_clear_locale(regnode_ssc *ssc)
1843 /* Set the SSC 'ssc' to not match any locale things */
1844 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1846 assert(is_ANYOF_SYNTHETIC(ssc));
1848 ANYOF_POSIXL_ZERO(ssc);
1849 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1852 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1855 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1857 /* The synthetic start class is used to hopefully quickly winnow down
1858 * places where a pattern could start a match in the target string. If it
1859 * doesn't really narrow things down that much, there isn't much point to
1860 * having the overhead of using it. This function uses some very crude
1861 * heuristics to decide if to use the ssc or not.
1863 * It returns TRUE if 'ssc' rules out more than half what it considers to
1864 * be the "likely" possible matches, but of course it doesn't know what the
1865 * actual things being matched are going to be; these are only guesses
1867 * For /l matches, it assumes that the only likely matches are going to be
1868 * in the 0-255 range, uniformly distributed, so half of that is 127
1869 * For /a and /d matches, it assumes that the likely matches will be just
1870 * the ASCII range, so half of that is 63
1871 * For /u and there isn't anything matching above the Latin1 range, it
1872 * assumes that that is the only range likely to be matched, and uses
1873 * half that as the cut-off: 127. If anything matches above Latin1,
1874 * it assumes that all of Unicode could match (uniformly), except for
1875 * non-Unicode code points and things in the General Category "Other"
1876 * (unassigned, private use, surrogates, controls and formats). This
1877 * is a much large number. */
1879 U32 count = 0; /* Running total of number of code points matched by
1881 UV start, end; /* Start and end points of current range in inversion
1883 const U32 max_code_points = (LOC)
1885 : (( ! UNI_SEMANTICS
1886 || invlist_highest(ssc->invlist) < 256)
1889 const U32 max_match = max_code_points / 2;
1891 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1893 invlist_iterinit(ssc->invlist);
1894 while (invlist_iternext(ssc->invlist, &start, &end)) {
1895 if (start >= max_code_points) {
1898 end = MIN(end, max_code_points - 1);
1899 count += end - start + 1;
1900 if (count >= max_match) {
1901 invlist_iterfinish(ssc->invlist);
1911 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1913 /* The inversion list in the SSC is marked mortal; now we need a more
1914 * permanent copy, which is stored the same way that is done in a regular
1915 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1918 SV* invlist = invlist_clone(ssc->invlist);
1920 PERL_ARGS_ASSERT_SSC_FINALIZE;
1922 assert(is_ANYOF_SYNTHETIC(ssc));
1924 /* The code in this file assumes that all but these flags aren't relevant
1925 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1926 * by the time we reach here */
1927 assert(! (ANYOF_FLAGS(ssc)
1928 & ~( ANYOF_COMMON_FLAGS
1929 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1930 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
1932 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1934 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1935 NULL, NULL, NULL, FALSE);
1937 /* Make sure is clone-safe */
1938 ssc->invlist = NULL;
1940 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1941 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1944 if (RExC_contains_locale) {
1948 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1951 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1952 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1953 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1954 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1955 ? (TRIE_LIST_CUR( idx ) - 1) \
1961 dump_trie(trie,widecharmap,revcharmap)
1962 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1963 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1965 These routines dump out a trie in a somewhat readable format.
1966 The _interim_ variants are used for debugging the interim
1967 tables that are used to generate the final compressed
1968 representation which is what dump_trie expects.
1970 Part of the reason for their existence is to provide a form
1971 of documentation as to how the different representations function.
1976 Dumps the final compressed table form of the trie to Perl_debug_log.
1977 Used for debugging make_trie().
1981 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1982 AV *revcharmap, U32 depth)
1985 SV *sv=sv_newmortal();
1986 int colwidth= widecharmap ? 6 : 4;
1988 GET_RE_DEBUG_FLAGS_DECL;
1990 PERL_ARGS_ASSERT_DUMP_TRIE;
1992 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
1993 depth+1, "Match","Base","Ofs" );
1995 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1996 SV ** const tmp = av_fetch( revcharmap, state, 0);
1998 Perl_re_printf( aTHX_ "%*s",
2000 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2001 PL_colors[0], PL_colors[1],
2002 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2003 PERL_PV_ESCAPE_FIRSTCHAR
2008 Perl_re_printf( aTHX_ "\n");
2009 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2011 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2012 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2013 Perl_re_printf( aTHX_ "\n");
2015 for( state = 1 ; state < trie->statecount ; state++ ) {
2016 const U32 base = trie->states[ state ].trans.base;
2018 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2020 if ( trie->states[ state ].wordnum ) {
2021 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2023 Perl_re_printf( aTHX_ "%6s", "" );
2026 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2031 while( ( base + ofs < trie->uniquecharcount ) ||
2032 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2033 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2037 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2039 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2040 if ( ( base + ofs >= trie->uniquecharcount )
2041 && ( base + ofs - trie->uniquecharcount
2043 && trie->trans[ base + ofs
2044 - trie->uniquecharcount ].check == state )
2046 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2047 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2050 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
2054 Perl_re_printf( aTHX_ "]");
2057 Perl_re_printf( aTHX_ "\n" );
2059 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2061 for (word=1; word <= trie->wordcount; word++) {
2062 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2063 (int)word, (int)(trie->wordinfo[word].prev),
2064 (int)(trie->wordinfo[word].len));
2066 Perl_re_printf( aTHX_ "\n" );
2069 Dumps a fully constructed but uncompressed trie in list form.
2070 List tries normally only are used for construction when the number of
2071 possible chars (trie->uniquecharcount) is very high.
2072 Used for debugging make_trie().
2075 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2076 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2080 SV *sv=sv_newmortal();
2081 int colwidth= widecharmap ? 6 : 4;
2082 GET_RE_DEBUG_FLAGS_DECL;
2084 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2086 /* print out the table precompression. */
2087 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2089 Perl_re_indentf( aTHX_ "%s",
2090 depth+1, "------:-----+-----------------\n" );
2092 for( state=1 ; state < next_alloc ; state ++ ) {
2095 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2096 depth+1, (UV)state );
2097 if ( ! trie->states[ state ].wordnum ) {
2098 Perl_re_printf( aTHX_ "%5s| ","");
2100 Perl_re_printf( aTHX_ "W%4x| ",
2101 trie->states[ state ].wordnum
2104 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2105 SV ** const tmp = av_fetch( revcharmap,
2106 TRIE_LIST_ITEM(state,charid).forid, 0);
2108 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2110 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2112 PL_colors[0], PL_colors[1],
2113 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2114 | PERL_PV_ESCAPE_FIRSTCHAR
2116 TRIE_LIST_ITEM(state,charid).forid,
2117 (UV)TRIE_LIST_ITEM(state,charid).newstate
2120 Perl_re_printf( aTHX_ "\n%*s| ",
2121 (int)((depth * 2) + 14), "");
2124 Perl_re_printf( aTHX_ "\n");
2129 Dumps a fully constructed but uncompressed trie in table form.
2130 This is the normal DFA style state transition table, with a few
2131 twists to facilitate compression later.
2132 Used for debugging make_trie().
2135 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2136 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2141 SV *sv=sv_newmortal();
2142 int colwidth= widecharmap ? 6 : 4;
2143 GET_RE_DEBUG_FLAGS_DECL;
2145 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2148 print out the table precompression so that we can do a visual check
2149 that they are identical.
2152 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2154 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2155 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2157 Perl_re_printf( aTHX_ "%*s",
2159 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2160 PL_colors[0], PL_colors[1],
2161 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2162 PERL_PV_ESCAPE_FIRSTCHAR
2168 Perl_re_printf( aTHX_ "\n");
2169 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2171 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2172 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2175 Perl_re_printf( aTHX_ "\n" );
2177 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2179 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2181 (UV)TRIE_NODENUM( state ) );
2183 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2184 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2186 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2188 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2190 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2191 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2192 (UV)trie->trans[ state ].check );
2194 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2195 (UV)trie->trans[ state ].check,
2196 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2204 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2205 startbranch: the first branch in the whole branch sequence
2206 first : start branch of sequence of branch-exact nodes.
2207 May be the same as startbranch
2208 last : Thing following the last branch.
2209 May be the same as tail.
2210 tail : item following the branch sequence
2211 count : words in the sequence
2212 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2213 depth : indent depth
2215 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2217 A trie is an N'ary tree where the branches are determined by digital
2218 decomposition of the key. IE, at the root node you look up the 1st character and
2219 follow that branch repeat until you find the end of the branches. Nodes can be
2220 marked as "accepting" meaning they represent a complete word. Eg:
2224 would convert into the following structure. Numbers represent states, letters
2225 following numbers represent valid transitions on the letter from that state, if
2226 the number is in square brackets it represents an accepting state, otherwise it
2227 will be in parenthesis.
2229 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2233 (1) +-i->(6)-+-s->[7]
2235 +-s->(3)-+-h->(4)-+-e->[5]
2237 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2239 This shows that when matching against the string 'hers' we will begin at state 1
2240 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2241 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2242 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2243 single traverse. We store a mapping from accepting to state to which word was
2244 matched, and then when we have multiple possibilities we try to complete the
2245 rest of the regex in the order in which they occurred in the alternation.
2247 The only prior NFA like behaviour that would be changed by the TRIE support is
2248 the silent ignoring of duplicate alternations which are of the form:
2250 / (DUPE|DUPE) X? (?{ ... }) Y /x
2252 Thus EVAL blocks following a trie may be called a different number of times with
2253 and without the optimisation. With the optimisations dupes will be silently
2254 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2255 the following demonstrates:
2257 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2259 which prints out 'word' three times, but
2261 'words'=~/(word|word|word)(?{ print $1 })S/
2263 which doesnt print it out at all. This is due to other optimisations kicking in.
2265 Example of what happens on a structural level:
2267 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2269 1: CURLYM[1] {1,32767}(18)
2280 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2281 and should turn into:
2283 1: CURLYM[1] {1,32767}(18)
2285 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2293 Cases where tail != last would be like /(?foo|bar)baz/:
2303 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2304 and would end up looking like:
2307 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2314 d = uvchr_to_utf8_flags(d, uv, 0);
2316 is the recommended Unicode-aware way of saying
2321 #define TRIE_STORE_REVCHAR(val) \
2324 SV *zlopp = newSV(UTF8_MAXBYTES); \
2325 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2326 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2327 SvCUR_set(zlopp, kapow - flrbbbbb); \
2330 av_push(revcharmap, zlopp); \
2332 char ooooff = (char)val; \
2333 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2337 /* This gets the next character from the input, folding it if not already
2339 #define TRIE_READ_CHAR STMT_START { \
2342 /* if it is UTF then it is either already folded, or does not need \
2344 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2346 else if (folder == PL_fold_latin1) { \
2347 /* This folder implies Unicode rules, which in the range expressible \
2348 * by not UTF is the lower case, with the two exceptions, one of \
2349 * which should have been taken care of before calling this */ \
2350 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2351 uvc = toLOWER_L1(*uc); \
2352 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2355 /* raw data, will be folded later if needed */ \
2363 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2364 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2365 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2366 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2368 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2369 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2370 TRIE_LIST_CUR( state )++; \
2373 #define TRIE_LIST_NEW(state) STMT_START { \
2374 Newxz( trie->states[ state ].trans.list, \
2375 4, reg_trie_trans_le ); \
2376 TRIE_LIST_CUR( state ) = 1; \
2377 TRIE_LIST_LEN( state ) = 4; \
2380 #define TRIE_HANDLE_WORD(state) STMT_START { \
2381 U16 dupe= trie->states[ state ].wordnum; \
2382 regnode * const noper_next = regnext( noper ); \
2385 /* store the word for dumping */ \
2387 if (OP(noper) != NOTHING) \
2388 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2390 tmp = newSVpvn_utf8( "", 0, UTF ); \
2391 av_push( trie_words, tmp ); \
2395 trie->wordinfo[curword].prev = 0; \
2396 trie->wordinfo[curword].len = wordlen; \
2397 trie->wordinfo[curword].accept = state; \
2399 if ( noper_next < tail ) { \
2401 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2403 trie->jump[curword] = (U16)(noper_next - convert); \
2405 jumper = noper_next; \
2407 nextbranch= regnext(cur); \
2411 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2412 /* chain, so that when the bits of chain are later */\
2413 /* linked together, the dups appear in the chain */\
2414 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2415 trie->wordinfo[dupe].prev = curword; \
2417 /* we haven't inserted this word yet. */ \
2418 trie->states[ state ].wordnum = curword; \
2423 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2424 ( ( base + charid >= ucharcount \
2425 && base + charid < ubound \
2426 && state == trie->trans[ base - ucharcount + charid ].check \
2427 && trie->trans[ base - ucharcount + charid ].next ) \
2428 ? trie->trans[ base - ucharcount + charid ].next \
2429 : ( state==1 ? special : 0 ) \
2432 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2434 TRIE_BITMAP_SET(trie, uvc); \
2435 /* store the folded codepoint */ \
2437 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2440 /* store first byte of utf8 representation of */ \
2441 /* variant codepoints */ \
2442 if (! UVCHR_IS_INVARIANT(uvc)) { \
2443 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2448 #define MADE_JUMP_TRIE 2
2449 #define MADE_EXACT_TRIE 4
2452 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2453 regnode *first, regnode *last, regnode *tail,
2454 U32 word_count, U32 flags, U32 depth)
2456 /* first pass, loop through and scan words */
2457 reg_trie_data *trie;
2458 HV *widecharmap = NULL;
2459 AV *revcharmap = newAV();
2465 regnode *jumper = NULL;
2466 regnode *nextbranch = NULL;
2467 regnode *convert = NULL;
2468 U32 *prev_states; /* temp array mapping each state to previous one */
2469 /* we just use folder as a flag in utf8 */
2470 const U8 * folder = NULL;
2473 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2474 AV *trie_words = NULL;
2475 /* along with revcharmap, this only used during construction but both are
2476 * useful during debugging so we store them in the struct when debugging.
2479 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2480 STRLEN trie_charcount=0;
2482 SV *re_trie_maxbuff;
2483 GET_RE_DEBUG_FLAGS_DECL;
2485 PERL_ARGS_ASSERT_MAKE_TRIE;
2487 PERL_UNUSED_ARG(depth);
2491 case EXACT: case EXACTL: break;
2495 case EXACTFLU8: folder = PL_fold_latin1; break;
2496 case EXACTF: folder = PL_fold; break;
2497 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2500 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2502 trie->startstate = 1;
2503 trie->wordcount = word_count;
2504 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2505 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2506 if (flags == EXACT || flags == EXACTL)
2507 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2508 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2509 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2512 trie_words = newAV();
2515 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2516 assert(re_trie_maxbuff);
2517 if (!SvIOK(re_trie_maxbuff)) {
2518 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2520 DEBUG_TRIE_COMPILE_r({
2521 Perl_re_indentf( aTHX_
2522 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2524 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2525 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2528 /* Find the node we are going to overwrite */
2529 if ( first == startbranch && OP( last ) != BRANCH ) {
2530 /* whole branch chain */
2533 /* branch sub-chain */
2534 convert = NEXTOPER( first );
2537 /* -- First loop and Setup --
2539 We first traverse the branches and scan each word to determine if it
2540 contains widechars, and how many unique chars there are, this is
2541 important as we have to build a table with at least as many columns as we
2544 We use an array of integers to represent the character codes 0..255
2545 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2546 the native representation of the character value as the key and IV's for
2549 *TODO* If we keep track of how many times each character is used we can
2550 remap the columns so that the table compression later on is more
2551 efficient in terms of memory by ensuring the most common value is in the
2552 middle and the least common are on the outside. IMO this would be better
2553 than a most to least common mapping as theres a decent chance the most
2554 common letter will share a node with the least common, meaning the node
2555 will not be compressible. With a middle is most common approach the worst
2556 case is when we have the least common nodes twice.
2560 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2561 regnode *noper = NEXTOPER( cur );
2565 U32 wordlen = 0; /* required init */
2566 STRLEN minchars = 0;
2567 STRLEN maxchars = 0;
2568 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2571 if (OP(noper) == NOTHING) {
2572 /* skip past a NOTHING at the start of an alternation
2573 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2575 regnode *noper_next= regnext(noper);
2576 if (noper_next < tail)
2580 if ( noper < tail &&
2582 OP(noper) == flags ||
2585 OP(noper) == EXACTFU_SS
2589 uc= (U8*)STRING(noper);
2590 e= uc + STR_LEN(noper);
2597 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2598 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2599 regardless of encoding */
2600 if (OP( noper ) == EXACTFU_SS) {
2601 /* false positives are ok, so just set this */
2602 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2606 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2608 TRIE_CHARCOUNT(trie)++;
2611 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2612 * is in effect. Under /i, this character can match itself, or
2613 * anything that folds to it. If not under /i, it can match just
2614 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2615 * all fold to k, and all are single characters. But some folds
2616 * expand to more than one character, so for example LATIN SMALL
2617 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2618 * the string beginning at 'uc' is 'ffi', it could be matched by
2619 * three characters, or just by the one ligature character. (It
2620 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2621 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2622 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2623 * match.) The trie needs to know the minimum and maximum number
2624 * of characters that could match so that it can use size alone to
2625 * quickly reject many match attempts. The max is simple: it is
2626 * the number of folded characters in this branch (since a fold is
2627 * never shorter than what folds to it. */
2631 /* And the min is equal to the max if not under /i (indicated by
2632 * 'folder' being NULL), or there are no multi-character folds. If
2633 * there is a multi-character fold, the min is incremented just
2634 * once, for the character that folds to the sequence. Each
2635 * character in the sequence needs to be added to the list below of
2636 * characters in the trie, but we count only the first towards the
2637 * min number of characters needed. This is done through the
2638 * variable 'foldlen', which is returned by the macros that look
2639 * for these sequences as the number of bytes the sequence
2640 * occupies. Each time through the loop, we decrement 'foldlen' by
2641 * how many bytes the current char occupies. Only when it reaches
2642 * 0 do we increment 'minchars' or look for another multi-character
2644 if (folder == NULL) {
2647 else if (foldlen > 0) {
2648 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2653 /* See if *uc is the beginning of a multi-character fold. If
2654 * so, we decrement the length remaining to look at, to account
2655 * for the current character this iteration. (We can use 'uc'
2656 * instead of the fold returned by TRIE_READ_CHAR because for
2657 * non-UTF, the latin1_safe macro is smart enough to account
2658 * for all the unfolded characters, and because for UTF, the
2659 * string will already have been folded earlier in the
2660 * compilation process */
2662 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2663 foldlen -= UTF8SKIP(uc);
2666 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2671 /* The current character (and any potential folds) should be added
2672 * to the possible matching characters for this position in this
2676 U8 folded= folder[ (U8) uvc ];
2677 if ( !trie->charmap[ folded ] ) {
2678 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2679 TRIE_STORE_REVCHAR( folded );
2682 if ( !trie->charmap[ uvc ] ) {
2683 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2684 TRIE_STORE_REVCHAR( uvc );
2687 /* store the codepoint in the bitmap, and its folded
2689 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2690 set_bit = 0; /* We've done our bit :-) */
2694 /* XXX We could come up with the list of code points that fold
2695 * to this using PL_utf8_foldclosures, except not for
2696 * multi-char folds, as there may be multiple combinations
2697 * there that could work, which needs to wait until runtime to
2698 * resolve (The comment about LIGATURE FFI above is such an
2703 widecharmap = newHV();
2705 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2708 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2710 if ( !SvTRUE( *svpp ) ) {
2711 sv_setiv( *svpp, ++trie->uniquecharcount );
2712 TRIE_STORE_REVCHAR(uvc);
2715 } /* end loop through characters in this branch of the trie */
2717 /* We take the min and max for this branch and combine to find the min
2718 * and max for all branches processed so far */
2719 if( cur == first ) {
2720 trie->minlen = minchars;
2721 trie->maxlen = maxchars;
2722 } else if (minchars < trie->minlen) {
2723 trie->minlen = minchars;
2724 } else if (maxchars > trie->maxlen) {
2725 trie->maxlen = maxchars;
2727 } /* end first pass */
2728 DEBUG_TRIE_COMPILE_r(
2729 Perl_re_indentf( aTHX_
2730 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2732 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2733 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2734 (int)trie->minlen, (int)trie->maxlen )
2738 We now know what we are dealing with in terms of unique chars and
2739 string sizes so we can calculate how much memory a naive
2740 representation using a flat table will take. If it's over a reasonable
2741 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2742 conservative but potentially much slower representation using an array
2745 At the end we convert both representations into the same compressed
2746 form that will be used in regexec.c for matching with. The latter
2747 is a form that cannot be used to construct with but has memory
2748 properties similar to the list form and access properties similar
2749 to the table form making it both suitable for fast searches and
2750 small enough that its feasable to store for the duration of a program.
2752 See the comment in the code where the compressed table is produced
2753 inplace from the flat tabe representation for an explanation of how
2754 the compression works.
2759 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2762 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2763 > SvIV(re_trie_maxbuff) )
2766 Second Pass -- Array Of Lists Representation
2768 Each state will be represented by a list of charid:state records
2769 (reg_trie_trans_le) the first such element holds the CUR and LEN
2770 points of the allocated array. (See defines above).
2772 We build the initial structure using the lists, and then convert
2773 it into the compressed table form which allows faster lookups
2774 (but cant be modified once converted).
2777 STRLEN transcount = 1;
2779 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2782 trie->states = (reg_trie_state *)
2783 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2784 sizeof(reg_trie_state) );
2788 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2790 regnode *noper = NEXTOPER( cur );
2791 U32 state = 1; /* required init */
2792 U16 charid = 0; /* sanity init */
2793 U32 wordlen = 0; /* required init */
2795 if (OP(noper) == NOTHING) {
2796 regnode *noper_next= regnext(noper);
2797 if (noper_next < tail)
2801 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2802 const U8 *uc= (U8*)STRING(noper);
2803 const U8 *e= uc + STR_LEN(noper);
2805 for ( ; uc < e ; uc += len ) {
2810 charid = trie->charmap[ uvc ];
2812 SV** const svpp = hv_fetch( widecharmap,
2819 charid=(U16)SvIV( *svpp );
2822 /* charid is now 0 if we dont know the char read, or
2823 * nonzero if we do */
2830 if ( !trie->states[ state ].trans.list ) {
2831 TRIE_LIST_NEW( state );
2834 check <= TRIE_LIST_USED( state );
2837 if ( TRIE_LIST_ITEM( state, check ).forid
2840 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2845 newstate = next_alloc++;
2846 prev_states[newstate] = state;
2847 TRIE_LIST_PUSH( state, charid, newstate );
2852 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
2856 TRIE_HANDLE_WORD(state);
2858 } /* end second pass */
2860 /* next alloc is the NEXT state to be allocated */
2861 trie->statecount = next_alloc;
2862 trie->states = (reg_trie_state *)
2863 PerlMemShared_realloc( trie->states,
2865 * sizeof(reg_trie_state) );
2867 /* and now dump it out before we compress it */
2868 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2869 revcharmap, next_alloc,
2873 trie->trans = (reg_trie_trans *)
2874 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2881 for( state=1 ; state < next_alloc ; state ++ ) {
2885 DEBUG_TRIE_COMPILE_MORE_r(
2886 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
2890 if (trie->states[state].trans.list) {
2891 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2895 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2896 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2897 if ( forid < minid ) {
2899 } else if ( forid > maxid ) {
2903 if ( transcount < tp + maxid - minid + 1) {
2905 trie->trans = (reg_trie_trans *)
2906 PerlMemShared_realloc( trie->trans,
2908 * sizeof(reg_trie_trans) );
2909 Zero( trie->trans + (transcount / 2),
2913 base = trie->uniquecharcount + tp - minid;
2914 if ( maxid == minid ) {
2916 for ( ; zp < tp ; zp++ ) {
2917 if ( ! trie->trans[ zp ].next ) {
2918 base = trie->uniquecharcount + zp - minid;
2919 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2921 trie->trans[ zp ].check = state;
2927 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2929 trie->trans[ tp ].check = state;
2934 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2935 const U32 tid = base
2936 - trie->uniquecharcount
2937 + TRIE_LIST_ITEM( state, idx ).forid;
2938 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2940 trie->trans[ tid ].check = state;
2942 tp += ( maxid - minid + 1 );
2944 Safefree(trie->states[ state ].trans.list);
2947 DEBUG_TRIE_COMPILE_MORE_r(
2948 Perl_re_printf( aTHX_ " base: %d\n",base);
2951 trie->states[ state ].trans.base=base;
2953 trie->lasttrans = tp + 1;
2957 Second Pass -- Flat Table Representation.
2959 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2960 each. We know that we will need Charcount+1 trans at most to store
2961 the data (one row per char at worst case) So we preallocate both
2962 structures assuming worst case.
2964 We then construct the trie using only the .next slots of the entry
2967 We use the .check field of the first entry of the node temporarily
2968 to make compression both faster and easier by keeping track of how
2969 many non zero fields are in the node.
2971 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2974 There are two terms at use here: state as a TRIE_NODEIDX() which is
2975 a number representing the first entry of the node, and state as a
2976 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2977 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2978 if there are 2 entrys per node. eg:
2986 The table is internally in the right hand, idx form. However as we
2987 also have to deal with the states array which is indexed by nodenum
2988 we have to use TRIE_NODENUM() to convert.
2991 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
2994 trie->trans = (reg_trie_trans *)
2995 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2996 * trie->uniquecharcount + 1,
2997 sizeof(reg_trie_trans) );
2998 trie->states = (reg_trie_state *)
2999 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3000 sizeof(reg_trie_state) );
3001 next_alloc = trie->uniquecharcount + 1;
3004 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3006 regnode *noper = NEXTOPER( cur );
3008 U32 state = 1; /* required init */
3010 U16 charid = 0; /* sanity init */
3011 U32 accept_state = 0; /* sanity init */
3013 U32 wordlen = 0; /* required init */
3015 if (OP(noper) == NOTHING) {
3016 regnode *noper_next= regnext(noper);
3017 if (noper_next < tail)
3021 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3022 const U8 *uc= (U8*)STRING(noper);
3023 const U8 *e= uc + STR_LEN(noper);
3025 for ( ; uc < e ; uc += len ) {
3030 charid = trie->charmap[ uvc ];
3032 SV* const * const svpp = hv_fetch( widecharmap,
3036 charid = svpp ? (U16)SvIV(*svpp) : 0;
3040 if ( !trie->trans[ state + charid ].next ) {
3041 trie->trans[ state + charid ].next = next_alloc;
3042 trie->trans[ state ].check++;
3043 prev_states[TRIE_NODENUM(next_alloc)]
3044 = TRIE_NODENUM(state);
3045 next_alloc += trie->uniquecharcount;
3047 state = trie->trans[ state + charid ].next;
3049 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3051 /* charid is now 0 if we dont know the char read, or
3052 * nonzero if we do */
3055 accept_state = TRIE_NODENUM( state );
3056 TRIE_HANDLE_WORD(accept_state);
3058 } /* end second pass */
3060 /* and now dump it out before we compress it */
3061 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3063 next_alloc, depth+1));
3067 * Inplace compress the table.*
3069 For sparse data sets the table constructed by the trie algorithm will
3070 be mostly 0/FAIL transitions or to put it another way mostly empty.
3071 (Note that leaf nodes will not contain any transitions.)
3073 This algorithm compresses the tables by eliminating most such
3074 transitions, at the cost of a modest bit of extra work during lookup:
3076 - Each states[] entry contains a .base field which indicates the
3077 index in the state[] array wheres its transition data is stored.
3079 - If .base is 0 there are no valid transitions from that node.
3081 - If .base is nonzero then charid is added to it to find an entry in
3084 -If trans[states[state].base+charid].check!=state then the
3085 transition is taken to be a 0/Fail transition. Thus if there are fail
3086 transitions at the front of the node then the .base offset will point
3087 somewhere inside the previous nodes data (or maybe even into a node
3088 even earlier), but the .check field determines if the transition is
3092 The following process inplace converts the table to the compressed
3093 table: We first do not compress the root node 1,and mark all its
3094 .check pointers as 1 and set its .base pointer as 1 as well. This
3095 allows us to do a DFA construction from the compressed table later,
3096 and ensures that any .base pointers we calculate later are greater
3099 - We set 'pos' to indicate the first entry of the second node.
3101 - We then iterate over the columns of the node, finding the first and
3102 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3103 and set the .check pointers accordingly, and advance pos
3104 appropriately and repreat for the next node. Note that when we copy
3105 the next pointers we have to convert them from the original
3106 NODEIDX form to NODENUM form as the former is not valid post
3109 - If a node has no transitions used we mark its base as 0 and do not
3110 advance the pos pointer.
3112 - If a node only has one transition we use a second pointer into the
3113 structure to fill in allocated fail transitions from other states.
3114 This pointer is independent of the main pointer and scans forward
3115 looking for null transitions that are allocated to a state. When it
3116 finds one it writes the single transition into the "hole". If the
3117 pointer doesnt find one the single transition is appended as normal.
3119 - Once compressed we can Renew/realloc the structures to release the
3122 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3123 specifically Fig 3.47 and the associated pseudocode.
3127 const U32 laststate = TRIE_NODENUM( next_alloc );
3130 trie->statecount = laststate;
3132 for ( state = 1 ; state < laststate ; state++ ) {
3134 const U32 stateidx = TRIE_NODEIDX( state );
3135 const U32 o_used = trie->trans[ stateidx ].check;
3136 U32 used = trie->trans[ stateidx ].check;
3137 trie->trans[ stateidx ].check = 0;
3140 used && charid < trie->uniquecharcount;
3143 if ( flag || trie->trans[ stateidx + charid ].next ) {
3144 if ( trie->trans[ stateidx + charid ].next ) {
3146 for ( ; zp < pos ; zp++ ) {
3147 if ( ! trie->trans[ zp ].next ) {
3151 trie->states[ state ].trans.base
3153 + trie->uniquecharcount
3155 trie->trans[ zp ].next
3156 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3158 trie->trans[ zp ].check = state;
3159 if ( ++zp > pos ) pos = zp;
3166 trie->states[ state ].trans.base
3167 = pos + trie->uniquecharcount - charid ;
3169 trie->trans[ pos ].next
3170 = SAFE_TRIE_NODENUM(
3171 trie->trans[ stateidx + charid ].next );
3172 trie->trans[ pos ].check = state;
3177 trie->lasttrans = pos + 1;
3178 trie->states = (reg_trie_state *)
3179 PerlMemShared_realloc( trie->states, laststate
3180 * sizeof(reg_trie_state) );
3181 DEBUG_TRIE_COMPILE_MORE_r(
3182 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3184 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3188 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3191 } /* end table compress */
3193 DEBUG_TRIE_COMPILE_MORE_r(
3194 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3196 (UV)trie->statecount,
3197 (UV)trie->lasttrans)
3199 /* resize the trans array to remove unused space */
3200 trie->trans = (reg_trie_trans *)
3201 PerlMemShared_realloc( trie->trans, trie->lasttrans
3202 * sizeof(reg_trie_trans) );
3204 { /* Modify the program and insert the new TRIE node */
3205 U8 nodetype =(U8)(flags & 0xFF);
3209 regnode *optimize = NULL;
3210 #ifdef RE_TRACK_PATTERN_OFFSETS
3213 U32 mjd_nodelen = 0;
3214 #endif /* RE_TRACK_PATTERN_OFFSETS */
3215 #endif /* DEBUGGING */
3217 This means we convert either the first branch or the first Exact,
3218 depending on whether the thing following (in 'last') is a branch
3219 or not and whther first is the startbranch (ie is it a sub part of
3220 the alternation or is it the whole thing.)
3221 Assuming its a sub part we convert the EXACT otherwise we convert
3222 the whole branch sequence, including the first.
3224 /* Find the node we are going to overwrite */
3225 if ( first != startbranch || OP( last ) == BRANCH ) {
3226 /* branch sub-chain */
3227 NEXT_OFF( first ) = (U16)(last - first);
3228 #ifdef RE_TRACK_PATTERN_OFFSETS
3230 mjd_offset= Node_Offset((convert));
3231 mjd_nodelen= Node_Length((convert));
3234 /* whole branch chain */
3236 #ifdef RE_TRACK_PATTERN_OFFSETS
3239 const regnode *nop = NEXTOPER( convert );
3240 mjd_offset= Node_Offset((nop));
3241 mjd_nodelen= Node_Length((nop));
3245 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3247 (UV)mjd_offset, (UV)mjd_nodelen)
3250 /* But first we check to see if there is a common prefix we can
3251 split out as an EXACT and put in front of the TRIE node. */
3252 trie->startstate= 1;
3253 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3254 /* we want to find the first state that has more than
3255 * one transition, if that state is not the first state
3256 * then we have a common prefix which we can remove.
3259 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3261 I32 first_ofs = -1; /* keeps track of the ofs of the first
3262 transition, -1 means none */
3264 const U32 base = trie->states[ state ].trans.base;
3266 /* does this state terminate an alternation? */
3267 if ( trie->states[state].wordnum )
3270 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3271 if ( ( base + ofs >= trie->uniquecharcount ) &&
3272 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3273 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3275 if ( ++count > 1 ) {
3276 /* we have more than one transition */
3279 /* if this is the first state there is no common prefix
3280 * to extract, so we can exit */
3281 if ( state == 1 ) break;
3282 tmp = av_fetch( revcharmap, ofs, 0);
3283 ch = (U8*)SvPV_nolen_const( *tmp );
3285 /* if we are on count 2 then we need to initialize the
3286 * bitmap, and store the previous char if there was one
3289 /* clear the bitmap */
3290 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3292 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3295 if (first_ofs >= 0) {
3296 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3297 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3299 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3301 Perl_re_printf( aTHX_ "%s", (char*)ch)
3305 /* store the current firstchar in the bitmap */
3306 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3307 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3313 /* This state has only one transition, its transition is part
3314 * of a common prefix - we need to concatenate the char it
3315 * represents to what we have so far. */
3316 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3318 char *ch = SvPV( *tmp, len );
3320 SV *sv=sv_newmortal();
3321 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3323 (UV)state, (UV)first_ofs,
3324 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3325 PL_colors[0], PL_colors[1],
3326 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3327 PERL_PV_ESCAPE_FIRSTCHAR
3332 OP( convert ) = nodetype;
3333 str=STRING(convert);
3336 STR_LEN(convert) += len;
3342 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3347 trie->prefixlen = (state-1);
3349 regnode *n = convert+NODE_SZ_STR(convert);
3350 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3351 trie->startstate = state;
3352 trie->minlen -= (state - 1);
3353 trie->maxlen -= (state - 1);
3355 /* At least the UNICOS C compiler choked on this
3356 * being argument to DEBUG_r(), so let's just have
3359 #ifdef PERL_EXT_RE_BUILD
3365 regnode *fix = convert;
3366 U32 word = trie->wordcount;
3368 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3369 while( ++fix < n ) {
3370 Set_Node_Offset_Length(fix, 0, 0);
3373 SV ** const tmp = av_fetch( trie_words, word, 0 );
3375 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3376 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3378 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3386 NEXT_OFF(convert) = (U16)(tail - convert);
3387 DEBUG_r(optimize= n);
3393 if ( trie->maxlen ) {
3394 NEXT_OFF( convert ) = (U16)(tail - convert);
3395 ARG_SET( convert, data_slot );
3396 /* Store the offset to the first unabsorbed branch in
3397 jump[0], which is otherwise unused by the jump logic.
3398 We use this when dumping a trie and during optimisation. */
3400 trie->jump[0] = (U16)(nextbranch - convert);
3402 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3403 * and there is a bitmap
3404 * and the first "jump target" node we found leaves enough room
3405 * then convert the TRIE node into a TRIEC node, with the bitmap
3406 * embedded inline in the opcode - this is hypothetically faster.
3408 if ( !trie->states[trie->startstate].wordnum
3410 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3412 OP( convert ) = TRIEC;
3413 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3414 PerlMemShared_free(trie->bitmap);
3417 OP( convert ) = TRIE;
3419 /* store the type in the flags */
3420 convert->flags = nodetype;
3424 + regarglen[ OP( convert ) ];
3426 /* XXX We really should free up the resource in trie now,
3427 as we won't use them - (which resources?) dmq */
3429 /* needed for dumping*/
3430 DEBUG_r(if (optimize) {
3431 regnode *opt = convert;
3433 while ( ++opt < optimize) {
3434 Set_Node_Offset_Length(opt,0,0);
3437 Try to clean up some of the debris left after the
3440 while( optimize < jumper ) {
3441 mjd_nodelen += Node_Length((optimize));
3442 OP( optimize ) = OPTIMIZED;
3443 Set_Node_Offset_Length(optimize,0,0);
3446 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3448 } /* end node insert */
3450 /* Finish populating the prev field of the wordinfo array. Walk back
3451 * from each accept state until we find another accept state, and if
3452 * so, point the first word's .prev field at the second word. If the
3453 * second already has a .prev field set, stop now. This will be the
3454 * case either if we've already processed that word's accept state,
3455 * or that state had multiple words, and the overspill words were
3456 * already linked up earlier.
3463 for (word=1; word <= trie->wordcount; word++) {
3465 if (trie->wordinfo[word].prev)
3467 state = trie->wordinfo[word].accept;
3469 state = prev_states[state];
3472 prev = trie->states[state].wordnum;
3476 trie->wordinfo[word].prev = prev;
3478 Safefree(prev_states);
3482 /* and now dump out the compressed format */
3483 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3485 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3487 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3488 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3490 SvREFCNT_dec_NN(revcharmap);
3494 : trie->startstate>1
3500 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3502 /* The Trie is constructed and compressed now so we can build a fail array if
3505 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3507 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3511 We find the fail state for each state in the trie, this state is the longest
3512 proper suffix of the current state's 'word' that is also a proper prefix of
3513 another word in our trie. State 1 represents the word '' and is thus the
3514 default fail state. This allows the DFA not to have to restart after its
3515 tried and failed a word at a given point, it simply continues as though it
3516 had been matching the other word in the first place.
3518 'abcdgu'=~/abcdefg|cdgu/
3519 When we get to 'd' we are still matching the first word, we would encounter
3520 'g' which would fail, which would bring us to the state representing 'd' in
3521 the second word where we would try 'g' and succeed, proceeding to match
3524 /* add a fail transition */
3525 const U32 trie_offset = ARG(source);
3526 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3528 const U32 ucharcount = trie->uniquecharcount;
3529 const U32 numstates = trie->statecount;
3530 const U32 ubound = trie->lasttrans + ucharcount;
3534 U32 base = trie->states[ 1 ].trans.base;
3537 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3539 GET_RE_DEBUG_FLAGS_DECL;
3541 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3542 PERL_UNUSED_CONTEXT;
3544 PERL_UNUSED_ARG(depth);
3547 if ( OP(source) == TRIE ) {
3548 struct regnode_1 *op = (struct regnode_1 *)
3549 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3550 StructCopy(source,op,struct regnode_1);
3551 stclass = (regnode *)op;
3553 struct regnode_charclass *op = (struct regnode_charclass *)
3554 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3555 StructCopy(source,op,struct regnode_charclass);
3556 stclass = (regnode *)op;
3558 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3560 ARG_SET( stclass, data_slot );
3561 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3562 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3563 aho->trie=trie_offset;
3564 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3565 Copy( trie->states, aho->states, numstates, reg_trie_state );
3566 Newxz( q, numstates, U32);
3567 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3570 /* initialize fail[0..1] to be 1 so that we always have
3571 a valid final fail state */
3572 fail[ 0 ] = fail[ 1 ] = 1;
3574 for ( charid = 0; charid < ucharcount ; charid++ ) {
3575 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3577 q[ q_write ] = newstate;
3578 /* set to point at the root */
3579 fail[ q[ q_write++ ] ]=1;
3582 while ( q_read < q_write) {
3583 const U32 cur = q[ q_read++ % numstates ];
3584 base = trie->states[ cur ].trans.base;
3586 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3587 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3589 U32 fail_state = cur;
3592 fail_state = fail[ fail_state ];
3593 fail_base = aho->states[ fail_state ].trans.base;
3594 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3596 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3597 fail[ ch_state ] = fail_state;
3598 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3600 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3602 q[ q_write++ % numstates] = ch_state;
3606 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3607 when we fail in state 1, this allows us to use the
3608 charclass scan to find a valid start char. This is based on the principle
3609 that theres a good chance the string being searched contains lots of stuff
3610 that cant be a start char.
3612 fail[ 0 ] = fail[ 1 ] = 0;
3613 DEBUG_TRIE_COMPILE_r({
3614 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3615 depth, (UV)numstates
3617 for( q_read=1; q_read<numstates; q_read++ ) {
3618 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3620 Perl_re_printf( aTHX_ "\n");
3623 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3628 #define DEBUG_PEEP(str,scan,depth) \
3629 DEBUG_OPTIMISE_r({if (scan){ \
3630 regnode *Next = regnext(scan); \
3631 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);\
3632 Perl_re_indentf( aTHX_ "" str ">%3d: %s (%d)", \
3633 depth, REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3634 Next ? (REG_NODE_NUM(Next)) : 0 );\
3635 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3636 Perl_re_printf( aTHX_ "\n"); \
3639 /* The below joins as many adjacent EXACTish nodes as possible into a single
3640 * one. The regop may be changed if the node(s) contain certain sequences that
3641 * require special handling. The joining is only done if:
3642 * 1) there is room in the current conglomerated node to entirely contain the
3644 * 2) they are the exact same node type
3646 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3647 * these get optimized out
3649 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3650 * as possible, even if that means splitting an existing node so that its first
3651 * part is moved to the preceeding node. This would maximise the efficiency of
3652 * memEQ during matching. Elsewhere in this file, khw proposes splitting
3653 * EXACTFish nodes into portions that don't change under folding vs those that
3654 * do. Those portions that don't change may be the only things in the pattern that
3655 * could be used to find fixed and floating strings.
3657 * If a node is to match under /i (folded), the number of characters it matches
3658 * can be different than its character length if it contains a multi-character
3659 * fold. *min_subtract is set to the total delta number of characters of the
3662 * And *unfolded_multi_char is set to indicate whether or not the node contains
3663 * an unfolded multi-char fold. This happens when whether the fold is valid or
3664 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3665 * SMALL LETTER SHARP S, as only if the target string being matched against
3666 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3667 * folding rules depend on the locale in force at runtime. (Multi-char folds
3668 * whose components are all above the Latin1 range are not run-time locale
3669 * dependent, and have already been folded by the time this function is
3672 * This is as good a place as any to discuss the design of handling these
3673 * multi-character fold sequences. It's been wrong in Perl for a very long
3674 * time. There are three code points in Unicode whose multi-character folds
3675 * were long ago discovered to mess things up. The previous designs for
3676 * dealing with these involved assigning a special node for them. This
3677 * approach doesn't always work, as evidenced by this example:
3678 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3679 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3680 * would match just the \xDF, it won't be able to handle the case where a
3681 * successful match would have to cross the node's boundary. The new approach
3682 * that hopefully generally solves the problem generates an EXACTFU_SS node
3683 * that is "sss" in this case.
3685 * It turns out that there are problems with all multi-character folds, and not
3686 * just these three. Now the code is general, for all such cases. The
3687 * approach taken is:
3688 * 1) This routine examines each EXACTFish node that could contain multi-
3689 * character folded sequences. Since a single character can fold into
3690 * such a sequence, the minimum match length for this node is less than
3691 * the number of characters in the node. This routine returns in
3692 * *min_subtract how many characters to subtract from the the actual
3693 * length of the string to get a real minimum match length; it is 0 if
3694 * there are no multi-char foldeds. This delta is used by the caller to
3695 * adjust the min length of the match, and the delta between min and max,
3696 * so that the optimizer doesn't reject these possibilities based on size
3698 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3699 * is used for an EXACTFU node that contains at least one "ss" sequence in
3700 * it. For non-UTF-8 patterns and strings, this is the only case where
3701 * there is a possible fold length change. That means that a regular
3702 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3703 * with length changes, and so can be processed faster. regexec.c takes
3704 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3705 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3706 * known until runtime). This saves effort in regex matching. However,
3707 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3708 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3709 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3710 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3711 * possibilities for the non-UTF8 patterns are quite simple, except for
3712 * the sharp s. All the ones that don't involve a UTF-8 target string are
3713 * members of a fold-pair, and arrays are set up for all of them so that
3714 * the other member of the pair can be found quickly. Code elsewhere in
3715 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3716 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3717 * described in the next item.
3718 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3719 * validity of the fold won't be known until runtime, and so must remain
3720 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3721 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3722 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3723 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3724 * The reason this is a problem is that the optimizer part of regexec.c
3725 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3726 * that a character in the pattern corresponds to at most a single
3727 * character in the target string. (And I do mean character, and not byte
3728 * here, unlike other parts of the documentation that have never been
3729 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3730 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3731 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3732 * nodes, violate the assumption, and they are the only instances where it
3733 * is violated. I'm reluctant to try to change the assumption, as the
3734 * code involved is impenetrable to me (khw), so instead the code here
3735 * punts. This routine examines EXACTFL nodes, and (when the pattern
3736 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3737 * boolean indicating whether or not the node contains such a fold. When
3738 * it is true, the caller sets a flag that later causes the optimizer in
3739 * this file to not set values for the floating and fixed string lengths,
3740 * and thus avoids the optimizer code in regexec.c that makes the invalid
3741 * assumption. Thus, there is no optimization based on string lengths for
3742 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3743 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3744 * assumption is wrong only in these cases is that all other non-UTF-8
3745 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3746 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3747 * EXACTF nodes because we don't know at compile time if it actually
3748 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3749 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3750 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3751 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3752 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3753 * string would require the pattern to be forced into UTF-8, the overhead
3754 * of which we want to avoid. Similarly the unfolded multi-char folds in
3755 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3758 * Similarly, the code that generates tries doesn't currently handle
3759 * not-already-folded multi-char folds, and it looks like a pain to change
3760 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3761 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3762 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3763 * using /iaa matching will be doing so almost entirely with ASCII
3764 * strings, so this should rarely be encountered in practice */
3766 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3767 if (PL_regkind[OP(scan)] == EXACT) \
3768 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3771 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3772 UV *min_subtract, bool *unfolded_multi_char,
3773 U32 flags,regnode *val, U32 depth)
3775 /* Merge several consecutive EXACTish nodes into one. */
3776 regnode *n = regnext(scan);
3778 regnode *next = scan + NODE_SZ_STR(scan);
3782 regnode *stop = scan;
3783 GET_RE_DEBUG_FLAGS_DECL;
3785 PERL_UNUSED_ARG(depth);
3788 PERL_ARGS_ASSERT_JOIN_EXACT;
3789 #ifndef EXPERIMENTAL_INPLACESCAN
3790 PERL_UNUSED_ARG(flags);
3791 PERL_UNUSED_ARG(val);
3793 DEBUG_PEEP("join",scan,depth);
3795 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3796 * EXACT ones that are mergeable to the current one. */
3798 && (PL_regkind[OP(n)] == NOTHING
3799 || (stringok && OP(n) == OP(scan)))
3801 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3804 if (OP(n) == TAIL || n > next)
3806 if (PL_regkind[OP(n)] == NOTHING) {
3807 DEBUG_PEEP("skip:",n,depth);
3808 NEXT_OFF(scan) += NEXT_OFF(n);
3809 next = n + NODE_STEP_REGNODE;
3816 else if (stringok) {
3817 const unsigned int oldl = STR_LEN(scan);
3818 regnode * const nnext = regnext(n);
3820 /* XXX I (khw) kind of doubt that this works on platforms (should
3821 * Perl ever run on one) where U8_MAX is above 255 because of lots
3822 * of other assumptions */
3823 /* Don't join if the sum can't fit into a single node */
3824 if (oldl + STR_LEN(n) > U8_MAX)
3827 DEBUG_PEEP("merg",n,depth);
3830 NEXT_OFF(scan) += NEXT_OFF(n);
3831 STR_LEN(scan) += STR_LEN(n);
3832 next = n + NODE_SZ_STR(n);
3833 /* Now we can overwrite *n : */
3834 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3842 #ifdef EXPERIMENTAL_INPLACESCAN
3843 if (flags && !NEXT_OFF(n)) {
3844 DEBUG_PEEP("atch", val, depth);
3845 if (reg_off_by_arg[OP(n)]) {
3846 ARG_SET(n, val - n);
3849 NEXT_OFF(n) = val - n;
3857 *unfolded_multi_char = FALSE;
3859 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3860 * can now analyze for sequences of problematic code points. (Prior to
3861 * this final joining, sequences could have been split over boundaries, and
3862 * hence missed). The sequences only happen in folding, hence for any
3863 * non-EXACT EXACTish node */
3864 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3865 U8* s0 = (U8*) STRING(scan);
3867 U8* s_end = s0 + STR_LEN(scan);
3869 int total_count_delta = 0; /* Total delta number of characters that
3870 multi-char folds expand to */
3872 /* One pass is made over the node's string looking for all the
3873 * possibilities. To avoid some tests in the loop, there are two main
3874 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3879 if (OP(scan) == EXACTFL) {
3882 /* An EXACTFL node would already have been changed to another
3883 * node type unless there is at least one character in it that
3884 * is problematic; likely a character whose fold definition
3885 * won't be known until runtime, and so has yet to be folded.
3886 * For all but the UTF-8 locale, folds are 1-1 in length, but
3887 * to handle the UTF-8 case, we need to create a temporary
3888 * folded copy using UTF-8 locale rules in order to analyze it.
3889 * This is because our macros that look to see if a sequence is
3890 * a multi-char fold assume everything is folded (otherwise the
3891 * tests in those macros would be too complicated and slow).
3892 * Note that here, the non-problematic folds will have already
3893 * been done, so we can just copy such characters. We actually
3894 * don't completely fold the EXACTFL string. We skip the
3895 * unfolded multi-char folds, as that would just create work
3896 * below to figure out the size they already are */
3898 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3901 STRLEN s_len = UTF8SKIP(s);
3902 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3903 Copy(s, d, s_len, U8);
3906 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3907 *unfolded_multi_char = TRUE;
3908 Copy(s, d, s_len, U8);
3911 else if (isASCII(*s)) {
3912 *(d++) = toFOLD(*s);
3916 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
3922 /* Point the remainder of the routine to look at our temporary
3926 } /* End of creating folded copy of EXACTFL string */
3928 /* Examine the string for a multi-character fold sequence. UTF-8
3929 * patterns have all characters pre-folded by the time this code is
3931 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3932 length sequence we are looking for is 2 */
3934 int count = 0; /* How many characters in a multi-char fold */
3935 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3936 if (! len) { /* Not a multi-char fold: get next char */
3941 /* Nodes with 'ss' require special handling, except for
3942 * EXACTFA-ish for which there is no multi-char fold to this */
3943 if (len == 2 && *s == 's' && *(s+1) == 's'
3944 && OP(scan) != EXACTFA
3945 && OP(scan) != EXACTFA_NO_TRIE)
3948 if (OP(scan) != EXACTFL) {
3949 OP(scan) = EXACTFU_SS;
3953 else { /* Here is a generic multi-char fold. */
3954 U8* multi_end = s + len;
3956 /* Count how many characters are in it. In the case of
3957 * /aa, no folds which contain ASCII code points are
3958 * allowed, so check for those, and skip if found. */
3959 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3960 count = utf8_length(s, multi_end);
3964 while (s < multi_end) {
3967 goto next_iteration;
3977 /* The delta is how long the sequence is minus 1 (1 is how long
3978 * the character that folds to the sequence is) */
3979 total_count_delta += count - 1;
3983 /* We created a temporary folded copy of the string in EXACTFL
3984 * nodes. Therefore we need to be sure it doesn't go below zero,
3985 * as the real string could be shorter */
3986 if (OP(scan) == EXACTFL) {
3987 int total_chars = utf8_length((U8*) STRING(scan),
3988 (U8*) STRING(scan) + STR_LEN(scan));
3989 if (total_count_delta > total_chars) {
3990 total_count_delta = total_chars;
3994 *min_subtract += total_count_delta;
3997 else if (OP(scan) == EXACTFA) {
3999 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
4000 * fold to the ASCII range (and there are no existing ones in the
4001 * upper latin1 range). But, as outlined in the comments preceding
4002 * this function, we need to flag any occurrences of the sharp s.
4003 * This character forbids trie formation (because of added
4005 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4006 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4007 || UNICODE_DOT_DOT_VERSION > 0)
4009 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4010 OP(scan) = EXACTFA_NO_TRIE;
4011 *unfolded_multi_char = TRUE;
4019 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
4020 * folds that are all Latin1. As explained in the comments
4021 * preceding this function, we look also for the sharp s in EXACTF
4022 * and EXACTFL nodes; it can be in the final position. Otherwise
4023 * we can stop looking 1 byte earlier because have to find at least
4024 * two characters for a multi-fold */
4025 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4030 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4031 if (! len) { /* Not a multi-char fold. */
4032 if (*s == LATIN_SMALL_LETTER_SHARP_S
4033 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4035 *unfolded_multi_char = TRUE;
4042 && isALPHA_FOLD_EQ(*s, 's')
4043 && isALPHA_FOLD_EQ(*(s+1), 's'))
4046 /* EXACTF nodes need to know that the minimum length
4047 * changed so that a sharp s in the string can match this
4048 * ss in the pattern, but they remain EXACTF nodes, as they
4049 * won't match this unless the target string is is UTF-8,
4050 * which we don't know until runtime. EXACTFL nodes can't
4051 * transform into EXACTFU nodes */
4052 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4053 OP(scan) = EXACTFU_SS;
4057 *min_subtract += len - 1;
4065 /* Allow dumping but overwriting the collection of skipped
4066 * ops and/or strings with fake optimized ops */
4067 n = scan + NODE_SZ_STR(scan);
4075 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
4079 /* REx optimizer. Converts nodes into quicker variants "in place".
4080 Finds fixed substrings. */
4082 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4083 to the position after last scanned or to NULL. */
4085 #define INIT_AND_WITHP \
4086 assert(!and_withp); \
4087 Newx(and_withp,1, regnode_ssc); \
4088 SAVEFREEPV(and_withp)
4092 S_unwind_scan_frames(pTHX_ const void *p)
4094 scan_frame *f= (scan_frame *)p;
4096 scan_frame *n= f->next_frame;
4104 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4105 SSize_t *minlenp, SSize_t *deltap,
4110 regnode_ssc *and_withp,
4111 U32 flags, U32 depth)
4112 /* scanp: Start here (read-write). */
4113 /* deltap: Write maxlen-minlen here. */
4114 /* last: Stop before this one. */
4115 /* data: string data about the pattern */
4116 /* stopparen: treat close N as END */
4117 /* recursed: which subroutines have we recursed into */
4118 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4120 /* There must be at least this number of characters to match */
4123 regnode *scan = *scanp, *next;
4125 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4126 int is_inf_internal = 0; /* The studied chunk is infinite */
4127 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4128 scan_data_t data_fake;
4129 SV *re_trie_maxbuff = NULL;
4130 regnode *first_non_open = scan;
4131 SSize_t stopmin = SSize_t_MAX;
4132 scan_frame *frame = NULL;
4133 GET_RE_DEBUG_FLAGS_DECL;
4135 PERL_ARGS_ASSERT_STUDY_CHUNK;
4136 RExC_study_started= 1;
4140 while (first_non_open && OP(first_non_open) == OPEN)
4141 first_non_open=regnext(first_non_open);
4147 RExC_study_chunk_recursed_count++;
4149 DEBUG_OPTIMISE_MORE_r(
4151 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4152 depth, (long)stopparen,
4153 (unsigned long)RExC_study_chunk_recursed_count,
4154 (unsigned long)depth, (unsigned long)recursed_depth,
4157 if (recursed_depth) {
4160 for ( j = 0 ; j < recursed_depth ; j++ ) {
4161 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
4163 PAREN_TEST(RExC_study_chunk_recursed +
4164 ( j * RExC_study_chunk_recursed_bytes), i )
4167 !PAREN_TEST(RExC_study_chunk_recursed +
4168 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
4171 Perl_re_printf( aTHX_ " %d",(int)i);
4175 if ( j + 1 < recursed_depth ) {
4176 Perl_re_printf( aTHX_ ",");
4180 Perl_re_printf( aTHX_ "\n");
4183 while ( scan && OP(scan) != END && scan < last ){
4184 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4185 node length to get a real minimum (because
4186 the folded version may be shorter) */
4187 bool unfolded_multi_char = FALSE;
4188 /* Peephole optimizer: */
4189 DEBUG_STUDYDATA("Peep:", data, depth);
4190 DEBUG_PEEP("Peep", scan, depth);
4193 /* The reason we do this here is that we need to deal with things like
4194 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4195 * parsing code, as each (?:..) is handled by a different invocation of
4198 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
4200 /* Follow the next-chain of the current node and optimize
4201 away all the NOTHINGs from it. */
4202 if (OP(scan) != CURLYX) {
4203 const int max = (reg_off_by_arg[OP(scan)]
4205 /* I32 may be smaller than U16 on CRAYs! */
4206 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4207 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
4211 /* Skip NOTHING and LONGJMP. */
4212 while ((n = regnext(n))
4213 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4214 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
4215 && off + noff < max)
4217 if (reg_off_by_arg[OP(scan)])
4220 NEXT_OFF(scan) = off;
4223 /* The principal pseudo-switch. Cannot be a switch, since we
4224 look into several different things. */
4225 if ( OP(scan) == DEFINEP ) {
4227 SSize_t deltanext = 0;
4228 SSize_t fake_last_close = 0;
4229 I32 f = SCF_IN_DEFINE;
4231 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4232 scan = regnext(scan);
4233 assert( OP(scan) == IFTHEN );
4234 DEBUG_PEEP("expect IFTHEN", scan, depth);
4236 data_fake.last_closep= &fake_last_close;
4238 next = regnext(scan);
4239 scan = NEXTOPER(NEXTOPER(scan));
4240 DEBUG_PEEP("scan", scan, depth);
4241 DEBUG_PEEP("next", next, depth);
4243 /* we suppose the run is continuous, last=next...
4244 * NOTE we dont use the return here! */
4245 (void)study_chunk(pRExC_state, &scan, &minlen,
4246 &deltanext, next, &data_fake, stopparen,
4247 recursed_depth, NULL, f, depth+1);
4252 OP(scan) == BRANCH ||
4253 OP(scan) == BRANCHJ ||
4256 next = regnext(scan);
4259 /* The op(next)==code check below is to see if we
4260 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4261 * IFTHEN is special as it might not appear in pairs.
4262 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4263 * we dont handle it cleanly. */
4264 if (OP(next) == code || code == IFTHEN) {
4265 /* NOTE - There is similar code to this block below for
4266 * handling TRIE nodes on a re-study. If you change stuff here
4267 * check there too. */
4268 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
4270 regnode * const startbranch=scan;
4272 if (flags & SCF_DO_SUBSTR) {
4273 /* Cannot merge strings after this. */
4274 scan_commit(pRExC_state, data, minlenp, is_inf);
4277 if (flags & SCF_DO_STCLASS)
4278 ssc_init_zero(pRExC_state, &accum);
4280 while (OP(scan) == code) {
4281 SSize_t deltanext, minnext, fake;
4283 regnode_ssc this_class;
4285 DEBUG_PEEP("Branch", scan, depth);
4288 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4290 data_fake.whilem_c = data->whilem_c;
4291 data_fake.last_closep = data->last_closep;
4294 data_fake.last_closep = &fake;
4296 data_fake.pos_delta = delta;
4297 next = regnext(scan);
4299 scan = NEXTOPER(scan); /* everything */
4300 if (code != BRANCH) /* everything but BRANCH */
4301 scan = NEXTOPER(scan);
4303 if (flags & SCF_DO_STCLASS) {
4304 ssc_init(pRExC_state, &this_class);
4305 data_fake.start_class = &this_class;
4306 f = SCF_DO_STCLASS_AND;
4308 if (flags & SCF_WHILEM_VISITED_POS)
4309 f |= SCF_WHILEM_VISITED_POS;
4311 /* we suppose the run is continuous, last=next...*/
4312 minnext = study_chunk(pRExC_state, &scan, minlenp,
4313 &deltanext, next, &data_fake, stopparen,
4314 recursed_depth, NULL, f,depth+1);
4318 if (deltanext == SSize_t_MAX) {
4319 is_inf = is_inf_internal = 1;
4321 } else if (max1 < minnext + deltanext)
4322 max1 = minnext + deltanext;
4324 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4326 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4327 if ( stopmin > minnext)
4328 stopmin = min + min1;
4329 flags &= ~SCF_DO_SUBSTR;
4331 data->flags |= SCF_SEEN_ACCEPT;
4334 if (data_fake.flags & SF_HAS_EVAL)
4335 data->flags |= SF_HAS_EVAL;
4336 data->whilem_c = data_fake.whilem_c;
4338 if (flags & SCF_DO_STCLASS)
4339 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4341 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4343 if (flags & SCF_DO_SUBSTR) {
4344 data->pos_min += min1;
4345 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
4346 data->pos_delta = SSize_t_MAX;
4348 data->pos_delta += max1 - min1;
4349 if (max1 != min1 || is_inf)
4350 data->longest = &(data->longest_float);
4353 if (delta == SSize_t_MAX
4354 || SSize_t_MAX - delta - (max1 - min1) < 0)
4355 delta = SSize_t_MAX;
4357 delta += max1 - min1;
4358 if (flags & SCF_DO_STCLASS_OR) {
4359 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4361 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4362 flags &= ~SCF_DO_STCLASS;
4365 else if (flags & SCF_DO_STCLASS_AND) {
4367 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4368 flags &= ~SCF_DO_STCLASS;
4371 /* Switch to OR mode: cache the old value of
4372 * data->start_class */
4374 StructCopy(data->start_class, and_withp, regnode_ssc);
4375 flags &= ~SCF_DO_STCLASS_AND;
4376 StructCopy(&accum, data->start_class, regnode_ssc);
4377 flags |= SCF_DO_STCLASS_OR;
4381 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4382 OP( startbranch ) == BRANCH )
4386 Assuming this was/is a branch we are dealing with: 'scan'
4387 now points at the item that follows the branch sequence,
4388 whatever it is. We now start at the beginning of the
4389 sequence and look for subsequences of
4395 which would be constructed from a pattern like
4398 If we can find such a subsequence we need to turn the first
4399 element into a trie and then add the subsequent branch exact
4400 strings to the trie.
4404 1. patterns where the whole set of branches can be
4407 2. patterns where only a subset can be converted.
4409 In case 1 we can replace the whole set with a single regop
4410 for the trie. In case 2 we need to keep the start and end
4413 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4414 becomes BRANCH TRIE; BRANCH X;
4416 There is an additional case, that being where there is a
4417 common prefix, which gets split out into an EXACT like node
4418 preceding the TRIE node.
4420 If x(1..n)==tail then we can do a simple trie, if not we make
4421 a "jump" trie, such that when we match the appropriate word
4422 we "jump" to the appropriate tail node. Essentially we turn
4423 a nested if into a case structure of sorts.
4428 if (!re_trie_maxbuff) {
4429 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4430 if (!SvIOK(re_trie_maxbuff))
4431 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4433 if ( SvIV(re_trie_maxbuff)>=0 ) {
4435 regnode *first = (regnode *)NULL;
4436 regnode *last = (regnode *)NULL;
4437 regnode *tail = scan;
4441 /* var tail is used because there may be a TAIL
4442 regop in the way. Ie, the exacts will point to the
4443 thing following the TAIL, but the last branch will
4444 point at the TAIL. So we advance tail. If we
4445 have nested (?:) we may have to move through several
4449 while ( OP( tail ) == TAIL ) {
4450 /* this is the TAIL generated by (?:) */
4451 tail = regnext( tail );
4455 DEBUG_TRIE_COMPILE_r({
4456 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4457 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4459 "Looking for TRIE'able sequences. Tail node is ",
4460 (UV)(tail - RExC_emit_start),
4461 SvPV_nolen_const( RExC_mysv )
4467 Step through the branches
4468 cur represents each branch,
4469 noper is the first thing to be matched as part
4471 noper_next is the regnext() of that node.
4473 We normally handle a case like this
4474 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4475 support building with NOJUMPTRIE, which restricts
4476 the trie logic to structures like /FOO|BAR/.
4478 If noper is a trieable nodetype then the branch is
4479 a possible optimization target. If we are building
4480 under NOJUMPTRIE then we require that noper_next is
4481 the same as scan (our current position in the regex
4484 Once we have two or more consecutive such branches
4485 we can create a trie of the EXACT's contents and
4486 stitch it in place into the program.
4488 If the sequence represents all of the branches in
4489 the alternation we replace the entire thing with a
4492 Otherwise when it is a subsequence we need to
4493 stitch it in place and replace only the relevant
4494 branches. This means the first branch has to remain
4495 as it is used by the alternation logic, and its
4496 next pointer, and needs to be repointed at the item
4497 on the branch chain following the last branch we
4498 have optimized away.
4500 This could be either a BRANCH, in which case the
4501 subsequence is internal, or it could be the item
4502 following the branch sequence in which case the
4503 subsequence is at the end (which does not
4504 necessarily mean the first node is the start of the
4507 TRIE_TYPE(X) is a define which maps the optype to a
4511 ----------------+-----------
4515 EXACTFU_SS | EXACTFU
4518 EXACTFLU8 | EXACTFLU8
4522 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4524 : ( EXACT == (X) ) \
4526 : ( EXACTFU == (X) || EXACTFU_SS == (X) ) \
4528 : ( EXACTFA == (X) ) \
4530 : ( EXACTL == (X) ) \
4532 : ( EXACTFLU8 == (X) ) \
4536 /* dont use tail as the end marker for this traverse */
4537 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4538 regnode * const noper = NEXTOPER( cur );
4539 U8 noper_type = OP( noper );
4540 U8 noper_trietype = TRIE_TYPE( noper_type );
4541 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4542 regnode * const noper_next = regnext( noper );
4543 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4544 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4547 DEBUG_TRIE_COMPILE_r({
4548 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4549 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4551 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4553 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4554 Perl_re_printf( aTHX_ " -> %d:%s",
4555 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
4558 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4559 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
4560 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
4562 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
4563 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4564 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4568 /* Is noper a trieable nodetype that can be merged
4569 * with the current trie (if there is one)? */
4573 ( noper_trietype == NOTHING )
4574 || ( trietype == NOTHING )
4575 || ( trietype == noper_trietype )
4578 && noper_next >= tail
4582 /* Handle mergable triable node Either we are
4583 * the first node in a new trieable sequence,
4584 * in which case we do some bookkeeping,
4585 * otherwise we update the end pointer. */
4588 if ( noper_trietype == NOTHING ) {
4589 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4590 regnode * const noper_next = regnext( noper );
4591 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4592 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4595 if ( noper_next_trietype ) {
4596 trietype = noper_next_trietype;
4597 } else if (noper_next_type) {
4598 /* a NOTHING regop is 1 regop wide.
4599 * We need at least two for a trie
4600 * so we can't merge this in */
4604 trietype = noper_trietype;
4607 if ( trietype == NOTHING )
4608 trietype = noper_trietype;
4613 } /* end handle mergable triable node */
4615 /* handle unmergable node -
4616 * noper may either be a triable node which can
4617 * not be tried together with the current trie,
4618 * or a non triable node */
4620 /* If last is set and trietype is not
4621 * NOTHING then we have found at least two
4622 * triable branch sequences in a row of a
4623 * similar trietype so we can turn them
4624 * into a trie. If/when we allow NOTHING to
4625 * start a trie sequence this condition
4626 * will be required, and it isn't expensive
4627 * so we leave it in for now. */
4628 if ( trietype && trietype != NOTHING )
4629 make_trie( pRExC_state,
4630 startbranch, first, cur, tail,
4631 count, trietype, depth+1 );
4632 last = NULL; /* note: we clear/update
4633 first, trietype etc below,
4634 so we dont do it here */
4638 && noper_next >= tail
4641 /* noper is triable, so we can start a new
4645 trietype = noper_trietype;
4647 /* if we already saw a first but the
4648 * current node is not triable then we have
4649 * to reset the first information. */
4654 } /* end handle unmergable node */
4655 } /* loop over branches */
4656 DEBUG_TRIE_COMPILE_r({
4657 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4658 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
4659 depth+1, SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4660 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
4661 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4662 PL_reg_name[trietype]
4666 if ( last && trietype ) {
4667 if ( trietype != NOTHING ) {
4668 /* the last branch of the sequence was part of
4669 * a trie, so we have to construct it here
4670 * outside of the loop */
4671 made= make_trie( pRExC_state, startbranch,
4672 first, scan, tail, count,
4673 trietype, depth+1 );
4674 #ifdef TRIE_STUDY_OPT
4675 if ( ((made == MADE_EXACT_TRIE &&
4676 startbranch == first)
4677 || ( first_non_open == first )) &&
4679 flags |= SCF_TRIE_RESTUDY;
4680 if ( startbranch == first
4683 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4688 /* at this point we know whatever we have is a
4689 * NOTHING sequence/branch AND if 'startbranch'
4690 * is 'first' then we can turn the whole thing
4693 if ( startbranch == first ) {
4695 /* the entire thing is a NOTHING sequence,
4696 * something like this: (?:|) So we can
4697 * turn it into a plain NOTHING op. */
4698 DEBUG_TRIE_COMPILE_r({
4699 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4700 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
4702 SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4705 OP(startbranch)= NOTHING;
4706 NEXT_OFF(startbranch)= tail - startbranch;
4707 for ( opt= startbranch + 1; opt < tail ; opt++ )
4711 } /* end if ( last) */
4712 } /* TRIE_MAXBUF is non zero */
4717 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4718 scan = NEXTOPER(NEXTOPER(scan));
4719 } else /* single branch is optimized. */
4720 scan = NEXTOPER(scan);
4722 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
4724 regnode *start = NULL;
4725 regnode *end = NULL;
4726 U32 my_recursed_depth= recursed_depth;
4728 if (OP(scan) != SUSPEND) { /* GOSUB */
4729 /* Do setup, note this code has side effects beyond
4730 * the rest of this block. Specifically setting
4731 * RExC_recurse[] must happen at least once during
4734 RExC_recurse[ARG2L(scan)] = scan;
4735 start = RExC_open_parens[paren];
4736 end = RExC_close_parens[paren];
4738 /* NOTE we MUST always execute the above code, even
4739 * if we do nothing with a GOSUB */
4741 ( flags & SCF_IN_DEFINE )
4744 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
4746 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
4749 /* no need to do anything here if we are in a define. */
4750 /* or we are after some kind of infinite construct
4751 * so we can skip recursing into this item.
4752 * Since it is infinite we will not change the maxlen
4753 * or delta, and if we miss something that might raise
4754 * the minlen it will merely pessimise a little.
4756 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
4757 * might result in a minlen of 1 and not of 4,
4758 * but this doesn't make us mismatch, just try a bit
4759 * harder than we should.
4761 scan= regnext(scan);
4768 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4770 /* it is quite possible that there are more efficient ways
4771 * to do this. We maintain a bitmap per level of recursion
4772 * of which patterns we have entered so we can detect if a
4773 * pattern creates a possible infinite loop. When we
4774 * recurse down a level we copy the previous levels bitmap
4775 * down. When we are at recursion level 0 we zero the top
4776 * level bitmap. It would be nice to implement a different
4777 * more efficient way of doing this. In particular the top
4778 * level bitmap may be unnecessary.
4780 if (!recursed_depth) {
4781 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4783 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4784 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4785 RExC_study_chunk_recursed_bytes, U8);
4787 /* we havent recursed into this paren yet, so recurse into it */
4788 DEBUG_STUDYDATA("gosub-set:", data,depth);
4789 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4790 my_recursed_depth= recursed_depth + 1;
4792 DEBUG_STUDYDATA("gosub-inf:", data,depth);
4793 /* some form of infinite recursion, assume infinite length
4795 if (flags & SCF_DO_SUBSTR) {
4796 scan_commit(pRExC_state, data, minlenp, is_inf);
4797 data->longest = &(data->longest_float);
4799 is_inf = is_inf_internal = 1;
4800 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4801 ssc_anything(data->start_class);
4802 flags &= ~SCF_DO_STCLASS;
4804 start= NULL; /* reset start so we dont recurse later on. */
4809 end = regnext(scan);
4812 scan_frame *newframe;
4814 if (!RExC_frame_last) {
4815 Newxz(newframe, 1, scan_frame);
4816 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
4817 RExC_frame_head= newframe;
4819 } else if (!RExC_frame_last->next_frame) {
4820 Newxz(newframe,1,scan_frame);
4821 RExC_frame_last->next_frame= newframe;
4822 newframe->prev_frame= RExC_frame_last;
4825 newframe= RExC_frame_last->next_frame;
4827 RExC_frame_last= newframe;
4829 newframe->next_regnode = regnext(scan);
4830 newframe->last_regnode = last;
4831 newframe->stopparen = stopparen;
4832 newframe->prev_recursed_depth = recursed_depth;
4833 newframe->this_prev_frame= frame;
4835 DEBUG_STUDYDATA("frame-new:",data,depth);
4836 DEBUG_PEEP("fnew", scan, depth);
4843 recursed_depth= my_recursed_depth;
4848 else if (OP(scan) == EXACT || OP(scan) == EXACTL) {
4849 SSize_t l = STR_LEN(scan);
4852 const U8 * const s = (U8*)STRING(scan);
4853 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4854 l = utf8_length(s, s + l);
4856 uc = *((U8*)STRING(scan));
4859 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4860 /* The code below prefers earlier match for fixed
4861 offset, later match for variable offset. */
4862 if (data->last_end == -1) { /* Update the start info. */
4863 data->last_start_min = data->pos_min;
4864 data->last_start_max = is_inf
4865 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4867 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4869 SvUTF8_on(data->last_found);
4871 SV * const sv = data->last_found;
4872 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4873 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4874 if (mg && mg->mg_len >= 0)
4875 mg->mg_len += utf8_length((U8*)STRING(scan),
4876 (U8*)STRING(scan)+STR_LEN(scan));
4878 data->last_end = data->pos_min + l;
4879 data->pos_min += l; /* As in the first entry. */
4880 data->flags &= ~SF_BEFORE_EOL;
4883 /* ANDing the code point leaves at most it, and not in locale, and
4884 * can't match null string */
4885 if (flags & SCF_DO_STCLASS_AND) {
4886 ssc_cp_and(data->start_class, uc);
4887 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4888 ssc_clear_locale(data->start_class);
4890 else if (flags & SCF_DO_STCLASS_OR) {
4891 ssc_add_cp(data->start_class, uc);
4892 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4894 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4895 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4897 flags &= ~SCF_DO_STCLASS;
4899 else if (PL_regkind[OP(scan)] == EXACT) {
4900 /* But OP != EXACT!, so is EXACTFish */
4901 SSize_t l = STR_LEN(scan);
4902 const U8 * s = (U8*)STRING(scan);
4904 /* Search for fixed substrings supports EXACT only. */
4905 if (flags & SCF_DO_SUBSTR) {
4907 scan_commit(pRExC_state, data, minlenp, is_inf);
4910 l = utf8_length(s, s + l);
4912 if (unfolded_multi_char) {
4913 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4915 min += l - min_subtract;
4917 delta += min_subtract;
4918 if (flags & SCF_DO_SUBSTR) {
4919 data->pos_min += l - min_subtract;
4920 if (data->pos_min < 0) {
4923 data->pos_delta += min_subtract;
4925 data->longest = &(data->longest_float);
4929 if (flags & SCF_DO_STCLASS) {
4930 SV* EXACTF_invlist = _make_exactf_invlist(pRExC_state, scan);
4932 assert(EXACTF_invlist);
4933 if (flags & SCF_DO_STCLASS_AND) {
4934 if (OP(scan) != EXACTFL)
4935 ssc_clear_locale(data->start_class);
4936 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4937 ANYOF_POSIXL_ZERO(data->start_class);
4938 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4940 else { /* SCF_DO_STCLASS_OR */
4941 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4942 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4944 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4945 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4947 flags &= ~SCF_DO_STCLASS;
4948 SvREFCNT_dec(EXACTF_invlist);
4951 else if (REGNODE_VARIES(OP(scan))) {
4952 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4953 I32 fl = 0, f = flags;
4954 regnode * const oscan = scan;
4955 regnode_ssc this_class;
4956 regnode_ssc *oclass = NULL;
4957 I32 next_is_eval = 0;
4959 switch (PL_regkind[OP(scan)]) {
4960 case WHILEM: /* End of (?:...)* . */
4961 scan = NEXTOPER(scan);
4964 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4965 next = NEXTOPER(scan);
4966 if (OP(next) == EXACT
4967 || OP(next) == EXACTL
4968 || (flags & SCF_DO_STCLASS))
4971 maxcount = REG_INFTY;
4972 next = regnext(scan);
4973 scan = NEXTOPER(scan);
4977 if (flags & SCF_DO_SUBSTR)
4982 if (flags & SCF_DO_STCLASS) {
4984 maxcount = REG_INFTY;
4985 next = regnext(scan);
4986 scan = NEXTOPER(scan);
4989 if (flags & SCF_DO_SUBSTR) {
4990 scan_commit(pRExC_state, data, minlenp, is_inf);
4991 /* Cannot extend fixed substrings */
4992 data->longest = &(data->longest_float);
4994 is_inf = is_inf_internal = 1;
4995 scan = regnext(scan);
4996 goto optimize_curly_tail;
4998 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
4999 && (scan->flags == stopparen))
5004 mincount = ARG1(scan);
5005 maxcount = ARG2(scan);
5007 next = regnext(scan);
5008 if (OP(scan) == CURLYX) {
5009 I32 lp = (data ? *(data->last_closep) : 0);
5010 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5012 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5013 next_is_eval = (OP(scan) == EVAL);
5015 if (flags & SCF_DO_SUBSTR) {
5017 scan_commit(pRExC_state, data, minlenp, is_inf);
5018 /* Cannot extend fixed substrings */
5019 pos_before = data->pos_min;
5023 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5025 data->flags |= SF_IS_INF;
5027 if (flags & SCF_DO_STCLASS) {
5028 ssc_init(pRExC_state, &this_class);
5029 oclass = data->start_class;
5030 data->start_class = &this_class;
5031 f |= SCF_DO_STCLASS_AND;
5032 f &= ~SCF_DO_STCLASS_OR;
5034 /* Exclude from super-linear cache processing any {n,m}
5035 regops for which the combination of input pos and regex
5036 pos is not enough information to determine if a match
5039 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5040 regex pos at the \s*, the prospects for a match depend not
5041 only on the input position but also on how many (bar\s*)
5042 repeats into the {4,8} we are. */
5043 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5044 f &= ~SCF_WHILEM_VISITED_POS;
5046 /* This will finish on WHILEM, setting scan, or on NULL: */
5047 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5048 last, data, stopparen, recursed_depth, NULL,
5050 ? (f & ~SCF_DO_SUBSTR)
5054 if (flags & SCF_DO_STCLASS)
5055 data->start_class = oclass;
5056 if (mincount == 0 || minnext == 0) {
5057 if (flags & SCF_DO_STCLASS_OR) {
5058 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5060 else if (flags & SCF_DO_STCLASS_AND) {
5061 /* Switch to OR mode: cache the old value of
5062 * data->start_class */
5064 StructCopy(data->start_class, and_withp, regnode_ssc);
5065 flags &= ~SCF_DO_STCLASS_AND;
5066 StructCopy(&this_class, data->start_class, regnode_ssc);
5067 flags |= SCF_DO_STCLASS_OR;
5068 ANYOF_FLAGS(data->start_class)
5069 |= SSC_MATCHES_EMPTY_STRING;
5071 } else { /* Non-zero len */
5072 if (flags & SCF_DO_STCLASS_OR) {
5073 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5074 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5076 else if (flags & SCF_DO_STCLASS_AND)
5077 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5078 flags &= ~SCF_DO_STCLASS;
5080 if (!scan) /* It was not CURLYX, but CURLY. */
5082 if (!(flags & SCF_TRIE_DOING_RESTUDY)
5083 /* ? quantifier ok, except for (?{ ... }) */
5084 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5085 && (minnext == 0) && (deltanext == 0)
5086 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5087 && maxcount <= REG_INFTY/3) /* Complement check for big
5090 /* Fatal warnings may leak the regexp without this: */
5091 SAVEFREESV(RExC_rx_sv);
5092 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5093 "Quantifier unexpected on zero-length expression "
5094 "in regex m/%" UTF8f "/",
5095 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5097 (void)ReREFCNT_inc(RExC_rx_sv);
5100 min += minnext * mincount;
5101 is_inf_internal |= deltanext == SSize_t_MAX
5102 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5103 is_inf |= is_inf_internal;
5105 delta = SSize_t_MAX;
5107 delta += (minnext + deltanext) * maxcount
5108 - minnext * mincount;
5110 /* Try powerful optimization CURLYX => CURLYN. */
5111 if ( OP(oscan) == CURLYX && data
5112 && data->flags & SF_IN_PAR
5113 && !(data->flags & SF_HAS_EVAL)
5114 && !deltanext && minnext == 1 ) {
5115 /* Try to optimize to CURLYN. */
5116 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5117 regnode * const nxt1 = nxt;
5124 if (!REGNODE_SIMPLE(OP(nxt))
5125 && !(PL_regkind[OP(nxt)] == EXACT
5126 && STR_LEN(nxt) == 1))
5132 if (OP(nxt) != CLOSE)
5134 if (RExC_open_parens) {
5135 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5136 RExC_close_parens[ARG(nxt1)]=nxt+2; /*close->while*/
5138 /* Now we know that nxt2 is the only contents: */
5139 oscan->flags = (U8)ARG(nxt);
5141 OP(nxt1) = NOTHING; /* was OPEN. */
5144 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5145 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5146 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5147 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5148 OP(nxt + 1) = OPTIMIZED; /* was count. */
5149 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5154 /* Try optimization CURLYX => CURLYM. */
5155 if ( OP(oscan) == CURLYX && data
5156 && !(data->flags & SF_HAS_PAR)
5157 && !(data->flags & SF_HAS_EVAL)
5158 && !deltanext /* atom is fixed width */
5159 && minnext != 0 /* CURLYM can't handle zero width */
5161 /* Nor characters whose fold at run-time may be
5162 * multi-character */
5163 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5165 /* XXXX How to optimize if data == 0? */
5166 /* Optimize to a simpler form. */
5167 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5171 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5172 && (OP(nxt2) != WHILEM))
5174 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5175 /* Need to optimize away parenths. */
5176 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5177 /* Set the parenth number. */
5178 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5180 oscan->flags = (U8)ARG(nxt);
5181 if (RExC_open_parens) {
5182 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5183 RExC_close_parens[ARG(nxt1)]=nxt2+1; /*close->NOTHING*/
5185 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5186 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5189 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5190 OP(nxt + 1) = OPTIMIZED; /* was count. */
5191 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5192 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5195 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5196 regnode *nnxt = regnext(nxt1);
5198 if (reg_off_by_arg[OP(nxt1)])
5199 ARG_SET(nxt1, nxt2 - nxt1);
5200 else if (nxt2 - nxt1 < U16_MAX)
5201 NEXT_OFF(nxt1) = nxt2 - nxt1;
5203 OP(nxt) = NOTHING; /* Cannot beautify */
5208 /* Optimize again: */
5209 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5210 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
5215 else if ((OP(oscan) == CURLYX)
5216 && (flags & SCF_WHILEM_VISITED_POS)
5217 /* See the comment on a similar expression above.
5218 However, this time it's not a subexpression
5219 we care about, but the expression itself. */
5220 && (maxcount == REG_INFTY)
5222 /* This stays as CURLYX, we can put the count/of pair. */
5223 /* Find WHILEM (as in regexec.c) */
5224 regnode *nxt = oscan + NEXT_OFF(oscan);
5226 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5228 nxt = PREVOPER(nxt);
5229 if (nxt->flags & 0xf) {
5230 /* we've already set whilem count on this node */
5231 } else if (++data->whilem_c < 16) {
5232 assert(data->whilem_c <= RExC_whilem_seen);
5233 nxt->flags = (U8)(data->whilem_c
5234 | (RExC_whilem_seen << 4)); /* On WHILEM */
5237 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5239 if (flags & SCF_DO_SUBSTR) {
5240 SV *last_str = NULL;
5241 STRLEN last_chrs = 0;
5242 int counted = mincount != 0;
5244 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5246 SSize_t b = pos_before >= data->last_start_min
5247 ? pos_before : data->last_start_min;
5249 const char * const s = SvPV_const(data->last_found, l);
5250 SSize_t old = b - data->last_start_min;
5253 old = utf8_hop((U8*)s, old) - (U8*)s;
5255 /* Get the added string: */
5256 last_str = newSVpvn_utf8(s + old, l, UTF);
5257 last_chrs = UTF ? utf8_length((U8*)(s + old),
5258 (U8*)(s + old + l)) : l;
5259 if (deltanext == 0 && pos_before == b) {
5260 /* What was added is a constant string */
5263 SvGROW(last_str, (mincount * l) + 1);
5264 repeatcpy(SvPVX(last_str) + l,
5265 SvPVX_const(last_str), l,
5267 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5268 /* Add additional parts. */
5269 SvCUR_set(data->last_found,
5270 SvCUR(data->last_found) - l);
5271 sv_catsv(data->last_found, last_str);
5273 SV * sv = data->last_found;
5275 SvUTF8(sv) && SvMAGICAL(sv) ?
5276 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5277 if (mg && mg->mg_len >= 0)
5278 mg->mg_len += last_chrs * (mincount-1);
5280 last_chrs *= mincount;
5281 data->last_end += l * (mincount - 1);
5284 /* start offset must point into the last copy */
5285 data->last_start_min += minnext * (mincount - 1);
5286 data->last_start_max =
5289 : data->last_start_max +
5290 (maxcount - 1) * (minnext + data->pos_delta);
5293 /* It is counted once already... */
5294 data->pos_min += minnext * (mincount - counted);
5296 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5297 " SSize_t_MAX=%" UVuf " minnext=%" UVuf
5298 " maxcount=%" UVuf " mincount=%" UVuf "\n",
5299 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
5301 if (deltanext != SSize_t_MAX)
5302 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5303 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5304 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
5306 if (deltanext == SSize_t_MAX
5307 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
5308 data->pos_delta = SSize_t_MAX;
5310 data->pos_delta += - counted * deltanext +
5311 (minnext + deltanext) * maxcount - minnext * mincount;
5312 if (mincount != maxcount) {
5313 /* Cannot extend fixed substrings found inside
5315 scan_commit(pRExC_state, data, minlenp, is_inf);
5316 if (mincount && last_str) {
5317 SV * const sv = data->last_found;
5318 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5319 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5323 sv_setsv(sv, last_str);
5324 data->last_end = data->pos_min;
5325 data->last_start_min = data->pos_min - last_chrs;
5326 data->last_start_max = is_inf
5328 : data->pos_min + data->pos_delta - last_chrs;
5330 data->longest = &(data->longest_float);
5332 SvREFCNT_dec(last_str);
5334 if (data && (fl & SF_HAS_EVAL))
5335 data->flags |= SF_HAS_EVAL;
5336 optimize_curly_tail:
5337 if (OP(oscan) != CURLYX) {
5338 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
5340 NEXT_OFF(oscan) += NEXT_OFF(next);
5346 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5351 if (flags & SCF_DO_SUBSTR) {
5352 /* Cannot expect anything... */
5353 scan_commit(pRExC_state, data, minlenp, is_inf);
5354 data->longest = &(data->longest_float);
5356 is_inf = is_inf_internal = 1;
5357 if (flags & SCF_DO_STCLASS_OR) {
5358 if (OP(scan) == CLUMP) {
5359 /* Actually is any start char, but very few code points
5360 * aren't start characters */
5361 ssc_match_all_cp(data->start_class);
5364 ssc_anything(data->start_class);
5367 flags &= ~SCF_DO_STCLASS;
5371 else if (OP(scan) == LNBREAK) {
5372 if (flags & SCF_DO_STCLASS) {
5373 if (flags & SCF_DO_STCLASS_AND) {
5374 ssc_intersection(data->start_class,
5375 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5376 ssc_clear_locale(data->start_class);
5377 ANYOF_FLAGS(data->start_class)
5378 &= ~SSC_MATCHES_EMPTY_STRING;
5380 else if (flags & SCF_DO_STCLASS_OR) {
5381 ssc_union(data->start_class,
5382 PL_XPosix_ptrs[_CC_VERTSPACE],
5384 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5386 /* See commit msg for
5387 * 749e076fceedeb708a624933726e7989f2302f6a */
5388 ANYOF_FLAGS(data->start_class)
5389 &= ~SSC_MATCHES_EMPTY_STRING;
5391 flags &= ~SCF_DO_STCLASS;
5394 if (delta != SSize_t_MAX)
5395 delta++; /* Because of the 2 char string cr-lf */
5396 if (flags & SCF_DO_SUBSTR) {
5397 /* Cannot expect anything... */
5398 scan_commit(pRExC_state, data, minlenp, is_inf);
5400 data->pos_delta += 1;
5401 data->longest = &(data->longest_float);
5404 else if (REGNODE_SIMPLE(OP(scan))) {
5406 if (flags & SCF_DO_SUBSTR) {
5407 scan_commit(pRExC_state, data, minlenp, is_inf);
5411 if (flags & SCF_DO_STCLASS) {
5413 SV* my_invlist = NULL;
5416 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5417 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5419 /* Some of the logic below assumes that switching
5420 locale on will only add false positives. */
5425 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5429 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5430 ssc_match_all_cp(data->start_class);
5435 SV* REG_ANY_invlist = _new_invlist(2);
5436 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5438 if (flags & SCF_DO_STCLASS_OR) {
5439 ssc_union(data->start_class,
5441 TRUE /* TRUE => invert, hence all but \n
5445 else if (flags & SCF_DO_STCLASS_AND) {
5446 ssc_intersection(data->start_class,
5448 TRUE /* TRUE => invert */
5450 ssc_clear_locale(data->start_class);
5452 SvREFCNT_dec_NN(REG_ANY_invlist);
5459 if (flags & SCF_DO_STCLASS_AND)
5460 ssc_and(pRExC_state, data->start_class,
5461 (regnode_charclass *) scan);
5463 ssc_or(pRExC_state, data->start_class,
5464 (regnode_charclass *) scan);
5472 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5473 if (flags & SCF_DO_STCLASS_AND) {
5474 bool was_there = cBOOL(
5475 ANYOF_POSIXL_TEST(data->start_class,
5477 ANYOF_POSIXL_ZERO(data->start_class);
5478 if (was_there) { /* Do an AND */
5479 ANYOF_POSIXL_SET(data->start_class, namedclass);
5481 /* No individual code points can now match */
5482 data->start_class->invlist
5483 = sv_2mortal(_new_invlist(0));
5486 int complement = namedclass + ((invert) ? -1 : 1);
5488 assert(flags & SCF_DO_STCLASS_OR);
5490 /* If the complement of this class was already there,
5491 * the result is that they match all code points,
5492 * (\d + \D == everything). Remove the classes from
5493 * future consideration. Locale is not relevant in
5495 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5496 ssc_match_all_cp(data->start_class);
5497 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5498 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5500 else { /* The usual case; just add this class to the
5502 ANYOF_POSIXL_SET(data->start_class, namedclass);
5507 case NPOSIXA: /* For these, we always know the exact set of
5512 if (FLAGS(scan) == _CC_ASCII) {
5513 my_invlist = invlist_clone(PL_XPosix_ptrs[_CC_ASCII]);
5516 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
5517 PL_XPosix_ptrs[_CC_ASCII],
5528 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5530 /* NPOSIXD matches all upper Latin1 code points unless the
5531 * target string being matched is UTF-8, which is
5532 * unknowable until match time. Since we are going to
5533 * invert, we want to get rid of all of them so that the
5534 * inversion will match all */
5535 if (OP(scan) == NPOSIXD) {
5536 _invlist_subtract(my_invlist, PL_UpperLatin1,
5542 if (flags & SCF_DO_STCLASS_AND) {
5543 ssc_intersection(data->start_class, my_invlist, invert);
5544 ssc_clear_locale(data->start_class);
5547 assert(flags & SCF_DO_STCLASS_OR);
5548 ssc_union(data->start_class, my_invlist, invert);
5550 SvREFCNT_dec(my_invlist);
5552 if (flags & SCF_DO_STCLASS_OR)
5553 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5554 flags &= ~SCF_DO_STCLASS;
5557 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5558 data->flags |= (OP(scan) == MEOL
5561 scan_commit(pRExC_state, data, minlenp, is_inf);
5564 else if ( PL_regkind[OP(scan)] == BRANCHJ
5565 /* Lookbehind, or need to calculate parens/evals/stclass: */
5566 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5567 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5569 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5570 || OP(scan) == UNLESSM )
5572 /* Negative Lookahead/lookbehind
5573 In this case we can't do fixed string optimisation.
5576 SSize_t deltanext, minnext, fake = 0;
5581 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5583 data_fake.whilem_c = data->whilem_c;
5584 data_fake.last_closep = data->last_closep;
5587 data_fake.last_closep = &fake;
5588 data_fake.pos_delta = delta;
5589 if ( flags & SCF_DO_STCLASS && !scan->flags
5590 && OP(scan) == IFMATCH ) { /* Lookahead */
5591 ssc_init(pRExC_state, &intrnl);
5592 data_fake.start_class = &intrnl;
5593 f |= SCF_DO_STCLASS_AND;
5595 if (flags & SCF_WHILEM_VISITED_POS)
5596 f |= SCF_WHILEM_VISITED_POS;
5597 next = regnext(scan);
5598 nscan = NEXTOPER(NEXTOPER(scan));
5599 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5600 last, &data_fake, stopparen,
5601 recursed_depth, NULL, f, depth+1);
5604 FAIL("Variable length lookbehind not implemented");
5606 else if (minnext > (I32)U8_MAX) {
5607 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5610 scan->flags = (U8)minnext;
5613 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5615 if (data_fake.flags & SF_HAS_EVAL)
5616 data->flags |= SF_HAS_EVAL;
5617 data->whilem_c = data_fake.whilem_c;
5619 if (f & SCF_DO_STCLASS_AND) {
5620 if (flags & SCF_DO_STCLASS_OR) {
5621 /* OR before, AND after: ideally we would recurse with
5622 * data_fake to get the AND applied by study of the
5623 * remainder of the pattern, and then derecurse;
5624 * *** HACK *** for now just treat as "no information".
5625 * See [perl #56690].
5627 ssc_init(pRExC_state, data->start_class);
5629 /* AND before and after: combine and continue. These
5630 * assertions are zero-length, so can match an EMPTY
5632 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5633 ANYOF_FLAGS(data->start_class)
5634 |= SSC_MATCHES_EMPTY_STRING;
5638 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5640 /* Positive Lookahead/lookbehind
5641 In this case we can do fixed string optimisation,
5642 but we must be careful about it. Note in the case of
5643 lookbehind the positions will be offset by the minimum
5644 length of the pattern, something we won't know about
5645 until after the recurse.
5647 SSize_t deltanext, fake = 0;
5651 /* We use SAVEFREEPV so that when the full compile
5652 is finished perl will clean up the allocated
5653 minlens when it's all done. This way we don't
5654 have to worry about freeing them when we know
5655 they wont be used, which would be a pain.
5658 Newx( minnextp, 1, SSize_t );
5659 SAVEFREEPV(minnextp);
5662 StructCopy(data, &data_fake, scan_data_t);
5663 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5666 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5667 data_fake.last_found=newSVsv(data->last_found);
5671 data_fake.last_closep = &fake;
5672 data_fake.flags = 0;
5673 data_fake.pos_delta = delta;
5675 data_fake.flags |= SF_IS_INF;
5676 if ( flags & SCF_DO_STCLASS && !scan->flags
5677 && OP(scan) == IFMATCH ) { /* Lookahead */
5678 ssc_init(pRExC_state, &intrnl);
5679 data_fake.start_class = &intrnl;
5680 f |= SCF_DO_STCLASS_AND;
5682 if (flags & SCF_WHILEM_VISITED_POS)
5683 f |= SCF_WHILEM_VISITED_POS;
5684 next = regnext(scan);
5685 nscan = NEXTOPER(NEXTOPER(scan));
5687 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5688 &deltanext, last, &data_fake,
5689 stopparen, recursed_depth, NULL,
5693 FAIL("Variable length lookbehind not implemented");
5695 else if (*minnextp > (I32)U8_MAX) {
5696 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5699 scan->flags = (U8)*minnextp;
5704 if (f & SCF_DO_STCLASS_AND) {
5705 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5706 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
5709 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5711 if (data_fake.flags & SF_HAS_EVAL)
5712 data->flags |= SF_HAS_EVAL;
5713 data->whilem_c = data_fake.whilem_c;
5714 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5715 if (RExC_rx->minlen<*minnextp)
5716 RExC_rx->minlen=*minnextp;
5717 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5718 SvREFCNT_dec_NN(data_fake.last_found);
5720 if ( data_fake.minlen_fixed != minlenp )
5722 data->offset_fixed= data_fake.offset_fixed;
5723 data->minlen_fixed= data_fake.minlen_fixed;
5724 data->lookbehind_fixed+= scan->flags;
5726 if ( data_fake.minlen_float != minlenp )
5728 data->minlen_float= data_fake.minlen_float;
5729 data->offset_float_min=data_fake.offset_float_min;
5730 data->offset_float_max=data_fake.offset_float_max;
5731 data->lookbehind_float+= scan->flags;
5738 else if (OP(scan) == OPEN) {
5739 if (stopparen != (I32)ARG(scan))
5742 else if (OP(scan) == CLOSE) {
5743 if (stopparen == (I32)ARG(scan)) {
5746 if ((I32)ARG(scan) == is_par) {
5747 next = regnext(scan);
5749 if ( next && (OP(next) != WHILEM) && next < last)
5750 is_par = 0; /* Disable optimization */
5753 *(data->last_closep) = ARG(scan);
5755 else if (OP(scan) == EVAL) {
5757 data->flags |= SF_HAS_EVAL;
5759 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5760 if (flags & SCF_DO_SUBSTR) {
5761 scan_commit(pRExC_state, data, minlenp, is_inf);
5762 flags &= ~SCF_DO_SUBSTR;
5764 if (data && OP(scan)==ACCEPT) {
5765 data->flags |= SCF_SEEN_ACCEPT;
5770 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5772 if (flags & SCF_DO_SUBSTR) {
5773 scan_commit(pRExC_state, data, minlenp, is_inf);
5774 data->longest = &(data->longest_float);
5776 is_inf = is_inf_internal = 1;
5777 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5778 ssc_anything(data->start_class);
5779 flags &= ~SCF_DO_STCLASS;
5781 else if (OP(scan) == GPOS) {
5782 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5783 !(delta || is_inf || (data && data->pos_delta)))
5785 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5786 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5787 if (RExC_rx->gofs < (STRLEN)min)
5788 RExC_rx->gofs = min;
5790 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5794 #ifdef TRIE_STUDY_OPT
5795 #ifdef FULL_TRIE_STUDY
5796 else if (PL_regkind[OP(scan)] == TRIE) {
5797 /* NOTE - There is similar code to this block above for handling
5798 BRANCH nodes on the initial study. If you change stuff here
5800 regnode *trie_node= scan;
5801 regnode *tail= regnext(scan);
5802 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5803 SSize_t max1 = 0, min1 = SSize_t_MAX;
5806 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5807 /* Cannot merge strings after this. */
5808 scan_commit(pRExC_state, data, minlenp, is_inf);
5810 if (flags & SCF_DO_STCLASS)
5811 ssc_init_zero(pRExC_state, &accum);
5817 const regnode *nextbranch= NULL;
5820 for ( word=1 ; word <= trie->wordcount ; word++)
5822 SSize_t deltanext=0, minnext=0, f = 0, fake;
5823 regnode_ssc this_class;
5825 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5827 data_fake.whilem_c = data->whilem_c;
5828 data_fake.last_closep = data->last_closep;
5831 data_fake.last_closep = &fake;
5832 data_fake.pos_delta = delta;
5833 if (flags & SCF_DO_STCLASS) {
5834 ssc_init(pRExC_state, &this_class);
5835 data_fake.start_class = &this_class;
5836 f = SCF_DO_STCLASS_AND;
5838 if (flags & SCF_WHILEM_VISITED_POS)
5839 f |= SCF_WHILEM_VISITED_POS;
5841 if (trie->jump[word]) {
5843 nextbranch = trie_node + trie->jump[0];
5844 scan= trie_node + trie->jump[word];
5845 /* We go from the jump point to the branch that follows
5846 it. Note this means we need the vestigal unused
5847 branches even though they arent otherwise used. */
5848 minnext = study_chunk(pRExC_state, &scan, minlenp,
5849 &deltanext, (regnode *)nextbranch, &data_fake,
5850 stopparen, recursed_depth, NULL, f,depth+1);
5852 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5853 nextbranch= regnext((regnode*)nextbranch);
5855 if (min1 > (SSize_t)(minnext + trie->minlen))
5856 min1 = minnext + trie->minlen;
5857 if (deltanext == SSize_t_MAX) {
5858 is_inf = is_inf_internal = 1;
5860 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5861 max1 = minnext + deltanext + trie->maxlen;
5863 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5865 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5866 if ( stopmin > min + min1)
5867 stopmin = min + min1;
5868 flags &= ~SCF_DO_SUBSTR;
5870 data->flags |= SCF_SEEN_ACCEPT;
5873 if (data_fake.flags & SF_HAS_EVAL)
5874 data->flags |= SF_HAS_EVAL;
5875 data->whilem_c = data_fake.whilem_c;
5877 if (flags & SCF_DO_STCLASS)
5878 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5881 if (flags & SCF_DO_SUBSTR) {
5882 data->pos_min += min1;
5883 data->pos_delta += max1 - min1;
5884 if (max1 != min1 || is_inf)
5885 data->longest = &(data->longest_float);
5888 if (delta != SSize_t_MAX)
5889 delta += max1 - min1;
5890 if (flags & SCF_DO_STCLASS_OR) {
5891 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5893 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5894 flags &= ~SCF_DO_STCLASS;
5897 else if (flags & SCF_DO_STCLASS_AND) {
5899 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5900 flags &= ~SCF_DO_STCLASS;
5903 /* Switch to OR mode: cache the old value of
5904 * data->start_class */
5906 StructCopy(data->start_class, and_withp, regnode_ssc);
5907 flags &= ~SCF_DO_STCLASS_AND;
5908 StructCopy(&accum, data->start_class, regnode_ssc);
5909 flags |= SCF_DO_STCLASS_OR;
5916 else if (PL_regkind[OP(scan)] == TRIE) {
5917 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5920 min += trie->minlen;
5921 delta += (trie->maxlen - trie->minlen);
5922 flags &= ~SCF_DO_STCLASS; /* xxx */
5923 if (flags & SCF_DO_SUBSTR) {
5924 /* Cannot expect anything... */
5925 scan_commit(pRExC_state, data, minlenp, is_inf);
5926 data->pos_min += trie->minlen;
5927 data->pos_delta += (trie->maxlen - trie->minlen);
5928 if (trie->maxlen != trie->minlen)
5929 data->longest = &(data->longest_float);
5931 if (trie->jump) /* no more substrings -- for now /grr*/
5932 flags &= ~SCF_DO_SUBSTR;
5934 #endif /* old or new */
5935 #endif /* TRIE_STUDY_OPT */
5937 /* Else: zero-length, ignore. */
5938 scan = regnext(scan);
5943 /* we need to unwind recursion. */
5946 DEBUG_STUDYDATA("frame-end:",data,depth);
5947 DEBUG_PEEP("fend", scan, depth);
5949 /* restore previous context */
5950 last = frame->last_regnode;
5951 scan = frame->next_regnode;
5952 stopparen = frame->stopparen;
5953 recursed_depth = frame->prev_recursed_depth;
5955 RExC_frame_last = frame->prev_frame;
5956 frame = frame->this_prev_frame;
5957 goto fake_study_recurse;
5961 DEBUG_STUDYDATA("pre-fin:",data,depth);
5964 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5966 if (flags & SCF_DO_SUBSTR && is_inf)
5967 data->pos_delta = SSize_t_MAX - data->pos_min;
5968 if (is_par > (I32)U8_MAX)
5970 if (is_par && pars==1 && data) {
5971 data->flags |= SF_IN_PAR;
5972 data->flags &= ~SF_HAS_PAR;
5974 else if (pars && data) {
5975 data->flags |= SF_HAS_PAR;
5976 data->flags &= ~SF_IN_PAR;
5978 if (flags & SCF_DO_STCLASS_OR)
5979 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5980 if (flags & SCF_TRIE_RESTUDY)
5981 data->flags |= SCF_TRIE_RESTUDY;
5983 DEBUG_STUDYDATA("post-fin:",data,depth);
5986 SSize_t final_minlen= min < stopmin ? min : stopmin;
5988 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
5989 if (final_minlen > SSize_t_MAX - delta)
5990 RExC_maxlen = SSize_t_MAX;
5991 else if (RExC_maxlen < final_minlen + delta)
5992 RExC_maxlen = final_minlen + delta;
5994 return final_minlen;
5996 NOT_REACHED; /* NOTREACHED */
6000 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
6002 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
6004 PERL_ARGS_ASSERT_ADD_DATA;
6006 Renewc(RExC_rxi->data,
6007 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6008 char, struct reg_data);
6010 Renew(RExC_rxi->data->what, count + n, U8);
6012 Newx(RExC_rxi->data->what, n, U8);
6013 RExC_rxi->data->count = count + n;
6014 Copy(s, RExC_rxi->data->what + count, n, U8);
6018 /*XXX: todo make this not included in a non debugging perl, but appears to be
6019 * used anyway there, in 'use re' */
6020 #ifndef PERL_IN_XSUB_RE
6022 Perl_reginitcolors(pTHX)
6024 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6026 char *t = savepv(s);
6030 t = strchr(t, '\t');
6036 PL_colors[i] = t = (char *)"";
6041 PL_colors[i++] = (char *)"";
6048 #ifdef TRIE_STUDY_OPT
6049 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6052 (data.flags & SCF_TRIE_RESTUDY) \
6060 #define CHECK_RESTUDY_GOTO_butfirst
6064 * pregcomp - compile a regular expression into internal code
6066 * Decides which engine's compiler to call based on the hint currently in
6070 #ifndef PERL_IN_XSUB_RE
6072 /* return the currently in-scope regex engine (or the default if none) */
6074 regexp_engine const *
6075 Perl_current_re_engine(pTHX)
6077 if (IN_PERL_COMPILETIME) {
6078 HV * const table = GvHV(PL_hintgv);
6081 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6082 return &PL_core_reg_engine;
6083 ptr = hv_fetchs(table, "regcomp", FALSE);
6084 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6085 return &PL_core_reg_engine;
6086 return INT2PTR(regexp_engine*,SvIV(*ptr));
6090 if (!PL_curcop->cop_hints_hash)
6091 return &PL_core_reg_engine;
6092 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6093 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6094 return &PL_core_reg_engine;
6095 return INT2PTR(regexp_engine*,SvIV(ptr));
6101 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6103 regexp_engine const *eng = current_re_engine();
6104 GET_RE_DEBUG_FLAGS_DECL;
6106 PERL_ARGS_ASSERT_PREGCOMP;
6108 /* Dispatch a request to compile a regexp to correct regexp engine. */
6110 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6113 return CALLREGCOMP_ENG(eng, pattern, flags);
6117 /* public(ish) entry point for the perl core's own regex compiling code.
6118 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6119 * pattern rather than a list of OPs, and uses the internal engine rather
6120 * than the current one */
6123 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6125 SV *pat = pattern; /* defeat constness! */
6126 PERL_ARGS_ASSERT_RE_COMPILE;
6127 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6128 #ifdef PERL_IN_XSUB_RE
6131 &PL_core_reg_engine,
6133 NULL, NULL, rx_flags, 0);
6138 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6142 if (--cbs->refcnt > 0)
6144 for (n = 0; n < cbs->count; n++) {
6145 REGEXP *rx = cbs->cb[n].src_regex;
6146 cbs->cb[n].src_regex = NULL;
6154 static struct reg_code_blocks *
6155 S_alloc_code_blocks(pTHX_ int ncode)
6157 struct reg_code_blocks *cbs;
6158 Newx(cbs, 1, struct reg_code_blocks);
6161 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6163 Newx(cbs->cb, ncode, struct reg_code_block);
6170 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6171 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6172 * point to the realloced string and length.
6174 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6178 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6179 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6181 U8 *const src = (U8*)*pat_p;
6186 GET_RE_DEBUG_FLAGS_DECL;
6188 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6189 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6191 Newx(dst, *plen_p * 2 + 1, U8);
6194 while (s < *plen_p) {
6195 append_utf8_from_native_byte(src[s], &d);
6197 if (n < num_code_blocks) {
6198 assert(pRExC_state->code_blocks);
6199 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6200 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6201 assert(*(d - 1) == '(');
6204 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6205 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6206 assert(*(d - 1) == ')');
6215 *pat_p = (char*) dst;
6217 RExC_orig_utf8 = RExC_utf8 = 1;
6222 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6223 * while recording any code block indices, and handling overloading,
6224 * nested qr// objects etc. If pat is null, it will allocate a new
6225 * string, or just return the first arg, if there's only one.
6227 * Returns the malloced/updated pat.
6228 * patternp and pat_count is the array of SVs to be concatted;
6229 * oplist is the optional list of ops that generated the SVs;
6230 * recompile_p is a pointer to a boolean that will be set if
6231 * the regex will need to be recompiled.
6232 * delim, if non-null is an SV that will be inserted between each element
6236 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6237 SV *pat, SV ** const patternp, int pat_count,
6238 OP *oplist, bool *recompile_p, SV *delim)
6242 bool use_delim = FALSE;
6243 bool alloced = FALSE;
6245 /* if we know we have at least two args, create an empty string,
6246 * then concatenate args to that. For no args, return an empty string */
6247 if (!pat && pat_count != 1) {
6253 for (svp = patternp; svp < patternp + pat_count; svp++) {
6256 STRLEN orig_patlen = 0;
6258 SV *msv = use_delim ? delim : *svp;
6259 if (!msv) msv = &PL_sv_undef;
6261 /* if we've got a delimiter, we go round the loop twice for each
6262 * svp slot (except the last), using the delimiter the second
6271 if (SvTYPE(msv) == SVt_PVAV) {
6272 /* we've encountered an interpolated array within
6273 * the pattern, e.g. /...@a..../. Expand the list of elements,
6274 * then recursively append elements.
6275 * The code in this block is based on S_pushav() */
6277 AV *const av = (AV*)msv;
6278 const SSize_t maxarg = AvFILL(av) + 1;
6282 assert(oplist->op_type == OP_PADAV
6283 || oplist->op_type == OP_RV2AV);
6284 oplist = OpSIBLING(oplist);
6287 if (SvRMAGICAL(av)) {
6290 Newx(array, maxarg, SV*);
6292 for (i=0; i < maxarg; i++) {
6293 SV ** const svp = av_fetch(av, i, FALSE);
6294 array[i] = svp ? *svp : &PL_sv_undef;
6298 array = AvARRAY(av);
6300 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6301 array, maxarg, NULL, recompile_p,
6303 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6309 /* we make the assumption here that each op in the list of
6310 * op_siblings maps to one SV pushed onto the stack,
6311 * except for code blocks, with have both an OP_NULL and
6313 * This allows us to match up the list of SVs against the
6314 * list of OPs to find the next code block.
6316 * Note that PUSHMARK PADSV PADSV ..
6318 * PADRANGE PADSV PADSV ..
6319 * so the alignment still works. */
6322 if (oplist->op_type == OP_NULL
6323 && (oplist->op_flags & OPf_SPECIAL))
6325 assert(n < pRExC_state->code_blocks->count);
6326 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6327 pRExC_state->code_blocks->cb[n].block = oplist;
6328 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6331 oplist = OpSIBLING(oplist); /* skip CONST */
6334 oplist = OpSIBLING(oplist);;
6337 /* apply magic and QR overloading to arg */
6340 if (SvROK(msv) && SvAMAGIC(msv)) {
6341 SV *sv = AMG_CALLunary(msv, regexp_amg);
6345 if (SvTYPE(sv) != SVt_REGEXP)
6346 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6351 /* try concatenation overload ... */
6352 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6353 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6356 /* overloading involved: all bets are off over literal
6357 * code. Pretend we haven't seen it */
6359 pRExC_state->code_blocks->count -= n;
6363 /* ... or failing that, try "" overload */
6364 while (SvAMAGIC(msv)
6365 && (sv = AMG_CALLunary(msv, string_amg))
6369 && SvRV(msv) == SvRV(sv))
6374 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6378 /* this is a partially unrolled
6379 * sv_catsv_nomg(pat, msv);
6380 * that allows us to adjust code block indices if
6383 char *dst = SvPV_force_nomg(pat, dlen);
6385 if (SvUTF8(msv) && !SvUTF8(pat)) {
6386 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6387 sv_setpvn(pat, dst, dlen);
6390 sv_catsv_nomg(pat, msv);
6394 /* We have only one SV to process, but we need to verify
6395 * it is properly null terminated or we will fail asserts
6396 * later. In theory we probably shouldn't get such SV's,
6397 * but if we do we should handle it gracefully. */
6398 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) ) {
6399 /* not a string, or a string with a trailing null */
6402 /* a string with no trailing null, we need to copy it
6403 * so it we have a trailing null */
6409 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
6412 /* extract any code blocks within any embedded qr//'s */
6413 if (rx && SvTYPE(rx) == SVt_REGEXP
6414 && RX_ENGINE((REGEXP*)rx)->op_comp)
6417 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6418 if (ri->code_blocks && ri->code_blocks->count) {
6420 /* the presence of an embedded qr// with code means
6421 * we should always recompile: the text of the
6422 * qr// may not have changed, but it may be a
6423 * different closure than last time */
6425 if (pRExC_state->code_blocks) {
6426 pRExC_state->code_blocks->count += ri->code_blocks->count;
6427 Renew(pRExC_state->code_blocks->cb,
6428 pRExC_state->code_blocks->count,
6429 struct reg_code_block);
6432 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
6433 ri->code_blocks->count);
6435 for (i=0; i < ri->code_blocks->count; i++) {
6436 struct reg_code_block *src, *dst;
6437 STRLEN offset = orig_patlen
6438 + ReANY((REGEXP *)rx)->pre_prefix;
6439 assert(n < pRExC_state->code_blocks->count);
6440 src = &ri->code_blocks->cb[i];
6441 dst = &pRExC_state->code_blocks->cb[n];
6442 dst->start = src->start + offset;
6443 dst->end = src->end + offset;
6444 dst->block = src->block;
6445 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6454 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6463 /* see if there are any run-time code blocks in the pattern.
6464 * False positives are allowed */
6467 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6468 char *pat, STRLEN plen)
6473 PERL_UNUSED_CONTEXT;
6475 for (s = 0; s < plen; s++) {
6476 if ( pRExC_state->code_blocks
6477 && n < pRExC_state->code_blocks->count
6478 && s == pRExC_state->code_blocks->cb[n].start)
6480 s = pRExC_state->code_blocks->cb[n].end;
6484 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6486 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6488 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6495 /* Handle run-time code blocks. We will already have compiled any direct
6496 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6497 * copy of it, but with any literal code blocks blanked out and
6498 * appropriate chars escaped; then feed it into
6500 * eval "qr'modified_pattern'"
6504 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6508 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6510 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6511 * and merge them with any code blocks of the original regexp.
6513 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6514 * instead, just save the qr and return FALSE; this tells our caller that
6515 * the original pattern needs upgrading to utf8.
6519 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6520 char *pat, STRLEN plen)
6524 GET_RE_DEBUG_FLAGS_DECL;
6526 if (pRExC_state->runtime_code_qr) {
6527 /* this is the second time we've been called; this should
6528 * only happen if the main pattern got upgraded to utf8
6529 * during compilation; re-use the qr we compiled first time
6530 * round (which should be utf8 too)
6532 qr = pRExC_state->runtime_code_qr;
6533 pRExC_state->runtime_code_qr = NULL;
6534 assert(RExC_utf8 && SvUTF8(qr));
6540 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
6544 /* determine how many extra chars we need for ' and \ escaping */
6545 for (s = 0; s < plen; s++) {
6546 if (pat[s] == '\'' || pat[s] == '\\')
6550 Newx(newpat, newlen, char);
6552 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6554 for (s = 0; s < plen; s++) {
6555 if ( pRExC_state->code_blocks
6556 && n < pRExC_state->code_blocks->count
6557 && s == pRExC_state->code_blocks->cb[n].start)
6559 /* blank out literal code block */
6560 assert(pat[s] == '(');
6561 while (s <= pRExC_state->code_blocks->cb[n].end) {
6569 if (pat[s] == '\'' || pat[s] == '\\')
6574 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
6576 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
6582 Perl_re_printf( aTHX_
6583 "%sre-parsing pattern for runtime code:%s %s\n",
6584 PL_colors[4],PL_colors[5],newpat);
6587 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6593 PUSHSTACKi(PERLSI_REQUIRE);
6594 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6595 * parsing qr''; normally only q'' does this. It also alters
6597 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6598 SvREFCNT_dec_NN(sv);
6603 SV * const errsv = ERRSV;
6604 if (SvTRUE_NN(errsv))
6605 /* use croak_sv ? */
6606 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
6608 assert(SvROK(qr_ref));
6610 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6611 /* the leaving below frees the tmp qr_ref.
6612 * Give qr a life of its own */
6620 if (!RExC_utf8 && SvUTF8(qr)) {
6621 /* first time through; the pattern got upgraded; save the
6622 * qr for the next time through */
6623 assert(!pRExC_state->runtime_code_qr);
6624 pRExC_state->runtime_code_qr = qr;
6629 /* extract any code blocks within the returned qr// */
6632 /* merge the main (r1) and run-time (r2) code blocks into one */
6634 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6635 struct reg_code_block *new_block, *dst;
6636 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6640 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
6642 SvREFCNT_dec_NN(qr);
6646 if (!r1->code_blocks)
6647 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
6649 r1c = r1->code_blocks->count;
6650 r2c = r2->code_blocks->count;
6652 Newx(new_block, r1c + r2c, struct reg_code_block);
6656 while (i1 < r1c || i2 < r2c) {
6657 struct reg_code_block *src;
6661 src = &r2->code_blocks->cb[i2++];
6665 src = &r1->code_blocks->cb[i1++];
6666 else if ( r1->code_blocks->cb[i1].start
6667 < r2->code_blocks->cb[i2].start)
6669 src = &r1->code_blocks->cb[i1++];
6670 assert(src->end < r2->code_blocks->cb[i2].start);
6673 assert( r1->code_blocks->cb[i1].start
6674 > r2->code_blocks->cb[i2].start);
6675 src = &r2->code_blocks->cb[i2++];
6677 assert(src->end < r1->code_blocks->cb[i1].start);
6680 assert(pat[src->start] == '(');
6681 assert(pat[src->end] == ')');
6682 dst->start = src->start;
6683 dst->end = src->end;
6684 dst->block = src->block;
6685 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6689 r1->code_blocks->count += r2c;
6690 Safefree(r1->code_blocks->cb);
6691 r1->code_blocks->cb = new_block;
6694 SvREFCNT_dec_NN(qr);
6700 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6701 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6702 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6703 STRLEN longest_length, bool eol, bool meol)
6705 /* This is the common code for setting up the floating and fixed length
6706 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6707 * as to whether succeeded or not */
6712 if (! (longest_length
6713 || (eol /* Can't have SEOL and MULTI */
6714 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6716 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6717 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6722 /* copy the information about the longest from the reg_scan_data
6723 over to the program. */
6724 if (SvUTF8(sv_longest)) {
6725 *rx_utf8 = sv_longest;
6728 *rx_substr = sv_longest;
6731 /* end_shift is how many chars that must be matched that
6732 follow this item. We calculate it ahead of time as once the
6733 lookbehind offset is added in we lose the ability to correctly
6735 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6736 *rx_end_shift = ml - offset
6738 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
6740 + (SvTAIL(sv_longest) != 0)
6744 t = (eol/* Can't have SEOL and MULTI */
6745 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6746 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6752 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6753 * regular expression into internal code.
6754 * The pattern may be passed either as:
6755 * a list of SVs (patternp plus pat_count)
6756 * a list of OPs (expr)
6757 * If both are passed, the SV list is used, but the OP list indicates
6758 * which SVs are actually pre-compiled code blocks
6760 * The SVs in the list have magic and qr overloading applied to them (and
6761 * the list may be modified in-place with replacement SVs in the latter
6764 * If the pattern hasn't changed from old_re, then old_re will be
6767 * eng is the current engine. If that engine has an op_comp method, then
6768 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6769 * do the initial concatenation of arguments and pass on to the external
6772 * If is_bare_re is not null, set it to a boolean indicating whether the
6773 * arg list reduced (after overloading) to a single bare regex which has
6774 * been returned (i.e. /$qr/).
6776 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6778 * pm_flags contains the PMf_* flags, typically based on those from the
6779 * pm_flags field of the related PMOP. Currently we're only interested in
6780 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6782 * We can't allocate space until we know how big the compiled form will be,
6783 * but we can't compile it (and thus know how big it is) until we've got a
6784 * place to put the code. So we cheat: we compile it twice, once with code
6785 * generation turned off and size counting turned on, and once "for real".
6786 * This also means that we don't allocate space until we are sure that the
6787 * thing really will compile successfully, and we never have to move the
6788 * code and thus invalidate pointers into it. (Note that it has to be in
6789 * one piece because free() must be able to free it all.) [NB: not true in perl]
6791 * Beware that the optimization-preparation code in here knows about some
6792 * of the structure of the compiled regexp. [I'll say.]
6796 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6797 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6798 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6802 regexp_internal *ri;
6810 SV** new_patternp = patternp;
6812 /* these are all flags - maybe they should be turned
6813 * into a single int with different bit masks */
6814 I32 sawlookahead = 0;
6819 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6821 bool runtime_code = 0;
6823 RExC_state_t RExC_state;
6824 RExC_state_t * const pRExC_state = &RExC_state;
6825 #ifdef TRIE_STUDY_OPT
6827 RExC_state_t copyRExC_state;
6829 GET_RE_DEBUG_FLAGS_DECL;
6831 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6833 DEBUG_r(if (!PL_colorset) reginitcolors());
6835 /* Initialize these here instead of as-needed, as is quick and avoids
6836 * having to test them each time otherwise */
6837 if (! PL_AboveLatin1) {
6839 char * dump_len_string;
6842 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6843 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6844 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6845 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6846 PL_HasMultiCharFold =
6847 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6849 /* This is calculated here, because the Perl program that generates the
6850 * static global ones doesn't currently have access to
6851 * NUM_ANYOF_CODE_POINTS */
6852 PL_InBitmap = _new_invlist(2);
6853 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
6854 NUM_ANYOF_CODE_POINTS - 1);
6856 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
6857 if ( ! dump_len_string
6858 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
6860 PL_dump_re_max_len = 0;
6865 pRExC_state->warn_text = NULL;
6866 pRExC_state->code_blocks = NULL;
6869 *is_bare_re = FALSE;
6871 if (expr && (expr->op_type == OP_LIST ||
6872 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6873 /* allocate code_blocks if needed */
6877 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
6878 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6879 ncode++; /* count of DO blocks */
6882 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
6886 /* compile-time pattern with just OP_CONSTs and DO blocks */
6891 /* find how many CONSTs there are */
6894 if (expr->op_type == OP_CONST)
6897 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6898 if (o->op_type == OP_CONST)
6902 /* fake up an SV array */
6904 assert(!new_patternp);
6905 Newx(new_patternp, n, SV*);
6906 SAVEFREEPV(new_patternp);
6910 if (expr->op_type == OP_CONST)
6911 new_patternp[n] = cSVOPx_sv(expr);
6913 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6914 if (o->op_type == OP_CONST)
6915 new_patternp[n++] = cSVOPo_sv;
6920 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6921 "Assembling pattern from %d elements%s\n", pat_count,
6922 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6924 /* set expr to the first arg op */
6926 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
6927 && expr->op_type != OP_CONST)
6929 expr = cLISTOPx(expr)->op_first;
6930 assert( expr->op_type == OP_PUSHMARK
6931 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6932 || expr->op_type == OP_PADRANGE);
6933 expr = OpSIBLING(expr);
6936 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6937 expr, &recompile, NULL);
6939 /* handle bare (possibly after overloading) regex: foo =~ $re */
6944 if (SvTYPE(re) == SVt_REGEXP) {
6948 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6949 "Precompiled pattern%s\n",
6950 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6956 exp = SvPV_nomg(pat, plen);
6958 if (!eng->op_comp) {
6959 if ((SvUTF8(pat) && IN_BYTES)
6960 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6962 /* make a temporary copy; either to convert to bytes,
6963 * or to avoid repeating get-magic / overloaded stringify */
6964 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6965 (IN_BYTES ? 0 : SvUTF8(pat)));
6967 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6970 /* ignore the utf8ness if the pattern is 0 length */
6971 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6973 RExC_uni_semantics = 0;
6974 RExC_seen_unfolded_sharp_s = 0;
6975 RExC_contains_locale = 0;
6976 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
6977 RExC_study_started = 0;
6978 pRExC_state->runtime_code_qr = NULL;
6979 RExC_frame_head= NULL;
6980 RExC_frame_last= NULL;
6981 RExC_frame_count= 0;
6984 RExC_mysv1= sv_newmortal();
6985 RExC_mysv2= sv_newmortal();
6988 SV *dsv= sv_newmortal();
6989 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6990 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
6991 PL_colors[4],PL_colors[5],s);
6995 /* we jump here if we have to recompile, e.g., from upgrading the pattern
6998 if ((pm_flags & PMf_USE_RE_EVAL)
6999 /* this second condition covers the non-regex literal case,
7000 * i.e. $foo =~ '(?{})'. */
7001 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
7003 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
7005 /* return old regex if pattern hasn't changed */
7006 /* XXX: note in the below we have to check the flags as well as the
7009 * Things get a touch tricky as we have to compare the utf8 flag
7010 * independently from the compile flags. */
7014 && !!RX_UTF8(old_re) == !!RExC_utf8
7015 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7016 && RX_PRECOMP(old_re)
7017 && RX_PRELEN(old_re) == plen
7018 && memEQ(RX_PRECOMP(old_re), exp, plen)
7019 && !runtime_code /* with runtime code, always recompile */ )
7024 rx_flags = orig_rx_flags;
7026 if ( initial_charset == REGEX_DEPENDS_CHARSET
7027 && (RExC_utf8 ||RExC_uni_semantics))
7030 /* Set to use unicode semantics if the pattern is in utf8 and has the
7031 * 'depends' charset specified, as it means unicode when utf8 */
7032 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7036 RExC_precomp_adj = 0;
7037 RExC_flags = rx_flags;
7038 RExC_pm_flags = pm_flags;
7041 assert(TAINTING_get || !TAINT_get);
7043 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7045 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7046 /* whoops, we have a non-utf8 pattern, whilst run-time code
7047 * got compiled as utf8. Try again with a utf8 pattern */
7048 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7049 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7050 goto redo_first_pass;
7053 assert(!pRExC_state->runtime_code_qr);
7059 RExC_in_lookbehind = 0;
7060 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7063 RExC_recode_x_to_native = 0;
7065 RExC_in_multi_char_class = 0;
7067 /* First pass: determine size, legality. */
7069 RExC_start = RExC_adjusted_start = exp;
7070 RExC_end = exp + plen;
7071 RExC_precomp_end = RExC_end;
7076 RExC_emit = (regnode *) &RExC_emit_dummy;
7077 RExC_whilem_seen = 0;
7078 RExC_open_parens = NULL;
7079 RExC_close_parens = NULL;
7081 RExC_paren_names = NULL;
7083 RExC_paren_name_list = NULL;
7085 RExC_recurse = NULL;
7086 RExC_study_chunk_recursed = NULL;
7087 RExC_study_chunk_recursed_bytes= 0;
7088 RExC_recurse_count = 0;
7089 pRExC_state->code_index = 0;
7091 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7092 * code makes sure the final byte is an uncounted NUL. But should this
7093 * ever not be the case, lots of things could read beyond the end of the
7094 * buffer: loops like
7095 * while(isFOO(*RExC_parse)) RExC_parse++;
7096 * strchr(RExC_parse, "foo");
7097 * etc. So it is worth noting. */
7098 assert(*RExC_end == '\0');
7101 Perl_re_printf( aTHX_ "Starting first pass (sizing)\n");
7103 RExC_lastparse=NULL;
7106 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7107 /* It's possible to write a regexp in ascii that represents Unicode
7108 codepoints outside of the byte range, such as via \x{100}. If we
7109 detect such a sequence we have to convert the entire pattern to utf8
7110 and then recompile, as our sizing calculation will have been based
7111 on 1 byte == 1 character, but we will need to use utf8 to encode
7112 at least some part of the pattern, and therefore must convert the whole
7115 if (flags & RESTART_PASS1) {
7116 if (flags & NEED_UTF8) {
7117 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7118 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7121 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7122 "Need to redo pass 1\n"));
7125 goto redo_first_pass;
7127 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#" UVxf, (UV) flags);
7131 Perl_re_printf( aTHX_
7132 "Required size %" IVdf " nodes\n"
7133 "Starting second pass (creation)\n",
7136 RExC_lastparse=NULL;
7139 /* The first pass could have found things that force Unicode semantics */
7140 if ((RExC_utf8 || RExC_uni_semantics)
7141 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
7143 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7146 /* Small enough for pointer-storage convention?
7147 If extralen==0, this means that we will not need long jumps. */
7148 if (RExC_size >= 0x10000L && RExC_extralen)
7149 RExC_size += RExC_extralen;
7152 if (RExC_whilem_seen > 15)
7153 RExC_whilem_seen = 15;
7155 /* Allocate space and zero-initialize. Note, the two step process
7156 of zeroing when in debug mode, thus anything assigned has to
7157 happen after that */
7158 rx = (REGEXP*) newSV_type(SVt_REGEXP);
7160 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7161 char, regexp_internal);
7162 if ( r == NULL || ri == NULL )
7163 FAIL("Regexp out of space");
7165 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
7166 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7169 /* bulk initialize base fields with 0. */
7170 Zero(ri, sizeof(regexp_internal), char);
7173 /* non-zero initialization begins here */
7176 r->extflags = rx_flags;
7177 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7179 if (pm_flags & PMf_IS_QR) {
7180 ri->code_blocks = pRExC_state->code_blocks;
7181 if (ri->code_blocks)
7182 ri->code_blocks->refcnt++;
7186 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7187 bool has_charset = (get_regex_charset(r->extflags)
7188 != REGEX_DEPENDS_CHARSET);
7190 /* The caret is output if there are any defaults: if not all the STD
7191 * flags are set, or if no character set specifier is needed */
7193 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7195 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7196 == REG_RUN_ON_COMMENT_SEEN);
7197 U8 reganch = (U8)((r->extflags & RXf_PMf_STD_PMMOD)
7198 >> RXf_PMf_STD_PMMOD_SHIFT);
7199 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7202 /* We output all the necessary flags; we never output a minus, as all
7203 * those are defaults, so are
7204 * covered by the caret */
7205 const STRLEN wraplen = plen + has_p + has_runon
7206 + has_default /* If needs a caret */
7207 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7209 /* If needs a character set specifier */
7210 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7211 + (sizeof("(?:)") - 1);
7213 /* make sure PL_bitcount bounds not exceeded */
7214 assert(sizeof(STD_PAT_MODS) <= 8);
7216 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
7217 r->xpv_len_u.xpvlenu_pv = p;
7219 SvFLAGS(rx) |= SVf_UTF8;
7222 /* If a default, cover it using the caret */
7224 *p++= DEFAULT_PAT_MOD;
7228 const char* const name = get_regex_charset_name(r->extflags, &len);
7229 Copy(name, p, len, char);
7233 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7236 while((ch = *fptr++)) {
7244 Copy(RExC_precomp, p, plen, char);
7245 assert ((RX_WRAPPED(rx) - p) < 16);
7246 r->pre_prefix = p - RX_WRAPPED(rx);
7252 SvCUR_set(rx, p - RX_WRAPPED(rx));
7256 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
7258 /* Useful during FAIL. */
7259 #ifdef RE_TRACK_PATTERN_OFFSETS
7260 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
7261 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7262 "%s %" UVuf " bytes for offset annotations.\n",
7263 ri->u.offsets ? "Got" : "Couldn't get",
7264 (UV)((2*RExC_size+1) * sizeof(U32))));
7266 SetProgLen(ri,RExC_size);
7271 /* Second pass: emit code. */
7272 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7273 RExC_pm_flags = pm_flags;
7275 RExC_end = exp + plen;
7277 RExC_emit_start = ri->program;
7278 RExC_emit = ri->program;
7279 RExC_emit_bound = ri->program + RExC_size + 1;
7280 pRExC_state->code_index = 0;
7282 *((char*) RExC_emit++) = (char) REG_MAGIC;
7283 /* setup various meta data about recursion, this all requires
7284 * RExC_npar to be correctly set, and a bit later on we clear it */
7285 if (RExC_seen & REG_RECURSE_SEEN) {
7286 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
7287 "%*s%*s Setting up open/close parens\n",
7288 22, "| |", (int)(0 * 2 + 1), ""));
7290 /* setup RExC_open_parens, which holds the address of each
7291 * OPEN tag, and to make things simpler for the 0 index
7292 * the start of the program - this is used later for offsets */
7293 Newxz(RExC_open_parens, RExC_npar,regnode *);
7294 SAVEFREEPV(RExC_open_parens);
7295 RExC_open_parens[0] = RExC_emit;
7297 /* setup RExC_close_parens, which holds the address of each
7298 * CLOSE tag, and to make things simpler for the 0 index
7299 * the end of the program - this is used later for offsets */
7300 Newxz(RExC_close_parens, RExC_npar,regnode *);
7301 SAVEFREEPV(RExC_close_parens);
7302 /* we dont know where end op starts yet, so we dont
7303 * need to set RExC_close_parens[0] like we do RExC_open_parens[0] above */
7305 /* Note, RExC_npar is 1 + the number of parens in a pattern.
7306 * So its 1 if there are no parens. */
7307 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
7308 ((RExC_npar & 0x07) != 0);
7309 Newx(RExC_study_chunk_recursed,
7310 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7311 SAVEFREEPV(RExC_study_chunk_recursed);
7314 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7316 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#" UVxf, (UV) flags);
7319 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7322 /* XXXX To minimize changes to RE engine we always allocate
7323 3-units-long substrs field. */
7324 Newx(r->substrs, 1, struct reg_substr_data);
7325 if (RExC_recurse_count) {
7326 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
7327 SAVEFREEPV(RExC_recurse);
7331 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7333 RExC_study_chunk_recursed_count= 0;
7335 Zero(r->substrs, 1, struct reg_substr_data);
7336 if (RExC_study_chunk_recursed) {
7337 Zero(RExC_study_chunk_recursed,
7338 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7342 #ifdef TRIE_STUDY_OPT
7344 StructCopy(&zero_scan_data, &data, scan_data_t);
7345 copyRExC_state = RExC_state;
7348 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
7350 RExC_state = copyRExC_state;
7351 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
7352 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
7354 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
7355 StructCopy(&zero_scan_data, &data, scan_data_t);
7358 StructCopy(&zero_scan_data, &data, scan_data_t);
7361 /* Dig out information for optimizations. */
7362 r->extflags = RExC_flags; /* was pm_op */
7363 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
7366 SvUTF8_on(rx); /* Unicode in it? */
7367 ri->regstclass = NULL;
7368 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
7369 r->intflags |= PREGf_NAUGHTY;
7370 scan = ri->program + 1; /* First BRANCH. */
7372 /* testing for BRANCH here tells us whether there is "must appear"
7373 data in the pattern. If there is then we can use it for optimisations */
7374 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7377 STRLEN longest_float_length, longest_fixed_length;
7378 regnode_ssc ch_class; /* pointed to by data */
7380 SSize_t last_close = 0; /* pointed to by data */
7381 regnode *first= scan;
7382 regnode *first_next= regnext(first);
7384 * Skip introductions and multiplicators >= 1
7385 * so that we can extract the 'meat' of the pattern that must
7386 * match in the large if() sequence following.
7387 * NOTE that EXACT is NOT covered here, as it is normally
7388 * picked up by the optimiser separately.
7390 * This is unfortunate as the optimiser isnt handling lookahead
7391 * properly currently.
7394 while ((OP(first) == OPEN && (sawopen = 1)) ||
7395 /* An OR of *one* alternative - should not happen now. */
7396 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7397 /* for now we can't handle lookbehind IFMATCH*/
7398 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7399 (OP(first) == PLUS) ||
7400 (OP(first) == MINMOD) ||
7401 /* An {n,m} with n>0 */
7402 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7403 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7406 * the only op that could be a regnode is PLUS, all the rest
7407 * will be regnode_1 or regnode_2.
7409 * (yves doesn't think this is true)
7411 if (OP(first) == PLUS)
7414 if (OP(first) == MINMOD)
7416 first += regarglen[OP(first)];
7418 first = NEXTOPER(first);
7419 first_next= regnext(first);
7422 /* Starting-point info. */
7424 DEBUG_PEEP("first:",first,0);
7425 /* Ignore EXACT as we deal with it later. */
7426 if (PL_regkind[OP(first)] == EXACT) {
7427 if (OP(first) == EXACT || OP(first) == EXACTL)
7428 NOOP; /* Empty, get anchored substr later. */
7430 ri->regstclass = first;
7433 else if (PL_regkind[OP(first)] == TRIE &&
7434 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
7436 /* this can happen only on restudy */
7437 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
7440 else if (REGNODE_SIMPLE(OP(first)))
7441 ri->regstclass = first;
7442 else if (PL_regkind[OP(first)] == BOUND ||
7443 PL_regkind[OP(first)] == NBOUND)
7444 ri->regstclass = first;
7445 else if (PL_regkind[OP(first)] == BOL) {
7446 r->intflags |= (OP(first) == MBOL
7449 first = NEXTOPER(first);
7452 else if (OP(first) == GPOS) {
7453 r->intflags |= PREGf_ANCH_GPOS;
7454 first = NEXTOPER(first);
7457 else if ((!sawopen || !RExC_sawback) &&
7459 (OP(first) == STAR &&
7460 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7461 !(r->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
7463 /* turn .* into ^.* with an implied $*=1 */
7465 (OP(NEXTOPER(first)) == REG_ANY)
7468 r->intflags |= (type | PREGf_IMPLICIT);
7469 first = NEXTOPER(first);
7472 if (sawplus && !sawminmod && !sawlookahead
7473 && (!sawopen || !RExC_sawback)
7474 && !pRExC_state->code_blocks) /* May examine pos and $& */
7475 /* x+ must match at the 1st pos of run of x's */
7476 r->intflags |= PREGf_SKIP;
7478 /* Scan is after the zeroth branch, first is atomic matcher. */
7479 #ifdef TRIE_STUDY_OPT
7482 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7483 (IV)(first - scan + 1))
7487 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7488 (IV)(first - scan + 1))
7494 * If there's something expensive in the r.e., find the
7495 * longest literal string that must appear and make it the
7496 * regmust. Resolve ties in favor of later strings, since
7497 * the regstart check works with the beginning of the r.e.
7498 * and avoiding duplication strengthens checking. Not a
7499 * strong reason, but sufficient in the absence of others.
7500 * [Now we resolve ties in favor of the earlier string if
7501 * it happens that c_offset_min has been invalidated, since the
7502 * earlier string may buy us something the later one won't.]
7505 data.longest_fixed = newSVpvs("");
7506 data.longest_float = newSVpvs("");
7507 data.last_found = newSVpvs("");
7508 data.longest = &(data.longest_fixed);
7509 ENTER_with_name("study_chunk");
7510 SAVEFREESV(data.longest_fixed);
7511 SAVEFREESV(data.longest_float);
7512 SAVEFREESV(data.last_found);
7514 if (!ri->regstclass) {
7515 ssc_init(pRExC_state, &ch_class);
7516 data.start_class = &ch_class;
7517 stclass_flag = SCF_DO_STCLASS_AND;
7518 } else /* XXXX Check for BOUND? */
7520 data.last_closep = &last_close;
7523 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
7524 scan + RExC_size, /* Up to end */
7526 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
7527 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
7531 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
7534 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
7535 && data.last_start_min == 0 && data.last_end > 0
7536 && !RExC_seen_zerolen
7537 && !(RExC_seen & REG_VERBARG_SEEN)
7538 && !(RExC_seen & REG_GPOS_SEEN)
7540 r->extflags |= RXf_CHECK_ALL;
7542 scan_commit(pRExC_state, &data,&minlen,0);
7544 longest_float_length = CHR_SVLEN(data.longest_float);
7546 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
7547 && data.offset_fixed == data.offset_float_min
7548 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
7549 && S_setup_longest (aTHX_ pRExC_state,
7553 &(r->float_end_shift),
7554 data.lookbehind_float,
7555 data.offset_float_min,
7557 longest_float_length,
7558 cBOOL(data.flags & SF_FL_BEFORE_EOL),
7559 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
7561 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
7562 r->float_max_offset = data.offset_float_max;
7563 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
7564 r->float_max_offset -= data.lookbehind_float;
7565 SvREFCNT_inc_simple_void_NN(data.longest_float);
7568 r->float_substr = r->float_utf8 = NULL;
7569 longest_float_length = 0;
7572 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
7574 if (S_setup_longest (aTHX_ pRExC_state,
7576 &(r->anchored_utf8),
7577 &(r->anchored_substr),
7578 &(r->anchored_end_shift),
7579 data.lookbehind_fixed,
7582 longest_fixed_length,
7583 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
7584 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
7586 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
7587 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
7590 r->anchored_substr = r->anchored_utf8 = NULL;
7591 longest_fixed_length = 0;
7593 LEAVE_with_name("study_chunk");
7596 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
7597 ri->regstclass = NULL;
7599 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
7601 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7602 && is_ssc_worth_it(pRExC_state, data.start_class))
7604 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7606 ssc_finalize(pRExC_state, data.start_class);
7608 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7609 StructCopy(data.start_class,
7610 (regnode_ssc*)RExC_rxi->data->data[n],
7612 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7613 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7614 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7615 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7616 Perl_re_printf( aTHX_
7617 "synthetic stclass \"%s\".\n",
7618 SvPVX_const(sv));});
7619 data.start_class = NULL;
7622 /* A temporary algorithm prefers floated substr to fixed one to dig
7624 if (longest_fixed_length > longest_float_length) {
7625 r->substrs->check_ix = 0;
7626 r->check_end_shift = r->anchored_end_shift;
7627 r->check_substr = r->anchored_substr;
7628 r->check_utf8 = r->anchored_utf8;
7629 r->check_offset_min = r->check_offset_max = r->anchored_offset;
7630 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
7631 r->intflags |= PREGf_NOSCAN;
7634 r->substrs->check_ix = 1;
7635 r->check_end_shift = r->float_end_shift;
7636 r->check_substr = r->float_substr;
7637 r->check_utf8 = r->float_utf8;
7638 r->check_offset_min = r->float_min_offset;
7639 r->check_offset_max = r->float_max_offset;
7641 if ((r->check_substr || r->check_utf8) ) {
7642 r->extflags |= RXf_USE_INTUIT;
7643 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7644 r->extflags |= RXf_INTUIT_TAIL;
7646 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
7648 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7649 if ( (STRLEN)minlen < longest_float_length )
7650 minlen= longest_float_length;
7651 if ( (STRLEN)minlen < longest_fixed_length )
7652 minlen= longest_fixed_length;
7656 /* Several toplevels. Best we can is to set minlen. */
7658 regnode_ssc ch_class;
7659 SSize_t last_close = 0;
7661 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
7663 scan = ri->program + 1;
7664 ssc_init(pRExC_state, &ch_class);
7665 data.start_class = &ch_class;
7666 data.last_closep = &last_close;
7669 minlen = study_chunk(pRExC_state,
7670 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7671 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7672 ? SCF_TRIE_DOING_RESTUDY
7676 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7678 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7679 = r->float_substr = r->float_utf8 = NULL;
7681 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7682 && is_ssc_worth_it(pRExC_state, data.start_class))
7684 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7686 ssc_finalize(pRExC_state, data.start_class);
7688 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7689 StructCopy(data.start_class,
7690 (regnode_ssc*)RExC_rxi->data->data[n],
7692 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7693 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7694 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7695 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7696 Perl_re_printf( aTHX_
7697 "synthetic stclass \"%s\".\n",
7698 SvPVX_const(sv));});
7699 data.start_class = NULL;
7703 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7704 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7705 r->maxlen = REG_INFTY;
7708 r->maxlen = RExC_maxlen;
7711 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7712 the "real" pattern. */
7714 Perl_re_printf( aTHX_ "minlen: %" IVdf " r->minlen:%" IVdf " maxlen:%" IVdf "\n",
7715 (IV)minlen, (IV)r->minlen, (IV)RExC_maxlen);
7717 r->minlenret = minlen;
7718 if (r->minlen < minlen)
7721 if (RExC_seen & REG_RECURSE_SEEN ) {
7722 r->intflags |= PREGf_RECURSE_SEEN;
7723 Newxz(r->recurse_locinput, r->nparens + 1, char *);
7725 if (RExC_seen & REG_GPOS_SEEN)
7726 r->intflags |= PREGf_GPOS_SEEN;
7727 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7728 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7730 if (pRExC_state->code_blocks)
7731 r->extflags |= RXf_EVAL_SEEN;
7732 if (RExC_seen & REG_VERBARG_SEEN)
7734 r->intflags |= PREGf_VERBARG_SEEN;
7735 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7737 if (RExC_seen & REG_CUTGROUP_SEEN)
7738 r->intflags |= PREGf_CUTGROUP_SEEN;
7739 if (pm_flags & PMf_USE_RE_EVAL)
7740 r->intflags |= PREGf_USE_RE_EVAL;
7741 if (RExC_paren_names)
7742 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7744 RXp_PAREN_NAMES(r) = NULL;
7746 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7747 * so it can be used in pp.c */
7748 if (r->intflags & PREGf_ANCH)
7749 r->extflags |= RXf_IS_ANCHORED;
7753 /* this is used to identify "special" patterns that might result
7754 * in Perl NOT calling the regex engine and instead doing the match "itself",
7755 * particularly special cases in split//. By having the regex compiler
7756 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7757 * we avoid weird issues with equivalent patterns resulting in different behavior,
7758 * AND we allow non Perl engines to get the same optimizations by the setting the
7759 * flags appropriately - Yves */
7760 regnode *first = ri->program + 1;
7762 regnode *next = regnext(first);
7765 if (PL_regkind[fop] == NOTHING && nop == END)
7766 r->extflags |= RXf_NULL;
7767 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
7768 /* when fop is SBOL first->flags will be true only when it was
7769 * produced by parsing /\A/, and not when parsing /^/. This is
7770 * very important for the split code as there we want to
7771 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
7772 * See rt #122761 for more details. -- Yves */
7773 r->extflags |= RXf_START_ONLY;
7774 else if (fop == PLUS
7775 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7777 r->extflags |= RXf_WHITE;
7778 else if ( r->extflags & RXf_SPLIT
7779 && (fop == EXACT || fop == EXACTL)
7780 && STR_LEN(first) == 1
7781 && *(STRING(first)) == ' '
7783 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7787 if (RExC_contains_locale) {
7788 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7792 if (RExC_paren_names) {
7793 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7794 ri->data->data[ri->name_list_idx]
7795 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7798 ri->name_list_idx = 0;
7800 while ( RExC_recurse_count > 0 ) {
7801 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
7803 * This data structure is set up in study_chunk() and is used
7804 * to calculate the distance between a GOSUB regopcode and
7805 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
7808 * If for some reason someone writes code that optimises
7809 * away a GOSUB opcode then the assert should be changed to
7810 * an if(scan) to guard the ARG2L_SET() - Yves
7813 assert(scan && OP(scan) == GOSUB);
7814 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - scan );
7817 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7818 /* assume we don't need to swap parens around before we match */
7820 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
7821 (unsigned long)RExC_study_chunk_recursed_count);
7825 Perl_re_printf( aTHX_ "Final program:\n");
7828 #ifdef RE_TRACK_PATTERN_OFFSETS
7829 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7830 const STRLEN len = ri->u.offsets[0];
7832 GET_RE_DEBUG_FLAGS_DECL;
7833 Perl_re_printf( aTHX_
7834 "Offsets: [%" UVuf "]\n\t", (UV)ri->u.offsets[0]);
7835 for (i = 1; i <= len; i++) {
7836 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7837 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7838 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7840 Perl_re_printf( aTHX_ "\n");
7845 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7846 * by setting the regexp SV to readonly-only instead. If the
7847 * pattern's been recompiled, the USEDness should remain. */
7848 if (old_re && SvREADONLY(old_re))
7856 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7859 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7861 PERL_UNUSED_ARG(value);
7863 if (flags & RXapif_FETCH) {
7864 return reg_named_buff_fetch(rx, key, flags);
7865 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7866 Perl_croak_no_modify();
7868 } else if (flags & RXapif_EXISTS) {
7869 return reg_named_buff_exists(rx, key, flags)
7872 } else if (flags & RXapif_REGNAMES) {
7873 return reg_named_buff_all(rx, flags);
7874 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7875 return reg_named_buff_scalar(rx, flags);
7877 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7883 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7886 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7887 PERL_UNUSED_ARG(lastkey);
7889 if (flags & RXapif_FIRSTKEY)
7890 return reg_named_buff_firstkey(rx, flags);
7891 else if (flags & RXapif_NEXTKEY)
7892 return reg_named_buff_nextkey(rx, flags);
7894 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7901 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7904 AV *retarray = NULL;
7906 struct regexp *const rx = ReANY(r);
7908 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7910 if (flags & RXapif_ALL)
7913 if (rx && RXp_PAREN_NAMES(rx)) {
7914 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7917 SV* sv_dat=HeVAL(he_str);
7918 I32 *nums=(I32*)SvPVX(sv_dat);
7919 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7920 if ((I32)(rx->nparens) >= nums[i]
7921 && rx->offs[nums[i]].start != -1
7922 && rx->offs[nums[i]].end != -1)
7925 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7930 ret = newSVsv(&PL_sv_undef);
7933 av_push(retarray, ret);
7936 return newRV_noinc(MUTABLE_SV(retarray));
7943 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7946 struct regexp *const rx = ReANY(r);
7948 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7950 if (rx && RXp_PAREN_NAMES(rx)) {
7951 if (flags & RXapif_ALL) {
7952 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7954 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7956 SvREFCNT_dec_NN(sv);
7968 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7970 struct regexp *const rx = ReANY(r);
7972 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7974 if ( rx && RXp_PAREN_NAMES(rx) ) {
7975 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7977 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7984 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7986 struct regexp *const rx = ReANY(r);
7987 GET_RE_DEBUG_FLAGS_DECL;
7989 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7991 if (rx && RXp_PAREN_NAMES(rx)) {
7992 HV *hv = RXp_PAREN_NAMES(rx);
7994 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7997 SV* sv_dat = HeVAL(temphe);
7998 I32 *nums = (I32*)SvPVX(sv_dat);
7999 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8000 if ((I32)(rx->lastparen) >= nums[i] &&
8001 rx->offs[nums[i]].start != -1 &&
8002 rx->offs[nums[i]].end != -1)
8008 if (parno || flags & RXapif_ALL) {
8009 return newSVhek(HeKEY_hek(temphe));
8017 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8022 struct regexp *const rx = ReANY(r);
8024 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8026 if (rx && RXp_PAREN_NAMES(rx)) {
8027 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8028 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8029 } else if (flags & RXapif_ONE) {
8030 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8031 av = MUTABLE_AV(SvRV(ret));
8032 length = av_tindex(av);
8033 SvREFCNT_dec_NN(ret);
8034 return newSViv(length + 1);
8036 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8041 return &PL_sv_undef;
8045 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8047 struct regexp *const rx = ReANY(r);
8050 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8052 if (rx && RXp_PAREN_NAMES(rx)) {
8053 HV *hv= RXp_PAREN_NAMES(rx);
8055 (void)hv_iterinit(hv);
8056 while ( (temphe = hv_iternext_flags(hv,0)) ) {
8059 SV* sv_dat = HeVAL(temphe);
8060 I32 *nums = (I32*)SvPVX(sv_dat);
8061 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8062 if ((I32)(rx->lastparen) >= nums[i] &&
8063 rx->offs[nums[i]].start != -1 &&
8064 rx->offs[nums[i]].end != -1)
8070 if (parno || flags & RXapif_ALL) {
8071 av_push(av, newSVhek(HeKEY_hek(temphe)));
8076 return newRV_noinc(MUTABLE_SV(av));
8080 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8083 struct regexp *const rx = ReANY(r);
8089 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8091 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8092 || n == RX_BUFF_IDX_CARET_FULLMATCH
8093 || n == RX_BUFF_IDX_CARET_POSTMATCH
8096 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8098 /* on something like
8101 * the KEEPCOPY is set on the PMOP rather than the regex */
8102 if (PL_curpm && r == PM_GETRE(PL_curpm))
8103 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8112 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8113 /* no need to distinguish between them any more */
8114 n = RX_BUFF_IDX_FULLMATCH;
8116 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8117 && rx->offs[0].start != -1)
8119 /* $`, ${^PREMATCH} */
8120 i = rx->offs[0].start;
8124 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8125 && rx->offs[0].end != -1)
8127 /* $', ${^POSTMATCH} */
8128 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8129 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8132 if ( 0 <= n && n <= (I32)rx->nparens &&
8133 (s1 = rx->offs[n].start) != -1 &&
8134 (t1 = rx->offs[n].end) != -1)
8136 /* $&, ${^MATCH}, $1 ... */
8138 s = rx->subbeg + s1 - rx->suboffset;
8143 assert(s >= rx->subbeg);
8144 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8146 #ifdef NO_TAINT_SUPPORT
8147 sv_setpvn(sv, s, i);
8149 const int oldtainted = TAINT_get;
8151 sv_setpvn(sv, s, i);
8152 TAINT_set(oldtainted);
8154 if (RXp_MATCH_UTF8(rx))
8159 if (RXp_MATCH_TAINTED(rx)) {
8160 if (SvTYPE(sv) >= SVt_PVMG) {
8161 MAGIC* const mg = SvMAGIC(sv);
8164 SvMAGIC_set(sv, mg->mg_moremagic);
8166 if ((mgt = SvMAGIC(sv))) {
8167 mg->mg_moremagic = mgt;
8168 SvMAGIC_set(sv, mg);
8185 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8186 SV const * const value)
8188 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8190 PERL_UNUSED_ARG(rx);
8191 PERL_UNUSED_ARG(paren);
8192 PERL_UNUSED_ARG(value);
8195 Perl_croak_no_modify();
8199 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8202 struct regexp *const rx = ReANY(r);
8206 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8208 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8209 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8210 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8213 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8215 /* on something like
8218 * the KEEPCOPY is set on the PMOP rather than the regex */
8219 if (PL_curpm && r == PM_GETRE(PL_curpm))
8220 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8226 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8228 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8229 case RX_BUFF_IDX_PREMATCH: /* $` */
8230 if (rx->offs[0].start != -1) {
8231 i = rx->offs[0].start;
8240 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8241 case RX_BUFF_IDX_POSTMATCH: /* $' */
8242 if (rx->offs[0].end != -1) {
8243 i = rx->sublen - rx->offs[0].end;
8245 s1 = rx->offs[0].end;
8252 default: /* $& / ${^MATCH}, $1, $2, ... */
8253 if (paren <= (I32)rx->nparens &&
8254 (s1 = rx->offs[paren].start) != -1 &&
8255 (t1 = rx->offs[paren].end) != -1)
8261 if (ckWARN(WARN_UNINITIALIZED))
8262 report_uninit((const SV *)sv);
8267 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8268 const char * const s = rx->subbeg - rx->suboffset + s1;
8273 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8280 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8282 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8283 PERL_UNUSED_ARG(rx);
8287 return newSVpvs("Regexp");
8290 /* Scans the name of a named buffer from the pattern.
8291 * If flags is REG_RSN_RETURN_NULL returns null.
8292 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8293 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8294 * to the parsed name as looked up in the RExC_paren_names hash.
8295 * If there is an error throws a vFAIL().. type exception.
8298 #define REG_RSN_RETURN_NULL 0
8299 #define REG_RSN_RETURN_NAME 1
8300 #define REG_RSN_RETURN_DATA 2
8303 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8305 char *name_start = RExC_parse;
8307 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8309 assert (RExC_parse <= RExC_end);
8310 if (RExC_parse == RExC_end) NOOP;
8311 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8312 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8313 * using do...while */
8316 RExC_parse += UTF8SKIP(RExC_parse);
8317 } while ( RExC_parse < RExC_end
8318 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8322 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8324 RExC_parse++; /* so the <- from the vFAIL is after the offending
8326 vFAIL("Group name must start with a non-digit word character");
8330 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8331 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8332 if ( flags == REG_RSN_RETURN_NAME)
8334 else if (flags==REG_RSN_RETURN_DATA) {
8337 if ( ! sv_name ) /* should not happen*/
8338 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8339 if (RExC_paren_names)
8340 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8342 sv_dat = HeVAL(he_str);
8344 vFAIL("Reference to nonexistent named group");
8348 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8349 (unsigned long) flags);
8351 NOT_REACHED; /* NOTREACHED */
8356 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8358 if (RExC_lastparse!=RExC_parse) { \
8359 Perl_re_printf( aTHX_ "%s", \
8360 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8361 RExC_end - RExC_parse, 16, \
8363 PERL_PV_ESCAPE_UNI_DETECT | \
8364 PERL_PV_PRETTY_ELLIPSES | \
8365 PERL_PV_PRETTY_LTGT | \
8366 PERL_PV_ESCAPE_RE | \
8367 PERL_PV_PRETTY_EXACTSIZE \
8371 Perl_re_printf( aTHX_ "%16s",""); \
8374 num = RExC_size + 1; \
8376 num=REG_NODE_NUM(RExC_emit); \
8377 if (RExC_lastnum!=num) \
8378 Perl_re_printf( aTHX_ "|%4d",num); \
8380 Perl_re_printf( aTHX_ "|%4s",""); \
8381 Perl_re_printf( aTHX_ "|%*s%-4s", \
8382 (int)((depth*2)), "", \
8386 RExC_lastparse=RExC_parse; \
8391 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8392 DEBUG_PARSE_MSG((funcname)); \
8393 Perl_re_printf( aTHX_ "%4s","\n"); \
8395 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8396 DEBUG_PARSE_MSG((funcname)); \
8397 Perl_re_printf( aTHX_ fmt "\n",args); \
8400 /* This section of code defines the inversion list object and its methods. The
8401 * interfaces are highly subject to change, so as much as possible is static to
8402 * this file. An inversion list is here implemented as a malloc'd C UV array
8403 * as an SVt_INVLIST scalar.
8405 * An inversion list for Unicode is an array of code points, sorted by ordinal
8406 * number. Each element gives the code point that begins a range that extends
8407 * up-to but not including the code point given by the next element. The final
8408 * element gives the first code point of a range that extends to the platform's
8409 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8410 * ...) give ranges whose code points are all in the inversion list. We say
8411 * that those ranges are in the set. The odd-numbered elements give ranges
8412 * whose code points are not in the inversion list, and hence not in the set.
8413 * Thus, element [0] is the first code point in the list. Element [1]
8414 * is the first code point beyond that not in the list; and element [2] is the
8415 * first code point beyond that that is in the list. In other words, the first
8416 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8417 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8418 * all code points in that range are not in the inversion list. The third
8419 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8420 * list, and so forth. Thus every element whose index is divisible by two
8421 * gives the beginning of a range that is in the list, and every element whose
8422 * index is not divisible by two gives the beginning of a range not in the
8423 * list. If the final element's index is divisible by two, the inversion list
8424 * extends to the platform's infinity; otherwise the highest code point in the
8425 * inversion list is the contents of that element minus 1.
8427 * A range that contains just a single code point N will look like
8429 * invlist[i+1] == N+1
8431 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8432 * impossible to represent, so element [i+1] is omitted. The single element
8434 * invlist[0] == UV_MAX
8435 * contains just UV_MAX, but is interpreted as matching to infinity.
8437 * Taking the complement (inverting) an inversion list is quite simple, if the
8438 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8439 * This implementation reserves an element at the beginning of each inversion
8440 * list to always contain 0; there is an additional flag in the header which
8441 * indicates if the list begins at the 0, or is offset to begin at the next
8442 * element. This means that the inversion list can be inverted without any
8443 * copying; just flip the flag.
8445 * More about inversion lists can be found in "Unicode Demystified"
8446 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8448 * The inversion list data structure is currently implemented as an SV pointing
8449 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8450 * array of UV whose memory management is automatically handled by the existing
8451 * facilities for SV's.
8453 * Some of the methods should always be private to the implementation, and some
8454 * should eventually be made public */
8456 /* The header definitions are in F<invlist_inline.h> */
8458 #ifndef PERL_IN_XSUB_RE
8460 PERL_STATIC_INLINE UV*
8461 S__invlist_array_init(SV* const invlist, const bool will_have_0)
8463 /* Returns a pointer to the first element in the inversion list's array.
8464 * This is called upon initialization of an inversion list. Where the
8465 * array begins depends on whether the list has the code point U+0000 in it
8466 * or not. The other parameter tells it whether the code that follows this
8467 * call is about to put a 0 in the inversion list or not. The first
8468 * element is either the element reserved for 0, if TRUE, or the element
8469 * after it, if FALSE */
8471 bool* offset = get_invlist_offset_addr(invlist);
8472 UV* zero_addr = (UV *) SvPVX(invlist);
8474 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8477 assert(! _invlist_len(invlist));
8481 /* 1^1 = 0; 1^0 = 1 */
8482 *offset = 1 ^ will_have_0;
8483 return zero_addr + *offset;
8488 PERL_STATIC_INLINE void
8489 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8491 /* Sets the current number of elements stored in the inversion list.
8492 * Updates SvCUR correspondingly */
8493 PERL_UNUSED_CONTEXT;
8494 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8496 assert(SvTYPE(invlist) == SVt_INVLIST);
8501 : TO_INTERNAL_SIZE(len + offset));
8502 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8505 #ifndef PERL_IN_XSUB_RE
8508 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
8510 /* Replaces the inversion list in 'dest' with the one from 'src'. It
8511 * steals the list from 'src', so 'src' is made to have a NULL list. This
8512 * is similar to what SvSetMagicSV() would do, if it were implemented on
8513 * inversion lists, though this routine avoids a copy */
8515 const UV src_len = _invlist_len(src);
8516 const bool src_offset = *get_invlist_offset_addr(src);
8517 const STRLEN src_byte_len = SvLEN(src);
8518 char * array = SvPVX(src);
8520 const int oldtainted = TAINT_get;
8522 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
8524 assert(SvTYPE(src) == SVt_INVLIST);
8525 assert(SvTYPE(dest) == SVt_INVLIST);
8526 assert(! invlist_is_iterating(src));
8527 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
8529 /* Make sure it ends in the right place with a NUL, as our inversion list
8530 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
8532 array[src_byte_len - 1] = '\0';
8534 TAINT_NOT; /* Otherwise it breaks */
8535 sv_usepvn_flags(dest,
8539 /* This flag is documented to cause a copy to be avoided */
8540 SV_HAS_TRAILING_NUL);
8541 TAINT_set(oldtainted);
8546 /* Finish up copying over the other fields in an inversion list */
8547 *get_invlist_offset_addr(dest) = src_offset;
8548 invlist_set_len(dest, src_len, src_offset);
8549 *get_invlist_previous_index_addr(dest) = 0;
8550 invlist_iterfinish(dest);
8553 PERL_STATIC_INLINE IV*
8554 S_get_invlist_previous_index_addr(SV* invlist)
8556 /* Return the address of the IV that is reserved to hold the cached index
8558 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8560 assert(SvTYPE(invlist) == SVt_INVLIST);
8562 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8565 PERL_STATIC_INLINE IV
8566 S_invlist_previous_index(SV* const invlist)
8568 /* Returns cached index of previous search */
8570 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8572 return *get_invlist_previous_index_addr(invlist);
8575 PERL_STATIC_INLINE void
8576 S_invlist_set_previous_index(SV* const invlist, const IV index)
8578 /* Caches <index> for later retrieval */
8580 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8582 assert(index == 0 || index < (int) _invlist_len(invlist));
8584 *get_invlist_previous_index_addr(invlist) = index;
8587 PERL_STATIC_INLINE void
8588 S_invlist_trim(SV* invlist)
8590 /* Free the not currently-being-used space in an inversion list */
8592 /* But don't free up the space needed for the 0 UV that is always at the
8593 * beginning of the list, nor the trailing NUL */
8594 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
8596 PERL_ARGS_ASSERT_INVLIST_TRIM;
8598 assert(SvTYPE(invlist) == SVt_INVLIST);
8600 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
8603 PERL_STATIC_INLINE void
8604 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
8606 PERL_ARGS_ASSERT_INVLIST_CLEAR;
8608 assert(SvTYPE(invlist) == SVt_INVLIST);
8610 invlist_set_len(invlist, 0, 0);
8611 invlist_trim(invlist);
8614 #endif /* ifndef PERL_IN_XSUB_RE */
8616 PERL_STATIC_INLINE bool
8617 S_invlist_is_iterating(SV* const invlist)
8619 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8621 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8624 #ifndef PERL_IN_XSUB_RE
8626 PERL_STATIC_INLINE UV
8627 S_invlist_max(SV* const invlist)
8629 /* Returns the maximum number of elements storable in the inversion list's
8630 * array, without having to realloc() */
8632 PERL_ARGS_ASSERT_INVLIST_MAX;
8634 assert(SvTYPE(invlist) == SVt_INVLIST);
8636 /* Assumes worst case, in which the 0 element is not counted in the
8637 * inversion list, so subtracts 1 for that */
8638 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8639 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8640 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8643 Perl__new_invlist(pTHX_ IV initial_size)
8646 /* Return a pointer to a newly constructed inversion list, with enough
8647 * space to store 'initial_size' elements. If that number is negative, a
8648 * system default is used instead */
8652 if (initial_size < 0) {
8656 /* Allocate the initial space */
8657 new_list = newSV_type(SVt_INVLIST);
8659 /* First 1 is in case the zero element isn't in the list; second 1 is for
8661 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8662 invlist_set_len(new_list, 0, 0);
8664 /* Force iterinit() to be used to get iteration to work */
8665 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8667 *get_invlist_previous_index_addr(new_list) = 0;
8673 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8675 /* Return a pointer to a newly constructed inversion list, initialized to
8676 * point to <list>, which has to be in the exact correct inversion list
8677 * form, including internal fields. Thus this is a dangerous routine that
8678 * should not be used in the wrong hands. The passed in 'list' contains
8679 * several header fields at the beginning that are not part of the
8680 * inversion list body proper */
8682 const STRLEN length = (STRLEN) list[0];
8683 const UV version_id = list[1];
8684 const bool offset = cBOOL(list[2]);
8685 #define HEADER_LENGTH 3
8686 /* If any of the above changes in any way, you must change HEADER_LENGTH
8687 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8688 * perl -E 'say int(rand 2**31-1)'
8690 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8691 data structure type, so that one being
8692 passed in can be validated to be an
8693 inversion list of the correct vintage.
8696 SV* invlist = newSV_type(SVt_INVLIST);
8698 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8700 if (version_id != INVLIST_VERSION_ID) {
8701 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8704 /* The generated array passed in includes header elements that aren't part
8705 * of the list proper, so start it just after them */
8706 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8708 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8709 shouldn't touch it */
8711 *(get_invlist_offset_addr(invlist)) = offset;
8713 /* The 'length' passed to us is the physical number of elements in the
8714 * inversion list. But if there is an offset the logical number is one
8716 invlist_set_len(invlist, length - offset, offset);
8718 invlist_set_previous_index(invlist, 0);
8720 /* Initialize the iteration pointer. */
8721 invlist_iterfinish(invlist);
8723 SvREADONLY_on(invlist);
8729 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8731 /* Grow the maximum size of an inversion list */
8733 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8735 assert(SvTYPE(invlist) == SVt_INVLIST);
8737 /* Add one to account for the zero element at the beginning which may not
8738 * be counted by the calling parameters */
8739 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8743 S__append_range_to_invlist(pTHX_ SV* const invlist,
8744 const UV start, const UV end)
8746 /* Subject to change or removal. Append the range from 'start' to 'end' at
8747 * the end of the inversion list. The range must be above any existing
8751 UV max = invlist_max(invlist);
8752 UV len = _invlist_len(invlist);
8755 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8757 if (len == 0) { /* Empty lists must be initialized */
8758 offset = start != 0;
8759 array = _invlist_array_init(invlist, ! offset);
8762 /* Here, the existing list is non-empty. The current max entry in the
8763 * list is generally the first value not in the set, except when the
8764 * set extends to the end of permissible values, in which case it is
8765 * the first entry in that final set, and so this call is an attempt to
8766 * append out-of-order */
8768 UV final_element = len - 1;
8769 array = invlist_array(invlist);
8770 if ( array[final_element] > start
8771 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8773 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",
8774 array[final_element], start,
8775 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8778 /* Here, it is a legal append. If the new range begins 1 above the end
8779 * of the range below it, it is extending the range below it, so the
8780 * new first value not in the set is one greater than the newly
8781 * extended range. */
8782 offset = *get_invlist_offset_addr(invlist);
8783 if (array[final_element] == start) {
8784 if (end != UV_MAX) {
8785 array[final_element] = end + 1;
8788 /* But if the end is the maximum representable on the machine,
8789 * assume that infinity was actually what was meant. Just let
8790 * the range that this would extend to have no end */
8791 invlist_set_len(invlist, len - 1, offset);
8797 /* Here the new range doesn't extend any existing set. Add it */
8799 len += 2; /* Includes an element each for the start and end of range */
8801 /* If wll overflow the existing space, extend, which may cause the array to
8804 invlist_extend(invlist, len);
8806 /* Have to set len here to avoid assert failure in invlist_array() */
8807 invlist_set_len(invlist, len, offset);
8809 array = invlist_array(invlist);
8812 invlist_set_len(invlist, len, offset);
8815 /* The next item on the list starts the range, the one after that is
8816 * one past the new range. */
8817 array[len - 2] = start;
8818 if (end != UV_MAX) {
8819 array[len - 1] = end + 1;
8822 /* But if the end is the maximum representable on the machine, just let
8823 * the range have no end */
8824 invlist_set_len(invlist, len - 1, offset);
8829 Perl__invlist_search(SV* const invlist, const UV cp)
8831 /* Searches the inversion list for the entry that contains the input code
8832 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8833 * return value is the index into the list's array of the range that
8834 * contains <cp>, that is, 'i' such that
8835 * array[i] <= cp < array[i+1]
8840 IV high = _invlist_len(invlist);
8841 const IV highest_element = high - 1;
8844 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8846 /* If list is empty, return failure. */
8851 /* (We can't get the array unless we know the list is non-empty) */
8852 array = invlist_array(invlist);
8854 mid = invlist_previous_index(invlist);
8856 if (mid > highest_element) {
8857 mid = highest_element;
8860 /* <mid> contains the cache of the result of the previous call to this
8861 * function (0 the first time). See if this call is for the same result,
8862 * or if it is for mid-1. This is under the theory that calls to this
8863 * function will often be for related code points that are near each other.
8864 * And benchmarks show that caching gives better results. We also test
8865 * here if the code point is within the bounds of the list. These tests
8866 * replace others that would have had to be made anyway to make sure that
8867 * the array bounds were not exceeded, and these give us extra information
8868 * at the same time */
8869 if (cp >= array[mid]) {
8870 if (cp >= array[highest_element]) {
8871 return highest_element;
8874 /* Here, array[mid] <= cp < array[highest_element]. This means that
8875 * the final element is not the answer, so can exclude it; it also
8876 * means that <mid> is not the final element, so can refer to 'mid + 1'
8878 if (cp < array[mid + 1]) {
8884 else { /* cp < aray[mid] */
8885 if (cp < array[0]) { /* Fail if outside the array */
8889 if (cp >= array[mid - 1]) {
8894 /* Binary search. What we are looking for is <i> such that
8895 * array[i] <= cp < array[i+1]
8896 * The loop below converges on the i+1. Note that there may not be an
8897 * (i+1)th element in the array, and things work nonetheless */
8898 while (low < high) {
8899 mid = (low + high) / 2;
8900 assert(mid <= highest_element);
8901 if (array[mid] <= cp) { /* cp >= array[mid] */
8904 /* We could do this extra test to exit the loop early.
8905 if (cp < array[low]) {
8910 else { /* cp < array[mid] */
8917 invlist_set_previous_index(invlist, high);
8922 Perl__invlist_populate_swatch(SV* const invlist,
8923 const UV start, const UV end, U8* swatch)
8925 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8926 * but is used when the swash has an inversion list. This makes this much
8927 * faster, as it uses a binary search instead of a linear one. This is
8928 * intimately tied to that function, and perhaps should be in utf8.c,
8929 * except it is intimately tied to inversion lists as well. It assumes
8930 * that <swatch> is all 0's on input */
8933 const IV len = _invlist_len(invlist);
8937 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8939 if (len == 0) { /* Empty inversion list */
8943 array = invlist_array(invlist);
8945 /* Find which element it is */
8946 i = _invlist_search(invlist, start);
8948 /* We populate from <start> to <end> */
8949 while (current < end) {
8952 /* The inversion list gives the results for every possible code point
8953 * after the first one in the list. Only those ranges whose index is
8954 * even are ones that the inversion list matches. For the odd ones,
8955 * and if the initial code point is not in the list, we have to skip
8956 * forward to the next element */
8957 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8959 if (i >= len) { /* Finished if beyond the end of the array */
8963 if (current >= end) { /* Finished if beyond the end of what we
8965 if (LIKELY(end < UV_MAX)) {
8969 /* We get here when the upper bound is the maximum
8970 * representable on the machine, and we are looking for just
8971 * that code point. Have to special case it */
8973 goto join_end_of_list;
8976 assert(current >= start);
8978 /* The current range ends one below the next one, except don't go past
8981 upper = (i < len && array[i] < end) ? array[i] : end;
8983 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8984 * for each code point in it */
8985 for (; current < upper; current++) {
8986 const STRLEN offset = (STRLEN)(current - start);
8987 swatch[offset >> 3] |= 1 << (offset & 7);
8992 /* Quit if at the end of the list */
8995 /* But first, have to deal with the highest possible code point on
8996 * the platform. The previous code assumes that <end> is one
8997 * beyond where we want to populate, but that is impossible at the
8998 * platform's infinity, so have to handle it specially */
8999 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
9001 const STRLEN offset = (STRLEN)(end - start);
9002 swatch[offset >> 3] |= 1 << (offset & 7);
9007 /* Advance to the next range, which will be for code points not in the
9016 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9017 const bool complement_b, SV** output)
9019 /* Take the union of two inversion lists and point '*output' to it. On
9020 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9021 * even 'a' or 'b'). If to an inversion list, the contents of the original
9022 * list will be replaced by the union. The first list, 'a', may be
9023 * NULL, in which case a copy of the second list is placed in '*output'.
9024 * If 'complement_b' is TRUE, the union is taken of the complement
9025 * (inversion) of 'b' instead of b itself.
9027 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9028 * Richard Gillam, published by Addison-Wesley, and explained at some
9029 * length there. The preface says to incorporate its examples into your
9030 * code at your own risk.
9032 * The algorithm is like a merge sort. */
9034 const UV* array_a; /* a's array */
9036 UV len_a; /* length of a's array */
9039 SV* u; /* the resulting union */
9043 UV i_a = 0; /* current index into a's array */
9047 /* running count, as explained in the algorithm source book; items are
9048 * stopped accumulating and are output when the count changes to/from 0.
9049 * The count is incremented when we start a range that's in an input's set,
9050 * and decremented when we start a range that's not in a set. So this
9051 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9052 * and hence nothing goes into the union; 1, just one of the inputs is in
9053 * its set (and its current range gets added to the union); and 2 when both
9054 * inputs are in their sets. */
9057 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9059 assert(*output == NULL || SvTYPE(*output) == SVt_INVLIST);
9061 len_b = _invlist_len(b);
9064 /* Here, 'b' is empty, hence it's complement is all possible code
9065 * points. So if the union includes the complement of 'b', it includes
9066 * everything, and we need not even look at 'a'. It's easiest to
9067 * create a new inversion list that matches everything. */
9069 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9071 if (*output == NULL) { /* If the output didn't exist, just point it
9073 *output = everything;
9075 else { /* Otherwise, replace its contents with the new list */
9076 invlist_replace_list_destroys_src(*output, everything);
9077 SvREFCNT_dec_NN(everything);
9083 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9084 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9085 * output will be empty */
9087 if (a == NULL || _invlist_len(a) == 0) {
9088 if (*output == NULL) {
9089 *output = _new_invlist(0);
9092 invlist_clear(*output);
9097 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9098 * union. We can just return a copy of 'a' if '*output' doesn't point
9099 * to an existing list */
9100 if (*output == NULL) {
9101 *output = invlist_clone(a);
9105 /* If the output is to overwrite 'a', we have a no-op, as it's
9111 /* Here, '*output' is to be overwritten by 'a' */
9112 u = invlist_clone(a);
9113 invlist_replace_list_destroys_src(*output, u);
9119 /* Here 'b' is not empty. See about 'a' */
9121 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9123 /* Here, 'a' is empty (and b is not). That means the union will come
9124 * entirely from 'b'. If '*output' is NULL, we can directly return a
9125 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9128 SV ** dest = (*output == NULL) ? output : &u;
9129 *dest = invlist_clone(b);
9131 _invlist_invert(*dest);
9135 invlist_replace_list_destroys_src(*output, u);
9142 /* Here both lists exist and are non-empty */
9143 array_a = invlist_array(a);
9144 array_b = invlist_array(b);
9146 /* If are to take the union of 'a' with the complement of b, set it
9147 * up so are looking at b's complement. */
9150 /* To complement, we invert: if the first element is 0, remove it. To
9151 * do this, we just pretend the array starts one later */
9152 if (array_b[0] == 0) {
9158 /* But if the first element is not zero, we pretend the list starts
9159 * at the 0 that is always stored immediately before the array. */
9165 /* Size the union for the worst case: that the sets are completely
9167 u = _new_invlist(len_a + len_b);
9169 /* Will contain U+0000 if either component does */
9170 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9171 || (len_b > 0 && array_b[0] == 0));
9173 /* Go through each input list item by item, stopping when have exhausted
9175 while (i_a < len_a && i_b < len_b) {
9176 UV cp; /* The element to potentially add to the union's array */
9177 bool cp_in_set; /* is it in the the input list's set or not */
9179 /* We need to take one or the other of the two inputs for the union.
9180 * Since we are merging two sorted lists, we take the smaller of the
9181 * next items. In case of a tie, we take first the one that is in its
9182 * set. If we first took the one not in its set, it would decrement
9183 * the count, possibly to 0 which would cause it to be output as ending
9184 * the range, and the next time through we would take the same number,
9185 * and output it again as beginning the next range. By doing it the
9186 * opposite way, there is no possibility that the count will be
9187 * momentarily decremented to 0, and thus the two adjoining ranges will
9188 * be seamlessly merged. (In a tie and both are in the set or both not
9189 * in the set, it doesn't matter which we take first.) */
9190 if ( array_a[i_a] < array_b[i_b]
9191 || ( array_a[i_a] == array_b[i_b]
9192 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9194 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9195 cp = array_a[i_a++];
9198 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9199 cp = array_b[i_b++];
9202 /* Here, have chosen which of the two inputs to look at. Only output
9203 * if the running count changes to/from 0, which marks the
9204 * beginning/end of a range that's in the set */
9207 array_u[i_u++] = cp;
9214 array_u[i_u++] = cp;
9220 /* The loop above increments the index into exactly one of the input lists
9221 * each iteration, and ends when either index gets to its list end. That
9222 * means the other index is lower than its end, and so something is
9223 * remaining in that one. We decrement 'count', as explained below, if
9224 * that list is in its set. (i_a and i_b each currently index the element
9225 * beyond the one we care about.) */
9226 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9227 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9232 /* Above we decremented 'count' if the list that had unexamined elements in
9233 * it was in its set. This has made it so that 'count' being non-zero
9234 * means there isn't anything left to output; and 'count' equal to 0 means
9235 * that what is left to output is precisely that which is left in the
9236 * non-exhausted input list.
9238 * To see why, note first that the exhausted input obviously has nothing
9239 * left to add to the union. If it was in its set at its end, that means
9240 * the set extends from here to the platform's infinity, and hence so does
9241 * the union and the non-exhausted set is irrelevant. The exhausted set
9242 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9243 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9244 * 'count' remains at 1. This is consistent with the decremented 'count'
9245 * != 0 meaning there's nothing left to add to the union.
9247 * But if the exhausted input wasn't in its set, it contributed 0 to
9248 * 'count', and the rest of the union will be whatever the other input is.
9249 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9250 * otherwise it gets decremented to 0. This is consistent with 'count'
9251 * == 0 meaning the remainder of the union is whatever is left in the
9252 * non-exhausted list. */
9257 IV copy_count = len_a - i_a;
9258 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9259 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9261 else { /* The non-exhausted input is b */
9262 copy_count = len_b - i_b;
9263 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9265 len_u = i_u + copy_count;
9268 /* Set the result to the final length, which can change the pointer to
9269 * array_u, so re-find it. (Note that it is unlikely that this will
9270 * change, as we are shrinking the space, not enlarging it) */
9271 if (len_u != _invlist_len(u)) {
9272 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9274 array_u = invlist_array(u);
9277 if (*output == NULL) { /* Simply return the new inversion list */
9281 /* Otherwise, overwrite the inversion list that was in '*output'. We
9282 * could instead free '*output', and then set it to 'u', but experience
9283 * has shown [perl #127392] that if the input is a mortal, we can get a
9284 * huge build-up of these during regex compilation before they get
9286 invlist_replace_list_destroys_src(*output, u);
9294 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9295 const bool complement_b, SV** i)
9297 /* Take the intersection of two inversion lists and point '*i' to it. On
9298 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9299 * even 'a' or 'b'). If to an inversion list, the contents of the original
9300 * list will be replaced by the intersection. The first list, 'a', may be
9301 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9302 * TRUE, the result will be the intersection of 'a' and the complement (or
9303 * inversion) of 'b' instead of 'b' directly.
9305 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9306 * Richard Gillam, published by Addison-Wesley, and explained at some
9307 * length there. The preface says to incorporate its examples into your
9308 * code at your own risk. In fact, it had bugs
9310 * The algorithm is like a merge sort, and is essentially the same as the
9314 const UV* array_a; /* a's array */
9316 UV len_a; /* length of a's array */
9319 SV* r; /* the resulting intersection */
9323 UV i_a = 0; /* current index into a's array */
9327 /* running count of how many of the two inputs are postitioned at ranges
9328 * that are in their sets. As explained in the algorithm source book,
9329 * items are stopped accumulating and are output when the count changes
9330 * to/from 2. The count is incremented when we start a range that's in an
9331 * input's set, and decremented when we start a range that's not in a set.
9332 * Only when it is 2 are we in the intersection. */
9335 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9337 assert(*i == NULL || SvTYPE(*i) == SVt_INVLIST);
9339 /* Special case if either one is empty */
9340 len_a = (a == NULL) ? 0 : _invlist_len(a);
9341 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9342 if (len_a != 0 && complement_b) {
9344 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9345 * must be empty. Here, also we are using 'b's complement, which
9346 * hence must be every possible code point. Thus the intersection
9349 if (*i == a) { /* No-op */
9354 *i = invlist_clone(a);
9358 r = invlist_clone(a);
9359 invlist_replace_list_destroys_src(*i, r);
9364 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9365 * intersection must be empty */
9367 *i = _new_invlist(0);
9375 /* Here both lists exist and are non-empty */
9376 array_a = invlist_array(a);
9377 array_b = invlist_array(b);
9379 /* If are to take the intersection of 'a' with the complement of b, set it
9380 * up so are looking at b's complement. */
9383 /* To complement, we invert: if the first element is 0, remove it. To
9384 * do this, we just pretend the array starts one later */
9385 if (array_b[0] == 0) {
9391 /* But if the first element is not zero, we pretend the list starts
9392 * at the 0 that is always stored immediately before the array. */
9398 /* Size the intersection for the worst case: that the intersection ends up
9399 * fragmenting everything to be completely disjoint */
9400 r= _new_invlist(len_a + len_b);
9402 /* Will contain U+0000 iff both components do */
9403 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9404 && len_b > 0 && array_b[0] == 0);
9406 /* Go through each list item by item, stopping when have exhausted one of
9408 while (i_a < len_a && i_b < len_b) {
9409 UV cp; /* The element to potentially add to the intersection's
9411 bool cp_in_set; /* Is it in the input list's set or not */
9413 /* We need to take one or the other of the two inputs for the
9414 * intersection. Since we are merging two sorted lists, we take the
9415 * smaller of the next items. In case of a tie, we take first the one
9416 * that is not in its set (a difference from the union algorithm). If
9417 * we first took the one in its set, it would increment the count,
9418 * possibly to 2 which would cause it to be output as starting a range
9419 * in the intersection, and the next time through we would take that
9420 * same number, and output it again as ending the set. By doing the
9421 * opposite of this, there is no possibility that the count will be
9422 * momentarily incremented to 2. (In a tie and both are in the set or
9423 * both not in the set, it doesn't matter which we take first.) */
9424 if ( array_a[i_a] < array_b[i_b]
9425 || ( array_a[i_a] == array_b[i_b]
9426 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9428 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9429 cp = array_a[i_a++];
9432 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9436 /* Here, have chosen which of the two inputs to look at. Only output
9437 * if the running count changes to/from 2, which marks the
9438 * beginning/end of a range that's in the intersection */
9442 array_r[i_r++] = cp;
9447 array_r[i_r++] = cp;
9454 /* The loop above increments the index into exactly one of the input lists
9455 * each iteration, and ends when either index gets to its list end. That
9456 * means the other index is lower than its end, and so something is
9457 * remaining in that one. We increment 'count', as explained below, if the
9458 * exhausted list was in its set. (i_a and i_b each currently index the
9459 * element beyond the one we care about.) */
9460 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9461 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9466 /* Above we incremented 'count' if the exhausted list was in its set. This
9467 * has made it so that 'count' being below 2 means there is nothing left to
9468 * output; otheriwse what's left to add to the intersection is precisely
9469 * that which is left in the non-exhausted input list.
9471 * To see why, note first that the exhausted input obviously has nothing
9472 * left to affect the intersection. If it was in its set at its end, that
9473 * means the set extends from here to the platform's infinity, and hence
9474 * anything in the non-exhausted's list will be in the intersection, and
9475 * anything not in it won't be. Hence, the rest of the intersection is
9476 * precisely what's in the non-exhausted list The exhausted set also
9477 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9478 * it means 'count' is now at least 2. This is consistent with the
9479 * incremented 'count' being >= 2 means to add the non-exhausted list to
9482 * But if the exhausted input wasn't in its set, it contributed 0 to
9483 * 'count', and the intersection can't include anything further; the
9484 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9485 * incremented. This is consistent with 'count' being < 2 meaning nothing
9486 * further to add to the intersection. */
9487 if (count < 2) { /* Nothing left to put in the intersection. */
9490 else { /* copy the non-exhausted list, unchanged. */
9491 IV copy_count = len_a - i_a;
9492 if (copy_count > 0) { /* a is the one with stuff left */
9493 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9495 else { /* b is the one with stuff left */
9496 copy_count = len_b - i_b;
9497 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9499 len_r = i_r + copy_count;
9502 /* Set the result to the final length, which can change the pointer to
9503 * array_r, so re-find it. (Note that it is unlikely that this will
9504 * change, as we are shrinking the space, not enlarging it) */
9505 if (len_r != _invlist_len(r)) {
9506 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9508 array_r = invlist_array(r);
9511 if (*i == NULL) { /* Simply return the calculated intersection */
9514 else { /* Otherwise, replace the existing inversion list in '*i'. We could
9515 instead free '*i', and then set it to 'r', but experience has
9516 shown [perl #127392] that if the input is a mortal, we can get a
9517 huge build-up of these during regex compilation before they get
9520 invlist_replace_list_destroys_src(*i, r);
9532 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9534 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9535 * set. A pointer to the inversion list is returned. This may actually be
9536 * a new list, in which case the passed in one has been destroyed. The
9537 * passed-in inversion list can be NULL, in which case a new one is created
9538 * with just the one range in it. The new list is not necessarily
9539 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9540 * result of this function. The gain would not be large, and in many
9541 * cases, this is called multiple times on a single inversion list, so
9542 * anything freed may almost immediately be needed again.
9544 * This used to mostly call the 'union' routine, but that is much more
9545 * heavyweight than really needed for a single range addition */
9547 UV* array; /* The array implementing the inversion list */
9548 UV len; /* How many elements in 'array' */
9549 SSize_t i_s; /* index into the invlist array where 'start'
9551 SSize_t i_e = 0; /* And the index where 'end' should go */
9552 UV cur_highest; /* The highest code point in the inversion list
9553 upon entry to this function */
9555 /* This range becomes the whole inversion list if none already existed */
9556 if (invlist == NULL) {
9557 invlist = _new_invlist(2);
9558 _append_range_to_invlist(invlist, start, end);
9562 /* Likewise, if the inversion list is currently empty */
9563 len = _invlist_len(invlist);
9565 _append_range_to_invlist(invlist, start, end);
9569 /* Starting here, we have to know the internals of the list */
9570 array = invlist_array(invlist);
9572 /* If the new range ends higher than the current highest ... */
9573 cur_highest = invlist_highest(invlist);
9574 if (end > cur_highest) {
9576 /* If the whole range is higher, we can just append it */
9577 if (start > cur_highest) {
9578 _append_range_to_invlist(invlist, start, end);
9582 /* Otherwise, add the portion that is higher ... */
9583 _append_range_to_invlist(invlist, cur_highest + 1, end);
9585 /* ... and continue on below to handle the rest. As a result of the
9586 * above append, we know that the index of the end of the range is the
9587 * final even numbered one of the array. Recall that the final element
9588 * always starts a range that extends to infinity. If that range is in
9589 * the set (meaning the set goes from here to infinity), it will be an
9590 * even index, but if it isn't in the set, it's odd, and the final
9591 * range in the set is one less, which is even. */
9592 if (end == UV_MAX) {
9600 /* We have dealt with appending, now see about prepending. If the new
9601 * range starts lower than the current lowest ... */
9602 if (start < array[0]) {
9604 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
9605 * Let the union code handle it, rather than having to know the
9606 * trickiness in two code places. */
9607 if (UNLIKELY(start == 0)) {
9610 range_invlist = _new_invlist(2);
9611 _append_range_to_invlist(range_invlist, start, end);
9613 _invlist_union(invlist, range_invlist, &invlist);
9615 SvREFCNT_dec_NN(range_invlist);
9620 /* If the whole new range comes before the first entry, and doesn't
9621 * extend it, we have to insert it as an additional range */
9622 if (end < array[0] - 1) {
9624 goto splice_in_new_range;
9627 /* Here the new range adjoins the existing first range, extending it
9631 /* And continue on below to handle the rest. We know that the index of
9632 * the beginning of the range is the first one of the array */
9635 else { /* Not prepending any part of the new range to the existing list.
9636 * Find where in the list it should go. This finds i_s, such that:
9637 * invlist[i_s] <= start < array[i_s+1]
9639 i_s = _invlist_search(invlist, start);
9642 /* At this point, any extending before the beginning of the inversion list
9643 * and/or after the end has been done. This has made it so that, in the
9644 * code below, each endpoint of the new range is either in a range that is
9645 * in the set, or is in a gap between two ranges that are. This means we
9646 * don't have to worry about exceeding the array bounds.
9648 * Find where in the list the new range ends (but we can skip this if we
9649 * have already determined what it is, or if it will be the same as i_s,
9650 * which we already have computed) */
9652 i_e = (start == end)
9654 : _invlist_search(invlist, end);
9657 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
9658 * is a range that goes to infinity there is no element at invlist[i_e+1],
9659 * so only the first relation holds. */
9661 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9663 /* Here, the ranges on either side of the beginning of the new range
9664 * are in the set, and this range starts in the gap between them.
9666 * The new range extends the range above it downwards if the new range
9667 * ends at or above that range's start */
9668 const bool extends_the_range_above = ( end == UV_MAX
9669 || end + 1 >= array[i_s+1]);
9671 /* The new range extends the range below it upwards if it begins just
9672 * after where that range ends */
9673 if (start == array[i_s]) {
9675 /* If the new range fills the entire gap between the other ranges,
9676 * they will get merged together. Other ranges may also get
9677 * merged, depending on how many of them the new range spans. In
9678 * the general case, we do the merge later, just once, after we
9679 * figure out how many to merge. But in the case where the new
9680 * range exactly spans just this one gap (possibly extending into
9681 * the one above), we do the merge here, and an early exit. This
9682 * is done here to avoid having to special case later. */
9683 if (i_e - i_s <= 1) {
9685 /* If i_e - i_s == 1, it means that the new range terminates
9686 * within the range above, and hence 'extends_the_range_above'
9687 * must be true. (If the range above it extends to infinity,
9688 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
9689 * will be 0, so no harm done.) */
9690 if (extends_the_range_above) {
9691 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
9692 invlist_set_len(invlist,
9694 *(get_invlist_offset_addr(invlist)));
9698 /* Here, i_e must == i_s. We keep them in sync, as they apply
9699 * to the same range, and below we are about to decrement i_s
9704 /* Here, the new range is adjacent to the one below. (It may also
9705 * span beyond the range above, but that will get resolved later.)
9706 * Extend the range below to include this one. */
9707 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
9711 else if (extends_the_range_above) {
9713 /* Here the new range only extends the range above it, but not the
9714 * one below. It merges with the one above. Again, we keep i_e
9715 * and i_s in sync if they point to the same range */
9724 /* Here, we've dealt with the new range start extending any adjoining
9727 * If the new range extends to infinity, it is now the final one,
9728 * regardless of what was there before */
9729 if (UNLIKELY(end == UV_MAX)) {
9730 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
9734 /* If i_e started as == i_s, it has also been dealt with,
9735 * and been updated to the new i_s, which will fail the following if */
9736 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
9738 /* Here, the ranges on either side of the end of the new range are in
9739 * the set, and this range ends in the gap between them.
9741 * If this range is adjacent to (hence extends) the range above it, it
9742 * becomes part of that range; likewise if it extends the range below,
9743 * it becomes part of that range */
9744 if (end + 1 == array[i_e+1]) {
9748 else if (start <= array[i_e]) {
9749 array[i_e] = end + 1;
9756 /* If the range fits entirely in an existing range (as possibly already
9757 * extended above), it doesn't add anything new */
9758 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9762 /* Here, no part of the range is in the list. Must add it. It will
9763 * occupy 2 more slots */
9764 splice_in_new_range:
9766 invlist_extend(invlist, len + 2);
9767 array = invlist_array(invlist);
9768 /* Move the rest of the array down two slots. Don't include any
9770 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
9772 /* Do the actual splice */
9773 array[i_e+1] = start;
9774 array[i_e+2] = end + 1;
9775 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
9779 /* Here the new range crossed the boundaries of a pre-existing range. The
9780 * code above has adjusted things so that both ends are in ranges that are
9781 * in the set. This means everything in between must also be in the set.
9782 * Just squash things together */
9783 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
9784 invlist_set_len(invlist,
9786 *(get_invlist_offset_addr(invlist)));
9792 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9793 UV** other_elements_ptr)
9795 /* Create and return an inversion list whose contents are to be populated
9796 * by the caller. The caller gives the number of elements (in 'size') and
9797 * the very first element ('element0'). This function will set
9798 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9801 * Obviously there is some trust involved that the caller will properly
9802 * fill in the other elements of the array.
9804 * (The first element needs to be passed in, as the underlying code does
9805 * things differently depending on whether it is zero or non-zero) */
9807 SV* invlist = _new_invlist(size);
9810 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9812 invlist = add_cp_to_invlist(invlist, element0);
9813 offset = *get_invlist_offset_addr(invlist);
9815 invlist_set_len(invlist, size, offset);
9816 *other_elements_ptr = invlist_array(invlist) + 1;
9822 PERL_STATIC_INLINE SV*
9823 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9824 return _add_range_to_invlist(invlist, cp, cp);
9827 #ifndef PERL_IN_XSUB_RE
9829 Perl__invlist_invert(pTHX_ SV* const invlist)
9831 /* Complement the input inversion list. This adds a 0 if the list didn't
9832 * have a zero; removes it otherwise. As described above, the data
9833 * structure is set up so that this is very efficient */
9835 PERL_ARGS_ASSERT__INVLIST_INVERT;
9837 assert(! invlist_is_iterating(invlist));
9839 /* The inverse of matching nothing is matching everything */
9840 if (_invlist_len(invlist) == 0) {
9841 _append_range_to_invlist(invlist, 0, UV_MAX);
9845 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9850 PERL_STATIC_INLINE SV*
9851 S_invlist_clone(pTHX_ SV* const invlist)
9854 /* Return a new inversion list that is a copy of the input one, which is
9855 * unchanged. The new list will not be mortal even if the old one was. */
9857 /* Need to allocate extra space to accommodate Perl's addition of a
9858 * trailing NUL to SvPV's, since it thinks they are always strings */
9859 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9860 STRLEN physical_length = SvCUR(invlist);
9861 bool offset = *(get_invlist_offset_addr(invlist));
9863 PERL_ARGS_ASSERT_INVLIST_CLONE;
9865 *(get_invlist_offset_addr(new_invlist)) = offset;
9866 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9867 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9872 PERL_STATIC_INLINE STRLEN*
9873 S_get_invlist_iter_addr(SV* invlist)
9875 /* Return the address of the UV that contains the current iteration
9878 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9880 assert(SvTYPE(invlist) == SVt_INVLIST);
9882 return &(((XINVLIST*) SvANY(invlist))->iterator);
9885 PERL_STATIC_INLINE void
9886 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9888 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9890 *get_invlist_iter_addr(invlist) = 0;
9893 PERL_STATIC_INLINE void
9894 S_invlist_iterfinish(SV* invlist)
9896 /* Terminate iterator for invlist. This is to catch development errors.
9897 * Any iteration that is interrupted before completed should call this
9898 * function. Functions that add code points anywhere else but to the end
9899 * of an inversion list assert that they are not in the middle of an
9900 * iteration. If they were, the addition would make the iteration
9901 * problematical: if the iteration hadn't reached the place where things
9902 * were being added, it would be ok */
9904 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9906 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9910 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9912 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9913 * This call sets in <*start> and <*end>, the next range in <invlist>.
9914 * Returns <TRUE> if successful and the next call will return the next
9915 * range; <FALSE> if was already at the end of the list. If the latter,
9916 * <*start> and <*end> are unchanged, and the next call to this function
9917 * will start over at the beginning of the list */
9919 STRLEN* pos = get_invlist_iter_addr(invlist);
9920 UV len = _invlist_len(invlist);
9923 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9926 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9930 array = invlist_array(invlist);
9932 *start = array[(*pos)++];
9938 *end = array[(*pos)++] - 1;
9944 PERL_STATIC_INLINE UV
9945 S_invlist_highest(SV* const invlist)
9947 /* Returns the highest code point that matches an inversion list. This API
9948 * has an ambiguity, as it returns 0 under either the highest is actually
9949 * 0, or if the list is empty. If this distinction matters to you, check
9950 * for emptiness before calling this function */
9952 UV len = _invlist_len(invlist);
9955 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9961 array = invlist_array(invlist);
9963 /* The last element in the array in the inversion list always starts a
9964 * range that goes to infinity. That range may be for code points that are
9965 * matched in the inversion list, or it may be for ones that aren't
9966 * matched. In the latter case, the highest code point in the set is one
9967 * less than the beginning of this range; otherwise it is the final element
9968 * of this range: infinity */
9969 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
9971 : array[len - 1] - 1;
9975 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
9977 /* Get the contents of an inversion list into a string SV so that they can
9978 * be printed out. If 'traditional_style' is TRUE, it uses the format
9979 * traditionally done for debug tracing; otherwise it uses a format
9980 * suitable for just copying to the output, with blanks between ranges and
9981 * a dash between range components */
9985 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
9986 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
9988 if (traditional_style) {
9989 output = newSVpvs("\n");
9992 output = newSVpvs("");
9995 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
9997 assert(! invlist_is_iterating(invlist));
9999 invlist_iterinit(invlist);
10000 while (invlist_iternext(invlist, &start, &end)) {
10001 if (end == UV_MAX) {
10002 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFINITY%c",
10003 start, intra_range_delimiter,
10004 inter_range_delimiter);
10006 else if (end != start) {
10007 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10009 intra_range_delimiter,
10010 end, inter_range_delimiter);
10013 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10014 start, inter_range_delimiter);
10018 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10019 SvCUR_set(output, SvCUR(output) - 1);
10025 #ifndef PERL_IN_XSUB_RE
10027 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10028 const char * const indent, SV* const invlist)
10030 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10031 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10032 * the string 'indent'. The output looks like this:
10033 [0] 0x000A .. 0x000D
10035 [4] 0x2028 .. 0x2029
10036 [6] 0x3104 .. INFINITY
10037 * This means that the first range of code points matched by the list are
10038 * 0xA through 0xD; the second range contains only the single code point
10039 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10040 * are used to define each range (except if the final range extends to
10041 * infinity, only a single element is needed). The array index of the
10042 * first element for the corresponding range is given in brackets. */
10047 PERL_ARGS_ASSERT__INVLIST_DUMP;
10049 if (invlist_is_iterating(invlist)) {
10050 Perl_dump_indent(aTHX_ level, file,
10051 "%sCan't dump inversion list because is in middle of iterating\n",
10056 invlist_iterinit(invlist);
10057 while (invlist_iternext(invlist, &start, &end)) {
10058 if (end == UV_MAX) {
10059 Perl_dump_indent(aTHX_ level, file,
10060 "%s[%" UVuf "] 0x%04" UVXf " .. INFINITY\n",
10061 indent, (UV)count, start);
10063 else if (end != start) {
10064 Perl_dump_indent(aTHX_ level, file,
10065 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10066 indent, (UV)count, start, end);
10069 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10070 indent, (UV)count, start);
10077 Perl__load_PL_utf8_foldclosures (pTHX)
10079 assert(! PL_utf8_foldclosures);
10081 /* If the folds haven't been read in, call a fold function
10083 if (! PL_utf8_tofold) {
10084 U8 dummy[UTF8_MAXBYTES_CASE+1];
10085 const U8 hyphen[] = HYPHEN_UTF8;
10087 /* This string is just a short named one above \xff */
10088 toFOLD_utf8_safe(hyphen, hyphen + sizeof(hyphen) - 1, dummy, NULL);
10089 assert(PL_utf8_tofold); /* Verify that worked */
10091 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
10095 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10097 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10099 /* Return a boolean as to if the two passed in inversion lists are
10100 * identical. The final argument, if TRUE, says to take the complement of
10101 * the second inversion list before doing the comparison */
10103 const UV* array_a = invlist_array(a);
10104 const UV* array_b = invlist_array(b);
10105 UV len_a = _invlist_len(a);
10106 UV len_b = _invlist_len(b);
10108 PERL_ARGS_ASSERT__INVLISTEQ;
10110 /* If are to compare 'a' with the complement of b, set it
10111 * up so are looking at b's complement. */
10112 if (complement_b) {
10114 /* The complement of nothing is everything, so <a> would have to have
10115 * just one element, starting at zero (ending at infinity) */
10117 return (len_a == 1 && array_a[0] == 0);
10119 else if (array_b[0] == 0) {
10121 /* Otherwise, to complement, we invert. Here, the first element is
10122 * 0, just remove it. To do this, we just pretend the array starts
10130 /* But if the first element is not zero, we pretend the list starts
10131 * at the 0 that is always stored immediately before the array. */
10137 return len_a == len_b
10138 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10144 * As best we can, determine the characters that can match the start of
10145 * the given EXACTF-ish node.
10147 * Returns the invlist as a new SV*; it is the caller's responsibility to
10148 * call SvREFCNT_dec() when done with it.
10151 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10153 const U8 * s = (U8*)STRING(node);
10154 SSize_t bytelen = STR_LEN(node);
10156 /* Start out big enough for 2 separate code points */
10157 SV* invlist = _new_invlist(4);
10159 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
10164 /* We punt and assume can match anything if the node begins
10165 * with a multi-character fold. Things are complicated. For
10166 * example, /ffi/i could match any of:
10167 * "\N{LATIN SMALL LIGATURE FFI}"
10168 * "\N{LATIN SMALL LIGATURE FF}I"
10169 * "F\N{LATIN SMALL LIGATURE FI}"
10170 * plus several other things; and making sure we have all the
10171 * possibilities is hard. */
10172 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10173 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10176 /* Any Latin1 range character can potentially match any
10177 * other depending on the locale */
10178 if (OP(node) == EXACTFL) {
10179 _invlist_union(invlist, PL_Latin1, &invlist);
10182 /* But otherwise, it matches at least itself. We can
10183 * quickly tell if it has a distinct fold, and if so,
10184 * it matches that as well */
10185 invlist = add_cp_to_invlist(invlist, uc);
10186 if (IS_IN_SOME_FOLD_L1(uc))
10187 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10190 /* Some characters match above-Latin1 ones under /i. This
10191 * is true of EXACTFL ones when the locale is UTF-8 */
10192 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10193 && (! isASCII(uc) || (OP(node) != EXACTFA
10194 && OP(node) != EXACTFA_NO_TRIE)))
10196 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10200 else { /* Pattern is UTF-8 */
10201 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10202 STRLEN foldlen = UTF8SKIP(s);
10203 const U8* e = s + bytelen;
10206 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10208 /* The only code points that aren't folded in a UTF EXACTFish
10209 * node are are the problematic ones in EXACTFL nodes */
10210 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10211 /* We need to check for the possibility that this EXACTFL
10212 * node begins with a multi-char fold. Therefore we fold
10213 * the first few characters of it so that we can make that
10218 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10220 *(d++) = (U8) toFOLD(*s);
10225 toFOLD_utf8_safe(s, e, d, &len);
10231 /* And set up so the code below that looks in this folded
10232 * buffer instead of the node's string */
10234 foldlen = UTF8SKIP(folded);
10238 /* When we reach here 's' points to the fold of the first
10239 * character(s) of the node; and 'e' points to far enough along
10240 * the folded string to be just past any possible multi-char
10241 * fold. 'foldlen' is the length in bytes of the first
10244 * Unlike the non-UTF-8 case, the macro for determining if a
10245 * string is a multi-char fold requires all the characters to
10246 * already be folded. This is because of all the complications
10247 * if not. Note that they are folded anyway, except in EXACTFL
10248 * nodes. Like the non-UTF case above, we punt if the node
10249 * begins with a multi-char fold */
10251 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10252 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10254 else { /* Single char fold */
10256 /* It matches all the things that fold to it, which are
10257 * found in PL_utf8_foldclosures (including itself) */
10258 invlist = add_cp_to_invlist(invlist, uc);
10259 if (! PL_utf8_foldclosures)
10260 _load_PL_utf8_foldclosures();
10261 if ((listp = hv_fetch(PL_utf8_foldclosures,
10262 (char *) s, foldlen, FALSE)))
10264 AV* list = (AV*) *listp;
10266 for (k = 0; k <= av_tindex_skip_len_mg(list); k++) {
10267 SV** c_p = av_fetch(list, k, FALSE);
10273 /* /aa doesn't allow folds between ASCII and non- */
10274 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
10275 && isASCII(c) != isASCII(uc))
10280 invlist = add_cp_to_invlist(invlist, c);
10289 #undef HEADER_LENGTH
10290 #undef TO_INTERNAL_SIZE
10291 #undef FROM_INTERNAL_SIZE
10292 #undef INVLIST_VERSION_ID
10294 /* End of inversion list object */
10297 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10299 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10300 * constructs, and updates RExC_flags with them. On input, RExC_parse
10301 * should point to the first flag; it is updated on output to point to the
10302 * final ')' or ':'. There needs to be at least one flag, or this will
10305 /* for (?g), (?gc), and (?o) warnings; warning
10306 about (?c) will warn about (?g) -- japhy */
10308 #define WASTED_O 0x01
10309 #define WASTED_G 0x02
10310 #define WASTED_C 0x04
10311 #define WASTED_GC (WASTED_G|WASTED_C)
10312 I32 wastedflags = 0x00;
10313 U32 posflags = 0, negflags = 0;
10314 U32 *flagsp = &posflags;
10315 char has_charset_modifier = '\0';
10317 bool has_use_defaults = FALSE;
10318 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10319 int x_mod_count = 0;
10321 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10323 /* '^' as an initial flag sets certain defaults */
10324 if (UCHARAT(RExC_parse) == '^') {
10326 has_use_defaults = TRUE;
10327 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10328 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
10329 ? REGEX_UNICODE_CHARSET
10330 : REGEX_DEPENDS_CHARSET);
10333 cs = get_regex_charset(RExC_flags);
10334 if (cs == REGEX_DEPENDS_CHARSET
10335 && (RExC_utf8 || RExC_uni_semantics))
10337 cs = REGEX_UNICODE_CHARSET;
10340 while (RExC_parse < RExC_end) {
10341 /* && strchr("iogcmsx", *RExC_parse) */
10342 /* (?g), (?gc) and (?o) are useless here
10343 and must be globally applied -- japhy */
10344 switch (*RExC_parse) {
10346 /* Code for the imsxn flags */
10347 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10349 case LOCALE_PAT_MOD:
10350 if (has_charset_modifier) {
10351 goto excess_modifier;
10353 else if (flagsp == &negflags) {
10356 cs = REGEX_LOCALE_CHARSET;
10357 has_charset_modifier = LOCALE_PAT_MOD;
10359 case UNICODE_PAT_MOD:
10360 if (has_charset_modifier) {
10361 goto excess_modifier;
10363 else if (flagsp == &negflags) {
10366 cs = REGEX_UNICODE_CHARSET;
10367 has_charset_modifier = UNICODE_PAT_MOD;
10369 case ASCII_RESTRICT_PAT_MOD:
10370 if (flagsp == &negflags) {
10373 if (has_charset_modifier) {
10374 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10375 goto excess_modifier;
10377 /* Doubled modifier implies more restricted */
10378 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10381 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10383 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10385 case DEPENDS_PAT_MOD:
10386 if (has_use_defaults) {
10387 goto fail_modifiers;
10389 else if (flagsp == &negflags) {
10392 else if (has_charset_modifier) {
10393 goto excess_modifier;
10396 /* The dual charset means unicode semantics if the
10397 * pattern (or target, not known until runtime) are
10398 * utf8, or something in the pattern indicates unicode
10400 cs = (RExC_utf8 || RExC_uni_semantics)
10401 ? REGEX_UNICODE_CHARSET
10402 : REGEX_DEPENDS_CHARSET;
10403 has_charset_modifier = DEPENDS_PAT_MOD;
10407 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10408 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10410 else if (has_charset_modifier == *(RExC_parse - 1)) {
10411 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10412 *(RExC_parse - 1));
10415 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10417 NOT_REACHED; /*NOTREACHED*/
10420 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10421 *(RExC_parse - 1));
10422 NOT_REACHED; /*NOTREACHED*/
10423 case ONCE_PAT_MOD: /* 'o' */
10424 case GLOBAL_PAT_MOD: /* 'g' */
10425 if (PASS2 && ckWARN(WARN_REGEXP)) {
10426 const I32 wflagbit = *RExC_parse == 'o'
10429 if (! (wastedflags & wflagbit) ) {
10430 wastedflags |= wflagbit;
10431 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10434 "Useless (%s%c) - %suse /%c modifier",
10435 flagsp == &negflags ? "?-" : "?",
10437 flagsp == &negflags ? "don't " : "",
10444 case CONTINUE_PAT_MOD: /* 'c' */
10445 if (PASS2 && ckWARN(WARN_REGEXP)) {
10446 if (! (wastedflags & WASTED_C) ) {
10447 wastedflags |= WASTED_GC;
10448 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10451 "Useless (%sc) - %suse /gc modifier",
10452 flagsp == &negflags ? "?-" : "?",
10453 flagsp == &negflags ? "don't " : ""
10458 case KEEPCOPY_PAT_MOD: /* 'p' */
10459 if (flagsp == &negflags) {
10461 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10463 *flagsp |= RXf_PMf_KEEPCOPY;
10467 /* A flag is a default iff it is following a minus, so
10468 * if there is a minus, it means will be trying to
10469 * re-specify a default which is an error */
10470 if (has_use_defaults || flagsp == &negflags) {
10471 goto fail_modifiers;
10473 flagsp = &negflags;
10474 wastedflags = 0; /* reset so (?g-c) warns twice */
10480 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
10481 negflags |= RXf_PMf_EXTENDED_MORE;
10483 RExC_flags |= posflags;
10485 if (negflags & RXf_PMf_EXTENDED) {
10486 negflags |= RXf_PMf_EXTENDED_MORE;
10488 RExC_flags &= ~negflags;
10489 set_regex_charset(&RExC_flags, cs);
10494 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10495 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10496 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
10497 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10498 NOT_REACHED; /*NOTREACHED*/
10501 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10504 vFAIL("Sequence (?... not terminated");
10508 - reg - regular expression, i.e. main body or parenthesized thing
10510 * Caller must absorb opening parenthesis.
10512 * Combining parenthesis handling with the base level of regular expression
10513 * is a trifle forced, but the need to tie the tails of the branches to what
10514 * follows makes it hard to avoid.
10516 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10518 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10520 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10523 PERL_STATIC_INLINE regnode *
10524 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10526 char * parse_start,
10531 char* name_start = RExC_parse;
10533 SV *sv_dat = reg_scan_name(pRExC_state, SIZE_ONLY
10534 ? REG_RSN_RETURN_NULL
10535 : REG_RSN_RETURN_DATA);
10536 GET_RE_DEBUG_FLAGS_DECL;
10538 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10540 if (RExC_parse == name_start || *RExC_parse != ch) {
10541 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10542 vFAIL2("Sequence %.3s... not terminated",parse_start);
10546 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10547 RExC_rxi->data->data[num]=(void*)sv_dat;
10548 SvREFCNT_inc_simple_void(sv_dat);
10551 ret = reganode(pRExC_state,
10554 : (ASCII_FOLD_RESTRICTED)
10556 : (AT_LEAST_UNI_SEMANTICS)
10562 *flagp |= HASWIDTH;
10564 Set_Node_Offset(ret, parse_start+1);
10565 Set_Node_Cur_Length(ret, parse_start);
10567 nextchar(pRExC_state);
10571 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
10572 flags. Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan
10573 needs to be restarted, or'd with NEED_UTF8 if the pattern needs to be
10574 upgraded to UTF-8. Otherwise would only return NULL if regbranch() returns
10575 NULL, which cannot happen. */
10577 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
10578 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10579 * 2 is like 1, but indicates that nextchar() has been called to advance
10580 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10581 * this flag alerts us to the need to check for that */
10583 regnode *ret; /* Will be the head of the group. */
10586 regnode *ender = NULL;
10589 U32 oregflags = RExC_flags;
10590 bool have_branch = 0;
10592 I32 freeze_paren = 0;
10593 I32 after_freeze = 0;
10594 I32 num; /* numeric backreferences */
10596 char * parse_start = RExC_parse; /* MJD */
10597 char * const oregcomp_parse = RExC_parse;
10599 GET_RE_DEBUG_FLAGS_DECL;
10601 PERL_ARGS_ASSERT_REG;
10602 DEBUG_PARSE("reg ");
10604 *flagp = 0; /* Tentatively. */
10606 /* Having this true makes it feasible to have a lot fewer tests for the
10607 * parse pointer being in scope. For example, we can write
10608 * while(isFOO(*RExC_parse)) RExC_parse++;
10610 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
10612 assert(*RExC_end == '\0');
10614 /* Make an OPEN node, if parenthesized. */
10617 /* Under /x, space and comments can be gobbled up between the '(' and
10618 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
10619 * intervening space, as the sequence is a token, and a token should be
10621 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
10623 if (RExC_parse >= RExC_end) {
10624 vFAIL("Unmatched (");
10627 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
10628 char *start_verb = RExC_parse + 1;
10630 char *start_arg = NULL;
10631 unsigned char op = 0;
10632 int arg_required = 0;
10633 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
10635 if (has_intervening_patws) {
10636 RExC_parse++; /* past the '*' */
10637 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
10639 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
10640 if ( *RExC_parse == ':' ) {
10641 start_arg = RExC_parse + 1;
10644 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10646 verb_len = RExC_parse - start_verb;
10648 if (RExC_parse >= RExC_end) {
10649 goto unterminated_verb_pattern;
10651 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10652 while ( RExC_parse < RExC_end && *RExC_parse != ')' )
10653 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10654 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10655 unterminated_verb_pattern:
10656 vFAIL("Unterminated verb pattern argument");
10657 if ( RExC_parse == start_arg )
10660 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10661 vFAIL("Unterminated verb pattern");
10664 /* Here, we know that RExC_parse < RExC_end */
10666 switch ( *start_verb ) {
10667 case 'A': /* (*ACCEPT) */
10668 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
10670 internal_argval = RExC_nestroot;
10673 case 'C': /* (*COMMIT) */
10674 if ( memEQs(start_verb,verb_len,"COMMIT") )
10677 case 'F': /* (*FAIL) */
10678 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
10682 case ':': /* (*:NAME) */
10683 case 'M': /* (*MARK:NAME) */
10684 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
10689 case 'P': /* (*PRUNE) */
10690 if ( memEQs(start_verb,verb_len,"PRUNE") )
10693 case 'S': /* (*SKIP) */
10694 if ( memEQs(start_verb,verb_len,"SKIP") )
10697 case 'T': /* (*THEN) */
10698 /* [19:06] <TimToady> :: is then */
10699 if ( memEQs(start_verb,verb_len,"THEN") ) {
10701 RExC_seen |= REG_CUTGROUP_SEEN;
10706 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10708 "Unknown verb pattern '%" UTF8f "'",
10709 UTF8fARG(UTF, verb_len, start_verb));
10711 if ( arg_required && !start_arg ) {
10712 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
10713 verb_len, start_verb);
10715 if (internal_argval == -1) {
10716 ret = reganode(pRExC_state, op, 0);
10718 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
10720 RExC_seen |= REG_VERBARG_SEEN;
10721 if ( ! SIZE_ONLY ) {
10723 SV *sv = newSVpvn( start_arg,
10724 RExC_parse - start_arg);
10725 ARG(ret) = add_data( pRExC_state,
10726 STR_WITH_LEN("S"));
10727 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
10732 if ( internal_argval != -1 )
10733 ARG2L_SET(ret, internal_argval);
10735 nextchar(pRExC_state);
10738 else if (*RExC_parse == '?') { /* (?...) */
10739 bool is_logical = 0;
10740 const char * const seqstart = RExC_parse;
10741 const char * endptr;
10742 if (has_intervening_patws) {
10744 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
10747 RExC_parse++; /* past the '?' */
10748 paren = *RExC_parse; /* might be a trailing NUL, if not
10750 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10751 if (RExC_parse > RExC_end) {
10754 ret = NULL; /* For look-ahead/behind. */
10757 case 'P': /* (?P...) variants for those used to PCRE/Python */
10758 paren = *RExC_parse;
10759 if ( paren == '<') { /* (?P<...>) named capture */
10761 if (RExC_parse >= RExC_end) {
10762 vFAIL("Sequence (?P<... not terminated");
10764 goto named_capture;
10766 else if (paren == '>') { /* (?P>name) named recursion */
10768 if (RExC_parse >= RExC_end) {
10769 vFAIL("Sequence (?P>... not terminated");
10771 goto named_recursion;
10773 else if (paren == '=') { /* (?P=...) named backref */
10775 return handle_named_backref(pRExC_state, flagp,
10778 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10779 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10780 vFAIL3("Sequence (%.*s...) not recognized",
10781 RExC_parse-seqstart, seqstart);
10782 NOT_REACHED; /*NOTREACHED*/
10783 case '<': /* (?<...) */
10784 if (*RExC_parse == '!')
10786 else if (*RExC_parse != '=')
10793 case '\'': /* (?'...') */
10794 name_start = RExC_parse;
10795 svname = reg_scan_name(pRExC_state,
10796 SIZE_ONLY /* reverse test from the others */
10797 ? REG_RSN_RETURN_NAME
10798 : REG_RSN_RETURN_NULL);
10799 if ( RExC_parse == name_start
10800 || RExC_parse >= RExC_end
10801 || *RExC_parse != paren)
10803 vFAIL2("Sequence (?%c... not terminated",
10804 paren=='>' ? '<' : paren);
10809 if (!svname) /* shouldn't happen */
10811 "panic: reg_scan_name returned NULL");
10812 if (!RExC_paren_names) {
10813 RExC_paren_names= newHV();
10814 sv_2mortal(MUTABLE_SV(RExC_paren_names));
10816 RExC_paren_name_list= newAV();
10817 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10820 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10822 sv_dat = HeVAL(he_str);
10824 /* croak baby croak */
10826 "panic: paren_name hash element allocation failed");
10827 } else if ( SvPOK(sv_dat) ) {
10828 /* (?|...) can mean we have dupes so scan to check
10829 its already been stored. Maybe a flag indicating
10830 we are inside such a construct would be useful,
10831 but the arrays are likely to be quite small, so
10832 for now we punt -- dmq */
10833 IV count = SvIV(sv_dat);
10834 I32 *pv = (I32*)SvPVX(sv_dat);
10836 for ( i = 0 ; i < count ; i++ ) {
10837 if ( pv[i] == RExC_npar ) {
10843 pv = (I32*)SvGROW(sv_dat,
10844 SvCUR(sv_dat) + sizeof(I32)+1);
10845 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10846 pv[count] = RExC_npar;
10847 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10850 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10851 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10854 SvIV_set(sv_dat, 1);
10857 /* Yes this does cause a memory leak in debugging Perls
10859 if (!av_store(RExC_paren_name_list,
10860 RExC_npar, SvREFCNT_inc(svname)))
10861 SvREFCNT_dec_NN(svname);
10864 /*sv_dump(sv_dat);*/
10866 nextchar(pRExC_state);
10868 goto capturing_parens;
10870 RExC_seen |= REG_LOOKBEHIND_SEEN;
10871 RExC_in_lookbehind++;
10873 if (RExC_parse >= RExC_end) {
10874 vFAIL("Sequence (?... not terminated");
10878 case '=': /* (?=...) */
10879 RExC_seen_zerolen++;
10881 case '!': /* (?!...) */
10882 RExC_seen_zerolen++;
10883 /* check if we're really just a "FAIL" assertion */
10884 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
10885 FALSE /* Don't force to /x */ );
10886 if (*RExC_parse == ')') {
10887 ret=reganode(pRExC_state, OPFAIL, 0);
10888 nextchar(pRExC_state);
10892 case '|': /* (?|...) */
10893 /* branch reset, behave like a (?:...) except that
10894 buffers in alternations share the same numbers */
10896 after_freeze = freeze_paren = RExC_npar;
10898 case ':': /* (?:...) */
10899 case '>': /* (?>...) */
10901 case '$': /* (?$...) */
10902 case '@': /* (?@...) */
10903 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10905 case '0' : /* (?0) */
10906 case 'R' : /* (?R) */
10907 if (RExC_parse == RExC_end || *RExC_parse != ')')
10908 FAIL("Sequence (?R) not terminated");
10910 RExC_seen |= REG_RECURSE_SEEN;
10911 *flagp |= POSTPONED;
10912 goto gen_recurse_regop;
10914 /* named and numeric backreferences */
10915 case '&': /* (?&NAME) */
10916 parse_start = RExC_parse - 1;
10919 SV *sv_dat = reg_scan_name(pRExC_state,
10920 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10921 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10923 if (RExC_parse >= RExC_end || *RExC_parse != ')')
10924 vFAIL("Sequence (?&... not terminated");
10925 goto gen_recurse_regop;
10928 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10930 vFAIL("Illegal pattern");
10932 goto parse_recursion;
10934 case '-': /* (?-1) */
10935 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10936 RExC_parse--; /* rewind to let it be handled later */
10940 case '1': case '2': case '3': case '4': /* (?1) */
10941 case '5': case '6': case '7': case '8': case '9':
10942 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
10945 bool is_neg = FALSE;
10947 parse_start = RExC_parse - 1; /* MJD */
10948 if (*RExC_parse == '-') {
10952 if (grok_atoUV(RExC_parse, &unum, &endptr)
10956 RExC_parse = (char*)endptr;
10960 /* Some limit for num? */
10964 if (*RExC_parse!=')')
10965 vFAIL("Expecting close bracket");
10968 if ( paren == '-' ) {
10970 Diagram of capture buffer numbering.
10971 Top line is the normal capture buffer numbers
10972 Bottom line is the negative indexing as from
10976 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
10980 num = RExC_npar + num;
10983 vFAIL("Reference to nonexistent group");
10985 } else if ( paren == '+' ) {
10986 num = RExC_npar + num - 1;
10988 /* We keep track how many GOSUB items we have produced.
10989 To start off the ARG2L() of the GOSUB holds its "id",
10990 which is used later in conjunction with RExC_recurse
10991 to calculate the offset we need to jump for the GOSUB,
10992 which it will store in the final representation.
10993 We have to defer the actual calculation until much later
10994 as the regop may move.
10997 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
10999 if (num > (I32)RExC_rx->nparens) {
11001 vFAIL("Reference to nonexistent group");
11003 RExC_recurse_count++;
11004 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11005 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
11006 22, "| |", (int)(depth * 2 + 1), "",
11007 (UV)ARG(ret), (IV)ARG2L(ret)));
11009 RExC_seen |= REG_RECURSE_SEEN;
11011 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
11012 Set_Node_Offset(ret, parse_start); /* MJD */
11014 *flagp |= POSTPONED;
11015 assert(*RExC_parse == ')');
11016 nextchar(pRExC_state);
11021 case '?': /* (??...) */
11023 if (*RExC_parse != '{') {
11024 RExC_parse += SKIP_IF_CHAR(RExC_parse);
11025 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11027 "Sequence (%" UTF8f "...) not recognized",
11028 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11029 NOT_REACHED; /*NOTREACHED*/
11031 *flagp |= POSTPONED;
11035 case '{': /* (?{...}) */
11038 struct reg_code_block *cb;
11040 RExC_seen_zerolen++;
11042 if ( !pRExC_state->code_blocks
11043 || pRExC_state->code_index
11044 >= pRExC_state->code_blocks->count
11045 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11046 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11049 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11050 FAIL("panic: Sequence (?{...}): no code block found\n");
11051 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11053 /* this is a pre-compiled code block (?{...}) */
11054 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11055 RExC_parse = RExC_start + cb->end;
11058 if (cb->src_regex) {
11059 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11060 RExC_rxi->data->data[n] =
11061 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11062 RExC_rxi->data->data[n+1] = (void*)o;
11065 n = add_data(pRExC_state,
11066 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11067 RExC_rxi->data->data[n] = (void*)o;
11070 pRExC_state->code_index++;
11071 nextchar(pRExC_state);
11075 ret = reg_node(pRExC_state, LOGICAL);
11077 eval = reg2Lanode(pRExC_state, EVAL,
11080 /* for later propagation into (??{})
11082 RExC_flags & RXf_PMf_COMPILETIME
11087 REGTAIL(pRExC_state, ret, eval);
11088 /* deal with the length of this later - MJD */
11091 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11092 Set_Node_Length(ret, RExC_parse - parse_start + 1);
11093 Set_Node_Offset(ret, parse_start);
11096 case '(': /* (?(?{...})...) and (?(?=...)...) */
11099 const int DEFINE_len = sizeof("DEFINE") - 1;
11100 if (RExC_parse[0] == '?') { /* (?(?...)) */
11101 if ( RExC_parse < RExC_end - 1
11102 && ( RExC_parse[1] == '='
11103 || RExC_parse[1] == '!'
11104 || RExC_parse[1] == '<'
11105 || RExC_parse[1] == '{')
11106 ) { /* Lookahead or eval. */
11110 ret = reg_node(pRExC_state, LOGICAL);
11114 tail = reg(pRExC_state, 1, &flag, depth+1);
11115 if (flag & (RESTART_PASS1|NEED_UTF8)) {
11116 *flagp = flag & (RESTART_PASS1|NEED_UTF8);
11119 REGTAIL(pRExC_state, ret, tail);
11122 /* Fall through to ‘Unknown switch condition’ at the
11123 end of the if/else chain. */
11125 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11126 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11128 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11129 char *name_start= RExC_parse++;
11131 SV *sv_dat=reg_scan_name(pRExC_state,
11132 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11133 if ( RExC_parse == name_start
11134 || RExC_parse >= RExC_end
11135 || *RExC_parse != ch)
11137 vFAIL2("Sequence (?(%c... not terminated",
11138 (ch == '>' ? '<' : ch));
11142 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11143 RExC_rxi->data->data[num]=(void*)sv_dat;
11144 SvREFCNT_inc_simple_void(sv_dat);
11146 ret = reganode(pRExC_state,NGROUPP,num);
11147 goto insert_if_check_paren;
11149 else if (RExC_end - RExC_parse >= DEFINE_len
11150 && strnEQ(RExC_parse, "DEFINE", DEFINE_len))
11152 ret = reganode(pRExC_state,DEFINEP,0);
11153 RExC_parse += DEFINE_len;
11155 goto insert_if_check_paren;
11157 else if (RExC_parse[0] == 'R') {
11159 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11160 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11161 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11164 if (RExC_parse[0] == '0') {
11168 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11170 if (grok_atoUV(RExC_parse, &uv, &endptr)
11173 parno = (I32)uv + 1;
11174 RExC_parse = (char*)endptr;
11176 /* else "Switch condition not recognized" below */
11177 } else if (RExC_parse[0] == '&') {
11180 sv_dat = reg_scan_name(pRExC_state,
11182 ? REG_RSN_RETURN_NULL
11183 : REG_RSN_RETURN_DATA);
11185 /* we should only have a false sv_dat when
11186 * SIZE_ONLY is true, and we always have false
11187 * sv_dat when SIZE_ONLY is true.
11188 * reg_scan_name() will VFAIL() if the name is
11189 * unknown when SIZE_ONLY is false, and otherwise
11190 * will return something, and when SIZE_ONLY is
11191 * true, reg_scan_name() just parses the string,
11192 * and doesnt return anything. (in theory) */
11193 assert(SIZE_ONLY ? !sv_dat : !!sv_dat);
11196 parno = 1 + *((I32 *)SvPVX(sv_dat));
11198 ret = reganode(pRExC_state,INSUBP,parno);
11199 goto insert_if_check_paren;
11201 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11205 if (grok_atoUV(RExC_parse, &uv, &endptr)
11209 RExC_parse = (char*)endptr;
11212 vFAIL("panic: grok_atoUV returned FALSE");
11214 ret = reganode(pRExC_state, GROUPP, parno);
11216 insert_if_check_paren:
11217 if (UCHARAT(RExC_parse) != ')') {
11218 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11219 vFAIL("Switch condition not recognized");
11221 nextchar(pRExC_state);
11223 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
11224 br = regbranch(pRExC_state, &flags, 1,depth+1);
11226 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11227 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11230 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11233 REGTAIL(pRExC_state, br, reganode(pRExC_state,
11235 c = UCHARAT(RExC_parse);
11236 nextchar(pRExC_state);
11237 if (flags&HASWIDTH)
11238 *flagp |= HASWIDTH;
11241 vFAIL("(?(DEFINE)....) does not allow branches");
11243 /* Fake one for optimizer. */
11244 lastbr = reganode(pRExC_state, IFTHEN, 0);
11246 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
11247 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11248 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11251 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11254 REGTAIL(pRExC_state, ret, lastbr);
11255 if (flags&HASWIDTH)
11256 *flagp |= HASWIDTH;
11257 c = UCHARAT(RExC_parse);
11258 nextchar(pRExC_state);
11263 if (RExC_parse >= RExC_end)
11264 vFAIL("Switch (?(condition)... not terminated");
11266 vFAIL("Switch (?(condition)... contains too many branches");
11268 ender = reg_node(pRExC_state, TAIL);
11269 REGTAIL(pRExC_state, br, ender);
11271 REGTAIL(pRExC_state, lastbr, ender);
11272 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11275 REGTAIL(pRExC_state, ret, ender);
11276 RExC_size++; /* XXX WHY do we need this?!!
11277 For large programs it seems to be required
11278 but I can't figure out why. -- dmq*/
11281 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11282 vFAIL("Unknown switch condition (?(...))");
11284 case '[': /* (?[ ... ]) */
11285 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
11287 case 0: /* A NUL */
11288 RExC_parse--; /* for vFAIL to print correctly */
11289 vFAIL("Sequence (? incomplete");
11291 default: /* e.g., (?i) */
11292 RExC_parse = (char *) seqstart + 1;
11294 parse_lparen_question_flags(pRExC_state);
11295 if (UCHARAT(RExC_parse) != ':') {
11296 if (RExC_parse < RExC_end)
11297 nextchar(pRExC_state);
11302 nextchar(pRExC_state);
11307 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11312 ret = reganode(pRExC_state, OPEN, parno);
11314 if (!RExC_nestroot)
11315 RExC_nestroot = parno;
11316 if (RExC_open_parens && !RExC_open_parens[parno])
11318 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11319 "%*s%*s Setting open paren #%" IVdf " to %d\n",
11320 22, "| |", (int)(depth * 2 + 1), "",
11321 (IV)parno, REG_NODE_NUM(ret)));
11322 RExC_open_parens[parno]= ret;
11325 Set_Node_Length(ret, 1); /* MJD */
11326 Set_Node_Offset(ret, RExC_parse); /* MJD */
11329 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
11338 /* Pick up the branches, linking them together. */
11339 parse_start = RExC_parse; /* MJD */
11340 br = regbranch(pRExC_state, &flags, 1,depth+1);
11342 /* branch_len = (paren != 0); */
11345 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11346 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11349 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11351 if (*RExC_parse == '|') {
11352 if (!SIZE_ONLY && RExC_extralen) {
11353 reginsert(pRExC_state, BRANCHJ, br, depth+1);
11356 reginsert(pRExC_state, BRANCH, br, depth+1);
11357 Set_Node_Length(br, paren != 0);
11358 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
11362 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
11364 else if (paren == ':') {
11365 *flagp |= flags&SIMPLE;
11367 if (is_open) { /* Starts with OPEN. */
11368 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
11370 else if (paren != '?') /* Not Conditional */
11372 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11374 while (*RExC_parse == '|') {
11375 if (!SIZE_ONLY && RExC_extralen) {
11376 ender = reganode(pRExC_state, LONGJMP,0);
11378 /* Append to the previous. */
11379 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11382 RExC_extralen += 2; /* Account for LONGJMP. */
11383 nextchar(pRExC_state);
11384 if (freeze_paren) {
11385 if (RExC_npar > after_freeze)
11386 after_freeze = RExC_npar;
11387 RExC_npar = freeze_paren;
11389 br = regbranch(pRExC_state, &flags, 0, depth+1);
11392 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11393 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11396 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11398 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
11400 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11403 if (have_branch || paren != ':') {
11404 /* Make a closing node, and hook it on the end. */
11407 ender = reg_node(pRExC_state, TAIL);
11410 ender = reganode(pRExC_state, CLOSE, parno);
11411 if ( RExC_close_parens ) {
11412 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11413 "%*s%*s Setting close paren #%" IVdf " to %d\n",
11414 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
11415 RExC_close_parens[parno]= ender;
11416 if (RExC_nestroot == parno)
11419 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
11420 Set_Node_Length(ender,1); /* MJD */
11426 *flagp &= ~HASWIDTH;
11429 ender = reg_node(pRExC_state, SUCCEED);
11432 ender = reg_node(pRExC_state, END);
11434 assert(!RExC_end_op); /* there can only be one! */
11435 RExC_end_op = ender;
11436 if (RExC_close_parens) {
11437 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11438 "%*s%*s Setting close paren #0 (END) to %d\n",
11439 22, "| |", (int)(depth * 2 + 1), "", REG_NODE_NUM(ender)));
11441 RExC_close_parens[0]= ender;
11446 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11447 DEBUG_PARSE_MSG("lsbr");
11448 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
11449 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11450 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11451 SvPV_nolen_const(RExC_mysv1),
11452 (IV)REG_NODE_NUM(lastbr),
11453 SvPV_nolen_const(RExC_mysv2),
11454 (IV)REG_NODE_NUM(ender),
11455 (IV)(ender - lastbr)
11458 REGTAIL(pRExC_state, lastbr, ender);
11460 if (have_branch && !SIZE_ONLY) {
11461 char is_nothing= 1;
11463 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
11465 /* Hook the tails of the branches to the closing node. */
11466 for (br = ret; br; br = regnext(br)) {
11467 const U8 op = PL_regkind[OP(br)];
11468 if (op == BRANCH) {
11469 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
11470 if ( OP(NEXTOPER(br)) != NOTHING
11471 || regnext(NEXTOPER(br)) != ender)
11474 else if (op == BRANCHJ) {
11475 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
11476 /* for now we always disable this optimisation * /
11477 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
11478 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
11484 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
11485 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11486 DEBUG_PARSE_MSG("NADA");
11487 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
11488 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11489 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11490 SvPV_nolen_const(RExC_mysv1),
11491 (IV)REG_NODE_NUM(ret),
11492 SvPV_nolen_const(RExC_mysv2),
11493 (IV)REG_NODE_NUM(ender),
11498 if (OP(ender) == TAIL) {
11503 for ( opt= br + 1; opt < ender ; opt++ )
11504 OP(opt)= OPTIMIZED;
11505 NEXT_OFF(br)= ender - br;
11513 static const char parens[] = "=!<,>";
11515 if (paren && (p = strchr(parens, paren))) {
11516 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
11517 int flag = (p - parens) > 1;
11520 node = SUSPEND, flag = 0;
11521 reginsert(pRExC_state, node,ret, depth+1);
11522 Set_Node_Cur_Length(ret, parse_start);
11523 Set_Node_Offset(ret, parse_start + 1);
11525 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
11529 /* Check for proper termination. */
11531 /* restore original flags, but keep (?p) and, if we've changed from /d
11532 * rules to /u, keep the /u */
11533 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
11534 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
11535 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
11537 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
11538 RExC_parse = oregcomp_parse;
11539 vFAIL("Unmatched (");
11541 nextchar(pRExC_state);
11543 else if (!paren && RExC_parse < RExC_end) {
11544 if (*RExC_parse == ')') {
11546 vFAIL("Unmatched )");
11549 FAIL("Junk on end of regexp"); /* "Can't happen". */
11550 NOT_REACHED; /* NOTREACHED */
11553 if (RExC_in_lookbehind) {
11554 RExC_in_lookbehind--;
11556 if (after_freeze > RExC_npar)
11557 RExC_npar = after_freeze;
11562 - regbranch - one alternative of an | operator
11564 * Implements the concatenation operator.
11566 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11567 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11570 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
11573 regnode *chain = NULL;
11575 I32 flags = 0, c = 0;
11576 GET_RE_DEBUG_FLAGS_DECL;
11578 PERL_ARGS_ASSERT_REGBRANCH;
11580 DEBUG_PARSE("brnc");
11585 if (!SIZE_ONLY && RExC_extralen)
11586 ret = reganode(pRExC_state, BRANCHJ,0);
11588 ret = reg_node(pRExC_state, BRANCH);
11589 Set_Node_Length(ret, 1);
11593 if (!first && SIZE_ONLY)
11594 RExC_extralen += 1; /* BRANCHJ */
11596 *flagp = WORST; /* Tentatively. */
11598 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11599 FALSE /* Don't force to /x */ );
11600 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
11601 flags &= ~TRYAGAIN;
11602 latest = regpiece(pRExC_state, &flags,depth+1);
11603 if (latest == NULL) {
11604 if (flags & TRYAGAIN)
11606 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11607 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11610 FAIL2("panic: regpiece returned NULL, flags=%#" UVxf, (UV) flags);
11612 else if (ret == NULL)
11614 *flagp |= flags&(HASWIDTH|POSTPONED);
11615 if (chain == NULL) /* First piece. */
11616 *flagp |= flags&SPSTART;
11618 /* FIXME adding one for every branch after the first is probably
11619 * excessive now we have TRIE support. (hv) */
11621 REGTAIL(pRExC_state, chain, latest);
11626 if (chain == NULL) { /* Loop ran zero times. */
11627 chain = reg_node(pRExC_state, NOTHING);
11632 *flagp |= flags&SIMPLE;
11639 - regpiece - something followed by possible quantifier * + ? {n,m}
11641 * Note that the branching code sequences used for ? and the general cases
11642 * of * and + are somewhat optimized: they use the same NOTHING node as
11643 * both the endmarker for their branch list and the body of the last branch.
11644 * It might seem that this node could be dispensed with entirely, but the
11645 * endmarker role is not redundant.
11647 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
11649 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11650 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11653 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11659 const char * const origparse = RExC_parse;
11661 I32 max = REG_INFTY;
11662 #ifdef RE_TRACK_PATTERN_OFFSETS
11665 const char *maxpos = NULL;
11668 /* Save the original in case we change the emitted regop to a FAIL. */
11669 regnode * const orig_emit = RExC_emit;
11671 GET_RE_DEBUG_FLAGS_DECL;
11673 PERL_ARGS_ASSERT_REGPIECE;
11675 DEBUG_PARSE("piec");
11677 ret = regatom(pRExC_state, &flags,depth+1);
11679 if (flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8))
11680 *flagp |= flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8);
11682 FAIL2("panic: regatom returned NULL, flags=%#" UVxf, (UV) flags);
11688 if (op == '{' && regcurly(RExC_parse)) {
11690 #ifdef RE_TRACK_PATTERN_OFFSETS
11691 parse_start = RExC_parse; /* MJD */
11693 next = RExC_parse + 1;
11694 while (isDIGIT(*next) || *next == ',') {
11695 if (*next == ',') {
11703 if (*next == '}') { /* got one */
11704 const char* endptr;
11708 if (isDIGIT(*RExC_parse)) {
11709 if (!grok_atoUV(RExC_parse, &uv, &endptr))
11710 vFAIL("Invalid quantifier in {,}");
11711 if (uv >= REG_INFTY)
11712 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11717 if (*maxpos == ',')
11720 maxpos = RExC_parse;
11721 if (isDIGIT(*maxpos)) {
11722 if (!grok_atoUV(maxpos, &uv, &endptr))
11723 vFAIL("Invalid quantifier in {,}");
11724 if (uv >= REG_INFTY)
11725 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11728 max = REG_INFTY; /* meaning "infinity" */
11731 nextchar(pRExC_state);
11732 if (max < min) { /* If can't match, warn and optimize to fail
11734 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
11736 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
11737 NEXT_OFF(orig_emit)= regarglen[OPFAIL] + NODE_STEP_REGNODE;
11741 else if (min == max && *RExC_parse == '?')
11744 ckWARN2reg(RExC_parse + 1,
11745 "Useless use of greediness modifier '%c'",
11751 if ((flags&SIMPLE)) {
11752 if (min == 0 && max == REG_INFTY) {
11753 reginsert(pRExC_state, STAR, ret, depth+1);
11756 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11759 if (min == 1 && max == REG_INFTY) {
11760 reginsert(pRExC_state, PLUS, ret, depth+1);
11763 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11766 MARK_NAUGHTY_EXP(2, 2);
11767 reginsert(pRExC_state, CURLY, ret, depth+1);
11768 Set_Node_Offset(ret, parse_start+1); /* MJD */
11769 Set_Node_Cur_Length(ret, parse_start);
11772 regnode * const w = reg_node(pRExC_state, WHILEM);
11775 REGTAIL(pRExC_state, ret, w);
11776 if (!SIZE_ONLY && RExC_extralen) {
11777 reginsert(pRExC_state, LONGJMP,ret, depth+1);
11778 reginsert(pRExC_state, NOTHING,ret, depth+1);
11779 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
11781 reginsert(pRExC_state, CURLYX,ret, depth+1);
11783 Set_Node_Offset(ret, parse_start+1);
11784 Set_Node_Length(ret,
11785 op == '{' ? (RExC_parse - parse_start) : 1);
11787 if (!SIZE_ONLY && RExC_extralen)
11788 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
11789 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
11791 RExC_whilem_seen++, RExC_extralen += 3;
11792 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
11799 *flagp |= HASWIDTH;
11801 ARG1_SET(ret, (U16)min);
11802 ARG2_SET(ret, (U16)max);
11804 if (max == REG_INFTY)
11805 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11811 if (!ISMULT1(op)) {
11816 #if 0 /* Now runtime fix should be reliable. */
11818 /* if this is reinstated, don't forget to put this back into perldiag:
11820 =item Regexp *+ operand could be empty at {#} in regex m/%s/
11822 (F) The part of the regexp subject to either the * or + quantifier
11823 could match an empty string. The {#} shows in the regular
11824 expression about where the problem was discovered.
11828 if (!(flags&HASWIDTH) && op != '?')
11829 vFAIL("Regexp *+ operand could be empty");
11832 #ifdef RE_TRACK_PATTERN_OFFSETS
11833 parse_start = RExC_parse;
11835 nextchar(pRExC_state);
11837 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
11843 else if (op == '+') {
11847 else if (op == '?') {
11852 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
11853 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
11854 ckWARN2reg(RExC_parse,
11855 "%" UTF8f " matches null string many times",
11856 UTF8fARG(UTF, (RExC_parse >= origparse
11857 ? RExC_parse - origparse
11860 (void)ReREFCNT_inc(RExC_rx_sv);
11863 if (*RExC_parse == '?') {
11864 nextchar(pRExC_state);
11865 reginsert(pRExC_state, MINMOD, ret, depth+1);
11866 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
11868 else if (*RExC_parse == '+') {
11870 nextchar(pRExC_state);
11871 ender = reg_node(pRExC_state, SUCCEED);
11872 REGTAIL(pRExC_state, ret, ender);
11873 reginsert(pRExC_state, SUSPEND, ret, depth+1);
11875 ender = reg_node(pRExC_state, TAIL);
11876 REGTAIL(pRExC_state, ret, ender);
11879 if (ISMULT2(RExC_parse)) {
11881 vFAIL("Nested quantifiers");
11888 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
11897 /* This routine teases apart the various meanings of \N and returns
11898 * accordingly. The input parameters constrain which meaning(s) is/are valid
11899 * in the current context.
11901 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
11903 * If <code_point_p> is not NULL, the context is expecting the result to be a
11904 * single code point. If this \N instance turns out to a single code point,
11905 * the function returns TRUE and sets *code_point_p to that code point.
11907 * If <node_p> is not NULL, the context is expecting the result to be one of
11908 * the things representable by a regnode. If this \N instance turns out to be
11909 * one such, the function generates the regnode, returns TRUE and sets *node_p
11910 * to point to that regnode.
11912 * If this instance of \N isn't legal in any context, this function will
11913 * generate a fatal error and not return.
11915 * On input, RExC_parse should point to the first char following the \N at the
11916 * time of the call. On successful return, RExC_parse will have been updated
11917 * to point to just after the sequence identified by this routine. Also
11918 * *flagp has been updated as needed.
11920 * When there is some problem with the current context and this \N instance,
11921 * the function returns FALSE, without advancing RExC_parse, nor setting
11922 * *node_p, nor *code_point_p, nor *flagp.
11924 * If <cp_count> is not NULL, the caller wants to know the length (in code
11925 * points) that this \N sequence matches. This is set even if the function
11926 * returns FALSE, as detailed below.
11928 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
11930 * Probably the most common case is for the \N to specify a single code point.
11931 * *cp_count will be set to 1, and *code_point_p will be set to that code
11934 * Another possibility is for the input to be an empty \N{}, which for
11935 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
11936 * will be set to a generated NOTHING node.
11938 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
11939 * set to 0. *node_p will be set to a generated REG_ANY node.
11941 * The fourth possibility is that \N resolves to a sequence of more than one
11942 * code points. *cp_count will be set to the number of code points in the
11943 * sequence. *node_p * will be set to a generated node returned by this
11944 * function calling S_reg().
11946 * The final possibility is that it is premature to be calling this function;
11947 * that pass1 needs to be restarted. This can happen when this changes from
11948 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
11949 * latter occurs only when the fourth possibility would otherwise be in
11950 * effect, and is because one of those code points requires the pattern to be
11951 * recompiled as UTF-8. The function returns FALSE, and sets the
11952 * RESTART_PASS1 and NEED_UTF8 flags in *flagp, as appropriate. When this
11953 * happens, the caller needs to desist from continuing parsing, and return
11954 * this information to its caller. This is not set for when there is only one
11955 * code point, as this can be called as part of an ANYOF node, and they can
11956 * store above-Latin1 code points without the pattern having to be in UTF-8.
11958 * For non-single-quoted regexes, the tokenizer has resolved character and
11959 * sequence names inside \N{...} into their Unicode values, normalizing the
11960 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
11961 * hex-represented code points in the sequence. This is done there because
11962 * the names can vary based on what charnames pragma is in scope at the time,
11963 * so we need a way to take a snapshot of what they resolve to at the time of
11964 * the original parse. [perl #56444].
11966 * That parsing is skipped for single-quoted regexes, so we may here get
11967 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
11968 * parser. But if the single-quoted regex is something like '\N{U+41}', that
11969 * is legal and handled here. The code point is Unicode, and has to be
11970 * translated into the native character set for non-ASCII platforms.
11973 char * endbrace; /* points to '}' following the name */
11974 char *endchar; /* Points to '.' or '}' ending cur char in the input
11976 char* p = RExC_parse; /* Temporary */
11978 GET_RE_DEBUG_FLAGS_DECL;
11980 PERL_ARGS_ASSERT_GROK_BSLASH_N;
11982 GET_RE_DEBUG_FLAGS;
11984 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
11985 assert(! (node_p && cp_count)); /* At most 1 should be set */
11987 if (cp_count) { /* Initialize return for the most common case */
11991 /* The [^\n] meaning of \N ignores spaces and comments under the /x
11992 * modifier. The other meanings do not, so use a temporary until we find
11993 * out which we are being called with */
11994 skip_to_be_ignored_text(pRExC_state, &p,
11995 FALSE /* Don't force to /x */ );
11997 /* Disambiguate between \N meaning a named character versus \N meaning
11998 * [^\n]. The latter is assumed when the {...} following the \N is a legal
11999 * quantifier, or there is no '{' at all */
12000 if (*p != '{' || regcurly(p)) {
12010 *node_p = reg_node(pRExC_state, REG_ANY);
12011 *flagp |= HASWIDTH|SIMPLE;
12013 Set_Node_Length(*node_p, 1); /* MJD */
12017 /* Here, we have decided it should be a named character or sequence */
12019 /* The test above made sure that the next real character is a '{', but
12020 * under the /x modifier, it could be separated by space (or a comment and
12021 * \n) and this is not allowed (for consistency with \x{...} and the
12022 * tokenizer handling of \N{NAME}). */
12023 if (*RExC_parse != '{') {
12024 vFAIL("Missing braces on \\N{}");
12027 RExC_parse++; /* Skip past the '{' */
12029 if (! (endbrace = strchr(RExC_parse, '}'))) { /* no trailing brace */
12030 vFAIL2("Missing right brace on \\%c{}", 'N');
12032 else if(!(endbrace == RExC_parse /* nothing between the {} */
12033 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked... */
12034 && strnEQ(RExC_parse, "U+", 2)))) /* ... below for a better
12037 RExC_parse = endbrace; /* position msg's '<--HERE' */
12038 vFAIL("\\N{NAME} must be resolved by the lexer");
12041 REQUIRE_UNI_RULES(flagp, FALSE); /* Unicode named chars imply Unicode
12044 if (endbrace == RExC_parse) { /* empty: \N{} */
12046 RExC_parse++; /* Position after the "}" */
12047 vFAIL("Zero length \\N{}");
12052 nextchar(pRExC_state);
12057 *node_p = reg_node(pRExC_state,NOTHING);
12061 RExC_parse += 2; /* Skip past the 'U+' */
12063 /* Because toke.c has generated a special construct for us guaranteed not
12064 * to have NULs, we can use a str function */
12065 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12067 /* Code points are separated by dots. If none, there is only one code
12068 * point, and is terminated by the brace */
12070 if (endchar >= endbrace) {
12071 STRLEN length_of_hex;
12072 I32 grok_hex_flags;
12074 /* Here, exactly one code point. If that isn't what is wanted, fail */
12075 if (! code_point_p) {
12080 /* Convert code point from hex */
12081 length_of_hex = (STRLEN)(endchar - RExC_parse);
12082 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
12083 | PERL_SCAN_DISALLOW_PREFIX
12085 /* No errors in the first pass (See [perl
12086 * #122671].) We let the code below find the
12087 * errors when there are multiple chars. */
12089 ? PERL_SCAN_SILENT_ILLDIGIT
12092 /* This routine is the one place where both single- and double-quotish
12093 * \N{U+xxxx} are evaluated. The value is a Unicode code point which
12094 * must be converted to native. */
12095 *code_point_p = UNI_TO_NATIVE(grok_hex(RExC_parse,
12100 /* The tokenizer should have guaranteed validity, but it's possible to
12101 * bypass it by using single quoting, so check. Don't do the check
12102 * here when there are multiple chars; we do it below anyway. */
12103 if (length_of_hex == 0
12104 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
12106 RExC_parse += length_of_hex; /* Includes all the valid */
12107 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12108 ? UTF8SKIP(RExC_parse)
12110 /* Guard against malformed utf8 */
12111 if (RExC_parse >= endchar) {
12112 RExC_parse = endchar;
12114 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12117 RExC_parse = endbrace + 1;
12120 else { /* Is a multiple character sequence */
12121 SV * substitute_parse;
12123 char *orig_end = RExC_end;
12124 char *save_start = RExC_start;
12127 /* Count the code points, if desired, in the sequence */
12130 while (RExC_parse < endbrace) {
12131 /* Point to the beginning of the next character in the sequence. */
12132 RExC_parse = endchar + 1;
12133 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12138 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12139 * But don't backup up the pointer if the caller want to know how many
12140 * code points there are (they can then handle things) */
12148 /* What is done here is to convert this to a sub-pattern of the form
12149 * \x{char1}\x{char2}... and then call reg recursively to parse it
12150 * (enclosing in "(?: ... )" ). That way, it retains its atomicness,
12151 * while not having to worry about special handling that some code
12152 * points may have. */
12154 substitute_parse = newSVpvs("?:");
12156 while (RExC_parse < endbrace) {
12158 /* Convert to notation the rest of the code understands */
12159 sv_catpv(substitute_parse, "\\x{");
12160 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
12161 sv_catpv(substitute_parse, "}");
12163 /* Point to the beginning of the next character in the sequence. */
12164 RExC_parse = endchar + 1;
12165 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12168 sv_catpv(substitute_parse, ")");
12170 RExC_parse = RExC_start = RExC_adjusted_start = SvPV(substitute_parse,
12173 /* Don't allow empty number */
12174 if (len < (STRLEN) 8) {
12175 RExC_parse = endbrace;
12176 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12178 RExC_end = RExC_parse + len;
12180 /* The values are Unicode, and therefore not subject to recoding, but
12181 * have to be converted to native on a non-Unicode (meaning non-ASCII)
12184 RExC_recode_x_to_native = 1;
12188 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
12189 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12190 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12193 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#" UVxf,
12196 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12199 /* Restore the saved values */
12200 RExC_start = RExC_adjusted_start = save_start;
12201 RExC_parse = endbrace;
12202 RExC_end = orig_end;
12204 RExC_recode_x_to_native = 0;
12207 SvREFCNT_dec_NN(substitute_parse);
12208 nextchar(pRExC_state);
12215 PERL_STATIC_INLINE U8
12216 S_compute_EXACTish(RExC_state_t *pRExC_state)
12220 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
12228 op = get_regex_charset(RExC_flags);
12229 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
12230 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
12231 been, so there is no hole */
12234 return op + EXACTF;
12237 PERL_STATIC_INLINE void
12238 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
12239 regnode *node, I32* flagp, STRLEN len, UV code_point,
12242 /* This knows the details about sizing an EXACTish node, setting flags for
12243 * it (by setting <*flagp>, and potentially populating it with a single
12246 * If <len> (the length in bytes) is non-zero, this function assumes that
12247 * the node has already been populated, and just does the sizing. In this
12248 * case <code_point> should be the final code point that has already been
12249 * placed into the node. This value will be ignored except that under some
12250 * circumstances <*flagp> is set based on it.
12252 * If <len> is zero, the function assumes that the node is to contain only
12253 * the single character given by <code_point> and calculates what <len>
12254 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
12255 * additionally will populate the node's STRING with <code_point> or its
12258 * In both cases <*flagp> is appropriately set
12260 * It knows that under FOLD, the Latin Sharp S and UTF characters above
12261 * 255, must be folded (the former only when the rules indicate it can
12264 * When it does the populating, it looks at the flag 'downgradable'. If
12265 * true with a node that folds, it checks if the single code point
12266 * participates in a fold, and if not downgrades the node to an EXACT.
12267 * This helps the optimizer */
12269 bool len_passed_in = cBOOL(len != 0);
12270 U8 character[UTF8_MAXBYTES_CASE+1];
12272 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
12274 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
12275 * sizing difference, and is extra work that is thrown away */
12276 if (downgradable && ! PASS2) {
12277 downgradable = FALSE;
12280 if (! len_passed_in) {
12282 if (UVCHR_IS_INVARIANT(code_point)) {
12283 if (LOC || ! FOLD) { /* /l defers folding until runtime */
12284 *character = (U8) code_point;
12286 else { /* Here is /i and not /l. (toFOLD() is defined on just
12287 ASCII, which isn't the same thing as INVARIANT on
12288 EBCDIC, but it works there, as the extra invariants
12289 fold to themselves) */
12290 *character = toFOLD((U8) code_point);
12292 /* We can downgrade to an EXACT node if this character
12293 * isn't a folding one. Note that this assumes that
12294 * nothing above Latin1 folds to some other invariant than
12295 * one of these alphabetics; otherwise we would also have
12297 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12298 * || ASCII_FOLD_RESTRICTED))
12300 if (downgradable && PL_fold[code_point] == code_point) {
12306 else if (FOLD && (! LOC
12307 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
12308 { /* Folding, and ok to do so now */
12309 UV folded = _to_uni_fold_flags(
12313 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12314 ? FOLD_FLAGS_NOMIX_ASCII
12317 && folded == code_point /* This quickly rules out many
12318 cases, avoiding the
12319 _invlist_contains_cp() overhead
12321 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
12328 else if (code_point <= MAX_UTF8_TWO_BYTE) {
12330 /* Not folding this cp, and can output it directly */
12331 *character = UTF8_TWO_BYTE_HI(code_point);
12332 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
12336 uvchr_to_utf8( character, code_point);
12337 len = UTF8SKIP(character);
12339 } /* Else pattern isn't UTF8. */
12341 *character = (U8) code_point;
12343 } /* Else is folded non-UTF8 */
12344 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12345 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12346 || UNICODE_DOT_DOT_VERSION > 0)
12347 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
12351 /* We don't fold any non-UTF8 except possibly the Sharp s (see
12352 * comments at join_exact()); */
12353 *character = (U8) code_point;
12356 /* Can turn into an EXACT node if we know the fold at compile time,
12357 * and it folds to itself and doesn't particpate in other folds */
12360 && PL_fold_latin1[code_point] == code_point
12361 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12362 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
12366 } /* else is Sharp s. May need to fold it */
12367 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
12369 *(character + 1) = 's';
12373 *character = LATIN_SMALL_LETTER_SHARP_S;
12379 RExC_size += STR_SZ(len);
12382 RExC_emit += STR_SZ(len);
12383 STR_LEN(node) = len;
12384 if (! len_passed_in) {
12385 Copy((char *) character, STRING(node), len, char);
12389 *flagp |= HASWIDTH;
12391 /* A single character node is SIMPLE, except for the special-cased SHARP S
12393 if ((len == 1 || (UTF && len == UVCHR_SKIP(code_point)))
12394 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12395 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12396 || UNICODE_DOT_DOT_VERSION > 0)
12397 && ( code_point != LATIN_SMALL_LETTER_SHARP_S
12398 || ! FOLD || ! DEPENDS_SEMANTICS)
12404 /* The OP may not be well defined in PASS1 */
12405 if (PASS2 && OP(node) == EXACTFL) {
12406 RExC_contains_locale = 1;
12411 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
12412 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
12415 S_backref_value(char *p)
12417 const char* endptr;
12419 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
12426 - regatom - the lowest level
12428 Try to identify anything special at the start of the current parse position.
12429 If there is, then handle it as required. This may involve generating a
12430 single regop, such as for an assertion; or it may involve recursing, such as
12431 to handle a () structure.
12433 If the string doesn't start with something special then we gobble up
12434 as much literal text as we can. If we encounter a quantifier, we have to
12435 back off the final literal character, as that quantifier applies to just it
12436 and not to the whole string of literals.
12438 Once we have been able to handle whatever type of thing started the
12439 sequence, we return.
12441 Note: we have to be careful with escapes, as they can be both literal
12442 and special, and in the case of \10 and friends, context determines which.
12444 A summary of the code structure is:
12446 switch (first_byte) {
12447 cases for each special:
12448 handle this special;
12451 switch (2nd byte) {
12452 cases for each unambiguous special:
12453 handle this special;
12455 cases for each ambigous special/literal:
12457 if (special) handle here
12459 default: // unambiguously literal:
12462 default: // is a literal char
12465 create EXACTish node for literal;
12466 while (more input and node isn't full) {
12467 switch (input_byte) {
12468 cases for each special;
12469 make sure parse pointer is set so that the next call to
12470 regatom will see this special first
12471 goto loopdone; // EXACTish node terminated by prev. char
12473 append char to EXACTISH node;
12475 get next input byte;
12479 return the generated node;
12481 Specifically there are two separate switches for handling
12482 escape sequences, with the one for handling literal escapes requiring
12483 a dummy entry for all of the special escapes that are actually handled
12486 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
12488 Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
12489 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
12490 Otherwise does not return NULL.
12494 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12496 regnode *ret = NULL;
12503 GET_RE_DEBUG_FLAGS_DECL;
12505 *flagp = WORST; /* Tentatively. */
12507 DEBUG_PARSE("atom");
12509 PERL_ARGS_ASSERT_REGATOM;
12512 parse_start = RExC_parse;
12513 assert(RExC_parse < RExC_end);
12514 switch ((U8)*RExC_parse) {
12516 RExC_seen_zerolen++;
12517 nextchar(pRExC_state);
12518 if (RExC_flags & RXf_PMf_MULTILINE)
12519 ret = reg_node(pRExC_state, MBOL);
12521 ret = reg_node(pRExC_state, SBOL);
12522 Set_Node_Length(ret, 1); /* MJD */
12525 nextchar(pRExC_state);
12527 RExC_seen_zerolen++;
12528 if (RExC_flags & RXf_PMf_MULTILINE)
12529 ret = reg_node(pRExC_state, MEOL);
12531 ret = reg_node(pRExC_state, SEOL);
12532 Set_Node_Length(ret, 1); /* MJD */
12535 nextchar(pRExC_state);
12536 if (RExC_flags & RXf_PMf_SINGLELINE)
12537 ret = reg_node(pRExC_state, SANY);
12539 ret = reg_node(pRExC_state, REG_ANY);
12540 *flagp |= HASWIDTH|SIMPLE;
12542 Set_Node_Length(ret, 1); /* MJD */
12546 char * const oregcomp_parse = ++RExC_parse;
12547 ret = regclass(pRExC_state, flagp,depth+1,
12548 FALSE, /* means parse the whole char class */
12549 TRUE, /* allow multi-char folds */
12550 FALSE, /* don't silence non-portable warnings. */
12551 (bool) RExC_strict,
12552 TRUE, /* Allow an optimized regnode result */
12556 if (*flagp & (RESTART_PASS1|NEED_UTF8))
12558 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12561 if (*RExC_parse != ']') {
12562 RExC_parse = oregcomp_parse;
12563 vFAIL("Unmatched [");
12565 nextchar(pRExC_state);
12566 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
12570 nextchar(pRExC_state);
12571 ret = reg(pRExC_state, 2, &flags,depth+1);
12573 if (flags & TRYAGAIN) {
12574 if (RExC_parse >= RExC_end) {
12575 /* Make parent create an empty node if needed. */
12576 *flagp |= TRYAGAIN;
12581 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12582 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12585 FAIL2("panic: reg returned NULL to regatom, flags=%#" UVxf,
12588 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12592 if (flags & TRYAGAIN) {
12593 *flagp |= TRYAGAIN;
12596 vFAIL("Internal urp");
12597 /* Supposed to be caught earlier. */
12603 vFAIL("Quantifier follows nothing");
12608 This switch handles escape sequences that resolve to some kind
12609 of special regop and not to literal text. Escape sequnces that
12610 resolve to literal text are handled below in the switch marked
12613 Every entry in this switch *must* have a corresponding entry
12614 in the literal escape switch. However, the opposite is not
12615 required, as the default for this switch is to jump to the
12616 literal text handling code.
12619 switch ((U8)*RExC_parse) {
12620 /* Special Escapes */
12622 RExC_seen_zerolen++;
12623 ret = reg_node(pRExC_state, SBOL);
12624 /* SBOL is shared with /^/ so we set the flags so we can tell
12625 * /\A/ from /^/ in split. We check ret because first pass we
12626 * have no regop struct to set the flags on. */
12630 goto finish_meta_pat;
12632 ret = reg_node(pRExC_state, GPOS);
12633 RExC_seen |= REG_GPOS_SEEN;
12635 goto finish_meta_pat;
12637 RExC_seen_zerolen++;
12638 ret = reg_node(pRExC_state, KEEPS);
12640 /* XXX:dmq : disabling in-place substitution seems to
12641 * be necessary here to avoid cases of memory corruption, as
12642 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
12644 RExC_seen |= REG_LOOKBEHIND_SEEN;
12645 goto finish_meta_pat;
12647 ret = reg_node(pRExC_state, SEOL);
12649 RExC_seen_zerolen++; /* Do not optimize RE away */
12650 goto finish_meta_pat;
12652 ret = reg_node(pRExC_state, EOS);
12654 RExC_seen_zerolen++; /* Do not optimize RE away */
12655 goto finish_meta_pat;
12657 vFAIL("\\C no longer supported");
12659 ret = reg_node(pRExC_state, CLUMP);
12660 *flagp |= HASWIDTH;
12661 goto finish_meta_pat;
12667 arg = ANYOF_WORDCHAR;
12675 regex_charset charset = get_regex_charset(RExC_flags);
12677 RExC_seen_zerolen++;
12678 RExC_seen |= REG_LOOKBEHIND_SEEN;
12679 op = BOUND + charset;
12681 if (op == BOUNDL) {
12682 RExC_contains_locale = 1;
12685 ret = reg_node(pRExC_state, op);
12687 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
12688 FLAGS(ret) = TRADITIONAL_BOUND;
12689 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
12695 char name = *RExC_parse;
12698 endbrace = strchr(RExC_parse, '}');
12701 vFAIL2("Missing right brace on \\%c{}", name);
12703 /* XXX Need to decide whether to take spaces or not. Should be
12704 * consistent with \p{}, but that currently is SPACE, which
12705 * means vertical too, which seems wrong
12706 * while (isBLANK(*RExC_parse)) {
12709 if (endbrace == RExC_parse) {
12710 RExC_parse++; /* After the '}' */
12711 vFAIL2("Empty \\%c{}", name);
12713 length = endbrace - RExC_parse;
12714 /*while (isBLANK(*(RExC_parse + length - 1))) {
12717 switch (*RExC_parse) {
12720 && (length != 3 || strnNE(RExC_parse + 1, "cb", 2)))
12722 goto bad_bound_type;
12724 FLAGS(ret) = GCB_BOUND;
12727 if (length != 2 || *(RExC_parse + 1) != 'b') {
12728 goto bad_bound_type;
12730 FLAGS(ret) = LB_BOUND;
12733 if (length != 2 || *(RExC_parse + 1) != 'b') {
12734 goto bad_bound_type;
12736 FLAGS(ret) = SB_BOUND;
12739 if (length != 2 || *(RExC_parse + 1) != 'b') {
12740 goto bad_bound_type;
12742 FLAGS(ret) = WB_BOUND;
12746 RExC_parse = endbrace;
12748 "'%" UTF8f "' is an unknown bound type",
12749 UTF8fARG(UTF, length, endbrace - length));
12750 NOT_REACHED; /*NOTREACHED*/
12752 RExC_parse = endbrace;
12753 REQUIRE_UNI_RULES(flagp, NULL);
12755 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
12759 /* Don't have to worry about UTF-8, in this message because
12760 * to get here the contents of the \b must be ASCII */
12761 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
12762 "Using /u for '%.*s' instead of /%s",
12764 endbrace - length + 1,
12765 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
12766 ? ASCII_RESTRICT_PAT_MODS
12767 : ASCII_MORE_RESTRICT_PAT_MODS);
12771 if (PASS2 && invert) {
12772 OP(ret) += NBOUND - BOUND;
12774 goto finish_meta_pat;
12782 if (! DEPENDS_SEMANTICS) {
12786 /* \d doesn't have any matches in the upper Latin1 range, hence /d
12787 * is equivalent to /u. Changing to /u saves some branches at
12790 goto join_posix_op_known;
12793 ret = reg_node(pRExC_state, LNBREAK);
12794 *flagp |= HASWIDTH|SIMPLE;
12795 goto finish_meta_pat;
12803 goto join_posix_op_known;
12809 arg = ANYOF_VERTWS;
12811 goto join_posix_op_known;
12821 op = POSIXD + get_regex_charset(RExC_flags);
12822 if (op > POSIXA) { /* /aa is same as /a */
12825 else if (op == POSIXL) {
12826 RExC_contains_locale = 1;
12829 join_posix_op_known:
12832 op += NPOSIXD - POSIXD;
12835 ret = reg_node(pRExC_state, op);
12837 FLAGS(ret) = namedclass_to_classnum(arg);
12840 *flagp |= HASWIDTH|SIMPLE;
12844 nextchar(pRExC_state);
12845 Set_Node_Length(ret, 2); /* MJD */
12851 ret = regclass(pRExC_state, flagp,depth+1,
12852 TRUE, /* means just parse this element */
12853 FALSE, /* don't allow multi-char folds */
12854 FALSE, /* don't silence non-portable warnings. It
12855 would be a bug if these returned
12857 (bool) RExC_strict,
12858 TRUE, /* Allow an optimized regnode result */
12861 if (*flagp & RESTART_PASS1)
12863 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
12864 * multi-char folds are allowed. */
12866 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12871 Set_Node_Offset(ret, parse_start);
12872 Set_Node_Cur_Length(ret, parse_start - 2);
12873 nextchar(pRExC_state);
12876 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
12877 * \N{...} evaluates to a sequence of more than one code points).
12878 * The function call below returns a regnode, which is our result.
12879 * The parameters cause it to fail if the \N{} evaluates to a
12880 * single code point; we handle those like any other literal. The
12881 * reason that the multicharacter case is handled here and not as
12882 * part of the EXACtish code is because of quantifiers. In
12883 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
12884 * this way makes that Just Happen. dmq.
12885 * join_exact() will join this up with adjacent EXACTish nodes
12886 * later on, if appropriate. */
12888 if (grok_bslash_N(pRExC_state,
12889 &ret, /* Want a regnode returned */
12890 NULL, /* Fail if evaluates to a single code
12892 NULL, /* Don't need a count of how many code
12901 if (*flagp & RESTART_PASS1)
12904 /* Here, evaluates to a single code point. Go get that */
12905 RExC_parse = parse_start;
12908 case 'k': /* Handle \k<NAME> and \k'NAME' */
12912 if ( RExC_parse >= RExC_end - 1
12913 || (( ch = RExC_parse[1]) != '<'
12918 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12919 vFAIL2("Sequence %.2s... not terminated",parse_start);
12922 ret = handle_named_backref(pRExC_state,
12934 case '1': case '2': case '3': case '4':
12935 case '5': case '6': case '7': case '8': case '9':
12940 if (*RExC_parse == 'g') {
12944 if (*RExC_parse == '{') {
12948 if (*RExC_parse == '-') {
12952 if (hasbrace && !isDIGIT(*RExC_parse)) {
12953 if (isrel) RExC_parse--;
12955 goto parse_named_seq;
12958 if (RExC_parse >= RExC_end) {
12959 goto unterminated_g;
12961 num = S_backref_value(RExC_parse);
12963 vFAIL("Reference to invalid group 0");
12964 else if (num == I32_MAX) {
12965 if (isDIGIT(*RExC_parse))
12966 vFAIL("Reference to nonexistent group");
12969 vFAIL("Unterminated \\g... pattern");
12973 num = RExC_npar - num;
12975 vFAIL("Reference to nonexistent or unclosed group");
12979 num = S_backref_value(RExC_parse);
12980 /* bare \NNN might be backref or octal - if it is larger
12981 * than or equal RExC_npar then it is assumed to be an
12982 * octal escape. Note RExC_npar is +1 from the actual
12983 * number of parens. */
12984 /* Note we do NOT check if num == I32_MAX here, as that is
12985 * handled by the RExC_npar check */
12988 /* any numeric escape < 10 is always a backref */
12990 /* any numeric escape < RExC_npar is a backref */
12991 && num >= RExC_npar
12992 /* cannot be an octal escape if it starts with 8 */
12993 && *RExC_parse != '8'
12994 /* cannot be an octal escape it it starts with 9 */
12995 && *RExC_parse != '9'
12998 /* Probably not a backref, instead likely to be an
12999 * octal character escape, e.g. \35 or \777.
13000 * The above logic should make it obvious why using
13001 * octal escapes in patterns is problematic. - Yves */
13002 RExC_parse = parse_start;
13007 /* At this point RExC_parse points at a numeric escape like
13008 * \12 or \88 or something similar, which we should NOT treat
13009 * as an octal escape. It may or may not be a valid backref
13010 * escape. For instance \88888888 is unlikely to be a valid
13012 while (isDIGIT(*RExC_parse))
13015 if (*RExC_parse != '}')
13016 vFAIL("Unterminated \\g{...} pattern");
13020 if (num > (I32)RExC_rx->nparens)
13021 vFAIL("Reference to nonexistent group");
13024 ret = reganode(pRExC_state,
13027 : (ASCII_FOLD_RESTRICTED)
13029 : (AT_LEAST_UNI_SEMANTICS)
13035 *flagp |= HASWIDTH;
13037 /* override incorrect value set in reganode MJD */
13038 Set_Node_Offset(ret, parse_start);
13039 Set_Node_Cur_Length(ret, parse_start-1);
13040 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13041 FALSE /* Don't force to /x */ );
13045 if (RExC_parse >= RExC_end)
13046 FAIL("Trailing \\");
13049 /* Do not generate "unrecognized" warnings here, we fall
13050 back into the quick-grab loop below */
13051 RExC_parse = parse_start;
13053 } /* end of switch on a \foo sequence */
13058 /* '#' comments should have been spaced over before this function was
13060 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13062 if (RExC_flags & RXf_PMf_EXTENDED) {
13063 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13064 if (RExC_parse < RExC_end)
13074 /* Here, we have determined that the next thing is probably a
13075 * literal character. RExC_parse points to the first byte of its
13076 * definition. (It still may be an escape sequence that evaluates
13077 * to a single character) */
13083 #define MAX_NODE_STRING_SIZE 127
13084 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
13086 U8 upper_parse = MAX_NODE_STRING_SIZE;
13087 U8 node_type = compute_EXACTish(pRExC_state);
13088 bool next_is_quantifier;
13089 char * oldp = NULL;
13091 /* We can convert EXACTF nodes to EXACTFU if they contain only
13092 * characters that match identically regardless of the target
13093 * string's UTF8ness. The reason to do this is that EXACTF is not
13094 * trie-able, EXACTFU is.
13096 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13097 * contain only above-Latin1 characters (hence must be in UTF8),
13098 * which don't participate in folds with Latin1-range characters,
13099 * as the latter's folds aren't known until runtime. (We don't
13100 * need to figure this out until pass 2) */
13101 bool maybe_exactfu = PASS2
13102 && (node_type == EXACTF || node_type == EXACTFL);
13104 /* If a folding node contains only code points that don't
13105 * participate in folds, it can be changed into an EXACT node,
13106 * which allows the optimizer more things to look for */
13109 ret = reg_node(pRExC_state, node_type);
13111 /* In pass1, folded, we use a temporary buffer instead of the
13112 * actual node, as the node doesn't exist yet */
13113 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
13119 /* We look for the EXACTFish to EXACT node optimizaton only if
13120 * folding. (And we don't need to figure this out until pass 2).
13121 * XXX It might actually make sense to split the node into portions
13122 * that are exact and ones that aren't, so that we could later use
13123 * the exact ones to find the longest fixed and floating strings.
13124 * One would want to join them back into a larger node. One could
13125 * use a pseudo regnode like 'EXACT_ORIG_FOLD' */
13126 maybe_exact = FOLD && PASS2;
13128 /* XXX The node can hold up to 255 bytes, yet this only goes to
13129 * 127. I (khw) do not know why. Keeping it somewhat less than
13130 * 255 allows us to not have to worry about overflow due to
13131 * converting to utf8 and fold expansion, but that value is
13132 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
13133 * split up by this limit into a single one using the real max of
13134 * 255. Even at 127, this breaks under rare circumstances. If
13135 * folding, we do not want to split a node at a character that is a
13136 * non-final in a multi-char fold, as an input string could just
13137 * happen to want to match across the node boundary. The join
13138 * would solve that problem if the join actually happens. But a
13139 * series of more than two nodes in a row each of 127 would cause
13140 * the first join to succeed to get to 254, but then there wouldn't
13141 * be room for the next one, which could at be one of those split
13142 * multi-char folds. I don't know of any fool-proof solution. One
13143 * could back off to end with only a code point that isn't such a
13144 * non-final, but it is possible for there not to be any in the
13147 assert( ! UTF /* Is at the beginning of a character */
13148 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13149 || UTF8_IS_START(UCHARAT(RExC_parse)));
13151 /* Here, we have a literal character. Find the maximal string of
13152 * them in the input that we can fit into a single EXACTish node.
13153 * We quit at the first non-literal or when the node gets full */
13154 for (p = RExC_parse;
13155 len < upper_parse && p < RExC_end;
13160 /* White space has already been ignored */
13161 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13162 || ! is_PATWS_safe((p), RExC_end, UTF));
13174 /* Literal Escapes Switch
13176 This switch is meant to handle escape sequences that
13177 resolve to a literal character.
13179 Every escape sequence that represents something
13180 else, like an assertion or a char class, is handled
13181 in the switch marked 'Special Escapes' above in this
13182 routine, but also has an entry here as anything that
13183 isn't explicitly mentioned here will be treated as
13184 an unescaped equivalent literal.
13187 switch ((U8)*++p) {
13188 /* These are all the special escapes. */
13189 case 'A': /* Start assertion */
13190 case 'b': case 'B': /* Word-boundary assertion*/
13191 case 'C': /* Single char !DANGEROUS! */
13192 case 'd': case 'D': /* digit class */
13193 case 'g': case 'G': /* generic-backref, pos assertion */
13194 case 'h': case 'H': /* HORIZWS */
13195 case 'k': case 'K': /* named backref, keep marker */
13196 case 'p': case 'P': /* Unicode property */
13197 case 'R': /* LNBREAK */
13198 case 's': case 'S': /* space class */
13199 case 'v': case 'V': /* VERTWS */
13200 case 'w': case 'W': /* word class */
13201 case 'X': /* eXtended Unicode "combining
13202 character sequence" */
13203 case 'z': case 'Z': /* End of line/string assertion */
13207 /* Anything after here is an escape that resolves to a
13208 literal. (Except digits, which may or may not)
13214 case 'N': /* Handle a single-code point named character. */
13215 RExC_parse = p + 1;
13216 if (! grok_bslash_N(pRExC_state,
13217 NULL, /* Fail if evaluates to
13218 anything other than a
13219 single code point */
13220 &ender, /* The returned single code
13222 NULL, /* Don't need a count of
13223 how many code points */
13228 if (*flagp & NEED_UTF8)
13229 FAIL("panic: grok_bslash_N set NEED_UTF8");
13230 if (*flagp & RESTART_PASS1)
13233 /* Here, it wasn't a single code point. Go close
13234 * up this EXACTish node. The switch() prior to
13235 * this switch handles the other cases */
13236 RExC_parse = p = oldp;
13240 if (ender > 0xff) {
13241 REQUIRE_UTF8(flagp);
13257 ender = ESC_NATIVE;
13267 const char* error_msg;
13269 bool valid = grok_bslash_o(&p,
13272 PASS2, /* out warnings */
13273 (bool) RExC_strict,
13274 TRUE, /* Output warnings
13279 RExC_parse = p; /* going to die anyway; point
13280 to exact spot of failure */
13284 if (ender > 0xff) {
13285 REQUIRE_UTF8(flagp);
13291 UV result = UV_MAX; /* initialize to erroneous
13293 const char* error_msg;
13295 bool valid = grok_bslash_x(&p,
13298 PASS2, /* out warnings */
13299 (bool) RExC_strict,
13300 TRUE, /* Silence warnings
13305 RExC_parse = p; /* going to die anyway; point
13306 to exact spot of failure */
13311 if (ender < 0x100) {
13313 if (RExC_recode_x_to_native) {
13314 ender = LATIN1_TO_NATIVE(ender);
13319 REQUIRE_UTF8(flagp);
13325 ender = grok_bslash_c(*p++, PASS2);
13327 case '8': case '9': /* must be a backreference */
13329 /* we have an escape like \8 which cannot be an octal escape
13330 * so we exit the loop, and let the outer loop handle this
13331 * escape which may or may not be a legitimate backref. */
13333 case '1': case '2': case '3':case '4':
13334 case '5': case '6': case '7':
13335 /* When we parse backslash escapes there is ambiguity
13336 * between backreferences and octal escapes. Any escape
13337 * from \1 - \9 is a backreference, any multi-digit
13338 * escape which does not start with 0 and which when
13339 * evaluated as decimal could refer to an already
13340 * parsed capture buffer is a back reference. Anything
13343 * Note this implies that \118 could be interpreted as
13344 * 118 OR as "\11" . "8" depending on whether there
13345 * were 118 capture buffers defined already in the
13348 /* NOTE, RExC_npar is 1 more than the actual number of
13349 * parens we have seen so far, hence the < RExC_npar below. */
13351 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
13352 { /* Not to be treated as an octal constant, go
13360 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13362 ender = grok_oct(p, &numlen, &flags, NULL);
13363 if (ender > 0xff) {
13364 REQUIRE_UTF8(flagp);
13367 if (PASS2 /* like \08, \178 */
13369 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
13371 reg_warn_non_literal_string(
13373 form_short_octal_warning(p, numlen));
13379 FAIL("Trailing \\");
13382 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
13383 /* Include any left brace following the alpha to emphasize
13384 * that it could be part of an escape at some point
13386 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
13387 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
13389 goto normal_default;
13390 } /* End of switch on '\' */
13393 /* Currently we don't care if the lbrace is at the start
13394 * of a construct. This catches it in the middle of a
13395 * literal string, or when it's the first thing after
13396 * something like "\b" */
13397 if (len || (p > RExC_start && isALPHA_A(*(p -1)))) {
13398 RExC_parse = p + 1;
13399 vFAIL("Unescaped left brace in regex is illegal here");
13401 goto normal_default;
13404 if (PASS2 && p > RExC_parse && RExC_strict) {
13405 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
13408 default: /* A literal character */
13410 if (! UTF8_IS_INVARIANT(*p) && UTF) {
13412 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
13413 &numlen, UTF8_ALLOW_DEFAULT);
13419 } /* End of switch on the literal */
13421 /* Here, have looked at the literal character and <ender>
13422 * contains its ordinal, <p> points to the character after it.
13423 * We need to check if the next non-ignored thing is a
13424 * quantifier. Move <p> to after anything that should be
13425 * ignored, which, as a side effect, positions <p> for the next
13426 * loop iteration */
13427 skip_to_be_ignored_text(pRExC_state, &p,
13428 FALSE /* Don't force to /x */ );
13430 /* If the next thing is a quantifier, it applies to this
13431 * character only, which means that this character has to be in
13432 * its own node and can't just be appended to the string in an
13433 * existing node, so if there are already other characters in
13434 * the node, close the node with just them, and set up to do
13435 * this character again next time through, when it will be the
13436 * only thing in its new node */
13438 if ((next_is_quantifier = ( LIKELY(p < RExC_end)
13439 && UNLIKELY(ISMULT2(p))))
13446 /* Ready to add 'ender' to the node */
13448 if (! FOLD) { /* The simple case, just append the literal */
13450 /* In the sizing pass, we need only the size of the
13451 * character we are appending, hence we can delay getting
13452 * its representation until PASS2. */
13455 const STRLEN unilen = UVCHR_SKIP(ender);
13458 /* We have to subtract 1 just below (and again in
13459 * the corresponding PASS2 code) because the loop
13460 * increments <len> each time, as all but this path
13461 * (and one other) through it add a single byte to
13462 * the EXACTish node. But these paths would change
13463 * len to be the correct final value, so cancel out
13464 * the increment that follows */
13470 } else { /* PASS2 */
13473 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
13474 len += (char *) new_s - s - 1;
13475 s = (char *) new_s;
13478 *(s++) = (char) ender;
13482 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
13484 /* Here are folding under /l, and the code point is
13485 * problematic. First, we know we can't simplify things */
13486 maybe_exact = FALSE;
13487 maybe_exactfu = FALSE;
13489 /* A problematic code point in this context means that its
13490 * fold isn't known until runtime, so we can't fold it now.
13491 * (The non-problematic code points are the above-Latin1
13492 * ones that fold to also all above-Latin1. Their folds
13493 * don't vary no matter what the locale is.) But here we
13494 * have characters whose fold depends on the locale.
13495 * Unlike the non-folding case above, we have to keep track
13496 * of these in the sizing pass, so that we can make sure we
13497 * don't split too-long nodes in the middle of a potential
13498 * multi-char fold. And unlike the regular fold case
13499 * handled in the else clauses below, we don't actually
13500 * fold and don't have special cases to consider. What we
13501 * do for both passes is the PASS2 code for non-folding */
13502 goto not_fold_common;
13504 else /* A regular FOLD code point */
13506 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13507 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13508 || UNICODE_DOT_DOT_VERSION > 0)
13509 /* See comments for join_exact() as to why we fold
13510 * this non-UTF at compile time */
13511 || ( node_type == EXACTFU
13512 && ender == LATIN_SMALL_LETTER_SHARP_S)
13515 /* Here, are folding and are not UTF-8 encoded; therefore
13516 * the character must be in the range 0-255, and is not /l
13517 * (Not /l because we already handled these under /l in
13518 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
13519 if (IS_IN_SOME_FOLD_L1(ender)) {
13520 maybe_exact = FALSE;
13522 /* See if the character's fold differs between /d and
13523 * /u. This includes the multi-char fold SHARP S to
13525 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
13526 RExC_seen_unfolded_sharp_s = 1;
13527 maybe_exactfu = FALSE;
13529 else if (maybe_exactfu
13530 && (PL_fold[ender] != PL_fold_latin1[ender]
13531 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13532 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13533 || UNICODE_DOT_DOT_VERSION > 0)
13535 && isALPHA_FOLD_EQ(ender, 's')
13536 && isALPHA_FOLD_EQ(*(s-1), 's'))
13539 maybe_exactfu = FALSE;
13543 /* Even when folding, we store just the input character, as
13544 * we have an array that finds its fold quickly */
13545 *(s++) = (char) ender;
13547 else { /* FOLD, and UTF (or sharp s) */
13548 /* Unlike the non-fold case, we do actually have to
13549 * calculate the results here in pass 1. This is for two
13550 * reasons, the folded length may be longer than the
13551 * unfolded, and we have to calculate how many EXACTish
13552 * nodes it will take; and we may run out of room in a node
13553 * in the middle of a potential multi-char fold, and have
13554 * to back off accordingly. */
13557 if (isASCII_uni(ender)) {
13558 folded = toFOLD(ender);
13559 *(s)++ = (U8) folded;
13564 folded = _to_uni_fold_flags(
13568 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
13569 ? FOLD_FLAGS_NOMIX_ASCII
13573 /* The loop increments <len> each time, as all but this
13574 * path (and one other) through it add a single byte to
13575 * the EXACTish node. But this one has changed len to
13576 * be the correct final value, so subtract one to
13577 * cancel out the increment that follows */
13578 len += foldlen - 1;
13580 /* If this node only contains non-folding code points so
13581 * far, see if this new one is also non-folding */
13583 if (folded != ender) {
13584 maybe_exact = FALSE;
13587 /* Here the fold is the original; we have to check
13588 * further to see if anything folds to it */
13589 if (_invlist_contains_cp(PL_utf8_foldable,
13592 maybe_exact = FALSE;
13599 if (next_is_quantifier) {
13601 /* Here, the next input is a quantifier, and to get here,
13602 * the current character is the only one in the node.
13603 * Also, here <len> doesn't include the final byte for this
13609 } /* End of loop through literal characters */
13611 /* Here we have either exhausted the input or ran out of room in
13612 * the node. (If we encountered a character that can't be in the
13613 * node, transfer is made directly to <loopdone>, and so we
13614 * wouldn't have fallen off the end of the loop.) In the latter
13615 * case, we artificially have to split the node into two, because
13616 * we just don't have enough space to hold everything. This
13617 * creates a problem if the final character participates in a
13618 * multi-character fold in the non-final position, as a match that
13619 * should have occurred won't, due to the way nodes are matched,
13620 * and our artificial boundary. So back off until we find a non-
13621 * problematic character -- one that isn't at the beginning or
13622 * middle of such a fold. (Either it doesn't participate in any
13623 * folds, or appears only in the final position of all the folds it
13624 * does participate in.) A better solution with far fewer false
13625 * positives, and that would fill the nodes more completely, would
13626 * be to actually have available all the multi-character folds to
13627 * test against, and to back-off only far enough to be sure that
13628 * this node isn't ending with a partial one. <upper_parse> is set
13629 * further below (if we need to reparse the node) to include just
13630 * up through that final non-problematic character that this code
13631 * identifies, so when it is set to less than the full node, we can
13632 * skip the rest of this */
13633 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
13635 const STRLEN full_len = len;
13637 assert(len >= MAX_NODE_STRING_SIZE);
13639 /* Here, <s> points to the final byte of the final character.
13640 * Look backwards through the string until find a non-
13641 * problematic character */
13645 /* This has no multi-char folds to non-UTF characters */
13646 if (ASCII_FOLD_RESTRICTED) {
13650 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
13654 if (! PL_NonL1NonFinalFold) {
13655 PL_NonL1NonFinalFold = _new_invlist_C_array(
13656 NonL1_Perl_Non_Final_Folds_invlist);
13659 /* Point to the first byte of the final character */
13660 s = (char *) utf8_hop((U8 *) s, -1);
13662 while (s >= s0) { /* Search backwards until find
13663 non-problematic char */
13664 if (UTF8_IS_INVARIANT(*s)) {
13666 /* There are no ascii characters that participate
13667 * in multi-char folds under /aa. In EBCDIC, the
13668 * non-ascii invariants are all control characters,
13669 * so don't ever participate in any folds. */
13670 if (ASCII_FOLD_RESTRICTED
13671 || ! IS_NON_FINAL_FOLD(*s))
13676 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
13677 if (! IS_NON_FINAL_FOLD(EIGHT_BIT_UTF8_TO_NATIVE(
13683 else if (! _invlist_contains_cp(
13684 PL_NonL1NonFinalFold,
13685 valid_utf8_to_uvchr((U8 *) s, NULL)))
13690 /* Here, the current character is problematic in that
13691 * it does occur in the non-final position of some
13692 * fold, so try the character before it, but have to
13693 * special case the very first byte in the string, so
13694 * we don't read outside the string */
13695 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
13696 } /* End of loop backwards through the string */
13698 /* If there were only problematic characters in the string,
13699 * <s> will point to before s0, in which case the length
13700 * should be 0, otherwise include the length of the
13701 * non-problematic character just found */
13702 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
13705 /* Here, have found the final character, if any, that is
13706 * non-problematic as far as ending the node without splitting
13707 * it across a potential multi-char fold. <len> contains the
13708 * number of bytes in the node up-to and including that
13709 * character, or is 0 if there is no such character, meaning
13710 * the whole node contains only problematic characters. In
13711 * this case, give up and just take the node as-is. We can't
13716 /* If the node ends in an 's' we make sure it stays EXACTF,
13717 * as if it turns into an EXACTFU, it could later get
13718 * joined with another 's' that would then wrongly match
13720 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
13722 maybe_exactfu = FALSE;
13726 /* Here, the node does contain some characters that aren't
13727 * problematic. If one such is the final character in the
13728 * node, we are done */
13729 if (len == full_len) {
13732 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
13734 /* If the final character is problematic, but the
13735 * penultimate is not, back-off that last character to
13736 * later start a new node with it */
13741 /* Here, the final non-problematic character is earlier
13742 * in the input than the penultimate character. What we do
13743 * is reparse from the beginning, going up only as far as
13744 * this final ok one, thus guaranteeing that the node ends
13745 * in an acceptable character. The reason we reparse is
13746 * that we know how far in the character is, but we don't
13747 * know how to correlate its position with the input parse.
13748 * An alternate implementation would be to build that
13749 * correlation as we go along during the original parse,
13750 * but that would entail extra work for every node, whereas
13751 * this code gets executed only when the string is too
13752 * large for the node, and the final two characters are
13753 * problematic, an infrequent occurrence. Yet another
13754 * possible strategy would be to save the tail of the
13755 * string, and the next time regatom is called, initialize
13756 * with that. The problem with this is that unless you
13757 * back off one more character, you won't be guaranteed
13758 * regatom will get called again, unless regbranch,
13759 * regpiece ... are also changed. If you do back off that
13760 * extra character, so that there is input guaranteed to
13761 * force calling regatom, you can't handle the case where
13762 * just the first character in the node is acceptable. I
13763 * (khw) decided to try this method which doesn't have that
13764 * pitfall; if performance issues are found, we can do a
13765 * combination of the current approach plus that one */
13771 } /* End of verifying node ends with an appropriate char */
13773 loopdone: /* Jumped to when encounters something that shouldn't be
13776 /* I (khw) don't know if you can get here with zero length, but the
13777 * old code handled this situation by creating a zero-length EXACT
13778 * node. Might as well be NOTHING instead */
13784 /* If 'maybe_exact' is still set here, means there are no
13785 * code points in the node that participate in folds;
13786 * similarly for 'maybe_exactfu' and code points that match
13787 * differently depending on UTF8ness of the target string
13788 * (for /u), or depending on locale for /l */
13794 else if (maybe_exactfu) {
13800 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
13801 FALSE /* Don't look to see if could
13802 be turned into an EXACT
13803 node, as we have already
13808 RExC_parse = p - 1;
13809 Set_Node_Cur_Length(ret, parse_start);
13812 /* len is STRLEN which is unsigned, need to copy to signed */
13815 vFAIL("Internal disaster");
13818 } /* End of label 'defchar:' */
13820 } /* End of giant switch on input character */
13822 /* Position parse to next real character */
13823 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13824 FALSE /* Don't force to /x */ );
13825 if (PASS2 && *RExC_parse == '{' && OP(ret) != SBOL && ! regcurly(RExC_parse)) {
13826 ckWARNregdep(RExC_parse + 1, "Unescaped left brace in regex is deprecated here (and will be fatal in Perl 5.30), passed through");
13834 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
13836 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
13837 * sets up the bitmap and any flags, removing those code points from the
13838 * inversion list, setting it to NULL should it become completely empty */
13840 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
13841 assert(PL_regkind[OP(node)] == ANYOF);
13843 ANYOF_BITMAP_ZERO(node);
13844 if (*invlist_ptr) {
13846 /* This gets set if we actually need to modify things */
13847 bool change_invlist = FALSE;
13851 /* Start looking through *invlist_ptr */
13852 invlist_iterinit(*invlist_ptr);
13853 while (invlist_iternext(*invlist_ptr, &start, &end)) {
13857 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
13858 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
13861 /* Quit if are above what we should change */
13862 if (start >= NUM_ANYOF_CODE_POINTS) {
13866 change_invlist = TRUE;
13868 /* Set all the bits in the range, up to the max that we are doing */
13869 high = (end < NUM_ANYOF_CODE_POINTS - 1)
13871 : NUM_ANYOF_CODE_POINTS - 1;
13872 for (i = start; i <= (int) high; i++) {
13873 if (! ANYOF_BITMAP_TEST(node, i)) {
13874 ANYOF_BITMAP_SET(node, i);
13878 invlist_iterfinish(*invlist_ptr);
13880 /* Done with loop; remove any code points that are in the bitmap from
13881 * *invlist_ptr; similarly for code points above the bitmap if we have
13882 * a flag to match all of them anyways */
13883 if (change_invlist) {
13884 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
13886 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
13887 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
13890 /* If have completely emptied it, remove it completely */
13891 if (_invlist_len(*invlist_ptr) == 0) {
13892 SvREFCNT_dec_NN(*invlist_ptr);
13893 *invlist_ptr = NULL;
13898 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
13899 Character classes ([:foo:]) can also be negated ([:^foo:]).
13900 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
13901 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
13902 but trigger failures because they are currently unimplemented. */
13904 #define POSIXCC_DONE(c) ((c) == ':')
13905 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
13906 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
13907 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
13909 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
13910 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
13911 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
13913 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
13915 /* 'posix_warnings' and 'warn_text' are names of variables in the following
13917 #define ADD_POSIX_WARNING(p, text) STMT_START { \
13918 if (posix_warnings) { \
13919 if (! RExC_warn_text ) RExC_warn_text = (AV *) sv_2mortal((SV *) newAV()); \
13920 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
13924 REPORT_LOCATION_ARGS(p))); \
13929 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
13931 const char * const s, /* Where the putative posix class begins.
13932 Normally, this is one past the '['. This
13933 parameter exists so it can be somewhere
13934 besides RExC_parse. */
13935 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
13937 AV ** posix_warnings, /* Where to place any generated warnings, or
13939 const bool check_only /* Don't die if error */
13942 /* This parses what the caller thinks may be one of the three POSIX
13944 * 1) a character class, like [:blank:]
13945 * 2) a collating symbol, like [. .]
13946 * 3) an equivalence class, like [= =]
13947 * In the latter two cases, it croaks if it finds a syntactically legal
13948 * one, as these are not handled by Perl.
13950 * The main purpose is to look for a POSIX character class. It returns:
13951 * a) the class number
13952 * if it is a completely syntactically and semantically legal class.
13953 * 'updated_parse_ptr', if not NULL, is set to point to just after the
13954 * closing ']' of the class
13955 * b) OOB_NAMEDCLASS
13956 * if it appears that one of the three POSIX constructs was meant, but
13957 * its specification was somehow defective. 'updated_parse_ptr', if
13958 * not NULL, is set to point to the character just after the end
13959 * character of the class. See below for handling of warnings.
13960 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
13961 * if it doesn't appear that a POSIX construct was intended.
13962 * 'updated_parse_ptr' is not changed. No warnings nor errors are
13965 * In b) there may be errors or warnings generated. If 'check_only' is
13966 * TRUE, then any errors are discarded. Warnings are returned to the
13967 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
13968 * instead it is NULL, warnings are suppressed. This is done in all
13969 * passes. The reason for this is that the rest of the parsing is heavily
13970 * dependent on whether this routine found a valid posix class or not. If
13971 * it did, the closing ']' is absorbed as part of the class. If no class,
13972 * or an invalid one is found, any ']' will be considered the terminator of
13973 * the outer bracketed character class, leading to very different results.
13974 * In particular, a '(?[ ])' construct will likely have a syntax error if
13975 * the class is parsed other than intended, and this will happen in pass1,
13976 * before the warnings would normally be output. This mechanism allows the
13977 * caller to output those warnings in pass1 just before dieing, giving a
13978 * much better clue as to what is wrong.
13980 * The reason for this function, and its complexity is that a bracketed
13981 * character class can contain just about anything. But it's easy to
13982 * mistype the very specific posix class syntax but yielding a valid
13983 * regular bracketed class, so it silently gets compiled into something
13984 * quite unintended.
13986 * The solution adopted here maintains backward compatibility except that
13987 * it adds a warning if it looks like a posix class was intended but
13988 * improperly specified. The warning is not raised unless what is input
13989 * very closely resembles one of the 14 legal posix classes. To do this,
13990 * it uses fuzzy parsing. It calculates how many single-character edits it
13991 * would take to transform what was input into a legal posix class. Only
13992 * if that number is quite small does it think that the intention was a
13993 * posix class. Obviously these are heuristics, and there will be cases
13994 * where it errs on one side or another, and they can be tweaked as
13995 * experience informs.
13997 * The syntax for a legal posix class is:
13999 * qr/(?xa: \[ : \^? [:lower:]{4,6} : \] )/
14001 * What this routine considers syntactically to be an intended posix class
14002 * is this (the comments indicate some restrictions that the pattern
14005 * qr/(?x: \[? # The left bracket, possibly
14007 * \h* # possibly followed by blanks
14008 * (?: \^ \h* )? # possibly a misplaced caret
14009 * [:;]? # The opening class character,
14010 * # possibly omitted. A typo
14011 * # semi-colon can also be used.
14013 * \^? # possibly a correctly placed
14014 * # caret, but not if there was also
14015 * # a misplaced one
14017 * .{3,15} # The class name. If there are
14018 * # deviations from the legal syntax,
14019 * # its edit distance must be close
14020 * # to a real class name in order
14021 * # for it to be considered to be
14022 * # an intended posix class.
14024 * [:punct:]? # The closing class character,
14025 * # possibly omitted. If not a colon
14026 * # nor semi colon, the class name
14027 * # must be even closer to a valid
14030 * \]? # The right bracket, possibly
14034 * In the above, \h must be ASCII-only.
14036 * These are heuristics, and can be tweaked as field experience dictates.
14037 * There will be cases when someone didn't intend to specify a posix class
14038 * that this warns as being so. The goal is to minimize these, while
14039 * maximizing the catching of things intended to be a posix class that
14040 * aren't parsed as such.
14044 const char * const e = RExC_end;
14045 unsigned complement = 0; /* If to complement the class */
14046 bool found_problem = FALSE; /* Assume OK until proven otherwise */
14047 bool has_opening_bracket = FALSE;
14048 bool has_opening_colon = FALSE;
14049 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
14051 const char * possible_end = NULL; /* used for a 2nd parse pass */
14052 const char* name_start; /* ptr to class name first char */
14054 /* If the number of single-character typos the input name is away from a
14055 * legal name is no more than this number, it is considered to have meant
14056 * the legal name */
14057 int max_distance = 2;
14059 /* to store the name. The size determines the maximum length before we
14060 * decide that no posix class was intended. Should be at least
14061 * sizeof("alphanumeric") */
14064 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
14066 if (posix_warnings && RExC_warn_text)
14067 av_clear(RExC_warn_text);
14070 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14073 if (*(p - 1) != '[') {
14074 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
14075 found_problem = TRUE;
14078 has_opening_bracket = TRUE;
14081 /* They could be confused and think you can put spaces between the
14084 found_problem = TRUE;
14088 } while (p < e && isBLANK(*p));
14090 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14093 /* For [. .] and [= =]. These are quite different internally from [: :],
14094 * so they are handled separately. */
14095 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
14096 and 1 for at least one char in it
14099 const char open_char = *p;
14100 const char * temp_ptr = p + 1;
14102 /* These two constructs are not handled by perl, and if we find a
14103 * syntactically valid one, we croak. khw, who wrote this code, finds
14104 * this explanation of them very unclear:
14105 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
14106 * And searching the rest of the internet wasn't very helpful either.
14107 * It looks like just about any byte can be in these constructs,
14108 * depending on the locale. But unless the pattern is being compiled
14109 * under /l, which is very rare, Perl runs under the C or POSIX locale.
14110 * In that case, it looks like [= =] isn't allowed at all, and that
14111 * [. .] could be any single code point, but for longer strings the
14112 * constituent characters would have to be the ASCII alphabetics plus
14113 * the minus-hyphen. Any sensible locale definition would limit itself
14114 * to these. And any portable one definitely should. Trying to parse
14115 * the general case is a nightmare (see [perl #127604]). So, this code
14116 * looks only for interiors of these constructs that match:
14118 * Using \w relaxes the apparent rules a little, without adding much
14119 * danger of mistaking something else for one of these constructs.
14121 * [. .] in some implementations described on the internet is usable to
14122 * escape a character that otherwise is special in bracketed character
14123 * classes. For example [.].] means a literal right bracket instead of
14124 * the ending of the class
14126 * [= =] can legitimately contain a [. .] construct, but we don't
14127 * handle this case, as that [. .] construct will later get parsed
14128 * itself and croak then. And [= =] is checked for even when not under
14129 * /l, as Perl has long done so.
14131 * The code below relies on there being a trailing NUL, so it doesn't
14132 * have to keep checking if the parse ptr < e.
14134 if (temp_ptr[1] == open_char) {
14137 else while ( temp_ptr < e
14138 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
14143 if (*temp_ptr == open_char) {
14145 if (*temp_ptr == ']') {
14147 if (! found_problem && ! check_only) {
14148 RExC_parse = (char *) temp_ptr;
14149 vFAIL3("POSIX syntax [%c %c] is reserved for future "
14150 "extensions", open_char, open_char);
14153 /* Here, the syntax wasn't completely valid, or else the call
14154 * is to check-only */
14155 if (updated_parse_ptr) {
14156 *updated_parse_ptr = (char *) temp_ptr;
14159 return OOB_NAMEDCLASS;
14163 /* If we find something that started out to look like one of these
14164 * constructs, but isn't, we continue below so that it can be checked
14165 * for being a class name with a typo of '.' or '=' instead of a colon.
14169 /* Here, we think there is a possibility that a [: :] class was meant, and
14170 * we have the first real character. It could be they think the '^' comes
14173 found_problem = TRUE;
14174 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
14179 found_problem = TRUE;
14183 } while (p < e && isBLANK(*p));
14185 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14189 /* But the first character should be a colon, which they could have easily
14190 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
14191 * distinguish from a colon, so treat that as a colon). */
14194 has_opening_colon = TRUE;
14196 else if (*p == ';') {
14197 found_problem = TRUE;
14199 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14200 has_opening_colon = TRUE;
14203 found_problem = TRUE;
14204 ADD_POSIX_WARNING(p, "there must be a starting ':'");
14206 /* Consider an initial punctuation (not one of the recognized ones) to
14207 * be a left terminator */
14208 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
14213 /* They may think that you can put spaces between the components */
14215 found_problem = TRUE;
14219 } while (p < e && isBLANK(*p));
14221 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14226 /* We consider something like [^:^alnum:]] to not have been intended to
14227 * be a posix class, but XXX maybe we should */
14229 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14236 /* Again, they may think that you can put spaces between the components */
14238 found_problem = TRUE;
14242 } while (p < e && isBLANK(*p));
14244 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14249 /* XXX This ']' may be a typo, and something else was meant. But
14250 * treating it as such creates enough complications, that that
14251 * possibility isn't currently considered here. So we assume that the
14252 * ']' is what is intended, and if we've already found an initial '[',
14253 * this leaves this construct looking like [:] or [:^], which almost
14254 * certainly weren't intended to be posix classes */
14255 if (has_opening_bracket) {
14256 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14259 /* But this function can be called when we parse the colon for
14260 * something like qr/[alpha:]]/, so we back up to look for the
14265 found_problem = TRUE;
14266 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14268 else if (*p != ':') {
14270 /* XXX We are currently very restrictive here, so this code doesn't
14271 * consider the possibility that, say, /[alpha.]]/ was intended to
14272 * be a posix class. */
14273 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14276 /* Here we have something like 'foo:]'. There was no initial colon,
14277 * and we back up over 'foo. XXX Unlike the going forward case, we
14278 * don't handle typos of non-word chars in the middle */
14279 has_opening_colon = FALSE;
14282 while (p > RExC_start && isWORDCHAR(*p)) {
14287 /* Here, we have positioned ourselves to where we think the first
14288 * character in the potential class is */
14291 /* Now the interior really starts. There are certain key characters that
14292 * can end the interior, or these could just be typos. To catch both
14293 * cases, we may have to do two passes. In the first pass, we keep on
14294 * going unless we come to a sequence that matches
14295 * qr/ [[:punct:]] [[:blank:]]* \] /xa
14296 * This means it takes a sequence to end the pass, so two typos in a row if
14297 * that wasn't what was intended. If the class is perfectly formed, just
14298 * this one pass is needed. We also stop if there are too many characters
14299 * being accumulated, but this number is deliberately set higher than any
14300 * real class. It is set high enough so that someone who thinks that
14301 * 'alphanumeric' is a correct name would get warned that it wasn't.
14302 * While doing the pass, we keep track of where the key characters were in
14303 * it. If we don't find an end to the class, and one of the key characters
14304 * was found, we redo the pass, but stop when we get to that character.
14305 * Thus the key character was considered a typo in the first pass, but a
14306 * terminator in the second. If two key characters are found, we stop at
14307 * the second one in the first pass. Again this can miss two typos, but
14308 * catches a single one
14310 * In the first pass, 'possible_end' starts as NULL, and then gets set to
14311 * point to the first key character. For the second pass, it starts as -1.
14317 bool has_blank = FALSE;
14318 bool has_upper = FALSE;
14319 bool has_terminating_colon = FALSE;
14320 bool has_terminating_bracket = FALSE;
14321 bool has_semi_colon = FALSE;
14322 unsigned int name_len = 0;
14323 int punct_count = 0;
14327 /* Squeeze out blanks when looking up the class name below */
14328 if (isBLANK(*p) ) {
14330 found_problem = TRUE;
14335 /* The name will end with a punctuation */
14337 const char * peek = p + 1;
14339 /* Treat any non-']' punctuation followed by a ']' (possibly
14340 * with intervening blanks) as trying to terminate the class.
14341 * ']]' is very likely to mean a class was intended (but
14342 * missing the colon), but the warning message that gets
14343 * generated shows the error position better if we exit the
14344 * loop at the bottom (eventually), so skip it here. */
14346 if (peek < e && isBLANK(*peek)) {
14348 found_problem = TRUE;
14351 } while (peek < e && isBLANK(*peek));
14354 if (peek < e && *peek == ']') {
14355 has_terminating_bracket = TRUE;
14357 has_terminating_colon = TRUE;
14359 else if (*p == ';') {
14360 has_semi_colon = TRUE;
14361 has_terminating_colon = TRUE;
14364 found_problem = TRUE;
14371 /* Here we have punctuation we thought didn't end the class.
14372 * Keep track of the position of the key characters that are
14373 * more likely to have been class-enders */
14374 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
14376 /* Allow just one such possible class-ender not actually
14377 * ending the class. */
14378 if (possible_end) {
14384 /* If we have too many punctuation characters, no use in
14386 if (++punct_count > max_distance) {
14390 /* Treat the punctuation as a typo. */
14391 input_text[name_len++] = *p;
14394 else if (isUPPER(*p)) { /* Use lowercase for lookup */
14395 input_text[name_len++] = toLOWER(*p);
14397 found_problem = TRUE;
14399 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
14400 input_text[name_len++] = *p;
14404 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
14408 /* The declaration of 'input_text' is how long we allow a potential
14409 * class name to be, before saying they didn't mean a class name at
14411 if (name_len >= C_ARRAY_LENGTH(input_text)) {
14416 /* We get to here when the possible class name hasn't been properly
14417 * terminated before:
14418 * 1) we ran off the end of the pattern; or
14419 * 2) found two characters, each of which might have been intended to
14420 * be the name's terminator
14421 * 3) found so many punctuation characters in the purported name,
14422 * that the edit distance to a valid one is exceeded
14423 * 4) we decided it was more characters than anyone could have
14424 * intended to be one. */
14426 found_problem = TRUE;
14428 /* In the final two cases, we know that looking up what we've
14429 * accumulated won't lead to a match, even a fuzzy one. */
14430 if ( name_len >= C_ARRAY_LENGTH(input_text)
14431 || punct_count > max_distance)
14433 /* If there was an intermediate key character that could have been
14434 * an intended end, redo the parse, but stop there */
14435 if (possible_end && possible_end != (char *) -1) {
14436 possible_end = (char *) -1; /* Special signal value to say
14437 we've done a first pass */
14442 /* Otherwise, it can't have meant to have been a class */
14443 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14446 /* If we ran off the end, and the final character was a punctuation
14447 * one, back up one, to look at that final one just below. Later, we
14448 * will restore the parse pointer if appropriate */
14449 if (name_len && p == e && isPUNCT(*(p-1))) {
14454 if (p < e && isPUNCT(*p)) {
14456 has_terminating_bracket = TRUE;
14458 /* If this is a 2nd ']', and the first one is just below this
14459 * one, consider that to be the real terminator. This gives a
14460 * uniform and better positioning for the warning message */
14462 && possible_end != (char *) -1
14463 && *possible_end == ']'
14464 && name_len && input_text[name_len - 1] == ']')
14469 /* And this is actually equivalent to having done the 2nd
14470 * pass now, so set it to not try again */
14471 possible_end = (char *) -1;
14476 has_terminating_colon = TRUE;
14478 else if (*p == ';') {
14479 has_semi_colon = TRUE;
14480 has_terminating_colon = TRUE;
14488 /* Here, we have a class name to look up. We can short circuit the
14489 * stuff below for short names that can't possibly be meant to be a
14490 * class name. (We can do this on the first pass, as any second pass
14491 * will yield an even shorter name) */
14492 if (name_len < 3) {
14493 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14496 /* Find which class it is. Initially switch on the length of the name.
14498 switch (name_len) {
14500 if (memEQ(name_start, "word", 4)) {
14501 /* this is not POSIX, this is the Perl \w */
14502 class_number = ANYOF_WORDCHAR;
14506 /* Names all of length 5: alnum alpha ascii blank cntrl digit
14507 * graph lower print punct space upper
14508 * Offset 4 gives the best switch position. */
14509 switch (name_start[4]) {
14511 if (memEQ(name_start, "alph", 4)) /* alpha */
14512 class_number = ANYOF_ALPHA;
14515 if (memEQ(name_start, "spac", 4)) /* space */
14516 class_number = ANYOF_SPACE;
14519 if (memEQ(name_start, "grap", 4)) /* graph */
14520 class_number = ANYOF_GRAPH;
14523 if (memEQ(name_start, "asci", 4)) /* ascii */
14524 class_number = ANYOF_ASCII;
14527 if (memEQ(name_start, "blan", 4)) /* blank */
14528 class_number = ANYOF_BLANK;
14531 if (memEQ(name_start, "cntr", 4)) /* cntrl */
14532 class_number = ANYOF_CNTRL;
14535 if (memEQ(name_start, "alnu", 4)) /* alnum */
14536 class_number = ANYOF_ALPHANUMERIC;
14539 if (memEQ(name_start, "lowe", 4)) /* lower */
14540 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
14541 else if (memEQ(name_start, "uppe", 4)) /* upper */
14542 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
14545 if (memEQ(name_start, "digi", 4)) /* digit */
14546 class_number = ANYOF_DIGIT;
14547 else if (memEQ(name_start, "prin", 4)) /* print */
14548 class_number = ANYOF_PRINT;
14549 else if (memEQ(name_start, "punc", 4)) /* punct */
14550 class_number = ANYOF_PUNCT;
14555 if (memEQ(name_start, "xdigit", 6))
14556 class_number = ANYOF_XDIGIT;
14560 /* If the name exactly matches a posix class name the class number will
14561 * here be set to it, and the input almost certainly was meant to be a
14562 * posix class, so we can skip further checking. If instead the syntax
14563 * is exactly correct, but the name isn't one of the legal ones, we
14564 * will return that as an error below. But if neither of these apply,
14565 * it could be that no posix class was intended at all, or that one
14566 * was, but there was a typo. We tease these apart by doing fuzzy
14567 * matching on the name */
14568 if (class_number == OOB_NAMEDCLASS && found_problem) {
14569 const UV posix_names[][6] = {
14570 { 'a', 'l', 'n', 'u', 'm' },
14571 { 'a', 'l', 'p', 'h', 'a' },
14572 { 'a', 's', 'c', 'i', 'i' },
14573 { 'b', 'l', 'a', 'n', 'k' },
14574 { 'c', 'n', 't', 'r', 'l' },
14575 { 'd', 'i', 'g', 'i', 't' },
14576 { 'g', 'r', 'a', 'p', 'h' },
14577 { 'l', 'o', 'w', 'e', 'r' },
14578 { 'p', 'r', 'i', 'n', 't' },
14579 { 'p', 'u', 'n', 'c', 't' },
14580 { 's', 'p', 'a', 'c', 'e' },
14581 { 'u', 'p', 'p', 'e', 'r' },
14582 { 'w', 'o', 'r', 'd' },
14583 { 'x', 'd', 'i', 'g', 'i', 't' }
14585 /* The names of the above all have added NULs to make them the same
14586 * size, so we need to also have the real lengths */
14587 const UV posix_name_lengths[] = {
14588 sizeof("alnum") - 1,
14589 sizeof("alpha") - 1,
14590 sizeof("ascii") - 1,
14591 sizeof("blank") - 1,
14592 sizeof("cntrl") - 1,
14593 sizeof("digit") - 1,
14594 sizeof("graph") - 1,
14595 sizeof("lower") - 1,
14596 sizeof("print") - 1,
14597 sizeof("punct") - 1,
14598 sizeof("space") - 1,
14599 sizeof("upper") - 1,
14600 sizeof("word") - 1,
14601 sizeof("xdigit")- 1
14604 int temp_max = max_distance; /* Use a temporary, so if we
14605 reparse, we haven't changed the
14608 /* Use a smaller max edit distance if we are missing one of the
14610 if ( has_opening_bracket + has_opening_colon < 2
14611 || has_terminating_bracket + has_terminating_colon < 2)
14616 /* See if the input name is close to a legal one */
14617 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
14619 /* Short circuit call if the lengths are too far apart to be
14621 if (abs( (int) (name_len - posix_name_lengths[i]))
14627 if (edit_distance(input_text,
14630 posix_name_lengths[i],
14634 { /* If it is close, it probably was intended to be a class */
14635 goto probably_meant_to_be;
14639 /* Here the input name is not close enough to a valid class name
14640 * for us to consider it to be intended to be a posix class. If
14641 * we haven't already done so, and the parse found a character that
14642 * could have been terminators for the name, but which we absorbed
14643 * as typos during the first pass, repeat the parse, signalling it
14644 * to stop at that character */
14645 if (possible_end && possible_end != (char *) -1) {
14646 possible_end = (char *) -1;
14651 /* Here neither pass found a close-enough class name */
14652 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14655 probably_meant_to_be:
14657 /* Here we think that a posix specification was intended. Update any
14659 if (updated_parse_ptr) {
14660 *updated_parse_ptr = (char *) p;
14663 /* If a posix class name was intended but incorrectly specified, we
14664 * output or return the warnings */
14665 if (found_problem) {
14667 /* We set flags for these issues in the parse loop above instead of
14668 * adding them to the list of warnings, because we can parse it
14669 * twice, and we only want one warning instance */
14671 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
14674 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14676 if (has_semi_colon) {
14677 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14679 else if (! has_terminating_colon) {
14680 ADD_POSIX_WARNING(p, "there is no terminating ':'");
14682 if (! has_terminating_bracket) {
14683 ADD_POSIX_WARNING(p, "there is no terminating ']'");
14686 if (posix_warnings && RExC_warn_text && av_top_index(RExC_warn_text) > -1) {
14687 *posix_warnings = RExC_warn_text;
14690 else if (class_number != OOB_NAMEDCLASS) {
14691 /* If it is a known class, return the class. The class number
14692 * #defines are structured so each complement is +1 to the normal
14694 return class_number + complement;
14696 else if (! check_only) {
14698 /* Here, it is an unrecognized class. This is an error (unless the
14699 * call is to check only, which we've already handled above) */
14700 const char * const complement_string = (complement)
14703 RExC_parse = (char *) p;
14704 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
14706 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
14710 return OOB_NAMEDCLASS;
14712 #undef ADD_POSIX_WARNING
14714 STATIC unsigned int
14715 S_regex_set_precedence(const U8 my_operator) {
14717 /* Returns the precedence in the (?[...]) construct of the input operator,
14718 * specified by its character representation. The precedence follows
14719 * general Perl rules, but it extends this so that ')' and ']' have (low)
14720 * precedence even though they aren't really operators */
14722 switch (my_operator) {
14738 NOT_REACHED; /* NOTREACHED */
14739 return 0; /* Silence compiler warning */
14743 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
14744 I32 *flagp, U32 depth,
14745 char * const oregcomp_parse)
14747 /* Handle the (?[...]) construct to do set operations */
14749 U8 curchar; /* Current character being parsed */
14750 UV start, end; /* End points of code point ranges */
14751 SV* final = NULL; /* The end result inversion list */
14752 SV* result_string; /* 'final' stringified */
14753 AV* stack; /* stack of operators and operands not yet
14755 AV* fence_stack = NULL; /* A stack containing the positions in
14756 'stack' of where the undealt-with left
14757 parens would be if they were actually
14759 /* The 'VOL' (expanding to 'volatile') is a workaround for an optimiser bug
14760 * in Solaris Studio 12.3. See RT #127455 */
14761 VOL IV fence = 0; /* Position of where most recent undealt-
14762 with left paren in stack is; -1 if none.
14764 STRLEN len; /* Temporary */
14765 regnode* node; /* Temporary, and final regnode returned by
14767 const bool save_fold = FOLD; /* Temporary */
14768 char *save_end, *save_parse; /* Temporaries */
14769 const bool in_locale = LOC; /* we turn off /l during processing */
14770 AV* posix_warnings = NULL;
14772 GET_RE_DEBUG_FLAGS_DECL;
14774 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
14777 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
14780 REQUIRE_UNI_RULES(flagp, NULL); /* The use of this operator implies /u.
14781 This is required so that the compile
14782 time values are valid in all runtime
14785 /* This will return only an ANYOF regnode, or (unlikely) something smaller
14786 * (such as EXACT). Thus we can skip most everything if just sizing. We
14787 * call regclass to handle '[]' so as to not have to reinvent its parsing
14788 * rules here (throwing away the size it computes each time). And, we exit
14789 * upon an unescaped ']' that isn't one ending a regclass. To do both
14790 * these things, we need to realize that something preceded by a backslash
14791 * is escaped, so we have to keep track of backslashes */
14793 UV depth = 0; /* how many nested (?[...]) constructs */
14795 while (RExC_parse < RExC_end) {
14796 SV* current = NULL;
14798 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14799 TRUE /* Force /x */ );
14801 switch (*RExC_parse) {
14803 if (RExC_parse[1] == '[') depth++, RExC_parse++;
14808 /* Skip past this, so the next character gets skipped, after
14811 if (*RExC_parse == 'c') {
14812 /* Skip the \cX notation for control characters */
14813 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14819 /* See if this is a [:posix:] class. */
14820 bool is_posix_class = (OOB_NAMEDCLASS
14821 < handle_possible_posix(pRExC_state,
14825 TRUE /* checking only */));
14826 /* If it is a posix class, leave the parse pointer at the
14827 * '[' to fool regclass() into thinking it is part of a
14828 * '[[:posix:]]'. */
14829 if (! is_posix_class) {
14833 /* regclass() can only return RESTART_PASS1 and NEED_UTF8
14834 * if multi-char folds are allowed. */
14835 if (!regclass(pRExC_state, flagp,depth+1,
14836 is_posix_class, /* parse the whole char
14837 class only if not a
14839 FALSE, /* don't allow multi-char folds */
14840 TRUE, /* silence non-portable warnings. */
14842 FALSE, /* Require return to be an ANYOF */
14846 FAIL2("panic: regclass returned NULL to handle_sets, "
14847 "flags=%#" UVxf, (UV) *flagp);
14849 /* function call leaves parse pointing to the ']', except
14850 * if we faked it */
14851 if (is_posix_class) {
14855 SvREFCNT_dec(current); /* In case it returned something */
14860 if (depth--) break;
14862 if (*RExC_parse == ')') {
14863 node = reganode(pRExC_state, ANYOF, 0);
14864 RExC_size += ANYOF_SKIP;
14865 nextchar(pRExC_state);
14866 Set_Node_Length(node,
14867 RExC_parse - oregcomp_parse + 1); /* MJD */
14869 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
14877 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14881 /* We output the messages even if warnings are off, because we'll fail
14882 * the very next thing, and these give a likely diagnosis for that */
14883 if (posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
14884 output_or_return_posix_warnings(pRExC_state, posix_warnings, NULL);
14887 FAIL("Syntax error in (?[...])");
14890 /* Pass 2 only after this. */
14891 Perl_ck_warner_d(aTHX_
14892 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
14893 "The regex_sets feature is experimental" REPORT_LOCATION,
14894 REPORT_LOCATION_ARGS(RExC_parse));
14896 /* Everything in this construct is a metacharacter. Operands begin with
14897 * either a '\' (for an escape sequence), or a '[' for a bracketed
14898 * character class. Any other character should be an operator, or
14899 * parenthesis for grouping. Both types of operands are handled by calling
14900 * regclass() to parse them. It is called with a parameter to indicate to
14901 * return the computed inversion list. The parsing here is implemented via
14902 * a stack. Each entry on the stack is a single character representing one
14903 * of the operators; or else a pointer to an operand inversion list. */
14905 #define IS_OPERATOR(a) SvIOK(a)
14906 #define IS_OPERAND(a) (! IS_OPERATOR(a))
14908 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
14909 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
14910 * with pronouncing it called it Reverse Polish instead, but now that YOU
14911 * know how to pronounce it you can use the correct term, thus giving due
14912 * credit to the person who invented it, and impressing your geek friends.
14913 * Wikipedia says that the pronounciation of "Ł" has been changing so that
14914 * it is now more like an English initial W (as in wonk) than an L.)
14916 * This means that, for example, 'a | b & c' is stored on the stack as
14924 * where the numbers in brackets give the stack [array] element number.
14925 * In this implementation, parentheses are not stored on the stack.
14926 * Instead a '(' creates a "fence" so that the part of the stack below the
14927 * fence is invisible except to the corresponding ')' (this allows us to
14928 * replace testing for parens, by using instead subtraction of the fence
14929 * position). As new operands are processed they are pushed onto the stack
14930 * (except as noted in the next paragraph). New operators of higher
14931 * precedence than the current final one are inserted on the stack before
14932 * the lhs operand (so that when the rhs is pushed next, everything will be
14933 * in the correct positions shown above. When an operator of equal or
14934 * lower precedence is encountered in parsing, all the stacked operations
14935 * of equal or higher precedence are evaluated, leaving the result as the
14936 * top entry on the stack. This makes higher precedence operations
14937 * evaluate before lower precedence ones, and causes operations of equal
14938 * precedence to left associate.
14940 * The only unary operator '!' is immediately pushed onto the stack when
14941 * encountered. When an operand is encountered, if the top of the stack is
14942 * a '!", the complement is immediately performed, and the '!' popped. The
14943 * resulting value is treated as a new operand, and the logic in the
14944 * previous paragraph is executed. Thus in the expression
14946 * the stack looks like
14952 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
14959 * A ')' is treated as an operator with lower precedence than all the
14960 * aforementioned ones, which causes all operations on the stack above the
14961 * corresponding '(' to be evaluated down to a single resultant operand.
14962 * Then the fence for the '(' is removed, and the operand goes through the
14963 * algorithm above, without the fence.
14965 * A separate stack is kept of the fence positions, so that the position of
14966 * the latest so-far unbalanced '(' is at the top of it.
14968 * The ']' ending the construct is treated as the lowest operator of all,
14969 * so that everything gets evaluated down to a single operand, which is the
14972 sv_2mortal((SV *)(stack = newAV()));
14973 sv_2mortal((SV *)(fence_stack = newAV()));
14975 while (RExC_parse < RExC_end) {
14976 I32 top_index; /* Index of top-most element in 'stack' */
14977 SV** top_ptr; /* Pointer to top 'stack' element */
14978 SV* current = NULL; /* To contain the current inversion list
14980 SV* only_to_avoid_leaks;
14982 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14983 TRUE /* Force /x */ );
14984 if (RExC_parse >= RExC_end) {
14985 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
14988 curchar = UCHARAT(RExC_parse);
14992 #ifdef ENABLE_REGEX_SETS_DEBUGGING
14993 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
14994 DEBUG_U(dump_regex_sets_structures(pRExC_state,
14995 stack, fence, fence_stack));
14998 top_index = av_tindex_skip_len_mg(stack);
15001 SV** stacked_ptr; /* Ptr to something already on 'stack' */
15002 char stacked_operator; /* The topmost operator on the 'stack'. */
15003 SV* lhs; /* Operand to the left of the operator */
15004 SV* rhs; /* Operand to the right of the operator */
15005 SV* fence_ptr; /* Pointer to top element of the fence
15010 if ( RExC_parse < RExC_end - 1
15011 && (UCHARAT(RExC_parse + 1) == '?'))
15013 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
15014 * This happens when we have some thing like
15016 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
15018 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
15020 * Here we would be handling the interpolated
15021 * '$thai_or_lao'. We handle this by a recursive call to
15022 * ourselves which returns the inversion list the
15023 * interpolated expression evaluates to. We use the flags
15024 * from the interpolated pattern. */
15025 U32 save_flags = RExC_flags;
15026 const char * save_parse;
15028 RExC_parse += 2; /* Skip past the '(?' */
15029 save_parse = RExC_parse;
15031 /* Parse any flags for the '(?' */
15032 parse_lparen_question_flags(pRExC_state);
15034 if (RExC_parse == save_parse /* Makes sure there was at
15035 least one flag (or else
15036 this embedding wasn't
15038 || RExC_parse >= RExC_end - 4
15039 || UCHARAT(RExC_parse) != ':'
15040 || UCHARAT(++RExC_parse) != '('
15041 || UCHARAT(++RExC_parse) != '?'
15042 || UCHARAT(++RExC_parse) != '[')
15045 /* In combination with the above, this moves the
15046 * pointer to the point just after the first erroneous
15047 * character (or if there are no flags, to where they
15048 * should have been) */
15049 if (RExC_parse >= RExC_end - 4) {
15050 RExC_parse = RExC_end;
15052 else if (RExC_parse != save_parse) {
15053 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15055 vFAIL("Expecting '(?flags:(?[...'");
15058 /* Recurse, with the meat of the embedded expression */
15060 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
15061 depth+1, oregcomp_parse);
15063 /* Here, 'current' contains the embedded expression's
15064 * inversion list, and RExC_parse points to the trailing
15065 * ']'; the next character should be the ')' */
15067 assert(UCHARAT(RExC_parse) == ')');
15069 /* Then the ')' matching the original '(' handled by this
15070 * case: statement */
15072 assert(UCHARAT(RExC_parse) == ')');
15075 RExC_flags = save_flags;
15076 goto handle_operand;
15079 /* A regular '('. Look behind for illegal syntax */
15080 if (top_index - fence >= 0) {
15081 /* If the top entry on the stack is an operator, it had
15082 * better be a '!', otherwise the entry below the top
15083 * operand should be an operator */
15084 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
15085 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
15086 || ( IS_OPERAND(*top_ptr)
15087 && ( top_index - fence < 1
15088 || ! (stacked_ptr = av_fetch(stack,
15091 || ! IS_OPERATOR(*stacked_ptr))))
15094 vFAIL("Unexpected '(' with no preceding operator");
15098 /* Stack the position of this undealt-with left paren */
15099 av_push(fence_stack, newSViv(fence));
15100 fence = top_index + 1;
15104 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15105 * multi-char folds are allowed. */
15106 if (!regclass(pRExC_state, flagp,depth+1,
15107 TRUE, /* means parse just the next thing */
15108 FALSE, /* don't allow multi-char folds */
15109 FALSE, /* don't silence non-portable warnings. */
15111 FALSE, /* Require return to be an ANYOF */
15115 FAIL2("panic: regclass returned NULL to handle_sets, "
15116 "flags=%#" UVxf, (UV) *flagp);
15119 /* regclass() will return with parsing just the \ sequence,
15120 * leaving the parse pointer at the next thing to parse */
15122 goto handle_operand;
15124 case '[': /* Is a bracketed character class */
15126 /* See if this is a [:posix:] class. */
15127 bool is_posix_class = (OOB_NAMEDCLASS
15128 < handle_possible_posix(pRExC_state,
15132 TRUE /* checking only */));
15133 /* If it is a posix class, leave the parse pointer at the '['
15134 * to fool regclass() into thinking it is part of a
15135 * '[[:posix:]]'. */
15136 if (! is_posix_class) {
15140 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15141 * multi-char folds are allowed. */
15142 if (!regclass(pRExC_state, flagp,depth+1,
15143 is_posix_class, /* parse the whole char
15144 class only if not a
15146 FALSE, /* don't allow multi-char folds */
15147 TRUE, /* silence non-portable warnings. */
15149 FALSE, /* Require return to be an ANYOF */
15154 FAIL2("panic: regclass returned NULL to handle_sets, "
15155 "flags=%#" UVxf, (UV) *flagp);
15158 /* function call leaves parse pointing to the ']', except if we
15160 if (is_posix_class) {
15164 goto handle_operand;
15168 if (top_index >= 1) {
15169 goto join_operators;
15172 /* Only a single operand on the stack: are done */
15176 if (av_tindex_skip_len_mg(fence_stack) < 0) {
15178 vFAIL("Unexpected ')'");
15181 /* If nothing after the fence, is missing an operand */
15182 if (top_index - fence < 0) {
15186 /* If at least two things on the stack, treat this as an
15188 if (top_index - fence >= 1) {
15189 goto join_operators;
15192 /* Here only a single thing on the fenced stack, and there is a
15193 * fence. Get rid of it */
15194 fence_ptr = av_pop(fence_stack);
15196 fence = SvIV(fence_ptr) - 1;
15197 SvREFCNT_dec_NN(fence_ptr);
15204 /* Having gotten rid of the fence, we pop the operand at the
15205 * stack top and process it as a newly encountered operand */
15206 current = av_pop(stack);
15207 if (IS_OPERAND(current)) {
15208 goto handle_operand;
15220 /* These binary operators should have a left operand already
15222 if ( top_index - fence < 0
15223 || top_index - fence == 1
15224 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
15225 || ! IS_OPERAND(*top_ptr))
15227 goto unexpected_binary;
15230 /* If only the one operand is on the part of the stack visible
15231 * to us, we just place this operator in the proper position */
15232 if (top_index - fence < 2) {
15234 /* Place the operator before the operand */
15236 SV* lhs = av_pop(stack);
15237 av_push(stack, newSVuv(curchar));
15238 av_push(stack, lhs);
15242 /* But if there is something else on the stack, we need to
15243 * process it before this new operator if and only if the
15244 * stacked operation has equal or higher precedence than the
15249 /* The operator on the stack is supposed to be below both its
15251 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
15252 || IS_OPERAND(*stacked_ptr))
15254 /* But if not, it's legal and indicates we are completely
15255 * done if and only if we're currently processing a ']',
15256 * which should be the final thing in the expression */
15257 if (curchar == ']') {
15263 vFAIL2("Unexpected binary operator '%c' with no "
15264 "preceding operand", curchar);
15266 stacked_operator = (char) SvUV(*stacked_ptr);
15268 if (regex_set_precedence(curchar)
15269 > regex_set_precedence(stacked_operator))
15271 /* Here, the new operator has higher precedence than the
15272 * stacked one. This means we need to add the new one to
15273 * the stack to await its rhs operand (and maybe more
15274 * stuff). We put it before the lhs operand, leaving
15275 * untouched the stacked operator and everything below it
15277 lhs = av_pop(stack);
15278 assert(IS_OPERAND(lhs));
15280 av_push(stack, newSVuv(curchar));
15281 av_push(stack, lhs);
15285 /* Here, the new operator has equal or lower precedence than
15286 * what's already there. This means the operation already
15287 * there should be performed now, before the new one. */
15289 rhs = av_pop(stack);
15290 if (! IS_OPERAND(rhs)) {
15292 /* This can happen when a ! is not followed by an operand,
15293 * like in /(?[\t &!])/ */
15297 lhs = av_pop(stack);
15299 if (! IS_OPERAND(lhs)) {
15301 /* This can happen when there is an empty (), like in
15302 * /(?[[0]+()+])/ */
15306 switch (stacked_operator) {
15308 _invlist_intersection(lhs, rhs, &rhs);
15313 _invlist_union(lhs, rhs, &rhs);
15317 _invlist_subtract(lhs, rhs, &rhs);
15320 case '^': /* The union minus the intersection */
15325 _invlist_union(lhs, rhs, &u);
15326 _invlist_intersection(lhs, rhs, &i);
15327 _invlist_subtract(u, i, &rhs);
15328 SvREFCNT_dec_NN(i);
15329 SvREFCNT_dec_NN(u);
15335 /* Here, the higher precedence operation has been done, and the
15336 * result is in 'rhs'. We overwrite the stacked operator with
15337 * the result. Then we redo this code to either push the new
15338 * operator onto the stack or perform any higher precedence
15339 * stacked operation */
15340 only_to_avoid_leaks = av_pop(stack);
15341 SvREFCNT_dec(only_to_avoid_leaks);
15342 av_push(stack, rhs);
15345 case '!': /* Highest priority, right associative */
15347 /* If what's already at the top of the stack is another '!",
15348 * they just cancel each other out */
15349 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
15350 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
15352 only_to_avoid_leaks = av_pop(stack);
15353 SvREFCNT_dec(only_to_avoid_leaks);
15355 else { /* Otherwise, since it's right associative, just push
15357 av_push(stack, newSVuv(curchar));
15362 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15363 vFAIL("Unexpected character");
15367 /* Here 'current' is the operand. If something is already on the
15368 * stack, we have to check if it is a !. But first, the code above
15369 * may have altered the stack in the time since we earlier set
15372 top_index = av_tindex_skip_len_mg(stack);
15373 if (top_index - fence >= 0) {
15374 /* If the top entry on the stack is an operator, it had better
15375 * be a '!', otherwise the entry below the top operand should
15376 * be an operator */
15377 top_ptr = av_fetch(stack, top_index, FALSE);
15379 if (IS_OPERATOR(*top_ptr)) {
15381 /* The only permissible operator at the top of the stack is
15382 * '!', which is applied immediately to this operand. */
15383 curchar = (char) SvUV(*top_ptr);
15384 if (curchar != '!') {
15385 SvREFCNT_dec(current);
15386 vFAIL2("Unexpected binary operator '%c' with no "
15387 "preceding operand", curchar);
15390 _invlist_invert(current);
15392 only_to_avoid_leaks = av_pop(stack);
15393 SvREFCNT_dec(only_to_avoid_leaks);
15395 /* And we redo with the inverted operand. This allows
15396 * handling multiple ! in a row */
15397 goto handle_operand;
15399 /* Single operand is ok only for the non-binary ')'
15401 else if ((top_index - fence == 0 && curchar != ')')
15402 || (top_index - fence > 0
15403 && (! (stacked_ptr = av_fetch(stack,
15406 || IS_OPERAND(*stacked_ptr))))
15408 SvREFCNT_dec(current);
15409 vFAIL("Operand with no preceding operator");
15413 /* Here there was nothing on the stack or the top element was
15414 * another operand. Just add this new one */
15415 av_push(stack, current);
15417 } /* End of switch on next parse token */
15419 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15420 } /* End of loop parsing through the construct */
15423 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
15424 vFAIL("Unmatched (");
15427 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
15428 || ((final = av_pop(stack)) == NULL)
15429 || ! IS_OPERAND(final)
15430 || SvTYPE(final) != SVt_INVLIST
15431 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
15434 SvREFCNT_dec(final);
15435 vFAIL("Incomplete expression within '(?[ ])'");
15438 /* Here, 'final' is the resultant inversion list from evaluating the
15439 * expression. Return it if so requested */
15440 if (return_invlist) {
15441 *return_invlist = final;
15445 /* Otherwise generate a resultant node, based on 'final'. regclass() is
15446 * expecting a string of ranges and individual code points */
15447 invlist_iterinit(final);
15448 result_string = newSVpvs("");
15449 while (invlist_iternext(final, &start, &end)) {
15450 if (start == end) {
15451 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
15454 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%" UVXf "}",
15459 /* About to generate an ANYOF (or similar) node from the inversion list we
15460 * have calculated */
15461 save_parse = RExC_parse;
15462 RExC_parse = SvPV(result_string, len);
15463 save_end = RExC_end;
15464 RExC_end = RExC_parse + len;
15466 /* We turn off folding around the call, as the class we have constructed
15467 * already has all folding taken into consideration, and we don't want
15468 * regclass() to add to that */
15469 RExC_flags &= ~RXf_PMf_FOLD;
15470 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if multi-char
15471 * folds are allowed. */
15472 node = regclass(pRExC_state, flagp,depth+1,
15473 FALSE, /* means parse the whole char class */
15474 FALSE, /* don't allow multi-char folds */
15475 TRUE, /* silence non-portable warnings. The above may very
15476 well have generated non-portable code points, but
15477 they're valid on this machine */
15478 FALSE, /* similarly, no need for strict */
15479 FALSE, /* Require return to be an ANYOF */
15484 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#" UVxf,
15487 /* Fix up the node type if we are in locale. (We have pretended we are
15488 * under /u for the purposes of regclass(), as this construct will only
15489 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
15490 * as to cause any warnings about bad locales to be output in regexec.c),
15491 * and add the flag that indicates to check if not in a UTF-8 locale. The
15492 * reason we above forbid optimization into something other than an ANYOF
15493 * node is simply to minimize the number of code changes in regexec.c.
15494 * Otherwise we would have to create new EXACTish node types and deal with
15495 * them. This decision could be revisited should this construct become
15498 * (One might think we could look at the resulting ANYOF node and suppress
15499 * the flag if everything is above 255, as those would be UTF-8 only,
15500 * but this isn't true, as the components that led to that result could
15501 * have been locale-affected, and just happen to cancel each other out
15502 * under UTF-8 locales.) */
15504 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
15506 assert(OP(node) == ANYOF);
15510 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
15514 RExC_flags |= RXf_PMf_FOLD;
15517 RExC_parse = save_parse + 1;
15518 RExC_end = save_end;
15519 SvREFCNT_dec_NN(final);
15520 SvREFCNT_dec_NN(result_string);
15522 nextchar(pRExC_state);
15523 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
15527 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15530 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
15531 AV * stack, const IV fence, AV * fence_stack)
15532 { /* Dumps the stacks in handle_regex_sets() */
15534 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
15535 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
15538 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
15540 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
15542 if (stack_top < 0) {
15543 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
15546 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
15547 for (i = stack_top; i >= 0; i--) {
15548 SV ** element_ptr = av_fetch(stack, i, FALSE);
15549 if (! element_ptr) {
15552 if (IS_OPERATOR(*element_ptr)) {
15553 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
15554 (int) i, (int) SvIV(*element_ptr));
15557 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
15558 sv_dump(*element_ptr);
15563 if (fence_stack_top < 0) {
15564 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
15567 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
15568 for (i = fence_stack_top; i >= 0; i--) {
15569 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
15570 if (! element_ptr) {
15573 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
15574 (int) i, (int) SvIV(*element_ptr));
15585 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
15587 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
15588 * innocent-looking character class, like /[ks]/i won't have to go out to
15589 * disk to find the possible matches.
15591 * This should be called only for a Latin1-range code points, cp, which is
15592 * known to be involved in a simple fold with other code points above
15593 * Latin1. It would give false results if /aa has been specified.
15594 * Multi-char folds are outside the scope of this, and must be handled
15597 * XXX It would be better to generate these via regen, in case a new
15598 * version of the Unicode standard adds new mappings, though that is not
15599 * really likely, and may be caught by the default: case of the switch
15602 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
15604 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
15610 add_cp_to_invlist(*invlist, KELVIN_SIGN);
15614 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
15617 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
15618 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
15620 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
15621 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
15622 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
15624 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
15625 *invlist = add_cp_to_invlist(*invlist,
15626 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
15629 #ifdef LATIN_CAPITAL_LETTER_SHARP_S /* not defined in early Unicode releases */
15631 case LATIN_SMALL_LETTER_SHARP_S:
15632 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
15637 #if UNICODE_MAJOR_VERSION < 3 \
15638 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0)
15640 /* In 3.0 and earlier, U+0130 folded simply to 'i'; and in 3.0.1 so did
15645 add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
15646 # if UNICODE_DOT_DOT_VERSION == 1
15647 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_DOTLESS_I);
15653 /* Use deprecated warning to increase the chances of this being
15656 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
15663 S_output_or_return_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings, AV** return_posix_warnings)
15665 /* If the final parameter is NULL, output the elements of the array given
15666 * by '*posix_warnings' as REGEXP warnings. Otherwise, the elements are
15667 * pushed onto it, (creating if necessary) */
15670 const bool first_is_fatal = ! return_posix_warnings
15671 && ckDEAD(packWARN(WARN_REGEXP));
15673 PERL_ARGS_ASSERT_OUTPUT_OR_RETURN_POSIX_WARNINGS;
15675 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
15676 if (return_posix_warnings) {
15677 if (! *return_posix_warnings) { /* mortalize to not leak if
15678 warnings are fatal */
15679 *return_posix_warnings = (AV *) sv_2mortal((SV *) newAV());
15681 av_push(*return_posix_warnings, msg);
15684 if (first_is_fatal) { /* Avoid leaking this */
15685 av_undef(posix_warnings); /* This isn't necessary if the
15686 array is mortal, but is a
15688 (void) sv_2mortal(msg);
15690 SAVEFREESV(RExC_rx_sv);
15693 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
15694 SvREFCNT_dec_NN(msg);
15700 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
15702 /* This adds the string scalar <multi_string> to the array
15703 * <multi_char_matches>. <multi_string> is known to have exactly
15704 * <cp_count> code points in it. This is used when constructing a
15705 * bracketed character class and we find something that needs to match more
15706 * than a single character.
15708 * <multi_char_matches> is actually an array of arrays. Each top-level
15709 * element is an array that contains all the strings known so far that are
15710 * the same length. And that length (in number of code points) is the same
15711 * as the index of the top-level array. Hence, the [2] element is an
15712 * array, each element thereof is a string containing TWO code points;
15713 * while element [3] is for strings of THREE characters, and so on. Since
15714 * this is for multi-char strings there can never be a [0] nor [1] element.
15716 * When we rewrite the character class below, we will do so such that the
15717 * longest strings are written first, so that it prefers the longest
15718 * matching strings first. This is done even if it turns out that any
15719 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
15720 * Christiansen has agreed that this is ok. This makes the test for the
15721 * ligature 'ffi' come before the test for 'ff', for example */
15724 AV** this_array_ptr;
15726 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
15728 if (! multi_char_matches) {
15729 multi_char_matches = newAV();
15732 if (av_exists(multi_char_matches, cp_count)) {
15733 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
15734 this_array = *this_array_ptr;
15737 this_array = newAV();
15738 av_store(multi_char_matches, cp_count,
15741 av_push(this_array, multi_string);
15743 return multi_char_matches;
15746 /* The names of properties whose definitions are not known at compile time are
15747 * stored in this SV, after a constant heading. So if the length has been
15748 * changed since initialization, then there is a run-time definition. */
15749 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
15750 (SvCUR(listsv) != initial_listsv_len)
15752 /* There is a restricted set of white space characters that are legal when
15753 * ignoring white space in a bracketed character class. This generates the
15754 * code to skip them.
15756 * There is a line below that uses the same white space criteria but is outside
15757 * this macro. Both here and there must use the same definition */
15758 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
15761 while (isBLANK_A(UCHARAT(p))) \
15769 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
15770 const bool stop_at_1, /* Just parse the next thing, don't
15771 look for a full character class */
15772 bool allow_multi_folds,
15773 const bool silence_non_portable, /* Don't output warnings
15777 bool optimizable, /* ? Allow a non-ANYOF return
15779 SV** ret_invlist, /* Return an inversion list, not a node */
15780 AV** return_posix_warnings
15783 /* parse a bracketed class specification. Most of these will produce an
15784 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
15785 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
15786 * under /i with multi-character folds: it will be rewritten following the
15787 * paradigm of this example, where the <multi-fold>s are characters which
15788 * fold to multiple character sequences:
15789 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
15790 * gets effectively rewritten as:
15791 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
15792 * reg() gets called (recursively) on the rewritten version, and this
15793 * function will return what it constructs. (Actually the <multi-fold>s
15794 * aren't physically removed from the [abcdefghi], it's just that they are
15795 * ignored in the recursion by means of a flag:
15796 * <RExC_in_multi_char_class>.)
15798 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
15799 * characters, with the corresponding bit set if that character is in the
15800 * list. For characters above this, a range list or swash is used. There
15801 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
15802 * determinable at compile time
15804 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs
15805 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded
15806 * to UTF-8. This can only happen if ret_invlist is non-NULL.
15809 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
15811 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
15814 int namedclass = OOB_NAMEDCLASS;
15815 char *rangebegin = NULL;
15816 bool need_class = 0;
15818 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
15819 than just initialized. */
15820 SV* properties = NULL; /* Code points that match \p{} \P{} */
15821 SV* posixes = NULL; /* Code points that match classes like [:word:],
15822 extended beyond the Latin1 range. These have to
15823 be kept separate from other code points for much
15824 of this function because their handling is
15825 different under /i, and for most classes under
15827 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
15828 separate for a while from the non-complemented
15829 versions because of complications with /d
15831 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
15832 treated more simply than the general case,
15833 leading to less compilation and execution
15835 UV element_count = 0; /* Number of distinct elements in the class.
15836 Optimizations may be possible if this is tiny */
15837 AV * multi_char_matches = NULL; /* Code points that fold to more than one
15838 character; used under /i */
15840 char * stop_ptr = RExC_end; /* where to stop parsing */
15842 /* ignore unescaped whitespace? */
15843 const bool skip_white = cBOOL( ret_invlist
15844 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
15846 /* Unicode properties are stored in a swash; this holds the current one
15847 * being parsed. If this swash is the only above-latin1 component of the
15848 * character class, an optimization is to pass it directly on to the
15849 * execution engine. Otherwise, it is set to NULL to indicate that there
15850 * are other things in the class that have to be dealt with at execution
15852 SV* swash = NULL; /* Code points that match \p{} \P{} */
15854 /* Set if a component of this character class is user-defined; just passed
15855 * on to the engine */
15856 bool has_user_defined_property = FALSE;
15858 /* inversion list of code points this node matches only when the target
15859 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
15861 SV* has_upper_latin1_only_utf8_matches = NULL;
15863 /* Inversion list of code points this node matches regardless of things
15864 * like locale, folding, utf8ness of the target string */
15865 SV* cp_list = NULL;
15867 /* Like cp_list, but code points on this list need to be checked for things
15868 * that fold to/from them under /i */
15869 SV* cp_foldable_list = NULL;
15871 /* Like cp_list, but code points on this list are valid only when the
15872 * runtime locale is UTF-8 */
15873 SV* only_utf8_locale_list = NULL;
15875 /* In a range, if one of the endpoints is non-character-set portable,
15876 * meaning that it hard-codes a code point that may mean a different
15877 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
15878 * mnemonic '\t' which each mean the same character no matter which
15879 * character set the platform is on. */
15880 unsigned int non_portable_endpoint = 0;
15882 /* Is the range unicode? which means on a platform that isn't 1-1 native
15883 * to Unicode (i.e. non-ASCII), each code point in it should be considered
15884 * to be a Unicode value. */
15885 bool unicode_range = FALSE;
15886 bool invert = FALSE; /* Is this class to be complemented */
15888 bool warn_super = ALWAYS_WARN_SUPER;
15890 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
15891 case we need to change the emitted regop to an EXACT. */
15892 const char * orig_parse = RExC_parse;
15893 const SSize_t orig_size = RExC_size;
15894 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
15896 /* This variable is used to mark where the end in the input is of something
15897 * that looks like a POSIX construct but isn't. During the parse, when
15898 * something looks like it could be such a construct is encountered, it is
15899 * checked for being one, but not if we've already checked this area of the
15900 * input. Only after this position is reached do we check again */
15901 char *not_posix_region_end = RExC_parse - 1;
15903 AV* posix_warnings = NULL;
15904 const bool do_posix_warnings = return_posix_warnings
15905 || (PASS2 && ckWARN(WARN_REGEXP));
15907 GET_RE_DEBUG_FLAGS_DECL;
15909 PERL_ARGS_ASSERT_REGCLASS;
15911 PERL_UNUSED_ARG(depth);
15914 DEBUG_PARSE("clas");
15916 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
15917 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
15918 && UNICODE_DOT_DOT_VERSION == 0)
15919 allow_multi_folds = FALSE;
15922 /* Assume we are going to generate an ANYOF node. */
15923 ret = reganode(pRExC_state,
15930 RExC_size += ANYOF_SKIP;
15931 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
15934 ANYOF_FLAGS(ret) = 0;
15936 RExC_emit += ANYOF_SKIP;
15937 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
15938 initial_listsv_len = SvCUR(listsv);
15939 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
15942 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15944 assert(RExC_parse <= RExC_end);
15946 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
15949 allow_multi_folds = FALSE;
15951 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15954 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
15955 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
15956 int maybe_class = handle_possible_posix(pRExC_state,
15958 ¬_posix_region_end,
15960 TRUE /* checking only */);
15961 if (PASS2 && maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
15962 SAVEFREESV(RExC_rx_sv);
15963 ckWARN4reg(not_posix_region_end,
15964 "POSIX syntax [%c %c] belongs inside character classes%s",
15965 *RExC_parse, *RExC_parse,
15966 (maybe_class == OOB_NAMEDCLASS)
15967 ? ((POSIXCC_NOTYET(*RExC_parse))
15968 ? " (but this one isn't implemented)"
15969 : " (but this one isn't fully valid)")
15972 (void)ReREFCNT_inc(RExC_rx_sv);
15976 /* If the caller wants us to just parse a single element, accomplish this
15977 * by faking the loop ending condition */
15978 if (stop_at_1 && RExC_end > RExC_parse) {
15979 stop_ptr = RExC_parse + 1;
15982 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
15983 if (UCHARAT(RExC_parse) == ']')
15984 goto charclassloop;
15988 if ( posix_warnings
15989 && av_tindex_skip_len_mg(posix_warnings) >= 0
15990 && RExC_parse > not_posix_region_end)
15992 /* Warnings about posix class issues are considered tentative until
15993 * we are far enough along in the parse that we can no longer
15994 * change our mind, at which point we either output them or add
15995 * them, if it has so specified, to what gets returned to the
15996 * caller. This is done each time through the loop so that a later
15997 * class won't zap them before they have been dealt with. */
15998 output_or_return_posix_warnings(pRExC_state, posix_warnings,
15999 return_posix_warnings);
16002 if (RExC_parse >= stop_ptr) {
16006 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16008 if (UCHARAT(RExC_parse) == ']') {
16014 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
16015 save_value = value;
16016 save_prevvalue = prevvalue;
16019 rangebegin = RExC_parse;
16021 non_portable_endpoint = 0;
16023 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
16024 value = utf8n_to_uvchr((U8*)RExC_parse,
16025 RExC_end - RExC_parse,
16026 &numlen, UTF8_ALLOW_DEFAULT);
16027 RExC_parse += numlen;
16030 value = UCHARAT(RExC_parse++);
16032 if (value == '[') {
16033 char * posix_class_end;
16034 namedclass = handle_possible_posix(pRExC_state,
16037 do_posix_warnings ? &posix_warnings : NULL,
16038 FALSE /* die if error */);
16039 if (namedclass > OOB_NAMEDCLASS) {
16041 /* If there was an earlier attempt to parse this particular
16042 * posix class, and it failed, it was a false alarm, as this
16043 * successful one proves */
16044 if ( posix_warnings
16045 && av_tindex_skip_len_mg(posix_warnings) >= 0
16046 && not_posix_region_end >= RExC_parse
16047 && not_posix_region_end <= posix_class_end)
16049 av_undef(posix_warnings);
16052 RExC_parse = posix_class_end;
16054 else if (namedclass == OOB_NAMEDCLASS) {
16055 not_posix_region_end = posix_class_end;
16058 namedclass = OOB_NAMEDCLASS;
16061 else if ( RExC_parse - 1 > not_posix_region_end
16062 && MAYBE_POSIXCC(value))
16064 (void) handle_possible_posix(
16066 RExC_parse - 1, /* -1 because parse has already been
16068 ¬_posix_region_end,
16069 do_posix_warnings ? &posix_warnings : NULL,
16070 TRUE /* checking only */);
16072 else if (value == '\\') {
16073 /* Is a backslash; get the code point of the char after it */
16075 if (RExC_parse >= RExC_end) {
16076 vFAIL("Unmatched [");
16079 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
16080 value = utf8n_to_uvchr((U8*)RExC_parse,
16081 RExC_end - RExC_parse,
16082 &numlen, UTF8_ALLOW_DEFAULT);
16083 RExC_parse += numlen;
16086 value = UCHARAT(RExC_parse++);
16088 /* Some compilers cannot handle switching on 64-bit integer
16089 * values, therefore value cannot be an UV. Yes, this will
16090 * be a problem later if we want switch on Unicode.
16091 * A similar issue a little bit later when switching on
16092 * namedclass. --jhi */
16094 /* If the \ is escaping white space when white space is being
16095 * skipped, it means that that white space is wanted literally, and
16096 * is already in 'value'. Otherwise, need to translate the escape
16097 * into what it signifies. */
16098 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
16100 case 'w': namedclass = ANYOF_WORDCHAR; break;
16101 case 'W': namedclass = ANYOF_NWORDCHAR; break;
16102 case 's': namedclass = ANYOF_SPACE; break;
16103 case 'S': namedclass = ANYOF_NSPACE; break;
16104 case 'd': namedclass = ANYOF_DIGIT; break;
16105 case 'D': namedclass = ANYOF_NDIGIT; break;
16106 case 'v': namedclass = ANYOF_VERTWS; break;
16107 case 'V': namedclass = ANYOF_NVERTWS; break;
16108 case 'h': namedclass = ANYOF_HORIZWS; break;
16109 case 'H': namedclass = ANYOF_NHORIZWS; break;
16110 case 'N': /* Handle \N{NAME} in class */
16112 const char * const backslash_N_beg = RExC_parse - 2;
16115 if (! grok_bslash_N(pRExC_state,
16116 NULL, /* No regnode */
16117 &value, /* Yes single value */
16118 &cp_count, /* Multiple code pt count */
16124 if (*flagp & NEED_UTF8)
16125 FAIL("panic: grok_bslash_N set NEED_UTF8");
16126 if (*flagp & RESTART_PASS1)
16129 if (cp_count < 0) {
16130 vFAIL("\\N in a character class must be a named character: \\N{...}");
16132 else if (cp_count == 0) {
16134 ckWARNreg(RExC_parse,
16135 "Ignoring zero length \\N{} in character class");
16138 else { /* cp_count > 1 */
16139 if (! RExC_in_multi_char_class) {
16140 if (invert || range || *RExC_parse == '-') {
16143 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
16146 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
16148 break; /* <value> contains the first code
16149 point. Drop out of the switch to
16153 SV * multi_char_N = newSVpvn(backslash_N_beg,
16154 RExC_parse - backslash_N_beg);
16156 = add_multi_match(multi_char_matches,
16161 } /* End of cp_count != 1 */
16163 /* This element should not be processed further in this
16166 value = save_value;
16167 prevvalue = save_prevvalue;
16168 continue; /* Back to top of loop to get next char */
16171 /* Here, is a single code point, and <value> contains it */
16172 unicode_range = TRUE; /* \N{} are Unicode */
16180 /* We will handle any undefined properties ourselves */
16181 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
16182 /* And we actually would prefer to get
16183 * the straight inversion list of the
16184 * swash, since we will be accessing it
16185 * anyway, to save a little time */
16186 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
16188 if (RExC_parse >= RExC_end)
16189 vFAIL2("Empty \\%c", (U8)value);
16190 if (*RExC_parse == '{') {
16191 const U8 c = (U8)value;
16192 e = strchr(RExC_parse, '}');
16195 vFAIL2("Missing right brace on \\%c{}", c);
16199 while (isSPACE(*RExC_parse)) {
16203 if (UCHARAT(RExC_parse) == '^') {
16205 /* toggle. (The rhs xor gets the single bit that
16206 * differs between P and p; the other xor inverts just
16208 value ^= 'P' ^ 'p';
16211 while (isSPACE(*RExC_parse)) {
16216 if (e == RExC_parse)
16217 vFAIL2("Empty \\%c{}", c);
16219 n = e - RExC_parse;
16220 while (isSPACE(*(RExC_parse + n - 1)))
16222 } /* The \p isn't immediately followed by a '{' */
16223 else if (! isALPHA(*RExC_parse)) {
16224 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16225 vFAIL2("Character following \\%c must be '{' or a "
16226 "single-character Unicode property name",
16236 char* base_name; /* name after any packages are stripped */
16237 char* lookup_name = NULL;
16238 const char * const colon_colon = "::";
16240 /* Try to get the definition of the property into
16241 * <invlist>. If /i is in effect, the effective property
16242 * will have its name be <__NAME_i>. The design is
16243 * discussed in commit
16244 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
16245 name = savepv(Perl_form(aTHX_ "%.*s", (int)n, RExC_parse));
16248 lookup_name = savepv(Perl_form(aTHX_ "__%s_i", name));
16250 /* The function call just below that uses this can fail
16251 * to return, leaking memory if we don't do this */
16252 SAVEFREEPV(lookup_name);
16255 /* Look up the property name, and get its swash and
16256 * inversion list, if the property is found */
16257 SvREFCNT_dec(swash); /* Free any left-overs */
16258 swash = _core_swash_init("utf8",
16265 NULL, /* No inversion list */
16268 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
16269 HV* curpkg = (IN_PERL_COMPILETIME)
16271 : CopSTASH(PL_curcop);
16275 if (swash) { /* Got a swash but no inversion list.
16276 Something is likely wrong that will
16277 be sorted-out later */
16278 SvREFCNT_dec_NN(swash);
16282 /* Here didn't find it. It could be a an error (like a
16283 * typo) in specifying a Unicode property, or it could
16284 * be a user-defined property that will be available at
16285 * run-time. The names of these must begin with 'In'
16286 * or 'Is' (after any packages are stripped off). So
16287 * if not one of those, or if we accept only
16288 * compile-time properties, is an error; otherwise add
16289 * it to the list for run-time look up. */
16290 if ((base_name = rninstr(name, name + n,
16291 colon_colon, colon_colon + 2)))
16292 { /* Has ::. We know this must be a user-defined
16295 final_n -= base_name - name;
16304 || base_name[0] != 'I'
16305 || (base_name[1] != 's' && base_name[1] != 'n')
16308 const char * const msg
16310 ? "Illegal user-defined property name"
16311 : "Can't find Unicode property definition";
16312 RExC_parse = e + 1;
16314 /* diag_listed_as: Can't find Unicode property definition "%s" */
16315 vFAIL3utf8f("%s \"%" UTF8f "\"",
16316 msg, UTF8fARG(UTF, n, name));
16319 /* If the property name doesn't already have a package
16320 * name, add the current one to it so that it can be
16321 * referred to outside it. [perl #121777] */
16322 if (! has_pkg && curpkg) {
16323 char* pkgname = HvNAME(curpkg);
16324 if (strNE(pkgname, "main")) {
16325 char* full_name = Perl_form(aTHX_
16329 n = strlen(full_name);
16330 name = savepvn(full_name, n);
16334 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s%" UTF8f "%s\n",
16335 (value == 'p' ? '+' : '!'),
16336 (FOLD) ? "__" : "",
16337 UTF8fARG(UTF, n, name),
16338 (FOLD) ? "_i" : "");
16339 has_user_defined_property = TRUE;
16340 optimizable = FALSE; /* Will have to leave this an
16343 /* We don't know yet what this matches, so have to flag
16345 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
16349 /* Here, did get the swash and its inversion list. If
16350 * the swash is from a user-defined property, then this
16351 * whole character class should be regarded as such */
16352 if (swash_init_flags
16353 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
16355 has_user_defined_property = TRUE;
16358 /* We warn on matching an above-Unicode code point
16359 * if the match would return true, except don't
16360 * warn for \p{All}, which has exactly one element
16362 (_invlist_contains_cp(invlist, 0x110000)
16363 && (! (_invlist_len(invlist) == 1
16364 && *invlist_array(invlist) == 0)))
16370 /* Invert if asking for the complement */
16371 if (value == 'P') {
16372 _invlist_union_complement_2nd(properties,
16376 /* The swash can't be used as-is, because we've
16377 * inverted things; delay removing it to here after
16378 * have copied its invlist above */
16379 SvREFCNT_dec_NN(swash);
16383 _invlist_union(properties, invlist, &properties);
16387 RExC_parse = e + 1;
16388 namedclass = ANYOF_UNIPROP; /* no official name, but it's
16391 /* \p means they want Unicode semantics */
16392 REQUIRE_UNI_RULES(flagp, NULL);
16395 case 'n': value = '\n'; break;
16396 case 'r': value = '\r'; break;
16397 case 't': value = '\t'; break;
16398 case 'f': value = '\f'; break;
16399 case 'b': value = '\b'; break;
16400 case 'e': value = ESC_NATIVE; break;
16401 case 'a': value = '\a'; break;
16403 RExC_parse--; /* function expects to be pointed at the 'o' */
16405 const char* error_msg;
16406 bool valid = grok_bslash_o(&RExC_parse,
16409 PASS2, /* warnings only in
16412 silence_non_portable,
16418 non_portable_endpoint++;
16421 RExC_parse--; /* function expects to be pointed at the 'x' */
16423 const char* error_msg;
16424 bool valid = grok_bslash_x(&RExC_parse,
16427 PASS2, /* Output warnings */
16429 silence_non_portable,
16435 non_portable_endpoint++;
16438 value = grok_bslash_c(*RExC_parse++, PASS2);
16439 non_portable_endpoint++;
16441 case '0': case '1': case '2': case '3': case '4':
16442 case '5': case '6': case '7':
16444 /* Take 1-3 octal digits */
16445 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
16446 numlen = (strict) ? 4 : 3;
16447 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
16448 RExC_parse += numlen;
16451 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16452 vFAIL("Need exactly 3 octal digits");
16454 else if (! SIZE_ONLY /* like \08, \178 */
16456 && RExC_parse < RExC_end
16457 && isDIGIT(*RExC_parse)
16458 && ckWARN(WARN_REGEXP))
16460 SAVEFREESV(RExC_rx_sv);
16461 reg_warn_non_literal_string(
16463 form_short_octal_warning(RExC_parse, numlen));
16464 (void)ReREFCNT_inc(RExC_rx_sv);
16467 non_portable_endpoint++;
16471 /* Allow \_ to not give an error */
16472 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
16474 vFAIL2("Unrecognized escape \\%c in character class",
16478 SAVEFREESV(RExC_rx_sv);
16479 ckWARN2reg(RExC_parse,
16480 "Unrecognized escape \\%c in character class passed through",
16482 (void)ReREFCNT_inc(RExC_rx_sv);
16486 } /* End of switch on char following backslash */
16487 } /* end of handling backslash escape sequences */
16489 /* Here, we have the current token in 'value' */
16491 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
16494 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
16495 * literal, as is the character that began the false range, i.e.
16496 * the 'a' in the examples */
16499 const int w = (RExC_parse >= rangebegin)
16500 ? RExC_parse - rangebegin
16504 "False [] range \"%" UTF8f "\"",
16505 UTF8fARG(UTF, w, rangebegin));
16508 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
16509 ckWARN2reg(RExC_parse,
16510 "False [] range \"%" UTF8f "\"",
16511 UTF8fARG(UTF, w, rangebegin));
16512 (void)ReREFCNT_inc(RExC_rx_sv);
16513 cp_list = add_cp_to_invlist(cp_list, '-');
16514 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
16519 range = 0; /* this was not a true range */
16520 element_count += 2; /* So counts for three values */
16523 classnum = namedclass_to_classnum(namedclass);
16525 if (LOC && namedclass < ANYOF_POSIXL_MAX
16526 #ifndef HAS_ISASCII
16527 && classnum != _CC_ASCII
16530 /* What the Posix classes (like \w, [:space:]) match in locale
16531 * isn't knowable under locale until actual match time. Room
16532 * must be reserved (one time per outer bracketed class) to
16533 * store such classes. The space will contain a bit for each
16534 * named class that is to be matched against. This isn't
16535 * needed for \p{} and pseudo-classes, as they are not affected
16536 * by locale, and hence are dealt with separately */
16537 if (! need_class) {
16540 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16543 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16545 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
16546 ANYOF_POSIXL_ZERO(ret);
16548 /* We can't change this into some other type of node
16549 * (unless this is the only element, in which case there
16550 * are nodes that mean exactly this) as has runtime
16552 optimizable = FALSE;
16555 /* Coverity thinks it is possible for this to be negative; both
16556 * jhi and khw think it's not, but be safer */
16557 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16558 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
16560 /* See if it already matches the complement of this POSIX
16562 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16563 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
16567 posixl_matches_all = TRUE;
16568 break; /* No need to continue. Since it matches both
16569 e.g., \w and \W, it matches everything, and the
16570 bracketed class can be optimized into qr/./s */
16573 /* Add this class to those that should be checked at runtime */
16574 ANYOF_POSIXL_SET(ret, namedclass);
16576 /* The above-Latin1 characters are not subject to locale rules.
16577 * Just add them, in the second pass, to the
16578 * unconditionally-matched list */
16580 SV* scratch_list = NULL;
16582 /* Get the list of the above-Latin1 code points this
16584 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
16585 PL_XPosix_ptrs[classnum],
16587 /* Odd numbers are complements, like
16588 * NDIGIT, NASCII, ... */
16589 namedclass % 2 != 0,
16591 /* Checking if 'cp_list' is NULL first saves an extra
16592 * clone. Its reference count will be decremented at the
16593 * next union, etc, or if this is the only instance, at the
16594 * end of the routine */
16596 cp_list = scratch_list;
16599 _invlist_union(cp_list, scratch_list, &cp_list);
16600 SvREFCNT_dec_NN(scratch_list);
16602 continue; /* Go get next character */
16605 else if (! SIZE_ONLY) {
16607 /* Here, not in pass1 (in that pass we skip calculating the
16608 * contents of this class), and is not /l, or is a POSIX class
16609 * for which /l doesn't matter (or is a Unicode property, which
16610 * is skipped here). */
16611 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
16612 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
16614 /* Here, should be \h, \H, \v, or \V. None of /d, /i
16615 * nor /l make a difference in what these match,
16616 * therefore we just add what they match to cp_list. */
16617 if (classnum != _CC_VERTSPACE) {
16618 assert( namedclass == ANYOF_HORIZWS
16619 || namedclass == ANYOF_NHORIZWS);
16621 /* It turns out that \h is just a synonym for
16623 classnum = _CC_BLANK;
16626 _invlist_union_maybe_complement_2nd(
16628 PL_XPosix_ptrs[classnum],
16629 namedclass % 2 != 0, /* Complement if odd
16630 (NHORIZWS, NVERTWS)
16635 else if ( UNI_SEMANTICS
16636 || classnum == _CC_ASCII
16637 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
16638 || classnum == _CC_XDIGIT)))
16640 /* We usually have to worry about /d and /a affecting what
16641 * POSIX classes match, with special code needed for /d
16642 * because we won't know until runtime what all matches.
16643 * But there is no extra work needed under /u, and
16644 * [:ascii:] is unaffected by /a and /d; and :digit: and
16645 * :xdigit: don't have runtime differences under /d. So we
16646 * can special case these, and avoid some extra work below,
16647 * and at runtime. */
16648 _invlist_union_maybe_complement_2nd(
16650 PL_XPosix_ptrs[classnum],
16651 namedclass % 2 != 0,
16654 else { /* Garden variety class. If is NUPPER, NALPHA, ...
16655 complement and use nposixes */
16656 SV** posixes_ptr = namedclass % 2 == 0
16659 _invlist_union_maybe_complement_2nd(
16661 PL_XPosix_ptrs[classnum],
16662 namedclass % 2 != 0,
16666 } /* end of namedclass \blah */
16668 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16670 /* If 'range' is set, 'value' is the ending of a range--check its
16671 * validity. (If value isn't a single code point in the case of a
16672 * range, we should have figured that out above in the code that
16673 * catches false ranges). Later, we will handle each individual code
16674 * point in the range. If 'range' isn't set, this could be the
16675 * beginning of a range, so check for that by looking ahead to see if
16676 * the next real character to be processed is the range indicator--the
16681 /* For unicode ranges, we have to test that the Unicode as opposed
16682 * to the native values are not decreasing. (Above 255, there is
16683 * no difference between native and Unicode) */
16684 if (unicode_range && prevvalue < 255 && value < 255) {
16685 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
16686 goto backwards_range;
16691 if (prevvalue > value) /* b-a */ {
16696 w = RExC_parse - rangebegin;
16698 "Invalid [] range \"%" UTF8f "\"",
16699 UTF8fARG(UTF, w, rangebegin));
16700 NOT_REACHED; /* NOTREACHED */
16704 prevvalue = value; /* save the beginning of the potential range */
16705 if (! stop_at_1 /* Can't be a range if parsing just one thing */
16706 && *RExC_parse == '-')
16708 char* next_char_ptr = RExC_parse + 1;
16710 /* Get the next real char after the '-' */
16711 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
16713 /* If the '-' is at the end of the class (just before the ']',
16714 * it is a literal minus; otherwise it is a range */
16715 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
16716 RExC_parse = next_char_ptr;
16718 /* a bad range like \w-, [:word:]- ? */
16719 if (namedclass > OOB_NAMEDCLASS) {
16720 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
16721 const int w = RExC_parse >= rangebegin
16722 ? RExC_parse - rangebegin
16725 vFAIL4("False [] range \"%*.*s\"",
16730 "False [] range \"%*.*s\"",
16735 cp_list = add_cp_to_invlist(cp_list, '-');
16739 range = 1; /* yeah, it's a range! */
16740 continue; /* but do it the next time */
16745 if (namedclass > OOB_NAMEDCLASS) {
16749 /* Here, we have a single value this time through the loop, and
16750 * <prevvalue> is the beginning of the range, if any; or <value> if
16753 /* non-Latin1 code point implies unicode semantics. Must be set in
16754 * pass1 so is there for the whole of pass 2 */
16756 REQUIRE_UNI_RULES(flagp, NULL);
16759 /* Ready to process either the single value, or the completed range.
16760 * For single-valued non-inverted ranges, we consider the possibility
16761 * of multi-char folds. (We made a conscious decision to not do this
16762 * for the other cases because it can often lead to non-intuitive
16763 * results. For example, you have the peculiar case that:
16764 * "s s" =~ /^[^\xDF]+$/i => Y
16765 * "ss" =~ /^[^\xDF]+$/i => N
16767 * See [perl #89750] */
16768 if (FOLD && allow_multi_folds && value == prevvalue) {
16769 if (value == LATIN_SMALL_LETTER_SHARP_S
16770 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
16773 /* Here <value> is indeed a multi-char fold. Get what it is */
16775 U8 foldbuf[UTF8_MAXBYTES_CASE];
16778 UV folded = _to_uni_fold_flags(
16782 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
16783 ? FOLD_FLAGS_NOMIX_ASCII
16787 /* Here, <folded> should be the first character of the
16788 * multi-char fold of <value>, with <foldbuf> containing the
16789 * whole thing. But, if this fold is not allowed (because of
16790 * the flags), <fold> will be the same as <value>, and should
16791 * be processed like any other character, so skip the special
16793 if (folded != value) {
16795 /* Skip if we are recursed, currently parsing the class
16796 * again. Otherwise add this character to the list of
16797 * multi-char folds. */
16798 if (! RExC_in_multi_char_class) {
16799 STRLEN cp_count = utf8_length(foldbuf,
16800 foldbuf + foldlen);
16801 SV* multi_fold = sv_2mortal(newSVpvs(""));
16803 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
16806 = add_multi_match(multi_char_matches,
16812 /* This element should not be processed further in this
16815 value = save_value;
16816 prevvalue = save_prevvalue;
16822 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
16825 /* If the range starts above 255, everything is portable and
16826 * likely to be so for any forseeable character set, so don't
16828 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
16829 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
16831 else if (prevvalue != value) {
16833 /* Under strict, ranges that stop and/or end in an ASCII
16834 * printable should have each end point be a portable value
16835 * for it (preferably like 'A', but we don't warn if it is
16836 * a (portable) Unicode name or code point), and the range
16837 * must be be all digits or all letters of the same case.
16838 * Otherwise, the range is non-portable and unclear as to
16839 * what it contains */
16840 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
16841 && ( non_portable_endpoint
16842 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
16843 || (isLOWER_A(prevvalue) && isLOWER_A(value))
16844 || (isUPPER_A(prevvalue) && isUPPER_A(value))
16846 vWARN(RExC_parse, "Ranges of ASCII printables should"
16847 " be some subset of \"0-9\","
16848 " \"A-Z\", or \"a-z\"");
16850 else if (prevvalue >= 0x660) { /* ARABIC_INDIC_DIGIT_ZERO */
16851 SSize_t index_start;
16852 SSize_t index_final;
16854 /* But the nature of Unicode and languages mean we
16855 * can't do the same checks for above-ASCII ranges,
16856 * except in the case of digit ones. These should
16857 * contain only digits from the same group of 10. The
16858 * ASCII case is handled just above. 0x660 is the
16859 * first digit character beyond ASCII. Hence here, the
16860 * range could be a range of digits. First some
16861 * unlikely special cases. Grandfather in that a range
16862 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
16863 * if its starting value is one of the 10 digits prior
16864 * to it. This is because it is an alternate way of
16865 * writing 19D1, and some people may expect it to be in
16866 * that group. But it is bad, because it won't give
16867 * the expected results. In Unicode 5.2 it was
16868 * considered to be in that group (of 11, hence), but
16869 * this was fixed in the next version */
16871 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
16872 goto warn_bad_digit_range;
16874 else if (UNLIKELY( prevvalue >= 0x1D7CE
16875 && value <= 0x1D7FF))
16877 /* This is the only other case currently in Unicode
16878 * where the algorithm below fails. The code
16879 * points just above are the end points of a single
16880 * range containing only decimal digits. It is 5
16881 * different series of 0-9. All other ranges of
16882 * digits currently in Unicode are just a single
16883 * series. (And mktables will notify us if a later
16884 * Unicode version breaks this.)
16886 * If the range being checked is at most 9 long,
16887 * and the digit values represented are in
16888 * numerical order, they are from the same series.
16890 if ( value - prevvalue > 9
16891 || ((( value - 0x1D7CE) % 10)
16892 <= (prevvalue - 0x1D7CE) % 10))
16894 goto warn_bad_digit_range;
16899 /* For all other ranges of digits in Unicode, the
16900 * algorithm is just to check if both end points
16901 * are in the same series, which is the same range.
16903 index_start = _invlist_search(
16904 PL_XPosix_ptrs[_CC_DIGIT],
16907 /* Warn if the range starts and ends with a digit,
16908 * and they are not in the same group of 10. */
16909 if ( index_start >= 0
16910 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
16912 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
16913 value)) != index_start
16914 && index_final >= 0
16915 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
16917 warn_bad_digit_range:
16918 vWARN(RExC_parse, "Ranges of digits should be"
16919 " from the same group of"
16926 if ((! range || prevvalue == value) && non_portable_endpoint) {
16927 if (isPRINT_A(value)) {
16930 if (isBACKSLASHED_PUNCT(value)) {
16931 literal[d++] = '\\';
16933 literal[d++] = (char) value;
16934 literal[d++] = '\0';
16936 vWARN4dep(RExC_parse,
16937 "\"%.*s\" is more clearly written simply as \"%s\". "
16938 "This will be a fatal error in Perl 5.28",
16939 (int) (RExC_parse - rangebegin),
16944 else if isMNEMONIC_CNTRL(value) {
16945 vWARN4dep(RExC_parse,
16946 "\"%.*s\" is more clearly written simply as \"%s\". "
16947 "This will be a fatal error in Perl 5.28",
16948 (int) (RExC_parse - rangebegin),
16950 cntrl_to_mnemonic((U8) value)
16956 /* Deal with this element of the class */
16960 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16963 /* On non-ASCII platforms, for ranges that span all of 0..255, and
16964 * ones that don't require special handling, we can just add the
16965 * range like we do for ASCII platforms */
16966 if ((UNLIKELY(prevvalue == 0) && value >= 255)
16967 || ! (prevvalue < 256
16969 || (! non_portable_endpoint
16970 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
16971 || (isUPPER_A(prevvalue)
16972 && isUPPER_A(value)))))))
16974 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16978 /* Here, requires special handling. This can be because it is
16979 * a range whose code points are considered to be Unicode, and
16980 * so must be individually translated into native, or because
16981 * its a subrange of 'A-Z' or 'a-z' which each aren't
16982 * contiguous in EBCDIC, but we have defined them to include
16983 * only the "expected" upper or lower case ASCII alphabetics.
16984 * Subranges above 255 are the same in native and Unicode, so
16985 * can be added as a range */
16986 U8 start = NATIVE_TO_LATIN1(prevvalue);
16988 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
16989 for (j = start; j <= end; j++) {
16990 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
16993 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17000 range = 0; /* this range (if it was one) is done now */
17001 } /* End of loop through all the text within the brackets */
17004 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
17005 output_or_return_posix_warnings(pRExC_state, posix_warnings,
17006 return_posix_warnings);
17009 /* If anything in the class expands to more than one character, we have to
17010 * deal with them by building up a substitute parse string, and recursively
17011 * calling reg() on it, instead of proceeding */
17012 if (multi_char_matches) {
17013 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
17016 char *save_end = RExC_end;
17017 char *save_parse = RExC_parse;
17018 char *save_start = RExC_start;
17019 STRLEN prefix_end = 0; /* We copy the character class after a
17020 prefix supplied here. This is the size
17021 + 1 of that prefix */
17022 bool first_time = TRUE; /* First multi-char occurrence doesn't get
17027 assert(RExC_precomp_adj == 0); /* Only one level of recursion allowed */
17029 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
17030 because too confusing */
17032 sv_catpv(substitute_parse, "(?:");
17036 /* Look at the longest folds first */
17037 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
17042 if (av_exists(multi_char_matches, cp_count)) {
17043 AV** this_array_ptr;
17046 this_array_ptr = (AV**) av_fetch(multi_char_matches,
17048 while ((this_sequence = av_pop(*this_array_ptr)) !=
17051 if (! first_time) {
17052 sv_catpv(substitute_parse, "|");
17054 first_time = FALSE;
17056 sv_catpv(substitute_parse, SvPVX(this_sequence));
17061 /* If the character class contains anything else besides these
17062 * multi-character folds, have to include it in recursive parsing */
17063 if (element_count) {
17064 sv_catpv(substitute_parse, "|[");
17065 prefix_end = SvCUR(substitute_parse);
17066 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
17068 /* Put in a closing ']' only if not going off the end, as otherwise
17069 * we are adding something that really isn't there */
17070 if (RExC_parse < RExC_end) {
17071 sv_catpv(substitute_parse, "]");
17075 sv_catpv(substitute_parse, ")");
17078 /* This is a way to get the parse to skip forward a whole named
17079 * sequence instead of matching the 2nd character when it fails the
17081 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
17085 /* Set up the data structure so that any errors will be properly
17086 * reported. See the comments at the definition of
17087 * REPORT_LOCATION_ARGS for details */
17088 RExC_precomp_adj = orig_parse - RExC_precomp;
17089 RExC_start = RExC_parse = SvPV(substitute_parse, len);
17090 RExC_adjusted_start = RExC_start + prefix_end;
17091 RExC_end = RExC_parse + len;
17092 RExC_in_multi_char_class = 1;
17093 RExC_emit = (regnode *)orig_emit;
17095 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17097 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PASS1|NEED_UTF8);
17099 /* And restore so can parse the rest of the pattern */
17100 RExC_parse = save_parse;
17101 RExC_start = RExC_adjusted_start = save_start;
17102 RExC_precomp_adj = 0;
17103 RExC_end = save_end;
17104 RExC_in_multi_char_class = 0;
17105 SvREFCNT_dec_NN(multi_char_matches);
17109 /* Here, we've gone through the entire class and dealt with multi-char
17110 * folds. We are now in a position that we can do some checks to see if we
17111 * can optimize this ANYOF node into a simpler one, even in Pass 1.
17112 * Currently we only do two checks:
17113 * 1) is in the unlikely event that the user has specified both, eg. \w and
17114 * \W under /l, then the class matches everything. (This optimization
17115 * is done only to make the optimizer code run later work.)
17116 * 2) if the character class contains only a single element (including a
17117 * single range), we see if there is an equivalent node for it.
17118 * Other checks are possible */
17120 && ! ret_invlist /* Can't optimize if returning the constructed
17122 && (UNLIKELY(posixl_matches_all) || element_count == 1))
17127 if (UNLIKELY(posixl_matches_all)) {
17130 else if (namedclass > OOB_NAMEDCLASS) { /* this is a single named
17131 class, like \w or [:digit:]
17134 /* All named classes are mapped into POSIXish nodes, with its FLAG
17135 * argument giving which class it is */
17136 switch ((I32)namedclass) {
17137 case ANYOF_UNIPROP:
17140 /* These don't depend on the charset modifiers. They always
17141 * match under /u rules */
17142 case ANYOF_NHORIZWS:
17143 case ANYOF_HORIZWS:
17144 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
17147 case ANYOF_NVERTWS:
17152 /* The actual POSIXish node for all the rest depends on the
17153 * charset modifier. The ones in the first set depend only on
17154 * ASCII or, if available on this platform, also locale */
17158 op = (LOC) ? POSIXL : POSIXA;
17164 /* The following don't have any matches in the upper Latin1
17165 * range, hence /d is equivalent to /u for them. Making it /u
17166 * saves some branches at runtime */
17170 case ANYOF_NXDIGIT:
17171 if (! DEPENDS_SEMANTICS) {
17172 goto treat_as_default;
17178 /* The following change to CASED under /i */
17184 namedclass = ANYOF_CASED + (namedclass % 2);
17188 /* The rest have more possibilities depending on the charset.
17189 * We take advantage of the enum ordering of the charset
17190 * modifiers to get the exact node type, */
17193 op = POSIXD + get_regex_charset(RExC_flags);
17194 if (op > POSIXA) { /* /aa is same as /a */
17199 /* The odd numbered ones are the complements of the
17200 * next-lower even number one */
17201 if (namedclass % 2 == 1) {
17205 arg = namedclass_to_classnum(namedclass);
17209 else if (value == prevvalue) {
17211 /* Here, the class consists of just a single code point */
17214 if (! LOC && value == '\n') {
17215 op = REG_ANY; /* Optimize [^\n] */
17216 *flagp |= HASWIDTH|SIMPLE;
17220 else if (value < 256 || UTF) {
17222 /* Optimize a single value into an EXACTish node, but not if it
17223 * would require converting the pattern to UTF-8. */
17224 op = compute_EXACTish(pRExC_state);
17226 } /* Otherwise is a range */
17227 else if (! LOC) { /* locale could vary these */
17228 if (prevvalue == '0') {
17229 if (value == '9') {
17234 else if (! FOLD || ASCII_FOLD_RESTRICTED) {
17235 /* We can optimize A-Z or a-z, but not if they could match
17236 * something like the KELVIN SIGN under /i. */
17237 if (prevvalue == 'A') {
17240 && ! non_portable_endpoint
17243 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
17247 else if (prevvalue == 'a') {
17250 && ! non_portable_endpoint
17253 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
17260 /* Here, we have changed <op> away from its initial value iff we found
17261 * an optimization */
17264 /* Throw away this ANYOF regnode, and emit the calculated one,
17265 * which should correspond to the beginning, not current, state of
17267 const char * cur_parse = RExC_parse;
17268 RExC_parse = (char *)orig_parse;
17272 /* To get locale nodes to not use the full ANYOF size would
17273 * require moving the code above that writes the portions
17274 * of it that aren't in other nodes to after this point.
17275 * e.g. ANYOF_POSIXL_SET */
17276 RExC_size = orig_size;
17280 RExC_emit = (regnode *)orig_emit;
17281 if (PL_regkind[op] == POSIXD) {
17282 if (op == POSIXL) {
17283 RExC_contains_locale = 1;
17286 op += NPOSIXD - POSIXD;
17291 ret = reg_node(pRExC_state, op);
17293 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17297 *flagp |= HASWIDTH|SIMPLE;
17299 else if (PL_regkind[op] == EXACT) {
17300 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17301 TRUE /* downgradable to EXACT */
17305 RExC_parse = (char *) cur_parse;
17307 SvREFCNT_dec(posixes);
17308 SvREFCNT_dec(nposixes);
17309 SvREFCNT_dec(simple_posixes);
17310 SvREFCNT_dec(cp_list);
17311 SvREFCNT_dec(cp_foldable_list);
17318 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
17320 /* If folding, we calculate all characters that could fold to or from the
17321 * ones already on the list */
17322 if (cp_foldable_list) {
17324 UV start, end; /* End points of code point ranges */
17326 SV* fold_intersection = NULL;
17329 /* Our calculated list will be for Unicode rules. For locale
17330 * matching, we have to keep a separate list that is consulted at
17331 * runtime only when the locale indicates Unicode rules. For
17332 * non-locale, we just use the general list */
17334 use_list = &only_utf8_locale_list;
17337 use_list = &cp_list;
17340 /* Only the characters in this class that participate in folds need
17341 * be checked. Get the intersection of this class and all the
17342 * possible characters that are foldable. This can quickly narrow
17343 * down a large class */
17344 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
17345 &fold_intersection);
17347 /* The folds for all the Latin1 characters are hard-coded into this
17348 * program, but we have to go out to disk to get the others. */
17349 if (invlist_highest(cp_foldable_list) >= 256) {
17351 /* This is a hash that for a particular fold gives all
17352 * characters that are involved in it */
17353 if (! PL_utf8_foldclosures) {
17354 _load_PL_utf8_foldclosures();
17358 /* Now look at the foldable characters in this class individually */
17359 invlist_iterinit(fold_intersection);
17360 while (invlist_iternext(fold_intersection, &start, &end)) {
17363 /* Look at every character in the range */
17364 for (j = start; j <= end; j++) {
17365 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17371 if (IS_IN_SOME_FOLD_L1(j)) {
17373 /* ASCII is always matched; non-ASCII is matched
17374 * only under Unicode rules (which could happen
17375 * under /l if the locale is a UTF-8 one */
17376 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
17377 *use_list = add_cp_to_invlist(*use_list,
17378 PL_fold_latin1[j]);
17381 has_upper_latin1_only_utf8_matches
17382 = add_cp_to_invlist(
17383 has_upper_latin1_only_utf8_matches,
17384 PL_fold_latin1[j]);
17388 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
17389 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
17391 add_above_Latin1_folds(pRExC_state,
17398 /* Here is an above Latin1 character. We don't have the
17399 * rules hard-coded for it. First, get its fold. This is
17400 * the simple fold, as the multi-character folds have been
17401 * handled earlier and separated out */
17402 _to_uni_fold_flags(j, foldbuf, &foldlen,
17403 (ASCII_FOLD_RESTRICTED)
17404 ? FOLD_FLAGS_NOMIX_ASCII
17407 /* Single character fold of above Latin1. Add everything in
17408 * its fold closure to the list that this node should match.
17409 * The fold closures data structure is a hash with the keys
17410 * being the UTF-8 of every character that is folded to, like
17411 * 'k', and the values each an array of all code points that
17412 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
17413 * Multi-character folds are not included */
17414 if ((listp = hv_fetch(PL_utf8_foldclosures,
17415 (char *) foldbuf, foldlen, FALSE)))
17417 AV* list = (AV*) *listp;
17419 for (k = 0; k <= av_tindex_skip_len_mg(list); k++) {
17420 SV** c_p = av_fetch(list, k, FALSE);
17426 /* /aa doesn't allow folds between ASCII and non- */
17427 if ((ASCII_FOLD_RESTRICTED
17428 && (isASCII(c) != isASCII(j))))
17433 /* Folds under /l which cross the 255/256 boundary
17434 * are added to a separate list. (These are valid
17435 * only when the locale is UTF-8.) */
17436 if (c < 256 && LOC) {
17437 *use_list = add_cp_to_invlist(*use_list, c);
17441 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
17443 cp_list = add_cp_to_invlist(cp_list, c);
17446 /* Similarly folds involving non-ascii Latin1
17447 * characters under /d are added to their list */
17448 has_upper_latin1_only_utf8_matches
17449 = add_cp_to_invlist(
17450 has_upper_latin1_only_utf8_matches,
17457 SvREFCNT_dec_NN(fold_intersection);
17460 /* Now that we have finished adding all the folds, there is no reason
17461 * to keep the foldable list separate */
17462 _invlist_union(cp_list, cp_foldable_list, &cp_list);
17463 SvREFCNT_dec_NN(cp_foldable_list);
17466 /* And combine the result (if any) with any inversion lists from posix
17467 * classes. The lists are kept separate up to now because we don't want to
17468 * fold the classes (folding of those is automatically handled by the swash
17469 * fetching code) */
17470 if (simple_posixes) { /* These are the classes known to be unaffected by
17473 _invlist_union(cp_list, simple_posixes, &cp_list);
17474 SvREFCNT_dec_NN(simple_posixes);
17477 cp_list = simple_posixes;
17480 if (posixes || nposixes) {
17482 /* We have to adjust /a and /aa */
17483 if (AT_LEAST_ASCII_RESTRICTED) {
17485 /* Under /a and /aa, nothing above ASCII matches these */
17487 _invlist_intersection(posixes,
17488 PL_XPosix_ptrs[_CC_ASCII],
17492 /* Under /a and /aa, everything above ASCII matches these
17495 _invlist_union_complement_2nd(nposixes,
17496 PL_XPosix_ptrs[_CC_ASCII],
17501 if (! DEPENDS_SEMANTICS) {
17503 /* For everything but /d, we can just add the current 'posixes' and
17504 * 'nposixes' to the main list */
17507 _invlist_union(cp_list, posixes, &cp_list);
17508 SvREFCNT_dec_NN(posixes);
17516 _invlist_union(cp_list, nposixes, &cp_list);
17517 SvREFCNT_dec_NN(nposixes);
17520 cp_list = nposixes;
17525 /* Under /d, things like \w match upper Latin1 characters only if
17526 * the target string is in UTF-8. But things like \W match all the
17527 * upper Latin1 characters if the target string is not in UTF-8.
17529 * Handle the case where there something like \W separately */
17531 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1);
17533 /* A complemented posix class matches all upper Latin1
17534 * characters if not in UTF-8. And it matches just certain
17535 * ones when in UTF-8. That means those certain ones are
17536 * matched regardless, so can just be added to the
17537 * unconditional list */
17539 _invlist_union(cp_list, nposixes, &cp_list);
17540 SvREFCNT_dec_NN(nposixes);
17544 cp_list = nposixes;
17547 /* Likewise for 'posixes' */
17548 _invlist_union(posixes, cp_list, &cp_list);
17550 /* Likewise for anything else in the range that matched only
17552 if (has_upper_latin1_only_utf8_matches) {
17553 _invlist_union(cp_list,
17554 has_upper_latin1_only_utf8_matches,
17556 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17557 has_upper_latin1_only_utf8_matches = NULL;
17560 /* If we don't match all the upper Latin1 characters regardless
17561 * of UTF-8ness, we have to set a flag to match the rest when
17563 _invlist_subtract(only_non_utf8_list, cp_list,
17564 &only_non_utf8_list);
17565 if (_invlist_len(only_non_utf8_list) != 0) {
17566 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17570 /* Here there were no complemented posix classes. That means
17571 * the upper Latin1 characters in 'posixes' match only when the
17572 * target string is in UTF-8. So we have to add them to the
17573 * list of those types of code points, while adding the
17574 * remainder to the unconditional list.
17576 * First calculate what they are */
17577 SV* nonascii_but_latin1_properties = NULL;
17578 _invlist_intersection(posixes, PL_UpperLatin1,
17579 &nonascii_but_latin1_properties);
17581 /* And add them to the final list of such characters. */
17582 _invlist_union(has_upper_latin1_only_utf8_matches,
17583 nonascii_but_latin1_properties,
17584 &has_upper_latin1_only_utf8_matches);
17586 /* Remove them from what now becomes the unconditional list */
17587 _invlist_subtract(posixes, nonascii_but_latin1_properties,
17590 /* And add those unconditional ones to the final list */
17592 _invlist_union(cp_list, posixes, &cp_list);
17593 SvREFCNT_dec_NN(posixes);
17600 SvREFCNT_dec(nonascii_but_latin1_properties);
17602 /* Get rid of any characters that we now know are matched
17603 * unconditionally from the conditional list, which may make
17604 * that list empty */
17605 _invlist_subtract(has_upper_latin1_only_utf8_matches,
17607 &has_upper_latin1_only_utf8_matches);
17608 if (_invlist_len(has_upper_latin1_only_utf8_matches) == 0) {
17609 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17610 has_upper_latin1_only_utf8_matches = NULL;
17616 /* And combine the result (if any) with any inversion list from properties.
17617 * The lists are kept separate up to now so that we can distinguish the two
17618 * in regards to matching above-Unicode. A run-time warning is generated
17619 * if a Unicode property is matched against a non-Unicode code point. But,
17620 * we allow user-defined properties to match anything, without any warning,
17621 * and we also suppress the warning if there is a portion of the character
17622 * class that isn't a Unicode property, and which matches above Unicode, \W
17623 * or [\x{110000}] for example.
17624 * (Note that in this case, unlike the Posix one above, there is no
17625 * <has_upper_latin1_only_utf8_matches>, because having a Unicode property
17626 * forces Unicode semantics */
17630 /* If it matters to the final outcome, see if a non-property
17631 * component of the class matches above Unicode. If so, the
17632 * warning gets suppressed. This is true even if just a single
17633 * such code point is specified, as, though not strictly correct if
17634 * another such code point is matched against, the fact that they
17635 * are using above-Unicode code points indicates they should know
17636 * the issues involved */
17638 warn_super = ! (invert
17639 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
17642 _invlist_union(properties, cp_list, &cp_list);
17643 SvREFCNT_dec_NN(properties);
17646 cp_list = properties;
17651 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17653 /* Because an ANYOF node is the only one that warns, this node
17654 * can't be optimized into something else */
17655 optimizable = FALSE;
17659 /* Here, we have calculated what code points should be in the character
17662 * Now we can see about various optimizations. Fold calculation (which we
17663 * did above) needs to take place before inversion. Otherwise /[^k]/i
17664 * would invert to include K, which under /i would match k, which it
17665 * shouldn't. Therefore we can't invert folded locale now, as it won't be
17666 * folded until runtime */
17668 /* If we didn't do folding, it's because some information isn't available
17669 * until runtime; set the run-time fold flag for these. (We don't have to
17670 * worry about properties folding, as that is taken care of by the swash
17671 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
17672 * locales, or the class matches at least one 0-255 range code point */
17675 /* Some things on the list might be unconditionally included because of
17676 * other components. Remove them, and clean up the list if it goes to
17678 if (only_utf8_locale_list && cp_list) {
17679 _invlist_subtract(only_utf8_locale_list, cp_list,
17680 &only_utf8_locale_list);
17682 if (_invlist_len(only_utf8_locale_list) == 0) {
17683 SvREFCNT_dec_NN(only_utf8_locale_list);
17684 only_utf8_locale_list = NULL;
17687 if (only_utf8_locale_list) {
17690 |ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17692 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
17694 invlist_iterinit(cp_list);
17695 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
17696 ANYOF_FLAGS(ret) |= ANYOFL_FOLD;
17698 invlist_iterfinish(cp_list);
17701 else if ( DEPENDS_SEMANTICS
17702 && ( has_upper_latin1_only_utf8_matches
17703 || (ANYOF_FLAGS(ret) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
17706 optimizable = FALSE;
17710 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
17711 * at compile time. Besides not inverting folded locale now, we can't
17712 * invert if there are things such as \w, which aren't known until runtime
17716 && OP(ret) != ANYOFD
17717 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
17718 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17720 _invlist_invert(cp_list);
17722 /* Any swash can't be used as-is, because we've inverted things */
17724 SvREFCNT_dec_NN(swash);
17728 /* Clear the invert flag since have just done it here */
17735 *ret_invlist = cp_list;
17736 SvREFCNT_dec(swash);
17738 /* Discard the generated node */
17740 RExC_size = orig_size;
17743 RExC_emit = orig_emit;
17748 /* Some character classes are equivalent to other nodes. Such nodes take
17749 * up less room and generally fewer operations to execute than ANYOF nodes.
17750 * Above, we checked for and optimized into some such equivalents for
17751 * certain common classes that are easy to test. Getting to this point in
17752 * the code means that the class didn't get optimized there. Since this
17753 * code is only executed in Pass 2, it is too late to save space--it has
17754 * been allocated in Pass 1, and currently isn't given back. But turning
17755 * things into an EXACTish node can allow the optimizer to join it to any
17756 * adjacent such nodes. And if the class is equivalent to things like /./,
17757 * expensive run-time swashes can be avoided. Now that we have more
17758 * complete information, we can find things necessarily missed by the
17759 * earlier code. Another possible "optimization" that isn't done is that
17760 * something like [Ee] could be changed into an EXACTFU. khw tried this
17761 * and found that the ANYOF is faster, including for code points not in the
17762 * bitmap. This still might make sense to do, provided it got joined with
17763 * an adjacent node(s) to create a longer EXACTFU one. This could be
17764 * accomplished by creating a pseudo ANYOF_EXACTFU node type that the join
17765 * routine would know is joinable. If that didn't happen, the node type
17766 * could then be made a straight ANYOF */
17768 if (optimizable && cp_list && ! invert) {
17770 U8 op = END; /* The optimzation node-type */
17771 int posix_class = -1; /* Illegal value */
17772 const char * cur_parse= RExC_parse;
17774 invlist_iterinit(cp_list);
17775 if (! invlist_iternext(cp_list, &start, &end)) {
17777 /* Here, the list is empty. This happens, for example, when a
17778 * Unicode property that doesn't match anything is the only element
17779 * in the character class (perluniprops.pod notes such properties).
17782 *flagp |= HASWIDTH|SIMPLE;
17784 else if (start == end) { /* The range is a single code point */
17785 if (! invlist_iternext(cp_list, &start, &end)
17787 /* Don't do this optimization if it would require changing
17788 * the pattern to UTF-8 */
17789 && (start < 256 || UTF))
17791 /* Here, the list contains a single code point. Can optimize
17792 * into an EXACTish node */
17803 /* A locale node under folding with one code point can be
17804 * an EXACTFL, as its fold won't be calculated until
17810 /* Here, we are generally folding, but there is only one
17811 * code point to match. If we have to, we use an EXACT
17812 * node, but it would be better for joining with adjacent
17813 * nodes in the optimization pass if we used the same
17814 * EXACTFish node that any such are likely to be. We can
17815 * do this iff the code point doesn't participate in any
17816 * folds. For example, an EXACTF of a colon is the same as
17817 * an EXACT one, since nothing folds to or from a colon. */
17819 if (IS_IN_SOME_FOLD_L1(value)) {
17824 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
17829 /* If we haven't found the node type, above, it means we
17830 * can use the prevailing one */
17832 op = compute_EXACTish(pRExC_state);
17836 } /* End of first range contains just a single code point */
17837 else if (start == 0) {
17838 if (end == UV_MAX) {
17840 *flagp |= HASWIDTH|SIMPLE;
17843 else if (end == '\n' - 1
17844 && invlist_iternext(cp_list, &start, &end)
17845 && start == '\n' + 1 && end == UV_MAX)
17848 *flagp |= HASWIDTH|SIMPLE;
17852 invlist_iterfinish(cp_list);
17855 const UV cp_list_len = _invlist_len(cp_list);
17856 const UV* cp_list_array = invlist_array(cp_list);
17858 /* Here, didn't find an optimization. See if this matches any of
17859 * the POSIX classes. These run slightly faster for above-Unicode
17860 * code points, so don't bother with POSIXA ones nor the 2 that
17861 * have no above-Unicode matches. We can avoid these checks unless
17862 * the ANYOF matches at least as high as the lowest POSIX one
17863 * (which was manually found to be \v. The actual code point may
17864 * increase in later Unicode releases, if a higher code point is
17865 * assigned to be \v, but this code will never break. It would
17866 * just mean we could execute the checks for posix optimizations
17867 * unnecessarily) */
17869 if (cp_list_array[cp_list_len-1] > 0x2029) {
17870 for (posix_class = 0;
17871 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
17875 if (posix_class == _CC_ASCII || posix_class == _CC_CNTRL) {
17878 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
17880 /* Check if matches normal or inverted */
17881 if (_invlistEQ(cp_list,
17882 PL_XPosix_ptrs[posix_class],
17885 op = (try_inverted)
17888 *flagp |= HASWIDTH|SIMPLE;
17898 RExC_parse = (char *)orig_parse;
17899 RExC_emit = (regnode *)orig_emit;
17901 if (regarglen[op]) {
17902 ret = reganode(pRExC_state, op, 0);
17904 ret = reg_node(pRExC_state, op);
17907 RExC_parse = (char *)cur_parse;
17909 if (PL_regkind[op] == EXACT) {
17910 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17911 TRUE /* downgradable to EXACT */
17914 else if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17915 FLAGS(ret) = posix_class;
17918 SvREFCNT_dec_NN(cp_list);
17923 /* Here, <cp_list> contains all the code points we can determine at
17924 * compile time that match under all conditions. Go through it, and
17925 * for things that belong in the bitmap, put them there, and delete from
17926 * <cp_list>. While we are at it, see if everything above 255 is in the
17927 * list, and if so, set a flag to speed up execution */
17929 populate_ANYOF_from_invlist(ret, &cp_list);
17932 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
17935 /* Here, the bitmap has been populated with all the Latin1 code points that
17936 * always match. Can now add to the overall list those that match only
17937 * when the target string is UTF-8 (<has_upper_latin1_only_utf8_matches>).
17939 if (has_upper_latin1_only_utf8_matches) {
17941 _invlist_union(cp_list,
17942 has_upper_latin1_only_utf8_matches,
17944 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17947 cp_list = has_upper_latin1_only_utf8_matches;
17949 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
17952 /* If there is a swash and more than one element, we can't use the swash in
17953 * the optimization below. */
17954 if (swash && element_count > 1) {
17955 SvREFCNT_dec_NN(swash);
17959 /* Note that the optimization of using 'swash' if it is the only thing in
17960 * the class doesn't have us change swash at all, so it can include things
17961 * that are also in the bitmap; otherwise we have purposely deleted that
17962 * duplicate information */
17963 set_ANYOF_arg(pRExC_state, ret, cp_list,
17964 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17966 only_utf8_locale_list,
17967 swash, has_user_defined_property);
17969 *flagp |= HASWIDTH|SIMPLE;
17971 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
17972 RExC_contains_locale = 1;
17978 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
17981 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
17982 regnode* const node,
17984 SV* const runtime_defns,
17985 SV* const only_utf8_locale_list,
17987 const bool has_user_defined_property)
17989 /* Sets the arg field of an ANYOF-type node 'node', using information about
17990 * the node passed-in. If there is nothing outside the node's bitmap, the
17991 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
17992 * the count returned by add_data(), having allocated and stored an array,
17993 * av, that that count references, as follows:
17994 * av[0] stores the character class description in its textual form.
17995 * This is used later (regexec.c:Perl_regclass_swash()) to
17996 * initialize the appropriate swash, and is also useful for dumping
17997 * the regnode. This is set to &PL_sv_undef if the textual
17998 * description is not needed at run-time (as happens if the other
17999 * elements completely define the class)
18000 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
18001 * computed from av[0]. But if no further computation need be done,
18002 * the swash is stored here now (and av[0] is &PL_sv_undef).
18003 * av[2] stores the inversion list of code points that match only if the
18004 * current locale is UTF-8
18005 * av[3] stores the cp_list inversion list for use in addition or instead
18006 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
18007 * (Otherwise everything needed is already in av[0] and av[1])
18008 * av[4] is set if any component of the class is from a user-defined
18009 * property; used only if av[3] exists */
18013 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
18015 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
18016 assert(! (ANYOF_FLAGS(node)
18017 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
18018 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
18021 AV * const av = newAV();
18024 av_store(av, 0, (runtime_defns)
18025 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
18028 av_store(av, 1, swash);
18029 SvREFCNT_dec_NN(cp_list);
18032 av_store(av, 1, &PL_sv_undef);
18034 av_store(av, 3, cp_list);
18035 av_store(av, 4, newSVuv(has_user_defined_property));
18039 if (only_utf8_locale_list) {
18040 av_store(av, 2, only_utf8_locale_list);
18043 av_store(av, 2, &PL_sv_undef);
18046 rv = newRV_noinc(MUTABLE_SV(av));
18047 n = add_data(pRExC_state, STR_WITH_LEN("s"));
18048 RExC_rxi->data->data[n] = (void*)rv;
18053 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
18055 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
18056 const regnode* node,
18059 SV** only_utf8_locale_ptr,
18060 SV** output_invlist)
18063 /* For internal core use only.
18064 * Returns the swash for the input 'node' in the regex 'prog'.
18065 * If <doinit> is 'true', will attempt to create the swash if not already
18067 * If <listsvp> is non-null, will return the printable contents of the
18068 * swash. This can be used to get debugging information even before the
18069 * swash exists, by calling this function with 'doinit' set to false, in
18070 * which case the components that will be used to eventually create the
18071 * swash are returned (in a printable form).
18072 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
18073 * store an inversion list of code points that should match only if the
18074 * execution-time locale is a UTF-8 one.
18075 * If <output_invlist> is not NULL, it is where this routine is to store an
18076 * inversion list of the code points that would be instead returned in
18077 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
18078 * when this parameter is used, is just the non-code point data that
18079 * will go into creating the swash. This currently should be just
18080 * user-defined properties whose definitions were not known at compile
18081 * time. Using this parameter allows for easier manipulation of the
18082 * swash's data by the caller. It is illegal to call this function with
18083 * this parameter set, but not <listsvp>
18085 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
18086 * that, in spite of this function's name, the swash it returns may include
18087 * the bitmap data as well */
18090 SV *si = NULL; /* Input swash initialization string */
18091 SV* invlist = NULL;
18093 RXi_GET_DECL(prog,progi);
18094 const struct reg_data * const data = prog ? progi->data : NULL;
18096 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
18097 assert(! output_invlist || listsvp);
18099 if (data && data->count) {
18100 const U32 n = ARG(node);
18102 if (data->what[n] == 's') {
18103 SV * const rv = MUTABLE_SV(data->data[n]);
18104 AV * const av = MUTABLE_AV(SvRV(rv));
18105 SV **const ary = AvARRAY(av);
18106 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
18108 si = *ary; /* ary[0] = the string to initialize the swash with */
18110 if (av_tindex_skip_len_mg(av) >= 2) {
18111 if (only_utf8_locale_ptr
18113 && ary[2] != &PL_sv_undef)
18115 *only_utf8_locale_ptr = ary[2];
18118 assert(only_utf8_locale_ptr);
18119 *only_utf8_locale_ptr = NULL;
18122 /* Elements 3 and 4 are either both present or both absent. [3]
18123 * is any inversion list generated at compile time; [4]
18124 * indicates if that inversion list has any user-defined
18125 * properties in it. */
18126 if (av_tindex_skip_len_mg(av) >= 3) {
18128 if (SvUV(ary[4])) {
18129 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
18137 /* Element [1] is reserved for the set-up swash. If already there,
18138 * return it; if not, create it and store it there */
18139 if (ary[1] && SvROK(ary[1])) {
18142 else if (doinit && ((si && si != &PL_sv_undef)
18143 || (invlist && invlist != &PL_sv_undef))) {
18145 sw = _core_swash_init("utf8", /* the utf8 package */
18149 0, /* not from tr/// */
18151 &swash_init_flags);
18152 (void)av_store(av, 1, sw);
18157 /* If requested, return a printable version of what this swash matches */
18159 SV* matches_string = NULL;
18161 /* The swash should be used, if possible, to get the data, as it
18162 * contains the resolved data. But this function can be called at
18163 * compile-time, before everything gets resolved, in which case we
18164 * return the currently best available information, which is the string
18165 * that will eventually be used to do that resolving, 'si' */
18166 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
18167 && (si && si != &PL_sv_undef))
18169 /* Here, we only have 'si' (and possibly some passed-in data in
18170 * 'invlist', which is handled below) If the caller only wants
18171 * 'si', use that. */
18172 if (! output_invlist) {
18173 matches_string = newSVsv(si);
18176 /* But if the caller wants an inversion list of the node, we
18177 * need to parse 'si' and place as much as possible in the
18178 * desired output inversion list, making 'matches_string' only
18179 * contain the currently unresolvable things */
18180 const char *si_string = SvPVX(si);
18181 STRLEN remaining = SvCUR(si);
18185 /* Ignore everything before the first new-line */
18186 while (*si_string != '\n' && remaining > 0) {
18190 assert(remaining > 0);
18195 while (remaining > 0) {
18197 /* The data consists of just strings defining user-defined
18198 * property names, but in prior incarnations, and perhaps
18199 * somehow from pluggable regex engines, it could still
18200 * hold hex code point definitions. Each component of a
18201 * range would be separated by a tab, and each range by a
18202 * new-line. If these are found, instead add them to the
18203 * inversion list */
18204 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
18205 |PERL_SCAN_SILENT_NON_PORTABLE;
18206 STRLEN len = remaining;
18207 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
18209 /* If the hex decode routine found something, it should go
18210 * up to the next \n */
18211 if ( *(si_string + len) == '\n') {
18212 if (count) { /* 2nd code point on line */
18213 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
18216 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
18219 goto prepare_for_next_iteration;
18222 /* If the hex decode was instead for the lower range limit,
18223 * save it, and go parse the upper range limit */
18224 if (*(si_string + len) == '\t') {
18225 assert(count == 0);
18229 prepare_for_next_iteration:
18230 si_string += len + 1;
18231 remaining -= len + 1;
18235 /* Here, didn't find a legal hex number. Just add it from
18236 * here to the next \n */
18239 while (*(si_string + len) != '\n' && remaining > 0) {
18243 if (*(si_string + len) == '\n') {
18247 if (matches_string) {
18248 sv_catpvn(matches_string, si_string, len - 1);
18251 matches_string = newSVpvn(si_string, len - 1);
18254 sv_catpvs(matches_string, " ");
18255 } /* end of loop through the text */
18257 assert(matches_string);
18258 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
18259 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
18261 } /* end of has an 'si' but no swash */
18264 /* If we have a swash in place, its equivalent inversion list was above
18265 * placed into 'invlist'. If not, this variable may contain a stored
18266 * inversion list which is information beyond what is in 'si' */
18269 /* Again, if the caller doesn't want the output inversion list, put
18270 * everything in 'matches-string' */
18271 if (! output_invlist) {
18272 if ( ! matches_string) {
18273 matches_string = newSVpvs("\n");
18275 sv_catsv(matches_string, invlist_contents(invlist,
18276 TRUE /* traditional style */
18279 else if (! *output_invlist) {
18280 *output_invlist = invlist_clone(invlist);
18283 _invlist_union(*output_invlist, invlist, output_invlist);
18287 *listsvp = matches_string;
18292 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
18294 /* reg_skipcomment()
18296 Absorbs an /x style # comment from the input stream,
18297 returning a pointer to the first character beyond the comment, or if the
18298 comment terminates the pattern without anything following it, this returns
18299 one past the final character of the pattern (in other words, RExC_end) and
18300 sets the REG_RUN_ON_COMMENT_SEEN flag.
18302 Note it's the callers responsibility to ensure that we are
18303 actually in /x mode
18307 PERL_STATIC_INLINE char*
18308 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
18310 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
18314 while (p < RExC_end) {
18315 if (*(++p) == '\n') {
18320 /* we ran off the end of the pattern without ending the comment, so we have
18321 * to add an \n when wrapping */
18322 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
18327 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
18329 const bool force_to_xmod
18332 /* If the text at the current parse position '*p' is a '(?#...)' comment,
18333 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
18334 * is /x whitespace, advance '*p' so that on exit it points to the first
18335 * byte past all such white space and comments */
18337 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
18339 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
18341 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
18344 if (RExC_end - (*p) >= 3
18346 && *(*p + 1) == '?'
18347 && *(*p + 2) == '#')
18349 while (*(*p) != ')') {
18350 if ((*p) == RExC_end)
18351 FAIL("Sequence (?#... not terminated");
18359 const char * save_p = *p;
18360 while ((*p) < RExC_end) {
18362 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
18365 else if (*(*p) == '#') {
18366 (*p) = reg_skipcomment(pRExC_state, (*p));
18372 if (*p != save_p) {
18385 Advances the parse position by one byte, unless that byte is the beginning
18386 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
18387 those two cases, the parse position is advanced beyond all such comments and
18390 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
18394 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
18396 PERL_ARGS_ASSERT_NEXTCHAR;
18398 if (RExC_parse < RExC_end) {
18400 || UTF8_IS_INVARIANT(*RExC_parse)
18401 || UTF8_IS_START(*RExC_parse));
18403 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
18405 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
18406 FALSE /* Don't force /x */ );
18411 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
18413 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
18414 * space. In pass1, it aligns and increments RExC_size; in pass2,
18417 regnode * const ret = RExC_emit;
18418 GET_RE_DEBUG_FLAGS_DECL;
18420 PERL_ARGS_ASSERT_REGNODE_GUTS;
18422 assert(extra_size >= regarglen[op]);
18425 SIZE_ALIGN(RExC_size);
18426 RExC_size += 1 + extra_size;
18429 if (RExC_emit >= RExC_emit_bound)
18430 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
18431 op, (void*)RExC_emit, (void*)RExC_emit_bound);
18433 NODE_ALIGN_FILL(ret);
18434 #ifndef RE_TRACK_PATTERN_OFFSETS
18435 PERL_UNUSED_ARG(name);
18437 if (RExC_offsets) { /* MJD */
18439 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
18442 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
18443 ? "Overwriting end of array!\n" : "OK",
18444 (UV)(RExC_emit - RExC_emit_start),
18445 (UV)(RExC_parse - RExC_start),
18446 (UV)RExC_offsets[0]));
18447 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
18454 - reg_node - emit a node
18456 STATIC regnode * /* Location. */
18457 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
18459 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
18461 PERL_ARGS_ASSERT_REG_NODE;
18463 assert(regarglen[op] == 0);
18466 regnode *ptr = ret;
18467 FILL_ADVANCE_NODE(ptr, op);
18474 - reganode - emit a node with an argument
18476 STATIC regnode * /* Location. */
18477 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
18479 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
18481 PERL_ARGS_ASSERT_REGANODE;
18483 assert(regarglen[op] == 1);
18486 regnode *ptr = ret;
18487 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
18494 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
18496 /* emit a node with U32 and I32 arguments */
18498 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
18500 PERL_ARGS_ASSERT_REG2LANODE;
18502 assert(regarglen[op] == 2);
18505 regnode *ptr = ret;
18506 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
18513 - reginsert - insert an operator in front of already-emitted operand
18515 * Means relocating the operand.
18517 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
18518 * set up NEXT_OFF() of the inserted node if needed. Something like this:
18520 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
18522 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
18526 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *operand, U32 depth)
18531 const int offset = regarglen[(U8)op];
18532 const int size = NODE_STEP_REGNODE + offset;
18533 GET_RE_DEBUG_FLAGS_DECL;
18535 PERL_ARGS_ASSERT_REGINSERT;
18536 PERL_UNUSED_CONTEXT;
18537 PERL_UNUSED_ARG(depth);
18538 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
18539 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
18544 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
18545 studying. If this is wrong then we need to adjust RExC_recurse
18546 below like we do with RExC_open_parens/RExC_close_parens. */
18550 if (RExC_open_parens) {
18552 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
18553 /* remember that RExC_npar is rex->nparens + 1,
18554 * iow it is 1 more than the number of parens seen in
18555 * the pattern so far. */
18556 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
18557 /* note, RExC_open_parens[0] is the start of the
18558 * regex, it can't move. RExC_close_parens[0] is the end
18559 * of the regex, it *can* move. */
18560 if ( paren && RExC_open_parens[paren] >= operand ) {
18561 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
18562 RExC_open_parens[paren] += size;
18564 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
18566 if ( RExC_close_parens[paren] >= operand ) {
18567 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
18568 RExC_close_parens[paren] += size;
18570 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
18575 RExC_end_op += size;
18577 while (src > operand) {
18578 StructCopy(--src, --dst, regnode);
18579 #ifdef RE_TRACK_PATTERN_OFFSETS
18580 if (RExC_offsets) { /* MJD 20010112 */
18582 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
18586 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
18587 ? "Overwriting end of array!\n" : "OK",
18588 (UV)(src - RExC_emit_start),
18589 (UV)(dst - RExC_emit_start),
18590 (UV)RExC_offsets[0]));
18591 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
18592 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
18598 place = operand; /* Op node, where operand used to be. */
18599 #ifdef RE_TRACK_PATTERN_OFFSETS
18600 if (RExC_offsets) { /* MJD */
18602 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
18606 (UV)(place - RExC_emit_start) > RExC_offsets[0]
18607 ? "Overwriting end of array!\n" : "OK",
18608 (UV)(place - RExC_emit_start),
18609 (UV)(RExC_parse - RExC_start),
18610 (UV)RExC_offsets[0]));
18611 Set_Node_Offset(place, RExC_parse);
18612 Set_Node_Length(place, 1);
18615 src = NEXTOPER(place);
18616 FILL_ADVANCE_NODE(place, op);
18617 Zero(src, offset, regnode);
18621 - regtail - set the next-pointer at the end of a node chain of p to val.
18622 - SEE ALSO: regtail_study
18625 S_regtail(pTHX_ RExC_state_t * pRExC_state,
18626 const regnode * const p,
18627 const regnode * const val,
18631 GET_RE_DEBUG_FLAGS_DECL;
18633 PERL_ARGS_ASSERT_REGTAIL;
18635 PERL_UNUSED_ARG(depth);
18641 /* Find last node. */
18642 scan = (regnode *) p;
18644 regnode * const temp = regnext(scan);
18646 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
18647 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18648 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
18649 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
18650 (temp == NULL ? "->" : ""),
18651 (temp == NULL ? PL_reg_name[OP(val)] : "")
18659 if (reg_off_by_arg[OP(scan)]) {
18660 ARG_SET(scan, val - scan);
18663 NEXT_OFF(scan) = val - scan;
18669 - regtail_study - set the next-pointer at the end of a node chain of p to val.
18670 - Look for optimizable sequences at the same time.
18671 - currently only looks for EXACT chains.
18673 This is experimental code. The idea is to use this routine to perform
18674 in place optimizations on branches and groups as they are constructed,
18675 with the long term intention of removing optimization from study_chunk so
18676 that it is purely analytical.
18678 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
18679 to control which is which.
18682 /* TODO: All four parms should be const */
18685 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
18686 const regnode *val,U32 depth)
18690 #ifdef EXPERIMENTAL_INPLACESCAN
18693 GET_RE_DEBUG_FLAGS_DECL;
18695 PERL_ARGS_ASSERT_REGTAIL_STUDY;
18701 /* Find last node. */
18705 regnode * const temp = regnext(scan);
18706 #ifdef EXPERIMENTAL_INPLACESCAN
18707 if (PL_regkind[OP(scan)] == EXACT) {
18708 bool unfolded_multi_char; /* Unexamined in this routine */
18709 if (join_exact(pRExC_state, scan, &min,
18710 &unfolded_multi_char, 1, val, depth+1))
18715 switch (OP(scan)) {
18719 case EXACTFA_NO_TRIE:
18725 if( exact == PSEUDO )
18727 else if ( exact != OP(scan) )
18736 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
18737 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18738 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
18739 SvPV_nolen_const(RExC_mysv),
18740 REG_NODE_NUM(scan),
18741 PL_reg_name[exact]);
18748 DEBUG_PARSE_MSG("");
18749 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
18750 Perl_re_printf( aTHX_
18751 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
18752 SvPV_nolen_const(RExC_mysv),
18753 (IV)REG_NODE_NUM(val),
18757 if (reg_off_by_arg[OP(scan)]) {
18758 ARG_SET(scan, val - scan);
18761 NEXT_OFF(scan) = val - scan;
18769 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
18774 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
18779 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18781 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
18782 if (flags & (1<<bit)) {
18783 if (!set++ && lead)
18784 Perl_re_printf( aTHX_ "%s",lead);
18785 Perl_re_printf( aTHX_ "%s ",PL_reg_intflags_name[bit]);
18790 Perl_re_printf( aTHX_ "\n");
18792 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18797 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
18803 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18805 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
18806 if (flags & (1<<bit)) {
18807 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
18810 if (!set++ && lead)
18811 Perl_re_printf( aTHX_ "%s",lead);
18812 Perl_re_printf( aTHX_ "%s ",PL_reg_extflags_name[bit]);
18815 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
18816 if (!set++ && lead) {
18817 Perl_re_printf( aTHX_ "%s",lead);
18820 case REGEX_UNICODE_CHARSET:
18821 Perl_re_printf( aTHX_ "UNICODE");
18823 case REGEX_LOCALE_CHARSET:
18824 Perl_re_printf( aTHX_ "LOCALE");
18826 case REGEX_ASCII_RESTRICTED_CHARSET:
18827 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
18829 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
18830 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
18833 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
18839 Perl_re_printf( aTHX_ "\n");
18841 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18847 Perl_regdump(pTHX_ const regexp *r)
18850 SV * const sv = sv_newmortal();
18851 SV *dsv= sv_newmortal();
18852 RXi_GET_DECL(r,ri);
18853 GET_RE_DEBUG_FLAGS_DECL;
18855 PERL_ARGS_ASSERT_REGDUMP;
18857 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
18859 /* Header fields of interest. */
18860 if (r->anchored_substr) {
18861 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
18862 RE_SV_DUMPLEN(r->anchored_substr), 30);
18863 Perl_re_printf( aTHX_
18864 "anchored %s%s at %" IVdf " ",
18865 s, RE_SV_TAIL(r->anchored_substr),
18866 (IV)r->anchored_offset);
18867 } else if (r->anchored_utf8) {
18868 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
18869 RE_SV_DUMPLEN(r->anchored_utf8), 30);
18870 Perl_re_printf( aTHX_
18871 "anchored utf8 %s%s at %" IVdf " ",
18872 s, RE_SV_TAIL(r->anchored_utf8),
18873 (IV)r->anchored_offset);
18875 if (r->float_substr) {
18876 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
18877 RE_SV_DUMPLEN(r->float_substr), 30);
18878 Perl_re_printf( aTHX_
18879 "floating %s%s at %" IVdf "..%" UVuf " ",
18880 s, RE_SV_TAIL(r->float_substr),
18881 (IV)r->float_min_offset, (UV)r->float_max_offset);
18882 } else if (r->float_utf8) {
18883 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
18884 RE_SV_DUMPLEN(r->float_utf8), 30);
18885 Perl_re_printf( aTHX_
18886 "floating utf8 %s%s at %" IVdf "..%" UVuf " ",
18887 s, RE_SV_TAIL(r->float_utf8),
18888 (IV)r->float_min_offset, (UV)r->float_max_offset);
18890 if (r->check_substr || r->check_utf8)
18891 Perl_re_printf( aTHX_
18893 (r->check_substr == r->float_substr
18894 && r->check_utf8 == r->float_utf8
18895 ? "(checking floating" : "(checking anchored"));
18896 if (r->intflags & PREGf_NOSCAN)
18897 Perl_re_printf( aTHX_ " noscan");
18898 if (r->extflags & RXf_CHECK_ALL)
18899 Perl_re_printf( aTHX_ " isall");
18900 if (r->check_substr || r->check_utf8)
18901 Perl_re_printf( aTHX_ ") ");
18903 if (ri->regstclass) {
18904 regprop(r, sv, ri->regstclass, NULL, NULL);
18905 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
18907 if (r->intflags & PREGf_ANCH) {
18908 Perl_re_printf( aTHX_ "anchored");
18909 if (r->intflags & PREGf_ANCH_MBOL)
18910 Perl_re_printf( aTHX_ "(MBOL)");
18911 if (r->intflags & PREGf_ANCH_SBOL)
18912 Perl_re_printf( aTHX_ "(SBOL)");
18913 if (r->intflags & PREGf_ANCH_GPOS)
18914 Perl_re_printf( aTHX_ "(GPOS)");
18915 Perl_re_printf( aTHX_ " ");
18917 if (r->intflags & PREGf_GPOS_SEEN)
18918 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
18919 if (r->intflags & PREGf_SKIP)
18920 Perl_re_printf( aTHX_ "plus ");
18921 if (r->intflags & PREGf_IMPLICIT)
18922 Perl_re_printf( aTHX_ "implicit ");
18923 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
18924 if (r->extflags & RXf_EVAL_SEEN)
18925 Perl_re_printf( aTHX_ "with eval ");
18926 Perl_re_printf( aTHX_ "\n");
18928 regdump_extflags("r->extflags: ",r->extflags);
18929 regdump_intflags("r->intflags: ",r->intflags);
18932 PERL_ARGS_ASSERT_REGDUMP;
18933 PERL_UNUSED_CONTEXT;
18934 PERL_UNUSED_ARG(r);
18935 #endif /* DEBUGGING */
18938 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
18941 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
18942 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
18943 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
18944 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
18945 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
18946 || _CC_VERTSPACE != 15
18947 # error Need to adjust order of anyofs[]
18949 static const char * const anyofs[] = {
18986 - regprop - printable representation of opcode, with run time support
18990 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
18994 RXi_GET_DECL(prog,progi);
18995 GET_RE_DEBUG_FLAGS_DECL;
18997 PERL_ARGS_ASSERT_REGPROP;
19001 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
19002 /* It would be nice to FAIL() here, but this may be called from
19003 regexec.c, and it would be hard to supply pRExC_state. */
19004 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
19005 (int)OP(o), (int)REGNODE_MAX);
19006 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
19008 k = PL_regkind[OP(o)];
19011 sv_catpvs(sv, " ");
19012 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
19013 * is a crude hack but it may be the best for now since
19014 * we have no flag "this EXACTish node was UTF-8"
19016 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
19017 PERL_PV_ESCAPE_UNI_DETECT |
19018 PERL_PV_ESCAPE_NONASCII |
19019 PERL_PV_PRETTY_ELLIPSES |
19020 PERL_PV_PRETTY_LTGT |
19021 PERL_PV_PRETTY_NOCLEAR
19023 } else if (k == TRIE) {
19024 /* print the details of the trie in dumpuntil instead, as
19025 * progi->data isn't available here */
19026 const char op = OP(o);
19027 const U32 n = ARG(o);
19028 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
19029 (reg_ac_data *)progi->data->data[n] :
19031 const reg_trie_data * const trie
19032 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
19034 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
19035 DEBUG_TRIE_COMPILE_r({
19037 sv_catpvs(sv, "(JUMP)");
19038 Perl_sv_catpvf(aTHX_ sv,
19039 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
19040 (UV)trie->startstate,
19041 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
19042 (UV)trie->wordcount,
19045 (UV)TRIE_CHARCOUNT(trie),
19046 (UV)trie->uniquecharcount
19049 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
19050 sv_catpvs(sv, "[");
19051 (void) put_charclass_bitmap_innards(sv,
19052 ((IS_ANYOF_TRIE(op))
19054 : TRIE_BITMAP(trie)),
19060 sv_catpvs(sv, "]");
19062 } else if (k == CURLY) {
19063 U32 lo = ARG1(o), hi = ARG2(o);
19064 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
19065 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
19066 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
19067 if (hi == REG_INFTY)
19068 sv_catpvs(sv, "INFTY");
19070 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
19071 sv_catpvs(sv, "}");
19073 else if (k == WHILEM && o->flags) /* Ordinal/of */
19074 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
19075 else if (k == REF || k == OPEN || k == CLOSE
19076 || k == GROUPP || OP(o)==ACCEPT)
19078 AV *name_list= NULL;
19079 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
19080 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
19081 if ( RXp_PAREN_NAMES(prog) ) {
19082 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19083 } else if ( pRExC_state ) {
19084 name_list= RExC_paren_name_list;
19087 if ( k != REF || (OP(o) < NREF)) {
19088 SV **name= av_fetch(name_list, parno, 0 );
19090 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19093 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
19094 I32 *nums=(I32*)SvPVX(sv_dat);
19095 SV **name= av_fetch(name_list, nums[0], 0 );
19098 for ( n=0; n<SvIVX(sv_dat); n++ ) {
19099 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
19100 (n ? "," : ""), (IV)nums[n]);
19102 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19106 if ( k == REF && reginfo) {
19107 U32 n = ARG(o); /* which paren pair */
19108 I32 ln = prog->offs[n].start;
19109 if (prog->lastparen < n || ln == -1)
19110 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
19111 else if (ln == prog->offs[n].end)
19112 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
19114 const char *s = reginfo->strbeg + ln;
19115 Perl_sv_catpvf(aTHX_ sv, ": ");
19116 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
19117 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
19120 } else if (k == GOSUB) {
19121 AV *name_list= NULL;
19122 if ( RXp_PAREN_NAMES(prog) ) {
19123 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19124 } else if ( pRExC_state ) {
19125 name_list= RExC_paren_name_list;
19128 /* Paren and offset */
19129 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
19130 (int)((o + (int)ARG2L(o)) - progi->program) );
19132 SV **name= av_fetch(name_list, ARG(o), 0 );
19134 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19137 else if (k == LOGICAL)
19138 /* 2: embedded, otherwise 1 */
19139 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
19140 else if (k == ANYOF) {
19141 const U8 flags = ANYOF_FLAGS(o);
19142 bool do_sep = FALSE; /* Do we need to separate various components of
19144 /* Set if there is still an unresolved user-defined property */
19145 SV *unresolved = NULL;
19147 /* Things that are ignored except when the runtime locale is UTF-8 */
19148 SV *only_utf8_locale_invlist = NULL;
19150 /* Code points that don't fit in the bitmap */
19151 SV *nonbitmap_invlist = NULL;
19153 /* And things that aren't in the bitmap, but are small enough to be */
19154 SV* bitmap_range_not_in_bitmap = NULL;
19156 const bool inverted = flags & ANYOF_INVERT;
19158 if (OP(o) == ANYOFL) {
19159 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
19160 sv_catpvs(sv, "{utf8-locale-reqd}");
19162 if (flags & ANYOFL_FOLD) {
19163 sv_catpvs(sv, "{i}");
19167 /* If there is stuff outside the bitmap, get it */
19168 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
19169 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
19171 &only_utf8_locale_invlist,
19172 &nonbitmap_invlist);
19173 /* The non-bitmap data may contain stuff that could fit in the
19174 * bitmap. This could come from a user-defined property being
19175 * finally resolved when this call was done; or much more likely
19176 * because there are matches that require UTF-8 to be valid, and so
19177 * aren't in the bitmap. This is teased apart later */
19178 _invlist_intersection(nonbitmap_invlist,
19180 &bitmap_range_not_in_bitmap);
19181 /* Leave just the things that don't fit into the bitmap */
19182 _invlist_subtract(nonbitmap_invlist,
19184 &nonbitmap_invlist);
19187 /* Obey this flag to add all above-the-bitmap code points */
19188 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
19189 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
19190 NUM_ANYOF_CODE_POINTS,
19194 /* Ready to start outputting. First, the initial left bracket */
19195 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
19197 /* Then all the things that could fit in the bitmap */
19198 do_sep = put_charclass_bitmap_innards(sv,
19200 bitmap_range_not_in_bitmap,
19201 only_utf8_locale_invlist,
19204 /* Can't try inverting for a
19205 * better display if there are
19206 * things that haven't been
19208 unresolved != NULL);
19209 SvREFCNT_dec(bitmap_range_not_in_bitmap);
19211 /* If there are user-defined properties which haven't been defined yet,
19212 * output them. If the result is not to be inverted, it is clearest to
19213 * output them in a separate [] from the bitmap range stuff. If the
19214 * result is to be complemented, we have to show everything in one [],
19215 * as the inversion applies to the whole thing. Use {braces} to
19216 * separate them from anything in the bitmap and anything above the
19220 if (! do_sep) { /* If didn't output anything in the bitmap */
19221 sv_catpvs(sv, "^");
19223 sv_catpvs(sv, "{");
19226 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19228 sv_catsv(sv, unresolved);
19230 sv_catpvs(sv, "}");
19232 do_sep = ! inverted;
19235 /* And, finally, add the above-the-bitmap stuff */
19236 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
19239 /* See if truncation size is overridden */
19240 const STRLEN dump_len = (PL_dump_re_max_len)
19241 ? PL_dump_re_max_len
19244 /* This is output in a separate [] */
19246 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19249 /* And, for easy of understanding, it is shown in the
19250 * uncomplemented form if possible. The one exception being if
19251 * there are unresolved items, where the inversion has to be
19252 * delayed until runtime */
19253 if (inverted && ! unresolved) {
19254 _invlist_invert(nonbitmap_invlist);
19255 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
19258 contents = invlist_contents(nonbitmap_invlist,
19259 FALSE /* output suitable for catsv */
19262 /* If the output is shorter than the permissible maximum, just do it. */
19263 if (SvCUR(contents) <= dump_len) {
19264 sv_catsv(sv, contents);
19267 const char * contents_string = SvPVX(contents);
19268 STRLEN i = dump_len;
19270 /* Otherwise, start at the permissible max and work back to the
19271 * first break possibility */
19272 while (i > 0 && contents_string[i] != ' ') {
19275 if (i == 0) { /* Fail-safe. Use the max if we couldn't
19276 find a legal break */
19280 sv_catpvn(sv, contents_string, i);
19281 sv_catpvs(sv, "...");
19284 SvREFCNT_dec_NN(contents);
19285 SvREFCNT_dec_NN(nonbitmap_invlist);
19288 /* And finally the matching, closing ']' */
19289 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
19291 SvREFCNT_dec(unresolved);
19293 else if (k == POSIXD || k == NPOSIXD) {
19294 U8 index = FLAGS(o) * 2;
19295 if (index < C_ARRAY_LENGTH(anyofs)) {
19296 if (*anyofs[index] != '[') {
19299 sv_catpv(sv, anyofs[index]);
19300 if (*anyofs[index] != '[') {
19305 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
19308 else if (k == BOUND || k == NBOUND) {
19309 /* Must be synced with order of 'bound_type' in regcomp.h */
19310 const char * const bounds[] = {
19311 "", /* Traditional */
19317 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
19318 sv_catpv(sv, bounds[FLAGS(o)]);
19320 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
19321 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
19322 else if (OP(o) == SBOL)
19323 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
19325 /* add on the verb argument if there is one */
19326 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
19327 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
19328 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
19331 PERL_UNUSED_CONTEXT;
19332 PERL_UNUSED_ARG(sv);
19333 PERL_UNUSED_ARG(o);
19334 PERL_UNUSED_ARG(prog);
19335 PERL_UNUSED_ARG(reginfo);
19336 PERL_UNUSED_ARG(pRExC_state);
19337 #endif /* DEBUGGING */
19343 Perl_re_intuit_string(pTHX_ REGEXP * const r)
19344 { /* Assume that RE_INTUIT is set */
19345 struct regexp *const prog = ReANY(r);
19346 GET_RE_DEBUG_FLAGS_DECL;
19348 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
19349 PERL_UNUSED_CONTEXT;
19353 const char * const s = SvPV_nolen_const(RX_UTF8(r)
19354 ? prog->check_utf8 : prog->check_substr);
19356 if (!PL_colorset) reginitcolors();
19357 Perl_re_printf( aTHX_
19358 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
19360 RX_UTF8(r) ? "utf8 " : "",
19361 PL_colors[5],PL_colors[0],
19364 (strlen(s) > 60 ? "..." : ""));
19367 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
19368 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
19374 handles refcounting and freeing the perl core regexp structure. When
19375 it is necessary to actually free the structure the first thing it
19376 does is call the 'free' method of the regexp_engine associated to
19377 the regexp, allowing the handling of the void *pprivate; member
19378 first. (This routine is not overridable by extensions, which is why
19379 the extensions free is called first.)
19381 See regdupe and regdupe_internal if you change anything here.
19383 #ifndef PERL_IN_XSUB_RE
19385 Perl_pregfree(pTHX_ REGEXP *r)
19391 Perl_pregfree2(pTHX_ REGEXP *rx)
19393 struct regexp *const r = ReANY(rx);
19394 GET_RE_DEBUG_FLAGS_DECL;
19396 PERL_ARGS_ASSERT_PREGFREE2;
19398 if (r->mother_re) {
19399 ReREFCNT_dec(r->mother_re);
19401 CALLREGFREE_PVT(rx); /* free the private data */
19402 SvREFCNT_dec(RXp_PAREN_NAMES(r));
19403 Safefree(r->xpv_len_u.xpvlenu_pv);
19406 SvREFCNT_dec(r->anchored_substr);
19407 SvREFCNT_dec(r->anchored_utf8);
19408 SvREFCNT_dec(r->float_substr);
19409 SvREFCNT_dec(r->float_utf8);
19410 Safefree(r->substrs);
19412 RX_MATCH_COPY_FREE(rx);
19413 #ifdef PERL_ANY_COW
19414 SvREFCNT_dec(r->saved_copy);
19417 SvREFCNT_dec(r->qr_anoncv);
19418 if (r->recurse_locinput)
19419 Safefree(r->recurse_locinput);
19420 rx->sv_u.svu_rx = 0;
19425 This is a hacky workaround to the structural issue of match results
19426 being stored in the regexp structure which is in turn stored in
19427 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
19428 could be PL_curpm in multiple contexts, and could require multiple
19429 result sets being associated with the pattern simultaneously, such
19430 as when doing a recursive match with (??{$qr})
19432 The solution is to make a lightweight copy of the regexp structure
19433 when a qr// is returned from the code executed by (??{$qr}) this
19434 lightweight copy doesn't actually own any of its data except for
19435 the starp/end and the actual regexp structure itself.
19441 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
19443 struct regexp *ret;
19444 struct regexp *const r = ReANY(rx);
19445 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
19447 PERL_ARGS_ASSERT_REG_TEMP_COPY;
19450 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
19452 SvOK_off((SV *)ret_x);
19454 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
19455 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
19456 made both spots point to the same regexp body.) */
19457 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
19458 assert(!SvPVX(ret_x));
19459 ret_x->sv_u.svu_rx = temp->sv_any;
19460 temp->sv_any = NULL;
19461 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
19462 SvREFCNT_dec_NN(temp);
19463 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
19464 ing below will not set it. */
19465 SvCUR_set(ret_x, SvCUR(rx));
19468 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
19469 sv_force_normal(sv) is called. */
19471 ret = ReANY(ret_x);
19473 SvFLAGS(ret_x) |= SvUTF8(rx);
19474 /* We share the same string buffer as the original regexp, on which we
19475 hold a reference count, incremented when mother_re is set below.
19476 The string pointer is copied here, being part of the regexp struct.
19478 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
19479 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
19481 const I32 npar = r->nparens+1;
19482 Newx(ret->offs, npar, regexp_paren_pair);
19483 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19486 Newx(ret->substrs, 1, struct reg_substr_data);
19487 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19489 SvREFCNT_inc_void(ret->anchored_substr);
19490 SvREFCNT_inc_void(ret->anchored_utf8);
19491 SvREFCNT_inc_void(ret->float_substr);
19492 SvREFCNT_inc_void(ret->float_utf8);
19494 /* check_substr and check_utf8, if non-NULL, point to either their
19495 anchored or float namesakes, and don't hold a second reference. */
19497 RX_MATCH_COPIED_off(ret_x);
19498 #ifdef PERL_ANY_COW
19499 ret->saved_copy = NULL;
19501 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
19502 SvREFCNT_inc_void(ret->qr_anoncv);
19503 if (r->recurse_locinput)
19504 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19510 /* regfree_internal()
19512 Free the private data in a regexp. This is overloadable by
19513 extensions. Perl takes care of the regexp structure in pregfree(),
19514 this covers the *pprivate pointer which technically perl doesn't
19515 know about, however of course we have to handle the
19516 regexp_internal structure when no extension is in use.
19518 Note this is called before freeing anything in the regexp
19523 Perl_regfree_internal(pTHX_ REGEXP * const rx)
19525 struct regexp *const r = ReANY(rx);
19526 RXi_GET_DECL(r,ri);
19527 GET_RE_DEBUG_FLAGS_DECL;
19529 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
19535 SV *dsv= sv_newmortal();
19536 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
19537 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
19538 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
19539 PL_colors[4],PL_colors[5],s);
19542 #ifdef RE_TRACK_PATTERN_OFFSETS
19544 Safefree(ri->u.offsets); /* 20010421 MJD */
19546 if (ri->code_blocks)
19547 S_free_codeblocks(aTHX_ ri->code_blocks);
19550 int n = ri->data->count;
19553 /* If you add a ->what type here, update the comment in regcomp.h */
19554 switch (ri->data->what[n]) {
19560 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
19563 Safefree(ri->data->data[n]);
19569 { /* Aho Corasick add-on structure for a trie node.
19570 Used in stclass optimization only */
19572 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
19573 #ifdef USE_ITHREADS
19577 refcount = --aho->refcount;
19580 PerlMemShared_free(aho->states);
19581 PerlMemShared_free(aho->fail);
19582 /* do this last!!!! */
19583 PerlMemShared_free(ri->data->data[n]);
19584 /* we should only ever get called once, so
19585 * assert as much, and also guard the free
19586 * which /might/ happen twice. At the least
19587 * it will make code anlyzers happy and it
19588 * doesn't cost much. - Yves */
19589 assert(ri->regstclass);
19590 if (ri->regstclass) {
19591 PerlMemShared_free(ri->regstclass);
19592 ri->regstclass = 0;
19599 /* trie structure. */
19601 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
19602 #ifdef USE_ITHREADS
19606 refcount = --trie->refcount;
19609 PerlMemShared_free(trie->charmap);
19610 PerlMemShared_free(trie->states);
19611 PerlMemShared_free(trie->trans);
19613 PerlMemShared_free(trie->bitmap);
19615 PerlMemShared_free(trie->jump);
19616 PerlMemShared_free(trie->wordinfo);
19617 /* do this last!!!! */
19618 PerlMemShared_free(ri->data->data[n]);
19623 Perl_croak(aTHX_ "panic: regfree data code '%c'",
19624 ri->data->what[n]);
19627 Safefree(ri->data->what);
19628 Safefree(ri->data);
19634 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
19635 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
19636 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
19639 re_dup_guts - duplicate a regexp.
19641 This routine is expected to clone a given regexp structure. It is only
19642 compiled under USE_ITHREADS.
19644 After all of the core data stored in struct regexp is duplicated
19645 the regexp_engine.dupe method is used to copy any private data
19646 stored in the *pprivate pointer. This allows extensions to handle
19647 any duplication it needs to do.
19649 See pregfree() and regfree_internal() if you change anything here.
19651 #if defined(USE_ITHREADS)
19652 #ifndef PERL_IN_XSUB_RE
19654 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
19658 const struct regexp *r = ReANY(sstr);
19659 struct regexp *ret = ReANY(dstr);
19661 PERL_ARGS_ASSERT_RE_DUP_GUTS;
19663 npar = r->nparens+1;
19664 Newx(ret->offs, npar, regexp_paren_pair);
19665 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19667 if (ret->substrs) {
19668 /* Do it this way to avoid reading from *r after the StructCopy().
19669 That way, if any of the sv_dup_inc()s dislodge *r from the L1
19670 cache, it doesn't matter. */
19671 const bool anchored = r->check_substr
19672 ? r->check_substr == r->anchored_substr
19673 : r->check_utf8 == r->anchored_utf8;
19674 Newx(ret->substrs, 1, struct reg_substr_data);
19675 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19677 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
19678 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
19679 ret->float_substr = sv_dup_inc(ret->float_substr, param);
19680 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
19682 /* check_substr and check_utf8, if non-NULL, point to either their
19683 anchored or float namesakes, and don't hold a second reference. */
19685 if (ret->check_substr) {
19687 assert(r->check_utf8 == r->anchored_utf8);
19688 ret->check_substr = ret->anchored_substr;
19689 ret->check_utf8 = ret->anchored_utf8;
19691 assert(r->check_substr == r->float_substr);
19692 assert(r->check_utf8 == r->float_utf8);
19693 ret->check_substr = ret->float_substr;
19694 ret->check_utf8 = ret->float_utf8;
19696 } else if (ret->check_utf8) {
19698 ret->check_utf8 = ret->anchored_utf8;
19700 ret->check_utf8 = ret->float_utf8;
19705 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
19706 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
19707 if (r->recurse_locinput)
19708 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19711 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
19713 if (RX_MATCH_COPIED(dstr))
19714 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
19716 ret->subbeg = NULL;
19717 #ifdef PERL_ANY_COW
19718 ret->saved_copy = NULL;
19721 /* Whether mother_re be set or no, we need to copy the string. We
19722 cannot refrain from copying it when the storage points directly to
19723 our mother regexp, because that's
19724 1: a buffer in a different thread
19725 2: something we no longer hold a reference on
19726 so we need to copy it locally. */
19727 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
19728 ret->mother_re = NULL;
19730 #endif /* PERL_IN_XSUB_RE */
19735 This is the internal complement to regdupe() which is used to copy
19736 the structure pointed to by the *pprivate pointer in the regexp.
19737 This is the core version of the extension overridable cloning hook.
19738 The regexp structure being duplicated will be copied by perl prior
19739 to this and will be provided as the regexp *r argument, however
19740 with the /old/ structures pprivate pointer value. Thus this routine
19741 may override any copying normally done by perl.
19743 It returns a pointer to the new regexp_internal structure.
19747 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
19750 struct regexp *const r = ReANY(rx);
19751 regexp_internal *reti;
19753 RXi_GET_DECL(r,ri);
19755 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
19759 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
19760 char, regexp_internal);
19761 Copy(ri->program, reti->program, len+1, regnode);
19764 if (ri->code_blocks) {
19766 Newx(reti->code_blocks, 1, struct reg_code_blocks);
19767 Newx(reti->code_blocks->cb, ri->code_blocks->count,
19768 struct reg_code_block);
19769 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
19770 ri->code_blocks->count, struct reg_code_block);
19771 for (n = 0; n < ri->code_blocks->count; n++)
19772 reti->code_blocks->cb[n].src_regex = (REGEXP*)
19773 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
19774 reti->code_blocks->count = ri->code_blocks->count;
19775 reti->code_blocks->refcnt = 1;
19778 reti->code_blocks = NULL;
19780 reti->regstclass = NULL;
19783 struct reg_data *d;
19784 const int count = ri->data->count;
19787 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
19788 char, struct reg_data);
19789 Newx(d->what, count, U8);
19792 for (i = 0; i < count; i++) {
19793 d->what[i] = ri->data->what[i];
19794 switch (d->what[i]) {
19795 /* see also regcomp.h and regfree_internal() */
19796 case 'a': /* actually an AV, but the dup function is identical. */
19800 case 'u': /* actually an HV, but the dup function is identical. */
19801 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
19804 /* This is cheating. */
19805 Newx(d->data[i], 1, regnode_ssc);
19806 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
19807 reti->regstclass = (regnode*)d->data[i];
19810 /* Trie stclasses are readonly and can thus be shared
19811 * without duplication. We free the stclass in pregfree
19812 * when the corresponding reg_ac_data struct is freed.
19814 reti->regstclass= ri->regstclass;
19818 ((reg_trie_data*)ri->data->data[i])->refcount++;
19823 d->data[i] = ri->data->data[i];
19826 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
19827 ri->data->what[i]);
19836 reti->name_list_idx = ri->name_list_idx;
19838 #ifdef RE_TRACK_PATTERN_OFFSETS
19839 if (ri->u.offsets) {
19840 Newx(reti->u.offsets, 2*len+1, U32);
19841 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
19844 SetProgLen(reti,len);
19847 return (void*)reti;
19850 #endif /* USE_ITHREADS */
19852 #ifndef PERL_IN_XSUB_RE
19855 - regnext - dig the "next" pointer out of a node
19858 Perl_regnext(pTHX_ regnode *p)
19865 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
19866 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
19867 (int)OP(p), (int)REGNODE_MAX);
19870 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
19879 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
19882 STRLEN l1 = strlen(pat1);
19883 STRLEN l2 = strlen(pat2);
19886 const char *message;
19888 PERL_ARGS_ASSERT_RE_CROAK2;
19894 Copy(pat1, buf, l1 , char);
19895 Copy(pat2, buf + l1, l2 , char);
19896 buf[l1 + l2] = '\n';
19897 buf[l1 + l2 + 1] = '\0';
19898 va_start(args, pat2);
19899 msv = vmess(buf, &args);
19901 message = SvPV_const(msv,l1);
19904 Copy(message, buf, l1 , char);
19905 /* l1-1 to avoid \n */
19906 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, l1-1, buf));
19909 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
19911 #ifndef PERL_IN_XSUB_RE
19913 Perl_save_re_context(pTHX)
19918 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
19921 const REGEXP * const rx = PM_GETRE(PL_curpm);
19923 nparens = RX_NPARENS(rx);
19926 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
19927 * that PL_curpm will be null, but that utf8.pm and the modules it
19928 * loads will only use $1..$3.
19929 * The t/porting/re_context.t test file checks this assumption.
19934 for (i = 1; i <= nparens; i++) {
19935 char digits[TYPE_CHARS(long)];
19936 const STRLEN len = my_snprintf(digits, sizeof(digits),
19938 GV *const *const gvp
19939 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
19942 GV * const gv = *gvp;
19943 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
19953 S_put_code_point(pTHX_ SV *sv, UV c)
19955 PERL_ARGS_ASSERT_PUT_CODE_POINT;
19958 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
19960 else if (isPRINT(c)) {
19961 const char string = (char) c;
19963 /* We use {phrase} as metanotation in the class, so also escape literal
19965 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
19966 sv_catpvs(sv, "\\");
19967 sv_catpvn(sv, &string, 1);
19969 else if (isMNEMONIC_CNTRL(c)) {
19970 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
19973 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
19977 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
19980 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
19982 /* Appends to 'sv' a displayable version of the range of code points from
19983 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
19984 * that have them, when they occur at the beginning or end of the range.
19985 * It uses hex to output the remaining code points, unless 'allow_literals'
19986 * is true, in which case the printable ASCII ones are output as-is (though
19987 * some of these will be escaped by put_code_point()).
19989 * NOTE: This is designed only for printing ranges of code points that fit
19990 * inside an ANYOF bitmap. Higher code points are simply suppressed
19993 const unsigned int min_range_count = 3;
19995 assert(start <= end);
19997 PERL_ARGS_ASSERT_PUT_RANGE;
19999 while (start <= end) {
20001 const char * format;
20003 if (end - start < min_range_count) {
20005 /* Output chars individually when they occur in short ranges */
20006 for (; start <= end; start++) {
20007 put_code_point(sv, start);
20012 /* If permitted by the input options, and there is a possibility that
20013 * this range contains a printable literal, look to see if there is
20015 if (allow_literals && start <= MAX_PRINT_A) {
20017 /* If the character at the beginning of the range isn't an ASCII
20018 * printable, effectively split the range into two parts:
20019 * 1) the portion before the first such printable,
20021 * and output them separately. */
20022 if (! isPRINT_A(start)) {
20023 UV temp_end = start + 1;
20025 /* There is no point looking beyond the final possible
20026 * printable, in MAX_PRINT_A */
20027 UV max = MIN(end, MAX_PRINT_A);
20029 while (temp_end <= max && ! isPRINT_A(temp_end)) {
20033 /* Here, temp_end points to one beyond the first printable if
20034 * found, or to one beyond 'max' if not. If none found, make
20035 * sure that we use the entire range */
20036 if (temp_end > MAX_PRINT_A) {
20037 temp_end = end + 1;
20040 /* Output the first part of the split range: the part that
20041 * doesn't have printables, with the parameter set to not look
20042 * for literals (otherwise we would infinitely recurse) */
20043 put_range(sv, start, temp_end - 1, FALSE);
20045 /* The 2nd part of the range (if any) starts here. */
20048 /* We do a continue, instead of dropping down, because even if
20049 * the 2nd part is non-empty, it could be so short that we want
20050 * to output it as individual characters, as tested for at the
20051 * top of this loop. */
20055 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
20056 * output a sub-range of just the digits or letters, then process
20057 * the remaining portion as usual. */
20058 if (isALPHANUMERIC_A(start)) {
20059 UV mask = (isDIGIT_A(start))
20064 UV temp_end = start + 1;
20066 /* Find the end of the sub-range that includes just the
20067 * characters in the same class as the first character in it */
20068 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
20073 /* For short ranges, don't duplicate the code above to output
20074 * them; just call recursively */
20075 if (temp_end - start < min_range_count) {
20076 put_range(sv, start, temp_end, FALSE);
20078 else { /* Output as a range */
20079 put_code_point(sv, start);
20080 sv_catpvs(sv, "-");
20081 put_code_point(sv, temp_end);
20083 start = temp_end + 1;
20087 /* We output any other printables as individual characters */
20088 if (isPUNCT_A(start) || isSPACE_A(start)) {
20089 while (start <= end && (isPUNCT_A(start)
20090 || isSPACE_A(start)))
20092 put_code_point(sv, start);
20097 } /* End of looking for literals */
20099 /* Here is not to output as a literal. Some control characters have
20100 * mnemonic names. Split off any of those at the beginning and end of
20101 * the range to print mnemonically. It isn't possible for many of
20102 * these to be in a row, so this won't overwhelm with output */
20104 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
20106 while (isMNEMONIC_CNTRL(start) && start <= end) {
20107 put_code_point(sv, start);
20111 /* If this didn't take care of the whole range ... */
20112 if (start <= end) {
20114 /* Look backwards from the end to find the final non-mnemonic
20117 while (isMNEMONIC_CNTRL(temp_end)) {
20121 /* And separately output the interior range that doesn't start
20122 * or end with mnemonics */
20123 put_range(sv, start, temp_end, FALSE);
20125 /* Then output the mnemonic trailing controls */
20126 start = temp_end + 1;
20127 while (start <= end) {
20128 put_code_point(sv, start);
20135 /* As a final resort, output the range or subrange as hex. */
20137 this_end = (end < NUM_ANYOF_CODE_POINTS)
20139 : NUM_ANYOF_CODE_POINTS - 1;
20140 #if NUM_ANYOF_CODE_POINTS > 256
20141 format = (this_end < 256)
20142 ? "\\x%02" UVXf "-\\x%02" UVXf
20143 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
20145 format = "\\x%02" UVXf "-\\x%02" UVXf;
20147 GCC_DIAG_IGNORE(-Wformat-nonliteral);
20148 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
20155 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
20157 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
20161 bool allow_literals = TRUE;
20163 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
20165 /* Generally, it is more readable if printable characters are output as
20166 * literals, but if a range (nearly) spans all of them, it's best to output
20167 * it as a single range. This code will use a single range if all but 2
20168 * ASCII printables are in it */
20169 invlist_iterinit(invlist);
20170 while (invlist_iternext(invlist, &start, &end)) {
20172 /* If the range starts beyond the final printable, it doesn't have any
20174 if (start > MAX_PRINT_A) {
20178 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
20179 * all but two, the range must start and end no later than 2 from
20181 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
20182 if (end > MAX_PRINT_A) {
20188 if (end - start >= MAX_PRINT_A - ' ' - 2) {
20189 allow_literals = FALSE;
20194 invlist_iterfinish(invlist);
20196 /* Here we have figured things out. Output each range */
20197 invlist_iterinit(invlist);
20198 while (invlist_iternext(invlist, &start, &end)) {
20199 if (start >= NUM_ANYOF_CODE_POINTS) {
20202 put_range(sv, start, end, allow_literals);
20204 invlist_iterfinish(invlist);
20210 S_put_charclass_bitmap_innards_common(pTHX_
20211 SV* invlist, /* The bitmap */
20212 SV* posixes, /* Under /l, things like [:word:], \S */
20213 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
20214 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
20215 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
20216 const bool invert /* Is the result to be inverted? */
20219 /* Create and return an SV containing a displayable version of the bitmap
20220 * and associated information determined by the input parameters. If the
20221 * output would have been only the inversion indicator '^', NULL is instead
20226 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
20229 output = newSVpvs("^");
20232 output = newSVpvs("");
20235 /* First, the code points in the bitmap that are unconditionally there */
20236 put_charclass_bitmap_innards_invlist(output, invlist);
20238 /* Traditionally, these have been placed after the main code points */
20240 sv_catsv(output, posixes);
20243 if (only_utf8 && _invlist_len(only_utf8)) {
20244 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
20245 put_charclass_bitmap_innards_invlist(output, only_utf8);
20248 if (not_utf8 && _invlist_len(not_utf8)) {
20249 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
20250 put_charclass_bitmap_innards_invlist(output, not_utf8);
20253 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
20254 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
20255 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
20257 /* This is the only list in this routine that can legally contain code
20258 * points outside the bitmap range. The call just above to
20259 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
20260 * output them here. There's about a half-dozen possible, and none in
20261 * contiguous ranges longer than 2 */
20262 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20264 SV* above_bitmap = NULL;
20266 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
20268 invlist_iterinit(above_bitmap);
20269 while (invlist_iternext(above_bitmap, &start, &end)) {
20272 for (i = start; i <= end; i++) {
20273 put_code_point(output, i);
20276 invlist_iterfinish(above_bitmap);
20277 SvREFCNT_dec_NN(above_bitmap);
20281 if (invert && SvCUR(output) == 1) {
20289 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
20291 SV *nonbitmap_invlist,
20292 SV *only_utf8_locale_invlist,
20293 const regnode * const node,
20294 const bool force_as_is_display)
20296 /* Appends to 'sv' a displayable version of the innards of the bracketed
20297 * character class defined by the other arguments:
20298 * 'bitmap' points to the bitmap.
20299 * 'nonbitmap_invlist' is an inversion list of the code points that are in
20300 * the bitmap range, but for some reason aren't in the bitmap; NULL if
20301 * none. The reasons for this could be that they require some
20302 * condition such as the target string being or not being in UTF-8
20303 * (under /d), or because they came from a user-defined property that
20304 * was not resolved at the time of the regex compilation (under /u)
20305 * 'only_utf8_locale_invlist' is an inversion list of the code points that
20306 * are valid only if the runtime locale is a UTF-8 one; NULL if none
20307 * 'node' is the regex pattern node. It is needed only when the above two
20308 * parameters are not null, and is passed so that this routine can
20309 * tease apart the various reasons for them.
20310 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
20311 * to invert things to see if that leads to a cleaner display. If
20312 * FALSE, this routine is free to use its judgment about doing this.
20314 * It returns TRUE if there was actually something output. (It may be that
20315 * the bitmap, etc is empty.)
20317 * When called for outputting the bitmap of a non-ANYOF node, just pass the
20318 * bitmap, with the succeeding parameters set to NULL, and the final one to
20322 /* In general, it tries to display the 'cleanest' representation of the
20323 * innards, choosing whether to display them inverted or not, regardless of
20324 * whether the class itself is to be inverted. However, there are some
20325 * cases where it can't try inverting, as what actually matches isn't known
20326 * until runtime, and hence the inversion isn't either. */
20327 bool inverting_allowed = ! force_as_is_display;
20330 STRLEN orig_sv_cur = SvCUR(sv);
20332 SV* invlist; /* Inversion list we accumulate of code points that
20333 are unconditionally matched */
20334 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
20336 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
20338 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
20339 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
20342 SV* as_is_display; /* The output string when we take the inputs
20344 SV* inverted_display; /* The output string when we invert the inputs */
20346 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
20348 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
20350 /* We are biased in favor of displaying things without them being inverted,
20351 * as that is generally easier to understand */
20352 const int bias = 5;
20354 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
20356 /* Start off with whatever code points are passed in. (We clone, so we
20357 * don't change the caller's list) */
20358 if (nonbitmap_invlist) {
20359 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
20360 invlist = invlist_clone(nonbitmap_invlist);
20362 else { /* Worst case size is every other code point is matched */
20363 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
20367 if (OP(node) == ANYOFD) {
20369 /* This flag indicates that the code points below 0x100 in the
20370 * nonbitmap list are precisely the ones that match only when the
20371 * target is UTF-8 (they should all be non-ASCII). */
20372 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
20374 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
20375 _invlist_subtract(invlist, only_utf8, &invlist);
20378 /* And this flag for matching all non-ASCII 0xFF and below */
20379 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
20381 not_utf8 = invlist_clone(PL_UpperLatin1);
20384 else if (OP(node) == ANYOFL) {
20386 /* If either of these flags are set, what matches isn't
20387 * determinable except during execution, so don't know enough here
20389 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
20390 inverting_allowed = FALSE;
20393 /* What the posix classes match also varies at runtime, so these
20394 * will be output symbolically. */
20395 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
20398 posixes = newSVpvs("");
20399 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
20400 if (ANYOF_POSIXL_TEST(node,i)) {
20401 sv_catpv(posixes, anyofs[i]);
20408 /* Accumulate the bit map into the unconditional match list */
20409 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
20410 if (BITMAP_TEST(bitmap, i)) {
20412 for (; i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i); i++) {
20415 invlist = _add_range_to_invlist(invlist, start, i-1);
20419 /* Make sure that the conditional match lists don't have anything in them
20420 * that match unconditionally; otherwise the output is quite confusing.
20421 * This could happen if the code that populates these misses some
20424 _invlist_subtract(only_utf8, invlist, &only_utf8);
20427 _invlist_subtract(not_utf8, invlist, ¬_utf8);
20430 if (only_utf8_locale_invlist) {
20432 /* Since this list is passed in, we have to make a copy before
20434 only_utf8_locale = invlist_clone(only_utf8_locale_invlist);
20436 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
20438 /* And, it can get really weird for us to try outputting an inverted
20439 * form of this list when it has things above the bitmap, so don't even
20441 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20442 inverting_allowed = FALSE;
20446 /* Calculate what the output would be if we take the input as-is */
20447 as_is_display = put_charclass_bitmap_innards_common(invlist,
20454 /* If have to take the output as-is, just do that */
20455 if (! inverting_allowed) {
20456 if (as_is_display) {
20457 sv_catsv(sv, as_is_display);
20458 SvREFCNT_dec_NN(as_is_display);
20461 else { /* But otherwise, create the output again on the inverted input, and
20462 use whichever version is shorter */
20464 int inverted_bias, as_is_bias;
20466 /* We will apply our bias to whichever of the the results doesn't have
20476 inverted_bias = bias;
20479 /* Now invert each of the lists that contribute to the output,
20480 * excluding from the result things outside the possible range */
20482 /* For the unconditional inversion list, we have to add in all the
20483 * conditional code points, so that when inverted, they will be gone
20485 _invlist_union(only_utf8, invlist, &invlist);
20486 _invlist_union(not_utf8, invlist, &invlist);
20487 _invlist_union(only_utf8_locale, invlist, &invlist);
20488 _invlist_invert(invlist);
20489 _invlist_intersection(invlist, PL_InBitmap, &invlist);
20492 _invlist_invert(only_utf8);
20493 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
20495 else if (not_utf8) {
20497 /* If a code point matches iff the target string is not in UTF-8,
20498 * then complementing the result has it not match iff not in UTF-8,
20499 * which is the same thing as matching iff it is UTF-8. */
20500 only_utf8 = not_utf8;
20504 if (only_utf8_locale) {
20505 _invlist_invert(only_utf8_locale);
20506 _invlist_intersection(only_utf8_locale,
20508 &only_utf8_locale);
20511 inverted_display = put_charclass_bitmap_innards_common(
20516 only_utf8_locale, invert);
20518 /* Use the shortest representation, taking into account our bias
20519 * against showing it inverted */
20520 if ( inverted_display
20521 && ( ! as_is_display
20522 || ( SvCUR(inverted_display) + inverted_bias
20523 < SvCUR(as_is_display) + as_is_bias)))
20525 sv_catsv(sv, inverted_display);
20527 else if (as_is_display) {
20528 sv_catsv(sv, as_is_display);
20531 SvREFCNT_dec(as_is_display);
20532 SvREFCNT_dec(inverted_display);
20535 SvREFCNT_dec_NN(invlist);
20536 SvREFCNT_dec(only_utf8);
20537 SvREFCNT_dec(not_utf8);
20538 SvREFCNT_dec(posixes);
20539 SvREFCNT_dec(only_utf8_locale);
20541 return SvCUR(sv) > orig_sv_cur;
20544 #define CLEAR_OPTSTART \
20545 if (optstart) STMT_START { \
20546 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
20547 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
20551 #define DUMPUNTIL(b,e) \
20553 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
20555 STATIC const regnode *
20556 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
20557 const regnode *last, const regnode *plast,
20558 SV* sv, I32 indent, U32 depth)
20560 U8 op = PSEUDO; /* Arbitrary non-END op. */
20561 const regnode *next;
20562 const regnode *optstart= NULL;
20564 RXi_GET_DECL(r,ri);
20565 GET_RE_DEBUG_FLAGS_DECL;
20567 PERL_ARGS_ASSERT_DUMPUNTIL;
20569 #ifdef DEBUG_DUMPUNTIL
20570 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n",indent,node-start,
20571 last ? last-start : 0,plast ? plast-start : 0);
20574 if (plast && plast < last)
20577 while (PL_regkind[op] != END && (!last || node < last)) {
20579 /* While that wasn't END last time... */
20582 if (op == CLOSE || op == WHILEM)
20584 next = regnext((regnode *)node);
20587 if (OP(node) == OPTIMIZED) {
20588 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
20595 regprop(r, sv, node, NULL, NULL);
20596 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
20597 (int)(2*indent + 1), "", SvPVX_const(sv));
20599 if (OP(node) != OPTIMIZED) {
20600 if (next == NULL) /* Next ptr. */
20601 Perl_re_printf( aTHX_ " (0)");
20602 else if (PL_regkind[(U8)op] == BRANCH
20603 && PL_regkind[OP(next)] != BRANCH )
20604 Perl_re_printf( aTHX_ " (FAIL)");
20606 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
20607 Perl_re_printf( aTHX_ "\n");
20611 if (PL_regkind[(U8)op] == BRANCHJ) {
20614 const regnode *nnode = (OP(next) == LONGJMP
20615 ? regnext((regnode *)next)
20617 if (last && nnode > last)
20619 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
20622 else if (PL_regkind[(U8)op] == BRANCH) {
20624 DUMPUNTIL(NEXTOPER(node), next);
20626 else if ( PL_regkind[(U8)op] == TRIE ) {
20627 const regnode *this_trie = node;
20628 const char op = OP(node);
20629 const U32 n = ARG(node);
20630 const reg_ac_data * const ac = op>=AHOCORASICK ?
20631 (reg_ac_data *)ri->data->data[n] :
20633 const reg_trie_data * const trie =
20634 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
20636 AV *const trie_words
20637 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
20639 const regnode *nextbranch= NULL;
20642 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
20643 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
20645 Perl_re_indentf( aTHX_ "%s ",
20648 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
20649 SvCUR(*elem_ptr), 60,
20650 PL_colors[0], PL_colors[1],
20652 ? PERL_PV_ESCAPE_UNI
20654 | PERL_PV_PRETTY_ELLIPSES
20655 | PERL_PV_PRETTY_LTGT
20660 U16 dist= trie->jump[word_idx+1];
20661 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
20662 (UV)((dist ? this_trie + dist : next) - start));
20665 nextbranch= this_trie + trie->jump[0];
20666 DUMPUNTIL(this_trie + dist, nextbranch);
20668 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
20669 nextbranch= regnext((regnode *)nextbranch);
20671 Perl_re_printf( aTHX_ "\n");
20674 if (last && next > last)
20679 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
20680 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
20681 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
20683 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
20685 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
20687 else if ( op == PLUS || op == STAR) {
20688 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
20690 else if (PL_regkind[(U8)op] == ANYOF) {
20691 /* arglen 1 + class block */
20692 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
20693 ? ANYOF_POSIXL_SKIP
20695 node = NEXTOPER(node);
20697 else if (PL_regkind[(U8)op] == EXACT) {
20698 /* Literal string, where present. */
20699 node += NODE_SZ_STR(node) - 1;
20700 node = NEXTOPER(node);
20703 node = NEXTOPER(node);
20704 node += regarglen[(U8)op];
20706 if (op == CURLYX || op == OPEN)
20710 #ifdef DEBUG_DUMPUNTIL
20711 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
20716 #endif /* DEBUGGING */
20719 * ex: set ts=8 sts=4 sw=4 et: