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)
5221 && data && ++data->whilem_c < 16) {
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 PREVOPER(nxt)->flags = (U8)(data->whilem_c
5229 | (RExC_whilem_seen << 4)); /* On WHILEM */
5231 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5233 if (flags & SCF_DO_SUBSTR) {
5234 SV *last_str = NULL;
5235 STRLEN last_chrs = 0;
5236 int counted = mincount != 0;
5238 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5240 SSize_t b = pos_before >= data->last_start_min
5241 ? pos_before : data->last_start_min;
5243 const char * const s = SvPV_const(data->last_found, l);
5244 SSize_t old = b - data->last_start_min;
5247 old = utf8_hop((U8*)s, old) - (U8*)s;
5249 /* Get the added string: */
5250 last_str = newSVpvn_utf8(s + old, l, UTF);
5251 last_chrs = UTF ? utf8_length((U8*)(s + old),
5252 (U8*)(s + old + l)) : l;
5253 if (deltanext == 0 && pos_before == b) {
5254 /* What was added is a constant string */
5257 SvGROW(last_str, (mincount * l) + 1);
5258 repeatcpy(SvPVX(last_str) + l,
5259 SvPVX_const(last_str), l,
5261 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5262 /* Add additional parts. */
5263 SvCUR_set(data->last_found,
5264 SvCUR(data->last_found) - l);
5265 sv_catsv(data->last_found, last_str);
5267 SV * sv = data->last_found;
5269 SvUTF8(sv) && SvMAGICAL(sv) ?
5270 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5271 if (mg && mg->mg_len >= 0)
5272 mg->mg_len += last_chrs * (mincount-1);
5274 last_chrs *= mincount;
5275 data->last_end += l * (mincount - 1);
5278 /* start offset must point into the last copy */
5279 data->last_start_min += minnext * (mincount - 1);
5280 data->last_start_max =
5283 : data->last_start_max +
5284 (maxcount - 1) * (minnext + data->pos_delta);
5287 /* It is counted once already... */
5288 data->pos_min += minnext * (mincount - counted);
5290 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5291 " SSize_t_MAX=%" UVuf " minnext=%" UVuf
5292 " maxcount=%" UVuf " mincount=%" UVuf "\n",
5293 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
5295 if (deltanext != SSize_t_MAX)
5296 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5297 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5298 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
5300 if (deltanext == SSize_t_MAX
5301 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
5302 data->pos_delta = SSize_t_MAX;
5304 data->pos_delta += - counted * deltanext +
5305 (minnext + deltanext) * maxcount - minnext * mincount;
5306 if (mincount != maxcount) {
5307 /* Cannot extend fixed substrings found inside
5309 scan_commit(pRExC_state, data, minlenp, is_inf);
5310 if (mincount && last_str) {
5311 SV * const sv = data->last_found;
5312 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5313 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5317 sv_setsv(sv, last_str);
5318 data->last_end = data->pos_min;
5319 data->last_start_min = data->pos_min - last_chrs;
5320 data->last_start_max = is_inf
5322 : data->pos_min + data->pos_delta - last_chrs;
5324 data->longest = &(data->longest_float);
5326 SvREFCNT_dec(last_str);
5328 if (data && (fl & SF_HAS_EVAL))
5329 data->flags |= SF_HAS_EVAL;
5330 optimize_curly_tail:
5331 if (OP(oscan) != CURLYX) {
5332 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
5334 NEXT_OFF(oscan) += NEXT_OFF(next);
5340 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5345 if (flags & SCF_DO_SUBSTR) {
5346 /* Cannot expect anything... */
5347 scan_commit(pRExC_state, data, minlenp, is_inf);
5348 data->longest = &(data->longest_float);
5350 is_inf = is_inf_internal = 1;
5351 if (flags & SCF_DO_STCLASS_OR) {
5352 if (OP(scan) == CLUMP) {
5353 /* Actually is any start char, but very few code points
5354 * aren't start characters */
5355 ssc_match_all_cp(data->start_class);
5358 ssc_anything(data->start_class);
5361 flags &= ~SCF_DO_STCLASS;
5365 else if (OP(scan) == LNBREAK) {
5366 if (flags & SCF_DO_STCLASS) {
5367 if (flags & SCF_DO_STCLASS_AND) {
5368 ssc_intersection(data->start_class,
5369 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5370 ssc_clear_locale(data->start_class);
5371 ANYOF_FLAGS(data->start_class)
5372 &= ~SSC_MATCHES_EMPTY_STRING;
5374 else if (flags & SCF_DO_STCLASS_OR) {
5375 ssc_union(data->start_class,
5376 PL_XPosix_ptrs[_CC_VERTSPACE],
5378 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5380 /* See commit msg for
5381 * 749e076fceedeb708a624933726e7989f2302f6a */
5382 ANYOF_FLAGS(data->start_class)
5383 &= ~SSC_MATCHES_EMPTY_STRING;
5385 flags &= ~SCF_DO_STCLASS;
5388 if (delta != SSize_t_MAX)
5389 delta++; /* Because of the 2 char string cr-lf */
5390 if (flags & SCF_DO_SUBSTR) {
5391 /* Cannot expect anything... */
5392 scan_commit(pRExC_state, data, minlenp, is_inf);
5394 data->pos_delta += 1;
5395 data->longest = &(data->longest_float);
5398 else if (REGNODE_SIMPLE(OP(scan))) {
5400 if (flags & SCF_DO_SUBSTR) {
5401 scan_commit(pRExC_state, data, minlenp, is_inf);
5405 if (flags & SCF_DO_STCLASS) {
5407 SV* my_invlist = NULL;
5410 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5411 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5413 /* Some of the logic below assumes that switching
5414 locale on will only add false positives. */
5419 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5423 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5424 ssc_match_all_cp(data->start_class);
5429 SV* REG_ANY_invlist = _new_invlist(2);
5430 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5432 if (flags & SCF_DO_STCLASS_OR) {
5433 ssc_union(data->start_class,
5435 TRUE /* TRUE => invert, hence all but \n
5439 else if (flags & SCF_DO_STCLASS_AND) {
5440 ssc_intersection(data->start_class,
5442 TRUE /* TRUE => invert */
5444 ssc_clear_locale(data->start_class);
5446 SvREFCNT_dec_NN(REG_ANY_invlist);
5453 if (flags & SCF_DO_STCLASS_AND)
5454 ssc_and(pRExC_state, data->start_class,
5455 (regnode_charclass *) scan);
5457 ssc_or(pRExC_state, data->start_class,
5458 (regnode_charclass *) scan);
5466 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5467 if (flags & SCF_DO_STCLASS_AND) {
5468 bool was_there = cBOOL(
5469 ANYOF_POSIXL_TEST(data->start_class,
5471 ANYOF_POSIXL_ZERO(data->start_class);
5472 if (was_there) { /* Do an AND */
5473 ANYOF_POSIXL_SET(data->start_class, namedclass);
5475 /* No individual code points can now match */
5476 data->start_class->invlist
5477 = sv_2mortal(_new_invlist(0));
5480 int complement = namedclass + ((invert) ? -1 : 1);
5482 assert(flags & SCF_DO_STCLASS_OR);
5484 /* If the complement of this class was already there,
5485 * the result is that they match all code points,
5486 * (\d + \D == everything). Remove the classes from
5487 * future consideration. Locale is not relevant in
5489 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5490 ssc_match_all_cp(data->start_class);
5491 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5492 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5494 else { /* The usual case; just add this class to the
5496 ANYOF_POSIXL_SET(data->start_class, namedclass);
5501 case NPOSIXA: /* For these, we always know the exact set of
5506 if (FLAGS(scan) == _CC_ASCII) {
5507 my_invlist = invlist_clone(PL_XPosix_ptrs[_CC_ASCII]);
5510 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
5511 PL_XPosix_ptrs[_CC_ASCII],
5522 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5524 /* NPOSIXD matches all upper Latin1 code points unless the
5525 * target string being matched is UTF-8, which is
5526 * unknowable until match time. Since we are going to
5527 * invert, we want to get rid of all of them so that the
5528 * inversion will match all */
5529 if (OP(scan) == NPOSIXD) {
5530 _invlist_subtract(my_invlist, PL_UpperLatin1,
5536 if (flags & SCF_DO_STCLASS_AND) {
5537 ssc_intersection(data->start_class, my_invlist, invert);
5538 ssc_clear_locale(data->start_class);
5541 assert(flags & SCF_DO_STCLASS_OR);
5542 ssc_union(data->start_class, my_invlist, invert);
5544 SvREFCNT_dec(my_invlist);
5546 if (flags & SCF_DO_STCLASS_OR)
5547 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5548 flags &= ~SCF_DO_STCLASS;
5551 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5552 data->flags |= (OP(scan) == MEOL
5555 scan_commit(pRExC_state, data, minlenp, is_inf);
5558 else if ( PL_regkind[OP(scan)] == BRANCHJ
5559 /* Lookbehind, or need to calculate parens/evals/stclass: */
5560 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5561 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5563 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5564 || OP(scan) == UNLESSM )
5566 /* Negative Lookahead/lookbehind
5567 In this case we can't do fixed string optimisation.
5570 SSize_t deltanext, minnext, fake = 0;
5575 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5577 data_fake.whilem_c = data->whilem_c;
5578 data_fake.last_closep = data->last_closep;
5581 data_fake.last_closep = &fake;
5582 data_fake.pos_delta = delta;
5583 if ( flags & SCF_DO_STCLASS && !scan->flags
5584 && OP(scan) == IFMATCH ) { /* Lookahead */
5585 ssc_init(pRExC_state, &intrnl);
5586 data_fake.start_class = &intrnl;
5587 f |= SCF_DO_STCLASS_AND;
5589 if (flags & SCF_WHILEM_VISITED_POS)
5590 f |= SCF_WHILEM_VISITED_POS;
5591 next = regnext(scan);
5592 nscan = NEXTOPER(NEXTOPER(scan));
5593 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5594 last, &data_fake, stopparen,
5595 recursed_depth, NULL, f, depth+1);
5598 FAIL("Variable length lookbehind not implemented");
5600 else if (minnext > (I32)U8_MAX) {
5601 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5604 scan->flags = (U8)minnext;
5607 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5609 if (data_fake.flags & SF_HAS_EVAL)
5610 data->flags |= SF_HAS_EVAL;
5611 data->whilem_c = data_fake.whilem_c;
5613 if (f & SCF_DO_STCLASS_AND) {
5614 if (flags & SCF_DO_STCLASS_OR) {
5615 /* OR before, AND after: ideally we would recurse with
5616 * data_fake to get the AND applied by study of the
5617 * remainder of the pattern, and then derecurse;
5618 * *** HACK *** for now just treat as "no information".
5619 * See [perl #56690].
5621 ssc_init(pRExC_state, data->start_class);
5623 /* AND before and after: combine and continue. These
5624 * assertions are zero-length, so can match an EMPTY
5626 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5627 ANYOF_FLAGS(data->start_class)
5628 |= SSC_MATCHES_EMPTY_STRING;
5632 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5634 /* Positive Lookahead/lookbehind
5635 In this case we can do fixed string optimisation,
5636 but we must be careful about it. Note in the case of
5637 lookbehind the positions will be offset by the minimum
5638 length of the pattern, something we won't know about
5639 until after the recurse.
5641 SSize_t deltanext, fake = 0;
5645 /* We use SAVEFREEPV so that when the full compile
5646 is finished perl will clean up the allocated
5647 minlens when it's all done. This way we don't
5648 have to worry about freeing them when we know
5649 they wont be used, which would be a pain.
5652 Newx( minnextp, 1, SSize_t );
5653 SAVEFREEPV(minnextp);
5656 StructCopy(data, &data_fake, scan_data_t);
5657 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5660 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5661 data_fake.last_found=newSVsv(data->last_found);
5665 data_fake.last_closep = &fake;
5666 data_fake.flags = 0;
5667 data_fake.pos_delta = delta;
5669 data_fake.flags |= SF_IS_INF;
5670 if ( flags & SCF_DO_STCLASS && !scan->flags
5671 && OP(scan) == IFMATCH ) { /* Lookahead */
5672 ssc_init(pRExC_state, &intrnl);
5673 data_fake.start_class = &intrnl;
5674 f |= SCF_DO_STCLASS_AND;
5676 if (flags & SCF_WHILEM_VISITED_POS)
5677 f |= SCF_WHILEM_VISITED_POS;
5678 next = regnext(scan);
5679 nscan = NEXTOPER(NEXTOPER(scan));
5681 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5682 &deltanext, last, &data_fake,
5683 stopparen, recursed_depth, NULL,
5687 FAIL("Variable length lookbehind not implemented");
5689 else if (*minnextp > (I32)U8_MAX) {
5690 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5693 scan->flags = (U8)*minnextp;
5698 if (f & SCF_DO_STCLASS_AND) {
5699 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5700 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
5703 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5705 if (data_fake.flags & SF_HAS_EVAL)
5706 data->flags |= SF_HAS_EVAL;
5707 data->whilem_c = data_fake.whilem_c;
5708 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5709 if (RExC_rx->minlen<*minnextp)
5710 RExC_rx->minlen=*minnextp;
5711 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5712 SvREFCNT_dec_NN(data_fake.last_found);
5714 if ( data_fake.minlen_fixed != minlenp )
5716 data->offset_fixed= data_fake.offset_fixed;
5717 data->minlen_fixed= data_fake.minlen_fixed;
5718 data->lookbehind_fixed+= scan->flags;
5720 if ( data_fake.minlen_float != minlenp )
5722 data->minlen_float= data_fake.minlen_float;
5723 data->offset_float_min=data_fake.offset_float_min;
5724 data->offset_float_max=data_fake.offset_float_max;
5725 data->lookbehind_float+= scan->flags;
5732 else if (OP(scan) == OPEN) {
5733 if (stopparen != (I32)ARG(scan))
5736 else if (OP(scan) == CLOSE) {
5737 if (stopparen == (I32)ARG(scan)) {
5740 if ((I32)ARG(scan) == is_par) {
5741 next = regnext(scan);
5743 if ( next && (OP(next) != WHILEM) && next < last)
5744 is_par = 0; /* Disable optimization */
5747 *(data->last_closep) = ARG(scan);
5749 else if (OP(scan) == EVAL) {
5751 data->flags |= SF_HAS_EVAL;
5753 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5754 if (flags & SCF_DO_SUBSTR) {
5755 scan_commit(pRExC_state, data, minlenp, is_inf);
5756 flags &= ~SCF_DO_SUBSTR;
5758 if (data && OP(scan)==ACCEPT) {
5759 data->flags |= SCF_SEEN_ACCEPT;
5764 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5766 if (flags & SCF_DO_SUBSTR) {
5767 scan_commit(pRExC_state, data, minlenp, is_inf);
5768 data->longest = &(data->longest_float);
5770 is_inf = is_inf_internal = 1;
5771 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5772 ssc_anything(data->start_class);
5773 flags &= ~SCF_DO_STCLASS;
5775 else if (OP(scan) == GPOS) {
5776 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5777 !(delta || is_inf || (data && data->pos_delta)))
5779 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5780 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5781 if (RExC_rx->gofs < (STRLEN)min)
5782 RExC_rx->gofs = min;
5784 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5788 #ifdef TRIE_STUDY_OPT
5789 #ifdef FULL_TRIE_STUDY
5790 else if (PL_regkind[OP(scan)] == TRIE) {
5791 /* NOTE - There is similar code to this block above for handling
5792 BRANCH nodes on the initial study. If you change stuff here
5794 regnode *trie_node= scan;
5795 regnode *tail= regnext(scan);
5796 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5797 SSize_t max1 = 0, min1 = SSize_t_MAX;
5800 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5801 /* Cannot merge strings after this. */
5802 scan_commit(pRExC_state, data, minlenp, is_inf);
5804 if (flags & SCF_DO_STCLASS)
5805 ssc_init_zero(pRExC_state, &accum);
5811 const regnode *nextbranch= NULL;
5814 for ( word=1 ; word <= trie->wordcount ; word++)
5816 SSize_t deltanext=0, minnext=0, f = 0, fake;
5817 regnode_ssc this_class;
5819 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5821 data_fake.whilem_c = data->whilem_c;
5822 data_fake.last_closep = data->last_closep;
5825 data_fake.last_closep = &fake;
5826 data_fake.pos_delta = delta;
5827 if (flags & SCF_DO_STCLASS) {
5828 ssc_init(pRExC_state, &this_class);
5829 data_fake.start_class = &this_class;
5830 f = SCF_DO_STCLASS_AND;
5832 if (flags & SCF_WHILEM_VISITED_POS)
5833 f |= SCF_WHILEM_VISITED_POS;
5835 if (trie->jump[word]) {
5837 nextbranch = trie_node + trie->jump[0];
5838 scan= trie_node + trie->jump[word];
5839 /* We go from the jump point to the branch that follows
5840 it. Note this means we need the vestigal unused
5841 branches even though they arent otherwise used. */
5842 minnext = study_chunk(pRExC_state, &scan, minlenp,
5843 &deltanext, (regnode *)nextbranch, &data_fake,
5844 stopparen, recursed_depth, NULL, f,depth+1);
5846 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5847 nextbranch= regnext((regnode*)nextbranch);
5849 if (min1 > (SSize_t)(minnext + trie->minlen))
5850 min1 = minnext + trie->minlen;
5851 if (deltanext == SSize_t_MAX) {
5852 is_inf = is_inf_internal = 1;
5854 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5855 max1 = minnext + deltanext + trie->maxlen;
5857 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5859 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5860 if ( stopmin > min + min1)
5861 stopmin = min + min1;
5862 flags &= ~SCF_DO_SUBSTR;
5864 data->flags |= SCF_SEEN_ACCEPT;
5867 if (data_fake.flags & SF_HAS_EVAL)
5868 data->flags |= SF_HAS_EVAL;
5869 data->whilem_c = data_fake.whilem_c;
5871 if (flags & SCF_DO_STCLASS)
5872 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5875 if (flags & SCF_DO_SUBSTR) {
5876 data->pos_min += min1;
5877 data->pos_delta += max1 - min1;
5878 if (max1 != min1 || is_inf)
5879 data->longest = &(data->longest_float);
5882 if (delta != SSize_t_MAX)
5883 delta += max1 - min1;
5884 if (flags & SCF_DO_STCLASS_OR) {
5885 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5887 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5888 flags &= ~SCF_DO_STCLASS;
5891 else if (flags & SCF_DO_STCLASS_AND) {
5893 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5894 flags &= ~SCF_DO_STCLASS;
5897 /* Switch to OR mode: cache the old value of
5898 * data->start_class */
5900 StructCopy(data->start_class, and_withp, regnode_ssc);
5901 flags &= ~SCF_DO_STCLASS_AND;
5902 StructCopy(&accum, data->start_class, regnode_ssc);
5903 flags |= SCF_DO_STCLASS_OR;
5910 else if (PL_regkind[OP(scan)] == TRIE) {
5911 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5914 min += trie->minlen;
5915 delta += (trie->maxlen - trie->minlen);
5916 flags &= ~SCF_DO_STCLASS; /* xxx */
5917 if (flags & SCF_DO_SUBSTR) {
5918 /* Cannot expect anything... */
5919 scan_commit(pRExC_state, data, minlenp, is_inf);
5920 data->pos_min += trie->minlen;
5921 data->pos_delta += (trie->maxlen - trie->minlen);
5922 if (trie->maxlen != trie->minlen)
5923 data->longest = &(data->longest_float);
5925 if (trie->jump) /* no more substrings -- for now /grr*/
5926 flags &= ~SCF_DO_SUBSTR;
5928 #endif /* old or new */
5929 #endif /* TRIE_STUDY_OPT */
5931 /* Else: zero-length, ignore. */
5932 scan = regnext(scan);
5937 /* we need to unwind recursion. */
5940 DEBUG_STUDYDATA("frame-end:",data,depth);
5941 DEBUG_PEEP("fend", scan, depth);
5943 /* restore previous context */
5944 last = frame->last_regnode;
5945 scan = frame->next_regnode;
5946 stopparen = frame->stopparen;
5947 recursed_depth = frame->prev_recursed_depth;
5949 RExC_frame_last = frame->prev_frame;
5950 frame = frame->this_prev_frame;
5951 goto fake_study_recurse;
5955 DEBUG_STUDYDATA("pre-fin:",data,depth);
5958 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5960 if (flags & SCF_DO_SUBSTR && is_inf)
5961 data->pos_delta = SSize_t_MAX - data->pos_min;
5962 if (is_par > (I32)U8_MAX)
5964 if (is_par && pars==1 && data) {
5965 data->flags |= SF_IN_PAR;
5966 data->flags &= ~SF_HAS_PAR;
5968 else if (pars && data) {
5969 data->flags |= SF_HAS_PAR;
5970 data->flags &= ~SF_IN_PAR;
5972 if (flags & SCF_DO_STCLASS_OR)
5973 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5974 if (flags & SCF_TRIE_RESTUDY)
5975 data->flags |= SCF_TRIE_RESTUDY;
5977 DEBUG_STUDYDATA("post-fin:",data,depth);
5980 SSize_t final_minlen= min < stopmin ? min : stopmin;
5982 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
5983 if (final_minlen > SSize_t_MAX - delta)
5984 RExC_maxlen = SSize_t_MAX;
5985 else if (RExC_maxlen < final_minlen + delta)
5986 RExC_maxlen = final_minlen + delta;
5988 return final_minlen;
5990 NOT_REACHED; /* NOTREACHED */
5994 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5996 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5998 PERL_ARGS_ASSERT_ADD_DATA;
6000 Renewc(RExC_rxi->data,
6001 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6002 char, struct reg_data);
6004 Renew(RExC_rxi->data->what, count + n, U8);
6006 Newx(RExC_rxi->data->what, n, U8);
6007 RExC_rxi->data->count = count + n;
6008 Copy(s, RExC_rxi->data->what + count, n, U8);
6012 /*XXX: todo make this not included in a non debugging perl, but appears to be
6013 * used anyway there, in 'use re' */
6014 #ifndef PERL_IN_XSUB_RE
6016 Perl_reginitcolors(pTHX)
6018 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6020 char *t = savepv(s);
6024 t = strchr(t, '\t');
6030 PL_colors[i] = t = (char *)"";
6035 PL_colors[i++] = (char *)"";
6042 #ifdef TRIE_STUDY_OPT
6043 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6046 (data.flags & SCF_TRIE_RESTUDY) \
6054 #define CHECK_RESTUDY_GOTO_butfirst
6058 * pregcomp - compile a regular expression into internal code
6060 * Decides which engine's compiler to call based on the hint currently in
6064 #ifndef PERL_IN_XSUB_RE
6066 /* return the currently in-scope regex engine (or the default if none) */
6068 regexp_engine const *
6069 Perl_current_re_engine(pTHX)
6071 if (IN_PERL_COMPILETIME) {
6072 HV * const table = GvHV(PL_hintgv);
6075 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6076 return &PL_core_reg_engine;
6077 ptr = hv_fetchs(table, "regcomp", FALSE);
6078 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6079 return &PL_core_reg_engine;
6080 return INT2PTR(regexp_engine*,SvIV(*ptr));
6084 if (!PL_curcop->cop_hints_hash)
6085 return &PL_core_reg_engine;
6086 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6087 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6088 return &PL_core_reg_engine;
6089 return INT2PTR(regexp_engine*,SvIV(ptr));
6095 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6097 regexp_engine const *eng = current_re_engine();
6098 GET_RE_DEBUG_FLAGS_DECL;
6100 PERL_ARGS_ASSERT_PREGCOMP;
6102 /* Dispatch a request to compile a regexp to correct regexp engine. */
6104 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6107 return CALLREGCOMP_ENG(eng, pattern, flags);
6111 /* public(ish) entry point for the perl core's own regex compiling code.
6112 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6113 * pattern rather than a list of OPs, and uses the internal engine rather
6114 * than the current one */
6117 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6119 SV *pat = pattern; /* defeat constness! */
6120 PERL_ARGS_ASSERT_RE_COMPILE;
6121 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6122 #ifdef PERL_IN_XSUB_RE
6125 &PL_core_reg_engine,
6127 NULL, NULL, rx_flags, 0);
6132 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6138 for (n = 0; n < cbs->count; n++)
6139 SvREFCNT_dec(cbs->cb[n].src_regex);
6145 static struct reg_code_blocks *
6146 S_alloc_code_blocks(pTHX_ int ncode)
6148 struct reg_code_blocks *cbs;
6149 Newx(cbs, 1, struct reg_code_blocks);
6151 cbs->attached = FALSE;
6152 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6154 Newx(cbs->cb, ncode, struct reg_code_block);
6161 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6162 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6163 * point to the realloced string and length.
6165 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6169 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6170 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6172 U8 *const src = (U8*)*pat_p;
6177 GET_RE_DEBUG_FLAGS_DECL;
6179 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6180 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6182 Newx(dst, *plen_p * 2 + 1, U8);
6185 while (s < *plen_p) {
6186 append_utf8_from_native_byte(src[s], &d);
6187 if (n < num_code_blocks) {
6188 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6189 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6190 assert(*(d - 1) == '(');
6193 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6194 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6195 assert(*(d - 1) == ')');
6204 *pat_p = (char*) dst;
6206 RExC_orig_utf8 = RExC_utf8 = 1;
6211 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6212 * while recording any code block indices, and handling overloading,
6213 * nested qr// objects etc. If pat is null, it will allocate a new
6214 * string, or just return the first arg, if there's only one.
6216 * Returns the malloced/updated pat.
6217 * patternp and pat_count is the array of SVs to be concatted;
6218 * oplist is the optional list of ops that generated the SVs;
6219 * recompile_p is a pointer to a boolean that will be set if
6220 * the regex will need to be recompiled.
6221 * delim, if non-null is an SV that will be inserted between each element
6225 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6226 SV *pat, SV ** const patternp, int pat_count,
6227 OP *oplist, bool *recompile_p, SV *delim)
6231 bool use_delim = FALSE;
6232 bool alloced = FALSE;
6234 /* if we know we have at least two args, create an empty string,
6235 * then concatenate args to that. For no args, return an empty string */
6236 if (!pat && pat_count != 1) {
6242 for (svp = patternp; svp < patternp + pat_count; svp++) {
6245 STRLEN orig_patlen = 0;
6247 SV *msv = use_delim ? delim : *svp;
6248 if (!msv) msv = &PL_sv_undef;
6250 /* if we've got a delimiter, we go round the loop twice for each
6251 * svp slot (except the last), using the delimiter the second
6260 if (SvTYPE(msv) == SVt_PVAV) {
6261 /* we've encountered an interpolated array within
6262 * the pattern, e.g. /...@a..../. Expand the list of elements,
6263 * then recursively append elements.
6264 * The code in this block is based on S_pushav() */
6266 AV *const av = (AV*)msv;
6267 const SSize_t maxarg = AvFILL(av) + 1;
6271 assert(oplist->op_type == OP_PADAV
6272 || oplist->op_type == OP_RV2AV);
6273 oplist = OpSIBLING(oplist);
6276 if (SvRMAGICAL(av)) {
6279 Newx(array, maxarg, SV*);
6281 for (i=0; i < maxarg; i++) {
6282 SV ** const svp = av_fetch(av, i, FALSE);
6283 array[i] = svp ? *svp : &PL_sv_undef;
6287 array = AvARRAY(av);
6289 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6290 array, maxarg, NULL, recompile_p,
6292 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6298 /* we make the assumption here that each op in the list of
6299 * op_siblings maps to one SV pushed onto the stack,
6300 * except for code blocks, with have both an OP_NULL and
6302 * This allows us to match up the list of SVs against the
6303 * list of OPs to find the next code block.
6305 * Note that PUSHMARK PADSV PADSV ..
6307 * PADRANGE PADSV PADSV ..
6308 * so the alignment still works. */
6311 if (oplist->op_type == OP_NULL
6312 && (oplist->op_flags & OPf_SPECIAL))
6314 assert(n < pRExC_state->code_blocks->count);
6315 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6316 pRExC_state->code_blocks->cb[n].block = oplist;
6317 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6320 oplist = OpSIBLING(oplist); /* skip CONST */
6323 oplist = OpSIBLING(oplist);;
6326 /* apply magic and QR overloading to arg */
6329 if (SvROK(msv) && SvAMAGIC(msv)) {
6330 SV *sv = AMG_CALLunary(msv, regexp_amg);
6334 if (SvTYPE(sv) != SVt_REGEXP)
6335 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6340 /* try concatenation overload ... */
6341 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6342 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6345 /* overloading involved: all bets are off over literal
6346 * code. Pretend we haven't seen it */
6348 pRExC_state->code_blocks->count -= n;
6352 /* ... or failing that, try "" overload */
6353 while (SvAMAGIC(msv)
6354 && (sv = AMG_CALLunary(msv, string_amg))
6358 && SvRV(msv) == SvRV(sv))
6363 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6367 /* this is a partially unrolled
6368 * sv_catsv_nomg(pat, msv);
6369 * that allows us to adjust code block indices if
6372 char *dst = SvPV_force_nomg(pat, dlen);
6374 if (SvUTF8(msv) && !SvUTF8(pat)) {
6375 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6376 sv_setpvn(pat, dst, dlen);
6379 sv_catsv_nomg(pat, msv);
6383 /* We have only one SV to process, but we need to verify
6384 * it is properly null terminated or we will fail asserts
6385 * later. In theory we probably shouldn't get such SV's,
6386 * but if we do we should handle it gracefully. */
6387 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) ) {
6388 /* not a string, or a string with a trailing null */
6391 /* a string with no trailing null, we need to copy it
6392 * so it we have a trailing null */
6398 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
6401 /* extract any code blocks within any embedded qr//'s */
6402 if (rx && SvTYPE(rx) == SVt_REGEXP
6403 && RX_ENGINE((REGEXP*)rx)->op_comp)
6406 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6407 if (ri->code_blocks && ri->code_blocks->count) {
6409 /* the presence of an embedded qr// with code means
6410 * we should always recompile: the text of the
6411 * qr// may not have changed, but it may be a
6412 * different closure than last time */
6414 if (pRExC_state->code_blocks) {
6415 pRExC_state->code_blocks->count += ri->code_blocks->count;
6416 Renew(pRExC_state->code_blocks->cb,
6417 pRExC_state->code_blocks->count,
6418 struct reg_code_block);
6421 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
6422 ri->code_blocks->count);
6424 for (i=0; i < ri->code_blocks->count; i++) {
6425 struct reg_code_block *src, *dst;
6426 STRLEN offset = orig_patlen
6427 + ReANY((REGEXP *)rx)->pre_prefix;
6428 assert(n < pRExC_state->code_blocks->count);
6429 src = &ri->code_blocks->cb[i];
6430 dst = &pRExC_state->code_blocks->cb[n];
6431 dst->start = src->start + offset;
6432 dst->end = src->end + offset;
6433 dst->block = src->block;
6434 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6443 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6452 /* see if there are any run-time code blocks in the pattern.
6453 * False positives are allowed */
6456 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6457 char *pat, STRLEN plen)
6462 PERL_UNUSED_CONTEXT;
6464 for (s = 0; s < plen; s++) {
6465 if ( pRExC_state->code_blocks
6466 && n < pRExC_state->code_blocks->count
6467 && s == pRExC_state->code_blocks->cb[n].start)
6469 s = pRExC_state->code_blocks->cb[n].end;
6473 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6475 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6477 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6484 /* Handle run-time code blocks. We will already have compiled any direct
6485 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6486 * copy of it, but with any literal code blocks blanked out and
6487 * appropriate chars escaped; then feed it into
6489 * eval "qr'modified_pattern'"
6493 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6497 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6499 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6500 * and merge them with any code blocks of the original regexp.
6502 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6503 * instead, just save the qr and return FALSE; this tells our caller that
6504 * the original pattern needs upgrading to utf8.
6508 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6509 char *pat, STRLEN plen)
6513 GET_RE_DEBUG_FLAGS_DECL;
6515 if (pRExC_state->runtime_code_qr) {
6516 /* this is the second time we've been called; this should
6517 * only happen if the main pattern got upgraded to utf8
6518 * during compilation; re-use the qr we compiled first time
6519 * round (which should be utf8 too)
6521 qr = pRExC_state->runtime_code_qr;
6522 pRExC_state->runtime_code_qr = NULL;
6523 assert(RExC_utf8 && SvUTF8(qr));
6529 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
6533 /* determine how many extra chars we need for ' and \ escaping */
6534 for (s = 0; s < plen; s++) {
6535 if (pat[s] == '\'' || pat[s] == '\\')
6539 Newx(newpat, newlen, char);
6541 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6543 for (s = 0; s < plen; s++) {
6544 if ( pRExC_state->code_blocks
6545 && n < pRExC_state->code_blocks->count
6546 && s == pRExC_state->code_blocks->cb[n].start)
6548 /* blank out literal code block */
6549 assert(pat[s] == '(');
6550 while (s <= pRExC_state->code_blocks->cb[n].end) {
6558 if (pat[s] == '\'' || pat[s] == '\\')
6563 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
6565 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
6571 Perl_re_printf( aTHX_
6572 "%sre-parsing pattern for runtime code:%s %s\n",
6573 PL_colors[4],PL_colors[5],newpat);
6576 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6582 PUSHSTACKi(PERLSI_REQUIRE);
6583 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6584 * parsing qr''; normally only q'' does this. It also alters
6586 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6587 SvREFCNT_dec_NN(sv);
6592 SV * const errsv = ERRSV;
6593 if (SvTRUE_NN(errsv))
6594 /* use croak_sv ? */
6595 Perl_croak_nocontext("%"SVf, SVfARG(errsv));
6597 assert(SvROK(qr_ref));
6599 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6600 /* the leaving below frees the tmp qr_ref.
6601 * Give qr a life of its own */
6609 if (!RExC_utf8 && SvUTF8(qr)) {
6610 /* first time through; the pattern got upgraded; save the
6611 * qr for the next time through */
6612 assert(!pRExC_state->runtime_code_qr);
6613 pRExC_state->runtime_code_qr = qr;
6618 /* extract any code blocks within the returned qr// */
6621 /* merge the main (r1) and run-time (r2) code blocks into one */
6623 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6624 struct reg_code_block *new_block, *dst;
6625 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6629 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
6631 SvREFCNT_dec_NN(qr);
6635 if (!r1->code_blocks)
6636 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
6638 r1c = r1->code_blocks->count;
6639 r2c = r2->code_blocks->count;
6641 Newx(new_block, r1c + r2c, struct reg_code_block);
6645 while (i1 < r1c || i2 < r2c) {
6646 struct reg_code_block *src;
6650 src = &r2->code_blocks->cb[i2++];
6654 src = &r1->code_blocks->cb[i1++];
6655 else if ( r1->code_blocks->cb[i1].start
6656 < r2->code_blocks->cb[i2].start)
6658 src = &r1->code_blocks->cb[i1++];
6659 assert(src->end < r2->code_blocks->cb[i2].start);
6662 assert( r1->code_blocks->cb[i1].start
6663 > r2->code_blocks->cb[i2].start);
6664 src = &r2->code_blocks->cb[i2++];
6666 assert(src->end < r1->code_blocks->cb[i1].start);
6669 assert(pat[src->start] == '(');
6670 assert(pat[src->end] == ')');
6671 dst->start = src->start;
6672 dst->end = src->end;
6673 dst->block = src->block;
6674 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6678 r1->code_blocks->count += r2c;
6679 Safefree(r1->code_blocks->cb);
6680 r1->code_blocks->cb = new_block;
6683 SvREFCNT_dec_NN(qr);
6689 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6690 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6691 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6692 STRLEN longest_length, bool eol, bool meol)
6694 /* This is the common code for setting up the floating and fixed length
6695 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6696 * as to whether succeeded or not */
6701 if (! (longest_length
6702 || (eol /* Can't have SEOL and MULTI */
6703 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6705 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6706 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6711 /* copy the information about the longest from the reg_scan_data
6712 over to the program. */
6713 if (SvUTF8(sv_longest)) {
6714 *rx_utf8 = sv_longest;
6717 *rx_substr = sv_longest;
6720 /* end_shift is how many chars that must be matched that
6721 follow this item. We calculate it ahead of time as once the
6722 lookbehind offset is added in we lose the ability to correctly
6724 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6725 *rx_end_shift = ml - offset
6727 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
6729 + (SvTAIL(sv_longest) != 0)
6733 t = (eol/* Can't have SEOL and MULTI */
6734 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6735 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6741 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6742 * regular expression into internal code.
6743 * The pattern may be passed either as:
6744 * a list of SVs (patternp plus pat_count)
6745 * a list of OPs (expr)
6746 * If both are passed, the SV list is used, but the OP list indicates
6747 * which SVs are actually pre-compiled code blocks
6749 * The SVs in the list have magic and qr overloading applied to them (and
6750 * the list may be modified in-place with replacement SVs in the latter
6753 * If the pattern hasn't changed from old_re, then old_re will be
6756 * eng is the current engine. If that engine has an op_comp method, then
6757 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6758 * do the initial concatenation of arguments and pass on to the external
6761 * If is_bare_re is not null, set it to a boolean indicating whether the
6762 * arg list reduced (after overloading) to a single bare regex which has
6763 * been returned (i.e. /$qr/).
6765 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6767 * pm_flags contains the PMf_* flags, typically based on those from the
6768 * pm_flags field of the related PMOP. Currently we're only interested in
6769 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6771 * We can't allocate space until we know how big the compiled form will be,
6772 * but we can't compile it (and thus know how big it is) until we've got a
6773 * place to put the code. So we cheat: we compile it twice, once with code
6774 * generation turned off and size counting turned on, and once "for real".
6775 * This also means that we don't allocate space until we are sure that the
6776 * thing really will compile successfully, and we never have to move the
6777 * code and thus invalidate pointers into it. (Note that it has to be in
6778 * one piece because free() must be able to free it all.) [NB: not true in perl]
6780 * Beware that the optimization-preparation code in here knows about some
6781 * of the structure of the compiled regexp. [I'll say.]
6785 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6786 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6787 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6791 regexp_internal *ri;
6799 SV** new_patternp = patternp;
6801 /* these are all flags - maybe they should be turned
6802 * into a single int with different bit masks */
6803 I32 sawlookahead = 0;
6808 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6810 bool runtime_code = 0;
6812 RExC_state_t RExC_state;
6813 RExC_state_t * const pRExC_state = &RExC_state;
6814 #ifdef TRIE_STUDY_OPT
6816 RExC_state_t copyRExC_state;
6818 GET_RE_DEBUG_FLAGS_DECL;
6820 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6822 DEBUG_r(if (!PL_colorset) reginitcolors());
6824 /* Initialize these here instead of as-needed, as is quick and avoids
6825 * having to test them each time otherwise */
6826 if (! PL_AboveLatin1) {
6828 char * dump_len_string;
6831 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6832 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6833 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6834 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6835 PL_HasMultiCharFold =
6836 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6838 /* This is calculated here, because the Perl program that generates the
6839 * static global ones doesn't currently have access to
6840 * NUM_ANYOF_CODE_POINTS */
6841 PL_InBitmap = _new_invlist(2);
6842 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
6843 NUM_ANYOF_CODE_POINTS - 1);
6845 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
6846 if ( ! dump_len_string
6847 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
6849 PL_dump_re_max_len = 0;
6854 pRExC_state->warn_text = NULL;
6855 pRExC_state->code_blocks = NULL;
6858 *is_bare_re = FALSE;
6860 if (expr && (expr->op_type == OP_LIST ||
6861 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6862 /* allocate code_blocks if needed */
6866 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
6867 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6868 ncode++; /* count of DO blocks */
6871 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
6875 /* compile-time pattern with just OP_CONSTs and DO blocks */
6880 /* find how many CONSTs there are */
6883 if (expr->op_type == OP_CONST)
6886 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6887 if (o->op_type == OP_CONST)
6891 /* fake up an SV array */
6893 assert(!new_patternp);
6894 Newx(new_patternp, n, SV*);
6895 SAVEFREEPV(new_patternp);
6899 if (expr->op_type == OP_CONST)
6900 new_patternp[n] = cSVOPx_sv(expr);
6902 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6903 if (o->op_type == OP_CONST)
6904 new_patternp[n++] = cSVOPo_sv;
6909 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6910 "Assembling pattern from %d elements%s\n", pat_count,
6911 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6913 /* set expr to the first arg op */
6915 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
6916 && expr->op_type != OP_CONST)
6918 expr = cLISTOPx(expr)->op_first;
6919 assert( expr->op_type == OP_PUSHMARK
6920 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6921 || expr->op_type == OP_PADRANGE);
6922 expr = OpSIBLING(expr);
6925 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6926 expr, &recompile, NULL);
6928 /* handle bare (possibly after overloading) regex: foo =~ $re */
6933 if (SvTYPE(re) == SVt_REGEXP) {
6937 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6938 "Precompiled pattern%s\n",
6939 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6945 exp = SvPV_nomg(pat, plen);
6947 if (!eng->op_comp) {
6948 if ((SvUTF8(pat) && IN_BYTES)
6949 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6951 /* make a temporary copy; either to convert to bytes,
6952 * or to avoid repeating get-magic / overloaded stringify */
6953 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6954 (IN_BYTES ? 0 : SvUTF8(pat)));
6956 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6959 /* ignore the utf8ness if the pattern is 0 length */
6960 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6962 RExC_uni_semantics = 0;
6963 RExC_seen_unfolded_sharp_s = 0;
6964 RExC_contains_locale = 0;
6965 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
6966 RExC_study_started = 0;
6967 pRExC_state->runtime_code_qr = NULL;
6968 RExC_frame_head= NULL;
6969 RExC_frame_last= NULL;
6970 RExC_frame_count= 0;
6973 RExC_mysv1= sv_newmortal();
6974 RExC_mysv2= sv_newmortal();
6977 SV *dsv= sv_newmortal();
6978 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6979 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
6980 PL_colors[4],PL_colors[5],s);
6984 /* we jump here if we have to recompile, e.g., from upgrading the pattern
6987 if ((pm_flags & PMf_USE_RE_EVAL)
6988 /* this second condition covers the non-regex literal case,
6989 * i.e. $foo =~ '(?{})'. */
6990 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6992 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6994 /* return old regex if pattern hasn't changed */
6995 /* XXX: note in the below we have to check the flags as well as the
6998 * Things get a touch tricky as we have to compare the utf8 flag
6999 * independently from the compile flags. */
7003 && !!RX_UTF8(old_re) == !!RExC_utf8
7004 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7005 && RX_PRECOMP(old_re)
7006 && RX_PRELEN(old_re) == plen
7007 && memEQ(RX_PRECOMP(old_re), exp, plen)
7008 && !runtime_code /* with runtime code, always recompile */ )
7013 rx_flags = orig_rx_flags;
7015 if ( initial_charset == REGEX_DEPENDS_CHARSET
7016 && (RExC_utf8 ||RExC_uni_semantics))
7019 /* Set to use unicode semantics if the pattern is in utf8 and has the
7020 * 'depends' charset specified, as it means unicode when utf8 */
7021 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7025 RExC_precomp_adj = 0;
7026 RExC_flags = rx_flags;
7027 RExC_pm_flags = pm_flags;
7030 assert(TAINTING_get || !TAINT_get);
7032 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7034 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7035 /* whoops, we have a non-utf8 pattern, whilst run-time code
7036 * got compiled as utf8. Try again with a utf8 pattern */
7037 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7038 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7039 goto redo_first_pass;
7042 assert(!pRExC_state->runtime_code_qr);
7048 RExC_in_lookbehind = 0;
7049 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7052 RExC_recode_x_to_native = 0;
7054 RExC_in_multi_char_class = 0;
7056 /* First pass: determine size, legality. */
7058 RExC_start = RExC_adjusted_start = exp;
7059 RExC_end = exp + plen;
7060 RExC_precomp_end = RExC_end;
7065 RExC_emit = (regnode *) &RExC_emit_dummy;
7066 RExC_whilem_seen = 0;
7067 RExC_open_parens = NULL;
7068 RExC_close_parens = NULL;
7070 RExC_paren_names = NULL;
7072 RExC_paren_name_list = NULL;
7074 RExC_recurse = NULL;
7075 RExC_study_chunk_recursed = NULL;
7076 RExC_study_chunk_recursed_bytes= 0;
7077 RExC_recurse_count = 0;
7078 pRExC_state->code_index = 0;
7080 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7081 * code makes sure the final byte is an uncounted NUL. But should this
7082 * ever not be the case, lots of things could read beyond the end of the
7083 * buffer: loops like
7084 * while(isFOO(*RExC_parse)) RExC_parse++;
7085 * strchr(RExC_parse, "foo");
7086 * etc. So it is worth noting. */
7087 assert(*RExC_end == '\0');
7090 Perl_re_printf( aTHX_ "Starting first pass (sizing)\n");
7092 RExC_lastparse=NULL;
7095 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7096 /* It's possible to write a regexp in ascii that represents Unicode
7097 codepoints outside of the byte range, such as via \x{100}. If we
7098 detect such a sequence we have to convert the entire pattern to utf8
7099 and then recompile, as our sizing calculation will have been based
7100 on 1 byte == 1 character, but we will need to use utf8 to encode
7101 at least some part of the pattern, and therefore must convert the whole
7104 if (flags & RESTART_PASS1) {
7105 if (flags & NEED_UTF8) {
7106 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7107 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7110 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7111 "Need to redo pass 1\n"));
7114 goto redo_first_pass;
7116 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#" UVxf, (UV) flags);
7120 Perl_re_printf( aTHX_
7121 "Required size %" IVdf " nodes\n"
7122 "Starting second pass (creation)\n",
7125 RExC_lastparse=NULL;
7128 /* The first pass could have found things that force Unicode semantics */
7129 if ((RExC_utf8 || RExC_uni_semantics)
7130 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
7132 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7135 /* Small enough for pointer-storage convention?
7136 If extralen==0, this means that we will not need long jumps. */
7137 if (RExC_size >= 0x10000L && RExC_extralen)
7138 RExC_size += RExC_extralen;
7141 if (RExC_whilem_seen > 15)
7142 RExC_whilem_seen = 15;
7144 /* Allocate space and zero-initialize. Note, the two step process
7145 of zeroing when in debug mode, thus anything assigned has to
7146 happen after that */
7147 rx = (REGEXP*) newSV_type(SVt_REGEXP);
7149 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7150 char, regexp_internal);
7151 if ( r == NULL || ri == NULL )
7152 FAIL("Regexp out of space");
7154 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
7155 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7158 /* bulk initialize base fields with 0. */
7159 Zero(ri, sizeof(regexp_internal), char);
7162 /* non-zero initialization begins here */
7165 r->extflags = rx_flags;
7166 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7168 if (pm_flags & PMf_IS_QR) {
7169 ri->code_blocks = pRExC_state->code_blocks;
7170 if (ri->code_blocks)
7171 /* disarm earlier SAVEDESTRUCTOR_X */
7172 ri->code_blocks->attached = TRUE;
7176 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7177 bool has_charset = (get_regex_charset(r->extflags)
7178 != REGEX_DEPENDS_CHARSET);
7180 /* The caret is output if there are any defaults: if not all the STD
7181 * flags are set, or if no character set specifier is needed */
7183 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7185 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7186 == REG_RUN_ON_COMMENT_SEEN);
7187 U8 reganch = (U8)((r->extflags & RXf_PMf_STD_PMMOD)
7188 >> RXf_PMf_STD_PMMOD_SHIFT);
7189 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7192 /* We output all the necessary flags; we never output a minus, as all
7193 * those are defaults, so are
7194 * covered by the caret */
7195 const STRLEN wraplen = plen + has_p + has_runon
7196 + has_default /* If needs a caret */
7197 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7199 /* If needs a character set specifier */
7200 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7201 + (sizeof("(?:)") - 1);
7203 /* make sure PL_bitcount bounds not exceeded */
7204 assert(sizeof(STD_PAT_MODS) <= 8);
7206 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
7207 r->xpv_len_u.xpvlenu_pv = p;
7209 SvFLAGS(rx) |= SVf_UTF8;
7212 /* If a default, cover it using the caret */
7214 *p++= DEFAULT_PAT_MOD;
7218 const char* const name = get_regex_charset_name(r->extflags, &len);
7219 Copy(name, p, len, char);
7223 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7226 while((ch = *fptr++)) {
7234 Copy(RExC_precomp, p, plen, char);
7235 assert ((RX_WRAPPED(rx) - p) < 16);
7236 r->pre_prefix = p - RX_WRAPPED(rx);
7242 SvCUR_set(rx, p - RX_WRAPPED(rx));
7246 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
7248 /* Useful during FAIL. */
7249 #ifdef RE_TRACK_PATTERN_OFFSETS
7250 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
7251 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7252 "%s %" UVuf " bytes for offset annotations.\n",
7253 ri->u.offsets ? "Got" : "Couldn't get",
7254 (UV)((2*RExC_size+1) * sizeof(U32))));
7256 SetProgLen(ri,RExC_size);
7261 /* Second pass: emit code. */
7262 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7263 RExC_pm_flags = pm_flags;
7265 RExC_end = exp + plen;
7267 RExC_emit_start = ri->program;
7268 RExC_emit = ri->program;
7269 RExC_emit_bound = ri->program + RExC_size + 1;
7270 pRExC_state->code_index = 0;
7272 *((char*) RExC_emit++) = (char) REG_MAGIC;
7273 /* setup various meta data about recursion, this all requires
7274 * RExC_npar to be correctly set, and a bit later on we clear it */
7275 if (RExC_seen & REG_RECURSE_SEEN) {
7276 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
7277 "%*s%*s Setting up open/close parens\n",
7278 22, "| |", (int)(0 * 2 + 1), ""));
7280 /* setup RExC_open_parens, which holds the address of each
7281 * OPEN tag, and to make things simpler for the 0 index
7282 * the start of the program - this is used later for offsets */
7283 Newxz(RExC_open_parens, RExC_npar,regnode *);
7284 SAVEFREEPV(RExC_open_parens);
7285 RExC_open_parens[0] = RExC_emit;
7287 /* setup RExC_close_parens, which holds the address of each
7288 * CLOSE tag, and to make things simpler for the 0 index
7289 * the end of the program - this is used later for offsets */
7290 Newxz(RExC_close_parens, RExC_npar,regnode *);
7291 SAVEFREEPV(RExC_close_parens);
7292 /* we dont know where end op starts yet, so we dont
7293 * need to set RExC_close_parens[0] like we do RExC_open_parens[0] above */
7295 /* Note, RExC_npar is 1 + the number of parens in a pattern.
7296 * So its 1 if there are no parens. */
7297 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
7298 ((RExC_npar & 0x07) != 0);
7299 Newx(RExC_study_chunk_recursed,
7300 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7301 SAVEFREEPV(RExC_study_chunk_recursed);
7304 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7306 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#" UVxf, (UV) flags);
7309 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7312 /* XXXX To minimize changes to RE engine we always allocate
7313 3-units-long substrs field. */
7314 Newx(r->substrs, 1, struct reg_substr_data);
7315 if (RExC_recurse_count) {
7316 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
7317 SAVEFREEPV(RExC_recurse);
7321 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7323 RExC_study_chunk_recursed_count= 0;
7325 Zero(r->substrs, 1, struct reg_substr_data);
7326 if (RExC_study_chunk_recursed) {
7327 Zero(RExC_study_chunk_recursed,
7328 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7332 #ifdef TRIE_STUDY_OPT
7334 StructCopy(&zero_scan_data, &data, scan_data_t);
7335 copyRExC_state = RExC_state;
7338 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
7340 RExC_state = copyRExC_state;
7341 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
7342 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
7344 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
7345 StructCopy(&zero_scan_data, &data, scan_data_t);
7348 StructCopy(&zero_scan_data, &data, scan_data_t);
7351 /* Dig out information for optimizations. */
7352 r->extflags = RExC_flags; /* was pm_op */
7353 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
7356 SvUTF8_on(rx); /* Unicode in it? */
7357 ri->regstclass = NULL;
7358 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
7359 r->intflags |= PREGf_NAUGHTY;
7360 scan = ri->program + 1; /* First BRANCH. */
7362 /* testing for BRANCH here tells us whether there is "must appear"
7363 data in the pattern. If there is then we can use it for optimisations */
7364 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7367 STRLEN longest_float_length, longest_fixed_length;
7368 regnode_ssc ch_class; /* pointed to by data */
7370 SSize_t last_close = 0; /* pointed to by data */
7371 regnode *first= scan;
7372 regnode *first_next= regnext(first);
7374 * Skip introductions and multiplicators >= 1
7375 * so that we can extract the 'meat' of the pattern that must
7376 * match in the large if() sequence following.
7377 * NOTE that EXACT is NOT covered here, as it is normally
7378 * picked up by the optimiser separately.
7380 * This is unfortunate as the optimiser isnt handling lookahead
7381 * properly currently.
7384 while ((OP(first) == OPEN && (sawopen = 1)) ||
7385 /* An OR of *one* alternative - should not happen now. */
7386 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7387 /* for now we can't handle lookbehind IFMATCH*/
7388 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7389 (OP(first) == PLUS) ||
7390 (OP(first) == MINMOD) ||
7391 /* An {n,m} with n>0 */
7392 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7393 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7396 * the only op that could be a regnode is PLUS, all the rest
7397 * will be regnode_1 or regnode_2.
7399 * (yves doesn't think this is true)
7401 if (OP(first) == PLUS)
7404 if (OP(first) == MINMOD)
7406 first += regarglen[OP(first)];
7408 first = NEXTOPER(first);
7409 first_next= regnext(first);
7412 /* Starting-point info. */
7414 DEBUG_PEEP("first:",first,0);
7415 /* Ignore EXACT as we deal with it later. */
7416 if (PL_regkind[OP(first)] == EXACT) {
7417 if (OP(first) == EXACT || OP(first) == EXACTL)
7418 NOOP; /* Empty, get anchored substr later. */
7420 ri->regstclass = first;
7423 else if (PL_regkind[OP(first)] == TRIE &&
7424 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
7426 /* this can happen only on restudy */
7427 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
7430 else if (REGNODE_SIMPLE(OP(first)))
7431 ri->regstclass = first;
7432 else if (PL_regkind[OP(first)] == BOUND ||
7433 PL_regkind[OP(first)] == NBOUND)
7434 ri->regstclass = first;
7435 else if (PL_regkind[OP(first)] == BOL) {
7436 r->intflags |= (OP(first) == MBOL
7439 first = NEXTOPER(first);
7442 else if (OP(first) == GPOS) {
7443 r->intflags |= PREGf_ANCH_GPOS;
7444 first = NEXTOPER(first);
7447 else if ((!sawopen || !RExC_sawback) &&
7449 (OP(first) == STAR &&
7450 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7451 !(r->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
7453 /* turn .* into ^.* with an implied $*=1 */
7455 (OP(NEXTOPER(first)) == REG_ANY)
7458 r->intflags |= (type | PREGf_IMPLICIT);
7459 first = NEXTOPER(first);
7462 if (sawplus && !sawminmod && !sawlookahead
7463 && (!sawopen || !RExC_sawback)
7464 && !pRExC_state->code_blocks) /* May examine pos and $& */
7465 /* x+ must match at the 1st pos of run of x's */
7466 r->intflags |= PREGf_SKIP;
7468 /* Scan is after the zeroth branch, first is atomic matcher. */
7469 #ifdef TRIE_STUDY_OPT
7472 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7473 (IV)(first - scan + 1))
7477 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7478 (IV)(first - scan + 1))
7484 * If there's something expensive in the r.e., find the
7485 * longest literal string that must appear and make it the
7486 * regmust. Resolve ties in favor of later strings, since
7487 * the regstart check works with the beginning of the r.e.
7488 * and avoiding duplication strengthens checking. Not a
7489 * strong reason, but sufficient in the absence of others.
7490 * [Now we resolve ties in favor of the earlier string if
7491 * it happens that c_offset_min has been invalidated, since the
7492 * earlier string may buy us something the later one won't.]
7495 data.longest_fixed = newSVpvs("");
7496 data.longest_float = newSVpvs("");
7497 data.last_found = newSVpvs("");
7498 data.longest = &(data.longest_fixed);
7499 ENTER_with_name("study_chunk");
7500 SAVEFREESV(data.longest_fixed);
7501 SAVEFREESV(data.longest_float);
7502 SAVEFREESV(data.last_found);
7504 if (!ri->regstclass) {
7505 ssc_init(pRExC_state, &ch_class);
7506 data.start_class = &ch_class;
7507 stclass_flag = SCF_DO_STCLASS_AND;
7508 } else /* XXXX Check for BOUND? */
7510 data.last_closep = &last_close;
7513 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
7514 scan + RExC_size, /* Up to end */
7516 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
7517 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
7521 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
7524 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
7525 && data.last_start_min == 0 && data.last_end > 0
7526 && !RExC_seen_zerolen
7527 && !(RExC_seen & REG_VERBARG_SEEN)
7528 && !(RExC_seen & REG_GPOS_SEEN)
7530 r->extflags |= RXf_CHECK_ALL;
7532 scan_commit(pRExC_state, &data,&minlen,0);
7534 longest_float_length = CHR_SVLEN(data.longest_float);
7536 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
7537 && data.offset_fixed == data.offset_float_min
7538 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
7539 && S_setup_longest (aTHX_ pRExC_state,
7543 &(r->float_end_shift),
7544 data.lookbehind_float,
7545 data.offset_float_min,
7547 longest_float_length,
7548 cBOOL(data.flags & SF_FL_BEFORE_EOL),
7549 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
7551 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
7552 r->float_max_offset = data.offset_float_max;
7553 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
7554 r->float_max_offset -= data.lookbehind_float;
7555 SvREFCNT_inc_simple_void_NN(data.longest_float);
7558 r->float_substr = r->float_utf8 = NULL;
7559 longest_float_length = 0;
7562 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
7564 if (S_setup_longest (aTHX_ pRExC_state,
7566 &(r->anchored_utf8),
7567 &(r->anchored_substr),
7568 &(r->anchored_end_shift),
7569 data.lookbehind_fixed,
7572 longest_fixed_length,
7573 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
7574 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
7576 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
7577 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
7580 r->anchored_substr = r->anchored_utf8 = NULL;
7581 longest_fixed_length = 0;
7583 LEAVE_with_name("study_chunk");
7586 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
7587 ri->regstclass = NULL;
7589 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
7591 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7592 && is_ssc_worth_it(pRExC_state, data.start_class))
7594 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7596 ssc_finalize(pRExC_state, data.start_class);
7598 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7599 StructCopy(data.start_class,
7600 (regnode_ssc*)RExC_rxi->data->data[n],
7602 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7603 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7604 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7605 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7606 Perl_re_printf( aTHX_
7607 "synthetic stclass \"%s\".\n",
7608 SvPVX_const(sv));});
7609 data.start_class = NULL;
7612 /* A temporary algorithm prefers floated substr to fixed one to dig
7614 if (longest_fixed_length > longest_float_length) {
7615 r->substrs->check_ix = 0;
7616 r->check_end_shift = r->anchored_end_shift;
7617 r->check_substr = r->anchored_substr;
7618 r->check_utf8 = r->anchored_utf8;
7619 r->check_offset_min = r->check_offset_max = r->anchored_offset;
7620 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
7621 r->intflags |= PREGf_NOSCAN;
7624 r->substrs->check_ix = 1;
7625 r->check_end_shift = r->float_end_shift;
7626 r->check_substr = r->float_substr;
7627 r->check_utf8 = r->float_utf8;
7628 r->check_offset_min = r->float_min_offset;
7629 r->check_offset_max = r->float_max_offset;
7631 if ((r->check_substr || r->check_utf8) ) {
7632 r->extflags |= RXf_USE_INTUIT;
7633 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7634 r->extflags |= RXf_INTUIT_TAIL;
7636 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
7638 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7639 if ( (STRLEN)minlen < longest_float_length )
7640 minlen= longest_float_length;
7641 if ( (STRLEN)minlen < longest_fixed_length )
7642 minlen= longest_fixed_length;
7646 /* Several toplevels. Best we can is to set minlen. */
7648 regnode_ssc ch_class;
7649 SSize_t last_close = 0;
7651 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
7653 scan = ri->program + 1;
7654 ssc_init(pRExC_state, &ch_class);
7655 data.start_class = &ch_class;
7656 data.last_closep = &last_close;
7659 minlen = study_chunk(pRExC_state,
7660 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7661 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7662 ? SCF_TRIE_DOING_RESTUDY
7666 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7668 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7669 = r->float_substr = r->float_utf8 = NULL;
7671 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7672 && is_ssc_worth_it(pRExC_state, data.start_class))
7674 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7676 ssc_finalize(pRExC_state, data.start_class);
7678 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7679 StructCopy(data.start_class,
7680 (regnode_ssc*)RExC_rxi->data->data[n],
7682 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7683 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7684 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7685 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7686 Perl_re_printf( aTHX_
7687 "synthetic stclass \"%s\".\n",
7688 SvPVX_const(sv));});
7689 data.start_class = NULL;
7693 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7694 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7695 r->maxlen = REG_INFTY;
7698 r->maxlen = RExC_maxlen;
7701 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7702 the "real" pattern. */
7704 Perl_re_printf( aTHX_ "minlen: %" IVdf " r->minlen:%" IVdf " maxlen:%" IVdf "\n",
7705 (IV)minlen, (IV)r->minlen, (IV)RExC_maxlen);
7707 r->minlenret = minlen;
7708 if (r->minlen < minlen)
7711 if (RExC_seen & REG_RECURSE_SEEN ) {
7712 r->intflags |= PREGf_RECURSE_SEEN;
7713 Newxz(r->recurse_locinput, r->nparens + 1, char *);
7715 if (RExC_seen & REG_GPOS_SEEN)
7716 r->intflags |= PREGf_GPOS_SEEN;
7717 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7718 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7720 if (pRExC_state->code_blocks)
7721 r->extflags |= RXf_EVAL_SEEN;
7722 if (RExC_seen & REG_VERBARG_SEEN)
7724 r->intflags |= PREGf_VERBARG_SEEN;
7725 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7727 if (RExC_seen & REG_CUTGROUP_SEEN)
7728 r->intflags |= PREGf_CUTGROUP_SEEN;
7729 if (pm_flags & PMf_USE_RE_EVAL)
7730 r->intflags |= PREGf_USE_RE_EVAL;
7731 if (RExC_paren_names)
7732 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7734 RXp_PAREN_NAMES(r) = NULL;
7736 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7737 * so it can be used in pp.c */
7738 if (r->intflags & PREGf_ANCH)
7739 r->extflags |= RXf_IS_ANCHORED;
7743 /* this is used to identify "special" patterns that might result
7744 * in Perl NOT calling the regex engine and instead doing the match "itself",
7745 * particularly special cases in split//. By having the regex compiler
7746 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7747 * we avoid weird issues with equivalent patterns resulting in different behavior,
7748 * AND we allow non Perl engines to get the same optimizations by the setting the
7749 * flags appropriately - Yves */
7750 regnode *first = ri->program + 1;
7752 regnode *next = regnext(first);
7755 if (PL_regkind[fop] == NOTHING && nop == END)
7756 r->extflags |= RXf_NULL;
7757 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
7758 /* when fop is SBOL first->flags will be true only when it was
7759 * produced by parsing /\A/, and not when parsing /^/. This is
7760 * very important for the split code as there we want to
7761 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
7762 * See rt #122761 for more details. -- Yves */
7763 r->extflags |= RXf_START_ONLY;
7764 else if (fop == PLUS
7765 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7767 r->extflags |= RXf_WHITE;
7768 else if ( r->extflags & RXf_SPLIT
7769 && (fop == EXACT || fop == EXACTL)
7770 && STR_LEN(first) == 1
7771 && *(STRING(first)) == ' '
7773 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7777 if (RExC_contains_locale) {
7778 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7782 if (RExC_paren_names) {
7783 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7784 ri->data->data[ri->name_list_idx]
7785 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7788 ri->name_list_idx = 0;
7790 while ( RExC_recurse_count > 0 ) {
7791 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
7792 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - scan );
7795 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7796 /* assume we don't need to swap parens around before we match */
7798 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
7799 (unsigned long)RExC_study_chunk_recursed_count);
7803 Perl_re_printf( aTHX_ "Final program:\n");
7806 #ifdef RE_TRACK_PATTERN_OFFSETS
7807 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7808 const STRLEN len = ri->u.offsets[0];
7810 GET_RE_DEBUG_FLAGS_DECL;
7811 Perl_re_printf( aTHX_
7812 "Offsets: [%" UVuf "]\n\t", (UV)ri->u.offsets[0]);
7813 for (i = 1; i <= len; i++) {
7814 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7815 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7816 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7818 Perl_re_printf( aTHX_ "\n");
7823 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7824 * by setting the regexp SV to readonly-only instead. If the
7825 * pattern's been recompiled, the USEDness should remain. */
7826 if (old_re && SvREADONLY(old_re))
7834 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7837 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7839 PERL_UNUSED_ARG(value);
7841 if (flags & RXapif_FETCH) {
7842 return reg_named_buff_fetch(rx, key, flags);
7843 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7844 Perl_croak_no_modify();
7846 } else if (flags & RXapif_EXISTS) {
7847 return reg_named_buff_exists(rx, key, flags)
7850 } else if (flags & RXapif_REGNAMES) {
7851 return reg_named_buff_all(rx, flags);
7852 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7853 return reg_named_buff_scalar(rx, flags);
7855 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7861 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7864 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7865 PERL_UNUSED_ARG(lastkey);
7867 if (flags & RXapif_FIRSTKEY)
7868 return reg_named_buff_firstkey(rx, flags);
7869 else if (flags & RXapif_NEXTKEY)
7870 return reg_named_buff_nextkey(rx, flags);
7872 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7879 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7882 AV *retarray = NULL;
7884 struct regexp *const rx = ReANY(r);
7886 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7888 if (flags & RXapif_ALL)
7891 if (rx && RXp_PAREN_NAMES(rx)) {
7892 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7895 SV* sv_dat=HeVAL(he_str);
7896 I32 *nums=(I32*)SvPVX(sv_dat);
7897 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7898 if ((I32)(rx->nparens) >= nums[i]
7899 && rx->offs[nums[i]].start != -1
7900 && rx->offs[nums[i]].end != -1)
7903 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7908 ret = newSVsv(&PL_sv_undef);
7911 av_push(retarray, ret);
7914 return newRV_noinc(MUTABLE_SV(retarray));
7921 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7924 struct regexp *const rx = ReANY(r);
7926 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7928 if (rx && RXp_PAREN_NAMES(rx)) {
7929 if (flags & RXapif_ALL) {
7930 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7932 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7934 SvREFCNT_dec_NN(sv);
7946 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7948 struct regexp *const rx = ReANY(r);
7950 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7952 if ( rx && RXp_PAREN_NAMES(rx) ) {
7953 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7955 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7962 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7964 struct regexp *const rx = ReANY(r);
7965 GET_RE_DEBUG_FLAGS_DECL;
7967 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7969 if (rx && RXp_PAREN_NAMES(rx)) {
7970 HV *hv = RXp_PAREN_NAMES(rx);
7972 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7975 SV* sv_dat = HeVAL(temphe);
7976 I32 *nums = (I32*)SvPVX(sv_dat);
7977 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7978 if ((I32)(rx->lastparen) >= nums[i] &&
7979 rx->offs[nums[i]].start != -1 &&
7980 rx->offs[nums[i]].end != -1)
7986 if (parno || flags & RXapif_ALL) {
7987 return newSVhek(HeKEY_hek(temphe));
7995 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8000 struct regexp *const rx = ReANY(r);
8002 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8004 if (rx && RXp_PAREN_NAMES(rx)) {
8005 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8006 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8007 } else if (flags & RXapif_ONE) {
8008 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8009 av = MUTABLE_AV(SvRV(ret));
8010 length = av_tindex(av);
8011 SvREFCNT_dec_NN(ret);
8012 return newSViv(length + 1);
8014 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8019 return &PL_sv_undef;
8023 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8025 struct regexp *const rx = ReANY(r);
8028 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8030 if (rx && RXp_PAREN_NAMES(rx)) {
8031 HV *hv= RXp_PAREN_NAMES(rx);
8033 (void)hv_iterinit(hv);
8034 while ( (temphe = hv_iternext_flags(hv,0)) ) {
8037 SV* sv_dat = HeVAL(temphe);
8038 I32 *nums = (I32*)SvPVX(sv_dat);
8039 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8040 if ((I32)(rx->lastparen) >= nums[i] &&
8041 rx->offs[nums[i]].start != -1 &&
8042 rx->offs[nums[i]].end != -1)
8048 if (parno || flags & RXapif_ALL) {
8049 av_push(av, newSVhek(HeKEY_hek(temphe)));
8054 return newRV_noinc(MUTABLE_SV(av));
8058 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8061 struct regexp *const rx = ReANY(r);
8067 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8069 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8070 || n == RX_BUFF_IDX_CARET_FULLMATCH
8071 || n == RX_BUFF_IDX_CARET_POSTMATCH
8074 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8076 /* on something like
8079 * the KEEPCOPY is set on the PMOP rather than the regex */
8080 if (PL_curpm && r == PM_GETRE(PL_curpm))
8081 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8090 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8091 /* no need to distinguish between them any more */
8092 n = RX_BUFF_IDX_FULLMATCH;
8094 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8095 && rx->offs[0].start != -1)
8097 /* $`, ${^PREMATCH} */
8098 i = rx->offs[0].start;
8102 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8103 && rx->offs[0].end != -1)
8105 /* $', ${^POSTMATCH} */
8106 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8107 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8110 if ( 0 <= n && n <= (I32)rx->nparens &&
8111 (s1 = rx->offs[n].start) != -1 &&
8112 (t1 = rx->offs[n].end) != -1)
8114 /* $&, ${^MATCH}, $1 ... */
8116 s = rx->subbeg + s1 - rx->suboffset;
8121 assert(s >= rx->subbeg);
8122 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8124 #ifdef NO_TAINT_SUPPORT
8125 sv_setpvn(sv, s, i);
8127 const int oldtainted = TAINT_get;
8129 sv_setpvn(sv, s, i);
8130 TAINT_set(oldtainted);
8132 if (RXp_MATCH_UTF8(rx))
8137 if (RXp_MATCH_TAINTED(rx)) {
8138 if (SvTYPE(sv) >= SVt_PVMG) {
8139 MAGIC* const mg = SvMAGIC(sv);
8142 SvMAGIC_set(sv, mg->mg_moremagic);
8144 if ((mgt = SvMAGIC(sv))) {
8145 mg->mg_moremagic = mgt;
8146 SvMAGIC_set(sv, mg);
8163 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8164 SV const * const value)
8166 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8168 PERL_UNUSED_ARG(rx);
8169 PERL_UNUSED_ARG(paren);
8170 PERL_UNUSED_ARG(value);
8173 Perl_croak_no_modify();
8177 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8180 struct regexp *const rx = ReANY(r);
8184 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8186 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8187 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8188 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8191 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8193 /* on something like
8196 * the KEEPCOPY is set on the PMOP rather than the regex */
8197 if (PL_curpm && r == PM_GETRE(PL_curpm))
8198 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8204 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8206 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8207 case RX_BUFF_IDX_PREMATCH: /* $` */
8208 if (rx->offs[0].start != -1) {
8209 i = rx->offs[0].start;
8218 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8219 case RX_BUFF_IDX_POSTMATCH: /* $' */
8220 if (rx->offs[0].end != -1) {
8221 i = rx->sublen - rx->offs[0].end;
8223 s1 = rx->offs[0].end;
8230 default: /* $& / ${^MATCH}, $1, $2, ... */
8231 if (paren <= (I32)rx->nparens &&
8232 (s1 = rx->offs[paren].start) != -1 &&
8233 (t1 = rx->offs[paren].end) != -1)
8239 if (ckWARN(WARN_UNINITIALIZED))
8240 report_uninit((const SV *)sv);
8245 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8246 const char * const s = rx->subbeg - rx->suboffset + s1;
8251 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8258 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8260 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8261 PERL_UNUSED_ARG(rx);
8265 return newSVpvs("Regexp");
8268 /* Scans the name of a named buffer from the pattern.
8269 * If flags is REG_RSN_RETURN_NULL returns null.
8270 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8271 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8272 * to the parsed name as looked up in the RExC_paren_names hash.
8273 * If there is an error throws a vFAIL().. type exception.
8276 #define REG_RSN_RETURN_NULL 0
8277 #define REG_RSN_RETURN_NAME 1
8278 #define REG_RSN_RETURN_DATA 2
8281 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8283 char *name_start = RExC_parse;
8285 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8287 assert (RExC_parse <= RExC_end);
8288 if (RExC_parse == RExC_end) NOOP;
8289 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8290 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8291 * using do...while */
8294 RExC_parse += UTF8SKIP(RExC_parse);
8295 } while ( RExC_parse < RExC_end
8296 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8300 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8302 RExC_parse++; /* so the <- from the vFAIL is after the offending
8304 vFAIL("Group name must start with a non-digit word character");
8308 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8309 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8310 if ( flags == REG_RSN_RETURN_NAME)
8312 else if (flags==REG_RSN_RETURN_DATA) {
8315 if ( ! sv_name ) /* should not happen*/
8316 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8317 if (RExC_paren_names)
8318 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8320 sv_dat = HeVAL(he_str);
8322 vFAIL("Reference to nonexistent named group");
8326 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8327 (unsigned long) flags);
8329 NOT_REACHED; /* NOTREACHED */
8334 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8336 if (RExC_lastparse!=RExC_parse) { \
8337 Perl_re_printf( aTHX_ "%s", \
8338 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8339 RExC_end - RExC_parse, 16, \
8341 PERL_PV_ESCAPE_UNI_DETECT | \
8342 PERL_PV_PRETTY_ELLIPSES | \
8343 PERL_PV_PRETTY_LTGT | \
8344 PERL_PV_ESCAPE_RE | \
8345 PERL_PV_PRETTY_EXACTSIZE \
8349 Perl_re_printf( aTHX_ "%16s",""); \
8352 num = RExC_size + 1; \
8354 num=REG_NODE_NUM(RExC_emit); \
8355 if (RExC_lastnum!=num) \
8356 Perl_re_printf( aTHX_ "|%4d",num); \
8358 Perl_re_printf( aTHX_ "|%4s",""); \
8359 Perl_re_printf( aTHX_ "|%*s%-4s", \
8360 (int)((depth*2)), "", \
8364 RExC_lastparse=RExC_parse; \
8369 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8370 DEBUG_PARSE_MSG((funcname)); \
8371 Perl_re_printf( aTHX_ "%4s","\n"); \
8373 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8374 DEBUG_PARSE_MSG((funcname)); \
8375 Perl_re_printf( aTHX_ fmt "\n",args); \
8378 /* This section of code defines the inversion list object and its methods. The
8379 * interfaces are highly subject to change, so as much as possible is static to
8380 * this file. An inversion list is here implemented as a malloc'd C UV array
8381 * as an SVt_INVLIST scalar.
8383 * An inversion list for Unicode is an array of code points, sorted by ordinal
8384 * number. Each element gives the code point that begins a range that extends
8385 * up-to but not including the code point given by the next element. The final
8386 * element gives the first code point of a range that extends to the platform's
8387 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8388 * ...) give ranges whose code points are all in the inversion list. We say
8389 * that those ranges are in the set. The odd-numbered elements give ranges
8390 * whose code points are not in the inversion list, and hence not in the set.
8391 * Thus, element [0] is the first code point in the list. Element [1]
8392 * is the first code point beyond that not in the list; and element [2] is the
8393 * first code point beyond that that is in the list. In other words, the first
8394 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8395 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8396 * all code points in that range are not in the inversion list. The third
8397 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8398 * list, and so forth. Thus every element whose index is divisible by two
8399 * gives the beginning of a range that is in the list, and every element whose
8400 * index is not divisible by two gives the beginning of a range not in the
8401 * list. If the final element's index is divisible by two, the inversion list
8402 * extends to the platform's infinity; otherwise the highest code point in the
8403 * inversion list is the contents of that element minus 1.
8405 * A range that contains just a single code point N will look like
8407 * invlist[i+1] == N+1
8409 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8410 * impossible to represent, so element [i+1] is omitted. The single element
8412 * invlist[0] == UV_MAX
8413 * contains just UV_MAX, but is interpreted as matching to infinity.
8415 * Taking the complement (inverting) an inversion list is quite simple, if the
8416 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8417 * This implementation reserves an element at the beginning of each inversion
8418 * list to always contain 0; there is an additional flag in the header which
8419 * indicates if the list begins at the 0, or is offset to begin at the next
8420 * element. This means that the inversion list can be inverted without any
8421 * copying; just flip the flag.
8423 * More about inversion lists can be found in "Unicode Demystified"
8424 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8426 * The inversion list data structure is currently implemented as an SV pointing
8427 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8428 * array of UV whose memory management is automatically handled by the existing
8429 * facilities for SV's.
8431 * Some of the methods should always be private to the implementation, and some
8432 * should eventually be made public */
8434 /* The header definitions are in F<invlist_inline.h> */
8436 #ifndef PERL_IN_XSUB_RE
8438 PERL_STATIC_INLINE UV*
8439 S__invlist_array_init(SV* const invlist, const bool will_have_0)
8441 /* Returns a pointer to the first element in the inversion list's array.
8442 * This is called upon initialization of an inversion list. Where the
8443 * array begins depends on whether the list has the code point U+0000 in it
8444 * or not. The other parameter tells it whether the code that follows this
8445 * call is about to put a 0 in the inversion list or not. The first
8446 * element is either the element reserved for 0, if TRUE, or the element
8447 * after it, if FALSE */
8449 bool* offset = get_invlist_offset_addr(invlist);
8450 UV* zero_addr = (UV *) SvPVX(invlist);
8452 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8455 assert(! _invlist_len(invlist));
8459 /* 1^1 = 0; 1^0 = 1 */
8460 *offset = 1 ^ will_have_0;
8461 return zero_addr + *offset;
8466 PERL_STATIC_INLINE void
8467 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8469 /* Sets the current number of elements stored in the inversion list.
8470 * Updates SvCUR correspondingly */
8471 PERL_UNUSED_CONTEXT;
8472 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8474 assert(SvTYPE(invlist) == SVt_INVLIST);
8479 : TO_INTERNAL_SIZE(len + offset));
8480 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8483 #ifndef PERL_IN_XSUB_RE
8486 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
8488 /* Replaces the inversion list in 'dest' with the one from 'src'. It
8489 * steals the list from 'src', so 'src' is made to have a NULL list. This
8490 * is similar to what SvSetMagicSV() would do, if it were implemented on
8491 * inversion lists, though this routine avoids a copy */
8493 const UV src_len = _invlist_len(src);
8494 const bool src_offset = *get_invlist_offset_addr(src);
8495 const STRLEN src_byte_len = SvLEN(src);
8496 char * array = SvPVX(src);
8498 const int oldtainted = TAINT_get;
8500 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
8502 assert(SvTYPE(src) == SVt_INVLIST);
8503 assert(SvTYPE(dest) == SVt_INVLIST);
8504 assert(! invlist_is_iterating(src));
8505 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
8507 /* Make sure it ends in the right place with a NUL, as our inversion list
8508 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
8510 array[src_byte_len - 1] = '\0';
8512 TAINT_NOT; /* Otherwise it breaks */
8513 sv_usepvn_flags(dest,
8517 /* This flag is documented to cause a copy to be avoided */
8518 SV_HAS_TRAILING_NUL);
8519 TAINT_set(oldtainted);
8524 /* Finish up copying over the other fields in an inversion list */
8525 *get_invlist_offset_addr(dest) = src_offset;
8526 invlist_set_len(dest, src_len, src_offset);
8527 *get_invlist_previous_index_addr(dest) = 0;
8528 invlist_iterfinish(dest);
8531 PERL_STATIC_INLINE IV*
8532 S_get_invlist_previous_index_addr(SV* invlist)
8534 /* Return the address of the IV that is reserved to hold the cached index
8536 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8538 assert(SvTYPE(invlist) == SVt_INVLIST);
8540 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8543 PERL_STATIC_INLINE IV
8544 S_invlist_previous_index(SV* const invlist)
8546 /* Returns cached index of previous search */
8548 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8550 return *get_invlist_previous_index_addr(invlist);
8553 PERL_STATIC_INLINE void
8554 S_invlist_set_previous_index(SV* const invlist, const IV index)
8556 /* Caches <index> for later retrieval */
8558 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8560 assert(index == 0 || index < (int) _invlist_len(invlist));
8562 *get_invlist_previous_index_addr(invlist) = index;
8565 PERL_STATIC_INLINE void
8566 S_invlist_trim(SV* invlist)
8568 /* Free the not currently-being-used space in an inversion list */
8570 /* But don't free up the space needed for the 0 UV that is always at the
8571 * beginning of the list, nor the trailing NUL */
8572 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
8574 PERL_ARGS_ASSERT_INVLIST_TRIM;
8576 assert(SvTYPE(invlist) == SVt_INVLIST);
8578 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
8581 PERL_STATIC_INLINE void
8582 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
8584 PERL_ARGS_ASSERT_INVLIST_CLEAR;
8586 assert(SvTYPE(invlist) == SVt_INVLIST);
8588 invlist_set_len(invlist, 0, 0);
8589 invlist_trim(invlist);
8592 #endif /* ifndef PERL_IN_XSUB_RE */
8594 PERL_STATIC_INLINE bool
8595 S_invlist_is_iterating(SV* const invlist)
8597 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8599 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8602 #ifndef PERL_IN_XSUB_RE
8604 PERL_STATIC_INLINE UV
8605 S_invlist_max(SV* const invlist)
8607 /* Returns the maximum number of elements storable in the inversion list's
8608 * array, without having to realloc() */
8610 PERL_ARGS_ASSERT_INVLIST_MAX;
8612 assert(SvTYPE(invlist) == SVt_INVLIST);
8614 /* Assumes worst case, in which the 0 element is not counted in the
8615 * inversion list, so subtracts 1 for that */
8616 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8617 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8618 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8621 Perl__new_invlist(pTHX_ IV initial_size)
8624 /* Return a pointer to a newly constructed inversion list, with enough
8625 * space to store 'initial_size' elements. If that number is negative, a
8626 * system default is used instead */
8630 if (initial_size < 0) {
8634 /* Allocate the initial space */
8635 new_list = newSV_type(SVt_INVLIST);
8637 /* First 1 is in case the zero element isn't in the list; second 1 is for
8639 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8640 invlist_set_len(new_list, 0, 0);
8642 /* Force iterinit() to be used to get iteration to work */
8643 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8645 *get_invlist_previous_index_addr(new_list) = 0;
8651 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8653 /* Return a pointer to a newly constructed inversion list, initialized to
8654 * point to <list>, which has to be in the exact correct inversion list
8655 * form, including internal fields. Thus this is a dangerous routine that
8656 * should not be used in the wrong hands. The passed in 'list' contains
8657 * several header fields at the beginning that are not part of the
8658 * inversion list body proper */
8660 const STRLEN length = (STRLEN) list[0];
8661 const UV version_id = list[1];
8662 const bool offset = cBOOL(list[2]);
8663 #define HEADER_LENGTH 3
8664 /* If any of the above changes in any way, you must change HEADER_LENGTH
8665 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8666 * perl -E 'say int(rand 2**31-1)'
8668 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8669 data structure type, so that one being
8670 passed in can be validated to be an
8671 inversion list of the correct vintage.
8674 SV* invlist = newSV_type(SVt_INVLIST);
8676 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8678 if (version_id != INVLIST_VERSION_ID) {
8679 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8682 /* The generated array passed in includes header elements that aren't part
8683 * of the list proper, so start it just after them */
8684 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8686 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8687 shouldn't touch it */
8689 *(get_invlist_offset_addr(invlist)) = offset;
8691 /* The 'length' passed to us is the physical number of elements in the
8692 * inversion list. But if there is an offset the logical number is one
8694 invlist_set_len(invlist, length - offset, offset);
8696 invlist_set_previous_index(invlist, 0);
8698 /* Initialize the iteration pointer. */
8699 invlist_iterfinish(invlist);
8701 SvREADONLY_on(invlist);
8707 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8709 /* Grow the maximum size of an inversion list */
8711 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8713 assert(SvTYPE(invlist) == SVt_INVLIST);
8715 /* Add one to account for the zero element at the beginning which may not
8716 * be counted by the calling parameters */
8717 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8721 S__append_range_to_invlist(pTHX_ SV* const invlist,
8722 const UV start, const UV end)
8724 /* Subject to change or removal. Append the range from 'start' to 'end' at
8725 * the end of the inversion list. The range must be above any existing
8729 UV max = invlist_max(invlist);
8730 UV len = _invlist_len(invlist);
8733 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8735 if (len == 0) { /* Empty lists must be initialized */
8736 offset = start != 0;
8737 array = _invlist_array_init(invlist, ! offset);
8740 /* Here, the existing list is non-empty. The current max entry in the
8741 * list is generally the first value not in the set, except when the
8742 * set extends to the end of permissible values, in which case it is
8743 * the first entry in that final set, and so this call is an attempt to
8744 * append out-of-order */
8746 UV final_element = len - 1;
8747 array = invlist_array(invlist);
8748 if ( array[final_element] > start
8749 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8751 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",
8752 array[final_element], start,
8753 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8756 /* Here, it is a legal append. If the new range begins 1 above the end
8757 * of the range below it, it is extending the range below it, so the
8758 * new first value not in the set is one greater than the newly
8759 * extended range. */
8760 offset = *get_invlist_offset_addr(invlist);
8761 if (array[final_element] == start) {
8762 if (end != UV_MAX) {
8763 array[final_element] = end + 1;
8766 /* But if the end is the maximum representable on the machine,
8767 * assume that infinity was actually what was meant. Just let
8768 * the range that this would extend to have no end */
8769 invlist_set_len(invlist, len - 1, offset);
8775 /* Here the new range doesn't extend any existing set. Add it */
8777 len += 2; /* Includes an element each for the start and end of range */
8779 /* If wll overflow the existing space, extend, which may cause the array to
8782 invlist_extend(invlist, len);
8784 /* Have to set len here to avoid assert failure in invlist_array() */
8785 invlist_set_len(invlist, len, offset);
8787 array = invlist_array(invlist);
8790 invlist_set_len(invlist, len, offset);
8793 /* The next item on the list starts the range, the one after that is
8794 * one past the new range. */
8795 array[len - 2] = start;
8796 if (end != UV_MAX) {
8797 array[len - 1] = end + 1;
8800 /* But if the end is the maximum representable on the machine, just let
8801 * the range have no end */
8802 invlist_set_len(invlist, len - 1, offset);
8807 Perl__invlist_search(SV* const invlist, const UV cp)
8809 /* Searches the inversion list for the entry that contains the input code
8810 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8811 * return value is the index into the list's array of the range that
8812 * contains <cp>, that is, 'i' such that
8813 * array[i] <= cp < array[i+1]
8818 IV high = _invlist_len(invlist);
8819 const IV highest_element = high - 1;
8822 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8824 /* If list is empty, return failure. */
8829 /* (We can't get the array unless we know the list is non-empty) */
8830 array = invlist_array(invlist);
8832 mid = invlist_previous_index(invlist);
8834 if (mid > highest_element) {
8835 mid = highest_element;
8838 /* <mid> contains the cache of the result of the previous call to this
8839 * function (0 the first time). See if this call is for the same result,
8840 * or if it is for mid-1. This is under the theory that calls to this
8841 * function will often be for related code points that are near each other.
8842 * And benchmarks show that caching gives better results. We also test
8843 * here if the code point is within the bounds of the list. These tests
8844 * replace others that would have had to be made anyway to make sure that
8845 * the array bounds were not exceeded, and these give us extra information
8846 * at the same time */
8847 if (cp >= array[mid]) {
8848 if (cp >= array[highest_element]) {
8849 return highest_element;
8852 /* Here, array[mid] <= cp < array[highest_element]. This means that
8853 * the final element is not the answer, so can exclude it; it also
8854 * means that <mid> is not the final element, so can refer to 'mid + 1'
8856 if (cp < array[mid + 1]) {
8862 else { /* cp < aray[mid] */
8863 if (cp < array[0]) { /* Fail if outside the array */
8867 if (cp >= array[mid - 1]) {
8872 /* Binary search. What we are looking for is <i> such that
8873 * array[i] <= cp < array[i+1]
8874 * The loop below converges on the i+1. Note that there may not be an
8875 * (i+1)th element in the array, and things work nonetheless */
8876 while (low < high) {
8877 mid = (low + high) / 2;
8878 assert(mid <= highest_element);
8879 if (array[mid] <= cp) { /* cp >= array[mid] */
8882 /* We could do this extra test to exit the loop early.
8883 if (cp < array[low]) {
8888 else { /* cp < array[mid] */
8895 invlist_set_previous_index(invlist, high);
8900 Perl__invlist_populate_swatch(SV* const invlist,
8901 const UV start, const UV end, U8* swatch)
8903 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8904 * but is used when the swash has an inversion list. This makes this much
8905 * faster, as it uses a binary search instead of a linear one. This is
8906 * intimately tied to that function, and perhaps should be in utf8.c,
8907 * except it is intimately tied to inversion lists as well. It assumes
8908 * that <swatch> is all 0's on input */
8911 const IV len = _invlist_len(invlist);
8915 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8917 if (len == 0) { /* Empty inversion list */
8921 array = invlist_array(invlist);
8923 /* Find which element it is */
8924 i = _invlist_search(invlist, start);
8926 /* We populate from <start> to <end> */
8927 while (current < end) {
8930 /* The inversion list gives the results for every possible code point
8931 * after the first one in the list. Only those ranges whose index is
8932 * even are ones that the inversion list matches. For the odd ones,
8933 * and if the initial code point is not in the list, we have to skip
8934 * forward to the next element */
8935 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8937 if (i >= len) { /* Finished if beyond the end of the array */
8941 if (current >= end) { /* Finished if beyond the end of what we
8943 if (LIKELY(end < UV_MAX)) {
8947 /* We get here when the upper bound is the maximum
8948 * representable on the machine, and we are looking for just
8949 * that code point. Have to special case it */
8951 goto join_end_of_list;
8954 assert(current >= start);
8956 /* The current range ends one below the next one, except don't go past
8959 upper = (i < len && array[i] < end) ? array[i] : end;
8961 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8962 * for each code point in it */
8963 for (; current < upper; current++) {
8964 const STRLEN offset = (STRLEN)(current - start);
8965 swatch[offset >> 3] |= 1 << (offset & 7);
8970 /* Quit if at the end of the list */
8973 /* But first, have to deal with the highest possible code point on
8974 * the platform. The previous code assumes that <end> is one
8975 * beyond where we want to populate, but that is impossible at the
8976 * platform's infinity, so have to handle it specially */
8977 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8979 const STRLEN offset = (STRLEN)(end - start);
8980 swatch[offset >> 3] |= 1 << (offset & 7);
8985 /* Advance to the next range, which will be for code points not in the
8994 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8995 const bool complement_b, SV** output)
8997 /* Take the union of two inversion lists and point '*output' to it. On
8998 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
8999 * even 'a' or 'b'). If to an inversion list, the contents of the original
9000 * list will be replaced by the union. The first list, 'a', may be
9001 * NULL, in which case a copy of the second list is placed in '*output'.
9002 * If 'complement_b' is TRUE, the union is taken of the complement
9003 * (inversion) of 'b' instead of b itself.
9005 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9006 * Richard Gillam, published by Addison-Wesley, and explained at some
9007 * length there. The preface says to incorporate its examples into your
9008 * code at your own risk.
9010 * The algorithm is like a merge sort. */
9012 const UV* array_a; /* a's array */
9014 UV len_a; /* length of a's array */
9017 SV* u; /* the resulting union */
9021 UV i_a = 0; /* current index into a's array */
9025 /* running count, as explained in the algorithm source book; items are
9026 * stopped accumulating and are output when the count changes to/from 0.
9027 * The count is incremented when we start a range that's in an input's set,
9028 * and decremented when we start a range that's not in a set. So this
9029 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9030 * and hence nothing goes into the union; 1, just one of the inputs is in
9031 * its set (and its current range gets added to the union); and 2 when both
9032 * inputs are in their sets. */
9035 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9037 assert(*output == NULL || SvTYPE(*output) == SVt_INVLIST);
9039 len_b = _invlist_len(b);
9042 /* Here, 'b' is empty, hence it's complement is all possible code
9043 * points. So if the union includes the complement of 'b', it includes
9044 * everything, and we need not even look at 'a'. It's easiest to
9045 * create a new inversion list that matches everything. */
9047 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9049 if (*output == NULL) { /* If the output didn't exist, just point it
9051 *output = everything;
9053 else { /* Otherwise, replace its contents with the new list */
9054 invlist_replace_list_destroys_src(*output, everything);
9055 SvREFCNT_dec_NN(everything);
9061 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9062 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9063 * output will be empty */
9065 if (a == NULL || _invlist_len(a) == 0) {
9066 if (*output == NULL) {
9067 *output = _new_invlist(0);
9070 invlist_clear(*output);
9075 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9076 * union. We can just return a copy of 'a' if '*output' doesn't point
9077 * to an existing list */
9078 if (*output == NULL) {
9079 *output = invlist_clone(a);
9083 /* If the output is to overwrite 'a', we have a no-op, as it's
9089 /* Here, '*output' is to be overwritten by 'a' */
9090 u = invlist_clone(a);
9091 invlist_replace_list_destroys_src(*output, u);
9097 /* Here 'b' is not empty. See about 'a' */
9099 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9101 /* Here, 'a' is empty (and b is not). That means the union will come
9102 * entirely from 'b'. If '*output' is NULL, we can directly return a
9103 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9106 SV ** dest = (*output == NULL) ? output : &u;
9107 *dest = invlist_clone(b);
9109 _invlist_invert(*dest);
9113 invlist_replace_list_destroys_src(*output, u);
9120 /* Here both lists exist and are non-empty */
9121 array_a = invlist_array(a);
9122 array_b = invlist_array(b);
9124 /* If are to take the union of 'a' with the complement of b, set it
9125 * up so are looking at b's complement. */
9128 /* To complement, we invert: if the first element is 0, remove it. To
9129 * do this, we just pretend the array starts one later */
9130 if (array_b[0] == 0) {
9136 /* But if the first element is not zero, we pretend the list starts
9137 * at the 0 that is always stored immediately before the array. */
9143 /* Size the union for the worst case: that the sets are completely
9145 u = _new_invlist(len_a + len_b);
9147 /* Will contain U+0000 if either component does */
9148 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9149 || (len_b > 0 && array_b[0] == 0));
9151 /* Go through each input list item by item, stopping when have exhausted
9153 while (i_a < len_a && i_b < len_b) {
9154 UV cp; /* The element to potentially add to the union's array */
9155 bool cp_in_set; /* is it in the the input list's set or not */
9157 /* We need to take one or the other of the two inputs for the union.
9158 * Since we are merging two sorted lists, we take the smaller of the
9159 * next items. In case of a tie, we take first the one that is in its
9160 * set. If we first took the one not in its set, it would decrement
9161 * the count, possibly to 0 which would cause it to be output as ending
9162 * the range, and the next time through we would take the same number,
9163 * and output it again as beginning the next range. By doing it the
9164 * opposite way, there is no possibility that the count will be
9165 * momentarily decremented to 0, and thus the two adjoining ranges will
9166 * be seamlessly merged. (In a tie and both are in the set or both not
9167 * in the set, it doesn't matter which we take first.) */
9168 if ( array_a[i_a] < array_b[i_b]
9169 || ( array_a[i_a] == array_b[i_b]
9170 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9172 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9173 cp = array_a[i_a++];
9176 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9177 cp = array_b[i_b++];
9180 /* Here, have chosen which of the two inputs to look at. Only output
9181 * if the running count changes to/from 0, which marks the
9182 * beginning/end of a range that's in the set */
9185 array_u[i_u++] = cp;
9192 array_u[i_u++] = cp;
9198 /* The loop above increments the index into exactly one of the input lists
9199 * each iteration, and ends when either index gets to its list end. That
9200 * means the other index is lower than its end, and so something is
9201 * remaining in that one. We decrement 'count', as explained below, if
9202 * that list is in its set. (i_a and i_b each currently index the element
9203 * beyond the one we care about.) */
9204 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9205 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9210 /* Above we decremented 'count' if the list that had unexamined elements in
9211 * it was in its set. This has made it so that 'count' being non-zero
9212 * means there isn't anything left to output; and 'count' equal to 0 means
9213 * that what is left to output is precisely that which is left in the
9214 * non-exhausted input list.
9216 * To see why, note first that the exhausted input obviously has nothing
9217 * left to add to the union. If it was in its set at its end, that means
9218 * the set extends from here to the platform's infinity, and hence so does
9219 * the union and the non-exhausted set is irrelevant. The exhausted set
9220 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9221 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9222 * 'count' remains at 1. This is consistent with the decremented 'count'
9223 * != 0 meaning there's nothing left to add to the union.
9225 * But if the exhausted input wasn't in its set, it contributed 0 to
9226 * 'count', and the rest of the union will be whatever the other input is.
9227 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9228 * otherwise it gets decremented to 0. This is consistent with 'count'
9229 * == 0 meaning the remainder of the union is whatever is left in the
9230 * non-exhausted list. */
9235 IV copy_count = len_a - i_a;
9236 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9237 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9239 else { /* The non-exhausted input is b */
9240 copy_count = len_b - i_b;
9241 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9243 len_u = i_u + copy_count;
9246 /* Set the result to the final length, which can change the pointer to
9247 * array_u, so re-find it. (Note that it is unlikely that this will
9248 * change, as we are shrinking the space, not enlarging it) */
9249 if (len_u != _invlist_len(u)) {
9250 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9252 array_u = invlist_array(u);
9255 if (*output == NULL) { /* Simply return the new inversion list */
9259 /* Otherwise, overwrite the inversion list that was in '*output'. We
9260 * could instead free '*output', and then set it to 'u', but experience
9261 * has shown [perl #127392] that if the input is a mortal, we can get a
9262 * huge build-up of these during regex compilation before they get
9264 invlist_replace_list_destroys_src(*output, u);
9272 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9273 const bool complement_b, SV** i)
9275 /* Take the intersection of two inversion lists and point '*i' to it. On
9276 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9277 * even 'a' or 'b'). If to an inversion list, the contents of the original
9278 * list will be replaced by the intersection. The first list, 'a', may be
9279 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9280 * TRUE, the result will be the intersection of 'a' and the complement (or
9281 * inversion) of 'b' instead of 'b' directly.
9283 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9284 * Richard Gillam, published by Addison-Wesley, and explained at some
9285 * length there. The preface says to incorporate its examples into your
9286 * code at your own risk. In fact, it had bugs
9288 * The algorithm is like a merge sort, and is essentially the same as the
9292 const UV* array_a; /* a's array */
9294 UV len_a; /* length of a's array */
9297 SV* r; /* the resulting intersection */
9301 UV i_a = 0; /* current index into a's array */
9305 /* running count of how many of the two inputs are postitioned at ranges
9306 * that are in their sets. As explained in the algorithm source book,
9307 * items are stopped accumulating and are output when the count changes
9308 * to/from 2. The count is incremented when we start a range that's in an
9309 * input's set, and decremented when we start a range that's not in a set.
9310 * Only when it is 2 are we in the intersection. */
9313 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9315 assert(*i == NULL || SvTYPE(*i) == SVt_INVLIST);
9317 /* Special case if either one is empty */
9318 len_a = (a == NULL) ? 0 : _invlist_len(a);
9319 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9320 if (len_a != 0 && complement_b) {
9322 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9323 * must be empty. Here, also we are using 'b's complement, which
9324 * hence must be every possible code point. Thus the intersection
9327 if (*i == a) { /* No-op */
9332 *i = invlist_clone(a);
9336 r = invlist_clone(a);
9337 invlist_replace_list_destroys_src(*i, r);
9342 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9343 * intersection must be empty */
9345 *i = _new_invlist(0);
9353 /* Here both lists exist and are non-empty */
9354 array_a = invlist_array(a);
9355 array_b = invlist_array(b);
9357 /* If are to take the intersection of 'a' with the complement of b, set it
9358 * up so are looking at b's complement. */
9361 /* To complement, we invert: if the first element is 0, remove it. To
9362 * do this, we just pretend the array starts one later */
9363 if (array_b[0] == 0) {
9369 /* But if the first element is not zero, we pretend the list starts
9370 * at the 0 that is always stored immediately before the array. */
9376 /* Size the intersection for the worst case: that the intersection ends up
9377 * fragmenting everything to be completely disjoint */
9378 r= _new_invlist(len_a + len_b);
9380 /* Will contain U+0000 iff both components do */
9381 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9382 && len_b > 0 && array_b[0] == 0);
9384 /* Go through each list item by item, stopping when have exhausted one of
9386 while (i_a < len_a && i_b < len_b) {
9387 UV cp; /* The element to potentially add to the intersection's
9389 bool cp_in_set; /* Is it in the input list's set or not */
9391 /* We need to take one or the other of the two inputs for the
9392 * intersection. Since we are merging two sorted lists, we take the
9393 * smaller of the next items. In case of a tie, we take first the one
9394 * that is not in its set (a difference from the union algorithm). If
9395 * we first took the one in its set, it would increment the count,
9396 * possibly to 2 which would cause it to be output as starting a range
9397 * in the intersection, and the next time through we would take that
9398 * same number, and output it again as ending the set. By doing the
9399 * opposite of this, there is no possibility that the count will be
9400 * momentarily incremented to 2. (In a tie and both are in the set or
9401 * both not in the set, it doesn't matter which we take first.) */
9402 if ( array_a[i_a] < array_b[i_b]
9403 || ( array_a[i_a] == array_b[i_b]
9404 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9406 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9407 cp = array_a[i_a++];
9410 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9414 /* Here, have chosen which of the two inputs to look at. Only output
9415 * if the running count changes to/from 2, which marks the
9416 * beginning/end of a range that's in the intersection */
9420 array_r[i_r++] = cp;
9425 array_r[i_r++] = cp;
9432 /* The loop above increments the index into exactly one of the input lists
9433 * each iteration, and ends when either index gets to its list end. That
9434 * means the other index is lower than its end, and so something is
9435 * remaining in that one. We increment 'count', as explained below, if the
9436 * exhausted list was in its set. (i_a and i_b each currently index the
9437 * element beyond the one we care about.) */
9438 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9439 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9444 /* Above we incremented 'count' if the exhausted list was in its set. This
9445 * has made it so that 'count' being below 2 means there is nothing left to
9446 * output; otheriwse what's left to add to the intersection is precisely
9447 * that which is left in the non-exhausted input list.
9449 * To see why, note first that the exhausted input obviously has nothing
9450 * left to affect the intersection. If it was in its set at its end, that
9451 * means the set extends from here to the platform's infinity, and hence
9452 * anything in the non-exhausted's list will be in the intersection, and
9453 * anything not in it won't be. Hence, the rest of the intersection is
9454 * precisely what's in the non-exhausted list The exhausted set also
9455 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9456 * it means 'count' is now at least 2. This is consistent with the
9457 * incremented 'count' being >= 2 means to add the non-exhausted list to
9460 * But if the exhausted input wasn't in its set, it contributed 0 to
9461 * 'count', and the intersection can't include anything further; the
9462 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9463 * incremented. This is consistent with 'count' being < 2 meaning nothing
9464 * further to add to the intersection. */
9465 if (count < 2) { /* Nothing left to put in the intersection. */
9468 else { /* copy the non-exhausted list, unchanged. */
9469 IV copy_count = len_a - i_a;
9470 if (copy_count > 0) { /* a is the one with stuff left */
9471 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9473 else { /* b is the one with stuff left */
9474 copy_count = len_b - i_b;
9475 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9477 len_r = i_r + copy_count;
9480 /* Set the result to the final length, which can change the pointer to
9481 * array_r, so re-find it. (Note that it is unlikely that this will
9482 * change, as we are shrinking the space, not enlarging it) */
9483 if (len_r != _invlist_len(r)) {
9484 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9486 array_r = invlist_array(r);
9489 if (*i == NULL) { /* Simply return the calculated intersection */
9492 else { /* Otherwise, replace the existing inversion list in '*i'. We could
9493 instead free '*i', and then set it to 'r', but experience has
9494 shown [perl #127392] that if the input is a mortal, we can get a
9495 huge build-up of these during regex compilation before they get
9498 invlist_replace_list_destroys_src(*i, r);
9510 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9512 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9513 * set. A pointer to the inversion list is returned. This may actually be
9514 * a new list, in which case the passed in one has been destroyed. The
9515 * passed-in inversion list can be NULL, in which case a new one is created
9516 * with just the one range in it. The new list is not necessarily
9517 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9518 * result of this function. The gain would not be large, and in many
9519 * cases, this is called multiple times on a single inversion list, so
9520 * anything freed may almost immediately be needed again.
9522 * This used to mostly call the 'union' routine, but that is much more
9523 * heavyweight than really needed for a single range addition */
9525 UV* array; /* The array implementing the inversion list */
9526 UV len; /* How many elements in 'array' */
9527 SSize_t i_s; /* index into the invlist array where 'start'
9529 SSize_t i_e = 0; /* And the index where 'end' should go */
9530 UV cur_highest; /* The highest code point in the inversion list
9531 upon entry to this function */
9533 /* This range becomes the whole inversion list if none already existed */
9534 if (invlist == NULL) {
9535 invlist = _new_invlist(2);
9536 _append_range_to_invlist(invlist, start, end);
9540 /* Likewise, if the inversion list is currently empty */
9541 len = _invlist_len(invlist);
9543 _append_range_to_invlist(invlist, start, end);
9547 /* Starting here, we have to know the internals of the list */
9548 array = invlist_array(invlist);
9550 /* If the new range ends higher than the current highest ... */
9551 cur_highest = invlist_highest(invlist);
9552 if (end > cur_highest) {
9554 /* If the whole range is higher, we can just append it */
9555 if (start > cur_highest) {
9556 _append_range_to_invlist(invlist, start, end);
9560 /* Otherwise, add the portion that is higher ... */
9561 _append_range_to_invlist(invlist, cur_highest + 1, end);
9563 /* ... and continue on below to handle the rest. As a result of the
9564 * above append, we know that the index of the end of the range is the
9565 * final even numbered one of the array. Recall that the final element
9566 * always starts a range that extends to infinity. If that range is in
9567 * the set (meaning the set goes from here to infinity), it will be an
9568 * even index, but if it isn't in the set, it's odd, and the final
9569 * range in the set is one less, which is even. */
9570 if (end == UV_MAX) {
9578 /* We have dealt with appending, now see about prepending. If the new
9579 * range starts lower than the current lowest ... */
9580 if (start < array[0]) {
9582 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
9583 * Let the union code handle it, rather than having to know the
9584 * trickiness in two code places. */
9585 if (UNLIKELY(start == 0)) {
9588 range_invlist = _new_invlist(2);
9589 _append_range_to_invlist(range_invlist, start, end);
9591 _invlist_union(invlist, range_invlist, &invlist);
9593 SvREFCNT_dec_NN(range_invlist);
9598 /* If the whole new range comes before the first entry, and doesn't
9599 * extend it, we have to insert it as an additional range */
9600 if (end < array[0] - 1) {
9602 goto splice_in_new_range;
9605 /* Here the new range adjoins the existing first range, extending it
9609 /* And continue on below to handle the rest. We know that the index of
9610 * the beginning of the range is the first one of the array */
9613 else { /* Not prepending any part of the new range to the existing list.
9614 * Find where in the list it should go. This finds i_s, such that:
9615 * invlist[i_s] <= start < array[i_s+1]
9617 i_s = _invlist_search(invlist, start);
9620 /* At this point, any extending before the beginning of the inversion list
9621 * and/or after the end has been done. This has made it so that, in the
9622 * code below, each endpoint of the new range is either in a range that is
9623 * in the set, or is in a gap between two ranges that are. This means we
9624 * don't have to worry about exceeding the array bounds.
9626 * Find where in the list the new range ends (but we can skip this if we
9627 * have already determined what it is, or if it will be the same as i_s,
9628 * which we already have computed) */
9630 i_e = (start == end)
9632 : _invlist_search(invlist, end);
9635 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
9636 * is a range that goes to infinity there is no element at invlist[i_e+1],
9637 * so only the first relation holds. */
9639 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9641 /* Here, the ranges on either side of the beginning of the new range
9642 * are in the set, and this range starts in the gap between them.
9644 * The new range extends the range above it downwards if the new range
9645 * ends at or above that range's start */
9646 const bool extends_the_range_above = ( end == UV_MAX
9647 || end + 1 >= array[i_s+1]);
9649 /* The new range extends the range below it upwards if it begins just
9650 * after where that range ends */
9651 if (start == array[i_s]) {
9653 /* If the new range fills the entire gap between the other ranges,
9654 * they will get merged together. Other ranges may also get
9655 * merged, depending on how many of them the new range spans. In
9656 * the general case, we do the merge later, just once, after we
9657 * figure out how many to merge. But in the case where the new
9658 * range exactly spans just this one gap (possibly extending into
9659 * the one above), we do the merge here, and an early exit. This
9660 * is done here to avoid having to special case later. */
9661 if (i_e - i_s <= 1) {
9663 /* If i_e - i_s == 1, it means that the new range terminates
9664 * within the range above, and hence 'extends_the_range_above'
9665 * must be true. (If the range above it extends to infinity,
9666 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
9667 * will be 0, so no harm done.) */
9668 if (extends_the_range_above) {
9669 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
9670 invlist_set_len(invlist,
9672 *(get_invlist_offset_addr(invlist)));
9676 /* Here, i_e must == i_s. We keep them in sync, as they apply
9677 * to the same range, and below we are about to decrement i_s
9682 /* Here, the new range is adjacent to the one below. (It may also
9683 * span beyond the range above, but that will get resolved later.)
9684 * Extend the range below to include this one. */
9685 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
9689 else if (extends_the_range_above) {
9691 /* Here the new range only extends the range above it, but not the
9692 * one below. It merges with the one above. Again, we keep i_e
9693 * and i_s in sync if they point to the same range */
9702 /* Here, we've dealt with the new range start extending any adjoining
9705 * If the new range extends to infinity, it is now the final one,
9706 * regardless of what was there before */
9707 if (UNLIKELY(end == UV_MAX)) {
9708 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
9712 /* If i_e started as == i_s, it has also been dealt with,
9713 * and been updated to the new i_s, which will fail the following if */
9714 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
9716 /* Here, the ranges on either side of the end of the new range are in
9717 * the set, and this range ends in the gap between them.
9719 * If this range is adjacent to (hence extends) the range above it, it
9720 * becomes part of that range; likewise if it extends the range below,
9721 * it becomes part of that range */
9722 if (end + 1 == array[i_e+1]) {
9726 else if (start <= array[i_e]) {
9727 array[i_e] = end + 1;
9734 /* If the range fits entirely in an existing range (as possibly already
9735 * extended above), it doesn't add anything new */
9736 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9740 /* Here, no part of the range is in the list. Must add it. It will
9741 * occupy 2 more slots */
9742 splice_in_new_range:
9744 invlist_extend(invlist, len + 2);
9745 array = invlist_array(invlist);
9746 /* Move the rest of the array down two slots. Don't include any
9748 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
9750 /* Do the actual splice */
9751 array[i_e+1] = start;
9752 array[i_e+2] = end + 1;
9753 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
9757 /* Here the new range crossed the boundaries of a pre-existing range. The
9758 * code above has adjusted things so that both ends are in ranges that are
9759 * in the set. This means everything in between must also be in the set.
9760 * Just squash things together */
9761 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
9762 invlist_set_len(invlist,
9764 *(get_invlist_offset_addr(invlist)));
9770 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9771 UV** other_elements_ptr)
9773 /* Create and return an inversion list whose contents are to be populated
9774 * by the caller. The caller gives the number of elements (in 'size') and
9775 * the very first element ('element0'). This function will set
9776 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9779 * Obviously there is some trust involved that the caller will properly
9780 * fill in the other elements of the array.
9782 * (The first element needs to be passed in, as the underlying code does
9783 * things differently depending on whether it is zero or non-zero) */
9785 SV* invlist = _new_invlist(size);
9788 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9790 invlist = add_cp_to_invlist(invlist, element0);
9791 offset = *get_invlist_offset_addr(invlist);
9793 invlist_set_len(invlist, size, offset);
9794 *other_elements_ptr = invlist_array(invlist) + 1;
9800 PERL_STATIC_INLINE SV*
9801 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9802 return _add_range_to_invlist(invlist, cp, cp);
9805 #ifndef PERL_IN_XSUB_RE
9807 Perl__invlist_invert(pTHX_ SV* const invlist)
9809 /* Complement the input inversion list. This adds a 0 if the list didn't
9810 * have a zero; removes it otherwise. As described above, the data
9811 * structure is set up so that this is very efficient */
9813 PERL_ARGS_ASSERT__INVLIST_INVERT;
9815 assert(! invlist_is_iterating(invlist));
9817 /* The inverse of matching nothing is matching everything */
9818 if (_invlist_len(invlist) == 0) {
9819 _append_range_to_invlist(invlist, 0, UV_MAX);
9823 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9828 PERL_STATIC_INLINE SV*
9829 S_invlist_clone(pTHX_ SV* const invlist)
9832 /* Return a new inversion list that is a copy of the input one, which is
9833 * unchanged. The new list will not be mortal even if the old one was. */
9835 /* Need to allocate extra space to accommodate Perl's addition of a
9836 * trailing NUL to SvPV's, since it thinks they are always strings */
9837 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9838 STRLEN physical_length = SvCUR(invlist);
9839 bool offset = *(get_invlist_offset_addr(invlist));
9841 PERL_ARGS_ASSERT_INVLIST_CLONE;
9843 *(get_invlist_offset_addr(new_invlist)) = offset;
9844 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9845 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9850 PERL_STATIC_INLINE STRLEN*
9851 S_get_invlist_iter_addr(SV* invlist)
9853 /* Return the address of the UV that contains the current iteration
9856 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9858 assert(SvTYPE(invlist) == SVt_INVLIST);
9860 return &(((XINVLIST*) SvANY(invlist))->iterator);
9863 PERL_STATIC_INLINE void
9864 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9866 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9868 *get_invlist_iter_addr(invlist) = 0;
9871 PERL_STATIC_INLINE void
9872 S_invlist_iterfinish(SV* invlist)
9874 /* Terminate iterator for invlist. This is to catch development errors.
9875 * Any iteration that is interrupted before completed should call this
9876 * function. Functions that add code points anywhere else but to the end
9877 * of an inversion list assert that they are not in the middle of an
9878 * iteration. If they were, the addition would make the iteration
9879 * problematical: if the iteration hadn't reached the place where things
9880 * were being added, it would be ok */
9882 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9884 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9888 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9890 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9891 * This call sets in <*start> and <*end>, the next range in <invlist>.
9892 * Returns <TRUE> if successful and the next call will return the next
9893 * range; <FALSE> if was already at the end of the list. If the latter,
9894 * <*start> and <*end> are unchanged, and the next call to this function
9895 * will start over at the beginning of the list */
9897 STRLEN* pos = get_invlist_iter_addr(invlist);
9898 UV len = _invlist_len(invlist);
9901 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9904 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9908 array = invlist_array(invlist);
9910 *start = array[(*pos)++];
9916 *end = array[(*pos)++] - 1;
9922 PERL_STATIC_INLINE UV
9923 S_invlist_highest(SV* const invlist)
9925 /* Returns the highest code point that matches an inversion list. This API
9926 * has an ambiguity, as it returns 0 under either the highest is actually
9927 * 0, or if the list is empty. If this distinction matters to you, check
9928 * for emptiness before calling this function */
9930 UV len = _invlist_len(invlist);
9933 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9939 array = invlist_array(invlist);
9941 /* The last element in the array in the inversion list always starts a
9942 * range that goes to infinity. That range may be for code points that are
9943 * matched in the inversion list, or it may be for ones that aren't
9944 * matched. In the latter case, the highest code point in the set is one
9945 * less than the beginning of this range; otherwise it is the final element
9946 * of this range: infinity */
9947 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
9949 : array[len - 1] - 1;
9953 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
9955 /* Get the contents of an inversion list into a string SV so that they can
9956 * be printed out. If 'traditional_style' is TRUE, it uses the format
9957 * traditionally done for debug tracing; otherwise it uses a format
9958 * suitable for just copying to the output, with blanks between ranges and
9959 * a dash between range components */
9963 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
9964 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
9966 if (traditional_style) {
9967 output = newSVpvs("\n");
9970 output = newSVpvs("");
9973 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
9975 assert(! invlist_is_iterating(invlist));
9977 invlist_iterinit(invlist);
9978 while (invlist_iternext(invlist, &start, &end)) {
9979 if (end == UV_MAX) {
9980 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFINITY%c",
9981 start, intra_range_delimiter,
9982 inter_range_delimiter);
9984 else if (end != start) {
9985 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
9987 intra_range_delimiter,
9988 end, inter_range_delimiter);
9991 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
9992 start, inter_range_delimiter);
9996 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
9997 SvCUR_set(output, SvCUR(output) - 1);
10003 #ifndef PERL_IN_XSUB_RE
10005 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10006 const char * const indent, SV* const invlist)
10008 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10009 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10010 * the string 'indent'. The output looks like this:
10011 [0] 0x000A .. 0x000D
10013 [4] 0x2028 .. 0x2029
10014 [6] 0x3104 .. INFINITY
10015 * This means that the first range of code points matched by the list are
10016 * 0xA through 0xD; the second range contains only the single code point
10017 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10018 * are used to define each range (except if the final range extends to
10019 * infinity, only a single element is needed). The array index of the
10020 * first element for the corresponding range is given in brackets. */
10025 PERL_ARGS_ASSERT__INVLIST_DUMP;
10027 if (invlist_is_iterating(invlist)) {
10028 Perl_dump_indent(aTHX_ level, file,
10029 "%sCan't dump inversion list because is in middle of iterating\n",
10034 invlist_iterinit(invlist);
10035 while (invlist_iternext(invlist, &start, &end)) {
10036 if (end == UV_MAX) {
10037 Perl_dump_indent(aTHX_ level, file,
10038 "%s[%" UVuf "] 0x%04" UVXf " .. INFINITY\n",
10039 indent, (UV)count, start);
10041 else if (end != start) {
10042 Perl_dump_indent(aTHX_ level, file,
10043 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10044 indent, (UV)count, start, end);
10047 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10048 indent, (UV)count, start);
10055 Perl__load_PL_utf8_foldclosures (pTHX)
10057 assert(! PL_utf8_foldclosures);
10059 /* If the folds haven't been read in, call a fold function
10061 if (! PL_utf8_tofold) {
10062 U8 dummy[UTF8_MAXBYTES_CASE+1];
10063 const U8 hyphen[] = HYPHEN_UTF8;
10065 /* This string is just a short named one above \xff */
10066 toFOLD_utf8_safe(hyphen, hyphen + sizeof(hyphen) - 1, dummy, NULL);
10067 assert(PL_utf8_tofold); /* Verify that worked */
10069 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
10073 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10075 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10077 /* Return a boolean as to if the two passed in inversion lists are
10078 * identical. The final argument, if TRUE, says to take the complement of
10079 * the second inversion list before doing the comparison */
10081 const UV* array_a = invlist_array(a);
10082 const UV* array_b = invlist_array(b);
10083 UV len_a = _invlist_len(a);
10084 UV len_b = _invlist_len(b);
10086 PERL_ARGS_ASSERT__INVLISTEQ;
10088 /* If are to compare 'a' with the complement of b, set it
10089 * up so are looking at b's complement. */
10090 if (complement_b) {
10092 /* The complement of nothing is everything, so <a> would have to have
10093 * just one element, starting at zero (ending at infinity) */
10095 return (len_a == 1 && array_a[0] == 0);
10097 else if (array_b[0] == 0) {
10099 /* Otherwise, to complement, we invert. Here, the first element is
10100 * 0, just remove it. To do this, we just pretend the array starts
10108 /* But if the first element is not zero, we pretend the list starts
10109 * at the 0 that is always stored immediately before the array. */
10115 return len_a == len_b
10116 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10122 * As best we can, determine the characters that can match the start of
10123 * the given EXACTF-ish node.
10125 * Returns the invlist as a new SV*; it is the caller's responsibility to
10126 * call SvREFCNT_dec() when done with it.
10129 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10131 const U8 * s = (U8*)STRING(node);
10132 SSize_t bytelen = STR_LEN(node);
10134 /* Start out big enough for 2 separate code points */
10135 SV* invlist = _new_invlist(4);
10137 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
10142 /* We punt and assume can match anything if the node begins
10143 * with a multi-character fold. Things are complicated. For
10144 * example, /ffi/i could match any of:
10145 * "\N{LATIN SMALL LIGATURE FFI}"
10146 * "\N{LATIN SMALL LIGATURE FF}I"
10147 * "F\N{LATIN SMALL LIGATURE FI}"
10148 * plus several other things; and making sure we have all the
10149 * possibilities is hard. */
10150 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10151 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10154 /* Any Latin1 range character can potentially match any
10155 * other depending on the locale */
10156 if (OP(node) == EXACTFL) {
10157 _invlist_union(invlist, PL_Latin1, &invlist);
10160 /* But otherwise, it matches at least itself. We can
10161 * quickly tell if it has a distinct fold, and if so,
10162 * it matches that as well */
10163 invlist = add_cp_to_invlist(invlist, uc);
10164 if (IS_IN_SOME_FOLD_L1(uc))
10165 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10168 /* Some characters match above-Latin1 ones under /i. This
10169 * is true of EXACTFL ones when the locale is UTF-8 */
10170 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10171 && (! isASCII(uc) || (OP(node) != EXACTFA
10172 && OP(node) != EXACTFA_NO_TRIE)))
10174 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10178 else { /* Pattern is UTF-8 */
10179 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10180 STRLEN foldlen = UTF8SKIP(s);
10181 const U8* e = s + bytelen;
10184 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10186 /* The only code points that aren't folded in a UTF EXACTFish
10187 * node are are the problematic ones in EXACTFL nodes */
10188 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10189 /* We need to check for the possibility that this EXACTFL
10190 * node begins with a multi-char fold. Therefore we fold
10191 * the first few characters of it so that we can make that
10196 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10198 *(d++) = (U8) toFOLD(*s);
10203 toFOLD_utf8_safe(s, e, d, &len);
10209 /* And set up so the code below that looks in this folded
10210 * buffer instead of the node's string */
10212 foldlen = UTF8SKIP(folded);
10216 /* When we reach here 's' points to the fold of the first
10217 * character(s) of the node; and 'e' points to far enough along
10218 * the folded string to be just past any possible multi-char
10219 * fold. 'foldlen' is the length in bytes of the first
10222 * Unlike the non-UTF-8 case, the macro for determining if a
10223 * string is a multi-char fold requires all the characters to
10224 * already be folded. This is because of all the complications
10225 * if not. Note that they are folded anyway, except in EXACTFL
10226 * nodes. Like the non-UTF case above, we punt if the node
10227 * begins with a multi-char fold */
10229 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10230 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10232 else { /* Single char fold */
10234 /* It matches all the things that fold to it, which are
10235 * found in PL_utf8_foldclosures (including itself) */
10236 invlist = add_cp_to_invlist(invlist, uc);
10237 if (! PL_utf8_foldclosures)
10238 _load_PL_utf8_foldclosures();
10239 if ((listp = hv_fetch(PL_utf8_foldclosures,
10240 (char *) s, foldlen, FALSE)))
10242 AV* list = (AV*) *listp;
10244 for (k = 0; k <= av_tindex_nomg(list); k++) {
10245 SV** c_p = av_fetch(list, k, FALSE);
10251 /* /aa doesn't allow folds between ASCII and non- */
10252 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
10253 && isASCII(c) != isASCII(uc))
10258 invlist = add_cp_to_invlist(invlist, c);
10267 #undef HEADER_LENGTH
10268 #undef TO_INTERNAL_SIZE
10269 #undef FROM_INTERNAL_SIZE
10270 #undef INVLIST_VERSION_ID
10272 /* End of inversion list object */
10275 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10277 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10278 * constructs, and updates RExC_flags with them. On input, RExC_parse
10279 * should point to the first flag; it is updated on output to point to the
10280 * final ')' or ':'. There needs to be at least one flag, or this will
10283 /* for (?g), (?gc), and (?o) warnings; warning
10284 about (?c) will warn about (?g) -- japhy */
10286 #define WASTED_O 0x01
10287 #define WASTED_G 0x02
10288 #define WASTED_C 0x04
10289 #define WASTED_GC (WASTED_G|WASTED_C)
10290 I32 wastedflags = 0x00;
10291 U32 posflags = 0, negflags = 0;
10292 U32 *flagsp = &posflags;
10293 char has_charset_modifier = '\0';
10295 bool has_use_defaults = FALSE;
10296 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10297 int x_mod_count = 0;
10299 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10301 /* '^' as an initial flag sets certain defaults */
10302 if (UCHARAT(RExC_parse) == '^') {
10304 has_use_defaults = TRUE;
10305 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10306 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
10307 ? REGEX_UNICODE_CHARSET
10308 : REGEX_DEPENDS_CHARSET);
10311 cs = get_regex_charset(RExC_flags);
10312 if (cs == REGEX_DEPENDS_CHARSET
10313 && (RExC_utf8 || RExC_uni_semantics))
10315 cs = REGEX_UNICODE_CHARSET;
10318 while (RExC_parse < RExC_end) {
10319 /* && strchr("iogcmsx", *RExC_parse) */
10320 /* (?g), (?gc) and (?o) are useless here
10321 and must be globally applied -- japhy */
10322 switch (*RExC_parse) {
10324 /* Code for the imsxn flags */
10325 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10327 case LOCALE_PAT_MOD:
10328 if (has_charset_modifier) {
10329 goto excess_modifier;
10331 else if (flagsp == &negflags) {
10334 cs = REGEX_LOCALE_CHARSET;
10335 has_charset_modifier = LOCALE_PAT_MOD;
10337 case UNICODE_PAT_MOD:
10338 if (has_charset_modifier) {
10339 goto excess_modifier;
10341 else if (flagsp == &negflags) {
10344 cs = REGEX_UNICODE_CHARSET;
10345 has_charset_modifier = UNICODE_PAT_MOD;
10347 case ASCII_RESTRICT_PAT_MOD:
10348 if (flagsp == &negflags) {
10351 if (has_charset_modifier) {
10352 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10353 goto excess_modifier;
10355 /* Doubled modifier implies more restricted */
10356 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10359 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10361 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10363 case DEPENDS_PAT_MOD:
10364 if (has_use_defaults) {
10365 goto fail_modifiers;
10367 else if (flagsp == &negflags) {
10370 else if (has_charset_modifier) {
10371 goto excess_modifier;
10374 /* The dual charset means unicode semantics if the
10375 * pattern (or target, not known until runtime) are
10376 * utf8, or something in the pattern indicates unicode
10378 cs = (RExC_utf8 || RExC_uni_semantics)
10379 ? REGEX_UNICODE_CHARSET
10380 : REGEX_DEPENDS_CHARSET;
10381 has_charset_modifier = DEPENDS_PAT_MOD;
10385 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10386 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10388 else if (has_charset_modifier == *(RExC_parse - 1)) {
10389 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10390 *(RExC_parse - 1));
10393 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10395 NOT_REACHED; /*NOTREACHED*/
10398 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10399 *(RExC_parse - 1));
10400 NOT_REACHED; /*NOTREACHED*/
10401 case ONCE_PAT_MOD: /* 'o' */
10402 case GLOBAL_PAT_MOD: /* 'g' */
10403 if (PASS2 && ckWARN(WARN_REGEXP)) {
10404 const I32 wflagbit = *RExC_parse == 'o'
10407 if (! (wastedflags & wflagbit) ) {
10408 wastedflags |= wflagbit;
10409 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10412 "Useless (%s%c) - %suse /%c modifier",
10413 flagsp == &negflags ? "?-" : "?",
10415 flagsp == &negflags ? "don't " : "",
10422 case CONTINUE_PAT_MOD: /* 'c' */
10423 if (PASS2 && ckWARN(WARN_REGEXP)) {
10424 if (! (wastedflags & WASTED_C) ) {
10425 wastedflags |= WASTED_GC;
10426 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10429 "Useless (%sc) - %suse /gc modifier",
10430 flagsp == &negflags ? "?-" : "?",
10431 flagsp == &negflags ? "don't " : ""
10436 case KEEPCOPY_PAT_MOD: /* 'p' */
10437 if (flagsp == &negflags) {
10439 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10441 *flagsp |= RXf_PMf_KEEPCOPY;
10445 /* A flag is a default iff it is following a minus, so
10446 * if there is a minus, it means will be trying to
10447 * re-specify a default which is an error */
10448 if (has_use_defaults || flagsp == &negflags) {
10449 goto fail_modifiers;
10451 flagsp = &negflags;
10452 wastedflags = 0; /* reset so (?g-c) warns twice */
10458 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
10459 negflags |= RXf_PMf_EXTENDED_MORE;
10461 RExC_flags |= posflags;
10463 if (negflags & RXf_PMf_EXTENDED) {
10464 negflags |= RXf_PMf_EXTENDED_MORE;
10466 RExC_flags &= ~negflags;
10467 set_regex_charset(&RExC_flags, cs);
10472 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10473 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10474 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
10475 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10476 NOT_REACHED; /*NOTREACHED*/
10479 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10482 vFAIL("Sequence (?... not terminated");
10486 - reg - regular expression, i.e. main body or parenthesized thing
10488 * Caller must absorb opening parenthesis.
10490 * Combining parenthesis handling with the base level of regular expression
10491 * is a trifle forced, but the need to tie the tails of the branches to what
10492 * follows makes it hard to avoid.
10494 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10496 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10498 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10501 PERL_STATIC_INLINE regnode *
10502 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10504 char * parse_start,
10509 char* name_start = RExC_parse;
10511 SV *sv_dat = reg_scan_name(pRExC_state, SIZE_ONLY
10512 ? REG_RSN_RETURN_NULL
10513 : REG_RSN_RETURN_DATA);
10514 GET_RE_DEBUG_FLAGS_DECL;
10516 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10518 if (RExC_parse == name_start || *RExC_parse != ch) {
10519 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10520 vFAIL2("Sequence %.3s... not terminated",parse_start);
10524 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10525 RExC_rxi->data->data[num]=(void*)sv_dat;
10526 SvREFCNT_inc_simple_void(sv_dat);
10529 ret = reganode(pRExC_state,
10532 : (ASCII_FOLD_RESTRICTED)
10534 : (AT_LEAST_UNI_SEMANTICS)
10540 *flagp |= HASWIDTH;
10542 Set_Node_Offset(ret, parse_start+1);
10543 Set_Node_Cur_Length(ret, parse_start);
10545 nextchar(pRExC_state);
10549 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
10550 flags. Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan
10551 needs to be restarted, or'd with NEED_UTF8 if the pattern needs to be
10552 upgraded to UTF-8. Otherwise would only return NULL if regbranch() returns
10553 NULL, which cannot happen. */
10555 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
10556 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10557 * 2 is like 1, but indicates that nextchar() has been called to advance
10558 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10559 * this flag alerts us to the need to check for that */
10561 regnode *ret; /* Will be the head of the group. */
10564 regnode *ender = NULL;
10567 U32 oregflags = RExC_flags;
10568 bool have_branch = 0;
10570 I32 freeze_paren = 0;
10571 I32 after_freeze = 0;
10572 I32 num; /* numeric backreferences */
10574 char * parse_start = RExC_parse; /* MJD */
10575 char * const oregcomp_parse = RExC_parse;
10577 GET_RE_DEBUG_FLAGS_DECL;
10579 PERL_ARGS_ASSERT_REG;
10580 DEBUG_PARSE("reg ");
10582 *flagp = 0; /* Tentatively. */
10584 /* Having this true makes it feasible to have a lot fewer tests for the
10585 * parse pointer being in scope. For example, we can write
10586 * while(isFOO(*RExC_parse)) RExC_parse++;
10588 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
10590 assert(*RExC_end == '\0');
10592 /* Make an OPEN node, if parenthesized. */
10595 /* Under /x, space and comments can be gobbled up between the '(' and
10596 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
10597 * intervening space, as the sequence is a token, and a token should be
10599 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
10601 if (RExC_parse >= RExC_end) {
10602 vFAIL("Unmatched (");
10605 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
10606 char *start_verb = RExC_parse + 1;
10608 char *start_arg = NULL;
10609 unsigned char op = 0;
10610 int arg_required = 0;
10611 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
10613 if (has_intervening_patws) {
10614 RExC_parse++; /* past the '*' */
10615 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
10617 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
10618 if ( *RExC_parse == ':' ) {
10619 start_arg = RExC_parse + 1;
10622 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10624 verb_len = RExC_parse - start_verb;
10626 if (RExC_parse >= RExC_end) {
10627 goto unterminated_verb_pattern;
10629 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10630 while ( RExC_parse < RExC_end && *RExC_parse != ')' )
10631 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10632 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10633 unterminated_verb_pattern:
10634 vFAIL("Unterminated verb pattern argument");
10635 if ( RExC_parse == start_arg )
10638 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10639 vFAIL("Unterminated verb pattern");
10642 /* Here, we know that RExC_parse < RExC_end */
10644 switch ( *start_verb ) {
10645 case 'A': /* (*ACCEPT) */
10646 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
10648 internal_argval = RExC_nestroot;
10651 case 'C': /* (*COMMIT) */
10652 if ( memEQs(start_verb,verb_len,"COMMIT") )
10655 case 'F': /* (*FAIL) */
10656 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
10660 case ':': /* (*:NAME) */
10661 case 'M': /* (*MARK:NAME) */
10662 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
10667 case 'P': /* (*PRUNE) */
10668 if ( memEQs(start_verb,verb_len,"PRUNE") )
10671 case 'S': /* (*SKIP) */
10672 if ( memEQs(start_verb,verb_len,"SKIP") )
10675 case 'T': /* (*THEN) */
10676 /* [19:06] <TimToady> :: is then */
10677 if ( memEQs(start_verb,verb_len,"THEN") ) {
10679 RExC_seen |= REG_CUTGROUP_SEEN;
10684 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10686 "Unknown verb pattern '%" UTF8f "'",
10687 UTF8fARG(UTF, verb_len, start_verb));
10689 if ( arg_required && !start_arg ) {
10690 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
10691 verb_len, start_verb);
10693 if (internal_argval == -1) {
10694 ret = reganode(pRExC_state, op, 0);
10696 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
10698 RExC_seen |= REG_VERBARG_SEEN;
10699 if ( ! SIZE_ONLY ) {
10701 SV *sv = newSVpvn( start_arg,
10702 RExC_parse - start_arg);
10703 ARG(ret) = add_data( pRExC_state,
10704 STR_WITH_LEN("S"));
10705 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
10710 if ( internal_argval != -1 )
10711 ARG2L_SET(ret, internal_argval);
10713 nextchar(pRExC_state);
10716 else if (*RExC_parse == '?') { /* (?...) */
10717 bool is_logical = 0;
10718 const char * const seqstart = RExC_parse;
10719 const char * endptr;
10720 if (has_intervening_patws) {
10722 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
10725 RExC_parse++; /* past the '?' */
10726 paren = *RExC_parse; /* might be a trailing NUL, if not
10728 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10729 if (RExC_parse > RExC_end) {
10732 ret = NULL; /* For look-ahead/behind. */
10735 case 'P': /* (?P...) variants for those used to PCRE/Python */
10736 paren = *RExC_parse;
10737 if ( paren == '<') { /* (?P<...>) named capture */
10739 if (RExC_parse >= RExC_end) {
10740 vFAIL("Sequence (?P<... not terminated");
10742 goto named_capture;
10744 else if (paren == '>') { /* (?P>name) named recursion */
10746 if (RExC_parse >= RExC_end) {
10747 vFAIL("Sequence (?P>... not terminated");
10749 goto named_recursion;
10751 else if (paren == '=') { /* (?P=...) named backref */
10753 return handle_named_backref(pRExC_state, flagp,
10756 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10757 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10758 vFAIL3("Sequence (%.*s...) not recognized",
10759 RExC_parse-seqstart, seqstart);
10760 NOT_REACHED; /*NOTREACHED*/
10761 case '<': /* (?<...) */
10762 if (*RExC_parse == '!')
10764 else if (*RExC_parse != '=')
10771 case '\'': /* (?'...') */
10772 name_start = RExC_parse;
10773 svname = reg_scan_name(pRExC_state,
10774 SIZE_ONLY /* reverse test from the others */
10775 ? REG_RSN_RETURN_NAME
10776 : REG_RSN_RETURN_NULL);
10777 if ( RExC_parse == name_start
10778 || RExC_parse >= RExC_end
10779 || *RExC_parse != paren)
10781 vFAIL2("Sequence (?%c... not terminated",
10782 paren=='>' ? '<' : paren);
10787 if (!svname) /* shouldn't happen */
10789 "panic: reg_scan_name returned NULL");
10790 if (!RExC_paren_names) {
10791 RExC_paren_names= newHV();
10792 sv_2mortal(MUTABLE_SV(RExC_paren_names));
10794 RExC_paren_name_list= newAV();
10795 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10798 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10800 sv_dat = HeVAL(he_str);
10802 /* croak baby croak */
10804 "panic: paren_name hash element allocation failed");
10805 } else if ( SvPOK(sv_dat) ) {
10806 /* (?|...) can mean we have dupes so scan to check
10807 its already been stored. Maybe a flag indicating
10808 we are inside such a construct would be useful,
10809 but the arrays are likely to be quite small, so
10810 for now we punt -- dmq */
10811 IV count = SvIV(sv_dat);
10812 I32 *pv = (I32*)SvPVX(sv_dat);
10814 for ( i = 0 ; i < count ; i++ ) {
10815 if ( pv[i] == RExC_npar ) {
10821 pv = (I32*)SvGROW(sv_dat,
10822 SvCUR(sv_dat) + sizeof(I32)+1);
10823 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10824 pv[count] = RExC_npar;
10825 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10828 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10829 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10832 SvIV_set(sv_dat, 1);
10835 /* Yes this does cause a memory leak in debugging Perls
10837 if (!av_store(RExC_paren_name_list,
10838 RExC_npar, SvREFCNT_inc(svname)))
10839 SvREFCNT_dec_NN(svname);
10842 /*sv_dump(sv_dat);*/
10844 nextchar(pRExC_state);
10846 goto capturing_parens;
10848 RExC_seen |= REG_LOOKBEHIND_SEEN;
10849 RExC_in_lookbehind++;
10851 if (RExC_parse >= RExC_end) {
10852 vFAIL("Sequence (?... not terminated");
10856 case '=': /* (?=...) */
10857 RExC_seen_zerolen++;
10859 case '!': /* (?!...) */
10860 RExC_seen_zerolen++;
10861 /* check if we're really just a "FAIL" assertion */
10862 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
10863 FALSE /* Don't force to /x */ );
10864 if (*RExC_parse == ')') {
10865 ret=reganode(pRExC_state, OPFAIL, 0);
10866 nextchar(pRExC_state);
10870 case '|': /* (?|...) */
10871 /* branch reset, behave like a (?:...) except that
10872 buffers in alternations share the same numbers */
10874 after_freeze = freeze_paren = RExC_npar;
10876 case ':': /* (?:...) */
10877 case '>': /* (?>...) */
10879 case '$': /* (?$...) */
10880 case '@': /* (?@...) */
10881 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10883 case '0' : /* (?0) */
10884 case 'R' : /* (?R) */
10885 if (RExC_parse == RExC_end || *RExC_parse != ')')
10886 FAIL("Sequence (?R) not terminated");
10888 RExC_seen |= REG_RECURSE_SEEN;
10889 *flagp |= POSTPONED;
10890 goto gen_recurse_regop;
10892 /* named and numeric backreferences */
10893 case '&': /* (?&NAME) */
10894 parse_start = RExC_parse - 1;
10897 SV *sv_dat = reg_scan_name(pRExC_state,
10898 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10899 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10901 if (RExC_parse >= RExC_end || *RExC_parse != ')')
10902 vFAIL("Sequence (?&... not terminated");
10903 goto gen_recurse_regop;
10906 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10908 vFAIL("Illegal pattern");
10910 goto parse_recursion;
10912 case '-': /* (?-1) */
10913 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10914 RExC_parse--; /* rewind to let it be handled later */
10918 case '1': case '2': case '3': case '4': /* (?1) */
10919 case '5': case '6': case '7': case '8': case '9':
10920 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
10923 bool is_neg = FALSE;
10925 parse_start = RExC_parse - 1; /* MJD */
10926 if (*RExC_parse == '-') {
10930 if (grok_atoUV(RExC_parse, &unum, &endptr)
10934 RExC_parse = (char*)endptr;
10938 /* Some limit for num? */
10942 if (*RExC_parse!=')')
10943 vFAIL("Expecting close bracket");
10946 if ( paren == '-' ) {
10948 Diagram of capture buffer numbering.
10949 Top line is the normal capture buffer numbers
10950 Bottom line is the negative indexing as from
10954 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
10958 num = RExC_npar + num;
10961 vFAIL("Reference to nonexistent group");
10963 } else if ( paren == '+' ) {
10964 num = RExC_npar + num - 1;
10966 /* We keep track how many GOSUB items we have produced.
10967 To start off the ARG2L() of the GOSUB holds its "id",
10968 which is used later in conjunction with RExC_recurse
10969 to calculate the offset we need to jump for the GOSUB,
10970 which it will store in the final representation.
10971 We have to defer the actual calculation until much later
10972 as the regop may move.
10975 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
10977 if (num > (I32)RExC_rx->nparens) {
10979 vFAIL("Reference to nonexistent group");
10981 RExC_recurse_count++;
10982 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
10983 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
10984 22, "| |", (int)(depth * 2 + 1), "",
10985 (UV)ARG(ret), (IV)ARG2L(ret)));
10987 RExC_seen |= REG_RECURSE_SEEN;
10989 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
10990 Set_Node_Offset(ret, parse_start); /* MJD */
10992 *flagp |= POSTPONED;
10993 assert(*RExC_parse == ')');
10994 nextchar(pRExC_state);
10999 case '?': /* (??...) */
11001 if (*RExC_parse != '{') {
11002 RExC_parse += SKIP_IF_CHAR(RExC_parse);
11003 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11005 "Sequence (%" UTF8f "...) not recognized",
11006 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11007 NOT_REACHED; /*NOTREACHED*/
11009 *flagp |= POSTPONED;
11013 case '{': /* (?{...}) */
11016 struct reg_code_block *cb;
11018 RExC_seen_zerolen++;
11020 if ( !pRExC_state->code_blocks
11021 || pRExC_state->code_index
11022 >= pRExC_state->code_blocks->count
11023 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11024 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11027 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11028 FAIL("panic: Sequence (?{...}): no code block found\n");
11029 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11031 /* this is a pre-compiled code block (?{...}) */
11032 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11033 RExC_parse = RExC_start + cb->end;
11036 if (cb->src_regex) {
11037 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11038 RExC_rxi->data->data[n] =
11039 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11040 RExC_rxi->data->data[n+1] = (void*)o;
11043 n = add_data(pRExC_state,
11044 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11045 RExC_rxi->data->data[n] = (void*)o;
11048 pRExC_state->code_index++;
11049 nextchar(pRExC_state);
11053 ret = reg_node(pRExC_state, LOGICAL);
11055 eval = reg2Lanode(pRExC_state, EVAL,
11058 /* for later propagation into (??{})
11060 RExC_flags & RXf_PMf_COMPILETIME
11065 REGTAIL(pRExC_state, ret, eval);
11066 /* deal with the length of this later - MJD */
11069 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11070 Set_Node_Length(ret, RExC_parse - parse_start + 1);
11071 Set_Node_Offset(ret, parse_start);
11074 case '(': /* (?(?{...})...) and (?(?=...)...) */
11077 const int DEFINE_len = sizeof("DEFINE") - 1;
11078 if (RExC_parse[0] == '?') { /* (?(?...)) */
11079 if ( RExC_parse < RExC_end - 1
11080 && ( RExC_parse[1] == '='
11081 || RExC_parse[1] == '!'
11082 || RExC_parse[1] == '<'
11083 || RExC_parse[1] == '{')
11084 ) { /* Lookahead or eval. */
11088 ret = reg_node(pRExC_state, LOGICAL);
11092 tail = reg(pRExC_state, 1, &flag, depth+1);
11093 if (flag & (RESTART_PASS1|NEED_UTF8)) {
11094 *flagp = flag & (RESTART_PASS1|NEED_UTF8);
11097 REGTAIL(pRExC_state, ret, tail);
11100 /* Fall through to ‘Unknown switch condition’ at the
11101 end of the if/else chain. */
11103 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11104 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11106 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11107 char *name_start= RExC_parse++;
11109 SV *sv_dat=reg_scan_name(pRExC_state,
11110 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11111 if ( RExC_parse == name_start
11112 || RExC_parse >= RExC_end
11113 || *RExC_parse != ch)
11115 vFAIL2("Sequence (?(%c... not terminated",
11116 (ch == '>' ? '<' : ch));
11120 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11121 RExC_rxi->data->data[num]=(void*)sv_dat;
11122 SvREFCNT_inc_simple_void(sv_dat);
11124 ret = reganode(pRExC_state,NGROUPP,num);
11125 goto insert_if_check_paren;
11127 else if (RExC_end - RExC_parse >= DEFINE_len
11128 && strnEQ(RExC_parse, "DEFINE", DEFINE_len))
11130 ret = reganode(pRExC_state,DEFINEP,0);
11131 RExC_parse += DEFINE_len;
11133 goto insert_if_check_paren;
11135 else if (RExC_parse[0] == 'R') {
11137 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11138 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11139 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11142 if (RExC_parse[0] == '0') {
11146 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11148 if (grok_atoUV(RExC_parse, &uv, &endptr)
11151 parno = (I32)uv + 1;
11152 RExC_parse = (char*)endptr;
11154 /* else "Switch condition not recognized" below */
11155 } else if (RExC_parse[0] == '&') {
11158 sv_dat = reg_scan_name(pRExC_state,
11160 ? REG_RSN_RETURN_NULL
11161 : REG_RSN_RETURN_DATA);
11163 /* we should only have a false sv_dat when
11164 * SIZE_ONLY is true, and we always have false
11165 * sv_dat when SIZE_ONLY is true.
11166 * reg_scan_name() will VFAIL() if the name is
11167 * unknown when SIZE_ONLY is false, and otherwise
11168 * will return something, and when SIZE_ONLY is
11169 * true, reg_scan_name() just parses the string,
11170 * and doesnt return anything. (in theory) */
11171 assert(SIZE_ONLY ? !sv_dat : !!sv_dat);
11174 parno = 1 + *((I32 *)SvPVX(sv_dat));
11176 ret = reganode(pRExC_state,INSUBP,parno);
11177 goto insert_if_check_paren;
11179 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11183 if (grok_atoUV(RExC_parse, &uv, &endptr)
11187 RExC_parse = (char*)endptr;
11190 vFAIL("panic: grok_atoUV returned FALSE");
11192 ret = reganode(pRExC_state, GROUPP, parno);
11194 insert_if_check_paren:
11195 if (UCHARAT(RExC_parse) != ')') {
11196 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11197 vFAIL("Switch condition not recognized");
11199 nextchar(pRExC_state);
11201 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
11202 br = regbranch(pRExC_state, &flags, 1,depth+1);
11204 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11205 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11208 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11211 REGTAIL(pRExC_state, br, reganode(pRExC_state,
11213 c = UCHARAT(RExC_parse);
11214 nextchar(pRExC_state);
11215 if (flags&HASWIDTH)
11216 *flagp |= HASWIDTH;
11219 vFAIL("(?(DEFINE)....) does not allow branches");
11221 /* Fake one for optimizer. */
11222 lastbr = reganode(pRExC_state, IFTHEN, 0);
11224 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
11225 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11226 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11229 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11232 REGTAIL(pRExC_state, ret, lastbr);
11233 if (flags&HASWIDTH)
11234 *flagp |= HASWIDTH;
11235 c = UCHARAT(RExC_parse);
11236 nextchar(pRExC_state);
11241 if (RExC_parse >= RExC_end)
11242 vFAIL("Switch (?(condition)... not terminated");
11244 vFAIL("Switch (?(condition)... contains too many branches");
11246 ender = reg_node(pRExC_state, TAIL);
11247 REGTAIL(pRExC_state, br, ender);
11249 REGTAIL(pRExC_state, lastbr, ender);
11250 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11253 REGTAIL(pRExC_state, ret, ender);
11254 RExC_size++; /* XXX WHY do we need this?!!
11255 For large programs it seems to be required
11256 but I can't figure out why. -- dmq*/
11259 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11260 vFAIL("Unknown switch condition (?(...))");
11262 case '[': /* (?[ ... ]) */
11263 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
11265 case 0: /* A NUL */
11266 RExC_parse--; /* for vFAIL to print correctly */
11267 vFAIL("Sequence (? incomplete");
11269 default: /* e.g., (?i) */
11270 RExC_parse = (char *) seqstart + 1;
11272 parse_lparen_question_flags(pRExC_state);
11273 if (UCHARAT(RExC_parse) != ':') {
11274 if (RExC_parse < RExC_end)
11275 nextchar(pRExC_state);
11280 nextchar(pRExC_state);
11285 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11290 ret = reganode(pRExC_state, OPEN, parno);
11292 if (!RExC_nestroot)
11293 RExC_nestroot = parno;
11294 if (RExC_open_parens && !RExC_open_parens[parno])
11296 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11297 "%*s%*s Setting open paren #%" IVdf " to %d\n",
11298 22, "| |", (int)(depth * 2 + 1), "",
11299 (IV)parno, REG_NODE_NUM(ret)));
11300 RExC_open_parens[parno]= ret;
11303 Set_Node_Length(ret, 1); /* MJD */
11304 Set_Node_Offset(ret, RExC_parse); /* MJD */
11307 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
11316 /* Pick up the branches, linking them together. */
11317 parse_start = RExC_parse; /* MJD */
11318 br = regbranch(pRExC_state, &flags, 1,depth+1);
11320 /* branch_len = (paren != 0); */
11323 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11324 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11327 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11329 if (*RExC_parse == '|') {
11330 if (!SIZE_ONLY && RExC_extralen) {
11331 reginsert(pRExC_state, BRANCHJ, br, depth+1);
11334 reginsert(pRExC_state, BRANCH, br, depth+1);
11335 Set_Node_Length(br, paren != 0);
11336 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
11340 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
11342 else if (paren == ':') {
11343 *flagp |= flags&SIMPLE;
11345 if (is_open) { /* Starts with OPEN. */
11346 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
11348 else if (paren != '?') /* Not Conditional */
11350 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11352 while (*RExC_parse == '|') {
11353 if (!SIZE_ONLY && RExC_extralen) {
11354 ender = reganode(pRExC_state, LONGJMP,0);
11356 /* Append to the previous. */
11357 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11360 RExC_extralen += 2; /* Account for LONGJMP. */
11361 nextchar(pRExC_state);
11362 if (freeze_paren) {
11363 if (RExC_npar > after_freeze)
11364 after_freeze = RExC_npar;
11365 RExC_npar = freeze_paren;
11367 br = regbranch(pRExC_state, &flags, 0, depth+1);
11370 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11371 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11374 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11376 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
11378 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11381 if (have_branch || paren != ':') {
11382 /* Make a closing node, and hook it on the end. */
11385 ender = reg_node(pRExC_state, TAIL);
11388 ender = reganode(pRExC_state, CLOSE, parno);
11389 if ( RExC_close_parens ) {
11390 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11391 "%*s%*s Setting close paren #%" IVdf " to %d\n",
11392 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
11393 RExC_close_parens[parno]= ender;
11394 if (RExC_nestroot == parno)
11397 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
11398 Set_Node_Length(ender,1); /* MJD */
11404 *flagp &= ~HASWIDTH;
11407 ender = reg_node(pRExC_state, SUCCEED);
11410 ender = reg_node(pRExC_state, END);
11412 assert(!RExC_end_op); /* there can only be one! */
11413 RExC_end_op = ender;
11414 if (RExC_close_parens) {
11415 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11416 "%*s%*s Setting close paren #0 (END) to %d\n",
11417 22, "| |", (int)(depth * 2 + 1), "", REG_NODE_NUM(ender)));
11419 RExC_close_parens[0]= ender;
11424 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11425 DEBUG_PARSE_MSG("lsbr");
11426 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
11427 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11428 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11429 SvPV_nolen_const(RExC_mysv1),
11430 (IV)REG_NODE_NUM(lastbr),
11431 SvPV_nolen_const(RExC_mysv2),
11432 (IV)REG_NODE_NUM(ender),
11433 (IV)(ender - lastbr)
11436 REGTAIL(pRExC_state, lastbr, ender);
11438 if (have_branch && !SIZE_ONLY) {
11439 char is_nothing= 1;
11441 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
11443 /* Hook the tails of the branches to the closing node. */
11444 for (br = ret; br; br = regnext(br)) {
11445 const U8 op = PL_regkind[OP(br)];
11446 if (op == BRANCH) {
11447 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
11448 if ( OP(NEXTOPER(br)) != NOTHING
11449 || regnext(NEXTOPER(br)) != ender)
11452 else if (op == BRANCHJ) {
11453 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
11454 /* for now we always disable this optimisation * /
11455 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
11456 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
11462 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
11463 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11464 DEBUG_PARSE_MSG("NADA");
11465 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
11466 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11467 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11468 SvPV_nolen_const(RExC_mysv1),
11469 (IV)REG_NODE_NUM(ret),
11470 SvPV_nolen_const(RExC_mysv2),
11471 (IV)REG_NODE_NUM(ender),
11476 if (OP(ender) == TAIL) {
11481 for ( opt= br + 1; opt < ender ; opt++ )
11482 OP(opt)= OPTIMIZED;
11483 NEXT_OFF(br)= ender - br;
11491 static const char parens[] = "=!<,>";
11493 if (paren && (p = strchr(parens, paren))) {
11494 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
11495 int flag = (p - parens) > 1;
11498 node = SUSPEND, flag = 0;
11499 reginsert(pRExC_state, node,ret, depth+1);
11500 Set_Node_Cur_Length(ret, parse_start);
11501 Set_Node_Offset(ret, parse_start + 1);
11503 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
11507 /* Check for proper termination. */
11509 /* restore original flags, but keep (?p) and, if we've changed from /d
11510 * rules to /u, keep the /u */
11511 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
11512 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
11513 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
11515 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
11516 RExC_parse = oregcomp_parse;
11517 vFAIL("Unmatched (");
11519 nextchar(pRExC_state);
11521 else if (!paren && RExC_parse < RExC_end) {
11522 if (*RExC_parse == ')') {
11524 vFAIL("Unmatched )");
11527 FAIL("Junk on end of regexp"); /* "Can't happen". */
11528 NOT_REACHED; /* NOTREACHED */
11531 if (RExC_in_lookbehind) {
11532 RExC_in_lookbehind--;
11534 if (after_freeze > RExC_npar)
11535 RExC_npar = after_freeze;
11540 - regbranch - one alternative of an | operator
11542 * Implements the concatenation operator.
11544 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11545 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11548 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
11551 regnode *chain = NULL;
11553 I32 flags = 0, c = 0;
11554 GET_RE_DEBUG_FLAGS_DECL;
11556 PERL_ARGS_ASSERT_REGBRANCH;
11558 DEBUG_PARSE("brnc");
11563 if (!SIZE_ONLY && RExC_extralen)
11564 ret = reganode(pRExC_state, BRANCHJ,0);
11566 ret = reg_node(pRExC_state, BRANCH);
11567 Set_Node_Length(ret, 1);
11571 if (!first && SIZE_ONLY)
11572 RExC_extralen += 1; /* BRANCHJ */
11574 *flagp = WORST; /* Tentatively. */
11576 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11577 FALSE /* Don't force to /x */ );
11578 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
11579 flags &= ~TRYAGAIN;
11580 latest = regpiece(pRExC_state, &flags,depth+1);
11581 if (latest == NULL) {
11582 if (flags & TRYAGAIN)
11584 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11585 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11588 FAIL2("panic: regpiece returned NULL, flags=%#" UVxf, (UV) flags);
11590 else if (ret == NULL)
11592 *flagp |= flags&(HASWIDTH|POSTPONED);
11593 if (chain == NULL) /* First piece. */
11594 *flagp |= flags&SPSTART;
11596 /* FIXME adding one for every branch after the first is probably
11597 * excessive now we have TRIE support. (hv) */
11599 REGTAIL(pRExC_state, chain, latest);
11604 if (chain == NULL) { /* Loop ran zero times. */
11605 chain = reg_node(pRExC_state, NOTHING);
11610 *flagp |= flags&SIMPLE;
11617 - regpiece - something followed by possible quantifier * + ? {n,m}
11619 * Note that the branching code sequences used for ? and the general cases
11620 * of * and + are somewhat optimized: they use the same NOTHING node as
11621 * both the endmarker for their branch list and the body of the last branch.
11622 * It might seem that this node could be dispensed with entirely, but the
11623 * endmarker role is not redundant.
11625 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
11627 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11628 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11631 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11637 const char * const origparse = RExC_parse;
11639 I32 max = REG_INFTY;
11640 #ifdef RE_TRACK_PATTERN_OFFSETS
11643 const char *maxpos = NULL;
11646 /* Save the original in case we change the emitted regop to a FAIL. */
11647 regnode * const orig_emit = RExC_emit;
11649 GET_RE_DEBUG_FLAGS_DECL;
11651 PERL_ARGS_ASSERT_REGPIECE;
11653 DEBUG_PARSE("piec");
11655 ret = regatom(pRExC_state, &flags,depth+1);
11657 if (flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8))
11658 *flagp |= flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8);
11660 FAIL2("panic: regatom returned NULL, flags=%#" UVxf, (UV) flags);
11666 if (op == '{' && regcurly(RExC_parse)) {
11668 #ifdef RE_TRACK_PATTERN_OFFSETS
11669 parse_start = RExC_parse; /* MJD */
11671 next = RExC_parse + 1;
11672 while (isDIGIT(*next) || *next == ',') {
11673 if (*next == ',') {
11681 if (*next == '}') { /* got one */
11682 const char* endptr;
11686 if (isDIGIT(*RExC_parse)) {
11687 if (!grok_atoUV(RExC_parse, &uv, &endptr))
11688 vFAIL("Invalid quantifier in {,}");
11689 if (uv >= REG_INFTY)
11690 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11695 if (*maxpos == ',')
11698 maxpos = RExC_parse;
11699 if (isDIGIT(*maxpos)) {
11700 if (!grok_atoUV(maxpos, &uv, &endptr))
11701 vFAIL("Invalid quantifier in {,}");
11702 if (uv >= REG_INFTY)
11703 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11706 max = REG_INFTY; /* meaning "infinity" */
11709 nextchar(pRExC_state);
11710 if (max < min) { /* If can't match, warn and optimize to fail
11714 /* We can't back off the size because we have to reserve
11715 * enough space for all the things we are about to throw
11716 * away, but we can shrink it by the amount we are about
11717 * to re-use here */
11718 RExC_size += PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
11721 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
11722 RExC_emit = orig_emit;
11724 ret = reganode(pRExC_state, OPFAIL, 0);
11727 else if (min == max && *RExC_parse == '?')
11730 ckWARN2reg(RExC_parse + 1,
11731 "Useless use of greediness modifier '%c'",
11737 if ((flags&SIMPLE)) {
11738 if (min == 0 && max == REG_INFTY) {
11739 reginsert(pRExC_state, STAR, ret, depth+1);
11742 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11745 if (min == 1 && max == REG_INFTY) {
11746 reginsert(pRExC_state, PLUS, ret, depth+1);
11749 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11752 MARK_NAUGHTY_EXP(2, 2);
11753 reginsert(pRExC_state, CURLY, ret, depth+1);
11754 Set_Node_Offset(ret, parse_start+1); /* MJD */
11755 Set_Node_Cur_Length(ret, parse_start);
11758 regnode * const w = reg_node(pRExC_state, WHILEM);
11761 REGTAIL(pRExC_state, ret, w);
11762 if (!SIZE_ONLY && RExC_extralen) {
11763 reginsert(pRExC_state, LONGJMP,ret, depth+1);
11764 reginsert(pRExC_state, NOTHING,ret, depth+1);
11765 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
11767 reginsert(pRExC_state, CURLYX,ret, depth+1);
11769 Set_Node_Offset(ret, parse_start+1);
11770 Set_Node_Length(ret,
11771 op == '{' ? (RExC_parse - parse_start) : 1);
11773 if (!SIZE_ONLY && RExC_extralen)
11774 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
11775 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
11777 RExC_whilem_seen++, RExC_extralen += 3;
11778 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
11785 *flagp |= HASWIDTH;
11787 ARG1_SET(ret, (U16)min);
11788 ARG2_SET(ret, (U16)max);
11790 if (max == REG_INFTY)
11791 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11797 if (!ISMULT1(op)) {
11802 #if 0 /* Now runtime fix should be reliable. */
11804 /* if this is reinstated, don't forget to put this back into perldiag:
11806 =item Regexp *+ operand could be empty at {#} in regex m/%s/
11808 (F) The part of the regexp subject to either the * or + quantifier
11809 could match an empty string. The {#} shows in the regular
11810 expression about where the problem was discovered.
11814 if (!(flags&HASWIDTH) && op != '?')
11815 vFAIL("Regexp *+ operand could be empty");
11818 #ifdef RE_TRACK_PATTERN_OFFSETS
11819 parse_start = RExC_parse;
11821 nextchar(pRExC_state);
11823 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
11829 else if (op == '+') {
11833 else if (op == '?') {
11838 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
11839 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
11840 ckWARN2reg(RExC_parse,
11841 "%" UTF8f " matches null string many times",
11842 UTF8fARG(UTF, (RExC_parse >= origparse
11843 ? RExC_parse - origparse
11846 (void)ReREFCNT_inc(RExC_rx_sv);
11849 if (*RExC_parse == '?') {
11850 nextchar(pRExC_state);
11851 reginsert(pRExC_state, MINMOD, ret, depth+1);
11852 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
11854 else if (*RExC_parse == '+') {
11856 nextchar(pRExC_state);
11857 ender = reg_node(pRExC_state, SUCCEED);
11858 REGTAIL(pRExC_state, ret, ender);
11859 reginsert(pRExC_state, SUSPEND, ret, depth+1);
11861 ender = reg_node(pRExC_state, TAIL);
11862 REGTAIL(pRExC_state, ret, ender);
11865 if (ISMULT2(RExC_parse)) {
11867 vFAIL("Nested quantifiers");
11874 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
11883 /* This routine teases apart the various meanings of \N and returns
11884 * accordingly. The input parameters constrain which meaning(s) is/are valid
11885 * in the current context.
11887 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
11889 * If <code_point_p> is not NULL, the context is expecting the result to be a
11890 * single code point. If this \N instance turns out to a single code point,
11891 * the function returns TRUE and sets *code_point_p to that code point.
11893 * If <node_p> is not NULL, the context is expecting the result to be one of
11894 * the things representable by a regnode. If this \N instance turns out to be
11895 * one such, the function generates the regnode, returns TRUE and sets *node_p
11896 * to point to that regnode.
11898 * If this instance of \N isn't legal in any context, this function will
11899 * generate a fatal error and not return.
11901 * On input, RExC_parse should point to the first char following the \N at the
11902 * time of the call. On successful return, RExC_parse will have been updated
11903 * to point to just after the sequence identified by this routine. Also
11904 * *flagp has been updated as needed.
11906 * When there is some problem with the current context and this \N instance,
11907 * the function returns FALSE, without advancing RExC_parse, nor setting
11908 * *node_p, nor *code_point_p, nor *flagp.
11910 * If <cp_count> is not NULL, the caller wants to know the length (in code
11911 * points) that this \N sequence matches. This is set even if the function
11912 * returns FALSE, as detailed below.
11914 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
11916 * Probably the most common case is for the \N to specify a single code point.
11917 * *cp_count will be set to 1, and *code_point_p will be set to that code
11920 * Another possibility is for the input to be an empty \N{}, which for
11921 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
11922 * will be set to a generated NOTHING node.
11924 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
11925 * set to 0. *node_p will be set to a generated REG_ANY node.
11927 * The fourth possibility is that \N resolves to a sequence of more than one
11928 * code points. *cp_count will be set to the number of code points in the
11929 * sequence. *node_p * will be set to a generated node returned by this
11930 * function calling S_reg().
11932 * The final possibility is that it is premature to be calling this function;
11933 * that pass1 needs to be restarted. This can happen when this changes from
11934 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
11935 * latter occurs only when the fourth possibility would otherwise be in
11936 * effect, and is because one of those code points requires the pattern to be
11937 * recompiled as UTF-8. The function returns FALSE, and sets the
11938 * RESTART_PASS1 and NEED_UTF8 flags in *flagp, as appropriate. When this
11939 * happens, the caller needs to desist from continuing parsing, and return
11940 * this information to its caller. This is not set for when there is only one
11941 * code point, as this can be called as part of an ANYOF node, and they can
11942 * store above-Latin1 code points without the pattern having to be in UTF-8.
11944 * For non-single-quoted regexes, the tokenizer has resolved character and
11945 * sequence names inside \N{...} into their Unicode values, normalizing the
11946 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
11947 * hex-represented code points in the sequence. This is done there because
11948 * the names can vary based on what charnames pragma is in scope at the time,
11949 * so we need a way to take a snapshot of what they resolve to at the time of
11950 * the original parse. [perl #56444].
11952 * That parsing is skipped for single-quoted regexes, so we may here get
11953 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
11954 * parser. But if the single-quoted regex is something like '\N{U+41}', that
11955 * is legal and handled here. The code point is Unicode, and has to be
11956 * translated into the native character set for non-ASCII platforms.
11959 char * endbrace; /* points to '}' following the name */
11960 char *endchar; /* Points to '.' or '}' ending cur char in the input
11962 char* p = RExC_parse; /* Temporary */
11964 GET_RE_DEBUG_FLAGS_DECL;
11966 PERL_ARGS_ASSERT_GROK_BSLASH_N;
11968 GET_RE_DEBUG_FLAGS;
11970 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
11971 assert(! (node_p && cp_count)); /* At most 1 should be set */
11973 if (cp_count) { /* Initialize return for the most common case */
11977 /* The [^\n] meaning of \N ignores spaces and comments under the /x
11978 * modifier. The other meanings do not, so use a temporary until we find
11979 * out which we are being called with */
11980 skip_to_be_ignored_text(pRExC_state, &p,
11981 FALSE /* Don't force to /x */ );
11983 /* Disambiguate between \N meaning a named character versus \N meaning
11984 * [^\n]. The latter is assumed when the {...} following the \N is a legal
11985 * quantifier, or there is no '{' at all */
11986 if (*p != '{' || regcurly(p)) {
11996 *node_p = reg_node(pRExC_state, REG_ANY);
11997 *flagp |= HASWIDTH|SIMPLE;
11999 Set_Node_Length(*node_p, 1); /* MJD */
12003 /* Here, we have decided it should be a named character or sequence */
12005 /* The test above made sure that the next real character is a '{', but
12006 * under the /x modifier, it could be separated by space (or a comment and
12007 * \n) and this is not allowed (for consistency with \x{...} and the
12008 * tokenizer handling of \N{NAME}). */
12009 if (*RExC_parse != '{') {
12010 vFAIL("Missing braces on \\N{}");
12013 RExC_parse++; /* Skip past the '{' */
12015 if (! (endbrace = strchr(RExC_parse, '}'))) { /* no trailing brace */
12016 vFAIL2("Missing right brace on \\%c{}", 'N');
12018 else if(!(endbrace == RExC_parse /* nothing between the {} */
12019 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked... */
12020 && strnEQ(RExC_parse, "U+", 2)))) /* ... below for a better
12023 RExC_parse = endbrace; /* position msg's '<--HERE' */
12024 vFAIL("\\N{NAME} must be resolved by the lexer");
12027 REQUIRE_UNI_RULES(flagp, FALSE); /* Unicode named chars imply Unicode
12030 if (endbrace == RExC_parse) { /* empty: \N{} */
12032 RExC_parse++; /* Position after the "}" */
12033 vFAIL("Zero length \\N{}");
12038 nextchar(pRExC_state);
12043 *node_p = reg_node(pRExC_state,NOTHING);
12047 RExC_parse += 2; /* Skip past the 'U+' */
12049 /* Because toke.c has generated a special construct for us guaranteed not
12050 * to have NULs, we can use a str function */
12051 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12053 /* Code points are separated by dots. If none, there is only one code
12054 * point, and is terminated by the brace */
12056 if (endchar >= endbrace) {
12057 STRLEN length_of_hex;
12058 I32 grok_hex_flags;
12060 /* Here, exactly one code point. If that isn't what is wanted, fail */
12061 if (! code_point_p) {
12066 /* Convert code point from hex */
12067 length_of_hex = (STRLEN)(endchar - RExC_parse);
12068 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
12069 | PERL_SCAN_DISALLOW_PREFIX
12071 /* No errors in the first pass (See [perl
12072 * #122671].) We let the code below find the
12073 * errors when there are multiple chars. */
12075 ? PERL_SCAN_SILENT_ILLDIGIT
12078 /* This routine is the one place where both single- and double-quotish
12079 * \N{U+xxxx} are evaluated. The value is a Unicode code point which
12080 * must be converted to native. */
12081 *code_point_p = UNI_TO_NATIVE(grok_hex(RExC_parse,
12086 /* The tokenizer should have guaranteed validity, but it's possible to
12087 * bypass it by using single quoting, so check. Don't do the check
12088 * here when there are multiple chars; we do it below anyway. */
12089 if (length_of_hex == 0
12090 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
12092 RExC_parse += length_of_hex; /* Includes all the valid */
12093 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12094 ? UTF8SKIP(RExC_parse)
12096 /* Guard against malformed utf8 */
12097 if (RExC_parse >= endchar) {
12098 RExC_parse = endchar;
12100 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12103 RExC_parse = endbrace + 1;
12106 else { /* Is a multiple character sequence */
12107 SV * substitute_parse;
12109 char *orig_end = RExC_end;
12110 char *save_start = RExC_start;
12113 /* Count the code points, if desired, in the sequence */
12116 while (RExC_parse < endbrace) {
12117 /* Point to the beginning of the next character in the sequence. */
12118 RExC_parse = endchar + 1;
12119 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12124 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12125 * But don't backup up the pointer if the caller want to know how many
12126 * code points there are (they can then handle things) */
12134 /* What is done here is to convert this to a sub-pattern of the form
12135 * \x{char1}\x{char2}... and then call reg recursively to parse it
12136 * (enclosing in "(?: ... )" ). That way, it retains its atomicness,
12137 * while not having to worry about special handling that some code
12138 * points may have. */
12140 substitute_parse = newSVpvs("?:");
12142 while (RExC_parse < endbrace) {
12144 /* Convert to notation the rest of the code understands */
12145 sv_catpv(substitute_parse, "\\x{");
12146 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
12147 sv_catpv(substitute_parse, "}");
12149 /* Point to the beginning of the next character in the sequence. */
12150 RExC_parse = endchar + 1;
12151 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12154 sv_catpv(substitute_parse, ")");
12156 RExC_parse = RExC_start = RExC_adjusted_start = SvPV(substitute_parse,
12159 /* Don't allow empty number */
12160 if (len < (STRLEN) 8) {
12161 RExC_parse = endbrace;
12162 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12164 RExC_end = RExC_parse + len;
12166 /* The values are Unicode, and therefore not subject to recoding, but
12167 * have to be converted to native on a non-Unicode (meaning non-ASCII)
12170 RExC_recode_x_to_native = 1;
12174 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
12175 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12176 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12179 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#" UVxf,
12182 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12185 /* Restore the saved values */
12186 RExC_start = RExC_adjusted_start = save_start;
12187 RExC_parse = endbrace;
12188 RExC_end = orig_end;
12190 RExC_recode_x_to_native = 0;
12193 SvREFCNT_dec_NN(substitute_parse);
12194 nextchar(pRExC_state);
12201 PERL_STATIC_INLINE U8
12202 S_compute_EXACTish(RExC_state_t *pRExC_state)
12206 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
12214 op = get_regex_charset(RExC_flags);
12215 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
12216 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
12217 been, so there is no hole */
12220 return op + EXACTF;
12223 PERL_STATIC_INLINE void
12224 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
12225 regnode *node, I32* flagp, STRLEN len, UV code_point,
12228 /* This knows the details about sizing an EXACTish node, setting flags for
12229 * it (by setting <*flagp>, and potentially populating it with a single
12232 * If <len> (the length in bytes) is non-zero, this function assumes that
12233 * the node has already been populated, and just does the sizing. In this
12234 * case <code_point> should be the final code point that has already been
12235 * placed into the node. This value will be ignored except that under some
12236 * circumstances <*flagp> is set based on it.
12238 * If <len> is zero, the function assumes that the node is to contain only
12239 * the single character given by <code_point> and calculates what <len>
12240 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
12241 * additionally will populate the node's STRING with <code_point> or its
12244 * In both cases <*flagp> is appropriately set
12246 * It knows that under FOLD, the Latin Sharp S and UTF characters above
12247 * 255, must be folded (the former only when the rules indicate it can
12250 * When it does the populating, it looks at the flag 'downgradable'. If
12251 * true with a node that folds, it checks if the single code point
12252 * participates in a fold, and if not downgrades the node to an EXACT.
12253 * This helps the optimizer */
12255 bool len_passed_in = cBOOL(len != 0);
12256 U8 character[UTF8_MAXBYTES_CASE+1];
12258 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
12260 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
12261 * sizing difference, and is extra work that is thrown away */
12262 if (downgradable && ! PASS2) {
12263 downgradable = FALSE;
12266 if (! len_passed_in) {
12268 if (UVCHR_IS_INVARIANT(code_point)) {
12269 if (LOC || ! FOLD) { /* /l defers folding until runtime */
12270 *character = (U8) code_point;
12272 else { /* Here is /i and not /l. (toFOLD() is defined on just
12273 ASCII, which isn't the same thing as INVARIANT on
12274 EBCDIC, but it works there, as the extra invariants
12275 fold to themselves) */
12276 *character = toFOLD((U8) code_point);
12278 /* We can downgrade to an EXACT node if this character
12279 * isn't a folding one. Note that this assumes that
12280 * nothing above Latin1 folds to some other invariant than
12281 * one of these alphabetics; otherwise we would also have
12283 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12284 * || ASCII_FOLD_RESTRICTED))
12286 if (downgradable && PL_fold[code_point] == code_point) {
12292 else if (FOLD && (! LOC
12293 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
12294 { /* Folding, and ok to do so now */
12295 UV folded = _to_uni_fold_flags(
12299 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12300 ? FOLD_FLAGS_NOMIX_ASCII
12303 && folded == code_point /* This quickly rules out many
12304 cases, avoiding the
12305 _invlist_contains_cp() overhead
12307 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
12314 else if (code_point <= MAX_UTF8_TWO_BYTE) {
12316 /* Not folding this cp, and can output it directly */
12317 *character = UTF8_TWO_BYTE_HI(code_point);
12318 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
12322 uvchr_to_utf8( character, code_point);
12323 len = UTF8SKIP(character);
12325 } /* Else pattern isn't UTF8. */
12327 *character = (U8) code_point;
12329 } /* Else is folded non-UTF8 */
12330 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12331 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12332 || UNICODE_DOT_DOT_VERSION > 0)
12333 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
12337 /* We don't fold any non-UTF8 except possibly the Sharp s (see
12338 * comments at join_exact()); */
12339 *character = (U8) code_point;
12342 /* Can turn into an EXACT node if we know the fold at compile time,
12343 * and it folds to itself and doesn't particpate in other folds */
12346 && PL_fold_latin1[code_point] == code_point
12347 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12348 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
12352 } /* else is Sharp s. May need to fold it */
12353 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
12355 *(character + 1) = 's';
12359 *character = LATIN_SMALL_LETTER_SHARP_S;
12365 RExC_size += STR_SZ(len);
12368 RExC_emit += STR_SZ(len);
12369 STR_LEN(node) = len;
12370 if (! len_passed_in) {
12371 Copy((char *) character, STRING(node), len, char);
12375 *flagp |= HASWIDTH;
12377 /* A single character node is SIMPLE, except for the special-cased SHARP S
12379 if ((len == 1 || (UTF && len == UVCHR_SKIP(code_point)))
12380 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12381 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12382 || UNICODE_DOT_DOT_VERSION > 0)
12383 && ( code_point != LATIN_SMALL_LETTER_SHARP_S
12384 || ! FOLD || ! DEPENDS_SEMANTICS)
12390 /* The OP may not be well defined in PASS1 */
12391 if (PASS2 && OP(node) == EXACTFL) {
12392 RExC_contains_locale = 1;
12397 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
12398 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
12401 S_backref_value(char *p)
12403 const char* endptr;
12405 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
12412 - regatom - the lowest level
12414 Try to identify anything special at the start of the current parse position.
12415 If there is, then handle it as required. This may involve generating a
12416 single regop, such as for an assertion; or it may involve recursing, such as
12417 to handle a () structure.
12419 If the string doesn't start with something special then we gobble up
12420 as much literal text as we can. If we encounter a quantifier, we have to
12421 back off the final literal character, as that quantifier applies to just it
12422 and not to the whole string of literals.
12424 Once we have been able to handle whatever type of thing started the
12425 sequence, we return.
12427 Note: we have to be careful with escapes, as they can be both literal
12428 and special, and in the case of \10 and friends, context determines which.
12430 A summary of the code structure is:
12432 switch (first_byte) {
12433 cases for each special:
12434 handle this special;
12437 switch (2nd byte) {
12438 cases for each unambiguous special:
12439 handle this special;
12441 cases for each ambigous special/literal:
12443 if (special) handle here
12445 default: // unambiguously literal:
12448 default: // is a literal char
12451 create EXACTish node for literal;
12452 while (more input and node isn't full) {
12453 switch (input_byte) {
12454 cases for each special;
12455 make sure parse pointer is set so that the next call to
12456 regatom will see this special first
12457 goto loopdone; // EXACTish node terminated by prev. char
12459 append char to EXACTISH node;
12461 get next input byte;
12465 return the generated node;
12467 Specifically there are two separate switches for handling
12468 escape sequences, with the one for handling literal escapes requiring
12469 a dummy entry for all of the special escapes that are actually handled
12472 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
12474 Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
12475 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
12476 Otherwise does not return NULL.
12480 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12482 regnode *ret = NULL;
12489 GET_RE_DEBUG_FLAGS_DECL;
12491 *flagp = WORST; /* Tentatively. */
12493 DEBUG_PARSE("atom");
12495 PERL_ARGS_ASSERT_REGATOM;
12498 parse_start = RExC_parse;
12499 assert(RExC_parse < RExC_end);
12500 switch ((U8)*RExC_parse) {
12502 RExC_seen_zerolen++;
12503 nextchar(pRExC_state);
12504 if (RExC_flags & RXf_PMf_MULTILINE)
12505 ret = reg_node(pRExC_state, MBOL);
12507 ret = reg_node(pRExC_state, SBOL);
12508 Set_Node_Length(ret, 1); /* MJD */
12511 nextchar(pRExC_state);
12513 RExC_seen_zerolen++;
12514 if (RExC_flags & RXf_PMf_MULTILINE)
12515 ret = reg_node(pRExC_state, MEOL);
12517 ret = reg_node(pRExC_state, SEOL);
12518 Set_Node_Length(ret, 1); /* MJD */
12521 nextchar(pRExC_state);
12522 if (RExC_flags & RXf_PMf_SINGLELINE)
12523 ret = reg_node(pRExC_state, SANY);
12525 ret = reg_node(pRExC_state, REG_ANY);
12526 *flagp |= HASWIDTH|SIMPLE;
12528 Set_Node_Length(ret, 1); /* MJD */
12532 char * const oregcomp_parse = ++RExC_parse;
12533 ret = regclass(pRExC_state, flagp,depth+1,
12534 FALSE, /* means parse the whole char class */
12535 TRUE, /* allow multi-char folds */
12536 FALSE, /* don't silence non-portable warnings. */
12537 (bool) RExC_strict,
12538 TRUE, /* Allow an optimized regnode result */
12542 if (*flagp & (RESTART_PASS1|NEED_UTF8))
12544 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12547 if (*RExC_parse != ']') {
12548 RExC_parse = oregcomp_parse;
12549 vFAIL("Unmatched [");
12551 nextchar(pRExC_state);
12552 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
12556 nextchar(pRExC_state);
12557 ret = reg(pRExC_state, 2, &flags,depth+1);
12559 if (flags & TRYAGAIN) {
12560 if (RExC_parse >= RExC_end) {
12561 /* Make parent create an empty node if needed. */
12562 *flagp |= TRYAGAIN;
12567 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12568 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12571 FAIL2("panic: reg returned NULL to regatom, flags=%#" UVxf,
12574 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12578 if (flags & TRYAGAIN) {
12579 *flagp |= TRYAGAIN;
12582 vFAIL("Internal urp");
12583 /* Supposed to be caught earlier. */
12589 vFAIL("Quantifier follows nothing");
12594 This switch handles escape sequences that resolve to some kind
12595 of special regop and not to literal text. Escape sequnces that
12596 resolve to literal text are handled below in the switch marked
12599 Every entry in this switch *must* have a corresponding entry
12600 in the literal escape switch. However, the opposite is not
12601 required, as the default for this switch is to jump to the
12602 literal text handling code.
12605 switch ((U8)*RExC_parse) {
12606 /* Special Escapes */
12608 RExC_seen_zerolen++;
12609 ret = reg_node(pRExC_state, SBOL);
12610 /* SBOL is shared with /^/ so we set the flags so we can tell
12611 * /\A/ from /^/ in split. We check ret because first pass we
12612 * have no regop struct to set the flags on. */
12616 goto finish_meta_pat;
12618 ret = reg_node(pRExC_state, GPOS);
12619 RExC_seen |= REG_GPOS_SEEN;
12621 goto finish_meta_pat;
12623 RExC_seen_zerolen++;
12624 ret = reg_node(pRExC_state, KEEPS);
12626 /* XXX:dmq : disabling in-place substitution seems to
12627 * be necessary here to avoid cases of memory corruption, as
12628 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
12630 RExC_seen |= REG_LOOKBEHIND_SEEN;
12631 goto finish_meta_pat;
12633 ret = reg_node(pRExC_state, SEOL);
12635 RExC_seen_zerolen++; /* Do not optimize RE away */
12636 goto finish_meta_pat;
12638 ret = reg_node(pRExC_state, EOS);
12640 RExC_seen_zerolen++; /* Do not optimize RE away */
12641 goto finish_meta_pat;
12643 vFAIL("\\C no longer supported");
12645 ret = reg_node(pRExC_state, CLUMP);
12646 *flagp |= HASWIDTH;
12647 goto finish_meta_pat;
12653 arg = ANYOF_WORDCHAR;
12661 regex_charset charset = get_regex_charset(RExC_flags);
12663 RExC_seen_zerolen++;
12664 RExC_seen |= REG_LOOKBEHIND_SEEN;
12665 op = BOUND + charset;
12667 if (op == BOUNDL) {
12668 RExC_contains_locale = 1;
12671 ret = reg_node(pRExC_state, op);
12673 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
12674 FLAGS(ret) = TRADITIONAL_BOUND;
12675 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
12681 char name = *RExC_parse;
12684 endbrace = strchr(RExC_parse, '}');
12687 vFAIL2("Missing right brace on \\%c{}", name);
12689 /* XXX Need to decide whether to take spaces or not. Should be
12690 * consistent with \p{}, but that currently is SPACE, which
12691 * means vertical too, which seems wrong
12692 * while (isBLANK(*RExC_parse)) {
12695 if (endbrace == RExC_parse) {
12696 RExC_parse++; /* After the '}' */
12697 vFAIL2("Empty \\%c{}", name);
12699 length = endbrace - RExC_parse;
12700 /*while (isBLANK(*(RExC_parse + length - 1))) {
12703 switch (*RExC_parse) {
12706 && (length != 3 || strnNE(RExC_parse + 1, "cb", 2)))
12708 goto bad_bound_type;
12710 FLAGS(ret) = GCB_BOUND;
12713 if (length != 2 || *(RExC_parse + 1) != 'b') {
12714 goto bad_bound_type;
12716 FLAGS(ret) = LB_BOUND;
12719 if (length != 2 || *(RExC_parse + 1) != 'b') {
12720 goto bad_bound_type;
12722 FLAGS(ret) = SB_BOUND;
12725 if (length != 2 || *(RExC_parse + 1) != 'b') {
12726 goto bad_bound_type;
12728 FLAGS(ret) = WB_BOUND;
12732 RExC_parse = endbrace;
12734 "'%" UTF8f "' is an unknown bound type",
12735 UTF8fARG(UTF, length, endbrace - length));
12736 NOT_REACHED; /*NOTREACHED*/
12738 RExC_parse = endbrace;
12739 REQUIRE_UNI_RULES(flagp, NULL);
12741 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
12745 /* Don't have to worry about UTF-8, in this message because
12746 * to get here the contents of the \b must be ASCII */
12747 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
12748 "Using /u for '%.*s' instead of /%s",
12750 endbrace - length + 1,
12751 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
12752 ? ASCII_RESTRICT_PAT_MODS
12753 : ASCII_MORE_RESTRICT_PAT_MODS);
12757 if (PASS2 && invert) {
12758 OP(ret) += NBOUND - BOUND;
12760 goto finish_meta_pat;
12768 if (! DEPENDS_SEMANTICS) {
12772 /* \d doesn't have any matches in the upper Latin1 range, hence /d
12773 * is equivalent to /u. Changing to /u saves some branches at
12776 goto join_posix_op_known;
12779 ret = reg_node(pRExC_state, LNBREAK);
12780 *flagp |= HASWIDTH|SIMPLE;
12781 goto finish_meta_pat;
12789 goto join_posix_op_known;
12795 arg = ANYOF_VERTWS;
12797 goto join_posix_op_known;
12807 op = POSIXD + get_regex_charset(RExC_flags);
12808 if (op > POSIXA) { /* /aa is same as /a */
12811 else if (op == POSIXL) {
12812 RExC_contains_locale = 1;
12815 join_posix_op_known:
12818 op += NPOSIXD - POSIXD;
12821 ret = reg_node(pRExC_state, op);
12823 FLAGS(ret) = namedclass_to_classnum(arg);
12826 *flagp |= HASWIDTH|SIMPLE;
12830 nextchar(pRExC_state);
12831 Set_Node_Length(ret, 2); /* MJD */
12837 ret = regclass(pRExC_state, flagp,depth+1,
12838 TRUE, /* means just parse this element */
12839 FALSE, /* don't allow multi-char folds */
12840 FALSE, /* don't silence non-portable warnings. It
12841 would be a bug if these returned
12843 (bool) RExC_strict,
12844 TRUE, /* Allow an optimized regnode result */
12847 if (*flagp & RESTART_PASS1)
12849 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
12850 * multi-char folds are allowed. */
12852 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12857 Set_Node_Offset(ret, parse_start);
12858 Set_Node_Cur_Length(ret, parse_start - 2);
12859 nextchar(pRExC_state);
12862 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
12863 * \N{...} evaluates to a sequence of more than one code points).
12864 * The function call below returns a regnode, which is our result.
12865 * The parameters cause it to fail if the \N{} evaluates to a
12866 * single code point; we handle those like any other literal. The
12867 * reason that the multicharacter case is handled here and not as
12868 * part of the EXACtish code is because of quantifiers. In
12869 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
12870 * this way makes that Just Happen. dmq.
12871 * join_exact() will join this up with adjacent EXACTish nodes
12872 * later on, if appropriate. */
12874 if (grok_bslash_N(pRExC_state,
12875 &ret, /* Want a regnode returned */
12876 NULL, /* Fail if evaluates to a single code
12878 NULL, /* Don't need a count of how many code
12887 if (*flagp & RESTART_PASS1)
12890 /* Here, evaluates to a single code point. Go get that */
12891 RExC_parse = parse_start;
12894 case 'k': /* Handle \k<NAME> and \k'NAME' */
12898 if ( RExC_parse >= RExC_end - 1
12899 || (( ch = RExC_parse[1]) != '<'
12904 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12905 vFAIL2("Sequence %.2s... not terminated",parse_start);
12908 ret = handle_named_backref(pRExC_state,
12920 case '1': case '2': case '3': case '4':
12921 case '5': case '6': case '7': case '8': case '9':
12926 if (*RExC_parse == 'g') {
12930 if (*RExC_parse == '{') {
12934 if (*RExC_parse == '-') {
12938 if (hasbrace && !isDIGIT(*RExC_parse)) {
12939 if (isrel) RExC_parse--;
12941 goto parse_named_seq;
12944 if (RExC_parse >= RExC_end) {
12945 goto unterminated_g;
12947 num = S_backref_value(RExC_parse);
12949 vFAIL("Reference to invalid group 0");
12950 else if (num == I32_MAX) {
12951 if (isDIGIT(*RExC_parse))
12952 vFAIL("Reference to nonexistent group");
12955 vFAIL("Unterminated \\g... pattern");
12959 num = RExC_npar - num;
12961 vFAIL("Reference to nonexistent or unclosed group");
12965 num = S_backref_value(RExC_parse);
12966 /* bare \NNN might be backref or octal - if it is larger
12967 * than or equal RExC_npar then it is assumed to be an
12968 * octal escape. Note RExC_npar is +1 from the actual
12969 * number of parens. */
12970 /* Note we do NOT check if num == I32_MAX here, as that is
12971 * handled by the RExC_npar check */
12974 /* any numeric escape < 10 is always a backref */
12976 /* any numeric escape < RExC_npar is a backref */
12977 && num >= RExC_npar
12978 /* cannot be an octal escape if it starts with 8 */
12979 && *RExC_parse != '8'
12980 /* cannot be an octal escape it it starts with 9 */
12981 && *RExC_parse != '9'
12984 /* Probably not a backref, instead likely to be an
12985 * octal character escape, e.g. \35 or \777.
12986 * The above logic should make it obvious why using
12987 * octal escapes in patterns is problematic. - Yves */
12988 RExC_parse = parse_start;
12993 /* At this point RExC_parse points at a numeric escape like
12994 * \12 or \88 or something similar, which we should NOT treat
12995 * as an octal escape. It may or may not be a valid backref
12996 * escape. For instance \88888888 is unlikely to be a valid
12998 while (isDIGIT(*RExC_parse))
13001 if (*RExC_parse != '}')
13002 vFAIL("Unterminated \\g{...} pattern");
13006 if (num > (I32)RExC_rx->nparens)
13007 vFAIL("Reference to nonexistent group");
13010 ret = reganode(pRExC_state,
13013 : (ASCII_FOLD_RESTRICTED)
13015 : (AT_LEAST_UNI_SEMANTICS)
13021 *flagp |= HASWIDTH;
13023 /* override incorrect value set in reganode MJD */
13024 Set_Node_Offset(ret, parse_start);
13025 Set_Node_Cur_Length(ret, parse_start-1);
13026 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13027 FALSE /* Don't force to /x */ );
13031 if (RExC_parse >= RExC_end)
13032 FAIL("Trailing \\");
13035 /* Do not generate "unrecognized" warnings here, we fall
13036 back into the quick-grab loop below */
13037 RExC_parse = parse_start;
13039 } /* end of switch on a \foo sequence */
13044 /* '#' comments should have been spaced over before this function was
13046 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13048 if (RExC_flags & RXf_PMf_EXTENDED) {
13049 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13050 if (RExC_parse < RExC_end)
13060 /* Here, we have determined that the next thing is probably a
13061 * literal character. RExC_parse points to the first byte of its
13062 * definition. (It still may be an escape sequence that evaluates
13063 * to a single character) */
13069 #define MAX_NODE_STRING_SIZE 127
13070 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
13072 U8 upper_parse = MAX_NODE_STRING_SIZE;
13073 U8 node_type = compute_EXACTish(pRExC_state);
13074 bool next_is_quantifier;
13075 char * oldp = NULL;
13077 /* We can convert EXACTF nodes to EXACTFU if they contain only
13078 * characters that match identically regardless of the target
13079 * string's UTF8ness. The reason to do this is that EXACTF is not
13080 * trie-able, EXACTFU is.
13082 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13083 * contain only above-Latin1 characters (hence must be in UTF8),
13084 * which don't participate in folds with Latin1-range characters,
13085 * as the latter's folds aren't known until runtime. (We don't
13086 * need to figure this out until pass 2) */
13087 bool maybe_exactfu = PASS2
13088 && (node_type == EXACTF || node_type == EXACTFL);
13090 /* If a folding node contains only code points that don't
13091 * participate in folds, it can be changed into an EXACT node,
13092 * which allows the optimizer more things to look for */
13095 ret = reg_node(pRExC_state, node_type);
13097 /* In pass1, folded, we use a temporary buffer instead of the
13098 * actual node, as the node doesn't exist yet */
13099 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
13105 /* We look for the EXACTFish to EXACT node optimizaton only if
13106 * folding. (And we don't need to figure this out until pass 2).
13107 * XXX It might actually make sense to split the node into portions
13108 * that are exact and ones that aren't, so that we could later use
13109 * the exact ones to find the longest fixed and floating strings.
13110 * One would want to join them back into a larger node. One could
13111 * use a pseudo regnode like 'EXACT_ORIG_FOLD' */
13112 maybe_exact = FOLD && PASS2;
13114 /* XXX The node can hold up to 255 bytes, yet this only goes to
13115 * 127. I (khw) do not know why. Keeping it somewhat less than
13116 * 255 allows us to not have to worry about overflow due to
13117 * converting to utf8 and fold expansion, but that value is
13118 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
13119 * split up by this limit into a single one using the real max of
13120 * 255. Even at 127, this breaks under rare circumstances. If
13121 * folding, we do not want to split a node at a character that is a
13122 * non-final in a multi-char fold, as an input string could just
13123 * happen to want to match across the node boundary. The join
13124 * would solve that problem if the join actually happens. But a
13125 * series of more than two nodes in a row each of 127 would cause
13126 * the first join to succeed to get to 254, but then there wouldn't
13127 * be room for the next one, which could at be one of those split
13128 * multi-char folds. I don't know of any fool-proof solution. One
13129 * could back off to end with only a code point that isn't such a
13130 * non-final, but it is possible for there not to be any in the
13133 assert( ! UTF /* Is at the beginning of a character */
13134 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13135 || UTF8_IS_START(UCHARAT(RExC_parse)));
13137 /* Here, we have a literal character. Find the maximal string of
13138 * them in the input that we can fit into a single EXACTish node.
13139 * We quit at the first non-literal or when the node gets full */
13140 for (p = RExC_parse;
13141 len < upper_parse && p < RExC_end;
13146 /* White space has already been ignored */
13147 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13148 || ! is_PATWS_safe((p), RExC_end, UTF));
13160 /* Literal Escapes Switch
13162 This switch is meant to handle escape sequences that
13163 resolve to a literal character.
13165 Every escape sequence that represents something
13166 else, like an assertion or a char class, is handled
13167 in the switch marked 'Special Escapes' above in this
13168 routine, but also has an entry here as anything that
13169 isn't explicitly mentioned here will be treated as
13170 an unescaped equivalent literal.
13173 switch ((U8)*++p) {
13174 /* These are all the special escapes. */
13175 case 'A': /* Start assertion */
13176 case 'b': case 'B': /* Word-boundary assertion*/
13177 case 'C': /* Single char !DANGEROUS! */
13178 case 'd': case 'D': /* digit class */
13179 case 'g': case 'G': /* generic-backref, pos assertion */
13180 case 'h': case 'H': /* HORIZWS */
13181 case 'k': case 'K': /* named backref, keep marker */
13182 case 'p': case 'P': /* Unicode property */
13183 case 'R': /* LNBREAK */
13184 case 's': case 'S': /* space class */
13185 case 'v': case 'V': /* VERTWS */
13186 case 'w': case 'W': /* word class */
13187 case 'X': /* eXtended Unicode "combining
13188 character sequence" */
13189 case 'z': case 'Z': /* End of line/string assertion */
13193 /* Anything after here is an escape that resolves to a
13194 literal. (Except digits, which may or may not)
13200 case 'N': /* Handle a single-code point named character. */
13201 RExC_parse = p + 1;
13202 if (! grok_bslash_N(pRExC_state,
13203 NULL, /* Fail if evaluates to
13204 anything other than a
13205 single code point */
13206 &ender, /* The returned single code
13208 NULL, /* Don't need a count of
13209 how many code points */
13214 if (*flagp & NEED_UTF8)
13215 FAIL("panic: grok_bslash_N set NEED_UTF8");
13216 if (*flagp & RESTART_PASS1)
13219 /* Here, it wasn't a single code point. Go close
13220 * up this EXACTish node. The switch() prior to
13221 * this switch handles the other cases */
13222 RExC_parse = p = oldp;
13226 if (ender > 0xff) {
13227 REQUIRE_UTF8(flagp);
13243 ender = ESC_NATIVE;
13253 const char* error_msg;
13255 bool valid = grok_bslash_o(&p,
13258 PASS2, /* out warnings */
13259 (bool) RExC_strict,
13260 TRUE, /* Output warnings
13265 RExC_parse = p; /* going to die anyway; point
13266 to exact spot of failure */
13270 if (ender > 0xff) {
13271 REQUIRE_UTF8(flagp);
13277 UV result = UV_MAX; /* initialize to erroneous
13279 const char* error_msg;
13281 bool valid = grok_bslash_x(&p,
13284 PASS2, /* out warnings */
13285 (bool) RExC_strict,
13286 TRUE, /* Silence warnings
13291 RExC_parse = p; /* going to die anyway; point
13292 to exact spot of failure */
13297 if (ender < 0x100) {
13299 if (RExC_recode_x_to_native) {
13300 ender = LATIN1_TO_NATIVE(ender);
13305 REQUIRE_UTF8(flagp);
13311 ender = grok_bslash_c(*p++, PASS2);
13313 case '8': case '9': /* must be a backreference */
13315 /* we have an escape like \8 which cannot be an octal escape
13316 * so we exit the loop, and let the outer loop handle this
13317 * escape which may or may not be a legitimate backref. */
13319 case '1': case '2': case '3':case '4':
13320 case '5': case '6': case '7':
13321 /* When we parse backslash escapes there is ambiguity
13322 * between backreferences and octal escapes. Any escape
13323 * from \1 - \9 is a backreference, any multi-digit
13324 * escape which does not start with 0 and which when
13325 * evaluated as decimal could refer to an already
13326 * parsed capture buffer is a back reference. Anything
13329 * Note this implies that \118 could be interpreted as
13330 * 118 OR as "\11" . "8" depending on whether there
13331 * were 118 capture buffers defined already in the
13334 /* NOTE, RExC_npar is 1 more than the actual number of
13335 * parens we have seen so far, hence the < RExC_npar below. */
13337 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
13338 { /* Not to be treated as an octal constant, go
13346 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13348 ender = grok_oct(p, &numlen, &flags, NULL);
13349 if (ender > 0xff) {
13350 REQUIRE_UTF8(flagp);
13353 if (PASS2 /* like \08, \178 */
13355 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
13357 reg_warn_non_literal_string(
13359 form_short_octal_warning(p, numlen));
13365 FAIL("Trailing \\");
13368 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
13369 /* Include any left brace following the alpha to emphasize
13370 * that it could be part of an escape at some point
13372 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
13373 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
13375 goto normal_default;
13376 } /* End of switch on '\' */
13379 /* Currently we don't care if the lbrace is at the start
13380 * of a construct. This catches it in the middle of a
13381 * literal string, or when it's the first thing after
13382 * something like "\b" */
13383 if (len || (p > RExC_start && isALPHA_A(*(p -1)))) {
13384 RExC_parse = p + 1;
13385 vFAIL("Unescaped left brace in regex is illegal here");
13387 goto normal_default;
13390 if (PASS2 && p > RExC_parse && RExC_strict) {
13391 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
13394 default: /* A literal character */
13396 if (! UTF8_IS_INVARIANT(*p) && UTF) {
13398 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
13399 &numlen, UTF8_ALLOW_DEFAULT);
13405 } /* End of switch on the literal */
13407 /* Here, have looked at the literal character and <ender>
13408 * contains its ordinal, <p> points to the character after it.
13409 * We need to check if the next non-ignored thing is a
13410 * quantifier. Move <p> to after anything that should be
13411 * ignored, which, as a side effect, positions <p> for the next
13412 * loop iteration */
13413 skip_to_be_ignored_text(pRExC_state, &p,
13414 FALSE /* Don't force to /x */ );
13416 /* If the next thing is a quantifier, it applies to this
13417 * character only, which means that this character has to be in
13418 * its own node and can't just be appended to the string in an
13419 * existing node, so if there are already other characters in
13420 * the node, close the node with just them, and set up to do
13421 * this character again next time through, when it will be the
13422 * only thing in its new node */
13424 if ((next_is_quantifier = ( LIKELY(p < RExC_end)
13425 && UNLIKELY(ISMULT2(p))))
13432 /* Ready to add 'ender' to the node */
13434 if (! FOLD) { /* The simple case, just append the literal */
13436 /* In the sizing pass, we need only the size of the
13437 * character we are appending, hence we can delay getting
13438 * its representation until PASS2. */
13441 const STRLEN unilen = UVCHR_SKIP(ender);
13444 /* We have to subtract 1 just below (and again in
13445 * the corresponding PASS2 code) because the loop
13446 * increments <len> each time, as all but this path
13447 * (and one other) through it add a single byte to
13448 * the EXACTish node. But these paths would change
13449 * len to be the correct final value, so cancel out
13450 * the increment that follows */
13456 } else { /* PASS2 */
13459 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
13460 len += (char *) new_s - s - 1;
13461 s = (char *) new_s;
13464 *(s++) = (char) ender;
13468 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
13470 /* Here are folding under /l, and the code point is
13471 * problematic. First, we know we can't simplify things */
13472 maybe_exact = FALSE;
13473 maybe_exactfu = FALSE;
13475 /* A problematic code point in this context means that its
13476 * fold isn't known until runtime, so we can't fold it now.
13477 * (The non-problematic code points are the above-Latin1
13478 * ones that fold to also all above-Latin1. Their folds
13479 * don't vary no matter what the locale is.) But here we
13480 * have characters whose fold depends on the locale.
13481 * Unlike the non-folding case above, we have to keep track
13482 * of these in the sizing pass, so that we can make sure we
13483 * don't split too-long nodes in the middle of a potential
13484 * multi-char fold. And unlike the regular fold case
13485 * handled in the else clauses below, we don't actually
13486 * fold and don't have special cases to consider. What we
13487 * do for both passes is the PASS2 code for non-folding */
13488 goto not_fold_common;
13490 else /* A regular FOLD code point */
13492 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13493 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13494 || UNICODE_DOT_DOT_VERSION > 0)
13495 /* See comments for join_exact() as to why we fold
13496 * this non-UTF at compile time */
13497 || ( node_type == EXACTFU
13498 && ender == LATIN_SMALL_LETTER_SHARP_S)
13501 /* Here, are folding and are not UTF-8 encoded; therefore
13502 * the character must be in the range 0-255, and is not /l
13503 * (Not /l because we already handled these under /l in
13504 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
13505 if (IS_IN_SOME_FOLD_L1(ender)) {
13506 maybe_exact = FALSE;
13508 /* See if the character's fold differs between /d and
13509 * /u. This includes the multi-char fold SHARP S to
13511 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
13512 RExC_seen_unfolded_sharp_s = 1;
13513 maybe_exactfu = FALSE;
13515 else if (maybe_exactfu
13516 && (PL_fold[ender] != PL_fold_latin1[ender]
13517 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13518 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13519 || UNICODE_DOT_DOT_VERSION > 0)
13521 && isALPHA_FOLD_EQ(ender, 's')
13522 && isALPHA_FOLD_EQ(*(s-1), 's'))
13525 maybe_exactfu = FALSE;
13529 /* Even when folding, we store just the input character, as
13530 * we have an array that finds its fold quickly */
13531 *(s++) = (char) ender;
13533 else { /* FOLD, and UTF (or sharp s) */
13534 /* Unlike the non-fold case, we do actually have to
13535 * calculate the results here in pass 1. This is for two
13536 * reasons, the folded length may be longer than the
13537 * unfolded, and we have to calculate how many EXACTish
13538 * nodes it will take; and we may run out of room in a node
13539 * in the middle of a potential multi-char fold, and have
13540 * to back off accordingly. */
13543 if (isASCII_uni(ender)) {
13544 folded = toFOLD(ender);
13545 *(s)++ = (U8) folded;
13550 folded = _to_uni_fold_flags(
13554 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
13555 ? FOLD_FLAGS_NOMIX_ASCII
13559 /* The loop increments <len> each time, as all but this
13560 * path (and one other) through it add a single byte to
13561 * the EXACTish node. But this one has changed len to
13562 * be the correct final value, so subtract one to
13563 * cancel out the increment that follows */
13564 len += foldlen - 1;
13566 /* If this node only contains non-folding code points so
13567 * far, see if this new one is also non-folding */
13569 if (folded != ender) {
13570 maybe_exact = FALSE;
13573 /* Here the fold is the original; we have to check
13574 * further to see if anything folds to it */
13575 if (_invlist_contains_cp(PL_utf8_foldable,
13578 maybe_exact = FALSE;
13585 if (next_is_quantifier) {
13587 /* Here, the next input is a quantifier, and to get here,
13588 * the current character is the only one in the node.
13589 * Also, here <len> doesn't include the final byte for this
13595 } /* End of loop through literal characters */
13597 /* Here we have either exhausted the input or ran out of room in
13598 * the node. (If we encountered a character that can't be in the
13599 * node, transfer is made directly to <loopdone>, and so we
13600 * wouldn't have fallen off the end of the loop.) In the latter
13601 * case, we artificially have to split the node into two, because
13602 * we just don't have enough space to hold everything. This
13603 * creates a problem if the final character participates in a
13604 * multi-character fold in the non-final position, as a match that
13605 * should have occurred won't, due to the way nodes are matched,
13606 * and our artificial boundary. So back off until we find a non-
13607 * problematic character -- one that isn't at the beginning or
13608 * middle of such a fold. (Either it doesn't participate in any
13609 * folds, or appears only in the final position of all the folds it
13610 * does participate in.) A better solution with far fewer false
13611 * positives, and that would fill the nodes more completely, would
13612 * be to actually have available all the multi-character folds to
13613 * test against, and to back-off only far enough to be sure that
13614 * this node isn't ending with a partial one. <upper_parse> is set
13615 * further below (if we need to reparse the node) to include just
13616 * up through that final non-problematic character that this code
13617 * identifies, so when it is set to less than the full node, we can
13618 * skip the rest of this */
13619 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
13621 const STRLEN full_len = len;
13623 assert(len >= MAX_NODE_STRING_SIZE);
13625 /* Here, <s> points to the final byte of the final character.
13626 * Look backwards through the string until find a non-
13627 * problematic character */
13631 /* This has no multi-char folds to non-UTF characters */
13632 if (ASCII_FOLD_RESTRICTED) {
13636 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
13640 if (! PL_NonL1NonFinalFold) {
13641 PL_NonL1NonFinalFold = _new_invlist_C_array(
13642 NonL1_Perl_Non_Final_Folds_invlist);
13645 /* Point to the first byte of the final character */
13646 s = (char *) utf8_hop((U8 *) s, -1);
13648 while (s >= s0) { /* Search backwards until find
13649 non-problematic char */
13650 if (UTF8_IS_INVARIANT(*s)) {
13652 /* There are no ascii characters that participate
13653 * in multi-char folds under /aa. In EBCDIC, the
13654 * non-ascii invariants are all control characters,
13655 * so don't ever participate in any folds. */
13656 if (ASCII_FOLD_RESTRICTED
13657 || ! IS_NON_FINAL_FOLD(*s))
13662 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
13663 if (! IS_NON_FINAL_FOLD(EIGHT_BIT_UTF8_TO_NATIVE(
13669 else if (! _invlist_contains_cp(
13670 PL_NonL1NonFinalFold,
13671 valid_utf8_to_uvchr((U8 *) s, NULL)))
13676 /* Here, the current character is problematic in that
13677 * it does occur in the non-final position of some
13678 * fold, so try the character before it, but have to
13679 * special case the very first byte in the string, so
13680 * we don't read outside the string */
13681 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
13682 } /* End of loop backwards through the string */
13684 /* If there were only problematic characters in the string,
13685 * <s> will point to before s0, in which case the length
13686 * should be 0, otherwise include the length of the
13687 * non-problematic character just found */
13688 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
13691 /* Here, have found the final character, if any, that is
13692 * non-problematic as far as ending the node without splitting
13693 * it across a potential multi-char fold. <len> contains the
13694 * number of bytes in the node up-to and including that
13695 * character, or is 0 if there is no such character, meaning
13696 * the whole node contains only problematic characters. In
13697 * this case, give up and just take the node as-is. We can't
13702 /* If the node ends in an 's' we make sure it stays EXACTF,
13703 * as if it turns into an EXACTFU, it could later get
13704 * joined with another 's' that would then wrongly match
13706 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
13708 maybe_exactfu = FALSE;
13712 /* Here, the node does contain some characters that aren't
13713 * problematic. If one such is the final character in the
13714 * node, we are done */
13715 if (len == full_len) {
13718 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
13720 /* If the final character is problematic, but the
13721 * penultimate is not, back-off that last character to
13722 * later start a new node with it */
13727 /* Here, the final non-problematic character is earlier
13728 * in the input than the penultimate character. What we do
13729 * is reparse from the beginning, going up only as far as
13730 * this final ok one, thus guaranteeing that the node ends
13731 * in an acceptable character. The reason we reparse is
13732 * that we know how far in the character is, but we don't
13733 * know how to correlate its position with the input parse.
13734 * An alternate implementation would be to build that
13735 * correlation as we go along during the original parse,
13736 * but that would entail extra work for every node, whereas
13737 * this code gets executed only when the string is too
13738 * large for the node, and the final two characters are
13739 * problematic, an infrequent occurrence. Yet another
13740 * possible strategy would be to save the tail of the
13741 * string, and the next time regatom is called, initialize
13742 * with that. The problem with this is that unless you
13743 * back off one more character, you won't be guaranteed
13744 * regatom will get called again, unless regbranch,
13745 * regpiece ... are also changed. If you do back off that
13746 * extra character, so that there is input guaranteed to
13747 * force calling regatom, you can't handle the case where
13748 * just the first character in the node is acceptable. I
13749 * (khw) decided to try this method which doesn't have that
13750 * pitfall; if performance issues are found, we can do a
13751 * combination of the current approach plus that one */
13757 } /* End of verifying node ends with an appropriate char */
13759 loopdone: /* Jumped to when encounters something that shouldn't be
13762 /* I (khw) don't know if you can get here with zero length, but the
13763 * old code handled this situation by creating a zero-length EXACT
13764 * node. Might as well be NOTHING instead */
13770 /* If 'maybe_exact' is still set here, means there are no
13771 * code points in the node that participate in folds;
13772 * similarly for 'maybe_exactfu' and code points that match
13773 * differently depending on UTF8ness of the target string
13774 * (for /u), or depending on locale for /l */
13780 else if (maybe_exactfu) {
13786 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
13787 FALSE /* Don't look to see if could
13788 be turned into an EXACT
13789 node, as we have already
13794 RExC_parse = p - 1;
13795 Set_Node_Cur_Length(ret, parse_start);
13798 /* len is STRLEN which is unsigned, need to copy to signed */
13801 vFAIL("Internal disaster");
13804 } /* End of label 'defchar:' */
13806 } /* End of giant switch on input character */
13808 /* Position parse to next real character */
13809 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13810 FALSE /* Don't force to /x */ );
13811 if (PASS2 && *RExC_parse == '{' && OP(ret) != SBOL && ! regcurly(RExC_parse)) {
13812 ckWARNregdep(RExC_parse + 1, "Unescaped left brace in regex is deprecated here (and will be fatal in Perl 5.30), passed through");
13820 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
13822 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
13823 * sets up the bitmap and any flags, removing those code points from the
13824 * inversion list, setting it to NULL should it become completely empty */
13826 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
13827 assert(PL_regkind[OP(node)] == ANYOF);
13829 ANYOF_BITMAP_ZERO(node);
13830 if (*invlist_ptr) {
13832 /* This gets set if we actually need to modify things */
13833 bool change_invlist = FALSE;
13837 /* Start looking through *invlist_ptr */
13838 invlist_iterinit(*invlist_ptr);
13839 while (invlist_iternext(*invlist_ptr, &start, &end)) {
13843 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
13844 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
13847 /* Quit if are above what we should change */
13848 if (start >= NUM_ANYOF_CODE_POINTS) {
13852 change_invlist = TRUE;
13854 /* Set all the bits in the range, up to the max that we are doing */
13855 high = (end < NUM_ANYOF_CODE_POINTS - 1)
13857 : NUM_ANYOF_CODE_POINTS - 1;
13858 for (i = start; i <= (int) high; i++) {
13859 if (! ANYOF_BITMAP_TEST(node, i)) {
13860 ANYOF_BITMAP_SET(node, i);
13864 invlist_iterfinish(*invlist_ptr);
13866 /* Done with loop; remove any code points that are in the bitmap from
13867 * *invlist_ptr; similarly for code points above the bitmap if we have
13868 * a flag to match all of them anyways */
13869 if (change_invlist) {
13870 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
13872 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
13873 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
13876 /* If have completely emptied it, remove it completely */
13877 if (_invlist_len(*invlist_ptr) == 0) {
13878 SvREFCNT_dec_NN(*invlist_ptr);
13879 *invlist_ptr = NULL;
13884 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
13885 Character classes ([:foo:]) can also be negated ([:^foo:]).
13886 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
13887 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
13888 but trigger failures because they are currently unimplemented. */
13890 #define POSIXCC_DONE(c) ((c) == ':')
13891 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
13892 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
13893 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
13895 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
13896 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
13897 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
13899 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
13901 /* 'posix_warnings' and 'warn_text' are names of variables in the following
13903 #define ADD_POSIX_WARNING(p, text) STMT_START { \
13904 if (posix_warnings) { \
13905 if (! RExC_warn_text ) RExC_warn_text = (AV *) sv_2mortal((SV *) newAV()); \
13906 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
13910 REPORT_LOCATION_ARGS(p))); \
13915 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
13917 const char * const s, /* Where the putative posix class begins.
13918 Normally, this is one past the '['. This
13919 parameter exists so it can be somewhere
13920 besides RExC_parse. */
13921 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
13923 AV ** posix_warnings, /* Where to place any generated warnings, or
13925 const bool check_only /* Don't die if error */
13928 /* This parses what the caller thinks may be one of the three POSIX
13930 * 1) a character class, like [:blank:]
13931 * 2) a collating symbol, like [. .]
13932 * 3) an equivalence class, like [= =]
13933 * In the latter two cases, it croaks if it finds a syntactically legal
13934 * one, as these are not handled by Perl.
13936 * The main purpose is to look for a POSIX character class. It returns:
13937 * a) the class number
13938 * if it is a completely syntactically and semantically legal class.
13939 * 'updated_parse_ptr', if not NULL, is set to point to just after the
13940 * closing ']' of the class
13941 * b) OOB_NAMEDCLASS
13942 * if it appears that one of the three POSIX constructs was meant, but
13943 * its specification was somehow defective. 'updated_parse_ptr', if
13944 * not NULL, is set to point to the character just after the end
13945 * character of the class. See below for handling of warnings.
13946 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
13947 * if it doesn't appear that a POSIX construct was intended.
13948 * 'updated_parse_ptr' is not changed. No warnings nor errors are
13951 * In b) there may be errors or warnings generated. If 'check_only' is
13952 * TRUE, then any errors are discarded. Warnings are returned to the
13953 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
13954 * instead it is NULL, warnings are suppressed. This is done in all
13955 * passes. The reason for this is that the rest of the parsing is heavily
13956 * dependent on whether this routine found a valid posix class or not. If
13957 * it did, the closing ']' is absorbed as part of the class. If no class,
13958 * or an invalid one is found, any ']' will be considered the terminator of
13959 * the outer bracketed character class, leading to very different results.
13960 * In particular, a '(?[ ])' construct will likely have a syntax error if
13961 * the class is parsed other than intended, and this will happen in pass1,
13962 * before the warnings would normally be output. This mechanism allows the
13963 * caller to output those warnings in pass1 just before dieing, giving a
13964 * much better clue as to what is wrong.
13966 * The reason for this function, and its complexity is that a bracketed
13967 * character class can contain just about anything. But it's easy to
13968 * mistype the very specific posix class syntax but yielding a valid
13969 * regular bracketed class, so it silently gets compiled into something
13970 * quite unintended.
13972 * The solution adopted here maintains backward compatibility except that
13973 * it adds a warning if it looks like a posix class was intended but
13974 * improperly specified. The warning is not raised unless what is input
13975 * very closely resembles one of the 14 legal posix classes. To do this,
13976 * it uses fuzzy parsing. It calculates how many single-character edits it
13977 * would take to transform what was input into a legal posix class. Only
13978 * if that number is quite small does it think that the intention was a
13979 * posix class. Obviously these are heuristics, and there will be cases
13980 * where it errs on one side or another, and they can be tweaked as
13981 * experience informs.
13983 * The syntax for a legal posix class is:
13985 * qr/(?xa: \[ : \^? [:lower:]{4,6} : \] )/
13987 * What this routine considers syntactically to be an intended posix class
13988 * is this (the comments indicate some restrictions that the pattern
13991 * qr/(?x: \[? # The left bracket, possibly
13993 * \h* # possibly followed by blanks
13994 * (?: \^ \h* )? # possibly a misplaced caret
13995 * [:;]? # The opening class character,
13996 * # possibly omitted. A typo
13997 * # semi-colon can also be used.
13999 * \^? # possibly a correctly placed
14000 * # caret, but not if there was also
14001 * # a misplaced one
14003 * .{3,15} # The class name. If there are
14004 * # deviations from the legal syntax,
14005 * # its edit distance must be close
14006 * # to a real class name in order
14007 * # for it to be considered to be
14008 * # an intended posix class.
14010 * [:punct:]? # The closing class character,
14011 * # possibly omitted. If not a colon
14012 * # nor semi colon, the class name
14013 * # must be even closer to a valid
14016 * \]? # The right bracket, possibly
14020 * In the above, \h must be ASCII-only.
14022 * These are heuristics, and can be tweaked as field experience dictates.
14023 * There will be cases when someone didn't intend to specify a posix class
14024 * that this warns as being so. The goal is to minimize these, while
14025 * maximizing the catching of things intended to be a posix class that
14026 * aren't parsed as such.
14030 const char * const e = RExC_end;
14031 unsigned complement = 0; /* If to complement the class */
14032 bool found_problem = FALSE; /* Assume OK until proven otherwise */
14033 bool has_opening_bracket = FALSE;
14034 bool has_opening_colon = FALSE;
14035 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
14037 const char * possible_end = NULL; /* used for a 2nd parse pass */
14038 const char* name_start; /* ptr to class name first char */
14040 /* If the number of single-character typos the input name is away from a
14041 * legal name is no more than this number, it is considered to have meant
14042 * the legal name */
14043 int max_distance = 2;
14045 /* to store the name. The size determines the maximum length before we
14046 * decide that no posix class was intended. Should be at least
14047 * sizeof("alphanumeric") */
14050 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
14052 if (posix_warnings && RExC_warn_text)
14053 av_clear(RExC_warn_text);
14056 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14059 if (*(p - 1) != '[') {
14060 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
14061 found_problem = TRUE;
14064 has_opening_bracket = TRUE;
14067 /* They could be confused and think you can put spaces between the
14070 found_problem = TRUE;
14074 } while (p < e && isBLANK(*p));
14076 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14079 /* For [. .] and [= =]. These are quite different internally from [: :],
14080 * so they are handled separately. */
14081 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
14082 and 1 for at least one char in it
14085 const char open_char = *p;
14086 const char * temp_ptr = p + 1;
14088 /* These two constructs are not handled by perl, and if we find a
14089 * syntactically valid one, we croak. khw, who wrote this code, finds
14090 * this explanation of them very unclear:
14091 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
14092 * And searching the rest of the internet wasn't very helpful either.
14093 * It looks like just about any byte can be in these constructs,
14094 * depending on the locale. But unless the pattern is being compiled
14095 * under /l, which is very rare, Perl runs under the C or POSIX locale.
14096 * In that case, it looks like [= =] isn't allowed at all, and that
14097 * [. .] could be any single code point, but for longer strings the
14098 * constituent characters would have to be the ASCII alphabetics plus
14099 * the minus-hyphen. Any sensible locale definition would limit itself
14100 * to these. And any portable one definitely should. Trying to parse
14101 * the general case is a nightmare (see [perl #127604]). So, this code
14102 * looks only for interiors of these constructs that match:
14104 * Using \w relaxes the apparent rules a little, without adding much
14105 * danger of mistaking something else for one of these constructs.
14107 * [. .] in some implementations described on the internet is usable to
14108 * escape a character that otherwise is special in bracketed character
14109 * classes. For example [.].] means a literal right bracket instead of
14110 * the ending of the class
14112 * [= =] can legitimately contain a [. .] construct, but we don't
14113 * handle this case, as that [. .] construct will later get parsed
14114 * itself and croak then. And [= =] is checked for even when not under
14115 * /l, as Perl has long done so.
14117 * The code below relies on there being a trailing NUL, so it doesn't
14118 * have to keep checking if the parse ptr < e.
14120 if (temp_ptr[1] == open_char) {
14123 else while ( temp_ptr < e
14124 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
14129 if (*temp_ptr == open_char) {
14131 if (*temp_ptr == ']') {
14133 if (! found_problem && ! check_only) {
14134 RExC_parse = (char *) temp_ptr;
14135 vFAIL3("POSIX syntax [%c %c] is reserved for future "
14136 "extensions", open_char, open_char);
14139 /* Here, the syntax wasn't completely valid, or else the call
14140 * is to check-only */
14141 if (updated_parse_ptr) {
14142 *updated_parse_ptr = (char *) temp_ptr;
14145 return OOB_NAMEDCLASS;
14149 /* If we find something that started out to look like one of these
14150 * constructs, but isn't, we continue below so that it can be checked
14151 * for being a class name with a typo of '.' or '=' instead of a colon.
14155 /* Here, we think there is a possibility that a [: :] class was meant, and
14156 * we have the first real character. It could be they think the '^' comes
14159 found_problem = TRUE;
14160 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
14165 found_problem = TRUE;
14169 } while (p < e && isBLANK(*p));
14171 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14175 /* But the first character should be a colon, which they could have easily
14176 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
14177 * distinguish from a colon, so treat that as a colon). */
14180 has_opening_colon = TRUE;
14182 else if (*p == ';') {
14183 found_problem = TRUE;
14185 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14186 has_opening_colon = TRUE;
14189 found_problem = TRUE;
14190 ADD_POSIX_WARNING(p, "there must be a starting ':'");
14192 /* Consider an initial punctuation (not one of the recognized ones) to
14193 * be a left terminator */
14194 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
14199 /* They may think that you can put spaces between the components */
14201 found_problem = TRUE;
14205 } while (p < e && isBLANK(*p));
14207 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14212 /* We consider something like [^:^alnum:]] to not have been intended to
14213 * be a posix class, but XXX maybe we should */
14215 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14222 /* Again, they may think that you can put spaces between the components */
14224 found_problem = TRUE;
14228 } while (p < e && isBLANK(*p));
14230 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14235 /* XXX This ']' may be a typo, and something else was meant. But
14236 * treating it as such creates enough complications, that that
14237 * possibility isn't currently considered here. So we assume that the
14238 * ']' is what is intended, and if we've already found an initial '[',
14239 * this leaves this construct looking like [:] or [:^], which almost
14240 * certainly weren't intended to be posix classes */
14241 if (has_opening_bracket) {
14242 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14245 /* But this function can be called when we parse the colon for
14246 * something like qr/[alpha:]]/, so we back up to look for the
14251 found_problem = TRUE;
14252 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14254 else if (*p != ':') {
14256 /* XXX We are currently very restrictive here, so this code doesn't
14257 * consider the possibility that, say, /[alpha.]]/ was intended to
14258 * be a posix class. */
14259 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14262 /* Here we have something like 'foo:]'. There was no initial colon,
14263 * and we back up over 'foo. XXX Unlike the going forward case, we
14264 * don't handle typos of non-word chars in the middle */
14265 has_opening_colon = FALSE;
14268 while (p > RExC_start && isWORDCHAR(*p)) {
14273 /* Here, we have positioned ourselves to where we think the first
14274 * character in the potential class is */
14277 /* Now the interior really starts. There are certain key characters that
14278 * can end the interior, or these could just be typos. To catch both
14279 * cases, we may have to do two passes. In the first pass, we keep on
14280 * going unless we come to a sequence that matches
14281 * qr/ [[:punct:]] [[:blank:]]* \] /xa
14282 * This means it takes a sequence to end the pass, so two typos in a row if
14283 * that wasn't what was intended. If the class is perfectly formed, just
14284 * this one pass is needed. We also stop if there are too many characters
14285 * being accumulated, but this number is deliberately set higher than any
14286 * real class. It is set high enough so that someone who thinks that
14287 * 'alphanumeric' is a correct name would get warned that it wasn't.
14288 * While doing the pass, we keep track of where the key characters were in
14289 * it. If we don't find an end to the class, and one of the key characters
14290 * was found, we redo the pass, but stop when we get to that character.
14291 * Thus the key character was considered a typo in the first pass, but a
14292 * terminator in the second. If two key characters are found, we stop at
14293 * the second one in the first pass. Again this can miss two typos, but
14294 * catches a single one
14296 * In the first pass, 'possible_end' starts as NULL, and then gets set to
14297 * point to the first key character. For the second pass, it starts as -1.
14303 bool has_blank = FALSE;
14304 bool has_upper = FALSE;
14305 bool has_terminating_colon = FALSE;
14306 bool has_terminating_bracket = FALSE;
14307 bool has_semi_colon = FALSE;
14308 unsigned int name_len = 0;
14309 int punct_count = 0;
14313 /* Squeeze out blanks when looking up the class name below */
14314 if (isBLANK(*p) ) {
14316 found_problem = TRUE;
14321 /* The name will end with a punctuation */
14323 const char * peek = p + 1;
14325 /* Treat any non-']' punctuation followed by a ']' (possibly
14326 * with intervening blanks) as trying to terminate the class.
14327 * ']]' is very likely to mean a class was intended (but
14328 * missing the colon), but the warning message that gets
14329 * generated shows the error position better if we exit the
14330 * loop at the bottom (eventually), so skip it here. */
14332 if (peek < e && isBLANK(*peek)) {
14334 found_problem = TRUE;
14337 } while (peek < e && isBLANK(*peek));
14340 if (peek < e && *peek == ']') {
14341 has_terminating_bracket = TRUE;
14343 has_terminating_colon = TRUE;
14345 else if (*p == ';') {
14346 has_semi_colon = TRUE;
14347 has_terminating_colon = TRUE;
14350 found_problem = TRUE;
14357 /* Here we have punctuation we thought didn't end the class.
14358 * Keep track of the position of the key characters that are
14359 * more likely to have been class-enders */
14360 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
14362 /* Allow just one such possible class-ender not actually
14363 * ending the class. */
14364 if (possible_end) {
14370 /* If we have too many punctuation characters, no use in
14372 if (++punct_count > max_distance) {
14376 /* Treat the punctuation as a typo. */
14377 input_text[name_len++] = *p;
14380 else if (isUPPER(*p)) { /* Use lowercase for lookup */
14381 input_text[name_len++] = toLOWER(*p);
14383 found_problem = TRUE;
14385 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
14386 input_text[name_len++] = *p;
14390 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
14394 /* The declaration of 'input_text' is how long we allow a potential
14395 * class name to be, before saying they didn't mean a class name at
14397 if (name_len >= C_ARRAY_LENGTH(input_text)) {
14402 /* We get to here when the possible class name hasn't been properly
14403 * terminated before:
14404 * 1) we ran off the end of the pattern; or
14405 * 2) found two characters, each of which might have been intended to
14406 * be the name's terminator
14407 * 3) found so many punctuation characters in the purported name,
14408 * that the edit distance to a valid one is exceeded
14409 * 4) we decided it was more characters than anyone could have
14410 * intended to be one. */
14412 found_problem = TRUE;
14414 /* In the final two cases, we know that looking up what we've
14415 * accumulated won't lead to a match, even a fuzzy one. */
14416 if ( name_len >= C_ARRAY_LENGTH(input_text)
14417 || punct_count > max_distance)
14419 /* If there was an intermediate key character that could have been
14420 * an intended end, redo the parse, but stop there */
14421 if (possible_end && possible_end != (char *) -1) {
14422 possible_end = (char *) -1; /* Special signal value to say
14423 we've done a first pass */
14428 /* Otherwise, it can't have meant to have been a class */
14429 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14432 /* If we ran off the end, and the final character was a punctuation
14433 * one, back up one, to look at that final one just below. Later, we
14434 * will restore the parse pointer if appropriate */
14435 if (name_len && p == e && isPUNCT(*(p-1))) {
14440 if (p < e && isPUNCT(*p)) {
14442 has_terminating_bracket = TRUE;
14444 /* If this is a 2nd ']', and the first one is just below this
14445 * one, consider that to be the real terminator. This gives a
14446 * uniform and better positioning for the warning message */
14448 && possible_end != (char *) -1
14449 && *possible_end == ']'
14450 && name_len && input_text[name_len - 1] == ']')
14455 /* And this is actually equivalent to having done the 2nd
14456 * pass now, so set it to not try again */
14457 possible_end = (char *) -1;
14462 has_terminating_colon = TRUE;
14464 else if (*p == ';') {
14465 has_semi_colon = TRUE;
14466 has_terminating_colon = TRUE;
14474 /* Here, we have a class name to look up. We can short circuit the
14475 * stuff below for short names that can't possibly be meant to be a
14476 * class name. (We can do this on the first pass, as any second pass
14477 * will yield an even shorter name) */
14478 if (name_len < 3) {
14479 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14482 /* Find which class it is. Initially switch on the length of the name.
14484 switch (name_len) {
14486 if (memEQ(name_start, "word", 4)) {
14487 /* this is not POSIX, this is the Perl \w */
14488 class_number = ANYOF_WORDCHAR;
14492 /* Names all of length 5: alnum alpha ascii blank cntrl digit
14493 * graph lower print punct space upper
14494 * Offset 4 gives the best switch position. */
14495 switch (name_start[4]) {
14497 if (memEQ(name_start, "alph", 4)) /* alpha */
14498 class_number = ANYOF_ALPHA;
14501 if (memEQ(name_start, "spac", 4)) /* space */
14502 class_number = ANYOF_SPACE;
14505 if (memEQ(name_start, "grap", 4)) /* graph */
14506 class_number = ANYOF_GRAPH;
14509 if (memEQ(name_start, "asci", 4)) /* ascii */
14510 class_number = ANYOF_ASCII;
14513 if (memEQ(name_start, "blan", 4)) /* blank */
14514 class_number = ANYOF_BLANK;
14517 if (memEQ(name_start, "cntr", 4)) /* cntrl */
14518 class_number = ANYOF_CNTRL;
14521 if (memEQ(name_start, "alnu", 4)) /* alnum */
14522 class_number = ANYOF_ALPHANUMERIC;
14525 if (memEQ(name_start, "lowe", 4)) /* lower */
14526 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
14527 else if (memEQ(name_start, "uppe", 4)) /* upper */
14528 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
14531 if (memEQ(name_start, "digi", 4)) /* digit */
14532 class_number = ANYOF_DIGIT;
14533 else if (memEQ(name_start, "prin", 4)) /* print */
14534 class_number = ANYOF_PRINT;
14535 else if (memEQ(name_start, "punc", 4)) /* punct */
14536 class_number = ANYOF_PUNCT;
14541 if (memEQ(name_start, "xdigit", 6))
14542 class_number = ANYOF_XDIGIT;
14546 /* If the name exactly matches a posix class name the class number will
14547 * here be set to it, and the input almost certainly was meant to be a
14548 * posix class, so we can skip further checking. If instead the syntax
14549 * is exactly correct, but the name isn't one of the legal ones, we
14550 * will return that as an error below. But if neither of these apply,
14551 * it could be that no posix class was intended at all, or that one
14552 * was, but there was a typo. We tease these apart by doing fuzzy
14553 * matching on the name */
14554 if (class_number == OOB_NAMEDCLASS && found_problem) {
14555 const UV posix_names[][6] = {
14556 { 'a', 'l', 'n', 'u', 'm' },
14557 { 'a', 'l', 'p', 'h', 'a' },
14558 { 'a', 's', 'c', 'i', 'i' },
14559 { 'b', 'l', 'a', 'n', 'k' },
14560 { 'c', 'n', 't', 'r', 'l' },
14561 { 'd', 'i', 'g', 'i', 't' },
14562 { 'g', 'r', 'a', 'p', 'h' },
14563 { 'l', 'o', 'w', 'e', 'r' },
14564 { 'p', 'r', 'i', 'n', 't' },
14565 { 'p', 'u', 'n', 'c', 't' },
14566 { 's', 'p', 'a', 'c', 'e' },
14567 { 'u', 'p', 'p', 'e', 'r' },
14568 { 'w', 'o', 'r', 'd' },
14569 { 'x', 'd', 'i', 'g', 'i', 't' }
14571 /* The names of the above all have added NULs to make them the same
14572 * size, so we need to also have the real lengths */
14573 const UV posix_name_lengths[] = {
14574 sizeof("alnum") - 1,
14575 sizeof("alpha") - 1,
14576 sizeof("ascii") - 1,
14577 sizeof("blank") - 1,
14578 sizeof("cntrl") - 1,
14579 sizeof("digit") - 1,
14580 sizeof("graph") - 1,
14581 sizeof("lower") - 1,
14582 sizeof("print") - 1,
14583 sizeof("punct") - 1,
14584 sizeof("space") - 1,
14585 sizeof("upper") - 1,
14586 sizeof("word") - 1,
14587 sizeof("xdigit")- 1
14590 int temp_max = max_distance; /* Use a temporary, so if we
14591 reparse, we haven't changed the
14594 /* Use a smaller max edit distance if we are missing one of the
14596 if ( has_opening_bracket + has_opening_colon < 2
14597 || has_terminating_bracket + has_terminating_colon < 2)
14602 /* See if the input name is close to a legal one */
14603 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
14605 /* Short circuit call if the lengths are too far apart to be
14607 if (abs( (int) (name_len - posix_name_lengths[i]))
14613 if (edit_distance(input_text,
14616 posix_name_lengths[i],
14620 { /* If it is close, it probably was intended to be a class */
14621 goto probably_meant_to_be;
14625 /* Here the input name is not close enough to a valid class name
14626 * for us to consider it to be intended to be a posix class. If
14627 * we haven't already done so, and the parse found a character that
14628 * could have been terminators for the name, but which we absorbed
14629 * as typos during the first pass, repeat the parse, signalling it
14630 * to stop at that character */
14631 if (possible_end && possible_end != (char *) -1) {
14632 possible_end = (char *) -1;
14637 /* Here neither pass found a close-enough class name */
14638 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14641 probably_meant_to_be:
14643 /* Here we think that a posix specification was intended. Update any
14645 if (updated_parse_ptr) {
14646 *updated_parse_ptr = (char *) p;
14649 /* If a posix class name was intended but incorrectly specified, we
14650 * output or return the warnings */
14651 if (found_problem) {
14653 /* We set flags for these issues in the parse loop above instead of
14654 * adding them to the list of warnings, because we can parse it
14655 * twice, and we only want one warning instance */
14657 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
14660 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14662 if (has_semi_colon) {
14663 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14665 else if (! has_terminating_colon) {
14666 ADD_POSIX_WARNING(p, "there is no terminating ':'");
14668 if (! has_terminating_bracket) {
14669 ADD_POSIX_WARNING(p, "there is no terminating ']'");
14672 if (posix_warnings && RExC_warn_text && av_top_index(RExC_warn_text) > -1) {
14673 *posix_warnings = RExC_warn_text;
14676 else if (class_number != OOB_NAMEDCLASS) {
14677 /* If it is a known class, return the class. The class number
14678 * #defines are structured so each complement is +1 to the normal
14680 return class_number + complement;
14682 else if (! check_only) {
14684 /* Here, it is an unrecognized class. This is an error (unless the
14685 * call is to check only, which we've already handled above) */
14686 const char * const complement_string = (complement)
14689 RExC_parse = (char *) p;
14690 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
14692 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
14696 return OOB_NAMEDCLASS;
14698 #undef ADD_POSIX_WARNING
14700 STATIC unsigned int
14701 S_regex_set_precedence(const U8 my_operator) {
14703 /* Returns the precedence in the (?[...]) construct of the input operator,
14704 * specified by its character representation. The precedence follows
14705 * general Perl rules, but it extends this so that ')' and ']' have (low)
14706 * precedence even though they aren't really operators */
14708 switch (my_operator) {
14724 NOT_REACHED; /* NOTREACHED */
14725 return 0; /* Silence compiler warning */
14729 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
14730 I32 *flagp, U32 depth,
14731 char * const oregcomp_parse)
14733 /* Handle the (?[...]) construct to do set operations */
14735 U8 curchar; /* Current character being parsed */
14736 UV start, end; /* End points of code point ranges */
14737 SV* final = NULL; /* The end result inversion list */
14738 SV* result_string; /* 'final' stringified */
14739 AV* stack; /* stack of operators and operands not yet
14741 AV* fence_stack = NULL; /* A stack containing the positions in
14742 'stack' of where the undealt-with left
14743 parens would be if they were actually
14745 /* The 'VOL' (expanding to 'volatile') is a workaround for an optimiser bug
14746 * in Solaris Studio 12.3. See RT #127455 */
14747 VOL IV fence = 0; /* Position of where most recent undealt-
14748 with left paren in stack is; -1 if none.
14750 STRLEN len; /* Temporary */
14751 regnode* node; /* Temporary, and final regnode returned by
14753 const bool save_fold = FOLD; /* Temporary */
14754 char *save_end, *save_parse; /* Temporaries */
14755 const bool in_locale = LOC; /* we turn off /l during processing */
14756 AV* posix_warnings = NULL;
14758 GET_RE_DEBUG_FLAGS_DECL;
14760 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
14763 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
14766 REQUIRE_UNI_RULES(flagp, NULL); /* The use of this operator implies /u.
14767 This is required so that the compile
14768 time values are valid in all runtime
14771 /* This will return only an ANYOF regnode, or (unlikely) something smaller
14772 * (such as EXACT). Thus we can skip most everything if just sizing. We
14773 * call regclass to handle '[]' so as to not have to reinvent its parsing
14774 * rules here (throwing away the size it computes each time). And, we exit
14775 * upon an unescaped ']' that isn't one ending a regclass. To do both
14776 * these things, we need to realize that something preceded by a backslash
14777 * is escaped, so we have to keep track of backslashes */
14779 UV depth = 0; /* how many nested (?[...]) constructs */
14781 while (RExC_parse < RExC_end) {
14782 SV* current = NULL;
14784 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14785 TRUE /* Force /x */ );
14787 switch (*RExC_parse) {
14789 if (RExC_parse[1] == '[') depth++, RExC_parse++;
14794 /* Skip past this, so the next character gets skipped, after
14797 if (*RExC_parse == 'c') {
14798 /* Skip the \cX notation for control characters */
14799 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14805 /* See if this is a [:posix:] class. */
14806 bool is_posix_class = (OOB_NAMEDCLASS
14807 < handle_possible_posix(pRExC_state,
14811 TRUE /* checking only */));
14812 /* If it is a posix class, leave the parse pointer at the
14813 * '[' to fool regclass() into thinking it is part of a
14814 * '[[:posix:]]'. */
14815 if (! is_posix_class) {
14819 /* regclass() can only return RESTART_PASS1 and NEED_UTF8
14820 * if multi-char folds are allowed. */
14821 if (!regclass(pRExC_state, flagp,depth+1,
14822 is_posix_class, /* parse the whole char
14823 class only if not a
14825 FALSE, /* don't allow multi-char folds */
14826 TRUE, /* silence non-portable warnings. */
14828 FALSE, /* Require return to be an ANYOF */
14832 FAIL2("panic: regclass returned NULL to handle_sets, "
14833 "flags=%#" UVxf, (UV) *flagp);
14835 /* function call leaves parse pointing to the ']', except
14836 * if we faked it */
14837 if (is_posix_class) {
14841 SvREFCNT_dec(current); /* In case it returned something */
14846 if (depth--) break;
14848 if (*RExC_parse == ')') {
14849 node = reganode(pRExC_state, ANYOF, 0);
14850 RExC_size += ANYOF_SKIP;
14851 nextchar(pRExC_state);
14852 Set_Node_Length(node,
14853 RExC_parse - oregcomp_parse + 1); /* MJD */
14855 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
14863 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14867 /* We output the messages even if warnings are off, because we'll fail
14868 * the very next thing, and these give a likely diagnosis for that */
14869 if (posix_warnings && av_tindex_nomg(posix_warnings) >= 0) {
14870 output_or_return_posix_warnings(pRExC_state, posix_warnings, NULL);
14873 FAIL("Syntax error in (?[...])");
14876 /* Pass 2 only after this. */
14877 Perl_ck_warner_d(aTHX_
14878 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
14879 "The regex_sets feature is experimental" REPORT_LOCATION,
14880 REPORT_LOCATION_ARGS(RExC_parse));
14882 /* Everything in this construct is a metacharacter. Operands begin with
14883 * either a '\' (for an escape sequence), or a '[' for a bracketed
14884 * character class. Any other character should be an operator, or
14885 * parenthesis for grouping. Both types of operands are handled by calling
14886 * regclass() to parse them. It is called with a parameter to indicate to
14887 * return the computed inversion list. The parsing here is implemented via
14888 * a stack. Each entry on the stack is a single character representing one
14889 * of the operators; or else a pointer to an operand inversion list. */
14891 #define IS_OPERATOR(a) SvIOK(a)
14892 #define IS_OPERAND(a) (! IS_OPERATOR(a))
14894 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
14895 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
14896 * with pronouncing it called it Reverse Polish instead, but now that YOU
14897 * know how to pronounce it you can use the correct term, thus giving due
14898 * credit to the person who invented it, and impressing your geek friends.
14899 * Wikipedia says that the pronounciation of "Ł" has been changing so that
14900 * it is now more like an English initial W (as in wonk) than an L.)
14902 * This means that, for example, 'a | b & c' is stored on the stack as
14910 * where the numbers in brackets give the stack [array] element number.
14911 * In this implementation, parentheses are not stored on the stack.
14912 * Instead a '(' creates a "fence" so that the part of the stack below the
14913 * fence is invisible except to the corresponding ')' (this allows us to
14914 * replace testing for parens, by using instead subtraction of the fence
14915 * position). As new operands are processed they are pushed onto the stack
14916 * (except as noted in the next paragraph). New operators of higher
14917 * precedence than the current final one are inserted on the stack before
14918 * the lhs operand (so that when the rhs is pushed next, everything will be
14919 * in the correct positions shown above. When an operator of equal or
14920 * lower precedence is encountered in parsing, all the stacked operations
14921 * of equal or higher precedence are evaluated, leaving the result as the
14922 * top entry on the stack. This makes higher precedence operations
14923 * evaluate before lower precedence ones, and causes operations of equal
14924 * precedence to left associate.
14926 * The only unary operator '!' is immediately pushed onto the stack when
14927 * encountered. When an operand is encountered, if the top of the stack is
14928 * a '!", the complement is immediately performed, and the '!' popped. The
14929 * resulting value is treated as a new operand, and the logic in the
14930 * previous paragraph is executed. Thus in the expression
14932 * the stack looks like
14938 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
14945 * A ')' is treated as an operator with lower precedence than all the
14946 * aforementioned ones, which causes all operations on the stack above the
14947 * corresponding '(' to be evaluated down to a single resultant operand.
14948 * Then the fence for the '(' is removed, and the operand goes through the
14949 * algorithm above, without the fence.
14951 * A separate stack is kept of the fence positions, so that the position of
14952 * the latest so-far unbalanced '(' is at the top of it.
14954 * The ']' ending the construct is treated as the lowest operator of all,
14955 * so that everything gets evaluated down to a single operand, which is the
14958 sv_2mortal((SV *)(stack = newAV()));
14959 sv_2mortal((SV *)(fence_stack = newAV()));
14961 while (RExC_parse < RExC_end) {
14962 I32 top_index; /* Index of top-most element in 'stack' */
14963 SV** top_ptr; /* Pointer to top 'stack' element */
14964 SV* current = NULL; /* To contain the current inversion list
14966 SV* only_to_avoid_leaks;
14968 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14969 TRUE /* Force /x */ );
14970 if (RExC_parse >= RExC_end) {
14971 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
14974 curchar = UCHARAT(RExC_parse);
14978 #ifdef ENABLE_REGEX_SETS_DEBUGGING
14979 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
14980 DEBUG_U(dump_regex_sets_structures(pRExC_state,
14981 stack, fence, fence_stack));
14984 top_index = av_tindex_nomg(stack);
14987 SV** stacked_ptr; /* Ptr to something already on 'stack' */
14988 char stacked_operator; /* The topmost operator on the 'stack'. */
14989 SV* lhs; /* Operand to the left of the operator */
14990 SV* rhs; /* Operand to the right of the operator */
14991 SV* fence_ptr; /* Pointer to top element of the fence
14996 if ( RExC_parse < RExC_end - 1
14997 && (UCHARAT(RExC_parse + 1) == '?'))
14999 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
15000 * This happens when we have some thing like
15002 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
15004 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
15006 * Here we would be handling the interpolated
15007 * '$thai_or_lao'. We handle this by a recursive call to
15008 * ourselves which returns the inversion list the
15009 * interpolated expression evaluates to. We use the flags
15010 * from the interpolated pattern. */
15011 U32 save_flags = RExC_flags;
15012 const char * save_parse;
15014 RExC_parse += 2; /* Skip past the '(?' */
15015 save_parse = RExC_parse;
15017 /* Parse any flags for the '(?' */
15018 parse_lparen_question_flags(pRExC_state);
15020 if (RExC_parse == save_parse /* Makes sure there was at
15021 least one flag (or else
15022 this embedding wasn't
15024 || RExC_parse >= RExC_end - 4
15025 || UCHARAT(RExC_parse) != ':'
15026 || UCHARAT(++RExC_parse) != '('
15027 || UCHARAT(++RExC_parse) != '?'
15028 || UCHARAT(++RExC_parse) != '[')
15031 /* In combination with the above, this moves the
15032 * pointer to the point just after the first erroneous
15033 * character (or if there are no flags, to where they
15034 * should have been) */
15035 if (RExC_parse >= RExC_end - 4) {
15036 RExC_parse = RExC_end;
15038 else if (RExC_parse != save_parse) {
15039 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15041 vFAIL("Expecting '(?flags:(?[...'");
15044 /* Recurse, with the meat of the embedded expression */
15046 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
15047 depth+1, oregcomp_parse);
15049 /* Here, 'current' contains the embedded expression's
15050 * inversion list, and RExC_parse points to the trailing
15051 * ']'; the next character should be the ')' */
15053 assert(UCHARAT(RExC_parse) == ')');
15055 /* Then the ')' matching the original '(' handled by this
15056 * case: statement */
15058 assert(UCHARAT(RExC_parse) == ')');
15061 RExC_flags = save_flags;
15062 goto handle_operand;
15065 /* A regular '('. Look behind for illegal syntax */
15066 if (top_index - fence >= 0) {
15067 /* If the top entry on the stack is an operator, it had
15068 * better be a '!', otherwise the entry below the top
15069 * operand should be an operator */
15070 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
15071 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
15072 || ( IS_OPERAND(*top_ptr)
15073 && ( top_index - fence < 1
15074 || ! (stacked_ptr = av_fetch(stack,
15077 || ! IS_OPERATOR(*stacked_ptr))))
15080 vFAIL("Unexpected '(' with no preceding operator");
15084 /* Stack the position of this undealt-with left paren */
15085 av_push(fence_stack, newSViv(fence));
15086 fence = top_index + 1;
15090 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15091 * multi-char folds are allowed. */
15092 if (!regclass(pRExC_state, flagp,depth+1,
15093 TRUE, /* means parse just the next thing */
15094 FALSE, /* don't allow multi-char folds */
15095 FALSE, /* don't silence non-portable warnings. */
15097 FALSE, /* Require return to be an ANYOF */
15101 FAIL2("panic: regclass returned NULL to handle_sets, "
15102 "flags=%#" UVxf, (UV) *flagp);
15105 /* regclass() will return with parsing just the \ sequence,
15106 * leaving the parse pointer at the next thing to parse */
15108 goto handle_operand;
15110 case '[': /* Is a bracketed character class */
15112 /* See if this is a [:posix:] class. */
15113 bool is_posix_class = (OOB_NAMEDCLASS
15114 < handle_possible_posix(pRExC_state,
15118 TRUE /* checking only */));
15119 /* If it is a posix class, leave the parse pointer at the '['
15120 * to fool regclass() into thinking it is part of a
15121 * '[[:posix:]]'. */
15122 if (! is_posix_class) {
15126 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15127 * multi-char folds are allowed. */
15128 if (!regclass(pRExC_state, flagp,depth+1,
15129 is_posix_class, /* parse the whole char
15130 class only if not a
15132 FALSE, /* don't allow multi-char folds */
15133 TRUE, /* silence non-portable warnings. */
15135 FALSE, /* Require return to be an ANYOF */
15140 FAIL2("panic: regclass returned NULL to handle_sets, "
15141 "flags=%#" UVxf, (UV) *flagp);
15144 /* function call leaves parse pointing to the ']', except if we
15146 if (is_posix_class) {
15150 goto handle_operand;
15154 if (top_index >= 1) {
15155 goto join_operators;
15158 /* Only a single operand on the stack: are done */
15162 if (av_tindex_nomg(fence_stack) < 0) {
15164 vFAIL("Unexpected ')'");
15167 /* If nothing after the fence, is missing an operand */
15168 if (top_index - fence < 0) {
15172 /* If at least two things on the stack, treat this as an
15174 if (top_index - fence >= 1) {
15175 goto join_operators;
15178 /* Here only a single thing on the fenced stack, and there is a
15179 * fence. Get rid of it */
15180 fence_ptr = av_pop(fence_stack);
15182 fence = SvIV(fence_ptr) - 1;
15183 SvREFCNT_dec_NN(fence_ptr);
15190 /* Having gotten rid of the fence, we pop the operand at the
15191 * stack top and process it as a newly encountered operand */
15192 current = av_pop(stack);
15193 if (IS_OPERAND(current)) {
15194 goto handle_operand;
15206 /* These binary operators should have a left operand already
15208 if ( top_index - fence < 0
15209 || top_index - fence == 1
15210 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
15211 || ! IS_OPERAND(*top_ptr))
15213 goto unexpected_binary;
15216 /* If only the one operand is on the part of the stack visible
15217 * to us, we just place this operator in the proper position */
15218 if (top_index - fence < 2) {
15220 /* Place the operator before the operand */
15222 SV* lhs = av_pop(stack);
15223 av_push(stack, newSVuv(curchar));
15224 av_push(stack, lhs);
15228 /* But if there is something else on the stack, we need to
15229 * process it before this new operator if and only if the
15230 * stacked operation has equal or higher precedence than the
15235 /* The operator on the stack is supposed to be below both its
15237 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
15238 || IS_OPERAND(*stacked_ptr))
15240 /* But if not, it's legal and indicates we are completely
15241 * done if and only if we're currently processing a ']',
15242 * which should be the final thing in the expression */
15243 if (curchar == ']') {
15249 vFAIL2("Unexpected binary operator '%c' with no "
15250 "preceding operand", curchar);
15252 stacked_operator = (char) SvUV(*stacked_ptr);
15254 if (regex_set_precedence(curchar)
15255 > regex_set_precedence(stacked_operator))
15257 /* Here, the new operator has higher precedence than the
15258 * stacked one. This means we need to add the new one to
15259 * the stack to await its rhs operand (and maybe more
15260 * stuff). We put it before the lhs operand, leaving
15261 * untouched the stacked operator and everything below it
15263 lhs = av_pop(stack);
15264 assert(IS_OPERAND(lhs));
15266 av_push(stack, newSVuv(curchar));
15267 av_push(stack, lhs);
15271 /* Here, the new operator has equal or lower precedence than
15272 * what's already there. This means the operation already
15273 * there should be performed now, before the new one. */
15275 rhs = av_pop(stack);
15276 if (! IS_OPERAND(rhs)) {
15278 /* This can happen when a ! is not followed by an operand,
15279 * like in /(?[\t &!])/ */
15283 lhs = av_pop(stack);
15285 if (! IS_OPERAND(lhs)) {
15287 /* This can happen when there is an empty (), like in
15288 * /(?[[0]+()+])/ */
15292 switch (stacked_operator) {
15294 _invlist_intersection(lhs, rhs, &rhs);
15299 _invlist_union(lhs, rhs, &rhs);
15303 _invlist_subtract(lhs, rhs, &rhs);
15306 case '^': /* The union minus the intersection */
15311 _invlist_union(lhs, rhs, &u);
15312 _invlist_intersection(lhs, rhs, &i);
15313 _invlist_subtract(u, i, &rhs);
15314 SvREFCNT_dec_NN(i);
15315 SvREFCNT_dec_NN(u);
15321 /* Here, the higher precedence operation has been done, and the
15322 * result is in 'rhs'. We overwrite the stacked operator with
15323 * the result. Then we redo this code to either push the new
15324 * operator onto the stack or perform any higher precedence
15325 * stacked operation */
15326 only_to_avoid_leaks = av_pop(stack);
15327 SvREFCNT_dec(only_to_avoid_leaks);
15328 av_push(stack, rhs);
15331 case '!': /* Highest priority, right associative */
15333 /* If what's already at the top of the stack is another '!",
15334 * they just cancel each other out */
15335 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
15336 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
15338 only_to_avoid_leaks = av_pop(stack);
15339 SvREFCNT_dec(only_to_avoid_leaks);
15341 else { /* Otherwise, since it's right associative, just push
15343 av_push(stack, newSVuv(curchar));
15348 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15349 vFAIL("Unexpected character");
15353 /* Here 'current' is the operand. If something is already on the
15354 * stack, we have to check if it is a !. But first, the code above
15355 * may have altered the stack in the time since we earlier set
15358 top_index = av_tindex_nomg(stack);
15359 if (top_index - fence >= 0) {
15360 /* If the top entry on the stack is an operator, it had better
15361 * be a '!', otherwise the entry below the top operand should
15362 * be an operator */
15363 top_ptr = av_fetch(stack, top_index, FALSE);
15365 if (IS_OPERATOR(*top_ptr)) {
15367 /* The only permissible operator at the top of the stack is
15368 * '!', which is applied immediately to this operand. */
15369 curchar = (char) SvUV(*top_ptr);
15370 if (curchar != '!') {
15371 SvREFCNT_dec(current);
15372 vFAIL2("Unexpected binary operator '%c' with no "
15373 "preceding operand", curchar);
15376 _invlist_invert(current);
15378 only_to_avoid_leaks = av_pop(stack);
15379 SvREFCNT_dec(only_to_avoid_leaks);
15381 /* And we redo with the inverted operand. This allows
15382 * handling multiple ! in a row */
15383 goto handle_operand;
15385 /* Single operand is ok only for the non-binary ')'
15387 else if ((top_index - fence == 0 && curchar != ')')
15388 || (top_index - fence > 0
15389 && (! (stacked_ptr = av_fetch(stack,
15392 || IS_OPERAND(*stacked_ptr))))
15394 SvREFCNT_dec(current);
15395 vFAIL("Operand with no preceding operator");
15399 /* Here there was nothing on the stack or the top element was
15400 * another operand. Just add this new one */
15401 av_push(stack, current);
15403 } /* End of switch on next parse token */
15405 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15406 } /* End of loop parsing through the construct */
15409 if (av_tindex_nomg(fence_stack) >= 0) {
15410 vFAIL("Unmatched (");
15413 if (av_tindex_nomg(stack) < 0 /* Was empty */
15414 || ((final = av_pop(stack)) == NULL)
15415 || ! IS_OPERAND(final)
15416 || SvTYPE(final) != SVt_INVLIST
15417 || av_tindex_nomg(stack) >= 0) /* More left on stack */
15420 SvREFCNT_dec(final);
15421 vFAIL("Incomplete expression within '(?[ ])'");
15424 /* Here, 'final' is the resultant inversion list from evaluating the
15425 * expression. Return it if so requested */
15426 if (return_invlist) {
15427 *return_invlist = final;
15431 /* Otherwise generate a resultant node, based on 'final'. regclass() is
15432 * expecting a string of ranges and individual code points */
15433 invlist_iterinit(final);
15434 result_string = newSVpvs("");
15435 while (invlist_iternext(final, &start, &end)) {
15436 if (start == end) {
15437 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
15440 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%" UVXf "}",
15445 /* About to generate an ANYOF (or similar) node from the inversion list we
15446 * have calculated */
15447 save_parse = RExC_parse;
15448 RExC_parse = SvPV(result_string, len);
15449 save_end = RExC_end;
15450 RExC_end = RExC_parse + len;
15452 /* We turn off folding around the call, as the class we have constructed
15453 * already has all folding taken into consideration, and we don't want
15454 * regclass() to add to that */
15455 RExC_flags &= ~RXf_PMf_FOLD;
15456 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if multi-char
15457 * folds are allowed. */
15458 node = regclass(pRExC_state, flagp,depth+1,
15459 FALSE, /* means parse the whole char class */
15460 FALSE, /* don't allow multi-char folds */
15461 TRUE, /* silence non-portable warnings. The above may very
15462 well have generated non-portable code points, but
15463 they're valid on this machine */
15464 FALSE, /* similarly, no need for strict */
15465 FALSE, /* Require return to be an ANYOF */
15470 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#" UVxf,
15473 /* Fix up the node type if we are in locale. (We have pretended we are
15474 * under /u for the purposes of regclass(), as this construct will only
15475 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
15476 * as to cause any warnings about bad locales to be output in regexec.c),
15477 * and add the flag that indicates to check if not in a UTF-8 locale. The
15478 * reason we above forbid optimization into something other than an ANYOF
15479 * node is simply to minimize the number of code changes in regexec.c.
15480 * Otherwise we would have to create new EXACTish node types and deal with
15481 * them. This decision could be revisited should this construct become
15484 * (One might think we could look at the resulting ANYOF node and suppress
15485 * the flag if everything is above 255, as those would be UTF-8 only,
15486 * but this isn't true, as the components that led to that result could
15487 * have been locale-affected, and just happen to cancel each other out
15488 * under UTF-8 locales.) */
15490 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
15492 assert(OP(node) == ANYOF);
15496 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
15500 RExC_flags |= RXf_PMf_FOLD;
15503 RExC_parse = save_parse + 1;
15504 RExC_end = save_end;
15505 SvREFCNT_dec_NN(final);
15506 SvREFCNT_dec_NN(result_string);
15508 nextchar(pRExC_state);
15509 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
15513 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15516 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
15517 AV * stack, const IV fence, AV * fence_stack)
15518 { /* Dumps the stacks in handle_regex_sets() */
15520 const SSize_t stack_top = av_tindex_nomg(stack);
15521 const SSize_t fence_stack_top = av_tindex_nomg(fence_stack);
15524 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
15526 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
15528 if (stack_top < 0) {
15529 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
15532 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
15533 for (i = stack_top; i >= 0; i--) {
15534 SV ** element_ptr = av_fetch(stack, i, FALSE);
15535 if (! element_ptr) {
15538 if (IS_OPERATOR(*element_ptr)) {
15539 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
15540 (int) i, (int) SvIV(*element_ptr));
15543 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
15544 sv_dump(*element_ptr);
15549 if (fence_stack_top < 0) {
15550 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
15553 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
15554 for (i = fence_stack_top; i >= 0; i--) {
15555 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
15556 if (! element_ptr) {
15559 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
15560 (int) i, (int) SvIV(*element_ptr));
15571 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
15573 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
15574 * innocent-looking character class, like /[ks]/i won't have to go out to
15575 * disk to find the possible matches.
15577 * This should be called only for a Latin1-range code points, cp, which is
15578 * known to be involved in a simple fold with other code points above
15579 * Latin1. It would give false results if /aa has been specified.
15580 * Multi-char folds are outside the scope of this, and must be handled
15583 * XXX It would be better to generate these via regen, in case a new
15584 * version of the Unicode standard adds new mappings, though that is not
15585 * really likely, and may be caught by the default: case of the switch
15588 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
15590 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
15596 add_cp_to_invlist(*invlist, KELVIN_SIGN);
15600 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
15603 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
15604 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
15606 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
15607 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
15608 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
15610 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
15611 *invlist = add_cp_to_invlist(*invlist,
15612 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
15615 #ifdef LATIN_CAPITAL_LETTER_SHARP_S /* not defined in early Unicode releases */
15617 case LATIN_SMALL_LETTER_SHARP_S:
15618 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
15623 #if UNICODE_MAJOR_VERSION < 3 \
15624 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0)
15626 /* In 3.0 and earlier, U+0130 folded simply to 'i'; and in 3.0.1 so did
15631 add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
15632 # if UNICODE_DOT_DOT_VERSION == 1
15633 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_DOTLESS_I);
15639 /* Use deprecated warning to increase the chances of this being
15642 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
15649 S_output_or_return_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings, AV** return_posix_warnings)
15651 /* If the final parameter is NULL, output the elements of the array given
15652 * by '*posix_warnings' as REGEXP warnings. Otherwise, the elements are
15653 * pushed onto it, (creating if necessary) */
15656 const bool first_is_fatal = ! return_posix_warnings
15657 && ckDEAD(packWARN(WARN_REGEXP));
15659 PERL_ARGS_ASSERT_OUTPUT_OR_RETURN_POSIX_WARNINGS;
15661 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
15662 if (return_posix_warnings) {
15663 if (! *return_posix_warnings) { /* mortalize to not leak if
15664 warnings are fatal */
15665 *return_posix_warnings = (AV *) sv_2mortal((SV *) newAV());
15667 av_push(*return_posix_warnings, msg);
15670 if (first_is_fatal) { /* Avoid leaking this */
15671 av_undef(posix_warnings); /* This isn't necessary if the
15672 array is mortal, but is a
15674 (void) sv_2mortal(msg);
15676 SAVEFREESV(RExC_rx_sv);
15679 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
15680 SvREFCNT_dec_NN(msg);
15686 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
15688 /* This adds the string scalar <multi_string> to the array
15689 * <multi_char_matches>. <multi_string> is known to have exactly
15690 * <cp_count> code points in it. This is used when constructing a
15691 * bracketed character class and we find something that needs to match more
15692 * than a single character.
15694 * <multi_char_matches> is actually an array of arrays. Each top-level
15695 * element is an array that contains all the strings known so far that are
15696 * the same length. And that length (in number of code points) is the same
15697 * as the index of the top-level array. Hence, the [2] element is an
15698 * array, each element thereof is a string containing TWO code points;
15699 * while element [3] is for strings of THREE characters, and so on. Since
15700 * this is for multi-char strings there can never be a [0] nor [1] element.
15702 * When we rewrite the character class below, we will do so such that the
15703 * longest strings are written first, so that it prefers the longest
15704 * matching strings first. This is done even if it turns out that any
15705 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
15706 * Christiansen has agreed that this is ok. This makes the test for the
15707 * ligature 'ffi' come before the test for 'ff', for example */
15710 AV** this_array_ptr;
15712 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
15714 if (! multi_char_matches) {
15715 multi_char_matches = newAV();
15718 if (av_exists(multi_char_matches, cp_count)) {
15719 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
15720 this_array = *this_array_ptr;
15723 this_array = newAV();
15724 av_store(multi_char_matches, cp_count,
15727 av_push(this_array, multi_string);
15729 return multi_char_matches;
15732 /* The names of properties whose definitions are not known at compile time are
15733 * stored in this SV, after a constant heading. So if the length has been
15734 * changed since initialization, then there is a run-time definition. */
15735 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
15736 (SvCUR(listsv) != initial_listsv_len)
15738 /* There is a restricted set of white space characters that are legal when
15739 * ignoring white space in a bracketed character class. This generates the
15740 * code to skip them.
15742 * There is a line below that uses the same white space criteria but is outside
15743 * this macro. Both here and there must use the same definition */
15744 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
15747 while (isBLANK_A(UCHARAT(p))) \
15755 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
15756 const bool stop_at_1, /* Just parse the next thing, don't
15757 look for a full character class */
15758 bool allow_multi_folds,
15759 const bool silence_non_portable, /* Don't output warnings
15763 bool optimizable, /* ? Allow a non-ANYOF return
15765 SV** ret_invlist, /* Return an inversion list, not a node */
15766 AV** return_posix_warnings
15769 /* parse a bracketed class specification. Most of these will produce an
15770 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
15771 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
15772 * under /i with multi-character folds: it will be rewritten following the
15773 * paradigm of this example, where the <multi-fold>s are characters which
15774 * fold to multiple character sequences:
15775 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
15776 * gets effectively rewritten as:
15777 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
15778 * reg() gets called (recursively) on the rewritten version, and this
15779 * function will return what it constructs. (Actually the <multi-fold>s
15780 * aren't physically removed from the [abcdefghi], it's just that they are
15781 * ignored in the recursion by means of a flag:
15782 * <RExC_in_multi_char_class>.)
15784 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
15785 * characters, with the corresponding bit set if that character is in the
15786 * list. For characters above this, a range list or swash is used. There
15787 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
15788 * determinable at compile time
15790 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs
15791 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded
15792 * to UTF-8. This can only happen if ret_invlist is non-NULL.
15795 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
15797 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
15800 int namedclass = OOB_NAMEDCLASS;
15801 char *rangebegin = NULL;
15802 bool need_class = 0;
15804 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
15805 than just initialized. */
15806 SV* properties = NULL; /* Code points that match \p{} \P{} */
15807 SV* posixes = NULL; /* Code points that match classes like [:word:],
15808 extended beyond the Latin1 range. These have to
15809 be kept separate from other code points for much
15810 of this function because their handling is
15811 different under /i, and for most classes under
15813 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
15814 separate for a while from the non-complemented
15815 versions because of complications with /d
15817 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
15818 treated more simply than the general case,
15819 leading to less compilation and execution
15821 UV element_count = 0; /* Number of distinct elements in the class.
15822 Optimizations may be possible if this is tiny */
15823 AV * multi_char_matches = NULL; /* Code points that fold to more than one
15824 character; used under /i */
15826 char * stop_ptr = RExC_end; /* where to stop parsing */
15828 /* ignore unescaped whitespace? */
15829 const bool skip_white = cBOOL( ret_invlist
15830 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
15832 /* Unicode properties are stored in a swash; this holds the current one
15833 * being parsed. If this swash is the only above-latin1 component of the
15834 * character class, an optimization is to pass it directly on to the
15835 * execution engine. Otherwise, it is set to NULL to indicate that there
15836 * are other things in the class that have to be dealt with at execution
15838 SV* swash = NULL; /* Code points that match \p{} \P{} */
15840 /* Set if a component of this character class is user-defined; just passed
15841 * on to the engine */
15842 bool has_user_defined_property = FALSE;
15844 /* inversion list of code points this node matches only when the target
15845 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
15847 SV* has_upper_latin1_only_utf8_matches = NULL;
15849 /* Inversion list of code points this node matches regardless of things
15850 * like locale, folding, utf8ness of the target string */
15851 SV* cp_list = NULL;
15853 /* Like cp_list, but code points on this list need to be checked for things
15854 * that fold to/from them under /i */
15855 SV* cp_foldable_list = NULL;
15857 /* Like cp_list, but code points on this list are valid only when the
15858 * runtime locale is UTF-8 */
15859 SV* only_utf8_locale_list = NULL;
15861 /* In a range, if one of the endpoints is non-character-set portable,
15862 * meaning that it hard-codes a code point that may mean a different
15863 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
15864 * mnemonic '\t' which each mean the same character no matter which
15865 * character set the platform is on. */
15866 unsigned int non_portable_endpoint = 0;
15868 /* Is the range unicode? which means on a platform that isn't 1-1 native
15869 * to Unicode (i.e. non-ASCII), each code point in it should be considered
15870 * to be a Unicode value. */
15871 bool unicode_range = FALSE;
15872 bool invert = FALSE; /* Is this class to be complemented */
15874 bool warn_super = ALWAYS_WARN_SUPER;
15876 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
15877 case we need to change the emitted regop to an EXACT. */
15878 const char * orig_parse = RExC_parse;
15879 const SSize_t orig_size = RExC_size;
15880 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
15882 /* This variable is used to mark where the end in the input is of something
15883 * that looks like a POSIX construct but isn't. During the parse, when
15884 * something looks like it could be such a construct is encountered, it is
15885 * checked for being one, but not if we've already checked this area of the
15886 * input. Only after this position is reached do we check again */
15887 char *not_posix_region_end = RExC_parse - 1;
15889 AV* posix_warnings = NULL;
15890 const bool do_posix_warnings = return_posix_warnings
15891 || (PASS2 && ckWARN(WARN_REGEXP));
15893 GET_RE_DEBUG_FLAGS_DECL;
15895 PERL_ARGS_ASSERT_REGCLASS;
15897 PERL_UNUSED_ARG(depth);
15900 DEBUG_PARSE("clas");
15902 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
15903 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
15904 && UNICODE_DOT_DOT_VERSION == 0)
15905 allow_multi_folds = FALSE;
15908 /* Assume we are going to generate an ANYOF node. */
15909 ret = reganode(pRExC_state,
15916 RExC_size += ANYOF_SKIP;
15917 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
15920 ANYOF_FLAGS(ret) = 0;
15922 RExC_emit += ANYOF_SKIP;
15923 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
15924 initial_listsv_len = SvCUR(listsv);
15925 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
15928 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15930 assert(RExC_parse <= RExC_end);
15932 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
15935 allow_multi_folds = FALSE;
15937 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15940 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
15941 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
15942 int maybe_class = handle_possible_posix(pRExC_state,
15944 ¬_posix_region_end,
15946 TRUE /* checking only */);
15947 if (PASS2 && maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
15948 SAVEFREESV(RExC_rx_sv);
15949 ckWARN4reg(not_posix_region_end,
15950 "POSIX syntax [%c %c] belongs inside character classes%s",
15951 *RExC_parse, *RExC_parse,
15952 (maybe_class == OOB_NAMEDCLASS)
15953 ? ((POSIXCC_NOTYET(*RExC_parse))
15954 ? " (but this one isn't implemented)"
15955 : " (but this one isn't fully valid)")
15958 (void)ReREFCNT_inc(RExC_rx_sv);
15962 /* If the caller wants us to just parse a single element, accomplish this
15963 * by faking the loop ending condition */
15964 if (stop_at_1 && RExC_end > RExC_parse) {
15965 stop_ptr = RExC_parse + 1;
15968 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
15969 if (UCHARAT(RExC_parse) == ']')
15970 goto charclassloop;
15974 if ( posix_warnings
15975 && av_tindex_nomg(posix_warnings) >= 0
15976 && RExC_parse > not_posix_region_end)
15978 /* Warnings about posix class issues are considered tentative until
15979 * we are far enough along in the parse that we can no longer
15980 * change our mind, at which point we either output them or add
15981 * them, if it has so specified, to what gets returned to the
15982 * caller. This is done each time through the loop so that a later
15983 * class won't zap them before they have been dealt with. */
15984 output_or_return_posix_warnings(pRExC_state, posix_warnings,
15985 return_posix_warnings);
15988 if (RExC_parse >= stop_ptr) {
15992 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15994 if (UCHARAT(RExC_parse) == ']') {
16000 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
16001 save_value = value;
16002 save_prevvalue = prevvalue;
16005 rangebegin = RExC_parse;
16007 non_portable_endpoint = 0;
16009 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
16010 value = utf8n_to_uvchr((U8*)RExC_parse,
16011 RExC_end - RExC_parse,
16012 &numlen, UTF8_ALLOW_DEFAULT);
16013 RExC_parse += numlen;
16016 value = UCHARAT(RExC_parse++);
16018 if (value == '[') {
16019 char * posix_class_end;
16020 namedclass = handle_possible_posix(pRExC_state,
16023 do_posix_warnings ? &posix_warnings : NULL,
16024 FALSE /* die if error */);
16025 if (namedclass > OOB_NAMEDCLASS) {
16027 /* If there was an earlier attempt to parse this particular
16028 * posix class, and it failed, it was a false alarm, as this
16029 * successful one proves */
16030 if ( posix_warnings
16031 && av_tindex_nomg(posix_warnings) >= 0
16032 && not_posix_region_end >= RExC_parse
16033 && not_posix_region_end <= posix_class_end)
16035 av_undef(posix_warnings);
16038 RExC_parse = posix_class_end;
16040 else if (namedclass == OOB_NAMEDCLASS) {
16041 not_posix_region_end = posix_class_end;
16044 namedclass = OOB_NAMEDCLASS;
16047 else if ( RExC_parse - 1 > not_posix_region_end
16048 && MAYBE_POSIXCC(value))
16050 (void) handle_possible_posix(
16052 RExC_parse - 1, /* -1 because parse has already been
16054 ¬_posix_region_end,
16055 do_posix_warnings ? &posix_warnings : NULL,
16056 TRUE /* checking only */);
16058 else if (value == '\\') {
16059 /* Is a backslash; get the code point of the char after it */
16061 if (RExC_parse >= RExC_end) {
16062 vFAIL("Unmatched [");
16065 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
16066 value = utf8n_to_uvchr((U8*)RExC_parse,
16067 RExC_end - RExC_parse,
16068 &numlen, UTF8_ALLOW_DEFAULT);
16069 RExC_parse += numlen;
16072 value = UCHARAT(RExC_parse++);
16074 /* Some compilers cannot handle switching on 64-bit integer
16075 * values, therefore value cannot be an UV. Yes, this will
16076 * be a problem later if we want switch on Unicode.
16077 * A similar issue a little bit later when switching on
16078 * namedclass. --jhi */
16080 /* If the \ is escaping white space when white space is being
16081 * skipped, it means that that white space is wanted literally, and
16082 * is already in 'value'. Otherwise, need to translate the escape
16083 * into what it signifies. */
16084 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
16086 case 'w': namedclass = ANYOF_WORDCHAR; break;
16087 case 'W': namedclass = ANYOF_NWORDCHAR; break;
16088 case 's': namedclass = ANYOF_SPACE; break;
16089 case 'S': namedclass = ANYOF_NSPACE; break;
16090 case 'd': namedclass = ANYOF_DIGIT; break;
16091 case 'D': namedclass = ANYOF_NDIGIT; break;
16092 case 'v': namedclass = ANYOF_VERTWS; break;
16093 case 'V': namedclass = ANYOF_NVERTWS; break;
16094 case 'h': namedclass = ANYOF_HORIZWS; break;
16095 case 'H': namedclass = ANYOF_NHORIZWS; break;
16096 case 'N': /* Handle \N{NAME} in class */
16098 const char * const backslash_N_beg = RExC_parse - 2;
16101 if (! grok_bslash_N(pRExC_state,
16102 NULL, /* No regnode */
16103 &value, /* Yes single value */
16104 &cp_count, /* Multiple code pt count */
16110 if (*flagp & NEED_UTF8)
16111 FAIL("panic: grok_bslash_N set NEED_UTF8");
16112 if (*flagp & RESTART_PASS1)
16115 if (cp_count < 0) {
16116 vFAIL("\\N in a character class must be a named character: \\N{...}");
16118 else if (cp_count == 0) {
16120 ckWARNreg(RExC_parse,
16121 "Ignoring zero length \\N{} in character class");
16124 else { /* cp_count > 1 */
16125 if (! RExC_in_multi_char_class) {
16126 if (invert || range || *RExC_parse == '-') {
16129 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
16132 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
16134 break; /* <value> contains the first code
16135 point. Drop out of the switch to
16139 SV * multi_char_N = newSVpvn(backslash_N_beg,
16140 RExC_parse - backslash_N_beg);
16142 = add_multi_match(multi_char_matches,
16147 } /* End of cp_count != 1 */
16149 /* This element should not be processed further in this
16152 value = save_value;
16153 prevvalue = save_prevvalue;
16154 continue; /* Back to top of loop to get next char */
16157 /* Here, is a single code point, and <value> contains it */
16158 unicode_range = TRUE; /* \N{} are Unicode */
16166 /* We will handle any undefined properties ourselves */
16167 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
16168 /* And we actually would prefer to get
16169 * the straight inversion list of the
16170 * swash, since we will be accessing it
16171 * anyway, to save a little time */
16172 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
16174 if (RExC_parse >= RExC_end)
16175 vFAIL2("Empty \\%c", (U8)value);
16176 if (*RExC_parse == '{') {
16177 const U8 c = (U8)value;
16178 e = strchr(RExC_parse, '}');
16181 vFAIL2("Missing right brace on \\%c{}", c);
16185 while (isSPACE(*RExC_parse)) {
16189 if (UCHARAT(RExC_parse) == '^') {
16191 /* toggle. (The rhs xor gets the single bit that
16192 * differs between P and p; the other xor inverts just
16194 value ^= 'P' ^ 'p';
16197 while (isSPACE(*RExC_parse)) {
16202 if (e == RExC_parse)
16203 vFAIL2("Empty \\%c{}", c);
16205 n = e - RExC_parse;
16206 while (isSPACE(*(RExC_parse + n - 1)))
16208 } /* The \p isn't immediately followed by a '{' */
16209 else if (! isALPHA(*RExC_parse)) {
16210 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16211 vFAIL2("Character following \\%c must be '{' or a "
16212 "single-character Unicode property name",
16222 char* base_name; /* name after any packages are stripped */
16223 char* lookup_name = NULL;
16224 const char * const colon_colon = "::";
16226 /* Try to get the definition of the property into
16227 * <invlist>. If /i is in effect, the effective property
16228 * will have its name be <__NAME_i>. The design is
16229 * discussed in commit
16230 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
16231 name = savepv(Perl_form(aTHX_ "%.*s", (int)n, RExC_parse));
16234 lookup_name = savepv(Perl_form(aTHX_ "__%s_i", name));
16236 /* The function call just below that uses this can fail
16237 * to return, leaking memory if we don't do this */
16238 SAVEFREEPV(lookup_name);
16241 /* Look up the property name, and get its swash and
16242 * inversion list, if the property is found */
16243 SvREFCNT_dec(swash); /* Free any left-overs */
16244 swash = _core_swash_init("utf8",
16251 NULL, /* No inversion list */
16254 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
16255 HV* curpkg = (IN_PERL_COMPILETIME)
16257 : CopSTASH(PL_curcop);
16261 if (swash) { /* Got a swash but no inversion list.
16262 Something is likely wrong that will
16263 be sorted-out later */
16264 SvREFCNT_dec_NN(swash);
16268 /* Here didn't find it. It could be a an error (like a
16269 * typo) in specifying a Unicode property, or it could
16270 * be a user-defined property that will be available at
16271 * run-time. The names of these must begin with 'In'
16272 * or 'Is' (after any packages are stripped off). So
16273 * if not one of those, or if we accept only
16274 * compile-time properties, is an error; otherwise add
16275 * it to the list for run-time look up. */
16276 if ((base_name = rninstr(name, name + n,
16277 colon_colon, colon_colon + 2)))
16278 { /* Has ::. We know this must be a user-defined
16281 final_n -= base_name - name;
16290 || base_name[0] != 'I'
16291 || (base_name[1] != 's' && base_name[1] != 'n')
16294 const char * const msg
16296 ? "Illegal user-defined property name"
16297 : "Can't find Unicode property definition";
16298 RExC_parse = e + 1;
16300 /* diag_listed_as: Can't find Unicode property definition "%s" */
16301 vFAIL3utf8f("%s \"%" UTF8f "\"",
16302 msg, UTF8fARG(UTF, n, name));
16305 /* If the property name doesn't already have a package
16306 * name, add the current one to it so that it can be
16307 * referred to outside it. [perl #121777] */
16308 if (! has_pkg && curpkg) {
16309 char* pkgname = HvNAME(curpkg);
16310 if (strNE(pkgname, "main")) {
16311 char* full_name = Perl_form(aTHX_
16315 n = strlen(full_name);
16316 name = savepvn(full_name, n);
16320 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s%" UTF8f "%s\n",
16321 (value == 'p' ? '+' : '!'),
16322 (FOLD) ? "__" : "",
16323 UTF8fARG(UTF, n, name),
16324 (FOLD) ? "_i" : "");
16325 has_user_defined_property = TRUE;
16326 optimizable = FALSE; /* Will have to leave this an
16329 /* We don't know yet what this matches, so have to flag
16331 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
16335 /* Here, did get the swash and its inversion list. If
16336 * the swash is from a user-defined property, then this
16337 * whole character class should be regarded as such */
16338 if (swash_init_flags
16339 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
16341 has_user_defined_property = TRUE;
16344 /* We warn on matching an above-Unicode code point
16345 * if the match would return true, except don't
16346 * warn for \p{All}, which has exactly one element
16348 (_invlist_contains_cp(invlist, 0x110000)
16349 && (! (_invlist_len(invlist) == 1
16350 && *invlist_array(invlist) == 0)))
16356 /* Invert if asking for the complement */
16357 if (value == 'P') {
16358 _invlist_union_complement_2nd(properties,
16362 /* The swash can't be used as-is, because we've
16363 * inverted things; delay removing it to here after
16364 * have copied its invlist above */
16365 SvREFCNT_dec_NN(swash);
16369 _invlist_union(properties, invlist, &properties);
16373 RExC_parse = e + 1;
16374 namedclass = ANYOF_UNIPROP; /* no official name, but it's
16377 /* \p means they want Unicode semantics */
16378 REQUIRE_UNI_RULES(flagp, NULL);
16381 case 'n': value = '\n'; break;
16382 case 'r': value = '\r'; break;
16383 case 't': value = '\t'; break;
16384 case 'f': value = '\f'; break;
16385 case 'b': value = '\b'; break;
16386 case 'e': value = ESC_NATIVE; break;
16387 case 'a': value = '\a'; break;
16389 RExC_parse--; /* function expects to be pointed at the 'o' */
16391 const char* error_msg;
16392 bool valid = grok_bslash_o(&RExC_parse,
16395 PASS2, /* warnings only in
16398 silence_non_portable,
16404 non_portable_endpoint++;
16407 RExC_parse--; /* function expects to be pointed at the 'x' */
16409 const char* error_msg;
16410 bool valid = grok_bslash_x(&RExC_parse,
16413 PASS2, /* Output warnings */
16415 silence_non_portable,
16421 non_portable_endpoint++;
16424 value = grok_bslash_c(*RExC_parse++, PASS2);
16425 non_portable_endpoint++;
16427 case '0': case '1': case '2': case '3': case '4':
16428 case '5': case '6': case '7':
16430 /* Take 1-3 octal digits */
16431 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
16432 numlen = (strict) ? 4 : 3;
16433 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
16434 RExC_parse += numlen;
16437 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16438 vFAIL("Need exactly 3 octal digits");
16440 else if (! SIZE_ONLY /* like \08, \178 */
16442 && RExC_parse < RExC_end
16443 && isDIGIT(*RExC_parse)
16444 && ckWARN(WARN_REGEXP))
16446 SAVEFREESV(RExC_rx_sv);
16447 reg_warn_non_literal_string(
16449 form_short_octal_warning(RExC_parse, numlen));
16450 (void)ReREFCNT_inc(RExC_rx_sv);
16453 non_portable_endpoint++;
16457 /* Allow \_ to not give an error */
16458 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
16460 vFAIL2("Unrecognized escape \\%c in character class",
16464 SAVEFREESV(RExC_rx_sv);
16465 ckWARN2reg(RExC_parse,
16466 "Unrecognized escape \\%c in character class passed through",
16468 (void)ReREFCNT_inc(RExC_rx_sv);
16472 } /* End of switch on char following backslash */
16473 } /* end of handling backslash escape sequences */
16475 /* Here, we have the current token in 'value' */
16477 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
16480 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
16481 * literal, as is the character that began the false range, i.e.
16482 * the 'a' in the examples */
16485 const int w = (RExC_parse >= rangebegin)
16486 ? RExC_parse - rangebegin
16490 "False [] range \"%" UTF8f "\"",
16491 UTF8fARG(UTF, w, rangebegin));
16494 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
16495 ckWARN2reg(RExC_parse,
16496 "False [] range \"%" UTF8f "\"",
16497 UTF8fARG(UTF, w, rangebegin));
16498 (void)ReREFCNT_inc(RExC_rx_sv);
16499 cp_list = add_cp_to_invlist(cp_list, '-');
16500 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
16505 range = 0; /* this was not a true range */
16506 element_count += 2; /* So counts for three values */
16509 classnum = namedclass_to_classnum(namedclass);
16511 if (LOC && namedclass < ANYOF_POSIXL_MAX
16512 #ifndef HAS_ISASCII
16513 && classnum != _CC_ASCII
16516 /* What the Posix classes (like \w, [:space:]) match in locale
16517 * isn't knowable under locale until actual match time. Room
16518 * must be reserved (one time per outer bracketed class) to
16519 * store such classes. The space will contain a bit for each
16520 * named class that is to be matched against. This isn't
16521 * needed for \p{} and pseudo-classes, as they are not affected
16522 * by locale, and hence are dealt with separately */
16523 if (! need_class) {
16526 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16529 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16531 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
16532 ANYOF_POSIXL_ZERO(ret);
16534 /* We can't change this into some other type of node
16535 * (unless this is the only element, in which case there
16536 * are nodes that mean exactly this) as has runtime
16538 optimizable = FALSE;
16541 /* Coverity thinks it is possible for this to be negative; both
16542 * jhi and khw think it's not, but be safer */
16543 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16544 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
16546 /* See if it already matches the complement of this POSIX
16548 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16549 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
16553 posixl_matches_all = TRUE;
16554 break; /* No need to continue. Since it matches both
16555 e.g., \w and \W, it matches everything, and the
16556 bracketed class can be optimized into qr/./s */
16559 /* Add this class to those that should be checked at runtime */
16560 ANYOF_POSIXL_SET(ret, namedclass);
16562 /* The above-Latin1 characters are not subject to locale rules.
16563 * Just add them, in the second pass, to the
16564 * unconditionally-matched list */
16566 SV* scratch_list = NULL;
16568 /* Get the list of the above-Latin1 code points this
16570 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
16571 PL_XPosix_ptrs[classnum],
16573 /* Odd numbers are complements, like
16574 * NDIGIT, NASCII, ... */
16575 namedclass % 2 != 0,
16577 /* Checking if 'cp_list' is NULL first saves an extra
16578 * clone. Its reference count will be decremented at the
16579 * next union, etc, or if this is the only instance, at the
16580 * end of the routine */
16582 cp_list = scratch_list;
16585 _invlist_union(cp_list, scratch_list, &cp_list);
16586 SvREFCNT_dec_NN(scratch_list);
16588 continue; /* Go get next character */
16591 else if (! SIZE_ONLY) {
16593 /* Here, not in pass1 (in that pass we skip calculating the
16594 * contents of this class), and is not /l, or is a POSIX class
16595 * for which /l doesn't matter (or is a Unicode property, which
16596 * is skipped here). */
16597 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
16598 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
16600 /* Here, should be \h, \H, \v, or \V. None of /d, /i
16601 * nor /l make a difference in what these match,
16602 * therefore we just add what they match to cp_list. */
16603 if (classnum != _CC_VERTSPACE) {
16604 assert( namedclass == ANYOF_HORIZWS
16605 || namedclass == ANYOF_NHORIZWS);
16607 /* It turns out that \h is just a synonym for
16609 classnum = _CC_BLANK;
16612 _invlist_union_maybe_complement_2nd(
16614 PL_XPosix_ptrs[classnum],
16615 namedclass % 2 != 0, /* Complement if odd
16616 (NHORIZWS, NVERTWS)
16621 else if ( UNI_SEMANTICS
16622 || classnum == _CC_ASCII
16623 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
16624 || classnum == _CC_XDIGIT)))
16626 /* We usually have to worry about /d and /a affecting what
16627 * POSIX classes match, with special code needed for /d
16628 * because we won't know until runtime what all matches.
16629 * But there is no extra work needed under /u, and
16630 * [:ascii:] is unaffected by /a and /d; and :digit: and
16631 * :xdigit: don't have runtime differences under /d. So we
16632 * can special case these, and avoid some extra work below,
16633 * and at runtime. */
16634 _invlist_union_maybe_complement_2nd(
16636 PL_XPosix_ptrs[classnum],
16637 namedclass % 2 != 0,
16640 else { /* Garden variety class. If is NUPPER, NALPHA, ...
16641 complement and use nposixes */
16642 SV** posixes_ptr = namedclass % 2 == 0
16645 _invlist_union_maybe_complement_2nd(
16647 PL_XPosix_ptrs[classnum],
16648 namedclass % 2 != 0,
16652 } /* end of namedclass \blah */
16654 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16656 /* If 'range' is set, 'value' is the ending of a range--check its
16657 * validity. (If value isn't a single code point in the case of a
16658 * range, we should have figured that out above in the code that
16659 * catches false ranges). Later, we will handle each individual code
16660 * point in the range. If 'range' isn't set, this could be the
16661 * beginning of a range, so check for that by looking ahead to see if
16662 * the next real character to be processed is the range indicator--the
16667 /* For unicode ranges, we have to test that the Unicode as opposed
16668 * to the native values are not decreasing. (Above 255, there is
16669 * no difference between native and Unicode) */
16670 if (unicode_range && prevvalue < 255 && value < 255) {
16671 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
16672 goto backwards_range;
16677 if (prevvalue > value) /* b-a */ {
16682 w = RExC_parse - rangebegin;
16684 "Invalid [] range \"%" UTF8f "\"",
16685 UTF8fARG(UTF, w, rangebegin));
16686 NOT_REACHED; /* NOTREACHED */
16690 prevvalue = value; /* save the beginning of the potential range */
16691 if (! stop_at_1 /* Can't be a range if parsing just one thing */
16692 && *RExC_parse == '-')
16694 char* next_char_ptr = RExC_parse + 1;
16696 /* Get the next real char after the '-' */
16697 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
16699 /* If the '-' is at the end of the class (just before the ']',
16700 * it is a literal minus; otherwise it is a range */
16701 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
16702 RExC_parse = next_char_ptr;
16704 /* a bad range like \w-, [:word:]- ? */
16705 if (namedclass > OOB_NAMEDCLASS) {
16706 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
16707 const int w = RExC_parse >= rangebegin
16708 ? RExC_parse - rangebegin
16711 vFAIL4("False [] range \"%*.*s\"",
16716 "False [] range \"%*.*s\"",
16721 cp_list = add_cp_to_invlist(cp_list, '-');
16725 range = 1; /* yeah, it's a range! */
16726 continue; /* but do it the next time */
16731 if (namedclass > OOB_NAMEDCLASS) {
16735 /* Here, we have a single value this time through the loop, and
16736 * <prevvalue> is the beginning of the range, if any; or <value> if
16739 /* non-Latin1 code point implies unicode semantics. Must be set in
16740 * pass1 so is there for the whole of pass 2 */
16742 REQUIRE_UNI_RULES(flagp, NULL);
16745 /* Ready to process either the single value, or the completed range.
16746 * For single-valued non-inverted ranges, we consider the possibility
16747 * of multi-char folds. (We made a conscious decision to not do this
16748 * for the other cases because it can often lead to non-intuitive
16749 * results. For example, you have the peculiar case that:
16750 * "s s" =~ /^[^\xDF]+$/i => Y
16751 * "ss" =~ /^[^\xDF]+$/i => N
16753 * See [perl #89750] */
16754 if (FOLD && allow_multi_folds && value == prevvalue) {
16755 if (value == LATIN_SMALL_LETTER_SHARP_S
16756 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
16759 /* Here <value> is indeed a multi-char fold. Get what it is */
16761 U8 foldbuf[UTF8_MAXBYTES_CASE];
16764 UV folded = _to_uni_fold_flags(
16768 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
16769 ? FOLD_FLAGS_NOMIX_ASCII
16773 /* Here, <folded> should be the first character of the
16774 * multi-char fold of <value>, with <foldbuf> containing the
16775 * whole thing. But, if this fold is not allowed (because of
16776 * the flags), <fold> will be the same as <value>, and should
16777 * be processed like any other character, so skip the special
16779 if (folded != value) {
16781 /* Skip if we are recursed, currently parsing the class
16782 * again. Otherwise add this character to the list of
16783 * multi-char folds. */
16784 if (! RExC_in_multi_char_class) {
16785 STRLEN cp_count = utf8_length(foldbuf,
16786 foldbuf + foldlen);
16787 SV* multi_fold = sv_2mortal(newSVpvs(""));
16789 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
16792 = add_multi_match(multi_char_matches,
16798 /* This element should not be processed further in this
16801 value = save_value;
16802 prevvalue = save_prevvalue;
16808 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
16811 /* If the range starts above 255, everything is portable and
16812 * likely to be so for any forseeable character set, so don't
16814 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
16815 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
16817 else if (prevvalue != value) {
16819 /* Under strict, ranges that stop and/or end in an ASCII
16820 * printable should have each end point be a portable value
16821 * for it (preferably like 'A', but we don't warn if it is
16822 * a (portable) Unicode name or code point), and the range
16823 * must be be all digits or all letters of the same case.
16824 * Otherwise, the range is non-portable and unclear as to
16825 * what it contains */
16826 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
16827 && ( non_portable_endpoint
16828 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
16829 || (isLOWER_A(prevvalue) && isLOWER_A(value))
16830 || (isUPPER_A(prevvalue) && isUPPER_A(value))
16832 vWARN(RExC_parse, "Ranges of ASCII printables should"
16833 " be some subset of \"0-9\","
16834 " \"A-Z\", or \"a-z\"");
16836 else if (prevvalue >= 0x660) { /* ARABIC_INDIC_DIGIT_ZERO */
16837 SSize_t index_start;
16838 SSize_t index_final;
16840 /* But the nature of Unicode and languages mean we
16841 * can't do the same checks for above-ASCII ranges,
16842 * except in the case of digit ones. These should
16843 * contain only digits from the same group of 10. The
16844 * ASCII case is handled just above. 0x660 is the
16845 * first digit character beyond ASCII. Hence here, the
16846 * range could be a range of digits. First some
16847 * unlikely special cases. Grandfather in that a range
16848 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
16849 * if its starting value is one of the 10 digits prior
16850 * to it. This is because it is an alternate way of
16851 * writing 19D1, and some people may expect it to be in
16852 * that group. But it is bad, because it won't give
16853 * the expected results. In Unicode 5.2 it was
16854 * considered to be in that group (of 11, hence), but
16855 * this was fixed in the next version */
16857 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
16858 goto warn_bad_digit_range;
16860 else if (UNLIKELY( prevvalue >= 0x1D7CE
16861 && value <= 0x1D7FF))
16863 /* This is the only other case currently in Unicode
16864 * where the algorithm below fails. The code
16865 * points just above are the end points of a single
16866 * range containing only decimal digits. It is 5
16867 * different series of 0-9. All other ranges of
16868 * digits currently in Unicode are just a single
16869 * series. (And mktables will notify us if a later
16870 * Unicode version breaks this.)
16872 * If the range being checked is at most 9 long,
16873 * and the digit values represented are in
16874 * numerical order, they are from the same series.
16876 if ( value - prevvalue > 9
16877 || ((( value - 0x1D7CE) % 10)
16878 <= (prevvalue - 0x1D7CE) % 10))
16880 goto warn_bad_digit_range;
16885 /* For all other ranges of digits in Unicode, the
16886 * algorithm is just to check if both end points
16887 * are in the same series, which is the same range.
16889 index_start = _invlist_search(
16890 PL_XPosix_ptrs[_CC_DIGIT],
16893 /* Warn if the range starts and ends with a digit,
16894 * and they are not in the same group of 10. */
16895 if ( index_start >= 0
16896 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
16898 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
16899 value)) != index_start
16900 && index_final >= 0
16901 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
16903 warn_bad_digit_range:
16904 vWARN(RExC_parse, "Ranges of digits should be"
16905 " from the same group of"
16912 if ((! range || prevvalue == value) && non_portable_endpoint) {
16913 if (isPRINT_A(value)) {
16916 if (isBACKSLASHED_PUNCT(value)) {
16917 literal[d++] = '\\';
16919 literal[d++] = (char) value;
16920 literal[d++] = '\0';
16922 vWARN4dep(RExC_parse,
16923 "\"%.*s\" is more clearly written simply as \"%s\". "
16924 "This will be a fatal error in Perl 5.28",
16925 (int) (RExC_parse - rangebegin),
16930 else if isMNEMONIC_CNTRL(value) {
16931 vWARN4dep(RExC_parse,
16932 "\"%.*s\" is more clearly written simply as \"%s\". "
16933 "This will be a fatal error in Perl 5.28",
16934 (int) (RExC_parse - rangebegin),
16936 cntrl_to_mnemonic((U8) value)
16942 /* Deal with this element of the class */
16946 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16949 /* On non-ASCII platforms, for ranges that span all of 0..255, and
16950 * ones that don't require special handling, we can just add the
16951 * range like we do for ASCII platforms */
16952 if ((UNLIKELY(prevvalue == 0) && value >= 255)
16953 || ! (prevvalue < 256
16955 || (! non_portable_endpoint
16956 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
16957 || (isUPPER_A(prevvalue)
16958 && isUPPER_A(value)))))))
16960 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16964 /* Here, requires special handling. This can be because it is
16965 * a range whose code points are considered to be Unicode, and
16966 * so must be individually translated into native, or because
16967 * its a subrange of 'A-Z' or 'a-z' which each aren't
16968 * contiguous in EBCDIC, but we have defined them to include
16969 * only the "expected" upper or lower case ASCII alphabetics.
16970 * Subranges above 255 are the same in native and Unicode, so
16971 * can be added as a range */
16972 U8 start = NATIVE_TO_LATIN1(prevvalue);
16974 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
16975 for (j = start; j <= end; j++) {
16976 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
16979 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16986 range = 0; /* this range (if it was one) is done now */
16987 } /* End of loop through all the text within the brackets */
16990 if ( posix_warnings && av_tindex_nomg(posix_warnings) >= 0) {
16991 output_or_return_posix_warnings(pRExC_state, posix_warnings,
16992 return_posix_warnings);
16995 /* If anything in the class expands to more than one character, we have to
16996 * deal with them by building up a substitute parse string, and recursively
16997 * calling reg() on it, instead of proceeding */
16998 if (multi_char_matches) {
16999 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
17002 char *save_end = RExC_end;
17003 char *save_parse = RExC_parse;
17004 char *save_start = RExC_start;
17005 STRLEN prefix_end = 0; /* We copy the character class after a
17006 prefix supplied here. This is the size
17007 + 1 of that prefix */
17008 bool first_time = TRUE; /* First multi-char occurrence doesn't get
17013 assert(RExC_precomp_adj == 0); /* Only one level of recursion allowed */
17015 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
17016 because too confusing */
17018 sv_catpv(substitute_parse, "(?:");
17022 /* Look at the longest folds first */
17023 for (cp_count = av_tindex_nomg(multi_char_matches);
17028 if (av_exists(multi_char_matches, cp_count)) {
17029 AV** this_array_ptr;
17032 this_array_ptr = (AV**) av_fetch(multi_char_matches,
17034 while ((this_sequence = av_pop(*this_array_ptr)) !=
17037 if (! first_time) {
17038 sv_catpv(substitute_parse, "|");
17040 first_time = FALSE;
17042 sv_catpv(substitute_parse, SvPVX(this_sequence));
17047 /* If the character class contains anything else besides these
17048 * multi-character folds, have to include it in recursive parsing */
17049 if (element_count) {
17050 sv_catpv(substitute_parse, "|[");
17051 prefix_end = SvCUR(substitute_parse);
17052 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
17054 /* Put in a closing ']' only if not going off the end, as otherwise
17055 * we are adding something that really isn't there */
17056 if (RExC_parse < RExC_end) {
17057 sv_catpv(substitute_parse, "]");
17061 sv_catpv(substitute_parse, ")");
17064 /* This is a way to get the parse to skip forward a whole named
17065 * sequence instead of matching the 2nd character when it fails the
17067 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
17071 /* Set up the data structure so that any errors will be properly
17072 * reported. See the comments at the definition of
17073 * REPORT_LOCATION_ARGS for details */
17074 RExC_precomp_adj = orig_parse - RExC_precomp;
17075 RExC_start = RExC_parse = SvPV(substitute_parse, len);
17076 RExC_adjusted_start = RExC_start + prefix_end;
17077 RExC_end = RExC_parse + len;
17078 RExC_in_multi_char_class = 1;
17079 RExC_emit = (regnode *)orig_emit;
17081 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17083 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PASS1|NEED_UTF8);
17085 /* And restore so can parse the rest of the pattern */
17086 RExC_parse = save_parse;
17087 RExC_start = RExC_adjusted_start = save_start;
17088 RExC_precomp_adj = 0;
17089 RExC_end = save_end;
17090 RExC_in_multi_char_class = 0;
17091 SvREFCNT_dec_NN(multi_char_matches);
17095 /* Here, we've gone through the entire class and dealt with multi-char
17096 * folds. We are now in a position that we can do some checks to see if we
17097 * can optimize this ANYOF node into a simpler one, even in Pass 1.
17098 * Currently we only do two checks:
17099 * 1) is in the unlikely event that the user has specified both, eg. \w and
17100 * \W under /l, then the class matches everything. (This optimization
17101 * is done only to make the optimizer code run later work.)
17102 * 2) if the character class contains only a single element (including a
17103 * single range), we see if there is an equivalent node for it.
17104 * Other checks are possible */
17106 && ! ret_invlist /* Can't optimize if returning the constructed
17108 && (UNLIKELY(posixl_matches_all) || element_count == 1))
17113 if (UNLIKELY(posixl_matches_all)) {
17116 else if (namedclass > OOB_NAMEDCLASS) { /* this is a single named
17117 class, like \w or [:digit:]
17120 /* All named classes are mapped into POSIXish nodes, with its FLAG
17121 * argument giving which class it is */
17122 switch ((I32)namedclass) {
17123 case ANYOF_UNIPROP:
17126 /* These don't depend on the charset modifiers. They always
17127 * match under /u rules */
17128 case ANYOF_NHORIZWS:
17129 case ANYOF_HORIZWS:
17130 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
17133 case ANYOF_NVERTWS:
17138 /* The actual POSIXish node for all the rest depends on the
17139 * charset modifier. The ones in the first set depend only on
17140 * ASCII or, if available on this platform, also locale */
17144 op = (LOC) ? POSIXL : POSIXA;
17150 /* The following don't have any matches in the upper Latin1
17151 * range, hence /d is equivalent to /u for them. Making it /u
17152 * saves some branches at runtime */
17156 case ANYOF_NXDIGIT:
17157 if (! DEPENDS_SEMANTICS) {
17158 goto treat_as_default;
17164 /* The following change to CASED under /i */
17170 namedclass = ANYOF_CASED + (namedclass % 2);
17174 /* The rest have more possibilities depending on the charset.
17175 * We take advantage of the enum ordering of the charset
17176 * modifiers to get the exact node type, */
17179 op = POSIXD + get_regex_charset(RExC_flags);
17180 if (op > POSIXA) { /* /aa is same as /a */
17185 /* The odd numbered ones are the complements of the
17186 * next-lower even number one */
17187 if (namedclass % 2 == 1) {
17191 arg = namedclass_to_classnum(namedclass);
17195 else if (value == prevvalue) {
17197 /* Here, the class consists of just a single code point */
17200 if (! LOC && value == '\n') {
17201 op = REG_ANY; /* Optimize [^\n] */
17202 *flagp |= HASWIDTH|SIMPLE;
17206 else if (value < 256 || UTF) {
17208 /* Optimize a single value into an EXACTish node, but not if it
17209 * would require converting the pattern to UTF-8. */
17210 op = compute_EXACTish(pRExC_state);
17212 } /* Otherwise is a range */
17213 else if (! LOC) { /* locale could vary these */
17214 if (prevvalue == '0') {
17215 if (value == '9') {
17220 else if (! FOLD || ASCII_FOLD_RESTRICTED) {
17221 /* We can optimize A-Z or a-z, but not if they could match
17222 * something like the KELVIN SIGN under /i. */
17223 if (prevvalue == 'A') {
17226 && ! non_portable_endpoint
17229 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
17233 else if (prevvalue == 'a') {
17236 && ! non_portable_endpoint
17239 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
17246 /* Here, we have changed <op> away from its initial value iff we found
17247 * an optimization */
17250 /* Throw away this ANYOF regnode, and emit the calculated one,
17251 * which should correspond to the beginning, not current, state of
17253 const char * cur_parse = RExC_parse;
17254 RExC_parse = (char *)orig_parse;
17258 /* To get locale nodes to not use the full ANYOF size would
17259 * require moving the code above that writes the portions
17260 * of it that aren't in other nodes to after this point.
17261 * e.g. ANYOF_POSIXL_SET */
17262 RExC_size = orig_size;
17266 RExC_emit = (regnode *)orig_emit;
17267 if (PL_regkind[op] == POSIXD) {
17268 if (op == POSIXL) {
17269 RExC_contains_locale = 1;
17272 op += NPOSIXD - POSIXD;
17277 ret = reg_node(pRExC_state, op);
17279 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17283 *flagp |= HASWIDTH|SIMPLE;
17285 else if (PL_regkind[op] == EXACT) {
17286 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17287 TRUE /* downgradable to EXACT */
17291 RExC_parse = (char *) cur_parse;
17293 SvREFCNT_dec(posixes);
17294 SvREFCNT_dec(nposixes);
17295 SvREFCNT_dec(simple_posixes);
17296 SvREFCNT_dec(cp_list);
17297 SvREFCNT_dec(cp_foldable_list);
17304 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
17306 /* If folding, we calculate all characters that could fold to or from the
17307 * ones already on the list */
17308 if (cp_foldable_list) {
17310 UV start, end; /* End points of code point ranges */
17312 SV* fold_intersection = NULL;
17315 /* Our calculated list will be for Unicode rules. For locale
17316 * matching, we have to keep a separate list that is consulted at
17317 * runtime only when the locale indicates Unicode rules. For
17318 * non-locale, we just use the general list */
17320 use_list = &only_utf8_locale_list;
17323 use_list = &cp_list;
17326 /* Only the characters in this class that participate in folds need
17327 * be checked. Get the intersection of this class and all the
17328 * possible characters that are foldable. This can quickly narrow
17329 * down a large class */
17330 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
17331 &fold_intersection);
17333 /* The folds for all the Latin1 characters are hard-coded into this
17334 * program, but we have to go out to disk to get the others. */
17335 if (invlist_highest(cp_foldable_list) >= 256) {
17337 /* This is a hash that for a particular fold gives all
17338 * characters that are involved in it */
17339 if (! PL_utf8_foldclosures) {
17340 _load_PL_utf8_foldclosures();
17344 /* Now look at the foldable characters in this class individually */
17345 invlist_iterinit(fold_intersection);
17346 while (invlist_iternext(fold_intersection, &start, &end)) {
17349 /* Look at every character in the range */
17350 for (j = start; j <= end; j++) {
17351 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17357 if (IS_IN_SOME_FOLD_L1(j)) {
17359 /* ASCII is always matched; non-ASCII is matched
17360 * only under Unicode rules (which could happen
17361 * under /l if the locale is a UTF-8 one */
17362 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
17363 *use_list = add_cp_to_invlist(*use_list,
17364 PL_fold_latin1[j]);
17367 has_upper_latin1_only_utf8_matches
17368 = add_cp_to_invlist(
17369 has_upper_latin1_only_utf8_matches,
17370 PL_fold_latin1[j]);
17374 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
17375 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
17377 add_above_Latin1_folds(pRExC_state,
17384 /* Here is an above Latin1 character. We don't have the
17385 * rules hard-coded for it. First, get its fold. This is
17386 * the simple fold, as the multi-character folds have been
17387 * handled earlier and separated out */
17388 _to_uni_fold_flags(j, foldbuf, &foldlen,
17389 (ASCII_FOLD_RESTRICTED)
17390 ? FOLD_FLAGS_NOMIX_ASCII
17393 /* Single character fold of above Latin1. Add everything in
17394 * its fold closure to the list that this node should match.
17395 * The fold closures data structure is a hash with the keys
17396 * being the UTF-8 of every character that is folded to, like
17397 * 'k', and the values each an array of all code points that
17398 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
17399 * Multi-character folds are not included */
17400 if ((listp = hv_fetch(PL_utf8_foldclosures,
17401 (char *) foldbuf, foldlen, FALSE)))
17403 AV* list = (AV*) *listp;
17405 for (k = 0; k <= av_tindex_nomg(list); k++) {
17406 SV** c_p = av_fetch(list, k, FALSE);
17412 /* /aa doesn't allow folds between ASCII and non- */
17413 if ((ASCII_FOLD_RESTRICTED
17414 && (isASCII(c) != isASCII(j))))
17419 /* Folds under /l which cross the 255/256 boundary
17420 * are added to a separate list. (These are valid
17421 * only when the locale is UTF-8.) */
17422 if (c < 256 && LOC) {
17423 *use_list = add_cp_to_invlist(*use_list, c);
17427 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
17429 cp_list = add_cp_to_invlist(cp_list, c);
17432 /* Similarly folds involving non-ascii Latin1
17433 * characters under /d are added to their list */
17434 has_upper_latin1_only_utf8_matches
17435 = add_cp_to_invlist(
17436 has_upper_latin1_only_utf8_matches,
17443 SvREFCNT_dec_NN(fold_intersection);
17446 /* Now that we have finished adding all the folds, there is no reason
17447 * to keep the foldable list separate */
17448 _invlist_union(cp_list, cp_foldable_list, &cp_list);
17449 SvREFCNT_dec_NN(cp_foldable_list);
17452 /* And combine the result (if any) with any inversion lists from posix
17453 * classes. The lists are kept separate up to now because we don't want to
17454 * fold the classes (folding of those is automatically handled by the swash
17455 * fetching code) */
17456 if (simple_posixes) { /* These are the classes known to be unaffected by
17459 _invlist_union(cp_list, simple_posixes, &cp_list);
17460 SvREFCNT_dec_NN(simple_posixes);
17463 cp_list = simple_posixes;
17466 if (posixes || nposixes) {
17468 /* We have to adjust /a and /aa */
17469 if (AT_LEAST_ASCII_RESTRICTED) {
17471 /* Under /a and /aa, nothing above ASCII matches these */
17473 _invlist_intersection(posixes,
17474 PL_XPosix_ptrs[_CC_ASCII],
17478 /* Under /a and /aa, everything above ASCII matches these
17481 _invlist_union_complement_2nd(nposixes,
17482 PL_XPosix_ptrs[_CC_ASCII],
17487 if (! DEPENDS_SEMANTICS) {
17489 /* For everything but /d, we can just add the current 'posixes' and
17490 * 'nposixes' to the main list */
17493 _invlist_union(cp_list, posixes, &cp_list);
17494 SvREFCNT_dec_NN(posixes);
17502 _invlist_union(cp_list, nposixes, &cp_list);
17503 SvREFCNT_dec_NN(nposixes);
17506 cp_list = nposixes;
17511 /* Under /d, things like \w match upper Latin1 characters only if
17512 * the target string is in UTF-8. But things like \W match all the
17513 * upper Latin1 characters if the target string is not in UTF-8.
17515 * Handle the case where there something like \W separately */
17517 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1);
17519 /* A complemented posix class matches all upper Latin1
17520 * characters if not in UTF-8. And it matches just certain
17521 * ones when in UTF-8. That means those certain ones are
17522 * matched regardless, so can just be added to the
17523 * unconditional list */
17525 _invlist_union(cp_list, nposixes, &cp_list);
17526 SvREFCNT_dec_NN(nposixes);
17530 cp_list = nposixes;
17533 /* Likewise for 'posixes' */
17534 _invlist_union(posixes, cp_list, &cp_list);
17536 /* Likewise for anything else in the range that matched only
17538 if (has_upper_latin1_only_utf8_matches) {
17539 _invlist_union(cp_list,
17540 has_upper_latin1_only_utf8_matches,
17542 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17543 has_upper_latin1_only_utf8_matches = NULL;
17546 /* If we don't match all the upper Latin1 characters regardless
17547 * of UTF-8ness, we have to set a flag to match the rest when
17549 _invlist_subtract(only_non_utf8_list, cp_list,
17550 &only_non_utf8_list);
17551 if (_invlist_len(only_non_utf8_list) != 0) {
17552 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17556 /* Here there were no complemented posix classes. That means
17557 * the upper Latin1 characters in 'posixes' match only when the
17558 * target string is in UTF-8. So we have to add them to the
17559 * list of those types of code points, while adding the
17560 * remainder to the unconditional list.
17562 * First calculate what they are */
17563 SV* nonascii_but_latin1_properties = NULL;
17564 _invlist_intersection(posixes, PL_UpperLatin1,
17565 &nonascii_but_latin1_properties);
17567 /* And add them to the final list of such characters. */
17568 _invlist_union(has_upper_latin1_only_utf8_matches,
17569 nonascii_but_latin1_properties,
17570 &has_upper_latin1_only_utf8_matches);
17572 /* Remove them from what now becomes the unconditional list */
17573 _invlist_subtract(posixes, nonascii_but_latin1_properties,
17576 /* And add those unconditional ones to the final list */
17578 _invlist_union(cp_list, posixes, &cp_list);
17579 SvREFCNT_dec_NN(posixes);
17586 SvREFCNT_dec(nonascii_but_latin1_properties);
17588 /* Get rid of any characters that we now know are matched
17589 * unconditionally from the conditional list, which may make
17590 * that list empty */
17591 _invlist_subtract(has_upper_latin1_only_utf8_matches,
17593 &has_upper_latin1_only_utf8_matches);
17594 if (_invlist_len(has_upper_latin1_only_utf8_matches) == 0) {
17595 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17596 has_upper_latin1_only_utf8_matches = NULL;
17602 /* And combine the result (if any) with any inversion list from properties.
17603 * The lists are kept separate up to now so that we can distinguish the two
17604 * in regards to matching above-Unicode. A run-time warning is generated
17605 * if a Unicode property is matched against a non-Unicode code point. But,
17606 * we allow user-defined properties to match anything, without any warning,
17607 * and we also suppress the warning if there is a portion of the character
17608 * class that isn't a Unicode property, and which matches above Unicode, \W
17609 * or [\x{110000}] for example.
17610 * (Note that in this case, unlike the Posix one above, there is no
17611 * <has_upper_latin1_only_utf8_matches>, because having a Unicode property
17612 * forces Unicode semantics */
17616 /* If it matters to the final outcome, see if a non-property
17617 * component of the class matches above Unicode. If so, the
17618 * warning gets suppressed. This is true even if just a single
17619 * such code point is specified, as, though not strictly correct if
17620 * another such code point is matched against, the fact that they
17621 * are using above-Unicode code points indicates they should know
17622 * the issues involved */
17624 warn_super = ! (invert
17625 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
17628 _invlist_union(properties, cp_list, &cp_list);
17629 SvREFCNT_dec_NN(properties);
17632 cp_list = properties;
17637 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17639 /* Because an ANYOF node is the only one that warns, this node
17640 * can't be optimized into something else */
17641 optimizable = FALSE;
17645 /* Here, we have calculated what code points should be in the character
17648 * Now we can see about various optimizations. Fold calculation (which we
17649 * did above) needs to take place before inversion. Otherwise /[^k]/i
17650 * would invert to include K, which under /i would match k, which it
17651 * shouldn't. Therefore we can't invert folded locale now, as it won't be
17652 * folded until runtime */
17654 /* If we didn't do folding, it's because some information isn't available
17655 * until runtime; set the run-time fold flag for these. (We don't have to
17656 * worry about properties folding, as that is taken care of by the swash
17657 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
17658 * locales, or the class matches at least one 0-255 range code point */
17661 /* Some things on the list might be unconditionally included because of
17662 * other components. Remove them, and clean up the list if it goes to
17664 if (only_utf8_locale_list && cp_list) {
17665 _invlist_subtract(only_utf8_locale_list, cp_list,
17666 &only_utf8_locale_list);
17668 if (_invlist_len(only_utf8_locale_list) == 0) {
17669 SvREFCNT_dec_NN(only_utf8_locale_list);
17670 only_utf8_locale_list = NULL;
17673 if (only_utf8_locale_list) {
17676 |ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17678 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
17680 invlist_iterinit(cp_list);
17681 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
17682 ANYOF_FLAGS(ret) |= ANYOFL_FOLD;
17684 invlist_iterfinish(cp_list);
17687 else if ( DEPENDS_SEMANTICS
17688 && ( has_upper_latin1_only_utf8_matches
17689 || (ANYOF_FLAGS(ret) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
17692 optimizable = FALSE;
17696 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
17697 * at compile time. Besides not inverting folded locale now, we can't
17698 * invert if there are things such as \w, which aren't known until runtime
17702 && OP(ret) != ANYOFD
17703 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
17704 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17706 _invlist_invert(cp_list);
17708 /* Any swash can't be used as-is, because we've inverted things */
17710 SvREFCNT_dec_NN(swash);
17714 /* Clear the invert flag since have just done it here */
17721 *ret_invlist = cp_list;
17722 SvREFCNT_dec(swash);
17724 /* Discard the generated node */
17726 RExC_size = orig_size;
17729 RExC_emit = orig_emit;
17734 /* Some character classes are equivalent to other nodes. Such nodes take
17735 * up less room and generally fewer operations to execute than ANYOF nodes.
17736 * Above, we checked for and optimized into some such equivalents for
17737 * certain common classes that are easy to test. Getting to this point in
17738 * the code means that the class didn't get optimized there. Since this
17739 * code is only executed in Pass 2, it is too late to save space--it has
17740 * been allocated in Pass 1, and currently isn't given back. But turning
17741 * things into an EXACTish node can allow the optimizer to join it to any
17742 * adjacent such nodes. And if the class is equivalent to things like /./,
17743 * expensive run-time swashes can be avoided. Now that we have more
17744 * complete information, we can find things necessarily missed by the
17745 * earlier code. Another possible "optimization" that isn't done is that
17746 * something like [Ee] could be changed into an EXACTFU. khw tried this
17747 * and found that the ANYOF is faster, including for code points not in the
17748 * bitmap. This still might make sense to do, provided it got joined with
17749 * an adjacent node(s) to create a longer EXACTFU one. This could be
17750 * accomplished by creating a pseudo ANYOF_EXACTFU node type that the join
17751 * routine would know is joinable. If that didn't happen, the node type
17752 * could then be made a straight ANYOF */
17754 if (optimizable && cp_list && ! invert) {
17756 U8 op = END; /* The optimzation node-type */
17757 int posix_class = -1; /* Illegal value */
17758 const char * cur_parse= RExC_parse;
17760 invlist_iterinit(cp_list);
17761 if (! invlist_iternext(cp_list, &start, &end)) {
17763 /* Here, the list is empty. This happens, for example, when a
17764 * Unicode property that doesn't match anything is the only element
17765 * in the character class (perluniprops.pod notes such properties).
17768 *flagp |= HASWIDTH|SIMPLE;
17770 else if (start == end) { /* The range is a single code point */
17771 if (! invlist_iternext(cp_list, &start, &end)
17773 /* Don't do this optimization if it would require changing
17774 * the pattern to UTF-8 */
17775 && (start < 256 || UTF))
17777 /* Here, the list contains a single code point. Can optimize
17778 * into an EXACTish node */
17789 /* A locale node under folding with one code point can be
17790 * an EXACTFL, as its fold won't be calculated until
17796 /* Here, we are generally folding, but there is only one
17797 * code point to match. If we have to, we use an EXACT
17798 * node, but it would be better for joining with adjacent
17799 * nodes in the optimization pass if we used the same
17800 * EXACTFish node that any such are likely to be. We can
17801 * do this iff the code point doesn't participate in any
17802 * folds. For example, an EXACTF of a colon is the same as
17803 * an EXACT one, since nothing folds to or from a colon. */
17805 if (IS_IN_SOME_FOLD_L1(value)) {
17810 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
17815 /* If we haven't found the node type, above, it means we
17816 * can use the prevailing one */
17818 op = compute_EXACTish(pRExC_state);
17822 } /* End of first range contains just a single code point */
17823 else if (start == 0) {
17824 if (end == UV_MAX) {
17826 *flagp |= HASWIDTH|SIMPLE;
17829 else if (end == '\n' - 1
17830 && invlist_iternext(cp_list, &start, &end)
17831 && start == '\n' + 1 && end == UV_MAX)
17834 *flagp |= HASWIDTH|SIMPLE;
17838 invlist_iterfinish(cp_list);
17841 const UV cp_list_len = _invlist_len(cp_list);
17842 const UV* cp_list_array = invlist_array(cp_list);
17844 /* Here, didn't find an optimization. See if this matches any of
17845 * the POSIX classes. These run slightly faster for above-Unicode
17846 * code points, so don't bother with POSIXA ones nor the 2 that
17847 * have no above-Unicode matches. We can avoid these checks unless
17848 * the ANYOF matches at least as high as the lowest POSIX one
17849 * (which was manually found to be \v. The actual code point may
17850 * increase in later Unicode releases, if a higher code point is
17851 * assigned to be \v, but this code will never break. It would
17852 * just mean we could execute the checks for posix optimizations
17853 * unnecessarily) */
17855 if (cp_list_array[cp_list_len-1] > 0x2029) {
17856 for (posix_class = 0;
17857 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
17861 if (posix_class == _CC_ASCII || posix_class == _CC_CNTRL) {
17864 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
17866 /* Check if matches normal or inverted */
17867 if (_invlistEQ(cp_list,
17868 PL_XPosix_ptrs[posix_class],
17871 op = (try_inverted)
17874 *flagp |= HASWIDTH|SIMPLE;
17884 RExC_parse = (char *)orig_parse;
17885 RExC_emit = (regnode *)orig_emit;
17887 if (regarglen[op]) {
17888 ret = reganode(pRExC_state, op, 0);
17890 ret = reg_node(pRExC_state, op);
17893 RExC_parse = (char *)cur_parse;
17895 if (PL_regkind[op] == EXACT) {
17896 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17897 TRUE /* downgradable to EXACT */
17900 else if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17901 FLAGS(ret) = posix_class;
17904 SvREFCNT_dec_NN(cp_list);
17909 /* Here, <cp_list> contains all the code points we can determine at
17910 * compile time that match under all conditions. Go through it, and
17911 * for things that belong in the bitmap, put them there, and delete from
17912 * <cp_list>. While we are at it, see if everything above 255 is in the
17913 * list, and if so, set a flag to speed up execution */
17915 populate_ANYOF_from_invlist(ret, &cp_list);
17918 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
17921 /* Here, the bitmap has been populated with all the Latin1 code points that
17922 * always match. Can now add to the overall list those that match only
17923 * when the target string is UTF-8 (<has_upper_latin1_only_utf8_matches>).
17925 if (has_upper_latin1_only_utf8_matches) {
17927 _invlist_union(cp_list,
17928 has_upper_latin1_only_utf8_matches,
17930 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17933 cp_list = has_upper_latin1_only_utf8_matches;
17935 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
17938 /* If there is a swash and more than one element, we can't use the swash in
17939 * the optimization below. */
17940 if (swash && element_count > 1) {
17941 SvREFCNT_dec_NN(swash);
17945 /* Note that the optimization of using 'swash' if it is the only thing in
17946 * the class doesn't have us change swash at all, so it can include things
17947 * that are also in the bitmap; otherwise we have purposely deleted that
17948 * duplicate information */
17949 set_ANYOF_arg(pRExC_state, ret, cp_list,
17950 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17952 only_utf8_locale_list,
17953 swash, has_user_defined_property);
17955 *flagp |= HASWIDTH|SIMPLE;
17957 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
17958 RExC_contains_locale = 1;
17964 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
17967 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
17968 regnode* const node,
17970 SV* const runtime_defns,
17971 SV* const only_utf8_locale_list,
17973 const bool has_user_defined_property)
17975 /* Sets the arg field of an ANYOF-type node 'node', using information about
17976 * the node passed-in. If there is nothing outside the node's bitmap, the
17977 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
17978 * the count returned by add_data(), having allocated and stored an array,
17979 * av, that that count references, as follows:
17980 * av[0] stores the character class description in its textual form.
17981 * This is used later (regexec.c:Perl_regclass_swash()) to
17982 * initialize the appropriate swash, and is also useful for dumping
17983 * the regnode. This is set to &PL_sv_undef if the textual
17984 * description is not needed at run-time (as happens if the other
17985 * elements completely define the class)
17986 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
17987 * computed from av[0]. But if no further computation need be done,
17988 * the swash is stored here now (and av[0] is &PL_sv_undef).
17989 * av[2] stores the inversion list of code points that match only if the
17990 * current locale is UTF-8
17991 * av[3] stores the cp_list inversion list for use in addition or instead
17992 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
17993 * (Otherwise everything needed is already in av[0] and av[1])
17994 * av[4] is set if any component of the class is from a user-defined
17995 * property; used only if av[3] exists */
17999 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
18001 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
18002 assert(! (ANYOF_FLAGS(node)
18003 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
18004 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
18007 AV * const av = newAV();
18010 av_store(av, 0, (runtime_defns)
18011 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
18014 av_store(av, 1, swash);
18015 SvREFCNT_dec_NN(cp_list);
18018 av_store(av, 1, &PL_sv_undef);
18020 av_store(av, 3, cp_list);
18021 av_store(av, 4, newSVuv(has_user_defined_property));
18025 if (only_utf8_locale_list) {
18026 av_store(av, 2, only_utf8_locale_list);
18029 av_store(av, 2, &PL_sv_undef);
18032 rv = newRV_noinc(MUTABLE_SV(av));
18033 n = add_data(pRExC_state, STR_WITH_LEN("s"));
18034 RExC_rxi->data->data[n] = (void*)rv;
18039 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
18041 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
18042 const regnode* node,
18045 SV** only_utf8_locale_ptr,
18046 SV** output_invlist)
18049 /* For internal core use only.
18050 * Returns the swash for the input 'node' in the regex 'prog'.
18051 * If <doinit> is 'true', will attempt to create the swash if not already
18053 * If <listsvp> is non-null, will return the printable contents of the
18054 * swash. This can be used to get debugging information even before the
18055 * swash exists, by calling this function with 'doinit' set to false, in
18056 * which case the components that will be used to eventually create the
18057 * swash are returned (in a printable form).
18058 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
18059 * store an inversion list of code points that should match only if the
18060 * execution-time locale is a UTF-8 one.
18061 * If <output_invlist> is not NULL, it is where this routine is to store an
18062 * inversion list of the code points that would be instead returned in
18063 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
18064 * when this parameter is used, is just the non-code point data that
18065 * will go into creating the swash. This currently should be just
18066 * user-defined properties whose definitions were not known at compile
18067 * time. Using this parameter allows for easier manipulation of the
18068 * swash's data by the caller. It is illegal to call this function with
18069 * this parameter set, but not <listsvp>
18071 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
18072 * that, in spite of this function's name, the swash it returns may include
18073 * the bitmap data as well */
18076 SV *si = NULL; /* Input swash initialization string */
18077 SV* invlist = NULL;
18079 RXi_GET_DECL(prog,progi);
18080 const struct reg_data * const data = prog ? progi->data : NULL;
18082 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
18083 assert(! output_invlist || listsvp);
18085 if (data && data->count) {
18086 const U32 n = ARG(node);
18088 if (data->what[n] == 's') {
18089 SV * const rv = MUTABLE_SV(data->data[n]);
18090 AV * const av = MUTABLE_AV(SvRV(rv));
18091 SV **const ary = AvARRAY(av);
18092 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
18094 si = *ary; /* ary[0] = the string to initialize the swash with */
18096 if (av_tindex_nomg(av) >= 2) {
18097 if (only_utf8_locale_ptr
18099 && ary[2] != &PL_sv_undef)
18101 *only_utf8_locale_ptr = ary[2];
18104 assert(only_utf8_locale_ptr);
18105 *only_utf8_locale_ptr = NULL;
18108 /* Elements 3 and 4 are either both present or both absent. [3]
18109 * is any inversion list generated at compile time; [4]
18110 * indicates if that inversion list has any user-defined
18111 * properties in it. */
18112 if (av_tindex_nomg(av) >= 3) {
18114 if (SvUV(ary[4])) {
18115 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
18123 /* Element [1] is reserved for the set-up swash. If already there,
18124 * return it; if not, create it and store it there */
18125 if (ary[1] && SvROK(ary[1])) {
18128 else if (doinit && ((si && si != &PL_sv_undef)
18129 || (invlist && invlist != &PL_sv_undef))) {
18131 sw = _core_swash_init("utf8", /* the utf8 package */
18135 0, /* not from tr/// */
18137 &swash_init_flags);
18138 (void)av_store(av, 1, sw);
18143 /* If requested, return a printable version of what this swash matches */
18145 SV* matches_string = NULL;
18147 /* The swash should be used, if possible, to get the data, as it
18148 * contains the resolved data. But this function can be called at
18149 * compile-time, before everything gets resolved, in which case we
18150 * return the currently best available information, which is the string
18151 * that will eventually be used to do that resolving, 'si' */
18152 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
18153 && (si && si != &PL_sv_undef))
18155 /* Here, we only have 'si' (and possibly some passed-in data in
18156 * 'invlist', which is handled below) If the caller only wants
18157 * 'si', use that. */
18158 if (! output_invlist) {
18159 matches_string = newSVsv(si);
18162 /* But if the caller wants an inversion list of the node, we
18163 * need to parse 'si' and place as much as possible in the
18164 * desired output inversion list, making 'matches_string' only
18165 * contain the currently unresolvable things */
18166 const char *si_string = SvPVX(si);
18167 STRLEN remaining = SvCUR(si);
18171 /* Ignore everything before the first new-line */
18172 while (*si_string != '\n' && remaining > 0) {
18176 assert(remaining > 0);
18181 while (remaining > 0) {
18183 /* The data consists of just strings defining user-defined
18184 * property names, but in prior incarnations, and perhaps
18185 * somehow from pluggable regex engines, it could still
18186 * hold hex code point definitions. Each component of a
18187 * range would be separated by a tab, and each range by a
18188 * new-line. If these are found, instead add them to the
18189 * inversion list */
18190 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
18191 |PERL_SCAN_SILENT_NON_PORTABLE;
18192 STRLEN len = remaining;
18193 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
18195 /* If the hex decode routine found something, it should go
18196 * up to the next \n */
18197 if ( *(si_string + len) == '\n') {
18198 if (count) { /* 2nd code point on line */
18199 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
18202 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
18205 goto prepare_for_next_iteration;
18208 /* If the hex decode was instead for the lower range limit,
18209 * save it, and go parse the upper range limit */
18210 if (*(si_string + len) == '\t') {
18211 assert(count == 0);
18215 prepare_for_next_iteration:
18216 si_string += len + 1;
18217 remaining -= len + 1;
18221 /* Here, didn't find a legal hex number. Just add it from
18222 * here to the next \n */
18225 while (*(si_string + len) != '\n' && remaining > 0) {
18229 if (*(si_string + len) == '\n') {
18233 if (matches_string) {
18234 sv_catpvn(matches_string, si_string, len - 1);
18237 matches_string = newSVpvn(si_string, len - 1);
18240 sv_catpvs(matches_string, " ");
18241 } /* end of loop through the text */
18243 assert(matches_string);
18244 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
18245 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
18247 } /* end of has an 'si' but no swash */
18250 /* If we have a swash in place, its equivalent inversion list was above
18251 * placed into 'invlist'. If not, this variable may contain a stored
18252 * inversion list which is information beyond what is in 'si' */
18255 /* Again, if the caller doesn't want the output inversion list, put
18256 * everything in 'matches-string' */
18257 if (! output_invlist) {
18258 if ( ! matches_string) {
18259 matches_string = newSVpvs("\n");
18261 sv_catsv(matches_string, invlist_contents(invlist,
18262 TRUE /* traditional style */
18265 else if (! *output_invlist) {
18266 *output_invlist = invlist_clone(invlist);
18269 _invlist_union(*output_invlist, invlist, output_invlist);
18273 *listsvp = matches_string;
18278 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
18280 /* reg_skipcomment()
18282 Absorbs an /x style # comment from the input stream,
18283 returning a pointer to the first character beyond the comment, or if the
18284 comment terminates the pattern without anything following it, this returns
18285 one past the final character of the pattern (in other words, RExC_end) and
18286 sets the REG_RUN_ON_COMMENT_SEEN flag.
18288 Note it's the callers responsibility to ensure that we are
18289 actually in /x mode
18293 PERL_STATIC_INLINE char*
18294 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
18296 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
18300 while (p < RExC_end) {
18301 if (*(++p) == '\n') {
18306 /* we ran off the end of the pattern without ending the comment, so we have
18307 * to add an \n when wrapping */
18308 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
18313 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
18315 const bool force_to_xmod
18318 /* If the text at the current parse position '*p' is a '(?#...)' comment,
18319 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
18320 * is /x whitespace, advance '*p' so that on exit it points to the first
18321 * byte past all such white space and comments */
18323 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
18325 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
18327 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
18330 if (RExC_end - (*p) >= 3
18332 && *(*p + 1) == '?'
18333 && *(*p + 2) == '#')
18335 while (*(*p) != ')') {
18336 if ((*p) == RExC_end)
18337 FAIL("Sequence (?#... not terminated");
18345 const char * save_p = *p;
18346 while ((*p) < RExC_end) {
18348 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
18351 else if (*(*p) == '#') {
18352 (*p) = reg_skipcomment(pRExC_state, (*p));
18358 if (*p != save_p) {
18371 Advances the parse position by one byte, unless that byte is the beginning
18372 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
18373 those two cases, the parse position is advanced beyond all such comments and
18376 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
18380 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
18382 PERL_ARGS_ASSERT_NEXTCHAR;
18384 if (RExC_parse < RExC_end) {
18386 || UTF8_IS_INVARIANT(*RExC_parse)
18387 || UTF8_IS_START(*RExC_parse));
18389 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
18391 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
18392 FALSE /* Don't force /x */ );
18397 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
18399 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
18400 * space. In pass1, it aligns and increments RExC_size; in pass2,
18403 regnode * const ret = RExC_emit;
18404 GET_RE_DEBUG_FLAGS_DECL;
18406 PERL_ARGS_ASSERT_REGNODE_GUTS;
18408 assert(extra_size >= regarglen[op]);
18411 SIZE_ALIGN(RExC_size);
18412 RExC_size += 1 + extra_size;
18415 if (RExC_emit >= RExC_emit_bound)
18416 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
18417 op, (void*)RExC_emit, (void*)RExC_emit_bound);
18419 NODE_ALIGN_FILL(ret);
18420 #ifndef RE_TRACK_PATTERN_OFFSETS
18421 PERL_UNUSED_ARG(name);
18423 if (RExC_offsets) { /* MJD */
18425 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
18428 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
18429 ? "Overwriting end of array!\n" : "OK",
18430 (UV)(RExC_emit - RExC_emit_start),
18431 (UV)(RExC_parse - RExC_start),
18432 (UV)RExC_offsets[0]));
18433 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
18440 - reg_node - emit a node
18442 STATIC regnode * /* Location. */
18443 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
18445 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
18447 PERL_ARGS_ASSERT_REG_NODE;
18449 assert(regarglen[op] == 0);
18452 regnode *ptr = ret;
18453 FILL_ADVANCE_NODE(ptr, op);
18460 - reganode - emit a node with an argument
18462 STATIC regnode * /* Location. */
18463 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
18465 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
18467 PERL_ARGS_ASSERT_REGANODE;
18469 assert(regarglen[op] == 1);
18472 regnode *ptr = ret;
18473 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
18480 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
18482 /* emit a node with U32 and I32 arguments */
18484 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
18486 PERL_ARGS_ASSERT_REG2LANODE;
18488 assert(regarglen[op] == 2);
18491 regnode *ptr = ret;
18492 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
18499 - reginsert - insert an operator in front of already-emitted operand
18501 * Means relocating the operand.
18504 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
18509 const int offset = regarglen[(U8)op];
18510 const int size = NODE_STEP_REGNODE + offset;
18511 GET_RE_DEBUG_FLAGS_DECL;
18513 PERL_ARGS_ASSERT_REGINSERT;
18514 PERL_UNUSED_CONTEXT;
18515 PERL_UNUSED_ARG(depth);
18516 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
18517 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
18522 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
18523 studying. If this is wrong then we need to adjust RExC_recurse
18524 below like we do with RExC_open_parens/RExC_close_parens. */
18528 if (RExC_open_parens) {
18530 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
18531 /* remember that RExC_npar is rex->nparens + 1,
18532 * iow it is 1 more than the number of parens seen in
18533 * the pattern so far. */
18534 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
18535 /* note, RExC_open_parens[0] is the start of the
18536 * regex, it can't move. RExC_close_parens[0] is the end
18537 * of the regex, it *can* move. */
18538 if ( paren && RExC_open_parens[paren] >= opnd ) {
18539 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
18540 RExC_open_parens[paren] += size;
18542 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
18544 if ( RExC_close_parens[paren] >= opnd ) {
18545 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
18546 RExC_close_parens[paren] += size;
18548 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
18553 RExC_end_op += size;
18555 while (src > opnd) {
18556 StructCopy(--src, --dst, regnode);
18557 #ifdef RE_TRACK_PATTERN_OFFSETS
18558 if (RExC_offsets) { /* MJD 20010112 */
18560 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
18564 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
18565 ? "Overwriting end of array!\n" : "OK",
18566 (UV)(src - RExC_emit_start),
18567 (UV)(dst - RExC_emit_start),
18568 (UV)RExC_offsets[0]));
18569 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
18570 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
18576 place = opnd; /* Op node, where operand used to be. */
18577 #ifdef RE_TRACK_PATTERN_OFFSETS
18578 if (RExC_offsets) { /* MJD */
18580 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
18584 (UV)(place - RExC_emit_start) > RExC_offsets[0]
18585 ? "Overwriting end of array!\n" : "OK",
18586 (UV)(place - RExC_emit_start),
18587 (UV)(RExC_parse - RExC_start),
18588 (UV)RExC_offsets[0]));
18589 Set_Node_Offset(place, RExC_parse);
18590 Set_Node_Length(place, 1);
18593 src = NEXTOPER(place);
18594 FILL_ADVANCE_NODE(place, op);
18595 Zero(src, offset, regnode);
18599 - regtail - set the next-pointer at the end of a node chain of p to val.
18600 - SEE ALSO: regtail_study
18603 S_regtail(pTHX_ RExC_state_t * pRExC_state,
18604 const regnode * const p,
18605 const regnode * const val,
18609 GET_RE_DEBUG_FLAGS_DECL;
18611 PERL_ARGS_ASSERT_REGTAIL;
18613 PERL_UNUSED_ARG(depth);
18619 /* Find last node. */
18620 scan = (regnode *) p;
18622 regnode * const temp = regnext(scan);
18624 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
18625 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18626 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
18627 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
18628 (temp == NULL ? "->" : ""),
18629 (temp == NULL ? PL_reg_name[OP(val)] : "")
18637 if (reg_off_by_arg[OP(scan)]) {
18638 ARG_SET(scan, val - scan);
18641 NEXT_OFF(scan) = val - scan;
18647 - regtail_study - set the next-pointer at the end of a node chain of p to val.
18648 - Look for optimizable sequences at the same time.
18649 - currently only looks for EXACT chains.
18651 This is experimental code. The idea is to use this routine to perform
18652 in place optimizations on branches and groups as they are constructed,
18653 with the long term intention of removing optimization from study_chunk so
18654 that it is purely analytical.
18656 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
18657 to control which is which.
18660 /* TODO: All four parms should be const */
18663 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
18664 const regnode *val,U32 depth)
18668 #ifdef EXPERIMENTAL_INPLACESCAN
18671 GET_RE_DEBUG_FLAGS_DECL;
18673 PERL_ARGS_ASSERT_REGTAIL_STUDY;
18679 /* Find last node. */
18683 regnode * const temp = regnext(scan);
18684 #ifdef EXPERIMENTAL_INPLACESCAN
18685 if (PL_regkind[OP(scan)] == EXACT) {
18686 bool unfolded_multi_char; /* Unexamined in this routine */
18687 if (join_exact(pRExC_state, scan, &min,
18688 &unfolded_multi_char, 1, val, depth+1))
18693 switch (OP(scan)) {
18697 case EXACTFA_NO_TRIE:
18703 if( exact == PSEUDO )
18705 else if ( exact != OP(scan) )
18714 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
18715 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18716 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
18717 SvPV_nolen_const(RExC_mysv),
18718 REG_NODE_NUM(scan),
18719 PL_reg_name[exact]);
18726 DEBUG_PARSE_MSG("");
18727 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
18728 Perl_re_printf( aTHX_
18729 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
18730 SvPV_nolen_const(RExC_mysv),
18731 (IV)REG_NODE_NUM(val),
18735 if (reg_off_by_arg[OP(scan)]) {
18736 ARG_SET(scan, val - scan);
18739 NEXT_OFF(scan) = val - scan;
18747 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
18752 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
18757 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18759 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
18760 if (flags & (1<<bit)) {
18761 if (!set++ && lead)
18762 Perl_re_printf( aTHX_ "%s",lead);
18763 Perl_re_printf( aTHX_ "%s ",PL_reg_intflags_name[bit]);
18768 Perl_re_printf( aTHX_ "\n");
18770 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18775 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
18781 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18783 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
18784 if (flags & (1<<bit)) {
18785 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
18788 if (!set++ && lead)
18789 Perl_re_printf( aTHX_ "%s",lead);
18790 Perl_re_printf( aTHX_ "%s ",PL_reg_extflags_name[bit]);
18793 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
18794 if (!set++ && lead) {
18795 Perl_re_printf( aTHX_ "%s",lead);
18798 case REGEX_UNICODE_CHARSET:
18799 Perl_re_printf( aTHX_ "UNICODE");
18801 case REGEX_LOCALE_CHARSET:
18802 Perl_re_printf( aTHX_ "LOCALE");
18804 case REGEX_ASCII_RESTRICTED_CHARSET:
18805 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
18807 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
18808 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
18811 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
18817 Perl_re_printf( aTHX_ "\n");
18819 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18825 Perl_regdump(pTHX_ const regexp *r)
18828 SV * const sv = sv_newmortal();
18829 SV *dsv= sv_newmortal();
18830 RXi_GET_DECL(r,ri);
18831 GET_RE_DEBUG_FLAGS_DECL;
18833 PERL_ARGS_ASSERT_REGDUMP;
18835 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
18837 /* Header fields of interest. */
18838 if (r->anchored_substr) {
18839 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
18840 RE_SV_DUMPLEN(r->anchored_substr), 30);
18841 Perl_re_printf( aTHX_
18842 "anchored %s%s at %" IVdf " ",
18843 s, RE_SV_TAIL(r->anchored_substr),
18844 (IV)r->anchored_offset);
18845 } else if (r->anchored_utf8) {
18846 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
18847 RE_SV_DUMPLEN(r->anchored_utf8), 30);
18848 Perl_re_printf( aTHX_
18849 "anchored utf8 %s%s at %" IVdf " ",
18850 s, RE_SV_TAIL(r->anchored_utf8),
18851 (IV)r->anchored_offset);
18853 if (r->float_substr) {
18854 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
18855 RE_SV_DUMPLEN(r->float_substr), 30);
18856 Perl_re_printf( aTHX_
18857 "floating %s%s at %" IVdf "..%" UVuf " ",
18858 s, RE_SV_TAIL(r->float_substr),
18859 (IV)r->float_min_offset, (UV)r->float_max_offset);
18860 } else if (r->float_utf8) {
18861 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
18862 RE_SV_DUMPLEN(r->float_utf8), 30);
18863 Perl_re_printf( aTHX_
18864 "floating utf8 %s%s at %" IVdf "..%" UVuf " ",
18865 s, RE_SV_TAIL(r->float_utf8),
18866 (IV)r->float_min_offset, (UV)r->float_max_offset);
18868 if (r->check_substr || r->check_utf8)
18869 Perl_re_printf( aTHX_
18871 (r->check_substr == r->float_substr
18872 && r->check_utf8 == r->float_utf8
18873 ? "(checking floating" : "(checking anchored"));
18874 if (r->intflags & PREGf_NOSCAN)
18875 Perl_re_printf( aTHX_ " noscan");
18876 if (r->extflags & RXf_CHECK_ALL)
18877 Perl_re_printf( aTHX_ " isall");
18878 if (r->check_substr || r->check_utf8)
18879 Perl_re_printf( aTHX_ ") ");
18881 if (ri->regstclass) {
18882 regprop(r, sv, ri->regstclass, NULL, NULL);
18883 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
18885 if (r->intflags & PREGf_ANCH) {
18886 Perl_re_printf( aTHX_ "anchored");
18887 if (r->intflags & PREGf_ANCH_MBOL)
18888 Perl_re_printf( aTHX_ "(MBOL)");
18889 if (r->intflags & PREGf_ANCH_SBOL)
18890 Perl_re_printf( aTHX_ "(SBOL)");
18891 if (r->intflags & PREGf_ANCH_GPOS)
18892 Perl_re_printf( aTHX_ "(GPOS)");
18893 Perl_re_printf( aTHX_ " ");
18895 if (r->intflags & PREGf_GPOS_SEEN)
18896 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
18897 if (r->intflags & PREGf_SKIP)
18898 Perl_re_printf( aTHX_ "plus ");
18899 if (r->intflags & PREGf_IMPLICIT)
18900 Perl_re_printf( aTHX_ "implicit ");
18901 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
18902 if (r->extflags & RXf_EVAL_SEEN)
18903 Perl_re_printf( aTHX_ "with eval ");
18904 Perl_re_printf( aTHX_ "\n");
18906 regdump_extflags("r->extflags: ",r->extflags);
18907 regdump_intflags("r->intflags: ",r->intflags);
18910 PERL_ARGS_ASSERT_REGDUMP;
18911 PERL_UNUSED_CONTEXT;
18912 PERL_UNUSED_ARG(r);
18913 #endif /* DEBUGGING */
18916 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
18919 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
18920 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
18921 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
18922 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
18923 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
18924 || _CC_VERTSPACE != 15
18925 # error Need to adjust order of anyofs[]
18927 static const char * const anyofs[] = {
18964 - regprop - printable representation of opcode, with run time support
18968 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
18972 RXi_GET_DECL(prog,progi);
18973 GET_RE_DEBUG_FLAGS_DECL;
18975 PERL_ARGS_ASSERT_REGPROP;
18979 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
18980 /* It would be nice to FAIL() here, but this may be called from
18981 regexec.c, and it would be hard to supply pRExC_state. */
18982 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
18983 (int)OP(o), (int)REGNODE_MAX);
18984 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
18986 k = PL_regkind[OP(o)];
18989 sv_catpvs(sv, " ");
18990 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
18991 * is a crude hack but it may be the best for now since
18992 * we have no flag "this EXACTish node was UTF-8"
18994 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
18995 PERL_PV_ESCAPE_UNI_DETECT |
18996 PERL_PV_ESCAPE_NONASCII |
18997 PERL_PV_PRETTY_ELLIPSES |
18998 PERL_PV_PRETTY_LTGT |
18999 PERL_PV_PRETTY_NOCLEAR
19001 } else if (k == TRIE) {
19002 /* print the details of the trie in dumpuntil instead, as
19003 * progi->data isn't available here */
19004 const char op = OP(o);
19005 const U32 n = ARG(o);
19006 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
19007 (reg_ac_data *)progi->data->data[n] :
19009 const reg_trie_data * const trie
19010 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
19012 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
19013 DEBUG_TRIE_COMPILE_r({
19015 sv_catpvs(sv, "(JUMP)");
19016 Perl_sv_catpvf(aTHX_ sv,
19017 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
19018 (UV)trie->startstate,
19019 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
19020 (UV)trie->wordcount,
19023 (UV)TRIE_CHARCOUNT(trie),
19024 (UV)trie->uniquecharcount
19027 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
19028 sv_catpvs(sv, "[");
19029 (void) put_charclass_bitmap_innards(sv,
19030 ((IS_ANYOF_TRIE(op))
19032 : TRIE_BITMAP(trie)),
19038 sv_catpvs(sv, "]");
19040 } else if (k == CURLY) {
19041 U32 lo = ARG1(o), hi = ARG2(o);
19042 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
19043 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
19044 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
19045 if (hi == REG_INFTY)
19046 sv_catpvs(sv, "INFTY");
19048 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
19049 sv_catpvs(sv, "}");
19051 else if (k == WHILEM && o->flags) /* Ordinal/of */
19052 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
19053 else if (k == REF || k == OPEN || k == CLOSE
19054 || k == GROUPP || OP(o)==ACCEPT)
19056 AV *name_list= NULL;
19057 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
19058 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
19059 if ( RXp_PAREN_NAMES(prog) ) {
19060 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19061 } else if ( pRExC_state ) {
19062 name_list= RExC_paren_name_list;
19065 if ( k != REF || (OP(o) < NREF)) {
19066 SV **name= av_fetch(name_list, parno, 0 );
19068 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19071 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
19072 I32 *nums=(I32*)SvPVX(sv_dat);
19073 SV **name= av_fetch(name_list, nums[0], 0 );
19076 for ( n=0; n<SvIVX(sv_dat); n++ ) {
19077 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
19078 (n ? "," : ""), (IV)nums[n]);
19080 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19084 if ( k == REF && reginfo) {
19085 U32 n = ARG(o); /* which paren pair */
19086 I32 ln = prog->offs[n].start;
19087 if (prog->lastparen < n || ln == -1)
19088 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
19089 else if (ln == prog->offs[n].end)
19090 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
19092 const char *s = reginfo->strbeg + ln;
19093 Perl_sv_catpvf(aTHX_ sv, ": ");
19094 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
19095 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
19098 } else if (k == GOSUB) {
19099 AV *name_list= NULL;
19100 if ( RXp_PAREN_NAMES(prog) ) {
19101 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19102 } else if ( pRExC_state ) {
19103 name_list= RExC_paren_name_list;
19106 /* Paren and offset */
19107 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
19108 (int)((o + (int)ARG2L(o)) - progi->program) );
19110 SV **name= av_fetch(name_list, ARG(o), 0 );
19112 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19115 else if (k == LOGICAL)
19116 /* 2: embedded, otherwise 1 */
19117 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
19118 else if (k == ANYOF) {
19119 const U8 flags = ANYOF_FLAGS(o);
19120 bool do_sep = FALSE; /* Do we need to separate various components of
19122 /* Set if there is still an unresolved user-defined property */
19123 SV *unresolved = NULL;
19125 /* Things that are ignored except when the runtime locale is UTF-8 */
19126 SV *only_utf8_locale_invlist = NULL;
19128 /* Code points that don't fit in the bitmap */
19129 SV *nonbitmap_invlist = NULL;
19131 /* And things that aren't in the bitmap, but are small enough to be */
19132 SV* bitmap_range_not_in_bitmap = NULL;
19134 const bool inverted = flags & ANYOF_INVERT;
19136 if (OP(o) == ANYOFL) {
19137 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
19138 sv_catpvs(sv, "{utf8-locale-reqd}");
19140 if (flags & ANYOFL_FOLD) {
19141 sv_catpvs(sv, "{i}");
19145 /* If there is stuff outside the bitmap, get it */
19146 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
19147 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
19149 &only_utf8_locale_invlist,
19150 &nonbitmap_invlist);
19151 /* The non-bitmap data may contain stuff that could fit in the
19152 * bitmap. This could come from a user-defined property being
19153 * finally resolved when this call was done; or much more likely
19154 * because there are matches that require UTF-8 to be valid, and so
19155 * aren't in the bitmap. This is teased apart later */
19156 _invlist_intersection(nonbitmap_invlist,
19158 &bitmap_range_not_in_bitmap);
19159 /* Leave just the things that don't fit into the bitmap */
19160 _invlist_subtract(nonbitmap_invlist,
19162 &nonbitmap_invlist);
19165 /* Obey this flag to add all above-the-bitmap code points */
19166 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
19167 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
19168 NUM_ANYOF_CODE_POINTS,
19172 /* Ready to start outputting. First, the initial left bracket */
19173 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
19175 /* Then all the things that could fit in the bitmap */
19176 do_sep = put_charclass_bitmap_innards(sv,
19178 bitmap_range_not_in_bitmap,
19179 only_utf8_locale_invlist,
19182 /* Can't try inverting for a
19183 * better display if there are
19184 * things that haven't been
19186 unresolved != NULL);
19187 SvREFCNT_dec(bitmap_range_not_in_bitmap);
19189 /* If there are user-defined properties which haven't been defined yet,
19190 * output them. If the result is not to be inverted, it is clearest to
19191 * output them in a separate [] from the bitmap range stuff. If the
19192 * result is to be complemented, we have to show everything in one [],
19193 * as the inversion applies to the whole thing. Use {braces} to
19194 * separate them from anything in the bitmap and anything above the
19198 if (! do_sep) { /* If didn't output anything in the bitmap */
19199 sv_catpvs(sv, "^");
19201 sv_catpvs(sv, "{");
19204 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19206 sv_catsv(sv, unresolved);
19208 sv_catpvs(sv, "}");
19210 do_sep = ! inverted;
19213 /* And, finally, add the above-the-bitmap stuff */
19214 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
19217 /* See if truncation size is overridden */
19218 const STRLEN dump_len = (PL_dump_re_max_len)
19219 ? PL_dump_re_max_len
19222 /* This is output in a separate [] */
19224 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19227 /* And, for easy of understanding, it is shown in the
19228 * uncomplemented form if possible. The one exception being if
19229 * there are unresolved items, where the inversion has to be
19230 * delayed until runtime */
19231 if (inverted && ! unresolved) {
19232 _invlist_invert(nonbitmap_invlist);
19233 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
19236 contents = invlist_contents(nonbitmap_invlist,
19237 FALSE /* output suitable for catsv */
19240 /* If the output is shorter than the permissible maximum, just do it. */
19241 if (SvCUR(contents) <= dump_len) {
19242 sv_catsv(sv, contents);
19245 const char * contents_string = SvPVX(contents);
19246 STRLEN i = dump_len;
19248 /* Otherwise, start at the permissible max and work back to the
19249 * first break possibility */
19250 while (i > 0 && contents_string[i] != ' ') {
19253 if (i == 0) { /* Fail-safe. Use the max if we couldn't
19254 find a legal break */
19258 sv_catpvn(sv, contents_string, i);
19259 sv_catpvs(sv, "...");
19262 SvREFCNT_dec_NN(contents);
19263 SvREFCNT_dec_NN(nonbitmap_invlist);
19266 /* And finally the matching, closing ']' */
19267 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
19269 SvREFCNT_dec(unresolved);
19271 else if (k == POSIXD || k == NPOSIXD) {
19272 U8 index = FLAGS(o) * 2;
19273 if (index < C_ARRAY_LENGTH(anyofs)) {
19274 if (*anyofs[index] != '[') {
19277 sv_catpv(sv, anyofs[index]);
19278 if (*anyofs[index] != '[') {
19283 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
19286 else if (k == BOUND || k == NBOUND) {
19287 /* Must be synced with order of 'bound_type' in regcomp.h */
19288 const char * const bounds[] = {
19289 "", /* Traditional */
19295 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
19296 sv_catpv(sv, bounds[FLAGS(o)]);
19298 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
19299 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
19300 else if (OP(o) == SBOL)
19301 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
19303 /* add on the verb argument if there is one */
19304 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
19305 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
19306 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
19309 PERL_UNUSED_CONTEXT;
19310 PERL_UNUSED_ARG(sv);
19311 PERL_UNUSED_ARG(o);
19312 PERL_UNUSED_ARG(prog);
19313 PERL_UNUSED_ARG(reginfo);
19314 PERL_UNUSED_ARG(pRExC_state);
19315 #endif /* DEBUGGING */
19321 Perl_re_intuit_string(pTHX_ REGEXP * const r)
19322 { /* Assume that RE_INTUIT is set */
19323 struct regexp *const prog = ReANY(r);
19324 GET_RE_DEBUG_FLAGS_DECL;
19326 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
19327 PERL_UNUSED_CONTEXT;
19331 const char * const s = SvPV_nolen_const(RX_UTF8(r)
19332 ? prog->check_utf8 : prog->check_substr);
19334 if (!PL_colorset) reginitcolors();
19335 Perl_re_printf( aTHX_
19336 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
19338 RX_UTF8(r) ? "utf8 " : "",
19339 PL_colors[5],PL_colors[0],
19342 (strlen(s) > 60 ? "..." : ""));
19345 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
19346 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
19352 handles refcounting and freeing the perl core regexp structure. When
19353 it is necessary to actually free the structure the first thing it
19354 does is call the 'free' method of the regexp_engine associated to
19355 the regexp, allowing the handling of the void *pprivate; member
19356 first. (This routine is not overridable by extensions, which is why
19357 the extensions free is called first.)
19359 See regdupe and regdupe_internal if you change anything here.
19361 #ifndef PERL_IN_XSUB_RE
19363 Perl_pregfree(pTHX_ REGEXP *r)
19369 Perl_pregfree2(pTHX_ REGEXP *rx)
19371 struct regexp *const r = ReANY(rx);
19372 GET_RE_DEBUG_FLAGS_DECL;
19374 PERL_ARGS_ASSERT_PREGFREE2;
19376 if (r->mother_re) {
19377 ReREFCNT_dec(r->mother_re);
19379 CALLREGFREE_PVT(rx); /* free the private data */
19380 SvREFCNT_dec(RXp_PAREN_NAMES(r));
19381 Safefree(r->xpv_len_u.xpvlenu_pv);
19384 SvREFCNT_dec(r->anchored_substr);
19385 SvREFCNT_dec(r->anchored_utf8);
19386 SvREFCNT_dec(r->float_substr);
19387 SvREFCNT_dec(r->float_utf8);
19388 Safefree(r->substrs);
19390 RX_MATCH_COPY_FREE(rx);
19391 #ifdef PERL_ANY_COW
19392 SvREFCNT_dec(r->saved_copy);
19395 SvREFCNT_dec(r->qr_anoncv);
19396 if (r->recurse_locinput)
19397 Safefree(r->recurse_locinput);
19398 rx->sv_u.svu_rx = 0;
19403 This is a hacky workaround to the structural issue of match results
19404 being stored in the regexp structure which is in turn stored in
19405 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
19406 could be PL_curpm in multiple contexts, and could require multiple
19407 result sets being associated with the pattern simultaneously, such
19408 as when doing a recursive match with (??{$qr})
19410 The solution is to make a lightweight copy of the regexp structure
19411 when a qr// is returned from the code executed by (??{$qr}) this
19412 lightweight copy doesn't actually own any of its data except for
19413 the starp/end and the actual regexp structure itself.
19419 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
19421 struct regexp *ret;
19422 struct regexp *const r = ReANY(rx);
19423 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
19425 PERL_ARGS_ASSERT_REG_TEMP_COPY;
19428 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
19430 SvOK_off((SV *)ret_x);
19432 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
19433 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
19434 made both spots point to the same regexp body.) */
19435 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
19436 assert(!SvPVX(ret_x));
19437 ret_x->sv_u.svu_rx = temp->sv_any;
19438 temp->sv_any = NULL;
19439 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
19440 SvREFCNT_dec_NN(temp);
19441 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
19442 ing below will not set it. */
19443 SvCUR_set(ret_x, SvCUR(rx));
19446 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
19447 sv_force_normal(sv) is called. */
19449 ret = ReANY(ret_x);
19451 SvFLAGS(ret_x) |= SvUTF8(rx);
19452 /* We share the same string buffer as the original regexp, on which we
19453 hold a reference count, incremented when mother_re is set below.
19454 The string pointer is copied here, being part of the regexp struct.
19456 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
19457 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
19459 const I32 npar = r->nparens+1;
19460 Newx(ret->offs, npar, regexp_paren_pair);
19461 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19464 Newx(ret->substrs, 1, struct reg_substr_data);
19465 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19467 SvREFCNT_inc_void(ret->anchored_substr);
19468 SvREFCNT_inc_void(ret->anchored_utf8);
19469 SvREFCNT_inc_void(ret->float_substr);
19470 SvREFCNT_inc_void(ret->float_utf8);
19472 /* check_substr and check_utf8, if non-NULL, point to either their
19473 anchored or float namesakes, and don't hold a second reference. */
19475 RX_MATCH_COPIED_off(ret_x);
19476 #ifdef PERL_ANY_COW
19477 ret->saved_copy = NULL;
19479 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
19480 SvREFCNT_inc_void(ret->qr_anoncv);
19481 if (r->recurse_locinput)
19482 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19488 /* regfree_internal()
19490 Free the private data in a regexp. This is overloadable by
19491 extensions. Perl takes care of the regexp structure in pregfree(),
19492 this covers the *pprivate pointer which technically perl doesn't
19493 know about, however of course we have to handle the
19494 regexp_internal structure when no extension is in use.
19496 Note this is called before freeing anything in the regexp
19501 Perl_regfree_internal(pTHX_ REGEXP * const rx)
19503 struct regexp *const r = ReANY(rx);
19504 RXi_GET_DECL(r,ri);
19505 GET_RE_DEBUG_FLAGS_DECL;
19507 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
19513 SV *dsv= sv_newmortal();
19514 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
19515 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
19516 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
19517 PL_colors[4],PL_colors[5],s);
19520 #ifdef RE_TRACK_PATTERN_OFFSETS
19522 Safefree(ri->u.offsets); /* 20010421 MJD */
19524 if (ri->code_blocks) {
19525 ri->code_blocks->attached = FALSE;
19526 S_free_codeblocks(aTHX_ ri->code_blocks);
19530 int n = ri->data->count;
19533 /* If you add a ->what type here, update the comment in regcomp.h */
19534 switch (ri->data->what[n]) {
19540 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
19543 Safefree(ri->data->data[n]);
19549 { /* Aho Corasick add-on structure for a trie node.
19550 Used in stclass optimization only */
19552 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
19553 #ifdef USE_ITHREADS
19557 refcount = --aho->refcount;
19560 PerlMemShared_free(aho->states);
19561 PerlMemShared_free(aho->fail);
19562 /* do this last!!!! */
19563 PerlMemShared_free(ri->data->data[n]);
19564 /* we should only ever get called once, so
19565 * assert as much, and also guard the free
19566 * which /might/ happen twice. At the least
19567 * it will make code anlyzers happy and it
19568 * doesn't cost much. - Yves */
19569 assert(ri->regstclass);
19570 if (ri->regstclass) {
19571 PerlMemShared_free(ri->regstclass);
19572 ri->regstclass = 0;
19579 /* trie structure. */
19581 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
19582 #ifdef USE_ITHREADS
19586 refcount = --trie->refcount;
19589 PerlMemShared_free(trie->charmap);
19590 PerlMemShared_free(trie->states);
19591 PerlMemShared_free(trie->trans);
19593 PerlMemShared_free(trie->bitmap);
19595 PerlMemShared_free(trie->jump);
19596 PerlMemShared_free(trie->wordinfo);
19597 /* do this last!!!! */
19598 PerlMemShared_free(ri->data->data[n]);
19603 Perl_croak(aTHX_ "panic: regfree data code '%c'",
19604 ri->data->what[n]);
19607 Safefree(ri->data->what);
19608 Safefree(ri->data);
19614 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
19615 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
19616 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
19619 re_dup_guts - duplicate a regexp.
19621 This routine is expected to clone a given regexp structure. It is only
19622 compiled under USE_ITHREADS.
19624 After all of the core data stored in struct regexp is duplicated
19625 the regexp_engine.dupe method is used to copy any private data
19626 stored in the *pprivate pointer. This allows extensions to handle
19627 any duplication it needs to do.
19629 See pregfree() and regfree_internal() if you change anything here.
19631 #if defined(USE_ITHREADS)
19632 #ifndef PERL_IN_XSUB_RE
19634 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
19638 const struct regexp *r = ReANY(sstr);
19639 struct regexp *ret = ReANY(dstr);
19641 PERL_ARGS_ASSERT_RE_DUP_GUTS;
19643 npar = r->nparens+1;
19644 Newx(ret->offs, npar, regexp_paren_pair);
19645 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19647 if (ret->substrs) {
19648 /* Do it this way to avoid reading from *r after the StructCopy().
19649 That way, if any of the sv_dup_inc()s dislodge *r from the L1
19650 cache, it doesn't matter. */
19651 const bool anchored = r->check_substr
19652 ? r->check_substr == r->anchored_substr
19653 : r->check_utf8 == r->anchored_utf8;
19654 Newx(ret->substrs, 1, struct reg_substr_data);
19655 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19657 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
19658 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
19659 ret->float_substr = sv_dup_inc(ret->float_substr, param);
19660 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
19662 /* check_substr and check_utf8, if non-NULL, point to either their
19663 anchored or float namesakes, and don't hold a second reference. */
19665 if (ret->check_substr) {
19667 assert(r->check_utf8 == r->anchored_utf8);
19668 ret->check_substr = ret->anchored_substr;
19669 ret->check_utf8 = ret->anchored_utf8;
19671 assert(r->check_substr == r->float_substr);
19672 assert(r->check_utf8 == r->float_utf8);
19673 ret->check_substr = ret->float_substr;
19674 ret->check_utf8 = ret->float_utf8;
19676 } else if (ret->check_utf8) {
19678 ret->check_utf8 = ret->anchored_utf8;
19680 ret->check_utf8 = ret->float_utf8;
19685 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
19686 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
19687 if (r->recurse_locinput)
19688 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19691 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
19693 if (RX_MATCH_COPIED(dstr))
19694 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
19696 ret->subbeg = NULL;
19697 #ifdef PERL_ANY_COW
19698 ret->saved_copy = NULL;
19701 /* Whether mother_re be set or no, we need to copy the string. We
19702 cannot refrain from copying it when the storage points directly to
19703 our mother regexp, because that's
19704 1: a buffer in a different thread
19705 2: something we no longer hold a reference on
19706 so we need to copy it locally. */
19707 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
19708 ret->mother_re = NULL;
19710 #endif /* PERL_IN_XSUB_RE */
19715 This is the internal complement to regdupe() which is used to copy
19716 the structure pointed to by the *pprivate pointer in the regexp.
19717 This is the core version of the extension overridable cloning hook.
19718 The regexp structure being duplicated will be copied by perl prior
19719 to this and will be provided as the regexp *r argument, however
19720 with the /old/ structures pprivate pointer value. Thus this routine
19721 may override any copying normally done by perl.
19723 It returns a pointer to the new regexp_internal structure.
19727 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
19730 struct regexp *const r = ReANY(rx);
19731 regexp_internal *reti;
19733 RXi_GET_DECL(r,ri);
19735 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
19739 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
19740 char, regexp_internal);
19741 Copy(ri->program, reti->program, len+1, regnode);
19744 if (ri->code_blocks) {
19746 Newx(reti->code_blocks, 1, struct reg_code_blocks);
19747 Newx(reti->code_blocks->cb, ri->code_blocks->count,
19748 struct reg_code_block);
19749 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
19750 ri->code_blocks->count, struct reg_code_block);
19751 for (n = 0; n < ri->code_blocks->count; n++)
19752 reti->code_blocks->cb[n].src_regex = (REGEXP*)
19753 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
19754 reti->code_blocks->count = ri->code_blocks->count;
19755 reti->code_blocks->attached = TRUE;
19758 reti->code_blocks = NULL;
19760 reti->regstclass = NULL;
19763 struct reg_data *d;
19764 const int count = ri->data->count;
19767 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
19768 char, struct reg_data);
19769 Newx(d->what, count, U8);
19772 for (i = 0; i < count; i++) {
19773 d->what[i] = ri->data->what[i];
19774 switch (d->what[i]) {
19775 /* see also regcomp.h and regfree_internal() */
19776 case 'a': /* actually an AV, but the dup function is identical. */
19780 case 'u': /* actually an HV, but the dup function is identical. */
19781 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
19784 /* This is cheating. */
19785 Newx(d->data[i], 1, regnode_ssc);
19786 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
19787 reti->regstclass = (regnode*)d->data[i];
19790 /* Trie stclasses are readonly and can thus be shared
19791 * without duplication. We free the stclass in pregfree
19792 * when the corresponding reg_ac_data struct is freed.
19794 reti->regstclass= ri->regstclass;
19798 ((reg_trie_data*)ri->data->data[i])->refcount++;
19803 d->data[i] = ri->data->data[i];
19806 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
19807 ri->data->what[i]);
19816 reti->name_list_idx = ri->name_list_idx;
19818 #ifdef RE_TRACK_PATTERN_OFFSETS
19819 if (ri->u.offsets) {
19820 Newx(reti->u.offsets, 2*len+1, U32);
19821 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
19824 SetProgLen(reti,len);
19827 return (void*)reti;
19830 #endif /* USE_ITHREADS */
19832 #ifndef PERL_IN_XSUB_RE
19835 - regnext - dig the "next" pointer out of a node
19838 Perl_regnext(pTHX_ regnode *p)
19845 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
19846 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
19847 (int)OP(p), (int)REGNODE_MAX);
19850 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
19859 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
19862 STRLEN l1 = strlen(pat1);
19863 STRLEN l2 = strlen(pat2);
19866 const char *message;
19868 PERL_ARGS_ASSERT_RE_CROAK2;
19874 Copy(pat1, buf, l1 , char);
19875 Copy(pat2, buf + l1, l2 , char);
19876 buf[l1 + l2] = '\n';
19877 buf[l1 + l2 + 1] = '\0';
19878 va_start(args, pat2);
19879 msv = vmess(buf, &args);
19881 message = SvPV_const(msv,l1);
19884 Copy(message, buf, l1 , char);
19885 /* l1-1 to avoid \n */
19886 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, l1-1, buf));
19889 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
19891 #ifndef PERL_IN_XSUB_RE
19893 Perl_save_re_context(pTHX)
19898 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
19901 const REGEXP * const rx = PM_GETRE(PL_curpm);
19903 nparens = RX_NPARENS(rx);
19906 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
19907 * that PL_curpm will be null, but that utf8.pm and the modules it
19908 * loads will only use $1..$3.
19909 * The t/porting/re_context.t test file checks this assumption.
19914 for (i = 1; i <= nparens; i++) {
19915 char digits[TYPE_CHARS(long)];
19916 const STRLEN len = my_snprintf(digits, sizeof(digits),
19918 GV *const *const gvp
19919 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
19922 GV * const gv = *gvp;
19923 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
19933 S_put_code_point(pTHX_ SV *sv, UV c)
19935 PERL_ARGS_ASSERT_PUT_CODE_POINT;
19938 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
19940 else if (isPRINT(c)) {
19941 const char string = (char) c;
19943 /* We use {phrase} as metanotation in the class, so also escape literal
19945 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
19946 sv_catpvs(sv, "\\");
19947 sv_catpvn(sv, &string, 1);
19949 else if (isMNEMONIC_CNTRL(c)) {
19950 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
19953 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
19957 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
19960 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
19962 /* Appends to 'sv' a displayable version of the range of code points from
19963 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
19964 * that have them, when they occur at the beginning or end of the range.
19965 * It uses hex to output the remaining code points, unless 'allow_literals'
19966 * is true, in which case the printable ASCII ones are output as-is (though
19967 * some of these will be escaped by put_code_point()).
19969 * NOTE: This is designed only for printing ranges of code points that fit
19970 * inside an ANYOF bitmap. Higher code points are simply suppressed
19973 const unsigned int min_range_count = 3;
19975 assert(start <= end);
19977 PERL_ARGS_ASSERT_PUT_RANGE;
19979 while (start <= end) {
19981 const char * format;
19983 if (end - start < min_range_count) {
19985 /* Output chars individually when they occur in short ranges */
19986 for (; start <= end; start++) {
19987 put_code_point(sv, start);
19992 /* If permitted by the input options, and there is a possibility that
19993 * this range contains a printable literal, look to see if there is
19995 if (allow_literals && start <= MAX_PRINT_A) {
19997 /* If the character at the beginning of the range isn't an ASCII
19998 * printable, effectively split the range into two parts:
19999 * 1) the portion before the first such printable,
20001 * and output them separately. */
20002 if (! isPRINT_A(start)) {
20003 UV temp_end = start + 1;
20005 /* There is no point looking beyond the final possible
20006 * printable, in MAX_PRINT_A */
20007 UV max = MIN(end, MAX_PRINT_A);
20009 while (temp_end <= max && ! isPRINT_A(temp_end)) {
20013 /* Here, temp_end points to one beyond the first printable if
20014 * found, or to one beyond 'max' if not. If none found, make
20015 * sure that we use the entire range */
20016 if (temp_end > MAX_PRINT_A) {
20017 temp_end = end + 1;
20020 /* Output the first part of the split range: the part that
20021 * doesn't have printables, with the parameter set to not look
20022 * for literals (otherwise we would infinitely recurse) */
20023 put_range(sv, start, temp_end - 1, FALSE);
20025 /* The 2nd part of the range (if any) starts here. */
20028 /* We do a continue, instead of dropping down, because even if
20029 * the 2nd part is non-empty, it could be so short that we want
20030 * to output it as individual characters, as tested for at the
20031 * top of this loop. */
20035 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
20036 * output a sub-range of just the digits or letters, then process
20037 * the remaining portion as usual. */
20038 if (isALPHANUMERIC_A(start)) {
20039 UV mask = (isDIGIT_A(start))
20044 UV temp_end = start + 1;
20046 /* Find the end of the sub-range that includes just the
20047 * characters in the same class as the first character in it */
20048 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
20053 /* For short ranges, don't duplicate the code above to output
20054 * them; just call recursively */
20055 if (temp_end - start < min_range_count) {
20056 put_range(sv, start, temp_end, FALSE);
20058 else { /* Output as a range */
20059 put_code_point(sv, start);
20060 sv_catpvs(sv, "-");
20061 put_code_point(sv, temp_end);
20063 start = temp_end + 1;
20067 /* We output any other printables as individual characters */
20068 if (isPUNCT_A(start) || isSPACE_A(start)) {
20069 while (start <= end && (isPUNCT_A(start)
20070 || isSPACE_A(start)))
20072 put_code_point(sv, start);
20077 } /* End of looking for literals */
20079 /* Here is not to output as a literal. Some control characters have
20080 * mnemonic names. Split off any of those at the beginning and end of
20081 * the range to print mnemonically. It isn't possible for many of
20082 * these to be in a row, so this won't overwhelm with output */
20084 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
20086 while (isMNEMONIC_CNTRL(start) && start <= end) {
20087 put_code_point(sv, start);
20091 /* If this didn't take care of the whole range ... */
20092 if (start <= end) {
20094 /* Look backwards from the end to find the final non-mnemonic
20097 while (isMNEMONIC_CNTRL(temp_end)) {
20101 /* And separately output the interior range that doesn't start
20102 * or end with mnemonics */
20103 put_range(sv, start, temp_end, FALSE);
20105 /* Then output the mnemonic trailing controls */
20106 start = temp_end + 1;
20107 while (start <= end) {
20108 put_code_point(sv, start);
20115 /* As a final resort, output the range or subrange as hex. */
20117 this_end = (end < NUM_ANYOF_CODE_POINTS)
20119 : NUM_ANYOF_CODE_POINTS - 1;
20120 #if NUM_ANYOF_CODE_POINTS > 256
20121 format = (this_end < 256)
20122 ? "\\x%02" UVXf "-\\x%02" UVXf
20123 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
20125 format = "\\x%02" UVXf "-\\x%02" UVXf;
20127 GCC_DIAG_IGNORE(-Wformat-nonliteral);
20128 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
20135 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
20137 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
20141 bool allow_literals = TRUE;
20143 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
20145 /* Generally, it is more readable if printable characters are output as
20146 * literals, but if a range (nearly) spans all of them, it's best to output
20147 * it as a single range. This code will use a single range if all but 2
20148 * ASCII printables are in it */
20149 invlist_iterinit(invlist);
20150 while (invlist_iternext(invlist, &start, &end)) {
20152 /* If the range starts beyond the final printable, it doesn't have any
20154 if (start > MAX_PRINT_A) {
20158 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
20159 * all but two, the range must start and end no later than 2 from
20161 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
20162 if (end > MAX_PRINT_A) {
20168 if (end - start >= MAX_PRINT_A - ' ' - 2) {
20169 allow_literals = FALSE;
20174 invlist_iterfinish(invlist);
20176 /* Here we have figured things out. Output each range */
20177 invlist_iterinit(invlist);
20178 while (invlist_iternext(invlist, &start, &end)) {
20179 if (start >= NUM_ANYOF_CODE_POINTS) {
20182 put_range(sv, start, end, allow_literals);
20184 invlist_iterfinish(invlist);
20190 S_put_charclass_bitmap_innards_common(pTHX_
20191 SV* invlist, /* The bitmap */
20192 SV* posixes, /* Under /l, things like [:word:], \S */
20193 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
20194 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
20195 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
20196 const bool invert /* Is the result to be inverted? */
20199 /* Create and return an SV containing a displayable version of the bitmap
20200 * and associated information determined by the input parameters. If the
20201 * output would have been only the inversion indicator '^', NULL is instead
20206 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
20209 output = newSVpvs("^");
20212 output = newSVpvs("");
20215 /* First, the code points in the bitmap that are unconditionally there */
20216 put_charclass_bitmap_innards_invlist(output, invlist);
20218 /* Traditionally, these have been placed after the main code points */
20220 sv_catsv(output, posixes);
20223 if (only_utf8 && _invlist_len(only_utf8)) {
20224 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
20225 put_charclass_bitmap_innards_invlist(output, only_utf8);
20228 if (not_utf8 && _invlist_len(not_utf8)) {
20229 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
20230 put_charclass_bitmap_innards_invlist(output, not_utf8);
20233 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
20234 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
20235 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
20237 /* This is the only list in this routine that can legally contain code
20238 * points outside the bitmap range. The call just above to
20239 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
20240 * output them here. There's about a half-dozen possible, and none in
20241 * contiguous ranges longer than 2 */
20242 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20244 SV* above_bitmap = NULL;
20246 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
20248 invlist_iterinit(above_bitmap);
20249 while (invlist_iternext(above_bitmap, &start, &end)) {
20252 for (i = start; i <= end; i++) {
20253 put_code_point(output, i);
20256 invlist_iterfinish(above_bitmap);
20257 SvREFCNT_dec_NN(above_bitmap);
20261 if (invert && SvCUR(output) == 1) {
20269 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
20271 SV *nonbitmap_invlist,
20272 SV *only_utf8_locale_invlist,
20273 const regnode * const node,
20274 const bool force_as_is_display)
20276 /* Appends to 'sv' a displayable version of the innards of the bracketed
20277 * character class defined by the other arguments:
20278 * 'bitmap' points to the bitmap.
20279 * 'nonbitmap_invlist' is an inversion list of the code points that are in
20280 * the bitmap range, but for some reason aren't in the bitmap; NULL if
20281 * none. The reasons for this could be that they require some
20282 * condition such as the target string being or not being in UTF-8
20283 * (under /d), or because they came from a user-defined property that
20284 * was not resolved at the time of the regex compilation (under /u)
20285 * 'only_utf8_locale_invlist' is an inversion list of the code points that
20286 * are valid only if the runtime locale is a UTF-8 one; NULL if none
20287 * 'node' is the regex pattern node. It is needed only when the above two
20288 * parameters are not null, and is passed so that this routine can
20289 * tease apart the various reasons for them.
20290 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
20291 * to invert things to see if that leads to a cleaner display. If
20292 * FALSE, this routine is free to use its judgment about doing this.
20294 * It returns TRUE if there was actually something output. (It may be that
20295 * the bitmap, etc is empty.)
20297 * When called for outputting the bitmap of a non-ANYOF node, just pass the
20298 * bitmap, with the succeeding parameters set to NULL, and the final one to
20302 /* In general, it tries to display the 'cleanest' representation of the
20303 * innards, choosing whether to display them inverted or not, regardless of
20304 * whether the class itself is to be inverted. However, there are some
20305 * cases where it can't try inverting, as what actually matches isn't known
20306 * until runtime, and hence the inversion isn't either. */
20307 bool inverting_allowed = ! force_as_is_display;
20310 STRLEN orig_sv_cur = SvCUR(sv);
20312 SV* invlist; /* Inversion list we accumulate of code points that
20313 are unconditionally matched */
20314 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
20316 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
20318 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
20319 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
20322 SV* as_is_display; /* The output string when we take the inputs
20324 SV* inverted_display; /* The output string when we invert the inputs */
20326 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
20328 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
20330 /* We are biased in favor of displaying things without them being inverted,
20331 * as that is generally easier to understand */
20332 const int bias = 5;
20334 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
20336 /* Start off with whatever code points are passed in. (We clone, so we
20337 * don't change the caller's list) */
20338 if (nonbitmap_invlist) {
20339 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
20340 invlist = invlist_clone(nonbitmap_invlist);
20342 else { /* Worst case size is every other code point is matched */
20343 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
20347 if (OP(node) == ANYOFD) {
20349 /* This flag indicates that the code points below 0x100 in the
20350 * nonbitmap list are precisely the ones that match only when the
20351 * target is UTF-8 (they should all be non-ASCII). */
20352 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
20354 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
20355 _invlist_subtract(invlist, only_utf8, &invlist);
20358 /* And this flag for matching all non-ASCII 0xFF and below */
20359 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
20361 not_utf8 = invlist_clone(PL_UpperLatin1);
20364 else if (OP(node) == ANYOFL) {
20366 /* If either of these flags are set, what matches isn't
20367 * determinable except during execution, so don't know enough here
20369 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
20370 inverting_allowed = FALSE;
20373 /* What the posix classes match also varies at runtime, so these
20374 * will be output symbolically. */
20375 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
20378 posixes = newSVpvs("");
20379 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
20380 if (ANYOF_POSIXL_TEST(node,i)) {
20381 sv_catpv(posixes, anyofs[i]);
20388 /* Accumulate the bit map into the unconditional match list */
20389 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
20390 if (BITMAP_TEST(bitmap, i)) {
20392 for (; i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i); i++) {
20395 invlist = _add_range_to_invlist(invlist, start, i-1);
20399 /* Make sure that the conditional match lists don't have anything in them
20400 * that match unconditionally; otherwise the output is quite confusing.
20401 * This could happen if the code that populates these misses some
20404 _invlist_subtract(only_utf8, invlist, &only_utf8);
20407 _invlist_subtract(not_utf8, invlist, ¬_utf8);
20410 if (only_utf8_locale_invlist) {
20412 /* Since this list is passed in, we have to make a copy before
20414 only_utf8_locale = invlist_clone(only_utf8_locale_invlist);
20416 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
20418 /* And, it can get really weird for us to try outputting an inverted
20419 * form of this list when it has things above the bitmap, so don't even
20421 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20422 inverting_allowed = FALSE;
20426 /* Calculate what the output would be if we take the input as-is */
20427 as_is_display = put_charclass_bitmap_innards_common(invlist,
20434 /* If have to take the output as-is, just do that */
20435 if (! inverting_allowed) {
20436 if (as_is_display) {
20437 sv_catsv(sv, as_is_display);
20438 SvREFCNT_dec_NN(as_is_display);
20441 else { /* But otherwise, create the output again on the inverted input, and
20442 use whichever version is shorter */
20444 int inverted_bias, as_is_bias;
20446 /* We will apply our bias to whichever of the the results doesn't have
20456 inverted_bias = bias;
20459 /* Now invert each of the lists that contribute to the output,
20460 * excluding from the result things outside the possible range */
20462 /* For the unconditional inversion list, we have to add in all the
20463 * conditional code points, so that when inverted, they will be gone
20465 _invlist_union(only_utf8, invlist, &invlist);
20466 _invlist_union(not_utf8, invlist, &invlist);
20467 _invlist_union(only_utf8_locale, invlist, &invlist);
20468 _invlist_invert(invlist);
20469 _invlist_intersection(invlist, PL_InBitmap, &invlist);
20472 _invlist_invert(only_utf8);
20473 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
20475 else if (not_utf8) {
20477 /* If a code point matches iff the target string is not in UTF-8,
20478 * then complementing the result has it not match iff not in UTF-8,
20479 * which is the same thing as matching iff it is UTF-8. */
20480 only_utf8 = not_utf8;
20484 if (only_utf8_locale) {
20485 _invlist_invert(only_utf8_locale);
20486 _invlist_intersection(only_utf8_locale,
20488 &only_utf8_locale);
20491 inverted_display = put_charclass_bitmap_innards_common(
20496 only_utf8_locale, invert);
20498 /* Use the shortest representation, taking into account our bias
20499 * against showing it inverted */
20500 if ( inverted_display
20501 && ( ! as_is_display
20502 || ( SvCUR(inverted_display) + inverted_bias
20503 < SvCUR(as_is_display) + as_is_bias)))
20505 sv_catsv(sv, inverted_display);
20507 else if (as_is_display) {
20508 sv_catsv(sv, as_is_display);
20511 SvREFCNT_dec(as_is_display);
20512 SvREFCNT_dec(inverted_display);
20515 SvREFCNT_dec_NN(invlist);
20516 SvREFCNT_dec(only_utf8);
20517 SvREFCNT_dec(not_utf8);
20518 SvREFCNT_dec(posixes);
20519 SvREFCNT_dec(only_utf8_locale);
20521 return SvCUR(sv) > orig_sv_cur;
20524 #define CLEAR_OPTSTART \
20525 if (optstart) STMT_START { \
20526 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
20527 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
20531 #define DUMPUNTIL(b,e) \
20533 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
20535 STATIC const regnode *
20536 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
20537 const regnode *last, const regnode *plast,
20538 SV* sv, I32 indent, U32 depth)
20540 U8 op = PSEUDO; /* Arbitrary non-END op. */
20541 const regnode *next;
20542 const regnode *optstart= NULL;
20544 RXi_GET_DECL(r,ri);
20545 GET_RE_DEBUG_FLAGS_DECL;
20547 PERL_ARGS_ASSERT_DUMPUNTIL;
20549 #ifdef DEBUG_DUMPUNTIL
20550 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n",indent,node-start,
20551 last ? last-start : 0,plast ? plast-start : 0);
20554 if (plast && plast < last)
20557 while (PL_regkind[op] != END && (!last || node < last)) {
20559 /* While that wasn't END last time... */
20562 if (op == CLOSE || op == WHILEM)
20564 next = regnext((regnode *)node);
20567 if (OP(node) == OPTIMIZED) {
20568 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
20575 regprop(r, sv, node, NULL, NULL);
20576 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
20577 (int)(2*indent + 1), "", SvPVX_const(sv));
20579 if (OP(node) != OPTIMIZED) {
20580 if (next == NULL) /* Next ptr. */
20581 Perl_re_printf( aTHX_ " (0)");
20582 else if (PL_regkind[(U8)op] == BRANCH
20583 && PL_regkind[OP(next)] != BRANCH )
20584 Perl_re_printf( aTHX_ " (FAIL)");
20586 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
20587 Perl_re_printf( aTHX_ "\n");
20591 if (PL_regkind[(U8)op] == BRANCHJ) {
20594 const regnode *nnode = (OP(next) == LONGJMP
20595 ? regnext((regnode *)next)
20597 if (last && nnode > last)
20599 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
20602 else if (PL_regkind[(U8)op] == BRANCH) {
20604 DUMPUNTIL(NEXTOPER(node), next);
20606 else if ( PL_regkind[(U8)op] == TRIE ) {
20607 const regnode *this_trie = node;
20608 const char op = OP(node);
20609 const U32 n = ARG(node);
20610 const reg_ac_data * const ac = op>=AHOCORASICK ?
20611 (reg_ac_data *)ri->data->data[n] :
20613 const reg_trie_data * const trie =
20614 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
20616 AV *const trie_words
20617 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
20619 const regnode *nextbranch= NULL;
20622 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
20623 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
20625 Perl_re_indentf( aTHX_ "%s ",
20628 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
20629 SvCUR(*elem_ptr), 60,
20630 PL_colors[0], PL_colors[1],
20632 ? PERL_PV_ESCAPE_UNI
20634 | PERL_PV_PRETTY_ELLIPSES
20635 | PERL_PV_PRETTY_LTGT
20640 U16 dist= trie->jump[word_idx+1];
20641 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
20642 (UV)((dist ? this_trie + dist : next) - start));
20645 nextbranch= this_trie + trie->jump[0];
20646 DUMPUNTIL(this_trie + dist, nextbranch);
20648 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
20649 nextbranch= regnext((regnode *)nextbranch);
20651 Perl_re_printf( aTHX_ "\n");
20654 if (last && next > last)
20659 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
20660 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
20661 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
20663 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
20665 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
20667 else if ( op == PLUS || op == STAR) {
20668 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
20670 else if (PL_regkind[(U8)op] == ANYOF) {
20671 /* arglen 1 + class block */
20672 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
20673 ? ANYOF_POSIXL_SKIP
20675 node = NEXTOPER(node);
20677 else if (PL_regkind[(U8)op] == EXACT) {
20678 /* Literal string, where present. */
20679 node += NODE_SZ_STR(node) - 1;
20680 node = NEXTOPER(node);
20683 node = NEXTOPER(node);
20684 node += regarglen[(U8)op];
20686 if (op == CURLYX || op == OPEN)
20690 #ifdef DEBUG_DUMPUNTIL
20691 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
20696 #endif /* DEBUGGING */
20699 * ex: set ts=8 sts=4 sw=4 et: