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 + 7; /* allow for "qr''xx\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 ];
7793 * This data structure is set up in study_chunk() and is used
7794 * to calculate the distance between a GOSUB regopcode and
7795 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
7798 * If for some reason someone writes code that optimises
7799 * away a GOSUB opcode then the assert should be changed to
7800 * an if(scan) to guard the ARG2L_SET() - Yves
7803 assert(scan && OP(scan) == GOSUB);
7804 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - scan );
7807 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7808 /* assume we don't need to swap parens around before we match */
7810 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
7811 (unsigned long)RExC_study_chunk_recursed_count);
7815 Perl_re_printf( aTHX_ "Final program:\n");
7818 #ifdef RE_TRACK_PATTERN_OFFSETS
7819 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7820 const STRLEN len = ri->u.offsets[0];
7822 GET_RE_DEBUG_FLAGS_DECL;
7823 Perl_re_printf( aTHX_
7824 "Offsets: [%" UVuf "]\n\t", (UV)ri->u.offsets[0]);
7825 for (i = 1; i <= len; i++) {
7826 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7827 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7828 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7830 Perl_re_printf( aTHX_ "\n");
7835 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7836 * by setting the regexp SV to readonly-only instead. If the
7837 * pattern's been recompiled, the USEDness should remain. */
7838 if (old_re && SvREADONLY(old_re))
7846 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7849 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7851 PERL_UNUSED_ARG(value);
7853 if (flags & RXapif_FETCH) {
7854 return reg_named_buff_fetch(rx, key, flags);
7855 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7856 Perl_croak_no_modify();
7858 } else if (flags & RXapif_EXISTS) {
7859 return reg_named_buff_exists(rx, key, flags)
7862 } else if (flags & RXapif_REGNAMES) {
7863 return reg_named_buff_all(rx, flags);
7864 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7865 return reg_named_buff_scalar(rx, flags);
7867 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7873 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7876 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7877 PERL_UNUSED_ARG(lastkey);
7879 if (flags & RXapif_FIRSTKEY)
7880 return reg_named_buff_firstkey(rx, flags);
7881 else if (flags & RXapif_NEXTKEY)
7882 return reg_named_buff_nextkey(rx, flags);
7884 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7891 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7894 AV *retarray = NULL;
7896 struct regexp *const rx = ReANY(r);
7898 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7900 if (flags & RXapif_ALL)
7903 if (rx && RXp_PAREN_NAMES(rx)) {
7904 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7907 SV* sv_dat=HeVAL(he_str);
7908 I32 *nums=(I32*)SvPVX(sv_dat);
7909 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7910 if ((I32)(rx->nparens) >= nums[i]
7911 && rx->offs[nums[i]].start != -1
7912 && rx->offs[nums[i]].end != -1)
7915 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7920 ret = newSVsv(&PL_sv_undef);
7923 av_push(retarray, ret);
7926 return newRV_noinc(MUTABLE_SV(retarray));
7933 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7936 struct regexp *const rx = ReANY(r);
7938 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7940 if (rx && RXp_PAREN_NAMES(rx)) {
7941 if (flags & RXapif_ALL) {
7942 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7944 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7946 SvREFCNT_dec_NN(sv);
7958 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7960 struct regexp *const rx = ReANY(r);
7962 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7964 if ( rx && RXp_PAREN_NAMES(rx) ) {
7965 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7967 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7974 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7976 struct regexp *const rx = ReANY(r);
7977 GET_RE_DEBUG_FLAGS_DECL;
7979 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7981 if (rx && RXp_PAREN_NAMES(rx)) {
7982 HV *hv = RXp_PAREN_NAMES(rx);
7984 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7987 SV* sv_dat = HeVAL(temphe);
7988 I32 *nums = (I32*)SvPVX(sv_dat);
7989 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7990 if ((I32)(rx->lastparen) >= nums[i] &&
7991 rx->offs[nums[i]].start != -1 &&
7992 rx->offs[nums[i]].end != -1)
7998 if (parno || flags & RXapif_ALL) {
7999 return newSVhek(HeKEY_hek(temphe));
8007 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8012 struct regexp *const rx = ReANY(r);
8014 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8016 if (rx && RXp_PAREN_NAMES(rx)) {
8017 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8018 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8019 } else if (flags & RXapif_ONE) {
8020 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8021 av = MUTABLE_AV(SvRV(ret));
8022 length = av_tindex(av);
8023 SvREFCNT_dec_NN(ret);
8024 return newSViv(length + 1);
8026 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8031 return &PL_sv_undef;
8035 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8037 struct regexp *const rx = ReANY(r);
8040 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8042 if (rx && RXp_PAREN_NAMES(rx)) {
8043 HV *hv= RXp_PAREN_NAMES(rx);
8045 (void)hv_iterinit(hv);
8046 while ( (temphe = hv_iternext_flags(hv,0)) ) {
8049 SV* sv_dat = HeVAL(temphe);
8050 I32 *nums = (I32*)SvPVX(sv_dat);
8051 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8052 if ((I32)(rx->lastparen) >= nums[i] &&
8053 rx->offs[nums[i]].start != -1 &&
8054 rx->offs[nums[i]].end != -1)
8060 if (parno || flags & RXapif_ALL) {
8061 av_push(av, newSVhek(HeKEY_hek(temphe)));
8066 return newRV_noinc(MUTABLE_SV(av));
8070 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8073 struct regexp *const rx = ReANY(r);
8079 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8081 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8082 || n == RX_BUFF_IDX_CARET_FULLMATCH
8083 || n == RX_BUFF_IDX_CARET_POSTMATCH
8086 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8088 /* on something like
8091 * the KEEPCOPY is set on the PMOP rather than the regex */
8092 if (PL_curpm && r == PM_GETRE(PL_curpm))
8093 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8102 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8103 /* no need to distinguish between them any more */
8104 n = RX_BUFF_IDX_FULLMATCH;
8106 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8107 && rx->offs[0].start != -1)
8109 /* $`, ${^PREMATCH} */
8110 i = rx->offs[0].start;
8114 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8115 && rx->offs[0].end != -1)
8117 /* $', ${^POSTMATCH} */
8118 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8119 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8122 if ( 0 <= n && n <= (I32)rx->nparens &&
8123 (s1 = rx->offs[n].start) != -1 &&
8124 (t1 = rx->offs[n].end) != -1)
8126 /* $&, ${^MATCH}, $1 ... */
8128 s = rx->subbeg + s1 - rx->suboffset;
8133 assert(s >= rx->subbeg);
8134 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8136 #ifdef NO_TAINT_SUPPORT
8137 sv_setpvn(sv, s, i);
8139 const int oldtainted = TAINT_get;
8141 sv_setpvn(sv, s, i);
8142 TAINT_set(oldtainted);
8144 if (RXp_MATCH_UTF8(rx))
8149 if (RXp_MATCH_TAINTED(rx)) {
8150 if (SvTYPE(sv) >= SVt_PVMG) {
8151 MAGIC* const mg = SvMAGIC(sv);
8154 SvMAGIC_set(sv, mg->mg_moremagic);
8156 if ((mgt = SvMAGIC(sv))) {
8157 mg->mg_moremagic = mgt;
8158 SvMAGIC_set(sv, mg);
8175 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8176 SV const * const value)
8178 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8180 PERL_UNUSED_ARG(rx);
8181 PERL_UNUSED_ARG(paren);
8182 PERL_UNUSED_ARG(value);
8185 Perl_croak_no_modify();
8189 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8192 struct regexp *const rx = ReANY(r);
8196 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8198 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8199 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8200 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8203 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8205 /* on something like
8208 * the KEEPCOPY is set on the PMOP rather than the regex */
8209 if (PL_curpm && r == PM_GETRE(PL_curpm))
8210 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8216 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8218 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8219 case RX_BUFF_IDX_PREMATCH: /* $` */
8220 if (rx->offs[0].start != -1) {
8221 i = rx->offs[0].start;
8230 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8231 case RX_BUFF_IDX_POSTMATCH: /* $' */
8232 if (rx->offs[0].end != -1) {
8233 i = rx->sublen - rx->offs[0].end;
8235 s1 = rx->offs[0].end;
8242 default: /* $& / ${^MATCH}, $1, $2, ... */
8243 if (paren <= (I32)rx->nparens &&
8244 (s1 = rx->offs[paren].start) != -1 &&
8245 (t1 = rx->offs[paren].end) != -1)
8251 if (ckWARN(WARN_UNINITIALIZED))
8252 report_uninit((const SV *)sv);
8257 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8258 const char * const s = rx->subbeg - rx->suboffset + s1;
8263 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8270 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8272 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8273 PERL_UNUSED_ARG(rx);
8277 return newSVpvs("Regexp");
8280 /* Scans the name of a named buffer from the pattern.
8281 * If flags is REG_RSN_RETURN_NULL returns null.
8282 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8283 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8284 * to the parsed name as looked up in the RExC_paren_names hash.
8285 * If there is an error throws a vFAIL().. type exception.
8288 #define REG_RSN_RETURN_NULL 0
8289 #define REG_RSN_RETURN_NAME 1
8290 #define REG_RSN_RETURN_DATA 2
8293 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8295 char *name_start = RExC_parse;
8297 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8299 assert (RExC_parse <= RExC_end);
8300 if (RExC_parse == RExC_end) NOOP;
8301 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8302 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8303 * using do...while */
8306 RExC_parse += UTF8SKIP(RExC_parse);
8307 } while ( RExC_parse < RExC_end
8308 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8312 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8314 RExC_parse++; /* so the <- from the vFAIL is after the offending
8316 vFAIL("Group name must start with a non-digit word character");
8320 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8321 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8322 if ( flags == REG_RSN_RETURN_NAME)
8324 else if (flags==REG_RSN_RETURN_DATA) {
8327 if ( ! sv_name ) /* should not happen*/
8328 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8329 if (RExC_paren_names)
8330 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8332 sv_dat = HeVAL(he_str);
8334 vFAIL("Reference to nonexistent named group");
8338 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8339 (unsigned long) flags);
8341 NOT_REACHED; /* NOTREACHED */
8346 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8348 if (RExC_lastparse!=RExC_parse) { \
8349 Perl_re_printf( aTHX_ "%s", \
8350 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8351 RExC_end - RExC_parse, 16, \
8353 PERL_PV_ESCAPE_UNI_DETECT | \
8354 PERL_PV_PRETTY_ELLIPSES | \
8355 PERL_PV_PRETTY_LTGT | \
8356 PERL_PV_ESCAPE_RE | \
8357 PERL_PV_PRETTY_EXACTSIZE \
8361 Perl_re_printf( aTHX_ "%16s",""); \
8364 num = RExC_size + 1; \
8366 num=REG_NODE_NUM(RExC_emit); \
8367 if (RExC_lastnum!=num) \
8368 Perl_re_printf( aTHX_ "|%4d",num); \
8370 Perl_re_printf( aTHX_ "|%4s",""); \
8371 Perl_re_printf( aTHX_ "|%*s%-4s", \
8372 (int)((depth*2)), "", \
8376 RExC_lastparse=RExC_parse; \
8381 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8382 DEBUG_PARSE_MSG((funcname)); \
8383 Perl_re_printf( aTHX_ "%4s","\n"); \
8385 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8386 DEBUG_PARSE_MSG((funcname)); \
8387 Perl_re_printf( aTHX_ fmt "\n",args); \
8390 /* This section of code defines the inversion list object and its methods. The
8391 * interfaces are highly subject to change, so as much as possible is static to
8392 * this file. An inversion list is here implemented as a malloc'd C UV array
8393 * as an SVt_INVLIST scalar.
8395 * An inversion list for Unicode is an array of code points, sorted by ordinal
8396 * number. Each element gives the code point that begins a range that extends
8397 * up-to but not including the code point given by the next element. The final
8398 * element gives the first code point of a range that extends to the platform's
8399 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8400 * ...) give ranges whose code points are all in the inversion list. We say
8401 * that those ranges are in the set. The odd-numbered elements give ranges
8402 * whose code points are not in the inversion list, and hence not in the set.
8403 * Thus, element [0] is the first code point in the list. Element [1]
8404 * is the first code point beyond that not in the list; and element [2] is the
8405 * first code point beyond that that is in the list. In other words, the first
8406 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8407 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8408 * all code points in that range are not in the inversion list. The third
8409 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8410 * list, and so forth. Thus every element whose index is divisible by two
8411 * gives the beginning of a range that is in the list, and every element whose
8412 * index is not divisible by two gives the beginning of a range not in the
8413 * list. If the final element's index is divisible by two, the inversion list
8414 * extends to the platform's infinity; otherwise the highest code point in the
8415 * inversion list is the contents of that element minus 1.
8417 * A range that contains just a single code point N will look like
8419 * invlist[i+1] == N+1
8421 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8422 * impossible to represent, so element [i+1] is omitted. The single element
8424 * invlist[0] == UV_MAX
8425 * contains just UV_MAX, but is interpreted as matching to infinity.
8427 * Taking the complement (inverting) an inversion list is quite simple, if the
8428 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8429 * This implementation reserves an element at the beginning of each inversion
8430 * list to always contain 0; there is an additional flag in the header which
8431 * indicates if the list begins at the 0, or is offset to begin at the next
8432 * element. This means that the inversion list can be inverted without any
8433 * copying; just flip the flag.
8435 * More about inversion lists can be found in "Unicode Demystified"
8436 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8438 * The inversion list data structure is currently implemented as an SV pointing
8439 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8440 * array of UV whose memory management is automatically handled by the existing
8441 * facilities for SV's.
8443 * Some of the methods should always be private to the implementation, and some
8444 * should eventually be made public */
8446 /* The header definitions are in F<invlist_inline.h> */
8448 #ifndef PERL_IN_XSUB_RE
8450 PERL_STATIC_INLINE UV*
8451 S__invlist_array_init(SV* const invlist, const bool will_have_0)
8453 /* Returns a pointer to the first element in the inversion list's array.
8454 * This is called upon initialization of an inversion list. Where the
8455 * array begins depends on whether the list has the code point U+0000 in it
8456 * or not. The other parameter tells it whether the code that follows this
8457 * call is about to put a 0 in the inversion list or not. The first
8458 * element is either the element reserved for 0, if TRUE, or the element
8459 * after it, if FALSE */
8461 bool* offset = get_invlist_offset_addr(invlist);
8462 UV* zero_addr = (UV *) SvPVX(invlist);
8464 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8467 assert(! _invlist_len(invlist));
8471 /* 1^1 = 0; 1^0 = 1 */
8472 *offset = 1 ^ will_have_0;
8473 return zero_addr + *offset;
8478 PERL_STATIC_INLINE void
8479 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8481 /* Sets the current number of elements stored in the inversion list.
8482 * Updates SvCUR correspondingly */
8483 PERL_UNUSED_CONTEXT;
8484 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8486 assert(SvTYPE(invlist) == SVt_INVLIST);
8491 : TO_INTERNAL_SIZE(len + offset));
8492 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8495 #ifndef PERL_IN_XSUB_RE
8498 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
8500 /* Replaces the inversion list in 'dest' with the one from 'src'. It
8501 * steals the list from 'src', so 'src' is made to have a NULL list. This
8502 * is similar to what SvSetMagicSV() would do, if it were implemented on
8503 * inversion lists, though this routine avoids a copy */
8505 const UV src_len = _invlist_len(src);
8506 const bool src_offset = *get_invlist_offset_addr(src);
8507 const STRLEN src_byte_len = SvLEN(src);
8508 char * array = SvPVX(src);
8510 const int oldtainted = TAINT_get;
8512 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
8514 assert(SvTYPE(src) == SVt_INVLIST);
8515 assert(SvTYPE(dest) == SVt_INVLIST);
8516 assert(! invlist_is_iterating(src));
8517 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
8519 /* Make sure it ends in the right place with a NUL, as our inversion list
8520 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
8522 array[src_byte_len - 1] = '\0';
8524 TAINT_NOT; /* Otherwise it breaks */
8525 sv_usepvn_flags(dest,
8529 /* This flag is documented to cause a copy to be avoided */
8530 SV_HAS_TRAILING_NUL);
8531 TAINT_set(oldtainted);
8536 /* Finish up copying over the other fields in an inversion list */
8537 *get_invlist_offset_addr(dest) = src_offset;
8538 invlist_set_len(dest, src_len, src_offset);
8539 *get_invlist_previous_index_addr(dest) = 0;
8540 invlist_iterfinish(dest);
8543 PERL_STATIC_INLINE IV*
8544 S_get_invlist_previous_index_addr(SV* invlist)
8546 /* Return the address of the IV that is reserved to hold the cached index
8548 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8550 assert(SvTYPE(invlist) == SVt_INVLIST);
8552 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8555 PERL_STATIC_INLINE IV
8556 S_invlist_previous_index(SV* const invlist)
8558 /* Returns cached index of previous search */
8560 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8562 return *get_invlist_previous_index_addr(invlist);
8565 PERL_STATIC_INLINE void
8566 S_invlist_set_previous_index(SV* const invlist, const IV index)
8568 /* Caches <index> for later retrieval */
8570 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8572 assert(index == 0 || index < (int) _invlist_len(invlist));
8574 *get_invlist_previous_index_addr(invlist) = index;
8577 PERL_STATIC_INLINE void
8578 S_invlist_trim(SV* invlist)
8580 /* Free the not currently-being-used space in an inversion list */
8582 /* But don't free up the space needed for the 0 UV that is always at the
8583 * beginning of the list, nor the trailing NUL */
8584 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
8586 PERL_ARGS_ASSERT_INVLIST_TRIM;
8588 assert(SvTYPE(invlist) == SVt_INVLIST);
8590 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
8593 PERL_STATIC_INLINE void
8594 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
8596 PERL_ARGS_ASSERT_INVLIST_CLEAR;
8598 assert(SvTYPE(invlist) == SVt_INVLIST);
8600 invlist_set_len(invlist, 0, 0);
8601 invlist_trim(invlist);
8604 #endif /* ifndef PERL_IN_XSUB_RE */
8606 PERL_STATIC_INLINE bool
8607 S_invlist_is_iterating(SV* const invlist)
8609 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8611 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8614 #ifndef PERL_IN_XSUB_RE
8616 PERL_STATIC_INLINE UV
8617 S_invlist_max(SV* const invlist)
8619 /* Returns the maximum number of elements storable in the inversion list's
8620 * array, without having to realloc() */
8622 PERL_ARGS_ASSERT_INVLIST_MAX;
8624 assert(SvTYPE(invlist) == SVt_INVLIST);
8626 /* Assumes worst case, in which the 0 element is not counted in the
8627 * inversion list, so subtracts 1 for that */
8628 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8629 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8630 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8633 Perl__new_invlist(pTHX_ IV initial_size)
8636 /* Return a pointer to a newly constructed inversion list, with enough
8637 * space to store 'initial_size' elements. If that number is negative, a
8638 * system default is used instead */
8642 if (initial_size < 0) {
8646 /* Allocate the initial space */
8647 new_list = newSV_type(SVt_INVLIST);
8649 /* First 1 is in case the zero element isn't in the list; second 1 is for
8651 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8652 invlist_set_len(new_list, 0, 0);
8654 /* Force iterinit() to be used to get iteration to work */
8655 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8657 *get_invlist_previous_index_addr(new_list) = 0;
8663 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8665 /* Return a pointer to a newly constructed inversion list, initialized to
8666 * point to <list>, which has to be in the exact correct inversion list
8667 * form, including internal fields. Thus this is a dangerous routine that
8668 * should not be used in the wrong hands. The passed in 'list' contains
8669 * several header fields at the beginning that are not part of the
8670 * inversion list body proper */
8672 const STRLEN length = (STRLEN) list[0];
8673 const UV version_id = list[1];
8674 const bool offset = cBOOL(list[2]);
8675 #define HEADER_LENGTH 3
8676 /* If any of the above changes in any way, you must change HEADER_LENGTH
8677 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8678 * perl -E 'say int(rand 2**31-1)'
8680 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8681 data structure type, so that one being
8682 passed in can be validated to be an
8683 inversion list of the correct vintage.
8686 SV* invlist = newSV_type(SVt_INVLIST);
8688 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8690 if (version_id != INVLIST_VERSION_ID) {
8691 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8694 /* The generated array passed in includes header elements that aren't part
8695 * of the list proper, so start it just after them */
8696 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8698 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8699 shouldn't touch it */
8701 *(get_invlist_offset_addr(invlist)) = offset;
8703 /* The 'length' passed to us is the physical number of elements in the
8704 * inversion list. But if there is an offset the logical number is one
8706 invlist_set_len(invlist, length - offset, offset);
8708 invlist_set_previous_index(invlist, 0);
8710 /* Initialize the iteration pointer. */
8711 invlist_iterfinish(invlist);
8713 SvREADONLY_on(invlist);
8719 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8721 /* Grow the maximum size of an inversion list */
8723 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8725 assert(SvTYPE(invlist) == SVt_INVLIST);
8727 /* Add one to account for the zero element at the beginning which may not
8728 * be counted by the calling parameters */
8729 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8733 S__append_range_to_invlist(pTHX_ SV* const invlist,
8734 const UV start, const UV end)
8736 /* Subject to change or removal. Append the range from 'start' to 'end' at
8737 * the end of the inversion list. The range must be above any existing
8741 UV max = invlist_max(invlist);
8742 UV len = _invlist_len(invlist);
8745 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8747 if (len == 0) { /* Empty lists must be initialized */
8748 offset = start != 0;
8749 array = _invlist_array_init(invlist, ! offset);
8752 /* Here, the existing list is non-empty. The current max entry in the
8753 * list is generally the first value not in the set, except when the
8754 * set extends to the end of permissible values, in which case it is
8755 * the first entry in that final set, and so this call is an attempt to
8756 * append out-of-order */
8758 UV final_element = len - 1;
8759 array = invlist_array(invlist);
8760 if ( array[final_element] > start
8761 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8763 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",
8764 array[final_element], start,
8765 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8768 /* Here, it is a legal append. If the new range begins 1 above the end
8769 * of the range below it, it is extending the range below it, so the
8770 * new first value not in the set is one greater than the newly
8771 * extended range. */
8772 offset = *get_invlist_offset_addr(invlist);
8773 if (array[final_element] == start) {
8774 if (end != UV_MAX) {
8775 array[final_element] = end + 1;
8778 /* But if the end is the maximum representable on the machine,
8779 * assume that infinity was actually what was meant. Just let
8780 * the range that this would extend to have no end */
8781 invlist_set_len(invlist, len - 1, offset);
8787 /* Here the new range doesn't extend any existing set. Add it */
8789 len += 2; /* Includes an element each for the start and end of range */
8791 /* If wll overflow the existing space, extend, which may cause the array to
8794 invlist_extend(invlist, len);
8796 /* Have to set len here to avoid assert failure in invlist_array() */
8797 invlist_set_len(invlist, len, offset);
8799 array = invlist_array(invlist);
8802 invlist_set_len(invlist, len, offset);
8805 /* The next item on the list starts the range, the one after that is
8806 * one past the new range. */
8807 array[len - 2] = start;
8808 if (end != UV_MAX) {
8809 array[len - 1] = end + 1;
8812 /* But if the end is the maximum representable on the machine, just let
8813 * the range have no end */
8814 invlist_set_len(invlist, len - 1, offset);
8819 Perl__invlist_search(SV* const invlist, const UV cp)
8821 /* Searches the inversion list for the entry that contains the input code
8822 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8823 * return value is the index into the list's array of the range that
8824 * contains <cp>, that is, 'i' such that
8825 * array[i] <= cp < array[i+1]
8830 IV high = _invlist_len(invlist);
8831 const IV highest_element = high - 1;
8834 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8836 /* If list is empty, return failure. */
8841 /* (We can't get the array unless we know the list is non-empty) */
8842 array = invlist_array(invlist);
8844 mid = invlist_previous_index(invlist);
8846 if (mid > highest_element) {
8847 mid = highest_element;
8850 /* <mid> contains the cache of the result of the previous call to this
8851 * function (0 the first time). See if this call is for the same result,
8852 * or if it is for mid-1. This is under the theory that calls to this
8853 * function will often be for related code points that are near each other.
8854 * And benchmarks show that caching gives better results. We also test
8855 * here if the code point is within the bounds of the list. These tests
8856 * replace others that would have had to be made anyway to make sure that
8857 * the array bounds were not exceeded, and these give us extra information
8858 * at the same time */
8859 if (cp >= array[mid]) {
8860 if (cp >= array[highest_element]) {
8861 return highest_element;
8864 /* Here, array[mid] <= cp < array[highest_element]. This means that
8865 * the final element is not the answer, so can exclude it; it also
8866 * means that <mid> is not the final element, so can refer to 'mid + 1'
8868 if (cp < array[mid + 1]) {
8874 else { /* cp < aray[mid] */
8875 if (cp < array[0]) { /* Fail if outside the array */
8879 if (cp >= array[mid - 1]) {
8884 /* Binary search. What we are looking for is <i> such that
8885 * array[i] <= cp < array[i+1]
8886 * The loop below converges on the i+1. Note that there may not be an
8887 * (i+1)th element in the array, and things work nonetheless */
8888 while (low < high) {
8889 mid = (low + high) / 2;
8890 assert(mid <= highest_element);
8891 if (array[mid] <= cp) { /* cp >= array[mid] */
8894 /* We could do this extra test to exit the loop early.
8895 if (cp < array[low]) {
8900 else { /* cp < array[mid] */
8907 invlist_set_previous_index(invlist, high);
8912 Perl__invlist_populate_swatch(SV* const invlist,
8913 const UV start, const UV end, U8* swatch)
8915 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8916 * but is used when the swash has an inversion list. This makes this much
8917 * faster, as it uses a binary search instead of a linear one. This is
8918 * intimately tied to that function, and perhaps should be in utf8.c,
8919 * except it is intimately tied to inversion lists as well. It assumes
8920 * that <swatch> is all 0's on input */
8923 const IV len = _invlist_len(invlist);
8927 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8929 if (len == 0) { /* Empty inversion list */
8933 array = invlist_array(invlist);
8935 /* Find which element it is */
8936 i = _invlist_search(invlist, start);
8938 /* We populate from <start> to <end> */
8939 while (current < end) {
8942 /* The inversion list gives the results for every possible code point
8943 * after the first one in the list. Only those ranges whose index is
8944 * even are ones that the inversion list matches. For the odd ones,
8945 * and if the initial code point is not in the list, we have to skip
8946 * forward to the next element */
8947 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8949 if (i >= len) { /* Finished if beyond the end of the array */
8953 if (current >= end) { /* Finished if beyond the end of what we
8955 if (LIKELY(end < UV_MAX)) {
8959 /* We get here when the upper bound is the maximum
8960 * representable on the machine, and we are looking for just
8961 * that code point. Have to special case it */
8963 goto join_end_of_list;
8966 assert(current >= start);
8968 /* The current range ends one below the next one, except don't go past
8971 upper = (i < len && array[i] < end) ? array[i] : end;
8973 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8974 * for each code point in it */
8975 for (; current < upper; current++) {
8976 const STRLEN offset = (STRLEN)(current - start);
8977 swatch[offset >> 3] |= 1 << (offset & 7);
8982 /* Quit if at the end of the list */
8985 /* But first, have to deal with the highest possible code point on
8986 * the platform. The previous code assumes that <end> is one
8987 * beyond where we want to populate, but that is impossible at the
8988 * platform's infinity, so have to handle it specially */
8989 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8991 const STRLEN offset = (STRLEN)(end - start);
8992 swatch[offset >> 3] |= 1 << (offset & 7);
8997 /* Advance to the next range, which will be for code points not in the
9006 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9007 const bool complement_b, SV** output)
9009 /* Take the union of two inversion lists and point '*output' to it. On
9010 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9011 * even 'a' or 'b'). If to an inversion list, the contents of the original
9012 * list will be replaced by the union. The first list, 'a', may be
9013 * NULL, in which case a copy of the second list is placed in '*output'.
9014 * If 'complement_b' is TRUE, the union is taken of the complement
9015 * (inversion) of 'b' instead of b itself.
9017 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9018 * Richard Gillam, published by Addison-Wesley, and explained at some
9019 * length there. The preface says to incorporate its examples into your
9020 * code at your own risk.
9022 * The algorithm is like a merge sort. */
9024 const UV* array_a; /* a's array */
9026 UV len_a; /* length of a's array */
9029 SV* u; /* the resulting union */
9033 UV i_a = 0; /* current index into a's array */
9037 /* running count, as explained in the algorithm source book; items are
9038 * stopped accumulating and are output when the count changes to/from 0.
9039 * The count is incremented when we start a range that's in an input's set,
9040 * and decremented when we start a range that's not in a set. So this
9041 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9042 * and hence nothing goes into the union; 1, just one of the inputs is in
9043 * its set (and its current range gets added to the union); and 2 when both
9044 * inputs are in their sets. */
9047 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9049 assert(*output == NULL || SvTYPE(*output) == SVt_INVLIST);
9051 len_b = _invlist_len(b);
9054 /* Here, 'b' is empty, hence it's complement is all possible code
9055 * points. So if the union includes the complement of 'b', it includes
9056 * everything, and we need not even look at 'a'. It's easiest to
9057 * create a new inversion list that matches everything. */
9059 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9061 if (*output == NULL) { /* If the output didn't exist, just point it
9063 *output = everything;
9065 else { /* Otherwise, replace its contents with the new list */
9066 invlist_replace_list_destroys_src(*output, everything);
9067 SvREFCNT_dec_NN(everything);
9073 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9074 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9075 * output will be empty */
9077 if (a == NULL || _invlist_len(a) == 0) {
9078 if (*output == NULL) {
9079 *output = _new_invlist(0);
9082 invlist_clear(*output);
9087 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9088 * union. We can just return a copy of 'a' if '*output' doesn't point
9089 * to an existing list */
9090 if (*output == NULL) {
9091 *output = invlist_clone(a);
9095 /* If the output is to overwrite 'a', we have a no-op, as it's
9101 /* Here, '*output' is to be overwritten by 'a' */
9102 u = invlist_clone(a);
9103 invlist_replace_list_destroys_src(*output, u);
9109 /* Here 'b' is not empty. See about 'a' */
9111 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9113 /* Here, 'a' is empty (and b is not). That means the union will come
9114 * entirely from 'b'. If '*output' is NULL, we can directly return a
9115 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9118 SV ** dest = (*output == NULL) ? output : &u;
9119 *dest = invlist_clone(b);
9121 _invlist_invert(*dest);
9125 invlist_replace_list_destroys_src(*output, u);
9132 /* Here both lists exist and are non-empty */
9133 array_a = invlist_array(a);
9134 array_b = invlist_array(b);
9136 /* If are to take the union of 'a' with the complement of b, set it
9137 * up so are looking at b's complement. */
9140 /* To complement, we invert: if the first element is 0, remove it. To
9141 * do this, we just pretend the array starts one later */
9142 if (array_b[0] == 0) {
9148 /* But if the first element is not zero, we pretend the list starts
9149 * at the 0 that is always stored immediately before the array. */
9155 /* Size the union for the worst case: that the sets are completely
9157 u = _new_invlist(len_a + len_b);
9159 /* Will contain U+0000 if either component does */
9160 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9161 || (len_b > 0 && array_b[0] == 0));
9163 /* Go through each input list item by item, stopping when have exhausted
9165 while (i_a < len_a && i_b < len_b) {
9166 UV cp; /* The element to potentially add to the union's array */
9167 bool cp_in_set; /* is it in the the input list's set or not */
9169 /* We need to take one or the other of the two inputs for the union.
9170 * Since we are merging two sorted lists, we take the smaller of the
9171 * next items. In case of a tie, we take first the one that is in its
9172 * set. If we first took the one not in its set, it would decrement
9173 * the count, possibly to 0 which would cause it to be output as ending
9174 * the range, and the next time through we would take the same number,
9175 * and output it again as beginning the next range. By doing it the
9176 * opposite way, there is no possibility that the count will be
9177 * momentarily decremented to 0, and thus the two adjoining ranges will
9178 * be seamlessly merged. (In a tie and both are in the set or both not
9179 * in the set, it doesn't matter which we take first.) */
9180 if ( array_a[i_a] < array_b[i_b]
9181 || ( array_a[i_a] == array_b[i_b]
9182 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9184 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9185 cp = array_a[i_a++];
9188 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9189 cp = array_b[i_b++];
9192 /* Here, have chosen which of the two inputs to look at. Only output
9193 * if the running count changes to/from 0, which marks the
9194 * beginning/end of a range that's in the set */
9197 array_u[i_u++] = cp;
9204 array_u[i_u++] = cp;
9210 /* The loop above increments the index into exactly one of the input lists
9211 * each iteration, and ends when either index gets to its list end. That
9212 * means the other index is lower than its end, and so something is
9213 * remaining in that one. We decrement 'count', as explained below, if
9214 * that list is in its set. (i_a and i_b each currently index the element
9215 * beyond the one we care about.) */
9216 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9217 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9222 /* Above we decremented 'count' if the list that had unexamined elements in
9223 * it was in its set. This has made it so that 'count' being non-zero
9224 * means there isn't anything left to output; and 'count' equal to 0 means
9225 * that what is left to output is precisely that which is left in the
9226 * non-exhausted input list.
9228 * To see why, note first that the exhausted input obviously has nothing
9229 * left to add to the union. If it was in its set at its end, that means
9230 * the set extends from here to the platform's infinity, and hence so does
9231 * the union and the non-exhausted set is irrelevant. The exhausted set
9232 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9233 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9234 * 'count' remains at 1. This is consistent with the decremented 'count'
9235 * != 0 meaning there's nothing left to add to the union.
9237 * But if the exhausted input wasn't in its set, it contributed 0 to
9238 * 'count', and the rest of the union will be whatever the other input is.
9239 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9240 * otherwise it gets decremented to 0. This is consistent with 'count'
9241 * == 0 meaning the remainder of the union is whatever is left in the
9242 * non-exhausted list. */
9247 IV copy_count = len_a - i_a;
9248 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9249 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9251 else { /* The non-exhausted input is b */
9252 copy_count = len_b - i_b;
9253 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9255 len_u = i_u + copy_count;
9258 /* Set the result to the final length, which can change the pointer to
9259 * array_u, so re-find it. (Note that it is unlikely that this will
9260 * change, as we are shrinking the space, not enlarging it) */
9261 if (len_u != _invlist_len(u)) {
9262 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9264 array_u = invlist_array(u);
9267 if (*output == NULL) { /* Simply return the new inversion list */
9271 /* Otherwise, overwrite the inversion list that was in '*output'. We
9272 * could instead free '*output', and then set it to 'u', but experience
9273 * has shown [perl #127392] that if the input is a mortal, we can get a
9274 * huge build-up of these during regex compilation before they get
9276 invlist_replace_list_destroys_src(*output, u);
9284 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9285 const bool complement_b, SV** i)
9287 /* Take the intersection of two inversion lists and point '*i' to it. On
9288 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9289 * even 'a' or 'b'). If to an inversion list, the contents of the original
9290 * list will be replaced by the intersection. The first list, 'a', may be
9291 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9292 * TRUE, the result will be the intersection of 'a' and the complement (or
9293 * inversion) of 'b' instead of 'b' directly.
9295 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9296 * Richard Gillam, published by Addison-Wesley, and explained at some
9297 * length there. The preface says to incorporate its examples into your
9298 * code at your own risk. In fact, it had bugs
9300 * The algorithm is like a merge sort, and is essentially the same as the
9304 const UV* array_a; /* a's array */
9306 UV len_a; /* length of a's array */
9309 SV* r; /* the resulting intersection */
9313 UV i_a = 0; /* current index into a's array */
9317 /* running count of how many of the two inputs are postitioned at ranges
9318 * that are in their sets. As explained in the algorithm source book,
9319 * items are stopped accumulating and are output when the count changes
9320 * to/from 2. The count is incremented when we start a range that's in an
9321 * input's set, and decremented when we start a range that's not in a set.
9322 * Only when it is 2 are we in the intersection. */
9325 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9327 assert(*i == NULL || SvTYPE(*i) == SVt_INVLIST);
9329 /* Special case if either one is empty */
9330 len_a = (a == NULL) ? 0 : _invlist_len(a);
9331 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9332 if (len_a != 0 && complement_b) {
9334 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9335 * must be empty. Here, also we are using 'b's complement, which
9336 * hence must be every possible code point. Thus the intersection
9339 if (*i == a) { /* No-op */
9344 *i = invlist_clone(a);
9348 r = invlist_clone(a);
9349 invlist_replace_list_destroys_src(*i, r);
9354 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9355 * intersection must be empty */
9357 *i = _new_invlist(0);
9365 /* Here both lists exist and are non-empty */
9366 array_a = invlist_array(a);
9367 array_b = invlist_array(b);
9369 /* If are to take the intersection of 'a' with the complement of b, set it
9370 * up so are looking at b's complement. */
9373 /* To complement, we invert: if the first element is 0, remove it. To
9374 * do this, we just pretend the array starts one later */
9375 if (array_b[0] == 0) {
9381 /* But if the first element is not zero, we pretend the list starts
9382 * at the 0 that is always stored immediately before the array. */
9388 /* Size the intersection for the worst case: that the intersection ends up
9389 * fragmenting everything to be completely disjoint */
9390 r= _new_invlist(len_a + len_b);
9392 /* Will contain U+0000 iff both components do */
9393 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9394 && len_b > 0 && array_b[0] == 0);
9396 /* Go through each list item by item, stopping when have exhausted one of
9398 while (i_a < len_a && i_b < len_b) {
9399 UV cp; /* The element to potentially add to the intersection's
9401 bool cp_in_set; /* Is it in the input list's set or not */
9403 /* We need to take one or the other of the two inputs for the
9404 * intersection. Since we are merging two sorted lists, we take the
9405 * smaller of the next items. In case of a tie, we take first the one
9406 * that is not in its set (a difference from the union algorithm). If
9407 * we first took the one in its set, it would increment the count,
9408 * possibly to 2 which would cause it to be output as starting a range
9409 * in the intersection, and the next time through we would take that
9410 * same number, and output it again as ending the set. By doing the
9411 * opposite of this, there is no possibility that the count will be
9412 * momentarily incremented to 2. (In a tie and both are in the set or
9413 * both not in the set, it doesn't matter which we take first.) */
9414 if ( array_a[i_a] < array_b[i_b]
9415 || ( array_a[i_a] == array_b[i_b]
9416 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9418 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9419 cp = array_a[i_a++];
9422 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9426 /* Here, have chosen which of the two inputs to look at. Only output
9427 * if the running count changes to/from 2, which marks the
9428 * beginning/end of a range that's in the intersection */
9432 array_r[i_r++] = cp;
9437 array_r[i_r++] = cp;
9444 /* The loop above increments the index into exactly one of the input lists
9445 * each iteration, and ends when either index gets to its list end. That
9446 * means the other index is lower than its end, and so something is
9447 * remaining in that one. We increment 'count', as explained below, if the
9448 * exhausted list was in its set. (i_a and i_b each currently index the
9449 * element beyond the one we care about.) */
9450 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9451 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9456 /* Above we incremented 'count' if the exhausted list was in its set. This
9457 * has made it so that 'count' being below 2 means there is nothing left to
9458 * output; otheriwse what's left to add to the intersection is precisely
9459 * that which is left in the non-exhausted input list.
9461 * To see why, note first that the exhausted input obviously has nothing
9462 * left to affect the intersection. If it was in its set at its end, that
9463 * means the set extends from here to the platform's infinity, and hence
9464 * anything in the non-exhausted's list will be in the intersection, and
9465 * anything not in it won't be. Hence, the rest of the intersection is
9466 * precisely what's in the non-exhausted list The exhausted set also
9467 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9468 * it means 'count' is now at least 2. This is consistent with the
9469 * incremented 'count' being >= 2 means to add the non-exhausted list to
9472 * But if the exhausted input wasn't in its set, it contributed 0 to
9473 * 'count', and the intersection can't include anything further; the
9474 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9475 * incremented. This is consistent with 'count' being < 2 meaning nothing
9476 * further to add to the intersection. */
9477 if (count < 2) { /* Nothing left to put in the intersection. */
9480 else { /* copy the non-exhausted list, unchanged. */
9481 IV copy_count = len_a - i_a;
9482 if (copy_count > 0) { /* a is the one with stuff left */
9483 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9485 else { /* b is the one with stuff left */
9486 copy_count = len_b - i_b;
9487 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9489 len_r = i_r + copy_count;
9492 /* Set the result to the final length, which can change the pointer to
9493 * array_r, so re-find it. (Note that it is unlikely that this will
9494 * change, as we are shrinking the space, not enlarging it) */
9495 if (len_r != _invlist_len(r)) {
9496 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9498 array_r = invlist_array(r);
9501 if (*i == NULL) { /* Simply return the calculated intersection */
9504 else { /* Otherwise, replace the existing inversion list in '*i'. We could
9505 instead free '*i', and then set it to 'r', but experience has
9506 shown [perl #127392] that if the input is a mortal, we can get a
9507 huge build-up of these during regex compilation before they get
9510 invlist_replace_list_destroys_src(*i, r);
9522 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9524 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9525 * set. A pointer to the inversion list is returned. This may actually be
9526 * a new list, in which case the passed in one has been destroyed. The
9527 * passed-in inversion list can be NULL, in which case a new one is created
9528 * with just the one range in it. The new list is not necessarily
9529 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9530 * result of this function. The gain would not be large, and in many
9531 * cases, this is called multiple times on a single inversion list, so
9532 * anything freed may almost immediately be needed again.
9534 * This used to mostly call the 'union' routine, but that is much more
9535 * heavyweight than really needed for a single range addition */
9537 UV* array; /* The array implementing the inversion list */
9538 UV len; /* How many elements in 'array' */
9539 SSize_t i_s; /* index into the invlist array where 'start'
9541 SSize_t i_e = 0; /* And the index where 'end' should go */
9542 UV cur_highest; /* The highest code point in the inversion list
9543 upon entry to this function */
9545 /* This range becomes the whole inversion list if none already existed */
9546 if (invlist == NULL) {
9547 invlist = _new_invlist(2);
9548 _append_range_to_invlist(invlist, start, end);
9552 /* Likewise, if the inversion list is currently empty */
9553 len = _invlist_len(invlist);
9555 _append_range_to_invlist(invlist, start, end);
9559 /* Starting here, we have to know the internals of the list */
9560 array = invlist_array(invlist);
9562 /* If the new range ends higher than the current highest ... */
9563 cur_highest = invlist_highest(invlist);
9564 if (end > cur_highest) {
9566 /* If the whole range is higher, we can just append it */
9567 if (start > cur_highest) {
9568 _append_range_to_invlist(invlist, start, end);
9572 /* Otherwise, add the portion that is higher ... */
9573 _append_range_to_invlist(invlist, cur_highest + 1, end);
9575 /* ... and continue on below to handle the rest. As a result of the
9576 * above append, we know that the index of the end of the range is the
9577 * final even numbered one of the array. Recall that the final element
9578 * always starts a range that extends to infinity. If that range is in
9579 * the set (meaning the set goes from here to infinity), it will be an
9580 * even index, but if it isn't in the set, it's odd, and the final
9581 * range in the set is one less, which is even. */
9582 if (end == UV_MAX) {
9590 /* We have dealt with appending, now see about prepending. If the new
9591 * range starts lower than the current lowest ... */
9592 if (start < array[0]) {
9594 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
9595 * Let the union code handle it, rather than having to know the
9596 * trickiness in two code places. */
9597 if (UNLIKELY(start == 0)) {
9600 range_invlist = _new_invlist(2);
9601 _append_range_to_invlist(range_invlist, start, end);
9603 _invlist_union(invlist, range_invlist, &invlist);
9605 SvREFCNT_dec_NN(range_invlist);
9610 /* If the whole new range comes before the first entry, and doesn't
9611 * extend it, we have to insert it as an additional range */
9612 if (end < array[0] - 1) {
9614 goto splice_in_new_range;
9617 /* Here the new range adjoins the existing first range, extending it
9621 /* And continue on below to handle the rest. We know that the index of
9622 * the beginning of the range is the first one of the array */
9625 else { /* Not prepending any part of the new range to the existing list.
9626 * Find where in the list it should go. This finds i_s, such that:
9627 * invlist[i_s] <= start < array[i_s+1]
9629 i_s = _invlist_search(invlist, start);
9632 /* At this point, any extending before the beginning of the inversion list
9633 * and/or after the end has been done. This has made it so that, in the
9634 * code below, each endpoint of the new range is either in a range that is
9635 * in the set, or is in a gap between two ranges that are. This means we
9636 * don't have to worry about exceeding the array bounds.
9638 * Find where in the list the new range ends (but we can skip this if we
9639 * have already determined what it is, or if it will be the same as i_s,
9640 * which we already have computed) */
9642 i_e = (start == end)
9644 : _invlist_search(invlist, end);
9647 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
9648 * is a range that goes to infinity there is no element at invlist[i_e+1],
9649 * so only the first relation holds. */
9651 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9653 /* Here, the ranges on either side of the beginning of the new range
9654 * are in the set, and this range starts in the gap between them.
9656 * The new range extends the range above it downwards if the new range
9657 * ends at or above that range's start */
9658 const bool extends_the_range_above = ( end == UV_MAX
9659 || end + 1 >= array[i_s+1]);
9661 /* The new range extends the range below it upwards if it begins just
9662 * after where that range ends */
9663 if (start == array[i_s]) {
9665 /* If the new range fills the entire gap between the other ranges,
9666 * they will get merged together. Other ranges may also get
9667 * merged, depending on how many of them the new range spans. In
9668 * the general case, we do the merge later, just once, after we
9669 * figure out how many to merge. But in the case where the new
9670 * range exactly spans just this one gap (possibly extending into
9671 * the one above), we do the merge here, and an early exit. This
9672 * is done here to avoid having to special case later. */
9673 if (i_e - i_s <= 1) {
9675 /* If i_e - i_s == 1, it means that the new range terminates
9676 * within the range above, and hence 'extends_the_range_above'
9677 * must be true. (If the range above it extends to infinity,
9678 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
9679 * will be 0, so no harm done.) */
9680 if (extends_the_range_above) {
9681 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
9682 invlist_set_len(invlist,
9684 *(get_invlist_offset_addr(invlist)));
9688 /* Here, i_e must == i_s. We keep them in sync, as they apply
9689 * to the same range, and below we are about to decrement i_s
9694 /* Here, the new range is adjacent to the one below. (It may also
9695 * span beyond the range above, but that will get resolved later.)
9696 * Extend the range below to include this one. */
9697 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
9701 else if (extends_the_range_above) {
9703 /* Here the new range only extends the range above it, but not the
9704 * one below. It merges with the one above. Again, we keep i_e
9705 * and i_s in sync if they point to the same range */
9714 /* Here, we've dealt with the new range start extending any adjoining
9717 * If the new range extends to infinity, it is now the final one,
9718 * regardless of what was there before */
9719 if (UNLIKELY(end == UV_MAX)) {
9720 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
9724 /* If i_e started as == i_s, it has also been dealt with,
9725 * and been updated to the new i_s, which will fail the following if */
9726 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
9728 /* Here, the ranges on either side of the end of the new range are in
9729 * the set, and this range ends in the gap between them.
9731 * If this range is adjacent to (hence extends) the range above it, it
9732 * becomes part of that range; likewise if it extends the range below,
9733 * it becomes part of that range */
9734 if (end + 1 == array[i_e+1]) {
9738 else if (start <= array[i_e]) {
9739 array[i_e] = end + 1;
9746 /* If the range fits entirely in an existing range (as possibly already
9747 * extended above), it doesn't add anything new */
9748 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9752 /* Here, no part of the range is in the list. Must add it. It will
9753 * occupy 2 more slots */
9754 splice_in_new_range:
9756 invlist_extend(invlist, len + 2);
9757 array = invlist_array(invlist);
9758 /* Move the rest of the array down two slots. Don't include any
9760 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
9762 /* Do the actual splice */
9763 array[i_e+1] = start;
9764 array[i_e+2] = end + 1;
9765 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
9769 /* Here the new range crossed the boundaries of a pre-existing range. The
9770 * code above has adjusted things so that both ends are in ranges that are
9771 * in the set. This means everything in between must also be in the set.
9772 * Just squash things together */
9773 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
9774 invlist_set_len(invlist,
9776 *(get_invlist_offset_addr(invlist)));
9782 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9783 UV** other_elements_ptr)
9785 /* Create and return an inversion list whose contents are to be populated
9786 * by the caller. The caller gives the number of elements (in 'size') and
9787 * the very first element ('element0'). This function will set
9788 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9791 * Obviously there is some trust involved that the caller will properly
9792 * fill in the other elements of the array.
9794 * (The first element needs to be passed in, as the underlying code does
9795 * things differently depending on whether it is zero or non-zero) */
9797 SV* invlist = _new_invlist(size);
9800 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9802 invlist = add_cp_to_invlist(invlist, element0);
9803 offset = *get_invlist_offset_addr(invlist);
9805 invlist_set_len(invlist, size, offset);
9806 *other_elements_ptr = invlist_array(invlist) + 1;
9812 PERL_STATIC_INLINE SV*
9813 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9814 return _add_range_to_invlist(invlist, cp, cp);
9817 #ifndef PERL_IN_XSUB_RE
9819 Perl__invlist_invert(pTHX_ SV* const invlist)
9821 /* Complement the input inversion list. This adds a 0 if the list didn't
9822 * have a zero; removes it otherwise. As described above, the data
9823 * structure is set up so that this is very efficient */
9825 PERL_ARGS_ASSERT__INVLIST_INVERT;
9827 assert(! invlist_is_iterating(invlist));
9829 /* The inverse of matching nothing is matching everything */
9830 if (_invlist_len(invlist) == 0) {
9831 _append_range_to_invlist(invlist, 0, UV_MAX);
9835 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9840 PERL_STATIC_INLINE SV*
9841 S_invlist_clone(pTHX_ SV* const invlist)
9844 /* Return a new inversion list that is a copy of the input one, which is
9845 * unchanged. The new list will not be mortal even if the old one was. */
9847 /* Need to allocate extra space to accommodate Perl's addition of a
9848 * trailing NUL to SvPV's, since it thinks they are always strings */
9849 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9850 STRLEN physical_length = SvCUR(invlist);
9851 bool offset = *(get_invlist_offset_addr(invlist));
9853 PERL_ARGS_ASSERT_INVLIST_CLONE;
9855 *(get_invlist_offset_addr(new_invlist)) = offset;
9856 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9857 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9862 PERL_STATIC_INLINE STRLEN*
9863 S_get_invlist_iter_addr(SV* invlist)
9865 /* Return the address of the UV that contains the current iteration
9868 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9870 assert(SvTYPE(invlist) == SVt_INVLIST);
9872 return &(((XINVLIST*) SvANY(invlist))->iterator);
9875 PERL_STATIC_INLINE void
9876 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9878 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9880 *get_invlist_iter_addr(invlist) = 0;
9883 PERL_STATIC_INLINE void
9884 S_invlist_iterfinish(SV* invlist)
9886 /* Terminate iterator for invlist. This is to catch development errors.
9887 * Any iteration that is interrupted before completed should call this
9888 * function. Functions that add code points anywhere else but to the end
9889 * of an inversion list assert that they are not in the middle of an
9890 * iteration. If they were, the addition would make the iteration
9891 * problematical: if the iteration hadn't reached the place where things
9892 * were being added, it would be ok */
9894 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9896 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9900 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9902 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9903 * This call sets in <*start> and <*end>, the next range in <invlist>.
9904 * Returns <TRUE> if successful and the next call will return the next
9905 * range; <FALSE> if was already at the end of the list. If the latter,
9906 * <*start> and <*end> are unchanged, and the next call to this function
9907 * will start over at the beginning of the list */
9909 STRLEN* pos = get_invlist_iter_addr(invlist);
9910 UV len = _invlist_len(invlist);
9913 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9916 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9920 array = invlist_array(invlist);
9922 *start = array[(*pos)++];
9928 *end = array[(*pos)++] - 1;
9934 PERL_STATIC_INLINE UV
9935 S_invlist_highest(SV* const invlist)
9937 /* Returns the highest code point that matches an inversion list. This API
9938 * has an ambiguity, as it returns 0 under either the highest is actually
9939 * 0, or if the list is empty. If this distinction matters to you, check
9940 * for emptiness before calling this function */
9942 UV len = _invlist_len(invlist);
9945 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9951 array = invlist_array(invlist);
9953 /* The last element in the array in the inversion list always starts a
9954 * range that goes to infinity. That range may be for code points that are
9955 * matched in the inversion list, or it may be for ones that aren't
9956 * matched. In the latter case, the highest code point in the set is one
9957 * less than the beginning of this range; otherwise it is the final element
9958 * of this range: infinity */
9959 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
9961 : array[len - 1] - 1;
9965 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
9967 /* Get the contents of an inversion list into a string SV so that they can
9968 * be printed out. If 'traditional_style' is TRUE, it uses the format
9969 * traditionally done for debug tracing; otherwise it uses a format
9970 * suitable for just copying to the output, with blanks between ranges and
9971 * a dash between range components */
9975 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
9976 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
9978 if (traditional_style) {
9979 output = newSVpvs("\n");
9982 output = newSVpvs("");
9985 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
9987 assert(! invlist_is_iterating(invlist));
9989 invlist_iterinit(invlist);
9990 while (invlist_iternext(invlist, &start, &end)) {
9991 if (end == UV_MAX) {
9992 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFINITY%c",
9993 start, intra_range_delimiter,
9994 inter_range_delimiter);
9996 else if (end != start) {
9997 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
9999 intra_range_delimiter,
10000 end, inter_range_delimiter);
10003 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10004 start, inter_range_delimiter);
10008 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10009 SvCUR_set(output, SvCUR(output) - 1);
10015 #ifndef PERL_IN_XSUB_RE
10017 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10018 const char * const indent, SV* const invlist)
10020 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10021 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10022 * the string 'indent'. The output looks like this:
10023 [0] 0x000A .. 0x000D
10025 [4] 0x2028 .. 0x2029
10026 [6] 0x3104 .. INFINITY
10027 * This means that the first range of code points matched by the list are
10028 * 0xA through 0xD; the second range contains only the single code point
10029 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10030 * are used to define each range (except if the final range extends to
10031 * infinity, only a single element is needed). The array index of the
10032 * first element for the corresponding range is given in brackets. */
10037 PERL_ARGS_ASSERT__INVLIST_DUMP;
10039 if (invlist_is_iterating(invlist)) {
10040 Perl_dump_indent(aTHX_ level, file,
10041 "%sCan't dump inversion list because is in middle of iterating\n",
10046 invlist_iterinit(invlist);
10047 while (invlist_iternext(invlist, &start, &end)) {
10048 if (end == UV_MAX) {
10049 Perl_dump_indent(aTHX_ level, file,
10050 "%s[%" UVuf "] 0x%04" UVXf " .. INFINITY\n",
10051 indent, (UV)count, start);
10053 else if (end != start) {
10054 Perl_dump_indent(aTHX_ level, file,
10055 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10056 indent, (UV)count, start, end);
10059 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10060 indent, (UV)count, start);
10067 Perl__load_PL_utf8_foldclosures (pTHX)
10069 assert(! PL_utf8_foldclosures);
10071 /* If the folds haven't been read in, call a fold function
10073 if (! PL_utf8_tofold) {
10074 U8 dummy[UTF8_MAXBYTES_CASE+1];
10075 const U8 hyphen[] = HYPHEN_UTF8;
10077 /* This string is just a short named one above \xff */
10078 toFOLD_utf8_safe(hyphen, hyphen + sizeof(hyphen) - 1, dummy, NULL);
10079 assert(PL_utf8_tofold); /* Verify that worked */
10081 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
10085 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10087 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10089 /* Return a boolean as to if the two passed in inversion lists are
10090 * identical. The final argument, if TRUE, says to take the complement of
10091 * the second inversion list before doing the comparison */
10093 const UV* array_a = invlist_array(a);
10094 const UV* array_b = invlist_array(b);
10095 UV len_a = _invlist_len(a);
10096 UV len_b = _invlist_len(b);
10098 PERL_ARGS_ASSERT__INVLISTEQ;
10100 /* If are to compare 'a' with the complement of b, set it
10101 * up so are looking at b's complement. */
10102 if (complement_b) {
10104 /* The complement of nothing is everything, so <a> would have to have
10105 * just one element, starting at zero (ending at infinity) */
10107 return (len_a == 1 && array_a[0] == 0);
10109 else if (array_b[0] == 0) {
10111 /* Otherwise, to complement, we invert. Here, the first element is
10112 * 0, just remove it. To do this, we just pretend the array starts
10120 /* But if the first element is not zero, we pretend the list starts
10121 * at the 0 that is always stored immediately before the array. */
10127 return len_a == len_b
10128 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10134 * As best we can, determine the characters that can match the start of
10135 * the given EXACTF-ish node.
10137 * Returns the invlist as a new SV*; it is the caller's responsibility to
10138 * call SvREFCNT_dec() when done with it.
10141 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10143 const U8 * s = (U8*)STRING(node);
10144 SSize_t bytelen = STR_LEN(node);
10146 /* Start out big enough for 2 separate code points */
10147 SV* invlist = _new_invlist(4);
10149 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
10154 /* We punt and assume can match anything if the node begins
10155 * with a multi-character fold. Things are complicated. For
10156 * example, /ffi/i could match any of:
10157 * "\N{LATIN SMALL LIGATURE FFI}"
10158 * "\N{LATIN SMALL LIGATURE FF}I"
10159 * "F\N{LATIN SMALL LIGATURE FI}"
10160 * plus several other things; and making sure we have all the
10161 * possibilities is hard. */
10162 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10163 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10166 /* Any Latin1 range character can potentially match any
10167 * other depending on the locale */
10168 if (OP(node) == EXACTFL) {
10169 _invlist_union(invlist, PL_Latin1, &invlist);
10172 /* But otherwise, it matches at least itself. We can
10173 * quickly tell if it has a distinct fold, and if so,
10174 * it matches that as well */
10175 invlist = add_cp_to_invlist(invlist, uc);
10176 if (IS_IN_SOME_FOLD_L1(uc))
10177 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10180 /* Some characters match above-Latin1 ones under /i. This
10181 * is true of EXACTFL ones when the locale is UTF-8 */
10182 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10183 && (! isASCII(uc) || (OP(node) != EXACTFA
10184 && OP(node) != EXACTFA_NO_TRIE)))
10186 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10190 else { /* Pattern is UTF-8 */
10191 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10192 STRLEN foldlen = UTF8SKIP(s);
10193 const U8* e = s + bytelen;
10196 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10198 /* The only code points that aren't folded in a UTF EXACTFish
10199 * node are are the problematic ones in EXACTFL nodes */
10200 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10201 /* We need to check for the possibility that this EXACTFL
10202 * node begins with a multi-char fold. Therefore we fold
10203 * the first few characters of it so that we can make that
10208 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10210 *(d++) = (U8) toFOLD(*s);
10215 toFOLD_utf8_safe(s, e, d, &len);
10221 /* And set up so the code below that looks in this folded
10222 * buffer instead of the node's string */
10224 foldlen = UTF8SKIP(folded);
10228 /* When we reach here 's' points to the fold of the first
10229 * character(s) of the node; and 'e' points to far enough along
10230 * the folded string to be just past any possible multi-char
10231 * fold. 'foldlen' is the length in bytes of the first
10234 * Unlike the non-UTF-8 case, the macro for determining if a
10235 * string is a multi-char fold requires all the characters to
10236 * already be folded. This is because of all the complications
10237 * if not. Note that they are folded anyway, except in EXACTFL
10238 * nodes. Like the non-UTF case above, we punt if the node
10239 * begins with a multi-char fold */
10241 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10242 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10244 else { /* Single char fold */
10246 /* It matches all the things that fold to it, which are
10247 * found in PL_utf8_foldclosures (including itself) */
10248 invlist = add_cp_to_invlist(invlist, uc);
10249 if (! PL_utf8_foldclosures)
10250 _load_PL_utf8_foldclosures();
10251 if ((listp = hv_fetch(PL_utf8_foldclosures,
10252 (char *) s, foldlen, FALSE)))
10254 AV* list = (AV*) *listp;
10256 for (k = 0; k <= av_tindex_nomg(list); k++) {
10257 SV** c_p = av_fetch(list, k, FALSE);
10263 /* /aa doesn't allow folds between ASCII and non- */
10264 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
10265 && isASCII(c) != isASCII(uc))
10270 invlist = add_cp_to_invlist(invlist, c);
10279 #undef HEADER_LENGTH
10280 #undef TO_INTERNAL_SIZE
10281 #undef FROM_INTERNAL_SIZE
10282 #undef INVLIST_VERSION_ID
10284 /* End of inversion list object */
10287 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10289 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10290 * constructs, and updates RExC_flags with them. On input, RExC_parse
10291 * should point to the first flag; it is updated on output to point to the
10292 * final ')' or ':'. There needs to be at least one flag, or this will
10295 /* for (?g), (?gc), and (?o) warnings; warning
10296 about (?c) will warn about (?g) -- japhy */
10298 #define WASTED_O 0x01
10299 #define WASTED_G 0x02
10300 #define WASTED_C 0x04
10301 #define WASTED_GC (WASTED_G|WASTED_C)
10302 I32 wastedflags = 0x00;
10303 U32 posflags = 0, negflags = 0;
10304 U32 *flagsp = &posflags;
10305 char has_charset_modifier = '\0';
10307 bool has_use_defaults = FALSE;
10308 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10309 int x_mod_count = 0;
10311 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10313 /* '^' as an initial flag sets certain defaults */
10314 if (UCHARAT(RExC_parse) == '^') {
10316 has_use_defaults = TRUE;
10317 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10318 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
10319 ? REGEX_UNICODE_CHARSET
10320 : REGEX_DEPENDS_CHARSET);
10323 cs = get_regex_charset(RExC_flags);
10324 if (cs == REGEX_DEPENDS_CHARSET
10325 && (RExC_utf8 || RExC_uni_semantics))
10327 cs = REGEX_UNICODE_CHARSET;
10330 while (RExC_parse < RExC_end) {
10331 /* && strchr("iogcmsx", *RExC_parse) */
10332 /* (?g), (?gc) and (?o) are useless here
10333 and must be globally applied -- japhy */
10334 switch (*RExC_parse) {
10336 /* Code for the imsxn flags */
10337 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10339 case LOCALE_PAT_MOD:
10340 if (has_charset_modifier) {
10341 goto excess_modifier;
10343 else if (flagsp == &negflags) {
10346 cs = REGEX_LOCALE_CHARSET;
10347 has_charset_modifier = LOCALE_PAT_MOD;
10349 case UNICODE_PAT_MOD:
10350 if (has_charset_modifier) {
10351 goto excess_modifier;
10353 else if (flagsp == &negflags) {
10356 cs = REGEX_UNICODE_CHARSET;
10357 has_charset_modifier = UNICODE_PAT_MOD;
10359 case ASCII_RESTRICT_PAT_MOD:
10360 if (flagsp == &negflags) {
10363 if (has_charset_modifier) {
10364 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10365 goto excess_modifier;
10367 /* Doubled modifier implies more restricted */
10368 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10371 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10373 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10375 case DEPENDS_PAT_MOD:
10376 if (has_use_defaults) {
10377 goto fail_modifiers;
10379 else if (flagsp == &negflags) {
10382 else if (has_charset_modifier) {
10383 goto excess_modifier;
10386 /* The dual charset means unicode semantics if the
10387 * pattern (or target, not known until runtime) are
10388 * utf8, or something in the pattern indicates unicode
10390 cs = (RExC_utf8 || RExC_uni_semantics)
10391 ? REGEX_UNICODE_CHARSET
10392 : REGEX_DEPENDS_CHARSET;
10393 has_charset_modifier = DEPENDS_PAT_MOD;
10397 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10398 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10400 else if (has_charset_modifier == *(RExC_parse - 1)) {
10401 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10402 *(RExC_parse - 1));
10405 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10407 NOT_REACHED; /*NOTREACHED*/
10410 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10411 *(RExC_parse - 1));
10412 NOT_REACHED; /*NOTREACHED*/
10413 case ONCE_PAT_MOD: /* 'o' */
10414 case GLOBAL_PAT_MOD: /* 'g' */
10415 if (PASS2 && ckWARN(WARN_REGEXP)) {
10416 const I32 wflagbit = *RExC_parse == 'o'
10419 if (! (wastedflags & wflagbit) ) {
10420 wastedflags |= wflagbit;
10421 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10424 "Useless (%s%c) - %suse /%c modifier",
10425 flagsp == &negflags ? "?-" : "?",
10427 flagsp == &negflags ? "don't " : "",
10434 case CONTINUE_PAT_MOD: /* 'c' */
10435 if (PASS2 && ckWARN(WARN_REGEXP)) {
10436 if (! (wastedflags & WASTED_C) ) {
10437 wastedflags |= WASTED_GC;
10438 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10441 "Useless (%sc) - %suse /gc modifier",
10442 flagsp == &negflags ? "?-" : "?",
10443 flagsp == &negflags ? "don't " : ""
10448 case KEEPCOPY_PAT_MOD: /* 'p' */
10449 if (flagsp == &negflags) {
10451 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10453 *flagsp |= RXf_PMf_KEEPCOPY;
10457 /* A flag is a default iff it is following a minus, so
10458 * if there is a minus, it means will be trying to
10459 * re-specify a default which is an error */
10460 if (has_use_defaults || flagsp == &negflags) {
10461 goto fail_modifiers;
10463 flagsp = &negflags;
10464 wastedflags = 0; /* reset so (?g-c) warns twice */
10470 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
10471 negflags |= RXf_PMf_EXTENDED_MORE;
10473 RExC_flags |= posflags;
10475 if (negflags & RXf_PMf_EXTENDED) {
10476 negflags |= RXf_PMf_EXTENDED_MORE;
10478 RExC_flags &= ~negflags;
10479 set_regex_charset(&RExC_flags, cs);
10484 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10485 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10486 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
10487 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10488 NOT_REACHED; /*NOTREACHED*/
10491 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10494 vFAIL("Sequence (?... not terminated");
10498 - reg - regular expression, i.e. main body or parenthesized thing
10500 * Caller must absorb opening parenthesis.
10502 * Combining parenthesis handling with the base level of regular expression
10503 * is a trifle forced, but the need to tie the tails of the branches to what
10504 * follows makes it hard to avoid.
10506 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10508 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10510 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10513 PERL_STATIC_INLINE regnode *
10514 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10516 char * parse_start,
10521 char* name_start = RExC_parse;
10523 SV *sv_dat = reg_scan_name(pRExC_state, SIZE_ONLY
10524 ? REG_RSN_RETURN_NULL
10525 : REG_RSN_RETURN_DATA);
10526 GET_RE_DEBUG_FLAGS_DECL;
10528 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10530 if (RExC_parse == name_start || *RExC_parse != ch) {
10531 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10532 vFAIL2("Sequence %.3s... not terminated",parse_start);
10536 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10537 RExC_rxi->data->data[num]=(void*)sv_dat;
10538 SvREFCNT_inc_simple_void(sv_dat);
10541 ret = reganode(pRExC_state,
10544 : (ASCII_FOLD_RESTRICTED)
10546 : (AT_LEAST_UNI_SEMANTICS)
10552 *flagp |= HASWIDTH;
10554 Set_Node_Offset(ret, parse_start+1);
10555 Set_Node_Cur_Length(ret, parse_start);
10557 nextchar(pRExC_state);
10561 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
10562 flags. Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan
10563 needs to be restarted, or'd with NEED_UTF8 if the pattern needs to be
10564 upgraded to UTF-8. Otherwise would only return NULL if regbranch() returns
10565 NULL, which cannot happen. */
10567 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
10568 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10569 * 2 is like 1, but indicates that nextchar() has been called to advance
10570 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10571 * this flag alerts us to the need to check for that */
10573 regnode *ret; /* Will be the head of the group. */
10576 regnode *ender = NULL;
10579 U32 oregflags = RExC_flags;
10580 bool have_branch = 0;
10582 I32 freeze_paren = 0;
10583 I32 after_freeze = 0;
10584 I32 num; /* numeric backreferences */
10586 char * parse_start = RExC_parse; /* MJD */
10587 char * const oregcomp_parse = RExC_parse;
10589 GET_RE_DEBUG_FLAGS_DECL;
10591 PERL_ARGS_ASSERT_REG;
10592 DEBUG_PARSE("reg ");
10594 *flagp = 0; /* Tentatively. */
10596 /* Having this true makes it feasible to have a lot fewer tests for the
10597 * parse pointer being in scope. For example, we can write
10598 * while(isFOO(*RExC_parse)) RExC_parse++;
10600 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
10602 assert(*RExC_end == '\0');
10604 /* Make an OPEN node, if parenthesized. */
10607 /* Under /x, space and comments can be gobbled up between the '(' and
10608 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
10609 * intervening space, as the sequence is a token, and a token should be
10611 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
10613 if (RExC_parse >= RExC_end) {
10614 vFAIL("Unmatched (");
10617 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
10618 char *start_verb = RExC_parse + 1;
10620 char *start_arg = NULL;
10621 unsigned char op = 0;
10622 int arg_required = 0;
10623 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
10625 if (has_intervening_patws) {
10626 RExC_parse++; /* past the '*' */
10627 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
10629 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
10630 if ( *RExC_parse == ':' ) {
10631 start_arg = RExC_parse + 1;
10634 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10636 verb_len = RExC_parse - start_verb;
10638 if (RExC_parse >= RExC_end) {
10639 goto unterminated_verb_pattern;
10641 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10642 while ( RExC_parse < RExC_end && *RExC_parse != ')' )
10643 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10644 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10645 unterminated_verb_pattern:
10646 vFAIL("Unterminated verb pattern argument");
10647 if ( RExC_parse == start_arg )
10650 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10651 vFAIL("Unterminated verb pattern");
10654 /* Here, we know that RExC_parse < RExC_end */
10656 switch ( *start_verb ) {
10657 case 'A': /* (*ACCEPT) */
10658 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
10660 internal_argval = RExC_nestroot;
10663 case 'C': /* (*COMMIT) */
10664 if ( memEQs(start_verb,verb_len,"COMMIT") )
10667 case 'F': /* (*FAIL) */
10668 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
10672 case ':': /* (*:NAME) */
10673 case 'M': /* (*MARK:NAME) */
10674 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
10679 case 'P': /* (*PRUNE) */
10680 if ( memEQs(start_verb,verb_len,"PRUNE") )
10683 case 'S': /* (*SKIP) */
10684 if ( memEQs(start_verb,verb_len,"SKIP") )
10687 case 'T': /* (*THEN) */
10688 /* [19:06] <TimToady> :: is then */
10689 if ( memEQs(start_verb,verb_len,"THEN") ) {
10691 RExC_seen |= REG_CUTGROUP_SEEN;
10696 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10698 "Unknown verb pattern '%" UTF8f "'",
10699 UTF8fARG(UTF, verb_len, start_verb));
10701 if ( arg_required && !start_arg ) {
10702 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
10703 verb_len, start_verb);
10705 if (internal_argval == -1) {
10706 ret = reganode(pRExC_state, op, 0);
10708 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
10710 RExC_seen |= REG_VERBARG_SEEN;
10711 if ( ! SIZE_ONLY ) {
10713 SV *sv = newSVpvn( start_arg,
10714 RExC_parse - start_arg);
10715 ARG(ret) = add_data( pRExC_state,
10716 STR_WITH_LEN("S"));
10717 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
10722 if ( internal_argval != -1 )
10723 ARG2L_SET(ret, internal_argval);
10725 nextchar(pRExC_state);
10728 else if (*RExC_parse == '?') { /* (?...) */
10729 bool is_logical = 0;
10730 const char * const seqstart = RExC_parse;
10731 const char * endptr;
10732 if (has_intervening_patws) {
10734 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
10737 RExC_parse++; /* past the '?' */
10738 paren = *RExC_parse; /* might be a trailing NUL, if not
10740 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10741 if (RExC_parse > RExC_end) {
10744 ret = NULL; /* For look-ahead/behind. */
10747 case 'P': /* (?P...) variants for those used to PCRE/Python */
10748 paren = *RExC_parse;
10749 if ( paren == '<') { /* (?P<...>) named capture */
10751 if (RExC_parse >= RExC_end) {
10752 vFAIL("Sequence (?P<... not terminated");
10754 goto named_capture;
10756 else if (paren == '>') { /* (?P>name) named recursion */
10758 if (RExC_parse >= RExC_end) {
10759 vFAIL("Sequence (?P>... not terminated");
10761 goto named_recursion;
10763 else if (paren == '=') { /* (?P=...) named backref */
10765 return handle_named_backref(pRExC_state, flagp,
10768 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10769 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10770 vFAIL3("Sequence (%.*s...) not recognized",
10771 RExC_parse-seqstart, seqstart);
10772 NOT_REACHED; /*NOTREACHED*/
10773 case '<': /* (?<...) */
10774 if (*RExC_parse == '!')
10776 else if (*RExC_parse != '=')
10783 case '\'': /* (?'...') */
10784 name_start = RExC_parse;
10785 svname = reg_scan_name(pRExC_state,
10786 SIZE_ONLY /* reverse test from the others */
10787 ? REG_RSN_RETURN_NAME
10788 : REG_RSN_RETURN_NULL);
10789 if ( RExC_parse == name_start
10790 || RExC_parse >= RExC_end
10791 || *RExC_parse != paren)
10793 vFAIL2("Sequence (?%c... not terminated",
10794 paren=='>' ? '<' : paren);
10799 if (!svname) /* shouldn't happen */
10801 "panic: reg_scan_name returned NULL");
10802 if (!RExC_paren_names) {
10803 RExC_paren_names= newHV();
10804 sv_2mortal(MUTABLE_SV(RExC_paren_names));
10806 RExC_paren_name_list= newAV();
10807 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10810 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10812 sv_dat = HeVAL(he_str);
10814 /* croak baby croak */
10816 "panic: paren_name hash element allocation failed");
10817 } else if ( SvPOK(sv_dat) ) {
10818 /* (?|...) can mean we have dupes so scan to check
10819 its already been stored. Maybe a flag indicating
10820 we are inside such a construct would be useful,
10821 but the arrays are likely to be quite small, so
10822 for now we punt -- dmq */
10823 IV count = SvIV(sv_dat);
10824 I32 *pv = (I32*)SvPVX(sv_dat);
10826 for ( i = 0 ; i < count ; i++ ) {
10827 if ( pv[i] == RExC_npar ) {
10833 pv = (I32*)SvGROW(sv_dat,
10834 SvCUR(sv_dat) + sizeof(I32)+1);
10835 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10836 pv[count] = RExC_npar;
10837 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10840 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10841 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10844 SvIV_set(sv_dat, 1);
10847 /* Yes this does cause a memory leak in debugging Perls
10849 if (!av_store(RExC_paren_name_list,
10850 RExC_npar, SvREFCNT_inc(svname)))
10851 SvREFCNT_dec_NN(svname);
10854 /*sv_dump(sv_dat);*/
10856 nextchar(pRExC_state);
10858 goto capturing_parens;
10860 RExC_seen |= REG_LOOKBEHIND_SEEN;
10861 RExC_in_lookbehind++;
10863 if (RExC_parse >= RExC_end) {
10864 vFAIL("Sequence (?... not terminated");
10868 case '=': /* (?=...) */
10869 RExC_seen_zerolen++;
10871 case '!': /* (?!...) */
10872 RExC_seen_zerolen++;
10873 /* check if we're really just a "FAIL" assertion */
10874 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
10875 FALSE /* Don't force to /x */ );
10876 if (*RExC_parse == ')') {
10877 ret=reganode(pRExC_state, OPFAIL, 0);
10878 nextchar(pRExC_state);
10882 case '|': /* (?|...) */
10883 /* branch reset, behave like a (?:...) except that
10884 buffers in alternations share the same numbers */
10886 after_freeze = freeze_paren = RExC_npar;
10888 case ':': /* (?:...) */
10889 case '>': /* (?>...) */
10891 case '$': /* (?$...) */
10892 case '@': /* (?@...) */
10893 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10895 case '0' : /* (?0) */
10896 case 'R' : /* (?R) */
10897 if (RExC_parse == RExC_end || *RExC_parse != ')')
10898 FAIL("Sequence (?R) not terminated");
10900 RExC_seen |= REG_RECURSE_SEEN;
10901 *flagp |= POSTPONED;
10902 goto gen_recurse_regop;
10904 /* named and numeric backreferences */
10905 case '&': /* (?&NAME) */
10906 parse_start = RExC_parse - 1;
10909 SV *sv_dat = reg_scan_name(pRExC_state,
10910 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10911 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10913 if (RExC_parse >= RExC_end || *RExC_parse != ')')
10914 vFAIL("Sequence (?&... not terminated");
10915 goto gen_recurse_regop;
10918 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10920 vFAIL("Illegal pattern");
10922 goto parse_recursion;
10924 case '-': /* (?-1) */
10925 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10926 RExC_parse--; /* rewind to let it be handled later */
10930 case '1': case '2': case '3': case '4': /* (?1) */
10931 case '5': case '6': case '7': case '8': case '9':
10932 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
10935 bool is_neg = FALSE;
10937 parse_start = RExC_parse - 1; /* MJD */
10938 if (*RExC_parse == '-') {
10942 if (grok_atoUV(RExC_parse, &unum, &endptr)
10946 RExC_parse = (char*)endptr;
10950 /* Some limit for num? */
10954 if (*RExC_parse!=')')
10955 vFAIL("Expecting close bracket");
10958 if ( paren == '-' ) {
10960 Diagram of capture buffer numbering.
10961 Top line is the normal capture buffer numbers
10962 Bottom line is the negative indexing as from
10966 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
10970 num = RExC_npar + num;
10973 vFAIL("Reference to nonexistent group");
10975 } else if ( paren == '+' ) {
10976 num = RExC_npar + num - 1;
10978 /* We keep track how many GOSUB items we have produced.
10979 To start off the ARG2L() of the GOSUB holds its "id",
10980 which is used later in conjunction with RExC_recurse
10981 to calculate the offset we need to jump for the GOSUB,
10982 which it will store in the final representation.
10983 We have to defer the actual calculation until much later
10984 as the regop may move.
10987 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
10989 if (num > (I32)RExC_rx->nparens) {
10991 vFAIL("Reference to nonexistent group");
10993 RExC_recurse_count++;
10994 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
10995 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
10996 22, "| |", (int)(depth * 2 + 1), "",
10997 (UV)ARG(ret), (IV)ARG2L(ret)));
10999 RExC_seen |= REG_RECURSE_SEEN;
11001 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
11002 Set_Node_Offset(ret, parse_start); /* MJD */
11004 *flagp |= POSTPONED;
11005 assert(*RExC_parse == ')');
11006 nextchar(pRExC_state);
11011 case '?': /* (??...) */
11013 if (*RExC_parse != '{') {
11014 RExC_parse += SKIP_IF_CHAR(RExC_parse);
11015 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11017 "Sequence (%" UTF8f "...) not recognized",
11018 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11019 NOT_REACHED; /*NOTREACHED*/
11021 *flagp |= POSTPONED;
11025 case '{': /* (?{...}) */
11028 struct reg_code_block *cb;
11030 RExC_seen_zerolen++;
11032 if ( !pRExC_state->code_blocks
11033 || pRExC_state->code_index
11034 >= pRExC_state->code_blocks->count
11035 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11036 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11039 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11040 FAIL("panic: Sequence (?{...}): no code block found\n");
11041 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11043 /* this is a pre-compiled code block (?{...}) */
11044 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11045 RExC_parse = RExC_start + cb->end;
11048 if (cb->src_regex) {
11049 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11050 RExC_rxi->data->data[n] =
11051 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11052 RExC_rxi->data->data[n+1] = (void*)o;
11055 n = add_data(pRExC_state,
11056 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11057 RExC_rxi->data->data[n] = (void*)o;
11060 pRExC_state->code_index++;
11061 nextchar(pRExC_state);
11065 ret = reg_node(pRExC_state, LOGICAL);
11067 eval = reg2Lanode(pRExC_state, EVAL,
11070 /* for later propagation into (??{})
11072 RExC_flags & RXf_PMf_COMPILETIME
11077 REGTAIL(pRExC_state, ret, eval);
11078 /* deal with the length of this later - MJD */
11081 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11082 Set_Node_Length(ret, RExC_parse - parse_start + 1);
11083 Set_Node_Offset(ret, parse_start);
11086 case '(': /* (?(?{...})...) and (?(?=...)...) */
11089 const int DEFINE_len = sizeof("DEFINE") - 1;
11090 if (RExC_parse[0] == '?') { /* (?(?...)) */
11091 if ( RExC_parse < RExC_end - 1
11092 && ( RExC_parse[1] == '='
11093 || RExC_parse[1] == '!'
11094 || RExC_parse[1] == '<'
11095 || RExC_parse[1] == '{')
11096 ) { /* Lookahead or eval. */
11100 ret = reg_node(pRExC_state, LOGICAL);
11104 tail = reg(pRExC_state, 1, &flag, depth+1);
11105 if (flag & (RESTART_PASS1|NEED_UTF8)) {
11106 *flagp = flag & (RESTART_PASS1|NEED_UTF8);
11109 REGTAIL(pRExC_state, ret, tail);
11112 /* Fall through to ‘Unknown switch condition’ at the
11113 end of the if/else chain. */
11115 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11116 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11118 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11119 char *name_start= RExC_parse++;
11121 SV *sv_dat=reg_scan_name(pRExC_state,
11122 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11123 if ( RExC_parse == name_start
11124 || RExC_parse >= RExC_end
11125 || *RExC_parse != ch)
11127 vFAIL2("Sequence (?(%c... not terminated",
11128 (ch == '>' ? '<' : ch));
11132 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11133 RExC_rxi->data->data[num]=(void*)sv_dat;
11134 SvREFCNT_inc_simple_void(sv_dat);
11136 ret = reganode(pRExC_state,NGROUPP,num);
11137 goto insert_if_check_paren;
11139 else if (RExC_end - RExC_parse >= DEFINE_len
11140 && strnEQ(RExC_parse, "DEFINE", DEFINE_len))
11142 ret = reganode(pRExC_state,DEFINEP,0);
11143 RExC_parse += DEFINE_len;
11145 goto insert_if_check_paren;
11147 else if (RExC_parse[0] == 'R') {
11149 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11150 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11151 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11154 if (RExC_parse[0] == '0') {
11158 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11160 if (grok_atoUV(RExC_parse, &uv, &endptr)
11163 parno = (I32)uv + 1;
11164 RExC_parse = (char*)endptr;
11166 /* else "Switch condition not recognized" below */
11167 } else if (RExC_parse[0] == '&') {
11170 sv_dat = reg_scan_name(pRExC_state,
11172 ? REG_RSN_RETURN_NULL
11173 : REG_RSN_RETURN_DATA);
11175 /* we should only have a false sv_dat when
11176 * SIZE_ONLY is true, and we always have false
11177 * sv_dat when SIZE_ONLY is true.
11178 * reg_scan_name() will VFAIL() if the name is
11179 * unknown when SIZE_ONLY is false, and otherwise
11180 * will return something, and when SIZE_ONLY is
11181 * true, reg_scan_name() just parses the string,
11182 * and doesnt return anything. (in theory) */
11183 assert(SIZE_ONLY ? !sv_dat : !!sv_dat);
11186 parno = 1 + *((I32 *)SvPVX(sv_dat));
11188 ret = reganode(pRExC_state,INSUBP,parno);
11189 goto insert_if_check_paren;
11191 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11195 if (grok_atoUV(RExC_parse, &uv, &endptr)
11199 RExC_parse = (char*)endptr;
11202 vFAIL("panic: grok_atoUV returned FALSE");
11204 ret = reganode(pRExC_state, GROUPP, parno);
11206 insert_if_check_paren:
11207 if (UCHARAT(RExC_parse) != ')') {
11208 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11209 vFAIL("Switch condition not recognized");
11211 nextchar(pRExC_state);
11213 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
11214 br = regbranch(pRExC_state, &flags, 1,depth+1);
11216 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11217 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11220 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11223 REGTAIL(pRExC_state, br, reganode(pRExC_state,
11225 c = UCHARAT(RExC_parse);
11226 nextchar(pRExC_state);
11227 if (flags&HASWIDTH)
11228 *flagp |= HASWIDTH;
11231 vFAIL("(?(DEFINE)....) does not allow branches");
11233 /* Fake one for optimizer. */
11234 lastbr = reganode(pRExC_state, IFTHEN, 0);
11236 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
11237 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11238 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11241 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11244 REGTAIL(pRExC_state, ret, lastbr);
11245 if (flags&HASWIDTH)
11246 *flagp |= HASWIDTH;
11247 c = UCHARAT(RExC_parse);
11248 nextchar(pRExC_state);
11253 if (RExC_parse >= RExC_end)
11254 vFAIL("Switch (?(condition)... not terminated");
11256 vFAIL("Switch (?(condition)... contains too many branches");
11258 ender = reg_node(pRExC_state, TAIL);
11259 REGTAIL(pRExC_state, br, ender);
11261 REGTAIL(pRExC_state, lastbr, ender);
11262 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11265 REGTAIL(pRExC_state, ret, ender);
11266 RExC_size++; /* XXX WHY do we need this?!!
11267 For large programs it seems to be required
11268 but I can't figure out why. -- dmq*/
11271 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11272 vFAIL("Unknown switch condition (?(...))");
11274 case '[': /* (?[ ... ]) */
11275 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
11277 case 0: /* A NUL */
11278 RExC_parse--; /* for vFAIL to print correctly */
11279 vFAIL("Sequence (? incomplete");
11281 default: /* e.g., (?i) */
11282 RExC_parse = (char *) seqstart + 1;
11284 parse_lparen_question_flags(pRExC_state);
11285 if (UCHARAT(RExC_parse) != ':') {
11286 if (RExC_parse < RExC_end)
11287 nextchar(pRExC_state);
11292 nextchar(pRExC_state);
11297 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11302 ret = reganode(pRExC_state, OPEN, parno);
11304 if (!RExC_nestroot)
11305 RExC_nestroot = parno;
11306 if (RExC_open_parens && !RExC_open_parens[parno])
11308 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11309 "%*s%*s Setting open paren #%" IVdf " to %d\n",
11310 22, "| |", (int)(depth * 2 + 1), "",
11311 (IV)parno, REG_NODE_NUM(ret)));
11312 RExC_open_parens[parno]= ret;
11315 Set_Node_Length(ret, 1); /* MJD */
11316 Set_Node_Offset(ret, RExC_parse); /* MJD */
11319 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
11328 /* Pick up the branches, linking them together. */
11329 parse_start = RExC_parse; /* MJD */
11330 br = regbranch(pRExC_state, &flags, 1,depth+1);
11332 /* branch_len = (paren != 0); */
11335 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11336 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11339 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11341 if (*RExC_parse == '|') {
11342 if (!SIZE_ONLY && RExC_extralen) {
11343 reginsert(pRExC_state, BRANCHJ, br, depth+1);
11346 reginsert(pRExC_state, BRANCH, br, depth+1);
11347 Set_Node_Length(br, paren != 0);
11348 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
11352 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
11354 else if (paren == ':') {
11355 *flagp |= flags&SIMPLE;
11357 if (is_open) { /* Starts with OPEN. */
11358 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
11360 else if (paren != '?') /* Not Conditional */
11362 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11364 while (*RExC_parse == '|') {
11365 if (!SIZE_ONLY && RExC_extralen) {
11366 ender = reganode(pRExC_state, LONGJMP,0);
11368 /* Append to the previous. */
11369 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11372 RExC_extralen += 2; /* Account for LONGJMP. */
11373 nextchar(pRExC_state);
11374 if (freeze_paren) {
11375 if (RExC_npar > after_freeze)
11376 after_freeze = RExC_npar;
11377 RExC_npar = freeze_paren;
11379 br = regbranch(pRExC_state, &flags, 0, depth+1);
11382 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11383 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11386 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11388 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
11390 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11393 if (have_branch || paren != ':') {
11394 /* Make a closing node, and hook it on the end. */
11397 ender = reg_node(pRExC_state, TAIL);
11400 ender = reganode(pRExC_state, CLOSE, parno);
11401 if ( RExC_close_parens ) {
11402 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11403 "%*s%*s Setting close paren #%" IVdf " to %d\n",
11404 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
11405 RExC_close_parens[parno]= ender;
11406 if (RExC_nestroot == parno)
11409 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
11410 Set_Node_Length(ender,1); /* MJD */
11416 *flagp &= ~HASWIDTH;
11419 ender = reg_node(pRExC_state, SUCCEED);
11422 ender = reg_node(pRExC_state, END);
11424 assert(!RExC_end_op); /* there can only be one! */
11425 RExC_end_op = ender;
11426 if (RExC_close_parens) {
11427 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11428 "%*s%*s Setting close paren #0 (END) to %d\n",
11429 22, "| |", (int)(depth * 2 + 1), "", REG_NODE_NUM(ender)));
11431 RExC_close_parens[0]= ender;
11436 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11437 DEBUG_PARSE_MSG("lsbr");
11438 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
11439 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11440 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11441 SvPV_nolen_const(RExC_mysv1),
11442 (IV)REG_NODE_NUM(lastbr),
11443 SvPV_nolen_const(RExC_mysv2),
11444 (IV)REG_NODE_NUM(ender),
11445 (IV)(ender - lastbr)
11448 REGTAIL(pRExC_state, lastbr, ender);
11450 if (have_branch && !SIZE_ONLY) {
11451 char is_nothing= 1;
11453 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
11455 /* Hook the tails of the branches to the closing node. */
11456 for (br = ret; br; br = regnext(br)) {
11457 const U8 op = PL_regkind[OP(br)];
11458 if (op == BRANCH) {
11459 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
11460 if ( OP(NEXTOPER(br)) != NOTHING
11461 || regnext(NEXTOPER(br)) != ender)
11464 else if (op == BRANCHJ) {
11465 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
11466 /* for now we always disable this optimisation * /
11467 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
11468 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
11474 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
11475 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11476 DEBUG_PARSE_MSG("NADA");
11477 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
11478 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11479 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11480 SvPV_nolen_const(RExC_mysv1),
11481 (IV)REG_NODE_NUM(ret),
11482 SvPV_nolen_const(RExC_mysv2),
11483 (IV)REG_NODE_NUM(ender),
11488 if (OP(ender) == TAIL) {
11493 for ( opt= br + 1; opt < ender ; opt++ )
11494 OP(opt)= OPTIMIZED;
11495 NEXT_OFF(br)= ender - br;
11503 static const char parens[] = "=!<,>";
11505 if (paren && (p = strchr(parens, paren))) {
11506 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
11507 int flag = (p - parens) > 1;
11510 node = SUSPEND, flag = 0;
11511 reginsert(pRExC_state, node,ret, depth+1);
11512 Set_Node_Cur_Length(ret, parse_start);
11513 Set_Node_Offset(ret, parse_start + 1);
11515 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
11519 /* Check for proper termination. */
11521 /* restore original flags, but keep (?p) and, if we've changed from /d
11522 * rules to /u, keep the /u */
11523 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
11524 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
11525 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
11527 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
11528 RExC_parse = oregcomp_parse;
11529 vFAIL("Unmatched (");
11531 nextchar(pRExC_state);
11533 else if (!paren && RExC_parse < RExC_end) {
11534 if (*RExC_parse == ')') {
11536 vFAIL("Unmatched )");
11539 FAIL("Junk on end of regexp"); /* "Can't happen". */
11540 NOT_REACHED; /* NOTREACHED */
11543 if (RExC_in_lookbehind) {
11544 RExC_in_lookbehind--;
11546 if (after_freeze > RExC_npar)
11547 RExC_npar = after_freeze;
11552 - regbranch - one alternative of an | operator
11554 * Implements the concatenation operator.
11556 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11557 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11560 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
11563 regnode *chain = NULL;
11565 I32 flags = 0, c = 0;
11566 GET_RE_DEBUG_FLAGS_DECL;
11568 PERL_ARGS_ASSERT_REGBRANCH;
11570 DEBUG_PARSE("brnc");
11575 if (!SIZE_ONLY && RExC_extralen)
11576 ret = reganode(pRExC_state, BRANCHJ,0);
11578 ret = reg_node(pRExC_state, BRANCH);
11579 Set_Node_Length(ret, 1);
11583 if (!first && SIZE_ONLY)
11584 RExC_extralen += 1; /* BRANCHJ */
11586 *flagp = WORST; /* Tentatively. */
11588 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11589 FALSE /* Don't force to /x */ );
11590 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
11591 flags &= ~TRYAGAIN;
11592 latest = regpiece(pRExC_state, &flags,depth+1);
11593 if (latest == NULL) {
11594 if (flags & TRYAGAIN)
11596 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11597 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11600 FAIL2("panic: regpiece returned NULL, flags=%#" UVxf, (UV) flags);
11602 else if (ret == NULL)
11604 *flagp |= flags&(HASWIDTH|POSTPONED);
11605 if (chain == NULL) /* First piece. */
11606 *flagp |= flags&SPSTART;
11608 /* FIXME adding one for every branch after the first is probably
11609 * excessive now we have TRIE support. (hv) */
11611 REGTAIL(pRExC_state, chain, latest);
11616 if (chain == NULL) { /* Loop ran zero times. */
11617 chain = reg_node(pRExC_state, NOTHING);
11622 *flagp |= flags&SIMPLE;
11629 - regpiece - something followed by possible quantifier * + ? {n,m}
11631 * Note that the branching code sequences used for ? and the general cases
11632 * of * and + are somewhat optimized: they use the same NOTHING node as
11633 * both the endmarker for their branch list and the body of the last branch.
11634 * It might seem that this node could be dispensed with entirely, but the
11635 * endmarker role is not redundant.
11637 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
11639 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11640 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11643 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11649 const char * const origparse = RExC_parse;
11651 I32 max = REG_INFTY;
11652 #ifdef RE_TRACK_PATTERN_OFFSETS
11655 const char *maxpos = NULL;
11658 /* Save the original in case we change the emitted regop to a FAIL. */
11659 regnode * const orig_emit = RExC_emit;
11661 GET_RE_DEBUG_FLAGS_DECL;
11663 PERL_ARGS_ASSERT_REGPIECE;
11665 DEBUG_PARSE("piec");
11667 ret = regatom(pRExC_state, &flags,depth+1);
11669 if (flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8))
11670 *flagp |= flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8);
11672 FAIL2("panic: regatom returned NULL, flags=%#" UVxf, (UV) flags);
11678 if (op == '{' && regcurly(RExC_parse)) {
11680 #ifdef RE_TRACK_PATTERN_OFFSETS
11681 parse_start = RExC_parse; /* MJD */
11683 next = RExC_parse + 1;
11684 while (isDIGIT(*next) || *next == ',') {
11685 if (*next == ',') {
11693 if (*next == '}') { /* got one */
11694 const char* endptr;
11698 if (isDIGIT(*RExC_parse)) {
11699 if (!grok_atoUV(RExC_parse, &uv, &endptr))
11700 vFAIL("Invalid quantifier in {,}");
11701 if (uv >= REG_INFTY)
11702 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11707 if (*maxpos == ',')
11710 maxpos = RExC_parse;
11711 if (isDIGIT(*maxpos)) {
11712 if (!grok_atoUV(maxpos, &uv, &endptr))
11713 vFAIL("Invalid quantifier in {,}");
11714 if (uv >= REG_INFTY)
11715 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11718 max = REG_INFTY; /* meaning "infinity" */
11721 nextchar(pRExC_state);
11722 if (max < min) { /* If can't match, warn and optimize to fail
11724 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
11726 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
11727 NEXT_OFF(orig_emit)= regarglen[OPFAIL] + NODE_STEP_REGNODE;
11731 else if (min == max && *RExC_parse == '?')
11734 ckWARN2reg(RExC_parse + 1,
11735 "Useless use of greediness modifier '%c'",
11741 if ((flags&SIMPLE)) {
11742 if (min == 0 && max == REG_INFTY) {
11743 reginsert(pRExC_state, STAR, ret, depth+1);
11746 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11749 if (min == 1 && max == REG_INFTY) {
11750 reginsert(pRExC_state, PLUS, ret, depth+1);
11753 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11756 MARK_NAUGHTY_EXP(2, 2);
11757 reginsert(pRExC_state, CURLY, ret, depth+1);
11758 Set_Node_Offset(ret, parse_start+1); /* MJD */
11759 Set_Node_Cur_Length(ret, parse_start);
11762 regnode * const w = reg_node(pRExC_state, WHILEM);
11765 REGTAIL(pRExC_state, ret, w);
11766 if (!SIZE_ONLY && RExC_extralen) {
11767 reginsert(pRExC_state, LONGJMP,ret, depth+1);
11768 reginsert(pRExC_state, NOTHING,ret, depth+1);
11769 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
11771 reginsert(pRExC_state, CURLYX,ret, depth+1);
11773 Set_Node_Offset(ret, parse_start+1);
11774 Set_Node_Length(ret,
11775 op == '{' ? (RExC_parse - parse_start) : 1);
11777 if (!SIZE_ONLY && RExC_extralen)
11778 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
11779 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
11781 RExC_whilem_seen++, RExC_extralen += 3;
11782 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
11789 *flagp |= HASWIDTH;
11791 ARG1_SET(ret, (U16)min);
11792 ARG2_SET(ret, (U16)max);
11794 if (max == REG_INFTY)
11795 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11801 if (!ISMULT1(op)) {
11806 #if 0 /* Now runtime fix should be reliable. */
11808 /* if this is reinstated, don't forget to put this back into perldiag:
11810 =item Regexp *+ operand could be empty at {#} in regex m/%s/
11812 (F) The part of the regexp subject to either the * or + quantifier
11813 could match an empty string. The {#} shows in the regular
11814 expression about where the problem was discovered.
11818 if (!(flags&HASWIDTH) && op != '?')
11819 vFAIL("Regexp *+ operand could be empty");
11822 #ifdef RE_TRACK_PATTERN_OFFSETS
11823 parse_start = RExC_parse;
11825 nextchar(pRExC_state);
11827 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
11833 else if (op == '+') {
11837 else if (op == '?') {
11842 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
11843 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
11844 ckWARN2reg(RExC_parse,
11845 "%" UTF8f " matches null string many times",
11846 UTF8fARG(UTF, (RExC_parse >= origparse
11847 ? RExC_parse - origparse
11850 (void)ReREFCNT_inc(RExC_rx_sv);
11853 if (*RExC_parse == '?') {
11854 nextchar(pRExC_state);
11855 reginsert(pRExC_state, MINMOD, ret, depth+1);
11856 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
11858 else if (*RExC_parse == '+') {
11860 nextchar(pRExC_state);
11861 ender = reg_node(pRExC_state, SUCCEED);
11862 REGTAIL(pRExC_state, ret, ender);
11863 reginsert(pRExC_state, SUSPEND, ret, depth+1);
11865 ender = reg_node(pRExC_state, TAIL);
11866 REGTAIL(pRExC_state, ret, ender);
11869 if (ISMULT2(RExC_parse)) {
11871 vFAIL("Nested quantifiers");
11878 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
11887 /* This routine teases apart the various meanings of \N and returns
11888 * accordingly. The input parameters constrain which meaning(s) is/are valid
11889 * in the current context.
11891 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
11893 * If <code_point_p> is not NULL, the context is expecting the result to be a
11894 * single code point. If this \N instance turns out to a single code point,
11895 * the function returns TRUE and sets *code_point_p to that code point.
11897 * If <node_p> is not NULL, the context is expecting the result to be one of
11898 * the things representable by a regnode. If this \N instance turns out to be
11899 * one such, the function generates the regnode, returns TRUE and sets *node_p
11900 * to point to that regnode.
11902 * If this instance of \N isn't legal in any context, this function will
11903 * generate a fatal error and not return.
11905 * On input, RExC_parse should point to the first char following the \N at the
11906 * time of the call. On successful return, RExC_parse will have been updated
11907 * to point to just after the sequence identified by this routine. Also
11908 * *flagp has been updated as needed.
11910 * When there is some problem with the current context and this \N instance,
11911 * the function returns FALSE, without advancing RExC_parse, nor setting
11912 * *node_p, nor *code_point_p, nor *flagp.
11914 * If <cp_count> is not NULL, the caller wants to know the length (in code
11915 * points) that this \N sequence matches. This is set even if the function
11916 * returns FALSE, as detailed below.
11918 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
11920 * Probably the most common case is for the \N to specify a single code point.
11921 * *cp_count will be set to 1, and *code_point_p will be set to that code
11924 * Another possibility is for the input to be an empty \N{}, which for
11925 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
11926 * will be set to a generated NOTHING node.
11928 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
11929 * set to 0. *node_p will be set to a generated REG_ANY node.
11931 * The fourth possibility is that \N resolves to a sequence of more than one
11932 * code points. *cp_count will be set to the number of code points in the
11933 * sequence. *node_p * will be set to a generated node returned by this
11934 * function calling S_reg().
11936 * The final possibility is that it is premature to be calling this function;
11937 * that pass1 needs to be restarted. This can happen when this changes from
11938 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
11939 * latter occurs only when the fourth possibility would otherwise be in
11940 * effect, and is because one of those code points requires the pattern to be
11941 * recompiled as UTF-8. The function returns FALSE, and sets the
11942 * RESTART_PASS1 and NEED_UTF8 flags in *flagp, as appropriate. When this
11943 * happens, the caller needs to desist from continuing parsing, and return
11944 * this information to its caller. This is not set for when there is only one
11945 * code point, as this can be called as part of an ANYOF node, and they can
11946 * store above-Latin1 code points without the pattern having to be in UTF-8.
11948 * For non-single-quoted regexes, the tokenizer has resolved character and
11949 * sequence names inside \N{...} into their Unicode values, normalizing the
11950 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
11951 * hex-represented code points in the sequence. This is done there because
11952 * the names can vary based on what charnames pragma is in scope at the time,
11953 * so we need a way to take a snapshot of what they resolve to at the time of
11954 * the original parse. [perl #56444].
11956 * That parsing is skipped for single-quoted regexes, so we may here get
11957 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
11958 * parser. But if the single-quoted regex is something like '\N{U+41}', that
11959 * is legal and handled here. The code point is Unicode, and has to be
11960 * translated into the native character set for non-ASCII platforms.
11963 char * endbrace; /* points to '}' following the name */
11964 char *endchar; /* Points to '.' or '}' ending cur char in the input
11966 char* p = RExC_parse; /* Temporary */
11968 GET_RE_DEBUG_FLAGS_DECL;
11970 PERL_ARGS_ASSERT_GROK_BSLASH_N;
11972 GET_RE_DEBUG_FLAGS;
11974 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
11975 assert(! (node_p && cp_count)); /* At most 1 should be set */
11977 if (cp_count) { /* Initialize return for the most common case */
11981 /* The [^\n] meaning of \N ignores spaces and comments under the /x
11982 * modifier. The other meanings do not, so use a temporary until we find
11983 * out which we are being called with */
11984 skip_to_be_ignored_text(pRExC_state, &p,
11985 FALSE /* Don't force to /x */ );
11987 /* Disambiguate between \N meaning a named character versus \N meaning
11988 * [^\n]. The latter is assumed when the {...} following the \N is a legal
11989 * quantifier, or there is no '{' at all */
11990 if (*p != '{' || regcurly(p)) {
12000 *node_p = reg_node(pRExC_state, REG_ANY);
12001 *flagp |= HASWIDTH|SIMPLE;
12003 Set_Node_Length(*node_p, 1); /* MJD */
12007 /* Here, we have decided it should be a named character or sequence */
12009 /* The test above made sure that the next real character is a '{', but
12010 * under the /x modifier, it could be separated by space (or a comment and
12011 * \n) and this is not allowed (for consistency with \x{...} and the
12012 * tokenizer handling of \N{NAME}). */
12013 if (*RExC_parse != '{') {
12014 vFAIL("Missing braces on \\N{}");
12017 RExC_parse++; /* Skip past the '{' */
12019 if (! (endbrace = strchr(RExC_parse, '}'))) { /* no trailing brace */
12020 vFAIL2("Missing right brace on \\%c{}", 'N');
12022 else if(!(endbrace == RExC_parse /* nothing between the {} */
12023 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked... */
12024 && strnEQ(RExC_parse, "U+", 2)))) /* ... below for a better
12027 RExC_parse = endbrace; /* position msg's '<--HERE' */
12028 vFAIL("\\N{NAME} must be resolved by the lexer");
12031 REQUIRE_UNI_RULES(flagp, FALSE); /* Unicode named chars imply Unicode
12034 if (endbrace == RExC_parse) { /* empty: \N{} */
12036 RExC_parse++; /* Position after the "}" */
12037 vFAIL("Zero length \\N{}");
12042 nextchar(pRExC_state);
12047 *node_p = reg_node(pRExC_state,NOTHING);
12051 RExC_parse += 2; /* Skip past the 'U+' */
12053 /* Because toke.c has generated a special construct for us guaranteed not
12054 * to have NULs, we can use a str function */
12055 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12057 /* Code points are separated by dots. If none, there is only one code
12058 * point, and is terminated by the brace */
12060 if (endchar >= endbrace) {
12061 STRLEN length_of_hex;
12062 I32 grok_hex_flags;
12064 /* Here, exactly one code point. If that isn't what is wanted, fail */
12065 if (! code_point_p) {
12070 /* Convert code point from hex */
12071 length_of_hex = (STRLEN)(endchar - RExC_parse);
12072 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
12073 | PERL_SCAN_DISALLOW_PREFIX
12075 /* No errors in the first pass (See [perl
12076 * #122671].) We let the code below find the
12077 * errors when there are multiple chars. */
12079 ? PERL_SCAN_SILENT_ILLDIGIT
12082 /* This routine is the one place where both single- and double-quotish
12083 * \N{U+xxxx} are evaluated. The value is a Unicode code point which
12084 * must be converted to native. */
12085 *code_point_p = UNI_TO_NATIVE(grok_hex(RExC_parse,
12090 /* The tokenizer should have guaranteed validity, but it's possible to
12091 * bypass it by using single quoting, so check. Don't do the check
12092 * here when there are multiple chars; we do it below anyway. */
12093 if (length_of_hex == 0
12094 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
12096 RExC_parse += length_of_hex; /* Includes all the valid */
12097 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12098 ? UTF8SKIP(RExC_parse)
12100 /* Guard against malformed utf8 */
12101 if (RExC_parse >= endchar) {
12102 RExC_parse = endchar;
12104 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12107 RExC_parse = endbrace + 1;
12110 else { /* Is a multiple character sequence */
12111 SV * substitute_parse;
12113 char *orig_end = RExC_end;
12114 char *save_start = RExC_start;
12117 /* Count the code points, if desired, in the sequence */
12120 while (RExC_parse < endbrace) {
12121 /* Point to the beginning of the next character in the sequence. */
12122 RExC_parse = endchar + 1;
12123 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12128 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12129 * But don't backup up the pointer if the caller want to know how many
12130 * code points there are (they can then handle things) */
12138 /* What is done here is to convert this to a sub-pattern of the form
12139 * \x{char1}\x{char2}... and then call reg recursively to parse it
12140 * (enclosing in "(?: ... )" ). That way, it retains its atomicness,
12141 * while not having to worry about special handling that some code
12142 * points may have. */
12144 substitute_parse = newSVpvs("?:");
12146 while (RExC_parse < endbrace) {
12148 /* Convert to notation the rest of the code understands */
12149 sv_catpv(substitute_parse, "\\x{");
12150 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
12151 sv_catpv(substitute_parse, "}");
12153 /* Point to the beginning of the next character in the sequence. */
12154 RExC_parse = endchar + 1;
12155 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12158 sv_catpv(substitute_parse, ")");
12160 RExC_parse = RExC_start = RExC_adjusted_start = SvPV(substitute_parse,
12163 /* Don't allow empty number */
12164 if (len < (STRLEN) 8) {
12165 RExC_parse = endbrace;
12166 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12168 RExC_end = RExC_parse + len;
12170 /* The values are Unicode, and therefore not subject to recoding, but
12171 * have to be converted to native on a non-Unicode (meaning non-ASCII)
12174 RExC_recode_x_to_native = 1;
12178 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
12179 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12180 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12183 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#" UVxf,
12186 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12189 /* Restore the saved values */
12190 RExC_start = RExC_adjusted_start = save_start;
12191 RExC_parse = endbrace;
12192 RExC_end = orig_end;
12194 RExC_recode_x_to_native = 0;
12197 SvREFCNT_dec_NN(substitute_parse);
12198 nextchar(pRExC_state);
12205 PERL_STATIC_INLINE U8
12206 S_compute_EXACTish(RExC_state_t *pRExC_state)
12210 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
12218 op = get_regex_charset(RExC_flags);
12219 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
12220 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
12221 been, so there is no hole */
12224 return op + EXACTF;
12227 PERL_STATIC_INLINE void
12228 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
12229 regnode *node, I32* flagp, STRLEN len, UV code_point,
12232 /* This knows the details about sizing an EXACTish node, setting flags for
12233 * it (by setting <*flagp>, and potentially populating it with a single
12236 * If <len> (the length in bytes) is non-zero, this function assumes that
12237 * the node has already been populated, and just does the sizing. In this
12238 * case <code_point> should be the final code point that has already been
12239 * placed into the node. This value will be ignored except that under some
12240 * circumstances <*flagp> is set based on it.
12242 * If <len> is zero, the function assumes that the node is to contain only
12243 * the single character given by <code_point> and calculates what <len>
12244 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
12245 * additionally will populate the node's STRING with <code_point> or its
12248 * In both cases <*flagp> is appropriately set
12250 * It knows that under FOLD, the Latin Sharp S and UTF characters above
12251 * 255, must be folded (the former only when the rules indicate it can
12254 * When it does the populating, it looks at the flag 'downgradable'. If
12255 * true with a node that folds, it checks if the single code point
12256 * participates in a fold, and if not downgrades the node to an EXACT.
12257 * This helps the optimizer */
12259 bool len_passed_in = cBOOL(len != 0);
12260 U8 character[UTF8_MAXBYTES_CASE+1];
12262 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
12264 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
12265 * sizing difference, and is extra work that is thrown away */
12266 if (downgradable && ! PASS2) {
12267 downgradable = FALSE;
12270 if (! len_passed_in) {
12272 if (UVCHR_IS_INVARIANT(code_point)) {
12273 if (LOC || ! FOLD) { /* /l defers folding until runtime */
12274 *character = (U8) code_point;
12276 else { /* Here is /i and not /l. (toFOLD() is defined on just
12277 ASCII, which isn't the same thing as INVARIANT on
12278 EBCDIC, but it works there, as the extra invariants
12279 fold to themselves) */
12280 *character = toFOLD((U8) code_point);
12282 /* We can downgrade to an EXACT node if this character
12283 * isn't a folding one. Note that this assumes that
12284 * nothing above Latin1 folds to some other invariant than
12285 * one of these alphabetics; otherwise we would also have
12287 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12288 * || ASCII_FOLD_RESTRICTED))
12290 if (downgradable && PL_fold[code_point] == code_point) {
12296 else if (FOLD && (! LOC
12297 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
12298 { /* Folding, and ok to do so now */
12299 UV folded = _to_uni_fold_flags(
12303 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12304 ? FOLD_FLAGS_NOMIX_ASCII
12307 && folded == code_point /* This quickly rules out many
12308 cases, avoiding the
12309 _invlist_contains_cp() overhead
12311 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
12318 else if (code_point <= MAX_UTF8_TWO_BYTE) {
12320 /* Not folding this cp, and can output it directly */
12321 *character = UTF8_TWO_BYTE_HI(code_point);
12322 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
12326 uvchr_to_utf8( character, code_point);
12327 len = UTF8SKIP(character);
12329 } /* Else pattern isn't UTF8. */
12331 *character = (U8) code_point;
12333 } /* Else is folded non-UTF8 */
12334 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12335 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12336 || UNICODE_DOT_DOT_VERSION > 0)
12337 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
12341 /* We don't fold any non-UTF8 except possibly the Sharp s (see
12342 * comments at join_exact()); */
12343 *character = (U8) code_point;
12346 /* Can turn into an EXACT node if we know the fold at compile time,
12347 * and it folds to itself and doesn't particpate in other folds */
12350 && PL_fold_latin1[code_point] == code_point
12351 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12352 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
12356 } /* else is Sharp s. May need to fold it */
12357 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
12359 *(character + 1) = 's';
12363 *character = LATIN_SMALL_LETTER_SHARP_S;
12369 RExC_size += STR_SZ(len);
12372 RExC_emit += STR_SZ(len);
12373 STR_LEN(node) = len;
12374 if (! len_passed_in) {
12375 Copy((char *) character, STRING(node), len, char);
12379 *flagp |= HASWIDTH;
12381 /* A single character node is SIMPLE, except for the special-cased SHARP S
12383 if ((len == 1 || (UTF && len == UVCHR_SKIP(code_point)))
12384 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12385 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12386 || UNICODE_DOT_DOT_VERSION > 0)
12387 && ( code_point != LATIN_SMALL_LETTER_SHARP_S
12388 || ! FOLD || ! DEPENDS_SEMANTICS)
12394 /* The OP may not be well defined in PASS1 */
12395 if (PASS2 && OP(node) == EXACTFL) {
12396 RExC_contains_locale = 1;
12401 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
12402 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
12405 S_backref_value(char *p)
12407 const char* endptr;
12409 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
12416 - regatom - the lowest level
12418 Try to identify anything special at the start of the current parse position.
12419 If there is, then handle it as required. This may involve generating a
12420 single regop, such as for an assertion; or it may involve recursing, such as
12421 to handle a () structure.
12423 If the string doesn't start with something special then we gobble up
12424 as much literal text as we can. If we encounter a quantifier, we have to
12425 back off the final literal character, as that quantifier applies to just it
12426 and not to the whole string of literals.
12428 Once we have been able to handle whatever type of thing started the
12429 sequence, we return.
12431 Note: we have to be careful with escapes, as they can be both literal
12432 and special, and in the case of \10 and friends, context determines which.
12434 A summary of the code structure is:
12436 switch (first_byte) {
12437 cases for each special:
12438 handle this special;
12441 switch (2nd byte) {
12442 cases for each unambiguous special:
12443 handle this special;
12445 cases for each ambigous special/literal:
12447 if (special) handle here
12449 default: // unambiguously literal:
12452 default: // is a literal char
12455 create EXACTish node for literal;
12456 while (more input and node isn't full) {
12457 switch (input_byte) {
12458 cases for each special;
12459 make sure parse pointer is set so that the next call to
12460 regatom will see this special first
12461 goto loopdone; // EXACTish node terminated by prev. char
12463 append char to EXACTISH node;
12465 get next input byte;
12469 return the generated node;
12471 Specifically there are two separate switches for handling
12472 escape sequences, with the one for handling literal escapes requiring
12473 a dummy entry for all of the special escapes that are actually handled
12476 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
12478 Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
12479 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
12480 Otherwise does not return NULL.
12484 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12486 regnode *ret = NULL;
12493 GET_RE_DEBUG_FLAGS_DECL;
12495 *flagp = WORST; /* Tentatively. */
12497 DEBUG_PARSE("atom");
12499 PERL_ARGS_ASSERT_REGATOM;
12502 parse_start = RExC_parse;
12503 assert(RExC_parse < RExC_end);
12504 switch ((U8)*RExC_parse) {
12506 RExC_seen_zerolen++;
12507 nextchar(pRExC_state);
12508 if (RExC_flags & RXf_PMf_MULTILINE)
12509 ret = reg_node(pRExC_state, MBOL);
12511 ret = reg_node(pRExC_state, SBOL);
12512 Set_Node_Length(ret, 1); /* MJD */
12515 nextchar(pRExC_state);
12517 RExC_seen_zerolen++;
12518 if (RExC_flags & RXf_PMf_MULTILINE)
12519 ret = reg_node(pRExC_state, MEOL);
12521 ret = reg_node(pRExC_state, SEOL);
12522 Set_Node_Length(ret, 1); /* MJD */
12525 nextchar(pRExC_state);
12526 if (RExC_flags & RXf_PMf_SINGLELINE)
12527 ret = reg_node(pRExC_state, SANY);
12529 ret = reg_node(pRExC_state, REG_ANY);
12530 *flagp |= HASWIDTH|SIMPLE;
12532 Set_Node_Length(ret, 1); /* MJD */
12536 char * const oregcomp_parse = ++RExC_parse;
12537 ret = regclass(pRExC_state, flagp,depth+1,
12538 FALSE, /* means parse the whole char class */
12539 TRUE, /* allow multi-char folds */
12540 FALSE, /* don't silence non-portable warnings. */
12541 (bool) RExC_strict,
12542 TRUE, /* Allow an optimized regnode result */
12546 if (*flagp & (RESTART_PASS1|NEED_UTF8))
12548 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12551 if (*RExC_parse != ']') {
12552 RExC_parse = oregcomp_parse;
12553 vFAIL("Unmatched [");
12555 nextchar(pRExC_state);
12556 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
12560 nextchar(pRExC_state);
12561 ret = reg(pRExC_state, 2, &flags,depth+1);
12563 if (flags & TRYAGAIN) {
12564 if (RExC_parse >= RExC_end) {
12565 /* Make parent create an empty node if needed. */
12566 *flagp |= TRYAGAIN;
12571 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12572 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12575 FAIL2("panic: reg returned NULL to regatom, flags=%#" UVxf,
12578 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12582 if (flags & TRYAGAIN) {
12583 *flagp |= TRYAGAIN;
12586 vFAIL("Internal urp");
12587 /* Supposed to be caught earlier. */
12593 vFAIL("Quantifier follows nothing");
12598 This switch handles escape sequences that resolve to some kind
12599 of special regop and not to literal text. Escape sequnces that
12600 resolve to literal text are handled below in the switch marked
12603 Every entry in this switch *must* have a corresponding entry
12604 in the literal escape switch. However, the opposite is not
12605 required, as the default for this switch is to jump to the
12606 literal text handling code.
12609 switch ((U8)*RExC_parse) {
12610 /* Special Escapes */
12612 RExC_seen_zerolen++;
12613 ret = reg_node(pRExC_state, SBOL);
12614 /* SBOL is shared with /^/ so we set the flags so we can tell
12615 * /\A/ from /^/ in split. We check ret because first pass we
12616 * have no regop struct to set the flags on. */
12620 goto finish_meta_pat;
12622 ret = reg_node(pRExC_state, GPOS);
12623 RExC_seen |= REG_GPOS_SEEN;
12625 goto finish_meta_pat;
12627 RExC_seen_zerolen++;
12628 ret = reg_node(pRExC_state, KEEPS);
12630 /* XXX:dmq : disabling in-place substitution seems to
12631 * be necessary here to avoid cases of memory corruption, as
12632 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
12634 RExC_seen |= REG_LOOKBEHIND_SEEN;
12635 goto finish_meta_pat;
12637 ret = reg_node(pRExC_state, SEOL);
12639 RExC_seen_zerolen++; /* Do not optimize RE away */
12640 goto finish_meta_pat;
12642 ret = reg_node(pRExC_state, EOS);
12644 RExC_seen_zerolen++; /* Do not optimize RE away */
12645 goto finish_meta_pat;
12647 vFAIL("\\C no longer supported");
12649 ret = reg_node(pRExC_state, CLUMP);
12650 *flagp |= HASWIDTH;
12651 goto finish_meta_pat;
12657 arg = ANYOF_WORDCHAR;
12665 regex_charset charset = get_regex_charset(RExC_flags);
12667 RExC_seen_zerolen++;
12668 RExC_seen |= REG_LOOKBEHIND_SEEN;
12669 op = BOUND + charset;
12671 if (op == BOUNDL) {
12672 RExC_contains_locale = 1;
12675 ret = reg_node(pRExC_state, op);
12677 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
12678 FLAGS(ret) = TRADITIONAL_BOUND;
12679 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
12685 char name = *RExC_parse;
12688 endbrace = strchr(RExC_parse, '}');
12691 vFAIL2("Missing right brace on \\%c{}", name);
12693 /* XXX Need to decide whether to take spaces or not. Should be
12694 * consistent with \p{}, but that currently is SPACE, which
12695 * means vertical too, which seems wrong
12696 * while (isBLANK(*RExC_parse)) {
12699 if (endbrace == RExC_parse) {
12700 RExC_parse++; /* After the '}' */
12701 vFAIL2("Empty \\%c{}", name);
12703 length = endbrace - RExC_parse;
12704 /*while (isBLANK(*(RExC_parse + length - 1))) {
12707 switch (*RExC_parse) {
12710 && (length != 3 || strnNE(RExC_parse + 1, "cb", 2)))
12712 goto bad_bound_type;
12714 FLAGS(ret) = GCB_BOUND;
12717 if (length != 2 || *(RExC_parse + 1) != 'b') {
12718 goto bad_bound_type;
12720 FLAGS(ret) = LB_BOUND;
12723 if (length != 2 || *(RExC_parse + 1) != 'b') {
12724 goto bad_bound_type;
12726 FLAGS(ret) = SB_BOUND;
12729 if (length != 2 || *(RExC_parse + 1) != 'b') {
12730 goto bad_bound_type;
12732 FLAGS(ret) = WB_BOUND;
12736 RExC_parse = endbrace;
12738 "'%" UTF8f "' is an unknown bound type",
12739 UTF8fARG(UTF, length, endbrace - length));
12740 NOT_REACHED; /*NOTREACHED*/
12742 RExC_parse = endbrace;
12743 REQUIRE_UNI_RULES(flagp, NULL);
12745 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
12749 /* Don't have to worry about UTF-8, in this message because
12750 * to get here the contents of the \b must be ASCII */
12751 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
12752 "Using /u for '%.*s' instead of /%s",
12754 endbrace - length + 1,
12755 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
12756 ? ASCII_RESTRICT_PAT_MODS
12757 : ASCII_MORE_RESTRICT_PAT_MODS);
12761 if (PASS2 && invert) {
12762 OP(ret) += NBOUND - BOUND;
12764 goto finish_meta_pat;
12772 if (! DEPENDS_SEMANTICS) {
12776 /* \d doesn't have any matches in the upper Latin1 range, hence /d
12777 * is equivalent to /u. Changing to /u saves some branches at
12780 goto join_posix_op_known;
12783 ret = reg_node(pRExC_state, LNBREAK);
12784 *flagp |= HASWIDTH|SIMPLE;
12785 goto finish_meta_pat;
12793 goto join_posix_op_known;
12799 arg = ANYOF_VERTWS;
12801 goto join_posix_op_known;
12811 op = POSIXD + get_regex_charset(RExC_flags);
12812 if (op > POSIXA) { /* /aa is same as /a */
12815 else if (op == POSIXL) {
12816 RExC_contains_locale = 1;
12819 join_posix_op_known:
12822 op += NPOSIXD - POSIXD;
12825 ret = reg_node(pRExC_state, op);
12827 FLAGS(ret) = namedclass_to_classnum(arg);
12830 *flagp |= HASWIDTH|SIMPLE;
12834 nextchar(pRExC_state);
12835 Set_Node_Length(ret, 2); /* MJD */
12841 ret = regclass(pRExC_state, flagp,depth+1,
12842 TRUE, /* means just parse this element */
12843 FALSE, /* don't allow multi-char folds */
12844 FALSE, /* don't silence non-portable warnings. It
12845 would be a bug if these returned
12847 (bool) RExC_strict,
12848 TRUE, /* Allow an optimized regnode result */
12851 if (*flagp & RESTART_PASS1)
12853 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
12854 * multi-char folds are allowed. */
12856 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12861 Set_Node_Offset(ret, parse_start);
12862 Set_Node_Cur_Length(ret, parse_start - 2);
12863 nextchar(pRExC_state);
12866 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
12867 * \N{...} evaluates to a sequence of more than one code points).
12868 * The function call below returns a regnode, which is our result.
12869 * The parameters cause it to fail if the \N{} evaluates to a
12870 * single code point; we handle those like any other literal. The
12871 * reason that the multicharacter case is handled here and not as
12872 * part of the EXACtish code is because of quantifiers. In
12873 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
12874 * this way makes that Just Happen. dmq.
12875 * join_exact() will join this up with adjacent EXACTish nodes
12876 * later on, if appropriate. */
12878 if (grok_bslash_N(pRExC_state,
12879 &ret, /* Want a regnode returned */
12880 NULL, /* Fail if evaluates to a single code
12882 NULL, /* Don't need a count of how many code
12891 if (*flagp & RESTART_PASS1)
12894 /* Here, evaluates to a single code point. Go get that */
12895 RExC_parse = parse_start;
12898 case 'k': /* Handle \k<NAME> and \k'NAME' */
12902 if ( RExC_parse >= RExC_end - 1
12903 || (( ch = RExC_parse[1]) != '<'
12908 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12909 vFAIL2("Sequence %.2s... not terminated",parse_start);
12912 ret = handle_named_backref(pRExC_state,
12924 case '1': case '2': case '3': case '4':
12925 case '5': case '6': case '7': case '8': case '9':
12930 if (*RExC_parse == 'g') {
12934 if (*RExC_parse == '{') {
12938 if (*RExC_parse == '-') {
12942 if (hasbrace && !isDIGIT(*RExC_parse)) {
12943 if (isrel) RExC_parse--;
12945 goto parse_named_seq;
12948 if (RExC_parse >= RExC_end) {
12949 goto unterminated_g;
12951 num = S_backref_value(RExC_parse);
12953 vFAIL("Reference to invalid group 0");
12954 else if (num == I32_MAX) {
12955 if (isDIGIT(*RExC_parse))
12956 vFAIL("Reference to nonexistent group");
12959 vFAIL("Unterminated \\g... pattern");
12963 num = RExC_npar - num;
12965 vFAIL("Reference to nonexistent or unclosed group");
12969 num = S_backref_value(RExC_parse);
12970 /* bare \NNN might be backref or octal - if it is larger
12971 * than or equal RExC_npar then it is assumed to be an
12972 * octal escape. Note RExC_npar is +1 from the actual
12973 * number of parens. */
12974 /* Note we do NOT check if num == I32_MAX here, as that is
12975 * handled by the RExC_npar check */
12978 /* any numeric escape < 10 is always a backref */
12980 /* any numeric escape < RExC_npar is a backref */
12981 && num >= RExC_npar
12982 /* cannot be an octal escape if it starts with 8 */
12983 && *RExC_parse != '8'
12984 /* cannot be an octal escape it it starts with 9 */
12985 && *RExC_parse != '9'
12988 /* Probably not a backref, instead likely to be an
12989 * octal character escape, e.g. \35 or \777.
12990 * The above logic should make it obvious why using
12991 * octal escapes in patterns is problematic. - Yves */
12992 RExC_parse = parse_start;
12997 /* At this point RExC_parse points at a numeric escape like
12998 * \12 or \88 or something similar, which we should NOT treat
12999 * as an octal escape. It may or may not be a valid backref
13000 * escape. For instance \88888888 is unlikely to be a valid
13002 while (isDIGIT(*RExC_parse))
13005 if (*RExC_parse != '}')
13006 vFAIL("Unterminated \\g{...} pattern");
13010 if (num > (I32)RExC_rx->nparens)
13011 vFAIL("Reference to nonexistent group");
13014 ret = reganode(pRExC_state,
13017 : (ASCII_FOLD_RESTRICTED)
13019 : (AT_LEAST_UNI_SEMANTICS)
13025 *flagp |= HASWIDTH;
13027 /* override incorrect value set in reganode MJD */
13028 Set_Node_Offset(ret, parse_start);
13029 Set_Node_Cur_Length(ret, parse_start-1);
13030 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13031 FALSE /* Don't force to /x */ );
13035 if (RExC_parse >= RExC_end)
13036 FAIL("Trailing \\");
13039 /* Do not generate "unrecognized" warnings here, we fall
13040 back into the quick-grab loop below */
13041 RExC_parse = parse_start;
13043 } /* end of switch on a \foo sequence */
13048 /* '#' comments should have been spaced over before this function was
13050 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13052 if (RExC_flags & RXf_PMf_EXTENDED) {
13053 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13054 if (RExC_parse < RExC_end)
13064 /* Here, we have determined that the next thing is probably a
13065 * literal character. RExC_parse points to the first byte of its
13066 * definition. (It still may be an escape sequence that evaluates
13067 * to a single character) */
13073 #define MAX_NODE_STRING_SIZE 127
13074 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
13076 U8 upper_parse = MAX_NODE_STRING_SIZE;
13077 U8 node_type = compute_EXACTish(pRExC_state);
13078 bool next_is_quantifier;
13079 char * oldp = NULL;
13081 /* We can convert EXACTF nodes to EXACTFU if they contain only
13082 * characters that match identically regardless of the target
13083 * string's UTF8ness. The reason to do this is that EXACTF is not
13084 * trie-able, EXACTFU is.
13086 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13087 * contain only above-Latin1 characters (hence must be in UTF8),
13088 * which don't participate in folds with Latin1-range characters,
13089 * as the latter's folds aren't known until runtime. (We don't
13090 * need to figure this out until pass 2) */
13091 bool maybe_exactfu = PASS2
13092 && (node_type == EXACTF || node_type == EXACTFL);
13094 /* If a folding node contains only code points that don't
13095 * participate in folds, it can be changed into an EXACT node,
13096 * which allows the optimizer more things to look for */
13099 ret = reg_node(pRExC_state, node_type);
13101 /* In pass1, folded, we use a temporary buffer instead of the
13102 * actual node, as the node doesn't exist yet */
13103 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
13109 /* We look for the EXACTFish to EXACT node optimizaton only if
13110 * folding. (And we don't need to figure this out until pass 2).
13111 * XXX It might actually make sense to split the node into portions
13112 * that are exact and ones that aren't, so that we could later use
13113 * the exact ones to find the longest fixed and floating strings.
13114 * One would want to join them back into a larger node. One could
13115 * use a pseudo regnode like 'EXACT_ORIG_FOLD' */
13116 maybe_exact = FOLD && PASS2;
13118 /* XXX The node can hold up to 255 bytes, yet this only goes to
13119 * 127. I (khw) do not know why. Keeping it somewhat less than
13120 * 255 allows us to not have to worry about overflow due to
13121 * converting to utf8 and fold expansion, but that value is
13122 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
13123 * split up by this limit into a single one using the real max of
13124 * 255. Even at 127, this breaks under rare circumstances. If
13125 * folding, we do not want to split a node at a character that is a
13126 * non-final in a multi-char fold, as an input string could just
13127 * happen to want to match across the node boundary. The join
13128 * would solve that problem if the join actually happens. But a
13129 * series of more than two nodes in a row each of 127 would cause
13130 * the first join to succeed to get to 254, but then there wouldn't
13131 * be room for the next one, which could at be one of those split
13132 * multi-char folds. I don't know of any fool-proof solution. One
13133 * could back off to end with only a code point that isn't such a
13134 * non-final, but it is possible for there not to be any in the
13137 assert( ! UTF /* Is at the beginning of a character */
13138 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13139 || UTF8_IS_START(UCHARAT(RExC_parse)));
13141 /* Here, we have a literal character. Find the maximal string of
13142 * them in the input that we can fit into a single EXACTish node.
13143 * We quit at the first non-literal or when the node gets full */
13144 for (p = RExC_parse;
13145 len < upper_parse && p < RExC_end;
13150 /* White space has already been ignored */
13151 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13152 || ! is_PATWS_safe((p), RExC_end, UTF));
13164 /* Literal Escapes Switch
13166 This switch is meant to handle escape sequences that
13167 resolve to a literal character.
13169 Every escape sequence that represents something
13170 else, like an assertion or a char class, is handled
13171 in the switch marked 'Special Escapes' above in this
13172 routine, but also has an entry here as anything that
13173 isn't explicitly mentioned here will be treated as
13174 an unescaped equivalent literal.
13177 switch ((U8)*++p) {
13178 /* These are all the special escapes. */
13179 case 'A': /* Start assertion */
13180 case 'b': case 'B': /* Word-boundary assertion*/
13181 case 'C': /* Single char !DANGEROUS! */
13182 case 'd': case 'D': /* digit class */
13183 case 'g': case 'G': /* generic-backref, pos assertion */
13184 case 'h': case 'H': /* HORIZWS */
13185 case 'k': case 'K': /* named backref, keep marker */
13186 case 'p': case 'P': /* Unicode property */
13187 case 'R': /* LNBREAK */
13188 case 's': case 'S': /* space class */
13189 case 'v': case 'V': /* VERTWS */
13190 case 'w': case 'W': /* word class */
13191 case 'X': /* eXtended Unicode "combining
13192 character sequence" */
13193 case 'z': case 'Z': /* End of line/string assertion */
13197 /* Anything after here is an escape that resolves to a
13198 literal. (Except digits, which may or may not)
13204 case 'N': /* Handle a single-code point named character. */
13205 RExC_parse = p + 1;
13206 if (! grok_bslash_N(pRExC_state,
13207 NULL, /* Fail if evaluates to
13208 anything other than a
13209 single code point */
13210 &ender, /* The returned single code
13212 NULL, /* Don't need a count of
13213 how many code points */
13218 if (*flagp & NEED_UTF8)
13219 FAIL("panic: grok_bslash_N set NEED_UTF8");
13220 if (*flagp & RESTART_PASS1)
13223 /* Here, it wasn't a single code point. Go close
13224 * up this EXACTish node. The switch() prior to
13225 * this switch handles the other cases */
13226 RExC_parse = p = oldp;
13230 if (ender > 0xff) {
13231 REQUIRE_UTF8(flagp);
13247 ender = ESC_NATIVE;
13257 const char* error_msg;
13259 bool valid = grok_bslash_o(&p,
13262 PASS2, /* out warnings */
13263 (bool) RExC_strict,
13264 TRUE, /* Output warnings
13269 RExC_parse = p; /* going to die anyway; point
13270 to exact spot of failure */
13274 if (ender > 0xff) {
13275 REQUIRE_UTF8(flagp);
13281 UV result = UV_MAX; /* initialize to erroneous
13283 const char* error_msg;
13285 bool valid = grok_bslash_x(&p,
13288 PASS2, /* out warnings */
13289 (bool) RExC_strict,
13290 TRUE, /* Silence warnings
13295 RExC_parse = p; /* going to die anyway; point
13296 to exact spot of failure */
13301 if (ender < 0x100) {
13303 if (RExC_recode_x_to_native) {
13304 ender = LATIN1_TO_NATIVE(ender);
13309 REQUIRE_UTF8(flagp);
13315 ender = grok_bslash_c(*p++, PASS2);
13317 case '8': case '9': /* must be a backreference */
13319 /* we have an escape like \8 which cannot be an octal escape
13320 * so we exit the loop, and let the outer loop handle this
13321 * escape which may or may not be a legitimate backref. */
13323 case '1': case '2': case '3':case '4':
13324 case '5': case '6': case '7':
13325 /* When we parse backslash escapes there is ambiguity
13326 * between backreferences and octal escapes. Any escape
13327 * from \1 - \9 is a backreference, any multi-digit
13328 * escape which does not start with 0 and which when
13329 * evaluated as decimal could refer to an already
13330 * parsed capture buffer is a back reference. Anything
13333 * Note this implies that \118 could be interpreted as
13334 * 118 OR as "\11" . "8" depending on whether there
13335 * were 118 capture buffers defined already in the
13338 /* NOTE, RExC_npar is 1 more than the actual number of
13339 * parens we have seen so far, hence the < RExC_npar below. */
13341 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
13342 { /* Not to be treated as an octal constant, go
13350 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13352 ender = grok_oct(p, &numlen, &flags, NULL);
13353 if (ender > 0xff) {
13354 REQUIRE_UTF8(flagp);
13357 if (PASS2 /* like \08, \178 */
13359 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
13361 reg_warn_non_literal_string(
13363 form_short_octal_warning(p, numlen));
13369 FAIL("Trailing \\");
13372 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
13373 /* Include any left brace following the alpha to emphasize
13374 * that it could be part of an escape at some point
13376 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
13377 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
13379 goto normal_default;
13380 } /* End of switch on '\' */
13383 /* Currently we don't care if the lbrace is at the start
13384 * of a construct. This catches it in the middle of a
13385 * literal string, or when it's the first thing after
13386 * something like "\b" */
13387 if (len || (p > RExC_start && isALPHA_A(*(p -1)))) {
13388 RExC_parse = p + 1;
13389 vFAIL("Unescaped left brace in regex is illegal here");
13391 goto normal_default;
13394 if (PASS2 && p > RExC_parse && RExC_strict) {
13395 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
13398 default: /* A literal character */
13400 if (! UTF8_IS_INVARIANT(*p) && UTF) {
13402 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
13403 &numlen, UTF8_ALLOW_DEFAULT);
13409 } /* End of switch on the literal */
13411 /* Here, have looked at the literal character and <ender>
13412 * contains its ordinal, <p> points to the character after it.
13413 * We need to check if the next non-ignored thing is a
13414 * quantifier. Move <p> to after anything that should be
13415 * ignored, which, as a side effect, positions <p> for the next
13416 * loop iteration */
13417 skip_to_be_ignored_text(pRExC_state, &p,
13418 FALSE /* Don't force to /x */ );
13420 /* If the next thing is a quantifier, it applies to this
13421 * character only, which means that this character has to be in
13422 * its own node and can't just be appended to the string in an
13423 * existing node, so if there are already other characters in
13424 * the node, close the node with just them, and set up to do
13425 * this character again next time through, when it will be the
13426 * only thing in its new node */
13428 if ((next_is_quantifier = ( LIKELY(p < RExC_end)
13429 && UNLIKELY(ISMULT2(p))))
13436 /* Ready to add 'ender' to the node */
13438 if (! FOLD) { /* The simple case, just append the literal */
13440 /* In the sizing pass, we need only the size of the
13441 * character we are appending, hence we can delay getting
13442 * its representation until PASS2. */
13445 const STRLEN unilen = UVCHR_SKIP(ender);
13448 /* We have to subtract 1 just below (and again in
13449 * the corresponding PASS2 code) because the loop
13450 * increments <len> each time, as all but this path
13451 * (and one other) through it add a single byte to
13452 * the EXACTish node. But these paths would change
13453 * len to be the correct final value, so cancel out
13454 * the increment that follows */
13460 } else { /* PASS2 */
13463 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
13464 len += (char *) new_s - s - 1;
13465 s = (char *) new_s;
13468 *(s++) = (char) ender;
13472 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
13474 /* Here are folding under /l, and the code point is
13475 * problematic. First, we know we can't simplify things */
13476 maybe_exact = FALSE;
13477 maybe_exactfu = FALSE;
13479 /* A problematic code point in this context means that its
13480 * fold isn't known until runtime, so we can't fold it now.
13481 * (The non-problematic code points are the above-Latin1
13482 * ones that fold to also all above-Latin1. Their folds
13483 * don't vary no matter what the locale is.) But here we
13484 * have characters whose fold depends on the locale.
13485 * Unlike the non-folding case above, we have to keep track
13486 * of these in the sizing pass, so that we can make sure we
13487 * don't split too-long nodes in the middle of a potential
13488 * multi-char fold. And unlike the regular fold case
13489 * handled in the else clauses below, we don't actually
13490 * fold and don't have special cases to consider. What we
13491 * do for both passes is the PASS2 code for non-folding */
13492 goto not_fold_common;
13494 else /* A regular FOLD code point */
13496 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13497 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13498 || UNICODE_DOT_DOT_VERSION > 0)
13499 /* See comments for join_exact() as to why we fold
13500 * this non-UTF at compile time */
13501 || ( node_type == EXACTFU
13502 && ender == LATIN_SMALL_LETTER_SHARP_S)
13505 /* Here, are folding and are not UTF-8 encoded; therefore
13506 * the character must be in the range 0-255, and is not /l
13507 * (Not /l because we already handled these under /l in
13508 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
13509 if (IS_IN_SOME_FOLD_L1(ender)) {
13510 maybe_exact = FALSE;
13512 /* See if the character's fold differs between /d and
13513 * /u. This includes the multi-char fold SHARP S to
13515 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
13516 RExC_seen_unfolded_sharp_s = 1;
13517 maybe_exactfu = FALSE;
13519 else if (maybe_exactfu
13520 && (PL_fold[ender] != PL_fold_latin1[ender]
13521 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13522 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13523 || UNICODE_DOT_DOT_VERSION > 0)
13525 && isALPHA_FOLD_EQ(ender, 's')
13526 && isALPHA_FOLD_EQ(*(s-1), 's'))
13529 maybe_exactfu = FALSE;
13533 /* Even when folding, we store just the input character, as
13534 * we have an array that finds its fold quickly */
13535 *(s++) = (char) ender;
13537 else { /* FOLD, and UTF (or sharp s) */
13538 /* Unlike the non-fold case, we do actually have to
13539 * calculate the results here in pass 1. This is for two
13540 * reasons, the folded length may be longer than the
13541 * unfolded, and we have to calculate how many EXACTish
13542 * nodes it will take; and we may run out of room in a node
13543 * in the middle of a potential multi-char fold, and have
13544 * to back off accordingly. */
13547 if (isASCII_uni(ender)) {
13548 folded = toFOLD(ender);
13549 *(s)++ = (U8) folded;
13554 folded = _to_uni_fold_flags(
13558 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
13559 ? FOLD_FLAGS_NOMIX_ASCII
13563 /* The loop increments <len> each time, as all but this
13564 * path (and one other) through it add a single byte to
13565 * the EXACTish node. But this one has changed len to
13566 * be the correct final value, so subtract one to
13567 * cancel out the increment that follows */
13568 len += foldlen - 1;
13570 /* If this node only contains non-folding code points so
13571 * far, see if this new one is also non-folding */
13573 if (folded != ender) {
13574 maybe_exact = FALSE;
13577 /* Here the fold is the original; we have to check
13578 * further to see if anything folds to it */
13579 if (_invlist_contains_cp(PL_utf8_foldable,
13582 maybe_exact = FALSE;
13589 if (next_is_quantifier) {
13591 /* Here, the next input is a quantifier, and to get here,
13592 * the current character is the only one in the node.
13593 * Also, here <len> doesn't include the final byte for this
13599 } /* End of loop through literal characters */
13601 /* Here we have either exhausted the input or ran out of room in
13602 * the node. (If we encountered a character that can't be in the
13603 * node, transfer is made directly to <loopdone>, and so we
13604 * wouldn't have fallen off the end of the loop.) In the latter
13605 * case, we artificially have to split the node into two, because
13606 * we just don't have enough space to hold everything. This
13607 * creates a problem if the final character participates in a
13608 * multi-character fold in the non-final position, as a match that
13609 * should have occurred won't, due to the way nodes are matched,
13610 * and our artificial boundary. So back off until we find a non-
13611 * problematic character -- one that isn't at the beginning or
13612 * middle of such a fold. (Either it doesn't participate in any
13613 * folds, or appears only in the final position of all the folds it
13614 * does participate in.) A better solution with far fewer false
13615 * positives, and that would fill the nodes more completely, would
13616 * be to actually have available all the multi-character folds to
13617 * test against, and to back-off only far enough to be sure that
13618 * this node isn't ending with a partial one. <upper_parse> is set
13619 * further below (if we need to reparse the node) to include just
13620 * up through that final non-problematic character that this code
13621 * identifies, so when it is set to less than the full node, we can
13622 * skip the rest of this */
13623 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
13625 const STRLEN full_len = len;
13627 assert(len >= MAX_NODE_STRING_SIZE);
13629 /* Here, <s> points to the final byte of the final character.
13630 * Look backwards through the string until find a non-
13631 * problematic character */
13635 /* This has no multi-char folds to non-UTF characters */
13636 if (ASCII_FOLD_RESTRICTED) {
13640 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
13644 if (! PL_NonL1NonFinalFold) {
13645 PL_NonL1NonFinalFold = _new_invlist_C_array(
13646 NonL1_Perl_Non_Final_Folds_invlist);
13649 /* Point to the first byte of the final character */
13650 s = (char *) utf8_hop((U8 *) s, -1);
13652 while (s >= s0) { /* Search backwards until find
13653 non-problematic char */
13654 if (UTF8_IS_INVARIANT(*s)) {
13656 /* There are no ascii characters that participate
13657 * in multi-char folds under /aa. In EBCDIC, the
13658 * non-ascii invariants are all control characters,
13659 * so don't ever participate in any folds. */
13660 if (ASCII_FOLD_RESTRICTED
13661 || ! IS_NON_FINAL_FOLD(*s))
13666 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
13667 if (! IS_NON_FINAL_FOLD(EIGHT_BIT_UTF8_TO_NATIVE(
13673 else if (! _invlist_contains_cp(
13674 PL_NonL1NonFinalFold,
13675 valid_utf8_to_uvchr((U8 *) s, NULL)))
13680 /* Here, the current character is problematic in that
13681 * it does occur in the non-final position of some
13682 * fold, so try the character before it, but have to
13683 * special case the very first byte in the string, so
13684 * we don't read outside the string */
13685 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
13686 } /* End of loop backwards through the string */
13688 /* If there were only problematic characters in the string,
13689 * <s> will point to before s0, in which case the length
13690 * should be 0, otherwise include the length of the
13691 * non-problematic character just found */
13692 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
13695 /* Here, have found the final character, if any, that is
13696 * non-problematic as far as ending the node without splitting
13697 * it across a potential multi-char fold. <len> contains the
13698 * number of bytes in the node up-to and including that
13699 * character, or is 0 if there is no such character, meaning
13700 * the whole node contains only problematic characters. In
13701 * this case, give up and just take the node as-is. We can't
13706 /* If the node ends in an 's' we make sure it stays EXACTF,
13707 * as if it turns into an EXACTFU, it could later get
13708 * joined with another 's' that would then wrongly match
13710 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
13712 maybe_exactfu = FALSE;
13716 /* Here, the node does contain some characters that aren't
13717 * problematic. If one such is the final character in the
13718 * node, we are done */
13719 if (len == full_len) {
13722 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
13724 /* If the final character is problematic, but the
13725 * penultimate is not, back-off that last character to
13726 * later start a new node with it */
13731 /* Here, the final non-problematic character is earlier
13732 * in the input than the penultimate character. What we do
13733 * is reparse from the beginning, going up only as far as
13734 * this final ok one, thus guaranteeing that the node ends
13735 * in an acceptable character. The reason we reparse is
13736 * that we know how far in the character is, but we don't
13737 * know how to correlate its position with the input parse.
13738 * An alternate implementation would be to build that
13739 * correlation as we go along during the original parse,
13740 * but that would entail extra work for every node, whereas
13741 * this code gets executed only when the string is too
13742 * large for the node, and the final two characters are
13743 * problematic, an infrequent occurrence. Yet another
13744 * possible strategy would be to save the tail of the
13745 * string, and the next time regatom is called, initialize
13746 * with that. The problem with this is that unless you
13747 * back off one more character, you won't be guaranteed
13748 * regatom will get called again, unless regbranch,
13749 * regpiece ... are also changed. If you do back off that
13750 * extra character, so that there is input guaranteed to
13751 * force calling regatom, you can't handle the case where
13752 * just the first character in the node is acceptable. I
13753 * (khw) decided to try this method which doesn't have that
13754 * pitfall; if performance issues are found, we can do a
13755 * combination of the current approach plus that one */
13761 } /* End of verifying node ends with an appropriate char */
13763 loopdone: /* Jumped to when encounters something that shouldn't be
13766 /* I (khw) don't know if you can get here with zero length, but the
13767 * old code handled this situation by creating a zero-length EXACT
13768 * node. Might as well be NOTHING instead */
13774 /* If 'maybe_exact' is still set here, means there are no
13775 * code points in the node that participate in folds;
13776 * similarly for 'maybe_exactfu' and code points that match
13777 * differently depending on UTF8ness of the target string
13778 * (for /u), or depending on locale for /l */
13784 else if (maybe_exactfu) {
13790 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
13791 FALSE /* Don't look to see if could
13792 be turned into an EXACT
13793 node, as we have already
13798 RExC_parse = p - 1;
13799 Set_Node_Cur_Length(ret, parse_start);
13802 /* len is STRLEN which is unsigned, need to copy to signed */
13805 vFAIL("Internal disaster");
13808 } /* End of label 'defchar:' */
13810 } /* End of giant switch on input character */
13812 /* Position parse to next real character */
13813 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13814 FALSE /* Don't force to /x */ );
13815 if (PASS2 && *RExC_parse == '{' && OP(ret) != SBOL && ! regcurly(RExC_parse)) {
13816 ckWARNregdep(RExC_parse + 1, "Unescaped left brace in regex is deprecated here (and will be fatal in Perl 5.30), passed through");
13824 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
13826 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
13827 * sets up the bitmap and any flags, removing those code points from the
13828 * inversion list, setting it to NULL should it become completely empty */
13830 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
13831 assert(PL_regkind[OP(node)] == ANYOF);
13833 ANYOF_BITMAP_ZERO(node);
13834 if (*invlist_ptr) {
13836 /* This gets set if we actually need to modify things */
13837 bool change_invlist = FALSE;
13841 /* Start looking through *invlist_ptr */
13842 invlist_iterinit(*invlist_ptr);
13843 while (invlist_iternext(*invlist_ptr, &start, &end)) {
13847 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
13848 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
13851 /* Quit if are above what we should change */
13852 if (start >= NUM_ANYOF_CODE_POINTS) {
13856 change_invlist = TRUE;
13858 /* Set all the bits in the range, up to the max that we are doing */
13859 high = (end < NUM_ANYOF_CODE_POINTS - 1)
13861 : NUM_ANYOF_CODE_POINTS - 1;
13862 for (i = start; i <= (int) high; i++) {
13863 if (! ANYOF_BITMAP_TEST(node, i)) {
13864 ANYOF_BITMAP_SET(node, i);
13868 invlist_iterfinish(*invlist_ptr);
13870 /* Done with loop; remove any code points that are in the bitmap from
13871 * *invlist_ptr; similarly for code points above the bitmap if we have
13872 * a flag to match all of them anyways */
13873 if (change_invlist) {
13874 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
13876 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
13877 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
13880 /* If have completely emptied it, remove it completely */
13881 if (_invlist_len(*invlist_ptr) == 0) {
13882 SvREFCNT_dec_NN(*invlist_ptr);
13883 *invlist_ptr = NULL;
13888 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
13889 Character classes ([:foo:]) can also be negated ([:^foo:]).
13890 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
13891 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
13892 but trigger failures because they are currently unimplemented. */
13894 #define POSIXCC_DONE(c) ((c) == ':')
13895 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
13896 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
13897 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
13899 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
13900 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
13901 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
13903 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
13905 /* 'posix_warnings' and 'warn_text' are names of variables in the following
13907 #define ADD_POSIX_WARNING(p, text) STMT_START { \
13908 if (posix_warnings) { \
13909 if (! RExC_warn_text ) RExC_warn_text = (AV *) sv_2mortal((SV *) newAV()); \
13910 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
13914 REPORT_LOCATION_ARGS(p))); \
13919 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
13921 const char * const s, /* Where the putative posix class begins.
13922 Normally, this is one past the '['. This
13923 parameter exists so it can be somewhere
13924 besides RExC_parse. */
13925 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
13927 AV ** posix_warnings, /* Where to place any generated warnings, or
13929 const bool check_only /* Don't die if error */
13932 /* This parses what the caller thinks may be one of the three POSIX
13934 * 1) a character class, like [:blank:]
13935 * 2) a collating symbol, like [. .]
13936 * 3) an equivalence class, like [= =]
13937 * In the latter two cases, it croaks if it finds a syntactically legal
13938 * one, as these are not handled by Perl.
13940 * The main purpose is to look for a POSIX character class. It returns:
13941 * a) the class number
13942 * if it is a completely syntactically and semantically legal class.
13943 * 'updated_parse_ptr', if not NULL, is set to point to just after the
13944 * closing ']' of the class
13945 * b) OOB_NAMEDCLASS
13946 * if it appears that one of the three POSIX constructs was meant, but
13947 * its specification was somehow defective. 'updated_parse_ptr', if
13948 * not NULL, is set to point to the character just after the end
13949 * character of the class. See below for handling of warnings.
13950 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
13951 * if it doesn't appear that a POSIX construct was intended.
13952 * 'updated_parse_ptr' is not changed. No warnings nor errors are
13955 * In b) there may be errors or warnings generated. If 'check_only' is
13956 * TRUE, then any errors are discarded. Warnings are returned to the
13957 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
13958 * instead it is NULL, warnings are suppressed. This is done in all
13959 * passes. The reason for this is that the rest of the parsing is heavily
13960 * dependent on whether this routine found a valid posix class or not. If
13961 * it did, the closing ']' is absorbed as part of the class. If no class,
13962 * or an invalid one is found, any ']' will be considered the terminator of
13963 * the outer bracketed character class, leading to very different results.
13964 * In particular, a '(?[ ])' construct will likely have a syntax error if
13965 * the class is parsed other than intended, and this will happen in pass1,
13966 * before the warnings would normally be output. This mechanism allows the
13967 * caller to output those warnings in pass1 just before dieing, giving a
13968 * much better clue as to what is wrong.
13970 * The reason for this function, and its complexity is that a bracketed
13971 * character class can contain just about anything. But it's easy to
13972 * mistype the very specific posix class syntax but yielding a valid
13973 * regular bracketed class, so it silently gets compiled into something
13974 * quite unintended.
13976 * The solution adopted here maintains backward compatibility except that
13977 * it adds a warning if it looks like a posix class was intended but
13978 * improperly specified. The warning is not raised unless what is input
13979 * very closely resembles one of the 14 legal posix classes. To do this,
13980 * it uses fuzzy parsing. It calculates how many single-character edits it
13981 * would take to transform what was input into a legal posix class. Only
13982 * if that number is quite small does it think that the intention was a
13983 * posix class. Obviously these are heuristics, and there will be cases
13984 * where it errs on one side or another, and they can be tweaked as
13985 * experience informs.
13987 * The syntax for a legal posix class is:
13989 * qr/(?xa: \[ : \^? [:lower:]{4,6} : \] )/
13991 * What this routine considers syntactically to be an intended posix class
13992 * is this (the comments indicate some restrictions that the pattern
13995 * qr/(?x: \[? # The left bracket, possibly
13997 * \h* # possibly followed by blanks
13998 * (?: \^ \h* )? # possibly a misplaced caret
13999 * [:;]? # The opening class character,
14000 * # possibly omitted. A typo
14001 * # semi-colon can also be used.
14003 * \^? # possibly a correctly placed
14004 * # caret, but not if there was also
14005 * # a misplaced one
14007 * .{3,15} # The class name. If there are
14008 * # deviations from the legal syntax,
14009 * # its edit distance must be close
14010 * # to a real class name in order
14011 * # for it to be considered to be
14012 * # an intended posix class.
14014 * [:punct:]? # The closing class character,
14015 * # possibly omitted. If not a colon
14016 * # nor semi colon, the class name
14017 * # must be even closer to a valid
14020 * \]? # The right bracket, possibly
14024 * In the above, \h must be ASCII-only.
14026 * These are heuristics, and can be tweaked as field experience dictates.
14027 * There will be cases when someone didn't intend to specify a posix class
14028 * that this warns as being so. The goal is to minimize these, while
14029 * maximizing the catching of things intended to be a posix class that
14030 * aren't parsed as such.
14034 const char * const e = RExC_end;
14035 unsigned complement = 0; /* If to complement the class */
14036 bool found_problem = FALSE; /* Assume OK until proven otherwise */
14037 bool has_opening_bracket = FALSE;
14038 bool has_opening_colon = FALSE;
14039 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
14041 const char * possible_end = NULL; /* used for a 2nd parse pass */
14042 const char* name_start; /* ptr to class name first char */
14044 /* If the number of single-character typos the input name is away from a
14045 * legal name is no more than this number, it is considered to have meant
14046 * the legal name */
14047 int max_distance = 2;
14049 /* to store the name. The size determines the maximum length before we
14050 * decide that no posix class was intended. Should be at least
14051 * sizeof("alphanumeric") */
14054 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
14056 if (posix_warnings && RExC_warn_text)
14057 av_clear(RExC_warn_text);
14060 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14063 if (*(p - 1) != '[') {
14064 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
14065 found_problem = TRUE;
14068 has_opening_bracket = TRUE;
14071 /* They could be confused and think you can put spaces between the
14074 found_problem = TRUE;
14078 } while (p < e && isBLANK(*p));
14080 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14083 /* For [. .] and [= =]. These are quite different internally from [: :],
14084 * so they are handled separately. */
14085 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
14086 and 1 for at least one char in it
14089 const char open_char = *p;
14090 const char * temp_ptr = p + 1;
14092 /* These two constructs are not handled by perl, and if we find a
14093 * syntactically valid one, we croak. khw, who wrote this code, finds
14094 * this explanation of them very unclear:
14095 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
14096 * And searching the rest of the internet wasn't very helpful either.
14097 * It looks like just about any byte can be in these constructs,
14098 * depending on the locale. But unless the pattern is being compiled
14099 * under /l, which is very rare, Perl runs under the C or POSIX locale.
14100 * In that case, it looks like [= =] isn't allowed at all, and that
14101 * [. .] could be any single code point, but for longer strings the
14102 * constituent characters would have to be the ASCII alphabetics plus
14103 * the minus-hyphen. Any sensible locale definition would limit itself
14104 * to these. And any portable one definitely should. Trying to parse
14105 * the general case is a nightmare (see [perl #127604]). So, this code
14106 * looks only for interiors of these constructs that match:
14108 * Using \w relaxes the apparent rules a little, without adding much
14109 * danger of mistaking something else for one of these constructs.
14111 * [. .] in some implementations described on the internet is usable to
14112 * escape a character that otherwise is special in bracketed character
14113 * classes. For example [.].] means a literal right bracket instead of
14114 * the ending of the class
14116 * [= =] can legitimately contain a [. .] construct, but we don't
14117 * handle this case, as that [. .] construct will later get parsed
14118 * itself and croak then. And [= =] is checked for even when not under
14119 * /l, as Perl has long done so.
14121 * The code below relies on there being a trailing NUL, so it doesn't
14122 * have to keep checking if the parse ptr < e.
14124 if (temp_ptr[1] == open_char) {
14127 else while ( temp_ptr < e
14128 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
14133 if (*temp_ptr == open_char) {
14135 if (*temp_ptr == ']') {
14137 if (! found_problem && ! check_only) {
14138 RExC_parse = (char *) temp_ptr;
14139 vFAIL3("POSIX syntax [%c %c] is reserved for future "
14140 "extensions", open_char, open_char);
14143 /* Here, the syntax wasn't completely valid, or else the call
14144 * is to check-only */
14145 if (updated_parse_ptr) {
14146 *updated_parse_ptr = (char *) temp_ptr;
14149 return OOB_NAMEDCLASS;
14153 /* If we find something that started out to look like one of these
14154 * constructs, but isn't, we continue below so that it can be checked
14155 * for being a class name with a typo of '.' or '=' instead of a colon.
14159 /* Here, we think there is a possibility that a [: :] class was meant, and
14160 * we have the first real character. It could be they think the '^' comes
14163 found_problem = TRUE;
14164 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
14169 found_problem = TRUE;
14173 } while (p < e && isBLANK(*p));
14175 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14179 /* But the first character should be a colon, which they could have easily
14180 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
14181 * distinguish from a colon, so treat that as a colon). */
14184 has_opening_colon = TRUE;
14186 else if (*p == ';') {
14187 found_problem = TRUE;
14189 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14190 has_opening_colon = TRUE;
14193 found_problem = TRUE;
14194 ADD_POSIX_WARNING(p, "there must be a starting ':'");
14196 /* Consider an initial punctuation (not one of the recognized ones) to
14197 * be a left terminator */
14198 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
14203 /* They may think that you can put spaces between the components */
14205 found_problem = TRUE;
14209 } while (p < e && isBLANK(*p));
14211 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14216 /* We consider something like [^:^alnum:]] to not have been intended to
14217 * be a posix class, but XXX maybe we should */
14219 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14226 /* Again, they may think that you can put spaces between the components */
14228 found_problem = TRUE;
14232 } while (p < e && isBLANK(*p));
14234 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14239 /* XXX This ']' may be a typo, and something else was meant. But
14240 * treating it as such creates enough complications, that that
14241 * possibility isn't currently considered here. So we assume that the
14242 * ']' is what is intended, and if we've already found an initial '[',
14243 * this leaves this construct looking like [:] or [:^], which almost
14244 * certainly weren't intended to be posix classes */
14245 if (has_opening_bracket) {
14246 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14249 /* But this function can be called when we parse the colon for
14250 * something like qr/[alpha:]]/, so we back up to look for the
14255 found_problem = TRUE;
14256 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14258 else if (*p != ':') {
14260 /* XXX We are currently very restrictive here, so this code doesn't
14261 * consider the possibility that, say, /[alpha.]]/ was intended to
14262 * be a posix class. */
14263 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14266 /* Here we have something like 'foo:]'. There was no initial colon,
14267 * and we back up over 'foo. XXX Unlike the going forward case, we
14268 * don't handle typos of non-word chars in the middle */
14269 has_opening_colon = FALSE;
14272 while (p > RExC_start && isWORDCHAR(*p)) {
14277 /* Here, we have positioned ourselves to where we think the first
14278 * character in the potential class is */
14281 /* Now the interior really starts. There are certain key characters that
14282 * can end the interior, or these could just be typos. To catch both
14283 * cases, we may have to do two passes. In the first pass, we keep on
14284 * going unless we come to a sequence that matches
14285 * qr/ [[:punct:]] [[:blank:]]* \] /xa
14286 * This means it takes a sequence to end the pass, so two typos in a row if
14287 * that wasn't what was intended. If the class is perfectly formed, just
14288 * this one pass is needed. We also stop if there are too many characters
14289 * being accumulated, but this number is deliberately set higher than any
14290 * real class. It is set high enough so that someone who thinks that
14291 * 'alphanumeric' is a correct name would get warned that it wasn't.
14292 * While doing the pass, we keep track of where the key characters were in
14293 * it. If we don't find an end to the class, and one of the key characters
14294 * was found, we redo the pass, but stop when we get to that character.
14295 * Thus the key character was considered a typo in the first pass, but a
14296 * terminator in the second. If two key characters are found, we stop at
14297 * the second one in the first pass. Again this can miss two typos, but
14298 * catches a single one
14300 * In the first pass, 'possible_end' starts as NULL, and then gets set to
14301 * point to the first key character. For the second pass, it starts as -1.
14307 bool has_blank = FALSE;
14308 bool has_upper = FALSE;
14309 bool has_terminating_colon = FALSE;
14310 bool has_terminating_bracket = FALSE;
14311 bool has_semi_colon = FALSE;
14312 unsigned int name_len = 0;
14313 int punct_count = 0;
14317 /* Squeeze out blanks when looking up the class name below */
14318 if (isBLANK(*p) ) {
14320 found_problem = TRUE;
14325 /* The name will end with a punctuation */
14327 const char * peek = p + 1;
14329 /* Treat any non-']' punctuation followed by a ']' (possibly
14330 * with intervening blanks) as trying to terminate the class.
14331 * ']]' is very likely to mean a class was intended (but
14332 * missing the colon), but the warning message that gets
14333 * generated shows the error position better if we exit the
14334 * loop at the bottom (eventually), so skip it here. */
14336 if (peek < e && isBLANK(*peek)) {
14338 found_problem = TRUE;
14341 } while (peek < e && isBLANK(*peek));
14344 if (peek < e && *peek == ']') {
14345 has_terminating_bracket = TRUE;
14347 has_terminating_colon = TRUE;
14349 else if (*p == ';') {
14350 has_semi_colon = TRUE;
14351 has_terminating_colon = TRUE;
14354 found_problem = TRUE;
14361 /* Here we have punctuation we thought didn't end the class.
14362 * Keep track of the position of the key characters that are
14363 * more likely to have been class-enders */
14364 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
14366 /* Allow just one such possible class-ender not actually
14367 * ending the class. */
14368 if (possible_end) {
14374 /* If we have too many punctuation characters, no use in
14376 if (++punct_count > max_distance) {
14380 /* Treat the punctuation as a typo. */
14381 input_text[name_len++] = *p;
14384 else if (isUPPER(*p)) { /* Use lowercase for lookup */
14385 input_text[name_len++] = toLOWER(*p);
14387 found_problem = TRUE;
14389 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
14390 input_text[name_len++] = *p;
14394 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
14398 /* The declaration of 'input_text' is how long we allow a potential
14399 * class name to be, before saying they didn't mean a class name at
14401 if (name_len >= C_ARRAY_LENGTH(input_text)) {
14406 /* We get to here when the possible class name hasn't been properly
14407 * terminated before:
14408 * 1) we ran off the end of the pattern; or
14409 * 2) found two characters, each of which might have been intended to
14410 * be the name's terminator
14411 * 3) found so many punctuation characters in the purported name,
14412 * that the edit distance to a valid one is exceeded
14413 * 4) we decided it was more characters than anyone could have
14414 * intended to be one. */
14416 found_problem = TRUE;
14418 /* In the final two cases, we know that looking up what we've
14419 * accumulated won't lead to a match, even a fuzzy one. */
14420 if ( name_len >= C_ARRAY_LENGTH(input_text)
14421 || punct_count > max_distance)
14423 /* If there was an intermediate key character that could have been
14424 * an intended end, redo the parse, but stop there */
14425 if (possible_end && possible_end != (char *) -1) {
14426 possible_end = (char *) -1; /* Special signal value to say
14427 we've done a first pass */
14432 /* Otherwise, it can't have meant to have been a class */
14433 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14436 /* If we ran off the end, and the final character was a punctuation
14437 * one, back up one, to look at that final one just below. Later, we
14438 * will restore the parse pointer if appropriate */
14439 if (name_len && p == e && isPUNCT(*(p-1))) {
14444 if (p < e && isPUNCT(*p)) {
14446 has_terminating_bracket = TRUE;
14448 /* If this is a 2nd ']', and the first one is just below this
14449 * one, consider that to be the real terminator. This gives a
14450 * uniform and better positioning for the warning message */
14452 && possible_end != (char *) -1
14453 && *possible_end == ']'
14454 && name_len && input_text[name_len - 1] == ']')
14459 /* And this is actually equivalent to having done the 2nd
14460 * pass now, so set it to not try again */
14461 possible_end = (char *) -1;
14466 has_terminating_colon = TRUE;
14468 else if (*p == ';') {
14469 has_semi_colon = TRUE;
14470 has_terminating_colon = TRUE;
14478 /* Here, we have a class name to look up. We can short circuit the
14479 * stuff below for short names that can't possibly be meant to be a
14480 * class name. (We can do this on the first pass, as any second pass
14481 * will yield an even shorter name) */
14482 if (name_len < 3) {
14483 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14486 /* Find which class it is. Initially switch on the length of the name.
14488 switch (name_len) {
14490 if (memEQ(name_start, "word", 4)) {
14491 /* this is not POSIX, this is the Perl \w */
14492 class_number = ANYOF_WORDCHAR;
14496 /* Names all of length 5: alnum alpha ascii blank cntrl digit
14497 * graph lower print punct space upper
14498 * Offset 4 gives the best switch position. */
14499 switch (name_start[4]) {
14501 if (memEQ(name_start, "alph", 4)) /* alpha */
14502 class_number = ANYOF_ALPHA;
14505 if (memEQ(name_start, "spac", 4)) /* space */
14506 class_number = ANYOF_SPACE;
14509 if (memEQ(name_start, "grap", 4)) /* graph */
14510 class_number = ANYOF_GRAPH;
14513 if (memEQ(name_start, "asci", 4)) /* ascii */
14514 class_number = ANYOF_ASCII;
14517 if (memEQ(name_start, "blan", 4)) /* blank */
14518 class_number = ANYOF_BLANK;
14521 if (memEQ(name_start, "cntr", 4)) /* cntrl */
14522 class_number = ANYOF_CNTRL;
14525 if (memEQ(name_start, "alnu", 4)) /* alnum */
14526 class_number = ANYOF_ALPHANUMERIC;
14529 if (memEQ(name_start, "lowe", 4)) /* lower */
14530 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
14531 else if (memEQ(name_start, "uppe", 4)) /* upper */
14532 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
14535 if (memEQ(name_start, "digi", 4)) /* digit */
14536 class_number = ANYOF_DIGIT;
14537 else if (memEQ(name_start, "prin", 4)) /* print */
14538 class_number = ANYOF_PRINT;
14539 else if (memEQ(name_start, "punc", 4)) /* punct */
14540 class_number = ANYOF_PUNCT;
14545 if (memEQ(name_start, "xdigit", 6))
14546 class_number = ANYOF_XDIGIT;
14550 /* If the name exactly matches a posix class name the class number will
14551 * here be set to it, and the input almost certainly was meant to be a
14552 * posix class, so we can skip further checking. If instead the syntax
14553 * is exactly correct, but the name isn't one of the legal ones, we
14554 * will return that as an error below. But if neither of these apply,
14555 * it could be that no posix class was intended at all, or that one
14556 * was, but there was a typo. We tease these apart by doing fuzzy
14557 * matching on the name */
14558 if (class_number == OOB_NAMEDCLASS && found_problem) {
14559 const UV posix_names[][6] = {
14560 { 'a', 'l', 'n', 'u', 'm' },
14561 { 'a', 'l', 'p', 'h', 'a' },
14562 { 'a', 's', 'c', 'i', 'i' },
14563 { 'b', 'l', 'a', 'n', 'k' },
14564 { 'c', 'n', 't', 'r', 'l' },
14565 { 'd', 'i', 'g', 'i', 't' },
14566 { 'g', 'r', 'a', 'p', 'h' },
14567 { 'l', 'o', 'w', 'e', 'r' },
14568 { 'p', 'r', 'i', 'n', 't' },
14569 { 'p', 'u', 'n', 'c', 't' },
14570 { 's', 'p', 'a', 'c', 'e' },
14571 { 'u', 'p', 'p', 'e', 'r' },
14572 { 'w', 'o', 'r', 'd' },
14573 { 'x', 'd', 'i', 'g', 'i', 't' }
14575 /* The names of the above all have added NULs to make them the same
14576 * size, so we need to also have the real lengths */
14577 const UV posix_name_lengths[] = {
14578 sizeof("alnum") - 1,
14579 sizeof("alpha") - 1,
14580 sizeof("ascii") - 1,
14581 sizeof("blank") - 1,
14582 sizeof("cntrl") - 1,
14583 sizeof("digit") - 1,
14584 sizeof("graph") - 1,
14585 sizeof("lower") - 1,
14586 sizeof("print") - 1,
14587 sizeof("punct") - 1,
14588 sizeof("space") - 1,
14589 sizeof("upper") - 1,
14590 sizeof("word") - 1,
14591 sizeof("xdigit")- 1
14594 int temp_max = max_distance; /* Use a temporary, so if we
14595 reparse, we haven't changed the
14598 /* Use a smaller max edit distance if we are missing one of the
14600 if ( has_opening_bracket + has_opening_colon < 2
14601 || has_terminating_bracket + has_terminating_colon < 2)
14606 /* See if the input name is close to a legal one */
14607 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
14609 /* Short circuit call if the lengths are too far apart to be
14611 if (abs( (int) (name_len - posix_name_lengths[i]))
14617 if (edit_distance(input_text,
14620 posix_name_lengths[i],
14624 { /* If it is close, it probably was intended to be a class */
14625 goto probably_meant_to_be;
14629 /* Here the input name is not close enough to a valid class name
14630 * for us to consider it to be intended to be a posix class. If
14631 * we haven't already done so, and the parse found a character that
14632 * could have been terminators for the name, but which we absorbed
14633 * as typos during the first pass, repeat the parse, signalling it
14634 * to stop at that character */
14635 if (possible_end && possible_end != (char *) -1) {
14636 possible_end = (char *) -1;
14641 /* Here neither pass found a close-enough class name */
14642 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14645 probably_meant_to_be:
14647 /* Here we think that a posix specification was intended. Update any
14649 if (updated_parse_ptr) {
14650 *updated_parse_ptr = (char *) p;
14653 /* If a posix class name was intended but incorrectly specified, we
14654 * output or return the warnings */
14655 if (found_problem) {
14657 /* We set flags for these issues in the parse loop above instead of
14658 * adding them to the list of warnings, because we can parse it
14659 * twice, and we only want one warning instance */
14661 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
14664 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14666 if (has_semi_colon) {
14667 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14669 else if (! has_terminating_colon) {
14670 ADD_POSIX_WARNING(p, "there is no terminating ':'");
14672 if (! has_terminating_bracket) {
14673 ADD_POSIX_WARNING(p, "there is no terminating ']'");
14676 if (posix_warnings && RExC_warn_text && av_top_index(RExC_warn_text) > -1) {
14677 *posix_warnings = RExC_warn_text;
14680 else if (class_number != OOB_NAMEDCLASS) {
14681 /* If it is a known class, return the class. The class number
14682 * #defines are structured so each complement is +1 to the normal
14684 return class_number + complement;
14686 else if (! check_only) {
14688 /* Here, it is an unrecognized class. This is an error (unless the
14689 * call is to check only, which we've already handled above) */
14690 const char * const complement_string = (complement)
14693 RExC_parse = (char *) p;
14694 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
14696 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
14700 return OOB_NAMEDCLASS;
14702 #undef ADD_POSIX_WARNING
14704 STATIC unsigned int
14705 S_regex_set_precedence(const U8 my_operator) {
14707 /* Returns the precedence in the (?[...]) construct of the input operator,
14708 * specified by its character representation. The precedence follows
14709 * general Perl rules, but it extends this so that ')' and ']' have (low)
14710 * precedence even though they aren't really operators */
14712 switch (my_operator) {
14728 NOT_REACHED; /* NOTREACHED */
14729 return 0; /* Silence compiler warning */
14733 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
14734 I32 *flagp, U32 depth,
14735 char * const oregcomp_parse)
14737 /* Handle the (?[...]) construct to do set operations */
14739 U8 curchar; /* Current character being parsed */
14740 UV start, end; /* End points of code point ranges */
14741 SV* final = NULL; /* The end result inversion list */
14742 SV* result_string; /* 'final' stringified */
14743 AV* stack; /* stack of operators and operands not yet
14745 AV* fence_stack = NULL; /* A stack containing the positions in
14746 'stack' of where the undealt-with left
14747 parens would be if they were actually
14749 /* The 'VOL' (expanding to 'volatile') is a workaround for an optimiser bug
14750 * in Solaris Studio 12.3. See RT #127455 */
14751 VOL IV fence = 0; /* Position of where most recent undealt-
14752 with left paren in stack is; -1 if none.
14754 STRLEN len; /* Temporary */
14755 regnode* node; /* Temporary, and final regnode returned by
14757 const bool save_fold = FOLD; /* Temporary */
14758 char *save_end, *save_parse; /* Temporaries */
14759 const bool in_locale = LOC; /* we turn off /l during processing */
14760 AV* posix_warnings = NULL;
14762 GET_RE_DEBUG_FLAGS_DECL;
14764 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
14767 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
14770 REQUIRE_UNI_RULES(flagp, NULL); /* The use of this operator implies /u.
14771 This is required so that the compile
14772 time values are valid in all runtime
14775 /* This will return only an ANYOF regnode, or (unlikely) something smaller
14776 * (such as EXACT). Thus we can skip most everything if just sizing. We
14777 * call regclass to handle '[]' so as to not have to reinvent its parsing
14778 * rules here (throwing away the size it computes each time). And, we exit
14779 * upon an unescaped ']' that isn't one ending a regclass. To do both
14780 * these things, we need to realize that something preceded by a backslash
14781 * is escaped, so we have to keep track of backslashes */
14783 UV depth = 0; /* how many nested (?[...]) constructs */
14785 while (RExC_parse < RExC_end) {
14786 SV* current = NULL;
14788 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14789 TRUE /* Force /x */ );
14791 switch (*RExC_parse) {
14793 if (RExC_parse[1] == '[') depth++, RExC_parse++;
14798 /* Skip past this, so the next character gets skipped, after
14801 if (*RExC_parse == 'c') {
14802 /* Skip the \cX notation for control characters */
14803 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14809 /* See if this is a [:posix:] class. */
14810 bool is_posix_class = (OOB_NAMEDCLASS
14811 < handle_possible_posix(pRExC_state,
14815 TRUE /* checking only */));
14816 /* If it is a posix class, leave the parse pointer at the
14817 * '[' to fool regclass() into thinking it is part of a
14818 * '[[:posix:]]'. */
14819 if (! is_posix_class) {
14823 /* regclass() can only return RESTART_PASS1 and NEED_UTF8
14824 * if multi-char folds are allowed. */
14825 if (!regclass(pRExC_state, flagp,depth+1,
14826 is_posix_class, /* parse the whole char
14827 class only if not a
14829 FALSE, /* don't allow multi-char folds */
14830 TRUE, /* silence non-portable warnings. */
14832 FALSE, /* Require return to be an ANYOF */
14836 FAIL2("panic: regclass returned NULL to handle_sets, "
14837 "flags=%#" UVxf, (UV) *flagp);
14839 /* function call leaves parse pointing to the ']', except
14840 * if we faked it */
14841 if (is_posix_class) {
14845 SvREFCNT_dec(current); /* In case it returned something */
14850 if (depth--) break;
14852 if (*RExC_parse == ')') {
14853 node = reganode(pRExC_state, ANYOF, 0);
14854 RExC_size += ANYOF_SKIP;
14855 nextchar(pRExC_state);
14856 Set_Node_Length(node,
14857 RExC_parse - oregcomp_parse + 1); /* MJD */
14859 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
14867 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14871 /* We output the messages even if warnings are off, because we'll fail
14872 * the very next thing, and these give a likely diagnosis for that */
14873 if (posix_warnings && av_tindex_nomg(posix_warnings) >= 0) {
14874 output_or_return_posix_warnings(pRExC_state, posix_warnings, NULL);
14877 FAIL("Syntax error in (?[...])");
14880 /* Pass 2 only after this. */
14881 Perl_ck_warner_d(aTHX_
14882 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
14883 "The regex_sets feature is experimental" REPORT_LOCATION,
14884 REPORT_LOCATION_ARGS(RExC_parse));
14886 /* Everything in this construct is a metacharacter. Operands begin with
14887 * either a '\' (for an escape sequence), or a '[' for a bracketed
14888 * character class. Any other character should be an operator, or
14889 * parenthesis for grouping. Both types of operands are handled by calling
14890 * regclass() to parse them. It is called with a parameter to indicate to
14891 * return the computed inversion list. The parsing here is implemented via
14892 * a stack. Each entry on the stack is a single character representing one
14893 * of the operators; or else a pointer to an operand inversion list. */
14895 #define IS_OPERATOR(a) SvIOK(a)
14896 #define IS_OPERAND(a) (! IS_OPERATOR(a))
14898 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
14899 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
14900 * with pronouncing it called it Reverse Polish instead, but now that YOU
14901 * know how to pronounce it you can use the correct term, thus giving due
14902 * credit to the person who invented it, and impressing your geek friends.
14903 * Wikipedia says that the pronounciation of "Ł" has been changing so that
14904 * it is now more like an English initial W (as in wonk) than an L.)
14906 * This means that, for example, 'a | b & c' is stored on the stack as
14914 * where the numbers in brackets give the stack [array] element number.
14915 * In this implementation, parentheses are not stored on the stack.
14916 * Instead a '(' creates a "fence" so that the part of the stack below the
14917 * fence is invisible except to the corresponding ')' (this allows us to
14918 * replace testing for parens, by using instead subtraction of the fence
14919 * position). As new operands are processed they are pushed onto the stack
14920 * (except as noted in the next paragraph). New operators of higher
14921 * precedence than the current final one are inserted on the stack before
14922 * the lhs operand (so that when the rhs is pushed next, everything will be
14923 * in the correct positions shown above. When an operator of equal or
14924 * lower precedence is encountered in parsing, all the stacked operations
14925 * of equal or higher precedence are evaluated, leaving the result as the
14926 * top entry on the stack. This makes higher precedence operations
14927 * evaluate before lower precedence ones, and causes operations of equal
14928 * precedence to left associate.
14930 * The only unary operator '!' is immediately pushed onto the stack when
14931 * encountered. When an operand is encountered, if the top of the stack is
14932 * a '!", the complement is immediately performed, and the '!' popped. The
14933 * resulting value is treated as a new operand, and the logic in the
14934 * previous paragraph is executed. Thus in the expression
14936 * the stack looks like
14942 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
14949 * A ')' is treated as an operator with lower precedence than all the
14950 * aforementioned ones, which causes all operations on the stack above the
14951 * corresponding '(' to be evaluated down to a single resultant operand.
14952 * Then the fence for the '(' is removed, and the operand goes through the
14953 * algorithm above, without the fence.
14955 * A separate stack is kept of the fence positions, so that the position of
14956 * the latest so-far unbalanced '(' is at the top of it.
14958 * The ']' ending the construct is treated as the lowest operator of all,
14959 * so that everything gets evaluated down to a single operand, which is the
14962 sv_2mortal((SV *)(stack = newAV()));
14963 sv_2mortal((SV *)(fence_stack = newAV()));
14965 while (RExC_parse < RExC_end) {
14966 I32 top_index; /* Index of top-most element in 'stack' */
14967 SV** top_ptr; /* Pointer to top 'stack' element */
14968 SV* current = NULL; /* To contain the current inversion list
14970 SV* only_to_avoid_leaks;
14972 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14973 TRUE /* Force /x */ );
14974 if (RExC_parse >= RExC_end) {
14975 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
14978 curchar = UCHARAT(RExC_parse);
14982 #ifdef ENABLE_REGEX_SETS_DEBUGGING
14983 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
14984 DEBUG_U(dump_regex_sets_structures(pRExC_state,
14985 stack, fence, fence_stack));
14988 top_index = av_tindex_nomg(stack);
14991 SV** stacked_ptr; /* Ptr to something already on 'stack' */
14992 char stacked_operator; /* The topmost operator on the 'stack'. */
14993 SV* lhs; /* Operand to the left of the operator */
14994 SV* rhs; /* Operand to the right of the operator */
14995 SV* fence_ptr; /* Pointer to top element of the fence
15000 if ( RExC_parse < RExC_end - 1
15001 && (UCHARAT(RExC_parse + 1) == '?'))
15003 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
15004 * This happens when we have some thing like
15006 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
15008 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
15010 * Here we would be handling the interpolated
15011 * '$thai_or_lao'. We handle this by a recursive call to
15012 * ourselves which returns the inversion list the
15013 * interpolated expression evaluates to. We use the flags
15014 * from the interpolated pattern. */
15015 U32 save_flags = RExC_flags;
15016 const char * save_parse;
15018 RExC_parse += 2; /* Skip past the '(?' */
15019 save_parse = RExC_parse;
15021 /* Parse any flags for the '(?' */
15022 parse_lparen_question_flags(pRExC_state);
15024 if (RExC_parse == save_parse /* Makes sure there was at
15025 least one flag (or else
15026 this embedding wasn't
15028 || RExC_parse >= RExC_end - 4
15029 || UCHARAT(RExC_parse) != ':'
15030 || UCHARAT(++RExC_parse) != '('
15031 || UCHARAT(++RExC_parse) != '?'
15032 || UCHARAT(++RExC_parse) != '[')
15035 /* In combination with the above, this moves the
15036 * pointer to the point just after the first erroneous
15037 * character (or if there are no flags, to where they
15038 * should have been) */
15039 if (RExC_parse >= RExC_end - 4) {
15040 RExC_parse = RExC_end;
15042 else if (RExC_parse != save_parse) {
15043 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15045 vFAIL("Expecting '(?flags:(?[...'");
15048 /* Recurse, with the meat of the embedded expression */
15050 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
15051 depth+1, oregcomp_parse);
15053 /* Here, 'current' contains the embedded expression's
15054 * inversion list, and RExC_parse points to the trailing
15055 * ']'; the next character should be the ')' */
15057 assert(UCHARAT(RExC_parse) == ')');
15059 /* Then the ')' matching the original '(' handled by this
15060 * case: statement */
15062 assert(UCHARAT(RExC_parse) == ')');
15065 RExC_flags = save_flags;
15066 goto handle_operand;
15069 /* A regular '('. Look behind for illegal syntax */
15070 if (top_index - fence >= 0) {
15071 /* If the top entry on the stack is an operator, it had
15072 * better be a '!', otherwise the entry below the top
15073 * operand should be an operator */
15074 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
15075 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
15076 || ( IS_OPERAND(*top_ptr)
15077 && ( top_index - fence < 1
15078 || ! (stacked_ptr = av_fetch(stack,
15081 || ! IS_OPERATOR(*stacked_ptr))))
15084 vFAIL("Unexpected '(' with no preceding operator");
15088 /* Stack the position of this undealt-with left paren */
15089 av_push(fence_stack, newSViv(fence));
15090 fence = top_index + 1;
15094 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15095 * multi-char folds are allowed. */
15096 if (!regclass(pRExC_state, flagp,depth+1,
15097 TRUE, /* means parse just the next thing */
15098 FALSE, /* don't allow multi-char folds */
15099 FALSE, /* don't silence non-portable warnings. */
15101 FALSE, /* Require return to be an ANYOF */
15105 FAIL2("panic: regclass returned NULL to handle_sets, "
15106 "flags=%#" UVxf, (UV) *flagp);
15109 /* regclass() will return with parsing just the \ sequence,
15110 * leaving the parse pointer at the next thing to parse */
15112 goto handle_operand;
15114 case '[': /* Is a bracketed character class */
15116 /* See if this is a [:posix:] class. */
15117 bool is_posix_class = (OOB_NAMEDCLASS
15118 < handle_possible_posix(pRExC_state,
15122 TRUE /* checking only */));
15123 /* If it is a posix class, leave the parse pointer at the '['
15124 * to fool regclass() into thinking it is part of a
15125 * '[[:posix:]]'. */
15126 if (! is_posix_class) {
15130 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15131 * multi-char folds are allowed. */
15132 if (!regclass(pRExC_state, flagp,depth+1,
15133 is_posix_class, /* parse the whole char
15134 class only if not a
15136 FALSE, /* don't allow multi-char folds */
15137 TRUE, /* silence non-portable warnings. */
15139 FALSE, /* Require return to be an ANYOF */
15144 FAIL2("panic: regclass returned NULL to handle_sets, "
15145 "flags=%#" UVxf, (UV) *flagp);
15148 /* function call leaves parse pointing to the ']', except if we
15150 if (is_posix_class) {
15154 goto handle_operand;
15158 if (top_index >= 1) {
15159 goto join_operators;
15162 /* Only a single operand on the stack: are done */
15166 if (av_tindex_nomg(fence_stack) < 0) {
15168 vFAIL("Unexpected ')'");
15171 /* If nothing after the fence, is missing an operand */
15172 if (top_index - fence < 0) {
15176 /* If at least two things on the stack, treat this as an
15178 if (top_index - fence >= 1) {
15179 goto join_operators;
15182 /* Here only a single thing on the fenced stack, and there is a
15183 * fence. Get rid of it */
15184 fence_ptr = av_pop(fence_stack);
15186 fence = SvIV(fence_ptr) - 1;
15187 SvREFCNT_dec_NN(fence_ptr);
15194 /* Having gotten rid of the fence, we pop the operand at the
15195 * stack top and process it as a newly encountered operand */
15196 current = av_pop(stack);
15197 if (IS_OPERAND(current)) {
15198 goto handle_operand;
15210 /* These binary operators should have a left operand already
15212 if ( top_index - fence < 0
15213 || top_index - fence == 1
15214 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
15215 || ! IS_OPERAND(*top_ptr))
15217 goto unexpected_binary;
15220 /* If only the one operand is on the part of the stack visible
15221 * to us, we just place this operator in the proper position */
15222 if (top_index - fence < 2) {
15224 /* Place the operator before the operand */
15226 SV* lhs = av_pop(stack);
15227 av_push(stack, newSVuv(curchar));
15228 av_push(stack, lhs);
15232 /* But if there is something else on the stack, we need to
15233 * process it before this new operator if and only if the
15234 * stacked operation has equal or higher precedence than the
15239 /* The operator on the stack is supposed to be below both its
15241 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
15242 || IS_OPERAND(*stacked_ptr))
15244 /* But if not, it's legal and indicates we are completely
15245 * done if and only if we're currently processing a ']',
15246 * which should be the final thing in the expression */
15247 if (curchar == ']') {
15253 vFAIL2("Unexpected binary operator '%c' with no "
15254 "preceding operand", curchar);
15256 stacked_operator = (char) SvUV(*stacked_ptr);
15258 if (regex_set_precedence(curchar)
15259 > regex_set_precedence(stacked_operator))
15261 /* Here, the new operator has higher precedence than the
15262 * stacked one. This means we need to add the new one to
15263 * the stack to await its rhs operand (and maybe more
15264 * stuff). We put it before the lhs operand, leaving
15265 * untouched the stacked operator and everything below it
15267 lhs = av_pop(stack);
15268 assert(IS_OPERAND(lhs));
15270 av_push(stack, newSVuv(curchar));
15271 av_push(stack, lhs);
15275 /* Here, the new operator has equal or lower precedence than
15276 * what's already there. This means the operation already
15277 * there should be performed now, before the new one. */
15279 rhs = av_pop(stack);
15280 if (! IS_OPERAND(rhs)) {
15282 /* This can happen when a ! is not followed by an operand,
15283 * like in /(?[\t &!])/ */
15287 lhs = av_pop(stack);
15289 if (! IS_OPERAND(lhs)) {
15291 /* This can happen when there is an empty (), like in
15292 * /(?[[0]+()+])/ */
15296 switch (stacked_operator) {
15298 _invlist_intersection(lhs, rhs, &rhs);
15303 _invlist_union(lhs, rhs, &rhs);
15307 _invlist_subtract(lhs, rhs, &rhs);
15310 case '^': /* The union minus the intersection */
15315 _invlist_union(lhs, rhs, &u);
15316 _invlist_intersection(lhs, rhs, &i);
15317 _invlist_subtract(u, i, &rhs);
15318 SvREFCNT_dec_NN(i);
15319 SvREFCNT_dec_NN(u);
15325 /* Here, the higher precedence operation has been done, and the
15326 * result is in 'rhs'. We overwrite the stacked operator with
15327 * the result. Then we redo this code to either push the new
15328 * operator onto the stack or perform any higher precedence
15329 * stacked operation */
15330 only_to_avoid_leaks = av_pop(stack);
15331 SvREFCNT_dec(only_to_avoid_leaks);
15332 av_push(stack, rhs);
15335 case '!': /* Highest priority, right associative */
15337 /* If what's already at the top of the stack is another '!",
15338 * they just cancel each other out */
15339 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
15340 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
15342 only_to_avoid_leaks = av_pop(stack);
15343 SvREFCNT_dec(only_to_avoid_leaks);
15345 else { /* Otherwise, since it's right associative, just push
15347 av_push(stack, newSVuv(curchar));
15352 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15353 vFAIL("Unexpected character");
15357 /* Here 'current' is the operand. If something is already on the
15358 * stack, we have to check if it is a !. But first, the code above
15359 * may have altered the stack in the time since we earlier set
15362 top_index = av_tindex_nomg(stack);
15363 if (top_index - fence >= 0) {
15364 /* If the top entry on the stack is an operator, it had better
15365 * be a '!', otherwise the entry below the top operand should
15366 * be an operator */
15367 top_ptr = av_fetch(stack, top_index, FALSE);
15369 if (IS_OPERATOR(*top_ptr)) {
15371 /* The only permissible operator at the top of the stack is
15372 * '!', which is applied immediately to this operand. */
15373 curchar = (char) SvUV(*top_ptr);
15374 if (curchar != '!') {
15375 SvREFCNT_dec(current);
15376 vFAIL2("Unexpected binary operator '%c' with no "
15377 "preceding operand", curchar);
15380 _invlist_invert(current);
15382 only_to_avoid_leaks = av_pop(stack);
15383 SvREFCNT_dec(only_to_avoid_leaks);
15385 /* And we redo with the inverted operand. This allows
15386 * handling multiple ! in a row */
15387 goto handle_operand;
15389 /* Single operand is ok only for the non-binary ')'
15391 else if ((top_index - fence == 0 && curchar != ')')
15392 || (top_index - fence > 0
15393 && (! (stacked_ptr = av_fetch(stack,
15396 || IS_OPERAND(*stacked_ptr))))
15398 SvREFCNT_dec(current);
15399 vFAIL("Operand with no preceding operator");
15403 /* Here there was nothing on the stack or the top element was
15404 * another operand. Just add this new one */
15405 av_push(stack, current);
15407 } /* End of switch on next parse token */
15409 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15410 } /* End of loop parsing through the construct */
15413 if (av_tindex_nomg(fence_stack) >= 0) {
15414 vFAIL("Unmatched (");
15417 if (av_tindex_nomg(stack) < 0 /* Was empty */
15418 || ((final = av_pop(stack)) == NULL)
15419 || ! IS_OPERAND(final)
15420 || SvTYPE(final) != SVt_INVLIST
15421 || av_tindex_nomg(stack) >= 0) /* More left on stack */
15424 SvREFCNT_dec(final);
15425 vFAIL("Incomplete expression within '(?[ ])'");
15428 /* Here, 'final' is the resultant inversion list from evaluating the
15429 * expression. Return it if so requested */
15430 if (return_invlist) {
15431 *return_invlist = final;
15435 /* Otherwise generate a resultant node, based on 'final'. regclass() is
15436 * expecting a string of ranges and individual code points */
15437 invlist_iterinit(final);
15438 result_string = newSVpvs("");
15439 while (invlist_iternext(final, &start, &end)) {
15440 if (start == end) {
15441 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
15444 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%" UVXf "}",
15449 /* About to generate an ANYOF (or similar) node from the inversion list we
15450 * have calculated */
15451 save_parse = RExC_parse;
15452 RExC_parse = SvPV(result_string, len);
15453 save_end = RExC_end;
15454 RExC_end = RExC_parse + len;
15456 /* We turn off folding around the call, as the class we have constructed
15457 * already has all folding taken into consideration, and we don't want
15458 * regclass() to add to that */
15459 RExC_flags &= ~RXf_PMf_FOLD;
15460 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if multi-char
15461 * folds are allowed. */
15462 node = regclass(pRExC_state, flagp,depth+1,
15463 FALSE, /* means parse the whole char class */
15464 FALSE, /* don't allow multi-char folds */
15465 TRUE, /* silence non-portable warnings. The above may very
15466 well have generated non-portable code points, but
15467 they're valid on this machine */
15468 FALSE, /* similarly, no need for strict */
15469 FALSE, /* Require return to be an ANYOF */
15474 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#" UVxf,
15477 /* Fix up the node type if we are in locale. (We have pretended we are
15478 * under /u for the purposes of regclass(), as this construct will only
15479 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
15480 * as to cause any warnings about bad locales to be output in regexec.c),
15481 * and add the flag that indicates to check if not in a UTF-8 locale. The
15482 * reason we above forbid optimization into something other than an ANYOF
15483 * node is simply to minimize the number of code changes in regexec.c.
15484 * Otherwise we would have to create new EXACTish node types and deal with
15485 * them. This decision could be revisited should this construct become
15488 * (One might think we could look at the resulting ANYOF node and suppress
15489 * the flag if everything is above 255, as those would be UTF-8 only,
15490 * but this isn't true, as the components that led to that result could
15491 * have been locale-affected, and just happen to cancel each other out
15492 * under UTF-8 locales.) */
15494 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
15496 assert(OP(node) == ANYOF);
15500 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
15504 RExC_flags |= RXf_PMf_FOLD;
15507 RExC_parse = save_parse + 1;
15508 RExC_end = save_end;
15509 SvREFCNT_dec_NN(final);
15510 SvREFCNT_dec_NN(result_string);
15512 nextchar(pRExC_state);
15513 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
15517 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15520 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
15521 AV * stack, const IV fence, AV * fence_stack)
15522 { /* Dumps the stacks in handle_regex_sets() */
15524 const SSize_t stack_top = av_tindex_nomg(stack);
15525 const SSize_t fence_stack_top = av_tindex_nomg(fence_stack);
15528 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
15530 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
15532 if (stack_top < 0) {
15533 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
15536 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
15537 for (i = stack_top; i >= 0; i--) {
15538 SV ** element_ptr = av_fetch(stack, i, FALSE);
15539 if (! element_ptr) {
15542 if (IS_OPERATOR(*element_ptr)) {
15543 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
15544 (int) i, (int) SvIV(*element_ptr));
15547 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
15548 sv_dump(*element_ptr);
15553 if (fence_stack_top < 0) {
15554 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
15557 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
15558 for (i = fence_stack_top; i >= 0; i--) {
15559 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
15560 if (! element_ptr) {
15563 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
15564 (int) i, (int) SvIV(*element_ptr));
15575 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
15577 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
15578 * innocent-looking character class, like /[ks]/i won't have to go out to
15579 * disk to find the possible matches.
15581 * This should be called only for a Latin1-range code points, cp, which is
15582 * known to be involved in a simple fold with other code points above
15583 * Latin1. It would give false results if /aa has been specified.
15584 * Multi-char folds are outside the scope of this, and must be handled
15587 * XXX It would be better to generate these via regen, in case a new
15588 * version of the Unicode standard adds new mappings, though that is not
15589 * really likely, and may be caught by the default: case of the switch
15592 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
15594 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
15600 add_cp_to_invlist(*invlist, KELVIN_SIGN);
15604 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
15607 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
15608 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
15610 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
15611 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
15612 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
15614 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
15615 *invlist = add_cp_to_invlist(*invlist,
15616 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
15619 #ifdef LATIN_CAPITAL_LETTER_SHARP_S /* not defined in early Unicode releases */
15621 case LATIN_SMALL_LETTER_SHARP_S:
15622 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
15627 #if UNICODE_MAJOR_VERSION < 3 \
15628 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0)
15630 /* In 3.0 and earlier, U+0130 folded simply to 'i'; and in 3.0.1 so did
15635 add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
15636 # if UNICODE_DOT_DOT_VERSION == 1
15637 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_DOTLESS_I);
15643 /* Use deprecated warning to increase the chances of this being
15646 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
15653 S_output_or_return_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings, AV** return_posix_warnings)
15655 /* If the final parameter is NULL, output the elements of the array given
15656 * by '*posix_warnings' as REGEXP warnings. Otherwise, the elements are
15657 * pushed onto it, (creating if necessary) */
15660 const bool first_is_fatal = ! return_posix_warnings
15661 && ckDEAD(packWARN(WARN_REGEXP));
15663 PERL_ARGS_ASSERT_OUTPUT_OR_RETURN_POSIX_WARNINGS;
15665 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
15666 if (return_posix_warnings) {
15667 if (! *return_posix_warnings) { /* mortalize to not leak if
15668 warnings are fatal */
15669 *return_posix_warnings = (AV *) sv_2mortal((SV *) newAV());
15671 av_push(*return_posix_warnings, msg);
15674 if (first_is_fatal) { /* Avoid leaking this */
15675 av_undef(posix_warnings); /* This isn't necessary if the
15676 array is mortal, but is a
15678 (void) sv_2mortal(msg);
15680 SAVEFREESV(RExC_rx_sv);
15683 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
15684 SvREFCNT_dec_NN(msg);
15690 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
15692 /* This adds the string scalar <multi_string> to the array
15693 * <multi_char_matches>. <multi_string> is known to have exactly
15694 * <cp_count> code points in it. This is used when constructing a
15695 * bracketed character class and we find something that needs to match more
15696 * than a single character.
15698 * <multi_char_matches> is actually an array of arrays. Each top-level
15699 * element is an array that contains all the strings known so far that are
15700 * the same length. And that length (in number of code points) is the same
15701 * as the index of the top-level array. Hence, the [2] element is an
15702 * array, each element thereof is a string containing TWO code points;
15703 * while element [3] is for strings of THREE characters, and so on. Since
15704 * this is for multi-char strings there can never be a [0] nor [1] element.
15706 * When we rewrite the character class below, we will do so such that the
15707 * longest strings are written first, so that it prefers the longest
15708 * matching strings first. This is done even if it turns out that any
15709 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
15710 * Christiansen has agreed that this is ok. This makes the test for the
15711 * ligature 'ffi' come before the test for 'ff', for example */
15714 AV** this_array_ptr;
15716 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
15718 if (! multi_char_matches) {
15719 multi_char_matches = newAV();
15722 if (av_exists(multi_char_matches, cp_count)) {
15723 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
15724 this_array = *this_array_ptr;
15727 this_array = newAV();
15728 av_store(multi_char_matches, cp_count,
15731 av_push(this_array, multi_string);
15733 return multi_char_matches;
15736 /* The names of properties whose definitions are not known at compile time are
15737 * stored in this SV, after a constant heading. So if the length has been
15738 * changed since initialization, then there is a run-time definition. */
15739 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
15740 (SvCUR(listsv) != initial_listsv_len)
15742 /* There is a restricted set of white space characters that are legal when
15743 * ignoring white space in a bracketed character class. This generates the
15744 * code to skip them.
15746 * There is a line below that uses the same white space criteria but is outside
15747 * this macro. Both here and there must use the same definition */
15748 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
15751 while (isBLANK_A(UCHARAT(p))) \
15759 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
15760 const bool stop_at_1, /* Just parse the next thing, don't
15761 look for a full character class */
15762 bool allow_multi_folds,
15763 const bool silence_non_portable, /* Don't output warnings
15767 bool optimizable, /* ? Allow a non-ANYOF return
15769 SV** ret_invlist, /* Return an inversion list, not a node */
15770 AV** return_posix_warnings
15773 /* parse a bracketed class specification. Most of these will produce an
15774 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
15775 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
15776 * under /i with multi-character folds: it will be rewritten following the
15777 * paradigm of this example, where the <multi-fold>s are characters which
15778 * fold to multiple character sequences:
15779 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
15780 * gets effectively rewritten as:
15781 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
15782 * reg() gets called (recursively) on the rewritten version, and this
15783 * function will return what it constructs. (Actually the <multi-fold>s
15784 * aren't physically removed from the [abcdefghi], it's just that they are
15785 * ignored in the recursion by means of a flag:
15786 * <RExC_in_multi_char_class>.)
15788 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
15789 * characters, with the corresponding bit set if that character is in the
15790 * list. For characters above this, a range list or swash is used. There
15791 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
15792 * determinable at compile time
15794 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs
15795 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded
15796 * to UTF-8. This can only happen if ret_invlist is non-NULL.
15799 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
15801 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
15804 int namedclass = OOB_NAMEDCLASS;
15805 char *rangebegin = NULL;
15806 bool need_class = 0;
15808 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
15809 than just initialized. */
15810 SV* properties = NULL; /* Code points that match \p{} \P{} */
15811 SV* posixes = NULL; /* Code points that match classes like [:word:],
15812 extended beyond the Latin1 range. These have to
15813 be kept separate from other code points for much
15814 of this function because their handling is
15815 different under /i, and for most classes under
15817 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
15818 separate for a while from the non-complemented
15819 versions because of complications with /d
15821 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
15822 treated more simply than the general case,
15823 leading to less compilation and execution
15825 UV element_count = 0; /* Number of distinct elements in the class.
15826 Optimizations may be possible if this is tiny */
15827 AV * multi_char_matches = NULL; /* Code points that fold to more than one
15828 character; used under /i */
15830 char * stop_ptr = RExC_end; /* where to stop parsing */
15832 /* ignore unescaped whitespace? */
15833 const bool skip_white = cBOOL( ret_invlist
15834 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
15836 /* Unicode properties are stored in a swash; this holds the current one
15837 * being parsed. If this swash is the only above-latin1 component of the
15838 * character class, an optimization is to pass it directly on to the
15839 * execution engine. Otherwise, it is set to NULL to indicate that there
15840 * are other things in the class that have to be dealt with at execution
15842 SV* swash = NULL; /* Code points that match \p{} \P{} */
15844 /* Set if a component of this character class is user-defined; just passed
15845 * on to the engine */
15846 bool has_user_defined_property = FALSE;
15848 /* inversion list of code points this node matches only when the target
15849 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
15851 SV* has_upper_latin1_only_utf8_matches = NULL;
15853 /* Inversion list of code points this node matches regardless of things
15854 * like locale, folding, utf8ness of the target string */
15855 SV* cp_list = NULL;
15857 /* Like cp_list, but code points on this list need to be checked for things
15858 * that fold to/from them under /i */
15859 SV* cp_foldable_list = NULL;
15861 /* Like cp_list, but code points on this list are valid only when the
15862 * runtime locale is UTF-8 */
15863 SV* only_utf8_locale_list = NULL;
15865 /* In a range, if one of the endpoints is non-character-set portable,
15866 * meaning that it hard-codes a code point that may mean a different
15867 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
15868 * mnemonic '\t' which each mean the same character no matter which
15869 * character set the platform is on. */
15870 unsigned int non_portable_endpoint = 0;
15872 /* Is the range unicode? which means on a platform that isn't 1-1 native
15873 * to Unicode (i.e. non-ASCII), each code point in it should be considered
15874 * to be a Unicode value. */
15875 bool unicode_range = FALSE;
15876 bool invert = FALSE; /* Is this class to be complemented */
15878 bool warn_super = ALWAYS_WARN_SUPER;
15880 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
15881 case we need to change the emitted regop to an EXACT. */
15882 const char * orig_parse = RExC_parse;
15883 const SSize_t orig_size = RExC_size;
15884 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
15886 /* This variable is used to mark where the end in the input is of something
15887 * that looks like a POSIX construct but isn't. During the parse, when
15888 * something looks like it could be such a construct is encountered, it is
15889 * checked for being one, but not if we've already checked this area of the
15890 * input. Only after this position is reached do we check again */
15891 char *not_posix_region_end = RExC_parse - 1;
15893 AV* posix_warnings = NULL;
15894 const bool do_posix_warnings = return_posix_warnings
15895 || (PASS2 && ckWARN(WARN_REGEXP));
15897 GET_RE_DEBUG_FLAGS_DECL;
15899 PERL_ARGS_ASSERT_REGCLASS;
15901 PERL_UNUSED_ARG(depth);
15904 DEBUG_PARSE("clas");
15906 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
15907 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
15908 && UNICODE_DOT_DOT_VERSION == 0)
15909 allow_multi_folds = FALSE;
15912 /* Assume we are going to generate an ANYOF node. */
15913 ret = reganode(pRExC_state,
15920 RExC_size += ANYOF_SKIP;
15921 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
15924 ANYOF_FLAGS(ret) = 0;
15926 RExC_emit += ANYOF_SKIP;
15927 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
15928 initial_listsv_len = SvCUR(listsv);
15929 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
15932 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15934 assert(RExC_parse <= RExC_end);
15936 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
15939 allow_multi_folds = FALSE;
15941 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15944 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
15945 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
15946 int maybe_class = handle_possible_posix(pRExC_state,
15948 ¬_posix_region_end,
15950 TRUE /* checking only */);
15951 if (PASS2 && maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
15952 SAVEFREESV(RExC_rx_sv);
15953 ckWARN4reg(not_posix_region_end,
15954 "POSIX syntax [%c %c] belongs inside character classes%s",
15955 *RExC_parse, *RExC_parse,
15956 (maybe_class == OOB_NAMEDCLASS)
15957 ? ((POSIXCC_NOTYET(*RExC_parse))
15958 ? " (but this one isn't implemented)"
15959 : " (but this one isn't fully valid)")
15962 (void)ReREFCNT_inc(RExC_rx_sv);
15966 /* If the caller wants us to just parse a single element, accomplish this
15967 * by faking the loop ending condition */
15968 if (stop_at_1 && RExC_end > RExC_parse) {
15969 stop_ptr = RExC_parse + 1;
15972 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
15973 if (UCHARAT(RExC_parse) == ']')
15974 goto charclassloop;
15978 if ( posix_warnings
15979 && av_tindex_nomg(posix_warnings) >= 0
15980 && RExC_parse > not_posix_region_end)
15982 /* Warnings about posix class issues are considered tentative until
15983 * we are far enough along in the parse that we can no longer
15984 * change our mind, at which point we either output them or add
15985 * them, if it has so specified, to what gets returned to the
15986 * caller. This is done each time through the loop so that a later
15987 * class won't zap them before they have been dealt with. */
15988 output_or_return_posix_warnings(pRExC_state, posix_warnings,
15989 return_posix_warnings);
15992 if (RExC_parse >= stop_ptr) {
15996 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15998 if (UCHARAT(RExC_parse) == ']') {
16004 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
16005 save_value = value;
16006 save_prevvalue = prevvalue;
16009 rangebegin = RExC_parse;
16011 non_portable_endpoint = 0;
16013 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
16014 value = utf8n_to_uvchr((U8*)RExC_parse,
16015 RExC_end - RExC_parse,
16016 &numlen, UTF8_ALLOW_DEFAULT);
16017 RExC_parse += numlen;
16020 value = UCHARAT(RExC_parse++);
16022 if (value == '[') {
16023 char * posix_class_end;
16024 namedclass = handle_possible_posix(pRExC_state,
16027 do_posix_warnings ? &posix_warnings : NULL,
16028 FALSE /* die if error */);
16029 if (namedclass > OOB_NAMEDCLASS) {
16031 /* If there was an earlier attempt to parse this particular
16032 * posix class, and it failed, it was a false alarm, as this
16033 * successful one proves */
16034 if ( posix_warnings
16035 && av_tindex_nomg(posix_warnings) >= 0
16036 && not_posix_region_end >= RExC_parse
16037 && not_posix_region_end <= posix_class_end)
16039 av_undef(posix_warnings);
16042 RExC_parse = posix_class_end;
16044 else if (namedclass == OOB_NAMEDCLASS) {
16045 not_posix_region_end = posix_class_end;
16048 namedclass = OOB_NAMEDCLASS;
16051 else if ( RExC_parse - 1 > not_posix_region_end
16052 && MAYBE_POSIXCC(value))
16054 (void) handle_possible_posix(
16056 RExC_parse - 1, /* -1 because parse has already been
16058 ¬_posix_region_end,
16059 do_posix_warnings ? &posix_warnings : NULL,
16060 TRUE /* checking only */);
16062 else if (value == '\\') {
16063 /* Is a backslash; get the code point of the char after it */
16065 if (RExC_parse >= RExC_end) {
16066 vFAIL("Unmatched [");
16069 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
16070 value = utf8n_to_uvchr((U8*)RExC_parse,
16071 RExC_end - RExC_parse,
16072 &numlen, UTF8_ALLOW_DEFAULT);
16073 RExC_parse += numlen;
16076 value = UCHARAT(RExC_parse++);
16078 /* Some compilers cannot handle switching on 64-bit integer
16079 * values, therefore value cannot be an UV. Yes, this will
16080 * be a problem later if we want switch on Unicode.
16081 * A similar issue a little bit later when switching on
16082 * namedclass. --jhi */
16084 /* If the \ is escaping white space when white space is being
16085 * skipped, it means that that white space is wanted literally, and
16086 * is already in 'value'. Otherwise, need to translate the escape
16087 * into what it signifies. */
16088 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
16090 case 'w': namedclass = ANYOF_WORDCHAR; break;
16091 case 'W': namedclass = ANYOF_NWORDCHAR; break;
16092 case 's': namedclass = ANYOF_SPACE; break;
16093 case 'S': namedclass = ANYOF_NSPACE; break;
16094 case 'd': namedclass = ANYOF_DIGIT; break;
16095 case 'D': namedclass = ANYOF_NDIGIT; break;
16096 case 'v': namedclass = ANYOF_VERTWS; break;
16097 case 'V': namedclass = ANYOF_NVERTWS; break;
16098 case 'h': namedclass = ANYOF_HORIZWS; break;
16099 case 'H': namedclass = ANYOF_NHORIZWS; break;
16100 case 'N': /* Handle \N{NAME} in class */
16102 const char * const backslash_N_beg = RExC_parse - 2;
16105 if (! grok_bslash_N(pRExC_state,
16106 NULL, /* No regnode */
16107 &value, /* Yes single value */
16108 &cp_count, /* Multiple code pt count */
16114 if (*flagp & NEED_UTF8)
16115 FAIL("panic: grok_bslash_N set NEED_UTF8");
16116 if (*flagp & RESTART_PASS1)
16119 if (cp_count < 0) {
16120 vFAIL("\\N in a character class must be a named character: \\N{...}");
16122 else if (cp_count == 0) {
16124 ckWARNreg(RExC_parse,
16125 "Ignoring zero length \\N{} in character class");
16128 else { /* cp_count > 1 */
16129 if (! RExC_in_multi_char_class) {
16130 if (invert || range || *RExC_parse == '-') {
16133 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
16136 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
16138 break; /* <value> contains the first code
16139 point. Drop out of the switch to
16143 SV * multi_char_N = newSVpvn(backslash_N_beg,
16144 RExC_parse - backslash_N_beg);
16146 = add_multi_match(multi_char_matches,
16151 } /* End of cp_count != 1 */
16153 /* This element should not be processed further in this
16156 value = save_value;
16157 prevvalue = save_prevvalue;
16158 continue; /* Back to top of loop to get next char */
16161 /* Here, is a single code point, and <value> contains it */
16162 unicode_range = TRUE; /* \N{} are Unicode */
16170 /* We will handle any undefined properties ourselves */
16171 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
16172 /* And we actually would prefer to get
16173 * the straight inversion list of the
16174 * swash, since we will be accessing it
16175 * anyway, to save a little time */
16176 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
16178 if (RExC_parse >= RExC_end)
16179 vFAIL2("Empty \\%c", (U8)value);
16180 if (*RExC_parse == '{') {
16181 const U8 c = (U8)value;
16182 e = strchr(RExC_parse, '}');
16185 vFAIL2("Missing right brace on \\%c{}", c);
16189 while (isSPACE(*RExC_parse)) {
16193 if (UCHARAT(RExC_parse) == '^') {
16195 /* toggle. (The rhs xor gets the single bit that
16196 * differs between P and p; the other xor inverts just
16198 value ^= 'P' ^ 'p';
16201 while (isSPACE(*RExC_parse)) {
16206 if (e == RExC_parse)
16207 vFAIL2("Empty \\%c{}", c);
16209 n = e - RExC_parse;
16210 while (isSPACE(*(RExC_parse + n - 1)))
16212 } /* The \p isn't immediately followed by a '{' */
16213 else if (! isALPHA(*RExC_parse)) {
16214 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16215 vFAIL2("Character following \\%c must be '{' or a "
16216 "single-character Unicode property name",
16226 char* base_name; /* name after any packages are stripped */
16227 char* lookup_name = NULL;
16228 const char * const colon_colon = "::";
16230 /* Try to get the definition of the property into
16231 * <invlist>. If /i is in effect, the effective property
16232 * will have its name be <__NAME_i>. The design is
16233 * discussed in commit
16234 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
16235 name = savepv(Perl_form(aTHX_ "%.*s", (int)n, RExC_parse));
16238 lookup_name = savepv(Perl_form(aTHX_ "__%s_i", name));
16240 /* The function call just below that uses this can fail
16241 * to return, leaking memory if we don't do this */
16242 SAVEFREEPV(lookup_name);
16245 /* Look up the property name, and get its swash and
16246 * inversion list, if the property is found */
16247 SvREFCNT_dec(swash); /* Free any left-overs */
16248 swash = _core_swash_init("utf8",
16255 NULL, /* No inversion list */
16258 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
16259 HV* curpkg = (IN_PERL_COMPILETIME)
16261 : CopSTASH(PL_curcop);
16265 if (swash) { /* Got a swash but no inversion list.
16266 Something is likely wrong that will
16267 be sorted-out later */
16268 SvREFCNT_dec_NN(swash);
16272 /* Here didn't find it. It could be a an error (like a
16273 * typo) in specifying a Unicode property, or it could
16274 * be a user-defined property that will be available at
16275 * run-time. The names of these must begin with 'In'
16276 * or 'Is' (after any packages are stripped off). So
16277 * if not one of those, or if we accept only
16278 * compile-time properties, is an error; otherwise add
16279 * it to the list for run-time look up. */
16280 if ((base_name = rninstr(name, name + n,
16281 colon_colon, colon_colon + 2)))
16282 { /* Has ::. We know this must be a user-defined
16285 final_n -= base_name - name;
16294 || base_name[0] != 'I'
16295 || (base_name[1] != 's' && base_name[1] != 'n')
16298 const char * const msg
16300 ? "Illegal user-defined property name"
16301 : "Can't find Unicode property definition";
16302 RExC_parse = e + 1;
16304 /* diag_listed_as: Can't find Unicode property definition "%s" */
16305 vFAIL3utf8f("%s \"%" UTF8f "\"",
16306 msg, UTF8fARG(UTF, n, name));
16309 /* If the property name doesn't already have a package
16310 * name, add the current one to it so that it can be
16311 * referred to outside it. [perl #121777] */
16312 if (! has_pkg && curpkg) {
16313 char* pkgname = HvNAME(curpkg);
16314 if (strNE(pkgname, "main")) {
16315 char* full_name = Perl_form(aTHX_
16319 n = strlen(full_name);
16320 name = savepvn(full_name, n);
16324 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s%" UTF8f "%s\n",
16325 (value == 'p' ? '+' : '!'),
16326 (FOLD) ? "__" : "",
16327 UTF8fARG(UTF, n, name),
16328 (FOLD) ? "_i" : "");
16329 has_user_defined_property = TRUE;
16330 optimizable = FALSE; /* Will have to leave this an
16333 /* We don't know yet what this matches, so have to flag
16335 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
16339 /* Here, did get the swash and its inversion list. If
16340 * the swash is from a user-defined property, then this
16341 * whole character class should be regarded as such */
16342 if (swash_init_flags
16343 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
16345 has_user_defined_property = TRUE;
16348 /* We warn on matching an above-Unicode code point
16349 * if the match would return true, except don't
16350 * warn for \p{All}, which has exactly one element
16352 (_invlist_contains_cp(invlist, 0x110000)
16353 && (! (_invlist_len(invlist) == 1
16354 && *invlist_array(invlist) == 0)))
16360 /* Invert if asking for the complement */
16361 if (value == 'P') {
16362 _invlist_union_complement_2nd(properties,
16366 /* The swash can't be used as-is, because we've
16367 * inverted things; delay removing it to here after
16368 * have copied its invlist above */
16369 SvREFCNT_dec_NN(swash);
16373 _invlist_union(properties, invlist, &properties);
16377 RExC_parse = e + 1;
16378 namedclass = ANYOF_UNIPROP; /* no official name, but it's
16381 /* \p means they want Unicode semantics */
16382 REQUIRE_UNI_RULES(flagp, NULL);
16385 case 'n': value = '\n'; break;
16386 case 'r': value = '\r'; break;
16387 case 't': value = '\t'; break;
16388 case 'f': value = '\f'; break;
16389 case 'b': value = '\b'; break;
16390 case 'e': value = ESC_NATIVE; break;
16391 case 'a': value = '\a'; break;
16393 RExC_parse--; /* function expects to be pointed at the 'o' */
16395 const char* error_msg;
16396 bool valid = grok_bslash_o(&RExC_parse,
16399 PASS2, /* warnings only in
16402 silence_non_portable,
16408 non_portable_endpoint++;
16411 RExC_parse--; /* function expects to be pointed at the 'x' */
16413 const char* error_msg;
16414 bool valid = grok_bslash_x(&RExC_parse,
16417 PASS2, /* Output warnings */
16419 silence_non_portable,
16425 non_portable_endpoint++;
16428 value = grok_bslash_c(*RExC_parse++, PASS2);
16429 non_portable_endpoint++;
16431 case '0': case '1': case '2': case '3': case '4':
16432 case '5': case '6': case '7':
16434 /* Take 1-3 octal digits */
16435 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
16436 numlen = (strict) ? 4 : 3;
16437 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
16438 RExC_parse += numlen;
16441 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16442 vFAIL("Need exactly 3 octal digits");
16444 else if (! SIZE_ONLY /* like \08, \178 */
16446 && RExC_parse < RExC_end
16447 && isDIGIT(*RExC_parse)
16448 && ckWARN(WARN_REGEXP))
16450 SAVEFREESV(RExC_rx_sv);
16451 reg_warn_non_literal_string(
16453 form_short_octal_warning(RExC_parse, numlen));
16454 (void)ReREFCNT_inc(RExC_rx_sv);
16457 non_portable_endpoint++;
16461 /* Allow \_ to not give an error */
16462 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
16464 vFAIL2("Unrecognized escape \\%c in character class",
16468 SAVEFREESV(RExC_rx_sv);
16469 ckWARN2reg(RExC_parse,
16470 "Unrecognized escape \\%c in character class passed through",
16472 (void)ReREFCNT_inc(RExC_rx_sv);
16476 } /* End of switch on char following backslash */
16477 } /* end of handling backslash escape sequences */
16479 /* Here, we have the current token in 'value' */
16481 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
16484 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
16485 * literal, as is the character that began the false range, i.e.
16486 * the 'a' in the examples */
16489 const int w = (RExC_parse >= rangebegin)
16490 ? RExC_parse - rangebegin
16494 "False [] range \"%" UTF8f "\"",
16495 UTF8fARG(UTF, w, rangebegin));
16498 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
16499 ckWARN2reg(RExC_parse,
16500 "False [] range \"%" UTF8f "\"",
16501 UTF8fARG(UTF, w, rangebegin));
16502 (void)ReREFCNT_inc(RExC_rx_sv);
16503 cp_list = add_cp_to_invlist(cp_list, '-');
16504 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
16509 range = 0; /* this was not a true range */
16510 element_count += 2; /* So counts for three values */
16513 classnum = namedclass_to_classnum(namedclass);
16515 if (LOC && namedclass < ANYOF_POSIXL_MAX
16516 #ifndef HAS_ISASCII
16517 && classnum != _CC_ASCII
16520 /* What the Posix classes (like \w, [:space:]) match in locale
16521 * isn't knowable under locale until actual match time. Room
16522 * must be reserved (one time per outer bracketed class) to
16523 * store such classes. The space will contain a bit for each
16524 * named class that is to be matched against. This isn't
16525 * needed for \p{} and pseudo-classes, as they are not affected
16526 * by locale, and hence are dealt with separately */
16527 if (! need_class) {
16530 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16533 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16535 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
16536 ANYOF_POSIXL_ZERO(ret);
16538 /* We can't change this into some other type of node
16539 * (unless this is the only element, in which case there
16540 * are nodes that mean exactly this) as has runtime
16542 optimizable = FALSE;
16545 /* Coverity thinks it is possible for this to be negative; both
16546 * jhi and khw think it's not, but be safer */
16547 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16548 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
16550 /* See if it already matches the complement of this POSIX
16552 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16553 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
16557 posixl_matches_all = TRUE;
16558 break; /* No need to continue. Since it matches both
16559 e.g., \w and \W, it matches everything, and the
16560 bracketed class can be optimized into qr/./s */
16563 /* Add this class to those that should be checked at runtime */
16564 ANYOF_POSIXL_SET(ret, namedclass);
16566 /* The above-Latin1 characters are not subject to locale rules.
16567 * Just add them, in the second pass, to the
16568 * unconditionally-matched list */
16570 SV* scratch_list = NULL;
16572 /* Get the list of the above-Latin1 code points this
16574 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
16575 PL_XPosix_ptrs[classnum],
16577 /* Odd numbers are complements, like
16578 * NDIGIT, NASCII, ... */
16579 namedclass % 2 != 0,
16581 /* Checking if 'cp_list' is NULL first saves an extra
16582 * clone. Its reference count will be decremented at the
16583 * next union, etc, or if this is the only instance, at the
16584 * end of the routine */
16586 cp_list = scratch_list;
16589 _invlist_union(cp_list, scratch_list, &cp_list);
16590 SvREFCNT_dec_NN(scratch_list);
16592 continue; /* Go get next character */
16595 else if (! SIZE_ONLY) {
16597 /* Here, not in pass1 (in that pass we skip calculating the
16598 * contents of this class), and is not /l, or is a POSIX class
16599 * for which /l doesn't matter (or is a Unicode property, which
16600 * is skipped here). */
16601 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
16602 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
16604 /* Here, should be \h, \H, \v, or \V. None of /d, /i
16605 * nor /l make a difference in what these match,
16606 * therefore we just add what they match to cp_list. */
16607 if (classnum != _CC_VERTSPACE) {
16608 assert( namedclass == ANYOF_HORIZWS
16609 || namedclass == ANYOF_NHORIZWS);
16611 /* It turns out that \h is just a synonym for
16613 classnum = _CC_BLANK;
16616 _invlist_union_maybe_complement_2nd(
16618 PL_XPosix_ptrs[classnum],
16619 namedclass % 2 != 0, /* Complement if odd
16620 (NHORIZWS, NVERTWS)
16625 else if ( UNI_SEMANTICS
16626 || classnum == _CC_ASCII
16627 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
16628 || classnum == _CC_XDIGIT)))
16630 /* We usually have to worry about /d and /a affecting what
16631 * POSIX classes match, with special code needed for /d
16632 * because we won't know until runtime what all matches.
16633 * But there is no extra work needed under /u, and
16634 * [:ascii:] is unaffected by /a and /d; and :digit: and
16635 * :xdigit: don't have runtime differences under /d. So we
16636 * can special case these, and avoid some extra work below,
16637 * and at runtime. */
16638 _invlist_union_maybe_complement_2nd(
16640 PL_XPosix_ptrs[classnum],
16641 namedclass % 2 != 0,
16644 else { /* Garden variety class. If is NUPPER, NALPHA, ...
16645 complement and use nposixes */
16646 SV** posixes_ptr = namedclass % 2 == 0
16649 _invlist_union_maybe_complement_2nd(
16651 PL_XPosix_ptrs[classnum],
16652 namedclass % 2 != 0,
16656 } /* end of namedclass \blah */
16658 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16660 /* If 'range' is set, 'value' is the ending of a range--check its
16661 * validity. (If value isn't a single code point in the case of a
16662 * range, we should have figured that out above in the code that
16663 * catches false ranges). Later, we will handle each individual code
16664 * point in the range. If 'range' isn't set, this could be the
16665 * beginning of a range, so check for that by looking ahead to see if
16666 * the next real character to be processed is the range indicator--the
16671 /* For unicode ranges, we have to test that the Unicode as opposed
16672 * to the native values are not decreasing. (Above 255, there is
16673 * no difference between native and Unicode) */
16674 if (unicode_range && prevvalue < 255 && value < 255) {
16675 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
16676 goto backwards_range;
16681 if (prevvalue > value) /* b-a */ {
16686 w = RExC_parse - rangebegin;
16688 "Invalid [] range \"%" UTF8f "\"",
16689 UTF8fARG(UTF, w, rangebegin));
16690 NOT_REACHED; /* NOTREACHED */
16694 prevvalue = value; /* save the beginning of the potential range */
16695 if (! stop_at_1 /* Can't be a range if parsing just one thing */
16696 && *RExC_parse == '-')
16698 char* next_char_ptr = RExC_parse + 1;
16700 /* Get the next real char after the '-' */
16701 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
16703 /* If the '-' is at the end of the class (just before the ']',
16704 * it is a literal minus; otherwise it is a range */
16705 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
16706 RExC_parse = next_char_ptr;
16708 /* a bad range like \w-, [:word:]- ? */
16709 if (namedclass > OOB_NAMEDCLASS) {
16710 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
16711 const int w = RExC_parse >= rangebegin
16712 ? RExC_parse - rangebegin
16715 vFAIL4("False [] range \"%*.*s\"",
16720 "False [] range \"%*.*s\"",
16725 cp_list = add_cp_to_invlist(cp_list, '-');
16729 range = 1; /* yeah, it's a range! */
16730 continue; /* but do it the next time */
16735 if (namedclass > OOB_NAMEDCLASS) {
16739 /* Here, we have a single value this time through the loop, and
16740 * <prevvalue> is the beginning of the range, if any; or <value> if
16743 /* non-Latin1 code point implies unicode semantics. Must be set in
16744 * pass1 so is there for the whole of pass 2 */
16746 REQUIRE_UNI_RULES(flagp, NULL);
16749 /* Ready to process either the single value, or the completed range.
16750 * For single-valued non-inverted ranges, we consider the possibility
16751 * of multi-char folds. (We made a conscious decision to not do this
16752 * for the other cases because it can often lead to non-intuitive
16753 * results. For example, you have the peculiar case that:
16754 * "s s" =~ /^[^\xDF]+$/i => Y
16755 * "ss" =~ /^[^\xDF]+$/i => N
16757 * See [perl #89750] */
16758 if (FOLD && allow_multi_folds && value == prevvalue) {
16759 if (value == LATIN_SMALL_LETTER_SHARP_S
16760 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
16763 /* Here <value> is indeed a multi-char fold. Get what it is */
16765 U8 foldbuf[UTF8_MAXBYTES_CASE];
16768 UV folded = _to_uni_fold_flags(
16772 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
16773 ? FOLD_FLAGS_NOMIX_ASCII
16777 /* Here, <folded> should be the first character of the
16778 * multi-char fold of <value>, with <foldbuf> containing the
16779 * whole thing. But, if this fold is not allowed (because of
16780 * the flags), <fold> will be the same as <value>, and should
16781 * be processed like any other character, so skip the special
16783 if (folded != value) {
16785 /* Skip if we are recursed, currently parsing the class
16786 * again. Otherwise add this character to the list of
16787 * multi-char folds. */
16788 if (! RExC_in_multi_char_class) {
16789 STRLEN cp_count = utf8_length(foldbuf,
16790 foldbuf + foldlen);
16791 SV* multi_fold = sv_2mortal(newSVpvs(""));
16793 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
16796 = add_multi_match(multi_char_matches,
16802 /* This element should not be processed further in this
16805 value = save_value;
16806 prevvalue = save_prevvalue;
16812 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
16815 /* If the range starts above 255, everything is portable and
16816 * likely to be so for any forseeable character set, so don't
16818 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
16819 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
16821 else if (prevvalue != value) {
16823 /* Under strict, ranges that stop and/or end in an ASCII
16824 * printable should have each end point be a portable value
16825 * for it (preferably like 'A', but we don't warn if it is
16826 * a (portable) Unicode name or code point), and the range
16827 * must be be all digits or all letters of the same case.
16828 * Otherwise, the range is non-portable and unclear as to
16829 * what it contains */
16830 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
16831 && ( non_portable_endpoint
16832 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
16833 || (isLOWER_A(prevvalue) && isLOWER_A(value))
16834 || (isUPPER_A(prevvalue) && isUPPER_A(value))
16836 vWARN(RExC_parse, "Ranges of ASCII printables should"
16837 " be some subset of \"0-9\","
16838 " \"A-Z\", or \"a-z\"");
16840 else if (prevvalue >= 0x660) { /* ARABIC_INDIC_DIGIT_ZERO */
16841 SSize_t index_start;
16842 SSize_t index_final;
16844 /* But the nature of Unicode and languages mean we
16845 * can't do the same checks for above-ASCII ranges,
16846 * except in the case of digit ones. These should
16847 * contain only digits from the same group of 10. The
16848 * ASCII case is handled just above. 0x660 is the
16849 * first digit character beyond ASCII. Hence here, the
16850 * range could be a range of digits. First some
16851 * unlikely special cases. Grandfather in that a range
16852 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
16853 * if its starting value is one of the 10 digits prior
16854 * to it. This is because it is an alternate way of
16855 * writing 19D1, and some people may expect it to be in
16856 * that group. But it is bad, because it won't give
16857 * the expected results. In Unicode 5.2 it was
16858 * considered to be in that group (of 11, hence), but
16859 * this was fixed in the next version */
16861 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
16862 goto warn_bad_digit_range;
16864 else if (UNLIKELY( prevvalue >= 0x1D7CE
16865 && value <= 0x1D7FF))
16867 /* This is the only other case currently in Unicode
16868 * where the algorithm below fails. The code
16869 * points just above are the end points of a single
16870 * range containing only decimal digits. It is 5
16871 * different series of 0-9. All other ranges of
16872 * digits currently in Unicode are just a single
16873 * series. (And mktables will notify us if a later
16874 * Unicode version breaks this.)
16876 * If the range being checked is at most 9 long,
16877 * and the digit values represented are in
16878 * numerical order, they are from the same series.
16880 if ( value - prevvalue > 9
16881 || ((( value - 0x1D7CE) % 10)
16882 <= (prevvalue - 0x1D7CE) % 10))
16884 goto warn_bad_digit_range;
16889 /* For all other ranges of digits in Unicode, the
16890 * algorithm is just to check if both end points
16891 * are in the same series, which is the same range.
16893 index_start = _invlist_search(
16894 PL_XPosix_ptrs[_CC_DIGIT],
16897 /* Warn if the range starts and ends with a digit,
16898 * and they are not in the same group of 10. */
16899 if ( index_start >= 0
16900 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
16902 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
16903 value)) != index_start
16904 && index_final >= 0
16905 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
16907 warn_bad_digit_range:
16908 vWARN(RExC_parse, "Ranges of digits should be"
16909 " from the same group of"
16916 if ((! range || prevvalue == value) && non_portable_endpoint) {
16917 if (isPRINT_A(value)) {
16920 if (isBACKSLASHED_PUNCT(value)) {
16921 literal[d++] = '\\';
16923 literal[d++] = (char) value;
16924 literal[d++] = '\0';
16926 vWARN4dep(RExC_parse,
16927 "\"%.*s\" is more clearly written simply as \"%s\". "
16928 "This will be a fatal error in Perl 5.28",
16929 (int) (RExC_parse - rangebegin),
16934 else if isMNEMONIC_CNTRL(value) {
16935 vWARN4dep(RExC_parse,
16936 "\"%.*s\" is more clearly written simply as \"%s\". "
16937 "This will be a fatal error in Perl 5.28",
16938 (int) (RExC_parse - rangebegin),
16940 cntrl_to_mnemonic((U8) value)
16946 /* Deal with this element of the class */
16950 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16953 /* On non-ASCII platforms, for ranges that span all of 0..255, and
16954 * ones that don't require special handling, we can just add the
16955 * range like we do for ASCII platforms */
16956 if ((UNLIKELY(prevvalue == 0) && value >= 255)
16957 || ! (prevvalue < 256
16959 || (! non_portable_endpoint
16960 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
16961 || (isUPPER_A(prevvalue)
16962 && isUPPER_A(value)))))))
16964 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16968 /* Here, requires special handling. This can be because it is
16969 * a range whose code points are considered to be Unicode, and
16970 * so must be individually translated into native, or because
16971 * its a subrange of 'A-Z' or 'a-z' which each aren't
16972 * contiguous in EBCDIC, but we have defined them to include
16973 * only the "expected" upper or lower case ASCII alphabetics.
16974 * Subranges above 255 are the same in native and Unicode, so
16975 * can be added as a range */
16976 U8 start = NATIVE_TO_LATIN1(prevvalue);
16978 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
16979 for (j = start; j <= end; j++) {
16980 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
16983 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16990 range = 0; /* this range (if it was one) is done now */
16991 } /* End of loop through all the text within the brackets */
16994 if ( posix_warnings && av_tindex_nomg(posix_warnings) >= 0) {
16995 output_or_return_posix_warnings(pRExC_state, posix_warnings,
16996 return_posix_warnings);
16999 /* If anything in the class expands to more than one character, we have to
17000 * deal with them by building up a substitute parse string, and recursively
17001 * calling reg() on it, instead of proceeding */
17002 if (multi_char_matches) {
17003 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
17006 char *save_end = RExC_end;
17007 char *save_parse = RExC_parse;
17008 char *save_start = RExC_start;
17009 STRLEN prefix_end = 0; /* We copy the character class after a
17010 prefix supplied here. This is the size
17011 + 1 of that prefix */
17012 bool first_time = TRUE; /* First multi-char occurrence doesn't get
17017 assert(RExC_precomp_adj == 0); /* Only one level of recursion allowed */
17019 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
17020 because too confusing */
17022 sv_catpv(substitute_parse, "(?:");
17026 /* Look at the longest folds first */
17027 for (cp_count = av_tindex_nomg(multi_char_matches);
17032 if (av_exists(multi_char_matches, cp_count)) {
17033 AV** this_array_ptr;
17036 this_array_ptr = (AV**) av_fetch(multi_char_matches,
17038 while ((this_sequence = av_pop(*this_array_ptr)) !=
17041 if (! first_time) {
17042 sv_catpv(substitute_parse, "|");
17044 first_time = FALSE;
17046 sv_catpv(substitute_parse, SvPVX(this_sequence));
17051 /* If the character class contains anything else besides these
17052 * multi-character folds, have to include it in recursive parsing */
17053 if (element_count) {
17054 sv_catpv(substitute_parse, "|[");
17055 prefix_end = SvCUR(substitute_parse);
17056 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
17058 /* Put in a closing ']' only if not going off the end, as otherwise
17059 * we are adding something that really isn't there */
17060 if (RExC_parse < RExC_end) {
17061 sv_catpv(substitute_parse, "]");
17065 sv_catpv(substitute_parse, ")");
17068 /* This is a way to get the parse to skip forward a whole named
17069 * sequence instead of matching the 2nd character when it fails the
17071 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
17075 /* Set up the data structure so that any errors will be properly
17076 * reported. See the comments at the definition of
17077 * REPORT_LOCATION_ARGS for details */
17078 RExC_precomp_adj = orig_parse - RExC_precomp;
17079 RExC_start = RExC_parse = SvPV(substitute_parse, len);
17080 RExC_adjusted_start = RExC_start + prefix_end;
17081 RExC_end = RExC_parse + len;
17082 RExC_in_multi_char_class = 1;
17083 RExC_emit = (regnode *)orig_emit;
17085 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17087 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PASS1|NEED_UTF8);
17089 /* And restore so can parse the rest of the pattern */
17090 RExC_parse = save_parse;
17091 RExC_start = RExC_adjusted_start = save_start;
17092 RExC_precomp_adj = 0;
17093 RExC_end = save_end;
17094 RExC_in_multi_char_class = 0;
17095 SvREFCNT_dec_NN(multi_char_matches);
17099 /* Here, we've gone through the entire class and dealt with multi-char
17100 * folds. We are now in a position that we can do some checks to see if we
17101 * can optimize this ANYOF node into a simpler one, even in Pass 1.
17102 * Currently we only do two checks:
17103 * 1) is in the unlikely event that the user has specified both, eg. \w and
17104 * \W under /l, then the class matches everything. (This optimization
17105 * is done only to make the optimizer code run later work.)
17106 * 2) if the character class contains only a single element (including a
17107 * single range), we see if there is an equivalent node for it.
17108 * Other checks are possible */
17110 && ! ret_invlist /* Can't optimize if returning the constructed
17112 && (UNLIKELY(posixl_matches_all) || element_count == 1))
17117 if (UNLIKELY(posixl_matches_all)) {
17120 else if (namedclass > OOB_NAMEDCLASS) { /* this is a single named
17121 class, like \w or [:digit:]
17124 /* All named classes are mapped into POSIXish nodes, with its FLAG
17125 * argument giving which class it is */
17126 switch ((I32)namedclass) {
17127 case ANYOF_UNIPROP:
17130 /* These don't depend on the charset modifiers. They always
17131 * match under /u rules */
17132 case ANYOF_NHORIZWS:
17133 case ANYOF_HORIZWS:
17134 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
17137 case ANYOF_NVERTWS:
17142 /* The actual POSIXish node for all the rest depends on the
17143 * charset modifier. The ones in the first set depend only on
17144 * ASCII or, if available on this platform, also locale */
17148 op = (LOC) ? POSIXL : POSIXA;
17154 /* The following don't have any matches in the upper Latin1
17155 * range, hence /d is equivalent to /u for them. Making it /u
17156 * saves some branches at runtime */
17160 case ANYOF_NXDIGIT:
17161 if (! DEPENDS_SEMANTICS) {
17162 goto treat_as_default;
17168 /* The following change to CASED under /i */
17174 namedclass = ANYOF_CASED + (namedclass % 2);
17178 /* The rest have more possibilities depending on the charset.
17179 * We take advantage of the enum ordering of the charset
17180 * modifiers to get the exact node type, */
17183 op = POSIXD + get_regex_charset(RExC_flags);
17184 if (op > POSIXA) { /* /aa is same as /a */
17189 /* The odd numbered ones are the complements of the
17190 * next-lower even number one */
17191 if (namedclass % 2 == 1) {
17195 arg = namedclass_to_classnum(namedclass);
17199 else if (value == prevvalue) {
17201 /* Here, the class consists of just a single code point */
17204 if (! LOC && value == '\n') {
17205 op = REG_ANY; /* Optimize [^\n] */
17206 *flagp |= HASWIDTH|SIMPLE;
17210 else if (value < 256 || UTF) {
17212 /* Optimize a single value into an EXACTish node, but not if it
17213 * would require converting the pattern to UTF-8. */
17214 op = compute_EXACTish(pRExC_state);
17216 } /* Otherwise is a range */
17217 else if (! LOC) { /* locale could vary these */
17218 if (prevvalue == '0') {
17219 if (value == '9') {
17224 else if (! FOLD || ASCII_FOLD_RESTRICTED) {
17225 /* We can optimize A-Z or a-z, but not if they could match
17226 * something like the KELVIN SIGN under /i. */
17227 if (prevvalue == 'A') {
17230 && ! non_portable_endpoint
17233 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
17237 else if (prevvalue == 'a') {
17240 && ! non_portable_endpoint
17243 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
17250 /* Here, we have changed <op> away from its initial value iff we found
17251 * an optimization */
17254 /* Throw away this ANYOF regnode, and emit the calculated one,
17255 * which should correspond to the beginning, not current, state of
17257 const char * cur_parse = RExC_parse;
17258 RExC_parse = (char *)orig_parse;
17262 /* To get locale nodes to not use the full ANYOF size would
17263 * require moving the code above that writes the portions
17264 * of it that aren't in other nodes to after this point.
17265 * e.g. ANYOF_POSIXL_SET */
17266 RExC_size = orig_size;
17270 RExC_emit = (regnode *)orig_emit;
17271 if (PL_regkind[op] == POSIXD) {
17272 if (op == POSIXL) {
17273 RExC_contains_locale = 1;
17276 op += NPOSIXD - POSIXD;
17281 ret = reg_node(pRExC_state, op);
17283 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17287 *flagp |= HASWIDTH|SIMPLE;
17289 else if (PL_regkind[op] == EXACT) {
17290 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17291 TRUE /* downgradable to EXACT */
17295 RExC_parse = (char *) cur_parse;
17297 SvREFCNT_dec(posixes);
17298 SvREFCNT_dec(nposixes);
17299 SvREFCNT_dec(simple_posixes);
17300 SvREFCNT_dec(cp_list);
17301 SvREFCNT_dec(cp_foldable_list);
17308 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
17310 /* If folding, we calculate all characters that could fold to or from the
17311 * ones already on the list */
17312 if (cp_foldable_list) {
17314 UV start, end; /* End points of code point ranges */
17316 SV* fold_intersection = NULL;
17319 /* Our calculated list will be for Unicode rules. For locale
17320 * matching, we have to keep a separate list that is consulted at
17321 * runtime only when the locale indicates Unicode rules. For
17322 * non-locale, we just use the general list */
17324 use_list = &only_utf8_locale_list;
17327 use_list = &cp_list;
17330 /* Only the characters in this class that participate in folds need
17331 * be checked. Get the intersection of this class and all the
17332 * possible characters that are foldable. This can quickly narrow
17333 * down a large class */
17334 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
17335 &fold_intersection);
17337 /* The folds for all the Latin1 characters are hard-coded into this
17338 * program, but we have to go out to disk to get the others. */
17339 if (invlist_highest(cp_foldable_list) >= 256) {
17341 /* This is a hash that for a particular fold gives all
17342 * characters that are involved in it */
17343 if (! PL_utf8_foldclosures) {
17344 _load_PL_utf8_foldclosures();
17348 /* Now look at the foldable characters in this class individually */
17349 invlist_iterinit(fold_intersection);
17350 while (invlist_iternext(fold_intersection, &start, &end)) {
17353 /* Look at every character in the range */
17354 for (j = start; j <= end; j++) {
17355 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17361 if (IS_IN_SOME_FOLD_L1(j)) {
17363 /* ASCII is always matched; non-ASCII is matched
17364 * only under Unicode rules (which could happen
17365 * under /l if the locale is a UTF-8 one */
17366 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
17367 *use_list = add_cp_to_invlist(*use_list,
17368 PL_fold_latin1[j]);
17371 has_upper_latin1_only_utf8_matches
17372 = add_cp_to_invlist(
17373 has_upper_latin1_only_utf8_matches,
17374 PL_fold_latin1[j]);
17378 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
17379 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
17381 add_above_Latin1_folds(pRExC_state,
17388 /* Here is an above Latin1 character. We don't have the
17389 * rules hard-coded for it. First, get its fold. This is
17390 * the simple fold, as the multi-character folds have been
17391 * handled earlier and separated out */
17392 _to_uni_fold_flags(j, foldbuf, &foldlen,
17393 (ASCII_FOLD_RESTRICTED)
17394 ? FOLD_FLAGS_NOMIX_ASCII
17397 /* Single character fold of above Latin1. Add everything in
17398 * its fold closure to the list that this node should match.
17399 * The fold closures data structure is a hash with the keys
17400 * being the UTF-8 of every character that is folded to, like
17401 * 'k', and the values each an array of all code points that
17402 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
17403 * Multi-character folds are not included */
17404 if ((listp = hv_fetch(PL_utf8_foldclosures,
17405 (char *) foldbuf, foldlen, FALSE)))
17407 AV* list = (AV*) *listp;
17409 for (k = 0; k <= av_tindex_nomg(list); k++) {
17410 SV** c_p = av_fetch(list, k, FALSE);
17416 /* /aa doesn't allow folds between ASCII and non- */
17417 if ((ASCII_FOLD_RESTRICTED
17418 && (isASCII(c) != isASCII(j))))
17423 /* Folds under /l which cross the 255/256 boundary
17424 * are added to a separate list. (These are valid
17425 * only when the locale is UTF-8.) */
17426 if (c < 256 && LOC) {
17427 *use_list = add_cp_to_invlist(*use_list, c);
17431 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
17433 cp_list = add_cp_to_invlist(cp_list, c);
17436 /* Similarly folds involving non-ascii Latin1
17437 * characters under /d are added to their list */
17438 has_upper_latin1_only_utf8_matches
17439 = add_cp_to_invlist(
17440 has_upper_latin1_only_utf8_matches,
17447 SvREFCNT_dec_NN(fold_intersection);
17450 /* Now that we have finished adding all the folds, there is no reason
17451 * to keep the foldable list separate */
17452 _invlist_union(cp_list, cp_foldable_list, &cp_list);
17453 SvREFCNT_dec_NN(cp_foldable_list);
17456 /* And combine the result (if any) with any inversion lists from posix
17457 * classes. The lists are kept separate up to now because we don't want to
17458 * fold the classes (folding of those is automatically handled by the swash
17459 * fetching code) */
17460 if (simple_posixes) { /* These are the classes known to be unaffected by
17463 _invlist_union(cp_list, simple_posixes, &cp_list);
17464 SvREFCNT_dec_NN(simple_posixes);
17467 cp_list = simple_posixes;
17470 if (posixes || nposixes) {
17472 /* We have to adjust /a and /aa */
17473 if (AT_LEAST_ASCII_RESTRICTED) {
17475 /* Under /a and /aa, nothing above ASCII matches these */
17477 _invlist_intersection(posixes,
17478 PL_XPosix_ptrs[_CC_ASCII],
17482 /* Under /a and /aa, everything above ASCII matches these
17485 _invlist_union_complement_2nd(nposixes,
17486 PL_XPosix_ptrs[_CC_ASCII],
17491 if (! DEPENDS_SEMANTICS) {
17493 /* For everything but /d, we can just add the current 'posixes' and
17494 * 'nposixes' to the main list */
17497 _invlist_union(cp_list, posixes, &cp_list);
17498 SvREFCNT_dec_NN(posixes);
17506 _invlist_union(cp_list, nposixes, &cp_list);
17507 SvREFCNT_dec_NN(nposixes);
17510 cp_list = nposixes;
17515 /* Under /d, things like \w match upper Latin1 characters only if
17516 * the target string is in UTF-8. But things like \W match all the
17517 * upper Latin1 characters if the target string is not in UTF-8.
17519 * Handle the case where there something like \W separately */
17521 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1);
17523 /* A complemented posix class matches all upper Latin1
17524 * characters if not in UTF-8. And it matches just certain
17525 * ones when in UTF-8. That means those certain ones are
17526 * matched regardless, so can just be added to the
17527 * unconditional list */
17529 _invlist_union(cp_list, nposixes, &cp_list);
17530 SvREFCNT_dec_NN(nposixes);
17534 cp_list = nposixes;
17537 /* Likewise for 'posixes' */
17538 _invlist_union(posixes, cp_list, &cp_list);
17540 /* Likewise for anything else in the range that matched only
17542 if (has_upper_latin1_only_utf8_matches) {
17543 _invlist_union(cp_list,
17544 has_upper_latin1_only_utf8_matches,
17546 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17547 has_upper_latin1_only_utf8_matches = NULL;
17550 /* If we don't match all the upper Latin1 characters regardless
17551 * of UTF-8ness, we have to set a flag to match the rest when
17553 _invlist_subtract(only_non_utf8_list, cp_list,
17554 &only_non_utf8_list);
17555 if (_invlist_len(only_non_utf8_list) != 0) {
17556 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17560 /* Here there were no complemented posix classes. That means
17561 * the upper Latin1 characters in 'posixes' match only when the
17562 * target string is in UTF-8. So we have to add them to the
17563 * list of those types of code points, while adding the
17564 * remainder to the unconditional list.
17566 * First calculate what they are */
17567 SV* nonascii_but_latin1_properties = NULL;
17568 _invlist_intersection(posixes, PL_UpperLatin1,
17569 &nonascii_but_latin1_properties);
17571 /* And add them to the final list of such characters. */
17572 _invlist_union(has_upper_latin1_only_utf8_matches,
17573 nonascii_but_latin1_properties,
17574 &has_upper_latin1_only_utf8_matches);
17576 /* Remove them from what now becomes the unconditional list */
17577 _invlist_subtract(posixes, nonascii_but_latin1_properties,
17580 /* And add those unconditional ones to the final list */
17582 _invlist_union(cp_list, posixes, &cp_list);
17583 SvREFCNT_dec_NN(posixes);
17590 SvREFCNT_dec(nonascii_but_latin1_properties);
17592 /* Get rid of any characters that we now know are matched
17593 * unconditionally from the conditional list, which may make
17594 * that list empty */
17595 _invlist_subtract(has_upper_latin1_only_utf8_matches,
17597 &has_upper_latin1_only_utf8_matches);
17598 if (_invlist_len(has_upper_latin1_only_utf8_matches) == 0) {
17599 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17600 has_upper_latin1_only_utf8_matches = NULL;
17606 /* And combine the result (if any) with any inversion list from properties.
17607 * The lists are kept separate up to now so that we can distinguish the two
17608 * in regards to matching above-Unicode. A run-time warning is generated
17609 * if a Unicode property is matched against a non-Unicode code point. But,
17610 * we allow user-defined properties to match anything, without any warning,
17611 * and we also suppress the warning if there is a portion of the character
17612 * class that isn't a Unicode property, and which matches above Unicode, \W
17613 * or [\x{110000}] for example.
17614 * (Note that in this case, unlike the Posix one above, there is no
17615 * <has_upper_latin1_only_utf8_matches>, because having a Unicode property
17616 * forces Unicode semantics */
17620 /* If it matters to the final outcome, see if a non-property
17621 * component of the class matches above Unicode. If so, the
17622 * warning gets suppressed. This is true even if just a single
17623 * such code point is specified, as, though not strictly correct if
17624 * another such code point is matched against, the fact that they
17625 * are using above-Unicode code points indicates they should know
17626 * the issues involved */
17628 warn_super = ! (invert
17629 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
17632 _invlist_union(properties, cp_list, &cp_list);
17633 SvREFCNT_dec_NN(properties);
17636 cp_list = properties;
17641 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17643 /* Because an ANYOF node is the only one that warns, this node
17644 * can't be optimized into something else */
17645 optimizable = FALSE;
17649 /* Here, we have calculated what code points should be in the character
17652 * Now we can see about various optimizations. Fold calculation (which we
17653 * did above) needs to take place before inversion. Otherwise /[^k]/i
17654 * would invert to include K, which under /i would match k, which it
17655 * shouldn't. Therefore we can't invert folded locale now, as it won't be
17656 * folded until runtime */
17658 /* If we didn't do folding, it's because some information isn't available
17659 * until runtime; set the run-time fold flag for these. (We don't have to
17660 * worry about properties folding, as that is taken care of by the swash
17661 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
17662 * locales, or the class matches at least one 0-255 range code point */
17665 /* Some things on the list might be unconditionally included because of
17666 * other components. Remove them, and clean up the list if it goes to
17668 if (only_utf8_locale_list && cp_list) {
17669 _invlist_subtract(only_utf8_locale_list, cp_list,
17670 &only_utf8_locale_list);
17672 if (_invlist_len(only_utf8_locale_list) == 0) {
17673 SvREFCNT_dec_NN(only_utf8_locale_list);
17674 only_utf8_locale_list = NULL;
17677 if (only_utf8_locale_list) {
17680 |ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17682 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
17684 invlist_iterinit(cp_list);
17685 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
17686 ANYOF_FLAGS(ret) |= ANYOFL_FOLD;
17688 invlist_iterfinish(cp_list);
17691 else if ( DEPENDS_SEMANTICS
17692 && ( has_upper_latin1_only_utf8_matches
17693 || (ANYOF_FLAGS(ret) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
17696 optimizable = FALSE;
17700 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
17701 * at compile time. Besides not inverting folded locale now, we can't
17702 * invert if there are things such as \w, which aren't known until runtime
17706 && OP(ret) != ANYOFD
17707 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
17708 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17710 _invlist_invert(cp_list);
17712 /* Any swash can't be used as-is, because we've inverted things */
17714 SvREFCNT_dec_NN(swash);
17718 /* Clear the invert flag since have just done it here */
17725 *ret_invlist = cp_list;
17726 SvREFCNT_dec(swash);
17728 /* Discard the generated node */
17730 RExC_size = orig_size;
17733 RExC_emit = orig_emit;
17738 /* Some character classes are equivalent to other nodes. Such nodes take
17739 * up less room and generally fewer operations to execute than ANYOF nodes.
17740 * Above, we checked for and optimized into some such equivalents for
17741 * certain common classes that are easy to test. Getting to this point in
17742 * the code means that the class didn't get optimized there. Since this
17743 * code is only executed in Pass 2, it is too late to save space--it has
17744 * been allocated in Pass 1, and currently isn't given back. But turning
17745 * things into an EXACTish node can allow the optimizer to join it to any
17746 * adjacent such nodes. And if the class is equivalent to things like /./,
17747 * expensive run-time swashes can be avoided. Now that we have more
17748 * complete information, we can find things necessarily missed by the
17749 * earlier code. Another possible "optimization" that isn't done is that
17750 * something like [Ee] could be changed into an EXACTFU. khw tried this
17751 * and found that the ANYOF is faster, including for code points not in the
17752 * bitmap. This still might make sense to do, provided it got joined with
17753 * an adjacent node(s) to create a longer EXACTFU one. This could be
17754 * accomplished by creating a pseudo ANYOF_EXACTFU node type that the join
17755 * routine would know is joinable. If that didn't happen, the node type
17756 * could then be made a straight ANYOF */
17758 if (optimizable && cp_list && ! invert) {
17760 U8 op = END; /* The optimzation node-type */
17761 int posix_class = -1; /* Illegal value */
17762 const char * cur_parse= RExC_parse;
17764 invlist_iterinit(cp_list);
17765 if (! invlist_iternext(cp_list, &start, &end)) {
17767 /* Here, the list is empty. This happens, for example, when a
17768 * Unicode property that doesn't match anything is the only element
17769 * in the character class (perluniprops.pod notes such properties).
17772 *flagp |= HASWIDTH|SIMPLE;
17774 else if (start == end) { /* The range is a single code point */
17775 if (! invlist_iternext(cp_list, &start, &end)
17777 /* Don't do this optimization if it would require changing
17778 * the pattern to UTF-8 */
17779 && (start < 256 || UTF))
17781 /* Here, the list contains a single code point. Can optimize
17782 * into an EXACTish node */
17793 /* A locale node under folding with one code point can be
17794 * an EXACTFL, as its fold won't be calculated until
17800 /* Here, we are generally folding, but there is only one
17801 * code point to match. If we have to, we use an EXACT
17802 * node, but it would be better for joining with adjacent
17803 * nodes in the optimization pass if we used the same
17804 * EXACTFish node that any such are likely to be. We can
17805 * do this iff the code point doesn't participate in any
17806 * folds. For example, an EXACTF of a colon is the same as
17807 * an EXACT one, since nothing folds to or from a colon. */
17809 if (IS_IN_SOME_FOLD_L1(value)) {
17814 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
17819 /* If we haven't found the node type, above, it means we
17820 * can use the prevailing one */
17822 op = compute_EXACTish(pRExC_state);
17826 } /* End of first range contains just a single code point */
17827 else if (start == 0) {
17828 if (end == UV_MAX) {
17830 *flagp |= HASWIDTH|SIMPLE;
17833 else if (end == '\n' - 1
17834 && invlist_iternext(cp_list, &start, &end)
17835 && start == '\n' + 1 && end == UV_MAX)
17838 *flagp |= HASWIDTH|SIMPLE;
17842 invlist_iterfinish(cp_list);
17845 const UV cp_list_len = _invlist_len(cp_list);
17846 const UV* cp_list_array = invlist_array(cp_list);
17848 /* Here, didn't find an optimization. See if this matches any of
17849 * the POSIX classes. These run slightly faster for above-Unicode
17850 * code points, so don't bother with POSIXA ones nor the 2 that
17851 * have no above-Unicode matches. We can avoid these checks unless
17852 * the ANYOF matches at least as high as the lowest POSIX one
17853 * (which was manually found to be \v. The actual code point may
17854 * increase in later Unicode releases, if a higher code point is
17855 * assigned to be \v, but this code will never break. It would
17856 * just mean we could execute the checks for posix optimizations
17857 * unnecessarily) */
17859 if (cp_list_array[cp_list_len-1] > 0x2029) {
17860 for (posix_class = 0;
17861 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
17865 if (posix_class == _CC_ASCII || posix_class == _CC_CNTRL) {
17868 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
17870 /* Check if matches normal or inverted */
17871 if (_invlistEQ(cp_list,
17872 PL_XPosix_ptrs[posix_class],
17875 op = (try_inverted)
17878 *flagp |= HASWIDTH|SIMPLE;
17888 RExC_parse = (char *)orig_parse;
17889 RExC_emit = (regnode *)orig_emit;
17891 if (regarglen[op]) {
17892 ret = reganode(pRExC_state, op, 0);
17894 ret = reg_node(pRExC_state, op);
17897 RExC_parse = (char *)cur_parse;
17899 if (PL_regkind[op] == EXACT) {
17900 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17901 TRUE /* downgradable to EXACT */
17904 else if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17905 FLAGS(ret) = posix_class;
17908 SvREFCNT_dec_NN(cp_list);
17913 /* Here, <cp_list> contains all the code points we can determine at
17914 * compile time that match under all conditions. Go through it, and
17915 * for things that belong in the bitmap, put them there, and delete from
17916 * <cp_list>. While we are at it, see if everything above 255 is in the
17917 * list, and if so, set a flag to speed up execution */
17919 populate_ANYOF_from_invlist(ret, &cp_list);
17922 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
17925 /* Here, the bitmap has been populated with all the Latin1 code points that
17926 * always match. Can now add to the overall list those that match only
17927 * when the target string is UTF-8 (<has_upper_latin1_only_utf8_matches>).
17929 if (has_upper_latin1_only_utf8_matches) {
17931 _invlist_union(cp_list,
17932 has_upper_latin1_only_utf8_matches,
17934 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17937 cp_list = has_upper_latin1_only_utf8_matches;
17939 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
17942 /* If there is a swash and more than one element, we can't use the swash in
17943 * the optimization below. */
17944 if (swash && element_count > 1) {
17945 SvREFCNT_dec_NN(swash);
17949 /* Note that the optimization of using 'swash' if it is the only thing in
17950 * the class doesn't have us change swash at all, so it can include things
17951 * that are also in the bitmap; otherwise we have purposely deleted that
17952 * duplicate information */
17953 set_ANYOF_arg(pRExC_state, ret, cp_list,
17954 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17956 only_utf8_locale_list,
17957 swash, has_user_defined_property);
17959 *flagp |= HASWIDTH|SIMPLE;
17961 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
17962 RExC_contains_locale = 1;
17968 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
17971 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
17972 regnode* const node,
17974 SV* const runtime_defns,
17975 SV* const only_utf8_locale_list,
17977 const bool has_user_defined_property)
17979 /* Sets the arg field of an ANYOF-type node 'node', using information about
17980 * the node passed-in. If there is nothing outside the node's bitmap, the
17981 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
17982 * the count returned by add_data(), having allocated and stored an array,
17983 * av, that that count references, as follows:
17984 * av[0] stores the character class description in its textual form.
17985 * This is used later (regexec.c:Perl_regclass_swash()) to
17986 * initialize the appropriate swash, and is also useful for dumping
17987 * the regnode. This is set to &PL_sv_undef if the textual
17988 * description is not needed at run-time (as happens if the other
17989 * elements completely define the class)
17990 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
17991 * computed from av[0]. But if no further computation need be done,
17992 * the swash is stored here now (and av[0] is &PL_sv_undef).
17993 * av[2] stores the inversion list of code points that match only if the
17994 * current locale is UTF-8
17995 * av[3] stores the cp_list inversion list for use in addition or instead
17996 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
17997 * (Otherwise everything needed is already in av[0] and av[1])
17998 * av[4] is set if any component of the class is from a user-defined
17999 * property; used only if av[3] exists */
18003 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
18005 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
18006 assert(! (ANYOF_FLAGS(node)
18007 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
18008 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
18011 AV * const av = newAV();
18014 av_store(av, 0, (runtime_defns)
18015 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
18018 av_store(av, 1, swash);
18019 SvREFCNT_dec_NN(cp_list);
18022 av_store(av, 1, &PL_sv_undef);
18024 av_store(av, 3, cp_list);
18025 av_store(av, 4, newSVuv(has_user_defined_property));
18029 if (only_utf8_locale_list) {
18030 av_store(av, 2, only_utf8_locale_list);
18033 av_store(av, 2, &PL_sv_undef);
18036 rv = newRV_noinc(MUTABLE_SV(av));
18037 n = add_data(pRExC_state, STR_WITH_LEN("s"));
18038 RExC_rxi->data->data[n] = (void*)rv;
18043 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
18045 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
18046 const regnode* node,
18049 SV** only_utf8_locale_ptr,
18050 SV** output_invlist)
18053 /* For internal core use only.
18054 * Returns the swash for the input 'node' in the regex 'prog'.
18055 * If <doinit> is 'true', will attempt to create the swash if not already
18057 * If <listsvp> is non-null, will return the printable contents of the
18058 * swash. This can be used to get debugging information even before the
18059 * swash exists, by calling this function with 'doinit' set to false, in
18060 * which case the components that will be used to eventually create the
18061 * swash are returned (in a printable form).
18062 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
18063 * store an inversion list of code points that should match only if the
18064 * execution-time locale is a UTF-8 one.
18065 * If <output_invlist> is not NULL, it is where this routine is to store an
18066 * inversion list of the code points that would be instead returned in
18067 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
18068 * when this parameter is used, is just the non-code point data that
18069 * will go into creating the swash. This currently should be just
18070 * user-defined properties whose definitions were not known at compile
18071 * time. Using this parameter allows for easier manipulation of the
18072 * swash's data by the caller. It is illegal to call this function with
18073 * this parameter set, but not <listsvp>
18075 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
18076 * that, in spite of this function's name, the swash it returns may include
18077 * the bitmap data as well */
18080 SV *si = NULL; /* Input swash initialization string */
18081 SV* invlist = NULL;
18083 RXi_GET_DECL(prog,progi);
18084 const struct reg_data * const data = prog ? progi->data : NULL;
18086 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
18087 assert(! output_invlist || listsvp);
18089 if (data && data->count) {
18090 const U32 n = ARG(node);
18092 if (data->what[n] == 's') {
18093 SV * const rv = MUTABLE_SV(data->data[n]);
18094 AV * const av = MUTABLE_AV(SvRV(rv));
18095 SV **const ary = AvARRAY(av);
18096 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
18098 si = *ary; /* ary[0] = the string to initialize the swash with */
18100 if (av_tindex_nomg(av) >= 2) {
18101 if (only_utf8_locale_ptr
18103 && ary[2] != &PL_sv_undef)
18105 *only_utf8_locale_ptr = ary[2];
18108 assert(only_utf8_locale_ptr);
18109 *only_utf8_locale_ptr = NULL;
18112 /* Elements 3 and 4 are either both present or both absent. [3]
18113 * is any inversion list generated at compile time; [4]
18114 * indicates if that inversion list has any user-defined
18115 * properties in it. */
18116 if (av_tindex_nomg(av) >= 3) {
18118 if (SvUV(ary[4])) {
18119 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
18127 /* Element [1] is reserved for the set-up swash. If already there,
18128 * return it; if not, create it and store it there */
18129 if (ary[1] && SvROK(ary[1])) {
18132 else if (doinit && ((si && si != &PL_sv_undef)
18133 || (invlist && invlist != &PL_sv_undef))) {
18135 sw = _core_swash_init("utf8", /* the utf8 package */
18139 0, /* not from tr/// */
18141 &swash_init_flags);
18142 (void)av_store(av, 1, sw);
18147 /* If requested, return a printable version of what this swash matches */
18149 SV* matches_string = NULL;
18151 /* The swash should be used, if possible, to get the data, as it
18152 * contains the resolved data. But this function can be called at
18153 * compile-time, before everything gets resolved, in which case we
18154 * return the currently best available information, which is the string
18155 * that will eventually be used to do that resolving, 'si' */
18156 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
18157 && (si && si != &PL_sv_undef))
18159 /* Here, we only have 'si' (and possibly some passed-in data in
18160 * 'invlist', which is handled below) If the caller only wants
18161 * 'si', use that. */
18162 if (! output_invlist) {
18163 matches_string = newSVsv(si);
18166 /* But if the caller wants an inversion list of the node, we
18167 * need to parse 'si' and place as much as possible in the
18168 * desired output inversion list, making 'matches_string' only
18169 * contain the currently unresolvable things */
18170 const char *si_string = SvPVX(si);
18171 STRLEN remaining = SvCUR(si);
18175 /* Ignore everything before the first new-line */
18176 while (*si_string != '\n' && remaining > 0) {
18180 assert(remaining > 0);
18185 while (remaining > 0) {
18187 /* The data consists of just strings defining user-defined
18188 * property names, but in prior incarnations, and perhaps
18189 * somehow from pluggable regex engines, it could still
18190 * hold hex code point definitions. Each component of a
18191 * range would be separated by a tab, and each range by a
18192 * new-line. If these are found, instead add them to the
18193 * inversion list */
18194 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
18195 |PERL_SCAN_SILENT_NON_PORTABLE;
18196 STRLEN len = remaining;
18197 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
18199 /* If the hex decode routine found something, it should go
18200 * up to the next \n */
18201 if ( *(si_string + len) == '\n') {
18202 if (count) { /* 2nd code point on line */
18203 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
18206 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
18209 goto prepare_for_next_iteration;
18212 /* If the hex decode was instead for the lower range limit,
18213 * save it, and go parse the upper range limit */
18214 if (*(si_string + len) == '\t') {
18215 assert(count == 0);
18219 prepare_for_next_iteration:
18220 si_string += len + 1;
18221 remaining -= len + 1;
18225 /* Here, didn't find a legal hex number. Just add it from
18226 * here to the next \n */
18229 while (*(si_string + len) != '\n' && remaining > 0) {
18233 if (*(si_string + len) == '\n') {
18237 if (matches_string) {
18238 sv_catpvn(matches_string, si_string, len - 1);
18241 matches_string = newSVpvn(si_string, len - 1);
18244 sv_catpvs(matches_string, " ");
18245 } /* end of loop through the text */
18247 assert(matches_string);
18248 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
18249 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
18251 } /* end of has an 'si' but no swash */
18254 /* If we have a swash in place, its equivalent inversion list was above
18255 * placed into 'invlist'. If not, this variable may contain a stored
18256 * inversion list which is information beyond what is in 'si' */
18259 /* Again, if the caller doesn't want the output inversion list, put
18260 * everything in 'matches-string' */
18261 if (! output_invlist) {
18262 if ( ! matches_string) {
18263 matches_string = newSVpvs("\n");
18265 sv_catsv(matches_string, invlist_contents(invlist,
18266 TRUE /* traditional style */
18269 else if (! *output_invlist) {
18270 *output_invlist = invlist_clone(invlist);
18273 _invlist_union(*output_invlist, invlist, output_invlist);
18277 *listsvp = matches_string;
18282 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
18284 /* reg_skipcomment()
18286 Absorbs an /x style # comment from the input stream,
18287 returning a pointer to the first character beyond the comment, or if the
18288 comment terminates the pattern without anything following it, this returns
18289 one past the final character of the pattern (in other words, RExC_end) and
18290 sets the REG_RUN_ON_COMMENT_SEEN flag.
18292 Note it's the callers responsibility to ensure that we are
18293 actually in /x mode
18297 PERL_STATIC_INLINE char*
18298 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
18300 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
18304 while (p < RExC_end) {
18305 if (*(++p) == '\n') {
18310 /* we ran off the end of the pattern without ending the comment, so we have
18311 * to add an \n when wrapping */
18312 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
18317 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
18319 const bool force_to_xmod
18322 /* If the text at the current parse position '*p' is a '(?#...)' comment,
18323 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
18324 * is /x whitespace, advance '*p' so that on exit it points to the first
18325 * byte past all such white space and comments */
18327 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
18329 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
18331 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
18334 if (RExC_end - (*p) >= 3
18336 && *(*p + 1) == '?'
18337 && *(*p + 2) == '#')
18339 while (*(*p) != ')') {
18340 if ((*p) == RExC_end)
18341 FAIL("Sequence (?#... not terminated");
18349 const char * save_p = *p;
18350 while ((*p) < RExC_end) {
18352 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
18355 else if (*(*p) == '#') {
18356 (*p) = reg_skipcomment(pRExC_state, (*p));
18362 if (*p != save_p) {
18375 Advances the parse position by one byte, unless that byte is the beginning
18376 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
18377 those two cases, the parse position is advanced beyond all such comments and
18380 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
18384 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
18386 PERL_ARGS_ASSERT_NEXTCHAR;
18388 if (RExC_parse < RExC_end) {
18390 || UTF8_IS_INVARIANT(*RExC_parse)
18391 || UTF8_IS_START(*RExC_parse));
18393 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
18395 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
18396 FALSE /* Don't force /x */ );
18401 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
18403 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
18404 * space. In pass1, it aligns and increments RExC_size; in pass2,
18407 regnode * const ret = RExC_emit;
18408 GET_RE_DEBUG_FLAGS_DECL;
18410 PERL_ARGS_ASSERT_REGNODE_GUTS;
18412 assert(extra_size >= regarglen[op]);
18415 SIZE_ALIGN(RExC_size);
18416 RExC_size += 1 + extra_size;
18419 if (RExC_emit >= RExC_emit_bound)
18420 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
18421 op, (void*)RExC_emit, (void*)RExC_emit_bound);
18423 NODE_ALIGN_FILL(ret);
18424 #ifndef RE_TRACK_PATTERN_OFFSETS
18425 PERL_UNUSED_ARG(name);
18427 if (RExC_offsets) { /* MJD */
18429 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
18432 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
18433 ? "Overwriting end of array!\n" : "OK",
18434 (UV)(RExC_emit - RExC_emit_start),
18435 (UV)(RExC_parse - RExC_start),
18436 (UV)RExC_offsets[0]));
18437 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
18444 - reg_node - emit a node
18446 STATIC regnode * /* Location. */
18447 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
18449 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
18451 PERL_ARGS_ASSERT_REG_NODE;
18453 assert(regarglen[op] == 0);
18456 regnode *ptr = ret;
18457 FILL_ADVANCE_NODE(ptr, op);
18464 - reganode - emit a node with an argument
18466 STATIC regnode * /* Location. */
18467 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
18469 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
18471 PERL_ARGS_ASSERT_REGANODE;
18473 assert(regarglen[op] == 1);
18476 regnode *ptr = ret;
18477 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
18484 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
18486 /* emit a node with U32 and I32 arguments */
18488 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
18490 PERL_ARGS_ASSERT_REG2LANODE;
18492 assert(regarglen[op] == 2);
18495 regnode *ptr = ret;
18496 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
18503 - reginsert - insert an operator in front of already-emitted operand
18505 * Means relocating the operand.
18507 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
18508 * set up NEXT_OFF() of the inserted node if needed. Something like this:
18510 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
18512 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
18516 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *operand, U32 depth)
18521 const int offset = regarglen[(U8)op];
18522 const int size = NODE_STEP_REGNODE + offset;
18523 GET_RE_DEBUG_FLAGS_DECL;
18525 PERL_ARGS_ASSERT_REGINSERT;
18526 PERL_UNUSED_CONTEXT;
18527 PERL_UNUSED_ARG(depth);
18528 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
18529 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
18534 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
18535 studying. If this is wrong then we need to adjust RExC_recurse
18536 below like we do with RExC_open_parens/RExC_close_parens. */
18540 if (RExC_open_parens) {
18542 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
18543 /* remember that RExC_npar is rex->nparens + 1,
18544 * iow it is 1 more than the number of parens seen in
18545 * the pattern so far. */
18546 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
18547 /* note, RExC_open_parens[0] is the start of the
18548 * regex, it can't move. RExC_close_parens[0] is the end
18549 * of the regex, it *can* move. */
18550 if ( paren && RExC_open_parens[paren] >= operand ) {
18551 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
18552 RExC_open_parens[paren] += size;
18554 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
18556 if ( RExC_close_parens[paren] >= operand ) {
18557 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
18558 RExC_close_parens[paren] += size;
18560 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
18565 RExC_end_op += size;
18567 while (src > operand) {
18568 StructCopy(--src, --dst, regnode);
18569 #ifdef RE_TRACK_PATTERN_OFFSETS
18570 if (RExC_offsets) { /* MJD 20010112 */
18572 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
18576 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
18577 ? "Overwriting end of array!\n" : "OK",
18578 (UV)(src - RExC_emit_start),
18579 (UV)(dst - RExC_emit_start),
18580 (UV)RExC_offsets[0]));
18581 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
18582 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
18588 place = operand; /* Op node, where operand used to be. */
18589 #ifdef RE_TRACK_PATTERN_OFFSETS
18590 if (RExC_offsets) { /* MJD */
18592 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
18596 (UV)(place - RExC_emit_start) > RExC_offsets[0]
18597 ? "Overwriting end of array!\n" : "OK",
18598 (UV)(place - RExC_emit_start),
18599 (UV)(RExC_parse - RExC_start),
18600 (UV)RExC_offsets[0]));
18601 Set_Node_Offset(place, RExC_parse);
18602 Set_Node_Length(place, 1);
18605 src = NEXTOPER(place);
18606 FILL_ADVANCE_NODE(place, op);
18607 Zero(src, offset, regnode);
18611 - regtail - set the next-pointer at the end of a node chain of p to val.
18612 - SEE ALSO: regtail_study
18615 S_regtail(pTHX_ RExC_state_t * pRExC_state,
18616 const regnode * const p,
18617 const regnode * const val,
18621 GET_RE_DEBUG_FLAGS_DECL;
18623 PERL_ARGS_ASSERT_REGTAIL;
18625 PERL_UNUSED_ARG(depth);
18631 /* Find last node. */
18632 scan = (regnode *) p;
18634 regnode * const temp = regnext(scan);
18636 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
18637 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18638 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
18639 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
18640 (temp == NULL ? "->" : ""),
18641 (temp == NULL ? PL_reg_name[OP(val)] : "")
18649 if (reg_off_by_arg[OP(scan)]) {
18650 ARG_SET(scan, val - scan);
18653 NEXT_OFF(scan) = val - scan;
18659 - regtail_study - set the next-pointer at the end of a node chain of p to val.
18660 - Look for optimizable sequences at the same time.
18661 - currently only looks for EXACT chains.
18663 This is experimental code. The idea is to use this routine to perform
18664 in place optimizations on branches and groups as they are constructed,
18665 with the long term intention of removing optimization from study_chunk so
18666 that it is purely analytical.
18668 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
18669 to control which is which.
18672 /* TODO: All four parms should be const */
18675 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
18676 const regnode *val,U32 depth)
18680 #ifdef EXPERIMENTAL_INPLACESCAN
18683 GET_RE_DEBUG_FLAGS_DECL;
18685 PERL_ARGS_ASSERT_REGTAIL_STUDY;
18691 /* Find last node. */
18695 regnode * const temp = regnext(scan);
18696 #ifdef EXPERIMENTAL_INPLACESCAN
18697 if (PL_regkind[OP(scan)] == EXACT) {
18698 bool unfolded_multi_char; /* Unexamined in this routine */
18699 if (join_exact(pRExC_state, scan, &min,
18700 &unfolded_multi_char, 1, val, depth+1))
18705 switch (OP(scan)) {
18709 case EXACTFA_NO_TRIE:
18715 if( exact == PSEUDO )
18717 else if ( exact != OP(scan) )
18726 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
18727 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18728 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
18729 SvPV_nolen_const(RExC_mysv),
18730 REG_NODE_NUM(scan),
18731 PL_reg_name[exact]);
18738 DEBUG_PARSE_MSG("");
18739 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
18740 Perl_re_printf( aTHX_
18741 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
18742 SvPV_nolen_const(RExC_mysv),
18743 (IV)REG_NODE_NUM(val),
18747 if (reg_off_by_arg[OP(scan)]) {
18748 ARG_SET(scan, val - scan);
18751 NEXT_OFF(scan) = val - scan;
18759 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
18764 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
18769 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18771 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
18772 if (flags & (1<<bit)) {
18773 if (!set++ && lead)
18774 Perl_re_printf( aTHX_ "%s",lead);
18775 Perl_re_printf( aTHX_ "%s ",PL_reg_intflags_name[bit]);
18780 Perl_re_printf( aTHX_ "\n");
18782 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18787 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
18793 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18795 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
18796 if (flags & (1<<bit)) {
18797 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
18800 if (!set++ && lead)
18801 Perl_re_printf( aTHX_ "%s",lead);
18802 Perl_re_printf( aTHX_ "%s ",PL_reg_extflags_name[bit]);
18805 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
18806 if (!set++ && lead) {
18807 Perl_re_printf( aTHX_ "%s",lead);
18810 case REGEX_UNICODE_CHARSET:
18811 Perl_re_printf( aTHX_ "UNICODE");
18813 case REGEX_LOCALE_CHARSET:
18814 Perl_re_printf( aTHX_ "LOCALE");
18816 case REGEX_ASCII_RESTRICTED_CHARSET:
18817 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
18819 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
18820 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
18823 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
18829 Perl_re_printf( aTHX_ "\n");
18831 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18837 Perl_regdump(pTHX_ const regexp *r)
18840 SV * const sv = sv_newmortal();
18841 SV *dsv= sv_newmortal();
18842 RXi_GET_DECL(r,ri);
18843 GET_RE_DEBUG_FLAGS_DECL;
18845 PERL_ARGS_ASSERT_REGDUMP;
18847 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
18849 /* Header fields of interest. */
18850 if (r->anchored_substr) {
18851 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
18852 RE_SV_DUMPLEN(r->anchored_substr), 30);
18853 Perl_re_printf( aTHX_
18854 "anchored %s%s at %" IVdf " ",
18855 s, RE_SV_TAIL(r->anchored_substr),
18856 (IV)r->anchored_offset);
18857 } else if (r->anchored_utf8) {
18858 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
18859 RE_SV_DUMPLEN(r->anchored_utf8), 30);
18860 Perl_re_printf( aTHX_
18861 "anchored utf8 %s%s at %" IVdf " ",
18862 s, RE_SV_TAIL(r->anchored_utf8),
18863 (IV)r->anchored_offset);
18865 if (r->float_substr) {
18866 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
18867 RE_SV_DUMPLEN(r->float_substr), 30);
18868 Perl_re_printf( aTHX_
18869 "floating %s%s at %" IVdf "..%" UVuf " ",
18870 s, RE_SV_TAIL(r->float_substr),
18871 (IV)r->float_min_offset, (UV)r->float_max_offset);
18872 } else if (r->float_utf8) {
18873 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
18874 RE_SV_DUMPLEN(r->float_utf8), 30);
18875 Perl_re_printf( aTHX_
18876 "floating utf8 %s%s at %" IVdf "..%" UVuf " ",
18877 s, RE_SV_TAIL(r->float_utf8),
18878 (IV)r->float_min_offset, (UV)r->float_max_offset);
18880 if (r->check_substr || r->check_utf8)
18881 Perl_re_printf( aTHX_
18883 (r->check_substr == r->float_substr
18884 && r->check_utf8 == r->float_utf8
18885 ? "(checking floating" : "(checking anchored"));
18886 if (r->intflags & PREGf_NOSCAN)
18887 Perl_re_printf( aTHX_ " noscan");
18888 if (r->extflags & RXf_CHECK_ALL)
18889 Perl_re_printf( aTHX_ " isall");
18890 if (r->check_substr || r->check_utf8)
18891 Perl_re_printf( aTHX_ ") ");
18893 if (ri->regstclass) {
18894 regprop(r, sv, ri->regstclass, NULL, NULL);
18895 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
18897 if (r->intflags & PREGf_ANCH) {
18898 Perl_re_printf( aTHX_ "anchored");
18899 if (r->intflags & PREGf_ANCH_MBOL)
18900 Perl_re_printf( aTHX_ "(MBOL)");
18901 if (r->intflags & PREGf_ANCH_SBOL)
18902 Perl_re_printf( aTHX_ "(SBOL)");
18903 if (r->intflags & PREGf_ANCH_GPOS)
18904 Perl_re_printf( aTHX_ "(GPOS)");
18905 Perl_re_printf( aTHX_ " ");
18907 if (r->intflags & PREGf_GPOS_SEEN)
18908 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
18909 if (r->intflags & PREGf_SKIP)
18910 Perl_re_printf( aTHX_ "plus ");
18911 if (r->intflags & PREGf_IMPLICIT)
18912 Perl_re_printf( aTHX_ "implicit ");
18913 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
18914 if (r->extflags & RXf_EVAL_SEEN)
18915 Perl_re_printf( aTHX_ "with eval ");
18916 Perl_re_printf( aTHX_ "\n");
18918 regdump_extflags("r->extflags: ",r->extflags);
18919 regdump_intflags("r->intflags: ",r->intflags);
18922 PERL_ARGS_ASSERT_REGDUMP;
18923 PERL_UNUSED_CONTEXT;
18924 PERL_UNUSED_ARG(r);
18925 #endif /* DEBUGGING */
18928 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
18931 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
18932 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
18933 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
18934 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
18935 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
18936 || _CC_VERTSPACE != 15
18937 # error Need to adjust order of anyofs[]
18939 static const char * const anyofs[] = {
18976 - regprop - printable representation of opcode, with run time support
18980 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
18984 RXi_GET_DECL(prog,progi);
18985 GET_RE_DEBUG_FLAGS_DECL;
18987 PERL_ARGS_ASSERT_REGPROP;
18991 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
18992 /* It would be nice to FAIL() here, but this may be called from
18993 regexec.c, and it would be hard to supply pRExC_state. */
18994 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
18995 (int)OP(o), (int)REGNODE_MAX);
18996 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
18998 k = PL_regkind[OP(o)];
19001 sv_catpvs(sv, " ");
19002 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
19003 * is a crude hack but it may be the best for now since
19004 * we have no flag "this EXACTish node was UTF-8"
19006 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
19007 PERL_PV_ESCAPE_UNI_DETECT |
19008 PERL_PV_ESCAPE_NONASCII |
19009 PERL_PV_PRETTY_ELLIPSES |
19010 PERL_PV_PRETTY_LTGT |
19011 PERL_PV_PRETTY_NOCLEAR
19013 } else if (k == TRIE) {
19014 /* print the details of the trie in dumpuntil instead, as
19015 * progi->data isn't available here */
19016 const char op = OP(o);
19017 const U32 n = ARG(o);
19018 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
19019 (reg_ac_data *)progi->data->data[n] :
19021 const reg_trie_data * const trie
19022 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
19024 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
19025 DEBUG_TRIE_COMPILE_r({
19027 sv_catpvs(sv, "(JUMP)");
19028 Perl_sv_catpvf(aTHX_ sv,
19029 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
19030 (UV)trie->startstate,
19031 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
19032 (UV)trie->wordcount,
19035 (UV)TRIE_CHARCOUNT(trie),
19036 (UV)trie->uniquecharcount
19039 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
19040 sv_catpvs(sv, "[");
19041 (void) put_charclass_bitmap_innards(sv,
19042 ((IS_ANYOF_TRIE(op))
19044 : TRIE_BITMAP(trie)),
19050 sv_catpvs(sv, "]");
19052 } else if (k == CURLY) {
19053 U32 lo = ARG1(o), hi = ARG2(o);
19054 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
19055 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
19056 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
19057 if (hi == REG_INFTY)
19058 sv_catpvs(sv, "INFTY");
19060 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
19061 sv_catpvs(sv, "}");
19063 else if (k == WHILEM && o->flags) /* Ordinal/of */
19064 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
19065 else if (k == REF || k == OPEN || k == CLOSE
19066 || k == GROUPP || OP(o)==ACCEPT)
19068 AV *name_list= NULL;
19069 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
19070 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
19071 if ( RXp_PAREN_NAMES(prog) ) {
19072 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19073 } else if ( pRExC_state ) {
19074 name_list= RExC_paren_name_list;
19077 if ( k != REF || (OP(o) < NREF)) {
19078 SV **name= av_fetch(name_list, parno, 0 );
19080 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19083 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
19084 I32 *nums=(I32*)SvPVX(sv_dat);
19085 SV **name= av_fetch(name_list, nums[0], 0 );
19088 for ( n=0; n<SvIVX(sv_dat); n++ ) {
19089 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
19090 (n ? "," : ""), (IV)nums[n]);
19092 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19096 if ( k == REF && reginfo) {
19097 U32 n = ARG(o); /* which paren pair */
19098 I32 ln = prog->offs[n].start;
19099 if (prog->lastparen < n || ln == -1)
19100 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
19101 else if (ln == prog->offs[n].end)
19102 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
19104 const char *s = reginfo->strbeg + ln;
19105 Perl_sv_catpvf(aTHX_ sv, ": ");
19106 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
19107 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
19110 } else if (k == GOSUB) {
19111 AV *name_list= NULL;
19112 if ( RXp_PAREN_NAMES(prog) ) {
19113 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19114 } else if ( pRExC_state ) {
19115 name_list= RExC_paren_name_list;
19118 /* Paren and offset */
19119 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
19120 (int)((o + (int)ARG2L(o)) - progi->program) );
19122 SV **name= av_fetch(name_list, ARG(o), 0 );
19124 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19127 else if (k == LOGICAL)
19128 /* 2: embedded, otherwise 1 */
19129 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
19130 else if (k == ANYOF) {
19131 const U8 flags = ANYOF_FLAGS(o);
19132 bool do_sep = FALSE; /* Do we need to separate various components of
19134 /* Set if there is still an unresolved user-defined property */
19135 SV *unresolved = NULL;
19137 /* Things that are ignored except when the runtime locale is UTF-8 */
19138 SV *only_utf8_locale_invlist = NULL;
19140 /* Code points that don't fit in the bitmap */
19141 SV *nonbitmap_invlist = NULL;
19143 /* And things that aren't in the bitmap, but are small enough to be */
19144 SV* bitmap_range_not_in_bitmap = NULL;
19146 const bool inverted = flags & ANYOF_INVERT;
19148 if (OP(o) == ANYOFL) {
19149 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
19150 sv_catpvs(sv, "{utf8-locale-reqd}");
19152 if (flags & ANYOFL_FOLD) {
19153 sv_catpvs(sv, "{i}");
19157 /* If there is stuff outside the bitmap, get it */
19158 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
19159 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
19161 &only_utf8_locale_invlist,
19162 &nonbitmap_invlist);
19163 /* The non-bitmap data may contain stuff that could fit in the
19164 * bitmap. This could come from a user-defined property being
19165 * finally resolved when this call was done; or much more likely
19166 * because there are matches that require UTF-8 to be valid, and so
19167 * aren't in the bitmap. This is teased apart later */
19168 _invlist_intersection(nonbitmap_invlist,
19170 &bitmap_range_not_in_bitmap);
19171 /* Leave just the things that don't fit into the bitmap */
19172 _invlist_subtract(nonbitmap_invlist,
19174 &nonbitmap_invlist);
19177 /* Obey this flag to add all above-the-bitmap code points */
19178 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
19179 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
19180 NUM_ANYOF_CODE_POINTS,
19184 /* Ready to start outputting. First, the initial left bracket */
19185 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
19187 /* Then all the things that could fit in the bitmap */
19188 do_sep = put_charclass_bitmap_innards(sv,
19190 bitmap_range_not_in_bitmap,
19191 only_utf8_locale_invlist,
19194 /* Can't try inverting for a
19195 * better display if there are
19196 * things that haven't been
19198 unresolved != NULL);
19199 SvREFCNT_dec(bitmap_range_not_in_bitmap);
19201 /* If there are user-defined properties which haven't been defined yet,
19202 * output them. If the result is not to be inverted, it is clearest to
19203 * output them in a separate [] from the bitmap range stuff. If the
19204 * result is to be complemented, we have to show everything in one [],
19205 * as the inversion applies to the whole thing. Use {braces} to
19206 * separate them from anything in the bitmap and anything above the
19210 if (! do_sep) { /* If didn't output anything in the bitmap */
19211 sv_catpvs(sv, "^");
19213 sv_catpvs(sv, "{");
19216 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19218 sv_catsv(sv, unresolved);
19220 sv_catpvs(sv, "}");
19222 do_sep = ! inverted;
19225 /* And, finally, add the above-the-bitmap stuff */
19226 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
19229 /* See if truncation size is overridden */
19230 const STRLEN dump_len = (PL_dump_re_max_len)
19231 ? PL_dump_re_max_len
19234 /* This is output in a separate [] */
19236 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19239 /* And, for easy of understanding, it is shown in the
19240 * uncomplemented form if possible. The one exception being if
19241 * there are unresolved items, where the inversion has to be
19242 * delayed until runtime */
19243 if (inverted && ! unresolved) {
19244 _invlist_invert(nonbitmap_invlist);
19245 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
19248 contents = invlist_contents(nonbitmap_invlist,
19249 FALSE /* output suitable for catsv */
19252 /* If the output is shorter than the permissible maximum, just do it. */
19253 if (SvCUR(contents) <= dump_len) {
19254 sv_catsv(sv, contents);
19257 const char * contents_string = SvPVX(contents);
19258 STRLEN i = dump_len;
19260 /* Otherwise, start at the permissible max and work back to the
19261 * first break possibility */
19262 while (i > 0 && contents_string[i] != ' ') {
19265 if (i == 0) { /* Fail-safe. Use the max if we couldn't
19266 find a legal break */
19270 sv_catpvn(sv, contents_string, i);
19271 sv_catpvs(sv, "...");
19274 SvREFCNT_dec_NN(contents);
19275 SvREFCNT_dec_NN(nonbitmap_invlist);
19278 /* And finally the matching, closing ']' */
19279 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
19281 SvREFCNT_dec(unresolved);
19283 else if (k == POSIXD || k == NPOSIXD) {
19284 U8 index = FLAGS(o) * 2;
19285 if (index < C_ARRAY_LENGTH(anyofs)) {
19286 if (*anyofs[index] != '[') {
19289 sv_catpv(sv, anyofs[index]);
19290 if (*anyofs[index] != '[') {
19295 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
19298 else if (k == BOUND || k == NBOUND) {
19299 /* Must be synced with order of 'bound_type' in regcomp.h */
19300 const char * const bounds[] = {
19301 "", /* Traditional */
19307 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
19308 sv_catpv(sv, bounds[FLAGS(o)]);
19310 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
19311 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
19312 else if (OP(o) == SBOL)
19313 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
19315 /* add on the verb argument if there is one */
19316 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
19317 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
19318 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
19321 PERL_UNUSED_CONTEXT;
19322 PERL_UNUSED_ARG(sv);
19323 PERL_UNUSED_ARG(o);
19324 PERL_UNUSED_ARG(prog);
19325 PERL_UNUSED_ARG(reginfo);
19326 PERL_UNUSED_ARG(pRExC_state);
19327 #endif /* DEBUGGING */
19333 Perl_re_intuit_string(pTHX_ REGEXP * const r)
19334 { /* Assume that RE_INTUIT is set */
19335 struct regexp *const prog = ReANY(r);
19336 GET_RE_DEBUG_FLAGS_DECL;
19338 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
19339 PERL_UNUSED_CONTEXT;
19343 const char * const s = SvPV_nolen_const(RX_UTF8(r)
19344 ? prog->check_utf8 : prog->check_substr);
19346 if (!PL_colorset) reginitcolors();
19347 Perl_re_printf( aTHX_
19348 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
19350 RX_UTF8(r) ? "utf8 " : "",
19351 PL_colors[5],PL_colors[0],
19354 (strlen(s) > 60 ? "..." : ""));
19357 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
19358 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
19364 handles refcounting and freeing the perl core regexp structure. When
19365 it is necessary to actually free the structure the first thing it
19366 does is call the 'free' method of the regexp_engine associated to
19367 the regexp, allowing the handling of the void *pprivate; member
19368 first. (This routine is not overridable by extensions, which is why
19369 the extensions free is called first.)
19371 See regdupe and regdupe_internal if you change anything here.
19373 #ifndef PERL_IN_XSUB_RE
19375 Perl_pregfree(pTHX_ REGEXP *r)
19381 Perl_pregfree2(pTHX_ REGEXP *rx)
19383 struct regexp *const r = ReANY(rx);
19384 GET_RE_DEBUG_FLAGS_DECL;
19386 PERL_ARGS_ASSERT_PREGFREE2;
19388 if (r->mother_re) {
19389 ReREFCNT_dec(r->mother_re);
19391 CALLREGFREE_PVT(rx); /* free the private data */
19392 SvREFCNT_dec(RXp_PAREN_NAMES(r));
19393 Safefree(r->xpv_len_u.xpvlenu_pv);
19396 SvREFCNT_dec(r->anchored_substr);
19397 SvREFCNT_dec(r->anchored_utf8);
19398 SvREFCNT_dec(r->float_substr);
19399 SvREFCNT_dec(r->float_utf8);
19400 Safefree(r->substrs);
19402 RX_MATCH_COPY_FREE(rx);
19403 #ifdef PERL_ANY_COW
19404 SvREFCNT_dec(r->saved_copy);
19407 SvREFCNT_dec(r->qr_anoncv);
19408 if (r->recurse_locinput)
19409 Safefree(r->recurse_locinput);
19410 rx->sv_u.svu_rx = 0;
19415 This is a hacky workaround to the structural issue of match results
19416 being stored in the regexp structure which is in turn stored in
19417 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
19418 could be PL_curpm in multiple contexts, and could require multiple
19419 result sets being associated with the pattern simultaneously, such
19420 as when doing a recursive match with (??{$qr})
19422 The solution is to make a lightweight copy of the regexp structure
19423 when a qr// is returned from the code executed by (??{$qr}) this
19424 lightweight copy doesn't actually own any of its data except for
19425 the starp/end and the actual regexp structure itself.
19431 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
19433 struct regexp *ret;
19434 struct regexp *const r = ReANY(rx);
19435 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
19437 PERL_ARGS_ASSERT_REG_TEMP_COPY;
19440 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
19442 SvOK_off((SV *)ret_x);
19444 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
19445 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
19446 made both spots point to the same regexp body.) */
19447 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
19448 assert(!SvPVX(ret_x));
19449 ret_x->sv_u.svu_rx = temp->sv_any;
19450 temp->sv_any = NULL;
19451 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
19452 SvREFCNT_dec_NN(temp);
19453 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
19454 ing below will not set it. */
19455 SvCUR_set(ret_x, SvCUR(rx));
19458 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
19459 sv_force_normal(sv) is called. */
19461 ret = ReANY(ret_x);
19463 SvFLAGS(ret_x) |= SvUTF8(rx);
19464 /* We share the same string buffer as the original regexp, on which we
19465 hold a reference count, incremented when mother_re is set below.
19466 The string pointer is copied here, being part of the regexp struct.
19468 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
19469 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
19471 const I32 npar = r->nparens+1;
19472 Newx(ret->offs, npar, regexp_paren_pair);
19473 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19476 Newx(ret->substrs, 1, struct reg_substr_data);
19477 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19479 SvREFCNT_inc_void(ret->anchored_substr);
19480 SvREFCNT_inc_void(ret->anchored_utf8);
19481 SvREFCNT_inc_void(ret->float_substr);
19482 SvREFCNT_inc_void(ret->float_utf8);
19484 /* check_substr and check_utf8, if non-NULL, point to either their
19485 anchored or float namesakes, and don't hold a second reference. */
19487 RX_MATCH_COPIED_off(ret_x);
19488 #ifdef PERL_ANY_COW
19489 ret->saved_copy = NULL;
19491 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
19492 SvREFCNT_inc_void(ret->qr_anoncv);
19493 if (r->recurse_locinput)
19494 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19500 /* regfree_internal()
19502 Free the private data in a regexp. This is overloadable by
19503 extensions. Perl takes care of the regexp structure in pregfree(),
19504 this covers the *pprivate pointer which technically perl doesn't
19505 know about, however of course we have to handle the
19506 regexp_internal structure when no extension is in use.
19508 Note this is called before freeing anything in the regexp
19513 Perl_regfree_internal(pTHX_ REGEXP * const rx)
19515 struct regexp *const r = ReANY(rx);
19516 RXi_GET_DECL(r,ri);
19517 GET_RE_DEBUG_FLAGS_DECL;
19519 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
19525 SV *dsv= sv_newmortal();
19526 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
19527 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
19528 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
19529 PL_colors[4],PL_colors[5],s);
19532 #ifdef RE_TRACK_PATTERN_OFFSETS
19534 Safefree(ri->u.offsets); /* 20010421 MJD */
19536 if (ri->code_blocks) {
19537 ri->code_blocks->attached = FALSE;
19538 S_free_codeblocks(aTHX_ ri->code_blocks);
19542 int n = ri->data->count;
19545 /* If you add a ->what type here, update the comment in regcomp.h */
19546 switch (ri->data->what[n]) {
19552 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
19555 Safefree(ri->data->data[n]);
19561 { /* Aho Corasick add-on structure for a trie node.
19562 Used in stclass optimization only */
19564 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
19565 #ifdef USE_ITHREADS
19569 refcount = --aho->refcount;
19572 PerlMemShared_free(aho->states);
19573 PerlMemShared_free(aho->fail);
19574 /* do this last!!!! */
19575 PerlMemShared_free(ri->data->data[n]);
19576 /* we should only ever get called once, so
19577 * assert as much, and also guard the free
19578 * which /might/ happen twice. At the least
19579 * it will make code anlyzers happy and it
19580 * doesn't cost much. - Yves */
19581 assert(ri->regstclass);
19582 if (ri->regstclass) {
19583 PerlMemShared_free(ri->regstclass);
19584 ri->regstclass = 0;
19591 /* trie structure. */
19593 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
19594 #ifdef USE_ITHREADS
19598 refcount = --trie->refcount;
19601 PerlMemShared_free(trie->charmap);
19602 PerlMemShared_free(trie->states);
19603 PerlMemShared_free(trie->trans);
19605 PerlMemShared_free(trie->bitmap);
19607 PerlMemShared_free(trie->jump);
19608 PerlMemShared_free(trie->wordinfo);
19609 /* do this last!!!! */
19610 PerlMemShared_free(ri->data->data[n]);
19615 Perl_croak(aTHX_ "panic: regfree data code '%c'",
19616 ri->data->what[n]);
19619 Safefree(ri->data->what);
19620 Safefree(ri->data);
19626 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
19627 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
19628 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
19631 re_dup_guts - duplicate a regexp.
19633 This routine is expected to clone a given regexp structure. It is only
19634 compiled under USE_ITHREADS.
19636 After all of the core data stored in struct regexp is duplicated
19637 the regexp_engine.dupe method is used to copy any private data
19638 stored in the *pprivate pointer. This allows extensions to handle
19639 any duplication it needs to do.
19641 See pregfree() and regfree_internal() if you change anything here.
19643 #if defined(USE_ITHREADS)
19644 #ifndef PERL_IN_XSUB_RE
19646 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
19650 const struct regexp *r = ReANY(sstr);
19651 struct regexp *ret = ReANY(dstr);
19653 PERL_ARGS_ASSERT_RE_DUP_GUTS;
19655 npar = r->nparens+1;
19656 Newx(ret->offs, npar, regexp_paren_pair);
19657 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19659 if (ret->substrs) {
19660 /* Do it this way to avoid reading from *r after the StructCopy().
19661 That way, if any of the sv_dup_inc()s dislodge *r from the L1
19662 cache, it doesn't matter. */
19663 const bool anchored = r->check_substr
19664 ? r->check_substr == r->anchored_substr
19665 : r->check_utf8 == r->anchored_utf8;
19666 Newx(ret->substrs, 1, struct reg_substr_data);
19667 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19669 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
19670 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
19671 ret->float_substr = sv_dup_inc(ret->float_substr, param);
19672 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
19674 /* check_substr and check_utf8, if non-NULL, point to either their
19675 anchored or float namesakes, and don't hold a second reference. */
19677 if (ret->check_substr) {
19679 assert(r->check_utf8 == r->anchored_utf8);
19680 ret->check_substr = ret->anchored_substr;
19681 ret->check_utf8 = ret->anchored_utf8;
19683 assert(r->check_substr == r->float_substr);
19684 assert(r->check_utf8 == r->float_utf8);
19685 ret->check_substr = ret->float_substr;
19686 ret->check_utf8 = ret->float_utf8;
19688 } else if (ret->check_utf8) {
19690 ret->check_utf8 = ret->anchored_utf8;
19692 ret->check_utf8 = ret->float_utf8;
19697 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
19698 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
19699 if (r->recurse_locinput)
19700 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19703 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
19705 if (RX_MATCH_COPIED(dstr))
19706 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
19708 ret->subbeg = NULL;
19709 #ifdef PERL_ANY_COW
19710 ret->saved_copy = NULL;
19713 /* Whether mother_re be set or no, we need to copy the string. We
19714 cannot refrain from copying it when the storage points directly to
19715 our mother regexp, because that's
19716 1: a buffer in a different thread
19717 2: something we no longer hold a reference on
19718 so we need to copy it locally. */
19719 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
19720 ret->mother_re = NULL;
19722 #endif /* PERL_IN_XSUB_RE */
19727 This is the internal complement to regdupe() which is used to copy
19728 the structure pointed to by the *pprivate pointer in the regexp.
19729 This is the core version of the extension overridable cloning hook.
19730 The regexp structure being duplicated will be copied by perl prior
19731 to this and will be provided as the regexp *r argument, however
19732 with the /old/ structures pprivate pointer value. Thus this routine
19733 may override any copying normally done by perl.
19735 It returns a pointer to the new regexp_internal structure.
19739 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
19742 struct regexp *const r = ReANY(rx);
19743 regexp_internal *reti;
19745 RXi_GET_DECL(r,ri);
19747 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
19751 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
19752 char, regexp_internal);
19753 Copy(ri->program, reti->program, len+1, regnode);
19756 if (ri->code_blocks) {
19758 Newx(reti->code_blocks, 1, struct reg_code_blocks);
19759 Newx(reti->code_blocks->cb, ri->code_blocks->count,
19760 struct reg_code_block);
19761 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
19762 ri->code_blocks->count, struct reg_code_block);
19763 for (n = 0; n < ri->code_blocks->count; n++)
19764 reti->code_blocks->cb[n].src_regex = (REGEXP*)
19765 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
19766 reti->code_blocks->count = ri->code_blocks->count;
19767 reti->code_blocks->attached = TRUE;
19770 reti->code_blocks = NULL;
19772 reti->regstclass = NULL;
19775 struct reg_data *d;
19776 const int count = ri->data->count;
19779 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
19780 char, struct reg_data);
19781 Newx(d->what, count, U8);
19784 for (i = 0; i < count; i++) {
19785 d->what[i] = ri->data->what[i];
19786 switch (d->what[i]) {
19787 /* see also regcomp.h and regfree_internal() */
19788 case 'a': /* actually an AV, but the dup function is identical. */
19792 case 'u': /* actually an HV, but the dup function is identical. */
19793 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
19796 /* This is cheating. */
19797 Newx(d->data[i], 1, regnode_ssc);
19798 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
19799 reti->regstclass = (regnode*)d->data[i];
19802 /* Trie stclasses are readonly and can thus be shared
19803 * without duplication. We free the stclass in pregfree
19804 * when the corresponding reg_ac_data struct is freed.
19806 reti->regstclass= ri->regstclass;
19810 ((reg_trie_data*)ri->data->data[i])->refcount++;
19815 d->data[i] = ri->data->data[i];
19818 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
19819 ri->data->what[i]);
19828 reti->name_list_idx = ri->name_list_idx;
19830 #ifdef RE_TRACK_PATTERN_OFFSETS
19831 if (ri->u.offsets) {
19832 Newx(reti->u.offsets, 2*len+1, U32);
19833 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
19836 SetProgLen(reti,len);
19839 return (void*)reti;
19842 #endif /* USE_ITHREADS */
19844 #ifndef PERL_IN_XSUB_RE
19847 - regnext - dig the "next" pointer out of a node
19850 Perl_regnext(pTHX_ regnode *p)
19857 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
19858 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
19859 (int)OP(p), (int)REGNODE_MAX);
19862 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
19871 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
19874 STRLEN l1 = strlen(pat1);
19875 STRLEN l2 = strlen(pat2);
19878 const char *message;
19880 PERL_ARGS_ASSERT_RE_CROAK2;
19886 Copy(pat1, buf, l1 , char);
19887 Copy(pat2, buf + l1, l2 , char);
19888 buf[l1 + l2] = '\n';
19889 buf[l1 + l2 + 1] = '\0';
19890 va_start(args, pat2);
19891 msv = vmess(buf, &args);
19893 message = SvPV_const(msv,l1);
19896 Copy(message, buf, l1 , char);
19897 /* l1-1 to avoid \n */
19898 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, l1-1, buf));
19901 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
19903 #ifndef PERL_IN_XSUB_RE
19905 Perl_save_re_context(pTHX)
19910 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
19913 const REGEXP * const rx = PM_GETRE(PL_curpm);
19915 nparens = RX_NPARENS(rx);
19918 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
19919 * that PL_curpm will be null, but that utf8.pm and the modules it
19920 * loads will only use $1..$3.
19921 * The t/porting/re_context.t test file checks this assumption.
19926 for (i = 1; i <= nparens; i++) {
19927 char digits[TYPE_CHARS(long)];
19928 const STRLEN len = my_snprintf(digits, sizeof(digits),
19930 GV *const *const gvp
19931 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
19934 GV * const gv = *gvp;
19935 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
19945 S_put_code_point(pTHX_ SV *sv, UV c)
19947 PERL_ARGS_ASSERT_PUT_CODE_POINT;
19950 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
19952 else if (isPRINT(c)) {
19953 const char string = (char) c;
19955 /* We use {phrase} as metanotation in the class, so also escape literal
19957 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
19958 sv_catpvs(sv, "\\");
19959 sv_catpvn(sv, &string, 1);
19961 else if (isMNEMONIC_CNTRL(c)) {
19962 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
19965 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
19969 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
19972 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
19974 /* Appends to 'sv' a displayable version of the range of code points from
19975 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
19976 * that have them, when they occur at the beginning or end of the range.
19977 * It uses hex to output the remaining code points, unless 'allow_literals'
19978 * is true, in which case the printable ASCII ones are output as-is (though
19979 * some of these will be escaped by put_code_point()).
19981 * NOTE: This is designed only for printing ranges of code points that fit
19982 * inside an ANYOF bitmap. Higher code points are simply suppressed
19985 const unsigned int min_range_count = 3;
19987 assert(start <= end);
19989 PERL_ARGS_ASSERT_PUT_RANGE;
19991 while (start <= end) {
19993 const char * format;
19995 if (end - start < min_range_count) {
19997 /* Output chars individually when they occur in short ranges */
19998 for (; start <= end; start++) {
19999 put_code_point(sv, start);
20004 /* If permitted by the input options, and there is a possibility that
20005 * this range contains a printable literal, look to see if there is
20007 if (allow_literals && start <= MAX_PRINT_A) {
20009 /* If the character at the beginning of the range isn't an ASCII
20010 * printable, effectively split the range into two parts:
20011 * 1) the portion before the first such printable,
20013 * and output them separately. */
20014 if (! isPRINT_A(start)) {
20015 UV temp_end = start + 1;
20017 /* There is no point looking beyond the final possible
20018 * printable, in MAX_PRINT_A */
20019 UV max = MIN(end, MAX_PRINT_A);
20021 while (temp_end <= max && ! isPRINT_A(temp_end)) {
20025 /* Here, temp_end points to one beyond the first printable if
20026 * found, or to one beyond 'max' if not. If none found, make
20027 * sure that we use the entire range */
20028 if (temp_end > MAX_PRINT_A) {
20029 temp_end = end + 1;
20032 /* Output the first part of the split range: the part that
20033 * doesn't have printables, with the parameter set to not look
20034 * for literals (otherwise we would infinitely recurse) */
20035 put_range(sv, start, temp_end - 1, FALSE);
20037 /* The 2nd part of the range (if any) starts here. */
20040 /* We do a continue, instead of dropping down, because even if
20041 * the 2nd part is non-empty, it could be so short that we want
20042 * to output it as individual characters, as tested for at the
20043 * top of this loop. */
20047 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
20048 * output a sub-range of just the digits or letters, then process
20049 * the remaining portion as usual. */
20050 if (isALPHANUMERIC_A(start)) {
20051 UV mask = (isDIGIT_A(start))
20056 UV temp_end = start + 1;
20058 /* Find the end of the sub-range that includes just the
20059 * characters in the same class as the first character in it */
20060 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
20065 /* For short ranges, don't duplicate the code above to output
20066 * them; just call recursively */
20067 if (temp_end - start < min_range_count) {
20068 put_range(sv, start, temp_end, FALSE);
20070 else { /* Output as a range */
20071 put_code_point(sv, start);
20072 sv_catpvs(sv, "-");
20073 put_code_point(sv, temp_end);
20075 start = temp_end + 1;
20079 /* We output any other printables as individual characters */
20080 if (isPUNCT_A(start) || isSPACE_A(start)) {
20081 while (start <= end && (isPUNCT_A(start)
20082 || isSPACE_A(start)))
20084 put_code_point(sv, start);
20089 } /* End of looking for literals */
20091 /* Here is not to output as a literal. Some control characters have
20092 * mnemonic names. Split off any of those at the beginning and end of
20093 * the range to print mnemonically. It isn't possible for many of
20094 * these to be in a row, so this won't overwhelm with output */
20096 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
20098 while (isMNEMONIC_CNTRL(start) && start <= end) {
20099 put_code_point(sv, start);
20103 /* If this didn't take care of the whole range ... */
20104 if (start <= end) {
20106 /* Look backwards from the end to find the final non-mnemonic
20109 while (isMNEMONIC_CNTRL(temp_end)) {
20113 /* And separately output the interior range that doesn't start
20114 * or end with mnemonics */
20115 put_range(sv, start, temp_end, FALSE);
20117 /* Then output the mnemonic trailing controls */
20118 start = temp_end + 1;
20119 while (start <= end) {
20120 put_code_point(sv, start);
20127 /* As a final resort, output the range or subrange as hex. */
20129 this_end = (end < NUM_ANYOF_CODE_POINTS)
20131 : NUM_ANYOF_CODE_POINTS - 1;
20132 #if NUM_ANYOF_CODE_POINTS > 256
20133 format = (this_end < 256)
20134 ? "\\x%02" UVXf "-\\x%02" UVXf
20135 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
20137 format = "\\x%02" UVXf "-\\x%02" UVXf;
20139 GCC_DIAG_IGNORE(-Wformat-nonliteral);
20140 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
20147 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
20149 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
20153 bool allow_literals = TRUE;
20155 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
20157 /* Generally, it is more readable if printable characters are output as
20158 * literals, but if a range (nearly) spans all of them, it's best to output
20159 * it as a single range. This code will use a single range if all but 2
20160 * ASCII printables are in it */
20161 invlist_iterinit(invlist);
20162 while (invlist_iternext(invlist, &start, &end)) {
20164 /* If the range starts beyond the final printable, it doesn't have any
20166 if (start > MAX_PRINT_A) {
20170 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
20171 * all but two, the range must start and end no later than 2 from
20173 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
20174 if (end > MAX_PRINT_A) {
20180 if (end - start >= MAX_PRINT_A - ' ' - 2) {
20181 allow_literals = FALSE;
20186 invlist_iterfinish(invlist);
20188 /* Here we have figured things out. Output each range */
20189 invlist_iterinit(invlist);
20190 while (invlist_iternext(invlist, &start, &end)) {
20191 if (start >= NUM_ANYOF_CODE_POINTS) {
20194 put_range(sv, start, end, allow_literals);
20196 invlist_iterfinish(invlist);
20202 S_put_charclass_bitmap_innards_common(pTHX_
20203 SV* invlist, /* The bitmap */
20204 SV* posixes, /* Under /l, things like [:word:], \S */
20205 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
20206 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
20207 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
20208 const bool invert /* Is the result to be inverted? */
20211 /* Create and return an SV containing a displayable version of the bitmap
20212 * and associated information determined by the input parameters. If the
20213 * output would have been only the inversion indicator '^', NULL is instead
20218 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
20221 output = newSVpvs("^");
20224 output = newSVpvs("");
20227 /* First, the code points in the bitmap that are unconditionally there */
20228 put_charclass_bitmap_innards_invlist(output, invlist);
20230 /* Traditionally, these have been placed after the main code points */
20232 sv_catsv(output, posixes);
20235 if (only_utf8 && _invlist_len(only_utf8)) {
20236 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
20237 put_charclass_bitmap_innards_invlist(output, only_utf8);
20240 if (not_utf8 && _invlist_len(not_utf8)) {
20241 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
20242 put_charclass_bitmap_innards_invlist(output, not_utf8);
20245 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
20246 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
20247 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
20249 /* This is the only list in this routine that can legally contain code
20250 * points outside the bitmap range. The call just above to
20251 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
20252 * output them here. There's about a half-dozen possible, and none in
20253 * contiguous ranges longer than 2 */
20254 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20256 SV* above_bitmap = NULL;
20258 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
20260 invlist_iterinit(above_bitmap);
20261 while (invlist_iternext(above_bitmap, &start, &end)) {
20264 for (i = start; i <= end; i++) {
20265 put_code_point(output, i);
20268 invlist_iterfinish(above_bitmap);
20269 SvREFCNT_dec_NN(above_bitmap);
20273 if (invert && SvCUR(output) == 1) {
20281 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
20283 SV *nonbitmap_invlist,
20284 SV *only_utf8_locale_invlist,
20285 const regnode * const node,
20286 const bool force_as_is_display)
20288 /* Appends to 'sv' a displayable version of the innards of the bracketed
20289 * character class defined by the other arguments:
20290 * 'bitmap' points to the bitmap.
20291 * 'nonbitmap_invlist' is an inversion list of the code points that are in
20292 * the bitmap range, but for some reason aren't in the bitmap; NULL if
20293 * none. The reasons for this could be that they require some
20294 * condition such as the target string being or not being in UTF-8
20295 * (under /d), or because they came from a user-defined property that
20296 * was not resolved at the time of the regex compilation (under /u)
20297 * 'only_utf8_locale_invlist' is an inversion list of the code points that
20298 * are valid only if the runtime locale is a UTF-8 one; NULL if none
20299 * 'node' is the regex pattern node. It is needed only when the above two
20300 * parameters are not null, and is passed so that this routine can
20301 * tease apart the various reasons for them.
20302 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
20303 * to invert things to see if that leads to a cleaner display. If
20304 * FALSE, this routine is free to use its judgment about doing this.
20306 * It returns TRUE if there was actually something output. (It may be that
20307 * the bitmap, etc is empty.)
20309 * When called for outputting the bitmap of a non-ANYOF node, just pass the
20310 * bitmap, with the succeeding parameters set to NULL, and the final one to
20314 /* In general, it tries to display the 'cleanest' representation of the
20315 * innards, choosing whether to display them inverted or not, regardless of
20316 * whether the class itself is to be inverted. However, there are some
20317 * cases where it can't try inverting, as what actually matches isn't known
20318 * until runtime, and hence the inversion isn't either. */
20319 bool inverting_allowed = ! force_as_is_display;
20322 STRLEN orig_sv_cur = SvCUR(sv);
20324 SV* invlist; /* Inversion list we accumulate of code points that
20325 are unconditionally matched */
20326 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
20328 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
20330 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
20331 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
20334 SV* as_is_display; /* The output string when we take the inputs
20336 SV* inverted_display; /* The output string when we invert the inputs */
20338 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
20340 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
20342 /* We are biased in favor of displaying things without them being inverted,
20343 * as that is generally easier to understand */
20344 const int bias = 5;
20346 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
20348 /* Start off with whatever code points are passed in. (We clone, so we
20349 * don't change the caller's list) */
20350 if (nonbitmap_invlist) {
20351 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
20352 invlist = invlist_clone(nonbitmap_invlist);
20354 else { /* Worst case size is every other code point is matched */
20355 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
20359 if (OP(node) == ANYOFD) {
20361 /* This flag indicates that the code points below 0x100 in the
20362 * nonbitmap list are precisely the ones that match only when the
20363 * target is UTF-8 (they should all be non-ASCII). */
20364 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
20366 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
20367 _invlist_subtract(invlist, only_utf8, &invlist);
20370 /* And this flag for matching all non-ASCII 0xFF and below */
20371 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
20373 not_utf8 = invlist_clone(PL_UpperLatin1);
20376 else if (OP(node) == ANYOFL) {
20378 /* If either of these flags are set, what matches isn't
20379 * determinable except during execution, so don't know enough here
20381 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
20382 inverting_allowed = FALSE;
20385 /* What the posix classes match also varies at runtime, so these
20386 * will be output symbolically. */
20387 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
20390 posixes = newSVpvs("");
20391 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
20392 if (ANYOF_POSIXL_TEST(node,i)) {
20393 sv_catpv(posixes, anyofs[i]);
20400 /* Accumulate the bit map into the unconditional match list */
20401 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
20402 if (BITMAP_TEST(bitmap, i)) {
20404 for (; i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i); i++) {
20407 invlist = _add_range_to_invlist(invlist, start, i-1);
20411 /* Make sure that the conditional match lists don't have anything in them
20412 * that match unconditionally; otherwise the output is quite confusing.
20413 * This could happen if the code that populates these misses some
20416 _invlist_subtract(only_utf8, invlist, &only_utf8);
20419 _invlist_subtract(not_utf8, invlist, ¬_utf8);
20422 if (only_utf8_locale_invlist) {
20424 /* Since this list is passed in, we have to make a copy before
20426 only_utf8_locale = invlist_clone(only_utf8_locale_invlist);
20428 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
20430 /* And, it can get really weird for us to try outputting an inverted
20431 * form of this list when it has things above the bitmap, so don't even
20433 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20434 inverting_allowed = FALSE;
20438 /* Calculate what the output would be if we take the input as-is */
20439 as_is_display = put_charclass_bitmap_innards_common(invlist,
20446 /* If have to take the output as-is, just do that */
20447 if (! inverting_allowed) {
20448 if (as_is_display) {
20449 sv_catsv(sv, as_is_display);
20450 SvREFCNT_dec_NN(as_is_display);
20453 else { /* But otherwise, create the output again on the inverted input, and
20454 use whichever version is shorter */
20456 int inverted_bias, as_is_bias;
20458 /* We will apply our bias to whichever of the the results doesn't have
20468 inverted_bias = bias;
20471 /* Now invert each of the lists that contribute to the output,
20472 * excluding from the result things outside the possible range */
20474 /* For the unconditional inversion list, we have to add in all the
20475 * conditional code points, so that when inverted, they will be gone
20477 _invlist_union(only_utf8, invlist, &invlist);
20478 _invlist_union(not_utf8, invlist, &invlist);
20479 _invlist_union(only_utf8_locale, invlist, &invlist);
20480 _invlist_invert(invlist);
20481 _invlist_intersection(invlist, PL_InBitmap, &invlist);
20484 _invlist_invert(only_utf8);
20485 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
20487 else if (not_utf8) {
20489 /* If a code point matches iff the target string is not in UTF-8,
20490 * then complementing the result has it not match iff not in UTF-8,
20491 * which is the same thing as matching iff it is UTF-8. */
20492 only_utf8 = not_utf8;
20496 if (only_utf8_locale) {
20497 _invlist_invert(only_utf8_locale);
20498 _invlist_intersection(only_utf8_locale,
20500 &only_utf8_locale);
20503 inverted_display = put_charclass_bitmap_innards_common(
20508 only_utf8_locale, invert);
20510 /* Use the shortest representation, taking into account our bias
20511 * against showing it inverted */
20512 if ( inverted_display
20513 && ( ! as_is_display
20514 || ( SvCUR(inverted_display) + inverted_bias
20515 < SvCUR(as_is_display) + as_is_bias)))
20517 sv_catsv(sv, inverted_display);
20519 else if (as_is_display) {
20520 sv_catsv(sv, as_is_display);
20523 SvREFCNT_dec(as_is_display);
20524 SvREFCNT_dec(inverted_display);
20527 SvREFCNT_dec_NN(invlist);
20528 SvREFCNT_dec(only_utf8);
20529 SvREFCNT_dec(not_utf8);
20530 SvREFCNT_dec(posixes);
20531 SvREFCNT_dec(only_utf8_locale);
20533 return SvCUR(sv) > orig_sv_cur;
20536 #define CLEAR_OPTSTART \
20537 if (optstart) STMT_START { \
20538 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
20539 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
20543 #define DUMPUNTIL(b,e) \
20545 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
20547 STATIC const regnode *
20548 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
20549 const regnode *last, const regnode *plast,
20550 SV* sv, I32 indent, U32 depth)
20552 U8 op = PSEUDO; /* Arbitrary non-END op. */
20553 const regnode *next;
20554 const regnode *optstart= NULL;
20556 RXi_GET_DECL(r,ri);
20557 GET_RE_DEBUG_FLAGS_DECL;
20559 PERL_ARGS_ASSERT_DUMPUNTIL;
20561 #ifdef DEBUG_DUMPUNTIL
20562 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n",indent,node-start,
20563 last ? last-start : 0,plast ? plast-start : 0);
20566 if (plast && plast < last)
20569 while (PL_regkind[op] != END && (!last || node < last)) {
20571 /* While that wasn't END last time... */
20574 if (op == CLOSE || op == WHILEM)
20576 next = regnext((regnode *)node);
20579 if (OP(node) == OPTIMIZED) {
20580 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
20587 regprop(r, sv, node, NULL, NULL);
20588 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
20589 (int)(2*indent + 1), "", SvPVX_const(sv));
20591 if (OP(node) != OPTIMIZED) {
20592 if (next == NULL) /* Next ptr. */
20593 Perl_re_printf( aTHX_ " (0)");
20594 else if (PL_regkind[(U8)op] == BRANCH
20595 && PL_regkind[OP(next)] != BRANCH )
20596 Perl_re_printf( aTHX_ " (FAIL)");
20598 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
20599 Perl_re_printf( aTHX_ "\n");
20603 if (PL_regkind[(U8)op] == BRANCHJ) {
20606 const regnode *nnode = (OP(next) == LONGJMP
20607 ? regnext((regnode *)next)
20609 if (last && nnode > last)
20611 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
20614 else if (PL_regkind[(U8)op] == BRANCH) {
20616 DUMPUNTIL(NEXTOPER(node), next);
20618 else if ( PL_regkind[(U8)op] == TRIE ) {
20619 const regnode *this_trie = node;
20620 const char op = OP(node);
20621 const U32 n = ARG(node);
20622 const reg_ac_data * const ac = op>=AHOCORASICK ?
20623 (reg_ac_data *)ri->data->data[n] :
20625 const reg_trie_data * const trie =
20626 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
20628 AV *const trie_words
20629 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
20631 const regnode *nextbranch= NULL;
20634 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
20635 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
20637 Perl_re_indentf( aTHX_ "%s ",
20640 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
20641 SvCUR(*elem_ptr), 60,
20642 PL_colors[0], PL_colors[1],
20644 ? PERL_PV_ESCAPE_UNI
20646 | PERL_PV_PRETTY_ELLIPSES
20647 | PERL_PV_PRETTY_LTGT
20652 U16 dist= trie->jump[word_idx+1];
20653 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
20654 (UV)((dist ? this_trie + dist : next) - start));
20657 nextbranch= this_trie + trie->jump[0];
20658 DUMPUNTIL(this_trie + dist, nextbranch);
20660 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
20661 nextbranch= regnext((regnode *)nextbranch);
20663 Perl_re_printf( aTHX_ "\n");
20666 if (last && next > last)
20671 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
20672 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
20673 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
20675 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
20677 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
20679 else if ( op == PLUS || op == STAR) {
20680 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
20682 else if (PL_regkind[(U8)op] == ANYOF) {
20683 /* arglen 1 + class block */
20684 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
20685 ? ANYOF_POSIXL_SKIP
20687 node = NEXTOPER(node);
20689 else if (PL_regkind[(U8)op] == EXACT) {
20690 /* Literal string, where present. */
20691 node += NODE_SZ_STR(node) - 1;
20692 node = NEXTOPER(node);
20695 node = NEXTOPER(node);
20696 node += regarglen[(U8)op];
20698 if (op == CURLYX || op == OPEN)
20702 #ifdef DEBUG_DUMPUNTIL
20703 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
20708 #endif /* DEBUGGING */
20711 * ex: set ts=8 sts=4 sw=4 et: