5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
84 EXTERN_C const struct regexp_engine my_reg_engine;
89 #include "dquote_inline.h"
90 #include "invlist_inline.h"
91 #include "unicode_constants.h"
93 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
97 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
98 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
101 #define STATIC static
104 /* this is a chain of data about sub patterns we are processing that
105 need to be handled separately/specially in study_chunk. Its so
106 we can simulate recursion without losing state. */
108 typedef struct scan_frame {
109 regnode *last_regnode; /* last node to process in this frame */
110 regnode *next_regnode; /* next node to process when last is reached */
111 U32 prev_recursed_depth;
112 I32 stopparen; /* what stopparen do we use */
113 U32 is_top_frame; /* what flags do we use? */
115 struct scan_frame *this_prev_frame; /* this previous frame */
116 struct scan_frame *prev_frame; /* previous frame */
117 struct scan_frame *next_frame; /* next frame */
120 /* Certain characters are output as a sequence with the first being a
122 #define isBACKSLASHED_PUNCT(c) \
123 ((c) == '-' || (c) == ']' || (c) == '\\' || (c) == '^')
126 struct RExC_state_t {
127 U32 flags; /* RXf_* are we folding, multilining? */
128 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
129 char *precomp; /* uncompiled string. */
130 char *precomp_end; /* pointer to end of uncompiled string. */
131 REGEXP *rx_sv; /* The SV that is the regexp. */
132 regexp *rx; /* perl core regexp structure */
133 regexp_internal *rxi; /* internal data for regexp object
135 char *start; /* Start of input for compile */
136 char *end; /* End of input for compile */
137 char *parse; /* Input-scan pointer. */
138 char *adjusted_start; /* 'start', adjusted. See code use */
139 STRLEN precomp_adj; /* an offset beyond precomp. See code use */
140 SSize_t whilem_seen; /* number of WHILEM in this expr */
141 regnode *emit_start; /* Start of emitted-code area */
142 regnode *emit_bound; /* First regnode outside of the
144 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
145 implies compiling, so don't emit */
146 regnode_ssc emit_dummy; /* placeholder for emit to point to;
147 large enough for the largest
148 non-EXACTish node, so can use it as
150 I32 naughty; /* How bad is this pattern? */
151 I32 sawback; /* Did we see \1, ...? */
153 SSize_t size; /* Code size. */
154 I32 npar; /* Capture buffer count, (OPEN) plus
155 one. ("par" 0 is the whole
157 I32 nestroot; /* root parens we are in - used by
161 regnode **open_parens; /* pointers to open parens */
162 regnode **close_parens; /* pointers to close parens */
163 regnode *end_op; /* END node in program */
164 I32 utf8; /* whether the pattern is utf8 or not */
165 I32 orig_utf8; /* whether the pattern was originally in utf8 */
166 /* XXX use this for future optimisation of case
167 * where pattern must be upgraded to utf8. */
168 I32 uni_semantics; /* If a d charset modifier should use unicode
169 rules, even if the pattern is not in
171 HV *paren_names; /* Paren names */
173 regnode **recurse; /* Recurse regops */
174 I32 recurse_count; /* Number of recurse regops we have generated */
175 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
177 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
180 I32 override_recoding;
182 I32 recode_x_to_native;
184 I32 in_multi_char_class;
185 struct reg_code_block *code_blocks; /* positions of literal (?{})
187 int num_code_blocks; /* size of code_blocks[] */
188 int code_index; /* next code_blocks[] slot */
189 SSize_t maxlen; /* mininum possible number of chars in string to match */
190 scan_frame *frame_head;
191 scan_frame *frame_last;
194 #ifdef ADD_TO_REGEXEC
195 char *starttry; /* -Dr: where regtry was called. */
196 #define RExC_starttry (pRExC_state->starttry)
198 SV *runtime_code_qr; /* qr with the runtime code blocks */
200 const char *lastparse;
202 AV *paren_name_list; /* idx -> name */
203 U32 study_chunk_recursed_count;
206 #define RExC_lastparse (pRExC_state->lastparse)
207 #define RExC_lastnum (pRExC_state->lastnum)
208 #define RExC_paren_name_list (pRExC_state->paren_name_list)
209 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
210 #define RExC_mysv (pRExC_state->mysv1)
211 #define RExC_mysv1 (pRExC_state->mysv1)
212 #define RExC_mysv2 (pRExC_state->mysv2)
215 bool seen_unfolded_sharp_s;
220 #define RExC_flags (pRExC_state->flags)
221 #define RExC_pm_flags (pRExC_state->pm_flags)
222 #define RExC_precomp (pRExC_state->precomp)
223 #define RExC_precomp_adj (pRExC_state->precomp_adj)
224 #define RExC_adjusted_start (pRExC_state->adjusted_start)
225 #define RExC_precomp_end (pRExC_state->precomp_end)
226 #define RExC_rx_sv (pRExC_state->rx_sv)
227 #define RExC_rx (pRExC_state->rx)
228 #define RExC_rxi (pRExC_state->rxi)
229 #define RExC_start (pRExC_state->start)
230 #define RExC_end (pRExC_state->end)
231 #define RExC_parse (pRExC_state->parse)
232 #define RExC_whilem_seen (pRExC_state->whilem_seen)
234 /* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
235 * EXACTF node, hence was parsed under /di rules. If later in the parse,
236 * something forces the pattern into using /ui rules, the sharp s should be
237 * folded into the sequence 'ss', which takes up more space than previously
238 * calculated. This means that the sizing pass needs to be restarted. (The
239 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
240 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
241 * so there is no need to resize [perl #125990]. */
242 #define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
244 #ifdef RE_TRACK_PATTERN_OFFSETS
245 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
248 #define RExC_emit (pRExC_state->emit)
249 #define RExC_emit_dummy (pRExC_state->emit_dummy)
250 #define RExC_emit_start (pRExC_state->emit_start)
251 #define RExC_emit_bound (pRExC_state->emit_bound)
252 #define RExC_sawback (pRExC_state->sawback)
253 #define RExC_seen (pRExC_state->seen)
254 #define RExC_size (pRExC_state->size)
255 #define RExC_maxlen (pRExC_state->maxlen)
256 #define RExC_npar (pRExC_state->npar)
257 #define RExC_nestroot (pRExC_state->nestroot)
258 #define RExC_extralen (pRExC_state->extralen)
259 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
260 #define RExC_utf8 (pRExC_state->utf8)
261 #define RExC_uni_semantics (pRExC_state->uni_semantics)
262 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
263 #define RExC_open_parens (pRExC_state->open_parens)
264 #define RExC_close_parens (pRExC_state->close_parens)
265 #define RExC_end_op (pRExC_state->end_op)
266 #define RExC_paren_names (pRExC_state->paren_names)
267 #define RExC_recurse (pRExC_state->recurse)
268 #define RExC_recurse_count (pRExC_state->recurse_count)
269 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
270 #define RExC_study_chunk_recursed_bytes \
271 (pRExC_state->study_chunk_recursed_bytes)
272 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
273 #define RExC_contains_locale (pRExC_state->contains_locale)
274 #define RExC_override_recoding (pRExC_state->override_recoding)
276 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
278 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
279 #define RExC_frame_head (pRExC_state->frame_head)
280 #define RExC_frame_last (pRExC_state->frame_last)
281 #define RExC_frame_count (pRExC_state->frame_count)
282 #define RExC_strict (pRExC_state->strict)
283 #define RExC_study_started (pRExC_state->study_started)
284 #define RExC_warn_text (pRExC_state->warn_text)
286 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
287 * a flag to disable back-off on the fixed/floating substrings - if it's
288 * a high complexity pattern we assume the benefit of avoiding a full match
289 * is worth the cost of checking for the substrings even if they rarely help.
291 #define RExC_naughty (pRExC_state->naughty)
292 #define TOO_NAUGHTY (10)
293 #define MARK_NAUGHTY(add) \
294 if (RExC_naughty < TOO_NAUGHTY) \
295 RExC_naughty += (add)
296 #define MARK_NAUGHTY_EXP(exp, add) \
297 if (RExC_naughty < TOO_NAUGHTY) \
298 RExC_naughty += RExC_naughty / (exp) + (add)
300 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
301 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
302 ((*s) == '{' && regcurly(s)))
305 * Flags to be passed up and down.
307 #define WORST 0 /* Worst case. */
308 #define HASWIDTH 0x01 /* Known to match non-null strings. */
310 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
311 * character. (There needs to be a case: in the switch statement in regexec.c
312 * for any node marked SIMPLE.) Note that this is not the same thing as
315 #define SPSTART 0x04 /* Starts with * or + */
316 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
317 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
318 #define RESTART_PASS1 0x20 /* Need to restart sizing pass */
319 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
320 calcuate sizes as UTF-8 */
322 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
324 /* whether trie related optimizations are enabled */
325 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
326 #define TRIE_STUDY_OPT
327 #define FULL_TRIE_STUDY
333 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
334 #define PBITVAL(paren) (1 << ((paren) & 7))
335 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
336 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
337 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
339 #define REQUIRE_UTF8(flagp) STMT_START { \
342 *flagp = RESTART_PASS1|NEED_UTF8; \
347 /* Change from /d into /u rules, and restart the parse if we've already seen
348 * something whose size would increase as a result, by setting *flagp and
349 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
350 * we've change to /u during the parse. */
351 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
353 if (DEPENDS_SEMANTICS) { \
355 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
356 RExC_uni_semantics = 1; \
357 if (RExC_seen_unfolded_sharp_s) { \
358 *flagp |= RESTART_PASS1; \
359 return restart_retval; \
364 /* This converts the named class defined in regcomp.h to its equivalent class
365 * number defined in handy.h. */
366 #define namedclass_to_classnum(class) ((int) ((class) / 2))
367 #define classnum_to_namedclass(classnum) ((classnum) * 2)
369 #define _invlist_union_complement_2nd(a, b, output) \
370 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
371 #define _invlist_intersection_complement_2nd(a, b, output) \
372 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
374 /* About scan_data_t.
376 During optimisation we recurse through the regexp program performing
377 various inplace (keyhole style) optimisations. In addition study_chunk
378 and scan_commit populate this data structure with information about
379 what strings MUST appear in the pattern. We look for the longest
380 string that must appear at a fixed location, and we look for the
381 longest string that may appear at a floating location. So for instance
386 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
387 strings (because they follow a .* construct). study_chunk will identify
388 both FOO and BAR as being the longest fixed and floating strings respectively.
390 The strings can be composites, for instance
394 will result in a composite fixed substring 'foo'.
396 For each string some basic information is maintained:
398 - offset or min_offset
399 This is the position the string must appear at, or not before.
400 It also implicitly (when combined with minlenp) tells us how many
401 characters must match before the string we are searching for.
402 Likewise when combined with minlenp and the length of the string it
403 tells us how many characters must appear after the string we have
407 Only used for floating strings. This is the rightmost point that
408 the string can appear at. If set to SSize_t_MAX it indicates that the
409 string can occur infinitely far to the right.
412 A pointer to the minimum number of characters of the pattern that the
413 string was found inside. This is important as in the case of positive
414 lookahead or positive lookbehind we can have multiple patterns
419 The minimum length of the pattern overall is 3, the minimum length
420 of the lookahead part is 3, but the minimum length of the part that
421 will actually match is 1. So 'FOO's minimum length is 3, but the
422 minimum length for the F is 1. This is important as the minimum length
423 is used to determine offsets in front of and behind the string being
424 looked for. Since strings can be composites this is the length of the
425 pattern at the time it was committed with a scan_commit. Note that
426 the length is calculated by study_chunk, so that the minimum lengths
427 are not known until the full pattern has been compiled, thus the
428 pointer to the value.
432 In the case of lookbehind the string being searched for can be
433 offset past the start point of the final matching string.
434 If this value was just blithely removed from the min_offset it would
435 invalidate some of the calculations for how many chars must match
436 before or after (as they are derived from min_offset and minlen and
437 the length of the string being searched for).
438 When the final pattern is compiled and the data is moved from the
439 scan_data_t structure into the regexp structure the information
440 about lookbehind is factored in, with the information that would
441 have been lost precalculated in the end_shift field for the
444 The fields pos_min and pos_delta are used to store the minimum offset
445 and the delta to the maximum offset at the current point in the pattern.
449 typedef struct scan_data_t {
450 /*I32 len_min; unused */
451 /*I32 len_delta; unused */
455 SSize_t last_end; /* min value, <0 unless valid. */
456 SSize_t last_start_min;
457 SSize_t last_start_max;
458 SV **longest; /* Either &l_fixed, or &l_float. */
459 SV *longest_fixed; /* longest fixed string found in pattern */
460 SSize_t offset_fixed; /* offset where it starts */
461 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
462 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
463 SV *longest_float; /* longest floating string found in pattern */
464 SSize_t offset_float_min; /* earliest point in string it can appear */
465 SSize_t offset_float_max; /* latest point in string it can appear */
466 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
467 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
470 SSize_t *last_closep;
471 regnode_ssc *start_class;
475 * Forward declarations for pregcomp()'s friends.
478 static const scan_data_t zero_scan_data =
479 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
481 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
482 #define SF_BEFORE_SEOL 0x0001
483 #define SF_BEFORE_MEOL 0x0002
484 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
485 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
487 #define SF_FIX_SHIFT_EOL (+2)
488 #define SF_FL_SHIFT_EOL (+4)
490 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
491 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
493 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
494 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
495 #define SF_IS_INF 0x0040
496 #define SF_HAS_PAR 0x0080
497 #define SF_IN_PAR 0x0100
498 #define SF_HAS_EVAL 0x0200
501 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
502 * longest substring in the pattern. When it is not set the optimiser keeps
503 * track of position, but does not keep track of the actual strings seen,
505 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
508 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
509 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
510 * turned off because of the alternation (BRANCH). */
511 #define SCF_DO_SUBSTR 0x0400
513 #define SCF_DO_STCLASS_AND 0x0800
514 #define SCF_DO_STCLASS_OR 0x1000
515 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
516 #define SCF_WHILEM_VISITED_POS 0x2000
518 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
519 #define SCF_SEEN_ACCEPT 0x8000
520 #define SCF_TRIE_DOING_RESTUDY 0x10000
521 #define SCF_IN_DEFINE 0x20000
526 #define UTF cBOOL(RExC_utf8)
528 /* The enums for all these are ordered so things work out correctly */
529 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
530 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
531 == REGEX_DEPENDS_CHARSET)
532 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
533 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
534 >= REGEX_UNICODE_CHARSET)
535 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
536 == REGEX_ASCII_RESTRICTED_CHARSET)
537 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
538 >= REGEX_ASCII_RESTRICTED_CHARSET)
539 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
540 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
542 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
544 /* For programs that want to be strictly Unicode compatible by dying if any
545 * attempt is made to match a non-Unicode code point against a Unicode
547 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
549 #define OOB_NAMEDCLASS -1
551 /* There is no code point that is out-of-bounds, so this is problematic. But
552 * its only current use is to initialize a variable that is always set before
554 #define OOB_UNICODE 0xDEADBEEF
556 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
559 /* length of regex to show in messages that don't mark a position within */
560 #define RegexLengthToShowInErrorMessages 127
563 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
564 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
565 * op/pragma/warn/regcomp.
567 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
568 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
570 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
571 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
573 /* The code in this file in places uses one level of recursion with parsing
574 * rebased to an alternate string constructed by us in memory. This can take
575 * the form of something that is completely different from the input, or
576 * something that uses the input as part of the alternate. In the first case,
577 * there should be no possibility of an error, as we are in complete control of
578 * the alternate string. But in the second case we don't control the input
579 * portion, so there may be errors in that. Here's an example:
581 * is handled specially because \x{df} folds to a sequence of more than one
582 * character, 'ss'. What is done is to create and parse an alternate string,
583 * which looks like this:
584 * /(?:\x{DF}|[abc\x{DF}def])/ui
585 * where it uses the input unchanged in the middle of something it constructs,
586 * which is a branch for the DF outside the character class, and clustering
587 * parens around the whole thing. (It knows enough to skip the DF inside the
588 * class while in this substitute parse.) 'abc' and 'def' may have errors that
589 * need to be reported. The general situation looks like this:
592 * Input: ----------------------------------------------------
593 * Constructed: ---------------------------------------------------
596 * The input string sI..eI is the input pattern. The string sC..EC is the
597 * constructed substitute parse string. The portions sC..tC and eC..EC are
598 * constructed by us. The portion tC..eC is an exact duplicate of the input
599 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
600 * while parsing, we find an error at xC. We want to display a message showing
601 * the real input string. Thus we need to find the point xI in it which
602 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
603 * been constructed by us, and so shouldn't have errors. We get:
605 * xI = sI + (tI - sI) + (xC - tC)
607 * and, the offset into sI is:
609 * (xI - sI) = (tI - sI) + (xC - tC)
611 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
612 * and we save tC as RExC_adjusted_start.
614 * During normal processing of the input pattern, everything points to that,
615 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
618 #define tI_sI RExC_precomp_adj
619 #define tC RExC_adjusted_start
620 #define sC RExC_precomp
621 #define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
622 #define xI(xC) (sC + xI_offset(xC))
623 #define eC RExC_precomp_end
625 #define REPORT_LOCATION_ARGS(xC) \
627 (xI(xC) > eC) /* Don't run off end */ \
628 ? eC - sC /* Length before the <--HERE */ \
630 sC), /* The input pattern printed up to the <--HERE */ \
632 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
633 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
635 /* Used to point after bad bytes for an error message, but avoid skipping
636 * past a nul byte. */
637 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
640 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
641 * arg. Show regex, up to a maximum length. If it's too long, chop and add
644 #define _FAIL(code) STMT_START { \
645 const char *ellipses = ""; \
646 IV len = RExC_precomp_end - RExC_precomp; \
649 SAVEFREESV(RExC_rx_sv); \
650 if (len > RegexLengthToShowInErrorMessages) { \
651 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
652 len = RegexLengthToShowInErrorMessages - 10; \
658 #define FAIL(msg) _FAIL( \
659 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
660 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
662 #define FAIL2(msg,arg) _FAIL( \
663 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
664 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
667 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
669 #define Simple_vFAIL(m) STMT_START { \
670 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
671 m, REPORT_LOCATION_ARGS(RExC_parse)); \
675 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
677 #define vFAIL(m) STMT_START { \
679 SAVEFREESV(RExC_rx_sv); \
684 * Like Simple_vFAIL(), but accepts two arguments.
686 #define Simple_vFAIL2(m,a1) STMT_START { \
687 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
688 REPORT_LOCATION_ARGS(RExC_parse)); \
692 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
694 #define vFAIL2(m,a1) STMT_START { \
696 SAVEFREESV(RExC_rx_sv); \
697 Simple_vFAIL2(m, a1); \
702 * Like Simple_vFAIL(), but accepts three arguments.
704 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
705 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
706 REPORT_LOCATION_ARGS(RExC_parse)); \
710 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
712 #define vFAIL3(m,a1,a2) STMT_START { \
714 SAVEFREESV(RExC_rx_sv); \
715 Simple_vFAIL3(m, a1, a2); \
719 * Like Simple_vFAIL(), but accepts four arguments.
721 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
722 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
723 REPORT_LOCATION_ARGS(RExC_parse)); \
726 #define vFAIL4(m,a1,a2,a3) STMT_START { \
728 SAVEFREESV(RExC_rx_sv); \
729 Simple_vFAIL4(m, a1, a2, a3); \
732 /* A specialized version of vFAIL2 that works with UTF8f */
733 #define vFAIL2utf8f(m, a1) STMT_START { \
735 SAVEFREESV(RExC_rx_sv); \
736 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
737 REPORT_LOCATION_ARGS(RExC_parse)); \
740 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
742 SAVEFREESV(RExC_rx_sv); \
743 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
744 REPORT_LOCATION_ARGS(RExC_parse)); \
747 /* These have asserts in them because of [perl #122671] Many warnings in
748 * regcomp.c can occur twice. If they get output in pass1 and later in that
749 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
750 * would get output again. So they should be output in pass2, and these
751 * asserts make sure new warnings follow that paradigm. */
753 /* m is not necessarily a "literal string", in this macro */
754 #define reg_warn_non_literal_string(loc, m) STMT_START { \
755 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
756 "%s" REPORT_LOCATION, \
757 m, REPORT_LOCATION_ARGS(loc)); \
760 #define ckWARNreg(loc,m) STMT_START { \
761 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
763 REPORT_LOCATION_ARGS(loc)); \
766 #define vWARN(loc, m) STMT_START { \
767 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
769 REPORT_LOCATION_ARGS(loc)); \
772 #define vWARN_dep(loc, m) STMT_START { \
773 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
775 REPORT_LOCATION_ARGS(loc)); \
778 #define ckWARNdep(loc,m) STMT_START { \
779 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
781 REPORT_LOCATION_ARGS(loc)); \
784 #define ckWARNregdep(loc,m) STMT_START { \
785 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
788 REPORT_LOCATION_ARGS(loc)); \
791 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
792 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
794 a1, REPORT_LOCATION_ARGS(loc)); \
797 #define ckWARN2reg(loc, m, a1) STMT_START { \
798 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
800 a1, REPORT_LOCATION_ARGS(loc)); \
803 #define vWARN3(loc, m, a1, a2) STMT_START { \
804 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
806 a1, a2, REPORT_LOCATION_ARGS(loc)); \
809 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
810 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
813 REPORT_LOCATION_ARGS(loc)); \
816 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
817 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
820 REPORT_LOCATION_ARGS(loc)); \
823 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
824 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
827 REPORT_LOCATION_ARGS(loc)); \
830 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
831 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
834 REPORT_LOCATION_ARGS(loc)); \
837 /* Macros for recording node offsets. 20001227 mjd@plover.com
838 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
839 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
840 * Element 0 holds the number n.
841 * Position is 1 indexed.
843 #ifndef RE_TRACK_PATTERN_OFFSETS
844 #define Set_Node_Offset_To_R(node,byte)
845 #define Set_Node_Offset(node,byte)
846 #define Set_Cur_Node_Offset
847 #define Set_Node_Length_To_R(node,len)
848 #define Set_Node_Length(node,len)
849 #define Set_Node_Cur_Length(node,start)
850 #define Node_Offset(n)
851 #define Node_Length(n)
852 #define Set_Node_Offset_Length(node,offset,len)
853 #define ProgLen(ri) ri->u.proglen
854 #define SetProgLen(ri,x) ri->u.proglen = x
856 #define ProgLen(ri) ri->u.offsets[0]
857 #define SetProgLen(ri,x) ri->u.offsets[0] = x
858 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
860 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
861 __LINE__, (int)(node), (int)(byte))); \
863 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
866 RExC_offsets[2*(node)-1] = (byte); \
871 #define Set_Node_Offset(node,byte) \
872 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
873 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
875 #define Set_Node_Length_To_R(node,len) STMT_START { \
877 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
878 __LINE__, (int)(node), (int)(len))); \
880 Perl_croak(aTHX_ "value of node is %d in Length macro", \
883 RExC_offsets[2*(node)] = (len); \
888 #define Set_Node_Length(node,len) \
889 Set_Node_Length_To_R((node)-RExC_emit_start, len)
890 #define Set_Node_Cur_Length(node, start) \
891 Set_Node_Length(node, RExC_parse - start)
893 /* Get offsets and lengths */
894 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
895 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
897 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
898 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
899 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
903 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
904 #define EXPERIMENTAL_INPLACESCAN
905 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
909 Perl_re_printf(pTHX_ const char *fmt, ...)
913 PerlIO *f= Perl_debug_log;
914 PERL_ARGS_ASSERT_RE_PRINTF;
916 result = PerlIO_vprintf(f, fmt, ap);
922 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
926 PerlIO *f= Perl_debug_log;
927 PERL_ARGS_ASSERT_RE_INDENTF;
929 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
930 result = PerlIO_vprintf(f, fmt, ap);
934 #endif /* DEBUGGING */
936 #define DEBUG_RExC_seen() \
937 DEBUG_OPTIMISE_MORE_r({ \
938 Perl_re_printf( aTHX_ "RExC_seen: "); \
940 if (RExC_seen & REG_ZERO_LEN_SEEN) \
941 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
943 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
944 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
946 if (RExC_seen & REG_GPOS_SEEN) \
947 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
949 if (RExC_seen & REG_RECURSE_SEEN) \
950 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
952 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
953 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
955 if (RExC_seen & REG_VERBARG_SEEN) \
956 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
958 if (RExC_seen & REG_CUTGROUP_SEEN) \
959 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
961 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
962 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
964 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
965 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
967 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
968 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
970 Perl_re_printf( aTHX_ "\n"); \
973 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
974 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
976 #define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
978 Perl_re_printf( aTHX_ "%s", open_str); \
979 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
980 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
981 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
982 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
983 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
984 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
985 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
986 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
987 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
988 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
989 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
990 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
991 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
992 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
993 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
994 Perl_re_printf( aTHX_ "%s", close_str); \
998 #define DEBUG_STUDYDATA(str,data,depth) \
999 DEBUG_OPTIMISE_MORE_r(if(data){ \
1000 Perl_re_indentf( aTHX_ "" str "Pos:%" IVdf "/%" IVdf \
1001 " Flags: 0x%" UVXf, \
1003 (IV)((data)->pos_min), \
1004 (IV)((data)->pos_delta), \
1005 (UV)((data)->flags) \
1007 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
1008 Perl_re_printf( aTHX_ \
1009 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s", \
1010 (IV)((data)->whilem_c), \
1011 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
1012 is_inf ? "INF " : "" \
1014 if ((data)->last_found) \
1015 Perl_re_printf( aTHX_ \
1016 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf \
1017 " %sFixed:'%s' @ %" IVdf \
1018 " %sFloat: '%s' @ %" IVdf "/%" IVdf, \
1019 SvPVX_const((data)->last_found), \
1020 (IV)((data)->last_end), \
1021 (IV)((data)->last_start_min), \
1022 (IV)((data)->last_start_max), \
1023 ((data)->longest && \
1024 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
1025 SvPVX_const((data)->longest_fixed), \
1026 (IV)((data)->offset_fixed), \
1027 ((data)->longest && \
1028 (data)->longest==&((data)->longest_float)) ? "*" : "", \
1029 SvPVX_const((data)->longest_float), \
1030 (IV)((data)->offset_float_min), \
1031 (IV)((data)->offset_float_max) \
1033 Perl_re_printf( aTHX_ "\n"); \
1037 /* =========================================================
1038 * BEGIN edit_distance stuff.
1040 * This calculates how many single character changes of any type are needed to
1041 * transform a string into another one. It is taken from version 3.1 of
1043 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1046 /* Our unsorted dictionary linked list. */
1047 /* Note we use UVs, not chars. */
1052 struct dictionary* next;
1054 typedef struct dictionary item;
1057 PERL_STATIC_INLINE item*
1058 push(UV key,item* curr)
1061 Newxz(head, 1, item);
1069 PERL_STATIC_INLINE item*
1070 find(item* head, UV key)
1072 item* iterator = head;
1074 if (iterator->key == key){
1077 iterator = iterator->next;
1083 PERL_STATIC_INLINE item*
1084 uniquePush(item* head,UV key)
1086 item* iterator = head;
1089 if (iterator->key == key) {
1092 iterator = iterator->next;
1095 return push(key,head);
1098 PERL_STATIC_INLINE void
1099 dict_free(item* head)
1101 item* iterator = head;
1104 item* temp = iterator;
1105 iterator = iterator->next;
1112 /* End of Dictionary Stuff */
1114 /* All calculations/work are done here */
1116 S_edit_distance(const UV* src,
1118 const STRLEN x, /* length of src[] */
1119 const STRLEN y, /* length of tgt[] */
1120 const SSize_t maxDistance
1124 UV swapCount,swapScore,targetCharCount,i,j;
1126 UV score_ceil = x + y;
1128 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1130 /* intialize matrix start values */
1131 Newxz(scores, ( (x + 2) * (y + 2)), UV);
1132 scores[0] = score_ceil;
1133 scores[1 * (y + 2) + 0] = score_ceil;
1134 scores[0 * (y + 2) + 1] = score_ceil;
1135 scores[1 * (y + 2) + 1] = 0;
1136 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1141 for (i=1;i<=x;i++) {
1143 head = uniquePush(head,src[i]);
1144 scores[(i+1) * (y + 2) + 1] = i;
1145 scores[(i+1) * (y + 2) + 0] = score_ceil;
1148 for (j=1;j<=y;j++) {
1151 head = uniquePush(head,tgt[j]);
1152 scores[1 * (y + 2) + (j + 1)] = j;
1153 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1156 targetCharCount = find(head,tgt[j-1])->value;
1157 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1159 if (src[i-1] != tgt[j-1]){
1160 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));
1164 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1168 find(head,src[i-1])->value = i;
1172 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1175 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1179 /* END of edit_distance() stuff
1180 * ========================================================= */
1182 /* is c a control character for which we have a mnemonic? */
1183 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1186 S_cntrl_to_mnemonic(const U8 c)
1188 /* Returns the mnemonic string that represents character 'c', if one
1189 * exists; NULL otherwise. The only ones that exist for the purposes of
1190 * this routine are a few control characters */
1193 case '\a': return "\\a";
1194 case '\b': return "\\b";
1195 case ESC_NATIVE: return "\\e";
1196 case '\f': return "\\f";
1197 case '\n': return "\\n";
1198 case '\r': return "\\r";
1199 case '\t': return "\\t";
1205 /* Mark that we cannot extend a found fixed substring at this point.
1206 Update the longest found anchored substring and the longest found
1207 floating substrings if needed. */
1210 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1211 SSize_t *minlenp, int is_inf)
1213 const STRLEN l = CHR_SVLEN(data->last_found);
1214 const STRLEN old_l = CHR_SVLEN(*data->longest);
1215 GET_RE_DEBUG_FLAGS_DECL;
1217 PERL_ARGS_ASSERT_SCAN_COMMIT;
1219 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1220 SvSetMagicSV(*data->longest, data->last_found);
1221 if (*data->longest == data->longest_fixed) {
1222 data->offset_fixed = l ? data->last_start_min : data->pos_min;
1223 if (data->flags & SF_BEFORE_EOL)
1225 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
1227 data->flags &= ~SF_FIX_BEFORE_EOL;
1228 data->minlen_fixed=minlenp;
1229 data->lookbehind_fixed=0;
1231 else { /* *data->longest == data->longest_float */
1232 data->offset_float_min = l ? data->last_start_min : data->pos_min;
1233 data->offset_float_max = (l
1234 ? data->last_start_max
1235 : (data->pos_delta > SSize_t_MAX - data->pos_min
1237 : data->pos_min + data->pos_delta));
1239 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
1240 data->offset_float_max = SSize_t_MAX;
1241 if (data->flags & SF_BEFORE_EOL)
1243 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
1245 data->flags &= ~SF_FL_BEFORE_EOL;
1246 data->minlen_float=minlenp;
1247 data->lookbehind_float=0;
1250 SvCUR_set(data->last_found, 0);
1252 SV * const sv = data->last_found;
1253 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1254 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1259 data->last_end = -1;
1260 data->flags &= ~SF_BEFORE_EOL;
1261 DEBUG_STUDYDATA("commit: ",data,0);
1264 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1265 * list that describes which code points it matches */
1268 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1270 /* Set the SSC 'ssc' to match an empty string or any code point */
1272 PERL_ARGS_ASSERT_SSC_ANYTHING;
1274 assert(is_ANYOF_SYNTHETIC(ssc));
1276 /* mortalize so won't leak */
1277 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1278 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1282 S_ssc_is_anything(const regnode_ssc *ssc)
1284 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1285 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1286 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1287 * in any way, so there's no point in using it */
1292 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1294 assert(is_ANYOF_SYNTHETIC(ssc));
1296 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1300 /* See if the list consists solely of the range 0 - Infinity */
1301 invlist_iterinit(ssc->invlist);
1302 ret = invlist_iternext(ssc->invlist, &start, &end)
1306 invlist_iterfinish(ssc->invlist);
1312 /* If e.g., both \w and \W are set, matches everything */
1313 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1315 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1316 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1326 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1328 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1329 * string, any code point, or any posix class under locale */
1331 PERL_ARGS_ASSERT_SSC_INIT;
1333 Zero(ssc, 1, regnode_ssc);
1334 set_ANYOF_SYNTHETIC(ssc);
1335 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1338 /* If any portion of the regex is to operate under locale rules that aren't
1339 * fully known at compile time, initialization includes it. The reason
1340 * this isn't done for all regexes is that the optimizer was written under
1341 * the assumption that locale was all-or-nothing. Given the complexity and
1342 * lack of documentation in the optimizer, and that there are inadequate
1343 * test cases for locale, many parts of it may not work properly, it is
1344 * safest to avoid locale unless necessary. */
1345 if (RExC_contains_locale) {
1346 ANYOF_POSIXL_SETALL(ssc);
1349 ANYOF_POSIXL_ZERO(ssc);
1354 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1355 const regnode_ssc *ssc)
1357 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1358 * to the list of code points matched, and locale posix classes; hence does
1359 * not check its flags) */
1364 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1366 assert(is_ANYOF_SYNTHETIC(ssc));
1368 invlist_iterinit(ssc->invlist);
1369 ret = invlist_iternext(ssc->invlist, &start, &end)
1373 invlist_iterfinish(ssc->invlist);
1379 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1387 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1388 const regnode_charclass* const node)
1390 /* Returns a mortal inversion list defining which code points are matched
1391 * by 'node', which is of type ANYOF. Handles complementing the result if
1392 * appropriate. If some code points aren't knowable at this time, the
1393 * returned list must, and will, contain every code point that is a
1397 SV* only_utf8_locale_invlist = NULL;
1399 const U32 n = ARG(node);
1400 bool new_node_has_latin1 = FALSE;
1402 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1404 /* Look at the data structure created by S_set_ANYOF_arg() */
1405 if (n != ANYOF_ONLY_HAS_BITMAP) {
1406 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1407 AV * const av = MUTABLE_AV(SvRV(rv));
1408 SV **const ary = AvARRAY(av);
1409 assert(RExC_rxi->data->what[n] == 's');
1411 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1412 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1414 else if (ary[0] && ary[0] != &PL_sv_undef) {
1416 /* Here, no compile-time swash, and there are things that won't be
1417 * known until runtime -- we have to assume it could be anything */
1418 invlist = sv_2mortal(_new_invlist(1));
1419 return _add_range_to_invlist(invlist, 0, UV_MAX);
1421 else if (ary[3] && ary[3] != &PL_sv_undef) {
1423 /* Here no compile-time swash, and no run-time only data. Use the
1424 * node's inversion list */
1425 invlist = sv_2mortal(invlist_clone(ary[3]));
1428 /* Get the code points valid only under UTF-8 locales */
1429 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1430 && ary[2] && ary[2] != &PL_sv_undef)
1432 only_utf8_locale_invlist = ary[2];
1437 invlist = sv_2mortal(_new_invlist(0));
1440 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1441 * code points, and an inversion list for the others, but if there are code
1442 * points that should match only conditionally on the target string being
1443 * UTF-8, those are placed in the inversion list, and not the bitmap.
1444 * Since there are circumstances under which they could match, they are
1445 * included in the SSC. But if the ANYOF node is to be inverted, we have
1446 * to exclude them here, so that when we invert below, the end result
1447 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1448 * have to do this here before we add the unconditionally matched code
1450 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1451 _invlist_intersection_complement_2nd(invlist,
1456 /* Add in the points from the bit map */
1457 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1458 if (ANYOF_BITMAP_TEST(node, i)) {
1459 unsigned int start = i++;
1461 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1464 invlist = _add_range_to_invlist(invlist, start, i-1);
1465 new_node_has_latin1 = TRUE;
1469 /* If this can match all upper Latin1 code points, have to add them
1470 * as well. But don't add them if inverting, as when that gets done below,
1471 * it would exclude all these characters, including the ones it shouldn't
1472 * that were added just above */
1473 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1474 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1476 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1479 /* Similarly for these */
1480 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1481 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1484 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1485 _invlist_invert(invlist);
1487 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1489 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1490 * locale. We can skip this if there are no 0-255 at all. */
1491 _invlist_union(invlist, PL_Latin1, &invlist);
1494 /* Similarly add the UTF-8 locale possible matches. These have to be
1495 * deferred until after the non-UTF-8 locale ones are taken care of just
1496 * above, or it leads to wrong results under ANYOF_INVERT */
1497 if (only_utf8_locale_invlist) {
1498 _invlist_union_maybe_complement_2nd(invlist,
1499 only_utf8_locale_invlist,
1500 ANYOF_FLAGS(node) & ANYOF_INVERT,
1507 /* These two functions currently do the exact same thing */
1508 #define ssc_init_zero ssc_init
1510 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1511 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1513 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1514 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1515 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1518 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1519 const regnode_charclass *and_with)
1521 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1522 * another SSC or a regular ANYOF class. Can create false positives. */
1527 PERL_ARGS_ASSERT_SSC_AND;
1529 assert(is_ANYOF_SYNTHETIC(ssc));
1531 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1532 * the code point inversion list and just the relevant flags */
1533 if (is_ANYOF_SYNTHETIC(and_with)) {
1534 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1535 anded_flags = ANYOF_FLAGS(and_with);
1537 /* XXX This is a kludge around what appears to be deficiencies in the
1538 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1539 * there are paths through the optimizer where it doesn't get weeded
1540 * out when it should. And if we don't make some extra provision for
1541 * it like the code just below, it doesn't get added when it should.
1542 * This solution is to add it only when AND'ing, which is here, and
1543 * only when what is being AND'ed is the pristine, original node
1544 * matching anything. Thus it is like adding it to ssc_anything() but
1545 * only when the result is to be AND'ed. Probably the same solution
1546 * could be adopted for the same problem we have with /l matching,
1547 * which is solved differently in S_ssc_init(), and that would lead to
1548 * fewer false positives than that solution has. But if this solution
1549 * creates bugs, the consequences are only that a warning isn't raised
1550 * that should be; while the consequences for having /l bugs is
1551 * incorrect matches */
1552 if (ssc_is_anything((regnode_ssc *)and_with)) {
1553 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1557 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1558 if (OP(and_with) == ANYOFD) {
1559 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1562 anded_flags = ANYOF_FLAGS(and_with)
1563 &( ANYOF_COMMON_FLAGS
1564 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1565 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1566 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1568 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1573 ANYOF_FLAGS(ssc) &= anded_flags;
1575 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1576 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1577 * 'and_with' may be inverted. When not inverted, we have the situation of
1579 * (C1 | P1) & (C2 | P2)
1580 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1581 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1582 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1583 * <= ((C1 & C2) | P1 | P2)
1584 * Alternatively, the last few steps could be:
1585 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1586 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1587 * <= (C1 | C2 | (P1 & P2))
1588 * We favor the second approach if either P1 or P2 is non-empty. This is
1589 * because these components are a barrier to doing optimizations, as what
1590 * they match cannot be known until the moment of matching as they are
1591 * dependent on the current locale, 'AND"ing them likely will reduce or
1593 * But we can do better if we know that C1,P1 are in their initial state (a
1594 * frequent occurrence), each matching everything:
1595 * (<everything>) & (C2 | P2) = C2 | P2
1596 * Similarly, if C2,P2 are in their initial state (again a frequent
1597 * occurrence), the result is a no-op
1598 * (C1 | P1) & (<everything>) = C1 | P1
1601 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1602 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1603 * <= (C1 & ~C2) | (P1 & ~P2)
1606 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1607 && ! is_ANYOF_SYNTHETIC(and_with))
1611 ssc_intersection(ssc,
1613 FALSE /* Has already been inverted */
1616 /* If either P1 or P2 is empty, the intersection will be also; can skip
1618 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1619 ANYOF_POSIXL_ZERO(ssc);
1621 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1623 /* Note that the Posix class component P from 'and_with' actually
1625 * P = Pa | Pb | ... | Pn
1626 * where each component is one posix class, such as in [\w\s].
1628 * ~P = ~(Pa | Pb | ... | Pn)
1629 * = ~Pa & ~Pb & ... & ~Pn
1630 * <= ~Pa | ~Pb | ... | ~Pn
1631 * The last is something we can easily calculate, but unfortunately
1632 * is likely to have many false positives. We could do better
1633 * in some (but certainly not all) instances if two classes in
1634 * P have known relationships. For example
1635 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1637 * :lower: & :print: = :lower:
1638 * And similarly for classes that must be disjoint. For example,
1639 * since \s and \w can have no elements in common based on rules in
1640 * the POSIX standard,
1641 * \w & ^\S = nothing
1642 * Unfortunately, some vendor locales do not meet the Posix
1643 * standard, in particular almost everything by Microsoft.
1644 * The loop below just changes e.g., \w into \W and vice versa */
1646 regnode_charclass_posixl temp;
1647 int add = 1; /* To calculate the index of the complement */
1649 ANYOF_POSIXL_ZERO(&temp);
1650 for (i = 0; i < ANYOF_MAX; i++) {
1652 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1653 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1655 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1656 ANYOF_POSIXL_SET(&temp, i + add);
1658 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1660 ANYOF_POSIXL_AND(&temp, ssc);
1662 } /* else ssc already has no posixes */
1663 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1664 in its initial state */
1665 else if (! is_ANYOF_SYNTHETIC(and_with)
1666 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1668 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1669 * copy it over 'ssc' */
1670 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1671 if (is_ANYOF_SYNTHETIC(and_with)) {
1672 StructCopy(and_with, ssc, regnode_ssc);
1675 ssc->invlist = anded_cp_list;
1676 ANYOF_POSIXL_ZERO(ssc);
1677 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1678 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1682 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1683 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1685 /* One or the other of P1, P2 is non-empty. */
1686 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1687 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1689 ssc_union(ssc, anded_cp_list, FALSE);
1691 else { /* P1 = P2 = empty */
1692 ssc_intersection(ssc, anded_cp_list, FALSE);
1698 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1699 const regnode_charclass *or_with)
1701 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1702 * another SSC or a regular ANYOF class. Can create false positives if
1703 * 'or_with' is to be inverted. */
1708 PERL_ARGS_ASSERT_SSC_OR;
1710 assert(is_ANYOF_SYNTHETIC(ssc));
1712 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1713 * the code point inversion list and just the relevant flags */
1714 if (is_ANYOF_SYNTHETIC(or_with)) {
1715 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1716 ored_flags = ANYOF_FLAGS(or_with);
1719 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1720 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1721 if (OP(or_with) != ANYOFD) {
1723 |= ANYOF_FLAGS(or_with)
1724 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1725 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1726 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1728 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1733 ANYOF_FLAGS(ssc) |= ored_flags;
1735 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1736 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1737 * 'or_with' may be inverted. When not inverted, we have the simple
1738 * situation of computing:
1739 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1740 * If P1|P2 yields a situation with both a class and its complement are
1741 * set, like having both \w and \W, this matches all code points, and we
1742 * can delete these from the P component of the ssc going forward. XXX We
1743 * might be able to delete all the P components, but I (khw) am not certain
1744 * about this, and it is better to be safe.
1747 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1748 * <= (C1 | P1) | ~C2
1749 * <= (C1 | ~C2) | P1
1750 * (which results in actually simpler code than the non-inverted case)
1753 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1754 && ! is_ANYOF_SYNTHETIC(or_with))
1756 /* We ignore P2, leaving P1 going forward */
1757 } /* else Not inverted */
1758 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1759 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1760 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1762 for (i = 0; i < ANYOF_MAX; i += 2) {
1763 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1765 ssc_match_all_cp(ssc);
1766 ANYOF_POSIXL_CLEAR(ssc, i);
1767 ANYOF_POSIXL_CLEAR(ssc, i+1);
1775 FALSE /* Already has been inverted */
1779 PERL_STATIC_INLINE void
1780 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1782 PERL_ARGS_ASSERT_SSC_UNION;
1784 assert(is_ANYOF_SYNTHETIC(ssc));
1786 _invlist_union_maybe_complement_2nd(ssc->invlist,
1792 PERL_STATIC_INLINE void
1793 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1795 const bool invert2nd)
1797 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1799 assert(is_ANYOF_SYNTHETIC(ssc));
1801 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1807 PERL_STATIC_INLINE void
1808 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1810 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1812 assert(is_ANYOF_SYNTHETIC(ssc));
1814 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1817 PERL_STATIC_INLINE void
1818 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1820 /* AND just the single code point 'cp' into the SSC 'ssc' */
1822 SV* cp_list = _new_invlist(2);
1824 PERL_ARGS_ASSERT_SSC_CP_AND;
1826 assert(is_ANYOF_SYNTHETIC(ssc));
1828 cp_list = add_cp_to_invlist(cp_list, cp);
1829 ssc_intersection(ssc, cp_list,
1830 FALSE /* Not inverted */
1832 SvREFCNT_dec_NN(cp_list);
1835 PERL_STATIC_INLINE void
1836 S_ssc_clear_locale(regnode_ssc *ssc)
1838 /* Set the SSC 'ssc' to not match any locale things */
1839 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1841 assert(is_ANYOF_SYNTHETIC(ssc));
1843 ANYOF_POSIXL_ZERO(ssc);
1844 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1847 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1850 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1852 /* The synthetic start class is used to hopefully quickly winnow down
1853 * places where a pattern could start a match in the target string. If it
1854 * doesn't really narrow things down that much, there isn't much point to
1855 * having the overhead of using it. This function uses some very crude
1856 * heuristics to decide if to use the ssc or not.
1858 * It returns TRUE if 'ssc' rules out more than half what it considers to
1859 * be the "likely" possible matches, but of course it doesn't know what the
1860 * actual things being matched are going to be; these are only guesses
1862 * For /l matches, it assumes that the only likely matches are going to be
1863 * in the 0-255 range, uniformly distributed, so half of that is 127
1864 * For /a and /d matches, it assumes that the likely matches will be just
1865 * the ASCII range, so half of that is 63
1866 * For /u and there isn't anything matching above the Latin1 range, it
1867 * assumes that that is the only range likely to be matched, and uses
1868 * half that as the cut-off: 127. If anything matches above Latin1,
1869 * it assumes that all of Unicode could match (uniformly), except for
1870 * non-Unicode code points and things in the General Category "Other"
1871 * (unassigned, private use, surrogates, controls and formats). This
1872 * is a much large number. */
1874 U32 count = 0; /* Running total of number of code points matched by
1876 UV start, end; /* Start and end points of current range in inversion
1878 const U32 max_code_points = (LOC)
1880 : (( ! UNI_SEMANTICS
1881 || invlist_highest(ssc->invlist) < 256)
1884 const U32 max_match = max_code_points / 2;
1886 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1888 invlist_iterinit(ssc->invlist);
1889 while (invlist_iternext(ssc->invlist, &start, &end)) {
1890 if (start >= max_code_points) {
1893 end = MIN(end, max_code_points - 1);
1894 count += end - start + 1;
1895 if (count >= max_match) {
1896 invlist_iterfinish(ssc->invlist);
1906 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1908 /* The inversion list in the SSC is marked mortal; now we need a more
1909 * permanent copy, which is stored the same way that is done in a regular
1910 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1913 SV* invlist = invlist_clone(ssc->invlist);
1915 PERL_ARGS_ASSERT_SSC_FINALIZE;
1917 assert(is_ANYOF_SYNTHETIC(ssc));
1919 /* The code in this file assumes that all but these flags aren't relevant
1920 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1921 * by the time we reach here */
1922 assert(! (ANYOF_FLAGS(ssc)
1923 & ~( ANYOF_COMMON_FLAGS
1924 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1925 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
1927 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1929 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1930 NULL, NULL, NULL, FALSE);
1932 /* Make sure is clone-safe */
1933 ssc->invlist = NULL;
1935 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1936 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1939 if (RExC_contains_locale) {
1943 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1946 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1947 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1948 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1949 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1950 ? (TRIE_LIST_CUR( idx ) - 1) \
1956 dump_trie(trie,widecharmap,revcharmap)
1957 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1958 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1960 These routines dump out a trie in a somewhat readable format.
1961 The _interim_ variants are used for debugging the interim
1962 tables that are used to generate the final compressed
1963 representation which is what dump_trie expects.
1965 Part of the reason for their existence is to provide a form
1966 of documentation as to how the different representations function.
1971 Dumps the final compressed table form of the trie to Perl_debug_log.
1972 Used for debugging make_trie().
1976 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1977 AV *revcharmap, U32 depth)
1980 SV *sv=sv_newmortal();
1981 int colwidth= widecharmap ? 6 : 4;
1983 GET_RE_DEBUG_FLAGS_DECL;
1985 PERL_ARGS_ASSERT_DUMP_TRIE;
1987 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
1988 depth+1, "Match","Base","Ofs" );
1990 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1991 SV ** const tmp = av_fetch( revcharmap, state, 0);
1993 Perl_re_printf( aTHX_ "%*s",
1995 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1996 PL_colors[0], PL_colors[1],
1997 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1998 PERL_PV_ESCAPE_FIRSTCHAR
2003 Perl_re_printf( aTHX_ "\n");
2004 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2006 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2007 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2008 Perl_re_printf( aTHX_ "\n");
2010 for( state = 1 ; state < trie->statecount ; state++ ) {
2011 const U32 base = trie->states[ state ].trans.base;
2013 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2015 if ( trie->states[ state ].wordnum ) {
2016 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2018 Perl_re_printf( aTHX_ "%6s", "" );
2021 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2026 while( ( base + ofs < trie->uniquecharcount ) ||
2027 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2028 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2032 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2034 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2035 if ( ( base + ofs >= trie->uniquecharcount )
2036 && ( base + ofs - trie->uniquecharcount
2038 && trie->trans[ base + ofs
2039 - trie->uniquecharcount ].check == state )
2041 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2042 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2045 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
2049 Perl_re_printf( aTHX_ "]");
2052 Perl_re_printf( aTHX_ "\n" );
2054 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2056 for (word=1; word <= trie->wordcount; word++) {
2057 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2058 (int)word, (int)(trie->wordinfo[word].prev),
2059 (int)(trie->wordinfo[word].len));
2061 Perl_re_printf( aTHX_ "\n" );
2064 Dumps a fully constructed but uncompressed trie in list form.
2065 List tries normally only are used for construction when the number of
2066 possible chars (trie->uniquecharcount) is very high.
2067 Used for debugging make_trie().
2070 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2071 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2075 SV *sv=sv_newmortal();
2076 int colwidth= widecharmap ? 6 : 4;
2077 GET_RE_DEBUG_FLAGS_DECL;
2079 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2081 /* print out the table precompression. */
2082 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2084 Perl_re_indentf( aTHX_ "%s",
2085 depth+1, "------:-----+-----------------\n" );
2087 for( state=1 ; state < next_alloc ; state ++ ) {
2090 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2091 depth+1, (UV)state );
2092 if ( ! trie->states[ state ].wordnum ) {
2093 Perl_re_printf( aTHX_ "%5s| ","");
2095 Perl_re_printf( aTHX_ "W%4x| ",
2096 trie->states[ state ].wordnum
2099 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2100 SV ** const tmp = av_fetch( revcharmap,
2101 TRIE_LIST_ITEM(state,charid).forid, 0);
2103 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2105 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2107 PL_colors[0], PL_colors[1],
2108 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2109 | PERL_PV_ESCAPE_FIRSTCHAR
2111 TRIE_LIST_ITEM(state,charid).forid,
2112 (UV)TRIE_LIST_ITEM(state,charid).newstate
2115 Perl_re_printf( aTHX_ "\n%*s| ",
2116 (int)((depth * 2) + 14), "");
2119 Perl_re_printf( aTHX_ "\n");
2124 Dumps a fully constructed but uncompressed trie in table form.
2125 This is the normal DFA style state transition table, with a few
2126 twists to facilitate compression later.
2127 Used for debugging make_trie().
2130 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2131 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2136 SV *sv=sv_newmortal();
2137 int colwidth= widecharmap ? 6 : 4;
2138 GET_RE_DEBUG_FLAGS_DECL;
2140 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2143 print out the table precompression so that we can do a visual check
2144 that they are identical.
2147 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2149 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2150 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2152 Perl_re_printf( aTHX_ "%*s",
2154 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2155 PL_colors[0], PL_colors[1],
2156 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2157 PERL_PV_ESCAPE_FIRSTCHAR
2163 Perl_re_printf( aTHX_ "\n");
2164 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2166 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2167 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2170 Perl_re_printf( aTHX_ "\n" );
2172 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2174 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2176 (UV)TRIE_NODENUM( state ) );
2178 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2179 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2181 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2183 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2185 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2186 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2187 (UV)trie->trans[ state ].check );
2189 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2190 (UV)trie->trans[ state ].check,
2191 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2199 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2200 startbranch: the first branch in the whole branch sequence
2201 first : start branch of sequence of branch-exact nodes.
2202 May be the same as startbranch
2203 last : Thing following the last branch.
2204 May be the same as tail.
2205 tail : item following the branch sequence
2206 count : words in the sequence
2207 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2208 depth : indent depth
2210 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2212 A trie is an N'ary tree where the branches are determined by digital
2213 decomposition of the key. IE, at the root node you look up the 1st character and
2214 follow that branch repeat until you find the end of the branches. Nodes can be
2215 marked as "accepting" meaning they represent a complete word. Eg:
2219 would convert into the following structure. Numbers represent states, letters
2220 following numbers represent valid transitions on the letter from that state, if
2221 the number is in square brackets it represents an accepting state, otherwise it
2222 will be in parenthesis.
2224 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2228 (1) +-i->(6)-+-s->[7]
2230 +-s->(3)-+-h->(4)-+-e->[5]
2232 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2234 This shows that when matching against the string 'hers' we will begin at state 1
2235 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2236 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2237 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2238 single traverse. We store a mapping from accepting to state to which word was
2239 matched, and then when we have multiple possibilities we try to complete the
2240 rest of the regex in the order in which they occurred in the alternation.
2242 The only prior NFA like behaviour that would be changed by the TRIE support is
2243 the silent ignoring of duplicate alternations which are of the form:
2245 / (DUPE|DUPE) X? (?{ ... }) Y /x
2247 Thus EVAL blocks following a trie may be called a different number of times with
2248 and without the optimisation. With the optimisations dupes will be silently
2249 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2250 the following demonstrates:
2252 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2254 which prints out 'word' three times, but
2256 'words'=~/(word|word|word)(?{ print $1 })S/
2258 which doesnt print it out at all. This is due to other optimisations kicking in.
2260 Example of what happens on a structural level:
2262 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2264 1: CURLYM[1] {1,32767}(18)
2275 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2276 and should turn into:
2278 1: CURLYM[1] {1,32767}(18)
2280 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2288 Cases where tail != last would be like /(?foo|bar)baz/:
2298 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2299 and would end up looking like:
2302 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2309 d = uvchr_to_utf8_flags(d, uv, 0);
2311 is the recommended Unicode-aware way of saying
2316 #define TRIE_STORE_REVCHAR(val) \
2319 SV *zlopp = newSV(UTF8_MAXBYTES); \
2320 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2321 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2322 SvCUR_set(zlopp, kapow - flrbbbbb); \
2325 av_push(revcharmap, zlopp); \
2327 char ooooff = (char)val; \
2328 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2332 /* This gets the next character from the input, folding it if not already
2334 #define TRIE_READ_CHAR STMT_START { \
2337 /* if it is UTF then it is either already folded, or does not need \
2339 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2341 else if (folder == PL_fold_latin1) { \
2342 /* This folder implies Unicode rules, which in the range expressible \
2343 * by not UTF is the lower case, with the two exceptions, one of \
2344 * which should have been taken care of before calling this */ \
2345 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2346 uvc = toLOWER_L1(*uc); \
2347 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2350 /* raw data, will be folded later if needed */ \
2358 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2359 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2360 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2361 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2363 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2364 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2365 TRIE_LIST_CUR( state )++; \
2368 #define TRIE_LIST_NEW(state) STMT_START { \
2369 Newxz( trie->states[ state ].trans.list, \
2370 4, reg_trie_trans_le ); \
2371 TRIE_LIST_CUR( state ) = 1; \
2372 TRIE_LIST_LEN( state ) = 4; \
2375 #define TRIE_HANDLE_WORD(state) STMT_START { \
2376 U16 dupe= trie->states[ state ].wordnum; \
2377 regnode * const noper_next = regnext( noper ); \
2380 /* store the word for dumping */ \
2382 if (OP(noper) != NOTHING) \
2383 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2385 tmp = newSVpvn_utf8( "", 0, UTF ); \
2386 av_push( trie_words, tmp ); \
2390 trie->wordinfo[curword].prev = 0; \
2391 trie->wordinfo[curword].len = wordlen; \
2392 trie->wordinfo[curword].accept = state; \
2394 if ( noper_next < tail ) { \
2396 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2398 trie->jump[curword] = (U16)(noper_next - convert); \
2400 jumper = noper_next; \
2402 nextbranch= regnext(cur); \
2406 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2407 /* chain, so that when the bits of chain are later */\
2408 /* linked together, the dups appear in the chain */\
2409 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2410 trie->wordinfo[dupe].prev = curword; \
2412 /* we haven't inserted this word yet. */ \
2413 trie->states[ state ].wordnum = curword; \
2418 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2419 ( ( base + charid >= ucharcount \
2420 && base + charid < ubound \
2421 && state == trie->trans[ base - ucharcount + charid ].check \
2422 && trie->trans[ base - ucharcount + charid ].next ) \
2423 ? trie->trans[ base - ucharcount + charid ].next \
2424 : ( state==1 ? special : 0 ) \
2427 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2429 TRIE_BITMAP_SET(trie, uvc); \
2430 /* store the folded codepoint */ \
2432 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2435 /* store first byte of utf8 representation of */ \
2436 /* variant codepoints */ \
2437 if (! UVCHR_IS_INVARIANT(uvc)) { \
2438 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2443 #define MADE_JUMP_TRIE 2
2444 #define MADE_EXACT_TRIE 4
2447 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2448 regnode *first, regnode *last, regnode *tail,
2449 U32 word_count, U32 flags, U32 depth)
2451 /* first pass, loop through and scan words */
2452 reg_trie_data *trie;
2453 HV *widecharmap = NULL;
2454 AV *revcharmap = newAV();
2460 regnode *jumper = NULL;
2461 regnode *nextbranch = NULL;
2462 regnode *convert = NULL;
2463 U32 *prev_states; /* temp array mapping each state to previous one */
2464 /* we just use folder as a flag in utf8 */
2465 const U8 * folder = NULL;
2468 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2469 AV *trie_words = NULL;
2470 /* along with revcharmap, this only used during construction but both are
2471 * useful during debugging so we store them in the struct when debugging.
2474 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2475 STRLEN trie_charcount=0;
2477 SV *re_trie_maxbuff;
2478 GET_RE_DEBUG_FLAGS_DECL;
2480 PERL_ARGS_ASSERT_MAKE_TRIE;
2482 PERL_UNUSED_ARG(depth);
2486 case EXACT: case EXACTL: break;
2490 case EXACTFLU8: folder = PL_fold_latin1; break;
2491 case EXACTF: folder = PL_fold; break;
2492 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2495 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2497 trie->startstate = 1;
2498 trie->wordcount = word_count;
2499 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2500 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2501 if (flags == EXACT || flags == EXACTL)
2502 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2503 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2504 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2507 trie_words = newAV();
2510 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2511 assert(re_trie_maxbuff);
2512 if (!SvIOK(re_trie_maxbuff)) {
2513 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2515 DEBUG_TRIE_COMPILE_r({
2516 Perl_re_indentf( aTHX_
2517 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2519 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2520 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2523 /* Find the node we are going to overwrite */
2524 if ( first == startbranch && OP( last ) != BRANCH ) {
2525 /* whole branch chain */
2528 /* branch sub-chain */
2529 convert = NEXTOPER( first );
2532 /* -- First loop and Setup --
2534 We first traverse the branches and scan each word to determine if it
2535 contains widechars, and how many unique chars there are, this is
2536 important as we have to build a table with at least as many columns as we
2539 We use an array of integers to represent the character codes 0..255
2540 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2541 the native representation of the character value as the key and IV's for
2544 *TODO* If we keep track of how many times each character is used we can
2545 remap the columns so that the table compression later on is more
2546 efficient in terms of memory by ensuring the most common value is in the
2547 middle and the least common are on the outside. IMO this would be better
2548 than a most to least common mapping as theres a decent chance the most
2549 common letter will share a node with the least common, meaning the node
2550 will not be compressible. With a middle is most common approach the worst
2551 case is when we have the least common nodes twice.
2555 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2556 regnode *noper = NEXTOPER( cur );
2560 U32 wordlen = 0; /* required init */
2561 STRLEN minchars = 0;
2562 STRLEN maxchars = 0;
2563 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2566 if (OP(noper) == NOTHING) {
2567 /* skip past a NOTHING at the start of an alternation
2568 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2570 regnode *noper_next= regnext(noper);
2571 if (noper_next < tail)
2575 if ( noper < tail &&
2577 OP(noper) == flags ||
2580 OP(noper) == EXACTFU_SS
2584 uc= (U8*)STRING(noper);
2585 e= uc + STR_LEN(noper);
2592 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2593 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2594 regardless of encoding */
2595 if (OP( noper ) == EXACTFU_SS) {
2596 /* false positives are ok, so just set this */
2597 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2601 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2603 TRIE_CHARCOUNT(trie)++;
2606 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2607 * is in effect. Under /i, this character can match itself, or
2608 * anything that folds to it. If not under /i, it can match just
2609 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2610 * all fold to k, and all are single characters. But some folds
2611 * expand to more than one character, so for example LATIN SMALL
2612 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2613 * the string beginning at 'uc' is 'ffi', it could be matched by
2614 * three characters, or just by the one ligature character. (It
2615 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2616 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2617 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2618 * match.) The trie needs to know the minimum and maximum number
2619 * of characters that could match so that it can use size alone to
2620 * quickly reject many match attempts. The max is simple: it is
2621 * the number of folded characters in this branch (since a fold is
2622 * never shorter than what folds to it. */
2626 /* And the min is equal to the max if not under /i (indicated by
2627 * 'folder' being NULL), or there are no multi-character folds. If
2628 * there is a multi-character fold, the min is incremented just
2629 * once, for the character that folds to the sequence. Each
2630 * character in the sequence needs to be added to the list below of
2631 * characters in the trie, but we count only the first towards the
2632 * min number of characters needed. This is done through the
2633 * variable 'foldlen', which is returned by the macros that look
2634 * for these sequences as the number of bytes the sequence
2635 * occupies. Each time through the loop, we decrement 'foldlen' by
2636 * how many bytes the current char occupies. Only when it reaches
2637 * 0 do we increment 'minchars' or look for another multi-character
2639 if (folder == NULL) {
2642 else if (foldlen > 0) {
2643 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2648 /* See if *uc is the beginning of a multi-character fold. If
2649 * so, we decrement the length remaining to look at, to account
2650 * for the current character this iteration. (We can use 'uc'
2651 * instead of the fold returned by TRIE_READ_CHAR because for
2652 * non-UTF, the latin1_safe macro is smart enough to account
2653 * for all the unfolded characters, and because for UTF, the
2654 * string will already have been folded earlier in the
2655 * compilation process */
2657 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2658 foldlen -= UTF8SKIP(uc);
2661 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2666 /* The current character (and any potential folds) should be added
2667 * to the possible matching characters for this position in this
2671 U8 folded= folder[ (U8) uvc ];
2672 if ( !trie->charmap[ folded ] ) {
2673 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2674 TRIE_STORE_REVCHAR( folded );
2677 if ( !trie->charmap[ uvc ] ) {
2678 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2679 TRIE_STORE_REVCHAR( uvc );
2682 /* store the codepoint in the bitmap, and its folded
2684 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2685 set_bit = 0; /* We've done our bit :-) */
2689 /* XXX We could come up with the list of code points that fold
2690 * to this using PL_utf8_foldclosures, except not for
2691 * multi-char folds, as there may be multiple combinations
2692 * there that could work, which needs to wait until runtime to
2693 * resolve (The comment about LIGATURE FFI above is such an
2698 widecharmap = newHV();
2700 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2703 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2705 if ( !SvTRUE( *svpp ) ) {
2706 sv_setiv( *svpp, ++trie->uniquecharcount );
2707 TRIE_STORE_REVCHAR(uvc);
2710 } /* end loop through characters in this branch of the trie */
2712 /* We take the min and max for this branch and combine to find the min
2713 * and max for all branches processed so far */
2714 if( cur == first ) {
2715 trie->minlen = minchars;
2716 trie->maxlen = maxchars;
2717 } else if (minchars < trie->minlen) {
2718 trie->minlen = minchars;
2719 } else if (maxchars > trie->maxlen) {
2720 trie->maxlen = maxchars;
2722 } /* end first pass */
2723 DEBUG_TRIE_COMPILE_r(
2724 Perl_re_indentf( aTHX_
2725 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2727 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2728 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2729 (int)trie->minlen, (int)trie->maxlen )
2733 We now know what we are dealing with in terms of unique chars and
2734 string sizes so we can calculate how much memory a naive
2735 representation using a flat table will take. If it's over a reasonable
2736 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2737 conservative but potentially much slower representation using an array
2740 At the end we convert both representations into the same compressed
2741 form that will be used in regexec.c for matching with. The latter
2742 is a form that cannot be used to construct with but has memory
2743 properties similar to the list form and access properties similar
2744 to the table form making it both suitable for fast searches and
2745 small enough that its feasable to store for the duration of a program.
2747 See the comment in the code where the compressed table is produced
2748 inplace from the flat tabe representation for an explanation of how
2749 the compression works.
2754 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2757 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2758 > SvIV(re_trie_maxbuff) )
2761 Second Pass -- Array Of Lists Representation
2763 Each state will be represented by a list of charid:state records
2764 (reg_trie_trans_le) the first such element holds the CUR and LEN
2765 points of the allocated array. (See defines above).
2767 We build the initial structure using the lists, and then convert
2768 it into the compressed table form which allows faster lookups
2769 (but cant be modified once converted).
2772 STRLEN transcount = 1;
2774 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2777 trie->states = (reg_trie_state *)
2778 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2779 sizeof(reg_trie_state) );
2783 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2785 regnode *noper = NEXTOPER( cur );
2786 U32 state = 1; /* required init */
2787 U16 charid = 0; /* sanity init */
2788 U32 wordlen = 0; /* required init */
2790 if (OP(noper) == NOTHING) {
2791 regnode *noper_next= regnext(noper);
2792 if (noper_next < tail)
2796 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2797 const U8 *uc= (U8*)STRING(noper);
2798 const U8 *e= uc + STR_LEN(noper);
2800 for ( ; uc < e ; uc += len ) {
2805 charid = trie->charmap[ uvc ];
2807 SV** const svpp = hv_fetch( widecharmap,
2814 charid=(U16)SvIV( *svpp );
2817 /* charid is now 0 if we dont know the char read, or
2818 * nonzero if we do */
2825 if ( !trie->states[ state ].trans.list ) {
2826 TRIE_LIST_NEW( state );
2829 check <= TRIE_LIST_USED( state );
2832 if ( TRIE_LIST_ITEM( state, check ).forid
2835 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2840 newstate = next_alloc++;
2841 prev_states[newstate] = state;
2842 TRIE_LIST_PUSH( state, charid, newstate );
2847 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
2851 TRIE_HANDLE_WORD(state);
2853 } /* end second pass */
2855 /* next alloc is the NEXT state to be allocated */
2856 trie->statecount = next_alloc;
2857 trie->states = (reg_trie_state *)
2858 PerlMemShared_realloc( trie->states,
2860 * sizeof(reg_trie_state) );
2862 /* and now dump it out before we compress it */
2863 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2864 revcharmap, next_alloc,
2868 trie->trans = (reg_trie_trans *)
2869 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2876 for( state=1 ; state < next_alloc ; state ++ ) {
2880 DEBUG_TRIE_COMPILE_MORE_r(
2881 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
2885 if (trie->states[state].trans.list) {
2886 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2890 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2891 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2892 if ( forid < minid ) {
2894 } else if ( forid > maxid ) {
2898 if ( transcount < tp + maxid - minid + 1) {
2900 trie->trans = (reg_trie_trans *)
2901 PerlMemShared_realloc( trie->trans,
2903 * sizeof(reg_trie_trans) );
2904 Zero( trie->trans + (transcount / 2),
2908 base = trie->uniquecharcount + tp - minid;
2909 if ( maxid == minid ) {
2911 for ( ; zp < tp ; zp++ ) {
2912 if ( ! trie->trans[ zp ].next ) {
2913 base = trie->uniquecharcount + zp - minid;
2914 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2916 trie->trans[ zp ].check = state;
2922 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2924 trie->trans[ tp ].check = state;
2929 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2930 const U32 tid = base
2931 - trie->uniquecharcount
2932 + TRIE_LIST_ITEM( state, idx ).forid;
2933 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2935 trie->trans[ tid ].check = state;
2937 tp += ( maxid - minid + 1 );
2939 Safefree(trie->states[ state ].trans.list);
2942 DEBUG_TRIE_COMPILE_MORE_r(
2943 Perl_re_printf( aTHX_ " base: %d\n",base);
2946 trie->states[ state ].trans.base=base;
2948 trie->lasttrans = tp + 1;
2952 Second Pass -- Flat Table Representation.
2954 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2955 each. We know that we will need Charcount+1 trans at most to store
2956 the data (one row per char at worst case) So we preallocate both
2957 structures assuming worst case.
2959 We then construct the trie using only the .next slots of the entry
2962 We use the .check field of the first entry of the node temporarily
2963 to make compression both faster and easier by keeping track of how
2964 many non zero fields are in the node.
2966 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2969 There are two terms at use here: state as a TRIE_NODEIDX() which is
2970 a number representing the first entry of the node, and state as a
2971 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2972 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2973 if there are 2 entrys per node. eg:
2981 The table is internally in the right hand, idx form. However as we
2982 also have to deal with the states array which is indexed by nodenum
2983 we have to use TRIE_NODENUM() to convert.
2986 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
2989 trie->trans = (reg_trie_trans *)
2990 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2991 * trie->uniquecharcount + 1,
2992 sizeof(reg_trie_trans) );
2993 trie->states = (reg_trie_state *)
2994 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2995 sizeof(reg_trie_state) );
2996 next_alloc = trie->uniquecharcount + 1;
2999 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3001 regnode *noper = NEXTOPER( cur );
3003 U32 state = 1; /* required init */
3005 U16 charid = 0; /* sanity init */
3006 U32 accept_state = 0; /* sanity init */
3008 U32 wordlen = 0; /* required init */
3010 if (OP(noper) == NOTHING) {
3011 regnode *noper_next= regnext(noper);
3012 if (noper_next < tail)
3016 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3017 const U8 *uc= (U8*)STRING(noper);
3018 const U8 *e= uc + STR_LEN(noper);
3020 for ( ; uc < e ; uc += len ) {
3025 charid = trie->charmap[ uvc ];
3027 SV* const * const svpp = hv_fetch( widecharmap,
3031 charid = svpp ? (U16)SvIV(*svpp) : 0;
3035 if ( !trie->trans[ state + charid ].next ) {
3036 trie->trans[ state + charid ].next = next_alloc;
3037 trie->trans[ state ].check++;
3038 prev_states[TRIE_NODENUM(next_alloc)]
3039 = TRIE_NODENUM(state);
3040 next_alloc += trie->uniquecharcount;
3042 state = trie->trans[ state + charid ].next;
3044 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3046 /* charid is now 0 if we dont know the char read, or
3047 * nonzero if we do */
3050 accept_state = TRIE_NODENUM( state );
3051 TRIE_HANDLE_WORD(accept_state);
3053 } /* end second pass */
3055 /* and now dump it out before we compress it */
3056 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3058 next_alloc, depth+1));
3062 * Inplace compress the table.*
3064 For sparse data sets the table constructed by the trie algorithm will
3065 be mostly 0/FAIL transitions or to put it another way mostly empty.
3066 (Note that leaf nodes will not contain any transitions.)
3068 This algorithm compresses the tables by eliminating most such
3069 transitions, at the cost of a modest bit of extra work during lookup:
3071 - Each states[] entry contains a .base field which indicates the
3072 index in the state[] array wheres its transition data is stored.
3074 - If .base is 0 there are no valid transitions from that node.
3076 - If .base is nonzero then charid is added to it to find an entry in
3079 -If trans[states[state].base+charid].check!=state then the
3080 transition is taken to be a 0/Fail transition. Thus if there are fail
3081 transitions at the front of the node then the .base offset will point
3082 somewhere inside the previous nodes data (or maybe even into a node
3083 even earlier), but the .check field determines if the transition is
3087 The following process inplace converts the table to the compressed
3088 table: We first do not compress the root node 1,and mark all its
3089 .check pointers as 1 and set its .base pointer as 1 as well. This
3090 allows us to do a DFA construction from the compressed table later,
3091 and ensures that any .base pointers we calculate later are greater
3094 - We set 'pos' to indicate the first entry of the second node.
3096 - We then iterate over the columns of the node, finding the first and
3097 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3098 and set the .check pointers accordingly, and advance pos
3099 appropriately and repreat for the next node. Note that when we copy
3100 the next pointers we have to convert them from the original
3101 NODEIDX form to NODENUM form as the former is not valid post
3104 - If a node has no transitions used we mark its base as 0 and do not
3105 advance the pos pointer.
3107 - If a node only has one transition we use a second pointer into the
3108 structure to fill in allocated fail transitions from other states.
3109 This pointer is independent of the main pointer and scans forward
3110 looking for null transitions that are allocated to a state. When it
3111 finds one it writes the single transition into the "hole". If the
3112 pointer doesnt find one the single transition is appended as normal.
3114 - Once compressed we can Renew/realloc the structures to release the
3117 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3118 specifically Fig 3.47 and the associated pseudocode.
3122 const U32 laststate = TRIE_NODENUM( next_alloc );
3125 trie->statecount = laststate;
3127 for ( state = 1 ; state < laststate ; state++ ) {
3129 const U32 stateidx = TRIE_NODEIDX( state );
3130 const U32 o_used = trie->trans[ stateidx ].check;
3131 U32 used = trie->trans[ stateidx ].check;
3132 trie->trans[ stateidx ].check = 0;
3135 used && charid < trie->uniquecharcount;
3138 if ( flag || trie->trans[ stateidx + charid ].next ) {
3139 if ( trie->trans[ stateidx + charid ].next ) {
3141 for ( ; zp < pos ; zp++ ) {
3142 if ( ! trie->trans[ zp ].next ) {
3146 trie->states[ state ].trans.base
3148 + trie->uniquecharcount
3150 trie->trans[ zp ].next
3151 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3153 trie->trans[ zp ].check = state;
3154 if ( ++zp > pos ) pos = zp;
3161 trie->states[ state ].trans.base
3162 = pos + trie->uniquecharcount - charid ;
3164 trie->trans[ pos ].next
3165 = SAFE_TRIE_NODENUM(
3166 trie->trans[ stateidx + charid ].next );
3167 trie->trans[ pos ].check = state;
3172 trie->lasttrans = pos + 1;
3173 trie->states = (reg_trie_state *)
3174 PerlMemShared_realloc( trie->states, laststate
3175 * sizeof(reg_trie_state) );
3176 DEBUG_TRIE_COMPILE_MORE_r(
3177 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3179 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3183 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3186 } /* end table compress */
3188 DEBUG_TRIE_COMPILE_MORE_r(
3189 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3191 (UV)trie->statecount,
3192 (UV)trie->lasttrans)
3194 /* resize the trans array to remove unused space */
3195 trie->trans = (reg_trie_trans *)
3196 PerlMemShared_realloc( trie->trans, trie->lasttrans
3197 * sizeof(reg_trie_trans) );
3199 { /* Modify the program and insert the new TRIE node */
3200 U8 nodetype =(U8)(flags & 0xFF);
3204 regnode *optimize = NULL;
3205 #ifdef RE_TRACK_PATTERN_OFFSETS
3208 U32 mjd_nodelen = 0;
3209 #endif /* RE_TRACK_PATTERN_OFFSETS */
3210 #endif /* DEBUGGING */
3212 This means we convert either the first branch or the first Exact,
3213 depending on whether the thing following (in 'last') is a branch
3214 or not and whther first is the startbranch (ie is it a sub part of
3215 the alternation or is it the whole thing.)
3216 Assuming its a sub part we convert the EXACT otherwise we convert
3217 the whole branch sequence, including the first.
3219 /* Find the node we are going to overwrite */
3220 if ( first != startbranch || OP( last ) == BRANCH ) {
3221 /* branch sub-chain */
3222 NEXT_OFF( first ) = (U16)(last - first);
3223 #ifdef RE_TRACK_PATTERN_OFFSETS
3225 mjd_offset= Node_Offset((convert));
3226 mjd_nodelen= Node_Length((convert));
3229 /* whole branch chain */
3231 #ifdef RE_TRACK_PATTERN_OFFSETS
3234 const regnode *nop = NEXTOPER( convert );
3235 mjd_offset= Node_Offset((nop));
3236 mjd_nodelen= Node_Length((nop));
3240 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3242 (UV)mjd_offset, (UV)mjd_nodelen)
3245 /* But first we check to see if there is a common prefix we can
3246 split out as an EXACT and put in front of the TRIE node. */
3247 trie->startstate= 1;
3248 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3249 /* we want to find the first state that has more than
3250 * one transition, if that state is not the first state
3251 * then we have a common prefix which we can remove.
3254 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3256 I32 first_ofs = -1; /* keeps track of the ofs of the first
3257 transition, -1 means none */
3259 const U32 base = trie->states[ state ].trans.base;
3261 /* does this state terminate an alternation? */
3262 if ( trie->states[state].wordnum )
3265 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3266 if ( ( base + ofs >= trie->uniquecharcount ) &&
3267 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3268 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3270 if ( ++count > 1 ) {
3271 /* we have more than one transition */
3274 /* if this is the first state there is no common prefix
3275 * to extract, so we can exit */
3276 if ( state == 1 ) break;
3277 tmp = av_fetch( revcharmap, ofs, 0);
3278 ch = (U8*)SvPV_nolen_const( *tmp );
3280 /* if we are on count 2 then we need to initialize the
3281 * bitmap, and store the previous char if there was one
3284 /* clear the bitmap */
3285 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3287 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3290 if (first_ofs >= 0) {
3291 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3292 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3294 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3296 Perl_re_printf( aTHX_ "%s", (char*)ch)
3300 /* store the current firstchar in the bitmap */
3301 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3302 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3308 /* This state has only one transition, its transition is part
3309 * of a common prefix - we need to concatenate the char it
3310 * represents to what we have so far. */
3311 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3313 char *ch = SvPV( *tmp, len );
3315 SV *sv=sv_newmortal();
3316 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3318 (UV)state, (UV)first_ofs,
3319 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3320 PL_colors[0], PL_colors[1],
3321 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3322 PERL_PV_ESCAPE_FIRSTCHAR
3327 OP( convert ) = nodetype;
3328 str=STRING(convert);
3331 STR_LEN(convert) += len;
3337 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3342 trie->prefixlen = (state-1);
3344 regnode *n = convert+NODE_SZ_STR(convert);
3345 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3346 trie->startstate = state;
3347 trie->minlen -= (state - 1);
3348 trie->maxlen -= (state - 1);
3350 /* At least the UNICOS C compiler choked on this
3351 * being argument to DEBUG_r(), so let's just have
3354 #ifdef PERL_EXT_RE_BUILD
3360 regnode *fix = convert;
3361 U32 word = trie->wordcount;
3363 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3364 while( ++fix < n ) {
3365 Set_Node_Offset_Length(fix, 0, 0);
3368 SV ** const tmp = av_fetch( trie_words, word, 0 );
3370 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3371 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3373 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3381 NEXT_OFF(convert) = (U16)(tail - convert);
3382 DEBUG_r(optimize= n);
3388 if ( trie->maxlen ) {
3389 NEXT_OFF( convert ) = (U16)(tail - convert);
3390 ARG_SET( convert, data_slot );
3391 /* Store the offset to the first unabsorbed branch in
3392 jump[0], which is otherwise unused by the jump logic.
3393 We use this when dumping a trie and during optimisation. */
3395 trie->jump[0] = (U16)(nextbranch - convert);
3397 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3398 * and there is a bitmap
3399 * and the first "jump target" node we found leaves enough room
3400 * then convert the TRIE node into a TRIEC node, with the bitmap
3401 * embedded inline in the opcode - this is hypothetically faster.
3403 if ( !trie->states[trie->startstate].wordnum
3405 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3407 OP( convert ) = TRIEC;
3408 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3409 PerlMemShared_free(trie->bitmap);
3412 OP( convert ) = TRIE;
3414 /* store the type in the flags */
3415 convert->flags = nodetype;
3419 + regarglen[ OP( convert ) ];
3421 /* XXX We really should free up the resource in trie now,
3422 as we won't use them - (which resources?) dmq */
3424 /* needed for dumping*/
3425 DEBUG_r(if (optimize) {
3426 regnode *opt = convert;
3428 while ( ++opt < optimize) {
3429 Set_Node_Offset_Length(opt,0,0);
3432 Try to clean up some of the debris left after the
3435 while( optimize < jumper ) {
3436 mjd_nodelen += Node_Length((optimize));
3437 OP( optimize ) = OPTIMIZED;
3438 Set_Node_Offset_Length(optimize,0,0);
3441 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3443 } /* end node insert */
3445 /* Finish populating the prev field of the wordinfo array. Walk back
3446 * from each accept state until we find another accept state, and if
3447 * so, point the first word's .prev field at the second word. If the
3448 * second already has a .prev field set, stop now. This will be the
3449 * case either if we've already processed that word's accept state,
3450 * or that state had multiple words, and the overspill words were
3451 * already linked up earlier.
3458 for (word=1; word <= trie->wordcount; word++) {
3460 if (trie->wordinfo[word].prev)
3462 state = trie->wordinfo[word].accept;
3464 state = prev_states[state];
3467 prev = trie->states[state].wordnum;
3471 trie->wordinfo[word].prev = prev;
3473 Safefree(prev_states);
3477 /* and now dump out the compressed format */
3478 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3480 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3482 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3483 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3485 SvREFCNT_dec_NN(revcharmap);
3489 : trie->startstate>1
3495 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3497 /* The Trie is constructed and compressed now so we can build a fail array if
3500 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3502 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3506 We find the fail state for each state in the trie, this state is the longest
3507 proper suffix of the current state's 'word' that is also a proper prefix of
3508 another word in our trie. State 1 represents the word '' and is thus the
3509 default fail state. This allows the DFA not to have to restart after its
3510 tried and failed a word at a given point, it simply continues as though it
3511 had been matching the other word in the first place.
3513 'abcdgu'=~/abcdefg|cdgu/
3514 When we get to 'd' we are still matching the first word, we would encounter
3515 'g' which would fail, which would bring us to the state representing 'd' in
3516 the second word where we would try 'g' and succeed, proceeding to match
3519 /* add a fail transition */
3520 const U32 trie_offset = ARG(source);
3521 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3523 const U32 ucharcount = trie->uniquecharcount;
3524 const U32 numstates = trie->statecount;
3525 const U32 ubound = trie->lasttrans + ucharcount;
3529 U32 base = trie->states[ 1 ].trans.base;
3532 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3534 GET_RE_DEBUG_FLAGS_DECL;
3536 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3537 PERL_UNUSED_CONTEXT;
3539 PERL_UNUSED_ARG(depth);
3542 if ( OP(source) == TRIE ) {
3543 struct regnode_1 *op = (struct regnode_1 *)
3544 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3545 StructCopy(source,op,struct regnode_1);
3546 stclass = (regnode *)op;
3548 struct regnode_charclass *op = (struct regnode_charclass *)
3549 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3550 StructCopy(source,op,struct regnode_charclass);
3551 stclass = (regnode *)op;
3553 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3555 ARG_SET( stclass, data_slot );
3556 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3557 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3558 aho->trie=trie_offset;
3559 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3560 Copy( trie->states, aho->states, numstates, reg_trie_state );
3561 Newxz( q, numstates, U32);
3562 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3565 /* initialize fail[0..1] to be 1 so that we always have
3566 a valid final fail state */
3567 fail[ 0 ] = fail[ 1 ] = 1;
3569 for ( charid = 0; charid < ucharcount ; charid++ ) {
3570 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3572 q[ q_write ] = newstate;
3573 /* set to point at the root */
3574 fail[ q[ q_write++ ] ]=1;
3577 while ( q_read < q_write) {
3578 const U32 cur = q[ q_read++ % numstates ];
3579 base = trie->states[ cur ].trans.base;
3581 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3582 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3584 U32 fail_state = cur;
3587 fail_state = fail[ fail_state ];
3588 fail_base = aho->states[ fail_state ].trans.base;
3589 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3591 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3592 fail[ ch_state ] = fail_state;
3593 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3595 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3597 q[ q_write++ % numstates] = ch_state;
3601 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3602 when we fail in state 1, this allows us to use the
3603 charclass scan to find a valid start char. This is based on the principle
3604 that theres a good chance the string being searched contains lots of stuff
3605 that cant be a start char.
3607 fail[ 0 ] = fail[ 1 ] = 0;
3608 DEBUG_TRIE_COMPILE_r({
3609 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3610 depth, (UV)numstates
3612 for( q_read=1; q_read<numstates; q_read++ ) {
3613 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3615 Perl_re_printf( aTHX_ "\n");
3618 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3623 #define DEBUG_PEEP(str,scan,depth) \
3624 DEBUG_OPTIMISE_r({if (scan){ \
3625 regnode *Next = regnext(scan); \
3626 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);\
3627 Perl_re_indentf( aTHX_ "" str ">%3d: %s (%d)", \
3628 depth, REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3629 Next ? (REG_NODE_NUM(Next)) : 0 );\
3630 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3631 Perl_re_printf( aTHX_ "\n"); \
3634 /* The below joins as many adjacent EXACTish nodes as possible into a single
3635 * one. The regop may be changed if the node(s) contain certain sequences that
3636 * require special handling. The joining is only done if:
3637 * 1) there is room in the current conglomerated node to entirely contain the
3639 * 2) they are the exact same node type
3641 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3642 * these get optimized out
3644 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3645 * as possible, even if that means splitting an existing node so that its first
3646 * part is moved to the preceeding node. This would maximise the efficiency of
3647 * memEQ during matching. Elsewhere in this file, khw proposes splitting
3648 * EXACTFish nodes into portions that don't change under folding vs those that
3649 * do. Those portions that don't change may be the only things in the pattern that
3650 * could be used to find fixed and floating strings.
3652 * If a node is to match under /i (folded), the number of characters it matches
3653 * can be different than its character length if it contains a multi-character
3654 * fold. *min_subtract is set to the total delta number of characters of the
3657 * And *unfolded_multi_char is set to indicate whether or not the node contains
3658 * an unfolded multi-char fold. This happens when whether the fold is valid or
3659 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3660 * SMALL LETTER SHARP S, as only if the target string being matched against
3661 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3662 * folding rules depend on the locale in force at runtime. (Multi-char folds
3663 * whose components are all above the Latin1 range are not run-time locale
3664 * dependent, and have already been folded by the time this function is
3667 * This is as good a place as any to discuss the design of handling these
3668 * multi-character fold sequences. It's been wrong in Perl for a very long
3669 * time. There are three code points in Unicode whose multi-character folds
3670 * were long ago discovered to mess things up. The previous designs for
3671 * dealing with these involved assigning a special node for them. This
3672 * approach doesn't always work, as evidenced by this example:
3673 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3674 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3675 * would match just the \xDF, it won't be able to handle the case where a
3676 * successful match would have to cross the node's boundary. The new approach
3677 * that hopefully generally solves the problem generates an EXACTFU_SS node
3678 * that is "sss" in this case.
3680 * It turns out that there are problems with all multi-character folds, and not
3681 * just these three. Now the code is general, for all such cases. The
3682 * approach taken is:
3683 * 1) This routine examines each EXACTFish node that could contain multi-
3684 * character folded sequences. Since a single character can fold into
3685 * such a sequence, the minimum match length for this node is less than
3686 * the number of characters in the node. This routine returns in
3687 * *min_subtract how many characters to subtract from the the actual
3688 * length of the string to get a real minimum match length; it is 0 if
3689 * there are no multi-char foldeds. This delta is used by the caller to
3690 * adjust the min length of the match, and the delta between min and max,
3691 * so that the optimizer doesn't reject these possibilities based on size
3693 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3694 * is used for an EXACTFU node that contains at least one "ss" sequence in
3695 * it. For non-UTF-8 patterns and strings, this is the only case where
3696 * there is a possible fold length change. That means that a regular
3697 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3698 * with length changes, and so can be processed faster. regexec.c takes
3699 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3700 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3701 * known until runtime). This saves effort in regex matching. However,
3702 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3703 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3704 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3705 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3706 * possibilities for the non-UTF8 patterns are quite simple, except for
3707 * the sharp s. All the ones that don't involve a UTF-8 target string are
3708 * members of a fold-pair, and arrays are set up for all of them so that
3709 * the other member of the pair can be found quickly. Code elsewhere in
3710 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3711 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3712 * described in the next item.
3713 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3714 * validity of the fold won't be known until runtime, and so must remain
3715 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3716 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3717 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3718 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3719 * The reason this is a problem is that the optimizer part of regexec.c
3720 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3721 * that a character in the pattern corresponds to at most a single
3722 * character in the target string. (And I do mean character, and not byte
3723 * here, unlike other parts of the documentation that have never been
3724 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3725 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3726 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3727 * nodes, violate the assumption, and they are the only instances where it
3728 * is violated. I'm reluctant to try to change the assumption, as the
3729 * code involved is impenetrable to me (khw), so instead the code here
3730 * punts. This routine examines EXACTFL nodes, and (when the pattern
3731 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3732 * boolean indicating whether or not the node contains such a fold. When
3733 * it is true, the caller sets a flag that later causes the optimizer in
3734 * this file to not set values for the floating and fixed string lengths,
3735 * and thus avoids the optimizer code in regexec.c that makes the invalid
3736 * assumption. Thus, there is no optimization based on string lengths for
3737 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3738 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3739 * assumption is wrong only in these cases is that all other non-UTF-8
3740 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3741 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3742 * EXACTF nodes because we don't know at compile time if it actually
3743 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3744 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3745 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3746 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3747 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3748 * string would require the pattern to be forced into UTF-8, the overhead
3749 * of which we want to avoid. Similarly the unfolded multi-char folds in
3750 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3753 * Similarly, the code that generates tries doesn't currently handle
3754 * not-already-folded multi-char folds, and it looks like a pain to change
3755 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3756 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3757 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3758 * using /iaa matching will be doing so almost entirely with ASCII
3759 * strings, so this should rarely be encountered in practice */
3761 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3762 if (PL_regkind[OP(scan)] == EXACT) \
3763 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3766 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3767 UV *min_subtract, bool *unfolded_multi_char,
3768 U32 flags,regnode *val, U32 depth)
3770 /* Merge several consecutive EXACTish nodes into one. */
3771 regnode *n = regnext(scan);
3773 regnode *next = scan + NODE_SZ_STR(scan);
3777 regnode *stop = scan;
3778 GET_RE_DEBUG_FLAGS_DECL;
3780 PERL_UNUSED_ARG(depth);
3783 PERL_ARGS_ASSERT_JOIN_EXACT;
3784 #ifndef EXPERIMENTAL_INPLACESCAN
3785 PERL_UNUSED_ARG(flags);
3786 PERL_UNUSED_ARG(val);
3788 DEBUG_PEEP("join",scan,depth);
3790 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3791 * EXACT ones that are mergeable to the current one. */
3793 && (PL_regkind[OP(n)] == NOTHING
3794 || (stringok && OP(n) == OP(scan)))
3796 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3799 if (OP(n) == TAIL || n > next)
3801 if (PL_regkind[OP(n)] == NOTHING) {
3802 DEBUG_PEEP("skip:",n,depth);
3803 NEXT_OFF(scan) += NEXT_OFF(n);
3804 next = n + NODE_STEP_REGNODE;
3811 else if (stringok) {
3812 const unsigned int oldl = STR_LEN(scan);
3813 regnode * const nnext = regnext(n);
3815 /* XXX I (khw) kind of doubt that this works on platforms (should
3816 * Perl ever run on one) where U8_MAX is above 255 because of lots
3817 * of other assumptions */
3818 /* Don't join if the sum can't fit into a single node */
3819 if (oldl + STR_LEN(n) > U8_MAX)
3822 DEBUG_PEEP("merg",n,depth);
3825 NEXT_OFF(scan) += NEXT_OFF(n);
3826 STR_LEN(scan) += STR_LEN(n);
3827 next = n + NODE_SZ_STR(n);
3828 /* Now we can overwrite *n : */
3829 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3837 #ifdef EXPERIMENTAL_INPLACESCAN
3838 if (flags && !NEXT_OFF(n)) {
3839 DEBUG_PEEP("atch", val, depth);
3840 if (reg_off_by_arg[OP(n)]) {
3841 ARG_SET(n, val - n);
3844 NEXT_OFF(n) = val - n;
3852 *unfolded_multi_char = FALSE;
3854 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3855 * can now analyze for sequences of problematic code points. (Prior to
3856 * this final joining, sequences could have been split over boundaries, and
3857 * hence missed). The sequences only happen in folding, hence for any
3858 * non-EXACT EXACTish node */
3859 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3860 U8* s0 = (U8*) STRING(scan);
3862 U8* s_end = s0 + STR_LEN(scan);
3864 int total_count_delta = 0; /* Total delta number of characters that
3865 multi-char folds expand to */
3867 /* One pass is made over the node's string looking for all the
3868 * possibilities. To avoid some tests in the loop, there are two main
3869 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3874 if (OP(scan) == EXACTFL) {
3877 /* An EXACTFL node would already have been changed to another
3878 * node type unless there is at least one character in it that
3879 * is problematic; likely a character whose fold definition
3880 * won't be known until runtime, and so has yet to be folded.
3881 * For all but the UTF-8 locale, folds are 1-1 in length, but
3882 * to handle the UTF-8 case, we need to create a temporary
3883 * folded copy using UTF-8 locale rules in order to analyze it.
3884 * This is because our macros that look to see if a sequence is
3885 * a multi-char fold assume everything is folded (otherwise the
3886 * tests in those macros would be too complicated and slow).
3887 * Note that here, the non-problematic folds will have already
3888 * been done, so we can just copy such characters. We actually
3889 * don't completely fold the EXACTFL string. We skip the
3890 * unfolded multi-char folds, as that would just create work
3891 * below to figure out the size they already are */
3893 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3896 STRLEN s_len = UTF8SKIP(s);
3897 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3898 Copy(s, d, s_len, U8);
3901 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3902 *unfolded_multi_char = TRUE;
3903 Copy(s, d, s_len, U8);
3906 else if (isASCII(*s)) {
3907 *(d++) = toFOLD(*s);
3911 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
3917 /* Point the remainder of the routine to look at our temporary
3921 } /* End of creating folded copy of EXACTFL string */
3923 /* Examine the string for a multi-character fold sequence. UTF-8
3924 * patterns have all characters pre-folded by the time this code is
3926 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3927 length sequence we are looking for is 2 */
3929 int count = 0; /* How many characters in a multi-char fold */
3930 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3931 if (! len) { /* Not a multi-char fold: get next char */
3936 /* Nodes with 'ss' require special handling, except for
3937 * EXACTFA-ish for which there is no multi-char fold to this */
3938 if (len == 2 && *s == 's' && *(s+1) == 's'
3939 && OP(scan) != EXACTFA
3940 && OP(scan) != EXACTFA_NO_TRIE)
3943 if (OP(scan) != EXACTFL) {
3944 OP(scan) = EXACTFU_SS;
3948 else { /* Here is a generic multi-char fold. */
3949 U8* multi_end = s + len;
3951 /* Count how many characters are in it. In the case of
3952 * /aa, no folds which contain ASCII code points are
3953 * allowed, so check for those, and skip if found. */
3954 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3955 count = utf8_length(s, multi_end);
3959 while (s < multi_end) {
3962 goto next_iteration;
3972 /* The delta is how long the sequence is minus 1 (1 is how long
3973 * the character that folds to the sequence is) */
3974 total_count_delta += count - 1;
3978 /* We created a temporary folded copy of the string in EXACTFL
3979 * nodes. Therefore we need to be sure it doesn't go below zero,
3980 * as the real string could be shorter */
3981 if (OP(scan) == EXACTFL) {
3982 int total_chars = utf8_length((U8*) STRING(scan),
3983 (U8*) STRING(scan) + STR_LEN(scan));
3984 if (total_count_delta > total_chars) {
3985 total_count_delta = total_chars;
3989 *min_subtract += total_count_delta;
3992 else if (OP(scan) == EXACTFA) {
3994 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3995 * fold to the ASCII range (and there are no existing ones in the
3996 * upper latin1 range). But, as outlined in the comments preceding
3997 * this function, we need to flag any occurrences of the sharp s.
3998 * This character forbids trie formation (because of added
4000 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4001 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4002 || UNICODE_DOT_DOT_VERSION > 0)
4004 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4005 OP(scan) = EXACTFA_NO_TRIE;
4006 *unfolded_multi_char = TRUE;
4014 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
4015 * folds that are all Latin1. As explained in the comments
4016 * preceding this function, we look also for the sharp s in EXACTF
4017 * and EXACTFL nodes; it can be in the final position. Otherwise
4018 * we can stop looking 1 byte earlier because have to find at least
4019 * two characters for a multi-fold */
4020 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4025 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4026 if (! len) { /* Not a multi-char fold. */
4027 if (*s == LATIN_SMALL_LETTER_SHARP_S
4028 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4030 *unfolded_multi_char = TRUE;
4037 && isALPHA_FOLD_EQ(*s, 's')
4038 && isALPHA_FOLD_EQ(*(s+1), 's'))
4041 /* EXACTF nodes need to know that the minimum length
4042 * changed so that a sharp s in the string can match this
4043 * ss in the pattern, but they remain EXACTF nodes, as they
4044 * won't match this unless the target string is is UTF-8,
4045 * which we don't know until runtime. EXACTFL nodes can't
4046 * transform into EXACTFU nodes */
4047 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4048 OP(scan) = EXACTFU_SS;
4052 *min_subtract += len - 1;
4060 /* Allow dumping but overwriting the collection of skipped
4061 * ops and/or strings with fake optimized ops */
4062 n = scan + NODE_SZ_STR(scan);
4070 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
4074 /* REx optimizer. Converts nodes into quicker variants "in place".
4075 Finds fixed substrings. */
4077 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4078 to the position after last scanned or to NULL. */
4080 #define INIT_AND_WITHP \
4081 assert(!and_withp); \
4082 Newx(and_withp,1, regnode_ssc); \
4083 SAVEFREEPV(and_withp)
4087 S_unwind_scan_frames(pTHX_ const void *p)
4089 scan_frame *f= (scan_frame *)p;
4091 scan_frame *n= f->next_frame;
4099 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4100 SSize_t *minlenp, SSize_t *deltap,
4105 regnode_ssc *and_withp,
4106 U32 flags, U32 depth)
4107 /* scanp: Start here (read-write). */
4108 /* deltap: Write maxlen-minlen here. */
4109 /* last: Stop before this one. */
4110 /* data: string data about the pattern */
4111 /* stopparen: treat close N as END */
4112 /* recursed: which subroutines have we recursed into */
4113 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4115 /* There must be at least this number of characters to match */
4118 regnode *scan = *scanp, *next;
4120 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4121 int is_inf_internal = 0; /* The studied chunk is infinite */
4122 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4123 scan_data_t data_fake;
4124 SV *re_trie_maxbuff = NULL;
4125 regnode *first_non_open = scan;
4126 SSize_t stopmin = SSize_t_MAX;
4127 scan_frame *frame = NULL;
4128 GET_RE_DEBUG_FLAGS_DECL;
4130 PERL_ARGS_ASSERT_STUDY_CHUNK;
4131 RExC_study_started= 1;
4135 while (first_non_open && OP(first_non_open) == OPEN)
4136 first_non_open=regnext(first_non_open);
4142 RExC_study_chunk_recursed_count++;
4144 DEBUG_OPTIMISE_MORE_r(
4146 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4147 depth, (long)stopparen,
4148 (unsigned long)RExC_study_chunk_recursed_count,
4149 (unsigned long)depth, (unsigned long)recursed_depth,
4152 if (recursed_depth) {
4155 for ( j = 0 ; j < recursed_depth ; j++ ) {
4156 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
4158 PAREN_TEST(RExC_study_chunk_recursed +
4159 ( j * RExC_study_chunk_recursed_bytes), i )
4162 !PAREN_TEST(RExC_study_chunk_recursed +
4163 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
4166 Perl_re_printf( aTHX_ " %d",(int)i);
4170 if ( j + 1 < recursed_depth ) {
4171 Perl_re_printf( aTHX_ ",");
4175 Perl_re_printf( aTHX_ "\n");
4178 while ( scan && OP(scan) != END && scan < last ){
4179 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4180 node length to get a real minimum (because
4181 the folded version may be shorter) */
4182 bool unfolded_multi_char = FALSE;
4183 /* Peephole optimizer: */
4184 DEBUG_STUDYDATA("Peep:", data, depth);
4185 DEBUG_PEEP("Peep", scan, depth);
4188 /* The reason we do this here is that we need to deal with things like
4189 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4190 * parsing code, as each (?:..) is handled by a different invocation of
4193 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
4195 /* Follow the next-chain of the current node and optimize
4196 away all the NOTHINGs from it. */
4197 if (OP(scan) != CURLYX) {
4198 const int max = (reg_off_by_arg[OP(scan)]
4200 /* I32 may be smaller than U16 on CRAYs! */
4201 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4202 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
4206 /* Skip NOTHING and LONGJMP. */
4207 while ((n = regnext(n))
4208 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4209 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
4210 && off + noff < max)
4212 if (reg_off_by_arg[OP(scan)])
4215 NEXT_OFF(scan) = off;
4218 /* The principal pseudo-switch. Cannot be a switch, since we
4219 look into several different things. */
4220 if ( OP(scan) == DEFINEP ) {
4222 SSize_t deltanext = 0;
4223 SSize_t fake_last_close = 0;
4224 I32 f = SCF_IN_DEFINE;
4226 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4227 scan = regnext(scan);
4228 assert( OP(scan) == IFTHEN );
4229 DEBUG_PEEP("expect IFTHEN", scan, depth);
4231 data_fake.last_closep= &fake_last_close;
4233 next = regnext(scan);
4234 scan = NEXTOPER(NEXTOPER(scan));
4235 DEBUG_PEEP("scan", scan, depth);
4236 DEBUG_PEEP("next", next, depth);
4238 /* we suppose the run is continuous, last=next...
4239 * NOTE we dont use the return here! */
4240 (void)study_chunk(pRExC_state, &scan, &minlen,
4241 &deltanext, next, &data_fake, stopparen,
4242 recursed_depth, NULL, f, depth+1);
4247 OP(scan) == BRANCH ||
4248 OP(scan) == BRANCHJ ||
4251 next = regnext(scan);
4254 /* The op(next)==code check below is to see if we
4255 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4256 * IFTHEN is special as it might not appear in pairs.
4257 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4258 * we dont handle it cleanly. */
4259 if (OP(next) == code || code == IFTHEN) {
4260 /* NOTE - There is similar code to this block below for
4261 * handling TRIE nodes on a re-study. If you change stuff here
4262 * check there too. */
4263 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
4265 regnode * const startbranch=scan;
4267 if (flags & SCF_DO_SUBSTR) {
4268 /* Cannot merge strings after this. */
4269 scan_commit(pRExC_state, data, minlenp, is_inf);
4272 if (flags & SCF_DO_STCLASS)
4273 ssc_init_zero(pRExC_state, &accum);
4275 while (OP(scan) == code) {
4276 SSize_t deltanext, minnext, fake;
4278 regnode_ssc this_class;
4280 DEBUG_PEEP("Branch", scan, depth);
4283 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4285 data_fake.whilem_c = data->whilem_c;
4286 data_fake.last_closep = data->last_closep;
4289 data_fake.last_closep = &fake;
4291 data_fake.pos_delta = delta;
4292 next = regnext(scan);
4294 scan = NEXTOPER(scan); /* everything */
4295 if (code != BRANCH) /* everything but BRANCH */
4296 scan = NEXTOPER(scan);
4298 if (flags & SCF_DO_STCLASS) {
4299 ssc_init(pRExC_state, &this_class);
4300 data_fake.start_class = &this_class;
4301 f = SCF_DO_STCLASS_AND;
4303 if (flags & SCF_WHILEM_VISITED_POS)
4304 f |= SCF_WHILEM_VISITED_POS;
4306 /* we suppose the run is continuous, last=next...*/
4307 minnext = study_chunk(pRExC_state, &scan, minlenp,
4308 &deltanext, next, &data_fake, stopparen,
4309 recursed_depth, NULL, f,depth+1);
4313 if (deltanext == SSize_t_MAX) {
4314 is_inf = is_inf_internal = 1;
4316 } else if (max1 < minnext + deltanext)
4317 max1 = minnext + deltanext;
4319 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4321 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4322 if ( stopmin > minnext)
4323 stopmin = min + min1;
4324 flags &= ~SCF_DO_SUBSTR;
4326 data->flags |= SCF_SEEN_ACCEPT;
4329 if (data_fake.flags & SF_HAS_EVAL)
4330 data->flags |= SF_HAS_EVAL;
4331 data->whilem_c = data_fake.whilem_c;
4333 if (flags & SCF_DO_STCLASS)
4334 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4336 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4338 if (flags & SCF_DO_SUBSTR) {
4339 data->pos_min += min1;
4340 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
4341 data->pos_delta = SSize_t_MAX;
4343 data->pos_delta += max1 - min1;
4344 if (max1 != min1 || is_inf)
4345 data->longest = &(data->longest_float);
4348 if (delta == SSize_t_MAX
4349 || SSize_t_MAX - delta - (max1 - min1) < 0)
4350 delta = SSize_t_MAX;
4352 delta += max1 - min1;
4353 if (flags & SCF_DO_STCLASS_OR) {
4354 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4356 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4357 flags &= ~SCF_DO_STCLASS;
4360 else if (flags & SCF_DO_STCLASS_AND) {
4362 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4363 flags &= ~SCF_DO_STCLASS;
4366 /* Switch to OR mode: cache the old value of
4367 * data->start_class */
4369 StructCopy(data->start_class, and_withp, regnode_ssc);
4370 flags &= ~SCF_DO_STCLASS_AND;
4371 StructCopy(&accum, data->start_class, regnode_ssc);
4372 flags |= SCF_DO_STCLASS_OR;
4376 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4377 OP( startbranch ) == BRANCH )
4381 Assuming this was/is a branch we are dealing with: 'scan'
4382 now points at the item that follows the branch sequence,
4383 whatever it is. We now start at the beginning of the
4384 sequence and look for subsequences of
4390 which would be constructed from a pattern like
4393 If we can find such a subsequence we need to turn the first
4394 element into a trie and then add the subsequent branch exact
4395 strings to the trie.
4399 1. patterns where the whole set of branches can be
4402 2. patterns where only a subset can be converted.
4404 In case 1 we can replace the whole set with a single regop
4405 for the trie. In case 2 we need to keep the start and end
4408 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4409 becomes BRANCH TRIE; BRANCH X;
4411 There is an additional case, that being where there is a
4412 common prefix, which gets split out into an EXACT like node
4413 preceding the TRIE node.
4415 If x(1..n)==tail then we can do a simple trie, if not we make
4416 a "jump" trie, such that when we match the appropriate word
4417 we "jump" to the appropriate tail node. Essentially we turn
4418 a nested if into a case structure of sorts.
4423 if (!re_trie_maxbuff) {
4424 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4425 if (!SvIOK(re_trie_maxbuff))
4426 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4428 if ( SvIV(re_trie_maxbuff)>=0 ) {
4430 regnode *first = (regnode *)NULL;
4431 regnode *last = (regnode *)NULL;
4432 regnode *tail = scan;
4436 /* var tail is used because there may be a TAIL
4437 regop in the way. Ie, the exacts will point to the
4438 thing following the TAIL, but the last branch will
4439 point at the TAIL. So we advance tail. If we
4440 have nested (?:) we may have to move through several
4444 while ( OP( tail ) == TAIL ) {
4445 /* this is the TAIL generated by (?:) */
4446 tail = regnext( tail );
4450 DEBUG_TRIE_COMPILE_r({
4451 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4452 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4454 "Looking for TRIE'able sequences. Tail node is ",
4455 (UV)(tail - RExC_emit_start),
4456 SvPV_nolen_const( RExC_mysv )
4462 Step through the branches
4463 cur represents each branch,
4464 noper is the first thing to be matched as part
4466 noper_next is the regnext() of that node.
4468 We normally handle a case like this
4469 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4470 support building with NOJUMPTRIE, which restricts
4471 the trie logic to structures like /FOO|BAR/.
4473 If noper is a trieable nodetype then the branch is
4474 a possible optimization target. If we are building
4475 under NOJUMPTRIE then we require that noper_next is
4476 the same as scan (our current position in the regex
4479 Once we have two or more consecutive such branches
4480 we can create a trie of the EXACT's contents and
4481 stitch it in place into the program.
4483 If the sequence represents all of the branches in
4484 the alternation we replace the entire thing with a
4487 Otherwise when it is a subsequence we need to
4488 stitch it in place and replace only the relevant
4489 branches. This means the first branch has to remain
4490 as it is used by the alternation logic, and its
4491 next pointer, and needs to be repointed at the item
4492 on the branch chain following the last branch we
4493 have optimized away.
4495 This could be either a BRANCH, in which case the
4496 subsequence is internal, or it could be the item
4497 following the branch sequence in which case the
4498 subsequence is at the end (which does not
4499 necessarily mean the first node is the start of the
4502 TRIE_TYPE(X) is a define which maps the optype to a
4506 ----------------+-----------
4510 EXACTFU_SS | EXACTFU
4513 EXACTFLU8 | EXACTFLU8
4517 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4519 : ( EXACT == (X) ) \
4521 : ( EXACTFU == (X) || EXACTFU_SS == (X) ) \
4523 : ( EXACTFA == (X) ) \
4525 : ( EXACTL == (X) ) \
4527 : ( EXACTFLU8 == (X) ) \
4531 /* dont use tail as the end marker for this traverse */
4532 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4533 regnode * const noper = NEXTOPER( cur );
4534 U8 noper_type = OP( noper );
4535 U8 noper_trietype = TRIE_TYPE( noper_type );
4536 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4537 regnode * const noper_next = regnext( noper );
4538 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4539 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4542 DEBUG_TRIE_COMPILE_r({
4543 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4544 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4546 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4548 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4549 Perl_re_printf( aTHX_ " -> %d:%s",
4550 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
4553 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4554 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
4555 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
4557 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
4558 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4559 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4563 /* Is noper a trieable nodetype that can be merged
4564 * with the current trie (if there is one)? */
4568 ( noper_trietype == NOTHING )
4569 || ( trietype == NOTHING )
4570 || ( trietype == noper_trietype )
4573 && noper_next >= tail
4577 /* Handle mergable triable node Either we are
4578 * the first node in a new trieable sequence,
4579 * in which case we do some bookkeeping,
4580 * otherwise we update the end pointer. */
4583 if ( noper_trietype == NOTHING ) {
4584 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4585 regnode * const noper_next = regnext( noper );
4586 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4587 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4590 if ( noper_next_trietype ) {
4591 trietype = noper_next_trietype;
4592 } else if (noper_next_type) {
4593 /* a NOTHING regop is 1 regop wide.
4594 * We need at least two for a trie
4595 * so we can't merge this in */
4599 trietype = noper_trietype;
4602 if ( trietype == NOTHING )
4603 trietype = noper_trietype;
4608 } /* end handle mergable triable node */
4610 /* handle unmergable node -
4611 * noper may either be a triable node which can
4612 * not be tried together with the current trie,
4613 * or a non triable node */
4615 /* If last is set and trietype is not
4616 * NOTHING then we have found at least two
4617 * triable branch sequences in a row of a
4618 * similar trietype so we can turn them
4619 * into a trie. If/when we allow NOTHING to
4620 * start a trie sequence this condition
4621 * will be required, and it isn't expensive
4622 * so we leave it in for now. */
4623 if ( trietype && trietype != NOTHING )
4624 make_trie( pRExC_state,
4625 startbranch, first, cur, tail,
4626 count, trietype, depth+1 );
4627 last = NULL; /* note: we clear/update
4628 first, trietype etc below,
4629 so we dont do it here */
4633 && noper_next >= tail
4636 /* noper is triable, so we can start a new
4640 trietype = noper_trietype;
4642 /* if we already saw a first but the
4643 * current node is not triable then we have
4644 * to reset the first information. */
4649 } /* end handle unmergable node */
4650 } /* loop over branches */
4651 DEBUG_TRIE_COMPILE_r({
4652 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4653 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
4654 depth+1, SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4655 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
4656 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4657 PL_reg_name[trietype]
4661 if ( last && trietype ) {
4662 if ( trietype != NOTHING ) {
4663 /* the last branch of the sequence was part of
4664 * a trie, so we have to construct it here
4665 * outside of the loop */
4666 made= make_trie( pRExC_state, startbranch,
4667 first, scan, tail, count,
4668 trietype, depth+1 );
4669 #ifdef TRIE_STUDY_OPT
4670 if ( ((made == MADE_EXACT_TRIE &&
4671 startbranch == first)
4672 || ( first_non_open == first )) &&
4674 flags |= SCF_TRIE_RESTUDY;
4675 if ( startbranch == first
4678 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4683 /* at this point we know whatever we have is a
4684 * NOTHING sequence/branch AND if 'startbranch'
4685 * is 'first' then we can turn the whole thing
4688 if ( startbranch == first ) {
4690 /* the entire thing is a NOTHING sequence,
4691 * something like this: (?:|) So we can
4692 * turn it into a plain NOTHING op. */
4693 DEBUG_TRIE_COMPILE_r({
4694 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4695 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
4697 SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4700 OP(startbranch)= NOTHING;
4701 NEXT_OFF(startbranch)= tail - startbranch;
4702 for ( opt= startbranch + 1; opt < tail ; opt++ )
4706 } /* end if ( last) */
4707 } /* TRIE_MAXBUF is non zero */
4712 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4713 scan = NEXTOPER(NEXTOPER(scan));
4714 } else /* single branch is optimized. */
4715 scan = NEXTOPER(scan);
4717 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
4719 regnode *start = NULL;
4720 regnode *end = NULL;
4721 U32 my_recursed_depth= recursed_depth;
4723 if (OP(scan) != SUSPEND) { /* GOSUB */
4724 /* Do setup, note this code has side effects beyond
4725 * the rest of this block. Specifically setting
4726 * RExC_recurse[] must happen at least once during
4729 RExC_recurse[ARG2L(scan)] = scan;
4730 start = RExC_open_parens[paren];
4731 end = RExC_close_parens[paren];
4733 /* NOTE we MUST always execute the above code, even
4734 * if we do nothing with a GOSUB */
4736 ( flags & SCF_IN_DEFINE )
4739 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
4741 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
4744 /* no need to do anything here if we are in a define. */
4745 /* or we are after some kind of infinite construct
4746 * so we can skip recursing into this item.
4747 * Since it is infinite we will not change the maxlen
4748 * or delta, and if we miss something that might raise
4749 * the minlen it will merely pessimise a little.
4751 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
4752 * might result in a minlen of 1 and not of 4,
4753 * but this doesn't make us mismatch, just try a bit
4754 * harder than we should.
4756 scan= regnext(scan);
4763 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4765 /* it is quite possible that there are more efficient ways
4766 * to do this. We maintain a bitmap per level of recursion
4767 * of which patterns we have entered so we can detect if a
4768 * pattern creates a possible infinite loop. When we
4769 * recurse down a level we copy the previous levels bitmap
4770 * down. When we are at recursion level 0 we zero the top
4771 * level bitmap. It would be nice to implement a different
4772 * more efficient way of doing this. In particular the top
4773 * level bitmap may be unnecessary.
4775 if (!recursed_depth) {
4776 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4778 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4779 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4780 RExC_study_chunk_recursed_bytes, U8);
4782 /* we havent recursed into this paren yet, so recurse into it */
4783 DEBUG_STUDYDATA("gosub-set:", data,depth);
4784 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4785 my_recursed_depth= recursed_depth + 1;
4787 DEBUG_STUDYDATA("gosub-inf:", data,depth);
4788 /* some form of infinite recursion, assume infinite length
4790 if (flags & SCF_DO_SUBSTR) {
4791 scan_commit(pRExC_state, data, minlenp, is_inf);
4792 data->longest = &(data->longest_float);
4794 is_inf = is_inf_internal = 1;
4795 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4796 ssc_anything(data->start_class);
4797 flags &= ~SCF_DO_STCLASS;
4799 start= NULL; /* reset start so we dont recurse later on. */
4804 end = regnext(scan);
4807 scan_frame *newframe;
4809 if (!RExC_frame_last) {
4810 Newxz(newframe, 1, scan_frame);
4811 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
4812 RExC_frame_head= newframe;
4814 } else if (!RExC_frame_last->next_frame) {
4815 Newxz(newframe,1,scan_frame);
4816 RExC_frame_last->next_frame= newframe;
4817 newframe->prev_frame= RExC_frame_last;
4820 newframe= RExC_frame_last->next_frame;
4822 RExC_frame_last= newframe;
4824 newframe->next_regnode = regnext(scan);
4825 newframe->last_regnode = last;
4826 newframe->stopparen = stopparen;
4827 newframe->prev_recursed_depth = recursed_depth;
4828 newframe->this_prev_frame= frame;
4830 DEBUG_STUDYDATA("frame-new:",data,depth);
4831 DEBUG_PEEP("fnew", scan, depth);
4838 recursed_depth= my_recursed_depth;
4843 else if (OP(scan) == EXACT || OP(scan) == EXACTL) {
4844 SSize_t l = STR_LEN(scan);
4847 const U8 * const s = (U8*)STRING(scan);
4848 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4849 l = utf8_length(s, s + l);
4851 uc = *((U8*)STRING(scan));
4854 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4855 /* The code below prefers earlier match for fixed
4856 offset, later match for variable offset. */
4857 if (data->last_end == -1) { /* Update the start info. */
4858 data->last_start_min = data->pos_min;
4859 data->last_start_max = is_inf
4860 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4862 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4864 SvUTF8_on(data->last_found);
4866 SV * const sv = data->last_found;
4867 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4868 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4869 if (mg && mg->mg_len >= 0)
4870 mg->mg_len += utf8_length((U8*)STRING(scan),
4871 (U8*)STRING(scan)+STR_LEN(scan));
4873 data->last_end = data->pos_min + l;
4874 data->pos_min += l; /* As in the first entry. */
4875 data->flags &= ~SF_BEFORE_EOL;
4878 /* ANDing the code point leaves at most it, and not in locale, and
4879 * can't match null string */
4880 if (flags & SCF_DO_STCLASS_AND) {
4881 ssc_cp_and(data->start_class, uc);
4882 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4883 ssc_clear_locale(data->start_class);
4885 else if (flags & SCF_DO_STCLASS_OR) {
4886 ssc_add_cp(data->start_class, uc);
4887 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4889 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4890 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4892 flags &= ~SCF_DO_STCLASS;
4894 else if (PL_regkind[OP(scan)] == EXACT) {
4895 /* But OP != EXACT!, so is EXACTFish */
4896 SSize_t l = STR_LEN(scan);
4897 const U8 * s = (U8*)STRING(scan);
4899 /* Search for fixed substrings supports EXACT only. */
4900 if (flags & SCF_DO_SUBSTR) {
4902 scan_commit(pRExC_state, data, minlenp, is_inf);
4905 l = utf8_length(s, s + l);
4907 if (unfolded_multi_char) {
4908 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4910 min += l - min_subtract;
4912 delta += min_subtract;
4913 if (flags & SCF_DO_SUBSTR) {
4914 data->pos_min += l - min_subtract;
4915 if (data->pos_min < 0) {
4918 data->pos_delta += min_subtract;
4920 data->longest = &(data->longest_float);
4924 if (flags & SCF_DO_STCLASS) {
4925 SV* EXACTF_invlist = _make_exactf_invlist(pRExC_state, scan);
4927 assert(EXACTF_invlist);
4928 if (flags & SCF_DO_STCLASS_AND) {
4929 if (OP(scan) != EXACTFL)
4930 ssc_clear_locale(data->start_class);
4931 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4932 ANYOF_POSIXL_ZERO(data->start_class);
4933 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4935 else { /* SCF_DO_STCLASS_OR */
4936 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4937 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4939 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4940 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4942 flags &= ~SCF_DO_STCLASS;
4943 SvREFCNT_dec(EXACTF_invlist);
4946 else if (REGNODE_VARIES(OP(scan))) {
4947 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4948 I32 fl = 0, f = flags;
4949 regnode * const oscan = scan;
4950 regnode_ssc this_class;
4951 regnode_ssc *oclass = NULL;
4952 I32 next_is_eval = 0;
4954 switch (PL_regkind[OP(scan)]) {
4955 case WHILEM: /* End of (?:...)* . */
4956 scan = NEXTOPER(scan);
4959 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4960 next = NEXTOPER(scan);
4961 if (OP(next) == EXACT
4962 || OP(next) == EXACTL
4963 || (flags & SCF_DO_STCLASS))
4966 maxcount = REG_INFTY;
4967 next = regnext(scan);
4968 scan = NEXTOPER(scan);
4972 if (flags & SCF_DO_SUBSTR)
4977 if (flags & SCF_DO_STCLASS) {
4979 maxcount = REG_INFTY;
4980 next = regnext(scan);
4981 scan = NEXTOPER(scan);
4984 if (flags & SCF_DO_SUBSTR) {
4985 scan_commit(pRExC_state, data, minlenp, is_inf);
4986 /* Cannot extend fixed substrings */
4987 data->longest = &(data->longest_float);
4989 is_inf = is_inf_internal = 1;
4990 scan = regnext(scan);
4991 goto optimize_curly_tail;
4993 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
4994 && (scan->flags == stopparen))
4999 mincount = ARG1(scan);
5000 maxcount = ARG2(scan);
5002 next = regnext(scan);
5003 if (OP(scan) == CURLYX) {
5004 I32 lp = (data ? *(data->last_closep) : 0);
5005 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5007 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5008 next_is_eval = (OP(scan) == EVAL);
5010 if (flags & SCF_DO_SUBSTR) {
5012 scan_commit(pRExC_state, data, minlenp, is_inf);
5013 /* Cannot extend fixed substrings */
5014 pos_before = data->pos_min;
5018 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5020 data->flags |= SF_IS_INF;
5022 if (flags & SCF_DO_STCLASS) {
5023 ssc_init(pRExC_state, &this_class);
5024 oclass = data->start_class;
5025 data->start_class = &this_class;
5026 f |= SCF_DO_STCLASS_AND;
5027 f &= ~SCF_DO_STCLASS_OR;
5029 /* Exclude from super-linear cache processing any {n,m}
5030 regops for which the combination of input pos and regex
5031 pos is not enough information to determine if a match
5034 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5035 regex pos at the \s*, the prospects for a match depend not
5036 only on the input position but also on how many (bar\s*)
5037 repeats into the {4,8} we are. */
5038 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5039 f &= ~SCF_WHILEM_VISITED_POS;
5041 /* This will finish on WHILEM, setting scan, or on NULL: */
5042 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5043 last, data, stopparen, recursed_depth, NULL,
5045 ? (f & ~SCF_DO_SUBSTR)
5049 if (flags & SCF_DO_STCLASS)
5050 data->start_class = oclass;
5051 if (mincount == 0 || minnext == 0) {
5052 if (flags & SCF_DO_STCLASS_OR) {
5053 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5055 else if (flags & SCF_DO_STCLASS_AND) {
5056 /* Switch to OR mode: cache the old value of
5057 * data->start_class */
5059 StructCopy(data->start_class, and_withp, regnode_ssc);
5060 flags &= ~SCF_DO_STCLASS_AND;
5061 StructCopy(&this_class, data->start_class, regnode_ssc);
5062 flags |= SCF_DO_STCLASS_OR;
5063 ANYOF_FLAGS(data->start_class)
5064 |= SSC_MATCHES_EMPTY_STRING;
5066 } else { /* Non-zero len */
5067 if (flags & SCF_DO_STCLASS_OR) {
5068 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5069 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5071 else if (flags & SCF_DO_STCLASS_AND)
5072 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5073 flags &= ~SCF_DO_STCLASS;
5075 if (!scan) /* It was not CURLYX, but CURLY. */
5077 if (!(flags & SCF_TRIE_DOING_RESTUDY)
5078 /* ? quantifier ok, except for (?{ ... }) */
5079 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5080 && (minnext == 0) && (deltanext == 0)
5081 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5082 && maxcount <= REG_INFTY/3) /* Complement check for big
5085 /* Fatal warnings may leak the regexp without this: */
5086 SAVEFREESV(RExC_rx_sv);
5087 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5088 "Quantifier unexpected on zero-length expression "
5089 "in regex m/%" UTF8f "/",
5090 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5092 (void)ReREFCNT_inc(RExC_rx_sv);
5095 min += minnext * mincount;
5096 is_inf_internal |= deltanext == SSize_t_MAX
5097 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5098 is_inf |= is_inf_internal;
5100 delta = SSize_t_MAX;
5102 delta += (minnext + deltanext) * maxcount
5103 - minnext * mincount;
5105 /* Try powerful optimization CURLYX => CURLYN. */
5106 if ( OP(oscan) == CURLYX && data
5107 && data->flags & SF_IN_PAR
5108 && !(data->flags & SF_HAS_EVAL)
5109 && !deltanext && minnext == 1 ) {
5110 /* Try to optimize to CURLYN. */
5111 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5112 regnode * const nxt1 = nxt;
5119 if (!REGNODE_SIMPLE(OP(nxt))
5120 && !(PL_regkind[OP(nxt)] == EXACT
5121 && STR_LEN(nxt) == 1))
5127 if (OP(nxt) != CLOSE)
5129 if (RExC_open_parens) {
5130 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5131 RExC_close_parens[ARG(nxt1)]=nxt+2; /*close->while*/
5133 /* Now we know that nxt2 is the only contents: */
5134 oscan->flags = (U8)ARG(nxt);
5136 OP(nxt1) = NOTHING; /* was OPEN. */
5139 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5140 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5141 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5142 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5143 OP(nxt + 1) = OPTIMIZED; /* was count. */
5144 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5149 /* Try optimization CURLYX => CURLYM. */
5150 if ( OP(oscan) == CURLYX && data
5151 && !(data->flags & SF_HAS_PAR)
5152 && !(data->flags & SF_HAS_EVAL)
5153 && !deltanext /* atom is fixed width */
5154 && minnext != 0 /* CURLYM can't handle zero width */
5156 /* Nor characters whose fold at run-time may be
5157 * multi-character */
5158 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5160 /* XXXX How to optimize if data == 0? */
5161 /* Optimize to a simpler form. */
5162 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5166 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5167 && (OP(nxt2) != WHILEM))
5169 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5170 /* Need to optimize away parenths. */
5171 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5172 /* Set the parenth number. */
5173 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5175 oscan->flags = (U8)ARG(nxt);
5176 if (RExC_open_parens) {
5177 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5178 RExC_close_parens[ARG(nxt1)]=nxt2+1; /*close->NOTHING*/
5180 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5181 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5184 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5185 OP(nxt + 1) = OPTIMIZED; /* was count. */
5186 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5187 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5190 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5191 regnode *nnxt = regnext(nxt1);
5193 if (reg_off_by_arg[OP(nxt1)])
5194 ARG_SET(nxt1, nxt2 - nxt1);
5195 else if (nxt2 - nxt1 < U16_MAX)
5196 NEXT_OFF(nxt1) = nxt2 - nxt1;
5198 OP(nxt) = NOTHING; /* Cannot beautify */
5203 /* Optimize again: */
5204 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5205 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
5210 else if ((OP(oscan) == CURLYX)
5211 && (flags & SCF_WHILEM_VISITED_POS)
5212 /* See the comment on a similar expression above.
5213 However, this time it's not a subexpression
5214 we care about, but the expression itself. */
5215 && (maxcount == REG_INFTY)
5216 && data && ++data->whilem_c < 16) {
5217 /* This stays as CURLYX, we can put the count/of pair. */
5218 /* Find WHILEM (as in regexec.c) */
5219 regnode *nxt = oscan + NEXT_OFF(oscan);
5221 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5223 PREVOPER(nxt)->flags = (U8)(data->whilem_c
5224 | (RExC_whilem_seen << 4)); /* On WHILEM */
5226 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5228 if (flags & SCF_DO_SUBSTR) {
5229 SV *last_str = NULL;
5230 STRLEN last_chrs = 0;
5231 int counted = mincount != 0;
5233 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5235 SSize_t b = pos_before >= data->last_start_min
5236 ? pos_before : data->last_start_min;
5238 const char * const s = SvPV_const(data->last_found, l);
5239 SSize_t old = b - data->last_start_min;
5242 old = utf8_hop((U8*)s, old) - (U8*)s;
5244 /* Get the added string: */
5245 last_str = newSVpvn_utf8(s + old, l, UTF);
5246 last_chrs = UTF ? utf8_length((U8*)(s + old),
5247 (U8*)(s + old + l)) : l;
5248 if (deltanext == 0 && pos_before == b) {
5249 /* What was added is a constant string */
5252 SvGROW(last_str, (mincount * l) + 1);
5253 repeatcpy(SvPVX(last_str) + l,
5254 SvPVX_const(last_str), l,
5256 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5257 /* Add additional parts. */
5258 SvCUR_set(data->last_found,
5259 SvCUR(data->last_found) - l);
5260 sv_catsv(data->last_found, last_str);
5262 SV * sv = data->last_found;
5264 SvUTF8(sv) && SvMAGICAL(sv) ?
5265 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5266 if (mg && mg->mg_len >= 0)
5267 mg->mg_len += last_chrs * (mincount-1);
5269 last_chrs *= mincount;
5270 data->last_end += l * (mincount - 1);
5273 /* start offset must point into the last copy */
5274 data->last_start_min += minnext * (mincount - 1);
5275 data->last_start_max =
5278 : data->last_start_max +
5279 (maxcount - 1) * (minnext + data->pos_delta);
5282 /* It is counted once already... */
5283 data->pos_min += minnext * (mincount - counted);
5285 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5286 " SSize_t_MAX=%" UVuf " minnext=%" UVuf
5287 " maxcount=%" UVuf " mincount=%" UVuf "\n",
5288 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
5290 if (deltanext != SSize_t_MAX)
5291 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5292 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5293 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
5295 if (deltanext == SSize_t_MAX
5296 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
5297 data->pos_delta = SSize_t_MAX;
5299 data->pos_delta += - counted * deltanext +
5300 (minnext + deltanext) * maxcount - minnext * mincount;
5301 if (mincount != maxcount) {
5302 /* Cannot extend fixed substrings found inside
5304 scan_commit(pRExC_state, data, minlenp, is_inf);
5305 if (mincount && last_str) {
5306 SV * const sv = data->last_found;
5307 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5308 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5312 sv_setsv(sv, last_str);
5313 data->last_end = data->pos_min;
5314 data->last_start_min = data->pos_min - last_chrs;
5315 data->last_start_max = is_inf
5317 : data->pos_min + data->pos_delta - last_chrs;
5319 data->longest = &(data->longest_float);
5321 SvREFCNT_dec(last_str);
5323 if (data && (fl & SF_HAS_EVAL))
5324 data->flags |= SF_HAS_EVAL;
5325 optimize_curly_tail:
5326 if (OP(oscan) != CURLYX) {
5327 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
5329 NEXT_OFF(oscan) += NEXT_OFF(next);
5335 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5340 if (flags & SCF_DO_SUBSTR) {
5341 /* Cannot expect anything... */
5342 scan_commit(pRExC_state, data, minlenp, is_inf);
5343 data->longest = &(data->longest_float);
5345 is_inf = is_inf_internal = 1;
5346 if (flags & SCF_DO_STCLASS_OR) {
5347 if (OP(scan) == CLUMP) {
5348 /* Actually is any start char, but very few code points
5349 * aren't start characters */
5350 ssc_match_all_cp(data->start_class);
5353 ssc_anything(data->start_class);
5356 flags &= ~SCF_DO_STCLASS;
5360 else if (OP(scan) == LNBREAK) {
5361 if (flags & SCF_DO_STCLASS) {
5362 if (flags & SCF_DO_STCLASS_AND) {
5363 ssc_intersection(data->start_class,
5364 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5365 ssc_clear_locale(data->start_class);
5366 ANYOF_FLAGS(data->start_class)
5367 &= ~SSC_MATCHES_EMPTY_STRING;
5369 else if (flags & SCF_DO_STCLASS_OR) {
5370 ssc_union(data->start_class,
5371 PL_XPosix_ptrs[_CC_VERTSPACE],
5373 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5375 /* See commit msg for
5376 * 749e076fceedeb708a624933726e7989f2302f6a */
5377 ANYOF_FLAGS(data->start_class)
5378 &= ~SSC_MATCHES_EMPTY_STRING;
5380 flags &= ~SCF_DO_STCLASS;
5383 if (delta != SSize_t_MAX)
5384 delta++; /* Because of the 2 char string cr-lf */
5385 if (flags & SCF_DO_SUBSTR) {
5386 /* Cannot expect anything... */
5387 scan_commit(pRExC_state, data, minlenp, is_inf);
5389 data->pos_delta += 1;
5390 data->longest = &(data->longest_float);
5393 else if (REGNODE_SIMPLE(OP(scan))) {
5395 if (flags & SCF_DO_SUBSTR) {
5396 scan_commit(pRExC_state, data, minlenp, is_inf);
5400 if (flags & SCF_DO_STCLASS) {
5402 SV* my_invlist = NULL;
5405 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5406 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5408 /* Some of the logic below assumes that switching
5409 locale on will only add false positives. */
5414 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5418 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5419 ssc_match_all_cp(data->start_class);
5424 SV* REG_ANY_invlist = _new_invlist(2);
5425 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5427 if (flags & SCF_DO_STCLASS_OR) {
5428 ssc_union(data->start_class,
5430 TRUE /* TRUE => invert, hence all but \n
5434 else if (flags & SCF_DO_STCLASS_AND) {
5435 ssc_intersection(data->start_class,
5437 TRUE /* TRUE => invert */
5439 ssc_clear_locale(data->start_class);
5441 SvREFCNT_dec_NN(REG_ANY_invlist);
5448 if (flags & SCF_DO_STCLASS_AND)
5449 ssc_and(pRExC_state, data->start_class,
5450 (regnode_charclass *) scan);
5452 ssc_or(pRExC_state, data->start_class,
5453 (regnode_charclass *) scan);
5461 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5462 if (flags & SCF_DO_STCLASS_AND) {
5463 bool was_there = cBOOL(
5464 ANYOF_POSIXL_TEST(data->start_class,
5466 ANYOF_POSIXL_ZERO(data->start_class);
5467 if (was_there) { /* Do an AND */
5468 ANYOF_POSIXL_SET(data->start_class, namedclass);
5470 /* No individual code points can now match */
5471 data->start_class->invlist
5472 = sv_2mortal(_new_invlist(0));
5475 int complement = namedclass + ((invert) ? -1 : 1);
5477 assert(flags & SCF_DO_STCLASS_OR);
5479 /* If the complement of this class was already there,
5480 * the result is that they match all code points,
5481 * (\d + \D == everything). Remove the classes from
5482 * future consideration. Locale is not relevant in
5484 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5485 ssc_match_all_cp(data->start_class);
5486 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5487 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5489 else { /* The usual case; just add this class to the
5491 ANYOF_POSIXL_SET(data->start_class, namedclass);
5496 case NPOSIXA: /* For these, we always know the exact set of
5501 if (FLAGS(scan) == _CC_ASCII) {
5502 my_invlist = invlist_clone(PL_XPosix_ptrs[_CC_ASCII]);
5505 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
5506 PL_XPosix_ptrs[_CC_ASCII],
5517 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5519 /* NPOSIXD matches all upper Latin1 code points unless the
5520 * target string being matched is UTF-8, which is
5521 * unknowable until match time. Since we are going to
5522 * invert, we want to get rid of all of them so that the
5523 * inversion will match all */
5524 if (OP(scan) == NPOSIXD) {
5525 _invlist_subtract(my_invlist, PL_UpperLatin1,
5531 if (flags & SCF_DO_STCLASS_AND) {
5532 ssc_intersection(data->start_class, my_invlist, invert);
5533 ssc_clear_locale(data->start_class);
5536 assert(flags & SCF_DO_STCLASS_OR);
5537 ssc_union(data->start_class, my_invlist, invert);
5539 SvREFCNT_dec(my_invlist);
5541 if (flags & SCF_DO_STCLASS_OR)
5542 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5543 flags &= ~SCF_DO_STCLASS;
5546 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5547 data->flags |= (OP(scan) == MEOL
5550 scan_commit(pRExC_state, data, minlenp, is_inf);
5553 else if ( PL_regkind[OP(scan)] == BRANCHJ
5554 /* Lookbehind, or need to calculate parens/evals/stclass: */
5555 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5556 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5558 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5559 || OP(scan) == UNLESSM )
5561 /* Negative Lookahead/lookbehind
5562 In this case we can't do fixed string optimisation.
5565 SSize_t deltanext, minnext, fake = 0;
5570 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5572 data_fake.whilem_c = data->whilem_c;
5573 data_fake.last_closep = data->last_closep;
5576 data_fake.last_closep = &fake;
5577 data_fake.pos_delta = delta;
5578 if ( flags & SCF_DO_STCLASS && !scan->flags
5579 && OP(scan) == IFMATCH ) { /* Lookahead */
5580 ssc_init(pRExC_state, &intrnl);
5581 data_fake.start_class = &intrnl;
5582 f |= SCF_DO_STCLASS_AND;
5584 if (flags & SCF_WHILEM_VISITED_POS)
5585 f |= SCF_WHILEM_VISITED_POS;
5586 next = regnext(scan);
5587 nscan = NEXTOPER(NEXTOPER(scan));
5588 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5589 last, &data_fake, stopparen,
5590 recursed_depth, NULL, f, depth+1);
5593 FAIL("Variable length lookbehind not implemented");
5595 else if (minnext > (I32)U8_MAX) {
5596 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5599 scan->flags = (U8)minnext;
5602 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5604 if (data_fake.flags & SF_HAS_EVAL)
5605 data->flags |= SF_HAS_EVAL;
5606 data->whilem_c = data_fake.whilem_c;
5608 if (f & SCF_DO_STCLASS_AND) {
5609 if (flags & SCF_DO_STCLASS_OR) {
5610 /* OR before, AND after: ideally we would recurse with
5611 * data_fake to get the AND applied by study of the
5612 * remainder of the pattern, and then derecurse;
5613 * *** HACK *** for now just treat as "no information".
5614 * See [perl #56690].
5616 ssc_init(pRExC_state, data->start_class);
5618 /* AND before and after: combine and continue. These
5619 * assertions are zero-length, so can match an EMPTY
5621 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5622 ANYOF_FLAGS(data->start_class)
5623 |= SSC_MATCHES_EMPTY_STRING;
5627 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5629 /* Positive Lookahead/lookbehind
5630 In this case we can do fixed string optimisation,
5631 but we must be careful about it. Note in the case of
5632 lookbehind the positions will be offset by the minimum
5633 length of the pattern, something we won't know about
5634 until after the recurse.
5636 SSize_t deltanext, fake = 0;
5640 /* We use SAVEFREEPV so that when the full compile
5641 is finished perl will clean up the allocated
5642 minlens when it's all done. This way we don't
5643 have to worry about freeing them when we know
5644 they wont be used, which would be a pain.
5647 Newx( minnextp, 1, SSize_t );
5648 SAVEFREEPV(minnextp);
5651 StructCopy(data, &data_fake, scan_data_t);
5652 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5655 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5656 data_fake.last_found=newSVsv(data->last_found);
5660 data_fake.last_closep = &fake;
5661 data_fake.flags = 0;
5662 data_fake.pos_delta = delta;
5664 data_fake.flags |= SF_IS_INF;
5665 if ( flags & SCF_DO_STCLASS && !scan->flags
5666 && OP(scan) == IFMATCH ) { /* Lookahead */
5667 ssc_init(pRExC_state, &intrnl);
5668 data_fake.start_class = &intrnl;
5669 f |= SCF_DO_STCLASS_AND;
5671 if (flags & SCF_WHILEM_VISITED_POS)
5672 f |= SCF_WHILEM_VISITED_POS;
5673 next = regnext(scan);
5674 nscan = NEXTOPER(NEXTOPER(scan));
5676 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5677 &deltanext, last, &data_fake,
5678 stopparen, recursed_depth, NULL,
5682 FAIL("Variable length lookbehind not implemented");
5684 else if (*minnextp > (I32)U8_MAX) {
5685 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5688 scan->flags = (U8)*minnextp;
5693 if (f & SCF_DO_STCLASS_AND) {
5694 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5695 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
5698 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5700 if (data_fake.flags & SF_HAS_EVAL)
5701 data->flags |= SF_HAS_EVAL;
5702 data->whilem_c = data_fake.whilem_c;
5703 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5704 if (RExC_rx->minlen<*minnextp)
5705 RExC_rx->minlen=*minnextp;
5706 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5707 SvREFCNT_dec_NN(data_fake.last_found);
5709 if ( data_fake.minlen_fixed != minlenp )
5711 data->offset_fixed= data_fake.offset_fixed;
5712 data->minlen_fixed= data_fake.minlen_fixed;
5713 data->lookbehind_fixed+= scan->flags;
5715 if ( data_fake.minlen_float != minlenp )
5717 data->minlen_float= data_fake.minlen_float;
5718 data->offset_float_min=data_fake.offset_float_min;
5719 data->offset_float_max=data_fake.offset_float_max;
5720 data->lookbehind_float+= scan->flags;
5727 else if (OP(scan) == OPEN) {
5728 if (stopparen != (I32)ARG(scan))
5731 else if (OP(scan) == CLOSE) {
5732 if (stopparen == (I32)ARG(scan)) {
5735 if ((I32)ARG(scan) == is_par) {
5736 next = regnext(scan);
5738 if ( next && (OP(next) != WHILEM) && next < last)
5739 is_par = 0; /* Disable optimization */
5742 *(data->last_closep) = ARG(scan);
5744 else if (OP(scan) == EVAL) {
5746 data->flags |= SF_HAS_EVAL;
5748 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5749 if (flags & SCF_DO_SUBSTR) {
5750 scan_commit(pRExC_state, data, minlenp, is_inf);
5751 flags &= ~SCF_DO_SUBSTR;
5753 if (data && OP(scan)==ACCEPT) {
5754 data->flags |= SCF_SEEN_ACCEPT;
5759 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5761 if (flags & SCF_DO_SUBSTR) {
5762 scan_commit(pRExC_state, data, minlenp, is_inf);
5763 data->longest = &(data->longest_float);
5765 is_inf = is_inf_internal = 1;
5766 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5767 ssc_anything(data->start_class);
5768 flags &= ~SCF_DO_STCLASS;
5770 else if (OP(scan) == GPOS) {
5771 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5772 !(delta || is_inf || (data && data->pos_delta)))
5774 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5775 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5776 if (RExC_rx->gofs < (STRLEN)min)
5777 RExC_rx->gofs = min;
5779 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5783 #ifdef TRIE_STUDY_OPT
5784 #ifdef FULL_TRIE_STUDY
5785 else if (PL_regkind[OP(scan)] == TRIE) {
5786 /* NOTE - There is similar code to this block above for handling
5787 BRANCH nodes on the initial study. If you change stuff here
5789 regnode *trie_node= scan;
5790 regnode *tail= regnext(scan);
5791 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5792 SSize_t max1 = 0, min1 = SSize_t_MAX;
5795 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5796 /* Cannot merge strings after this. */
5797 scan_commit(pRExC_state, data, minlenp, is_inf);
5799 if (flags & SCF_DO_STCLASS)
5800 ssc_init_zero(pRExC_state, &accum);
5806 const regnode *nextbranch= NULL;
5809 for ( word=1 ; word <= trie->wordcount ; word++)
5811 SSize_t deltanext=0, minnext=0, f = 0, fake;
5812 regnode_ssc this_class;
5814 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5816 data_fake.whilem_c = data->whilem_c;
5817 data_fake.last_closep = data->last_closep;
5820 data_fake.last_closep = &fake;
5821 data_fake.pos_delta = delta;
5822 if (flags & SCF_DO_STCLASS) {
5823 ssc_init(pRExC_state, &this_class);
5824 data_fake.start_class = &this_class;
5825 f = SCF_DO_STCLASS_AND;
5827 if (flags & SCF_WHILEM_VISITED_POS)
5828 f |= SCF_WHILEM_VISITED_POS;
5830 if (trie->jump[word]) {
5832 nextbranch = trie_node + trie->jump[0];
5833 scan= trie_node + trie->jump[word];
5834 /* We go from the jump point to the branch that follows
5835 it. Note this means we need the vestigal unused
5836 branches even though they arent otherwise used. */
5837 minnext = study_chunk(pRExC_state, &scan, minlenp,
5838 &deltanext, (regnode *)nextbranch, &data_fake,
5839 stopparen, recursed_depth, NULL, f,depth+1);
5841 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5842 nextbranch= regnext((regnode*)nextbranch);
5844 if (min1 > (SSize_t)(minnext + trie->minlen))
5845 min1 = minnext + trie->minlen;
5846 if (deltanext == SSize_t_MAX) {
5847 is_inf = is_inf_internal = 1;
5849 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5850 max1 = minnext + deltanext + trie->maxlen;
5852 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5854 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5855 if ( stopmin > min + min1)
5856 stopmin = min + min1;
5857 flags &= ~SCF_DO_SUBSTR;
5859 data->flags |= SCF_SEEN_ACCEPT;
5862 if (data_fake.flags & SF_HAS_EVAL)
5863 data->flags |= SF_HAS_EVAL;
5864 data->whilem_c = data_fake.whilem_c;
5866 if (flags & SCF_DO_STCLASS)
5867 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5870 if (flags & SCF_DO_SUBSTR) {
5871 data->pos_min += min1;
5872 data->pos_delta += max1 - min1;
5873 if (max1 != min1 || is_inf)
5874 data->longest = &(data->longest_float);
5877 if (delta != SSize_t_MAX)
5878 delta += max1 - min1;
5879 if (flags & SCF_DO_STCLASS_OR) {
5880 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5882 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5883 flags &= ~SCF_DO_STCLASS;
5886 else if (flags & SCF_DO_STCLASS_AND) {
5888 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5889 flags &= ~SCF_DO_STCLASS;
5892 /* Switch to OR mode: cache the old value of
5893 * data->start_class */
5895 StructCopy(data->start_class, and_withp, regnode_ssc);
5896 flags &= ~SCF_DO_STCLASS_AND;
5897 StructCopy(&accum, data->start_class, regnode_ssc);
5898 flags |= SCF_DO_STCLASS_OR;
5905 else if (PL_regkind[OP(scan)] == TRIE) {
5906 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5909 min += trie->minlen;
5910 delta += (trie->maxlen - trie->minlen);
5911 flags &= ~SCF_DO_STCLASS; /* xxx */
5912 if (flags & SCF_DO_SUBSTR) {
5913 /* Cannot expect anything... */
5914 scan_commit(pRExC_state, data, minlenp, is_inf);
5915 data->pos_min += trie->minlen;
5916 data->pos_delta += (trie->maxlen - trie->minlen);
5917 if (trie->maxlen != trie->minlen)
5918 data->longest = &(data->longest_float);
5920 if (trie->jump) /* no more substrings -- for now /grr*/
5921 flags &= ~SCF_DO_SUBSTR;
5923 #endif /* old or new */
5924 #endif /* TRIE_STUDY_OPT */
5926 /* Else: zero-length, ignore. */
5927 scan = regnext(scan);
5932 /* we need to unwind recursion. */
5935 DEBUG_STUDYDATA("frame-end:",data,depth);
5936 DEBUG_PEEP("fend", scan, depth);
5938 /* restore previous context */
5939 last = frame->last_regnode;
5940 scan = frame->next_regnode;
5941 stopparen = frame->stopparen;
5942 recursed_depth = frame->prev_recursed_depth;
5944 RExC_frame_last = frame->prev_frame;
5945 frame = frame->this_prev_frame;
5946 goto fake_study_recurse;
5950 DEBUG_STUDYDATA("pre-fin:",data,depth);
5953 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5955 if (flags & SCF_DO_SUBSTR && is_inf)
5956 data->pos_delta = SSize_t_MAX - data->pos_min;
5957 if (is_par > (I32)U8_MAX)
5959 if (is_par && pars==1 && data) {
5960 data->flags |= SF_IN_PAR;
5961 data->flags &= ~SF_HAS_PAR;
5963 else if (pars && data) {
5964 data->flags |= SF_HAS_PAR;
5965 data->flags &= ~SF_IN_PAR;
5967 if (flags & SCF_DO_STCLASS_OR)
5968 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5969 if (flags & SCF_TRIE_RESTUDY)
5970 data->flags |= SCF_TRIE_RESTUDY;
5972 DEBUG_STUDYDATA("post-fin:",data,depth);
5975 SSize_t final_minlen= min < stopmin ? min : stopmin;
5977 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
5978 if (final_minlen > SSize_t_MAX - delta)
5979 RExC_maxlen = SSize_t_MAX;
5980 else if (RExC_maxlen < final_minlen + delta)
5981 RExC_maxlen = final_minlen + delta;
5983 return final_minlen;
5985 NOT_REACHED; /* NOTREACHED */
5989 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5991 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5993 PERL_ARGS_ASSERT_ADD_DATA;
5995 Renewc(RExC_rxi->data,
5996 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
5997 char, struct reg_data);
5999 Renew(RExC_rxi->data->what, count + n, U8);
6001 Newx(RExC_rxi->data->what, n, U8);
6002 RExC_rxi->data->count = count + n;
6003 Copy(s, RExC_rxi->data->what + count, n, U8);
6007 /*XXX: todo make this not included in a non debugging perl, but appears to be
6008 * used anyway there, in 'use re' */
6009 #ifndef PERL_IN_XSUB_RE
6011 Perl_reginitcolors(pTHX)
6013 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6015 char *t = savepv(s);
6019 t = strchr(t, '\t');
6025 PL_colors[i] = t = (char *)"";
6030 PL_colors[i++] = (char *)"";
6037 #ifdef TRIE_STUDY_OPT
6038 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6041 (data.flags & SCF_TRIE_RESTUDY) \
6049 #define CHECK_RESTUDY_GOTO_butfirst
6053 * pregcomp - compile a regular expression into internal code
6055 * Decides which engine's compiler to call based on the hint currently in
6059 #ifndef PERL_IN_XSUB_RE
6061 /* return the currently in-scope regex engine (or the default if none) */
6063 regexp_engine const *
6064 Perl_current_re_engine(pTHX)
6066 if (IN_PERL_COMPILETIME) {
6067 HV * const table = GvHV(PL_hintgv);
6070 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6071 return &PL_core_reg_engine;
6072 ptr = hv_fetchs(table, "regcomp", FALSE);
6073 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6074 return &PL_core_reg_engine;
6075 return INT2PTR(regexp_engine*,SvIV(*ptr));
6079 if (!PL_curcop->cop_hints_hash)
6080 return &PL_core_reg_engine;
6081 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6082 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6083 return &PL_core_reg_engine;
6084 return INT2PTR(regexp_engine*,SvIV(ptr));
6090 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6092 regexp_engine const *eng = current_re_engine();
6093 GET_RE_DEBUG_FLAGS_DECL;
6095 PERL_ARGS_ASSERT_PREGCOMP;
6097 /* Dispatch a request to compile a regexp to correct regexp engine. */
6099 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6102 return CALLREGCOMP_ENG(eng, pattern, flags);
6106 /* public(ish) entry point for the perl core's own regex compiling code.
6107 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6108 * pattern rather than a list of OPs, and uses the internal engine rather
6109 * than the current one */
6112 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6114 SV *pat = pattern; /* defeat constness! */
6115 PERL_ARGS_ASSERT_RE_COMPILE;
6116 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6117 #ifdef PERL_IN_XSUB_RE
6120 &PL_core_reg_engine,
6122 NULL, NULL, rx_flags, 0);
6126 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6127 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6128 * point to the realloced string and length.
6130 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6134 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6135 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6137 U8 *const src = (U8*)*pat_p;
6142 GET_RE_DEBUG_FLAGS_DECL;
6144 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6145 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6147 Newx(dst, *plen_p * 2 + 1, U8);
6150 while (s < *plen_p) {
6151 append_utf8_from_native_byte(src[s], &d);
6152 if (n < num_code_blocks) {
6153 if (!do_end && pRExC_state->code_blocks[n].start == s) {
6154 pRExC_state->code_blocks[n].start = d - dst - 1;
6155 assert(*(d - 1) == '(');
6158 else if (do_end && pRExC_state->code_blocks[n].end == s) {
6159 pRExC_state->code_blocks[n].end = d - dst - 1;
6160 assert(*(d - 1) == ')');
6169 *pat_p = (char*) dst;
6171 RExC_orig_utf8 = RExC_utf8 = 1;
6176 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6177 * while recording any code block indices, and handling overloading,
6178 * nested qr// objects etc. If pat is null, it will allocate a new
6179 * string, or just return the first arg, if there's only one.
6181 * Returns the malloced/updated pat.
6182 * patternp and pat_count is the array of SVs to be concatted;
6183 * oplist is the optional list of ops that generated the SVs;
6184 * recompile_p is a pointer to a boolean that will be set if
6185 * the regex will need to be recompiled.
6186 * delim, if non-null is an SV that will be inserted between each element
6190 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6191 SV *pat, SV ** const patternp, int pat_count,
6192 OP *oplist, bool *recompile_p, SV *delim)
6196 bool use_delim = FALSE;
6197 bool alloced = FALSE;
6199 /* if we know we have at least two args, create an empty string,
6200 * then concatenate args to that. For no args, return an empty string */
6201 if (!pat && pat_count != 1) {
6207 for (svp = patternp; svp < patternp + pat_count; svp++) {
6210 STRLEN orig_patlen = 0;
6212 SV *msv = use_delim ? delim : *svp;
6213 if (!msv) msv = &PL_sv_undef;
6215 /* if we've got a delimiter, we go round the loop twice for each
6216 * svp slot (except the last), using the delimiter the second
6225 if (SvTYPE(msv) == SVt_PVAV) {
6226 /* we've encountered an interpolated array within
6227 * the pattern, e.g. /...@a..../. Expand the list of elements,
6228 * then recursively append elements.
6229 * The code in this block is based on S_pushav() */
6231 AV *const av = (AV*)msv;
6232 const SSize_t maxarg = AvFILL(av) + 1;
6236 assert(oplist->op_type == OP_PADAV
6237 || oplist->op_type == OP_RV2AV);
6238 oplist = OpSIBLING(oplist);
6241 if (SvRMAGICAL(av)) {
6244 Newx(array, maxarg, SV*);
6246 for (i=0; i < maxarg; i++) {
6247 SV ** const svp = av_fetch(av, i, FALSE);
6248 array[i] = svp ? *svp : &PL_sv_undef;
6252 array = AvARRAY(av);
6254 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6255 array, maxarg, NULL, recompile_p,
6257 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6263 /* we make the assumption here that each op in the list of
6264 * op_siblings maps to one SV pushed onto the stack,
6265 * except for code blocks, with have both an OP_NULL and
6267 * This allows us to match up the list of SVs against the
6268 * list of OPs to find the next code block.
6270 * Note that PUSHMARK PADSV PADSV ..
6272 * PADRANGE PADSV PADSV ..
6273 * so the alignment still works. */
6276 if (oplist->op_type == OP_NULL
6277 && (oplist->op_flags & OPf_SPECIAL))
6279 assert(n < pRExC_state->num_code_blocks);
6280 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
6281 pRExC_state->code_blocks[n].block = oplist;
6282 pRExC_state->code_blocks[n].src_regex = NULL;
6285 oplist = OpSIBLING(oplist); /* skip CONST */
6288 oplist = OpSIBLING(oplist);;
6291 /* apply magic and QR overloading to arg */
6294 if (SvROK(msv) && SvAMAGIC(msv)) {
6295 SV *sv = AMG_CALLunary(msv, regexp_amg);
6299 if (SvTYPE(sv) != SVt_REGEXP)
6300 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6305 /* try concatenation overload ... */
6306 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6307 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6310 /* overloading involved: all bets are off over literal
6311 * code. Pretend we haven't seen it */
6312 pRExC_state->num_code_blocks -= n;
6316 /* ... or failing that, try "" overload */
6317 while (SvAMAGIC(msv)
6318 && (sv = AMG_CALLunary(msv, string_amg))
6322 && SvRV(msv) == SvRV(sv))
6327 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6331 /* this is a partially unrolled
6332 * sv_catsv_nomg(pat, msv);
6333 * that allows us to adjust code block indices if
6336 char *dst = SvPV_force_nomg(pat, dlen);
6338 if (SvUTF8(msv) && !SvUTF8(pat)) {
6339 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6340 sv_setpvn(pat, dst, dlen);
6343 sv_catsv_nomg(pat, msv);
6347 /* We have only one SV to process, but we need to verify
6348 * it is properly null terminated or we will fail asserts
6349 * later. In theory we probably shouldn't get such SV's,
6350 * but if we do we should handle it gracefully. */
6351 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) ) {
6352 /* not a string, or a string with a trailing null */
6355 /* a string with no trailing null, we need to copy it
6356 * so it we have a trailing null */
6362 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
6365 /* extract any code blocks within any embedded qr//'s */
6366 if (rx && SvTYPE(rx) == SVt_REGEXP
6367 && RX_ENGINE((REGEXP*)rx)->op_comp)
6370 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6371 if (ri->num_code_blocks) {
6373 /* the presence of an embedded qr// with code means
6374 * we should always recompile: the text of the
6375 * qr// may not have changed, but it may be a
6376 * different closure than last time */
6378 Renew(pRExC_state->code_blocks,
6379 pRExC_state->num_code_blocks + ri->num_code_blocks,
6380 struct reg_code_block);
6381 pRExC_state->num_code_blocks += ri->num_code_blocks;
6383 for (i=0; i < ri->num_code_blocks; i++) {
6384 struct reg_code_block *src, *dst;
6385 STRLEN offset = orig_patlen
6386 + ReANY((REGEXP *)rx)->pre_prefix;
6387 assert(n < pRExC_state->num_code_blocks);
6388 src = &ri->code_blocks[i];
6389 dst = &pRExC_state->code_blocks[n];
6390 dst->start = src->start + offset;
6391 dst->end = src->end + offset;
6392 dst->block = src->block;
6393 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6402 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6411 /* see if there are any run-time code blocks in the pattern.
6412 * False positives are allowed */
6415 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6416 char *pat, STRLEN plen)
6421 PERL_UNUSED_CONTEXT;
6423 for (s = 0; s < plen; s++) {
6424 if (n < pRExC_state->num_code_blocks
6425 && s == pRExC_state->code_blocks[n].start)
6427 s = pRExC_state->code_blocks[n].end;
6431 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6433 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6435 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6442 /* Handle run-time code blocks. We will already have compiled any direct
6443 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6444 * copy of it, but with any literal code blocks blanked out and
6445 * appropriate chars escaped; then feed it into
6447 * eval "qr'modified_pattern'"
6451 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6455 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6457 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6458 * and merge them with any code blocks of the original regexp.
6460 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6461 * instead, just save the qr and return FALSE; this tells our caller that
6462 * the original pattern needs upgrading to utf8.
6466 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6467 char *pat, STRLEN plen)
6471 GET_RE_DEBUG_FLAGS_DECL;
6473 if (pRExC_state->runtime_code_qr) {
6474 /* this is the second time we've been called; this should
6475 * only happen if the main pattern got upgraded to utf8
6476 * during compilation; re-use the qr we compiled first time
6477 * round (which should be utf8 too)
6479 qr = pRExC_state->runtime_code_qr;
6480 pRExC_state->runtime_code_qr = NULL;
6481 assert(RExC_utf8 && SvUTF8(qr));
6487 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
6491 /* determine how many extra chars we need for ' and \ escaping */
6492 for (s = 0; s < plen; s++) {
6493 if (pat[s] == '\'' || pat[s] == '\\')
6497 Newx(newpat, newlen, char);
6499 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6501 for (s = 0; s < plen; s++) {
6502 if (n < pRExC_state->num_code_blocks
6503 && s == pRExC_state->code_blocks[n].start)
6505 /* blank out literal code block */
6506 assert(pat[s] == '(');
6507 while (s <= pRExC_state->code_blocks[n].end) {
6515 if (pat[s] == '\'' || pat[s] == '\\')
6520 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
6524 Perl_re_printf( aTHX_
6525 "%sre-parsing pattern for runtime code:%s %s\n",
6526 PL_colors[4],PL_colors[5],newpat);
6529 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6535 PUSHSTACKi(PERLSI_REQUIRE);
6536 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6537 * parsing qr''; normally only q'' does this. It also alters
6539 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6540 SvREFCNT_dec_NN(sv);
6545 SV * const errsv = ERRSV;
6546 if (SvTRUE_NN(errsv))
6548 Safefree(pRExC_state->code_blocks);
6549 /* use croak_sv ? */
6550 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
6553 assert(SvROK(qr_ref));
6555 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6556 /* the leaving below frees the tmp qr_ref.
6557 * Give qr a life of its own */
6565 if (!RExC_utf8 && SvUTF8(qr)) {
6566 /* first time through; the pattern got upgraded; save the
6567 * qr for the next time through */
6568 assert(!pRExC_state->runtime_code_qr);
6569 pRExC_state->runtime_code_qr = qr;
6574 /* extract any code blocks within the returned qr// */
6577 /* merge the main (r1) and run-time (r2) code blocks into one */
6579 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6580 struct reg_code_block *new_block, *dst;
6581 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6584 if (!r2->num_code_blocks) /* we guessed wrong */
6586 SvREFCNT_dec_NN(qr);
6591 r1->num_code_blocks + r2->num_code_blocks,
6592 struct reg_code_block);
6595 while ( i1 < r1->num_code_blocks
6596 || i2 < r2->num_code_blocks)
6598 struct reg_code_block *src;
6601 if (i1 == r1->num_code_blocks) {
6602 src = &r2->code_blocks[i2++];
6605 else if (i2 == r2->num_code_blocks)
6606 src = &r1->code_blocks[i1++];
6607 else if ( r1->code_blocks[i1].start
6608 < r2->code_blocks[i2].start)
6610 src = &r1->code_blocks[i1++];
6611 assert(src->end < r2->code_blocks[i2].start);
6614 assert( r1->code_blocks[i1].start
6615 > r2->code_blocks[i2].start);
6616 src = &r2->code_blocks[i2++];
6618 assert(src->end < r1->code_blocks[i1].start);
6621 assert(pat[src->start] == '(');
6622 assert(pat[src->end] == ')');
6623 dst->start = src->start;
6624 dst->end = src->end;
6625 dst->block = src->block;
6626 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6630 r1->num_code_blocks += r2->num_code_blocks;
6631 Safefree(r1->code_blocks);
6632 r1->code_blocks = new_block;
6635 SvREFCNT_dec_NN(qr);
6641 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6642 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6643 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6644 STRLEN longest_length, bool eol, bool meol)
6646 /* This is the common code for setting up the floating and fixed length
6647 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6648 * as to whether succeeded or not */
6653 if (! (longest_length
6654 || (eol /* Can't have SEOL and MULTI */
6655 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6657 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6658 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6663 /* copy the information about the longest from the reg_scan_data
6664 over to the program. */
6665 if (SvUTF8(sv_longest)) {
6666 *rx_utf8 = sv_longest;
6669 *rx_substr = sv_longest;
6672 /* end_shift is how many chars that must be matched that
6673 follow this item. We calculate it ahead of time as once the
6674 lookbehind offset is added in we lose the ability to correctly
6676 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6677 *rx_end_shift = ml - offset
6679 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
6681 + (SvTAIL(sv_longest) != 0)
6685 t = (eol/* Can't have SEOL and MULTI */
6686 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6687 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6693 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6694 * regular expression into internal code.
6695 * The pattern may be passed either as:
6696 * a list of SVs (patternp plus pat_count)
6697 * a list of OPs (expr)
6698 * If both are passed, the SV list is used, but the OP list indicates
6699 * which SVs are actually pre-compiled code blocks
6701 * The SVs in the list have magic and qr overloading applied to them (and
6702 * the list may be modified in-place with replacement SVs in the latter
6705 * If the pattern hasn't changed from old_re, then old_re will be
6708 * eng is the current engine. If that engine has an op_comp method, then
6709 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6710 * do the initial concatenation of arguments and pass on to the external
6713 * If is_bare_re is not null, set it to a boolean indicating whether the
6714 * arg list reduced (after overloading) to a single bare regex which has
6715 * been returned (i.e. /$qr/).
6717 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6719 * pm_flags contains the PMf_* flags, typically based on those from the
6720 * pm_flags field of the related PMOP. Currently we're only interested in
6721 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6723 * We can't allocate space until we know how big the compiled form will be,
6724 * but we can't compile it (and thus know how big it is) until we've got a
6725 * place to put the code. So we cheat: we compile it twice, once with code
6726 * generation turned off and size counting turned on, and once "for real".
6727 * This also means that we don't allocate space until we are sure that the
6728 * thing really will compile successfully, and we never have to move the
6729 * code and thus invalidate pointers into it. (Note that it has to be in
6730 * one piece because free() must be able to free it all.) [NB: not true in perl]
6732 * Beware that the optimization-preparation code in here knows about some
6733 * of the structure of the compiled regexp. [I'll say.]
6737 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6738 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6739 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6743 regexp_internal *ri;
6751 SV *code_blocksv = NULL;
6752 SV** new_patternp = patternp;
6754 /* these are all flags - maybe they should be turned
6755 * into a single int with different bit masks */
6756 I32 sawlookahead = 0;
6761 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6763 bool runtime_code = 0;
6765 RExC_state_t RExC_state;
6766 RExC_state_t * const pRExC_state = &RExC_state;
6767 #ifdef TRIE_STUDY_OPT
6769 RExC_state_t copyRExC_state;
6771 GET_RE_DEBUG_FLAGS_DECL;
6773 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6775 DEBUG_r(if (!PL_colorset) reginitcolors());
6777 /* Initialize these here instead of as-needed, as is quick and avoids
6778 * having to test them each time otherwise */
6779 if (! PL_AboveLatin1) {
6781 char * dump_len_string;
6784 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6785 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6786 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6787 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6788 PL_HasMultiCharFold =
6789 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6791 /* This is calculated here, because the Perl program that generates the
6792 * static global ones doesn't currently have access to
6793 * NUM_ANYOF_CODE_POINTS */
6794 PL_InBitmap = _new_invlist(2);
6795 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
6796 NUM_ANYOF_CODE_POINTS - 1);
6798 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
6799 if ( ! dump_len_string
6800 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
6802 PL_dump_re_max_len = 0;
6807 pRExC_state->warn_text = NULL;
6808 pRExC_state->code_blocks = NULL;
6809 pRExC_state->num_code_blocks = 0;
6812 *is_bare_re = FALSE;
6814 if (expr && (expr->op_type == OP_LIST ||
6815 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6816 /* allocate code_blocks if needed */
6820 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
6821 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6822 ncode++; /* count of DO blocks */
6824 pRExC_state->num_code_blocks = ncode;
6825 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
6830 /* compile-time pattern with just OP_CONSTs and DO blocks */
6835 /* find how many CONSTs there are */
6838 if (expr->op_type == OP_CONST)
6841 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6842 if (o->op_type == OP_CONST)
6846 /* fake up an SV array */
6848 assert(!new_patternp);
6849 Newx(new_patternp, n, SV*);
6850 SAVEFREEPV(new_patternp);
6854 if (expr->op_type == OP_CONST)
6855 new_patternp[n] = cSVOPx_sv(expr);
6857 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6858 if (o->op_type == OP_CONST)
6859 new_patternp[n++] = cSVOPo_sv;
6864 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6865 "Assembling pattern from %d elements%s\n", pat_count,
6866 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6868 /* set expr to the first arg op */
6870 if (pRExC_state->num_code_blocks
6871 && expr->op_type != OP_CONST)
6873 expr = cLISTOPx(expr)->op_first;
6874 assert( expr->op_type == OP_PUSHMARK
6875 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6876 || expr->op_type == OP_PADRANGE);
6877 expr = OpSIBLING(expr);
6880 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6881 expr, &recompile, NULL);
6883 /* handle bare (possibly after overloading) regex: foo =~ $re */
6888 if (SvTYPE(re) == SVt_REGEXP) {
6892 Safefree(pRExC_state->code_blocks);
6893 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6894 "Precompiled pattern%s\n",
6895 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6901 exp = SvPV_nomg(pat, plen);
6903 if (!eng->op_comp) {
6904 if ((SvUTF8(pat) && IN_BYTES)
6905 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6907 /* make a temporary copy; either to convert to bytes,
6908 * or to avoid repeating get-magic / overloaded stringify */
6909 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6910 (IN_BYTES ? 0 : SvUTF8(pat)));
6912 Safefree(pRExC_state->code_blocks);
6913 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6916 /* ignore the utf8ness if the pattern is 0 length */
6917 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6919 RExC_uni_semantics = 0;
6920 RExC_seen_unfolded_sharp_s = 0;
6921 RExC_contains_locale = 0;
6922 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
6923 RExC_study_started = 0;
6924 pRExC_state->runtime_code_qr = NULL;
6925 RExC_frame_head= NULL;
6926 RExC_frame_last= NULL;
6927 RExC_frame_count= 0;
6930 RExC_mysv1= sv_newmortal();
6931 RExC_mysv2= sv_newmortal();
6934 SV *dsv= sv_newmortal();
6935 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6936 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
6937 PL_colors[4],PL_colors[5],s);
6941 /* we jump here if we have to recompile, e.g., from upgrading the pattern
6944 if ((pm_flags & PMf_USE_RE_EVAL)
6945 /* this second condition covers the non-regex literal case,
6946 * i.e. $foo =~ '(?{})'. */
6947 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6949 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6951 /* return old regex if pattern hasn't changed */
6952 /* XXX: note in the below we have to check the flags as well as the
6955 * Things get a touch tricky as we have to compare the utf8 flag
6956 * independently from the compile flags. */
6960 && !!RX_UTF8(old_re) == !!RExC_utf8
6961 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
6962 && RX_PRECOMP(old_re)
6963 && RX_PRELEN(old_re) == plen
6964 && memEQ(RX_PRECOMP(old_re), exp, plen)
6965 && !runtime_code /* with runtime code, always recompile */ )
6967 Safefree(pRExC_state->code_blocks);
6971 rx_flags = orig_rx_flags;
6973 if ( initial_charset == REGEX_DEPENDS_CHARSET
6974 && (RExC_utf8 ||RExC_uni_semantics))
6977 /* Set to use unicode semantics if the pattern is in utf8 and has the
6978 * 'depends' charset specified, as it means unicode when utf8 */
6979 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6983 RExC_precomp_adj = 0;
6984 RExC_flags = rx_flags;
6985 RExC_pm_flags = pm_flags;
6988 assert(TAINTING_get || !TAINT_get);
6990 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
6992 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
6993 /* whoops, we have a non-utf8 pattern, whilst run-time code
6994 * got compiled as utf8. Try again with a utf8 pattern */
6995 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6996 pRExC_state->num_code_blocks);
6997 goto redo_first_pass;
7000 assert(!pRExC_state->runtime_code_qr);
7006 RExC_in_lookbehind = 0;
7007 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7009 RExC_override_recoding = 0;
7011 RExC_recode_x_to_native = 0;
7013 RExC_in_multi_char_class = 0;
7015 /* First pass: determine size, legality. */
7017 RExC_start = RExC_adjusted_start = exp;
7018 RExC_end = exp + plen;
7019 RExC_precomp_end = RExC_end;
7024 RExC_emit = (regnode *) &RExC_emit_dummy;
7025 RExC_whilem_seen = 0;
7026 RExC_open_parens = NULL;
7027 RExC_close_parens = NULL;
7029 RExC_paren_names = NULL;
7031 RExC_paren_name_list = NULL;
7033 RExC_recurse = NULL;
7034 RExC_study_chunk_recursed = NULL;
7035 RExC_study_chunk_recursed_bytes= 0;
7036 RExC_recurse_count = 0;
7037 pRExC_state->code_index = 0;
7039 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7040 * code makes sure the final byte is an uncounted NUL. But should this
7041 * ever not be the case, lots of things could read beyond the end of the
7042 * buffer: loops like
7043 * while(isFOO(*RExC_parse)) RExC_parse++;
7044 * strchr(RExC_parse, "foo");
7045 * etc. So it is worth noting. */
7046 assert(*RExC_end == '\0');
7049 Perl_re_printf( aTHX_ "Starting first pass (sizing)\n");
7051 RExC_lastparse=NULL;
7053 /* reg may croak on us, not giving us a chance to free
7054 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
7055 need it to survive as long as the regexp (qr/(?{})/).
7056 We must check that code_blocksv is not already set, because we may
7057 have jumped back to restart the sizing pass. */
7058 if (pRExC_state->code_blocks && !code_blocksv) {
7059 code_blocksv = newSV_type(SVt_PV);
7060 SAVEFREESV(code_blocksv);
7061 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
7062 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
7064 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7065 /* It's possible to write a regexp in ascii that represents Unicode
7066 codepoints outside of the byte range, such as via \x{100}. If we
7067 detect such a sequence we have to convert the entire pattern to utf8
7068 and then recompile, as our sizing calculation will have been based
7069 on 1 byte == 1 character, but we will need to use utf8 to encode
7070 at least some part of the pattern, and therefore must convert the whole
7073 if (flags & RESTART_PASS1) {
7074 if (flags & NEED_UTF8) {
7075 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7076 pRExC_state->num_code_blocks);
7079 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7080 "Need to redo pass 1\n"));
7083 goto redo_first_pass;
7085 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#" UVxf, (UV) flags);
7088 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
7091 Perl_re_printf( aTHX_
7092 "Required size %" IVdf " nodes\n"
7093 "Starting second pass (creation)\n",
7096 RExC_lastparse=NULL;
7099 /* The first pass could have found things that force Unicode semantics */
7100 if ((RExC_utf8 || RExC_uni_semantics)
7101 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
7103 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7106 /* Small enough for pointer-storage convention?
7107 If extralen==0, this means that we will not need long jumps. */
7108 if (RExC_size >= 0x10000L && RExC_extralen)
7109 RExC_size += RExC_extralen;
7112 if (RExC_whilem_seen > 15)
7113 RExC_whilem_seen = 15;
7115 /* Allocate space and zero-initialize. Note, the two step process
7116 of zeroing when in debug mode, thus anything assigned has to
7117 happen after that */
7118 rx = (REGEXP*) newSV_type(SVt_REGEXP);
7120 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7121 char, regexp_internal);
7122 if ( r == NULL || ri == NULL )
7123 FAIL("Regexp out of space");
7125 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
7126 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7129 /* bulk initialize base fields with 0. */
7130 Zero(ri, sizeof(regexp_internal), char);
7133 /* non-zero initialization begins here */
7136 r->extflags = rx_flags;
7137 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7139 if (pm_flags & PMf_IS_QR) {
7140 ri->code_blocks = pRExC_state->code_blocks;
7141 ri->num_code_blocks = pRExC_state->num_code_blocks;
7146 for (n = 0; n < pRExC_state->num_code_blocks; n++)
7147 if (pRExC_state->code_blocks[n].src_regex)
7148 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
7149 if(pRExC_state->code_blocks)
7150 SAVEFREEPV(pRExC_state->code_blocks); /* often null */
7154 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7155 bool has_charset = (get_regex_charset(r->extflags)
7156 != REGEX_DEPENDS_CHARSET);
7158 /* The caret is output if there are any defaults: if not all the STD
7159 * flags are set, or if no character set specifier is needed */
7161 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7163 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7164 == REG_RUN_ON_COMMENT_SEEN);
7165 U8 reganch = (U8)((r->extflags & RXf_PMf_STD_PMMOD)
7166 >> RXf_PMf_STD_PMMOD_SHIFT);
7167 const char *fptr = STD_PAT_MODS; /*"msixn"*/
7170 /* We output all the necessary flags; we never output a minus, as all
7171 * those are defaults, so are
7172 * covered by the caret */
7173 const STRLEN wraplen = plen + has_p + has_runon
7174 + has_default /* If needs a caret */
7175 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7177 /* If needs a character set specifier */
7178 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7179 + (sizeof("(?:)") - 1);
7181 /* make sure PL_bitcount bounds not exceeded */
7182 assert(sizeof(STD_PAT_MODS) <= 8);
7184 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
7185 r->xpv_len_u.xpvlenu_pv = p;
7187 SvFLAGS(rx) |= SVf_UTF8;
7190 /* If a default, cover it using the caret */
7192 *p++= DEFAULT_PAT_MOD;
7196 const char* const name = get_regex_charset_name(r->extflags, &len);
7197 Copy(name, p, len, char);
7201 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7204 while((ch = *fptr++)) {
7212 Copy(RExC_precomp, p, plen, char);
7213 assert ((RX_WRAPPED(rx) - p) < 16);
7214 r->pre_prefix = p - RX_WRAPPED(rx);
7220 SvCUR_set(rx, p - RX_WRAPPED(rx));
7224 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
7226 /* Useful during FAIL. */
7227 #ifdef RE_TRACK_PATTERN_OFFSETS
7228 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
7229 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7230 "%s %" UVuf " bytes for offset annotations.\n",
7231 ri->u.offsets ? "Got" : "Couldn't get",
7232 (UV)((2*RExC_size+1) * sizeof(U32))));
7234 SetProgLen(ri,RExC_size);
7239 /* Second pass: emit code. */
7240 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7241 RExC_pm_flags = pm_flags;
7243 RExC_end = exp + plen;
7245 RExC_emit_start = ri->program;
7246 RExC_emit = ri->program;
7247 RExC_emit_bound = ri->program + RExC_size + 1;
7248 pRExC_state->code_index = 0;
7250 *((char*) RExC_emit++) = (char) REG_MAGIC;
7251 /* setup various meta data about recursion, this all requires
7252 * RExC_npar to be correctly set, and a bit later on we clear it */
7253 if (RExC_seen & REG_RECURSE_SEEN) {
7254 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
7255 "%*s%*s Setting up open/close parens\n",
7256 22, "| |", (int)(0 * 2 + 1), ""));
7258 /* setup RExC_open_parens, which holds the address of each
7259 * OPEN tag, and to make things simpler for the 0 index
7260 * the start of the program - this is used later for offsets */
7261 Newxz(RExC_open_parens, RExC_npar,regnode *);
7262 SAVEFREEPV(RExC_open_parens);
7263 RExC_open_parens[0] = RExC_emit;
7265 /* setup RExC_close_parens, which holds the address of each
7266 * CLOSE tag, and to make things simpler for the 0 index
7267 * the end of the program - this is used later for offsets */
7268 Newxz(RExC_close_parens, RExC_npar,regnode *);
7269 SAVEFREEPV(RExC_close_parens);
7270 /* we dont know where end op starts yet, so we dont
7271 * need to set RExC_close_parens[0] like we do RExC_open_parens[0] above */
7273 /* Note, RExC_npar is 1 + the number of parens in a pattern.
7274 * So its 1 if there are no parens. */
7275 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
7276 ((RExC_npar & 0x07) != 0);
7277 Newx(RExC_study_chunk_recursed,
7278 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7279 SAVEFREEPV(RExC_study_chunk_recursed);
7282 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7284 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#" UVxf, (UV) flags);
7287 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7290 /* XXXX To minimize changes to RE engine we always allocate
7291 3-units-long substrs field. */
7292 Newx(r->substrs, 1, struct reg_substr_data);
7293 if (RExC_recurse_count) {
7294 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
7295 SAVEFREEPV(RExC_recurse);
7299 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7301 RExC_study_chunk_recursed_count= 0;
7303 Zero(r->substrs, 1, struct reg_substr_data);
7304 if (RExC_study_chunk_recursed) {
7305 Zero(RExC_study_chunk_recursed,
7306 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7310 #ifdef TRIE_STUDY_OPT
7312 StructCopy(&zero_scan_data, &data, scan_data_t);
7313 copyRExC_state = RExC_state;
7316 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
7318 RExC_state = copyRExC_state;
7319 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
7320 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
7322 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
7323 StructCopy(&zero_scan_data, &data, scan_data_t);
7326 StructCopy(&zero_scan_data, &data, scan_data_t);
7329 /* Dig out information for optimizations. */
7330 r->extflags = RExC_flags; /* was pm_op */
7331 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
7334 SvUTF8_on(rx); /* Unicode in it? */
7335 ri->regstclass = NULL;
7336 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
7337 r->intflags |= PREGf_NAUGHTY;
7338 scan = ri->program + 1; /* First BRANCH. */
7340 /* testing for BRANCH here tells us whether there is "must appear"
7341 data in the pattern. If there is then we can use it for optimisations */
7342 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7345 STRLEN longest_float_length, longest_fixed_length;
7346 regnode_ssc ch_class; /* pointed to by data */
7348 SSize_t last_close = 0; /* pointed to by data */
7349 regnode *first= scan;
7350 regnode *first_next= regnext(first);
7352 * Skip introductions and multiplicators >= 1
7353 * so that we can extract the 'meat' of the pattern that must
7354 * match in the large if() sequence following.
7355 * NOTE that EXACT is NOT covered here, as it is normally
7356 * picked up by the optimiser separately.
7358 * This is unfortunate as the optimiser isnt handling lookahead
7359 * properly currently.
7362 while ((OP(first) == OPEN && (sawopen = 1)) ||
7363 /* An OR of *one* alternative - should not happen now. */
7364 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7365 /* for now we can't handle lookbehind IFMATCH*/
7366 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7367 (OP(first) == PLUS) ||
7368 (OP(first) == MINMOD) ||
7369 /* An {n,m} with n>0 */
7370 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7371 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7374 * the only op that could be a regnode is PLUS, all the rest
7375 * will be regnode_1 or regnode_2.
7377 * (yves doesn't think this is true)
7379 if (OP(first) == PLUS)
7382 if (OP(first) == MINMOD)
7384 first += regarglen[OP(first)];
7386 first = NEXTOPER(first);
7387 first_next= regnext(first);
7390 /* Starting-point info. */
7392 DEBUG_PEEP("first:",first,0);
7393 /* Ignore EXACT as we deal with it later. */
7394 if (PL_regkind[OP(first)] == EXACT) {
7395 if (OP(first) == EXACT || OP(first) == EXACTL)
7396 NOOP; /* Empty, get anchored substr later. */
7398 ri->regstclass = first;
7401 else if (PL_regkind[OP(first)] == TRIE &&
7402 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
7404 /* this can happen only on restudy */
7405 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
7408 else if (REGNODE_SIMPLE(OP(first)))
7409 ri->regstclass = first;
7410 else if (PL_regkind[OP(first)] == BOUND ||
7411 PL_regkind[OP(first)] == NBOUND)
7412 ri->regstclass = first;
7413 else if (PL_regkind[OP(first)] == BOL) {
7414 r->intflags |= (OP(first) == MBOL
7417 first = NEXTOPER(first);
7420 else if (OP(first) == GPOS) {
7421 r->intflags |= PREGf_ANCH_GPOS;
7422 first = NEXTOPER(first);
7425 else if ((!sawopen || !RExC_sawback) &&
7427 (OP(first) == STAR &&
7428 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7429 !(r->intflags & PREGf_ANCH) && !pRExC_state->num_code_blocks)
7431 /* turn .* into ^.* with an implied $*=1 */
7433 (OP(NEXTOPER(first)) == REG_ANY)
7436 r->intflags |= (type | PREGf_IMPLICIT);
7437 first = NEXTOPER(first);
7440 if (sawplus && !sawminmod && !sawlookahead
7441 && (!sawopen || !RExC_sawback)
7442 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
7443 /* x+ must match at the 1st pos of run of x's */
7444 r->intflags |= PREGf_SKIP;
7446 /* Scan is after the zeroth branch, first is atomic matcher. */
7447 #ifdef TRIE_STUDY_OPT
7450 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7451 (IV)(first - scan + 1))
7455 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7456 (IV)(first - scan + 1))
7462 * If there's something expensive in the r.e., find the
7463 * longest literal string that must appear and make it the
7464 * regmust. Resolve ties in favor of later strings, since
7465 * the regstart check works with the beginning of the r.e.
7466 * and avoiding duplication strengthens checking. Not a
7467 * strong reason, but sufficient in the absence of others.
7468 * [Now we resolve ties in favor of the earlier string if
7469 * it happens that c_offset_min has been invalidated, since the
7470 * earlier string may buy us something the later one won't.]
7473 data.longest_fixed = newSVpvs("");
7474 data.longest_float = newSVpvs("");
7475 data.last_found = newSVpvs("");
7476 data.longest = &(data.longest_fixed);
7477 ENTER_with_name("study_chunk");
7478 SAVEFREESV(data.longest_fixed);
7479 SAVEFREESV(data.longest_float);
7480 SAVEFREESV(data.last_found);
7482 if (!ri->regstclass) {
7483 ssc_init(pRExC_state, &ch_class);
7484 data.start_class = &ch_class;
7485 stclass_flag = SCF_DO_STCLASS_AND;
7486 } else /* XXXX Check for BOUND? */
7488 data.last_closep = &last_close;
7491 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
7492 scan + RExC_size, /* Up to end */
7494 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
7495 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
7499 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
7502 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
7503 && data.last_start_min == 0 && data.last_end > 0
7504 && !RExC_seen_zerolen
7505 && !(RExC_seen & REG_VERBARG_SEEN)
7506 && !(RExC_seen & REG_GPOS_SEEN)
7508 r->extflags |= RXf_CHECK_ALL;
7510 scan_commit(pRExC_state, &data,&minlen,0);
7512 longest_float_length = CHR_SVLEN(data.longest_float);
7514 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
7515 && data.offset_fixed == data.offset_float_min
7516 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
7517 && S_setup_longest (aTHX_ pRExC_state,
7521 &(r->float_end_shift),
7522 data.lookbehind_float,
7523 data.offset_float_min,
7525 longest_float_length,
7526 cBOOL(data.flags & SF_FL_BEFORE_EOL),
7527 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
7529 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
7530 r->float_max_offset = data.offset_float_max;
7531 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
7532 r->float_max_offset -= data.lookbehind_float;
7533 SvREFCNT_inc_simple_void_NN(data.longest_float);
7536 r->float_substr = r->float_utf8 = NULL;
7537 longest_float_length = 0;
7540 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
7542 if (S_setup_longest (aTHX_ pRExC_state,
7544 &(r->anchored_utf8),
7545 &(r->anchored_substr),
7546 &(r->anchored_end_shift),
7547 data.lookbehind_fixed,
7550 longest_fixed_length,
7551 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
7552 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
7554 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
7555 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
7558 r->anchored_substr = r->anchored_utf8 = NULL;
7559 longest_fixed_length = 0;
7561 LEAVE_with_name("study_chunk");
7564 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
7565 ri->regstclass = NULL;
7567 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
7569 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7570 && is_ssc_worth_it(pRExC_state, data.start_class))
7572 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7574 ssc_finalize(pRExC_state, data.start_class);
7576 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7577 StructCopy(data.start_class,
7578 (regnode_ssc*)RExC_rxi->data->data[n],
7580 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7581 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7582 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7583 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7584 Perl_re_printf( aTHX_
7585 "synthetic stclass \"%s\".\n",
7586 SvPVX_const(sv));});
7587 data.start_class = NULL;
7590 /* A temporary algorithm prefers floated substr to fixed one to dig
7592 if (longest_fixed_length > longest_float_length) {
7593 r->substrs->check_ix = 0;
7594 r->check_end_shift = r->anchored_end_shift;
7595 r->check_substr = r->anchored_substr;
7596 r->check_utf8 = r->anchored_utf8;
7597 r->check_offset_min = r->check_offset_max = r->anchored_offset;
7598 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
7599 r->intflags |= PREGf_NOSCAN;
7602 r->substrs->check_ix = 1;
7603 r->check_end_shift = r->float_end_shift;
7604 r->check_substr = r->float_substr;
7605 r->check_utf8 = r->float_utf8;
7606 r->check_offset_min = r->float_min_offset;
7607 r->check_offset_max = r->float_max_offset;
7609 if ((r->check_substr || r->check_utf8) ) {
7610 r->extflags |= RXf_USE_INTUIT;
7611 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7612 r->extflags |= RXf_INTUIT_TAIL;
7614 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
7616 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7617 if ( (STRLEN)minlen < longest_float_length )
7618 minlen= longest_float_length;
7619 if ( (STRLEN)minlen < longest_fixed_length )
7620 minlen= longest_fixed_length;
7624 /* Several toplevels. Best we can is to set minlen. */
7626 regnode_ssc ch_class;
7627 SSize_t last_close = 0;
7629 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
7631 scan = ri->program + 1;
7632 ssc_init(pRExC_state, &ch_class);
7633 data.start_class = &ch_class;
7634 data.last_closep = &last_close;
7637 minlen = study_chunk(pRExC_state,
7638 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7639 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7640 ? SCF_TRIE_DOING_RESTUDY
7644 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7646 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7647 = r->float_substr = r->float_utf8 = NULL;
7649 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7650 && is_ssc_worth_it(pRExC_state, data.start_class))
7652 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7654 ssc_finalize(pRExC_state, data.start_class);
7656 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7657 StructCopy(data.start_class,
7658 (regnode_ssc*)RExC_rxi->data->data[n],
7660 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7661 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7662 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7663 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7664 Perl_re_printf( aTHX_
7665 "synthetic stclass \"%s\".\n",
7666 SvPVX_const(sv));});
7667 data.start_class = NULL;
7671 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7672 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7673 r->maxlen = REG_INFTY;
7676 r->maxlen = RExC_maxlen;
7679 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7680 the "real" pattern. */
7682 Perl_re_printf( aTHX_ "minlen: %" IVdf " r->minlen:%" IVdf " maxlen:%" IVdf "\n",
7683 (IV)minlen, (IV)r->minlen, (IV)RExC_maxlen);
7685 r->minlenret = minlen;
7686 if (r->minlen < minlen)
7689 if (RExC_seen & REG_RECURSE_SEEN ) {
7690 r->intflags |= PREGf_RECURSE_SEEN;
7691 Newxz(r->recurse_locinput, r->nparens + 1, char *);
7693 if (RExC_seen & REG_GPOS_SEEN)
7694 r->intflags |= PREGf_GPOS_SEEN;
7695 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7696 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7698 if (pRExC_state->num_code_blocks)
7699 r->extflags |= RXf_EVAL_SEEN;
7700 if (RExC_seen & REG_VERBARG_SEEN)
7702 r->intflags |= PREGf_VERBARG_SEEN;
7703 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7705 if (RExC_seen & REG_CUTGROUP_SEEN)
7706 r->intflags |= PREGf_CUTGROUP_SEEN;
7707 if (pm_flags & PMf_USE_RE_EVAL)
7708 r->intflags |= PREGf_USE_RE_EVAL;
7709 if (RExC_paren_names)
7710 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7712 RXp_PAREN_NAMES(r) = NULL;
7714 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7715 * so it can be used in pp.c */
7716 if (r->intflags & PREGf_ANCH)
7717 r->extflags |= RXf_IS_ANCHORED;
7721 /* this is used to identify "special" patterns that might result
7722 * in Perl NOT calling the regex engine and instead doing the match "itself",
7723 * particularly special cases in split//. By having the regex compiler
7724 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7725 * we avoid weird issues with equivalent patterns resulting in different behavior,
7726 * AND we allow non Perl engines to get the same optimizations by the setting the
7727 * flags appropriately - Yves */
7728 regnode *first = ri->program + 1;
7730 regnode *next = regnext(first);
7733 if (PL_regkind[fop] == NOTHING && nop == END)
7734 r->extflags |= RXf_NULL;
7735 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
7736 /* when fop is SBOL first->flags will be true only when it was
7737 * produced by parsing /\A/, and not when parsing /^/. This is
7738 * very important for the split code as there we want to
7739 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
7740 * See rt #122761 for more details. -- Yves */
7741 r->extflags |= RXf_START_ONLY;
7742 else if (fop == PLUS
7743 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7745 r->extflags |= RXf_WHITE;
7746 else if ( r->extflags & RXf_SPLIT
7747 && (fop == EXACT || fop == EXACTL)
7748 && STR_LEN(first) == 1
7749 && *(STRING(first)) == ' '
7751 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7755 if (RExC_contains_locale) {
7756 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7760 if (RExC_paren_names) {
7761 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7762 ri->data->data[ri->name_list_idx]
7763 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7766 ri->name_list_idx = 0;
7768 while ( RExC_recurse_count > 0 ) {
7769 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
7770 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - scan );
7773 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7774 /* assume we don't need to swap parens around before we match */
7776 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
7777 (unsigned long)RExC_study_chunk_recursed_count);
7781 Perl_re_printf( aTHX_ "Final program:\n");
7784 #ifdef RE_TRACK_PATTERN_OFFSETS
7785 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7786 const STRLEN len = ri->u.offsets[0];
7788 GET_RE_DEBUG_FLAGS_DECL;
7789 Perl_re_printf( aTHX_
7790 "Offsets: [%" UVuf "]\n\t", (UV)ri->u.offsets[0]);
7791 for (i = 1; i <= len; i++) {
7792 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7793 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7794 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7796 Perl_re_printf( aTHX_ "\n");
7801 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7802 * by setting the regexp SV to readonly-only instead. If the
7803 * pattern's been recompiled, the USEDness should remain. */
7804 if (old_re && SvREADONLY(old_re))
7812 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7815 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7817 PERL_UNUSED_ARG(value);
7819 if (flags & RXapif_FETCH) {
7820 return reg_named_buff_fetch(rx, key, flags);
7821 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7822 Perl_croak_no_modify();
7824 } else if (flags & RXapif_EXISTS) {
7825 return reg_named_buff_exists(rx, key, flags)
7828 } else if (flags & RXapif_REGNAMES) {
7829 return reg_named_buff_all(rx, flags);
7830 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7831 return reg_named_buff_scalar(rx, flags);
7833 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7839 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7842 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7843 PERL_UNUSED_ARG(lastkey);
7845 if (flags & RXapif_FIRSTKEY)
7846 return reg_named_buff_firstkey(rx, flags);
7847 else if (flags & RXapif_NEXTKEY)
7848 return reg_named_buff_nextkey(rx, flags);
7850 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7857 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7860 AV *retarray = NULL;
7862 struct regexp *const rx = ReANY(r);
7864 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7866 if (flags & RXapif_ALL)
7869 if (rx && RXp_PAREN_NAMES(rx)) {
7870 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7873 SV* sv_dat=HeVAL(he_str);
7874 I32 *nums=(I32*)SvPVX(sv_dat);
7875 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7876 if ((I32)(rx->nparens) >= nums[i]
7877 && rx->offs[nums[i]].start != -1
7878 && rx->offs[nums[i]].end != -1)
7881 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7886 ret = newSVsv(&PL_sv_undef);
7889 av_push(retarray, ret);
7892 return newRV_noinc(MUTABLE_SV(retarray));
7899 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7902 struct regexp *const rx = ReANY(r);
7904 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7906 if (rx && RXp_PAREN_NAMES(rx)) {
7907 if (flags & RXapif_ALL) {
7908 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7910 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7912 SvREFCNT_dec_NN(sv);
7924 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7926 struct regexp *const rx = ReANY(r);
7928 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7930 if ( rx && RXp_PAREN_NAMES(rx) ) {
7931 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7933 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7940 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7942 struct regexp *const rx = ReANY(r);
7943 GET_RE_DEBUG_FLAGS_DECL;
7945 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7947 if (rx && RXp_PAREN_NAMES(rx)) {
7948 HV *hv = RXp_PAREN_NAMES(rx);
7950 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7953 SV* sv_dat = HeVAL(temphe);
7954 I32 *nums = (I32*)SvPVX(sv_dat);
7955 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7956 if ((I32)(rx->lastparen) >= nums[i] &&
7957 rx->offs[nums[i]].start != -1 &&
7958 rx->offs[nums[i]].end != -1)
7964 if (parno || flags & RXapif_ALL) {
7965 return newSVhek(HeKEY_hek(temphe));
7973 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
7978 struct regexp *const rx = ReANY(r);
7980 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
7982 if (rx && RXp_PAREN_NAMES(rx)) {
7983 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
7984 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
7985 } else if (flags & RXapif_ONE) {
7986 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
7987 av = MUTABLE_AV(SvRV(ret));
7988 length = av_tindex(av);
7989 SvREFCNT_dec_NN(ret);
7990 return newSViv(length + 1);
7992 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
7997 return &PL_sv_undef;
8001 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8003 struct regexp *const rx = ReANY(r);
8006 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8008 if (rx && RXp_PAREN_NAMES(rx)) {
8009 HV *hv= RXp_PAREN_NAMES(rx);
8011 (void)hv_iterinit(hv);
8012 while ( (temphe = hv_iternext_flags(hv,0)) ) {
8015 SV* sv_dat = HeVAL(temphe);
8016 I32 *nums = (I32*)SvPVX(sv_dat);
8017 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8018 if ((I32)(rx->lastparen) >= nums[i] &&
8019 rx->offs[nums[i]].start != -1 &&
8020 rx->offs[nums[i]].end != -1)
8026 if (parno || flags & RXapif_ALL) {
8027 av_push(av, newSVhek(HeKEY_hek(temphe)));
8032 return newRV_noinc(MUTABLE_SV(av));
8036 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8039 struct regexp *const rx = ReANY(r);
8045 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8047 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8048 || n == RX_BUFF_IDX_CARET_FULLMATCH
8049 || n == RX_BUFF_IDX_CARET_POSTMATCH
8052 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8054 /* on something like
8057 * the KEEPCOPY is set on the PMOP rather than the regex */
8058 if (PL_curpm && r == PM_GETRE(PL_curpm))
8059 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8068 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8069 /* no need to distinguish between them any more */
8070 n = RX_BUFF_IDX_FULLMATCH;
8072 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8073 && rx->offs[0].start != -1)
8075 /* $`, ${^PREMATCH} */
8076 i = rx->offs[0].start;
8080 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8081 && rx->offs[0].end != -1)
8083 /* $', ${^POSTMATCH} */
8084 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8085 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8088 if ( 0 <= n && n <= (I32)rx->nparens &&
8089 (s1 = rx->offs[n].start) != -1 &&
8090 (t1 = rx->offs[n].end) != -1)
8092 /* $&, ${^MATCH}, $1 ... */
8094 s = rx->subbeg + s1 - rx->suboffset;
8099 assert(s >= rx->subbeg);
8100 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8102 #ifdef NO_TAINT_SUPPORT
8103 sv_setpvn(sv, s, i);
8105 const int oldtainted = TAINT_get;
8107 sv_setpvn(sv, s, i);
8108 TAINT_set(oldtainted);
8110 if (RXp_MATCH_UTF8(rx))
8115 if (RXp_MATCH_TAINTED(rx)) {
8116 if (SvTYPE(sv) >= SVt_PVMG) {
8117 MAGIC* const mg = SvMAGIC(sv);
8120 SvMAGIC_set(sv, mg->mg_moremagic);
8122 if ((mgt = SvMAGIC(sv))) {
8123 mg->mg_moremagic = mgt;
8124 SvMAGIC_set(sv, mg);
8141 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8142 SV const * const value)
8144 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8146 PERL_UNUSED_ARG(rx);
8147 PERL_UNUSED_ARG(paren);
8148 PERL_UNUSED_ARG(value);
8151 Perl_croak_no_modify();
8155 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8158 struct regexp *const rx = ReANY(r);
8162 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8164 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8165 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8166 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8169 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8171 /* on something like
8174 * the KEEPCOPY is set on the PMOP rather than the regex */
8175 if (PL_curpm && r == PM_GETRE(PL_curpm))
8176 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8182 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8184 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8185 case RX_BUFF_IDX_PREMATCH: /* $` */
8186 if (rx->offs[0].start != -1) {
8187 i = rx->offs[0].start;
8196 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8197 case RX_BUFF_IDX_POSTMATCH: /* $' */
8198 if (rx->offs[0].end != -1) {
8199 i = rx->sublen - rx->offs[0].end;
8201 s1 = rx->offs[0].end;
8208 default: /* $& / ${^MATCH}, $1, $2, ... */
8209 if (paren <= (I32)rx->nparens &&
8210 (s1 = rx->offs[paren].start) != -1 &&
8211 (t1 = rx->offs[paren].end) != -1)
8217 if (ckWARN(WARN_UNINITIALIZED))
8218 report_uninit((const SV *)sv);
8223 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8224 const char * const s = rx->subbeg - rx->suboffset + s1;
8229 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8236 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8238 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8239 PERL_UNUSED_ARG(rx);
8243 return newSVpvs("Regexp");
8246 /* Scans the name of a named buffer from the pattern.
8247 * If flags is REG_RSN_RETURN_NULL returns null.
8248 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8249 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8250 * to the parsed name as looked up in the RExC_paren_names hash.
8251 * If there is an error throws a vFAIL().. type exception.
8254 #define REG_RSN_RETURN_NULL 0
8255 #define REG_RSN_RETURN_NAME 1
8256 #define REG_RSN_RETURN_DATA 2
8259 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8261 char *name_start = RExC_parse;
8263 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8265 assert (RExC_parse <= RExC_end);
8266 if (RExC_parse == RExC_end) NOOP;
8267 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8268 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8269 * using do...while */
8272 RExC_parse += UTF8SKIP(RExC_parse);
8273 } while ( RExC_parse < RExC_end
8274 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8278 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8280 RExC_parse++; /* so the <- from the vFAIL is after the offending
8282 vFAIL("Group name must start with a non-digit word character");
8286 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8287 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8288 if ( flags == REG_RSN_RETURN_NAME)
8290 else if (flags==REG_RSN_RETURN_DATA) {
8293 if ( ! sv_name ) /* should not happen*/
8294 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8295 if (RExC_paren_names)
8296 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8298 sv_dat = HeVAL(he_str);
8300 vFAIL("Reference to nonexistent named group");
8304 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8305 (unsigned long) flags);
8307 NOT_REACHED; /* NOTREACHED */
8312 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8314 if (RExC_lastparse!=RExC_parse) { \
8315 Perl_re_printf( aTHX_ "%s", \
8316 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8317 RExC_end - RExC_parse, 16, \
8319 PERL_PV_ESCAPE_UNI_DETECT | \
8320 PERL_PV_PRETTY_ELLIPSES | \
8321 PERL_PV_PRETTY_LTGT | \
8322 PERL_PV_ESCAPE_RE | \
8323 PERL_PV_PRETTY_EXACTSIZE \
8327 Perl_re_printf( aTHX_ "%16s",""); \
8330 num = RExC_size + 1; \
8332 num=REG_NODE_NUM(RExC_emit); \
8333 if (RExC_lastnum!=num) \
8334 Perl_re_printf( aTHX_ "|%4d",num); \
8336 Perl_re_printf( aTHX_ "|%4s",""); \
8337 Perl_re_printf( aTHX_ "|%*s%-4s", \
8338 (int)((depth*2)), "", \
8342 RExC_lastparse=RExC_parse; \
8347 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8348 DEBUG_PARSE_MSG((funcname)); \
8349 Perl_re_printf( aTHX_ "%4s","\n"); \
8351 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8352 DEBUG_PARSE_MSG((funcname)); \
8353 Perl_re_printf( aTHX_ fmt "\n",args); \
8356 /* This section of code defines the inversion list object and its methods. The
8357 * interfaces are highly subject to change, so as much as possible is static to
8358 * this file. An inversion list is here implemented as a malloc'd C UV array
8359 * as an SVt_INVLIST scalar.
8361 * An inversion list for Unicode is an array of code points, sorted by ordinal
8362 * number. Each element gives the code point that begins a range that extends
8363 * up-to but not including the code point given by the next element. The final
8364 * element gives the first code point of a range that extends to the platform's
8365 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8366 * ...) give ranges whose code points are all in the inversion list. We say
8367 * that those ranges are in the set. The odd-numbered elements give ranges
8368 * whose code points are not in the inversion list, and hence not in the set.
8369 * Thus, element [0] is the first code point in the list. Element [1]
8370 * is the first code point beyond that not in the list; and element [2] is the
8371 * first code point beyond that that is in the list. In other words, the first
8372 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8373 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8374 * all code points in that range are not in the inversion list. The third
8375 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8376 * list, and so forth. Thus every element whose index is divisible by two
8377 * gives the beginning of a range that is in the list, and every element whose
8378 * index is not divisible by two gives the beginning of a range not in the
8379 * list. If the final element's index is divisible by two, the inversion list
8380 * extends to the platform's infinity; otherwise the highest code point in the
8381 * inversion list is the contents of that element minus 1.
8383 * A range that contains just a single code point N will look like
8385 * invlist[i+1] == N+1
8387 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8388 * impossible to represent, so element [i+1] is omitted. The single element
8390 * invlist[0] == UV_MAX
8391 * contains just UV_MAX, but is interpreted as matching to infinity.
8393 * Taking the complement (inverting) an inversion list is quite simple, if the
8394 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8395 * This implementation reserves an element at the beginning of each inversion
8396 * list to always contain 0; there is an additional flag in the header which
8397 * indicates if the list begins at the 0, or is offset to begin at the next
8398 * element. This means that the inversion list can be inverted without any
8399 * copying; just flip the flag.
8401 * More about inversion lists can be found in "Unicode Demystified"
8402 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8404 * The inversion list data structure is currently implemented as an SV pointing
8405 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8406 * array of UV whose memory management is automatically handled by the existing
8407 * facilities for SV's.
8409 * Some of the methods should always be private to the implementation, and some
8410 * should eventually be made public */
8412 /* The header definitions are in F<invlist_inline.h> */
8414 #ifndef PERL_IN_XSUB_RE
8416 PERL_STATIC_INLINE UV*
8417 S__invlist_array_init(SV* const invlist, const bool will_have_0)
8419 /* Returns a pointer to the first element in the inversion list's array.
8420 * This is called upon initialization of an inversion list. Where the
8421 * array begins depends on whether the list has the code point U+0000 in it
8422 * or not. The other parameter tells it whether the code that follows this
8423 * call is about to put a 0 in the inversion list or not. The first
8424 * element is either the element reserved for 0, if TRUE, or the element
8425 * after it, if FALSE */
8427 bool* offset = get_invlist_offset_addr(invlist);
8428 UV* zero_addr = (UV *) SvPVX(invlist);
8430 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8433 assert(! _invlist_len(invlist));
8437 /* 1^1 = 0; 1^0 = 1 */
8438 *offset = 1 ^ will_have_0;
8439 return zero_addr + *offset;
8444 PERL_STATIC_INLINE void
8445 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8447 /* Sets the current number of elements stored in the inversion list.
8448 * Updates SvCUR correspondingly */
8449 PERL_UNUSED_CONTEXT;
8450 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8452 assert(SvTYPE(invlist) == SVt_INVLIST);
8457 : TO_INTERNAL_SIZE(len + offset));
8458 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8461 #ifndef PERL_IN_XSUB_RE
8464 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
8466 /* Replaces the inversion list in 'dest' with the one from 'src'. It
8467 * steals the list from 'src', so 'src' is made to have a NULL list. This
8468 * is similar to what SvSetMagicSV() would do, if it were implemented on
8469 * inversion lists, though this routine avoids a copy */
8471 const UV src_len = _invlist_len(src);
8472 const bool src_offset = *get_invlist_offset_addr(src);
8473 const STRLEN src_byte_len = SvLEN(src);
8474 char * array = SvPVX(src);
8476 const int oldtainted = TAINT_get;
8478 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
8480 assert(SvTYPE(src) == SVt_INVLIST);
8481 assert(SvTYPE(dest) == SVt_INVLIST);
8482 assert(! invlist_is_iterating(src));
8483 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
8485 /* Make sure it ends in the right place with a NUL, as our inversion list
8486 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
8488 array[src_byte_len - 1] = '\0';
8490 TAINT_NOT; /* Otherwise it breaks */
8491 sv_usepvn_flags(dest,
8495 /* This flag is documented to cause a copy to be avoided */
8496 SV_HAS_TRAILING_NUL);
8497 TAINT_set(oldtainted);
8502 /* Finish up copying over the other fields in an inversion list */
8503 *get_invlist_offset_addr(dest) = src_offset;
8504 invlist_set_len(dest, src_len, src_offset);
8505 *get_invlist_previous_index_addr(dest) = 0;
8506 invlist_iterfinish(dest);
8509 PERL_STATIC_INLINE IV*
8510 S_get_invlist_previous_index_addr(SV* invlist)
8512 /* Return the address of the IV that is reserved to hold the cached index
8514 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8516 assert(SvTYPE(invlist) == SVt_INVLIST);
8518 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8521 PERL_STATIC_INLINE IV
8522 S_invlist_previous_index(SV* const invlist)
8524 /* Returns cached index of previous search */
8526 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8528 return *get_invlist_previous_index_addr(invlist);
8531 PERL_STATIC_INLINE void
8532 S_invlist_set_previous_index(SV* const invlist, const IV index)
8534 /* Caches <index> for later retrieval */
8536 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8538 assert(index == 0 || index < (int) _invlist_len(invlist));
8540 *get_invlist_previous_index_addr(invlist) = index;
8543 PERL_STATIC_INLINE void
8544 S_invlist_trim(SV* invlist)
8546 /* Free the not currently-being-used space in an inversion list */
8548 /* But don't free up the space needed for the 0 UV that is always at the
8549 * beginning of the list, nor the trailing NUL */
8550 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
8552 PERL_ARGS_ASSERT_INVLIST_TRIM;
8554 assert(SvTYPE(invlist) == SVt_INVLIST);
8556 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
8559 PERL_STATIC_INLINE void
8560 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
8562 PERL_ARGS_ASSERT_INVLIST_CLEAR;
8564 assert(SvTYPE(invlist) == SVt_INVLIST);
8566 invlist_set_len(invlist, 0, 0);
8567 invlist_trim(invlist);
8570 #endif /* ifndef PERL_IN_XSUB_RE */
8572 PERL_STATIC_INLINE bool
8573 S_invlist_is_iterating(SV* const invlist)
8575 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8577 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8580 #ifndef PERL_IN_XSUB_RE
8582 PERL_STATIC_INLINE UV
8583 S_invlist_max(SV* const invlist)
8585 /* Returns the maximum number of elements storable in the inversion list's
8586 * array, without having to realloc() */
8588 PERL_ARGS_ASSERT_INVLIST_MAX;
8590 assert(SvTYPE(invlist) == SVt_INVLIST);
8592 /* Assumes worst case, in which the 0 element is not counted in the
8593 * inversion list, so subtracts 1 for that */
8594 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8595 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8596 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8599 Perl__new_invlist(pTHX_ IV initial_size)
8602 /* Return a pointer to a newly constructed inversion list, with enough
8603 * space to store 'initial_size' elements. If that number is negative, a
8604 * system default is used instead */
8608 if (initial_size < 0) {
8612 /* Allocate the initial space */
8613 new_list = newSV_type(SVt_INVLIST);
8615 /* First 1 is in case the zero element isn't in the list; second 1 is for
8617 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8618 invlist_set_len(new_list, 0, 0);
8620 /* Force iterinit() to be used to get iteration to work */
8621 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8623 *get_invlist_previous_index_addr(new_list) = 0;
8629 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8631 /* Return a pointer to a newly constructed inversion list, initialized to
8632 * point to <list>, which has to be in the exact correct inversion list
8633 * form, including internal fields. Thus this is a dangerous routine that
8634 * should not be used in the wrong hands. The passed in 'list' contains
8635 * several header fields at the beginning that are not part of the
8636 * inversion list body proper */
8638 const STRLEN length = (STRLEN) list[0];
8639 const UV version_id = list[1];
8640 const bool offset = cBOOL(list[2]);
8641 #define HEADER_LENGTH 3
8642 /* If any of the above changes in any way, you must change HEADER_LENGTH
8643 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8644 * perl -E 'say int(rand 2**31-1)'
8646 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8647 data structure type, so that one being
8648 passed in can be validated to be an
8649 inversion list of the correct vintage.
8652 SV* invlist = newSV_type(SVt_INVLIST);
8654 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8656 if (version_id != INVLIST_VERSION_ID) {
8657 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8660 /* The generated array passed in includes header elements that aren't part
8661 * of the list proper, so start it just after them */
8662 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8664 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8665 shouldn't touch it */
8667 *(get_invlist_offset_addr(invlist)) = offset;
8669 /* The 'length' passed to us is the physical number of elements in the
8670 * inversion list. But if there is an offset the logical number is one
8672 invlist_set_len(invlist, length - offset, offset);
8674 invlist_set_previous_index(invlist, 0);
8676 /* Initialize the iteration pointer. */
8677 invlist_iterfinish(invlist);
8679 SvREADONLY_on(invlist);
8685 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8687 /* Grow the maximum size of an inversion list */
8689 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8691 assert(SvTYPE(invlist) == SVt_INVLIST);
8693 /* Add one to account for the zero element at the beginning which may not
8694 * be counted by the calling parameters */
8695 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8699 S__append_range_to_invlist(pTHX_ SV* const invlist,
8700 const UV start, const UV end)
8702 /* Subject to change or removal. Append the range from 'start' to 'end' at
8703 * the end of the inversion list. The range must be above any existing
8707 UV max = invlist_max(invlist);
8708 UV len = _invlist_len(invlist);
8711 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8713 if (len == 0) { /* Empty lists must be initialized */
8714 offset = start != 0;
8715 array = _invlist_array_init(invlist, ! offset);
8718 /* Here, the existing list is non-empty. The current max entry in the
8719 * list is generally the first value not in the set, except when the
8720 * set extends to the end of permissible values, in which case it is
8721 * the first entry in that final set, and so this call is an attempt to
8722 * append out-of-order */
8724 UV final_element = len - 1;
8725 array = invlist_array(invlist);
8726 if ( array[final_element] > start
8727 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8729 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",
8730 array[final_element], start,
8731 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8734 /* Here, it is a legal append. If the new range begins 1 above the end
8735 * of the range below it, it is extending the range below it, so the
8736 * new first value not in the set is one greater than the newly
8737 * extended range. */
8738 offset = *get_invlist_offset_addr(invlist);
8739 if (array[final_element] == start) {
8740 if (end != UV_MAX) {
8741 array[final_element] = end + 1;
8744 /* But if the end is the maximum representable on the machine,
8745 * assume that infinity was actually what was meant. Just let
8746 * the range that this would extend to have no end */
8747 invlist_set_len(invlist, len - 1, offset);
8753 /* Here the new range doesn't extend any existing set. Add it */
8755 len += 2; /* Includes an element each for the start and end of range */
8757 /* If wll overflow the existing space, extend, which may cause the array to
8760 invlist_extend(invlist, len);
8762 /* Have to set len here to avoid assert failure in invlist_array() */
8763 invlist_set_len(invlist, len, offset);
8765 array = invlist_array(invlist);
8768 invlist_set_len(invlist, len, offset);
8771 /* The next item on the list starts the range, the one after that is
8772 * one past the new range. */
8773 array[len - 2] = start;
8774 if (end != UV_MAX) {
8775 array[len - 1] = end + 1;
8778 /* But if the end is the maximum representable on the machine, just let
8779 * the range have no end */
8780 invlist_set_len(invlist, len - 1, offset);
8785 Perl__invlist_search(SV* const invlist, const UV cp)
8787 /* Searches the inversion list for the entry that contains the input code
8788 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8789 * return value is the index into the list's array of the range that
8790 * contains <cp>, that is, 'i' such that
8791 * array[i] <= cp < array[i+1]
8796 IV high = _invlist_len(invlist);
8797 const IV highest_element = high - 1;
8800 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8802 /* If list is empty, return failure. */
8807 /* (We can't get the array unless we know the list is non-empty) */
8808 array = invlist_array(invlist);
8810 mid = invlist_previous_index(invlist);
8812 if (mid > highest_element) {
8813 mid = highest_element;
8816 /* <mid> contains the cache of the result of the previous call to this
8817 * function (0 the first time). See if this call is for the same result,
8818 * or if it is for mid-1. This is under the theory that calls to this
8819 * function will often be for related code points that are near each other.
8820 * And benchmarks show that caching gives better results. We also test
8821 * here if the code point is within the bounds of the list. These tests
8822 * replace others that would have had to be made anyway to make sure that
8823 * the array bounds were not exceeded, and these give us extra information
8824 * at the same time */
8825 if (cp >= array[mid]) {
8826 if (cp >= array[highest_element]) {
8827 return highest_element;
8830 /* Here, array[mid] <= cp < array[highest_element]. This means that
8831 * the final element is not the answer, so can exclude it; it also
8832 * means that <mid> is not the final element, so can refer to 'mid + 1'
8834 if (cp < array[mid + 1]) {
8840 else { /* cp < aray[mid] */
8841 if (cp < array[0]) { /* Fail if outside the array */
8845 if (cp >= array[mid - 1]) {
8850 /* Binary search. What we are looking for is <i> such that
8851 * array[i] <= cp < array[i+1]
8852 * The loop below converges on the i+1. Note that there may not be an
8853 * (i+1)th element in the array, and things work nonetheless */
8854 while (low < high) {
8855 mid = (low + high) / 2;
8856 assert(mid <= highest_element);
8857 if (array[mid] <= cp) { /* cp >= array[mid] */
8860 /* We could do this extra test to exit the loop early.
8861 if (cp < array[low]) {
8866 else { /* cp < array[mid] */
8873 invlist_set_previous_index(invlist, high);
8878 Perl__invlist_populate_swatch(SV* const invlist,
8879 const UV start, const UV end, U8* swatch)
8881 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8882 * but is used when the swash has an inversion list. This makes this much
8883 * faster, as it uses a binary search instead of a linear one. This is
8884 * intimately tied to that function, and perhaps should be in utf8.c,
8885 * except it is intimately tied to inversion lists as well. It assumes
8886 * that <swatch> is all 0's on input */
8889 const IV len = _invlist_len(invlist);
8893 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8895 if (len == 0) { /* Empty inversion list */
8899 array = invlist_array(invlist);
8901 /* Find which element it is */
8902 i = _invlist_search(invlist, start);
8904 /* We populate from <start> to <end> */
8905 while (current < end) {
8908 /* The inversion list gives the results for every possible code point
8909 * after the first one in the list. Only those ranges whose index is
8910 * even are ones that the inversion list matches. For the odd ones,
8911 * and if the initial code point is not in the list, we have to skip
8912 * forward to the next element */
8913 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8915 if (i >= len) { /* Finished if beyond the end of the array */
8919 if (current >= end) { /* Finished if beyond the end of what we
8921 if (LIKELY(end < UV_MAX)) {
8925 /* We get here when the upper bound is the maximum
8926 * representable on the machine, and we are looking for just
8927 * that code point. Have to special case it */
8929 goto join_end_of_list;
8932 assert(current >= start);
8934 /* The current range ends one below the next one, except don't go past
8937 upper = (i < len && array[i] < end) ? array[i] : end;
8939 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8940 * for each code point in it */
8941 for (; current < upper; current++) {
8942 const STRLEN offset = (STRLEN)(current - start);
8943 swatch[offset >> 3] |= 1 << (offset & 7);
8948 /* Quit if at the end of the list */
8951 /* But first, have to deal with the highest possible code point on
8952 * the platform. The previous code assumes that <end> is one
8953 * beyond where we want to populate, but that is impossible at the
8954 * platform's infinity, so have to handle it specially */
8955 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8957 const STRLEN offset = (STRLEN)(end - start);
8958 swatch[offset >> 3] |= 1 << (offset & 7);
8963 /* Advance to the next range, which will be for code points not in the
8972 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8973 const bool complement_b, SV** output)
8975 /* Take the union of two inversion lists and point '*output' to it. On
8976 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
8977 * even 'a' or 'b'). If to an inversion list, the contents of the original
8978 * list will be replaced by the union. The first list, 'a', may be
8979 * NULL, in which case a copy of the second list is placed in '*output'.
8980 * If 'complement_b' is TRUE, the union is taken of the complement
8981 * (inversion) of 'b' instead of b itself.
8983 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8984 * Richard Gillam, published by Addison-Wesley, and explained at some
8985 * length there. The preface says to incorporate its examples into your
8986 * code at your own risk.
8988 * The algorithm is like a merge sort. */
8990 const UV* array_a; /* a's array */
8992 UV len_a; /* length of a's array */
8995 SV* u; /* the resulting union */
8999 UV i_a = 0; /* current index into a's array */
9003 /* running count, as explained in the algorithm source book; items are
9004 * stopped accumulating and are output when the count changes to/from 0.
9005 * The count is incremented when we start a range that's in an input's set,
9006 * and decremented when we start a range that's not in a set. So this
9007 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9008 * and hence nothing goes into the union; 1, just one of the inputs is in
9009 * its set (and its current range gets added to the union); and 2 when both
9010 * inputs are in their sets. */
9013 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9015 assert(*output == NULL || SvTYPE(*output) == SVt_INVLIST);
9017 len_b = _invlist_len(b);
9020 /* Here, 'b' is empty, hence it's complement is all possible code
9021 * points. So if the union includes the complement of 'b', it includes
9022 * everything, and we need not even look at 'a'. It's easiest to
9023 * create a new inversion list that matches everything. */
9025 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9027 if (*output == NULL) { /* If the output didn't exist, just point it
9029 *output = everything;
9031 else { /* Otherwise, replace its contents with the new list */
9032 invlist_replace_list_destroys_src(*output, everything);
9033 SvREFCNT_dec_NN(everything);
9039 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9040 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9041 * output will be empty */
9043 if (a == NULL || _invlist_len(a) == 0) {
9044 if (*output == NULL) {
9045 *output = _new_invlist(0);
9048 invlist_clear(*output);
9053 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9054 * union. We can just return a copy of 'a' if '*output' doesn't point
9055 * to an existing list */
9056 if (*output == NULL) {
9057 *output = invlist_clone(a);
9061 /* If the output is to overwrite 'a', we have a no-op, as it's
9067 /* Here, '*output' is to be overwritten by 'a' */
9068 u = invlist_clone(a);
9069 invlist_replace_list_destroys_src(*output, u);
9075 /* Here 'b' is not empty. See about 'a' */
9077 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9079 /* Here, 'a' is empty (and b is not). That means the union will come
9080 * entirely from 'b'. If '*output' is NULL, we can directly return a
9081 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9084 SV ** dest = (*output == NULL) ? output : &u;
9085 *dest = invlist_clone(b);
9087 _invlist_invert(*dest);
9091 invlist_replace_list_destroys_src(*output, u);
9098 /* Here both lists exist and are non-empty */
9099 array_a = invlist_array(a);
9100 array_b = invlist_array(b);
9102 /* If are to take the union of 'a' with the complement of b, set it
9103 * up so are looking at b's complement. */
9106 /* To complement, we invert: if the first element is 0, remove it. To
9107 * do this, we just pretend the array starts one later */
9108 if (array_b[0] == 0) {
9114 /* But if the first element is not zero, we pretend the list starts
9115 * at the 0 that is always stored immediately before the array. */
9121 /* Size the union for the worst case: that the sets are completely
9123 u = _new_invlist(len_a + len_b);
9125 /* Will contain U+0000 if either component does */
9126 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9127 || (len_b > 0 && array_b[0] == 0));
9129 /* Go through each input list item by item, stopping when have exhausted
9131 while (i_a < len_a && i_b < len_b) {
9132 UV cp; /* The element to potentially add to the union's array */
9133 bool cp_in_set; /* is it in the the input list's set or not */
9135 /* We need to take one or the other of the two inputs for the union.
9136 * Since we are merging two sorted lists, we take the smaller of the
9137 * next items. In case of a tie, we take first the one that is in its
9138 * set. If we first took the one not in its set, it would decrement
9139 * the count, possibly to 0 which would cause it to be output as ending
9140 * the range, and the next time through we would take the same number,
9141 * and output it again as beginning the next range. By doing it the
9142 * opposite way, there is no possibility that the count will be
9143 * momentarily decremented to 0, and thus the two adjoining ranges will
9144 * be seamlessly merged. (In a tie and both are in the set or both not
9145 * in the set, it doesn't matter which we take first.) */
9146 if ( array_a[i_a] < array_b[i_b]
9147 || ( array_a[i_a] == array_b[i_b]
9148 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9150 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9151 cp = array_a[i_a++];
9154 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9155 cp = array_b[i_b++];
9158 /* Here, have chosen which of the two inputs to look at. Only output
9159 * if the running count changes to/from 0, which marks the
9160 * beginning/end of a range that's in the set */
9163 array_u[i_u++] = cp;
9170 array_u[i_u++] = cp;
9176 /* The loop above increments the index into exactly one of the input lists
9177 * each iteration, and ends when either index gets to its list end. That
9178 * means the other index is lower than its end, and so something is
9179 * remaining in that one. We decrement 'count', as explained below, if
9180 * that list is in its set. (i_a and i_b each currently index the element
9181 * beyond the one we care about.) */
9182 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9183 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9188 /* Above we decremented 'count' if the list that had unexamined elements in
9189 * it was in its set. This has made it so that 'count' being non-zero
9190 * means there isn't anything left to output; and 'count' equal to 0 means
9191 * that what is left to output is precisely that which is left in the
9192 * non-exhausted input list.
9194 * To see why, note first that the exhausted input obviously has nothing
9195 * left to add to the union. If it was in its set at its end, that means
9196 * the set extends from here to the platform's infinity, and hence so does
9197 * the union and the non-exhausted set is irrelevant. The exhausted set
9198 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9199 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9200 * 'count' remains at 1. This is consistent with the decremented 'count'
9201 * != 0 meaning there's nothing left to add to the union.
9203 * But if the exhausted input wasn't in its set, it contributed 0 to
9204 * 'count', and the rest of the union will be whatever the other input is.
9205 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9206 * otherwise it gets decremented to 0. This is consistent with 'count'
9207 * == 0 meaning the remainder of the union is whatever is left in the
9208 * non-exhausted list. */
9213 IV copy_count = len_a - i_a;
9214 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9215 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9217 else { /* The non-exhausted input is b */
9218 copy_count = len_b - i_b;
9219 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9221 len_u = i_u + copy_count;
9224 /* Set the result to the final length, which can change the pointer to
9225 * array_u, so re-find it. (Note that it is unlikely that this will
9226 * change, as we are shrinking the space, not enlarging it) */
9227 if (len_u != _invlist_len(u)) {
9228 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9230 array_u = invlist_array(u);
9233 if (*output == NULL) { /* Simply return the new inversion list */
9237 /* Otherwise, overwrite the inversion list that was in '*output'. We
9238 * could instead free '*output', and then set it to 'u', but experience
9239 * has shown [perl #127392] that if the input is a mortal, we can get a
9240 * huge build-up of these during regex compilation before they get
9242 invlist_replace_list_destroys_src(*output, u);
9250 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9251 const bool complement_b, SV** i)
9253 /* Take the intersection of two inversion lists and point '*i' to it. On
9254 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9255 * even 'a' or 'b'). If to an inversion list, the contents of the original
9256 * list will be replaced by the intersection. The first list, 'a', may be
9257 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9258 * TRUE, the result will be the intersection of 'a' and the complement (or
9259 * inversion) of 'b' instead of 'b' directly.
9261 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9262 * Richard Gillam, published by Addison-Wesley, and explained at some
9263 * length there. The preface says to incorporate its examples into your
9264 * code at your own risk. In fact, it had bugs
9266 * The algorithm is like a merge sort, and is essentially the same as the
9270 const UV* array_a; /* a's array */
9272 UV len_a; /* length of a's array */
9275 SV* r; /* the resulting intersection */
9279 UV i_a = 0; /* current index into a's array */
9283 /* running count of how many of the two inputs are postitioned at ranges
9284 * that are in their sets. As explained in the algorithm source book,
9285 * items are stopped accumulating and are output when the count changes
9286 * to/from 2. The count is incremented when we start a range that's in an
9287 * input's set, and decremented when we start a range that's not in a set.
9288 * Only when it is 2 are we in the intersection. */
9291 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9293 assert(*i == NULL || SvTYPE(*i) == SVt_INVLIST);
9295 /* Special case if either one is empty */
9296 len_a = (a == NULL) ? 0 : _invlist_len(a);
9297 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9298 if (len_a != 0 && complement_b) {
9300 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9301 * must be empty. Here, also we are using 'b's complement, which
9302 * hence must be every possible code point. Thus the intersection
9305 if (*i == a) { /* No-op */
9310 *i = invlist_clone(a);
9314 r = invlist_clone(a);
9315 invlist_replace_list_destroys_src(*i, r);
9320 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9321 * intersection must be empty */
9323 *i = _new_invlist(0);
9331 /* Here both lists exist and are non-empty */
9332 array_a = invlist_array(a);
9333 array_b = invlist_array(b);
9335 /* If are to take the intersection of 'a' with the complement of b, set it
9336 * up so are looking at b's complement. */
9339 /* To complement, we invert: if the first element is 0, remove it. To
9340 * do this, we just pretend the array starts one later */
9341 if (array_b[0] == 0) {
9347 /* But if the first element is not zero, we pretend the list starts
9348 * at the 0 that is always stored immediately before the array. */
9354 /* Size the intersection for the worst case: that the intersection ends up
9355 * fragmenting everything to be completely disjoint */
9356 r= _new_invlist(len_a + len_b);
9358 /* Will contain U+0000 iff both components do */
9359 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9360 && len_b > 0 && array_b[0] == 0);
9362 /* Go through each list item by item, stopping when have exhausted one of
9364 while (i_a < len_a && i_b < len_b) {
9365 UV cp; /* The element to potentially add to the intersection's
9367 bool cp_in_set; /* Is it in the input list's set or not */
9369 /* We need to take one or the other of the two inputs for the
9370 * intersection. Since we are merging two sorted lists, we take the
9371 * smaller of the next items. In case of a tie, we take first the one
9372 * that is not in its set (a difference from the union algorithm). If
9373 * we first took the one in its set, it would increment the count,
9374 * possibly to 2 which would cause it to be output as starting a range
9375 * in the intersection, and the next time through we would take that
9376 * same number, and output it again as ending the set. By doing the
9377 * opposite of this, there is no possibility that the count will be
9378 * momentarily incremented to 2. (In a tie and both are in the set or
9379 * both not in the set, it doesn't matter which we take first.) */
9380 if ( array_a[i_a] < array_b[i_b]
9381 || ( array_a[i_a] == array_b[i_b]
9382 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9384 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9385 cp = array_a[i_a++];
9388 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9392 /* Here, have chosen which of the two inputs to look at. Only output
9393 * if the running count changes to/from 2, which marks the
9394 * beginning/end of a range that's in the intersection */
9398 array_r[i_r++] = cp;
9403 array_r[i_r++] = cp;
9410 /* The loop above increments the index into exactly one of the input lists
9411 * each iteration, and ends when either index gets to its list end. That
9412 * means the other index is lower than its end, and so something is
9413 * remaining in that one. We increment 'count', as explained below, if the
9414 * exhausted list was in its set. (i_a and i_b each currently index the
9415 * element beyond the one we care about.) */
9416 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9417 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9422 /* Above we incremented 'count' if the exhausted list was in its set. This
9423 * has made it so that 'count' being below 2 means there is nothing left to
9424 * output; otheriwse what's left to add to the intersection is precisely
9425 * that which is left in the non-exhausted input list.
9427 * To see why, note first that the exhausted input obviously has nothing
9428 * left to affect the intersection. If it was in its set at its end, that
9429 * means the set extends from here to the platform's infinity, and hence
9430 * anything in the non-exhausted's list will be in the intersection, and
9431 * anything not in it won't be. Hence, the rest of the intersection is
9432 * precisely what's in the non-exhausted list The exhausted set also
9433 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9434 * it means 'count' is now at least 2. This is consistent with the
9435 * incremented 'count' being >= 2 means to add the non-exhausted list to
9438 * But if the exhausted input wasn't in its set, it contributed 0 to
9439 * 'count', and the intersection can't include anything further; the
9440 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9441 * incremented. This is consistent with 'count' being < 2 meaning nothing
9442 * further to add to the intersection. */
9443 if (count < 2) { /* Nothing left to put in the intersection. */
9446 else { /* copy the non-exhausted list, unchanged. */
9447 IV copy_count = len_a - i_a;
9448 if (copy_count > 0) { /* a is the one with stuff left */
9449 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9451 else { /* b is the one with stuff left */
9452 copy_count = len_b - i_b;
9453 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9455 len_r = i_r + copy_count;
9458 /* Set the result to the final length, which can change the pointer to
9459 * array_r, so re-find it. (Note that it is unlikely that this will
9460 * change, as we are shrinking the space, not enlarging it) */
9461 if (len_r != _invlist_len(r)) {
9462 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9464 array_r = invlist_array(r);
9467 if (*i == NULL) { /* Simply return the calculated intersection */
9470 else { /* Otherwise, replace the existing inversion list in '*i'. We could
9471 instead free '*i', and then set it to 'r', but experience has
9472 shown [perl #127392] that if the input is a mortal, we can get a
9473 huge build-up of these during regex compilation before they get
9476 invlist_replace_list_destroys_src(*i, r);
9488 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9490 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9491 * set. A pointer to the inversion list is returned. This may actually be
9492 * a new list, in which case the passed in one has been destroyed. The
9493 * passed-in inversion list can be NULL, in which case a new one is created
9494 * with just the one range in it. The new list is not necessarily
9495 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9496 * result of this function. The gain would not be large, and in many
9497 * cases, this is called multiple times on a single inversion list, so
9498 * anything freed may almost immediately be needed again.
9500 * This used to mostly call the 'union' routine, but that is much more
9501 * heavyweight than really needed for a single range addition */
9503 UV* array; /* The array implementing the inversion list */
9504 UV len; /* How many elements in 'array' */
9505 SSize_t i_s; /* index into the invlist array where 'start'
9507 SSize_t i_e = 0; /* And the index where 'end' should go */
9508 UV cur_highest; /* The highest code point in the inversion list
9509 upon entry to this function */
9511 /* This range becomes the whole inversion list if none already existed */
9512 if (invlist == NULL) {
9513 invlist = _new_invlist(2);
9514 _append_range_to_invlist(invlist, start, end);
9518 /* Likewise, if the inversion list is currently empty */
9519 len = _invlist_len(invlist);
9521 _append_range_to_invlist(invlist, start, end);
9525 /* Starting here, we have to know the internals of the list */
9526 array = invlist_array(invlist);
9528 /* If the new range ends higher than the current highest ... */
9529 cur_highest = invlist_highest(invlist);
9530 if (end > cur_highest) {
9532 /* If the whole range is higher, we can just append it */
9533 if (start > cur_highest) {
9534 _append_range_to_invlist(invlist, start, end);
9538 /* Otherwise, add the portion that is higher ... */
9539 _append_range_to_invlist(invlist, cur_highest + 1, end);
9541 /* ... and continue on below to handle the rest. As a result of the
9542 * above append, we know that the index of the end of the range is the
9543 * final even numbered one of the array. Recall that the final element
9544 * always starts a range that extends to infinity. If that range is in
9545 * the set (meaning the set goes from here to infinity), it will be an
9546 * even index, but if it isn't in the set, it's odd, and the final
9547 * range in the set is one less, which is even. */
9548 if (end == UV_MAX) {
9556 /* We have dealt with appending, now see about prepending. If the new
9557 * range starts lower than the current lowest ... */
9558 if (start < array[0]) {
9560 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
9561 * Let the union code handle it, rather than having to know the
9562 * trickiness in two code places. */
9563 if (UNLIKELY(start == 0)) {
9566 range_invlist = _new_invlist(2);
9567 _append_range_to_invlist(range_invlist, start, end);
9569 _invlist_union(invlist, range_invlist, &invlist);
9571 SvREFCNT_dec_NN(range_invlist);
9576 /* If the whole new range comes before the first entry, and doesn't
9577 * extend it, we have to insert it as an additional range */
9578 if (end < array[0] - 1) {
9580 goto splice_in_new_range;
9583 /* Here the new range adjoins the existing first range, extending it
9587 /* And continue on below to handle the rest. We know that the index of
9588 * the beginning of the range is the first one of the array */
9591 else { /* Not prepending any part of the new range to the existing list.
9592 * Find where in the list it should go. This finds i_s, such that:
9593 * invlist[i_s] <= start < array[i_s+1]
9595 i_s = _invlist_search(invlist, start);
9598 /* At this point, any extending before the beginning of the inversion list
9599 * and/or after the end has been done. This has made it so that, in the
9600 * code below, each endpoint of the new range is either in a range that is
9601 * in the set, or is in a gap between two ranges that are. This means we
9602 * don't have to worry about exceeding the array bounds.
9604 * Find where in the list the new range ends (but we can skip this if we
9605 * have already determined what it is, or if it will be the same as i_s,
9606 * which we already have computed) */
9608 i_e = (start == end)
9610 : _invlist_search(invlist, end);
9613 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
9614 * is a range that goes to infinity there is no element at invlist[i_e+1],
9615 * so only the first relation holds. */
9617 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9619 /* Here, the ranges on either side of the beginning of the new range
9620 * are in the set, and this range starts in the gap between them.
9622 * The new range extends the range above it downwards if the new range
9623 * ends at or above that range's start */
9624 const bool extends_the_range_above = ( end == UV_MAX
9625 || end + 1 >= array[i_s+1]);
9627 /* The new range extends the range below it upwards if it begins just
9628 * after where that range ends */
9629 if (start == array[i_s]) {
9631 /* If the new range fills the entire gap between the other ranges,
9632 * they will get merged together. Other ranges may also get
9633 * merged, depending on how many of them the new range spans. In
9634 * the general case, we do the merge later, just once, after we
9635 * figure out how many to merge. But in the case where the new
9636 * range exactly spans just this one gap (possibly extending into
9637 * the one above), we do the merge here, and an early exit. This
9638 * is done here to avoid having to special case later. */
9639 if (i_e - i_s <= 1) {
9641 /* If i_e - i_s == 1, it means that the new range terminates
9642 * within the range above, and hence 'extends_the_range_above'
9643 * must be true. (If the range above it extends to infinity,
9644 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
9645 * will be 0, so no harm done.) */
9646 if (extends_the_range_above) {
9647 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
9648 invlist_set_len(invlist,
9650 *(get_invlist_offset_addr(invlist)));
9654 /* Here, i_e must == i_s. We keep them in sync, as they apply
9655 * to the same range, and below we are about to decrement i_s
9660 /* Here, the new range is adjacent to the one below. (It may also
9661 * span beyond the range above, but that will get resolved later.)
9662 * Extend the range below to include this one. */
9663 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
9667 else if (extends_the_range_above) {
9669 /* Here the new range only extends the range above it, but not the
9670 * one below. It merges with the one above. Again, we keep i_e
9671 * and i_s in sync if they point to the same range */
9680 /* Here, we've dealt with the new range start extending any adjoining
9683 * If the new range extends to infinity, it is now the final one,
9684 * regardless of what was there before */
9685 if (UNLIKELY(end == UV_MAX)) {
9686 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
9690 /* If i_e started as == i_s, it has also been dealt with,
9691 * and been updated to the new i_s, which will fail the following if */
9692 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
9694 /* Here, the ranges on either side of the end of the new range are in
9695 * the set, and this range ends in the gap between them.
9697 * If this range is adjacent to (hence extends) the range above it, it
9698 * becomes part of that range; likewise if it extends the range below,
9699 * it becomes part of that range */
9700 if (end + 1 == array[i_e+1]) {
9704 else if (start <= array[i_e]) {
9705 array[i_e] = end + 1;
9712 /* If the range fits entirely in an existing range (as possibly already
9713 * extended above), it doesn't add anything new */
9714 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9718 /* Here, no part of the range is in the list. Must add it. It will
9719 * occupy 2 more slots */
9720 splice_in_new_range:
9722 invlist_extend(invlist, len + 2);
9723 array = invlist_array(invlist);
9724 /* Move the rest of the array down two slots. Don't include any
9726 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
9728 /* Do the actual splice */
9729 array[i_e+1] = start;
9730 array[i_e+2] = end + 1;
9731 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
9735 /* Here the new range crossed the boundaries of a pre-existing range. The
9736 * code above has adjusted things so that both ends are in ranges that are
9737 * in the set. This means everything in between must also be in the set.
9738 * Just squash things together */
9739 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
9740 invlist_set_len(invlist,
9742 *(get_invlist_offset_addr(invlist)));
9748 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9749 UV** other_elements_ptr)
9751 /* Create and return an inversion list whose contents are to be populated
9752 * by the caller. The caller gives the number of elements (in 'size') and
9753 * the very first element ('element0'). This function will set
9754 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9757 * Obviously there is some trust involved that the caller will properly
9758 * fill in the other elements of the array.
9760 * (The first element needs to be passed in, as the underlying code does
9761 * things differently depending on whether it is zero or non-zero) */
9763 SV* invlist = _new_invlist(size);
9766 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9768 invlist = add_cp_to_invlist(invlist, element0);
9769 offset = *get_invlist_offset_addr(invlist);
9771 invlist_set_len(invlist, size, offset);
9772 *other_elements_ptr = invlist_array(invlist) + 1;
9778 PERL_STATIC_INLINE SV*
9779 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9780 return _add_range_to_invlist(invlist, cp, cp);
9783 #ifndef PERL_IN_XSUB_RE
9785 Perl__invlist_invert(pTHX_ SV* const invlist)
9787 /* Complement the input inversion list. This adds a 0 if the list didn't
9788 * have a zero; removes it otherwise. As described above, the data
9789 * structure is set up so that this is very efficient */
9791 PERL_ARGS_ASSERT__INVLIST_INVERT;
9793 assert(! invlist_is_iterating(invlist));
9795 /* The inverse of matching nothing is matching everything */
9796 if (_invlist_len(invlist) == 0) {
9797 _append_range_to_invlist(invlist, 0, UV_MAX);
9801 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9806 PERL_STATIC_INLINE SV*
9807 S_invlist_clone(pTHX_ SV* const invlist)
9810 /* Return a new inversion list that is a copy of the input one, which is
9811 * unchanged. The new list will not be mortal even if the old one was. */
9813 /* Need to allocate extra space to accommodate Perl's addition of a
9814 * trailing NUL to SvPV's, since it thinks they are always strings */
9815 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9816 STRLEN physical_length = SvCUR(invlist);
9817 bool offset = *(get_invlist_offset_addr(invlist));
9819 PERL_ARGS_ASSERT_INVLIST_CLONE;
9821 *(get_invlist_offset_addr(new_invlist)) = offset;
9822 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9823 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9828 PERL_STATIC_INLINE STRLEN*
9829 S_get_invlist_iter_addr(SV* invlist)
9831 /* Return the address of the UV that contains the current iteration
9834 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9836 assert(SvTYPE(invlist) == SVt_INVLIST);
9838 return &(((XINVLIST*) SvANY(invlist))->iterator);
9841 PERL_STATIC_INLINE void
9842 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9844 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9846 *get_invlist_iter_addr(invlist) = 0;
9849 PERL_STATIC_INLINE void
9850 S_invlist_iterfinish(SV* invlist)
9852 /* Terminate iterator for invlist. This is to catch development errors.
9853 * Any iteration that is interrupted before completed should call this
9854 * function. Functions that add code points anywhere else but to the end
9855 * of an inversion list assert that they are not in the middle of an
9856 * iteration. If they were, the addition would make the iteration
9857 * problematical: if the iteration hadn't reached the place where things
9858 * were being added, it would be ok */
9860 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9862 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9866 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9868 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9869 * This call sets in <*start> and <*end>, the next range in <invlist>.
9870 * Returns <TRUE> if successful and the next call will return the next
9871 * range; <FALSE> if was already at the end of the list. If the latter,
9872 * <*start> and <*end> are unchanged, and the next call to this function
9873 * will start over at the beginning of the list */
9875 STRLEN* pos = get_invlist_iter_addr(invlist);
9876 UV len = _invlist_len(invlist);
9879 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9882 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9886 array = invlist_array(invlist);
9888 *start = array[(*pos)++];
9894 *end = array[(*pos)++] - 1;
9900 PERL_STATIC_INLINE UV
9901 S_invlist_highest(SV* const invlist)
9903 /* Returns the highest code point that matches an inversion list. This API
9904 * has an ambiguity, as it returns 0 under either the highest is actually
9905 * 0, or if the list is empty. If this distinction matters to you, check
9906 * for emptiness before calling this function */
9908 UV len = _invlist_len(invlist);
9911 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9917 array = invlist_array(invlist);
9919 /* The last element in the array in the inversion list always starts a
9920 * range that goes to infinity. That range may be for code points that are
9921 * matched in the inversion list, or it may be for ones that aren't
9922 * matched. In the latter case, the highest code point in the set is one
9923 * less than the beginning of this range; otherwise it is the final element
9924 * of this range: infinity */
9925 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
9927 : array[len - 1] - 1;
9931 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
9933 /* Get the contents of an inversion list into a string SV so that they can
9934 * be printed out. If 'traditional_style' is TRUE, it uses the format
9935 * traditionally done for debug tracing; otherwise it uses a format
9936 * suitable for just copying to the output, with blanks between ranges and
9937 * a dash between range components */
9941 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
9942 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
9944 if (traditional_style) {
9945 output = newSVpvs("\n");
9948 output = newSVpvs("");
9951 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
9953 assert(! invlist_is_iterating(invlist));
9955 invlist_iterinit(invlist);
9956 while (invlist_iternext(invlist, &start, &end)) {
9957 if (end == UV_MAX) {
9958 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFINITY%c",
9959 start, intra_range_delimiter,
9960 inter_range_delimiter);
9962 else if (end != start) {
9963 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
9965 intra_range_delimiter,
9966 end, inter_range_delimiter);
9969 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
9970 start, inter_range_delimiter);
9974 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
9975 SvCUR_set(output, SvCUR(output) - 1);
9981 #ifndef PERL_IN_XSUB_RE
9983 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
9984 const char * const indent, SV* const invlist)
9986 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
9987 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
9988 * the string 'indent'. The output looks like this:
9989 [0] 0x000A .. 0x000D
9991 [4] 0x2028 .. 0x2029
9992 [6] 0x3104 .. INFINITY
9993 * This means that the first range of code points matched by the list are
9994 * 0xA through 0xD; the second range contains only the single code point
9995 * 0x85, etc. An inversion list is an array of UVs. Two array elements
9996 * are used to define each range (except if the final range extends to
9997 * infinity, only a single element is needed). The array index of the
9998 * first element for the corresponding range is given in brackets. */
10003 PERL_ARGS_ASSERT__INVLIST_DUMP;
10005 if (invlist_is_iterating(invlist)) {
10006 Perl_dump_indent(aTHX_ level, file,
10007 "%sCan't dump inversion list because is in middle of iterating\n",
10012 invlist_iterinit(invlist);
10013 while (invlist_iternext(invlist, &start, &end)) {
10014 if (end == UV_MAX) {
10015 Perl_dump_indent(aTHX_ level, file,
10016 "%s[%" UVuf "] 0x%04" UVXf " .. INFINITY\n",
10017 indent, (UV)count, start);
10019 else if (end != start) {
10020 Perl_dump_indent(aTHX_ level, file,
10021 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10022 indent, (UV)count, start, end);
10025 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10026 indent, (UV)count, start);
10033 Perl__load_PL_utf8_foldclosures (pTHX)
10035 assert(! PL_utf8_foldclosures);
10037 /* If the folds haven't been read in, call a fold function
10039 if (! PL_utf8_tofold) {
10040 U8 dummy[UTF8_MAXBYTES_CASE+1];
10041 const U8 hyphen[] = HYPHEN_UTF8;
10043 /* This string is just a short named one above \xff */
10044 toFOLD_utf8_safe(hyphen, hyphen + sizeof(hyphen) - 1, dummy, NULL);
10045 assert(PL_utf8_tofold); /* Verify that worked */
10047 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
10051 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10053 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10055 /* Return a boolean as to if the two passed in inversion lists are
10056 * identical. The final argument, if TRUE, says to take the complement of
10057 * the second inversion list before doing the comparison */
10059 const UV* array_a = invlist_array(a);
10060 const UV* array_b = invlist_array(b);
10061 UV len_a = _invlist_len(a);
10062 UV len_b = _invlist_len(b);
10064 PERL_ARGS_ASSERT__INVLISTEQ;
10066 /* If are to compare 'a' with the complement of b, set it
10067 * up so are looking at b's complement. */
10068 if (complement_b) {
10070 /* The complement of nothing is everything, so <a> would have to have
10071 * just one element, starting at zero (ending at infinity) */
10073 return (len_a == 1 && array_a[0] == 0);
10075 else if (array_b[0] == 0) {
10077 /* Otherwise, to complement, we invert. Here, the first element is
10078 * 0, just remove it. To do this, we just pretend the array starts
10086 /* But if the first element is not zero, we pretend the list starts
10087 * at the 0 that is always stored immediately before the array. */
10093 return len_a == len_b
10094 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10100 * As best we can, determine the characters that can match the start of
10101 * the given EXACTF-ish node.
10103 * Returns the invlist as a new SV*; it is the caller's responsibility to
10104 * call SvREFCNT_dec() when done with it.
10107 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10109 const U8 * s = (U8*)STRING(node);
10110 SSize_t bytelen = STR_LEN(node);
10112 /* Start out big enough for 2 separate code points */
10113 SV* invlist = _new_invlist(4);
10115 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
10120 /* We punt and assume can match anything if the node begins
10121 * with a multi-character fold. Things are complicated. For
10122 * example, /ffi/i could match any of:
10123 * "\N{LATIN SMALL LIGATURE FFI}"
10124 * "\N{LATIN SMALL LIGATURE FF}I"
10125 * "F\N{LATIN SMALL LIGATURE FI}"
10126 * plus several other things; and making sure we have all the
10127 * possibilities is hard. */
10128 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10129 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10132 /* Any Latin1 range character can potentially match any
10133 * other depending on the locale */
10134 if (OP(node) == EXACTFL) {
10135 _invlist_union(invlist, PL_Latin1, &invlist);
10138 /* But otherwise, it matches at least itself. We can
10139 * quickly tell if it has a distinct fold, and if so,
10140 * it matches that as well */
10141 invlist = add_cp_to_invlist(invlist, uc);
10142 if (IS_IN_SOME_FOLD_L1(uc))
10143 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10146 /* Some characters match above-Latin1 ones under /i. This
10147 * is true of EXACTFL ones when the locale is UTF-8 */
10148 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10149 && (! isASCII(uc) || (OP(node) != EXACTFA
10150 && OP(node) != EXACTFA_NO_TRIE)))
10152 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10156 else { /* Pattern is UTF-8 */
10157 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10158 STRLEN foldlen = UTF8SKIP(s);
10159 const U8* e = s + bytelen;
10162 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10164 /* The only code points that aren't folded in a UTF EXACTFish
10165 * node are are the problematic ones in EXACTFL nodes */
10166 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10167 /* We need to check for the possibility that this EXACTFL
10168 * node begins with a multi-char fold. Therefore we fold
10169 * the first few characters of it so that we can make that
10174 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10176 *(d++) = (U8) toFOLD(*s);
10181 toFOLD_utf8_safe(s, e, d, &len);
10187 /* And set up so the code below that looks in this folded
10188 * buffer instead of the node's string */
10190 foldlen = UTF8SKIP(folded);
10194 /* When we reach here 's' points to the fold of the first
10195 * character(s) of the node; and 'e' points to far enough along
10196 * the folded string to be just past any possible multi-char
10197 * fold. 'foldlen' is the length in bytes of the first
10200 * Unlike the non-UTF-8 case, the macro for determining if a
10201 * string is a multi-char fold requires all the characters to
10202 * already be folded. This is because of all the complications
10203 * if not. Note that they are folded anyway, except in EXACTFL
10204 * nodes. Like the non-UTF case above, we punt if the node
10205 * begins with a multi-char fold */
10207 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10208 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10210 else { /* Single char fold */
10212 /* It matches all the things that fold to it, which are
10213 * found in PL_utf8_foldclosures (including itself) */
10214 invlist = add_cp_to_invlist(invlist, uc);
10215 if (! PL_utf8_foldclosures)
10216 _load_PL_utf8_foldclosures();
10217 if ((listp = hv_fetch(PL_utf8_foldclosures,
10218 (char *) s, foldlen, FALSE)))
10220 AV* list = (AV*) *listp;
10222 for (k = 0; k <= av_tindex_nomg(list); k++) {
10223 SV** c_p = av_fetch(list, k, FALSE);
10229 /* /aa doesn't allow folds between ASCII and non- */
10230 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
10231 && isASCII(c) != isASCII(uc))
10236 invlist = add_cp_to_invlist(invlist, c);
10245 #undef HEADER_LENGTH
10246 #undef TO_INTERNAL_SIZE
10247 #undef FROM_INTERNAL_SIZE
10248 #undef INVLIST_VERSION_ID
10250 /* End of inversion list object */
10253 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10255 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10256 * constructs, and updates RExC_flags with them. On input, RExC_parse
10257 * should point to the first flag; it is updated on output to point to the
10258 * final ')' or ':'. There needs to be at least one flag, or this will
10261 /* for (?g), (?gc), and (?o) warnings; warning
10262 about (?c) will warn about (?g) -- japhy */
10264 #define WASTED_O 0x01
10265 #define WASTED_G 0x02
10266 #define WASTED_C 0x04
10267 #define WASTED_GC (WASTED_G|WASTED_C)
10268 I32 wastedflags = 0x00;
10269 U32 posflags = 0, negflags = 0;
10270 U32 *flagsp = &posflags;
10271 char has_charset_modifier = '\0';
10273 bool has_use_defaults = FALSE;
10274 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10275 int x_mod_count = 0;
10277 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10279 /* '^' as an initial flag sets certain defaults */
10280 if (UCHARAT(RExC_parse) == '^') {
10282 has_use_defaults = TRUE;
10283 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10284 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
10285 ? REGEX_UNICODE_CHARSET
10286 : REGEX_DEPENDS_CHARSET);
10289 cs = get_regex_charset(RExC_flags);
10290 if (cs == REGEX_DEPENDS_CHARSET
10291 && (RExC_utf8 || RExC_uni_semantics))
10293 cs = REGEX_UNICODE_CHARSET;
10296 while (RExC_parse < RExC_end) {
10297 /* && strchr("iogcmsx", *RExC_parse) */
10298 /* (?g), (?gc) and (?o) are useless here
10299 and must be globally applied -- japhy */
10300 switch (*RExC_parse) {
10302 /* Code for the imsxn flags */
10303 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10305 case LOCALE_PAT_MOD:
10306 if (has_charset_modifier) {
10307 goto excess_modifier;
10309 else if (flagsp == &negflags) {
10312 cs = REGEX_LOCALE_CHARSET;
10313 has_charset_modifier = LOCALE_PAT_MOD;
10315 case UNICODE_PAT_MOD:
10316 if (has_charset_modifier) {
10317 goto excess_modifier;
10319 else if (flagsp == &negflags) {
10322 cs = REGEX_UNICODE_CHARSET;
10323 has_charset_modifier = UNICODE_PAT_MOD;
10325 case ASCII_RESTRICT_PAT_MOD:
10326 if (flagsp == &negflags) {
10329 if (has_charset_modifier) {
10330 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10331 goto excess_modifier;
10333 /* Doubled modifier implies more restricted */
10334 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10337 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10339 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10341 case DEPENDS_PAT_MOD:
10342 if (has_use_defaults) {
10343 goto fail_modifiers;
10345 else if (flagsp == &negflags) {
10348 else if (has_charset_modifier) {
10349 goto excess_modifier;
10352 /* The dual charset means unicode semantics if the
10353 * pattern (or target, not known until runtime) are
10354 * utf8, or something in the pattern indicates unicode
10356 cs = (RExC_utf8 || RExC_uni_semantics)
10357 ? REGEX_UNICODE_CHARSET
10358 : REGEX_DEPENDS_CHARSET;
10359 has_charset_modifier = DEPENDS_PAT_MOD;
10363 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10364 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10366 else if (has_charset_modifier == *(RExC_parse - 1)) {
10367 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10368 *(RExC_parse - 1));
10371 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10373 NOT_REACHED; /*NOTREACHED*/
10376 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10377 *(RExC_parse - 1));
10378 NOT_REACHED; /*NOTREACHED*/
10379 case ONCE_PAT_MOD: /* 'o' */
10380 case GLOBAL_PAT_MOD: /* 'g' */
10381 if (PASS2 && ckWARN(WARN_REGEXP)) {
10382 const I32 wflagbit = *RExC_parse == 'o'
10385 if (! (wastedflags & wflagbit) ) {
10386 wastedflags |= wflagbit;
10387 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10390 "Useless (%s%c) - %suse /%c modifier",
10391 flagsp == &negflags ? "?-" : "?",
10393 flagsp == &negflags ? "don't " : "",
10400 case CONTINUE_PAT_MOD: /* 'c' */
10401 if (PASS2 && ckWARN(WARN_REGEXP)) {
10402 if (! (wastedflags & WASTED_C) ) {
10403 wastedflags |= WASTED_GC;
10404 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10407 "Useless (%sc) - %suse /gc modifier",
10408 flagsp == &negflags ? "?-" : "?",
10409 flagsp == &negflags ? "don't " : ""
10414 case KEEPCOPY_PAT_MOD: /* 'p' */
10415 if (flagsp == &negflags) {
10417 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10419 *flagsp |= RXf_PMf_KEEPCOPY;
10423 /* A flag is a default iff it is following a minus, so
10424 * if there is a minus, it means will be trying to
10425 * re-specify a default which is an error */
10426 if (has_use_defaults || flagsp == &negflags) {
10427 goto fail_modifiers;
10429 flagsp = &negflags;
10430 wastedflags = 0; /* reset so (?g-c) warns twice */
10434 RExC_flags |= posflags;
10435 RExC_flags &= ~negflags;
10436 set_regex_charset(&RExC_flags, cs);
10438 if (UNLIKELY((x_mod_count) > 1)) {
10439 vFAIL("Only one /x regex modifier is allowed");
10445 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10446 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10447 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
10448 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10449 NOT_REACHED; /*NOTREACHED*/
10452 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10455 vFAIL("Sequence (?... not terminated");
10459 - reg - regular expression, i.e. main body or parenthesized thing
10461 * Caller must absorb opening parenthesis.
10463 * Combining parenthesis handling with the base level of regular expression
10464 * is a trifle forced, but the need to tie the tails of the branches to what
10465 * follows makes it hard to avoid.
10467 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10469 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10471 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10474 PERL_STATIC_INLINE regnode *
10475 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10477 char * parse_start,
10482 char* name_start = RExC_parse;
10484 SV *sv_dat = reg_scan_name(pRExC_state, SIZE_ONLY
10485 ? REG_RSN_RETURN_NULL
10486 : REG_RSN_RETURN_DATA);
10487 GET_RE_DEBUG_FLAGS_DECL;
10489 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10491 if (RExC_parse == name_start || *RExC_parse != ch) {
10492 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10493 vFAIL2("Sequence %.3s... not terminated",parse_start);
10497 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10498 RExC_rxi->data->data[num]=(void*)sv_dat;
10499 SvREFCNT_inc_simple_void(sv_dat);
10502 ret = reganode(pRExC_state,
10505 : (ASCII_FOLD_RESTRICTED)
10507 : (AT_LEAST_UNI_SEMANTICS)
10513 *flagp |= HASWIDTH;
10515 Set_Node_Offset(ret, parse_start+1);
10516 Set_Node_Cur_Length(ret, parse_start);
10518 nextchar(pRExC_state);
10522 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
10523 flags. Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan
10524 needs to be restarted, or'd with NEED_UTF8 if the pattern needs to be
10525 upgraded to UTF-8. Otherwise would only return NULL if regbranch() returns
10526 NULL, which cannot happen. */
10528 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
10529 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10530 * 2 is like 1, but indicates that nextchar() has been called to advance
10531 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10532 * this flag alerts us to the need to check for that */
10534 regnode *ret; /* Will be the head of the group. */
10537 regnode *ender = NULL;
10540 U32 oregflags = RExC_flags;
10541 bool have_branch = 0;
10543 I32 freeze_paren = 0;
10544 I32 after_freeze = 0;
10545 I32 num; /* numeric backreferences */
10547 char * parse_start = RExC_parse; /* MJD */
10548 char * const oregcomp_parse = RExC_parse;
10550 GET_RE_DEBUG_FLAGS_DECL;
10552 PERL_ARGS_ASSERT_REG;
10553 DEBUG_PARSE("reg ");
10555 *flagp = 0; /* Tentatively. */
10557 /* Having this true makes it feasible to have a lot fewer tests for the
10558 * parse pointer being in scope. For example, we can write
10559 * while(isFOO(*RExC_parse)) RExC_parse++;
10561 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
10563 assert(*RExC_end == '\0');
10565 /* Make an OPEN node, if parenthesized. */
10568 /* Under /x, space and comments can be gobbled up between the '(' and
10569 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
10570 * intervening space, as the sequence is a token, and a token should be
10572 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
10574 if (RExC_parse >= RExC_end) {
10575 vFAIL("Unmatched (");
10578 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
10579 char *start_verb = RExC_parse + 1;
10581 char *start_arg = NULL;
10582 unsigned char op = 0;
10583 int arg_required = 0;
10584 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
10586 if (has_intervening_patws) {
10587 RExC_parse++; /* past the '*' */
10588 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
10590 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
10591 if ( *RExC_parse == ':' ) {
10592 start_arg = RExC_parse + 1;
10595 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10597 verb_len = RExC_parse - start_verb;
10599 if (RExC_parse >= RExC_end) {
10600 goto unterminated_verb_pattern;
10602 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10603 while ( RExC_parse < RExC_end && *RExC_parse != ')' )
10604 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10605 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10606 unterminated_verb_pattern:
10607 vFAIL("Unterminated verb pattern argument");
10608 if ( RExC_parse == start_arg )
10611 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10612 vFAIL("Unterminated verb pattern");
10615 /* Here, we know that RExC_parse < RExC_end */
10617 switch ( *start_verb ) {
10618 case 'A': /* (*ACCEPT) */
10619 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
10621 internal_argval = RExC_nestroot;
10624 case 'C': /* (*COMMIT) */
10625 if ( memEQs(start_verb,verb_len,"COMMIT") )
10628 case 'F': /* (*FAIL) */
10629 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
10633 case ':': /* (*:NAME) */
10634 case 'M': /* (*MARK:NAME) */
10635 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
10640 case 'P': /* (*PRUNE) */
10641 if ( memEQs(start_verb,verb_len,"PRUNE") )
10644 case 'S': /* (*SKIP) */
10645 if ( memEQs(start_verb,verb_len,"SKIP") )
10648 case 'T': /* (*THEN) */
10649 /* [19:06] <TimToady> :: is then */
10650 if ( memEQs(start_verb,verb_len,"THEN") ) {
10652 RExC_seen |= REG_CUTGROUP_SEEN;
10657 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10659 "Unknown verb pattern '%" UTF8f "'",
10660 UTF8fARG(UTF, verb_len, start_verb));
10662 if ( arg_required && !start_arg ) {
10663 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
10664 verb_len, start_verb);
10666 if (internal_argval == -1) {
10667 ret = reganode(pRExC_state, op, 0);
10669 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
10671 RExC_seen |= REG_VERBARG_SEEN;
10672 if ( ! SIZE_ONLY ) {
10674 SV *sv = newSVpvn( start_arg,
10675 RExC_parse - start_arg);
10676 ARG(ret) = add_data( pRExC_state,
10677 STR_WITH_LEN("S"));
10678 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
10683 if ( internal_argval != -1 )
10684 ARG2L_SET(ret, internal_argval);
10686 nextchar(pRExC_state);
10689 else if (*RExC_parse == '?') { /* (?...) */
10690 bool is_logical = 0;
10691 const char * const seqstart = RExC_parse;
10692 const char * endptr;
10693 if (has_intervening_patws) {
10695 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
10698 RExC_parse++; /* past the '?' */
10699 paren = *RExC_parse; /* might be a trailing NUL, if not
10701 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10702 if (RExC_parse > RExC_end) {
10705 ret = NULL; /* For look-ahead/behind. */
10708 case 'P': /* (?P...) variants for those used to PCRE/Python */
10709 paren = *RExC_parse;
10710 if ( paren == '<') { /* (?P<...>) named capture */
10712 if (RExC_parse >= RExC_end) {
10713 vFAIL("Sequence (?P<... not terminated");
10715 goto named_capture;
10717 else if (paren == '>') { /* (?P>name) named recursion */
10719 if (RExC_parse >= RExC_end) {
10720 vFAIL("Sequence (?P>... not terminated");
10722 goto named_recursion;
10724 else if (paren == '=') { /* (?P=...) named backref */
10726 return handle_named_backref(pRExC_state, flagp,
10729 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10730 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10731 vFAIL3("Sequence (%.*s...) not recognized",
10732 RExC_parse-seqstart, seqstart);
10733 NOT_REACHED; /*NOTREACHED*/
10734 case '<': /* (?<...) */
10735 if (*RExC_parse == '!')
10737 else if (*RExC_parse != '=')
10744 case '\'': /* (?'...') */
10745 name_start = RExC_parse;
10746 svname = reg_scan_name(pRExC_state,
10747 SIZE_ONLY /* reverse test from the others */
10748 ? REG_RSN_RETURN_NAME
10749 : REG_RSN_RETURN_NULL);
10750 if ( RExC_parse == name_start
10751 || RExC_parse >= RExC_end
10752 || *RExC_parse != paren)
10754 vFAIL2("Sequence (?%c... not terminated",
10755 paren=='>' ? '<' : paren);
10760 if (!svname) /* shouldn't happen */
10762 "panic: reg_scan_name returned NULL");
10763 if (!RExC_paren_names) {
10764 RExC_paren_names= newHV();
10765 sv_2mortal(MUTABLE_SV(RExC_paren_names));
10767 RExC_paren_name_list= newAV();
10768 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10771 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10773 sv_dat = HeVAL(he_str);
10775 /* croak baby croak */
10777 "panic: paren_name hash element allocation failed");
10778 } else if ( SvPOK(sv_dat) ) {
10779 /* (?|...) can mean we have dupes so scan to check
10780 its already been stored. Maybe a flag indicating
10781 we are inside such a construct would be useful,
10782 but the arrays are likely to be quite small, so
10783 for now we punt -- dmq */
10784 IV count = SvIV(sv_dat);
10785 I32 *pv = (I32*)SvPVX(sv_dat);
10787 for ( i = 0 ; i < count ; i++ ) {
10788 if ( pv[i] == RExC_npar ) {
10794 pv = (I32*)SvGROW(sv_dat,
10795 SvCUR(sv_dat) + sizeof(I32)+1);
10796 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10797 pv[count] = RExC_npar;
10798 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10801 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10802 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10805 SvIV_set(sv_dat, 1);
10808 /* Yes this does cause a memory leak in debugging Perls
10810 if (!av_store(RExC_paren_name_list,
10811 RExC_npar, SvREFCNT_inc(svname)))
10812 SvREFCNT_dec_NN(svname);
10815 /*sv_dump(sv_dat);*/
10817 nextchar(pRExC_state);
10819 goto capturing_parens;
10821 RExC_seen |= REG_LOOKBEHIND_SEEN;
10822 RExC_in_lookbehind++;
10824 if (RExC_parse >= RExC_end) {
10825 vFAIL("Sequence (?... not terminated");
10829 case '=': /* (?=...) */
10830 RExC_seen_zerolen++;
10832 case '!': /* (?!...) */
10833 RExC_seen_zerolen++;
10834 /* check if we're really just a "FAIL" assertion */
10835 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
10836 FALSE /* Don't force to /x */ );
10837 if (*RExC_parse == ')') {
10838 ret=reganode(pRExC_state, OPFAIL, 0);
10839 nextchar(pRExC_state);
10843 case '|': /* (?|...) */
10844 /* branch reset, behave like a (?:...) except that
10845 buffers in alternations share the same numbers */
10847 after_freeze = freeze_paren = RExC_npar;
10849 case ':': /* (?:...) */
10850 case '>': /* (?>...) */
10852 case '$': /* (?$...) */
10853 case '@': /* (?@...) */
10854 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10856 case '0' : /* (?0) */
10857 case 'R' : /* (?R) */
10858 if (RExC_parse == RExC_end || *RExC_parse != ')')
10859 FAIL("Sequence (?R) not terminated");
10861 RExC_seen |= REG_RECURSE_SEEN;
10862 *flagp |= POSTPONED;
10863 goto gen_recurse_regop;
10865 /* named and numeric backreferences */
10866 case '&': /* (?&NAME) */
10867 parse_start = RExC_parse - 1;
10870 SV *sv_dat = reg_scan_name(pRExC_state,
10871 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10872 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10874 if (RExC_parse >= RExC_end || *RExC_parse != ')')
10875 vFAIL("Sequence (?&... not terminated");
10876 goto gen_recurse_regop;
10879 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10881 vFAIL("Illegal pattern");
10883 goto parse_recursion;
10885 case '-': /* (?-1) */
10886 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10887 RExC_parse--; /* rewind to let it be handled later */
10891 case '1': case '2': case '3': case '4': /* (?1) */
10892 case '5': case '6': case '7': case '8': case '9':
10893 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
10896 bool is_neg = FALSE;
10898 parse_start = RExC_parse - 1; /* MJD */
10899 if (*RExC_parse == '-') {
10903 if (grok_atoUV(RExC_parse, &unum, &endptr)
10907 RExC_parse = (char*)endptr;
10911 /* Some limit for num? */
10915 if (*RExC_parse!=')')
10916 vFAIL("Expecting close bracket");
10919 if ( paren == '-' ) {
10921 Diagram of capture buffer numbering.
10922 Top line is the normal capture buffer numbers
10923 Bottom line is the negative indexing as from
10927 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
10931 num = RExC_npar + num;
10934 vFAIL("Reference to nonexistent group");
10936 } else if ( paren == '+' ) {
10937 num = RExC_npar + num - 1;
10939 /* We keep track how many GOSUB items we have produced.
10940 To start off the ARG2L() of the GOSUB holds its "id",
10941 which is used later in conjunction with RExC_recurse
10942 to calculate the offset we need to jump for the GOSUB,
10943 which it will store in the final representation.
10944 We have to defer the actual calculation until much later
10945 as the regop may move.
10948 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
10950 if (num > (I32)RExC_rx->nparens) {
10952 vFAIL("Reference to nonexistent group");
10954 RExC_recurse_count++;
10955 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
10956 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
10957 22, "| |", (int)(depth * 2 + 1), "",
10958 (UV)ARG(ret), (IV)ARG2L(ret)));
10960 RExC_seen |= REG_RECURSE_SEEN;
10962 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
10963 Set_Node_Offset(ret, parse_start); /* MJD */
10965 *flagp |= POSTPONED;
10966 assert(*RExC_parse == ')');
10967 nextchar(pRExC_state);
10972 case '?': /* (??...) */
10974 if (*RExC_parse != '{') {
10975 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10976 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10978 "Sequence (%" UTF8f "...) not recognized",
10979 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10980 NOT_REACHED; /*NOTREACHED*/
10982 *flagp |= POSTPONED;
10986 case '{': /* (?{...}) */
10989 struct reg_code_block *cb;
10991 RExC_seen_zerolen++;
10993 if ( !pRExC_state->num_code_blocks
10994 || pRExC_state->code_index >= pRExC_state->num_code_blocks
10995 || pRExC_state->code_blocks[pRExC_state->code_index].start
10996 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
10999 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11000 FAIL("panic: Sequence (?{...}): no code block found\n");
11001 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11003 /* this is a pre-compiled code block (?{...}) */
11004 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
11005 RExC_parse = RExC_start + cb->end;
11008 if (cb->src_regex) {
11009 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11010 RExC_rxi->data->data[n] =
11011 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11012 RExC_rxi->data->data[n+1] = (void*)o;
11015 n = add_data(pRExC_state,
11016 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11017 RExC_rxi->data->data[n] = (void*)o;
11020 pRExC_state->code_index++;
11021 nextchar(pRExC_state);
11025 ret = reg_node(pRExC_state, LOGICAL);
11027 eval = reg2Lanode(pRExC_state, EVAL,
11030 /* for later propagation into (??{})
11032 RExC_flags & RXf_PMf_COMPILETIME
11037 REGTAIL(pRExC_state, ret, eval);
11038 /* deal with the length of this later - MJD */
11041 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11042 Set_Node_Length(ret, RExC_parse - parse_start + 1);
11043 Set_Node_Offset(ret, parse_start);
11046 case '(': /* (?(?{...})...) and (?(?=...)...) */
11049 const int DEFINE_len = sizeof("DEFINE") - 1;
11050 if (RExC_parse[0] == '?') { /* (?(?...)) */
11051 if ( RExC_parse < RExC_end - 1
11052 && ( RExC_parse[1] == '='
11053 || RExC_parse[1] == '!'
11054 || RExC_parse[1] == '<'
11055 || RExC_parse[1] == '{')
11056 ) { /* Lookahead or eval. */
11060 ret = reg_node(pRExC_state, LOGICAL);
11064 tail = reg(pRExC_state, 1, &flag, depth+1);
11065 if (flag & (RESTART_PASS1|NEED_UTF8)) {
11066 *flagp = flag & (RESTART_PASS1|NEED_UTF8);
11069 REGTAIL(pRExC_state, ret, tail);
11072 /* Fall through to ‘Unknown switch condition’ at the
11073 end of the if/else chain. */
11075 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11076 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11078 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11079 char *name_start= RExC_parse++;
11081 SV *sv_dat=reg_scan_name(pRExC_state,
11082 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11083 if ( RExC_parse == name_start
11084 || RExC_parse >= RExC_end
11085 || *RExC_parse != ch)
11087 vFAIL2("Sequence (?(%c... not terminated",
11088 (ch == '>' ? '<' : ch));
11092 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11093 RExC_rxi->data->data[num]=(void*)sv_dat;
11094 SvREFCNT_inc_simple_void(sv_dat);
11096 ret = reganode(pRExC_state,NGROUPP,num);
11097 goto insert_if_check_paren;
11099 else if (RExC_end - RExC_parse >= DEFINE_len
11100 && strnEQ(RExC_parse, "DEFINE", DEFINE_len))
11102 ret = reganode(pRExC_state,DEFINEP,0);
11103 RExC_parse += DEFINE_len;
11105 goto insert_if_check_paren;
11107 else if (RExC_parse[0] == 'R') {
11109 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11110 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11111 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11114 if (RExC_parse[0] == '0') {
11118 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11120 if (grok_atoUV(RExC_parse, &uv, &endptr)
11123 parno = (I32)uv + 1;
11124 RExC_parse = (char*)endptr;
11126 /* else "Switch condition not recognized" below */
11127 } else if (RExC_parse[0] == '&') {
11130 sv_dat = reg_scan_name(pRExC_state,
11132 ? REG_RSN_RETURN_NULL
11133 : REG_RSN_RETURN_DATA);
11135 /* we should only have a false sv_dat when
11136 * SIZE_ONLY is true, and we always have false
11137 * sv_dat when SIZE_ONLY is true.
11138 * reg_scan_name() will VFAIL() if the name is
11139 * unknown when SIZE_ONLY is false, and otherwise
11140 * will return something, and when SIZE_ONLY is
11141 * true, reg_scan_name() just parses the string,
11142 * and doesnt return anything. (in theory) */
11143 assert(SIZE_ONLY ? !sv_dat : !!sv_dat);
11146 parno = 1 + *((I32 *)SvPVX(sv_dat));
11148 ret = reganode(pRExC_state,INSUBP,parno);
11149 goto insert_if_check_paren;
11151 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11155 if (grok_atoUV(RExC_parse, &uv, &endptr)
11159 RExC_parse = (char*)endptr;
11162 vFAIL("panic: grok_atoUV returned FALSE");
11164 ret = reganode(pRExC_state, GROUPP, parno);
11166 insert_if_check_paren:
11167 if (UCHARAT(RExC_parse) != ')') {
11168 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11169 vFAIL("Switch condition not recognized");
11171 nextchar(pRExC_state);
11173 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
11174 br = regbranch(pRExC_state, &flags, 1,depth+1);
11176 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11177 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11180 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11183 REGTAIL(pRExC_state, br, reganode(pRExC_state,
11185 c = UCHARAT(RExC_parse);
11186 nextchar(pRExC_state);
11187 if (flags&HASWIDTH)
11188 *flagp |= HASWIDTH;
11191 vFAIL("(?(DEFINE)....) does not allow branches");
11193 /* Fake one for optimizer. */
11194 lastbr = reganode(pRExC_state, IFTHEN, 0);
11196 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
11197 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11198 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11201 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11204 REGTAIL(pRExC_state, ret, lastbr);
11205 if (flags&HASWIDTH)
11206 *flagp |= HASWIDTH;
11207 c = UCHARAT(RExC_parse);
11208 nextchar(pRExC_state);
11213 if (RExC_parse >= RExC_end)
11214 vFAIL("Switch (?(condition)... not terminated");
11216 vFAIL("Switch (?(condition)... contains too many branches");
11218 ender = reg_node(pRExC_state, TAIL);
11219 REGTAIL(pRExC_state, br, ender);
11221 REGTAIL(pRExC_state, lastbr, ender);
11222 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11225 REGTAIL(pRExC_state, ret, ender);
11226 RExC_size++; /* XXX WHY do we need this?!!
11227 For large programs it seems to be required
11228 but I can't figure out why. -- dmq*/
11231 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11232 vFAIL("Unknown switch condition (?(...))");
11234 case '[': /* (?[ ... ]) */
11235 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
11237 case 0: /* A NUL */
11238 RExC_parse--; /* for vFAIL to print correctly */
11239 vFAIL("Sequence (? incomplete");
11241 default: /* e.g., (?i) */
11242 RExC_parse = (char *) seqstart + 1;
11244 parse_lparen_question_flags(pRExC_state);
11245 if (UCHARAT(RExC_parse) != ':') {
11246 if (RExC_parse < RExC_end)
11247 nextchar(pRExC_state);
11252 nextchar(pRExC_state);
11257 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11262 ret = reganode(pRExC_state, OPEN, parno);
11264 if (!RExC_nestroot)
11265 RExC_nestroot = parno;
11266 if (RExC_open_parens && !RExC_open_parens[parno])
11268 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11269 "%*s%*s Setting open paren #%" IVdf " to %d\n",
11270 22, "| |", (int)(depth * 2 + 1), "",
11271 (IV)parno, REG_NODE_NUM(ret)));
11272 RExC_open_parens[parno]= ret;
11275 Set_Node_Length(ret, 1); /* MJD */
11276 Set_Node_Offset(ret, RExC_parse); /* MJD */
11279 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
11288 /* Pick up the branches, linking them together. */
11289 parse_start = RExC_parse; /* MJD */
11290 br = regbranch(pRExC_state, &flags, 1,depth+1);
11292 /* branch_len = (paren != 0); */
11295 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11296 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11299 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11301 if (*RExC_parse == '|') {
11302 if (!SIZE_ONLY && RExC_extralen) {
11303 reginsert(pRExC_state, BRANCHJ, br, depth+1);
11306 reginsert(pRExC_state, BRANCH, br, depth+1);
11307 Set_Node_Length(br, paren != 0);
11308 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
11312 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
11314 else if (paren == ':') {
11315 *flagp |= flags&SIMPLE;
11317 if (is_open) { /* Starts with OPEN. */
11318 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
11320 else if (paren != '?') /* Not Conditional */
11322 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11324 while (*RExC_parse == '|') {
11325 if (!SIZE_ONLY && RExC_extralen) {
11326 ender = reganode(pRExC_state, LONGJMP,0);
11328 /* Append to the previous. */
11329 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11332 RExC_extralen += 2; /* Account for LONGJMP. */
11333 nextchar(pRExC_state);
11334 if (freeze_paren) {
11335 if (RExC_npar > after_freeze)
11336 after_freeze = RExC_npar;
11337 RExC_npar = freeze_paren;
11339 br = regbranch(pRExC_state, &flags, 0, depth+1);
11342 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11343 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11346 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11348 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
11350 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11353 if (have_branch || paren != ':') {
11354 /* Make a closing node, and hook it on the end. */
11357 ender = reg_node(pRExC_state, TAIL);
11360 ender = reganode(pRExC_state, CLOSE, parno);
11361 if ( RExC_close_parens ) {
11362 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11363 "%*s%*s Setting close paren #%" IVdf " to %d\n",
11364 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
11365 RExC_close_parens[parno]= ender;
11366 if (RExC_nestroot == parno)
11369 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
11370 Set_Node_Length(ender,1); /* MJD */
11376 *flagp &= ~HASWIDTH;
11379 ender = reg_node(pRExC_state, SUCCEED);
11382 ender = reg_node(pRExC_state, END);
11384 assert(!RExC_end_op); /* there can only be one! */
11385 RExC_end_op = ender;
11386 if (RExC_close_parens) {
11387 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11388 "%*s%*s Setting close paren #0 (END) to %d\n",
11389 22, "| |", (int)(depth * 2 + 1), "", REG_NODE_NUM(ender)));
11391 RExC_close_parens[0]= ender;
11396 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11397 DEBUG_PARSE_MSG("lsbr");
11398 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
11399 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11400 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11401 SvPV_nolen_const(RExC_mysv1),
11402 (IV)REG_NODE_NUM(lastbr),
11403 SvPV_nolen_const(RExC_mysv2),
11404 (IV)REG_NODE_NUM(ender),
11405 (IV)(ender - lastbr)
11408 REGTAIL(pRExC_state, lastbr, ender);
11410 if (have_branch && !SIZE_ONLY) {
11411 char is_nothing= 1;
11413 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
11415 /* Hook the tails of the branches to the closing node. */
11416 for (br = ret; br; br = regnext(br)) {
11417 const U8 op = PL_regkind[OP(br)];
11418 if (op == BRANCH) {
11419 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
11420 if ( OP(NEXTOPER(br)) != NOTHING
11421 || regnext(NEXTOPER(br)) != ender)
11424 else if (op == BRANCHJ) {
11425 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
11426 /* for now we always disable this optimisation * /
11427 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
11428 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
11434 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
11435 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11436 DEBUG_PARSE_MSG("NADA");
11437 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
11438 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11439 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11440 SvPV_nolen_const(RExC_mysv1),
11441 (IV)REG_NODE_NUM(ret),
11442 SvPV_nolen_const(RExC_mysv2),
11443 (IV)REG_NODE_NUM(ender),
11448 if (OP(ender) == TAIL) {
11453 for ( opt= br + 1; opt < ender ; opt++ )
11454 OP(opt)= OPTIMIZED;
11455 NEXT_OFF(br)= ender - br;
11463 static const char parens[] = "=!<,>";
11465 if (paren && (p = strchr(parens, paren))) {
11466 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
11467 int flag = (p - parens) > 1;
11470 node = SUSPEND, flag = 0;
11471 reginsert(pRExC_state, node,ret, depth+1);
11472 Set_Node_Cur_Length(ret, parse_start);
11473 Set_Node_Offset(ret, parse_start + 1);
11475 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
11479 /* Check for proper termination. */
11481 /* restore original flags, but keep (?p) and, if we've changed from /d
11482 * rules to /u, keep the /u */
11483 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
11484 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
11485 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
11487 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
11488 RExC_parse = oregcomp_parse;
11489 vFAIL("Unmatched (");
11491 nextchar(pRExC_state);
11493 else if (!paren && RExC_parse < RExC_end) {
11494 if (*RExC_parse == ')') {
11496 vFAIL("Unmatched )");
11499 FAIL("Junk on end of regexp"); /* "Can't happen". */
11500 NOT_REACHED; /* NOTREACHED */
11503 if (RExC_in_lookbehind) {
11504 RExC_in_lookbehind--;
11506 if (after_freeze > RExC_npar)
11507 RExC_npar = after_freeze;
11512 - regbranch - one alternative of an | operator
11514 * Implements the concatenation operator.
11516 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11517 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11520 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
11523 regnode *chain = NULL;
11525 I32 flags = 0, c = 0;
11526 GET_RE_DEBUG_FLAGS_DECL;
11528 PERL_ARGS_ASSERT_REGBRANCH;
11530 DEBUG_PARSE("brnc");
11535 if (!SIZE_ONLY && RExC_extralen)
11536 ret = reganode(pRExC_state, BRANCHJ,0);
11538 ret = reg_node(pRExC_state, BRANCH);
11539 Set_Node_Length(ret, 1);
11543 if (!first && SIZE_ONLY)
11544 RExC_extralen += 1; /* BRANCHJ */
11546 *flagp = WORST; /* Tentatively. */
11548 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11549 FALSE /* Don't force to /x */ );
11550 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
11551 flags &= ~TRYAGAIN;
11552 latest = regpiece(pRExC_state, &flags,depth+1);
11553 if (latest == NULL) {
11554 if (flags & TRYAGAIN)
11556 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11557 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11560 FAIL2("panic: regpiece returned NULL, flags=%#" UVxf, (UV) flags);
11562 else if (ret == NULL)
11564 *flagp |= flags&(HASWIDTH|POSTPONED);
11565 if (chain == NULL) /* First piece. */
11566 *flagp |= flags&SPSTART;
11568 /* FIXME adding one for every branch after the first is probably
11569 * excessive now we have TRIE support. (hv) */
11571 REGTAIL(pRExC_state, chain, latest);
11576 if (chain == NULL) { /* Loop ran zero times. */
11577 chain = reg_node(pRExC_state, NOTHING);
11582 *flagp |= flags&SIMPLE;
11589 - regpiece - something followed by possible [*+?]
11591 * Note that the branching code sequences used for ? and the general cases
11592 * of * and + are somewhat optimized: they use the same NOTHING node as
11593 * both the endmarker for their branch list and the body of the last branch.
11594 * It might seem that this node could be dispensed with entirely, but the
11595 * endmarker role is not redundant.
11597 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
11599 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11600 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11603 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11609 const char * const origparse = RExC_parse;
11611 I32 max = REG_INFTY;
11612 #ifdef RE_TRACK_PATTERN_OFFSETS
11615 const char *maxpos = NULL;
11618 /* Save the original in case we change the emitted regop to a FAIL. */
11619 regnode * const orig_emit = RExC_emit;
11621 GET_RE_DEBUG_FLAGS_DECL;
11623 PERL_ARGS_ASSERT_REGPIECE;
11625 DEBUG_PARSE("piec");
11627 ret = regatom(pRExC_state, &flags,depth+1);
11629 if (flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8))
11630 *flagp |= flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8);
11632 FAIL2("panic: regatom returned NULL, flags=%#" UVxf, (UV) flags);
11638 if (op == '{' && regcurly(RExC_parse)) {
11640 #ifdef RE_TRACK_PATTERN_OFFSETS
11641 parse_start = RExC_parse; /* MJD */
11643 next = RExC_parse + 1;
11644 while (isDIGIT(*next) || *next == ',') {
11645 if (*next == ',') {
11653 if (*next == '}') { /* got one */
11654 const char* endptr;
11658 if (isDIGIT(*RExC_parse)) {
11659 if (!grok_atoUV(RExC_parse, &uv, &endptr))
11660 vFAIL("Invalid quantifier in {,}");
11661 if (uv >= REG_INFTY)
11662 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11667 if (*maxpos == ',')
11670 maxpos = RExC_parse;
11671 if (isDIGIT(*maxpos)) {
11672 if (!grok_atoUV(maxpos, &uv, &endptr))
11673 vFAIL("Invalid quantifier in {,}");
11674 if (uv >= REG_INFTY)
11675 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11678 max = REG_INFTY; /* meaning "infinity" */
11681 nextchar(pRExC_state);
11682 if (max < min) { /* If can't match, warn and optimize to fail
11686 /* We can't back off the size because we have to reserve
11687 * enough space for all the things we are about to throw
11688 * away, but we can shrink it by the amount we are about
11689 * to re-use here */
11690 RExC_size += PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
11693 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
11694 RExC_emit = orig_emit;
11696 ret = reganode(pRExC_state, OPFAIL, 0);
11699 else if (min == max && *RExC_parse == '?')
11702 ckWARN2reg(RExC_parse + 1,
11703 "Useless use of greediness modifier '%c'",
11709 if ((flags&SIMPLE)) {
11710 if (min == 0 && max == REG_INFTY) {
11711 reginsert(pRExC_state, STAR, ret, depth+1);
11714 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11717 if (min == 1 && max == REG_INFTY) {
11718 reginsert(pRExC_state, PLUS, ret, depth+1);
11721 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11724 MARK_NAUGHTY_EXP(2, 2);
11725 reginsert(pRExC_state, CURLY, ret, depth+1);
11726 Set_Node_Offset(ret, parse_start+1); /* MJD */
11727 Set_Node_Cur_Length(ret, parse_start);
11730 regnode * const w = reg_node(pRExC_state, WHILEM);
11733 REGTAIL(pRExC_state, ret, w);
11734 if (!SIZE_ONLY && RExC_extralen) {
11735 reginsert(pRExC_state, LONGJMP,ret, depth+1);
11736 reginsert(pRExC_state, NOTHING,ret, depth+1);
11737 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
11739 reginsert(pRExC_state, CURLYX,ret, depth+1);
11741 Set_Node_Offset(ret, parse_start+1);
11742 Set_Node_Length(ret,
11743 op == '{' ? (RExC_parse - parse_start) : 1);
11745 if (!SIZE_ONLY && RExC_extralen)
11746 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
11747 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
11749 RExC_whilem_seen++, RExC_extralen += 3;
11750 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
11757 *flagp |= HASWIDTH;
11759 ARG1_SET(ret, (U16)min);
11760 ARG2_SET(ret, (U16)max);
11762 if (max == REG_INFTY)
11763 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11769 if (!ISMULT1(op)) {
11774 #if 0 /* Now runtime fix should be reliable. */
11776 /* if this is reinstated, don't forget to put this back into perldiag:
11778 =item Regexp *+ operand could be empty at {#} in regex m/%s/
11780 (F) The part of the regexp subject to either the * or + quantifier
11781 could match an empty string. The {#} shows in the regular
11782 expression about where the problem was discovered.
11786 if (!(flags&HASWIDTH) && op != '?')
11787 vFAIL("Regexp *+ operand could be empty");
11790 #ifdef RE_TRACK_PATTERN_OFFSETS
11791 parse_start = RExC_parse;
11793 nextchar(pRExC_state);
11795 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
11801 else if (op == '+') {
11805 else if (op == '?') {
11810 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
11811 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
11812 ckWARN2reg(RExC_parse,
11813 "%" UTF8f " matches null string many times",
11814 UTF8fARG(UTF, (RExC_parse >= origparse
11815 ? RExC_parse - origparse
11818 (void)ReREFCNT_inc(RExC_rx_sv);
11821 if (*RExC_parse == '?') {
11822 nextchar(pRExC_state);
11823 reginsert(pRExC_state, MINMOD, ret, depth+1);
11824 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
11826 else if (*RExC_parse == '+') {
11828 nextchar(pRExC_state);
11829 ender = reg_node(pRExC_state, SUCCEED);
11830 REGTAIL(pRExC_state, ret, ender);
11831 reginsert(pRExC_state, SUSPEND, ret, depth+1);
11833 ender = reg_node(pRExC_state, TAIL);
11834 REGTAIL(pRExC_state, ret, ender);
11837 if (ISMULT2(RExC_parse)) {
11839 vFAIL("Nested quantifiers");
11846 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
11855 /* This routine teases apart the various meanings of \N and returns
11856 * accordingly. The input parameters constrain which meaning(s) is/are valid
11857 * in the current context.
11859 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
11861 * If <code_point_p> is not NULL, the context is expecting the result to be a
11862 * single code point. If this \N instance turns out to a single code point,
11863 * the function returns TRUE and sets *code_point_p to that code point.
11865 * If <node_p> is not NULL, the context is expecting the result to be one of
11866 * the things representable by a regnode. If this \N instance turns out to be
11867 * one such, the function generates the regnode, returns TRUE and sets *node_p
11868 * to point to that regnode.
11870 * If this instance of \N isn't legal in any context, this function will
11871 * generate a fatal error and not return.
11873 * On input, RExC_parse should point to the first char following the \N at the
11874 * time of the call. On successful return, RExC_parse will have been updated
11875 * to point to just after the sequence identified by this routine. Also
11876 * *flagp has been updated as needed.
11878 * When there is some problem with the current context and this \N instance,
11879 * the function returns FALSE, without advancing RExC_parse, nor setting
11880 * *node_p, nor *code_point_p, nor *flagp.
11882 * If <cp_count> is not NULL, the caller wants to know the length (in code
11883 * points) that this \N sequence matches. This is set even if the function
11884 * returns FALSE, as detailed below.
11886 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
11888 * Probably the most common case is for the \N to specify a single code point.
11889 * *cp_count will be set to 1, and *code_point_p will be set to that code
11892 * Another possibility is for the input to be an empty \N{}, which for
11893 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
11894 * will be set to a generated NOTHING node.
11896 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
11897 * set to 0. *node_p will be set to a generated REG_ANY node.
11899 * The fourth possibility is that \N resolves to a sequence of more than one
11900 * code points. *cp_count will be set to the number of code points in the
11901 * sequence. *node_p * will be set to a generated node returned by this
11902 * function calling S_reg().
11904 * The final possibility is that it is premature to be calling this function;
11905 * that pass1 needs to be restarted. This can happen when this changes from
11906 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
11907 * latter occurs only when the fourth possibility would otherwise be in
11908 * effect, and is because one of those code points requires the pattern to be
11909 * recompiled as UTF-8. The function returns FALSE, and sets the
11910 * RESTART_PASS1 and NEED_UTF8 flags in *flagp, as appropriate. When this
11911 * happens, the caller needs to desist from continuing parsing, and return
11912 * this information to its caller. This is not set for when there is only one
11913 * code point, as this can be called as part of an ANYOF node, and they can
11914 * store above-Latin1 code points without the pattern having to be in UTF-8.
11916 * For non-single-quoted regexes, the tokenizer has resolved character and
11917 * sequence names inside \N{...} into their Unicode values, normalizing the
11918 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
11919 * hex-represented code points in the sequence. This is done there because
11920 * the names can vary based on what charnames pragma is in scope at the time,
11921 * so we need a way to take a snapshot of what they resolve to at the time of
11922 * the original parse. [perl #56444].
11924 * That parsing is skipped for single-quoted regexes, so we may here get
11925 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
11926 * parser. But if the single-quoted regex is something like '\N{U+41}', that
11927 * is legal and handled here. The code point is Unicode, and has to be
11928 * translated into the native character set for non-ASCII platforms.
11931 char * endbrace; /* points to '}' following the name */
11932 char *endchar; /* Points to '.' or '}' ending cur char in the input
11934 char* p = RExC_parse; /* Temporary */
11936 GET_RE_DEBUG_FLAGS_DECL;
11938 PERL_ARGS_ASSERT_GROK_BSLASH_N;
11940 GET_RE_DEBUG_FLAGS;
11942 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
11943 assert(! (node_p && cp_count)); /* At most 1 should be set */
11945 if (cp_count) { /* Initialize return for the most common case */
11949 /* The [^\n] meaning of \N ignores spaces and comments under the /x
11950 * modifier. The other meanings do not, so use a temporary until we find
11951 * out which we are being called with */
11952 skip_to_be_ignored_text(pRExC_state, &p,
11953 FALSE /* Don't force to /x */ );
11955 /* Disambiguate between \N meaning a named character versus \N meaning
11956 * [^\n]. The latter is assumed when the {...} following the \N is a legal
11957 * quantifier, or there is no '{' at all */
11958 if (*p != '{' || regcurly(p)) {
11968 *node_p = reg_node(pRExC_state, REG_ANY);
11969 *flagp |= HASWIDTH|SIMPLE;
11971 Set_Node_Length(*node_p, 1); /* MJD */
11975 /* Here, we have decided it should be a named character or sequence */
11977 /* The test above made sure that the next real character is a '{', but
11978 * under the /x modifier, it could be separated by space (or a comment and
11979 * \n) and this is not allowed (for consistency with \x{...} and the
11980 * tokenizer handling of \N{NAME}). */
11981 if (*RExC_parse != '{') {
11982 vFAIL("Missing braces on \\N{}");
11985 RExC_parse++; /* Skip past the '{' */
11987 if (! (endbrace = strchr(RExC_parse, '}'))) { /* no trailing brace */
11988 vFAIL2("Missing right brace on \\%c{}", 'N');
11990 else if(!(endbrace == RExC_parse /* nothing between the {} */
11991 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked... */
11992 && strnEQ(RExC_parse, "U+", 2)))) /* ... below for a better
11995 RExC_parse = endbrace; /* position msg's '<--HERE' */
11996 vFAIL("\\N{NAME} must be resolved by the lexer");
11999 REQUIRE_UNI_RULES(flagp, FALSE); /* Unicode named chars imply Unicode
12002 if (endbrace == RExC_parse) { /* empty: \N{} */
12004 RExC_parse++; /* Position after the "}" */
12005 vFAIL("Zero length \\N{}");
12010 nextchar(pRExC_state);
12015 *node_p = reg_node(pRExC_state,NOTHING);
12019 RExC_parse += 2; /* Skip past the 'U+' */
12021 /* Because toke.c has generated a special construct for us guaranteed not
12022 * to have NULs, we can use a str function */
12023 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12025 /* Code points are separated by dots. If none, there is only one code
12026 * point, and is terminated by the brace */
12028 if (endchar >= endbrace) {
12029 STRLEN length_of_hex;
12030 I32 grok_hex_flags;
12032 /* Here, exactly one code point. If that isn't what is wanted, fail */
12033 if (! code_point_p) {
12038 /* Convert code point from hex */
12039 length_of_hex = (STRLEN)(endchar - RExC_parse);
12040 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
12041 | PERL_SCAN_DISALLOW_PREFIX
12043 /* No errors in the first pass (See [perl
12044 * #122671].) We let the code below find the
12045 * errors when there are multiple chars. */
12047 ? PERL_SCAN_SILENT_ILLDIGIT
12050 /* This routine is the one place where both single- and double-quotish
12051 * \N{U+xxxx} are evaluated. The value is a Unicode code point which
12052 * must be converted to native. */
12053 *code_point_p = UNI_TO_NATIVE(grok_hex(RExC_parse,
12058 /* The tokenizer should have guaranteed validity, but it's possible to
12059 * bypass it by using single quoting, so check. Don't do the check
12060 * here when there are multiple chars; we do it below anyway. */
12061 if (length_of_hex == 0
12062 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
12064 RExC_parse += length_of_hex; /* Includes all the valid */
12065 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12066 ? UTF8SKIP(RExC_parse)
12068 /* Guard against malformed utf8 */
12069 if (RExC_parse >= endchar) {
12070 RExC_parse = endchar;
12072 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12075 RExC_parse = endbrace + 1;
12078 else { /* Is a multiple character sequence */
12079 SV * substitute_parse;
12081 char *orig_end = RExC_end;
12082 char *save_start = RExC_start;
12085 /* Count the code points, if desired, in the sequence */
12088 while (RExC_parse < endbrace) {
12089 /* Point to the beginning of the next character in the sequence. */
12090 RExC_parse = endchar + 1;
12091 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12096 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12097 * But don't backup up the pointer if the caller want to know how many
12098 * code points there are (they can then handle things) */
12106 /* What is done here is to convert this to a sub-pattern of the form
12107 * \x{char1}\x{char2}... and then call reg recursively to parse it
12108 * (enclosing in "(?: ... )" ). That way, it retains its atomicness,
12109 * while not having to worry about special handling that some code
12110 * points may have. */
12112 substitute_parse = newSVpvs("?:");
12114 while (RExC_parse < endbrace) {
12116 /* Convert to notation the rest of the code understands */
12117 sv_catpv(substitute_parse, "\\x{");
12118 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
12119 sv_catpv(substitute_parse, "}");
12121 /* Point to the beginning of the next character in the sequence. */
12122 RExC_parse = endchar + 1;
12123 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12126 sv_catpv(substitute_parse, ")");
12128 RExC_parse = RExC_start = RExC_adjusted_start = SvPV(substitute_parse,
12131 /* Don't allow empty number */
12132 if (len < (STRLEN) 8) {
12133 RExC_parse = endbrace;
12134 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12136 RExC_end = RExC_parse + len;
12138 /* The values are Unicode, and therefore not subject to recoding, but
12139 * have to be converted to native on a non-Unicode (meaning non-ASCII)
12141 RExC_override_recoding = 1;
12143 RExC_recode_x_to_native = 1;
12147 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
12148 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12149 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12152 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#" UVxf,
12155 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12158 /* Restore the saved values */
12159 RExC_start = RExC_adjusted_start = save_start;
12160 RExC_parse = endbrace;
12161 RExC_end = orig_end;
12162 RExC_override_recoding = 0;
12164 RExC_recode_x_to_native = 0;
12167 SvREFCNT_dec_NN(substitute_parse);
12168 nextchar(pRExC_state);
12175 PERL_STATIC_INLINE U8
12176 S_compute_EXACTish(RExC_state_t *pRExC_state)
12180 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
12188 op = get_regex_charset(RExC_flags);
12189 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
12190 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
12191 been, so there is no hole */
12194 return op + EXACTF;
12197 PERL_STATIC_INLINE void
12198 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
12199 regnode *node, I32* flagp, STRLEN len, UV code_point,
12202 /* This knows the details about sizing an EXACTish node, setting flags for
12203 * it (by setting <*flagp>, and potentially populating it with a single
12206 * If <len> (the length in bytes) is non-zero, this function assumes that
12207 * the node has already been populated, and just does the sizing. In this
12208 * case <code_point> should be the final code point that has already been
12209 * placed into the node. This value will be ignored except that under some
12210 * circumstances <*flagp> is set based on it.
12212 * If <len> is zero, the function assumes that the node is to contain only
12213 * the single character given by <code_point> and calculates what <len>
12214 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
12215 * additionally will populate the node's STRING with <code_point> or its
12218 * In both cases <*flagp> is appropriately set
12220 * It knows that under FOLD, the Latin Sharp S and UTF characters above
12221 * 255, must be folded (the former only when the rules indicate it can
12224 * When it does the populating, it looks at the flag 'downgradable'. If
12225 * true with a node that folds, it checks if the single code point
12226 * participates in a fold, and if not downgrades the node to an EXACT.
12227 * This helps the optimizer */
12229 bool len_passed_in = cBOOL(len != 0);
12230 U8 character[UTF8_MAXBYTES_CASE+1];
12232 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
12234 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
12235 * sizing difference, and is extra work that is thrown away */
12236 if (downgradable && ! PASS2) {
12237 downgradable = FALSE;
12240 if (! len_passed_in) {
12242 if (UVCHR_IS_INVARIANT(code_point)) {
12243 if (LOC || ! FOLD) { /* /l defers folding until runtime */
12244 *character = (U8) code_point;
12246 else { /* Here is /i and not /l. (toFOLD() is defined on just
12247 ASCII, which isn't the same thing as INVARIANT on
12248 EBCDIC, but it works there, as the extra invariants
12249 fold to themselves) */
12250 *character = toFOLD((U8) code_point);
12252 /* We can downgrade to an EXACT node if this character
12253 * isn't a folding one. Note that this assumes that
12254 * nothing above Latin1 folds to some other invariant than
12255 * one of these alphabetics; otherwise we would also have
12257 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12258 * || ASCII_FOLD_RESTRICTED))
12260 if (downgradable && PL_fold[code_point] == code_point) {
12266 else if (FOLD && (! LOC
12267 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
12268 { /* Folding, and ok to do so now */
12269 UV folded = _to_uni_fold_flags(
12273 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12274 ? FOLD_FLAGS_NOMIX_ASCII
12277 && folded == code_point /* This quickly rules out many
12278 cases, avoiding the
12279 _invlist_contains_cp() overhead
12281 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
12288 else if (code_point <= MAX_UTF8_TWO_BYTE) {
12290 /* Not folding this cp, and can output it directly */
12291 *character = UTF8_TWO_BYTE_HI(code_point);
12292 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
12296 uvchr_to_utf8( character, code_point);
12297 len = UTF8SKIP(character);
12299 } /* Else pattern isn't UTF8. */
12301 *character = (U8) code_point;
12303 } /* Else is folded non-UTF8 */
12304 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12305 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12306 || UNICODE_DOT_DOT_VERSION > 0)
12307 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
12311 /* We don't fold any non-UTF8 except possibly the Sharp s (see
12312 * comments at join_exact()); */
12313 *character = (U8) code_point;
12316 /* Can turn into an EXACT node if we know the fold at compile time,
12317 * and it folds to itself and doesn't particpate in other folds */
12320 && PL_fold_latin1[code_point] == code_point
12321 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12322 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
12326 } /* else is Sharp s. May need to fold it */
12327 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
12329 *(character + 1) = 's';
12333 *character = LATIN_SMALL_LETTER_SHARP_S;
12339 RExC_size += STR_SZ(len);
12342 RExC_emit += STR_SZ(len);
12343 STR_LEN(node) = len;
12344 if (! len_passed_in) {
12345 Copy((char *) character, STRING(node), len, char);
12349 *flagp |= HASWIDTH;
12351 /* A single character node is SIMPLE, except for the special-cased SHARP S
12353 if ((len == 1 || (UTF && len == UVCHR_SKIP(code_point)))
12354 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12355 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12356 || UNICODE_DOT_DOT_VERSION > 0)
12357 && ( code_point != LATIN_SMALL_LETTER_SHARP_S
12358 || ! FOLD || ! DEPENDS_SEMANTICS)
12364 /* The OP may not be well defined in PASS1 */
12365 if (PASS2 && OP(node) == EXACTFL) {
12366 RExC_contains_locale = 1;
12371 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
12372 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
12375 S_backref_value(char *p)
12377 const char* endptr;
12379 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
12386 - regatom - the lowest level
12388 Try to identify anything special at the start of the current parse position.
12389 If there is, then handle it as required. This may involve generating a
12390 single regop, such as for an assertion; or it may involve recursing, such as
12391 to handle a () structure.
12393 If the string doesn't start with something special then we gobble up
12394 as much literal text as we can. If we encounter a quantifier, we have to
12395 back off the final literal character, as that quantifier applies to just it
12396 and not to the whole string of literals.
12398 Once we have been able to handle whatever type of thing started the
12399 sequence, we return.
12401 Note: we have to be careful with escapes, as they can be both literal
12402 and special, and in the case of \10 and friends, context determines which.
12404 A summary of the code structure is:
12406 switch (first_byte) {
12407 cases for each special:
12408 handle this special;
12411 switch (2nd byte) {
12412 cases for each unambiguous special:
12413 handle this special;
12415 cases for each ambigous special/literal:
12417 if (special) handle here
12419 default: // unambiguously literal:
12422 default: // is a literal char
12425 create EXACTish node for literal;
12426 while (more input and node isn't full) {
12427 switch (input_byte) {
12428 cases for each special;
12429 make sure parse pointer is set so that the next call to
12430 regatom will see this special first
12431 goto loopdone; // EXACTish node terminated by prev. char
12433 append char to EXACTISH node;
12435 get next input byte;
12439 return the generated node;
12441 Specifically there are two separate switches for handling
12442 escape sequences, with the one for handling literal escapes requiring
12443 a dummy entry for all of the special escapes that are actually handled
12446 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
12448 Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
12449 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
12450 Otherwise does not return NULL.
12454 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12456 regnode *ret = NULL;
12463 GET_RE_DEBUG_FLAGS_DECL;
12465 *flagp = WORST; /* Tentatively. */
12467 DEBUG_PARSE("atom");
12469 PERL_ARGS_ASSERT_REGATOM;
12472 parse_start = RExC_parse;
12473 assert(RExC_parse < RExC_end);
12474 switch ((U8)*RExC_parse) {
12476 RExC_seen_zerolen++;
12477 nextchar(pRExC_state);
12478 if (RExC_flags & RXf_PMf_MULTILINE)
12479 ret = reg_node(pRExC_state, MBOL);
12481 ret = reg_node(pRExC_state, SBOL);
12482 Set_Node_Length(ret, 1); /* MJD */
12485 nextchar(pRExC_state);
12487 RExC_seen_zerolen++;
12488 if (RExC_flags & RXf_PMf_MULTILINE)
12489 ret = reg_node(pRExC_state, MEOL);
12491 ret = reg_node(pRExC_state, SEOL);
12492 Set_Node_Length(ret, 1); /* MJD */
12495 nextchar(pRExC_state);
12496 if (RExC_flags & RXf_PMf_SINGLELINE)
12497 ret = reg_node(pRExC_state, SANY);
12499 ret = reg_node(pRExC_state, REG_ANY);
12500 *flagp |= HASWIDTH|SIMPLE;
12502 Set_Node_Length(ret, 1); /* MJD */
12506 char * const oregcomp_parse = ++RExC_parse;
12507 ret = regclass(pRExC_state, flagp,depth+1,
12508 FALSE, /* means parse the whole char class */
12509 TRUE, /* allow multi-char folds */
12510 FALSE, /* don't silence non-portable warnings. */
12511 (bool) RExC_strict,
12512 TRUE, /* Allow an optimized regnode result */
12516 if (*flagp & (RESTART_PASS1|NEED_UTF8))
12518 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12521 if (*RExC_parse != ']') {
12522 RExC_parse = oregcomp_parse;
12523 vFAIL("Unmatched [");
12525 nextchar(pRExC_state);
12526 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
12530 nextchar(pRExC_state);
12531 ret = reg(pRExC_state, 2, &flags,depth+1);
12533 if (flags & TRYAGAIN) {
12534 if (RExC_parse >= RExC_end) {
12535 /* Make parent create an empty node if needed. */
12536 *flagp |= TRYAGAIN;
12541 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12542 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12545 FAIL2("panic: reg returned NULL to regatom, flags=%#" UVxf,
12548 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12552 if (flags & TRYAGAIN) {
12553 *flagp |= TRYAGAIN;
12556 vFAIL("Internal urp");
12557 /* Supposed to be caught earlier. */
12563 vFAIL("Quantifier follows nothing");
12568 This switch handles escape sequences that resolve to some kind
12569 of special regop and not to literal text. Escape sequnces that
12570 resolve to literal text are handled below in the switch marked
12573 Every entry in this switch *must* have a corresponding entry
12574 in the literal escape switch. However, the opposite is not
12575 required, as the default for this switch is to jump to the
12576 literal text handling code.
12579 switch ((U8)*RExC_parse) {
12580 /* Special Escapes */
12582 RExC_seen_zerolen++;
12583 ret = reg_node(pRExC_state, SBOL);
12584 /* SBOL is shared with /^/ so we set the flags so we can tell
12585 * /\A/ from /^/ in split. We check ret because first pass we
12586 * have no regop struct to set the flags on. */
12590 goto finish_meta_pat;
12592 ret = reg_node(pRExC_state, GPOS);
12593 RExC_seen |= REG_GPOS_SEEN;
12595 goto finish_meta_pat;
12597 RExC_seen_zerolen++;
12598 ret = reg_node(pRExC_state, KEEPS);
12600 /* XXX:dmq : disabling in-place substitution seems to
12601 * be necessary here to avoid cases of memory corruption, as
12602 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
12604 RExC_seen |= REG_LOOKBEHIND_SEEN;
12605 goto finish_meta_pat;
12607 ret = reg_node(pRExC_state, SEOL);
12609 RExC_seen_zerolen++; /* Do not optimize RE away */
12610 goto finish_meta_pat;
12612 ret = reg_node(pRExC_state, EOS);
12614 RExC_seen_zerolen++; /* Do not optimize RE away */
12615 goto finish_meta_pat;
12617 vFAIL("\\C no longer supported");
12619 ret = reg_node(pRExC_state, CLUMP);
12620 *flagp |= HASWIDTH;
12621 goto finish_meta_pat;
12627 arg = ANYOF_WORDCHAR;
12635 regex_charset charset = get_regex_charset(RExC_flags);
12637 RExC_seen_zerolen++;
12638 RExC_seen |= REG_LOOKBEHIND_SEEN;
12639 op = BOUND + charset;
12641 if (op == BOUNDL) {
12642 RExC_contains_locale = 1;
12645 ret = reg_node(pRExC_state, op);
12647 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
12648 FLAGS(ret) = TRADITIONAL_BOUND;
12649 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
12655 char name = *RExC_parse;
12658 endbrace = strchr(RExC_parse, '}');
12661 vFAIL2("Missing right brace on \\%c{}", name);
12663 /* XXX Need to decide whether to take spaces or not. Should be
12664 * consistent with \p{}, but that currently is SPACE, which
12665 * means vertical too, which seems wrong
12666 * while (isBLANK(*RExC_parse)) {
12669 if (endbrace == RExC_parse) {
12670 RExC_parse++; /* After the '}' */
12671 vFAIL2("Empty \\%c{}", name);
12673 length = endbrace - RExC_parse;
12674 /*while (isBLANK(*(RExC_parse + length - 1))) {
12677 switch (*RExC_parse) {
12680 && (length != 3 || strnNE(RExC_parse + 1, "cb", 2)))
12682 goto bad_bound_type;
12684 FLAGS(ret) = GCB_BOUND;
12687 if (length != 2 || *(RExC_parse + 1) != 'b') {
12688 goto bad_bound_type;
12690 FLAGS(ret) = LB_BOUND;
12693 if (length != 2 || *(RExC_parse + 1) != 'b') {
12694 goto bad_bound_type;
12696 FLAGS(ret) = SB_BOUND;
12699 if (length != 2 || *(RExC_parse + 1) != 'b') {
12700 goto bad_bound_type;
12702 FLAGS(ret) = WB_BOUND;
12706 RExC_parse = endbrace;
12708 "'%" UTF8f "' is an unknown bound type",
12709 UTF8fARG(UTF, length, endbrace - length));
12710 NOT_REACHED; /*NOTREACHED*/
12712 RExC_parse = endbrace;
12713 REQUIRE_UNI_RULES(flagp, NULL);
12715 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
12719 /* Don't have to worry about UTF-8, in this message because
12720 * to get here the contents of the \b must be ASCII */
12721 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
12722 "Using /u for '%.*s' instead of /%s",
12724 endbrace - length + 1,
12725 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
12726 ? ASCII_RESTRICT_PAT_MODS
12727 : ASCII_MORE_RESTRICT_PAT_MODS);
12731 if (PASS2 && invert) {
12732 OP(ret) += NBOUND - BOUND;
12734 goto finish_meta_pat;
12742 if (! DEPENDS_SEMANTICS) {
12746 /* \d doesn't have any matches in the upper Latin1 range, hence /d
12747 * is equivalent to /u. Changing to /u saves some branches at
12750 goto join_posix_op_known;
12753 ret = reg_node(pRExC_state, LNBREAK);
12754 *flagp |= HASWIDTH|SIMPLE;
12755 goto finish_meta_pat;
12763 goto join_posix_op_known;
12769 arg = ANYOF_VERTWS;
12771 goto join_posix_op_known;
12781 op = POSIXD + get_regex_charset(RExC_flags);
12782 if (op > POSIXA) { /* /aa is same as /a */
12785 else if (op == POSIXL) {
12786 RExC_contains_locale = 1;
12789 join_posix_op_known:
12792 op += NPOSIXD - POSIXD;
12795 ret = reg_node(pRExC_state, op);
12797 FLAGS(ret) = namedclass_to_classnum(arg);
12800 *flagp |= HASWIDTH|SIMPLE;
12804 nextchar(pRExC_state);
12805 Set_Node_Length(ret, 2); /* MJD */
12811 ret = regclass(pRExC_state, flagp,depth+1,
12812 TRUE, /* means just parse this element */
12813 FALSE, /* don't allow multi-char folds */
12814 FALSE, /* don't silence non-portable warnings. It
12815 would be a bug if these returned
12817 (bool) RExC_strict,
12818 TRUE, /* Allow an optimized regnode result */
12821 if (*flagp & RESTART_PASS1)
12823 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
12824 * multi-char folds are allowed. */
12826 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12831 Set_Node_Offset(ret, parse_start);
12832 Set_Node_Cur_Length(ret, parse_start - 2);
12833 nextchar(pRExC_state);
12836 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
12837 * \N{...} evaluates to a sequence of more than one code points).
12838 * The function call below returns a regnode, which is our result.
12839 * The parameters cause it to fail if the \N{} evaluates to a
12840 * single code point; we handle those like any other literal. The
12841 * reason that the multicharacter case is handled here and not as
12842 * part of the EXACtish code is because of quantifiers. In
12843 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
12844 * this way makes that Just Happen. dmq.
12845 * join_exact() will join this up with adjacent EXACTish nodes
12846 * later on, if appropriate. */
12848 if (grok_bslash_N(pRExC_state,
12849 &ret, /* Want a regnode returned */
12850 NULL, /* Fail if evaluates to a single code
12852 NULL, /* Don't need a count of how many code
12861 if (*flagp & RESTART_PASS1)
12864 /* Here, evaluates to a single code point. Go get that */
12865 RExC_parse = parse_start;
12868 case 'k': /* Handle \k<NAME> and \k'NAME' */
12872 if ( RExC_parse >= RExC_end - 1
12873 || (( ch = RExC_parse[1]) != '<'
12878 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12879 vFAIL2("Sequence %.2s... not terminated",parse_start);
12882 ret = handle_named_backref(pRExC_state,
12894 case '1': case '2': case '3': case '4':
12895 case '5': case '6': case '7': case '8': case '9':
12900 if (*RExC_parse == 'g') {
12904 if (*RExC_parse == '{') {
12908 if (*RExC_parse == '-') {
12912 if (hasbrace && !isDIGIT(*RExC_parse)) {
12913 if (isrel) RExC_parse--;
12915 goto parse_named_seq;
12918 if (RExC_parse >= RExC_end) {
12919 goto unterminated_g;
12921 num = S_backref_value(RExC_parse);
12923 vFAIL("Reference to invalid group 0");
12924 else if (num == I32_MAX) {
12925 if (isDIGIT(*RExC_parse))
12926 vFAIL("Reference to nonexistent group");
12929 vFAIL("Unterminated \\g... pattern");
12933 num = RExC_npar - num;
12935 vFAIL("Reference to nonexistent or unclosed group");
12939 num = S_backref_value(RExC_parse);
12940 /* bare \NNN might be backref or octal - if it is larger
12941 * than or equal RExC_npar then it is assumed to be an
12942 * octal escape. Note RExC_npar is +1 from the actual
12943 * number of parens. */
12944 /* Note we do NOT check if num == I32_MAX here, as that is
12945 * handled by the RExC_npar check */
12948 /* any numeric escape < 10 is always a backref */
12950 /* any numeric escape < RExC_npar is a backref */
12951 && num >= RExC_npar
12952 /* cannot be an octal escape if it starts with 8 */
12953 && *RExC_parse != '8'
12954 /* cannot be an octal escape it it starts with 9 */
12955 && *RExC_parse != '9'
12958 /* Probably not a backref, instead likely to be an
12959 * octal character escape, e.g. \35 or \777.
12960 * The above logic should make it obvious why using
12961 * octal escapes in patterns is problematic. - Yves */
12962 RExC_parse = parse_start;
12967 /* At this point RExC_parse points at a numeric escape like
12968 * \12 or \88 or something similar, which we should NOT treat
12969 * as an octal escape. It may or may not be a valid backref
12970 * escape. For instance \88888888 is unlikely to be a valid
12972 while (isDIGIT(*RExC_parse))
12975 if (*RExC_parse != '}')
12976 vFAIL("Unterminated \\g{...} pattern");
12980 if (num > (I32)RExC_rx->nparens)
12981 vFAIL("Reference to nonexistent group");
12984 ret = reganode(pRExC_state,
12987 : (ASCII_FOLD_RESTRICTED)
12989 : (AT_LEAST_UNI_SEMANTICS)
12995 *flagp |= HASWIDTH;
12997 /* override incorrect value set in reganode MJD */
12998 Set_Node_Offset(ret, parse_start);
12999 Set_Node_Cur_Length(ret, parse_start-1);
13000 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13001 FALSE /* Don't force to /x */ );
13005 if (RExC_parse >= RExC_end)
13006 FAIL("Trailing \\");
13009 /* Do not generate "unrecognized" warnings here, we fall
13010 back into the quick-grab loop below */
13011 RExC_parse = parse_start;
13013 } /* end of switch on a \foo sequence */
13018 /* '#' comments should have been spaced over before this function was
13020 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13022 if (RExC_flags & RXf_PMf_EXTENDED) {
13023 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13024 if (RExC_parse < RExC_end)
13034 /* Here, we have determined that the next thing is probably a
13035 * literal character. RExC_parse points to the first byte of its
13036 * definition. (It still may be an escape sequence that evaluates
13037 * to a single character) */
13043 #define MAX_NODE_STRING_SIZE 127
13044 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
13046 U8 upper_parse = MAX_NODE_STRING_SIZE;
13047 U8 node_type = compute_EXACTish(pRExC_state);
13048 bool next_is_quantifier;
13049 char * oldp = NULL;
13051 /* We can convert EXACTF nodes to EXACTFU if they contain only
13052 * characters that match identically regardless of the target
13053 * string's UTF8ness. The reason to do this is that EXACTF is not
13054 * trie-able, EXACTFU is.
13056 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13057 * contain only above-Latin1 characters (hence must be in UTF8),
13058 * which don't participate in folds with Latin1-range characters,
13059 * as the latter's folds aren't known until runtime. (We don't
13060 * need to figure this out until pass 2) */
13061 bool maybe_exactfu = PASS2
13062 && (node_type == EXACTF || node_type == EXACTFL);
13064 /* If a folding node contains only code points that don't
13065 * participate in folds, it can be changed into an EXACT node,
13066 * which allows the optimizer more things to look for */
13069 ret = reg_node(pRExC_state, node_type);
13071 /* In pass1, folded, we use a temporary buffer instead of the
13072 * actual node, as the node doesn't exist yet */
13073 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
13079 /* We look for the EXACTFish to EXACT node optimizaton only if
13080 * folding. (And we don't need to figure this out until pass 2).
13081 * XXX It might actually make sense to split the node into portions
13082 * that are exact and ones that aren't, so that we could later use
13083 * the exact ones to find the longest fixed and floating strings.
13084 * One would want to join them back into a larger node. One could
13085 * use a pseudo regnode like 'EXACT_ORIG_FOLD' */
13086 maybe_exact = FOLD && PASS2;
13088 /* XXX The node can hold up to 255 bytes, yet this only goes to
13089 * 127. I (khw) do not know why. Keeping it somewhat less than
13090 * 255 allows us to not have to worry about overflow due to
13091 * converting to utf8 and fold expansion, but that value is
13092 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
13093 * split up by this limit into a single one using the real max of
13094 * 255. Even at 127, this breaks under rare circumstances. If
13095 * folding, we do not want to split a node at a character that is a
13096 * non-final in a multi-char fold, as an input string could just
13097 * happen to want to match across the node boundary. The join
13098 * would solve that problem if the join actually happens. But a
13099 * series of more than two nodes in a row each of 127 would cause
13100 * the first join to succeed to get to 254, but then there wouldn't
13101 * be room for the next one, which could at be one of those split
13102 * multi-char folds. I don't know of any fool-proof solution. One
13103 * could back off to end with only a code point that isn't such a
13104 * non-final, but it is possible for there not to be any in the
13107 assert( ! UTF /* Is at the beginning of a character */
13108 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13109 || UTF8_IS_START(UCHARAT(RExC_parse)));
13111 /* Here, we have a literal character. Find the maximal string of
13112 * them in the input that we can fit into a single EXACTish node.
13113 * We quit at the first non-literal or when the node gets full */
13114 for (p = RExC_parse;
13115 len < upper_parse && p < RExC_end;
13120 /* White space has already been ignored */
13121 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13122 || ! is_PATWS_safe((p), RExC_end, UTF));
13134 /* Literal Escapes Switch
13136 This switch is meant to handle escape sequences that
13137 resolve to a literal character.
13139 Every escape sequence that represents something
13140 else, like an assertion or a char class, is handled
13141 in the switch marked 'Special Escapes' above in this
13142 routine, but also has an entry here as anything that
13143 isn't explicitly mentioned here will be treated as
13144 an unescaped equivalent literal.
13147 switch ((U8)*++p) {
13148 /* These are all the special escapes. */
13149 case 'A': /* Start assertion */
13150 case 'b': case 'B': /* Word-boundary assertion*/
13151 case 'C': /* Single char !DANGEROUS! */
13152 case 'd': case 'D': /* digit class */
13153 case 'g': case 'G': /* generic-backref, pos assertion */
13154 case 'h': case 'H': /* HORIZWS */
13155 case 'k': case 'K': /* named backref, keep marker */
13156 case 'p': case 'P': /* Unicode property */
13157 case 'R': /* LNBREAK */
13158 case 's': case 'S': /* space class */
13159 case 'v': case 'V': /* VERTWS */
13160 case 'w': case 'W': /* word class */
13161 case 'X': /* eXtended Unicode "combining
13162 character sequence" */
13163 case 'z': case 'Z': /* End of line/string assertion */
13167 /* Anything after here is an escape that resolves to a
13168 literal. (Except digits, which may or may not)
13174 case 'N': /* Handle a single-code point named character. */
13175 RExC_parse = p + 1;
13176 if (! grok_bslash_N(pRExC_state,
13177 NULL, /* Fail if evaluates to
13178 anything other than a
13179 single code point */
13180 &ender, /* The returned single code
13182 NULL, /* Don't need a count of
13183 how many code points */
13188 if (*flagp & NEED_UTF8)
13189 FAIL("panic: grok_bslash_N set NEED_UTF8");
13190 if (*flagp & RESTART_PASS1)
13193 /* Here, it wasn't a single code point. Go close
13194 * up this EXACTish node. The switch() prior to
13195 * this switch handles the other cases */
13196 RExC_parse = p = oldp;
13200 if (ender > 0xff) {
13201 REQUIRE_UTF8(flagp);
13217 ender = ESC_NATIVE;
13227 const char* error_msg;
13229 bool valid = grok_bslash_o(&p,
13232 PASS2, /* out warnings */
13233 (bool) RExC_strict,
13234 TRUE, /* Output warnings
13239 RExC_parse = p; /* going to die anyway; point
13240 to exact spot of failure */
13244 if (ender > 0xff) {
13245 REQUIRE_UTF8(flagp);
13251 UV result = UV_MAX; /* initialize to erroneous
13253 const char* error_msg;
13255 bool valid = grok_bslash_x(&p,
13258 PASS2, /* out warnings */
13259 (bool) RExC_strict,
13260 TRUE, /* Silence warnings
13265 RExC_parse = p; /* going to die anyway; point
13266 to exact spot of failure */
13271 if (ender < 0x100) {
13273 if (RExC_recode_x_to_native) {
13274 ender = LATIN1_TO_NATIVE(ender);
13279 REQUIRE_UTF8(flagp);
13285 ender = grok_bslash_c(*p++, PASS2);
13287 case '8': case '9': /* must be a backreference */
13289 /* we have an escape like \8 which cannot be an octal escape
13290 * so we exit the loop, and let the outer loop handle this
13291 * escape which may or may not be a legitimate backref. */
13293 case '1': case '2': case '3':case '4':
13294 case '5': case '6': case '7':
13295 /* When we parse backslash escapes there is ambiguity
13296 * between backreferences and octal escapes. Any escape
13297 * from \1 - \9 is a backreference, any multi-digit
13298 * escape which does not start with 0 and which when
13299 * evaluated as decimal could refer to an already
13300 * parsed capture buffer is a back reference. Anything
13303 * Note this implies that \118 could be interpreted as
13304 * 118 OR as "\11" . "8" depending on whether there
13305 * were 118 capture buffers defined already in the
13308 /* NOTE, RExC_npar is 1 more than the actual number of
13309 * parens we have seen so far, hence the < RExC_npar below. */
13311 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
13312 { /* Not to be treated as an octal constant, go
13320 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13322 ender = grok_oct(p, &numlen, &flags, NULL);
13323 if (ender > 0xff) {
13324 REQUIRE_UTF8(flagp);
13327 if (PASS2 /* like \08, \178 */
13329 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
13331 reg_warn_non_literal_string(
13333 form_short_octal_warning(p, numlen));
13339 FAIL("Trailing \\");
13342 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
13343 /* Include any left brace following the alpha to emphasize
13344 * that it could be part of an escape at some point
13346 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
13347 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
13349 goto normal_default;
13350 } /* End of switch on '\' */
13353 /* Currently we don't care if the lbrace is at the start
13354 * of a construct. This catches it in the middle of a
13355 * literal string, or when it's the first thing after
13356 * something like "\b" */
13357 if (len || (p > RExC_start && isALPHA_A(*(p -1)))) {
13358 RExC_parse = p + 1;
13359 vFAIL("Unescaped left brace in regex is illegal here");
13362 default: /* A literal character */
13364 if (! UTF8_IS_INVARIANT(*p) && UTF) {
13366 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
13367 &numlen, UTF8_ALLOW_DEFAULT);
13373 } /* End of switch on the literal */
13375 /* Here, have looked at the literal character and <ender>
13376 * contains its ordinal, <p> points to the character after it.
13377 * We need to check if the next non-ignored thing is a
13378 * quantifier. Move <p> to after anything that should be
13379 * ignored, which, as a side effect, positions <p> for the next
13380 * loop iteration */
13381 skip_to_be_ignored_text(pRExC_state, &p,
13382 FALSE /* Don't force to /x */ );
13384 /* If the next thing is a quantifier, it applies to this
13385 * character only, which means that this character has to be in
13386 * its own node and can't just be appended to the string in an
13387 * existing node, so if there are already other characters in
13388 * the node, close the node with just them, and set up to do
13389 * this character again next time through, when it will be the
13390 * only thing in its new node */
13392 if ((next_is_quantifier = ( LIKELY(p < RExC_end)
13393 && UNLIKELY(ISMULT2(p))))
13400 /* Ready to add 'ender' to the node */
13402 if (! FOLD) { /* The simple case, just append the literal */
13404 /* In the sizing pass, we need only the size of the
13405 * character we are appending, hence we can delay getting
13406 * its representation until PASS2. */
13409 const STRLEN unilen = UVCHR_SKIP(ender);
13412 /* We have to subtract 1 just below (and again in
13413 * the corresponding PASS2 code) because the loop
13414 * increments <len> each time, as all but this path
13415 * (and one other) through it add a single byte to
13416 * the EXACTish node. But these paths would change
13417 * len to be the correct final value, so cancel out
13418 * the increment that follows */
13424 } else { /* PASS2 */
13427 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
13428 len += (char *) new_s - s - 1;
13429 s = (char *) new_s;
13432 *(s++) = (char) ender;
13436 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
13438 /* Here are folding under /l, and the code point is
13439 * problematic. First, we know we can't simplify things */
13440 maybe_exact = FALSE;
13441 maybe_exactfu = FALSE;
13443 /* A problematic code point in this context means that its
13444 * fold isn't known until runtime, so we can't fold it now.
13445 * (The non-problematic code points are the above-Latin1
13446 * ones that fold to also all above-Latin1. Their folds
13447 * don't vary no matter what the locale is.) But here we
13448 * have characters whose fold depends on the locale.
13449 * Unlike the non-folding case above, we have to keep track
13450 * of these in the sizing pass, so that we can make sure we
13451 * don't split too-long nodes in the middle of a potential
13452 * multi-char fold. And unlike the regular fold case
13453 * handled in the else clauses below, we don't actually
13454 * fold and don't have special cases to consider. What we
13455 * do for both passes is the PASS2 code for non-folding */
13456 goto not_fold_common;
13458 else /* A regular FOLD code point */
13460 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13461 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13462 || UNICODE_DOT_DOT_VERSION > 0)
13463 /* See comments for join_exact() as to why we fold
13464 * this non-UTF at compile time */
13465 || ( node_type == EXACTFU
13466 && ender == LATIN_SMALL_LETTER_SHARP_S)
13469 /* Here, are folding and are not UTF-8 encoded; therefore
13470 * the character must be in the range 0-255, and is not /l
13471 * (Not /l because we already handled these under /l in
13472 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
13473 if (IS_IN_SOME_FOLD_L1(ender)) {
13474 maybe_exact = FALSE;
13476 /* See if the character's fold differs between /d and
13477 * /u. This includes the multi-char fold SHARP S to
13479 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
13480 RExC_seen_unfolded_sharp_s = 1;
13481 maybe_exactfu = FALSE;
13483 else if (maybe_exactfu
13484 && (PL_fold[ender] != PL_fold_latin1[ender]
13485 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13486 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13487 || UNICODE_DOT_DOT_VERSION > 0)
13489 && isALPHA_FOLD_EQ(ender, 's')
13490 && isALPHA_FOLD_EQ(*(s-1), 's'))
13493 maybe_exactfu = FALSE;
13497 /* Even when folding, we store just the input character, as
13498 * we have an array that finds its fold quickly */
13499 *(s++) = (char) ender;
13501 else { /* FOLD, and UTF (or sharp s) */
13502 /* Unlike the non-fold case, we do actually have to
13503 * calculate the results here in pass 1. This is for two
13504 * reasons, the folded length may be longer than the
13505 * unfolded, and we have to calculate how many EXACTish
13506 * nodes it will take; and we may run out of room in a node
13507 * in the middle of a potential multi-char fold, and have
13508 * to back off accordingly. */
13511 if (isASCII_uni(ender)) {
13512 folded = toFOLD(ender);
13513 *(s)++ = (U8) folded;
13518 folded = _to_uni_fold_flags(
13522 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
13523 ? FOLD_FLAGS_NOMIX_ASCII
13527 /* The loop increments <len> each time, as all but this
13528 * path (and one other) through it add a single byte to
13529 * the EXACTish node. But this one has changed len to
13530 * be the correct final value, so subtract one to
13531 * cancel out the increment that follows */
13532 len += foldlen - 1;
13534 /* If this node only contains non-folding code points so
13535 * far, see if this new one is also non-folding */
13537 if (folded != ender) {
13538 maybe_exact = FALSE;
13541 /* Here the fold is the original; we have to check
13542 * further to see if anything folds to it */
13543 if (_invlist_contains_cp(PL_utf8_foldable,
13546 maybe_exact = FALSE;
13553 if (next_is_quantifier) {
13555 /* Here, the next input is a quantifier, and to get here,
13556 * the current character is the only one in the node.
13557 * Also, here <len> doesn't include the final byte for this
13563 } /* End of loop through literal characters */
13565 /* Here we have either exhausted the input or ran out of room in
13566 * the node. (If we encountered a character that can't be in the
13567 * node, transfer is made directly to <loopdone>, and so we
13568 * wouldn't have fallen off the end of the loop.) In the latter
13569 * case, we artificially have to split the node into two, because
13570 * we just don't have enough space to hold everything. This
13571 * creates a problem if the final character participates in a
13572 * multi-character fold in the non-final position, as a match that
13573 * should have occurred won't, due to the way nodes are matched,
13574 * and our artificial boundary. So back off until we find a non-
13575 * problematic character -- one that isn't at the beginning or
13576 * middle of such a fold. (Either it doesn't participate in any
13577 * folds, or appears only in the final position of all the folds it
13578 * does participate in.) A better solution with far fewer false
13579 * positives, and that would fill the nodes more completely, would
13580 * be to actually have available all the multi-character folds to
13581 * test against, and to back-off only far enough to be sure that
13582 * this node isn't ending with a partial one. <upper_parse> is set
13583 * further below (if we need to reparse the node) to include just
13584 * up through that final non-problematic character that this code
13585 * identifies, so when it is set to less than the full node, we can
13586 * skip the rest of this */
13587 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
13589 const STRLEN full_len = len;
13591 assert(len >= MAX_NODE_STRING_SIZE);
13593 /* Here, <s> points to the final byte of the final character.
13594 * Look backwards through the string until find a non-
13595 * problematic character */
13599 /* This has no multi-char folds to non-UTF characters */
13600 if (ASCII_FOLD_RESTRICTED) {
13604 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
13608 if (! PL_NonL1NonFinalFold) {
13609 PL_NonL1NonFinalFold = _new_invlist_C_array(
13610 NonL1_Perl_Non_Final_Folds_invlist);
13613 /* Point to the first byte of the final character */
13614 s = (char *) utf8_hop((U8 *) s, -1);
13616 while (s >= s0) { /* Search backwards until find
13617 non-problematic char */
13618 if (UTF8_IS_INVARIANT(*s)) {
13620 /* There are no ascii characters that participate
13621 * in multi-char folds under /aa. In EBCDIC, the
13622 * non-ascii invariants are all control characters,
13623 * so don't ever participate in any folds. */
13624 if (ASCII_FOLD_RESTRICTED
13625 || ! IS_NON_FINAL_FOLD(*s))
13630 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
13631 if (! IS_NON_FINAL_FOLD(EIGHT_BIT_UTF8_TO_NATIVE(
13637 else if (! _invlist_contains_cp(
13638 PL_NonL1NonFinalFold,
13639 valid_utf8_to_uvchr((U8 *) s, NULL)))
13644 /* Here, the current character is problematic in that
13645 * it does occur in the non-final position of some
13646 * fold, so try the character before it, but have to
13647 * special case the very first byte in the string, so
13648 * we don't read outside the string */
13649 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
13650 } /* End of loop backwards through the string */
13652 /* If there were only problematic characters in the string,
13653 * <s> will point to before s0, in which case the length
13654 * should be 0, otherwise include the length of the
13655 * non-problematic character just found */
13656 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
13659 /* Here, have found the final character, if any, that is
13660 * non-problematic as far as ending the node without splitting
13661 * it across a potential multi-char fold. <len> contains the
13662 * number of bytes in the node up-to and including that
13663 * character, or is 0 if there is no such character, meaning
13664 * the whole node contains only problematic characters. In
13665 * this case, give up and just take the node as-is. We can't
13670 /* If the node ends in an 's' we make sure it stays EXACTF,
13671 * as if it turns into an EXACTFU, it could later get
13672 * joined with another 's' that would then wrongly match
13674 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
13676 maybe_exactfu = FALSE;
13680 /* Here, the node does contain some characters that aren't
13681 * problematic. If one such is the final character in the
13682 * node, we are done */
13683 if (len == full_len) {
13686 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
13688 /* If the final character is problematic, but the
13689 * penultimate is not, back-off that last character to
13690 * later start a new node with it */
13695 /* Here, the final non-problematic character is earlier
13696 * in the input than the penultimate character. What we do
13697 * is reparse from the beginning, going up only as far as
13698 * this final ok one, thus guaranteeing that the node ends
13699 * in an acceptable character. The reason we reparse is
13700 * that we know how far in the character is, but we don't
13701 * know how to correlate its position with the input parse.
13702 * An alternate implementation would be to build that
13703 * correlation as we go along during the original parse,
13704 * but that would entail extra work for every node, whereas
13705 * this code gets executed only when the string is too
13706 * large for the node, and the final two characters are
13707 * problematic, an infrequent occurrence. Yet another
13708 * possible strategy would be to save the tail of the
13709 * string, and the next time regatom is called, initialize
13710 * with that. The problem with this is that unless you
13711 * back off one more character, you won't be guaranteed
13712 * regatom will get called again, unless regbranch,
13713 * regpiece ... are also changed. If you do back off that
13714 * extra character, so that there is input guaranteed to
13715 * force calling regatom, you can't handle the case where
13716 * just the first character in the node is acceptable. I
13717 * (khw) decided to try this method which doesn't have that
13718 * pitfall; if performance issues are found, we can do a
13719 * combination of the current approach plus that one */
13725 } /* End of verifying node ends with an appropriate char */
13727 loopdone: /* Jumped to when encounters something that shouldn't be
13730 /* I (khw) don't know if you can get here with zero length, but the
13731 * old code handled this situation by creating a zero-length EXACT
13732 * node. Might as well be NOTHING instead */
13738 /* If 'maybe_exact' is still set here, means there are no
13739 * code points in the node that participate in folds;
13740 * similarly for 'maybe_exactfu' and code points that match
13741 * differently depending on UTF8ness of the target string
13742 * (for /u), or depending on locale for /l */
13748 else if (maybe_exactfu) {
13754 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
13755 FALSE /* Don't look to see if could
13756 be turned into an EXACT
13757 node, as we have already
13762 RExC_parse = p - 1;
13763 Set_Node_Cur_Length(ret, parse_start);
13766 /* len is STRLEN which is unsigned, need to copy to signed */
13769 vFAIL("Internal disaster");
13772 } /* End of label 'defchar:' */
13774 } /* End of giant switch on input character */
13776 /* Position parse to next real character */
13777 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13778 FALSE /* Don't force to /x */ );
13779 if (PASS2 && *RExC_parse == '{' && OP(ret) != SBOL && ! regcurly(RExC_parse)) {
13780 ckWARNregdep(RExC_parse + 1, "Unescaped left brace in regex is deprecated here, passed through");
13788 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
13790 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
13791 * sets up the bitmap and any flags, removing those code points from the
13792 * inversion list, setting it to NULL should it become completely empty */
13794 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
13795 assert(PL_regkind[OP(node)] == ANYOF);
13797 ANYOF_BITMAP_ZERO(node);
13798 if (*invlist_ptr) {
13800 /* This gets set if we actually need to modify things */
13801 bool change_invlist = FALSE;
13805 /* Start looking through *invlist_ptr */
13806 invlist_iterinit(*invlist_ptr);
13807 while (invlist_iternext(*invlist_ptr, &start, &end)) {
13811 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
13812 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
13815 /* Quit if are above what we should change */
13816 if (start >= NUM_ANYOF_CODE_POINTS) {
13820 change_invlist = TRUE;
13822 /* Set all the bits in the range, up to the max that we are doing */
13823 high = (end < NUM_ANYOF_CODE_POINTS - 1)
13825 : NUM_ANYOF_CODE_POINTS - 1;
13826 for (i = start; i <= (int) high; i++) {
13827 if (! ANYOF_BITMAP_TEST(node, i)) {
13828 ANYOF_BITMAP_SET(node, i);
13832 invlist_iterfinish(*invlist_ptr);
13834 /* Done with loop; remove any code points that are in the bitmap from
13835 * *invlist_ptr; similarly for code points above the bitmap if we have
13836 * a flag to match all of them anyways */
13837 if (change_invlist) {
13838 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
13840 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
13841 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
13844 /* If have completely emptied it, remove it completely */
13845 if (_invlist_len(*invlist_ptr) == 0) {
13846 SvREFCNT_dec_NN(*invlist_ptr);
13847 *invlist_ptr = NULL;
13852 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
13853 Character classes ([:foo:]) can also be negated ([:^foo:]).
13854 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
13855 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
13856 but trigger failures because they are currently unimplemented. */
13858 #define POSIXCC_DONE(c) ((c) == ':')
13859 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
13860 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
13861 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
13863 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
13864 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
13865 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
13867 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
13869 /* 'posix_warnings' and 'warn_text' are names of variables in the following
13871 #define ADD_POSIX_WARNING(p, text) STMT_START { \
13872 if (posix_warnings) { \
13873 if (! RExC_warn_text ) RExC_warn_text = (AV *) sv_2mortal((SV *) newAV()); \
13874 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
13878 REPORT_LOCATION_ARGS(p))); \
13883 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
13885 const char * const s, /* Where the putative posix class begins.
13886 Normally, this is one past the '['. This
13887 parameter exists so it can be somewhere
13888 besides RExC_parse. */
13889 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
13891 AV ** posix_warnings, /* Where to place any generated warnings, or
13893 const bool check_only /* Don't die if error */
13896 /* This parses what the caller thinks may be one of the three POSIX
13898 * 1) a character class, like [:blank:]
13899 * 2) a collating symbol, like [. .]
13900 * 3) an equivalence class, like [= =]
13901 * In the latter two cases, it croaks if it finds a syntactically legal
13902 * one, as these are not handled by Perl.
13904 * The main purpose is to look for a POSIX character class. It returns:
13905 * a) the class number
13906 * if it is a completely syntactically and semantically legal class.
13907 * 'updated_parse_ptr', if not NULL, is set to point to just after the
13908 * closing ']' of the class
13909 * b) OOB_NAMEDCLASS
13910 * if it appears that one of the three POSIX constructs was meant, but
13911 * its specification was somehow defective. 'updated_parse_ptr', if
13912 * not NULL, is set to point to the character just after the end
13913 * character of the class. See below for handling of warnings.
13914 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
13915 * if it doesn't appear that a POSIX construct was intended.
13916 * 'updated_parse_ptr' is not changed. No warnings nor errors are
13919 * In b) there may be errors or warnings generated. If 'check_only' is
13920 * TRUE, then any errors are discarded. Warnings are returned to the
13921 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
13922 * instead it is NULL, warnings are suppressed. This is done in all
13923 * passes. The reason for this is that the rest of the parsing is heavily
13924 * dependent on whether this routine found a valid posix class or not. If
13925 * it did, the closing ']' is absorbed as part of the class. If no class,
13926 * or an invalid one is found, any ']' will be considered the terminator of
13927 * the outer bracketed character class, leading to very different results.
13928 * In particular, a '(?[ ])' construct will likely have a syntax error if
13929 * the class is parsed other than intended, and this will happen in pass1,
13930 * before the warnings would normally be output. This mechanism allows the
13931 * caller to output those warnings in pass1 just before dieing, giving a
13932 * much better clue as to what is wrong.
13934 * The reason for this function, and its complexity is that a bracketed
13935 * character class can contain just about anything. But it's easy to
13936 * mistype the very specific posix class syntax but yielding a valid
13937 * regular bracketed class, so it silently gets compiled into something
13938 * quite unintended.
13940 * The solution adopted here maintains backward compatibility except that
13941 * it adds a warning if it looks like a posix class was intended but
13942 * improperly specified. The warning is not raised unless what is input
13943 * very closely resembles one of the 14 legal posix classes. To do this,
13944 * it uses fuzzy parsing. It calculates how many single-character edits it
13945 * would take to transform what was input into a legal posix class. Only
13946 * if that number is quite small does it think that the intention was a
13947 * posix class. Obviously these are heuristics, and there will be cases
13948 * where it errs on one side or another, and they can be tweaked as
13949 * experience informs.
13951 * The syntax for a legal posix class is:
13953 * qr/(?xa: \[ : \^? [:lower:]{4,6} : \] )/
13955 * What this routine considers syntactically to be an intended posix class
13956 * is this (the comments indicate some restrictions that the pattern
13959 * qr/(?x: \[? # The left bracket, possibly
13961 * \h* # possibly followed by blanks
13962 * (?: \^ \h* )? # possibly a misplaced caret
13963 * [:;]? # The opening class character,
13964 * # possibly omitted. A typo
13965 * # semi-colon can also be used.
13967 * \^? # possibly a correctly placed
13968 * # caret, but not if there was also
13969 * # a misplaced one
13971 * .{3,15} # The class name. If there are
13972 * # deviations from the legal syntax,
13973 * # its edit distance must be close
13974 * # to a real class name in order
13975 * # for it to be considered to be
13976 * # an intended posix class.
13978 * [:punct:]? # The closing class character,
13979 * # possibly omitted. If not a colon
13980 * # nor semi colon, the class name
13981 * # must be even closer to a valid
13984 * \]? # The right bracket, possibly
13988 * In the above, \h must be ASCII-only.
13990 * These are heuristics, and can be tweaked as field experience dictates.
13991 * There will be cases when someone didn't intend to specify a posix class
13992 * that this warns as being so. The goal is to minimize these, while
13993 * maximizing the catching of things intended to be a posix class that
13994 * aren't parsed as such.
13998 const char * const e = RExC_end;
13999 unsigned complement = 0; /* If to complement the class */
14000 bool found_problem = FALSE; /* Assume OK until proven otherwise */
14001 bool has_opening_bracket = FALSE;
14002 bool has_opening_colon = FALSE;
14003 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
14005 const char * possible_end = NULL; /* used for a 2nd parse pass */
14006 const char* name_start; /* ptr to class name first char */
14008 /* If the number of single-character typos the input name is away from a
14009 * legal name is no more than this number, it is considered to have meant
14010 * the legal name */
14011 int max_distance = 2;
14013 /* to store the name. The size determines the maximum length before we
14014 * decide that no posix class was intended. Should be at least
14015 * sizeof("alphanumeric") */
14018 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
14020 if (posix_warnings && RExC_warn_text)
14021 av_clear(RExC_warn_text);
14024 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14027 if (*(p - 1) != '[') {
14028 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
14029 found_problem = TRUE;
14032 has_opening_bracket = TRUE;
14035 /* They could be confused and think you can put spaces between the
14038 found_problem = TRUE;
14042 } while (p < e && isBLANK(*p));
14044 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14047 /* For [. .] and [= =]. These are quite different internally from [: :],
14048 * so they are handled separately. */
14049 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
14050 and 1 for at least one char in it
14053 const char open_char = *p;
14054 const char * temp_ptr = p + 1;
14056 /* These two constructs are not handled by perl, and if we find a
14057 * syntactically valid one, we croak. khw, who wrote this code, finds
14058 * this explanation of them very unclear:
14059 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
14060 * And searching the rest of the internet wasn't very helpful either.
14061 * It looks like just about any byte can be in these constructs,
14062 * depending on the locale. But unless the pattern is being compiled
14063 * under /l, which is very rare, Perl runs under the C or POSIX locale.
14064 * In that case, it looks like [= =] isn't allowed at all, and that
14065 * [. .] could be any single code point, but for longer strings the
14066 * constituent characters would have to be the ASCII alphabetics plus
14067 * the minus-hyphen. Any sensible locale definition would limit itself
14068 * to these. And any portable one definitely should. Trying to parse
14069 * the general case is a nightmare (see [perl #127604]). So, this code
14070 * looks only for interiors of these constructs that match:
14072 * Using \w relaxes the apparent rules a little, without adding much
14073 * danger of mistaking something else for one of these constructs.
14075 * [. .] in some implementations described on the internet is usable to
14076 * escape a character that otherwise is special in bracketed character
14077 * classes. For example [.].] means a literal right bracket instead of
14078 * the ending of the class
14080 * [= =] can legitimately contain a [. .] construct, but we don't
14081 * handle this case, as that [. .] construct will later get parsed
14082 * itself and croak then. And [= =] is checked for even when not under
14083 * /l, as Perl has long done so.
14085 * The code below relies on there being a trailing NUL, so it doesn't
14086 * have to keep checking if the parse ptr < e.
14088 if (temp_ptr[1] == open_char) {
14091 else while ( temp_ptr < e
14092 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
14097 if (*temp_ptr == open_char) {
14099 if (*temp_ptr == ']') {
14101 if (! found_problem && ! check_only) {
14102 RExC_parse = (char *) temp_ptr;
14103 vFAIL3("POSIX syntax [%c %c] is reserved for future "
14104 "extensions", open_char, open_char);
14107 /* Here, the syntax wasn't completely valid, or else the call
14108 * is to check-only */
14109 if (updated_parse_ptr) {
14110 *updated_parse_ptr = (char *) temp_ptr;
14113 return OOB_NAMEDCLASS;
14117 /* If we find something that started out to look like one of these
14118 * constructs, but isn't, we continue below so that it can be checked
14119 * for being a class name with a typo of '.' or '=' instead of a colon.
14123 /* Here, we think there is a possibility that a [: :] class was meant, and
14124 * we have the first real character. It could be they think the '^' comes
14127 found_problem = TRUE;
14128 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
14133 found_problem = TRUE;
14137 } while (p < e && isBLANK(*p));
14139 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14143 /* But the first character should be a colon, which they could have easily
14144 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
14145 * distinguish from a colon, so treat that as a colon). */
14148 has_opening_colon = TRUE;
14150 else if (*p == ';') {
14151 found_problem = TRUE;
14153 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14154 has_opening_colon = TRUE;
14157 found_problem = TRUE;
14158 ADD_POSIX_WARNING(p, "there must be a starting ':'");
14160 /* Consider an initial punctuation (not one of the recognized ones) to
14161 * be a left terminator */
14162 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
14167 /* They may think that you can put spaces between the components */
14169 found_problem = TRUE;
14173 } while (p < e && isBLANK(*p));
14175 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14180 /* We consider something like [^:^alnum:]] to not have been intended to
14181 * be a posix class, but XXX maybe we should */
14183 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14190 /* Again, they may think that you can put spaces between the components */
14192 found_problem = TRUE;
14196 } while (p < e && isBLANK(*p));
14198 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14203 /* XXX This ']' may be a typo, and something else was meant. But
14204 * treating it as such creates enough complications, that that
14205 * possibility isn't currently considered here. So we assume that the
14206 * ']' is what is intended, and if we've already found an initial '[',
14207 * this leaves this construct looking like [:] or [:^], which almost
14208 * certainly weren't intended to be posix classes */
14209 if (has_opening_bracket) {
14210 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14213 /* But this function can be called when we parse the colon for
14214 * something like qr/[alpha:]]/, so we back up to look for the
14219 found_problem = TRUE;
14220 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14222 else if (*p != ':') {
14224 /* XXX We are currently very restrictive here, so this code doesn't
14225 * consider the possibility that, say, /[alpha.]]/ was intended to
14226 * be a posix class. */
14227 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14230 /* Here we have something like 'foo:]'. There was no initial colon,
14231 * and we back up over 'foo. XXX Unlike the going forward case, we
14232 * don't handle typos of non-word chars in the middle */
14233 has_opening_colon = FALSE;
14236 while (p > RExC_start && isWORDCHAR(*p)) {
14241 /* Here, we have positioned ourselves to where we think the first
14242 * character in the potential class is */
14245 /* Now the interior really starts. There are certain key characters that
14246 * can end the interior, or these could just be typos. To catch both
14247 * cases, we may have to do two passes. In the first pass, we keep on
14248 * going unless we come to a sequence that matches
14249 * qr/ [[:punct:]] [[:blank:]]* \] /xa
14250 * This means it takes a sequence to end the pass, so two typos in a row if
14251 * that wasn't what was intended. If the class is perfectly formed, just
14252 * this one pass is needed. We also stop if there are too many characters
14253 * being accumulated, but this number is deliberately set higher than any
14254 * real class. It is set high enough so that someone who thinks that
14255 * 'alphanumeric' is a correct name would get warned that it wasn't.
14256 * While doing the pass, we keep track of where the key characters were in
14257 * it. If we don't find an end to the class, and one of the key characters
14258 * was found, we redo the pass, but stop when we get to that character.
14259 * Thus the key character was considered a typo in the first pass, but a
14260 * terminator in the second. If two key characters are found, we stop at
14261 * the second one in the first pass. Again this can miss two typos, but
14262 * catches a single one
14264 * In the first pass, 'possible_end' starts as NULL, and then gets set to
14265 * point to the first key character. For the second pass, it starts as -1.
14271 bool has_blank = FALSE;
14272 bool has_upper = FALSE;
14273 bool has_terminating_colon = FALSE;
14274 bool has_terminating_bracket = FALSE;
14275 bool has_semi_colon = FALSE;
14276 unsigned int name_len = 0;
14277 int punct_count = 0;
14281 /* Squeeze out blanks when looking up the class name below */
14282 if (isBLANK(*p) ) {
14284 found_problem = TRUE;
14289 /* The name will end with a punctuation */
14291 const char * peek = p + 1;
14293 /* Treat any non-']' punctuation followed by a ']' (possibly
14294 * with intervening blanks) as trying to terminate the class.
14295 * ']]' is very likely to mean a class was intended (but
14296 * missing the colon), but the warning message that gets
14297 * generated shows the error position better if we exit the
14298 * loop at the bottom (eventually), so skip it here. */
14300 if (peek < e && isBLANK(*peek)) {
14302 found_problem = TRUE;
14305 } while (peek < e && isBLANK(*peek));
14308 if (peek < e && *peek == ']') {
14309 has_terminating_bracket = TRUE;
14311 has_terminating_colon = TRUE;
14313 else if (*p == ';') {
14314 has_semi_colon = TRUE;
14315 has_terminating_colon = TRUE;
14318 found_problem = TRUE;
14325 /* Here we have punctuation we thought didn't end the class.
14326 * Keep track of the position of the key characters that are
14327 * more likely to have been class-enders */
14328 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
14330 /* Allow just one such possible class-ender not actually
14331 * ending the class. */
14332 if (possible_end) {
14338 /* If we have too many punctuation characters, no use in
14340 if (++punct_count > max_distance) {
14344 /* Treat the punctuation as a typo. */
14345 input_text[name_len++] = *p;
14348 else if (isUPPER(*p)) { /* Use lowercase for lookup */
14349 input_text[name_len++] = toLOWER(*p);
14351 found_problem = TRUE;
14353 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
14354 input_text[name_len++] = *p;
14358 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
14362 /* The declaration of 'input_text' is how long we allow a potential
14363 * class name to be, before saying they didn't mean a class name at
14365 if (name_len >= C_ARRAY_LENGTH(input_text)) {
14370 /* We get to here when the possible class name hasn't been properly
14371 * terminated before:
14372 * 1) we ran off the end of the pattern; or
14373 * 2) found two characters, each of which might have been intended to
14374 * be the name's terminator
14375 * 3) found so many punctuation characters in the purported name,
14376 * that the edit distance to a valid one is exceeded
14377 * 4) we decided it was more characters than anyone could have
14378 * intended to be one. */
14380 found_problem = TRUE;
14382 /* In the final two cases, we know that looking up what we've
14383 * accumulated won't lead to a match, even a fuzzy one. */
14384 if ( name_len >= C_ARRAY_LENGTH(input_text)
14385 || punct_count > max_distance)
14387 /* If there was an intermediate key character that could have been
14388 * an intended end, redo the parse, but stop there */
14389 if (possible_end && possible_end != (char *) -1) {
14390 possible_end = (char *) -1; /* Special signal value to say
14391 we've done a first pass */
14396 /* Otherwise, it can't have meant to have been a class */
14397 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14400 /* If we ran off the end, and the final character was a punctuation
14401 * one, back up one, to look at that final one just below. Later, we
14402 * will restore the parse pointer if appropriate */
14403 if (name_len && p == e && isPUNCT(*(p-1))) {
14408 if (p < e && isPUNCT(*p)) {
14410 has_terminating_bracket = TRUE;
14412 /* If this is a 2nd ']', and the first one is just below this
14413 * one, consider that to be the real terminator. This gives a
14414 * uniform and better positioning for the warning message */
14416 && possible_end != (char *) -1
14417 && *possible_end == ']'
14418 && name_len && input_text[name_len - 1] == ']')
14423 /* And this is actually equivalent to having done the 2nd
14424 * pass now, so set it to not try again */
14425 possible_end = (char *) -1;
14430 has_terminating_colon = TRUE;
14432 else if (*p == ';') {
14433 has_semi_colon = TRUE;
14434 has_terminating_colon = TRUE;
14442 /* Here, we have a class name to look up. We can short circuit the
14443 * stuff below for short names that can't possibly be meant to be a
14444 * class name. (We can do this on the first pass, as any second pass
14445 * will yield an even shorter name) */
14446 if (name_len < 3) {
14447 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14450 /* Find which class it is. Initially switch on the length of the name.
14452 switch (name_len) {
14454 if (memEQ(name_start, "word", 4)) {
14455 /* this is not POSIX, this is the Perl \w */
14456 class_number = ANYOF_WORDCHAR;
14460 /* Names all of length 5: alnum alpha ascii blank cntrl digit
14461 * graph lower print punct space upper
14462 * Offset 4 gives the best switch position. */
14463 switch (name_start[4]) {
14465 if (memEQ(name_start, "alph", 4)) /* alpha */
14466 class_number = ANYOF_ALPHA;
14469 if (memEQ(name_start, "spac", 4)) /* space */
14470 class_number = ANYOF_SPACE;
14473 if (memEQ(name_start, "grap", 4)) /* graph */
14474 class_number = ANYOF_GRAPH;
14477 if (memEQ(name_start, "asci", 4)) /* ascii */
14478 class_number = ANYOF_ASCII;
14481 if (memEQ(name_start, "blan", 4)) /* blank */
14482 class_number = ANYOF_BLANK;
14485 if (memEQ(name_start, "cntr", 4)) /* cntrl */
14486 class_number = ANYOF_CNTRL;
14489 if (memEQ(name_start, "alnu", 4)) /* alnum */
14490 class_number = ANYOF_ALPHANUMERIC;
14493 if (memEQ(name_start, "lowe", 4)) /* lower */
14494 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
14495 else if (memEQ(name_start, "uppe", 4)) /* upper */
14496 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
14499 if (memEQ(name_start, "digi", 4)) /* digit */
14500 class_number = ANYOF_DIGIT;
14501 else if (memEQ(name_start, "prin", 4)) /* print */
14502 class_number = ANYOF_PRINT;
14503 else if (memEQ(name_start, "punc", 4)) /* punct */
14504 class_number = ANYOF_PUNCT;
14509 if (memEQ(name_start, "xdigit", 6))
14510 class_number = ANYOF_XDIGIT;
14514 /* If the name exactly matches a posix class name the class number will
14515 * here be set to it, and the input almost certainly was meant to be a
14516 * posix class, so we can skip further checking. If instead the syntax
14517 * is exactly correct, but the name isn't one of the legal ones, we
14518 * will return that as an error below. But if neither of these apply,
14519 * it could be that no posix class was intended at all, or that one
14520 * was, but there was a typo. We tease these apart by doing fuzzy
14521 * matching on the name */
14522 if (class_number == OOB_NAMEDCLASS && found_problem) {
14523 const UV posix_names[][6] = {
14524 { 'a', 'l', 'n', 'u', 'm' },
14525 { 'a', 'l', 'p', 'h', 'a' },
14526 { 'a', 's', 'c', 'i', 'i' },
14527 { 'b', 'l', 'a', 'n', 'k' },
14528 { 'c', 'n', 't', 'r', 'l' },
14529 { 'd', 'i', 'g', 'i', 't' },
14530 { 'g', 'r', 'a', 'p', 'h' },
14531 { 'l', 'o', 'w', 'e', 'r' },
14532 { 'p', 'r', 'i', 'n', 't' },
14533 { 'p', 'u', 'n', 'c', 't' },
14534 { 's', 'p', 'a', 'c', 'e' },
14535 { 'u', 'p', 'p', 'e', 'r' },
14536 { 'w', 'o', 'r', 'd' },
14537 { 'x', 'd', 'i', 'g', 'i', 't' }
14539 /* The names of the above all have added NULs to make them the same
14540 * size, so we need to also have the real lengths */
14541 const UV posix_name_lengths[] = {
14542 sizeof("alnum") - 1,
14543 sizeof("alpha") - 1,
14544 sizeof("ascii") - 1,
14545 sizeof("blank") - 1,
14546 sizeof("cntrl") - 1,
14547 sizeof("digit") - 1,
14548 sizeof("graph") - 1,
14549 sizeof("lower") - 1,
14550 sizeof("print") - 1,
14551 sizeof("punct") - 1,
14552 sizeof("space") - 1,
14553 sizeof("upper") - 1,
14554 sizeof("word") - 1,
14555 sizeof("xdigit")- 1
14558 int temp_max = max_distance; /* Use a temporary, so if we
14559 reparse, we haven't changed the
14562 /* Use a smaller max edit distance if we are missing one of the
14564 if ( has_opening_bracket + has_opening_colon < 2
14565 || has_terminating_bracket + has_terminating_colon < 2)
14570 /* See if the input name is close to a legal one */
14571 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
14573 /* Short circuit call if the lengths are too far apart to be
14575 if (abs( (int) (name_len - posix_name_lengths[i]))
14581 if (edit_distance(input_text,
14584 posix_name_lengths[i],
14588 { /* If it is close, it probably was intended to be a class */
14589 goto probably_meant_to_be;
14593 /* Here the input name is not close enough to a valid class name
14594 * for us to consider it to be intended to be a posix class. If
14595 * we haven't already done so, and the parse found a character that
14596 * could have been terminators for the name, but which we absorbed
14597 * as typos during the first pass, repeat the parse, signalling it
14598 * to stop at that character */
14599 if (possible_end && possible_end != (char *) -1) {
14600 possible_end = (char *) -1;
14605 /* Here neither pass found a close-enough class name */
14606 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14609 probably_meant_to_be:
14611 /* Here we think that a posix specification was intended. Update any
14613 if (updated_parse_ptr) {
14614 *updated_parse_ptr = (char *) p;
14617 /* If a posix class name was intended but incorrectly specified, we
14618 * output or return the warnings */
14619 if (found_problem) {
14621 /* We set flags for these issues in the parse loop above instead of
14622 * adding them to the list of warnings, because we can parse it
14623 * twice, and we only want one warning instance */
14625 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
14628 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14630 if (has_semi_colon) {
14631 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14633 else if (! has_terminating_colon) {
14634 ADD_POSIX_WARNING(p, "there is no terminating ':'");
14636 if (! has_terminating_bracket) {
14637 ADD_POSIX_WARNING(p, "there is no terminating ']'");
14640 if (posix_warnings && RExC_warn_text && av_top_index(RExC_warn_text) > -1) {
14641 *posix_warnings = RExC_warn_text;
14644 else if (class_number != OOB_NAMEDCLASS) {
14645 /* If it is a known class, return the class. The class number
14646 * #defines are structured so each complement is +1 to the normal
14648 return class_number + complement;
14650 else if (! check_only) {
14652 /* Here, it is an unrecognized class. This is an error (unless the
14653 * call is to check only, which we've already handled above) */
14654 const char * const complement_string = (complement)
14657 RExC_parse = (char *) p;
14658 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
14660 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
14664 return OOB_NAMEDCLASS;
14666 #undef ADD_POSIX_WARNING
14668 STATIC unsigned int
14669 S_regex_set_precedence(const U8 my_operator) {
14671 /* Returns the precedence in the (?[...]) construct of the input operator,
14672 * specified by its character representation. The precedence follows
14673 * general Perl rules, but it extends this so that ')' and ']' have (low)
14674 * precedence even though they aren't really operators */
14676 switch (my_operator) {
14692 NOT_REACHED; /* NOTREACHED */
14693 return 0; /* Silence compiler warning */
14697 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
14698 I32 *flagp, U32 depth,
14699 char * const oregcomp_parse)
14701 /* Handle the (?[...]) construct to do set operations */
14703 U8 curchar; /* Current character being parsed */
14704 UV start, end; /* End points of code point ranges */
14705 SV* final = NULL; /* The end result inversion list */
14706 SV* result_string; /* 'final' stringified */
14707 AV* stack; /* stack of operators and operands not yet
14709 AV* fence_stack = NULL; /* A stack containing the positions in
14710 'stack' of where the undealt-with left
14711 parens would be if they were actually
14713 /* The 'VOL' (expanding to 'volatile') is a workaround for an optimiser bug
14714 * in Solaris Studio 12.3. See RT #127455 */
14715 VOL IV fence = 0; /* Position of where most recent undealt-
14716 with left paren in stack is; -1 if none.
14718 STRLEN len; /* Temporary */
14719 regnode* node; /* Temporary, and final regnode returned by
14721 const bool save_fold = FOLD; /* Temporary */
14722 char *save_end, *save_parse; /* Temporaries */
14723 const bool in_locale = LOC; /* we turn off /l during processing */
14724 AV* posix_warnings = NULL;
14726 GET_RE_DEBUG_FLAGS_DECL;
14728 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
14731 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
14734 REQUIRE_UNI_RULES(flagp, NULL); /* The use of this operator implies /u.
14735 This is required so that the compile
14736 time values are valid in all runtime
14739 /* This will return only an ANYOF regnode, or (unlikely) something smaller
14740 * (such as EXACT). Thus we can skip most everything if just sizing. We
14741 * call regclass to handle '[]' so as to not have to reinvent its parsing
14742 * rules here (throwing away the size it computes each time). And, we exit
14743 * upon an unescaped ']' that isn't one ending a regclass. To do both
14744 * these things, we need to realize that something preceded by a backslash
14745 * is escaped, so we have to keep track of backslashes */
14747 UV depth = 0; /* how many nested (?[...]) constructs */
14749 while (RExC_parse < RExC_end) {
14750 SV* current = NULL;
14752 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14753 TRUE /* Force /x */ );
14755 switch (*RExC_parse) {
14757 if (RExC_parse[1] == '[') depth++, RExC_parse++;
14762 /* Skip past this, so the next character gets skipped, after
14765 if (*RExC_parse == 'c') {
14766 /* Skip the \cX notation for control characters */
14767 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14773 /* See if this is a [:posix:] class. */
14774 bool is_posix_class = (OOB_NAMEDCLASS
14775 < handle_possible_posix(pRExC_state,
14779 TRUE /* checking only */));
14780 /* If it is a posix class, leave the parse pointer at the
14781 * '[' to fool regclass() into thinking it is part of a
14782 * '[[:posix:]]'. */
14783 if (! is_posix_class) {
14787 /* regclass() can only return RESTART_PASS1 and NEED_UTF8
14788 * if multi-char folds are allowed. */
14789 if (!regclass(pRExC_state, flagp,depth+1,
14790 is_posix_class, /* parse the whole char
14791 class only if not a
14793 FALSE, /* don't allow multi-char folds */
14794 TRUE, /* silence non-portable warnings. */
14796 FALSE, /* Require return to be an ANYOF */
14800 FAIL2("panic: regclass returned NULL to handle_sets, "
14801 "flags=%#" UVxf, (UV) *flagp);
14803 /* function call leaves parse pointing to the ']', except
14804 * if we faked it */
14805 if (is_posix_class) {
14809 SvREFCNT_dec(current); /* In case it returned something */
14814 if (depth--) break;
14816 if (*RExC_parse == ')') {
14817 node = reganode(pRExC_state, ANYOF, 0);
14818 RExC_size += ANYOF_SKIP;
14819 nextchar(pRExC_state);
14820 Set_Node_Length(node,
14821 RExC_parse - oregcomp_parse + 1); /* MJD */
14823 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
14831 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14835 /* We output the messages even if warnings are off, because we'll fail
14836 * the very next thing, and these give a likely diagnosis for that */
14837 if (posix_warnings && av_tindex_nomg(posix_warnings) >= 0) {
14838 output_or_return_posix_warnings(pRExC_state, posix_warnings, NULL);
14841 FAIL("Syntax error in (?[...])");
14844 /* Pass 2 only after this. */
14845 Perl_ck_warner_d(aTHX_
14846 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
14847 "The regex_sets feature is experimental" REPORT_LOCATION,
14848 REPORT_LOCATION_ARGS(RExC_parse));
14850 /* Everything in this construct is a metacharacter. Operands begin with
14851 * either a '\' (for an escape sequence), or a '[' for a bracketed
14852 * character class. Any other character should be an operator, or
14853 * parenthesis for grouping. Both types of operands are handled by calling
14854 * regclass() to parse them. It is called with a parameter to indicate to
14855 * return the computed inversion list. The parsing here is implemented via
14856 * a stack. Each entry on the stack is a single character representing one
14857 * of the operators; or else a pointer to an operand inversion list. */
14859 #define IS_OPERATOR(a) SvIOK(a)
14860 #define IS_OPERAND(a) (! IS_OPERATOR(a))
14862 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
14863 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
14864 * with pronouncing it called it Reverse Polish instead, but now that YOU
14865 * know how to pronounce it you can use the correct term, thus giving due
14866 * credit to the person who invented it, and impressing your geek friends.
14867 * Wikipedia says that the pronounciation of "Ł" has been changing so that
14868 * it is now more like an English initial W (as in wonk) than an L.)
14870 * This means that, for example, 'a | b & c' is stored on the stack as
14878 * where the numbers in brackets give the stack [array] element number.
14879 * In this implementation, parentheses are not stored on the stack.
14880 * Instead a '(' creates a "fence" so that the part of the stack below the
14881 * fence is invisible except to the corresponding ')' (this allows us to
14882 * replace testing for parens, by using instead subtraction of the fence
14883 * position). As new operands are processed they are pushed onto the stack
14884 * (except as noted in the next paragraph). New operators of higher
14885 * precedence than the current final one are inserted on the stack before
14886 * the lhs operand (so that when the rhs is pushed next, everything will be
14887 * in the correct positions shown above. When an operator of equal or
14888 * lower precedence is encountered in parsing, all the stacked operations
14889 * of equal or higher precedence are evaluated, leaving the result as the
14890 * top entry on the stack. This makes higher precedence operations
14891 * evaluate before lower precedence ones, and causes operations of equal
14892 * precedence to left associate.
14894 * The only unary operator '!' is immediately pushed onto the stack when
14895 * encountered. When an operand is encountered, if the top of the stack is
14896 * a '!", the complement is immediately performed, and the '!' popped. The
14897 * resulting value is treated as a new operand, and the logic in the
14898 * previous paragraph is executed. Thus in the expression
14900 * the stack looks like
14906 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
14913 * A ')' is treated as an operator with lower precedence than all the
14914 * aforementioned ones, which causes all operations on the stack above the
14915 * corresponding '(' to be evaluated down to a single resultant operand.
14916 * Then the fence for the '(' is removed, and the operand goes through the
14917 * algorithm above, without the fence.
14919 * A separate stack is kept of the fence positions, so that the position of
14920 * the latest so-far unbalanced '(' is at the top of it.
14922 * The ']' ending the construct is treated as the lowest operator of all,
14923 * so that everything gets evaluated down to a single operand, which is the
14926 sv_2mortal((SV *)(stack = newAV()));
14927 sv_2mortal((SV *)(fence_stack = newAV()));
14929 while (RExC_parse < RExC_end) {
14930 I32 top_index; /* Index of top-most element in 'stack' */
14931 SV** top_ptr; /* Pointer to top 'stack' element */
14932 SV* current = NULL; /* To contain the current inversion list
14934 SV* only_to_avoid_leaks;
14936 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14937 TRUE /* Force /x */ );
14938 if (RExC_parse >= RExC_end) {
14939 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
14942 curchar = UCHARAT(RExC_parse);
14946 #ifdef ENABLE_REGEX_SETS_DEBUGGING
14947 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
14948 DEBUG_U(dump_regex_sets_structures(pRExC_state,
14949 stack, fence, fence_stack));
14952 top_index = av_tindex_nomg(stack);
14955 SV** stacked_ptr; /* Ptr to something already on 'stack' */
14956 char stacked_operator; /* The topmost operator on the 'stack'. */
14957 SV* lhs; /* Operand to the left of the operator */
14958 SV* rhs; /* Operand to the right of the operator */
14959 SV* fence_ptr; /* Pointer to top element of the fence
14964 if ( RExC_parse < RExC_end - 1
14965 && (UCHARAT(RExC_parse + 1) == '?'))
14967 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
14968 * This happens when we have some thing like
14970 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
14972 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
14974 * Here we would be handling the interpolated
14975 * '$thai_or_lao'. We handle this by a recursive call to
14976 * ourselves which returns the inversion list the
14977 * interpolated expression evaluates to. We use the flags
14978 * from the interpolated pattern. */
14979 U32 save_flags = RExC_flags;
14980 const char * save_parse;
14982 RExC_parse += 2; /* Skip past the '(?' */
14983 save_parse = RExC_parse;
14985 /* Parse any flags for the '(?' */
14986 parse_lparen_question_flags(pRExC_state);
14988 if (RExC_parse == save_parse /* Makes sure there was at
14989 least one flag (or else
14990 this embedding wasn't
14992 || RExC_parse >= RExC_end - 4
14993 || UCHARAT(RExC_parse) != ':'
14994 || UCHARAT(++RExC_parse) != '('
14995 || UCHARAT(++RExC_parse) != '?'
14996 || UCHARAT(++RExC_parse) != '[')
14999 /* In combination with the above, this moves the
15000 * pointer to the point just after the first erroneous
15001 * character (or if there are no flags, to where they
15002 * should have been) */
15003 if (RExC_parse >= RExC_end - 4) {
15004 RExC_parse = RExC_end;
15006 else if (RExC_parse != save_parse) {
15007 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15009 vFAIL("Expecting '(?flags:(?[...'");
15012 /* Recurse, with the meat of the embedded expression */
15014 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
15015 depth+1, oregcomp_parse);
15017 /* Here, 'current' contains the embedded expression's
15018 * inversion list, and RExC_parse points to the trailing
15019 * ']'; the next character should be the ')' */
15021 assert(UCHARAT(RExC_parse) == ')');
15023 /* Then the ')' matching the original '(' handled by this
15024 * case: statement */
15026 assert(UCHARAT(RExC_parse) == ')');
15029 RExC_flags = save_flags;
15030 goto handle_operand;
15033 /* A regular '('. Look behind for illegal syntax */
15034 if (top_index - fence >= 0) {
15035 /* If the top entry on the stack is an operator, it had
15036 * better be a '!', otherwise the entry below the top
15037 * operand should be an operator */
15038 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
15039 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
15040 || ( IS_OPERAND(*top_ptr)
15041 && ( top_index - fence < 1
15042 || ! (stacked_ptr = av_fetch(stack,
15045 || ! IS_OPERATOR(*stacked_ptr))))
15048 vFAIL("Unexpected '(' with no preceding operator");
15052 /* Stack the position of this undealt-with left paren */
15053 av_push(fence_stack, newSViv(fence));
15054 fence = top_index + 1;
15058 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15059 * multi-char folds are allowed. */
15060 if (!regclass(pRExC_state, flagp,depth+1,
15061 TRUE, /* means parse just the next thing */
15062 FALSE, /* don't allow multi-char folds */
15063 FALSE, /* don't silence non-portable warnings. */
15065 FALSE, /* Require return to be an ANYOF */
15069 FAIL2("panic: regclass returned NULL to handle_sets, "
15070 "flags=%#" UVxf, (UV) *flagp);
15073 /* regclass() will return with parsing just the \ sequence,
15074 * leaving the parse pointer at the next thing to parse */
15076 goto handle_operand;
15078 case '[': /* Is a bracketed character class */
15080 /* See if this is a [:posix:] class. */
15081 bool is_posix_class = (OOB_NAMEDCLASS
15082 < handle_possible_posix(pRExC_state,
15086 TRUE /* checking only */));
15087 /* If it is a posix class, leave the parse pointer at the '['
15088 * to fool regclass() into thinking it is part of a
15089 * '[[:posix:]]'. */
15090 if (! is_posix_class) {
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 is_posix_class, /* parse the whole char
15098 class only if not a
15100 FALSE, /* don't allow multi-char folds */
15101 TRUE, /* silence non-portable warnings. */
15103 FALSE, /* Require return to be an ANYOF */
15108 FAIL2("panic: regclass returned NULL to handle_sets, "
15109 "flags=%#" UVxf, (UV) *flagp);
15112 /* function call leaves parse pointing to the ']', except if we
15114 if (is_posix_class) {
15118 goto handle_operand;
15122 if (top_index >= 1) {
15123 goto join_operators;
15126 /* Only a single operand on the stack: are done */
15130 if (av_tindex_nomg(fence_stack) < 0) {
15132 vFAIL("Unexpected ')'");
15135 /* If nothing after the fence, is missing an operand */
15136 if (top_index - fence < 0) {
15140 /* If at least two things on the stack, treat this as an
15142 if (top_index - fence >= 1) {
15143 goto join_operators;
15146 /* Here only a single thing on the fenced stack, and there is a
15147 * fence. Get rid of it */
15148 fence_ptr = av_pop(fence_stack);
15150 fence = SvIV(fence_ptr) - 1;
15151 SvREFCNT_dec_NN(fence_ptr);
15158 /* Having gotten rid of the fence, we pop the operand at the
15159 * stack top and process it as a newly encountered operand */
15160 current = av_pop(stack);
15161 if (IS_OPERAND(current)) {
15162 goto handle_operand;
15174 /* These binary operators should have a left operand already
15176 if ( top_index - fence < 0
15177 || top_index - fence == 1
15178 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
15179 || ! IS_OPERAND(*top_ptr))
15181 goto unexpected_binary;
15184 /* If only the one operand is on the part of the stack visible
15185 * to us, we just place this operator in the proper position */
15186 if (top_index - fence < 2) {
15188 /* Place the operator before the operand */
15190 SV* lhs = av_pop(stack);
15191 av_push(stack, newSVuv(curchar));
15192 av_push(stack, lhs);
15196 /* But if there is something else on the stack, we need to
15197 * process it before this new operator if and only if the
15198 * stacked operation has equal or higher precedence than the
15203 /* The operator on the stack is supposed to be below both its
15205 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
15206 || IS_OPERAND(*stacked_ptr))
15208 /* But if not, it's legal and indicates we are completely
15209 * done if and only if we're currently processing a ']',
15210 * which should be the final thing in the expression */
15211 if (curchar == ']') {
15217 vFAIL2("Unexpected binary operator '%c' with no "
15218 "preceding operand", curchar);
15220 stacked_operator = (char) SvUV(*stacked_ptr);
15222 if (regex_set_precedence(curchar)
15223 > regex_set_precedence(stacked_operator))
15225 /* Here, the new operator has higher precedence than the
15226 * stacked one. This means we need to add the new one to
15227 * the stack to await its rhs operand (and maybe more
15228 * stuff). We put it before the lhs operand, leaving
15229 * untouched the stacked operator and everything below it
15231 lhs = av_pop(stack);
15232 assert(IS_OPERAND(lhs));
15234 av_push(stack, newSVuv(curchar));
15235 av_push(stack, lhs);
15239 /* Here, the new operator has equal or lower precedence than
15240 * what's already there. This means the operation already
15241 * there should be performed now, before the new one. */
15243 rhs = av_pop(stack);
15244 if (! IS_OPERAND(rhs)) {
15246 /* This can happen when a ! is not followed by an operand,
15247 * like in /(?[\t &!])/ */
15251 lhs = av_pop(stack);
15253 if (! IS_OPERAND(lhs)) {
15255 /* This can happen when there is an empty (), like in
15256 * /(?[[0]+()+])/ */
15260 switch (stacked_operator) {
15262 _invlist_intersection(lhs, rhs, &rhs);
15267 _invlist_union(lhs, rhs, &rhs);
15271 _invlist_subtract(lhs, rhs, &rhs);
15274 case '^': /* The union minus the intersection */
15279 _invlist_union(lhs, rhs, &u);
15280 _invlist_intersection(lhs, rhs, &i);
15281 _invlist_subtract(u, i, &rhs);
15282 SvREFCNT_dec_NN(i);
15283 SvREFCNT_dec_NN(u);
15289 /* Here, the higher precedence operation has been done, and the
15290 * result is in 'rhs'. We overwrite the stacked operator with
15291 * the result. Then we redo this code to either push the new
15292 * operator onto the stack or perform any higher precedence
15293 * stacked operation */
15294 only_to_avoid_leaks = av_pop(stack);
15295 SvREFCNT_dec(only_to_avoid_leaks);
15296 av_push(stack, rhs);
15299 case '!': /* Highest priority, right associative */
15301 /* If what's already at the top of the stack is another '!",
15302 * they just cancel each other out */
15303 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
15304 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
15306 only_to_avoid_leaks = av_pop(stack);
15307 SvREFCNT_dec(only_to_avoid_leaks);
15309 else { /* Otherwise, since it's right associative, just push
15311 av_push(stack, newSVuv(curchar));
15316 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15317 vFAIL("Unexpected character");
15321 /* Here 'current' is the operand. If something is already on the
15322 * stack, we have to check if it is a !. But first, the code above
15323 * may have altered the stack in the time since we earlier set
15326 top_index = av_tindex_nomg(stack);
15327 if (top_index - fence >= 0) {
15328 /* If the top entry on the stack is an operator, it had better
15329 * be a '!', otherwise the entry below the top operand should
15330 * be an operator */
15331 top_ptr = av_fetch(stack, top_index, FALSE);
15333 if (IS_OPERATOR(*top_ptr)) {
15335 /* The only permissible operator at the top of the stack is
15336 * '!', which is applied immediately to this operand. */
15337 curchar = (char) SvUV(*top_ptr);
15338 if (curchar != '!') {
15339 SvREFCNT_dec(current);
15340 vFAIL2("Unexpected binary operator '%c' with no "
15341 "preceding operand", curchar);
15344 _invlist_invert(current);
15346 only_to_avoid_leaks = av_pop(stack);
15347 SvREFCNT_dec(only_to_avoid_leaks);
15349 /* And we redo with the inverted operand. This allows
15350 * handling multiple ! in a row */
15351 goto handle_operand;
15353 /* Single operand is ok only for the non-binary ')'
15355 else if ((top_index - fence == 0 && curchar != ')')
15356 || (top_index - fence > 0
15357 && (! (stacked_ptr = av_fetch(stack,
15360 || IS_OPERAND(*stacked_ptr))))
15362 SvREFCNT_dec(current);
15363 vFAIL("Operand with no preceding operator");
15367 /* Here there was nothing on the stack or the top element was
15368 * another operand. Just add this new one */
15369 av_push(stack, current);
15371 } /* End of switch on next parse token */
15373 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15374 } /* End of loop parsing through the construct */
15377 if (av_tindex_nomg(fence_stack) >= 0) {
15378 vFAIL("Unmatched (");
15381 if (av_tindex_nomg(stack) < 0 /* Was empty */
15382 || ((final = av_pop(stack)) == NULL)
15383 || ! IS_OPERAND(final)
15384 || SvTYPE(final) != SVt_INVLIST
15385 || av_tindex_nomg(stack) >= 0) /* More left on stack */
15388 SvREFCNT_dec(final);
15389 vFAIL("Incomplete expression within '(?[ ])'");
15392 /* Here, 'final' is the resultant inversion list from evaluating the
15393 * expression. Return it if so requested */
15394 if (return_invlist) {
15395 *return_invlist = final;
15399 /* Otherwise generate a resultant node, based on 'final'. regclass() is
15400 * expecting a string of ranges and individual code points */
15401 invlist_iterinit(final);
15402 result_string = newSVpvs("");
15403 while (invlist_iternext(final, &start, &end)) {
15404 if (start == end) {
15405 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
15408 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%" UVXf "}",
15413 /* About to generate an ANYOF (or similar) node from the inversion list we
15414 * have calculated */
15415 save_parse = RExC_parse;
15416 RExC_parse = SvPV(result_string, len);
15417 save_end = RExC_end;
15418 RExC_end = RExC_parse + len;
15420 /* We turn off folding around the call, as the class we have constructed
15421 * already has all folding taken into consideration, and we don't want
15422 * regclass() to add to that */
15423 RExC_flags &= ~RXf_PMf_FOLD;
15424 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if multi-char
15425 * folds are allowed. */
15426 node = regclass(pRExC_state, flagp,depth+1,
15427 FALSE, /* means parse the whole char class */
15428 FALSE, /* don't allow multi-char folds */
15429 TRUE, /* silence non-portable warnings. The above may very
15430 well have generated non-portable code points, but
15431 they're valid on this machine */
15432 FALSE, /* similarly, no need for strict */
15433 FALSE, /* Require return to be an ANYOF */
15438 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#" UVxf,
15441 /* Fix up the node type if we are in locale. (We have pretended we are
15442 * under /u for the purposes of regclass(), as this construct will only
15443 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
15444 * as to cause any warnings about bad locales to be output in regexec.c),
15445 * and add the flag that indicates to check if not in a UTF-8 locale. The
15446 * reason we above forbid optimization into something other than an ANYOF
15447 * node is simply to minimize the number of code changes in regexec.c.
15448 * Otherwise we would have to create new EXACTish node types and deal with
15449 * them. This decision could be revisited should this construct become
15452 * (One might think we could look at the resulting ANYOF node and suppress
15453 * the flag if everything is above 255, as those would be UTF-8 only,
15454 * but this isn't true, as the components that led to that result could
15455 * have been locale-affected, and just happen to cancel each other out
15456 * under UTF-8 locales.) */
15458 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
15460 assert(OP(node) == ANYOF);
15464 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
15468 RExC_flags |= RXf_PMf_FOLD;
15471 RExC_parse = save_parse + 1;
15472 RExC_end = save_end;
15473 SvREFCNT_dec_NN(final);
15474 SvREFCNT_dec_NN(result_string);
15476 nextchar(pRExC_state);
15477 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
15481 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15484 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
15485 AV * stack, const IV fence, AV * fence_stack)
15486 { /* Dumps the stacks in handle_regex_sets() */
15488 const SSize_t stack_top = av_tindex_nomg(stack);
15489 const SSize_t fence_stack_top = av_tindex_nomg(fence_stack);
15492 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
15494 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
15496 if (stack_top < 0) {
15497 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
15500 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
15501 for (i = stack_top; i >= 0; i--) {
15502 SV ** element_ptr = av_fetch(stack, i, FALSE);
15503 if (! element_ptr) {
15506 if (IS_OPERATOR(*element_ptr)) {
15507 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
15508 (int) i, (int) SvIV(*element_ptr));
15511 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
15512 sv_dump(*element_ptr);
15517 if (fence_stack_top < 0) {
15518 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
15521 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
15522 for (i = fence_stack_top; i >= 0; i--) {
15523 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
15524 if (! element_ptr) {
15527 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
15528 (int) i, (int) SvIV(*element_ptr));
15539 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
15541 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
15542 * innocent-looking character class, like /[ks]/i won't have to go out to
15543 * disk to find the possible matches.
15545 * This should be called only for a Latin1-range code points, cp, which is
15546 * known to be involved in a simple fold with other code points above
15547 * Latin1. It would give false results if /aa has been specified.
15548 * Multi-char folds are outside the scope of this, and must be handled
15551 * XXX It would be better to generate these via regen, in case a new
15552 * version of the Unicode standard adds new mappings, though that is not
15553 * really likely, and may be caught by the default: case of the switch
15556 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
15558 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
15564 add_cp_to_invlist(*invlist, KELVIN_SIGN);
15568 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
15571 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
15572 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
15574 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
15575 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
15576 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
15578 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
15579 *invlist = add_cp_to_invlist(*invlist,
15580 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
15583 #ifdef LATIN_CAPITAL_LETTER_SHARP_S /* not defined in early Unicode releases */
15585 case LATIN_SMALL_LETTER_SHARP_S:
15586 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
15591 #if UNICODE_MAJOR_VERSION < 3 \
15592 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0)
15594 /* In 3.0 and earlier, U+0130 folded simply to 'i'; and in 3.0.1 so did
15599 add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
15600 # if UNICODE_DOT_DOT_VERSION == 1
15601 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_DOTLESS_I);
15607 /* Use deprecated warning to increase the chances of this being
15610 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
15617 S_output_or_return_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings, AV** return_posix_warnings)
15619 /* If the final parameter is NULL, output the elements of the array given
15620 * by '*posix_warnings' as REGEXP warnings. Otherwise, the elements are
15621 * pushed onto it, (creating if necessary) */
15624 const bool first_is_fatal = ! return_posix_warnings
15625 && ckDEAD(packWARN(WARN_REGEXP));
15627 PERL_ARGS_ASSERT_OUTPUT_OR_RETURN_POSIX_WARNINGS;
15629 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
15630 if (return_posix_warnings) {
15631 if (! *return_posix_warnings) { /* mortalize to not leak if
15632 warnings are fatal */
15633 *return_posix_warnings = (AV *) sv_2mortal((SV *) newAV());
15635 av_push(*return_posix_warnings, msg);
15638 if (first_is_fatal) { /* Avoid leaking this */
15639 av_undef(posix_warnings); /* This isn't necessary if the
15640 array is mortal, but is a
15642 (void) sv_2mortal(msg);
15644 SAVEFREESV(RExC_rx_sv);
15647 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
15648 SvREFCNT_dec_NN(msg);
15654 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
15656 /* This adds the string scalar <multi_string> to the array
15657 * <multi_char_matches>. <multi_string> is known to have exactly
15658 * <cp_count> code points in it. This is used when constructing a
15659 * bracketed character class and we find something that needs to match more
15660 * than a single character.
15662 * <multi_char_matches> is actually an array of arrays. Each top-level
15663 * element is an array that contains all the strings known so far that are
15664 * the same length. And that length (in number of code points) is the same
15665 * as the index of the top-level array. Hence, the [2] element is an
15666 * array, each element thereof is a string containing TWO code points;
15667 * while element [3] is for strings of THREE characters, and so on. Since
15668 * this is for multi-char strings there can never be a [0] nor [1] element.
15670 * When we rewrite the character class below, we will do so such that the
15671 * longest strings are written first, so that it prefers the longest
15672 * matching strings first. This is done even if it turns out that any
15673 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
15674 * Christiansen has agreed that this is ok. This makes the test for the
15675 * ligature 'ffi' come before the test for 'ff', for example */
15678 AV** this_array_ptr;
15680 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
15682 if (! multi_char_matches) {
15683 multi_char_matches = newAV();
15686 if (av_exists(multi_char_matches, cp_count)) {
15687 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
15688 this_array = *this_array_ptr;
15691 this_array = newAV();
15692 av_store(multi_char_matches, cp_count,
15695 av_push(this_array, multi_string);
15697 return multi_char_matches;
15700 /* The names of properties whose definitions are not known at compile time are
15701 * stored in this SV, after a constant heading. So if the length has been
15702 * changed since initialization, then there is a run-time definition. */
15703 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
15704 (SvCUR(listsv) != initial_listsv_len)
15706 /* There is a restricted set of white space characters that are legal when
15707 * ignoring white space in a bracketed character class. This generates the
15708 * code to skip them.
15710 * There is a line below that uses the same white space criteria but is outside
15711 * this macro. Both here and there must use the same definition */
15712 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
15715 while (isBLANK_A(UCHARAT(p))) \
15723 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
15724 const bool stop_at_1, /* Just parse the next thing, don't
15725 look for a full character class */
15726 bool allow_multi_folds,
15727 const bool silence_non_portable, /* Don't output warnings
15731 bool optimizable, /* ? Allow a non-ANYOF return
15733 SV** ret_invlist, /* Return an inversion list, not a node */
15734 AV** return_posix_warnings
15737 /* parse a bracketed class specification. Most of these will produce an
15738 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
15739 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
15740 * under /i with multi-character folds: it will be rewritten following the
15741 * paradigm of this example, where the <multi-fold>s are characters which
15742 * fold to multiple character sequences:
15743 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
15744 * gets effectively rewritten as:
15745 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
15746 * reg() gets called (recursively) on the rewritten version, and this
15747 * function will return what it constructs. (Actually the <multi-fold>s
15748 * aren't physically removed from the [abcdefghi], it's just that they are
15749 * ignored in the recursion by means of a flag:
15750 * <RExC_in_multi_char_class>.)
15752 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
15753 * characters, with the corresponding bit set if that character is in the
15754 * list. For characters above this, a range list or swash is used. There
15755 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
15756 * determinable at compile time
15758 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs
15759 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded
15760 * to UTF-8. This can only happen if ret_invlist is non-NULL.
15763 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
15765 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
15768 int namedclass = OOB_NAMEDCLASS;
15769 char *rangebegin = NULL;
15770 bool need_class = 0;
15772 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
15773 than just initialized. */
15774 SV* properties = NULL; /* Code points that match \p{} \P{} */
15775 SV* posixes = NULL; /* Code points that match classes like [:word:],
15776 extended beyond the Latin1 range. These have to
15777 be kept separate from other code points for much
15778 of this function because their handling is
15779 different under /i, and for most classes under
15781 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
15782 separate for a while from the non-complemented
15783 versions because of complications with /d
15785 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
15786 treated more simply than the general case,
15787 leading to less compilation and execution
15789 UV element_count = 0; /* Number of distinct elements in the class.
15790 Optimizations may be possible if this is tiny */
15791 AV * multi_char_matches = NULL; /* Code points that fold to more than one
15792 character; used under /i */
15794 char * stop_ptr = RExC_end; /* where to stop parsing */
15795 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
15798 /* Unicode properties are stored in a swash; this holds the current one
15799 * being parsed. If this swash is the only above-latin1 component of the
15800 * character class, an optimization is to pass it directly on to the
15801 * execution engine. Otherwise, it is set to NULL to indicate that there
15802 * are other things in the class that have to be dealt with at execution
15804 SV* swash = NULL; /* Code points that match \p{} \P{} */
15806 /* Set if a component of this character class is user-defined; just passed
15807 * on to the engine */
15808 bool has_user_defined_property = FALSE;
15810 /* inversion list of code points this node matches only when the target
15811 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
15813 SV* has_upper_latin1_only_utf8_matches = NULL;
15815 /* Inversion list of code points this node matches regardless of things
15816 * like locale, folding, utf8ness of the target string */
15817 SV* cp_list = NULL;
15819 /* Like cp_list, but code points on this list need to be checked for things
15820 * that fold to/from them under /i */
15821 SV* cp_foldable_list = NULL;
15823 /* Like cp_list, but code points on this list are valid only when the
15824 * runtime locale is UTF-8 */
15825 SV* only_utf8_locale_list = NULL;
15827 /* In a range, if one of the endpoints is non-character-set portable,
15828 * meaning that it hard-codes a code point that may mean a different
15829 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
15830 * mnemonic '\t' which each mean the same character no matter which
15831 * character set the platform is on. */
15832 unsigned int non_portable_endpoint = 0;
15834 /* Is the range unicode? which means on a platform that isn't 1-1 native
15835 * to Unicode (i.e. non-ASCII), each code point in it should be considered
15836 * to be a Unicode value. */
15837 bool unicode_range = FALSE;
15838 bool invert = FALSE; /* Is this class to be complemented */
15840 bool warn_super = ALWAYS_WARN_SUPER;
15842 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
15843 case we need to change the emitted regop to an EXACT. */
15844 const char * orig_parse = RExC_parse;
15845 const SSize_t orig_size = RExC_size;
15846 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
15848 /* This variable is used to mark where the end in the input is of something
15849 * that looks like a POSIX construct but isn't. During the parse, when
15850 * something looks like it could be such a construct is encountered, it is
15851 * checked for being one, but not if we've already checked this area of the
15852 * input. Only after this position is reached do we check again */
15853 char *not_posix_region_end = RExC_parse - 1;
15855 AV* posix_warnings = NULL;
15856 const bool do_posix_warnings = return_posix_warnings
15857 || (PASS2 && ckWARN(WARN_REGEXP));
15859 GET_RE_DEBUG_FLAGS_DECL;
15861 PERL_ARGS_ASSERT_REGCLASS;
15863 PERL_UNUSED_ARG(depth);
15866 DEBUG_PARSE("clas");
15868 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
15869 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
15870 && UNICODE_DOT_DOT_VERSION == 0)
15871 allow_multi_folds = FALSE;
15874 /* Assume we are going to generate an ANYOF node. */
15875 ret = reganode(pRExC_state,
15882 RExC_size += ANYOF_SKIP;
15883 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
15886 ANYOF_FLAGS(ret) = 0;
15888 RExC_emit += ANYOF_SKIP;
15889 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
15890 initial_listsv_len = SvCUR(listsv);
15891 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
15894 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15896 assert(RExC_parse <= RExC_end);
15898 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
15901 allow_multi_folds = FALSE;
15903 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15906 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
15907 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
15908 int maybe_class = handle_possible_posix(pRExC_state,
15910 ¬_posix_region_end,
15912 TRUE /* checking only */);
15913 if (PASS2 && maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
15914 SAVEFREESV(RExC_rx_sv);
15915 ckWARN4reg(not_posix_region_end,
15916 "POSIX syntax [%c %c] belongs inside character classes%s",
15917 *RExC_parse, *RExC_parse,
15918 (maybe_class == OOB_NAMEDCLASS)
15919 ? ((POSIXCC_NOTYET(*RExC_parse))
15920 ? " (but this one isn't implemented)"
15921 : " (but this one isn't fully valid)")
15924 (void)ReREFCNT_inc(RExC_rx_sv);
15928 /* If the caller wants us to just parse a single element, accomplish this
15929 * by faking the loop ending condition */
15930 if (stop_at_1 && RExC_end > RExC_parse) {
15931 stop_ptr = RExC_parse + 1;
15934 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
15935 if (UCHARAT(RExC_parse) == ']')
15936 goto charclassloop;
15940 if ( posix_warnings
15941 && av_tindex_nomg(posix_warnings) >= 0
15942 && RExC_parse > not_posix_region_end)
15944 /* Warnings about posix class issues are considered tentative until
15945 * we are far enough along in the parse that we can no longer
15946 * change our mind, at which point we either output them or add
15947 * them, if it has so specified, to what gets returned to the
15948 * caller. This is done each time through the loop so that a later
15949 * class won't zap them before they have been dealt with. */
15950 output_or_return_posix_warnings(pRExC_state, posix_warnings,
15951 return_posix_warnings);
15954 if (RExC_parse >= stop_ptr) {
15958 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15960 if (UCHARAT(RExC_parse) == ']') {
15966 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
15967 save_value = value;
15968 save_prevvalue = prevvalue;
15971 rangebegin = RExC_parse;
15973 non_portable_endpoint = 0;
15975 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
15976 value = utf8n_to_uvchr((U8*)RExC_parse,
15977 RExC_end - RExC_parse,
15978 &numlen, UTF8_ALLOW_DEFAULT);
15979 RExC_parse += numlen;
15982 value = UCHARAT(RExC_parse++);
15984 if (value == '[') {
15985 char * posix_class_end;
15986 namedclass = handle_possible_posix(pRExC_state,
15989 do_posix_warnings ? &posix_warnings : NULL,
15990 FALSE /* die if error */);
15991 if (namedclass > OOB_NAMEDCLASS) {
15993 /* If there was an earlier attempt to parse this particular
15994 * posix class, and it failed, it was a false alarm, as this
15995 * successful one proves */
15996 if ( posix_warnings
15997 && av_tindex_nomg(posix_warnings) >= 0
15998 && not_posix_region_end >= RExC_parse
15999 && not_posix_region_end <= posix_class_end)
16001 av_undef(posix_warnings);
16004 RExC_parse = posix_class_end;
16006 else if (namedclass == OOB_NAMEDCLASS) {
16007 not_posix_region_end = posix_class_end;
16010 namedclass = OOB_NAMEDCLASS;
16013 else if ( RExC_parse - 1 > not_posix_region_end
16014 && MAYBE_POSIXCC(value))
16016 (void) handle_possible_posix(
16018 RExC_parse - 1, /* -1 because parse has already been
16020 ¬_posix_region_end,
16021 do_posix_warnings ? &posix_warnings : NULL,
16022 TRUE /* checking only */);
16024 else if (value == '\\') {
16025 /* Is a backslash; get the code point of the char after it */
16027 if (RExC_parse >= RExC_end) {
16028 vFAIL("Unmatched [");
16031 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
16032 value = utf8n_to_uvchr((U8*)RExC_parse,
16033 RExC_end - RExC_parse,
16034 &numlen, UTF8_ALLOW_DEFAULT);
16035 RExC_parse += numlen;
16038 value = UCHARAT(RExC_parse++);
16040 /* Some compilers cannot handle switching on 64-bit integer
16041 * values, therefore value cannot be an UV. Yes, this will
16042 * be a problem later if we want switch on Unicode.
16043 * A similar issue a little bit later when switching on
16044 * namedclass. --jhi */
16046 /* If the \ is escaping white space when white space is being
16047 * skipped, it means that that white space is wanted literally, and
16048 * is already in 'value'. Otherwise, need to translate the escape
16049 * into what it signifies. */
16050 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
16052 case 'w': namedclass = ANYOF_WORDCHAR; break;
16053 case 'W': namedclass = ANYOF_NWORDCHAR; break;
16054 case 's': namedclass = ANYOF_SPACE; break;
16055 case 'S': namedclass = ANYOF_NSPACE; break;
16056 case 'd': namedclass = ANYOF_DIGIT; break;
16057 case 'D': namedclass = ANYOF_NDIGIT; break;
16058 case 'v': namedclass = ANYOF_VERTWS; break;
16059 case 'V': namedclass = ANYOF_NVERTWS; break;
16060 case 'h': namedclass = ANYOF_HORIZWS; break;
16061 case 'H': namedclass = ANYOF_NHORIZWS; break;
16062 case 'N': /* Handle \N{NAME} in class */
16064 const char * const backslash_N_beg = RExC_parse - 2;
16067 if (! grok_bslash_N(pRExC_state,
16068 NULL, /* No regnode */
16069 &value, /* Yes single value */
16070 &cp_count, /* Multiple code pt count */
16076 if (*flagp & NEED_UTF8)
16077 FAIL("panic: grok_bslash_N set NEED_UTF8");
16078 if (*flagp & RESTART_PASS1)
16081 if (cp_count < 0) {
16082 vFAIL("\\N in a character class must be a named character: \\N{...}");
16084 else if (cp_count == 0) {
16086 ckWARNreg(RExC_parse,
16087 "Ignoring zero length \\N{} in character class");
16090 else { /* cp_count > 1 */
16091 if (! RExC_in_multi_char_class) {
16092 if (invert || range || *RExC_parse == '-') {
16095 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
16098 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
16100 break; /* <value> contains the first code
16101 point. Drop out of the switch to
16105 SV * multi_char_N = newSVpvn(backslash_N_beg,
16106 RExC_parse - backslash_N_beg);
16108 = add_multi_match(multi_char_matches,
16113 } /* End of cp_count != 1 */
16115 /* This element should not be processed further in this
16118 value = save_value;
16119 prevvalue = save_prevvalue;
16120 continue; /* Back to top of loop to get next char */
16123 /* Here, is a single code point, and <value> contains it */
16124 unicode_range = TRUE; /* \N{} are Unicode */
16132 /* We will handle any undefined properties ourselves */
16133 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
16134 /* And we actually would prefer to get
16135 * the straight inversion list of the
16136 * swash, since we will be accessing it
16137 * anyway, to save a little time */
16138 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
16140 if (RExC_parse >= RExC_end)
16141 vFAIL2("Empty \\%c", (U8)value);
16142 if (*RExC_parse == '{') {
16143 const U8 c = (U8)value;
16144 e = strchr(RExC_parse, '}');
16147 vFAIL2("Missing right brace on \\%c{}", c);
16151 while (isSPACE(*RExC_parse)) {
16155 if (UCHARAT(RExC_parse) == '^') {
16157 /* toggle. (The rhs xor gets the single bit that
16158 * differs between P and p; the other xor inverts just
16160 value ^= 'P' ^ 'p';
16163 while (isSPACE(*RExC_parse)) {
16168 if (e == RExC_parse)
16169 vFAIL2("Empty \\%c{}", c);
16171 n = e - RExC_parse;
16172 while (isSPACE(*(RExC_parse + n - 1)))
16174 } /* The \p isn't immediately followed by a '{' */
16175 else if (! isALPHA(*RExC_parse)) {
16176 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16177 vFAIL2("Character following \\%c must be '{' or a "
16178 "single-character Unicode property name",
16188 char* base_name; /* name after any packages are stripped */
16189 char* lookup_name = NULL;
16190 const char * const colon_colon = "::";
16192 /* Try to get the definition of the property into
16193 * <invlist>. If /i is in effect, the effective property
16194 * will have its name be <__NAME_i>. The design is
16195 * discussed in commit
16196 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
16197 name = savepv(Perl_form(aTHX_ "%.*s", (int)n, RExC_parse));
16200 lookup_name = savepv(Perl_form(aTHX_ "__%s_i", name));
16202 /* The function call just below that uses this can fail
16203 * to return, leaking memory if we don't do this */
16204 SAVEFREEPV(lookup_name);
16207 /* Look up the property name, and get its swash and
16208 * inversion list, if the property is found */
16209 SvREFCNT_dec(swash); /* Free any left-overs */
16210 swash = _core_swash_init("utf8",
16217 NULL, /* No inversion list */
16220 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
16221 HV* curpkg = (IN_PERL_COMPILETIME)
16223 : CopSTASH(PL_curcop);
16227 if (swash) { /* Got a swash but no inversion list.
16228 Something is likely wrong that will
16229 be sorted-out later */
16230 SvREFCNT_dec_NN(swash);
16234 /* Here didn't find it. It could be a an error (like a
16235 * typo) in specifying a Unicode property, or it could
16236 * be a user-defined property that will be available at
16237 * run-time. The names of these must begin with 'In'
16238 * or 'Is' (after any packages are stripped off). So
16239 * if not one of those, or if we accept only
16240 * compile-time properties, is an error; otherwise add
16241 * it to the list for run-time look up. */
16242 if ((base_name = rninstr(name, name + n,
16243 colon_colon, colon_colon + 2)))
16244 { /* Has ::. We know this must be a user-defined
16247 final_n -= base_name - name;
16256 || base_name[0] != 'I'
16257 || (base_name[1] != 's' && base_name[1] != 'n')
16260 const char * const msg
16262 ? "Illegal user-defined property name"
16263 : "Can't find Unicode property definition";
16264 RExC_parse = e + 1;
16266 /* diag_listed_as: Can't find Unicode property definition "%s" */
16267 vFAIL3utf8f("%s \"%" UTF8f "\"",
16268 msg, UTF8fARG(UTF, n, name));
16271 /* If the property name doesn't already have a package
16272 * name, add the current one to it so that it can be
16273 * referred to outside it. [perl #121777] */
16274 if (! has_pkg && curpkg) {
16275 char* pkgname = HvNAME(curpkg);
16276 if (strNE(pkgname, "main")) {
16277 char* full_name = Perl_form(aTHX_
16281 n = strlen(full_name);
16282 name = savepvn(full_name, n);
16286 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s%" UTF8f "%s\n",
16287 (value == 'p' ? '+' : '!'),
16288 (FOLD) ? "__" : "",
16289 UTF8fARG(UTF, n, name),
16290 (FOLD) ? "_i" : "");
16291 has_user_defined_property = TRUE;
16292 optimizable = FALSE; /* Will have to leave this an
16295 /* We don't know yet what this matches, so have to flag
16297 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
16301 /* Here, did get the swash and its inversion list. If
16302 * the swash is from a user-defined property, then this
16303 * whole character class should be regarded as such */
16304 if (swash_init_flags
16305 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
16307 has_user_defined_property = TRUE;
16310 /* We warn on matching an above-Unicode code point
16311 * if the match would return true, except don't
16312 * warn for \p{All}, which has exactly one element
16314 (_invlist_contains_cp(invlist, 0x110000)
16315 && (! (_invlist_len(invlist) == 1
16316 && *invlist_array(invlist) == 0)))
16322 /* Invert if asking for the complement */
16323 if (value == 'P') {
16324 _invlist_union_complement_2nd(properties,
16328 /* The swash can't be used as-is, because we've
16329 * inverted things; delay removing it to here after
16330 * have copied its invlist above */
16331 SvREFCNT_dec_NN(swash);
16335 _invlist_union(properties, invlist, &properties);
16339 RExC_parse = e + 1;
16340 namedclass = ANYOF_UNIPROP; /* no official name, but it's
16343 /* \p means they want Unicode semantics */
16344 REQUIRE_UNI_RULES(flagp, NULL);
16347 case 'n': value = '\n'; break;
16348 case 'r': value = '\r'; break;
16349 case 't': value = '\t'; break;
16350 case 'f': value = '\f'; break;
16351 case 'b': value = '\b'; break;
16352 case 'e': value = ESC_NATIVE; break;
16353 case 'a': value = '\a'; break;
16355 RExC_parse--; /* function expects to be pointed at the 'o' */
16357 const char* error_msg;
16358 bool valid = grok_bslash_o(&RExC_parse,
16361 PASS2, /* warnings only in
16364 silence_non_portable,
16370 non_portable_endpoint++;
16373 RExC_parse--; /* function expects to be pointed at the 'x' */
16375 const char* error_msg;
16376 bool valid = grok_bslash_x(&RExC_parse,
16379 PASS2, /* Output warnings */
16381 silence_non_portable,
16387 non_portable_endpoint++;
16390 value = grok_bslash_c(*RExC_parse++, PASS2);
16391 non_portable_endpoint++;
16393 case '0': case '1': case '2': case '3': case '4':
16394 case '5': case '6': case '7':
16396 /* Take 1-3 octal digits */
16397 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
16398 numlen = (strict) ? 4 : 3;
16399 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
16400 RExC_parse += numlen;
16403 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16404 vFAIL("Need exactly 3 octal digits");
16406 else if (! SIZE_ONLY /* like \08, \178 */
16408 && RExC_parse < RExC_end
16409 && isDIGIT(*RExC_parse)
16410 && ckWARN(WARN_REGEXP))
16412 SAVEFREESV(RExC_rx_sv);
16413 reg_warn_non_literal_string(
16415 form_short_octal_warning(RExC_parse, numlen));
16416 (void)ReREFCNT_inc(RExC_rx_sv);
16419 non_portable_endpoint++;
16423 /* Allow \_ to not give an error */
16424 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
16426 vFAIL2("Unrecognized escape \\%c in character class",
16430 SAVEFREESV(RExC_rx_sv);
16431 ckWARN2reg(RExC_parse,
16432 "Unrecognized escape \\%c in character class passed through",
16434 (void)ReREFCNT_inc(RExC_rx_sv);
16438 } /* End of switch on char following backslash */
16439 } /* end of handling backslash escape sequences */
16441 /* Here, we have the current token in 'value' */
16443 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
16446 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
16447 * literal, as is the character that began the false range, i.e.
16448 * the 'a' in the examples */
16451 const int w = (RExC_parse >= rangebegin)
16452 ? RExC_parse - rangebegin
16456 "False [] range \"%" UTF8f "\"",
16457 UTF8fARG(UTF, w, rangebegin));
16460 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
16461 ckWARN2reg(RExC_parse,
16462 "False [] range \"%" UTF8f "\"",
16463 UTF8fARG(UTF, w, rangebegin));
16464 (void)ReREFCNT_inc(RExC_rx_sv);
16465 cp_list = add_cp_to_invlist(cp_list, '-');
16466 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
16471 range = 0; /* this was not a true range */
16472 element_count += 2; /* So counts for three values */
16475 classnum = namedclass_to_classnum(namedclass);
16477 if (LOC && namedclass < ANYOF_POSIXL_MAX
16478 #ifndef HAS_ISASCII
16479 && classnum != _CC_ASCII
16482 /* What the Posix classes (like \w, [:space:]) match in locale
16483 * isn't knowable under locale until actual match time. Room
16484 * must be reserved (one time per outer bracketed class) to
16485 * store such classes. The space will contain a bit for each
16486 * named class that is to be matched against. This isn't
16487 * needed for \p{} and pseudo-classes, as they are not affected
16488 * by locale, and hence are dealt with separately */
16489 if (! need_class) {
16492 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16495 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16497 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
16498 ANYOF_POSIXL_ZERO(ret);
16500 /* We can't change this into some other type of node
16501 * (unless this is the only element, in which case there
16502 * are nodes that mean exactly this) as has runtime
16504 optimizable = FALSE;
16507 /* Coverity thinks it is possible for this to be negative; both
16508 * jhi and khw think it's not, but be safer */
16509 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16510 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
16512 /* See if it already matches the complement of this POSIX
16514 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16515 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
16519 posixl_matches_all = TRUE;
16520 break; /* No need to continue. Since it matches both
16521 e.g., \w and \W, it matches everything, and the
16522 bracketed class can be optimized into qr/./s */
16525 /* Add this class to those that should be checked at runtime */
16526 ANYOF_POSIXL_SET(ret, namedclass);
16528 /* The above-Latin1 characters are not subject to locale rules.
16529 * Just add them, in the second pass, to the
16530 * unconditionally-matched list */
16532 SV* scratch_list = NULL;
16534 /* Get the list of the above-Latin1 code points this
16536 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
16537 PL_XPosix_ptrs[classnum],
16539 /* Odd numbers are complements, like
16540 * NDIGIT, NASCII, ... */
16541 namedclass % 2 != 0,
16543 /* Checking if 'cp_list' is NULL first saves an extra
16544 * clone. Its reference count will be decremented at the
16545 * next union, etc, or if this is the only instance, at the
16546 * end of the routine */
16548 cp_list = scratch_list;
16551 _invlist_union(cp_list, scratch_list, &cp_list);
16552 SvREFCNT_dec_NN(scratch_list);
16554 continue; /* Go get next character */
16557 else if (! SIZE_ONLY) {
16559 /* Here, not in pass1 (in that pass we skip calculating the
16560 * contents of this class), and is not /l, or is a POSIX class
16561 * for which /l doesn't matter (or is a Unicode property, which
16562 * is skipped here). */
16563 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
16564 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
16566 /* Here, should be \h, \H, \v, or \V. None of /d, /i
16567 * nor /l make a difference in what these match,
16568 * therefore we just add what they match to cp_list. */
16569 if (classnum != _CC_VERTSPACE) {
16570 assert( namedclass == ANYOF_HORIZWS
16571 || namedclass == ANYOF_NHORIZWS);
16573 /* It turns out that \h is just a synonym for
16575 classnum = _CC_BLANK;
16578 _invlist_union_maybe_complement_2nd(
16580 PL_XPosix_ptrs[classnum],
16581 namedclass % 2 != 0, /* Complement if odd
16582 (NHORIZWS, NVERTWS)
16587 else if ( UNI_SEMANTICS
16588 || classnum == _CC_ASCII
16589 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
16590 || classnum == _CC_XDIGIT)))
16592 /* We usually have to worry about /d and /a affecting what
16593 * POSIX classes match, with special code needed for /d
16594 * because we won't know until runtime what all matches.
16595 * But there is no extra work needed under /u, and
16596 * [:ascii:] is unaffected by /a and /d; and :digit: and
16597 * :xdigit: don't have runtime differences under /d. So we
16598 * can special case these, and avoid some extra work below,
16599 * and at runtime. */
16600 _invlist_union_maybe_complement_2nd(
16602 PL_XPosix_ptrs[classnum],
16603 namedclass % 2 != 0,
16606 else { /* Garden variety class. If is NUPPER, NALPHA, ...
16607 complement and use nposixes */
16608 SV** posixes_ptr = namedclass % 2 == 0
16611 _invlist_union_maybe_complement_2nd(
16613 PL_XPosix_ptrs[classnum],
16614 namedclass % 2 != 0,
16618 } /* end of namedclass \blah */
16620 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16622 /* If 'range' is set, 'value' is the ending of a range--check its
16623 * validity. (If value isn't a single code point in the case of a
16624 * range, we should have figured that out above in the code that
16625 * catches false ranges). Later, we will handle each individual code
16626 * point in the range. If 'range' isn't set, this could be the
16627 * beginning of a range, so check for that by looking ahead to see if
16628 * the next real character to be processed is the range indicator--the
16633 /* For unicode ranges, we have to test that the Unicode as opposed
16634 * to the native values are not decreasing. (Above 255, there is
16635 * no difference between native and Unicode) */
16636 if (unicode_range && prevvalue < 255 && value < 255) {
16637 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
16638 goto backwards_range;
16643 if (prevvalue > value) /* b-a */ {
16648 w = RExC_parse - rangebegin;
16650 "Invalid [] range \"%" UTF8f "\"",
16651 UTF8fARG(UTF, w, rangebegin));
16652 NOT_REACHED; /* NOTREACHED */
16656 prevvalue = value; /* save the beginning of the potential range */
16657 if (! stop_at_1 /* Can't be a range if parsing just one thing */
16658 && *RExC_parse == '-')
16660 char* next_char_ptr = RExC_parse + 1;
16662 /* Get the next real char after the '-' */
16663 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
16665 /* If the '-' is at the end of the class (just before the ']',
16666 * it is a literal minus; otherwise it is a range */
16667 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
16668 RExC_parse = next_char_ptr;
16670 /* a bad range like \w-, [:word:]- ? */
16671 if (namedclass > OOB_NAMEDCLASS) {
16672 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
16673 const int w = RExC_parse >= rangebegin
16674 ? RExC_parse - rangebegin
16677 vFAIL4("False [] range \"%*.*s\"",
16682 "False [] range \"%*.*s\"",
16687 cp_list = add_cp_to_invlist(cp_list, '-');
16691 range = 1; /* yeah, it's a range! */
16692 continue; /* but do it the next time */
16697 if (namedclass > OOB_NAMEDCLASS) {
16701 /* Here, we have a single value this time through the loop, and
16702 * <prevvalue> is the beginning of the range, if any; or <value> if
16705 /* non-Latin1 code point implies unicode semantics. Must be set in
16706 * pass1 so is there for the whole of pass 2 */
16708 REQUIRE_UNI_RULES(flagp, NULL);
16711 /* Ready to process either the single value, or the completed range.
16712 * For single-valued non-inverted ranges, we consider the possibility
16713 * of multi-char folds. (We made a conscious decision to not do this
16714 * for the other cases because it can often lead to non-intuitive
16715 * results. For example, you have the peculiar case that:
16716 * "s s" =~ /^[^\xDF]+$/i => Y
16717 * "ss" =~ /^[^\xDF]+$/i => N
16719 * See [perl #89750] */
16720 if (FOLD && allow_multi_folds && value == prevvalue) {
16721 if (value == LATIN_SMALL_LETTER_SHARP_S
16722 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
16725 /* Here <value> is indeed a multi-char fold. Get what it is */
16727 U8 foldbuf[UTF8_MAXBYTES_CASE];
16730 UV folded = _to_uni_fold_flags(
16734 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
16735 ? FOLD_FLAGS_NOMIX_ASCII
16739 /* Here, <folded> should be the first character of the
16740 * multi-char fold of <value>, with <foldbuf> containing the
16741 * whole thing. But, if this fold is not allowed (because of
16742 * the flags), <fold> will be the same as <value>, and should
16743 * be processed like any other character, so skip the special
16745 if (folded != value) {
16747 /* Skip if we are recursed, currently parsing the class
16748 * again. Otherwise add this character to the list of
16749 * multi-char folds. */
16750 if (! RExC_in_multi_char_class) {
16751 STRLEN cp_count = utf8_length(foldbuf,
16752 foldbuf + foldlen);
16753 SV* multi_fold = sv_2mortal(newSVpvs(""));
16755 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
16758 = add_multi_match(multi_char_matches,
16764 /* This element should not be processed further in this
16767 value = save_value;
16768 prevvalue = save_prevvalue;
16774 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
16777 /* If the range starts above 255, everything is portable and
16778 * likely to be so for any forseeable character set, so don't
16780 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
16781 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
16783 else if (prevvalue != value) {
16785 /* Under strict, ranges that stop and/or end in an ASCII
16786 * printable should have each end point be a portable value
16787 * for it (preferably like 'A', but we don't warn if it is
16788 * a (portable) Unicode name or code point), and the range
16789 * must be be all digits or all letters of the same case.
16790 * Otherwise, the range is non-portable and unclear as to
16791 * what it contains */
16792 if ((isPRINT_A(prevvalue) || isPRINT_A(value))
16793 && (non_portable_endpoint
16794 || ! ((isDIGIT_A(prevvalue) && isDIGIT_A(value))
16795 || (isLOWER_A(prevvalue) && isLOWER_A(value))
16796 || (isUPPER_A(prevvalue) && isUPPER_A(value)))))
16798 vWARN(RExC_parse, "Ranges of ASCII printables should be some subset of \"0-9\", \"A-Z\", or \"a-z\"");
16800 else if (prevvalue >= 0x660) { /* ARABIC_INDIC_DIGIT_ZERO */
16802 /* But the nature of Unicode and languages mean we
16803 * can't do the same checks for above-ASCII ranges,
16804 * except in the case of digit ones. These should
16805 * contain only digits from the same group of 10. The
16806 * ASCII case is handled just above. 0x660 is the
16807 * first digit character beyond ASCII. Hence here, the
16808 * range could be a range of digits. Find out. */
16809 IV index_start = _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
16811 IV index_final = _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
16814 /* If the range start and final points are in the same
16815 * inversion list element, it means that either both
16816 * are not digits, or both are digits in a consecutive
16817 * sequence of digits. (So far, Unicode has kept all
16818 * such sequences as distinct groups of 10, but assert
16819 * to make sure). If the end points are not in the
16820 * same element, neither should be a digit. */
16821 if (index_start == index_final) {
16822 assert(! ELEMENT_RANGE_MATCHES_INVLIST(index_start)
16823 || (invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start+1]
16824 - invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
16826 /* But actually Unicode did have one group of 11
16827 * 'digits' in 5.2, so in case we are operating
16828 * on that version, let that pass */
16829 || (invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start+1]
16830 - invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
16832 && invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
16836 else if ((index_start >= 0
16837 && ELEMENT_RANGE_MATCHES_INVLIST(index_start))
16838 || (index_final >= 0
16839 && ELEMENT_RANGE_MATCHES_INVLIST(index_final)))
16841 vWARN(RExC_parse, "Ranges of digits should be from the same group of 10");
16846 if ((! range || prevvalue == value) && non_portable_endpoint) {
16847 if (isPRINT_A(value)) {
16850 if (isBACKSLASHED_PUNCT(value)) {
16851 literal[d++] = '\\';
16853 literal[d++] = (char) value;
16854 literal[d++] = '\0';
16857 "\"%.*s\" is more clearly written simply as \"%s\"",
16858 (int) (RExC_parse - rangebegin),
16863 else if isMNEMONIC_CNTRL(value) {
16865 "\"%.*s\" is more clearly written simply as \"%s\"",
16866 (int) (RExC_parse - rangebegin),
16868 cntrl_to_mnemonic((U8) value)
16874 /* Deal with this element of the class */
16878 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16881 /* On non-ASCII platforms, for ranges that span all of 0..255, and
16882 * ones that don't require special handling, we can just add the
16883 * range like we do for ASCII platforms */
16884 if ((UNLIKELY(prevvalue == 0) && value >= 255)
16885 || ! (prevvalue < 256
16887 || (! non_portable_endpoint
16888 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
16889 || (isUPPER_A(prevvalue)
16890 && isUPPER_A(value)))))))
16892 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16896 /* Here, requires special handling. This can be because it is
16897 * a range whose code points are considered to be Unicode, and
16898 * so must be individually translated into native, or because
16899 * its a subrange of 'A-Z' or 'a-z' which each aren't
16900 * contiguous in EBCDIC, but we have defined them to include
16901 * only the "expected" upper or lower case ASCII alphabetics.
16902 * Subranges above 255 are the same in native and Unicode, so
16903 * can be added as a range */
16904 U8 start = NATIVE_TO_LATIN1(prevvalue);
16906 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
16907 for (j = start; j <= end; j++) {
16908 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
16911 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16918 range = 0; /* this range (if it was one) is done now */
16919 } /* End of loop through all the text within the brackets */
16922 if ( posix_warnings && av_tindex_nomg(posix_warnings) >= 0) {
16923 output_or_return_posix_warnings(pRExC_state, posix_warnings,
16924 return_posix_warnings);
16927 /* If anything in the class expands to more than one character, we have to
16928 * deal with them by building up a substitute parse string, and recursively
16929 * calling reg() on it, instead of proceeding */
16930 if (multi_char_matches) {
16931 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
16934 char *save_end = RExC_end;
16935 char *save_parse = RExC_parse;
16936 char *save_start = RExC_start;
16937 STRLEN prefix_end = 0; /* We copy the character class after a
16938 prefix supplied here. This is the size
16939 + 1 of that prefix */
16940 bool first_time = TRUE; /* First multi-char occurrence doesn't get
16945 assert(RExC_precomp_adj == 0); /* Only one level of recursion allowed */
16947 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
16948 because too confusing */
16950 sv_catpv(substitute_parse, "(?:");
16954 /* Look at the longest folds first */
16955 for (cp_count = av_tindex_nomg(multi_char_matches);
16960 if (av_exists(multi_char_matches, cp_count)) {
16961 AV** this_array_ptr;
16964 this_array_ptr = (AV**) av_fetch(multi_char_matches,
16966 while ((this_sequence = av_pop(*this_array_ptr)) !=
16969 if (! first_time) {
16970 sv_catpv(substitute_parse, "|");
16972 first_time = FALSE;
16974 sv_catpv(substitute_parse, SvPVX(this_sequence));
16979 /* If the character class contains anything else besides these
16980 * multi-character folds, have to include it in recursive parsing */
16981 if (element_count) {
16982 sv_catpv(substitute_parse, "|[");
16983 prefix_end = SvCUR(substitute_parse);
16984 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
16986 /* Put in a closing ']' only if not going off the end, as otherwise
16987 * we are adding something that really isn't there */
16988 if (RExC_parse < RExC_end) {
16989 sv_catpv(substitute_parse, "]");
16993 sv_catpv(substitute_parse, ")");
16996 /* This is a way to get the parse to skip forward a whole named
16997 * sequence instead of matching the 2nd character when it fails the
16999 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
17003 /* Set up the data structure so that any errors will be properly
17004 * reported. See the comments at the definition of
17005 * REPORT_LOCATION_ARGS for details */
17006 RExC_precomp_adj = orig_parse - RExC_precomp;
17007 RExC_start = RExC_parse = SvPV(substitute_parse, len);
17008 RExC_adjusted_start = RExC_start + prefix_end;
17009 RExC_end = RExC_parse + len;
17010 RExC_in_multi_char_class = 1;
17011 RExC_override_recoding = 1;
17012 RExC_emit = (regnode *)orig_emit;
17014 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17016 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PASS1|NEED_UTF8);
17018 /* And restore so can parse the rest of the pattern */
17019 RExC_parse = save_parse;
17020 RExC_start = RExC_adjusted_start = save_start;
17021 RExC_precomp_adj = 0;
17022 RExC_end = save_end;
17023 RExC_in_multi_char_class = 0;
17024 RExC_override_recoding = 0;
17025 SvREFCNT_dec_NN(multi_char_matches);
17029 /* Here, we've gone through the entire class and dealt with multi-char
17030 * folds. We are now in a position that we can do some checks to see if we
17031 * can optimize this ANYOF node into a simpler one, even in Pass 1.
17032 * Currently we only do two checks:
17033 * 1) is in the unlikely event that the user has specified both, eg. \w and
17034 * \W under /l, then the class matches everything. (This optimization
17035 * is done only to make the optimizer code run later work.)
17036 * 2) if the character class contains only a single element (including a
17037 * single range), we see if there is an equivalent node for it.
17038 * Other checks are possible */
17040 && ! ret_invlist /* Can't optimize if returning the constructed
17042 && (UNLIKELY(posixl_matches_all) || element_count == 1))
17047 if (UNLIKELY(posixl_matches_all)) {
17050 else if (namedclass > OOB_NAMEDCLASS) { /* this is a single named
17051 class, like \w or [:digit:]
17054 /* All named classes are mapped into POSIXish nodes, with its FLAG
17055 * argument giving which class it is */
17056 switch ((I32)namedclass) {
17057 case ANYOF_UNIPROP:
17060 /* These don't depend on the charset modifiers. They always
17061 * match under /u rules */
17062 case ANYOF_NHORIZWS:
17063 case ANYOF_HORIZWS:
17064 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
17067 case ANYOF_NVERTWS:
17072 /* The actual POSIXish node for all the rest depends on the
17073 * charset modifier. The ones in the first set depend only on
17074 * ASCII or, if available on this platform, also locale */
17078 op = (LOC) ? POSIXL : POSIXA;
17084 /* The following don't have any matches in the upper Latin1
17085 * range, hence /d is equivalent to /u for them. Making it /u
17086 * saves some branches at runtime */
17090 case ANYOF_NXDIGIT:
17091 if (! DEPENDS_SEMANTICS) {
17092 goto treat_as_default;
17098 /* The following change to CASED under /i */
17104 namedclass = ANYOF_CASED + (namedclass % 2);
17108 /* The rest have more possibilities depending on the charset.
17109 * We take advantage of the enum ordering of the charset
17110 * modifiers to get the exact node type, */
17113 op = POSIXD + get_regex_charset(RExC_flags);
17114 if (op > POSIXA) { /* /aa is same as /a */
17119 /* The odd numbered ones are the complements of the
17120 * next-lower even number one */
17121 if (namedclass % 2 == 1) {
17125 arg = namedclass_to_classnum(namedclass);
17129 else if (value == prevvalue) {
17131 /* Here, the class consists of just a single code point */
17134 if (! LOC && value == '\n') {
17135 op = REG_ANY; /* Optimize [^\n] */
17136 *flagp |= HASWIDTH|SIMPLE;
17140 else if (value < 256 || UTF) {
17142 /* Optimize a single value into an EXACTish node, but not if it
17143 * would require converting the pattern to UTF-8. */
17144 op = compute_EXACTish(pRExC_state);
17146 } /* Otherwise is a range */
17147 else if (! LOC) { /* locale could vary these */
17148 if (prevvalue == '0') {
17149 if (value == '9') {
17154 else if (! FOLD || ASCII_FOLD_RESTRICTED) {
17155 /* We can optimize A-Z or a-z, but not if they could match
17156 * something like the KELVIN SIGN under /i. */
17157 if (prevvalue == 'A') {
17160 && ! non_portable_endpoint
17163 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
17167 else if (prevvalue == 'a') {
17170 && ! non_portable_endpoint
17173 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
17180 /* Here, we have changed <op> away from its initial value iff we found
17181 * an optimization */
17184 /* Throw away this ANYOF regnode, and emit the calculated one,
17185 * which should correspond to the beginning, not current, state of
17187 const char * cur_parse = RExC_parse;
17188 RExC_parse = (char *)orig_parse;
17192 /* To get locale nodes to not use the full ANYOF size would
17193 * require moving the code above that writes the portions
17194 * of it that aren't in other nodes to after this point.
17195 * e.g. ANYOF_POSIXL_SET */
17196 RExC_size = orig_size;
17200 RExC_emit = (regnode *)orig_emit;
17201 if (PL_regkind[op] == POSIXD) {
17202 if (op == POSIXL) {
17203 RExC_contains_locale = 1;
17206 op += NPOSIXD - POSIXD;
17211 ret = reg_node(pRExC_state, op);
17213 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17217 *flagp |= HASWIDTH|SIMPLE;
17219 else if (PL_regkind[op] == EXACT) {
17220 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17221 TRUE /* downgradable to EXACT */
17225 RExC_parse = (char *) cur_parse;
17227 SvREFCNT_dec(posixes);
17228 SvREFCNT_dec(nposixes);
17229 SvREFCNT_dec(simple_posixes);
17230 SvREFCNT_dec(cp_list);
17231 SvREFCNT_dec(cp_foldable_list);
17238 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
17240 /* If folding, we calculate all characters that could fold to or from the
17241 * ones already on the list */
17242 if (cp_foldable_list) {
17244 UV start, end; /* End points of code point ranges */
17246 SV* fold_intersection = NULL;
17249 /* Our calculated list will be for Unicode rules. For locale
17250 * matching, we have to keep a separate list that is consulted at
17251 * runtime only when the locale indicates Unicode rules. For
17252 * non-locale, we just use the general list */
17254 use_list = &only_utf8_locale_list;
17257 use_list = &cp_list;
17260 /* Only the characters in this class that participate in folds need
17261 * be checked. Get the intersection of this class and all the
17262 * possible characters that are foldable. This can quickly narrow
17263 * down a large class */
17264 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
17265 &fold_intersection);
17267 /* The folds for all the Latin1 characters are hard-coded into this
17268 * program, but we have to go out to disk to get the others. */
17269 if (invlist_highest(cp_foldable_list) >= 256) {
17271 /* This is a hash that for a particular fold gives all
17272 * characters that are involved in it */
17273 if (! PL_utf8_foldclosures) {
17274 _load_PL_utf8_foldclosures();
17278 /* Now look at the foldable characters in this class individually */
17279 invlist_iterinit(fold_intersection);
17280 while (invlist_iternext(fold_intersection, &start, &end)) {
17283 /* Look at every character in the range */
17284 for (j = start; j <= end; j++) {
17285 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17291 if (IS_IN_SOME_FOLD_L1(j)) {
17293 /* ASCII is always matched; non-ASCII is matched
17294 * only under Unicode rules (which could happen
17295 * under /l if the locale is a UTF-8 one */
17296 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
17297 *use_list = add_cp_to_invlist(*use_list,
17298 PL_fold_latin1[j]);
17301 has_upper_latin1_only_utf8_matches
17302 = add_cp_to_invlist(
17303 has_upper_latin1_only_utf8_matches,
17304 PL_fold_latin1[j]);
17308 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
17309 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
17311 add_above_Latin1_folds(pRExC_state,
17318 /* Here is an above Latin1 character. We don't have the
17319 * rules hard-coded for it. First, get its fold. This is
17320 * the simple fold, as the multi-character folds have been
17321 * handled earlier and separated out */
17322 _to_uni_fold_flags(j, foldbuf, &foldlen,
17323 (ASCII_FOLD_RESTRICTED)
17324 ? FOLD_FLAGS_NOMIX_ASCII
17327 /* Single character fold of above Latin1. Add everything in
17328 * its fold closure to the list that this node should match.
17329 * The fold closures data structure is a hash with the keys
17330 * being the UTF-8 of every character that is folded to, like
17331 * 'k', and the values each an array of all code points that
17332 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
17333 * Multi-character folds are not included */
17334 if ((listp = hv_fetch(PL_utf8_foldclosures,
17335 (char *) foldbuf, foldlen, FALSE)))
17337 AV* list = (AV*) *listp;
17339 for (k = 0; k <= av_tindex_nomg(list); k++) {
17340 SV** c_p = av_fetch(list, k, FALSE);
17346 /* /aa doesn't allow folds between ASCII and non- */
17347 if ((ASCII_FOLD_RESTRICTED
17348 && (isASCII(c) != isASCII(j))))
17353 /* Folds under /l which cross the 255/256 boundary
17354 * are added to a separate list. (These are valid
17355 * only when the locale is UTF-8.) */
17356 if (c < 256 && LOC) {
17357 *use_list = add_cp_to_invlist(*use_list, c);
17361 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
17363 cp_list = add_cp_to_invlist(cp_list, c);
17366 /* Similarly folds involving non-ascii Latin1
17367 * characters under /d are added to their list */
17368 has_upper_latin1_only_utf8_matches
17369 = add_cp_to_invlist(
17370 has_upper_latin1_only_utf8_matches,
17377 SvREFCNT_dec_NN(fold_intersection);
17380 /* Now that we have finished adding all the folds, there is no reason
17381 * to keep the foldable list separate */
17382 _invlist_union(cp_list, cp_foldable_list, &cp_list);
17383 SvREFCNT_dec_NN(cp_foldable_list);
17386 /* And combine the result (if any) with any inversion lists from posix
17387 * classes. The lists are kept separate up to now because we don't want to
17388 * fold the classes (folding of those is automatically handled by the swash
17389 * fetching code) */
17390 if (simple_posixes) { /* These are the classes known to be unaffected by
17393 _invlist_union(cp_list, simple_posixes, &cp_list);
17394 SvREFCNT_dec_NN(simple_posixes);
17397 cp_list = simple_posixes;
17400 if (posixes || nposixes) {
17402 /* We have to adjust /a and /aa */
17403 if (AT_LEAST_ASCII_RESTRICTED) {
17405 /* Under /a and /aa, nothing above ASCII matches these */
17407 _invlist_intersection(posixes,
17408 PL_XPosix_ptrs[_CC_ASCII],
17412 /* Under /a and /aa, everything above ASCII matches these
17415 _invlist_union_complement_2nd(nposixes,
17416 PL_XPosix_ptrs[_CC_ASCII],
17421 if (! DEPENDS_SEMANTICS) {
17423 /* For everything but /d, we can just add the current 'posixes' and
17424 * 'nposixes' to the main list */
17427 _invlist_union(cp_list, posixes, &cp_list);
17428 SvREFCNT_dec_NN(posixes);
17436 _invlist_union(cp_list, nposixes, &cp_list);
17437 SvREFCNT_dec_NN(nposixes);
17440 cp_list = nposixes;
17445 /* Under /d, things like \w match upper Latin1 characters only if
17446 * the target string is in UTF-8. But things like \W match all the
17447 * upper Latin1 characters if the target string is not in UTF-8.
17449 * Handle the case where there something like \W separately */
17451 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1);
17453 /* A complemented posix class matches all upper Latin1
17454 * characters if not in UTF-8. And it matches just certain
17455 * ones when in UTF-8. That means those certain ones are
17456 * matched regardless, so can just be added to the
17457 * unconditional list */
17459 _invlist_union(cp_list, nposixes, &cp_list);
17460 SvREFCNT_dec_NN(nposixes);
17464 cp_list = nposixes;
17467 /* Likewise for 'posixes' */
17468 _invlist_union(posixes, cp_list, &cp_list);
17470 /* Likewise for anything else in the range that matched only
17472 if (has_upper_latin1_only_utf8_matches) {
17473 _invlist_union(cp_list,
17474 has_upper_latin1_only_utf8_matches,
17476 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17477 has_upper_latin1_only_utf8_matches = NULL;
17480 /* If we don't match all the upper Latin1 characters regardless
17481 * of UTF-8ness, we have to set a flag to match the rest when
17483 _invlist_subtract(only_non_utf8_list, cp_list,
17484 &only_non_utf8_list);
17485 if (_invlist_len(only_non_utf8_list) != 0) {
17486 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17490 /* Here there were no complemented posix classes. That means
17491 * the upper Latin1 characters in 'posixes' match only when the
17492 * target string is in UTF-8. So we have to add them to the
17493 * list of those types of code points, while adding the
17494 * remainder to the unconditional list.
17496 * First calculate what they are */
17497 SV* nonascii_but_latin1_properties = NULL;
17498 _invlist_intersection(posixes, PL_UpperLatin1,
17499 &nonascii_but_latin1_properties);
17501 /* And add them to the final list of such characters. */
17502 _invlist_union(has_upper_latin1_only_utf8_matches,
17503 nonascii_but_latin1_properties,
17504 &has_upper_latin1_only_utf8_matches);
17506 /* Remove them from what now becomes the unconditional list */
17507 _invlist_subtract(posixes, nonascii_but_latin1_properties,
17510 /* And add those unconditional ones to the final list */
17512 _invlist_union(cp_list, posixes, &cp_list);
17513 SvREFCNT_dec_NN(posixes);
17520 SvREFCNT_dec(nonascii_but_latin1_properties);
17522 /* Get rid of any characters that we now know are matched
17523 * unconditionally from the conditional list, which may make
17524 * that list empty */
17525 _invlist_subtract(has_upper_latin1_only_utf8_matches,
17527 &has_upper_latin1_only_utf8_matches);
17528 if (_invlist_len(has_upper_latin1_only_utf8_matches) == 0) {
17529 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17530 has_upper_latin1_only_utf8_matches = NULL;
17536 /* And combine the result (if any) with any inversion list from properties.
17537 * The lists are kept separate up to now so that we can distinguish the two
17538 * in regards to matching above-Unicode. A run-time warning is generated
17539 * if a Unicode property is matched against a non-Unicode code point. But,
17540 * we allow user-defined properties to match anything, without any warning,
17541 * and we also suppress the warning if there is a portion of the character
17542 * class that isn't a Unicode property, and which matches above Unicode, \W
17543 * or [\x{110000}] for example.
17544 * (Note that in this case, unlike the Posix one above, there is no
17545 * <has_upper_latin1_only_utf8_matches>, because having a Unicode property
17546 * forces Unicode semantics */
17550 /* If it matters to the final outcome, see if a non-property
17551 * component of the class matches above Unicode. If so, the
17552 * warning gets suppressed. This is true even if just a single
17553 * such code point is specified, as, though not strictly correct if
17554 * another such code point is matched against, the fact that they
17555 * are using above-Unicode code points indicates they should know
17556 * the issues involved */
17558 warn_super = ! (invert
17559 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
17562 _invlist_union(properties, cp_list, &cp_list);
17563 SvREFCNT_dec_NN(properties);
17566 cp_list = properties;
17571 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17573 /* Because an ANYOF node is the only one that warns, this node
17574 * can't be optimized into something else */
17575 optimizable = FALSE;
17579 /* Here, we have calculated what code points should be in the character
17582 * Now we can see about various optimizations. Fold calculation (which we
17583 * did above) needs to take place before inversion. Otherwise /[^k]/i
17584 * would invert to include K, which under /i would match k, which it
17585 * shouldn't. Therefore we can't invert folded locale now, as it won't be
17586 * folded until runtime */
17588 /* If we didn't do folding, it's because some information isn't available
17589 * until runtime; set the run-time fold flag for these. (We don't have to
17590 * worry about properties folding, as that is taken care of by the swash
17591 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
17592 * locales, or the class matches at least one 0-255 range code point */
17595 /* Some things on the list might be unconditionally included because of
17596 * other components. Remove them, and clean up the list if it goes to
17598 if (only_utf8_locale_list && cp_list) {
17599 _invlist_subtract(only_utf8_locale_list, cp_list,
17600 &only_utf8_locale_list);
17602 if (_invlist_len(only_utf8_locale_list) == 0) {
17603 SvREFCNT_dec_NN(only_utf8_locale_list);
17604 only_utf8_locale_list = NULL;
17607 if (only_utf8_locale_list) {
17610 |ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17612 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
17614 invlist_iterinit(cp_list);
17615 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
17616 ANYOF_FLAGS(ret) |= ANYOFL_FOLD;
17618 invlist_iterfinish(cp_list);
17621 else if ( DEPENDS_SEMANTICS
17622 && ( has_upper_latin1_only_utf8_matches
17623 || (ANYOF_FLAGS(ret) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
17626 optimizable = FALSE;
17630 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
17631 * at compile time. Besides not inverting folded locale now, we can't
17632 * invert if there are things such as \w, which aren't known until runtime
17636 && OP(ret) != ANYOFD
17637 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
17638 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17640 _invlist_invert(cp_list);
17642 /* Any swash can't be used as-is, because we've inverted things */
17644 SvREFCNT_dec_NN(swash);
17648 /* Clear the invert flag since have just done it here */
17655 *ret_invlist = cp_list;
17656 SvREFCNT_dec(swash);
17658 /* Discard the generated node */
17660 RExC_size = orig_size;
17663 RExC_emit = orig_emit;
17668 /* Some character classes are equivalent to other nodes. Such nodes take
17669 * up less room and generally fewer operations to execute than ANYOF nodes.
17670 * Above, we checked for and optimized into some such equivalents for
17671 * certain common classes that are easy to test. Getting to this point in
17672 * the code means that the class didn't get optimized there. Since this
17673 * code is only executed in Pass 2, it is too late to save space--it has
17674 * been allocated in Pass 1, and currently isn't given back. But turning
17675 * things into an EXACTish node can allow the optimizer to join it to any
17676 * adjacent such nodes. And if the class is equivalent to things like /./,
17677 * expensive run-time swashes can be avoided. Now that we have more
17678 * complete information, we can find things necessarily missed by the
17679 * earlier code. Another possible "optimization" that isn't done is that
17680 * something like [Ee] could be changed into an EXACTFU. khw tried this
17681 * and found that the ANYOF is faster, including for code points not in the
17682 * bitmap. This still might make sense to do, provided it got joined with
17683 * an adjacent node(s) to create a longer EXACTFU one. This could be
17684 * accomplished by creating a pseudo ANYOF_EXACTFU node type that the join
17685 * routine would know is joinable. If that didn't happen, the node type
17686 * could then be made a straight ANYOF */
17688 if (optimizable && cp_list && ! invert) {
17690 U8 op = END; /* The optimzation node-type */
17691 int posix_class = -1; /* Illegal value */
17692 const char * cur_parse= RExC_parse;
17694 invlist_iterinit(cp_list);
17695 if (! invlist_iternext(cp_list, &start, &end)) {
17697 /* Here, the list is empty. This happens, for example, when a
17698 * Unicode property that doesn't match anything is the only element
17699 * in the character class (perluniprops.pod notes such properties).
17702 *flagp |= HASWIDTH|SIMPLE;
17704 else if (start == end) { /* The range is a single code point */
17705 if (! invlist_iternext(cp_list, &start, &end)
17707 /* Don't do this optimization if it would require changing
17708 * the pattern to UTF-8 */
17709 && (start < 256 || UTF))
17711 /* Here, the list contains a single code point. Can optimize
17712 * into an EXACTish node */
17723 /* A locale node under folding with one code point can be
17724 * an EXACTFL, as its fold won't be calculated until
17730 /* Here, we are generally folding, but there is only one
17731 * code point to match. If we have to, we use an EXACT
17732 * node, but it would be better for joining with adjacent
17733 * nodes in the optimization pass if we used the same
17734 * EXACTFish node that any such are likely to be. We can
17735 * do this iff the code point doesn't participate in any
17736 * folds. For example, an EXACTF of a colon is the same as
17737 * an EXACT one, since nothing folds to or from a colon. */
17739 if (IS_IN_SOME_FOLD_L1(value)) {
17744 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
17749 /* If we haven't found the node type, above, it means we
17750 * can use the prevailing one */
17752 op = compute_EXACTish(pRExC_state);
17756 } /* End of first range contains just a single code point */
17757 else if (start == 0) {
17758 if (end == UV_MAX) {
17760 *flagp |= HASWIDTH|SIMPLE;
17763 else if (end == '\n' - 1
17764 && invlist_iternext(cp_list, &start, &end)
17765 && start == '\n' + 1 && end == UV_MAX)
17768 *flagp |= HASWIDTH|SIMPLE;
17772 invlist_iterfinish(cp_list);
17775 const UV cp_list_len = _invlist_len(cp_list);
17776 const UV* cp_list_array = invlist_array(cp_list);
17778 /* Here, didn't find an optimization. See if this matches any of
17779 * the POSIX classes. These run slightly faster for above-Unicode
17780 * code points, so don't bother with POSIXA ones nor the 2 that
17781 * have no above-Unicode matches. We can avoid these checks unless
17782 * the ANYOF matches at least as high as the lowest POSIX one
17783 * (which was manually found to be \v. The actual code point may
17784 * increase in later Unicode releases, if a higher code point is
17785 * assigned to be \v, but this code will never break. It would
17786 * just mean we could execute the checks for posix optimizations
17787 * unnecessarily) */
17789 if (cp_list_array[cp_list_len-1] > 0x2029) {
17790 for (posix_class = 0;
17791 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
17795 if (posix_class == _CC_ASCII || posix_class == _CC_CNTRL) {
17798 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
17800 /* Check if matches normal or inverted */
17801 if (_invlistEQ(cp_list,
17802 PL_XPosix_ptrs[posix_class],
17805 op = (try_inverted)
17808 *flagp |= HASWIDTH|SIMPLE;
17818 RExC_parse = (char *)orig_parse;
17819 RExC_emit = (regnode *)orig_emit;
17821 if (regarglen[op]) {
17822 ret = reganode(pRExC_state, op, 0);
17824 ret = reg_node(pRExC_state, op);
17827 RExC_parse = (char *)cur_parse;
17829 if (PL_regkind[op] == EXACT) {
17830 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17831 TRUE /* downgradable to EXACT */
17834 else if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17835 FLAGS(ret) = posix_class;
17838 SvREFCNT_dec_NN(cp_list);
17843 /* Here, <cp_list> contains all the code points we can determine at
17844 * compile time that match under all conditions. Go through it, and
17845 * for things that belong in the bitmap, put them there, and delete from
17846 * <cp_list>. While we are at it, see if everything above 255 is in the
17847 * list, and if so, set a flag to speed up execution */
17849 populate_ANYOF_from_invlist(ret, &cp_list);
17852 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
17855 /* Here, the bitmap has been populated with all the Latin1 code points that
17856 * always match. Can now add to the overall list those that match only
17857 * when the target string is UTF-8 (<has_upper_latin1_only_utf8_matches>).
17859 if (has_upper_latin1_only_utf8_matches) {
17861 _invlist_union(cp_list,
17862 has_upper_latin1_only_utf8_matches,
17864 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17867 cp_list = has_upper_latin1_only_utf8_matches;
17869 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
17872 /* If there is a swash and more than one element, we can't use the swash in
17873 * the optimization below. */
17874 if (swash && element_count > 1) {
17875 SvREFCNT_dec_NN(swash);
17879 /* Note that the optimization of using 'swash' if it is the only thing in
17880 * the class doesn't have us change swash at all, so it can include things
17881 * that are also in the bitmap; otherwise we have purposely deleted that
17882 * duplicate information */
17883 set_ANYOF_arg(pRExC_state, ret, cp_list,
17884 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17886 only_utf8_locale_list,
17887 swash, has_user_defined_property);
17889 *flagp |= HASWIDTH|SIMPLE;
17891 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
17892 RExC_contains_locale = 1;
17898 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
17901 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
17902 regnode* const node,
17904 SV* const runtime_defns,
17905 SV* const only_utf8_locale_list,
17907 const bool has_user_defined_property)
17909 /* Sets the arg field of an ANYOF-type node 'node', using information about
17910 * the node passed-in. If there is nothing outside the node's bitmap, the
17911 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
17912 * the count returned by add_data(), having allocated and stored an array,
17913 * av, that that count references, as follows:
17914 * av[0] stores the character class description in its textual form.
17915 * This is used later (regexec.c:Perl_regclass_swash()) to
17916 * initialize the appropriate swash, and is also useful for dumping
17917 * the regnode. This is set to &PL_sv_undef if the textual
17918 * description is not needed at run-time (as happens if the other
17919 * elements completely define the class)
17920 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
17921 * computed from av[0]. But if no further computation need be done,
17922 * the swash is stored here now (and av[0] is &PL_sv_undef).
17923 * av[2] stores the inversion list of code points that match only if the
17924 * current locale is UTF-8
17925 * av[3] stores the cp_list inversion list for use in addition or instead
17926 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
17927 * (Otherwise everything needed is already in av[0] and av[1])
17928 * av[4] is set if any component of the class is from a user-defined
17929 * property; used only if av[3] exists */
17933 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
17935 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
17936 assert(! (ANYOF_FLAGS(node)
17937 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
17938 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
17941 AV * const av = newAV();
17944 av_store(av, 0, (runtime_defns)
17945 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
17948 av_store(av, 1, swash);
17949 SvREFCNT_dec_NN(cp_list);
17952 av_store(av, 1, &PL_sv_undef);
17954 av_store(av, 3, cp_list);
17955 av_store(av, 4, newSVuv(has_user_defined_property));
17959 if (only_utf8_locale_list) {
17960 av_store(av, 2, only_utf8_locale_list);
17963 av_store(av, 2, &PL_sv_undef);
17966 rv = newRV_noinc(MUTABLE_SV(av));
17967 n = add_data(pRExC_state, STR_WITH_LEN("s"));
17968 RExC_rxi->data->data[n] = (void*)rv;
17973 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
17975 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
17976 const regnode* node,
17979 SV** only_utf8_locale_ptr,
17980 SV** output_invlist)
17983 /* For internal core use only.
17984 * Returns the swash for the input 'node' in the regex 'prog'.
17985 * If <doinit> is 'true', will attempt to create the swash if not already
17987 * If <listsvp> is non-null, will return the printable contents of the
17988 * swash. This can be used to get debugging information even before the
17989 * swash exists, by calling this function with 'doinit' set to false, in
17990 * which case the components that will be used to eventually create the
17991 * swash are returned (in a printable form).
17992 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
17993 * store an inversion list of code points that should match only if the
17994 * execution-time locale is a UTF-8 one.
17995 * If <output_invlist> is not NULL, it is where this routine is to store an
17996 * inversion list of the code points that would be instead returned in
17997 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
17998 * when this parameter is used, is just the non-code point data that
17999 * will go into creating the swash. This currently should be just
18000 * user-defined properties whose definitions were not known at compile
18001 * time. Using this parameter allows for easier manipulation of the
18002 * swash's data by the caller. It is illegal to call this function with
18003 * this parameter set, but not <listsvp>
18005 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
18006 * that, in spite of this function's name, the swash it returns may include
18007 * the bitmap data as well */
18010 SV *si = NULL; /* Input swash initialization string */
18011 SV* invlist = NULL;
18013 RXi_GET_DECL(prog,progi);
18014 const struct reg_data * const data = prog ? progi->data : NULL;
18016 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
18017 assert(! output_invlist || listsvp);
18019 if (data && data->count) {
18020 const U32 n = ARG(node);
18022 if (data->what[n] == 's') {
18023 SV * const rv = MUTABLE_SV(data->data[n]);
18024 AV * const av = MUTABLE_AV(SvRV(rv));
18025 SV **const ary = AvARRAY(av);
18026 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
18028 si = *ary; /* ary[0] = the string to initialize the swash with */
18030 if (av_tindex_nomg(av) >= 2) {
18031 if (only_utf8_locale_ptr
18033 && ary[2] != &PL_sv_undef)
18035 *only_utf8_locale_ptr = ary[2];
18038 assert(only_utf8_locale_ptr);
18039 *only_utf8_locale_ptr = NULL;
18042 /* Elements 3 and 4 are either both present or both absent. [3]
18043 * is any inversion list generated at compile time; [4]
18044 * indicates if that inversion list has any user-defined
18045 * properties in it. */
18046 if (av_tindex_nomg(av) >= 3) {
18048 if (SvUV(ary[4])) {
18049 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
18057 /* Element [1] is reserved for the set-up swash. If already there,
18058 * return it; if not, create it and store it there */
18059 if (ary[1] && SvROK(ary[1])) {
18062 else if (doinit && ((si && si != &PL_sv_undef)
18063 || (invlist && invlist != &PL_sv_undef))) {
18065 sw = _core_swash_init("utf8", /* the utf8 package */
18069 0, /* not from tr/// */
18071 &swash_init_flags);
18072 (void)av_store(av, 1, sw);
18077 /* If requested, return a printable version of what this swash matches */
18079 SV* matches_string = NULL;
18081 /* The swash should be used, if possible, to get the data, as it
18082 * contains the resolved data. But this function can be called at
18083 * compile-time, before everything gets resolved, in which case we
18084 * return the currently best available information, which is the string
18085 * that will eventually be used to do that resolving, 'si' */
18086 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
18087 && (si && si != &PL_sv_undef))
18089 /* Here, we only have 'si' (and possibly some passed-in data in
18090 * 'invlist', which is handled below) If the caller only wants
18091 * 'si', use that. */
18092 if (! output_invlist) {
18093 matches_string = newSVsv(si);
18096 /* But if the caller wants an inversion list of the node, we
18097 * need to parse 'si' and place as much as possible in the
18098 * desired output inversion list, making 'matches_string' only
18099 * contain the currently unresolvable things */
18100 const char *si_string = SvPVX(si);
18101 STRLEN remaining = SvCUR(si);
18105 /* Ignore everything before the first new-line */
18106 while (*si_string != '\n' && remaining > 0) {
18110 assert(remaining > 0);
18115 while (remaining > 0) {
18117 /* The data consists of just strings defining user-defined
18118 * property names, but in prior incarnations, and perhaps
18119 * somehow from pluggable regex engines, it could still
18120 * hold hex code point definitions. Each component of a
18121 * range would be separated by a tab, and each range by a
18122 * new-line. If these are found, instead add them to the
18123 * inversion list */
18124 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
18125 |PERL_SCAN_SILENT_NON_PORTABLE;
18126 STRLEN len = remaining;
18127 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
18129 /* If the hex decode routine found something, it should go
18130 * up to the next \n */
18131 if ( *(si_string + len) == '\n') {
18132 if (count) { /* 2nd code point on line */
18133 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
18136 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
18139 goto prepare_for_next_iteration;
18142 /* If the hex decode was instead for the lower range limit,
18143 * save it, and go parse the upper range limit */
18144 if (*(si_string + len) == '\t') {
18145 assert(count == 0);
18149 prepare_for_next_iteration:
18150 si_string += len + 1;
18151 remaining -= len + 1;
18155 /* Here, didn't find a legal hex number. Just add it from
18156 * here to the next \n */
18159 while (*(si_string + len) != '\n' && remaining > 0) {
18163 if (*(si_string + len) == '\n') {
18167 if (matches_string) {
18168 sv_catpvn(matches_string, si_string, len - 1);
18171 matches_string = newSVpvn(si_string, len - 1);
18174 sv_catpvs(matches_string, " ");
18175 } /* end of loop through the text */
18177 assert(matches_string);
18178 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
18179 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
18181 } /* end of has an 'si' but no swash */
18184 /* If we have a swash in place, its equivalent inversion list was above
18185 * placed into 'invlist'. If not, this variable may contain a stored
18186 * inversion list which is information beyond what is in 'si' */
18189 /* Again, if the caller doesn't want the output inversion list, put
18190 * everything in 'matches-string' */
18191 if (! output_invlist) {
18192 if ( ! matches_string) {
18193 matches_string = newSVpvs("\n");
18195 sv_catsv(matches_string, invlist_contents(invlist,
18196 TRUE /* traditional style */
18199 else if (! *output_invlist) {
18200 *output_invlist = invlist_clone(invlist);
18203 _invlist_union(*output_invlist, invlist, output_invlist);
18207 *listsvp = matches_string;
18212 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
18214 /* reg_skipcomment()
18216 Absorbs an /x style # comment from the input stream,
18217 returning a pointer to the first character beyond the comment, or if the
18218 comment terminates the pattern without anything following it, this returns
18219 one past the final character of the pattern (in other words, RExC_end) and
18220 sets the REG_RUN_ON_COMMENT_SEEN flag.
18222 Note it's the callers responsibility to ensure that we are
18223 actually in /x mode
18227 PERL_STATIC_INLINE char*
18228 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
18230 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
18234 while (p < RExC_end) {
18235 if (*(++p) == '\n') {
18240 /* we ran off the end of the pattern without ending the comment, so we have
18241 * to add an \n when wrapping */
18242 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
18247 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
18249 const bool force_to_xmod
18252 /* If the text at the current parse position '*p' is a '(?#...)' comment,
18253 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
18254 * is /x whitespace, advance '*p' so that on exit it points to the first
18255 * byte past all such white space and comments */
18257 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
18259 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
18261 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
18264 if (RExC_end - (*p) >= 3
18266 && *(*p + 1) == '?'
18267 && *(*p + 2) == '#')
18269 while (*(*p) != ')') {
18270 if ((*p) == RExC_end)
18271 FAIL("Sequence (?#... not terminated");
18279 const char * save_p = *p;
18280 while ((*p) < RExC_end) {
18282 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
18285 else if (*(*p) == '#') {
18286 (*p) = reg_skipcomment(pRExC_state, (*p));
18292 if (*p != save_p) {
18305 Advances the parse position by one byte, unless that byte is the beginning
18306 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
18307 those two cases, the parse position is advanced beyond all such comments and
18310 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
18314 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
18316 PERL_ARGS_ASSERT_NEXTCHAR;
18318 if (RExC_parse < RExC_end) {
18320 || UTF8_IS_INVARIANT(*RExC_parse)
18321 || UTF8_IS_START(*RExC_parse));
18323 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
18325 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
18326 FALSE /* Don't force /x */ );
18331 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
18333 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
18334 * space. In pass1, it aligns and increments RExC_size; in pass2,
18337 regnode * const ret = RExC_emit;
18338 GET_RE_DEBUG_FLAGS_DECL;
18340 PERL_ARGS_ASSERT_REGNODE_GUTS;
18342 assert(extra_size >= regarglen[op]);
18345 SIZE_ALIGN(RExC_size);
18346 RExC_size += 1 + extra_size;
18349 if (RExC_emit >= RExC_emit_bound)
18350 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
18351 op, (void*)RExC_emit, (void*)RExC_emit_bound);
18353 NODE_ALIGN_FILL(ret);
18354 #ifndef RE_TRACK_PATTERN_OFFSETS
18355 PERL_UNUSED_ARG(name);
18357 if (RExC_offsets) { /* MJD */
18359 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
18362 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
18363 ? "Overwriting end of array!\n" : "OK",
18364 (UV)(RExC_emit - RExC_emit_start),
18365 (UV)(RExC_parse - RExC_start),
18366 (UV)RExC_offsets[0]));
18367 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
18374 - reg_node - emit a node
18376 STATIC regnode * /* Location. */
18377 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
18379 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
18381 PERL_ARGS_ASSERT_REG_NODE;
18383 assert(regarglen[op] == 0);
18386 regnode *ptr = ret;
18387 FILL_ADVANCE_NODE(ptr, op);
18394 - reganode - emit a node with an argument
18396 STATIC regnode * /* Location. */
18397 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
18399 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
18401 PERL_ARGS_ASSERT_REGANODE;
18403 assert(regarglen[op] == 1);
18406 regnode *ptr = ret;
18407 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
18414 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
18416 /* emit a node with U32 and I32 arguments */
18418 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
18420 PERL_ARGS_ASSERT_REG2LANODE;
18422 assert(regarglen[op] == 2);
18425 regnode *ptr = ret;
18426 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
18433 - reginsert - insert an operator in front of already-emitted operand
18435 * Means relocating the operand.
18438 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
18443 const int offset = regarglen[(U8)op];
18444 const int size = NODE_STEP_REGNODE + offset;
18445 GET_RE_DEBUG_FLAGS_DECL;
18447 PERL_ARGS_ASSERT_REGINSERT;
18448 PERL_UNUSED_CONTEXT;
18449 PERL_UNUSED_ARG(depth);
18450 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
18451 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
18456 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
18457 studying. If this is wrong then we need to adjust RExC_recurse
18458 below like we do with RExC_open_parens/RExC_close_parens. */
18462 if (RExC_open_parens) {
18464 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
18465 /* remember that RExC_npar is rex->nparens + 1,
18466 * iow it is 1 more than the number of parens seen in
18467 * the pattern so far. */
18468 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
18469 /* note, RExC_open_parens[0] is the start of the
18470 * regex, it can't move. RExC_close_parens[0] is the end
18471 * of the regex, it *can* move. */
18472 if ( paren && RExC_open_parens[paren] >= opnd ) {
18473 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
18474 RExC_open_parens[paren] += size;
18476 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
18478 if ( RExC_close_parens[paren] >= opnd ) {
18479 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
18480 RExC_close_parens[paren] += size;
18482 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
18487 RExC_end_op += size;
18489 while (src > opnd) {
18490 StructCopy(--src, --dst, regnode);
18491 #ifdef RE_TRACK_PATTERN_OFFSETS
18492 if (RExC_offsets) { /* MJD 20010112 */
18494 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
18498 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
18499 ? "Overwriting end of array!\n" : "OK",
18500 (UV)(src - RExC_emit_start),
18501 (UV)(dst - RExC_emit_start),
18502 (UV)RExC_offsets[0]));
18503 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
18504 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
18510 place = opnd; /* Op node, where operand used to be. */
18511 #ifdef RE_TRACK_PATTERN_OFFSETS
18512 if (RExC_offsets) { /* MJD */
18514 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
18518 (UV)(place - RExC_emit_start) > RExC_offsets[0]
18519 ? "Overwriting end of array!\n" : "OK",
18520 (UV)(place - RExC_emit_start),
18521 (UV)(RExC_parse - RExC_start),
18522 (UV)RExC_offsets[0]));
18523 Set_Node_Offset(place, RExC_parse);
18524 Set_Node_Length(place, 1);
18527 src = NEXTOPER(place);
18528 FILL_ADVANCE_NODE(place, op);
18529 Zero(src, offset, regnode);
18533 - regtail - set the next-pointer at the end of a node chain of p to val.
18534 - SEE ALSO: regtail_study
18537 S_regtail(pTHX_ RExC_state_t * pRExC_state,
18538 const regnode * const p,
18539 const regnode * const val,
18543 GET_RE_DEBUG_FLAGS_DECL;
18545 PERL_ARGS_ASSERT_REGTAIL;
18547 PERL_UNUSED_ARG(depth);
18553 /* Find last node. */
18554 scan = (regnode *) p;
18556 regnode * const temp = regnext(scan);
18558 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
18559 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18560 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
18561 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
18562 (temp == NULL ? "->" : ""),
18563 (temp == NULL ? PL_reg_name[OP(val)] : "")
18571 if (reg_off_by_arg[OP(scan)]) {
18572 ARG_SET(scan, val - scan);
18575 NEXT_OFF(scan) = val - scan;
18581 - regtail_study - set the next-pointer at the end of a node chain of p to val.
18582 - Look for optimizable sequences at the same time.
18583 - currently only looks for EXACT chains.
18585 This is experimental code. The idea is to use this routine to perform
18586 in place optimizations on branches and groups as they are constructed,
18587 with the long term intention of removing optimization from study_chunk so
18588 that it is purely analytical.
18590 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
18591 to control which is which.
18594 /* TODO: All four parms should be const */
18597 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
18598 const regnode *val,U32 depth)
18602 #ifdef EXPERIMENTAL_INPLACESCAN
18605 GET_RE_DEBUG_FLAGS_DECL;
18607 PERL_ARGS_ASSERT_REGTAIL_STUDY;
18613 /* Find last node. */
18617 regnode * const temp = regnext(scan);
18618 #ifdef EXPERIMENTAL_INPLACESCAN
18619 if (PL_regkind[OP(scan)] == EXACT) {
18620 bool unfolded_multi_char; /* Unexamined in this routine */
18621 if (join_exact(pRExC_state, scan, &min,
18622 &unfolded_multi_char, 1, val, depth+1))
18627 switch (OP(scan)) {
18631 case EXACTFA_NO_TRIE:
18637 if( exact == PSEUDO )
18639 else if ( exact != OP(scan) )
18648 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
18649 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18650 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
18651 SvPV_nolen_const(RExC_mysv),
18652 REG_NODE_NUM(scan),
18653 PL_reg_name[exact]);
18660 DEBUG_PARSE_MSG("");
18661 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
18662 Perl_re_printf( aTHX_
18663 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
18664 SvPV_nolen_const(RExC_mysv),
18665 (IV)REG_NODE_NUM(val),
18669 if (reg_off_by_arg[OP(scan)]) {
18670 ARG_SET(scan, val - scan);
18673 NEXT_OFF(scan) = val - scan;
18681 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
18686 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
18691 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18693 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
18694 if (flags & (1<<bit)) {
18695 if (!set++ && lead)
18696 Perl_re_printf( aTHX_ "%s",lead);
18697 Perl_re_printf( aTHX_ "%s ",PL_reg_intflags_name[bit]);
18702 Perl_re_printf( aTHX_ "\n");
18704 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18709 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
18715 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18717 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
18718 if (flags & (1<<bit)) {
18719 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
18722 if (!set++ && lead)
18723 Perl_re_printf( aTHX_ "%s",lead);
18724 Perl_re_printf( aTHX_ "%s ",PL_reg_extflags_name[bit]);
18727 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
18728 if (!set++ && lead) {
18729 Perl_re_printf( aTHX_ "%s",lead);
18732 case REGEX_UNICODE_CHARSET:
18733 Perl_re_printf( aTHX_ "UNICODE");
18735 case REGEX_LOCALE_CHARSET:
18736 Perl_re_printf( aTHX_ "LOCALE");
18738 case REGEX_ASCII_RESTRICTED_CHARSET:
18739 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
18741 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
18742 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
18745 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
18751 Perl_re_printf( aTHX_ "\n");
18753 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18759 Perl_regdump(pTHX_ const regexp *r)
18762 SV * const sv = sv_newmortal();
18763 SV *dsv= sv_newmortal();
18764 RXi_GET_DECL(r,ri);
18765 GET_RE_DEBUG_FLAGS_DECL;
18767 PERL_ARGS_ASSERT_REGDUMP;
18769 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
18771 /* Header fields of interest. */
18772 if (r->anchored_substr) {
18773 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
18774 RE_SV_DUMPLEN(r->anchored_substr), 30);
18775 Perl_re_printf( aTHX_
18776 "anchored %s%s at %" IVdf " ",
18777 s, RE_SV_TAIL(r->anchored_substr),
18778 (IV)r->anchored_offset);
18779 } else if (r->anchored_utf8) {
18780 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
18781 RE_SV_DUMPLEN(r->anchored_utf8), 30);
18782 Perl_re_printf( aTHX_
18783 "anchored utf8 %s%s at %" IVdf " ",
18784 s, RE_SV_TAIL(r->anchored_utf8),
18785 (IV)r->anchored_offset);
18787 if (r->float_substr) {
18788 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
18789 RE_SV_DUMPLEN(r->float_substr), 30);
18790 Perl_re_printf( aTHX_
18791 "floating %s%s at %" IVdf "..%" UVuf " ",
18792 s, RE_SV_TAIL(r->float_substr),
18793 (IV)r->float_min_offset, (UV)r->float_max_offset);
18794 } else if (r->float_utf8) {
18795 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
18796 RE_SV_DUMPLEN(r->float_utf8), 30);
18797 Perl_re_printf( aTHX_
18798 "floating utf8 %s%s at %" IVdf "..%" UVuf " ",
18799 s, RE_SV_TAIL(r->float_utf8),
18800 (IV)r->float_min_offset, (UV)r->float_max_offset);
18802 if (r->check_substr || r->check_utf8)
18803 Perl_re_printf( aTHX_
18805 (r->check_substr == r->float_substr
18806 && r->check_utf8 == r->float_utf8
18807 ? "(checking floating" : "(checking anchored"));
18808 if (r->intflags & PREGf_NOSCAN)
18809 Perl_re_printf( aTHX_ " noscan");
18810 if (r->extflags & RXf_CHECK_ALL)
18811 Perl_re_printf( aTHX_ " isall");
18812 if (r->check_substr || r->check_utf8)
18813 Perl_re_printf( aTHX_ ") ");
18815 if (ri->regstclass) {
18816 regprop(r, sv, ri->regstclass, NULL, NULL);
18817 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
18819 if (r->intflags & PREGf_ANCH) {
18820 Perl_re_printf( aTHX_ "anchored");
18821 if (r->intflags & PREGf_ANCH_MBOL)
18822 Perl_re_printf( aTHX_ "(MBOL)");
18823 if (r->intflags & PREGf_ANCH_SBOL)
18824 Perl_re_printf( aTHX_ "(SBOL)");
18825 if (r->intflags & PREGf_ANCH_GPOS)
18826 Perl_re_printf( aTHX_ "(GPOS)");
18827 Perl_re_printf( aTHX_ " ");
18829 if (r->intflags & PREGf_GPOS_SEEN)
18830 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
18831 if (r->intflags & PREGf_SKIP)
18832 Perl_re_printf( aTHX_ "plus ");
18833 if (r->intflags & PREGf_IMPLICIT)
18834 Perl_re_printf( aTHX_ "implicit ");
18835 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
18836 if (r->extflags & RXf_EVAL_SEEN)
18837 Perl_re_printf( aTHX_ "with eval ");
18838 Perl_re_printf( aTHX_ "\n");
18840 regdump_extflags("r->extflags: ",r->extflags);
18841 regdump_intflags("r->intflags: ",r->intflags);
18844 PERL_ARGS_ASSERT_REGDUMP;
18845 PERL_UNUSED_CONTEXT;
18846 PERL_UNUSED_ARG(r);
18847 #endif /* DEBUGGING */
18850 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
18853 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
18854 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
18855 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
18856 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
18857 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
18858 || _CC_VERTSPACE != 15
18859 # error Need to adjust order of anyofs[]
18861 static const char * const anyofs[] = {
18898 - regprop - printable representation of opcode, with run time support
18902 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
18906 RXi_GET_DECL(prog,progi);
18907 GET_RE_DEBUG_FLAGS_DECL;
18909 PERL_ARGS_ASSERT_REGPROP;
18913 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
18914 /* It would be nice to FAIL() here, but this may be called from
18915 regexec.c, and it would be hard to supply pRExC_state. */
18916 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
18917 (int)OP(o), (int)REGNODE_MAX);
18918 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
18920 k = PL_regkind[OP(o)];
18923 sv_catpvs(sv, " ");
18924 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
18925 * is a crude hack but it may be the best for now since
18926 * we have no flag "this EXACTish node was UTF-8"
18928 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
18929 PERL_PV_ESCAPE_UNI_DETECT |
18930 PERL_PV_ESCAPE_NONASCII |
18931 PERL_PV_PRETTY_ELLIPSES |
18932 PERL_PV_PRETTY_LTGT |
18933 PERL_PV_PRETTY_NOCLEAR
18935 } else if (k == TRIE) {
18936 /* print the details of the trie in dumpuntil instead, as
18937 * progi->data isn't available here */
18938 const char op = OP(o);
18939 const U32 n = ARG(o);
18940 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
18941 (reg_ac_data *)progi->data->data[n] :
18943 const reg_trie_data * const trie
18944 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
18946 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
18947 DEBUG_TRIE_COMPILE_r({
18949 sv_catpvs(sv, "(JUMP)");
18950 Perl_sv_catpvf(aTHX_ sv,
18951 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
18952 (UV)trie->startstate,
18953 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
18954 (UV)trie->wordcount,
18957 (UV)TRIE_CHARCOUNT(trie),
18958 (UV)trie->uniquecharcount
18961 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
18962 sv_catpvs(sv, "[");
18963 (void) put_charclass_bitmap_innards(sv,
18964 ((IS_ANYOF_TRIE(op))
18966 : TRIE_BITMAP(trie)),
18972 sv_catpvs(sv, "]");
18974 } else if (k == CURLY) {
18975 U32 lo = ARG1(o), hi = ARG2(o);
18976 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
18977 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
18978 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
18979 if (hi == REG_INFTY)
18980 sv_catpvs(sv, "INFTY");
18982 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
18983 sv_catpvs(sv, "}");
18985 else if (k == WHILEM && o->flags) /* Ordinal/of */
18986 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
18987 else if (k == REF || k == OPEN || k == CLOSE
18988 || k == GROUPP || OP(o)==ACCEPT)
18990 AV *name_list= NULL;
18991 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
18992 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
18993 if ( RXp_PAREN_NAMES(prog) ) {
18994 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
18995 } else if ( pRExC_state ) {
18996 name_list= RExC_paren_name_list;
18999 if ( k != REF || (OP(o) < NREF)) {
19000 SV **name= av_fetch(name_list, parno, 0 );
19002 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19005 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
19006 I32 *nums=(I32*)SvPVX(sv_dat);
19007 SV **name= av_fetch(name_list, nums[0], 0 );
19010 for ( n=0; n<SvIVX(sv_dat); n++ ) {
19011 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
19012 (n ? "," : ""), (IV)nums[n]);
19014 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19018 if ( k == REF && reginfo) {
19019 U32 n = ARG(o); /* which paren pair */
19020 I32 ln = prog->offs[n].start;
19021 if (prog->lastparen < n || ln == -1)
19022 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
19023 else if (ln == prog->offs[n].end)
19024 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
19026 const char *s = reginfo->strbeg + ln;
19027 Perl_sv_catpvf(aTHX_ sv, ": ");
19028 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
19029 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
19032 } else if (k == GOSUB) {
19033 AV *name_list= NULL;
19034 if ( RXp_PAREN_NAMES(prog) ) {
19035 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19036 } else if ( pRExC_state ) {
19037 name_list= RExC_paren_name_list;
19040 /* Paren and offset */
19041 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
19042 (int)((o + (int)ARG2L(o)) - progi->program) );
19044 SV **name= av_fetch(name_list, ARG(o), 0 );
19046 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19049 else if (k == LOGICAL)
19050 /* 2: embedded, otherwise 1 */
19051 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
19052 else if (k == ANYOF) {
19053 const U8 flags = ANYOF_FLAGS(o);
19054 bool do_sep = FALSE; /* Do we need to separate various components of
19056 /* Set if there is still an unresolved user-defined property */
19057 SV *unresolved = NULL;
19059 /* Things that are ignored except when the runtime locale is UTF-8 */
19060 SV *only_utf8_locale_invlist = NULL;
19062 /* Code points that don't fit in the bitmap */
19063 SV *nonbitmap_invlist = NULL;
19065 /* And things that aren't in the bitmap, but are small enough to be */
19066 SV* bitmap_range_not_in_bitmap = NULL;
19068 const bool inverted = flags & ANYOF_INVERT;
19070 if (OP(o) == ANYOFL) {
19071 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
19072 sv_catpvs(sv, "{utf8-locale-reqd}");
19074 if (flags & ANYOFL_FOLD) {
19075 sv_catpvs(sv, "{i}");
19079 /* If there is stuff outside the bitmap, get it */
19080 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
19081 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
19083 &only_utf8_locale_invlist,
19084 &nonbitmap_invlist);
19085 /* The non-bitmap data may contain stuff that could fit in the
19086 * bitmap. This could come from a user-defined property being
19087 * finally resolved when this call was done; or much more likely
19088 * because there are matches that require UTF-8 to be valid, and so
19089 * aren't in the bitmap. This is teased apart later */
19090 _invlist_intersection(nonbitmap_invlist,
19092 &bitmap_range_not_in_bitmap);
19093 /* Leave just the things that don't fit into the bitmap */
19094 _invlist_subtract(nonbitmap_invlist,
19096 &nonbitmap_invlist);
19099 /* Obey this flag to add all above-the-bitmap code points */
19100 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
19101 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
19102 NUM_ANYOF_CODE_POINTS,
19106 /* Ready to start outputting. First, the initial left bracket */
19107 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
19109 /* Then all the things that could fit in the bitmap */
19110 do_sep = put_charclass_bitmap_innards(sv,
19112 bitmap_range_not_in_bitmap,
19113 only_utf8_locale_invlist,
19116 /* Can't try inverting for a
19117 * better display if there are
19118 * things that haven't been
19120 unresolved != NULL);
19121 SvREFCNT_dec(bitmap_range_not_in_bitmap);
19123 /* If there are user-defined properties which haven't been defined yet,
19124 * output them. If the result is not to be inverted, it is clearest to
19125 * output them in a separate [] from the bitmap range stuff. If the
19126 * result is to be complemented, we have to show everything in one [],
19127 * as the inversion applies to the whole thing. Use {braces} to
19128 * separate them from anything in the bitmap and anything above the
19132 if (! do_sep) { /* If didn't output anything in the bitmap */
19133 sv_catpvs(sv, "^");
19135 sv_catpvs(sv, "{");
19138 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19140 sv_catsv(sv, unresolved);
19142 sv_catpvs(sv, "}");
19144 do_sep = ! inverted;
19147 /* And, finally, add the above-the-bitmap stuff */
19148 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
19151 /* See if truncation size is overridden */
19152 const STRLEN dump_len = (PL_dump_re_max_len)
19153 ? PL_dump_re_max_len
19156 /* This is output in a separate [] */
19158 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19161 /* And, for easy of understanding, it is shown in the
19162 * uncomplemented form if possible. The one exception being if
19163 * there are unresolved items, where the inversion has to be
19164 * delayed until runtime */
19165 if (inverted && ! unresolved) {
19166 _invlist_invert(nonbitmap_invlist);
19167 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
19170 contents = invlist_contents(nonbitmap_invlist,
19171 FALSE /* output suitable for catsv */
19174 /* If the output is shorter than the permissible maximum, just do it. */
19175 if (SvCUR(contents) <= dump_len) {
19176 sv_catsv(sv, contents);
19179 const char * contents_string = SvPVX(contents);
19180 STRLEN i = dump_len;
19182 /* Otherwise, start at the permissible max and work back to the
19183 * first break possibility */
19184 while (i > 0 && contents_string[i] != ' ') {
19187 if (i == 0) { /* Fail-safe. Use the max if we couldn't
19188 find a legal break */
19192 sv_catpvn(sv, contents_string, i);
19193 sv_catpvs(sv, "...");
19196 SvREFCNT_dec_NN(contents);
19197 SvREFCNT_dec_NN(nonbitmap_invlist);
19200 /* And finally the matching, closing ']' */
19201 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
19203 SvREFCNT_dec(unresolved);
19205 else if (k == POSIXD || k == NPOSIXD) {
19206 U8 index = FLAGS(o) * 2;
19207 if (index < C_ARRAY_LENGTH(anyofs)) {
19208 if (*anyofs[index] != '[') {
19211 sv_catpv(sv, anyofs[index]);
19212 if (*anyofs[index] != '[') {
19217 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
19220 else if (k == BOUND || k == NBOUND) {
19221 /* Must be synced with order of 'bound_type' in regcomp.h */
19222 const char * const bounds[] = {
19223 "", /* Traditional */
19229 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
19230 sv_catpv(sv, bounds[FLAGS(o)]);
19232 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
19233 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
19234 else if (OP(o) == SBOL)
19235 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
19237 /* add on the verb argument if there is one */
19238 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
19239 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
19240 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
19243 PERL_UNUSED_CONTEXT;
19244 PERL_UNUSED_ARG(sv);
19245 PERL_UNUSED_ARG(o);
19246 PERL_UNUSED_ARG(prog);
19247 PERL_UNUSED_ARG(reginfo);
19248 PERL_UNUSED_ARG(pRExC_state);
19249 #endif /* DEBUGGING */
19255 Perl_re_intuit_string(pTHX_ REGEXP * const r)
19256 { /* Assume that RE_INTUIT is set */
19257 struct regexp *const prog = ReANY(r);
19258 GET_RE_DEBUG_FLAGS_DECL;
19260 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
19261 PERL_UNUSED_CONTEXT;
19265 const char * const s = SvPV_nolen_const(RX_UTF8(r)
19266 ? prog->check_utf8 : prog->check_substr);
19268 if (!PL_colorset) reginitcolors();
19269 Perl_re_printf( aTHX_
19270 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
19272 RX_UTF8(r) ? "utf8 " : "",
19273 PL_colors[5],PL_colors[0],
19276 (strlen(s) > 60 ? "..." : ""));
19279 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
19280 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
19286 handles refcounting and freeing the perl core regexp structure. When
19287 it is necessary to actually free the structure the first thing it
19288 does is call the 'free' method of the regexp_engine associated to
19289 the regexp, allowing the handling of the void *pprivate; member
19290 first. (This routine is not overridable by extensions, which is why
19291 the extensions free is called first.)
19293 See regdupe and regdupe_internal if you change anything here.
19295 #ifndef PERL_IN_XSUB_RE
19297 Perl_pregfree(pTHX_ REGEXP *r)
19303 Perl_pregfree2(pTHX_ REGEXP *rx)
19305 struct regexp *const r = ReANY(rx);
19306 GET_RE_DEBUG_FLAGS_DECL;
19308 PERL_ARGS_ASSERT_PREGFREE2;
19310 if (r->mother_re) {
19311 ReREFCNT_dec(r->mother_re);
19313 CALLREGFREE_PVT(rx); /* free the private data */
19314 SvREFCNT_dec(RXp_PAREN_NAMES(r));
19315 Safefree(r->xpv_len_u.xpvlenu_pv);
19318 SvREFCNT_dec(r->anchored_substr);
19319 SvREFCNT_dec(r->anchored_utf8);
19320 SvREFCNT_dec(r->float_substr);
19321 SvREFCNT_dec(r->float_utf8);
19322 Safefree(r->substrs);
19324 RX_MATCH_COPY_FREE(rx);
19325 #ifdef PERL_ANY_COW
19326 SvREFCNT_dec(r->saved_copy);
19329 SvREFCNT_dec(r->qr_anoncv);
19330 if (r->recurse_locinput)
19331 Safefree(r->recurse_locinput);
19332 rx->sv_u.svu_rx = 0;
19337 This is a hacky workaround to the structural issue of match results
19338 being stored in the regexp structure which is in turn stored in
19339 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
19340 could be PL_curpm in multiple contexts, and could require multiple
19341 result sets being associated with the pattern simultaneously, such
19342 as when doing a recursive match with (??{$qr})
19344 The solution is to make a lightweight copy of the regexp structure
19345 when a qr// is returned from the code executed by (??{$qr}) this
19346 lightweight copy doesn't actually own any of its data except for
19347 the starp/end and the actual regexp structure itself.
19353 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
19355 struct regexp *ret;
19356 struct regexp *const r = ReANY(rx);
19357 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
19359 PERL_ARGS_ASSERT_REG_TEMP_COPY;
19362 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
19364 SvOK_off((SV *)ret_x);
19366 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
19367 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
19368 made both spots point to the same regexp body.) */
19369 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
19370 assert(!SvPVX(ret_x));
19371 ret_x->sv_u.svu_rx = temp->sv_any;
19372 temp->sv_any = NULL;
19373 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
19374 SvREFCNT_dec_NN(temp);
19375 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
19376 ing below will not set it. */
19377 SvCUR_set(ret_x, SvCUR(rx));
19380 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
19381 sv_force_normal(sv) is called. */
19383 ret = ReANY(ret_x);
19385 SvFLAGS(ret_x) |= SvUTF8(rx);
19386 /* We share the same string buffer as the original regexp, on which we
19387 hold a reference count, incremented when mother_re is set below.
19388 The string pointer is copied here, being part of the regexp struct.
19390 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
19391 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
19393 const I32 npar = r->nparens+1;
19394 Newx(ret->offs, npar, regexp_paren_pair);
19395 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19398 Newx(ret->substrs, 1, struct reg_substr_data);
19399 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19401 SvREFCNT_inc_void(ret->anchored_substr);
19402 SvREFCNT_inc_void(ret->anchored_utf8);
19403 SvREFCNT_inc_void(ret->float_substr);
19404 SvREFCNT_inc_void(ret->float_utf8);
19406 /* check_substr and check_utf8, if non-NULL, point to either their
19407 anchored or float namesakes, and don't hold a second reference. */
19409 RX_MATCH_COPIED_off(ret_x);
19410 #ifdef PERL_ANY_COW
19411 ret->saved_copy = NULL;
19413 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
19414 SvREFCNT_inc_void(ret->qr_anoncv);
19415 if (r->recurse_locinput)
19416 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19422 /* regfree_internal()
19424 Free the private data in a regexp. This is overloadable by
19425 extensions. Perl takes care of the regexp structure in pregfree(),
19426 this covers the *pprivate pointer which technically perl doesn't
19427 know about, however of course we have to handle the
19428 regexp_internal structure when no extension is in use.
19430 Note this is called before freeing anything in the regexp
19435 Perl_regfree_internal(pTHX_ REGEXP * const rx)
19437 struct regexp *const r = ReANY(rx);
19438 RXi_GET_DECL(r,ri);
19439 GET_RE_DEBUG_FLAGS_DECL;
19441 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
19447 SV *dsv= sv_newmortal();
19448 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
19449 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
19450 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
19451 PL_colors[4],PL_colors[5],s);
19454 #ifdef RE_TRACK_PATTERN_OFFSETS
19456 Safefree(ri->u.offsets); /* 20010421 MJD */
19458 if (ri->code_blocks) {
19460 for (n = 0; n < ri->num_code_blocks; n++)
19461 SvREFCNT_dec(ri->code_blocks[n].src_regex);
19462 Safefree(ri->code_blocks);
19466 int n = ri->data->count;
19469 /* If you add a ->what type here, update the comment in regcomp.h */
19470 switch (ri->data->what[n]) {
19476 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
19479 Safefree(ri->data->data[n]);
19485 { /* Aho Corasick add-on structure for a trie node.
19486 Used in stclass optimization only */
19488 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
19489 #ifdef USE_ITHREADS
19493 refcount = --aho->refcount;
19496 PerlMemShared_free(aho->states);
19497 PerlMemShared_free(aho->fail);
19498 /* do this last!!!! */
19499 PerlMemShared_free(ri->data->data[n]);
19500 /* we should only ever get called once, so
19501 * assert as much, and also guard the free
19502 * which /might/ happen twice. At the least
19503 * it will make code anlyzers happy and it
19504 * doesn't cost much. - Yves */
19505 assert(ri->regstclass);
19506 if (ri->regstclass) {
19507 PerlMemShared_free(ri->regstclass);
19508 ri->regstclass = 0;
19515 /* trie structure. */
19517 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
19518 #ifdef USE_ITHREADS
19522 refcount = --trie->refcount;
19525 PerlMemShared_free(trie->charmap);
19526 PerlMemShared_free(trie->states);
19527 PerlMemShared_free(trie->trans);
19529 PerlMemShared_free(trie->bitmap);
19531 PerlMemShared_free(trie->jump);
19532 PerlMemShared_free(trie->wordinfo);
19533 /* do this last!!!! */
19534 PerlMemShared_free(ri->data->data[n]);
19539 Perl_croak(aTHX_ "panic: regfree data code '%c'",
19540 ri->data->what[n]);
19543 Safefree(ri->data->what);
19544 Safefree(ri->data);
19550 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
19551 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
19552 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
19555 re_dup_guts - duplicate a regexp.
19557 This routine is expected to clone a given regexp structure. It is only
19558 compiled under USE_ITHREADS.
19560 After all of the core data stored in struct regexp is duplicated
19561 the regexp_engine.dupe method is used to copy any private data
19562 stored in the *pprivate pointer. This allows extensions to handle
19563 any duplication it needs to do.
19565 See pregfree() and regfree_internal() if you change anything here.
19567 #if defined(USE_ITHREADS)
19568 #ifndef PERL_IN_XSUB_RE
19570 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
19574 const struct regexp *r = ReANY(sstr);
19575 struct regexp *ret = ReANY(dstr);
19577 PERL_ARGS_ASSERT_RE_DUP_GUTS;
19579 npar = r->nparens+1;
19580 Newx(ret->offs, npar, regexp_paren_pair);
19581 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19583 if (ret->substrs) {
19584 /* Do it this way to avoid reading from *r after the StructCopy().
19585 That way, if any of the sv_dup_inc()s dislodge *r from the L1
19586 cache, it doesn't matter. */
19587 const bool anchored = r->check_substr
19588 ? r->check_substr == r->anchored_substr
19589 : r->check_utf8 == r->anchored_utf8;
19590 Newx(ret->substrs, 1, struct reg_substr_data);
19591 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19593 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
19594 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
19595 ret->float_substr = sv_dup_inc(ret->float_substr, param);
19596 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
19598 /* check_substr and check_utf8, if non-NULL, point to either their
19599 anchored or float namesakes, and don't hold a second reference. */
19601 if (ret->check_substr) {
19603 assert(r->check_utf8 == r->anchored_utf8);
19604 ret->check_substr = ret->anchored_substr;
19605 ret->check_utf8 = ret->anchored_utf8;
19607 assert(r->check_substr == r->float_substr);
19608 assert(r->check_utf8 == r->float_utf8);
19609 ret->check_substr = ret->float_substr;
19610 ret->check_utf8 = ret->float_utf8;
19612 } else if (ret->check_utf8) {
19614 ret->check_utf8 = ret->anchored_utf8;
19616 ret->check_utf8 = ret->float_utf8;
19621 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
19622 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
19623 if (r->recurse_locinput)
19624 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19627 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
19629 if (RX_MATCH_COPIED(dstr))
19630 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
19632 ret->subbeg = NULL;
19633 #ifdef PERL_ANY_COW
19634 ret->saved_copy = NULL;
19637 /* Whether mother_re be set or no, we need to copy the string. We
19638 cannot refrain from copying it when the storage points directly to
19639 our mother regexp, because that's
19640 1: a buffer in a different thread
19641 2: something we no longer hold a reference on
19642 so we need to copy it locally. */
19643 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
19644 ret->mother_re = NULL;
19646 #endif /* PERL_IN_XSUB_RE */
19651 This is the internal complement to regdupe() which is used to copy
19652 the structure pointed to by the *pprivate pointer in the regexp.
19653 This is the core version of the extension overridable cloning hook.
19654 The regexp structure being duplicated will be copied by perl prior
19655 to this and will be provided as the regexp *r argument, however
19656 with the /old/ structures pprivate pointer value. Thus this routine
19657 may override any copying normally done by perl.
19659 It returns a pointer to the new regexp_internal structure.
19663 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
19666 struct regexp *const r = ReANY(rx);
19667 regexp_internal *reti;
19669 RXi_GET_DECL(r,ri);
19671 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
19675 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
19676 char, regexp_internal);
19677 Copy(ri->program, reti->program, len+1, regnode);
19680 reti->num_code_blocks = ri->num_code_blocks;
19681 if (ri->code_blocks) {
19683 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
19684 struct reg_code_block);
19685 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
19686 struct reg_code_block);
19687 for (n = 0; n < ri->num_code_blocks; n++)
19688 reti->code_blocks[n].src_regex = (REGEXP*)
19689 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
19692 reti->code_blocks = NULL;
19694 reti->regstclass = NULL;
19697 struct reg_data *d;
19698 const int count = ri->data->count;
19701 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
19702 char, struct reg_data);
19703 Newx(d->what, count, U8);
19706 for (i = 0; i < count; i++) {
19707 d->what[i] = ri->data->what[i];
19708 switch (d->what[i]) {
19709 /* see also regcomp.h and regfree_internal() */
19710 case 'a': /* actually an AV, but the dup function is identical. */
19714 case 'u': /* actually an HV, but the dup function is identical. */
19715 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
19718 /* This is cheating. */
19719 Newx(d->data[i], 1, regnode_ssc);
19720 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
19721 reti->regstclass = (regnode*)d->data[i];
19724 /* Trie stclasses are readonly and can thus be shared
19725 * without duplication. We free the stclass in pregfree
19726 * when the corresponding reg_ac_data struct is freed.
19728 reti->regstclass= ri->regstclass;
19732 ((reg_trie_data*)ri->data->data[i])->refcount++;
19737 d->data[i] = ri->data->data[i];
19740 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
19741 ri->data->what[i]);
19750 reti->name_list_idx = ri->name_list_idx;
19752 #ifdef RE_TRACK_PATTERN_OFFSETS
19753 if (ri->u.offsets) {
19754 Newx(reti->u.offsets, 2*len+1, U32);
19755 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
19758 SetProgLen(reti,len);
19761 return (void*)reti;
19764 #endif /* USE_ITHREADS */
19766 #ifndef PERL_IN_XSUB_RE
19769 - regnext - dig the "next" pointer out of a node
19772 Perl_regnext(pTHX_ regnode *p)
19779 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
19780 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
19781 (int)OP(p), (int)REGNODE_MAX);
19784 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
19793 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
19796 STRLEN l1 = strlen(pat1);
19797 STRLEN l2 = strlen(pat2);
19800 const char *message;
19802 PERL_ARGS_ASSERT_RE_CROAK2;
19808 Copy(pat1, buf, l1 , char);
19809 Copy(pat2, buf + l1, l2 , char);
19810 buf[l1 + l2] = '\n';
19811 buf[l1 + l2 + 1] = '\0';
19812 va_start(args, pat2);
19813 msv = vmess(buf, &args);
19815 message = SvPV_const(msv,l1);
19818 Copy(message, buf, l1 , char);
19819 /* l1-1 to avoid \n */
19820 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, l1-1, buf));
19823 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
19825 #ifndef PERL_IN_XSUB_RE
19827 Perl_save_re_context(pTHX)
19832 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
19835 const REGEXP * const rx = PM_GETRE(PL_curpm);
19837 nparens = RX_NPARENS(rx);
19840 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
19841 * that PL_curpm will be null, but that utf8.pm and the modules it
19842 * loads will only use $1..$3.
19843 * The t/porting/re_context.t test file checks this assumption.
19848 for (i = 1; i <= nparens; i++) {
19849 char digits[TYPE_CHARS(long)];
19850 const STRLEN len = my_snprintf(digits, sizeof(digits),
19852 GV *const *const gvp
19853 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
19856 GV * const gv = *gvp;
19857 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
19867 S_put_code_point(pTHX_ SV *sv, UV c)
19869 PERL_ARGS_ASSERT_PUT_CODE_POINT;
19872 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
19874 else if (isPRINT(c)) {
19875 const char string = (char) c;
19877 /* We use {phrase} as metanotation in the class, so also escape literal
19879 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
19880 sv_catpvs(sv, "\\");
19881 sv_catpvn(sv, &string, 1);
19883 else if (isMNEMONIC_CNTRL(c)) {
19884 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
19887 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
19891 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
19894 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
19896 /* Appends to 'sv' a displayable version of the range of code points from
19897 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
19898 * that have them, when they occur at the beginning or end of the range.
19899 * It uses hex to output the remaining code points, unless 'allow_literals'
19900 * is true, in which case the printable ASCII ones are output as-is (though
19901 * some of these will be escaped by put_code_point()).
19903 * NOTE: This is designed only for printing ranges of code points that fit
19904 * inside an ANYOF bitmap. Higher code points are simply suppressed
19907 const unsigned int min_range_count = 3;
19909 assert(start <= end);
19911 PERL_ARGS_ASSERT_PUT_RANGE;
19913 while (start <= end) {
19915 const char * format;
19917 if (end - start < min_range_count) {
19919 /* Output chars individually when they occur in short ranges */
19920 for (; start <= end; start++) {
19921 put_code_point(sv, start);
19926 /* If permitted by the input options, and there is a possibility that
19927 * this range contains a printable literal, look to see if there is
19929 if (allow_literals && start <= MAX_PRINT_A) {
19931 /* If the character at the beginning of the range isn't an ASCII
19932 * printable, effectively split the range into two parts:
19933 * 1) the portion before the first such printable,
19935 * and output them separately. */
19936 if (! isPRINT_A(start)) {
19937 UV temp_end = start + 1;
19939 /* There is no point looking beyond the final possible
19940 * printable, in MAX_PRINT_A */
19941 UV max = MIN(end, MAX_PRINT_A);
19943 while (temp_end <= max && ! isPRINT_A(temp_end)) {
19947 /* Here, temp_end points to one beyond the first printable if
19948 * found, or to one beyond 'max' if not. If none found, make
19949 * sure that we use the entire range */
19950 if (temp_end > MAX_PRINT_A) {
19951 temp_end = end + 1;
19954 /* Output the first part of the split range: the part that
19955 * doesn't have printables, with the parameter set to not look
19956 * for literals (otherwise we would infinitely recurse) */
19957 put_range(sv, start, temp_end - 1, FALSE);
19959 /* The 2nd part of the range (if any) starts here. */
19962 /* We do a continue, instead of dropping down, because even if
19963 * the 2nd part is non-empty, it could be so short that we want
19964 * to output it as individual characters, as tested for at the
19965 * top of this loop. */
19969 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
19970 * output a sub-range of just the digits or letters, then process
19971 * the remaining portion as usual. */
19972 if (isALPHANUMERIC_A(start)) {
19973 UV mask = (isDIGIT_A(start))
19978 UV temp_end = start + 1;
19980 /* Find the end of the sub-range that includes just the
19981 * characters in the same class as the first character in it */
19982 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
19987 /* For short ranges, don't duplicate the code above to output
19988 * them; just call recursively */
19989 if (temp_end - start < min_range_count) {
19990 put_range(sv, start, temp_end, FALSE);
19992 else { /* Output as a range */
19993 put_code_point(sv, start);
19994 sv_catpvs(sv, "-");
19995 put_code_point(sv, temp_end);
19997 start = temp_end + 1;
20001 /* We output any other printables as individual characters */
20002 if (isPUNCT_A(start) || isSPACE_A(start)) {
20003 while (start <= end && (isPUNCT_A(start)
20004 || isSPACE_A(start)))
20006 put_code_point(sv, start);
20011 } /* End of looking for literals */
20013 /* Here is not to output as a literal. Some control characters have
20014 * mnemonic names. Split off any of those at the beginning and end of
20015 * the range to print mnemonically. It isn't possible for many of
20016 * these to be in a row, so this won't overwhelm with output */
20018 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
20020 while (isMNEMONIC_CNTRL(start) && start <= end) {
20021 put_code_point(sv, start);
20025 /* If this didn't take care of the whole range ... */
20026 if (start <= end) {
20028 /* Look backwards from the end to find the final non-mnemonic
20031 while (isMNEMONIC_CNTRL(temp_end)) {
20035 /* And separately output the interior range that doesn't start
20036 * or end with mnemonics */
20037 put_range(sv, start, temp_end, FALSE);
20039 /* Then output the mnemonic trailing controls */
20040 start = temp_end + 1;
20041 while (start <= end) {
20042 put_code_point(sv, start);
20049 /* As a final resort, output the range or subrange as hex. */
20051 this_end = (end < NUM_ANYOF_CODE_POINTS)
20053 : NUM_ANYOF_CODE_POINTS - 1;
20054 #if NUM_ANYOF_CODE_POINTS > 256
20055 format = (this_end < 256)
20056 ? "\\x%02" UVXf "-\\x%02" UVXf
20057 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
20059 format = "\\x%02" UVXf "-\\x%02" UVXf;
20061 GCC_DIAG_IGNORE(-Wformat-nonliteral);
20062 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
20069 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
20071 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
20075 bool allow_literals = TRUE;
20077 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
20079 /* Generally, it is more readable if printable characters are output as
20080 * literals, but if a range (nearly) spans all of them, it's best to output
20081 * it as a single range. This code will use a single range if all but 2
20082 * ASCII printables are in it */
20083 invlist_iterinit(invlist);
20084 while (invlist_iternext(invlist, &start, &end)) {
20086 /* If the range starts beyond the final printable, it doesn't have any
20088 if (start > MAX_PRINT_A) {
20092 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
20093 * all but two, the range must start and end no later than 2 from
20095 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
20096 if (end > MAX_PRINT_A) {
20102 if (end - start >= MAX_PRINT_A - ' ' - 2) {
20103 allow_literals = FALSE;
20108 invlist_iterfinish(invlist);
20110 /* Here we have figured things out. Output each range */
20111 invlist_iterinit(invlist);
20112 while (invlist_iternext(invlist, &start, &end)) {
20113 if (start >= NUM_ANYOF_CODE_POINTS) {
20116 put_range(sv, start, end, allow_literals);
20118 invlist_iterfinish(invlist);
20124 S_put_charclass_bitmap_innards_common(pTHX_
20125 SV* invlist, /* The bitmap */
20126 SV* posixes, /* Under /l, things like [:word:], \S */
20127 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
20128 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
20129 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
20130 const bool invert /* Is the result to be inverted? */
20133 /* Create and return an SV containing a displayable version of the bitmap
20134 * and associated information determined by the input parameters. If the
20135 * output would have been only the inversion indicator '^', NULL is instead
20140 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
20143 output = newSVpvs("^");
20146 output = newSVpvs("");
20149 /* First, the code points in the bitmap that are unconditionally there */
20150 put_charclass_bitmap_innards_invlist(output, invlist);
20152 /* Traditionally, these have been placed after the main code points */
20154 sv_catsv(output, posixes);
20157 if (only_utf8 && _invlist_len(only_utf8)) {
20158 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
20159 put_charclass_bitmap_innards_invlist(output, only_utf8);
20162 if (not_utf8 && _invlist_len(not_utf8)) {
20163 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
20164 put_charclass_bitmap_innards_invlist(output, not_utf8);
20167 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
20168 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
20169 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
20171 /* This is the only list in this routine that can legally contain code
20172 * points outside the bitmap range. The call just above to
20173 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
20174 * output them here. There's about a half-dozen possible, and none in
20175 * contiguous ranges longer than 2 */
20176 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20178 SV* above_bitmap = NULL;
20180 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
20182 invlist_iterinit(above_bitmap);
20183 while (invlist_iternext(above_bitmap, &start, &end)) {
20186 for (i = start; i <= end; i++) {
20187 put_code_point(output, i);
20190 invlist_iterfinish(above_bitmap);
20191 SvREFCNT_dec_NN(above_bitmap);
20195 if (invert && SvCUR(output) == 1) {
20203 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
20205 SV *nonbitmap_invlist,
20206 SV *only_utf8_locale_invlist,
20207 const regnode * const node,
20208 const bool force_as_is_display)
20210 /* Appends to 'sv' a displayable version of the innards of the bracketed
20211 * character class defined by the other arguments:
20212 * 'bitmap' points to the bitmap.
20213 * 'nonbitmap_invlist' is an inversion list of the code points that are in
20214 * the bitmap range, but for some reason aren't in the bitmap; NULL if
20215 * none. The reasons for this could be that they require some
20216 * condition such as the target string being or not being in UTF-8
20217 * (under /d), or because they came from a user-defined property that
20218 * was not resolved at the time of the regex compilation (under /u)
20219 * 'only_utf8_locale_invlist' is an inversion list of the code points that
20220 * are valid only if the runtime locale is a UTF-8 one; NULL if none
20221 * 'node' is the regex pattern node. It is needed only when the above two
20222 * parameters are not null, and is passed so that this routine can
20223 * tease apart the various reasons for them.
20224 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
20225 * to invert things to see if that leads to a cleaner display. If
20226 * FALSE, this routine is free to use its judgment about doing this.
20228 * It returns TRUE if there was actually something output. (It may be that
20229 * the bitmap, etc is empty.)
20231 * When called for outputting the bitmap of a non-ANYOF node, just pass the
20232 * bitmap, with the succeeding parameters set to NULL, and the final one to
20236 /* In general, it tries to display the 'cleanest' representation of the
20237 * innards, choosing whether to display them inverted or not, regardless of
20238 * whether the class itself is to be inverted. However, there are some
20239 * cases where it can't try inverting, as what actually matches isn't known
20240 * until runtime, and hence the inversion isn't either. */
20241 bool inverting_allowed = ! force_as_is_display;
20244 STRLEN orig_sv_cur = SvCUR(sv);
20246 SV* invlist; /* Inversion list we accumulate of code points that
20247 are unconditionally matched */
20248 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
20250 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
20252 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
20253 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
20256 SV* as_is_display; /* The output string when we take the inputs
20258 SV* inverted_display; /* The output string when we invert the inputs */
20260 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
20262 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
20264 /* We are biased in favor of displaying things without them being inverted,
20265 * as that is generally easier to understand */
20266 const int bias = 5;
20268 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
20270 /* Start off with whatever code points are passed in. (We clone, so we
20271 * don't change the caller's list) */
20272 if (nonbitmap_invlist) {
20273 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
20274 invlist = invlist_clone(nonbitmap_invlist);
20276 else { /* Worst case size is every other code point is matched */
20277 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
20281 if (OP(node) == ANYOFD) {
20283 /* This flag indicates that the code points below 0x100 in the
20284 * nonbitmap list are precisely the ones that match only when the
20285 * target is UTF-8 (they should all be non-ASCII). */
20286 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
20288 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
20289 _invlist_subtract(invlist, only_utf8, &invlist);
20292 /* And this flag for matching all non-ASCII 0xFF and below */
20293 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
20295 not_utf8 = invlist_clone(PL_UpperLatin1);
20298 else if (OP(node) == ANYOFL) {
20300 /* If either of these flags are set, what matches isn't
20301 * determinable except during execution, so don't know enough here
20303 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
20304 inverting_allowed = FALSE;
20307 /* What the posix classes match also varies at runtime, so these
20308 * will be output symbolically. */
20309 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
20312 posixes = newSVpvs("");
20313 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
20314 if (ANYOF_POSIXL_TEST(node,i)) {
20315 sv_catpv(posixes, anyofs[i]);
20322 /* Accumulate the bit map into the unconditional match list */
20323 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
20324 if (BITMAP_TEST(bitmap, i)) {
20326 for (; i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i); i++) {
20329 invlist = _add_range_to_invlist(invlist, start, i-1);
20333 /* Make sure that the conditional match lists don't have anything in them
20334 * that match unconditionally; otherwise the output is quite confusing.
20335 * This could happen if the code that populates these misses some
20338 _invlist_subtract(only_utf8, invlist, &only_utf8);
20341 _invlist_subtract(not_utf8, invlist, ¬_utf8);
20344 if (only_utf8_locale_invlist) {
20346 /* Since this list is passed in, we have to make a copy before
20348 only_utf8_locale = invlist_clone(only_utf8_locale_invlist);
20350 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
20352 /* And, it can get really weird for us to try outputting an inverted
20353 * form of this list when it has things above the bitmap, so don't even
20355 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20356 inverting_allowed = FALSE;
20360 /* Calculate what the output would be if we take the input as-is */
20361 as_is_display = put_charclass_bitmap_innards_common(invlist,
20368 /* If have to take the output as-is, just do that */
20369 if (! inverting_allowed) {
20370 if (as_is_display) {
20371 sv_catsv(sv, as_is_display);
20372 SvREFCNT_dec_NN(as_is_display);
20375 else { /* But otherwise, create the output again on the inverted input, and
20376 use whichever version is shorter */
20378 int inverted_bias, as_is_bias;
20380 /* We will apply our bias to whichever of the the results doesn't have
20390 inverted_bias = bias;
20393 /* Now invert each of the lists that contribute to the output,
20394 * excluding from the result things outside the possible range */
20396 /* For the unconditional inversion list, we have to add in all the
20397 * conditional code points, so that when inverted, they will be gone
20399 _invlist_union(only_utf8, invlist, &invlist);
20400 _invlist_union(not_utf8, invlist, &invlist);
20401 _invlist_union(only_utf8_locale, invlist, &invlist);
20402 _invlist_invert(invlist);
20403 _invlist_intersection(invlist, PL_InBitmap, &invlist);
20406 _invlist_invert(only_utf8);
20407 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
20409 else if (not_utf8) {
20411 /* If a code point matches iff the target string is not in UTF-8,
20412 * then complementing the result has it not match iff not in UTF-8,
20413 * which is the same thing as matching iff it is UTF-8. */
20414 only_utf8 = not_utf8;
20418 if (only_utf8_locale) {
20419 _invlist_invert(only_utf8_locale);
20420 _invlist_intersection(only_utf8_locale,
20422 &only_utf8_locale);
20425 inverted_display = put_charclass_bitmap_innards_common(
20430 only_utf8_locale, invert);
20432 /* Use the shortest representation, taking into account our bias
20433 * against showing it inverted */
20434 if ( inverted_display
20435 && ( ! as_is_display
20436 || ( SvCUR(inverted_display) + inverted_bias
20437 < SvCUR(as_is_display) + as_is_bias)))
20439 sv_catsv(sv, inverted_display);
20441 else if (as_is_display) {
20442 sv_catsv(sv, as_is_display);
20445 SvREFCNT_dec(as_is_display);
20446 SvREFCNT_dec(inverted_display);
20449 SvREFCNT_dec_NN(invlist);
20450 SvREFCNT_dec(only_utf8);
20451 SvREFCNT_dec(not_utf8);
20452 SvREFCNT_dec(posixes);
20453 SvREFCNT_dec(only_utf8_locale);
20455 return SvCUR(sv) > orig_sv_cur;
20458 #define CLEAR_OPTSTART \
20459 if (optstart) STMT_START { \
20460 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
20461 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
20465 #define DUMPUNTIL(b,e) \
20467 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
20469 STATIC const regnode *
20470 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
20471 const regnode *last, const regnode *plast,
20472 SV* sv, I32 indent, U32 depth)
20474 U8 op = PSEUDO; /* Arbitrary non-END op. */
20475 const regnode *next;
20476 const regnode *optstart= NULL;
20478 RXi_GET_DECL(r,ri);
20479 GET_RE_DEBUG_FLAGS_DECL;
20481 PERL_ARGS_ASSERT_DUMPUNTIL;
20483 #ifdef DEBUG_DUMPUNTIL
20484 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n",indent,node-start,
20485 last ? last-start : 0,plast ? plast-start : 0);
20488 if (plast && plast < last)
20491 while (PL_regkind[op] != END && (!last || node < last)) {
20493 /* While that wasn't END last time... */
20496 if (op == CLOSE || op == WHILEM)
20498 next = regnext((regnode *)node);
20501 if (OP(node) == OPTIMIZED) {
20502 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
20509 regprop(r, sv, node, NULL, NULL);
20510 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
20511 (int)(2*indent + 1), "", SvPVX_const(sv));
20513 if (OP(node) != OPTIMIZED) {
20514 if (next == NULL) /* Next ptr. */
20515 Perl_re_printf( aTHX_ " (0)");
20516 else if (PL_regkind[(U8)op] == BRANCH
20517 && PL_regkind[OP(next)] != BRANCH )
20518 Perl_re_printf( aTHX_ " (FAIL)");
20520 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
20521 Perl_re_printf( aTHX_ "\n");
20525 if (PL_regkind[(U8)op] == BRANCHJ) {
20528 const regnode *nnode = (OP(next) == LONGJMP
20529 ? regnext((regnode *)next)
20531 if (last && nnode > last)
20533 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
20536 else if (PL_regkind[(U8)op] == BRANCH) {
20538 DUMPUNTIL(NEXTOPER(node), next);
20540 else if ( PL_regkind[(U8)op] == TRIE ) {
20541 const regnode *this_trie = node;
20542 const char op = OP(node);
20543 const U32 n = ARG(node);
20544 const reg_ac_data * const ac = op>=AHOCORASICK ?
20545 (reg_ac_data *)ri->data->data[n] :
20547 const reg_trie_data * const trie =
20548 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
20550 AV *const trie_words
20551 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
20553 const regnode *nextbranch= NULL;
20556 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
20557 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
20559 Perl_re_indentf( aTHX_ "%s ",
20562 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
20563 SvCUR(*elem_ptr), 60,
20564 PL_colors[0], PL_colors[1],
20566 ? PERL_PV_ESCAPE_UNI
20568 | PERL_PV_PRETTY_ELLIPSES
20569 | PERL_PV_PRETTY_LTGT
20574 U16 dist= trie->jump[word_idx+1];
20575 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
20576 (UV)((dist ? this_trie + dist : next) - start));
20579 nextbranch= this_trie + trie->jump[0];
20580 DUMPUNTIL(this_trie + dist, nextbranch);
20582 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
20583 nextbranch= regnext((regnode *)nextbranch);
20585 Perl_re_printf( aTHX_ "\n");
20588 if (last && next > last)
20593 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
20594 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
20595 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
20597 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
20599 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
20601 else if ( op == PLUS || op == STAR) {
20602 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
20604 else if (PL_regkind[(U8)op] == ANYOF) {
20605 /* arglen 1 + class block */
20606 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
20607 ? ANYOF_POSIXL_SKIP
20609 node = NEXTOPER(node);
20611 else if (PL_regkind[(U8)op] == EXACT) {
20612 /* Literal string, where present. */
20613 node += NODE_SZ_STR(node) - 1;
20614 node = NEXTOPER(node);
20617 node = NEXTOPER(node);
20618 node += regarglen[(U8)op];
20620 if (op == CURLYX || op == OPEN)
20624 #ifdef DEBUG_DUMPUNTIL
20625 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
20630 #endif /* DEBUGGING */
20633 * ex: set ts=8 sts=4 sw=4 et: