5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
84 EXTERN_C const struct regexp_engine my_reg_engine;
89 #include "dquote_inline.h"
90 #include "invlist_inline.h"
91 #include "unicode_constants.h"
93 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
97 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
98 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
101 #define STATIC static
104 /* this is a chain of data about sub patterns we are processing that
105 need to be handled separately/specially in study_chunk. Its so
106 we can simulate recursion without losing state. */
108 typedef struct scan_frame {
109 regnode *last_regnode; /* last node to process in this frame */
110 regnode *next_regnode; /* next node to process when last is reached */
111 U32 prev_recursed_depth;
112 I32 stopparen; /* what stopparen do we use */
113 U32 is_top_frame; /* what flags do we use? */
115 struct scan_frame *this_prev_frame; /* this previous frame */
116 struct scan_frame *prev_frame; /* previous frame */
117 struct scan_frame *next_frame; /* next frame */
120 /* Certain characters are output as a sequence with the first being a
122 #define isBACKSLASHED_PUNCT(c) \
123 ((c) == '-' || (c) == ']' || (c) == '\\' || (c) == '^')
126 struct RExC_state_t {
127 U32 flags; /* RXf_* are we folding, multilining? */
128 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
129 char *precomp; /* uncompiled string. */
130 char *precomp_end; /* pointer to end of uncompiled string. */
131 REGEXP *rx_sv; /* The SV that is the regexp. */
132 regexp *rx; /* perl core regexp structure */
133 regexp_internal *rxi; /* internal data for regexp object
135 char *start; /* Start of input for compile */
136 char *end; /* End of input for compile */
137 char *parse; /* Input-scan pointer. */
138 char *adjusted_start; /* 'start', adjusted. See code use */
139 STRLEN precomp_adj; /* an offset beyond precomp. See code use */
140 SSize_t whilem_seen; /* number of WHILEM in this expr */
141 regnode *emit_start; /* Start of emitted-code area */
142 regnode *emit_bound; /* First regnode outside of the
144 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
145 implies compiling, so don't emit */
146 regnode_ssc emit_dummy; /* placeholder for emit to point to;
147 large enough for the largest
148 non-EXACTish node, so can use it as
150 I32 naughty; /* How bad is this pattern? */
151 I32 sawback; /* Did we see \1, ...? */
153 SSize_t size; /* Code size. */
154 I32 npar; /* Capture buffer count, (OPEN) plus
155 one. ("par" 0 is the whole
157 I32 nestroot; /* root parens we are in - used by
161 regnode **open_parens; /* pointers to open parens */
162 regnode **close_parens; /* pointers to close parens */
163 regnode *end_op; /* END node in program */
164 I32 utf8; /* whether the pattern is utf8 or not */
165 I32 orig_utf8; /* whether the pattern was originally in utf8 */
166 /* XXX use this for future optimisation of case
167 * where pattern must be upgraded to utf8. */
168 I32 uni_semantics; /* If a d charset modifier should use unicode
169 rules, even if the pattern is not in
171 HV *paren_names; /* Paren names */
173 regnode **recurse; /* Recurse regops */
174 I32 recurse_count; /* Number of recurse regops we have generated */
175 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
177 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
180 I32 override_recoding;
182 I32 recode_x_to_native;
184 I32 in_multi_char_class;
185 struct reg_code_block *code_blocks; /* positions of literal (?{})
187 int num_code_blocks; /* size of code_blocks[] */
188 int code_index; /* next code_blocks[] slot */
189 SSize_t maxlen; /* mininum possible number of chars in string to match */
190 scan_frame *frame_head;
191 scan_frame *frame_last;
194 #ifdef ADD_TO_REGEXEC
195 char *starttry; /* -Dr: where regtry was called. */
196 #define RExC_starttry (pRExC_state->starttry)
198 SV *runtime_code_qr; /* qr with the runtime code blocks */
200 const char *lastparse;
202 AV *paren_name_list; /* idx -> name */
203 U32 study_chunk_recursed_count;
206 #define RExC_lastparse (pRExC_state->lastparse)
207 #define RExC_lastnum (pRExC_state->lastnum)
208 #define RExC_paren_name_list (pRExC_state->paren_name_list)
209 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
210 #define RExC_mysv (pRExC_state->mysv1)
211 #define RExC_mysv1 (pRExC_state->mysv1)
212 #define RExC_mysv2 (pRExC_state->mysv2)
215 bool seen_unfolded_sharp_s;
220 #define RExC_flags (pRExC_state->flags)
221 #define RExC_pm_flags (pRExC_state->pm_flags)
222 #define RExC_precomp (pRExC_state->precomp)
223 #define RExC_precomp_adj (pRExC_state->precomp_adj)
224 #define RExC_adjusted_start (pRExC_state->adjusted_start)
225 #define RExC_precomp_end (pRExC_state->precomp_end)
226 #define RExC_rx_sv (pRExC_state->rx_sv)
227 #define RExC_rx (pRExC_state->rx)
228 #define RExC_rxi (pRExC_state->rxi)
229 #define RExC_start (pRExC_state->start)
230 #define RExC_end (pRExC_state->end)
231 #define RExC_parse (pRExC_state->parse)
232 #define RExC_whilem_seen (pRExC_state->whilem_seen)
234 /* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
235 * EXACTF node, hence was parsed under /di rules. If later in the parse,
236 * something forces the pattern into using /ui rules, the sharp s should be
237 * folded into the sequence 'ss', which takes up more space than previously
238 * calculated. This means that the sizing pass needs to be restarted. (The
239 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
240 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
241 * so there is no need to resize [perl #125990]. */
242 #define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
244 #ifdef RE_TRACK_PATTERN_OFFSETS
245 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
248 #define RExC_emit (pRExC_state->emit)
249 #define RExC_emit_dummy (pRExC_state->emit_dummy)
250 #define RExC_emit_start (pRExC_state->emit_start)
251 #define RExC_emit_bound (pRExC_state->emit_bound)
252 #define RExC_sawback (pRExC_state->sawback)
253 #define RExC_seen (pRExC_state->seen)
254 #define RExC_size (pRExC_state->size)
255 #define RExC_maxlen (pRExC_state->maxlen)
256 #define RExC_npar (pRExC_state->npar)
257 #define RExC_nestroot (pRExC_state->nestroot)
258 #define RExC_extralen (pRExC_state->extralen)
259 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
260 #define RExC_utf8 (pRExC_state->utf8)
261 #define RExC_uni_semantics (pRExC_state->uni_semantics)
262 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
263 #define RExC_open_parens (pRExC_state->open_parens)
264 #define RExC_close_parens (pRExC_state->close_parens)
265 #define RExC_end_op (pRExC_state->end_op)
266 #define RExC_paren_names (pRExC_state->paren_names)
267 #define RExC_recurse (pRExC_state->recurse)
268 #define RExC_recurse_count (pRExC_state->recurse_count)
269 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
270 #define RExC_study_chunk_recursed_bytes \
271 (pRExC_state->study_chunk_recursed_bytes)
272 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
273 #define RExC_contains_locale (pRExC_state->contains_locale)
275 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
277 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
278 #define RExC_frame_head (pRExC_state->frame_head)
279 #define RExC_frame_last (pRExC_state->frame_last)
280 #define RExC_frame_count (pRExC_state->frame_count)
281 #define RExC_strict (pRExC_state->strict)
282 #define RExC_study_started (pRExC_state->study_started)
283 #define RExC_warn_text (pRExC_state->warn_text)
285 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
286 * a flag to disable back-off on the fixed/floating substrings - if it's
287 * a high complexity pattern we assume the benefit of avoiding a full match
288 * is worth the cost of checking for the substrings even if they rarely help.
290 #define RExC_naughty (pRExC_state->naughty)
291 #define TOO_NAUGHTY (10)
292 #define MARK_NAUGHTY(add) \
293 if (RExC_naughty < TOO_NAUGHTY) \
294 RExC_naughty += (add)
295 #define MARK_NAUGHTY_EXP(exp, add) \
296 if (RExC_naughty < TOO_NAUGHTY) \
297 RExC_naughty += RExC_naughty / (exp) + (add)
299 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
300 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
301 ((*s) == '{' && regcurly(s)))
304 * Flags to be passed up and down.
306 #define WORST 0 /* Worst case. */
307 #define HASWIDTH 0x01 /* Known to match non-null strings. */
309 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
310 * character. (There needs to be a case: in the switch statement in regexec.c
311 * for any node marked SIMPLE.) Note that this is not the same thing as
314 #define SPSTART 0x04 /* Starts with * or + */
315 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
316 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
317 #define RESTART_PASS1 0x20 /* Need to restart sizing pass */
318 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
319 calcuate sizes as UTF-8 */
321 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
323 /* whether trie related optimizations are enabled */
324 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
325 #define TRIE_STUDY_OPT
326 #define FULL_TRIE_STUDY
332 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
333 #define PBITVAL(paren) (1 << ((paren) & 7))
334 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
335 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
336 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
338 #define REQUIRE_UTF8(flagp) STMT_START { \
341 *flagp = RESTART_PASS1|NEED_UTF8; \
346 /* Change from /d into /u rules, and restart the parse if we've already seen
347 * something whose size would increase as a result, by setting *flagp and
348 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
349 * we've change to /u during the parse. */
350 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
352 if (DEPENDS_SEMANTICS) { \
354 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
355 RExC_uni_semantics = 1; \
356 if (RExC_seen_unfolded_sharp_s) { \
357 *flagp |= RESTART_PASS1; \
358 return restart_retval; \
363 /* This converts the named class defined in regcomp.h to its equivalent class
364 * number defined in handy.h. */
365 #define namedclass_to_classnum(class) ((int) ((class) / 2))
366 #define classnum_to_namedclass(classnum) ((classnum) * 2)
368 #define _invlist_union_complement_2nd(a, b, output) \
369 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
370 #define _invlist_intersection_complement_2nd(a, b, output) \
371 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
373 /* About scan_data_t.
375 During optimisation we recurse through the regexp program performing
376 various inplace (keyhole style) optimisations. In addition study_chunk
377 and scan_commit populate this data structure with information about
378 what strings MUST appear in the pattern. We look for the longest
379 string that must appear at a fixed location, and we look for the
380 longest string that may appear at a floating location. So for instance
385 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
386 strings (because they follow a .* construct). study_chunk will identify
387 both FOO and BAR as being the longest fixed and floating strings respectively.
389 The strings can be composites, for instance
393 will result in a composite fixed substring 'foo'.
395 For each string some basic information is maintained:
397 - offset or min_offset
398 This is the position the string must appear at, or not before.
399 It also implicitly (when combined with minlenp) tells us how many
400 characters must match before the string we are searching for.
401 Likewise when combined with minlenp and the length of the string it
402 tells us how many characters must appear after the string we have
406 Only used for floating strings. This is the rightmost point that
407 the string can appear at. If set to SSize_t_MAX it indicates that the
408 string can occur infinitely far to the right.
411 A pointer to the minimum number of characters of the pattern that the
412 string was found inside. This is important as in the case of positive
413 lookahead or positive lookbehind we can have multiple patterns
418 The minimum length of the pattern overall is 3, the minimum length
419 of the lookahead part is 3, but the minimum length of the part that
420 will actually match is 1. So 'FOO's minimum length is 3, but the
421 minimum length for the F is 1. This is important as the minimum length
422 is used to determine offsets in front of and behind the string being
423 looked for. Since strings can be composites this is the length of the
424 pattern at the time it was committed with a scan_commit. Note that
425 the length is calculated by study_chunk, so that the minimum lengths
426 are not known until the full pattern has been compiled, thus the
427 pointer to the value.
431 In the case of lookbehind the string being searched for can be
432 offset past the start point of the final matching string.
433 If this value was just blithely removed from the min_offset it would
434 invalidate some of the calculations for how many chars must match
435 before or after (as they are derived from min_offset and minlen and
436 the length of the string being searched for).
437 When the final pattern is compiled and the data is moved from the
438 scan_data_t structure into the regexp structure the information
439 about lookbehind is factored in, with the information that would
440 have been lost precalculated in the end_shift field for the
443 The fields pos_min and pos_delta are used to store the minimum offset
444 and the delta to the maximum offset at the current point in the pattern.
448 typedef struct scan_data_t {
449 /*I32 len_min; unused */
450 /*I32 len_delta; unused */
454 SSize_t last_end; /* min value, <0 unless valid. */
455 SSize_t last_start_min;
456 SSize_t last_start_max;
457 SV **longest; /* Either &l_fixed, or &l_float. */
458 SV *longest_fixed; /* longest fixed string found in pattern */
459 SSize_t offset_fixed; /* offset where it starts */
460 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
461 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
462 SV *longest_float; /* longest floating string found in pattern */
463 SSize_t offset_float_min; /* earliest point in string it can appear */
464 SSize_t offset_float_max; /* latest point in string it can appear */
465 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
466 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
469 SSize_t *last_closep;
470 regnode_ssc *start_class;
474 * Forward declarations for pregcomp()'s friends.
477 static const scan_data_t zero_scan_data =
478 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
480 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
481 #define SF_BEFORE_SEOL 0x0001
482 #define SF_BEFORE_MEOL 0x0002
483 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
484 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
486 #define SF_FIX_SHIFT_EOL (+2)
487 #define SF_FL_SHIFT_EOL (+4)
489 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
490 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
492 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
493 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
494 #define SF_IS_INF 0x0040
495 #define SF_HAS_PAR 0x0080
496 #define SF_IN_PAR 0x0100
497 #define SF_HAS_EVAL 0x0200
500 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
501 * longest substring in the pattern. When it is not set the optimiser keeps
502 * track of position, but does not keep track of the actual strings seen,
504 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
507 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
508 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
509 * turned off because of the alternation (BRANCH). */
510 #define SCF_DO_SUBSTR 0x0400
512 #define SCF_DO_STCLASS_AND 0x0800
513 #define SCF_DO_STCLASS_OR 0x1000
514 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
515 #define SCF_WHILEM_VISITED_POS 0x2000
517 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
518 #define SCF_SEEN_ACCEPT 0x8000
519 #define SCF_TRIE_DOING_RESTUDY 0x10000
520 #define SCF_IN_DEFINE 0x20000
525 #define UTF cBOOL(RExC_utf8)
527 /* The enums for all these are ordered so things work out correctly */
528 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
529 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
530 == REGEX_DEPENDS_CHARSET)
531 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
532 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
533 >= REGEX_UNICODE_CHARSET)
534 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
535 == REGEX_ASCII_RESTRICTED_CHARSET)
536 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
537 >= REGEX_ASCII_RESTRICTED_CHARSET)
538 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
539 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
541 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
543 /* For programs that want to be strictly Unicode compatible by dying if any
544 * attempt is made to match a non-Unicode code point against a Unicode
546 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
548 #define OOB_NAMEDCLASS -1
550 /* There is no code point that is out-of-bounds, so this is problematic. But
551 * its only current use is to initialize a variable that is always set before
553 #define OOB_UNICODE 0xDEADBEEF
555 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
558 /* length of regex to show in messages that don't mark a position within */
559 #define RegexLengthToShowInErrorMessages 127
562 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
563 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
564 * op/pragma/warn/regcomp.
566 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
567 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
569 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
570 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
572 /* The code in this file in places uses one level of recursion with parsing
573 * rebased to an alternate string constructed by us in memory. This can take
574 * the form of something that is completely different from the input, or
575 * something that uses the input as part of the alternate. In the first case,
576 * there should be no possibility of an error, as we are in complete control of
577 * the alternate string. But in the second case we don't control the input
578 * portion, so there may be errors in that. Here's an example:
580 * is handled specially because \x{df} folds to a sequence of more than one
581 * character, 'ss'. What is done is to create and parse an alternate string,
582 * which looks like this:
583 * /(?:\x{DF}|[abc\x{DF}def])/ui
584 * where it uses the input unchanged in the middle of something it constructs,
585 * which is a branch for the DF outside the character class, and clustering
586 * parens around the whole thing. (It knows enough to skip the DF inside the
587 * class while in this substitute parse.) 'abc' and 'def' may have errors that
588 * need to be reported. The general situation looks like this:
591 * Input: ----------------------------------------------------
592 * Constructed: ---------------------------------------------------
595 * The input string sI..eI is the input pattern. The string sC..EC is the
596 * constructed substitute parse string. The portions sC..tC and eC..EC are
597 * constructed by us. The portion tC..eC is an exact duplicate of the input
598 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
599 * while parsing, we find an error at xC. We want to display a message showing
600 * the real input string. Thus we need to find the point xI in it which
601 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
602 * been constructed by us, and so shouldn't have errors. We get:
604 * xI = sI + (tI - sI) + (xC - tC)
606 * and, the offset into sI is:
608 * (xI - sI) = (tI - sI) + (xC - tC)
610 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
611 * and we save tC as RExC_adjusted_start.
613 * During normal processing of the input pattern, everything points to that,
614 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
617 #define tI_sI RExC_precomp_adj
618 #define tC RExC_adjusted_start
619 #define sC RExC_precomp
620 #define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
621 #define xI(xC) (sC + xI_offset(xC))
622 #define eC RExC_precomp_end
624 #define REPORT_LOCATION_ARGS(xC) \
626 (xI(xC) > eC) /* Don't run off end */ \
627 ? eC - sC /* Length before the <--HERE */ \
629 sC), /* The input pattern printed up to the <--HERE */ \
631 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
632 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
634 /* Used to point after bad bytes for an error message, but avoid skipping
635 * past a nul byte. */
636 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
639 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
640 * arg. Show regex, up to a maximum length. If it's too long, chop and add
643 #define _FAIL(code) STMT_START { \
644 const char *ellipses = ""; \
645 IV len = RExC_precomp_end - RExC_precomp; \
648 SAVEFREESV(RExC_rx_sv); \
649 if (len > RegexLengthToShowInErrorMessages) { \
650 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
651 len = RegexLengthToShowInErrorMessages - 10; \
657 #define FAIL(msg) _FAIL( \
658 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
659 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
661 #define FAIL2(msg,arg) _FAIL( \
662 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
663 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
666 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
668 #define Simple_vFAIL(m) STMT_START { \
669 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
670 m, REPORT_LOCATION_ARGS(RExC_parse)); \
674 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
676 #define vFAIL(m) STMT_START { \
678 SAVEFREESV(RExC_rx_sv); \
683 * Like Simple_vFAIL(), but accepts two arguments.
685 #define Simple_vFAIL2(m,a1) STMT_START { \
686 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
687 REPORT_LOCATION_ARGS(RExC_parse)); \
691 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
693 #define vFAIL2(m,a1) STMT_START { \
695 SAVEFREESV(RExC_rx_sv); \
696 Simple_vFAIL2(m, a1); \
701 * Like Simple_vFAIL(), but accepts three arguments.
703 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
704 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
705 REPORT_LOCATION_ARGS(RExC_parse)); \
709 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
711 #define vFAIL3(m,a1,a2) STMT_START { \
713 SAVEFREESV(RExC_rx_sv); \
714 Simple_vFAIL3(m, a1, a2); \
718 * Like Simple_vFAIL(), but accepts four arguments.
720 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
721 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
722 REPORT_LOCATION_ARGS(RExC_parse)); \
725 #define vFAIL4(m,a1,a2,a3) STMT_START { \
727 SAVEFREESV(RExC_rx_sv); \
728 Simple_vFAIL4(m, a1, a2, a3); \
731 /* A specialized version of vFAIL2 that works with UTF8f */
732 #define vFAIL2utf8f(m, a1) STMT_START { \
734 SAVEFREESV(RExC_rx_sv); \
735 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
736 REPORT_LOCATION_ARGS(RExC_parse)); \
739 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
741 SAVEFREESV(RExC_rx_sv); \
742 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
743 REPORT_LOCATION_ARGS(RExC_parse)); \
746 /* These have asserts in them because of [perl #122671] Many warnings in
747 * regcomp.c can occur twice. If they get output in pass1 and later in that
748 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
749 * would get output again. So they should be output in pass2, and these
750 * asserts make sure new warnings follow that paradigm. */
752 /* m is not necessarily a "literal string", in this macro */
753 #define reg_warn_non_literal_string(loc, m) STMT_START { \
754 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
755 "%s" REPORT_LOCATION, \
756 m, REPORT_LOCATION_ARGS(loc)); \
759 #define ckWARNreg(loc,m) STMT_START { \
760 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
762 REPORT_LOCATION_ARGS(loc)); \
765 #define vWARN(loc, m) STMT_START { \
766 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
768 REPORT_LOCATION_ARGS(loc)); \
771 #define vWARN_dep(loc, m) STMT_START { \
772 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
774 REPORT_LOCATION_ARGS(loc)); \
777 #define ckWARNdep(loc,m) STMT_START { \
778 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
780 REPORT_LOCATION_ARGS(loc)); \
783 #define ckWARNregdep(loc,m) STMT_START { \
784 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
787 REPORT_LOCATION_ARGS(loc)); \
790 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
791 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
793 a1, REPORT_LOCATION_ARGS(loc)); \
796 #define ckWARN2reg(loc, m, a1) STMT_START { \
797 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
799 a1, REPORT_LOCATION_ARGS(loc)); \
802 #define vWARN3(loc, m, a1, a2) STMT_START { \
803 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
805 a1, a2, REPORT_LOCATION_ARGS(loc)); \
808 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
809 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
812 REPORT_LOCATION_ARGS(loc)); \
815 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
816 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
819 REPORT_LOCATION_ARGS(loc)); \
822 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
823 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
826 REPORT_LOCATION_ARGS(loc)); \
829 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
830 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
833 REPORT_LOCATION_ARGS(loc)); \
836 /* Macros for recording node offsets. 20001227 mjd@plover.com
837 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
838 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
839 * Element 0 holds the number n.
840 * Position is 1 indexed.
842 #ifndef RE_TRACK_PATTERN_OFFSETS
843 #define Set_Node_Offset_To_R(node,byte)
844 #define Set_Node_Offset(node,byte)
845 #define Set_Cur_Node_Offset
846 #define Set_Node_Length_To_R(node,len)
847 #define Set_Node_Length(node,len)
848 #define Set_Node_Cur_Length(node,start)
849 #define Node_Offset(n)
850 #define Node_Length(n)
851 #define Set_Node_Offset_Length(node,offset,len)
852 #define ProgLen(ri) ri->u.proglen
853 #define SetProgLen(ri,x) ri->u.proglen = x
855 #define ProgLen(ri) ri->u.offsets[0]
856 #define SetProgLen(ri,x) ri->u.offsets[0] = x
857 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
859 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
860 __LINE__, (int)(node), (int)(byte))); \
862 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
865 RExC_offsets[2*(node)-1] = (byte); \
870 #define Set_Node_Offset(node,byte) \
871 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
872 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
874 #define Set_Node_Length_To_R(node,len) STMT_START { \
876 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
877 __LINE__, (int)(node), (int)(len))); \
879 Perl_croak(aTHX_ "value of node is %d in Length macro", \
882 RExC_offsets[2*(node)] = (len); \
887 #define Set_Node_Length(node,len) \
888 Set_Node_Length_To_R((node)-RExC_emit_start, len)
889 #define Set_Node_Cur_Length(node, start) \
890 Set_Node_Length(node, RExC_parse - start)
892 /* Get offsets and lengths */
893 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
894 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
896 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
897 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
898 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
902 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
903 #define EXPERIMENTAL_INPLACESCAN
904 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
908 Perl_re_printf(pTHX_ const char *fmt, ...)
912 PerlIO *f= Perl_debug_log;
913 PERL_ARGS_ASSERT_RE_PRINTF;
915 result = PerlIO_vprintf(f, fmt, ap);
921 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
925 PerlIO *f= Perl_debug_log;
926 PERL_ARGS_ASSERT_RE_INDENTF;
928 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
929 result = PerlIO_vprintf(f, fmt, ap);
933 #endif /* DEBUGGING */
935 #define DEBUG_RExC_seen() \
936 DEBUG_OPTIMISE_MORE_r({ \
937 Perl_re_printf( aTHX_ "RExC_seen: "); \
939 if (RExC_seen & REG_ZERO_LEN_SEEN) \
940 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
942 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
943 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
945 if (RExC_seen & REG_GPOS_SEEN) \
946 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
948 if (RExC_seen & REG_RECURSE_SEEN) \
949 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
951 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
952 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
954 if (RExC_seen & REG_VERBARG_SEEN) \
955 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
957 if (RExC_seen & REG_CUTGROUP_SEEN) \
958 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
960 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
961 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
963 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
964 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
966 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
967 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
969 Perl_re_printf( aTHX_ "\n"); \
972 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
973 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
975 #define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
977 Perl_re_printf( aTHX_ "%s", open_str); \
978 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
979 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
980 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
981 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
982 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
983 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
984 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
985 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
986 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
987 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
988 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
989 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
990 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
991 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
992 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
993 Perl_re_printf( aTHX_ "%s", close_str); \
997 #define DEBUG_STUDYDATA(str,data,depth) \
998 DEBUG_OPTIMISE_MORE_r(if(data){ \
999 Perl_re_indentf( aTHX_ "" str "Pos:%" IVdf "/%" IVdf \
1000 " Flags: 0x%" UVXf, \
1002 (IV)((data)->pos_min), \
1003 (IV)((data)->pos_delta), \
1004 (UV)((data)->flags) \
1006 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
1007 Perl_re_printf( aTHX_ \
1008 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s", \
1009 (IV)((data)->whilem_c), \
1010 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
1011 is_inf ? "INF " : "" \
1013 if ((data)->last_found) \
1014 Perl_re_printf( aTHX_ \
1015 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf \
1016 " %sFixed:'%s' @ %" IVdf \
1017 " %sFloat: '%s' @ %" IVdf "/%" IVdf, \
1018 SvPVX_const((data)->last_found), \
1019 (IV)((data)->last_end), \
1020 (IV)((data)->last_start_min), \
1021 (IV)((data)->last_start_max), \
1022 ((data)->longest && \
1023 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
1024 SvPVX_const((data)->longest_fixed), \
1025 (IV)((data)->offset_fixed), \
1026 ((data)->longest && \
1027 (data)->longest==&((data)->longest_float)) ? "*" : "", \
1028 SvPVX_const((data)->longest_float), \
1029 (IV)((data)->offset_float_min), \
1030 (IV)((data)->offset_float_max) \
1032 Perl_re_printf( aTHX_ "\n"); \
1036 /* =========================================================
1037 * BEGIN edit_distance stuff.
1039 * This calculates how many single character changes of any type are needed to
1040 * transform a string into another one. It is taken from version 3.1 of
1042 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1045 /* Our unsorted dictionary linked list. */
1046 /* Note we use UVs, not chars. */
1051 struct dictionary* next;
1053 typedef struct dictionary item;
1056 PERL_STATIC_INLINE item*
1057 push(UV key,item* curr)
1060 Newxz(head, 1, item);
1068 PERL_STATIC_INLINE item*
1069 find(item* head, UV key)
1071 item* iterator = head;
1073 if (iterator->key == key){
1076 iterator = iterator->next;
1082 PERL_STATIC_INLINE item*
1083 uniquePush(item* head,UV key)
1085 item* iterator = head;
1088 if (iterator->key == key) {
1091 iterator = iterator->next;
1094 return push(key,head);
1097 PERL_STATIC_INLINE void
1098 dict_free(item* head)
1100 item* iterator = head;
1103 item* temp = iterator;
1104 iterator = iterator->next;
1111 /* End of Dictionary Stuff */
1113 /* All calculations/work are done here */
1115 S_edit_distance(const UV* src,
1117 const STRLEN x, /* length of src[] */
1118 const STRLEN y, /* length of tgt[] */
1119 const SSize_t maxDistance
1123 UV swapCount,swapScore,targetCharCount,i,j;
1125 UV score_ceil = x + y;
1127 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1129 /* intialize matrix start values */
1130 Newxz(scores, ( (x + 2) * (y + 2)), UV);
1131 scores[0] = score_ceil;
1132 scores[1 * (y + 2) + 0] = score_ceil;
1133 scores[0 * (y + 2) + 1] = score_ceil;
1134 scores[1 * (y + 2) + 1] = 0;
1135 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1140 for (i=1;i<=x;i++) {
1142 head = uniquePush(head,src[i]);
1143 scores[(i+1) * (y + 2) + 1] = i;
1144 scores[(i+1) * (y + 2) + 0] = score_ceil;
1147 for (j=1;j<=y;j++) {
1150 head = uniquePush(head,tgt[j]);
1151 scores[1 * (y + 2) + (j + 1)] = j;
1152 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1155 targetCharCount = find(head,tgt[j-1])->value;
1156 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1158 if (src[i-1] != tgt[j-1]){
1159 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));
1163 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1167 find(head,src[i-1])->value = i;
1171 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1174 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1178 /* END of edit_distance() stuff
1179 * ========================================================= */
1181 /* is c a control character for which we have a mnemonic? */
1182 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1185 S_cntrl_to_mnemonic(const U8 c)
1187 /* Returns the mnemonic string that represents character 'c', if one
1188 * exists; NULL otherwise. The only ones that exist for the purposes of
1189 * this routine are a few control characters */
1192 case '\a': return "\\a";
1193 case '\b': return "\\b";
1194 case ESC_NATIVE: return "\\e";
1195 case '\f': return "\\f";
1196 case '\n': return "\\n";
1197 case '\r': return "\\r";
1198 case '\t': return "\\t";
1204 /* Mark that we cannot extend a found fixed substring at this point.
1205 Update the longest found anchored substring and the longest found
1206 floating substrings if needed. */
1209 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1210 SSize_t *minlenp, int is_inf)
1212 const STRLEN l = CHR_SVLEN(data->last_found);
1213 const STRLEN old_l = CHR_SVLEN(*data->longest);
1214 GET_RE_DEBUG_FLAGS_DECL;
1216 PERL_ARGS_ASSERT_SCAN_COMMIT;
1218 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1219 SvSetMagicSV(*data->longest, data->last_found);
1220 if (*data->longest == data->longest_fixed) {
1221 data->offset_fixed = l ? data->last_start_min : data->pos_min;
1222 if (data->flags & SF_BEFORE_EOL)
1224 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
1226 data->flags &= ~SF_FIX_BEFORE_EOL;
1227 data->minlen_fixed=minlenp;
1228 data->lookbehind_fixed=0;
1230 else { /* *data->longest == data->longest_float */
1231 data->offset_float_min = l ? data->last_start_min : data->pos_min;
1232 data->offset_float_max = (l
1233 ? data->last_start_max
1234 : (data->pos_delta > SSize_t_MAX - data->pos_min
1236 : data->pos_min + data->pos_delta));
1238 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
1239 data->offset_float_max = SSize_t_MAX;
1240 if (data->flags & SF_BEFORE_EOL)
1242 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
1244 data->flags &= ~SF_FL_BEFORE_EOL;
1245 data->minlen_float=minlenp;
1246 data->lookbehind_float=0;
1249 SvCUR_set(data->last_found, 0);
1251 SV * const sv = data->last_found;
1252 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1253 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1258 data->last_end = -1;
1259 data->flags &= ~SF_BEFORE_EOL;
1260 DEBUG_STUDYDATA("commit: ",data,0);
1263 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1264 * list that describes which code points it matches */
1267 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1269 /* Set the SSC 'ssc' to match an empty string or any code point */
1271 PERL_ARGS_ASSERT_SSC_ANYTHING;
1273 assert(is_ANYOF_SYNTHETIC(ssc));
1275 /* mortalize so won't leak */
1276 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1277 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1281 S_ssc_is_anything(const regnode_ssc *ssc)
1283 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1284 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1285 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1286 * in any way, so there's no point in using it */
1291 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1293 assert(is_ANYOF_SYNTHETIC(ssc));
1295 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1299 /* See if the list consists solely of the range 0 - Infinity */
1300 invlist_iterinit(ssc->invlist);
1301 ret = invlist_iternext(ssc->invlist, &start, &end)
1305 invlist_iterfinish(ssc->invlist);
1311 /* If e.g., both \w and \W are set, matches everything */
1312 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1314 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1315 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1325 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1327 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1328 * string, any code point, or any posix class under locale */
1330 PERL_ARGS_ASSERT_SSC_INIT;
1332 Zero(ssc, 1, regnode_ssc);
1333 set_ANYOF_SYNTHETIC(ssc);
1334 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1337 /* If any portion of the regex is to operate under locale rules that aren't
1338 * fully known at compile time, initialization includes it. The reason
1339 * this isn't done for all regexes is that the optimizer was written under
1340 * the assumption that locale was all-or-nothing. Given the complexity and
1341 * lack of documentation in the optimizer, and that there are inadequate
1342 * test cases for locale, many parts of it may not work properly, it is
1343 * safest to avoid locale unless necessary. */
1344 if (RExC_contains_locale) {
1345 ANYOF_POSIXL_SETALL(ssc);
1348 ANYOF_POSIXL_ZERO(ssc);
1353 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1354 const regnode_ssc *ssc)
1356 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1357 * to the list of code points matched, and locale posix classes; hence does
1358 * not check its flags) */
1363 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1365 assert(is_ANYOF_SYNTHETIC(ssc));
1367 invlist_iterinit(ssc->invlist);
1368 ret = invlist_iternext(ssc->invlist, &start, &end)
1372 invlist_iterfinish(ssc->invlist);
1378 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1386 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1387 const regnode_charclass* const node)
1389 /* Returns a mortal inversion list defining which code points are matched
1390 * by 'node', which is of type ANYOF. Handles complementing the result if
1391 * appropriate. If some code points aren't knowable at this time, the
1392 * returned list must, and will, contain every code point that is a
1396 SV* only_utf8_locale_invlist = NULL;
1398 const U32 n = ARG(node);
1399 bool new_node_has_latin1 = FALSE;
1401 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1403 /* Look at the data structure created by S_set_ANYOF_arg() */
1404 if (n != ANYOF_ONLY_HAS_BITMAP) {
1405 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1406 AV * const av = MUTABLE_AV(SvRV(rv));
1407 SV **const ary = AvARRAY(av);
1408 assert(RExC_rxi->data->what[n] == 's');
1410 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1411 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1413 else if (ary[0] && ary[0] != &PL_sv_undef) {
1415 /* Here, no compile-time swash, and there are things that won't be
1416 * known until runtime -- we have to assume it could be anything */
1417 invlist = sv_2mortal(_new_invlist(1));
1418 return _add_range_to_invlist(invlist, 0, UV_MAX);
1420 else if (ary[3] && ary[3] != &PL_sv_undef) {
1422 /* Here no compile-time swash, and no run-time only data. Use the
1423 * node's inversion list */
1424 invlist = sv_2mortal(invlist_clone(ary[3]));
1427 /* Get the code points valid only under UTF-8 locales */
1428 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1429 && ary[2] && ary[2] != &PL_sv_undef)
1431 only_utf8_locale_invlist = ary[2];
1436 invlist = sv_2mortal(_new_invlist(0));
1439 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1440 * code points, and an inversion list for the others, but if there are code
1441 * points that should match only conditionally on the target string being
1442 * UTF-8, those are placed in the inversion list, and not the bitmap.
1443 * Since there are circumstances under which they could match, they are
1444 * included in the SSC. But if the ANYOF node is to be inverted, we have
1445 * to exclude them here, so that when we invert below, the end result
1446 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1447 * have to do this here before we add the unconditionally matched code
1449 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1450 _invlist_intersection_complement_2nd(invlist,
1455 /* Add in the points from the bit map */
1456 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1457 if (ANYOF_BITMAP_TEST(node, i)) {
1458 unsigned int start = i++;
1460 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1463 invlist = _add_range_to_invlist(invlist, start, i-1);
1464 new_node_has_latin1 = TRUE;
1468 /* If this can match all upper Latin1 code points, have to add them
1469 * as well. But don't add them if inverting, as when that gets done below,
1470 * it would exclude all these characters, including the ones it shouldn't
1471 * that were added just above */
1472 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1473 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1475 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1478 /* Similarly for these */
1479 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1480 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1483 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1484 _invlist_invert(invlist);
1486 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1488 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1489 * locale. We can skip this if there are no 0-255 at all. */
1490 _invlist_union(invlist, PL_Latin1, &invlist);
1493 /* Similarly add the UTF-8 locale possible matches. These have to be
1494 * deferred until after the non-UTF-8 locale ones are taken care of just
1495 * above, or it leads to wrong results under ANYOF_INVERT */
1496 if (only_utf8_locale_invlist) {
1497 _invlist_union_maybe_complement_2nd(invlist,
1498 only_utf8_locale_invlist,
1499 ANYOF_FLAGS(node) & ANYOF_INVERT,
1506 /* These two functions currently do the exact same thing */
1507 #define ssc_init_zero ssc_init
1509 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1510 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1512 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1513 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1514 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1517 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1518 const regnode_charclass *and_with)
1520 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1521 * another SSC or a regular ANYOF class. Can create false positives. */
1526 PERL_ARGS_ASSERT_SSC_AND;
1528 assert(is_ANYOF_SYNTHETIC(ssc));
1530 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1531 * the code point inversion list and just the relevant flags */
1532 if (is_ANYOF_SYNTHETIC(and_with)) {
1533 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1534 anded_flags = ANYOF_FLAGS(and_with);
1536 /* XXX This is a kludge around what appears to be deficiencies in the
1537 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1538 * there are paths through the optimizer where it doesn't get weeded
1539 * out when it should. And if we don't make some extra provision for
1540 * it like the code just below, it doesn't get added when it should.
1541 * This solution is to add it only when AND'ing, which is here, and
1542 * only when what is being AND'ed is the pristine, original node
1543 * matching anything. Thus it is like adding it to ssc_anything() but
1544 * only when the result is to be AND'ed. Probably the same solution
1545 * could be adopted for the same problem we have with /l matching,
1546 * which is solved differently in S_ssc_init(), and that would lead to
1547 * fewer false positives than that solution has. But if this solution
1548 * creates bugs, the consequences are only that a warning isn't raised
1549 * that should be; while the consequences for having /l bugs is
1550 * incorrect matches */
1551 if (ssc_is_anything((regnode_ssc *)and_with)) {
1552 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1556 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1557 if (OP(and_with) == ANYOFD) {
1558 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1561 anded_flags = ANYOF_FLAGS(and_with)
1562 &( ANYOF_COMMON_FLAGS
1563 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1564 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1565 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1567 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1572 ANYOF_FLAGS(ssc) &= anded_flags;
1574 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1575 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1576 * 'and_with' may be inverted. When not inverted, we have the situation of
1578 * (C1 | P1) & (C2 | P2)
1579 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1580 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1581 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1582 * <= ((C1 & C2) | P1 | P2)
1583 * Alternatively, the last few steps could be:
1584 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1585 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1586 * <= (C1 | C2 | (P1 & P2))
1587 * We favor the second approach if either P1 or P2 is non-empty. This is
1588 * because these components are a barrier to doing optimizations, as what
1589 * they match cannot be known until the moment of matching as they are
1590 * dependent on the current locale, 'AND"ing them likely will reduce or
1592 * But we can do better if we know that C1,P1 are in their initial state (a
1593 * frequent occurrence), each matching everything:
1594 * (<everything>) & (C2 | P2) = C2 | P2
1595 * Similarly, if C2,P2 are in their initial state (again a frequent
1596 * occurrence), the result is a no-op
1597 * (C1 | P1) & (<everything>) = C1 | P1
1600 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1601 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1602 * <= (C1 & ~C2) | (P1 & ~P2)
1605 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1606 && ! is_ANYOF_SYNTHETIC(and_with))
1610 ssc_intersection(ssc,
1612 FALSE /* Has already been inverted */
1615 /* If either P1 or P2 is empty, the intersection will be also; can skip
1617 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1618 ANYOF_POSIXL_ZERO(ssc);
1620 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1622 /* Note that the Posix class component P from 'and_with' actually
1624 * P = Pa | Pb | ... | Pn
1625 * where each component is one posix class, such as in [\w\s].
1627 * ~P = ~(Pa | Pb | ... | Pn)
1628 * = ~Pa & ~Pb & ... & ~Pn
1629 * <= ~Pa | ~Pb | ... | ~Pn
1630 * The last is something we can easily calculate, but unfortunately
1631 * is likely to have many false positives. We could do better
1632 * in some (but certainly not all) instances if two classes in
1633 * P have known relationships. For example
1634 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1636 * :lower: & :print: = :lower:
1637 * And similarly for classes that must be disjoint. For example,
1638 * since \s and \w can have no elements in common based on rules in
1639 * the POSIX standard,
1640 * \w & ^\S = nothing
1641 * Unfortunately, some vendor locales do not meet the Posix
1642 * standard, in particular almost everything by Microsoft.
1643 * The loop below just changes e.g., \w into \W and vice versa */
1645 regnode_charclass_posixl temp;
1646 int add = 1; /* To calculate the index of the complement */
1648 ANYOF_POSIXL_ZERO(&temp);
1649 for (i = 0; i < ANYOF_MAX; i++) {
1651 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1652 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1654 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1655 ANYOF_POSIXL_SET(&temp, i + add);
1657 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1659 ANYOF_POSIXL_AND(&temp, ssc);
1661 } /* else ssc already has no posixes */
1662 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1663 in its initial state */
1664 else if (! is_ANYOF_SYNTHETIC(and_with)
1665 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1667 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1668 * copy it over 'ssc' */
1669 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1670 if (is_ANYOF_SYNTHETIC(and_with)) {
1671 StructCopy(and_with, ssc, regnode_ssc);
1674 ssc->invlist = anded_cp_list;
1675 ANYOF_POSIXL_ZERO(ssc);
1676 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1677 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1681 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1682 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1684 /* One or the other of P1, P2 is non-empty. */
1685 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1686 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1688 ssc_union(ssc, anded_cp_list, FALSE);
1690 else { /* P1 = P2 = empty */
1691 ssc_intersection(ssc, anded_cp_list, FALSE);
1697 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1698 const regnode_charclass *or_with)
1700 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1701 * another SSC or a regular ANYOF class. Can create false positives if
1702 * 'or_with' is to be inverted. */
1707 PERL_ARGS_ASSERT_SSC_OR;
1709 assert(is_ANYOF_SYNTHETIC(ssc));
1711 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1712 * the code point inversion list and just the relevant flags */
1713 if (is_ANYOF_SYNTHETIC(or_with)) {
1714 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1715 ored_flags = ANYOF_FLAGS(or_with);
1718 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1719 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1720 if (OP(or_with) != ANYOFD) {
1722 |= ANYOF_FLAGS(or_with)
1723 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1724 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1725 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1727 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1732 ANYOF_FLAGS(ssc) |= ored_flags;
1734 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1735 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1736 * 'or_with' may be inverted. When not inverted, we have the simple
1737 * situation of computing:
1738 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1739 * If P1|P2 yields a situation with both a class and its complement are
1740 * set, like having both \w and \W, this matches all code points, and we
1741 * can delete these from the P component of the ssc going forward. XXX We
1742 * might be able to delete all the P components, but I (khw) am not certain
1743 * about this, and it is better to be safe.
1746 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1747 * <= (C1 | P1) | ~C2
1748 * <= (C1 | ~C2) | P1
1749 * (which results in actually simpler code than the non-inverted case)
1752 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1753 && ! is_ANYOF_SYNTHETIC(or_with))
1755 /* We ignore P2, leaving P1 going forward */
1756 } /* else Not inverted */
1757 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1758 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1759 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1761 for (i = 0; i < ANYOF_MAX; i += 2) {
1762 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1764 ssc_match_all_cp(ssc);
1765 ANYOF_POSIXL_CLEAR(ssc, i);
1766 ANYOF_POSIXL_CLEAR(ssc, i+1);
1774 FALSE /* Already has been inverted */
1778 PERL_STATIC_INLINE void
1779 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1781 PERL_ARGS_ASSERT_SSC_UNION;
1783 assert(is_ANYOF_SYNTHETIC(ssc));
1785 _invlist_union_maybe_complement_2nd(ssc->invlist,
1791 PERL_STATIC_INLINE void
1792 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1794 const bool invert2nd)
1796 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1798 assert(is_ANYOF_SYNTHETIC(ssc));
1800 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1806 PERL_STATIC_INLINE void
1807 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1809 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1811 assert(is_ANYOF_SYNTHETIC(ssc));
1813 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1816 PERL_STATIC_INLINE void
1817 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1819 /* AND just the single code point 'cp' into the SSC 'ssc' */
1821 SV* cp_list = _new_invlist(2);
1823 PERL_ARGS_ASSERT_SSC_CP_AND;
1825 assert(is_ANYOF_SYNTHETIC(ssc));
1827 cp_list = add_cp_to_invlist(cp_list, cp);
1828 ssc_intersection(ssc, cp_list,
1829 FALSE /* Not inverted */
1831 SvREFCNT_dec_NN(cp_list);
1834 PERL_STATIC_INLINE void
1835 S_ssc_clear_locale(regnode_ssc *ssc)
1837 /* Set the SSC 'ssc' to not match any locale things */
1838 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1840 assert(is_ANYOF_SYNTHETIC(ssc));
1842 ANYOF_POSIXL_ZERO(ssc);
1843 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1846 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1849 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1851 /* The synthetic start class is used to hopefully quickly winnow down
1852 * places where a pattern could start a match in the target string. If it
1853 * doesn't really narrow things down that much, there isn't much point to
1854 * having the overhead of using it. This function uses some very crude
1855 * heuristics to decide if to use the ssc or not.
1857 * It returns TRUE if 'ssc' rules out more than half what it considers to
1858 * be the "likely" possible matches, but of course it doesn't know what the
1859 * actual things being matched are going to be; these are only guesses
1861 * For /l matches, it assumes that the only likely matches are going to be
1862 * in the 0-255 range, uniformly distributed, so half of that is 127
1863 * For /a and /d matches, it assumes that the likely matches will be just
1864 * the ASCII range, so half of that is 63
1865 * For /u and there isn't anything matching above the Latin1 range, it
1866 * assumes that that is the only range likely to be matched, and uses
1867 * half that as the cut-off: 127. If anything matches above Latin1,
1868 * it assumes that all of Unicode could match (uniformly), except for
1869 * non-Unicode code points and things in the General Category "Other"
1870 * (unassigned, private use, surrogates, controls and formats). This
1871 * is a much large number. */
1873 U32 count = 0; /* Running total of number of code points matched by
1875 UV start, end; /* Start and end points of current range in inversion
1877 const U32 max_code_points = (LOC)
1879 : (( ! UNI_SEMANTICS
1880 || invlist_highest(ssc->invlist) < 256)
1883 const U32 max_match = max_code_points / 2;
1885 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1887 invlist_iterinit(ssc->invlist);
1888 while (invlist_iternext(ssc->invlist, &start, &end)) {
1889 if (start >= max_code_points) {
1892 end = MIN(end, max_code_points - 1);
1893 count += end - start + 1;
1894 if (count >= max_match) {
1895 invlist_iterfinish(ssc->invlist);
1905 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1907 /* The inversion list in the SSC is marked mortal; now we need a more
1908 * permanent copy, which is stored the same way that is done in a regular
1909 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1912 SV* invlist = invlist_clone(ssc->invlist);
1914 PERL_ARGS_ASSERT_SSC_FINALIZE;
1916 assert(is_ANYOF_SYNTHETIC(ssc));
1918 /* The code in this file assumes that all but these flags aren't relevant
1919 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1920 * by the time we reach here */
1921 assert(! (ANYOF_FLAGS(ssc)
1922 & ~( ANYOF_COMMON_FLAGS
1923 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1924 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
1926 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1928 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1929 NULL, NULL, NULL, FALSE);
1931 /* Make sure is clone-safe */
1932 ssc->invlist = NULL;
1934 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1935 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1938 if (RExC_contains_locale) {
1942 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1945 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1946 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1947 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1948 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1949 ? (TRIE_LIST_CUR( idx ) - 1) \
1955 dump_trie(trie,widecharmap,revcharmap)
1956 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1957 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1959 These routines dump out a trie in a somewhat readable format.
1960 The _interim_ variants are used for debugging the interim
1961 tables that are used to generate the final compressed
1962 representation which is what dump_trie expects.
1964 Part of the reason for their existence is to provide a form
1965 of documentation as to how the different representations function.
1970 Dumps the final compressed table form of the trie to Perl_debug_log.
1971 Used for debugging make_trie().
1975 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1976 AV *revcharmap, U32 depth)
1979 SV *sv=sv_newmortal();
1980 int colwidth= widecharmap ? 6 : 4;
1982 GET_RE_DEBUG_FLAGS_DECL;
1984 PERL_ARGS_ASSERT_DUMP_TRIE;
1986 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
1987 depth+1, "Match","Base","Ofs" );
1989 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1990 SV ** const tmp = av_fetch( revcharmap, state, 0);
1992 Perl_re_printf( aTHX_ "%*s",
1994 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1995 PL_colors[0], PL_colors[1],
1996 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1997 PERL_PV_ESCAPE_FIRSTCHAR
2002 Perl_re_printf( aTHX_ "\n");
2003 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2005 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2006 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2007 Perl_re_printf( aTHX_ "\n");
2009 for( state = 1 ; state < trie->statecount ; state++ ) {
2010 const U32 base = trie->states[ state ].trans.base;
2012 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2014 if ( trie->states[ state ].wordnum ) {
2015 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2017 Perl_re_printf( aTHX_ "%6s", "" );
2020 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2025 while( ( base + ofs < trie->uniquecharcount ) ||
2026 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2027 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2031 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2033 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2034 if ( ( base + ofs >= trie->uniquecharcount )
2035 && ( base + ofs - trie->uniquecharcount
2037 && trie->trans[ base + ofs
2038 - trie->uniquecharcount ].check == state )
2040 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2041 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2044 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
2048 Perl_re_printf( aTHX_ "]");
2051 Perl_re_printf( aTHX_ "\n" );
2053 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2055 for (word=1; word <= trie->wordcount; word++) {
2056 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2057 (int)word, (int)(trie->wordinfo[word].prev),
2058 (int)(trie->wordinfo[word].len));
2060 Perl_re_printf( aTHX_ "\n" );
2063 Dumps a fully constructed but uncompressed trie in list form.
2064 List tries normally only are used for construction when the number of
2065 possible chars (trie->uniquecharcount) is very high.
2066 Used for debugging make_trie().
2069 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2070 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2074 SV *sv=sv_newmortal();
2075 int colwidth= widecharmap ? 6 : 4;
2076 GET_RE_DEBUG_FLAGS_DECL;
2078 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2080 /* print out the table precompression. */
2081 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2083 Perl_re_indentf( aTHX_ "%s",
2084 depth+1, "------:-----+-----------------\n" );
2086 for( state=1 ; state < next_alloc ; state ++ ) {
2089 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2090 depth+1, (UV)state );
2091 if ( ! trie->states[ state ].wordnum ) {
2092 Perl_re_printf( aTHX_ "%5s| ","");
2094 Perl_re_printf( aTHX_ "W%4x| ",
2095 trie->states[ state ].wordnum
2098 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2099 SV ** const tmp = av_fetch( revcharmap,
2100 TRIE_LIST_ITEM(state,charid).forid, 0);
2102 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2104 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2106 PL_colors[0], PL_colors[1],
2107 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2108 | PERL_PV_ESCAPE_FIRSTCHAR
2110 TRIE_LIST_ITEM(state,charid).forid,
2111 (UV)TRIE_LIST_ITEM(state,charid).newstate
2114 Perl_re_printf( aTHX_ "\n%*s| ",
2115 (int)((depth * 2) + 14), "");
2118 Perl_re_printf( aTHX_ "\n");
2123 Dumps a fully constructed but uncompressed trie in table form.
2124 This is the normal DFA style state transition table, with a few
2125 twists to facilitate compression later.
2126 Used for debugging make_trie().
2129 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2130 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2135 SV *sv=sv_newmortal();
2136 int colwidth= widecharmap ? 6 : 4;
2137 GET_RE_DEBUG_FLAGS_DECL;
2139 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2142 print out the table precompression so that we can do a visual check
2143 that they are identical.
2146 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2148 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2149 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2151 Perl_re_printf( aTHX_ "%*s",
2153 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2154 PL_colors[0], PL_colors[1],
2155 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2156 PERL_PV_ESCAPE_FIRSTCHAR
2162 Perl_re_printf( aTHX_ "\n");
2163 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2165 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2166 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2169 Perl_re_printf( aTHX_ "\n" );
2171 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2173 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2175 (UV)TRIE_NODENUM( state ) );
2177 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2178 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2180 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2182 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2184 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2185 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2186 (UV)trie->trans[ state ].check );
2188 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2189 (UV)trie->trans[ state ].check,
2190 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2198 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2199 startbranch: the first branch in the whole branch sequence
2200 first : start branch of sequence of branch-exact nodes.
2201 May be the same as startbranch
2202 last : Thing following the last branch.
2203 May be the same as tail.
2204 tail : item following the branch sequence
2205 count : words in the sequence
2206 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2207 depth : indent depth
2209 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2211 A trie is an N'ary tree where the branches are determined by digital
2212 decomposition of the key. IE, at the root node you look up the 1st character and
2213 follow that branch repeat until you find the end of the branches. Nodes can be
2214 marked as "accepting" meaning they represent a complete word. Eg:
2218 would convert into the following structure. Numbers represent states, letters
2219 following numbers represent valid transitions on the letter from that state, if
2220 the number is in square brackets it represents an accepting state, otherwise it
2221 will be in parenthesis.
2223 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2227 (1) +-i->(6)-+-s->[7]
2229 +-s->(3)-+-h->(4)-+-e->[5]
2231 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2233 This shows that when matching against the string 'hers' we will begin at state 1
2234 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2235 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2236 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2237 single traverse. We store a mapping from accepting to state to which word was
2238 matched, and then when we have multiple possibilities we try to complete the
2239 rest of the regex in the order in which they occurred in the alternation.
2241 The only prior NFA like behaviour that would be changed by the TRIE support is
2242 the silent ignoring of duplicate alternations which are of the form:
2244 / (DUPE|DUPE) X? (?{ ... }) Y /x
2246 Thus EVAL blocks following a trie may be called a different number of times with
2247 and without the optimisation. With the optimisations dupes will be silently
2248 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2249 the following demonstrates:
2251 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2253 which prints out 'word' three times, but
2255 'words'=~/(word|word|word)(?{ print $1 })S/
2257 which doesnt print it out at all. This is due to other optimisations kicking in.
2259 Example of what happens on a structural level:
2261 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2263 1: CURLYM[1] {1,32767}(18)
2274 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2275 and should turn into:
2277 1: CURLYM[1] {1,32767}(18)
2279 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2287 Cases where tail != last would be like /(?foo|bar)baz/:
2297 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2298 and would end up looking like:
2301 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2308 d = uvchr_to_utf8_flags(d, uv, 0);
2310 is the recommended Unicode-aware way of saying
2315 #define TRIE_STORE_REVCHAR(val) \
2318 SV *zlopp = newSV(UTF8_MAXBYTES); \
2319 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2320 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2321 SvCUR_set(zlopp, kapow - flrbbbbb); \
2324 av_push(revcharmap, zlopp); \
2326 char ooooff = (char)val; \
2327 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2331 /* This gets the next character from the input, folding it if not already
2333 #define TRIE_READ_CHAR STMT_START { \
2336 /* if it is UTF then it is either already folded, or does not need \
2338 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2340 else if (folder == PL_fold_latin1) { \
2341 /* This folder implies Unicode rules, which in the range expressible \
2342 * by not UTF is the lower case, with the two exceptions, one of \
2343 * which should have been taken care of before calling this */ \
2344 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2345 uvc = toLOWER_L1(*uc); \
2346 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2349 /* raw data, will be folded later if needed */ \
2357 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2358 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2359 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2360 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2362 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2363 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2364 TRIE_LIST_CUR( state )++; \
2367 #define TRIE_LIST_NEW(state) STMT_START { \
2368 Newxz( trie->states[ state ].trans.list, \
2369 4, reg_trie_trans_le ); \
2370 TRIE_LIST_CUR( state ) = 1; \
2371 TRIE_LIST_LEN( state ) = 4; \
2374 #define TRIE_HANDLE_WORD(state) STMT_START { \
2375 U16 dupe= trie->states[ state ].wordnum; \
2376 regnode * const noper_next = regnext( noper ); \
2379 /* store the word for dumping */ \
2381 if (OP(noper) != NOTHING) \
2382 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2384 tmp = newSVpvn_utf8( "", 0, UTF ); \
2385 av_push( trie_words, tmp ); \
2389 trie->wordinfo[curword].prev = 0; \
2390 trie->wordinfo[curword].len = wordlen; \
2391 trie->wordinfo[curword].accept = state; \
2393 if ( noper_next < tail ) { \
2395 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2397 trie->jump[curword] = (U16)(noper_next - convert); \
2399 jumper = noper_next; \
2401 nextbranch= regnext(cur); \
2405 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2406 /* chain, so that when the bits of chain are later */\
2407 /* linked together, the dups appear in the chain */\
2408 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2409 trie->wordinfo[dupe].prev = curword; \
2411 /* we haven't inserted this word yet. */ \
2412 trie->states[ state ].wordnum = curword; \
2417 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2418 ( ( base + charid >= ucharcount \
2419 && base + charid < ubound \
2420 && state == trie->trans[ base - ucharcount + charid ].check \
2421 && trie->trans[ base - ucharcount + charid ].next ) \
2422 ? trie->trans[ base - ucharcount + charid ].next \
2423 : ( state==1 ? special : 0 ) \
2426 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2428 TRIE_BITMAP_SET(trie, uvc); \
2429 /* store the folded codepoint */ \
2431 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2434 /* store first byte of utf8 representation of */ \
2435 /* variant codepoints */ \
2436 if (! UVCHR_IS_INVARIANT(uvc)) { \
2437 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2442 #define MADE_JUMP_TRIE 2
2443 #define MADE_EXACT_TRIE 4
2446 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2447 regnode *first, regnode *last, regnode *tail,
2448 U32 word_count, U32 flags, U32 depth)
2450 /* first pass, loop through and scan words */
2451 reg_trie_data *trie;
2452 HV *widecharmap = NULL;
2453 AV *revcharmap = newAV();
2459 regnode *jumper = NULL;
2460 regnode *nextbranch = NULL;
2461 regnode *convert = NULL;
2462 U32 *prev_states; /* temp array mapping each state to previous one */
2463 /* we just use folder as a flag in utf8 */
2464 const U8 * folder = NULL;
2467 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2468 AV *trie_words = NULL;
2469 /* along with revcharmap, this only used during construction but both are
2470 * useful during debugging so we store them in the struct when debugging.
2473 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2474 STRLEN trie_charcount=0;
2476 SV *re_trie_maxbuff;
2477 GET_RE_DEBUG_FLAGS_DECL;
2479 PERL_ARGS_ASSERT_MAKE_TRIE;
2481 PERL_UNUSED_ARG(depth);
2485 case EXACT: case EXACTL: break;
2489 case EXACTFLU8: folder = PL_fold_latin1; break;
2490 case EXACTF: folder = PL_fold; break;
2491 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2494 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2496 trie->startstate = 1;
2497 trie->wordcount = word_count;
2498 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2499 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2500 if (flags == EXACT || flags == EXACTL)
2501 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2502 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2503 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2506 trie_words = newAV();
2509 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2510 assert(re_trie_maxbuff);
2511 if (!SvIOK(re_trie_maxbuff)) {
2512 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2514 DEBUG_TRIE_COMPILE_r({
2515 Perl_re_indentf( aTHX_
2516 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2518 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2519 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2522 /* Find the node we are going to overwrite */
2523 if ( first == startbranch && OP( last ) != BRANCH ) {
2524 /* whole branch chain */
2527 /* branch sub-chain */
2528 convert = NEXTOPER( first );
2531 /* -- First loop and Setup --
2533 We first traverse the branches and scan each word to determine if it
2534 contains widechars, and how many unique chars there are, this is
2535 important as we have to build a table with at least as many columns as we
2538 We use an array of integers to represent the character codes 0..255
2539 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2540 the native representation of the character value as the key and IV's for
2543 *TODO* If we keep track of how many times each character is used we can
2544 remap the columns so that the table compression later on is more
2545 efficient in terms of memory by ensuring the most common value is in the
2546 middle and the least common are on the outside. IMO this would be better
2547 than a most to least common mapping as theres a decent chance the most
2548 common letter will share a node with the least common, meaning the node
2549 will not be compressible. With a middle is most common approach the worst
2550 case is when we have the least common nodes twice.
2554 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2555 regnode *noper = NEXTOPER( cur );
2559 U32 wordlen = 0; /* required init */
2560 STRLEN minchars = 0;
2561 STRLEN maxchars = 0;
2562 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2565 if (OP(noper) == NOTHING) {
2566 /* skip past a NOTHING at the start of an alternation
2567 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2569 regnode *noper_next= regnext(noper);
2570 if (noper_next < tail)
2574 if ( noper < tail &&
2576 OP(noper) == flags ||
2579 OP(noper) == EXACTFU_SS
2583 uc= (U8*)STRING(noper);
2584 e= uc + STR_LEN(noper);
2591 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2592 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2593 regardless of encoding */
2594 if (OP( noper ) == EXACTFU_SS) {
2595 /* false positives are ok, so just set this */
2596 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2600 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2602 TRIE_CHARCOUNT(trie)++;
2605 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2606 * is in effect. Under /i, this character can match itself, or
2607 * anything that folds to it. If not under /i, it can match just
2608 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2609 * all fold to k, and all are single characters. But some folds
2610 * expand to more than one character, so for example LATIN SMALL
2611 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2612 * the string beginning at 'uc' is 'ffi', it could be matched by
2613 * three characters, or just by the one ligature character. (It
2614 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2615 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2616 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2617 * match.) The trie needs to know the minimum and maximum number
2618 * of characters that could match so that it can use size alone to
2619 * quickly reject many match attempts. The max is simple: it is
2620 * the number of folded characters in this branch (since a fold is
2621 * never shorter than what folds to it. */
2625 /* And the min is equal to the max if not under /i (indicated by
2626 * 'folder' being NULL), or there are no multi-character folds. If
2627 * there is a multi-character fold, the min is incremented just
2628 * once, for the character that folds to the sequence. Each
2629 * character in the sequence needs to be added to the list below of
2630 * characters in the trie, but we count only the first towards the
2631 * min number of characters needed. This is done through the
2632 * variable 'foldlen', which is returned by the macros that look
2633 * for these sequences as the number of bytes the sequence
2634 * occupies. Each time through the loop, we decrement 'foldlen' by
2635 * how many bytes the current char occupies. Only when it reaches
2636 * 0 do we increment 'minchars' or look for another multi-character
2638 if (folder == NULL) {
2641 else if (foldlen > 0) {
2642 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2647 /* See if *uc is the beginning of a multi-character fold. If
2648 * so, we decrement the length remaining to look at, to account
2649 * for the current character this iteration. (We can use 'uc'
2650 * instead of the fold returned by TRIE_READ_CHAR because for
2651 * non-UTF, the latin1_safe macro is smart enough to account
2652 * for all the unfolded characters, and because for UTF, the
2653 * string will already have been folded earlier in the
2654 * compilation process */
2656 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2657 foldlen -= UTF8SKIP(uc);
2660 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2665 /* The current character (and any potential folds) should be added
2666 * to the possible matching characters for this position in this
2670 U8 folded= folder[ (U8) uvc ];
2671 if ( !trie->charmap[ folded ] ) {
2672 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2673 TRIE_STORE_REVCHAR( folded );
2676 if ( !trie->charmap[ uvc ] ) {
2677 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2678 TRIE_STORE_REVCHAR( uvc );
2681 /* store the codepoint in the bitmap, and its folded
2683 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2684 set_bit = 0; /* We've done our bit :-) */
2688 /* XXX We could come up with the list of code points that fold
2689 * to this using PL_utf8_foldclosures, except not for
2690 * multi-char folds, as there may be multiple combinations
2691 * there that could work, which needs to wait until runtime to
2692 * resolve (The comment about LIGATURE FFI above is such an
2697 widecharmap = newHV();
2699 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2702 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2704 if ( !SvTRUE( *svpp ) ) {
2705 sv_setiv( *svpp, ++trie->uniquecharcount );
2706 TRIE_STORE_REVCHAR(uvc);
2709 } /* end loop through characters in this branch of the trie */
2711 /* We take the min and max for this branch and combine to find the min
2712 * and max for all branches processed so far */
2713 if( cur == first ) {
2714 trie->minlen = minchars;
2715 trie->maxlen = maxchars;
2716 } else if (minchars < trie->minlen) {
2717 trie->minlen = minchars;
2718 } else if (maxchars > trie->maxlen) {
2719 trie->maxlen = maxchars;
2721 } /* end first pass */
2722 DEBUG_TRIE_COMPILE_r(
2723 Perl_re_indentf( aTHX_
2724 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2726 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2727 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2728 (int)trie->minlen, (int)trie->maxlen )
2732 We now know what we are dealing with in terms of unique chars and
2733 string sizes so we can calculate how much memory a naive
2734 representation using a flat table will take. If it's over a reasonable
2735 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2736 conservative but potentially much slower representation using an array
2739 At the end we convert both representations into the same compressed
2740 form that will be used in regexec.c for matching with. The latter
2741 is a form that cannot be used to construct with but has memory
2742 properties similar to the list form and access properties similar
2743 to the table form making it both suitable for fast searches and
2744 small enough that its feasable to store for the duration of a program.
2746 See the comment in the code where the compressed table is produced
2747 inplace from the flat tabe representation for an explanation of how
2748 the compression works.
2753 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2756 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2757 > SvIV(re_trie_maxbuff) )
2760 Second Pass -- Array Of Lists Representation
2762 Each state will be represented by a list of charid:state records
2763 (reg_trie_trans_le) the first such element holds the CUR and LEN
2764 points of the allocated array. (See defines above).
2766 We build the initial structure using the lists, and then convert
2767 it into the compressed table form which allows faster lookups
2768 (but cant be modified once converted).
2771 STRLEN transcount = 1;
2773 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2776 trie->states = (reg_trie_state *)
2777 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2778 sizeof(reg_trie_state) );
2782 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2784 regnode *noper = NEXTOPER( cur );
2785 U32 state = 1; /* required init */
2786 U16 charid = 0; /* sanity init */
2787 U32 wordlen = 0; /* required init */
2789 if (OP(noper) == NOTHING) {
2790 regnode *noper_next= regnext(noper);
2791 if (noper_next < tail)
2795 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2796 const U8 *uc= (U8*)STRING(noper);
2797 const U8 *e= uc + STR_LEN(noper);
2799 for ( ; uc < e ; uc += len ) {
2804 charid = trie->charmap[ uvc ];
2806 SV** const svpp = hv_fetch( widecharmap,
2813 charid=(U16)SvIV( *svpp );
2816 /* charid is now 0 if we dont know the char read, or
2817 * nonzero if we do */
2824 if ( !trie->states[ state ].trans.list ) {
2825 TRIE_LIST_NEW( state );
2828 check <= TRIE_LIST_USED( state );
2831 if ( TRIE_LIST_ITEM( state, check ).forid
2834 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2839 newstate = next_alloc++;
2840 prev_states[newstate] = state;
2841 TRIE_LIST_PUSH( state, charid, newstate );
2846 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
2850 TRIE_HANDLE_WORD(state);
2852 } /* end second pass */
2854 /* next alloc is the NEXT state to be allocated */
2855 trie->statecount = next_alloc;
2856 trie->states = (reg_trie_state *)
2857 PerlMemShared_realloc( trie->states,
2859 * sizeof(reg_trie_state) );
2861 /* and now dump it out before we compress it */
2862 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2863 revcharmap, next_alloc,
2867 trie->trans = (reg_trie_trans *)
2868 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2875 for( state=1 ; state < next_alloc ; state ++ ) {
2879 DEBUG_TRIE_COMPILE_MORE_r(
2880 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
2884 if (trie->states[state].trans.list) {
2885 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2889 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2890 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2891 if ( forid < minid ) {
2893 } else if ( forid > maxid ) {
2897 if ( transcount < tp + maxid - minid + 1) {
2899 trie->trans = (reg_trie_trans *)
2900 PerlMemShared_realloc( trie->trans,
2902 * sizeof(reg_trie_trans) );
2903 Zero( trie->trans + (transcount / 2),
2907 base = trie->uniquecharcount + tp - minid;
2908 if ( maxid == minid ) {
2910 for ( ; zp < tp ; zp++ ) {
2911 if ( ! trie->trans[ zp ].next ) {
2912 base = trie->uniquecharcount + zp - minid;
2913 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2915 trie->trans[ zp ].check = state;
2921 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2923 trie->trans[ tp ].check = state;
2928 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2929 const U32 tid = base
2930 - trie->uniquecharcount
2931 + TRIE_LIST_ITEM( state, idx ).forid;
2932 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2934 trie->trans[ tid ].check = state;
2936 tp += ( maxid - minid + 1 );
2938 Safefree(trie->states[ state ].trans.list);
2941 DEBUG_TRIE_COMPILE_MORE_r(
2942 Perl_re_printf( aTHX_ " base: %d\n",base);
2945 trie->states[ state ].trans.base=base;
2947 trie->lasttrans = tp + 1;
2951 Second Pass -- Flat Table Representation.
2953 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2954 each. We know that we will need Charcount+1 trans at most to store
2955 the data (one row per char at worst case) So we preallocate both
2956 structures assuming worst case.
2958 We then construct the trie using only the .next slots of the entry
2961 We use the .check field of the first entry of the node temporarily
2962 to make compression both faster and easier by keeping track of how
2963 many non zero fields are in the node.
2965 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2968 There are two terms at use here: state as a TRIE_NODEIDX() which is
2969 a number representing the first entry of the node, and state as a
2970 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2971 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2972 if there are 2 entrys per node. eg:
2980 The table is internally in the right hand, idx form. However as we
2981 also have to deal with the states array which is indexed by nodenum
2982 we have to use TRIE_NODENUM() to convert.
2985 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
2988 trie->trans = (reg_trie_trans *)
2989 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2990 * trie->uniquecharcount + 1,
2991 sizeof(reg_trie_trans) );
2992 trie->states = (reg_trie_state *)
2993 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2994 sizeof(reg_trie_state) );
2995 next_alloc = trie->uniquecharcount + 1;
2998 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3000 regnode *noper = NEXTOPER( cur );
3002 U32 state = 1; /* required init */
3004 U16 charid = 0; /* sanity init */
3005 U32 accept_state = 0; /* sanity init */
3007 U32 wordlen = 0; /* required init */
3009 if (OP(noper) == NOTHING) {
3010 regnode *noper_next= regnext(noper);
3011 if (noper_next < tail)
3015 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3016 const U8 *uc= (U8*)STRING(noper);
3017 const U8 *e= uc + STR_LEN(noper);
3019 for ( ; uc < e ; uc += len ) {
3024 charid = trie->charmap[ uvc ];
3026 SV* const * const svpp = hv_fetch( widecharmap,
3030 charid = svpp ? (U16)SvIV(*svpp) : 0;
3034 if ( !trie->trans[ state + charid ].next ) {
3035 trie->trans[ state + charid ].next = next_alloc;
3036 trie->trans[ state ].check++;
3037 prev_states[TRIE_NODENUM(next_alloc)]
3038 = TRIE_NODENUM(state);
3039 next_alloc += trie->uniquecharcount;
3041 state = trie->trans[ state + charid ].next;
3043 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3045 /* charid is now 0 if we dont know the char read, or
3046 * nonzero if we do */
3049 accept_state = TRIE_NODENUM( state );
3050 TRIE_HANDLE_WORD(accept_state);
3052 } /* end second pass */
3054 /* and now dump it out before we compress it */
3055 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3057 next_alloc, depth+1));
3061 * Inplace compress the table.*
3063 For sparse data sets the table constructed by the trie algorithm will
3064 be mostly 0/FAIL transitions or to put it another way mostly empty.
3065 (Note that leaf nodes will not contain any transitions.)
3067 This algorithm compresses the tables by eliminating most such
3068 transitions, at the cost of a modest bit of extra work during lookup:
3070 - Each states[] entry contains a .base field which indicates the
3071 index in the state[] array wheres its transition data is stored.
3073 - If .base is 0 there are no valid transitions from that node.
3075 - If .base is nonzero then charid is added to it to find an entry in
3078 -If trans[states[state].base+charid].check!=state then the
3079 transition is taken to be a 0/Fail transition. Thus if there are fail
3080 transitions at the front of the node then the .base offset will point
3081 somewhere inside the previous nodes data (or maybe even into a node
3082 even earlier), but the .check field determines if the transition is
3086 The following process inplace converts the table to the compressed
3087 table: We first do not compress the root node 1,and mark all its
3088 .check pointers as 1 and set its .base pointer as 1 as well. This
3089 allows us to do a DFA construction from the compressed table later,
3090 and ensures that any .base pointers we calculate later are greater
3093 - We set 'pos' to indicate the first entry of the second node.
3095 - We then iterate over the columns of the node, finding the first and
3096 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3097 and set the .check pointers accordingly, and advance pos
3098 appropriately and repreat for the next node. Note that when we copy
3099 the next pointers we have to convert them from the original
3100 NODEIDX form to NODENUM form as the former is not valid post
3103 - If a node has no transitions used we mark its base as 0 and do not
3104 advance the pos pointer.
3106 - If a node only has one transition we use a second pointer into the
3107 structure to fill in allocated fail transitions from other states.
3108 This pointer is independent of the main pointer and scans forward
3109 looking for null transitions that are allocated to a state. When it
3110 finds one it writes the single transition into the "hole". If the
3111 pointer doesnt find one the single transition is appended as normal.
3113 - Once compressed we can Renew/realloc the structures to release the
3116 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3117 specifically Fig 3.47 and the associated pseudocode.
3121 const U32 laststate = TRIE_NODENUM( next_alloc );
3124 trie->statecount = laststate;
3126 for ( state = 1 ; state < laststate ; state++ ) {
3128 const U32 stateidx = TRIE_NODEIDX( state );
3129 const U32 o_used = trie->trans[ stateidx ].check;
3130 U32 used = trie->trans[ stateidx ].check;
3131 trie->trans[ stateidx ].check = 0;
3134 used && charid < trie->uniquecharcount;
3137 if ( flag || trie->trans[ stateidx + charid ].next ) {
3138 if ( trie->trans[ stateidx + charid ].next ) {
3140 for ( ; zp < pos ; zp++ ) {
3141 if ( ! trie->trans[ zp ].next ) {
3145 trie->states[ state ].trans.base
3147 + trie->uniquecharcount
3149 trie->trans[ zp ].next
3150 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3152 trie->trans[ zp ].check = state;
3153 if ( ++zp > pos ) pos = zp;
3160 trie->states[ state ].trans.base
3161 = pos + trie->uniquecharcount - charid ;
3163 trie->trans[ pos ].next
3164 = SAFE_TRIE_NODENUM(
3165 trie->trans[ stateidx + charid ].next );
3166 trie->trans[ pos ].check = state;
3171 trie->lasttrans = pos + 1;
3172 trie->states = (reg_trie_state *)
3173 PerlMemShared_realloc( trie->states, laststate
3174 * sizeof(reg_trie_state) );
3175 DEBUG_TRIE_COMPILE_MORE_r(
3176 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3178 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3182 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3185 } /* end table compress */
3187 DEBUG_TRIE_COMPILE_MORE_r(
3188 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3190 (UV)trie->statecount,
3191 (UV)trie->lasttrans)
3193 /* resize the trans array to remove unused space */
3194 trie->trans = (reg_trie_trans *)
3195 PerlMemShared_realloc( trie->trans, trie->lasttrans
3196 * sizeof(reg_trie_trans) );
3198 { /* Modify the program and insert the new TRIE node */
3199 U8 nodetype =(U8)(flags & 0xFF);
3203 regnode *optimize = NULL;
3204 #ifdef RE_TRACK_PATTERN_OFFSETS
3207 U32 mjd_nodelen = 0;
3208 #endif /* RE_TRACK_PATTERN_OFFSETS */
3209 #endif /* DEBUGGING */
3211 This means we convert either the first branch or the first Exact,
3212 depending on whether the thing following (in 'last') is a branch
3213 or not and whther first is the startbranch (ie is it a sub part of
3214 the alternation or is it the whole thing.)
3215 Assuming its a sub part we convert the EXACT otherwise we convert
3216 the whole branch sequence, including the first.
3218 /* Find the node we are going to overwrite */
3219 if ( first != startbranch || OP( last ) == BRANCH ) {
3220 /* branch sub-chain */
3221 NEXT_OFF( first ) = (U16)(last - first);
3222 #ifdef RE_TRACK_PATTERN_OFFSETS
3224 mjd_offset= Node_Offset((convert));
3225 mjd_nodelen= Node_Length((convert));
3228 /* whole branch chain */
3230 #ifdef RE_TRACK_PATTERN_OFFSETS
3233 const regnode *nop = NEXTOPER( convert );
3234 mjd_offset= Node_Offset((nop));
3235 mjd_nodelen= Node_Length((nop));
3239 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3241 (UV)mjd_offset, (UV)mjd_nodelen)
3244 /* But first we check to see if there is a common prefix we can
3245 split out as an EXACT and put in front of the TRIE node. */
3246 trie->startstate= 1;
3247 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3248 /* we want to find the first state that has more than
3249 * one transition, if that state is not the first state
3250 * then we have a common prefix which we can remove.
3253 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3255 I32 first_ofs = -1; /* keeps track of the ofs of the first
3256 transition, -1 means none */
3258 const U32 base = trie->states[ state ].trans.base;
3260 /* does this state terminate an alternation? */
3261 if ( trie->states[state].wordnum )
3264 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3265 if ( ( base + ofs >= trie->uniquecharcount ) &&
3266 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3267 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3269 if ( ++count > 1 ) {
3270 /* we have more than one transition */
3273 /* if this is the first state there is no common prefix
3274 * to extract, so we can exit */
3275 if ( state == 1 ) break;
3276 tmp = av_fetch( revcharmap, ofs, 0);
3277 ch = (U8*)SvPV_nolen_const( *tmp );
3279 /* if we are on count 2 then we need to initialize the
3280 * bitmap, and store the previous char if there was one
3283 /* clear the bitmap */
3284 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3286 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3289 if (first_ofs >= 0) {
3290 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3291 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3293 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3295 Perl_re_printf( aTHX_ "%s", (char*)ch)
3299 /* store the current firstchar in the bitmap */
3300 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3301 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3307 /* This state has only one transition, its transition is part
3308 * of a common prefix - we need to concatenate the char it
3309 * represents to what we have so far. */
3310 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3312 char *ch = SvPV( *tmp, len );
3314 SV *sv=sv_newmortal();
3315 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3317 (UV)state, (UV)first_ofs,
3318 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3319 PL_colors[0], PL_colors[1],
3320 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3321 PERL_PV_ESCAPE_FIRSTCHAR
3326 OP( convert ) = nodetype;
3327 str=STRING(convert);
3330 STR_LEN(convert) += len;
3336 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3341 trie->prefixlen = (state-1);
3343 regnode *n = convert+NODE_SZ_STR(convert);
3344 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3345 trie->startstate = state;
3346 trie->minlen -= (state - 1);
3347 trie->maxlen -= (state - 1);
3349 /* At least the UNICOS C compiler choked on this
3350 * being argument to DEBUG_r(), so let's just have
3353 #ifdef PERL_EXT_RE_BUILD
3359 regnode *fix = convert;
3360 U32 word = trie->wordcount;
3362 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3363 while( ++fix < n ) {
3364 Set_Node_Offset_Length(fix, 0, 0);
3367 SV ** const tmp = av_fetch( trie_words, word, 0 );
3369 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3370 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3372 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3380 NEXT_OFF(convert) = (U16)(tail - convert);
3381 DEBUG_r(optimize= n);
3387 if ( trie->maxlen ) {
3388 NEXT_OFF( convert ) = (U16)(tail - convert);
3389 ARG_SET( convert, data_slot );
3390 /* Store the offset to the first unabsorbed branch in
3391 jump[0], which is otherwise unused by the jump logic.
3392 We use this when dumping a trie and during optimisation. */
3394 trie->jump[0] = (U16)(nextbranch - convert);
3396 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3397 * and there is a bitmap
3398 * and the first "jump target" node we found leaves enough room
3399 * then convert the TRIE node into a TRIEC node, with the bitmap
3400 * embedded inline in the opcode - this is hypothetically faster.
3402 if ( !trie->states[trie->startstate].wordnum
3404 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3406 OP( convert ) = TRIEC;
3407 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3408 PerlMemShared_free(trie->bitmap);
3411 OP( convert ) = TRIE;
3413 /* store the type in the flags */
3414 convert->flags = nodetype;
3418 + regarglen[ OP( convert ) ];
3420 /* XXX We really should free up the resource in trie now,
3421 as we won't use them - (which resources?) dmq */
3423 /* needed for dumping*/
3424 DEBUG_r(if (optimize) {
3425 regnode *opt = convert;
3427 while ( ++opt < optimize) {
3428 Set_Node_Offset_Length(opt,0,0);
3431 Try to clean up some of the debris left after the
3434 while( optimize < jumper ) {
3435 mjd_nodelen += Node_Length((optimize));
3436 OP( optimize ) = OPTIMIZED;
3437 Set_Node_Offset_Length(optimize,0,0);
3440 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3442 } /* end node insert */
3444 /* Finish populating the prev field of the wordinfo array. Walk back
3445 * from each accept state until we find another accept state, and if
3446 * so, point the first word's .prev field at the second word. If the
3447 * second already has a .prev field set, stop now. This will be the
3448 * case either if we've already processed that word's accept state,
3449 * or that state had multiple words, and the overspill words were
3450 * already linked up earlier.
3457 for (word=1; word <= trie->wordcount; word++) {
3459 if (trie->wordinfo[word].prev)
3461 state = trie->wordinfo[word].accept;
3463 state = prev_states[state];
3466 prev = trie->states[state].wordnum;
3470 trie->wordinfo[word].prev = prev;
3472 Safefree(prev_states);
3476 /* and now dump out the compressed format */
3477 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3479 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3481 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3482 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3484 SvREFCNT_dec_NN(revcharmap);
3488 : trie->startstate>1
3494 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3496 /* The Trie is constructed and compressed now so we can build a fail array if
3499 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3501 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3505 We find the fail state for each state in the trie, this state is the longest
3506 proper suffix of the current state's 'word' that is also a proper prefix of
3507 another word in our trie. State 1 represents the word '' and is thus the
3508 default fail state. This allows the DFA not to have to restart after its
3509 tried and failed a word at a given point, it simply continues as though it
3510 had been matching the other word in the first place.
3512 'abcdgu'=~/abcdefg|cdgu/
3513 When we get to 'd' we are still matching the first word, we would encounter
3514 'g' which would fail, which would bring us to the state representing 'd' in
3515 the second word where we would try 'g' and succeed, proceeding to match
3518 /* add a fail transition */
3519 const U32 trie_offset = ARG(source);
3520 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3522 const U32 ucharcount = trie->uniquecharcount;
3523 const U32 numstates = trie->statecount;
3524 const U32 ubound = trie->lasttrans + ucharcount;
3528 U32 base = trie->states[ 1 ].trans.base;
3531 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3533 GET_RE_DEBUG_FLAGS_DECL;
3535 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3536 PERL_UNUSED_CONTEXT;
3538 PERL_UNUSED_ARG(depth);
3541 if ( OP(source) == TRIE ) {
3542 struct regnode_1 *op = (struct regnode_1 *)
3543 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3544 StructCopy(source,op,struct regnode_1);
3545 stclass = (regnode *)op;
3547 struct regnode_charclass *op = (struct regnode_charclass *)
3548 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3549 StructCopy(source,op,struct regnode_charclass);
3550 stclass = (regnode *)op;
3552 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3554 ARG_SET( stclass, data_slot );
3555 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3556 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3557 aho->trie=trie_offset;
3558 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3559 Copy( trie->states, aho->states, numstates, reg_trie_state );
3560 Newxz( q, numstates, U32);
3561 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3564 /* initialize fail[0..1] to be 1 so that we always have
3565 a valid final fail state */
3566 fail[ 0 ] = fail[ 1 ] = 1;
3568 for ( charid = 0; charid < ucharcount ; charid++ ) {
3569 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3571 q[ q_write ] = newstate;
3572 /* set to point at the root */
3573 fail[ q[ q_write++ ] ]=1;
3576 while ( q_read < q_write) {
3577 const U32 cur = q[ q_read++ % numstates ];
3578 base = trie->states[ cur ].trans.base;
3580 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3581 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3583 U32 fail_state = cur;
3586 fail_state = fail[ fail_state ];
3587 fail_base = aho->states[ fail_state ].trans.base;
3588 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3590 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3591 fail[ ch_state ] = fail_state;
3592 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3594 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3596 q[ q_write++ % numstates] = ch_state;
3600 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3601 when we fail in state 1, this allows us to use the
3602 charclass scan to find a valid start char. This is based on the principle
3603 that theres a good chance the string being searched contains lots of stuff
3604 that cant be a start char.
3606 fail[ 0 ] = fail[ 1 ] = 0;
3607 DEBUG_TRIE_COMPILE_r({
3608 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3609 depth, (UV)numstates
3611 for( q_read=1; q_read<numstates; q_read++ ) {
3612 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3614 Perl_re_printf( aTHX_ "\n");
3617 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3622 #define DEBUG_PEEP(str,scan,depth) \
3623 DEBUG_OPTIMISE_r({if (scan){ \
3624 regnode *Next = regnext(scan); \
3625 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);\
3626 Perl_re_indentf( aTHX_ "" str ">%3d: %s (%d)", \
3627 depth, REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3628 Next ? (REG_NODE_NUM(Next)) : 0 );\
3629 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3630 Perl_re_printf( aTHX_ "\n"); \
3633 /* The below joins as many adjacent EXACTish nodes as possible into a single
3634 * one. The regop may be changed if the node(s) contain certain sequences that
3635 * require special handling. The joining is only done if:
3636 * 1) there is room in the current conglomerated node to entirely contain the
3638 * 2) they are the exact same node type
3640 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3641 * these get optimized out
3643 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3644 * as possible, even if that means splitting an existing node so that its first
3645 * part is moved to the preceeding node. This would maximise the efficiency of
3646 * memEQ during matching. Elsewhere in this file, khw proposes splitting
3647 * EXACTFish nodes into portions that don't change under folding vs those that
3648 * do. Those portions that don't change may be the only things in the pattern that
3649 * could be used to find fixed and floating strings.
3651 * If a node is to match under /i (folded), the number of characters it matches
3652 * can be different than its character length if it contains a multi-character
3653 * fold. *min_subtract is set to the total delta number of characters of the
3656 * And *unfolded_multi_char is set to indicate whether or not the node contains
3657 * an unfolded multi-char fold. This happens when whether the fold is valid or
3658 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3659 * SMALL LETTER SHARP S, as only if the target string being matched against
3660 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3661 * folding rules depend on the locale in force at runtime. (Multi-char folds
3662 * whose components are all above the Latin1 range are not run-time locale
3663 * dependent, and have already been folded by the time this function is
3666 * This is as good a place as any to discuss the design of handling these
3667 * multi-character fold sequences. It's been wrong in Perl for a very long
3668 * time. There are three code points in Unicode whose multi-character folds
3669 * were long ago discovered to mess things up. The previous designs for
3670 * dealing with these involved assigning a special node for them. This
3671 * approach doesn't always work, as evidenced by this example:
3672 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3673 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3674 * would match just the \xDF, it won't be able to handle the case where a
3675 * successful match would have to cross the node's boundary. The new approach
3676 * that hopefully generally solves the problem generates an EXACTFU_SS node
3677 * that is "sss" in this case.
3679 * It turns out that there are problems with all multi-character folds, and not
3680 * just these three. Now the code is general, for all such cases. The
3681 * approach taken is:
3682 * 1) This routine examines each EXACTFish node that could contain multi-
3683 * character folded sequences. Since a single character can fold into
3684 * such a sequence, the minimum match length for this node is less than
3685 * the number of characters in the node. This routine returns in
3686 * *min_subtract how many characters to subtract from the the actual
3687 * length of the string to get a real minimum match length; it is 0 if
3688 * there are no multi-char foldeds. This delta is used by the caller to
3689 * adjust the min length of the match, and the delta between min and max,
3690 * so that the optimizer doesn't reject these possibilities based on size
3692 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3693 * is used for an EXACTFU node that contains at least one "ss" sequence in
3694 * it. For non-UTF-8 patterns and strings, this is the only case where
3695 * there is a possible fold length change. That means that a regular
3696 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3697 * with length changes, and so can be processed faster. regexec.c takes
3698 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3699 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3700 * known until runtime). This saves effort in regex matching. However,
3701 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3702 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3703 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3704 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3705 * possibilities for the non-UTF8 patterns are quite simple, except for
3706 * the sharp s. All the ones that don't involve a UTF-8 target string are
3707 * members of a fold-pair, and arrays are set up for all of them so that
3708 * the other member of the pair can be found quickly. Code elsewhere in
3709 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3710 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3711 * described in the next item.
3712 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3713 * validity of the fold won't be known until runtime, and so must remain
3714 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3715 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3716 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3717 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3718 * The reason this is a problem is that the optimizer part of regexec.c
3719 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3720 * that a character in the pattern corresponds to at most a single
3721 * character in the target string. (And I do mean character, and not byte
3722 * here, unlike other parts of the documentation that have never been
3723 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3724 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3725 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3726 * nodes, violate the assumption, and they are the only instances where it
3727 * is violated. I'm reluctant to try to change the assumption, as the
3728 * code involved is impenetrable to me (khw), so instead the code here
3729 * punts. This routine examines EXACTFL nodes, and (when the pattern
3730 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3731 * boolean indicating whether or not the node contains such a fold. When
3732 * it is true, the caller sets a flag that later causes the optimizer in
3733 * this file to not set values for the floating and fixed string lengths,
3734 * and thus avoids the optimizer code in regexec.c that makes the invalid
3735 * assumption. Thus, there is no optimization based on string lengths for
3736 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3737 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3738 * assumption is wrong only in these cases is that all other non-UTF-8
3739 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3740 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3741 * EXACTF nodes because we don't know at compile time if it actually
3742 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3743 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3744 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3745 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3746 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3747 * string would require the pattern to be forced into UTF-8, the overhead
3748 * of which we want to avoid. Similarly the unfolded multi-char folds in
3749 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3752 * Similarly, the code that generates tries doesn't currently handle
3753 * not-already-folded multi-char folds, and it looks like a pain to change
3754 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3755 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3756 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3757 * using /iaa matching will be doing so almost entirely with ASCII
3758 * strings, so this should rarely be encountered in practice */
3760 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3761 if (PL_regkind[OP(scan)] == EXACT) \
3762 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3765 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3766 UV *min_subtract, bool *unfolded_multi_char,
3767 U32 flags,regnode *val, U32 depth)
3769 /* Merge several consecutive EXACTish nodes into one. */
3770 regnode *n = regnext(scan);
3772 regnode *next = scan + NODE_SZ_STR(scan);
3776 regnode *stop = scan;
3777 GET_RE_DEBUG_FLAGS_DECL;
3779 PERL_UNUSED_ARG(depth);
3782 PERL_ARGS_ASSERT_JOIN_EXACT;
3783 #ifndef EXPERIMENTAL_INPLACESCAN
3784 PERL_UNUSED_ARG(flags);
3785 PERL_UNUSED_ARG(val);
3787 DEBUG_PEEP("join",scan,depth);
3789 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3790 * EXACT ones that are mergeable to the current one. */
3792 && (PL_regkind[OP(n)] == NOTHING
3793 || (stringok && OP(n) == OP(scan)))
3795 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3798 if (OP(n) == TAIL || n > next)
3800 if (PL_regkind[OP(n)] == NOTHING) {
3801 DEBUG_PEEP("skip:",n,depth);
3802 NEXT_OFF(scan) += NEXT_OFF(n);
3803 next = n + NODE_STEP_REGNODE;
3810 else if (stringok) {
3811 const unsigned int oldl = STR_LEN(scan);
3812 regnode * const nnext = regnext(n);
3814 /* XXX I (khw) kind of doubt that this works on platforms (should
3815 * Perl ever run on one) where U8_MAX is above 255 because of lots
3816 * of other assumptions */
3817 /* Don't join if the sum can't fit into a single node */
3818 if (oldl + STR_LEN(n) > U8_MAX)
3821 DEBUG_PEEP("merg",n,depth);
3824 NEXT_OFF(scan) += NEXT_OFF(n);
3825 STR_LEN(scan) += STR_LEN(n);
3826 next = n + NODE_SZ_STR(n);
3827 /* Now we can overwrite *n : */
3828 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3836 #ifdef EXPERIMENTAL_INPLACESCAN
3837 if (flags && !NEXT_OFF(n)) {
3838 DEBUG_PEEP("atch", val, depth);
3839 if (reg_off_by_arg[OP(n)]) {
3840 ARG_SET(n, val - n);
3843 NEXT_OFF(n) = val - n;
3851 *unfolded_multi_char = FALSE;
3853 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3854 * can now analyze for sequences of problematic code points. (Prior to
3855 * this final joining, sequences could have been split over boundaries, and
3856 * hence missed). The sequences only happen in folding, hence for any
3857 * non-EXACT EXACTish node */
3858 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3859 U8* s0 = (U8*) STRING(scan);
3861 U8* s_end = s0 + STR_LEN(scan);
3863 int total_count_delta = 0; /* Total delta number of characters that
3864 multi-char folds expand to */
3866 /* One pass is made over the node's string looking for all the
3867 * possibilities. To avoid some tests in the loop, there are two main
3868 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3873 if (OP(scan) == EXACTFL) {
3876 /* An EXACTFL node would already have been changed to another
3877 * node type unless there is at least one character in it that
3878 * is problematic; likely a character whose fold definition
3879 * won't be known until runtime, and so has yet to be folded.
3880 * For all but the UTF-8 locale, folds are 1-1 in length, but
3881 * to handle the UTF-8 case, we need to create a temporary
3882 * folded copy using UTF-8 locale rules in order to analyze it.
3883 * This is because our macros that look to see if a sequence is
3884 * a multi-char fold assume everything is folded (otherwise the
3885 * tests in those macros would be too complicated and slow).
3886 * Note that here, the non-problematic folds will have already
3887 * been done, so we can just copy such characters. We actually
3888 * don't completely fold the EXACTFL string. We skip the
3889 * unfolded multi-char folds, as that would just create work
3890 * below to figure out the size they already are */
3892 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3895 STRLEN s_len = UTF8SKIP(s);
3896 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3897 Copy(s, d, s_len, U8);
3900 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3901 *unfolded_multi_char = TRUE;
3902 Copy(s, d, s_len, U8);
3905 else if (isASCII(*s)) {
3906 *(d++) = toFOLD(*s);
3910 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
3916 /* Point the remainder of the routine to look at our temporary
3920 } /* End of creating folded copy of EXACTFL string */
3922 /* Examine the string for a multi-character fold sequence. UTF-8
3923 * patterns have all characters pre-folded by the time this code is
3925 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3926 length sequence we are looking for is 2 */
3928 int count = 0; /* How many characters in a multi-char fold */
3929 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3930 if (! len) { /* Not a multi-char fold: get next char */
3935 /* Nodes with 'ss' require special handling, except for
3936 * EXACTFA-ish for which there is no multi-char fold to this */
3937 if (len == 2 && *s == 's' && *(s+1) == 's'
3938 && OP(scan) != EXACTFA
3939 && OP(scan) != EXACTFA_NO_TRIE)
3942 if (OP(scan) != EXACTFL) {
3943 OP(scan) = EXACTFU_SS;
3947 else { /* Here is a generic multi-char fold. */
3948 U8* multi_end = s + len;
3950 /* Count how many characters are in it. In the case of
3951 * /aa, no folds which contain ASCII code points are
3952 * allowed, so check for those, and skip if found. */
3953 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3954 count = utf8_length(s, multi_end);
3958 while (s < multi_end) {
3961 goto next_iteration;
3971 /* The delta is how long the sequence is minus 1 (1 is how long
3972 * the character that folds to the sequence is) */
3973 total_count_delta += count - 1;
3977 /* We created a temporary folded copy of the string in EXACTFL
3978 * nodes. Therefore we need to be sure it doesn't go below zero,
3979 * as the real string could be shorter */
3980 if (OP(scan) == EXACTFL) {
3981 int total_chars = utf8_length((U8*) STRING(scan),
3982 (U8*) STRING(scan) + STR_LEN(scan));
3983 if (total_count_delta > total_chars) {
3984 total_count_delta = total_chars;
3988 *min_subtract += total_count_delta;
3991 else if (OP(scan) == EXACTFA) {
3993 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3994 * fold to the ASCII range (and there are no existing ones in the
3995 * upper latin1 range). But, as outlined in the comments preceding
3996 * this function, we need to flag any occurrences of the sharp s.
3997 * This character forbids trie formation (because of added
3999 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4000 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4001 || UNICODE_DOT_DOT_VERSION > 0)
4003 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4004 OP(scan) = EXACTFA_NO_TRIE;
4005 *unfolded_multi_char = TRUE;
4013 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
4014 * folds that are all Latin1. As explained in the comments
4015 * preceding this function, we look also for the sharp s in EXACTF
4016 * and EXACTFL nodes; it can be in the final position. Otherwise
4017 * we can stop looking 1 byte earlier because have to find at least
4018 * two characters for a multi-fold */
4019 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4024 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4025 if (! len) { /* Not a multi-char fold. */
4026 if (*s == LATIN_SMALL_LETTER_SHARP_S
4027 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4029 *unfolded_multi_char = TRUE;
4036 && isALPHA_FOLD_EQ(*s, 's')
4037 && isALPHA_FOLD_EQ(*(s+1), 's'))
4040 /* EXACTF nodes need to know that the minimum length
4041 * changed so that a sharp s in the string can match this
4042 * ss in the pattern, but they remain EXACTF nodes, as they
4043 * won't match this unless the target string is is UTF-8,
4044 * which we don't know until runtime. EXACTFL nodes can't
4045 * transform into EXACTFU nodes */
4046 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4047 OP(scan) = EXACTFU_SS;
4051 *min_subtract += len - 1;
4059 /* Allow dumping but overwriting the collection of skipped
4060 * ops and/or strings with fake optimized ops */
4061 n = scan + NODE_SZ_STR(scan);
4069 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
4073 /* REx optimizer. Converts nodes into quicker variants "in place".
4074 Finds fixed substrings. */
4076 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4077 to the position after last scanned or to NULL. */
4079 #define INIT_AND_WITHP \
4080 assert(!and_withp); \
4081 Newx(and_withp,1, regnode_ssc); \
4082 SAVEFREEPV(and_withp)
4086 S_unwind_scan_frames(pTHX_ const void *p)
4088 scan_frame *f= (scan_frame *)p;
4090 scan_frame *n= f->next_frame;
4098 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4099 SSize_t *minlenp, SSize_t *deltap,
4104 regnode_ssc *and_withp,
4105 U32 flags, U32 depth)
4106 /* scanp: Start here (read-write). */
4107 /* deltap: Write maxlen-minlen here. */
4108 /* last: Stop before this one. */
4109 /* data: string data about the pattern */
4110 /* stopparen: treat close N as END */
4111 /* recursed: which subroutines have we recursed into */
4112 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4114 /* There must be at least this number of characters to match */
4117 regnode *scan = *scanp, *next;
4119 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4120 int is_inf_internal = 0; /* The studied chunk is infinite */
4121 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4122 scan_data_t data_fake;
4123 SV *re_trie_maxbuff = NULL;
4124 regnode *first_non_open = scan;
4125 SSize_t stopmin = SSize_t_MAX;
4126 scan_frame *frame = NULL;
4127 GET_RE_DEBUG_FLAGS_DECL;
4129 PERL_ARGS_ASSERT_STUDY_CHUNK;
4130 RExC_study_started= 1;
4134 while (first_non_open && OP(first_non_open) == OPEN)
4135 first_non_open=regnext(first_non_open);
4141 RExC_study_chunk_recursed_count++;
4143 DEBUG_OPTIMISE_MORE_r(
4145 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4146 depth, (long)stopparen,
4147 (unsigned long)RExC_study_chunk_recursed_count,
4148 (unsigned long)depth, (unsigned long)recursed_depth,
4151 if (recursed_depth) {
4154 for ( j = 0 ; j < recursed_depth ; j++ ) {
4155 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
4157 PAREN_TEST(RExC_study_chunk_recursed +
4158 ( j * RExC_study_chunk_recursed_bytes), i )
4161 !PAREN_TEST(RExC_study_chunk_recursed +
4162 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
4165 Perl_re_printf( aTHX_ " %d",(int)i);
4169 if ( j + 1 < recursed_depth ) {
4170 Perl_re_printf( aTHX_ ",");
4174 Perl_re_printf( aTHX_ "\n");
4177 while ( scan && OP(scan) != END && scan < last ){
4178 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4179 node length to get a real minimum (because
4180 the folded version may be shorter) */
4181 bool unfolded_multi_char = FALSE;
4182 /* Peephole optimizer: */
4183 DEBUG_STUDYDATA("Peep:", data, depth);
4184 DEBUG_PEEP("Peep", scan, depth);
4187 /* The reason we do this here is that we need to deal with things like
4188 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4189 * parsing code, as each (?:..) is handled by a different invocation of
4192 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
4194 /* Follow the next-chain of the current node and optimize
4195 away all the NOTHINGs from it. */
4196 if (OP(scan) != CURLYX) {
4197 const int max = (reg_off_by_arg[OP(scan)]
4199 /* I32 may be smaller than U16 on CRAYs! */
4200 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4201 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
4205 /* Skip NOTHING and LONGJMP. */
4206 while ((n = regnext(n))
4207 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4208 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
4209 && off + noff < max)
4211 if (reg_off_by_arg[OP(scan)])
4214 NEXT_OFF(scan) = off;
4217 /* The principal pseudo-switch. Cannot be a switch, since we
4218 look into several different things. */
4219 if ( OP(scan) == DEFINEP ) {
4221 SSize_t deltanext = 0;
4222 SSize_t fake_last_close = 0;
4223 I32 f = SCF_IN_DEFINE;
4225 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4226 scan = regnext(scan);
4227 assert( OP(scan) == IFTHEN );
4228 DEBUG_PEEP("expect IFTHEN", scan, depth);
4230 data_fake.last_closep= &fake_last_close;
4232 next = regnext(scan);
4233 scan = NEXTOPER(NEXTOPER(scan));
4234 DEBUG_PEEP("scan", scan, depth);
4235 DEBUG_PEEP("next", next, depth);
4237 /* we suppose the run is continuous, last=next...
4238 * NOTE we dont use the return here! */
4239 (void)study_chunk(pRExC_state, &scan, &minlen,
4240 &deltanext, next, &data_fake, stopparen,
4241 recursed_depth, NULL, f, depth+1);
4246 OP(scan) == BRANCH ||
4247 OP(scan) == BRANCHJ ||
4250 next = regnext(scan);
4253 /* The op(next)==code check below is to see if we
4254 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4255 * IFTHEN is special as it might not appear in pairs.
4256 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4257 * we dont handle it cleanly. */
4258 if (OP(next) == code || code == IFTHEN) {
4259 /* NOTE - There is similar code to this block below for
4260 * handling TRIE nodes on a re-study. If you change stuff here
4261 * check there too. */
4262 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
4264 regnode * const startbranch=scan;
4266 if (flags & SCF_DO_SUBSTR) {
4267 /* Cannot merge strings after this. */
4268 scan_commit(pRExC_state, data, minlenp, is_inf);
4271 if (flags & SCF_DO_STCLASS)
4272 ssc_init_zero(pRExC_state, &accum);
4274 while (OP(scan) == code) {
4275 SSize_t deltanext, minnext, fake;
4277 regnode_ssc this_class;
4279 DEBUG_PEEP("Branch", scan, depth);
4282 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4284 data_fake.whilem_c = data->whilem_c;
4285 data_fake.last_closep = data->last_closep;
4288 data_fake.last_closep = &fake;
4290 data_fake.pos_delta = delta;
4291 next = regnext(scan);
4293 scan = NEXTOPER(scan); /* everything */
4294 if (code != BRANCH) /* everything but BRANCH */
4295 scan = NEXTOPER(scan);
4297 if (flags & SCF_DO_STCLASS) {
4298 ssc_init(pRExC_state, &this_class);
4299 data_fake.start_class = &this_class;
4300 f = SCF_DO_STCLASS_AND;
4302 if (flags & SCF_WHILEM_VISITED_POS)
4303 f |= SCF_WHILEM_VISITED_POS;
4305 /* we suppose the run is continuous, last=next...*/
4306 minnext = study_chunk(pRExC_state, &scan, minlenp,
4307 &deltanext, next, &data_fake, stopparen,
4308 recursed_depth, NULL, f,depth+1);
4312 if (deltanext == SSize_t_MAX) {
4313 is_inf = is_inf_internal = 1;
4315 } else if (max1 < minnext + deltanext)
4316 max1 = minnext + deltanext;
4318 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4320 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4321 if ( stopmin > minnext)
4322 stopmin = min + min1;
4323 flags &= ~SCF_DO_SUBSTR;
4325 data->flags |= SCF_SEEN_ACCEPT;
4328 if (data_fake.flags & SF_HAS_EVAL)
4329 data->flags |= SF_HAS_EVAL;
4330 data->whilem_c = data_fake.whilem_c;
4332 if (flags & SCF_DO_STCLASS)
4333 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4335 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4337 if (flags & SCF_DO_SUBSTR) {
4338 data->pos_min += min1;
4339 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
4340 data->pos_delta = SSize_t_MAX;
4342 data->pos_delta += max1 - min1;
4343 if (max1 != min1 || is_inf)
4344 data->longest = &(data->longest_float);
4347 if (delta == SSize_t_MAX
4348 || SSize_t_MAX - delta - (max1 - min1) < 0)
4349 delta = SSize_t_MAX;
4351 delta += max1 - min1;
4352 if (flags & SCF_DO_STCLASS_OR) {
4353 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4355 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4356 flags &= ~SCF_DO_STCLASS;
4359 else if (flags & SCF_DO_STCLASS_AND) {
4361 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4362 flags &= ~SCF_DO_STCLASS;
4365 /* Switch to OR mode: cache the old value of
4366 * data->start_class */
4368 StructCopy(data->start_class, and_withp, regnode_ssc);
4369 flags &= ~SCF_DO_STCLASS_AND;
4370 StructCopy(&accum, data->start_class, regnode_ssc);
4371 flags |= SCF_DO_STCLASS_OR;
4375 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4376 OP( startbranch ) == BRANCH )
4380 Assuming this was/is a branch we are dealing with: 'scan'
4381 now points at the item that follows the branch sequence,
4382 whatever it is. We now start at the beginning of the
4383 sequence and look for subsequences of
4389 which would be constructed from a pattern like
4392 If we can find such a subsequence we need to turn the first
4393 element into a trie and then add the subsequent branch exact
4394 strings to the trie.
4398 1. patterns where the whole set of branches can be
4401 2. patterns where only a subset can be converted.
4403 In case 1 we can replace the whole set with a single regop
4404 for the trie. In case 2 we need to keep the start and end
4407 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4408 becomes BRANCH TRIE; BRANCH X;
4410 There is an additional case, that being where there is a
4411 common prefix, which gets split out into an EXACT like node
4412 preceding the TRIE node.
4414 If x(1..n)==tail then we can do a simple trie, if not we make
4415 a "jump" trie, such that when we match the appropriate word
4416 we "jump" to the appropriate tail node. Essentially we turn
4417 a nested if into a case structure of sorts.
4422 if (!re_trie_maxbuff) {
4423 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4424 if (!SvIOK(re_trie_maxbuff))
4425 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4427 if ( SvIV(re_trie_maxbuff)>=0 ) {
4429 regnode *first = (regnode *)NULL;
4430 regnode *last = (regnode *)NULL;
4431 regnode *tail = scan;
4435 /* var tail is used because there may be a TAIL
4436 regop in the way. Ie, the exacts will point to the
4437 thing following the TAIL, but the last branch will
4438 point at the TAIL. So we advance tail. If we
4439 have nested (?:) we may have to move through several
4443 while ( OP( tail ) == TAIL ) {
4444 /* this is the TAIL generated by (?:) */
4445 tail = regnext( tail );
4449 DEBUG_TRIE_COMPILE_r({
4450 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4451 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4453 "Looking for TRIE'able sequences. Tail node is ",
4454 (UV)(tail - RExC_emit_start),
4455 SvPV_nolen_const( RExC_mysv )
4461 Step through the branches
4462 cur represents each branch,
4463 noper is the first thing to be matched as part
4465 noper_next is the regnext() of that node.
4467 We normally handle a case like this
4468 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4469 support building with NOJUMPTRIE, which restricts
4470 the trie logic to structures like /FOO|BAR/.
4472 If noper is a trieable nodetype then the branch is
4473 a possible optimization target. If we are building
4474 under NOJUMPTRIE then we require that noper_next is
4475 the same as scan (our current position in the regex
4478 Once we have two or more consecutive such branches
4479 we can create a trie of the EXACT's contents and
4480 stitch it in place into the program.
4482 If the sequence represents all of the branches in
4483 the alternation we replace the entire thing with a
4486 Otherwise when it is a subsequence we need to
4487 stitch it in place and replace only the relevant
4488 branches. This means the first branch has to remain
4489 as it is used by the alternation logic, and its
4490 next pointer, and needs to be repointed at the item
4491 on the branch chain following the last branch we
4492 have optimized away.
4494 This could be either a BRANCH, in which case the
4495 subsequence is internal, or it could be the item
4496 following the branch sequence in which case the
4497 subsequence is at the end (which does not
4498 necessarily mean the first node is the start of the
4501 TRIE_TYPE(X) is a define which maps the optype to a
4505 ----------------+-----------
4509 EXACTFU_SS | EXACTFU
4512 EXACTFLU8 | EXACTFLU8
4516 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4518 : ( EXACT == (X) ) \
4520 : ( EXACTFU == (X) || EXACTFU_SS == (X) ) \
4522 : ( EXACTFA == (X) ) \
4524 : ( EXACTL == (X) ) \
4526 : ( EXACTFLU8 == (X) ) \
4530 /* dont use tail as the end marker for this traverse */
4531 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4532 regnode * const noper = NEXTOPER( cur );
4533 U8 noper_type = OP( noper );
4534 U8 noper_trietype = TRIE_TYPE( noper_type );
4535 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4536 regnode * const noper_next = regnext( noper );
4537 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4538 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4541 DEBUG_TRIE_COMPILE_r({
4542 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4543 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4545 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4547 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4548 Perl_re_printf( aTHX_ " -> %d:%s",
4549 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
4552 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4553 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
4554 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
4556 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
4557 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4558 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4562 /* Is noper a trieable nodetype that can be merged
4563 * with the current trie (if there is one)? */
4567 ( noper_trietype == NOTHING )
4568 || ( trietype == NOTHING )
4569 || ( trietype == noper_trietype )
4572 && noper_next >= tail
4576 /* Handle mergable triable node Either we are
4577 * the first node in a new trieable sequence,
4578 * in which case we do some bookkeeping,
4579 * otherwise we update the end pointer. */
4582 if ( noper_trietype == NOTHING ) {
4583 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4584 regnode * const noper_next = regnext( noper );
4585 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4586 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4589 if ( noper_next_trietype ) {
4590 trietype = noper_next_trietype;
4591 } else if (noper_next_type) {
4592 /* a NOTHING regop is 1 regop wide.
4593 * We need at least two for a trie
4594 * so we can't merge this in */
4598 trietype = noper_trietype;
4601 if ( trietype == NOTHING )
4602 trietype = noper_trietype;
4607 } /* end handle mergable triable node */
4609 /* handle unmergable node -
4610 * noper may either be a triable node which can
4611 * not be tried together with the current trie,
4612 * or a non triable node */
4614 /* If last is set and trietype is not
4615 * NOTHING then we have found at least two
4616 * triable branch sequences in a row of a
4617 * similar trietype so we can turn them
4618 * into a trie. If/when we allow NOTHING to
4619 * start a trie sequence this condition
4620 * will be required, and it isn't expensive
4621 * so we leave it in for now. */
4622 if ( trietype && trietype != NOTHING )
4623 make_trie( pRExC_state,
4624 startbranch, first, cur, tail,
4625 count, trietype, depth+1 );
4626 last = NULL; /* note: we clear/update
4627 first, trietype etc below,
4628 so we dont do it here */
4632 && noper_next >= tail
4635 /* noper is triable, so we can start a new
4639 trietype = noper_trietype;
4641 /* if we already saw a first but the
4642 * current node is not triable then we have
4643 * to reset the first information. */
4648 } /* end handle unmergable node */
4649 } /* loop over branches */
4650 DEBUG_TRIE_COMPILE_r({
4651 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4652 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
4653 depth+1, SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4654 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
4655 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4656 PL_reg_name[trietype]
4660 if ( last && trietype ) {
4661 if ( trietype != NOTHING ) {
4662 /* the last branch of the sequence was part of
4663 * a trie, so we have to construct it here
4664 * outside of the loop */
4665 made= make_trie( pRExC_state, startbranch,
4666 first, scan, tail, count,
4667 trietype, depth+1 );
4668 #ifdef TRIE_STUDY_OPT
4669 if ( ((made == MADE_EXACT_TRIE &&
4670 startbranch == first)
4671 || ( first_non_open == first )) &&
4673 flags |= SCF_TRIE_RESTUDY;
4674 if ( startbranch == first
4677 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4682 /* at this point we know whatever we have is a
4683 * NOTHING sequence/branch AND if 'startbranch'
4684 * is 'first' then we can turn the whole thing
4687 if ( startbranch == first ) {
4689 /* the entire thing is a NOTHING sequence,
4690 * something like this: (?:|) So we can
4691 * turn it into a plain NOTHING op. */
4692 DEBUG_TRIE_COMPILE_r({
4693 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4694 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
4696 SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4699 OP(startbranch)= NOTHING;
4700 NEXT_OFF(startbranch)= tail - startbranch;
4701 for ( opt= startbranch + 1; opt < tail ; opt++ )
4705 } /* end if ( last) */
4706 } /* TRIE_MAXBUF is non zero */
4711 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4712 scan = NEXTOPER(NEXTOPER(scan));
4713 } else /* single branch is optimized. */
4714 scan = NEXTOPER(scan);
4716 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
4718 regnode *start = NULL;
4719 regnode *end = NULL;
4720 U32 my_recursed_depth= recursed_depth;
4722 if (OP(scan) != SUSPEND) { /* GOSUB */
4723 /* Do setup, note this code has side effects beyond
4724 * the rest of this block. Specifically setting
4725 * RExC_recurse[] must happen at least once during
4728 RExC_recurse[ARG2L(scan)] = scan;
4729 start = RExC_open_parens[paren];
4730 end = RExC_close_parens[paren];
4732 /* NOTE we MUST always execute the above code, even
4733 * if we do nothing with a GOSUB */
4735 ( flags & SCF_IN_DEFINE )
4738 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
4740 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
4743 /* no need to do anything here if we are in a define. */
4744 /* or we are after some kind of infinite construct
4745 * so we can skip recursing into this item.
4746 * Since it is infinite we will not change the maxlen
4747 * or delta, and if we miss something that might raise
4748 * the minlen it will merely pessimise a little.
4750 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
4751 * might result in a minlen of 1 and not of 4,
4752 * but this doesn't make us mismatch, just try a bit
4753 * harder than we should.
4755 scan= regnext(scan);
4762 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4764 /* it is quite possible that there are more efficient ways
4765 * to do this. We maintain a bitmap per level of recursion
4766 * of which patterns we have entered so we can detect if a
4767 * pattern creates a possible infinite loop. When we
4768 * recurse down a level we copy the previous levels bitmap
4769 * down. When we are at recursion level 0 we zero the top
4770 * level bitmap. It would be nice to implement a different
4771 * more efficient way of doing this. In particular the top
4772 * level bitmap may be unnecessary.
4774 if (!recursed_depth) {
4775 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4777 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4778 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4779 RExC_study_chunk_recursed_bytes, U8);
4781 /* we havent recursed into this paren yet, so recurse into it */
4782 DEBUG_STUDYDATA("gosub-set:", data,depth);
4783 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4784 my_recursed_depth= recursed_depth + 1;
4786 DEBUG_STUDYDATA("gosub-inf:", data,depth);
4787 /* some form of infinite recursion, assume infinite length
4789 if (flags & SCF_DO_SUBSTR) {
4790 scan_commit(pRExC_state, data, minlenp, is_inf);
4791 data->longest = &(data->longest_float);
4793 is_inf = is_inf_internal = 1;
4794 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4795 ssc_anything(data->start_class);
4796 flags &= ~SCF_DO_STCLASS;
4798 start= NULL; /* reset start so we dont recurse later on. */
4803 end = regnext(scan);
4806 scan_frame *newframe;
4808 if (!RExC_frame_last) {
4809 Newxz(newframe, 1, scan_frame);
4810 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
4811 RExC_frame_head= newframe;
4813 } else if (!RExC_frame_last->next_frame) {
4814 Newxz(newframe,1,scan_frame);
4815 RExC_frame_last->next_frame= newframe;
4816 newframe->prev_frame= RExC_frame_last;
4819 newframe= RExC_frame_last->next_frame;
4821 RExC_frame_last= newframe;
4823 newframe->next_regnode = regnext(scan);
4824 newframe->last_regnode = last;
4825 newframe->stopparen = stopparen;
4826 newframe->prev_recursed_depth = recursed_depth;
4827 newframe->this_prev_frame= frame;
4829 DEBUG_STUDYDATA("frame-new:",data,depth);
4830 DEBUG_PEEP("fnew", scan, depth);
4837 recursed_depth= my_recursed_depth;
4842 else if (OP(scan) == EXACT || OP(scan) == EXACTL) {
4843 SSize_t l = STR_LEN(scan);
4846 const U8 * const s = (U8*)STRING(scan);
4847 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4848 l = utf8_length(s, s + l);
4850 uc = *((U8*)STRING(scan));
4853 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4854 /* The code below prefers earlier match for fixed
4855 offset, later match for variable offset. */
4856 if (data->last_end == -1) { /* Update the start info. */
4857 data->last_start_min = data->pos_min;
4858 data->last_start_max = is_inf
4859 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4861 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4863 SvUTF8_on(data->last_found);
4865 SV * const sv = data->last_found;
4866 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4867 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4868 if (mg && mg->mg_len >= 0)
4869 mg->mg_len += utf8_length((U8*)STRING(scan),
4870 (U8*)STRING(scan)+STR_LEN(scan));
4872 data->last_end = data->pos_min + l;
4873 data->pos_min += l; /* As in the first entry. */
4874 data->flags &= ~SF_BEFORE_EOL;
4877 /* ANDing the code point leaves at most it, and not in locale, and
4878 * can't match null string */
4879 if (flags & SCF_DO_STCLASS_AND) {
4880 ssc_cp_and(data->start_class, uc);
4881 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4882 ssc_clear_locale(data->start_class);
4884 else if (flags & SCF_DO_STCLASS_OR) {
4885 ssc_add_cp(data->start_class, uc);
4886 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4888 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4889 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4891 flags &= ~SCF_DO_STCLASS;
4893 else if (PL_regkind[OP(scan)] == EXACT) {
4894 /* But OP != EXACT!, so is EXACTFish */
4895 SSize_t l = STR_LEN(scan);
4896 const U8 * s = (U8*)STRING(scan);
4898 /* Search for fixed substrings supports EXACT only. */
4899 if (flags & SCF_DO_SUBSTR) {
4901 scan_commit(pRExC_state, data, minlenp, is_inf);
4904 l = utf8_length(s, s + l);
4906 if (unfolded_multi_char) {
4907 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4909 min += l - min_subtract;
4911 delta += min_subtract;
4912 if (flags & SCF_DO_SUBSTR) {
4913 data->pos_min += l - min_subtract;
4914 if (data->pos_min < 0) {
4917 data->pos_delta += min_subtract;
4919 data->longest = &(data->longest_float);
4923 if (flags & SCF_DO_STCLASS) {
4924 SV* EXACTF_invlist = _make_exactf_invlist(pRExC_state, scan);
4926 assert(EXACTF_invlist);
4927 if (flags & SCF_DO_STCLASS_AND) {
4928 if (OP(scan) != EXACTFL)
4929 ssc_clear_locale(data->start_class);
4930 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4931 ANYOF_POSIXL_ZERO(data->start_class);
4932 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4934 else { /* SCF_DO_STCLASS_OR */
4935 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4936 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4938 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4939 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4941 flags &= ~SCF_DO_STCLASS;
4942 SvREFCNT_dec(EXACTF_invlist);
4945 else if (REGNODE_VARIES(OP(scan))) {
4946 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4947 I32 fl = 0, f = flags;
4948 regnode * const oscan = scan;
4949 regnode_ssc this_class;
4950 regnode_ssc *oclass = NULL;
4951 I32 next_is_eval = 0;
4953 switch (PL_regkind[OP(scan)]) {
4954 case WHILEM: /* End of (?:...)* . */
4955 scan = NEXTOPER(scan);
4958 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4959 next = NEXTOPER(scan);
4960 if (OP(next) == EXACT
4961 || OP(next) == EXACTL
4962 || (flags & SCF_DO_STCLASS))
4965 maxcount = REG_INFTY;
4966 next = regnext(scan);
4967 scan = NEXTOPER(scan);
4971 if (flags & SCF_DO_SUBSTR)
4976 if (flags & SCF_DO_STCLASS) {
4978 maxcount = REG_INFTY;
4979 next = regnext(scan);
4980 scan = NEXTOPER(scan);
4983 if (flags & SCF_DO_SUBSTR) {
4984 scan_commit(pRExC_state, data, minlenp, is_inf);
4985 /* Cannot extend fixed substrings */
4986 data->longest = &(data->longest_float);
4988 is_inf = is_inf_internal = 1;
4989 scan = regnext(scan);
4990 goto optimize_curly_tail;
4992 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
4993 && (scan->flags == stopparen))
4998 mincount = ARG1(scan);
4999 maxcount = ARG2(scan);
5001 next = regnext(scan);
5002 if (OP(scan) == CURLYX) {
5003 I32 lp = (data ? *(data->last_closep) : 0);
5004 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5006 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5007 next_is_eval = (OP(scan) == EVAL);
5009 if (flags & SCF_DO_SUBSTR) {
5011 scan_commit(pRExC_state, data, minlenp, is_inf);
5012 /* Cannot extend fixed substrings */
5013 pos_before = data->pos_min;
5017 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5019 data->flags |= SF_IS_INF;
5021 if (flags & SCF_DO_STCLASS) {
5022 ssc_init(pRExC_state, &this_class);
5023 oclass = data->start_class;
5024 data->start_class = &this_class;
5025 f |= SCF_DO_STCLASS_AND;
5026 f &= ~SCF_DO_STCLASS_OR;
5028 /* Exclude from super-linear cache processing any {n,m}
5029 regops for which the combination of input pos and regex
5030 pos is not enough information to determine if a match
5033 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5034 regex pos at the \s*, the prospects for a match depend not
5035 only on the input position but also on how many (bar\s*)
5036 repeats into the {4,8} we are. */
5037 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5038 f &= ~SCF_WHILEM_VISITED_POS;
5040 /* This will finish on WHILEM, setting scan, or on NULL: */
5041 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5042 last, data, stopparen, recursed_depth, NULL,
5044 ? (f & ~SCF_DO_SUBSTR)
5048 if (flags & SCF_DO_STCLASS)
5049 data->start_class = oclass;
5050 if (mincount == 0 || minnext == 0) {
5051 if (flags & SCF_DO_STCLASS_OR) {
5052 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5054 else if (flags & SCF_DO_STCLASS_AND) {
5055 /* Switch to OR mode: cache the old value of
5056 * data->start_class */
5058 StructCopy(data->start_class, and_withp, regnode_ssc);
5059 flags &= ~SCF_DO_STCLASS_AND;
5060 StructCopy(&this_class, data->start_class, regnode_ssc);
5061 flags |= SCF_DO_STCLASS_OR;
5062 ANYOF_FLAGS(data->start_class)
5063 |= SSC_MATCHES_EMPTY_STRING;
5065 } else { /* Non-zero len */
5066 if (flags & SCF_DO_STCLASS_OR) {
5067 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5068 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5070 else if (flags & SCF_DO_STCLASS_AND)
5071 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5072 flags &= ~SCF_DO_STCLASS;
5074 if (!scan) /* It was not CURLYX, but CURLY. */
5076 if (!(flags & SCF_TRIE_DOING_RESTUDY)
5077 /* ? quantifier ok, except for (?{ ... }) */
5078 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5079 && (minnext == 0) && (deltanext == 0)
5080 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5081 && maxcount <= REG_INFTY/3) /* Complement check for big
5084 /* Fatal warnings may leak the regexp without this: */
5085 SAVEFREESV(RExC_rx_sv);
5086 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5087 "Quantifier unexpected on zero-length expression "
5088 "in regex m/%" UTF8f "/",
5089 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5091 (void)ReREFCNT_inc(RExC_rx_sv);
5094 min += minnext * mincount;
5095 is_inf_internal |= deltanext == SSize_t_MAX
5096 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5097 is_inf |= is_inf_internal;
5099 delta = SSize_t_MAX;
5101 delta += (minnext + deltanext) * maxcount
5102 - minnext * mincount;
5104 /* Try powerful optimization CURLYX => CURLYN. */
5105 if ( OP(oscan) == CURLYX && data
5106 && data->flags & SF_IN_PAR
5107 && !(data->flags & SF_HAS_EVAL)
5108 && !deltanext && minnext == 1 ) {
5109 /* Try to optimize to CURLYN. */
5110 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5111 regnode * const nxt1 = nxt;
5118 if (!REGNODE_SIMPLE(OP(nxt))
5119 && !(PL_regkind[OP(nxt)] == EXACT
5120 && STR_LEN(nxt) == 1))
5126 if (OP(nxt) != CLOSE)
5128 if (RExC_open_parens) {
5129 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5130 RExC_close_parens[ARG(nxt1)]=nxt+2; /*close->while*/
5132 /* Now we know that nxt2 is the only contents: */
5133 oscan->flags = (U8)ARG(nxt);
5135 OP(nxt1) = NOTHING; /* was OPEN. */
5138 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5139 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5140 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5141 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5142 OP(nxt + 1) = OPTIMIZED; /* was count. */
5143 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5148 /* Try optimization CURLYX => CURLYM. */
5149 if ( OP(oscan) == CURLYX && data
5150 && !(data->flags & SF_HAS_PAR)
5151 && !(data->flags & SF_HAS_EVAL)
5152 && !deltanext /* atom is fixed width */
5153 && minnext != 0 /* CURLYM can't handle zero width */
5155 /* Nor characters whose fold at run-time may be
5156 * multi-character */
5157 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5159 /* XXXX How to optimize if data == 0? */
5160 /* Optimize to a simpler form. */
5161 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5165 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5166 && (OP(nxt2) != WHILEM))
5168 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5169 /* Need to optimize away parenths. */
5170 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5171 /* Set the parenth number. */
5172 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5174 oscan->flags = (U8)ARG(nxt);
5175 if (RExC_open_parens) {
5176 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5177 RExC_close_parens[ARG(nxt1)]=nxt2+1; /*close->NOTHING*/
5179 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5180 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5183 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5184 OP(nxt + 1) = OPTIMIZED; /* was count. */
5185 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5186 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5189 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5190 regnode *nnxt = regnext(nxt1);
5192 if (reg_off_by_arg[OP(nxt1)])
5193 ARG_SET(nxt1, nxt2 - nxt1);
5194 else if (nxt2 - nxt1 < U16_MAX)
5195 NEXT_OFF(nxt1) = nxt2 - nxt1;
5197 OP(nxt) = NOTHING; /* Cannot beautify */
5202 /* Optimize again: */
5203 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5204 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
5209 else if ((OP(oscan) == CURLYX)
5210 && (flags & SCF_WHILEM_VISITED_POS)
5211 /* See the comment on a similar expression above.
5212 However, this time it's not a subexpression
5213 we care about, but the expression itself. */
5214 && (maxcount == REG_INFTY)
5215 && data && ++data->whilem_c < 16) {
5216 /* This stays as CURLYX, we can put the count/of pair. */
5217 /* Find WHILEM (as in regexec.c) */
5218 regnode *nxt = oscan + NEXT_OFF(oscan);
5220 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5222 PREVOPER(nxt)->flags = (U8)(data->whilem_c
5223 | (RExC_whilem_seen << 4)); /* On WHILEM */
5225 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5227 if (flags & SCF_DO_SUBSTR) {
5228 SV *last_str = NULL;
5229 STRLEN last_chrs = 0;
5230 int counted = mincount != 0;
5232 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5234 SSize_t b = pos_before >= data->last_start_min
5235 ? pos_before : data->last_start_min;
5237 const char * const s = SvPV_const(data->last_found, l);
5238 SSize_t old = b - data->last_start_min;
5241 old = utf8_hop((U8*)s, old) - (U8*)s;
5243 /* Get the added string: */
5244 last_str = newSVpvn_utf8(s + old, l, UTF);
5245 last_chrs = UTF ? utf8_length((U8*)(s + old),
5246 (U8*)(s + old + l)) : l;
5247 if (deltanext == 0 && pos_before == b) {
5248 /* What was added is a constant string */
5251 SvGROW(last_str, (mincount * l) + 1);
5252 repeatcpy(SvPVX(last_str) + l,
5253 SvPVX_const(last_str), l,
5255 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5256 /* Add additional parts. */
5257 SvCUR_set(data->last_found,
5258 SvCUR(data->last_found) - l);
5259 sv_catsv(data->last_found, last_str);
5261 SV * sv = data->last_found;
5263 SvUTF8(sv) && SvMAGICAL(sv) ?
5264 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5265 if (mg && mg->mg_len >= 0)
5266 mg->mg_len += last_chrs * (mincount-1);
5268 last_chrs *= mincount;
5269 data->last_end += l * (mincount - 1);
5272 /* start offset must point into the last copy */
5273 data->last_start_min += minnext * (mincount - 1);
5274 data->last_start_max =
5277 : data->last_start_max +
5278 (maxcount - 1) * (minnext + data->pos_delta);
5281 /* It is counted once already... */
5282 data->pos_min += minnext * (mincount - counted);
5284 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5285 " SSize_t_MAX=%" UVuf " minnext=%" UVuf
5286 " maxcount=%" UVuf " mincount=%" UVuf "\n",
5287 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
5289 if (deltanext != SSize_t_MAX)
5290 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5291 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5292 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
5294 if (deltanext == SSize_t_MAX
5295 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
5296 data->pos_delta = SSize_t_MAX;
5298 data->pos_delta += - counted * deltanext +
5299 (minnext + deltanext) * maxcount - minnext * mincount;
5300 if (mincount != maxcount) {
5301 /* Cannot extend fixed substrings found inside
5303 scan_commit(pRExC_state, data, minlenp, is_inf);
5304 if (mincount && last_str) {
5305 SV * const sv = data->last_found;
5306 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5307 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5311 sv_setsv(sv, last_str);
5312 data->last_end = data->pos_min;
5313 data->last_start_min = data->pos_min - last_chrs;
5314 data->last_start_max = is_inf
5316 : data->pos_min + data->pos_delta - last_chrs;
5318 data->longest = &(data->longest_float);
5320 SvREFCNT_dec(last_str);
5322 if (data && (fl & SF_HAS_EVAL))
5323 data->flags |= SF_HAS_EVAL;
5324 optimize_curly_tail:
5325 if (OP(oscan) != CURLYX) {
5326 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
5328 NEXT_OFF(oscan) += NEXT_OFF(next);
5334 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5339 if (flags & SCF_DO_SUBSTR) {
5340 /* Cannot expect anything... */
5341 scan_commit(pRExC_state, data, minlenp, is_inf);
5342 data->longest = &(data->longest_float);
5344 is_inf = is_inf_internal = 1;
5345 if (flags & SCF_DO_STCLASS_OR) {
5346 if (OP(scan) == CLUMP) {
5347 /* Actually is any start char, but very few code points
5348 * aren't start characters */
5349 ssc_match_all_cp(data->start_class);
5352 ssc_anything(data->start_class);
5355 flags &= ~SCF_DO_STCLASS;
5359 else if (OP(scan) == LNBREAK) {
5360 if (flags & SCF_DO_STCLASS) {
5361 if (flags & SCF_DO_STCLASS_AND) {
5362 ssc_intersection(data->start_class,
5363 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5364 ssc_clear_locale(data->start_class);
5365 ANYOF_FLAGS(data->start_class)
5366 &= ~SSC_MATCHES_EMPTY_STRING;
5368 else if (flags & SCF_DO_STCLASS_OR) {
5369 ssc_union(data->start_class,
5370 PL_XPosix_ptrs[_CC_VERTSPACE],
5372 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5374 /* See commit msg for
5375 * 749e076fceedeb708a624933726e7989f2302f6a */
5376 ANYOF_FLAGS(data->start_class)
5377 &= ~SSC_MATCHES_EMPTY_STRING;
5379 flags &= ~SCF_DO_STCLASS;
5382 if (delta != SSize_t_MAX)
5383 delta++; /* Because of the 2 char string cr-lf */
5384 if (flags & SCF_DO_SUBSTR) {
5385 /* Cannot expect anything... */
5386 scan_commit(pRExC_state, data, minlenp, is_inf);
5388 data->pos_delta += 1;
5389 data->longest = &(data->longest_float);
5392 else if (REGNODE_SIMPLE(OP(scan))) {
5394 if (flags & SCF_DO_SUBSTR) {
5395 scan_commit(pRExC_state, data, minlenp, is_inf);
5399 if (flags & SCF_DO_STCLASS) {
5401 SV* my_invlist = NULL;
5404 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5405 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5407 /* Some of the logic below assumes that switching
5408 locale on will only add false positives. */
5413 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5417 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5418 ssc_match_all_cp(data->start_class);
5423 SV* REG_ANY_invlist = _new_invlist(2);
5424 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5426 if (flags & SCF_DO_STCLASS_OR) {
5427 ssc_union(data->start_class,
5429 TRUE /* TRUE => invert, hence all but \n
5433 else if (flags & SCF_DO_STCLASS_AND) {
5434 ssc_intersection(data->start_class,
5436 TRUE /* TRUE => invert */
5438 ssc_clear_locale(data->start_class);
5440 SvREFCNT_dec_NN(REG_ANY_invlist);
5447 if (flags & SCF_DO_STCLASS_AND)
5448 ssc_and(pRExC_state, data->start_class,
5449 (regnode_charclass *) scan);
5451 ssc_or(pRExC_state, data->start_class,
5452 (regnode_charclass *) scan);
5460 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5461 if (flags & SCF_DO_STCLASS_AND) {
5462 bool was_there = cBOOL(
5463 ANYOF_POSIXL_TEST(data->start_class,
5465 ANYOF_POSIXL_ZERO(data->start_class);
5466 if (was_there) { /* Do an AND */
5467 ANYOF_POSIXL_SET(data->start_class, namedclass);
5469 /* No individual code points can now match */
5470 data->start_class->invlist
5471 = sv_2mortal(_new_invlist(0));
5474 int complement = namedclass + ((invert) ? -1 : 1);
5476 assert(flags & SCF_DO_STCLASS_OR);
5478 /* If the complement of this class was already there,
5479 * the result is that they match all code points,
5480 * (\d + \D == everything). Remove the classes from
5481 * future consideration. Locale is not relevant in
5483 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5484 ssc_match_all_cp(data->start_class);
5485 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5486 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5488 else { /* The usual case; just add this class to the
5490 ANYOF_POSIXL_SET(data->start_class, namedclass);
5495 case NPOSIXA: /* For these, we always know the exact set of
5500 if (FLAGS(scan) == _CC_ASCII) {
5501 my_invlist = invlist_clone(PL_XPosix_ptrs[_CC_ASCII]);
5504 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
5505 PL_XPosix_ptrs[_CC_ASCII],
5516 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5518 /* NPOSIXD matches all upper Latin1 code points unless the
5519 * target string being matched is UTF-8, which is
5520 * unknowable until match time. Since we are going to
5521 * invert, we want to get rid of all of them so that the
5522 * inversion will match all */
5523 if (OP(scan) == NPOSIXD) {
5524 _invlist_subtract(my_invlist, PL_UpperLatin1,
5530 if (flags & SCF_DO_STCLASS_AND) {
5531 ssc_intersection(data->start_class, my_invlist, invert);
5532 ssc_clear_locale(data->start_class);
5535 assert(flags & SCF_DO_STCLASS_OR);
5536 ssc_union(data->start_class, my_invlist, invert);
5538 SvREFCNT_dec(my_invlist);
5540 if (flags & SCF_DO_STCLASS_OR)
5541 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5542 flags &= ~SCF_DO_STCLASS;
5545 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5546 data->flags |= (OP(scan) == MEOL
5549 scan_commit(pRExC_state, data, minlenp, is_inf);
5552 else if ( PL_regkind[OP(scan)] == BRANCHJ
5553 /* Lookbehind, or need to calculate parens/evals/stclass: */
5554 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5555 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5557 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5558 || OP(scan) == UNLESSM )
5560 /* Negative Lookahead/lookbehind
5561 In this case we can't do fixed string optimisation.
5564 SSize_t deltanext, minnext, fake = 0;
5569 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5571 data_fake.whilem_c = data->whilem_c;
5572 data_fake.last_closep = data->last_closep;
5575 data_fake.last_closep = &fake;
5576 data_fake.pos_delta = delta;
5577 if ( flags & SCF_DO_STCLASS && !scan->flags
5578 && OP(scan) == IFMATCH ) { /* Lookahead */
5579 ssc_init(pRExC_state, &intrnl);
5580 data_fake.start_class = &intrnl;
5581 f |= SCF_DO_STCLASS_AND;
5583 if (flags & SCF_WHILEM_VISITED_POS)
5584 f |= SCF_WHILEM_VISITED_POS;
5585 next = regnext(scan);
5586 nscan = NEXTOPER(NEXTOPER(scan));
5587 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5588 last, &data_fake, stopparen,
5589 recursed_depth, NULL, f, depth+1);
5592 FAIL("Variable length lookbehind not implemented");
5594 else if (minnext > (I32)U8_MAX) {
5595 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5598 scan->flags = (U8)minnext;
5601 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5603 if (data_fake.flags & SF_HAS_EVAL)
5604 data->flags |= SF_HAS_EVAL;
5605 data->whilem_c = data_fake.whilem_c;
5607 if (f & SCF_DO_STCLASS_AND) {
5608 if (flags & SCF_DO_STCLASS_OR) {
5609 /* OR before, AND after: ideally we would recurse with
5610 * data_fake to get the AND applied by study of the
5611 * remainder of the pattern, and then derecurse;
5612 * *** HACK *** for now just treat as "no information".
5613 * See [perl #56690].
5615 ssc_init(pRExC_state, data->start_class);
5617 /* AND before and after: combine and continue. These
5618 * assertions are zero-length, so can match an EMPTY
5620 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5621 ANYOF_FLAGS(data->start_class)
5622 |= SSC_MATCHES_EMPTY_STRING;
5626 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5628 /* Positive Lookahead/lookbehind
5629 In this case we can do fixed string optimisation,
5630 but we must be careful about it. Note in the case of
5631 lookbehind the positions will be offset by the minimum
5632 length of the pattern, something we won't know about
5633 until after the recurse.
5635 SSize_t deltanext, fake = 0;
5639 /* We use SAVEFREEPV so that when the full compile
5640 is finished perl will clean up the allocated
5641 minlens when it's all done. This way we don't
5642 have to worry about freeing them when we know
5643 they wont be used, which would be a pain.
5646 Newx( minnextp, 1, SSize_t );
5647 SAVEFREEPV(minnextp);
5650 StructCopy(data, &data_fake, scan_data_t);
5651 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5654 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5655 data_fake.last_found=newSVsv(data->last_found);
5659 data_fake.last_closep = &fake;
5660 data_fake.flags = 0;
5661 data_fake.pos_delta = delta;
5663 data_fake.flags |= SF_IS_INF;
5664 if ( flags & SCF_DO_STCLASS && !scan->flags
5665 && OP(scan) == IFMATCH ) { /* Lookahead */
5666 ssc_init(pRExC_state, &intrnl);
5667 data_fake.start_class = &intrnl;
5668 f |= SCF_DO_STCLASS_AND;
5670 if (flags & SCF_WHILEM_VISITED_POS)
5671 f |= SCF_WHILEM_VISITED_POS;
5672 next = regnext(scan);
5673 nscan = NEXTOPER(NEXTOPER(scan));
5675 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5676 &deltanext, last, &data_fake,
5677 stopparen, recursed_depth, NULL,
5681 FAIL("Variable length lookbehind not implemented");
5683 else if (*minnextp > (I32)U8_MAX) {
5684 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5687 scan->flags = (U8)*minnextp;
5692 if (f & SCF_DO_STCLASS_AND) {
5693 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5694 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
5697 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5699 if (data_fake.flags & SF_HAS_EVAL)
5700 data->flags |= SF_HAS_EVAL;
5701 data->whilem_c = data_fake.whilem_c;
5702 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5703 if (RExC_rx->minlen<*minnextp)
5704 RExC_rx->minlen=*minnextp;
5705 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5706 SvREFCNT_dec_NN(data_fake.last_found);
5708 if ( data_fake.minlen_fixed != minlenp )
5710 data->offset_fixed= data_fake.offset_fixed;
5711 data->minlen_fixed= data_fake.minlen_fixed;
5712 data->lookbehind_fixed+= scan->flags;
5714 if ( data_fake.minlen_float != minlenp )
5716 data->minlen_float= data_fake.minlen_float;
5717 data->offset_float_min=data_fake.offset_float_min;
5718 data->offset_float_max=data_fake.offset_float_max;
5719 data->lookbehind_float+= scan->flags;
5726 else if (OP(scan) == OPEN) {
5727 if (stopparen != (I32)ARG(scan))
5730 else if (OP(scan) == CLOSE) {
5731 if (stopparen == (I32)ARG(scan)) {
5734 if ((I32)ARG(scan) == is_par) {
5735 next = regnext(scan);
5737 if ( next && (OP(next) != WHILEM) && next < last)
5738 is_par = 0; /* Disable optimization */
5741 *(data->last_closep) = ARG(scan);
5743 else if (OP(scan) == EVAL) {
5745 data->flags |= SF_HAS_EVAL;
5747 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5748 if (flags & SCF_DO_SUBSTR) {
5749 scan_commit(pRExC_state, data, minlenp, is_inf);
5750 flags &= ~SCF_DO_SUBSTR;
5752 if (data && OP(scan)==ACCEPT) {
5753 data->flags |= SCF_SEEN_ACCEPT;
5758 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5760 if (flags & SCF_DO_SUBSTR) {
5761 scan_commit(pRExC_state, data, minlenp, is_inf);
5762 data->longest = &(data->longest_float);
5764 is_inf = is_inf_internal = 1;
5765 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5766 ssc_anything(data->start_class);
5767 flags &= ~SCF_DO_STCLASS;
5769 else if (OP(scan) == GPOS) {
5770 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5771 !(delta || is_inf || (data && data->pos_delta)))
5773 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5774 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5775 if (RExC_rx->gofs < (STRLEN)min)
5776 RExC_rx->gofs = min;
5778 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5782 #ifdef TRIE_STUDY_OPT
5783 #ifdef FULL_TRIE_STUDY
5784 else if (PL_regkind[OP(scan)] == TRIE) {
5785 /* NOTE - There is similar code to this block above for handling
5786 BRANCH nodes on the initial study. If you change stuff here
5788 regnode *trie_node= scan;
5789 regnode *tail= regnext(scan);
5790 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5791 SSize_t max1 = 0, min1 = SSize_t_MAX;
5794 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5795 /* Cannot merge strings after this. */
5796 scan_commit(pRExC_state, data, minlenp, is_inf);
5798 if (flags & SCF_DO_STCLASS)
5799 ssc_init_zero(pRExC_state, &accum);
5805 const regnode *nextbranch= NULL;
5808 for ( word=1 ; word <= trie->wordcount ; word++)
5810 SSize_t deltanext=0, minnext=0, f = 0, fake;
5811 regnode_ssc this_class;
5813 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5815 data_fake.whilem_c = data->whilem_c;
5816 data_fake.last_closep = data->last_closep;
5819 data_fake.last_closep = &fake;
5820 data_fake.pos_delta = delta;
5821 if (flags & SCF_DO_STCLASS) {
5822 ssc_init(pRExC_state, &this_class);
5823 data_fake.start_class = &this_class;
5824 f = SCF_DO_STCLASS_AND;
5826 if (flags & SCF_WHILEM_VISITED_POS)
5827 f |= SCF_WHILEM_VISITED_POS;
5829 if (trie->jump[word]) {
5831 nextbranch = trie_node + trie->jump[0];
5832 scan= trie_node + trie->jump[word];
5833 /* We go from the jump point to the branch that follows
5834 it. Note this means we need the vestigal unused
5835 branches even though they arent otherwise used. */
5836 minnext = study_chunk(pRExC_state, &scan, minlenp,
5837 &deltanext, (regnode *)nextbranch, &data_fake,
5838 stopparen, recursed_depth, NULL, f,depth+1);
5840 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5841 nextbranch= regnext((regnode*)nextbranch);
5843 if (min1 > (SSize_t)(minnext + trie->minlen))
5844 min1 = minnext + trie->minlen;
5845 if (deltanext == SSize_t_MAX) {
5846 is_inf = is_inf_internal = 1;
5848 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5849 max1 = minnext + deltanext + trie->maxlen;
5851 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5853 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5854 if ( stopmin > min + min1)
5855 stopmin = min + min1;
5856 flags &= ~SCF_DO_SUBSTR;
5858 data->flags |= SCF_SEEN_ACCEPT;
5861 if (data_fake.flags & SF_HAS_EVAL)
5862 data->flags |= SF_HAS_EVAL;
5863 data->whilem_c = data_fake.whilem_c;
5865 if (flags & SCF_DO_STCLASS)
5866 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5869 if (flags & SCF_DO_SUBSTR) {
5870 data->pos_min += min1;
5871 data->pos_delta += max1 - min1;
5872 if (max1 != min1 || is_inf)
5873 data->longest = &(data->longest_float);
5876 if (delta != SSize_t_MAX)
5877 delta += max1 - min1;
5878 if (flags & SCF_DO_STCLASS_OR) {
5879 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5881 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5882 flags &= ~SCF_DO_STCLASS;
5885 else if (flags & SCF_DO_STCLASS_AND) {
5887 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5888 flags &= ~SCF_DO_STCLASS;
5891 /* Switch to OR mode: cache the old value of
5892 * data->start_class */
5894 StructCopy(data->start_class, and_withp, regnode_ssc);
5895 flags &= ~SCF_DO_STCLASS_AND;
5896 StructCopy(&accum, data->start_class, regnode_ssc);
5897 flags |= SCF_DO_STCLASS_OR;
5904 else if (PL_regkind[OP(scan)] == TRIE) {
5905 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5908 min += trie->minlen;
5909 delta += (trie->maxlen - trie->minlen);
5910 flags &= ~SCF_DO_STCLASS; /* xxx */
5911 if (flags & SCF_DO_SUBSTR) {
5912 /* Cannot expect anything... */
5913 scan_commit(pRExC_state, data, minlenp, is_inf);
5914 data->pos_min += trie->minlen;
5915 data->pos_delta += (trie->maxlen - trie->minlen);
5916 if (trie->maxlen != trie->minlen)
5917 data->longest = &(data->longest_float);
5919 if (trie->jump) /* no more substrings -- for now /grr*/
5920 flags &= ~SCF_DO_SUBSTR;
5922 #endif /* old or new */
5923 #endif /* TRIE_STUDY_OPT */
5925 /* Else: zero-length, ignore. */
5926 scan = regnext(scan);
5931 /* we need to unwind recursion. */
5934 DEBUG_STUDYDATA("frame-end:",data,depth);
5935 DEBUG_PEEP("fend", scan, depth);
5937 /* restore previous context */
5938 last = frame->last_regnode;
5939 scan = frame->next_regnode;
5940 stopparen = frame->stopparen;
5941 recursed_depth = frame->prev_recursed_depth;
5943 RExC_frame_last = frame->prev_frame;
5944 frame = frame->this_prev_frame;
5945 goto fake_study_recurse;
5949 DEBUG_STUDYDATA("pre-fin:",data,depth);
5952 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5954 if (flags & SCF_DO_SUBSTR && is_inf)
5955 data->pos_delta = SSize_t_MAX - data->pos_min;
5956 if (is_par > (I32)U8_MAX)
5958 if (is_par && pars==1 && data) {
5959 data->flags |= SF_IN_PAR;
5960 data->flags &= ~SF_HAS_PAR;
5962 else if (pars && data) {
5963 data->flags |= SF_HAS_PAR;
5964 data->flags &= ~SF_IN_PAR;
5966 if (flags & SCF_DO_STCLASS_OR)
5967 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5968 if (flags & SCF_TRIE_RESTUDY)
5969 data->flags |= SCF_TRIE_RESTUDY;
5971 DEBUG_STUDYDATA("post-fin:",data,depth);
5974 SSize_t final_minlen= min < stopmin ? min : stopmin;
5976 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
5977 if (final_minlen > SSize_t_MAX - delta)
5978 RExC_maxlen = SSize_t_MAX;
5979 else if (RExC_maxlen < final_minlen + delta)
5980 RExC_maxlen = final_minlen + delta;
5982 return final_minlen;
5984 NOT_REACHED; /* NOTREACHED */
5988 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5990 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5992 PERL_ARGS_ASSERT_ADD_DATA;
5994 Renewc(RExC_rxi->data,
5995 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
5996 char, struct reg_data);
5998 Renew(RExC_rxi->data->what, count + n, U8);
6000 Newx(RExC_rxi->data->what, n, U8);
6001 RExC_rxi->data->count = count + n;
6002 Copy(s, RExC_rxi->data->what + count, n, U8);
6006 /*XXX: todo make this not included in a non debugging perl, but appears to be
6007 * used anyway there, in 'use re' */
6008 #ifndef PERL_IN_XSUB_RE
6010 Perl_reginitcolors(pTHX)
6012 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6014 char *t = savepv(s);
6018 t = strchr(t, '\t');
6024 PL_colors[i] = t = (char *)"";
6029 PL_colors[i++] = (char *)"";
6036 #ifdef TRIE_STUDY_OPT
6037 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6040 (data.flags & SCF_TRIE_RESTUDY) \
6048 #define CHECK_RESTUDY_GOTO_butfirst
6052 * pregcomp - compile a regular expression into internal code
6054 * Decides which engine's compiler to call based on the hint currently in
6058 #ifndef PERL_IN_XSUB_RE
6060 /* return the currently in-scope regex engine (or the default if none) */
6062 regexp_engine const *
6063 Perl_current_re_engine(pTHX)
6065 if (IN_PERL_COMPILETIME) {
6066 HV * const table = GvHV(PL_hintgv);
6069 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6070 return &PL_core_reg_engine;
6071 ptr = hv_fetchs(table, "regcomp", FALSE);
6072 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6073 return &PL_core_reg_engine;
6074 return INT2PTR(regexp_engine*,SvIV(*ptr));
6078 if (!PL_curcop->cop_hints_hash)
6079 return &PL_core_reg_engine;
6080 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6081 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6082 return &PL_core_reg_engine;
6083 return INT2PTR(regexp_engine*,SvIV(ptr));
6089 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6091 regexp_engine const *eng = current_re_engine();
6092 GET_RE_DEBUG_FLAGS_DECL;
6094 PERL_ARGS_ASSERT_PREGCOMP;
6096 /* Dispatch a request to compile a regexp to correct regexp engine. */
6098 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6101 return CALLREGCOMP_ENG(eng, pattern, flags);
6105 /* public(ish) entry point for the perl core's own regex compiling code.
6106 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6107 * pattern rather than a list of OPs, and uses the internal engine rather
6108 * than the current one */
6111 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6113 SV *pat = pattern; /* defeat constness! */
6114 PERL_ARGS_ASSERT_RE_COMPILE;
6115 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6116 #ifdef PERL_IN_XSUB_RE
6119 &PL_core_reg_engine,
6121 NULL, NULL, rx_flags, 0);
6125 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6126 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6127 * point to the realloced string and length.
6129 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6133 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6134 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6136 U8 *const src = (U8*)*pat_p;
6141 GET_RE_DEBUG_FLAGS_DECL;
6143 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6144 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6146 Newx(dst, *plen_p * 2 + 1, U8);
6149 while (s < *plen_p) {
6150 append_utf8_from_native_byte(src[s], &d);
6151 if (n < num_code_blocks) {
6152 if (!do_end && pRExC_state->code_blocks[n].start == s) {
6153 pRExC_state->code_blocks[n].start = d - dst - 1;
6154 assert(*(d - 1) == '(');
6157 else if (do_end && pRExC_state->code_blocks[n].end == s) {
6158 pRExC_state->code_blocks[n].end = d - dst - 1;
6159 assert(*(d - 1) == ')');
6168 *pat_p = (char*) dst;
6170 RExC_orig_utf8 = RExC_utf8 = 1;
6175 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6176 * while recording any code block indices, and handling overloading,
6177 * nested qr// objects etc. If pat is null, it will allocate a new
6178 * string, or just return the first arg, if there's only one.
6180 * Returns the malloced/updated pat.
6181 * patternp and pat_count is the array of SVs to be concatted;
6182 * oplist is the optional list of ops that generated the SVs;
6183 * recompile_p is a pointer to a boolean that will be set if
6184 * the regex will need to be recompiled.
6185 * delim, if non-null is an SV that will be inserted between each element
6189 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6190 SV *pat, SV ** const patternp, int pat_count,
6191 OP *oplist, bool *recompile_p, SV *delim)
6195 bool use_delim = FALSE;
6196 bool alloced = FALSE;
6198 /* if we know we have at least two args, create an empty string,
6199 * then concatenate args to that. For no args, return an empty string */
6200 if (!pat && pat_count != 1) {
6206 for (svp = patternp; svp < patternp + pat_count; svp++) {
6209 STRLEN orig_patlen = 0;
6211 SV *msv = use_delim ? delim : *svp;
6212 if (!msv) msv = &PL_sv_undef;
6214 /* if we've got a delimiter, we go round the loop twice for each
6215 * svp slot (except the last), using the delimiter the second
6224 if (SvTYPE(msv) == SVt_PVAV) {
6225 /* we've encountered an interpolated array within
6226 * the pattern, e.g. /...@a..../. Expand the list of elements,
6227 * then recursively append elements.
6228 * The code in this block is based on S_pushav() */
6230 AV *const av = (AV*)msv;
6231 const SSize_t maxarg = AvFILL(av) + 1;
6235 assert(oplist->op_type == OP_PADAV
6236 || oplist->op_type == OP_RV2AV);
6237 oplist = OpSIBLING(oplist);
6240 if (SvRMAGICAL(av)) {
6243 Newx(array, maxarg, SV*);
6245 for (i=0; i < maxarg; i++) {
6246 SV ** const svp = av_fetch(av, i, FALSE);
6247 array[i] = svp ? *svp : &PL_sv_undef;
6251 array = AvARRAY(av);
6253 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6254 array, maxarg, NULL, recompile_p,
6256 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6262 /* we make the assumption here that each op in the list of
6263 * op_siblings maps to one SV pushed onto the stack,
6264 * except for code blocks, with have both an OP_NULL and
6266 * This allows us to match up the list of SVs against the
6267 * list of OPs to find the next code block.
6269 * Note that PUSHMARK PADSV PADSV ..
6271 * PADRANGE PADSV PADSV ..
6272 * so the alignment still works. */
6275 if (oplist->op_type == OP_NULL
6276 && (oplist->op_flags & OPf_SPECIAL))
6278 assert(n < pRExC_state->num_code_blocks);
6279 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
6280 pRExC_state->code_blocks[n].block = oplist;
6281 pRExC_state->code_blocks[n].src_regex = NULL;
6284 oplist = OpSIBLING(oplist); /* skip CONST */
6287 oplist = OpSIBLING(oplist);;
6290 /* apply magic and QR overloading to arg */
6293 if (SvROK(msv) && SvAMAGIC(msv)) {
6294 SV *sv = AMG_CALLunary(msv, regexp_amg);
6298 if (SvTYPE(sv) != SVt_REGEXP)
6299 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6304 /* try concatenation overload ... */
6305 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6306 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6309 /* overloading involved: all bets are off over literal
6310 * code. Pretend we haven't seen it */
6311 pRExC_state->num_code_blocks -= n;
6315 /* ... or failing that, try "" overload */
6316 while (SvAMAGIC(msv)
6317 && (sv = AMG_CALLunary(msv, string_amg))
6321 && SvRV(msv) == SvRV(sv))
6326 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6330 /* this is a partially unrolled
6331 * sv_catsv_nomg(pat, msv);
6332 * that allows us to adjust code block indices if
6335 char *dst = SvPV_force_nomg(pat, dlen);
6337 if (SvUTF8(msv) && !SvUTF8(pat)) {
6338 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6339 sv_setpvn(pat, dst, dlen);
6342 sv_catsv_nomg(pat, msv);
6346 /* We have only one SV to process, but we need to verify
6347 * it is properly null terminated or we will fail asserts
6348 * later. In theory we probably shouldn't get such SV's,
6349 * but if we do we should handle it gracefully. */
6350 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) ) {
6351 /* not a string, or a string with a trailing null */
6354 /* a string with no trailing null, we need to copy it
6355 * so it we have a trailing null */
6361 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
6364 /* extract any code blocks within any embedded qr//'s */
6365 if (rx && SvTYPE(rx) == SVt_REGEXP
6366 && RX_ENGINE((REGEXP*)rx)->op_comp)
6369 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6370 if (ri->num_code_blocks) {
6372 /* the presence of an embedded qr// with code means
6373 * we should always recompile: the text of the
6374 * qr// may not have changed, but it may be a
6375 * different closure than last time */
6377 Renew(pRExC_state->code_blocks,
6378 pRExC_state->num_code_blocks + ri->num_code_blocks,
6379 struct reg_code_block);
6380 pRExC_state->num_code_blocks += ri->num_code_blocks;
6382 for (i=0; i < ri->num_code_blocks; i++) {
6383 struct reg_code_block *src, *dst;
6384 STRLEN offset = orig_patlen
6385 + ReANY((REGEXP *)rx)->pre_prefix;
6386 assert(n < pRExC_state->num_code_blocks);
6387 src = &ri->code_blocks[i];
6388 dst = &pRExC_state->code_blocks[n];
6389 dst->start = src->start + offset;
6390 dst->end = src->end + offset;
6391 dst->block = src->block;
6392 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6401 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6410 /* see if there are any run-time code blocks in the pattern.
6411 * False positives are allowed */
6414 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6415 char *pat, STRLEN plen)
6420 PERL_UNUSED_CONTEXT;
6422 for (s = 0; s < plen; s++) {
6423 if (n < pRExC_state->num_code_blocks
6424 && s == pRExC_state->code_blocks[n].start)
6426 s = pRExC_state->code_blocks[n].end;
6430 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6432 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6434 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6441 /* Handle run-time code blocks. We will already have compiled any direct
6442 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6443 * copy of it, but with any literal code blocks blanked out and
6444 * appropriate chars escaped; then feed it into
6446 * eval "qr'modified_pattern'"
6450 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6454 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6456 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6457 * and merge them with any code blocks of the original regexp.
6459 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6460 * instead, just save the qr and return FALSE; this tells our caller that
6461 * the original pattern needs upgrading to utf8.
6465 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6466 char *pat, STRLEN plen)
6470 GET_RE_DEBUG_FLAGS_DECL;
6472 if (pRExC_state->runtime_code_qr) {
6473 /* this is the second time we've been called; this should
6474 * only happen if the main pattern got upgraded to utf8
6475 * during compilation; re-use the qr we compiled first time
6476 * round (which should be utf8 too)
6478 qr = pRExC_state->runtime_code_qr;
6479 pRExC_state->runtime_code_qr = NULL;
6480 assert(RExC_utf8 && SvUTF8(qr));
6486 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
6490 /* determine how many extra chars we need for ' and \ escaping */
6491 for (s = 0; s < plen; s++) {
6492 if (pat[s] == '\'' || pat[s] == '\\')
6496 Newx(newpat, newlen, char);
6498 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6500 for (s = 0; s < plen; s++) {
6501 if (n < pRExC_state->num_code_blocks
6502 && s == pRExC_state->code_blocks[n].start)
6504 /* blank out literal code block */
6505 assert(pat[s] == '(');
6506 while (s <= pRExC_state->code_blocks[n].end) {
6514 if (pat[s] == '\'' || pat[s] == '\\')
6519 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
6521 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
6527 Perl_re_printf( aTHX_
6528 "%sre-parsing pattern for runtime code:%s %s\n",
6529 PL_colors[4],PL_colors[5],newpat);
6532 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6538 PUSHSTACKi(PERLSI_REQUIRE);
6539 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6540 * parsing qr''; normally only q'' does this. It also alters
6542 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6543 SvREFCNT_dec_NN(sv);
6548 SV * const errsv = ERRSV;
6549 if (SvTRUE_NN(errsv))
6551 Safefree(pRExC_state->code_blocks);
6552 /* use croak_sv ? */
6553 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
6556 assert(SvROK(qr_ref));
6558 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6559 /* the leaving below frees the tmp qr_ref.
6560 * Give qr a life of its own */
6568 if (!RExC_utf8 && SvUTF8(qr)) {
6569 /* first time through; the pattern got upgraded; save the
6570 * qr for the next time through */
6571 assert(!pRExC_state->runtime_code_qr);
6572 pRExC_state->runtime_code_qr = qr;
6577 /* extract any code blocks within the returned qr// */
6580 /* merge the main (r1) and run-time (r2) code blocks into one */
6582 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6583 struct reg_code_block *new_block, *dst;
6584 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6587 if (!r2->num_code_blocks) /* we guessed wrong */
6589 SvREFCNT_dec_NN(qr);
6594 r1->num_code_blocks + r2->num_code_blocks,
6595 struct reg_code_block);
6598 while ( i1 < r1->num_code_blocks
6599 || i2 < r2->num_code_blocks)
6601 struct reg_code_block *src;
6604 if (i1 == r1->num_code_blocks) {
6605 src = &r2->code_blocks[i2++];
6608 else if (i2 == r2->num_code_blocks)
6609 src = &r1->code_blocks[i1++];
6610 else if ( r1->code_blocks[i1].start
6611 < r2->code_blocks[i2].start)
6613 src = &r1->code_blocks[i1++];
6614 assert(src->end < r2->code_blocks[i2].start);
6617 assert( r1->code_blocks[i1].start
6618 > r2->code_blocks[i2].start);
6619 src = &r2->code_blocks[i2++];
6621 assert(src->end < r1->code_blocks[i1].start);
6624 assert(pat[src->start] == '(');
6625 assert(pat[src->end] == ')');
6626 dst->start = src->start;
6627 dst->end = src->end;
6628 dst->block = src->block;
6629 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6633 r1->num_code_blocks += r2->num_code_blocks;
6634 Safefree(r1->code_blocks);
6635 r1->code_blocks = new_block;
6638 SvREFCNT_dec_NN(qr);
6644 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6645 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6646 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6647 STRLEN longest_length, bool eol, bool meol)
6649 /* This is the common code for setting up the floating and fixed length
6650 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6651 * as to whether succeeded or not */
6656 if (! (longest_length
6657 || (eol /* Can't have SEOL and MULTI */
6658 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6660 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6661 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6666 /* copy the information about the longest from the reg_scan_data
6667 over to the program. */
6668 if (SvUTF8(sv_longest)) {
6669 *rx_utf8 = sv_longest;
6672 *rx_substr = sv_longest;
6675 /* end_shift is how many chars that must be matched that
6676 follow this item. We calculate it ahead of time as once the
6677 lookbehind offset is added in we lose the ability to correctly
6679 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6680 *rx_end_shift = ml - offset
6682 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
6684 + (SvTAIL(sv_longest) != 0)
6688 t = (eol/* Can't have SEOL and MULTI */
6689 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6690 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6696 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6697 * regular expression into internal code.
6698 * The pattern may be passed either as:
6699 * a list of SVs (patternp plus pat_count)
6700 * a list of OPs (expr)
6701 * If both are passed, the SV list is used, but the OP list indicates
6702 * which SVs are actually pre-compiled code blocks
6704 * The SVs in the list have magic and qr overloading applied to them (and
6705 * the list may be modified in-place with replacement SVs in the latter
6708 * If the pattern hasn't changed from old_re, then old_re will be
6711 * eng is the current engine. If that engine has an op_comp method, then
6712 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6713 * do the initial concatenation of arguments and pass on to the external
6716 * If is_bare_re is not null, set it to a boolean indicating whether the
6717 * arg list reduced (after overloading) to a single bare regex which has
6718 * been returned (i.e. /$qr/).
6720 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6722 * pm_flags contains the PMf_* flags, typically based on those from the
6723 * pm_flags field of the related PMOP. Currently we're only interested in
6724 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6726 * We can't allocate space until we know how big the compiled form will be,
6727 * but we can't compile it (and thus know how big it is) until we've got a
6728 * place to put the code. So we cheat: we compile it twice, once with code
6729 * generation turned off and size counting turned on, and once "for real".
6730 * This also means that we don't allocate space until we are sure that the
6731 * thing really will compile successfully, and we never have to move the
6732 * code and thus invalidate pointers into it. (Note that it has to be in
6733 * one piece because free() must be able to free it all.) [NB: not true in perl]
6735 * Beware that the optimization-preparation code in here knows about some
6736 * of the structure of the compiled regexp. [I'll say.]
6740 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6741 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6742 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6746 regexp_internal *ri;
6754 SV *code_blocksv = NULL;
6755 SV** new_patternp = patternp;
6757 /* these are all flags - maybe they should be turned
6758 * into a single int with different bit masks */
6759 I32 sawlookahead = 0;
6764 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6766 bool runtime_code = 0;
6768 RExC_state_t RExC_state;
6769 RExC_state_t * const pRExC_state = &RExC_state;
6770 #ifdef TRIE_STUDY_OPT
6772 RExC_state_t copyRExC_state;
6774 GET_RE_DEBUG_FLAGS_DECL;
6776 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6778 DEBUG_r(if (!PL_colorset) reginitcolors());
6780 /* Initialize these here instead of as-needed, as is quick and avoids
6781 * having to test them each time otherwise */
6782 if (! PL_AboveLatin1) {
6784 char * dump_len_string;
6787 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6788 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6789 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6790 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6791 PL_HasMultiCharFold =
6792 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6794 /* This is calculated here, because the Perl program that generates the
6795 * static global ones doesn't currently have access to
6796 * NUM_ANYOF_CODE_POINTS */
6797 PL_InBitmap = _new_invlist(2);
6798 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
6799 NUM_ANYOF_CODE_POINTS - 1);
6801 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
6802 if ( ! dump_len_string
6803 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
6805 PL_dump_re_max_len = 0;
6810 pRExC_state->warn_text = NULL;
6811 pRExC_state->code_blocks = NULL;
6812 pRExC_state->num_code_blocks = 0;
6815 *is_bare_re = FALSE;
6817 if (expr && (expr->op_type == OP_LIST ||
6818 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6819 /* allocate code_blocks if needed */
6823 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
6824 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6825 ncode++; /* count of DO blocks */
6827 pRExC_state->num_code_blocks = ncode;
6828 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
6833 /* compile-time pattern with just OP_CONSTs and DO blocks */
6838 /* find how many CONSTs there are */
6841 if (expr->op_type == OP_CONST)
6844 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6845 if (o->op_type == OP_CONST)
6849 /* fake up an SV array */
6851 assert(!new_patternp);
6852 Newx(new_patternp, n, SV*);
6853 SAVEFREEPV(new_patternp);
6857 if (expr->op_type == OP_CONST)
6858 new_patternp[n] = cSVOPx_sv(expr);
6860 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6861 if (o->op_type == OP_CONST)
6862 new_patternp[n++] = cSVOPo_sv;
6867 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6868 "Assembling pattern from %d elements%s\n", pat_count,
6869 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6871 /* set expr to the first arg op */
6873 if (pRExC_state->num_code_blocks
6874 && expr->op_type != OP_CONST)
6876 expr = cLISTOPx(expr)->op_first;
6877 assert( expr->op_type == OP_PUSHMARK
6878 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6879 || expr->op_type == OP_PADRANGE);
6880 expr = OpSIBLING(expr);
6883 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6884 expr, &recompile, NULL);
6886 /* handle bare (possibly after overloading) regex: foo =~ $re */
6891 if (SvTYPE(re) == SVt_REGEXP) {
6895 Safefree(pRExC_state->code_blocks);
6896 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6897 "Precompiled pattern%s\n",
6898 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6904 exp = SvPV_nomg(pat, plen);
6906 if (!eng->op_comp) {
6907 if ((SvUTF8(pat) && IN_BYTES)
6908 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6910 /* make a temporary copy; either to convert to bytes,
6911 * or to avoid repeating get-magic / overloaded stringify */
6912 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6913 (IN_BYTES ? 0 : SvUTF8(pat)));
6915 Safefree(pRExC_state->code_blocks);
6916 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6919 /* ignore the utf8ness if the pattern is 0 length */
6920 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6922 RExC_uni_semantics = 0;
6923 RExC_seen_unfolded_sharp_s = 0;
6924 RExC_contains_locale = 0;
6925 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
6926 RExC_study_started = 0;
6927 pRExC_state->runtime_code_qr = NULL;
6928 RExC_frame_head= NULL;
6929 RExC_frame_last= NULL;
6930 RExC_frame_count= 0;
6933 RExC_mysv1= sv_newmortal();
6934 RExC_mysv2= sv_newmortal();
6937 SV *dsv= sv_newmortal();
6938 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6939 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
6940 PL_colors[4],PL_colors[5],s);
6944 /* we jump here if we have to recompile, e.g., from upgrading the pattern
6947 if ((pm_flags & PMf_USE_RE_EVAL)
6948 /* this second condition covers the non-regex literal case,
6949 * i.e. $foo =~ '(?{})'. */
6950 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6952 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6954 /* return old regex if pattern hasn't changed */
6955 /* XXX: note in the below we have to check the flags as well as the
6958 * Things get a touch tricky as we have to compare the utf8 flag
6959 * independently from the compile flags. */
6963 && !!RX_UTF8(old_re) == !!RExC_utf8
6964 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
6965 && RX_PRECOMP(old_re)
6966 && RX_PRELEN(old_re) == plen
6967 && memEQ(RX_PRECOMP(old_re), exp, plen)
6968 && !runtime_code /* with runtime code, always recompile */ )
6970 Safefree(pRExC_state->code_blocks);
6974 rx_flags = orig_rx_flags;
6976 if ( initial_charset == REGEX_DEPENDS_CHARSET
6977 && (RExC_utf8 ||RExC_uni_semantics))
6980 /* Set to use unicode semantics if the pattern is in utf8 and has the
6981 * 'depends' charset specified, as it means unicode when utf8 */
6982 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6986 RExC_precomp_adj = 0;
6987 RExC_flags = rx_flags;
6988 RExC_pm_flags = pm_flags;
6991 assert(TAINTING_get || !TAINT_get);
6993 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
6995 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
6996 /* whoops, we have a non-utf8 pattern, whilst run-time code
6997 * got compiled as utf8. Try again with a utf8 pattern */
6998 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6999 pRExC_state->num_code_blocks);
7000 goto redo_first_pass;
7003 assert(!pRExC_state->runtime_code_qr);
7009 RExC_in_lookbehind = 0;
7010 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7013 RExC_recode_x_to_native = 0;
7015 RExC_in_multi_char_class = 0;
7017 /* First pass: determine size, legality. */
7019 RExC_start = RExC_adjusted_start = exp;
7020 RExC_end = exp + plen;
7021 RExC_precomp_end = RExC_end;
7026 RExC_emit = (regnode *) &RExC_emit_dummy;
7027 RExC_whilem_seen = 0;
7028 RExC_open_parens = NULL;
7029 RExC_close_parens = NULL;
7031 RExC_paren_names = NULL;
7033 RExC_paren_name_list = NULL;
7035 RExC_recurse = NULL;
7036 RExC_study_chunk_recursed = NULL;
7037 RExC_study_chunk_recursed_bytes= 0;
7038 RExC_recurse_count = 0;
7039 pRExC_state->code_index = 0;
7041 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7042 * code makes sure the final byte is an uncounted NUL. But should this
7043 * ever not be the case, lots of things could read beyond the end of the
7044 * buffer: loops like
7045 * while(isFOO(*RExC_parse)) RExC_parse++;
7046 * strchr(RExC_parse, "foo");
7047 * etc. So it is worth noting. */
7048 assert(*RExC_end == '\0');
7051 Perl_re_printf( aTHX_ "Starting first pass (sizing)\n");
7053 RExC_lastparse=NULL;
7055 /* reg may croak on us, not giving us a chance to free
7056 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
7057 need it to survive as long as the regexp (qr/(?{})/).
7058 We must check that code_blocksv is not already set, because we may
7059 have jumped back to restart the sizing pass. */
7060 if (pRExC_state->code_blocks && !code_blocksv) {
7061 code_blocksv = newSV_type(SVt_PV);
7062 SAVEFREESV(code_blocksv);
7063 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
7064 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
7066 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7067 /* It's possible to write a regexp in ascii that represents Unicode
7068 codepoints outside of the byte range, such as via \x{100}. If we
7069 detect such a sequence we have to convert the entire pattern to utf8
7070 and then recompile, as our sizing calculation will have been based
7071 on 1 byte == 1 character, but we will need to use utf8 to encode
7072 at least some part of the pattern, and therefore must convert the whole
7075 if (flags & RESTART_PASS1) {
7076 if (flags & NEED_UTF8) {
7077 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7078 pRExC_state->num_code_blocks);
7081 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7082 "Need to redo pass 1\n"));
7085 goto redo_first_pass;
7087 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#" UVxf, (UV) flags);
7090 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
7093 Perl_re_printf( aTHX_
7094 "Required size %" IVdf " nodes\n"
7095 "Starting second pass (creation)\n",
7098 RExC_lastparse=NULL;
7101 /* The first pass could have found things that force Unicode semantics */
7102 if ((RExC_utf8 || RExC_uni_semantics)
7103 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
7105 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7108 /* Small enough for pointer-storage convention?
7109 If extralen==0, this means that we will not need long jumps. */
7110 if (RExC_size >= 0x10000L && RExC_extralen)
7111 RExC_size += RExC_extralen;
7114 if (RExC_whilem_seen > 15)
7115 RExC_whilem_seen = 15;
7117 /* Allocate space and zero-initialize. Note, the two step process
7118 of zeroing when in debug mode, thus anything assigned has to
7119 happen after that */
7120 rx = (REGEXP*) newSV_type(SVt_REGEXP);
7122 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7123 char, regexp_internal);
7124 if ( r == NULL || ri == NULL )
7125 FAIL("Regexp out of space");
7127 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
7128 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7131 /* bulk initialize base fields with 0. */
7132 Zero(ri, sizeof(regexp_internal), char);
7135 /* non-zero initialization begins here */
7138 r->extflags = rx_flags;
7139 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7141 if (pm_flags & PMf_IS_QR) {
7142 ri->code_blocks = pRExC_state->code_blocks;
7143 ri->num_code_blocks = pRExC_state->num_code_blocks;
7148 for (n = 0; n < pRExC_state->num_code_blocks; n++)
7149 if (pRExC_state->code_blocks[n].src_regex)
7150 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
7151 if(pRExC_state->code_blocks)
7152 SAVEFREEPV(pRExC_state->code_blocks); /* often null */
7156 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7157 bool has_charset = (get_regex_charset(r->extflags)
7158 != REGEX_DEPENDS_CHARSET);
7160 /* The caret is output if there are any defaults: if not all the STD
7161 * flags are set, or if no character set specifier is needed */
7163 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7165 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7166 == REG_RUN_ON_COMMENT_SEEN);
7167 U8 reganch = (U8)((r->extflags & RXf_PMf_STD_PMMOD)
7168 >> RXf_PMf_STD_PMMOD_SHIFT);
7169 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7172 /* We output all the necessary flags; we never output a minus, as all
7173 * those are defaults, so are
7174 * covered by the caret */
7175 const STRLEN wraplen = plen + has_p + has_runon
7176 + has_default /* If needs a caret */
7177 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7179 /* If needs a character set specifier */
7180 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7181 + (sizeof("(?:)") - 1);
7183 /* make sure PL_bitcount bounds not exceeded */
7184 assert(sizeof(STD_PAT_MODS) <= 8);
7186 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
7187 r->xpv_len_u.xpvlenu_pv = p;
7189 SvFLAGS(rx) |= SVf_UTF8;
7192 /* If a default, cover it using the caret */
7194 *p++= DEFAULT_PAT_MOD;
7198 const char* const name = get_regex_charset_name(r->extflags, &len);
7199 Copy(name, p, len, char);
7203 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7206 while((ch = *fptr++)) {
7214 Copy(RExC_precomp, p, plen, char);
7215 assert ((RX_WRAPPED(rx) - p) < 16);
7216 r->pre_prefix = p - RX_WRAPPED(rx);
7222 SvCUR_set(rx, p - RX_WRAPPED(rx));
7226 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
7228 /* Useful during FAIL. */
7229 #ifdef RE_TRACK_PATTERN_OFFSETS
7230 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
7231 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7232 "%s %" UVuf " bytes for offset annotations.\n",
7233 ri->u.offsets ? "Got" : "Couldn't get",
7234 (UV)((2*RExC_size+1) * sizeof(U32))));
7236 SetProgLen(ri,RExC_size);
7241 /* Second pass: emit code. */
7242 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7243 RExC_pm_flags = pm_flags;
7245 RExC_end = exp + plen;
7247 RExC_emit_start = ri->program;
7248 RExC_emit = ri->program;
7249 RExC_emit_bound = ri->program + RExC_size + 1;
7250 pRExC_state->code_index = 0;
7252 *((char*) RExC_emit++) = (char) REG_MAGIC;
7253 /* setup various meta data about recursion, this all requires
7254 * RExC_npar to be correctly set, and a bit later on we clear it */
7255 if (RExC_seen & REG_RECURSE_SEEN) {
7256 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
7257 "%*s%*s Setting up open/close parens\n",
7258 22, "| |", (int)(0 * 2 + 1), ""));
7260 /* setup RExC_open_parens, which holds the address of each
7261 * OPEN tag, and to make things simpler for the 0 index
7262 * the start of the program - this is used later for offsets */
7263 Newxz(RExC_open_parens, RExC_npar,regnode *);
7264 SAVEFREEPV(RExC_open_parens);
7265 RExC_open_parens[0] = RExC_emit;
7267 /* setup RExC_close_parens, which holds the address of each
7268 * CLOSE tag, and to make things simpler for the 0 index
7269 * the end of the program - this is used later for offsets */
7270 Newxz(RExC_close_parens, RExC_npar,regnode *);
7271 SAVEFREEPV(RExC_close_parens);
7272 /* we dont know where end op starts yet, so we dont
7273 * need to set RExC_close_parens[0] like we do RExC_open_parens[0] above */
7275 /* Note, RExC_npar is 1 + the number of parens in a pattern.
7276 * So its 1 if there are no parens. */
7277 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
7278 ((RExC_npar & 0x07) != 0);
7279 Newx(RExC_study_chunk_recursed,
7280 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7281 SAVEFREEPV(RExC_study_chunk_recursed);
7284 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7286 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#" UVxf, (UV) flags);
7289 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7292 /* XXXX To minimize changes to RE engine we always allocate
7293 3-units-long substrs field. */
7294 Newx(r->substrs, 1, struct reg_substr_data);
7295 if (RExC_recurse_count) {
7296 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
7297 SAVEFREEPV(RExC_recurse);
7301 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7303 RExC_study_chunk_recursed_count= 0;
7305 Zero(r->substrs, 1, struct reg_substr_data);
7306 if (RExC_study_chunk_recursed) {
7307 Zero(RExC_study_chunk_recursed,
7308 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7312 #ifdef TRIE_STUDY_OPT
7314 StructCopy(&zero_scan_data, &data, scan_data_t);
7315 copyRExC_state = RExC_state;
7318 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
7320 RExC_state = copyRExC_state;
7321 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
7322 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
7324 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
7325 StructCopy(&zero_scan_data, &data, scan_data_t);
7328 StructCopy(&zero_scan_data, &data, scan_data_t);
7331 /* Dig out information for optimizations. */
7332 r->extflags = RExC_flags; /* was pm_op */
7333 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
7336 SvUTF8_on(rx); /* Unicode in it? */
7337 ri->regstclass = NULL;
7338 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
7339 r->intflags |= PREGf_NAUGHTY;
7340 scan = ri->program + 1; /* First BRANCH. */
7342 /* testing for BRANCH here tells us whether there is "must appear"
7343 data in the pattern. If there is then we can use it for optimisations */
7344 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7347 STRLEN longest_float_length, longest_fixed_length;
7348 regnode_ssc ch_class; /* pointed to by data */
7350 SSize_t last_close = 0; /* pointed to by data */
7351 regnode *first= scan;
7352 regnode *first_next= regnext(first);
7354 * Skip introductions and multiplicators >= 1
7355 * so that we can extract the 'meat' of the pattern that must
7356 * match in the large if() sequence following.
7357 * NOTE that EXACT is NOT covered here, as it is normally
7358 * picked up by the optimiser separately.
7360 * This is unfortunate as the optimiser isnt handling lookahead
7361 * properly currently.
7364 while ((OP(first) == OPEN && (sawopen = 1)) ||
7365 /* An OR of *one* alternative - should not happen now. */
7366 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7367 /* for now we can't handle lookbehind IFMATCH*/
7368 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7369 (OP(first) == PLUS) ||
7370 (OP(first) == MINMOD) ||
7371 /* An {n,m} with n>0 */
7372 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7373 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7376 * the only op that could be a regnode is PLUS, all the rest
7377 * will be regnode_1 or regnode_2.
7379 * (yves doesn't think this is true)
7381 if (OP(first) == PLUS)
7384 if (OP(first) == MINMOD)
7386 first += regarglen[OP(first)];
7388 first = NEXTOPER(first);
7389 first_next= regnext(first);
7392 /* Starting-point info. */
7394 DEBUG_PEEP("first:",first,0);
7395 /* Ignore EXACT as we deal with it later. */
7396 if (PL_regkind[OP(first)] == EXACT) {
7397 if (OP(first) == EXACT || OP(first) == EXACTL)
7398 NOOP; /* Empty, get anchored substr later. */
7400 ri->regstclass = first;
7403 else if (PL_regkind[OP(first)] == TRIE &&
7404 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
7406 /* this can happen only on restudy */
7407 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
7410 else if (REGNODE_SIMPLE(OP(first)))
7411 ri->regstclass = first;
7412 else if (PL_regkind[OP(first)] == BOUND ||
7413 PL_regkind[OP(first)] == NBOUND)
7414 ri->regstclass = first;
7415 else if (PL_regkind[OP(first)] == BOL) {
7416 r->intflags |= (OP(first) == MBOL
7419 first = NEXTOPER(first);
7422 else if (OP(first) == GPOS) {
7423 r->intflags |= PREGf_ANCH_GPOS;
7424 first = NEXTOPER(first);
7427 else if ((!sawopen || !RExC_sawback) &&
7429 (OP(first) == STAR &&
7430 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7431 !(r->intflags & PREGf_ANCH) && !pRExC_state->num_code_blocks)
7433 /* turn .* into ^.* with an implied $*=1 */
7435 (OP(NEXTOPER(first)) == REG_ANY)
7438 r->intflags |= (type | PREGf_IMPLICIT);
7439 first = NEXTOPER(first);
7442 if (sawplus && !sawminmod && !sawlookahead
7443 && (!sawopen || !RExC_sawback)
7444 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
7445 /* x+ must match at the 1st pos of run of x's */
7446 r->intflags |= PREGf_SKIP;
7448 /* Scan is after the zeroth branch, first is atomic matcher. */
7449 #ifdef TRIE_STUDY_OPT
7452 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7453 (IV)(first - scan + 1))
7457 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7458 (IV)(first - scan + 1))
7464 * If there's something expensive in the r.e., find the
7465 * longest literal string that must appear and make it the
7466 * regmust. Resolve ties in favor of later strings, since
7467 * the regstart check works with the beginning of the r.e.
7468 * and avoiding duplication strengthens checking. Not a
7469 * strong reason, but sufficient in the absence of others.
7470 * [Now we resolve ties in favor of the earlier string if
7471 * it happens that c_offset_min has been invalidated, since the
7472 * earlier string may buy us something the later one won't.]
7475 data.longest_fixed = newSVpvs("");
7476 data.longest_float = newSVpvs("");
7477 data.last_found = newSVpvs("");
7478 data.longest = &(data.longest_fixed);
7479 ENTER_with_name("study_chunk");
7480 SAVEFREESV(data.longest_fixed);
7481 SAVEFREESV(data.longest_float);
7482 SAVEFREESV(data.last_found);
7484 if (!ri->regstclass) {
7485 ssc_init(pRExC_state, &ch_class);
7486 data.start_class = &ch_class;
7487 stclass_flag = SCF_DO_STCLASS_AND;
7488 } else /* XXXX Check for BOUND? */
7490 data.last_closep = &last_close;
7493 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
7494 scan + RExC_size, /* Up to end */
7496 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
7497 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
7501 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
7504 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
7505 && data.last_start_min == 0 && data.last_end > 0
7506 && !RExC_seen_zerolen
7507 && !(RExC_seen & REG_VERBARG_SEEN)
7508 && !(RExC_seen & REG_GPOS_SEEN)
7510 r->extflags |= RXf_CHECK_ALL;
7512 scan_commit(pRExC_state, &data,&minlen,0);
7514 longest_float_length = CHR_SVLEN(data.longest_float);
7516 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
7517 && data.offset_fixed == data.offset_float_min
7518 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
7519 && S_setup_longest (aTHX_ pRExC_state,
7523 &(r->float_end_shift),
7524 data.lookbehind_float,
7525 data.offset_float_min,
7527 longest_float_length,
7528 cBOOL(data.flags & SF_FL_BEFORE_EOL),
7529 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
7531 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
7532 r->float_max_offset = data.offset_float_max;
7533 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
7534 r->float_max_offset -= data.lookbehind_float;
7535 SvREFCNT_inc_simple_void_NN(data.longest_float);
7538 r->float_substr = r->float_utf8 = NULL;
7539 longest_float_length = 0;
7542 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
7544 if (S_setup_longest (aTHX_ pRExC_state,
7546 &(r->anchored_utf8),
7547 &(r->anchored_substr),
7548 &(r->anchored_end_shift),
7549 data.lookbehind_fixed,
7552 longest_fixed_length,
7553 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
7554 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
7556 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
7557 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
7560 r->anchored_substr = r->anchored_utf8 = NULL;
7561 longest_fixed_length = 0;
7563 LEAVE_with_name("study_chunk");
7566 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
7567 ri->regstclass = NULL;
7569 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
7571 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7572 && is_ssc_worth_it(pRExC_state, data.start_class))
7574 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7576 ssc_finalize(pRExC_state, data.start_class);
7578 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7579 StructCopy(data.start_class,
7580 (regnode_ssc*)RExC_rxi->data->data[n],
7582 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7583 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7584 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7585 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7586 Perl_re_printf( aTHX_
7587 "synthetic stclass \"%s\".\n",
7588 SvPVX_const(sv));});
7589 data.start_class = NULL;
7592 /* A temporary algorithm prefers floated substr to fixed one to dig
7594 if (longest_fixed_length > longest_float_length) {
7595 r->substrs->check_ix = 0;
7596 r->check_end_shift = r->anchored_end_shift;
7597 r->check_substr = r->anchored_substr;
7598 r->check_utf8 = r->anchored_utf8;
7599 r->check_offset_min = r->check_offset_max = r->anchored_offset;
7600 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
7601 r->intflags |= PREGf_NOSCAN;
7604 r->substrs->check_ix = 1;
7605 r->check_end_shift = r->float_end_shift;
7606 r->check_substr = r->float_substr;
7607 r->check_utf8 = r->float_utf8;
7608 r->check_offset_min = r->float_min_offset;
7609 r->check_offset_max = r->float_max_offset;
7611 if ((r->check_substr || r->check_utf8) ) {
7612 r->extflags |= RXf_USE_INTUIT;
7613 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7614 r->extflags |= RXf_INTUIT_TAIL;
7616 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
7618 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7619 if ( (STRLEN)minlen < longest_float_length )
7620 minlen= longest_float_length;
7621 if ( (STRLEN)minlen < longest_fixed_length )
7622 minlen= longest_fixed_length;
7626 /* Several toplevels. Best we can is to set minlen. */
7628 regnode_ssc ch_class;
7629 SSize_t last_close = 0;
7631 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
7633 scan = ri->program + 1;
7634 ssc_init(pRExC_state, &ch_class);
7635 data.start_class = &ch_class;
7636 data.last_closep = &last_close;
7639 minlen = study_chunk(pRExC_state,
7640 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7641 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7642 ? SCF_TRIE_DOING_RESTUDY
7646 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7648 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7649 = r->float_substr = r->float_utf8 = NULL;
7651 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7652 && is_ssc_worth_it(pRExC_state, data.start_class))
7654 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7656 ssc_finalize(pRExC_state, data.start_class);
7658 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7659 StructCopy(data.start_class,
7660 (regnode_ssc*)RExC_rxi->data->data[n],
7662 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7663 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7664 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7665 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7666 Perl_re_printf( aTHX_
7667 "synthetic stclass \"%s\".\n",
7668 SvPVX_const(sv));});
7669 data.start_class = NULL;
7673 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7674 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7675 r->maxlen = REG_INFTY;
7678 r->maxlen = RExC_maxlen;
7681 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7682 the "real" pattern. */
7684 Perl_re_printf( aTHX_ "minlen: %" IVdf " r->minlen:%" IVdf " maxlen:%" IVdf "\n",
7685 (IV)minlen, (IV)r->minlen, (IV)RExC_maxlen);
7687 r->minlenret = minlen;
7688 if (r->minlen < minlen)
7691 if (RExC_seen & REG_RECURSE_SEEN ) {
7692 r->intflags |= PREGf_RECURSE_SEEN;
7693 Newxz(r->recurse_locinput, r->nparens + 1, char *);
7695 if (RExC_seen & REG_GPOS_SEEN)
7696 r->intflags |= PREGf_GPOS_SEEN;
7697 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7698 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7700 if (pRExC_state->num_code_blocks)
7701 r->extflags |= RXf_EVAL_SEEN;
7702 if (RExC_seen & REG_VERBARG_SEEN)
7704 r->intflags |= PREGf_VERBARG_SEEN;
7705 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7707 if (RExC_seen & REG_CUTGROUP_SEEN)
7708 r->intflags |= PREGf_CUTGROUP_SEEN;
7709 if (pm_flags & PMf_USE_RE_EVAL)
7710 r->intflags |= PREGf_USE_RE_EVAL;
7711 if (RExC_paren_names)
7712 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7714 RXp_PAREN_NAMES(r) = NULL;
7716 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7717 * so it can be used in pp.c */
7718 if (r->intflags & PREGf_ANCH)
7719 r->extflags |= RXf_IS_ANCHORED;
7723 /* this is used to identify "special" patterns that might result
7724 * in Perl NOT calling the regex engine and instead doing the match "itself",
7725 * particularly special cases in split//. By having the regex compiler
7726 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7727 * we avoid weird issues with equivalent patterns resulting in different behavior,
7728 * AND we allow non Perl engines to get the same optimizations by the setting the
7729 * flags appropriately - Yves */
7730 regnode *first = ri->program + 1;
7732 regnode *next = regnext(first);
7735 if (PL_regkind[fop] == NOTHING && nop == END)
7736 r->extflags |= RXf_NULL;
7737 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
7738 /* when fop is SBOL first->flags will be true only when it was
7739 * produced by parsing /\A/, and not when parsing /^/. This is
7740 * very important for the split code as there we want to
7741 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
7742 * See rt #122761 for more details. -- Yves */
7743 r->extflags |= RXf_START_ONLY;
7744 else if (fop == PLUS
7745 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7747 r->extflags |= RXf_WHITE;
7748 else if ( r->extflags & RXf_SPLIT
7749 && (fop == EXACT || fop == EXACTL)
7750 && STR_LEN(first) == 1
7751 && *(STRING(first)) == ' '
7753 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7757 if (RExC_contains_locale) {
7758 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7762 if (RExC_paren_names) {
7763 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7764 ri->data->data[ri->name_list_idx]
7765 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7768 ri->name_list_idx = 0;
7770 while ( RExC_recurse_count > 0 ) {
7771 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
7772 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - scan );
7775 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7776 /* assume we don't need to swap parens around before we match */
7778 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
7779 (unsigned long)RExC_study_chunk_recursed_count);
7783 Perl_re_printf( aTHX_ "Final program:\n");
7786 #ifdef RE_TRACK_PATTERN_OFFSETS
7787 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7788 const STRLEN len = ri->u.offsets[0];
7790 GET_RE_DEBUG_FLAGS_DECL;
7791 Perl_re_printf( aTHX_
7792 "Offsets: [%" UVuf "]\n\t", (UV)ri->u.offsets[0]);
7793 for (i = 1; i <= len; i++) {
7794 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7795 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7796 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7798 Perl_re_printf( aTHX_ "\n");
7803 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7804 * by setting the regexp SV to readonly-only instead. If the
7805 * pattern's been recompiled, the USEDness should remain. */
7806 if (old_re && SvREADONLY(old_re))
7814 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7817 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7819 PERL_UNUSED_ARG(value);
7821 if (flags & RXapif_FETCH) {
7822 return reg_named_buff_fetch(rx, key, flags);
7823 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7824 Perl_croak_no_modify();
7826 } else if (flags & RXapif_EXISTS) {
7827 return reg_named_buff_exists(rx, key, flags)
7830 } else if (flags & RXapif_REGNAMES) {
7831 return reg_named_buff_all(rx, flags);
7832 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7833 return reg_named_buff_scalar(rx, flags);
7835 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7841 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7844 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7845 PERL_UNUSED_ARG(lastkey);
7847 if (flags & RXapif_FIRSTKEY)
7848 return reg_named_buff_firstkey(rx, flags);
7849 else if (flags & RXapif_NEXTKEY)
7850 return reg_named_buff_nextkey(rx, flags);
7852 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7859 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7862 AV *retarray = NULL;
7864 struct regexp *const rx = ReANY(r);
7866 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7868 if (flags & RXapif_ALL)
7871 if (rx && RXp_PAREN_NAMES(rx)) {
7872 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7875 SV* sv_dat=HeVAL(he_str);
7876 I32 *nums=(I32*)SvPVX(sv_dat);
7877 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7878 if ((I32)(rx->nparens) >= nums[i]
7879 && rx->offs[nums[i]].start != -1
7880 && rx->offs[nums[i]].end != -1)
7883 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7888 ret = newSVsv(&PL_sv_undef);
7891 av_push(retarray, ret);
7894 return newRV_noinc(MUTABLE_SV(retarray));
7901 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7904 struct regexp *const rx = ReANY(r);
7906 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7908 if (rx && RXp_PAREN_NAMES(rx)) {
7909 if (flags & RXapif_ALL) {
7910 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7912 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7914 SvREFCNT_dec_NN(sv);
7926 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7928 struct regexp *const rx = ReANY(r);
7930 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7932 if ( rx && RXp_PAREN_NAMES(rx) ) {
7933 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7935 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7942 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7944 struct regexp *const rx = ReANY(r);
7945 GET_RE_DEBUG_FLAGS_DECL;
7947 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7949 if (rx && RXp_PAREN_NAMES(rx)) {
7950 HV *hv = RXp_PAREN_NAMES(rx);
7952 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7955 SV* sv_dat = HeVAL(temphe);
7956 I32 *nums = (I32*)SvPVX(sv_dat);
7957 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7958 if ((I32)(rx->lastparen) >= nums[i] &&
7959 rx->offs[nums[i]].start != -1 &&
7960 rx->offs[nums[i]].end != -1)
7966 if (parno || flags & RXapif_ALL) {
7967 return newSVhek(HeKEY_hek(temphe));
7975 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
7980 struct regexp *const rx = ReANY(r);
7982 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
7984 if (rx && RXp_PAREN_NAMES(rx)) {
7985 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
7986 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
7987 } else if (flags & RXapif_ONE) {
7988 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
7989 av = MUTABLE_AV(SvRV(ret));
7990 length = av_tindex(av);
7991 SvREFCNT_dec_NN(ret);
7992 return newSViv(length + 1);
7994 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
7999 return &PL_sv_undef;
8003 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8005 struct regexp *const rx = ReANY(r);
8008 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8010 if (rx && RXp_PAREN_NAMES(rx)) {
8011 HV *hv= RXp_PAREN_NAMES(rx);
8013 (void)hv_iterinit(hv);
8014 while ( (temphe = hv_iternext_flags(hv,0)) ) {
8017 SV* sv_dat = HeVAL(temphe);
8018 I32 *nums = (I32*)SvPVX(sv_dat);
8019 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8020 if ((I32)(rx->lastparen) >= nums[i] &&
8021 rx->offs[nums[i]].start != -1 &&
8022 rx->offs[nums[i]].end != -1)
8028 if (parno || flags & RXapif_ALL) {
8029 av_push(av, newSVhek(HeKEY_hek(temphe)));
8034 return newRV_noinc(MUTABLE_SV(av));
8038 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8041 struct regexp *const rx = ReANY(r);
8047 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8049 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8050 || n == RX_BUFF_IDX_CARET_FULLMATCH
8051 || n == RX_BUFF_IDX_CARET_POSTMATCH
8054 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8056 /* on something like
8059 * the KEEPCOPY is set on the PMOP rather than the regex */
8060 if (PL_curpm && r == PM_GETRE(PL_curpm))
8061 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8070 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8071 /* no need to distinguish between them any more */
8072 n = RX_BUFF_IDX_FULLMATCH;
8074 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8075 && rx->offs[0].start != -1)
8077 /* $`, ${^PREMATCH} */
8078 i = rx->offs[0].start;
8082 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8083 && rx->offs[0].end != -1)
8085 /* $', ${^POSTMATCH} */
8086 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8087 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8090 if ( 0 <= n && n <= (I32)rx->nparens &&
8091 (s1 = rx->offs[n].start) != -1 &&
8092 (t1 = rx->offs[n].end) != -1)
8094 /* $&, ${^MATCH}, $1 ... */
8096 s = rx->subbeg + s1 - rx->suboffset;
8101 assert(s >= rx->subbeg);
8102 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8104 #ifdef NO_TAINT_SUPPORT
8105 sv_setpvn(sv, s, i);
8107 const int oldtainted = TAINT_get;
8109 sv_setpvn(sv, s, i);
8110 TAINT_set(oldtainted);
8112 if (RXp_MATCH_UTF8(rx))
8117 if (RXp_MATCH_TAINTED(rx)) {
8118 if (SvTYPE(sv) >= SVt_PVMG) {
8119 MAGIC* const mg = SvMAGIC(sv);
8122 SvMAGIC_set(sv, mg->mg_moremagic);
8124 if ((mgt = SvMAGIC(sv))) {
8125 mg->mg_moremagic = mgt;
8126 SvMAGIC_set(sv, mg);
8143 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8144 SV const * const value)
8146 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8148 PERL_UNUSED_ARG(rx);
8149 PERL_UNUSED_ARG(paren);
8150 PERL_UNUSED_ARG(value);
8153 Perl_croak_no_modify();
8157 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8160 struct regexp *const rx = ReANY(r);
8164 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8166 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8167 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8168 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8171 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8173 /* on something like
8176 * the KEEPCOPY is set on the PMOP rather than the regex */
8177 if (PL_curpm && r == PM_GETRE(PL_curpm))
8178 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8184 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8186 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8187 case RX_BUFF_IDX_PREMATCH: /* $` */
8188 if (rx->offs[0].start != -1) {
8189 i = rx->offs[0].start;
8198 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8199 case RX_BUFF_IDX_POSTMATCH: /* $' */
8200 if (rx->offs[0].end != -1) {
8201 i = rx->sublen - rx->offs[0].end;
8203 s1 = rx->offs[0].end;
8210 default: /* $& / ${^MATCH}, $1, $2, ... */
8211 if (paren <= (I32)rx->nparens &&
8212 (s1 = rx->offs[paren].start) != -1 &&
8213 (t1 = rx->offs[paren].end) != -1)
8219 if (ckWARN(WARN_UNINITIALIZED))
8220 report_uninit((const SV *)sv);
8225 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8226 const char * const s = rx->subbeg - rx->suboffset + s1;
8231 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8238 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8240 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8241 PERL_UNUSED_ARG(rx);
8245 return newSVpvs("Regexp");
8248 /* Scans the name of a named buffer from the pattern.
8249 * If flags is REG_RSN_RETURN_NULL returns null.
8250 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8251 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8252 * to the parsed name as looked up in the RExC_paren_names hash.
8253 * If there is an error throws a vFAIL().. type exception.
8256 #define REG_RSN_RETURN_NULL 0
8257 #define REG_RSN_RETURN_NAME 1
8258 #define REG_RSN_RETURN_DATA 2
8261 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8263 char *name_start = RExC_parse;
8265 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8267 assert (RExC_parse <= RExC_end);
8268 if (RExC_parse == RExC_end) NOOP;
8269 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8270 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8271 * using do...while */
8274 RExC_parse += UTF8SKIP(RExC_parse);
8275 } while ( RExC_parse < RExC_end
8276 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8280 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8282 RExC_parse++; /* so the <- from the vFAIL is after the offending
8284 vFAIL("Group name must start with a non-digit word character");
8288 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8289 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8290 if ( flags == REG_RSN_RETURN_NAME)
8292 else if (flags==REG_RSN_RETURN_DATA) {
8295 if ( ! sv_name ) /* should not happen*/
8296 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8297 if (RExC_paren_names)
8298 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8300 sv_dat = HeVAL(he_str);
8302 vFAIL("Reference to nonexistent named group");
8306 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8307 (unsigned long) flags);
8309 NOT_REACHED; /* NOTREACHED */
8314 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8316 if (RExC_lastparse!=RExC_parse) { \
8317 Perl_re_printf( aTHX_ "%s", \
8318 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8319 RExC_end - RExC_parse, 16, \
8321 PERL_PV_ESCAPE_UNI_DETECT | \
8322 PERL_PV_PRETTY_ELLIPSES | \
8323 PERL_PV_PRETTY_LTGT | \
8324 PERL_PV_ESCAPE_RE | \
8325 PERL_PV_PRETTY_EXACTSIZE \
8329 Perl_re_printf( aTHX_ "%16s",""); \
8332 num = RExC_size + 1; \
8334 num=REG_NODE_NUM(RExC_emit); \
8335 if (RExC_lastnum!=num) \
8336 Perl_re_printf( aTHX_ "|%4d",num); \
8338 Perl_re_printf( aTHX_ "|%4s",""); \
8339 Perl_re_printf( aTHX_ "|%*s%-4s", \
8340 (int)((depth*2)), "", \
8344 RExC_lastparse=RExC_parse; \
8349 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8350 DEBUG_PARSE_MSG((funcname)); \
8351 Perl_re_printf( aTHX_ "%4s","\n"); \
8353 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8354 DEBUG_PARSE_MSG((funcname)); \
8355 Perl_re_printf( aTHX_ fmt "\n",args); \
8358 /* This section of code defines the inversion list object and its methods. The
8359 * interfaces are highly subject to change, so as much as possible is static to
8360 * this file. An inversion list is here implemented as a malloc'd C UV array
8361 * as an SVt_INVLIST scalar.
8363 * An inversion list for Unicode is an array of code points, sorted by ordinal
8364 * number. Each element gives the code point that begins a range that extends
8365 * up-to but not including the code point given by the next element. The final
8366 * element gives the first code point of a range that extends to the platform's
8367 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8368 * ...) give ranges whose code points are all in the inversion list. We say
8369 * that those ranges are in the set. The odd-numbered elements give ranges
8370 * whose code points are not in the inversion list, and hence not in the set.
8371 * Thus, element [0] is the first code point in the list. Element [1]
8372 * is the first code point beyond that not in the list; and element [2] is the
8373 * first code point beyond that that is in the list. In other words, the first
8374 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8375 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8376 * all code points in that range are not in the inversion list. The third
8377 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8378 * list, and so forth. Thus every element whose index is divisible by two
8379 * gives the beginning of a range that is in the list, and every element whose
8380 * index is not divisible by two gives the beginning of a range not in the
8381 * list. If the final element's index is divisible by two, the inversion list
8382 * extends to the platform's infinity; otherwise the highest code point in the
8383 * inversion list is the contents of that element minus 1.
8385 * A range that contains just a single code point N will look like
8387 * invlist[i+1] == N+1
8389 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8390 * impossible to represent, so element [i+1] is omitted. The single element
8392 * invlist[0] == UV_MAX
8393 * contains just UV_MAX, but is interpreted as matching to infinity.
8395 * Taking the complement (inverting) an inversion list is quite simple, if the
8396 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8397 * This implementation reserves an element at the beginning of each inversion
8398 * list to always contain 0; there is an additional flag in the header which
8399 * indicates if the list begins at the 0, or is offset to begin at the next
8400 * element. This means that the inversion list can be inverted without any
8401 * copying; just flip the flag.
8403 * More about inversion lists can be found in "Unicode Demystified"
8404 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8406 * The inversion list data structure is currently implemented as an SV pointing
8407 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8408 * array of UV whose memory management is automatically handled by the existing
8409 * facilities for SV's.
8411 * Some of the methods should always be private to the implementation, and some
8412 * should eventually be made public */
8414 /* The header definitions are in F<invlist_inline.h> */
8416 #ifndef PERL_IN_XSUB_RE
8418 PERL_STATIC_INLINE UV*
8419 S__invlist_array_init(SV* const invlist, const bool will_have_0)
8421 /* Returns a pointer to the first element in the inversion list's array.
8422 * This is called upon initialization of an inversion list. Where the
8423 * array begins depends on whether the list has the code point U+0000 in it
8424 * or not. The other parameter tells it whether the code that follows this
8425 * call is about to put a 0 in the inversion list or not. The first
8426 * element is either the element reserved for 0, if TRUE, or the element
8427 * after it, if FALSE */
8429 bool* offset = get_invlist_offset_addr(invlist);
8430 UV* zero_addr = (UV *) SvPVX(invlist);
8432 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8435 assert(! _invlist_len(invlist));
8439 /* 1^1 = 0; 1^0 = 1 */
8440 *offset = 1 ^ will_have_0;
8441 return zero_addr + *offset;
8446 PERL_STATIC_INLINE void
8447 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8449 /* Sets the current number of elements stored in the inversion list.
8450 * Updates SvCUR correspondingly */
8451 PERL_UNUSED_CONTEXT;
8452 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8454 assert(SvTYPE(invlist) == SVt_INVLIST);
8459 : TO_INTERNAL_SIZE(len + offset));
8460 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8463 #ifndef PERL_IN_XSUB_RE
8466 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
8468 /* Replaces the inversion list in 'dest' with the one from 'src'. It
8469 * steals the list from 'src', so 'src' is made to have a NULL list. This
8470 * is similar to what SvSetMagicSV() would do, if it were implemented on
8471 * inversion lists, though this routine avoids a copy */
8473 const UV src_len = _invlist_len(src);
8474 const bool src_offset = *get_invlist_offset_addr(src);
8475 const STRLEN src_byte_len = SvLEN(src);
8476 char * array = SvPVX(src);
8478 const int oldtainted = TAINT_get;
8480 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
8482 assert(SvTYPE(src) == SVt_INVLIST);
8483 assert(SvTYPE(dest) == SVt_INVLIST);
8484 assert(! invlist_is_iterating(src));
8485 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
8487 /* Make sure it ends in the right place with a NUL, as our inversion list
8488 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
8490 array[src_byte_len - 1] = '\0';
8492 TAINT_NOT; /* Otherwise it breaks */
8493 sv_usepvn_flags(dest,
8497 /* This flag is documented to cause a copy to be avoided */
8498 SV_HAS_TRAILING_NUL);
8499 TAINT_set(oldtainted);
8504 /* Finish up copying over the other fields in an inversion list */
8505 *get_invlist_offset_addr(dest) = src_offset;
8506 invlist_set_len(dest, src_len, src_offset);
8507 *get_invlist_previous_index_addr(dest) = 0;
8508 invlist_iterfinish(dest);
8511 PERL_STATIC_INLINE IV*
8512 S_get_invlist_previous_index_addr(SV* invlist)
8514 /* Return the address of the IV that is reserved to hold the cached index
8516 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8518 assert(SvTYPE(invlist) == SVt_INVLIST);
8520 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8523 PERL_STATIC_INLINE IV
8524 S_invlist_previous_index(SV* const invlist)
8526 /* Returns cached index of previous search */
8528 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8530 return *get_invlist_previous_index_addr(invlist);
8533 PERL_STATIC_INLINE void
8534 S_invlist_set_previous_index(SV* const invlist, const IV index)
8536 /* Caches <index> for later retrieval */
8538 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8540 assert(index == 0 || index < (int) _invlist_len(invlist));
8542 *get_invlist_previous_index_addr(invlist) = index;
8545 PERL_STATIC_INLINE void
8546 S_invlist_trim(SV* invlist)
8548 /* Free the not currently-being-used space in an inversion list */
8550 /* But don't free up the space needed for the 0 UV that is always at the
8551 * beginning of the list, nor the trailing NUL */
8552 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
8554 PERL_ARGS_ASSERT_INVLIST_TRIM;
8556 assert(SvTYPE(invlist) == SVt_INVLIST);
8558 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
8561 PERL_STATIC_INLINE void
8562 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
8564 PERL_ARGS_ASSERT_INVLIST_CLEAR;
8566 assert(SvTYPE(invlist) == SVt_INVLIST);
8568 invlist_set_len(invlist, 0, 0);
8569 invlist_trim(invlist);
8572 #endif /* ifndef PERL_IN_XSUB_RE */
8574 PERL_STATIC_INLINE bool
8575 S_invlist_is_iterating(SV* const invlist)
8577 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8579 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8582 #ifndef PERL_IN_XSUB_RE
8584 PERL_STATIC_INLINE UV
8585 S_invlist_max(SV* const invlist)
8587 /* Returns the maximum number of elements storable in the inversion list's
8588 * array, without having to realloc() */
8590 PERL_ARGS_ASSERT_INVLIST_MAX;
8592 assert(SvTYPE(invlist) == SVt_INVLIST);
8594 /* Assumes worst case, in which the 0 element is not counted in the
8595 * inversion list, so subtracts 1 for that */
8596 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8597 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8598 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8601 Perl__new_invlist(pTHX_ IV initial_size)
8604 /* Return a pointer to a newly constructed inversion list, with enough
8605 * space to store 'initial_size' elements. If that number is negative, a
8606 * system default is used instead */
8610 if (initial_size < 0) {
8614 /* Allocate the initial space */
8615 new_list = newSV_type(SVt_INVLIST);
8617 /* First 1 is in case the zero element isn't in the list; second 1 is for
8619 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8620 invlist_set_len(new_list, 0, 0);
8622 /* Force iterinit() to be used to get iteration to work */
8623 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8625 *get_invlist_previous_index_addr(new_list) = 0;
8631 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8633 /* Return a pointer to a newly constructed inversion list, initialized to
8634 * point to <list>, which has to be in the exact correct inversion list
8635 * form, including internal fields. Thus this is a dangerous routine that
8636 * should not be used in the wrong hands. The passed in 'list' contains
8637 * several header fields at the beginning that are not part of the
8638 * inversion list body proper */
8640 const STRLEN length = (STRLEN) list[0];
8641 const UV version_id = list[1];
8642 const bool offset = cBOOL(list[2]);
8643 #define HEADER_LENGTH 3
8644 /* If any of the above changes in any way, you must change HEADER_LENGTH
8645 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8646 * perl -E 'say int(rand 2**31-1)'
8648 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8649 data structure type, so that one being
8650 passed in can be validated to be an
8651 inversion list of the correct vintage.
8654 SV* invlist = newSV_type(SVt_INVLIST);
8656 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8658 if (version_id != INVLIST_VERSION_ID) {
8659 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8662 /* The generated array passed in includes header elements that aren't part
8663 * of the list proper, so start it just after them */
8664 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8666 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8667 shouldn't touch it */
8669 *(get_invlist_offset_addr(invlist)) = offset;
8671 /* The 'length' passed to us is the physical number of elements in the
8672 * inversion list. But if there is an offset the logical number is one
8674 invlist_set_len(invlist, length - offset, offset);
8676 invlist_set_previous_index(invlist, 0);
8678 /* Initialize the iteration pointer. */
8679 invlist_iterfinish(invlist);
8681 SvREADONLY_on(invlist);
8687 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8689 /* Grow the maximum size of an inversion list */
8691 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8693 assert(SvTYPE(invlist) == SVt_INVLIST);
8695 /* Add one to account for the zero element at the beginning which may not
8696 * be counted by the calling parameters */
8697 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8701 S__append_range_to_invlist(pTHX_ SV* const invlist,
8702 const UV start, const UV end)
8704 /* Subject to change or removal. Append the range from 'start' to 'end' at
8705 * the end of the inversion list. The range must be above any existing
8709 UV max = invlist_max(invlist);
8710 UV len = _invlist_len(invlist);
8713 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8715 if (len == 0) { /* Empty lists must be initialized */
8716 offset = start != 0;
8717 array = _invlist_array_init(invlist, ! offset);
8720 /* Here, the existing list is non-empty. The current max entry in the
8721 * list is generally the first value not in the set, except when the
8722 * set extends to the end of permissible values, in which case it is
8723 * the first entry in that final set, and so this call is an attempt to
8724 * append out-of-order */
8726 UV final_element = len - 1;
8727 array = invlist_array(invlist);
8728 if ( array[final_element] > start
8729 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8731 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",
8732 array[final_element], start,
8733 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8736 /* Here, it is a legal append. If the new range begins 1 above the end
8737 * of the range below it, it is extending the range below it, so the
8738 * new first value not in the set is one greater than the newly
8739 * extended range. */
8740 offset = *get_invlist_offset_addr(invlist);
8741 if (array[final_element] == start) {
8742 if (end != UV_MAX) {
8743 array[final_element] = end + 1;
8746 /* But if the end is the maximum representable on the machine,
8747 * assume that infinity was actually what was meant. Just let
8748 * the range that this would extend to have no end */
8749 invlist_set_len(invlist, len - 1, offset);
8755 /* Here the new range doesn't extend any existing set. Add it */
8757 len += 2; /* Includes an element each for the start and end of range */
8759 /* If wll overflow the existing space, extend, which may cause the array to
8762 invlist_extend(invlist, len);
8764 /* Have to set len here to avoid assert failure in invlist_array() */
8765 invlist_set_len(invlist, len, offset);
8767 array = invlist_array(invlist);
8770 invlist_set_len(invlist, len, offset);
8773 /* The next item on the list starts the range, the one after that is
8774 * one past the new range. */
8775 array[len - 2] = start;
8776 if (end != UV_MAX) {
8777 array[len - 1] = end + 1;
8780 /* But if the end is the maximum representable on the machine, just let
8781 * the range have no end */
8782 invlist_set_len(invlist, len - 1, offset);
8787 Perl__invlist_search(SV* const invlist, const UV cp)
8789 /* Searches the inversion list for the entry that contains the input code
8790 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8791 * return value is the index into the list's array of the range that
8792 * contains <cp>, that is, 'i' such that
8793 * array[i] <= cp < array[i+1]
8798 IV high = _invlist_len(invlist);
8799 const IV highest_element = high - 1;
8802 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8804 /* If list is empty, return failure. */
8809 /* (We can't get the array unless we know the list is non-empty) */
8810 array = invlist_array(invlist);
8812 mid = invlist_previous_index(invlist);
8814 if (mid > highest_element) {
8815 mid = highest_element;
8818 /* <mid> contains the cache of the result of the previous call to this
8819 * function (0 the first time). See if this call is for the same result,
8820 * or if it is for mid-1. This is under the theory that calls to this
8821 * function will often be for related code points that are near each other.
8822 * And benchmarks show that caching gives better results. We also test
8823 * here if the code point is within the bounds of the list. These tests
8824 * replace others that would have had to be made anyway to make sure that
8825 * the array bounds were not exceeded, and these give us extra information
8826 * at the same time */
8827 if (cp >= array[mid]) {
8828 if (cp >= array[highest_element]) {
8829 return highest_element;
8832 /* Here, array[mid] <= cp < array[highest_element]. This means that
8833 * the final element is not the answer, so can exclude it; it also
8834 * means that <mid> is not the final element, so can refer to 'mid + 1'
8836 if (cp < array[mid + 1]) {
8842 else { /* cp < aray[mid] */
8843 if (cp < array[0]) { /* Fail if outside the array */
8847 if (cp >= array[mid - 1]) {
8852 /* Binary search. What we are looking for is <i> such that
8853 * array[i] <= cp < array[i+1]
8854 * The loop below converges on the i+1. Note that there may not be an
8855 * (i+1)th element in the array, and things work nonetheless */
8856 while (low < high) {
8857 mid = (low + high) / 2;
8858 assert(mid <= highest_element);
8859 if (array[mid] <= cp) { /* cp >= array[mid] */
8862 /* We could do this extra test to exit the loop early.
8863 if (cp < array[low]) {
8868 else { /* cp < array[mid] */
8875 invlist_set_previous_index(invlist, high);
8880 Perl__invlist_populate_swatch(SV* const invlist,
8881 const UV start, const UV end, U8* swatch)
8883 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8884 * but is used when the swash has an inversion list. This makes this much
8885 * faster, as it uses a binary search instead of a linear one. This is
8886 * intimately tied to that function, and perhaps should be in utf8.c,
8887 * except it is intimately tied to inversion lists as well. It assumes
8888 * that <swatch> is all 0's on input */
8891 const IV len = _invlist_len(invlist);
8895 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8897 if (len == 0) { /* Empty inversion list */
8901 array = invlist_array(invlist);
8903 /* Find which element it is */
8904 i = _invlist_search(invlist, start);
8906 /* We populate from <start> to <end> */
8907 while (current < end) {
8910 /* The inversion list gives the results for every possible code point
8911 * after the first one in the list. Only those ranges whose index is
8912 * even are ones that the inversion list matches. For the odd ones,
8913 * and if the initial code point is not in the list, we have to skip
8914 * forward to the next element */
8915 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8917 if (i >= len) { /* Finished if beyond the end of the array */
8921 if (current >= end) { /* Finished if beyond the end of what we
8923 if (LIKELY(end < UV_MAX)) {
8927 /* We get here when the upper bound is the maximum
8928 * representable on the machine, and we are looking for just
8929 * that code point. Have to special case it */
8931 goto join_end_of_list;
8934 assert(current >= start);
8936 /* The current range ends one below the next one, except don't go past
8939 upper = (i < len && array[i] < end) ? array[i] : end;
8941 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8942 * for each code point in it */
8943 for (; current < upper; current++) {
8944 const STRLEN offset = (STRLEN)(current - start);
8945 swatch[offset >> 3] |= 1 << (offset & 7);
8950 /* Quit if at the end of the list */
8953 /* But first, have to deal with the highest possible code point on
8954 * the platform. The previous code assumes that <end> is one
8955 * beyond where we want to populate, but that is impossible at the
8956 * platform's infinity, so have to handle it specially */
8957 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8959 const STRLEN offset = (STRLEN)(end - start);
8960 swatch[offset >> 3] |= 1 << (offset & 7);
8965 /* Advance to the next range, which will be for code points not in the
8974 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8975 const bool complement_b, SV** output)
8977 /* Take the union of two inversion lists and point '*output' to it. On
8978 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
8979 * even 'a' or 'b'). If to an inversion list, the contents of the original
8980 * list will be replaced by the union. The first list, 'a', may be
8981 * NULL, in which case a copy of the second list is placed in '*output'.
8982 * If 'complement_b' is TRUE, the union is taken of the complement
8983 * (inversion) of 'b' instead of b itself.
8985 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8986 * Richard Gillam, published by Addison-Wesley, and explained at some
8987 * length there. The preface says to incorporate its examples into your
8988 * code at your own risk.
8990 * The algorithm is like a merge sort. */
8992 const UV* array_a; /* a's array */
8994 UV len_a; /* length of a's array */
8997 SV* u; /* the resulting union */
9001 UV i_a = 0; /* current index into a's array */
9005 /* running count, as explained in the algorithm source book; items are
9006 * stopped accumulating and are output when the count changes to/from 0.
9007 * The count is incremented when we start a range that's in an input's set,
9008 * and decremented when we start a range that's not in a set. So this
9009 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9010 * and hence nothing goes into the union; 1, just one of the inputs is in
9011 * its set (and its current range gets added to the union); and 2 when both
9012 * inputs are in their sets. */
9015 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9017 assert(*output == NULL || SvTYPE(*output) == SVt_INVLIST);
9019 len_b = _invlist_len(b);
9022 /* Here, 'b' is empty, hence it's complement is all possible code
9023 * points. So if the union includes the complement of 'b', it includes
9024 * everything, and we need not even look at 'a'. It's easiest to
9025 * create a new inversion list that matches everything. */
9027 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9029 if (*output == NULL) { /* If the output didn't exist, just point it
9031 *output = everything;
9033 else { /* Otherwise, replace its contents with the new list */
9034 invlist_replace_list_destroys_src(*output, everything);
9035 SvREFCNT_dec_NN(everything);
9041 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9042 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9043 * output will be empty */
9045 if (a == NULL || _invlist_len(a) == 0) {
9046 if (*output == NULL) {
9047 *output = _new_invlist(0);
9050 invlist_clear(*output);
9055 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9056 * union. We can just return a copy of 'a' if '*output' doesn't point
9057 * to an existing list */
9058 if (*output == NULL) {
9059 *output = invlist_clone(a);
9063 /* If the output is to overwrite 'a', we have a no-op, as it's
9069 /* Here, '*output' is to be overwritten by 'a' */
9070 u = invlist_clone(a);
9071 invlist_replace_list_destroys_src(*output, u);
9077 /* Here 'b' is not empty. See about 'a' */
9079 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9081 /* Here, 'a' is empty (and b is not). That means the union will come
9082 * entirely from 'b'. If '*output' is NULL, we can directly return a
9083 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9086 SV ** dest = (*output == NULL) ? output : &u;
9087 *dest = invlist_clone(b);
9089 _invlist_invert(*dest);
9093 invlist_replace_list_destroys_src(*output, u);
9100 /* Here both lists exist and are non-empty */
9101 array_a = invlist_array(a);
9102 array_b = invlist_array(b);
9104 /* If are to take the union of 'a' with the complement of b, set it
9105 * up so are looking at b's complement. */
9108 /* To complement, we invert: if the first element is 0, remove it. To
9109 * do this, we just pretend the array starts one later */
9110 if (array_b[0] == 0) {
9116 /* But if the first element is not zero, we pretend the list starts
9117 * at the 0 that is always stored immediately before the array. */
9123 /* Size the union for the worst case: that the sets are completely
9125 u = _new_invlist(len_a + len_b);
9127 /* Will contain U+0000 if either component does */
9128 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9129 || (len_b > 0 && array_b[0] == 0));
9131 /* Go through each input list item by item, stopping when have exhausted
9133 while (i_a < len_a && i_b < len_b) {
9134 UV cp; /* The element to potentially add to the union's array */
9135 bool cp_in_set; /* is it in the the input list's set or not */
9137 /* We need to take one or the other of the two inputs for the union.
9138 * Since we are merging two sorted lists, we take the smaller of the
9139 * next items. In case of a tie, we take first the one that is in its
9140 * set. If we first took the one not in its set, it would decrement
9141 * the count, possibly to 0 which would cause it to be output as ending
9142 * the range, and the next time through we would take the same number,
9143 * and output it again as beginning the next range. By doing it the
9144 * opposite way, there is no possibility that the count will be
9145 * momentarily decremented to 0, and thus the two adjoining ranges will
9146 * be seamlessly merged. (In a tie and both are in the set or both not
9147 * in the set, it doesn't matter which we take first.) */
9148 if ( array_a[i_a] < array_b[i_b]
9149 || ( array_a[i_a] == array_b[i_b]
9150 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9152 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9153 cp = array_a[i_a++];
9156 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9157 cp = array_b[i_b++];
9160 /* Here, have chosen which of the two inputs to look at. Only output
9161 * if the running count changes to/from 0, which marks the
9162 * beginning/end of a range that's in the set */
9165 array_u[i_u++] = cp;
9172 array_u[i_u++] = cp;
9178 /* The loop above increments the index into exactly one of the input lists
9179 * each iteration, and ends when either index gets to its list end. That
9180 * means the other index is lower than its end, and so something is
9181 * remaining in that one. We decrement 'count', as explained below, if
9182 * that list is in its set. (i_a and i_b each currently index the element
9183 * beyond the one we care about.) */
9184 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9185 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9190 /* Above we decremented 'count' if the list that had unexamined elements in
9191 * it was in its set. This has made it so that 'count' being non-zero
9192 * means there isn't anything left to output; and 'count' equal to 0 means
9193 * that what is left to output is precisely that which is left in the
9194 * non-exhausted input list.
9196 * To see why, note first that the exhausted input obviously has nothing
9197 * left to add to the union. If it was in its set at its end, that means
9198 * the set extends from here to the platform's infinity, and hence so does
9199 * the union and the non-exhausted set is irrelevant. The exhausted set
9200 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9201 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9202 * 'count' remains at 1. This is consistent with the decremented 'count'
9203 * != 0 meaning there's nothing left to add to the union.
9205 * But if the exhausted input wasn't in its set, it contributed 0 to
9206 * 'count', and the rest of the union will be whatever the other input is.
9207 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9208 * otherwise it gets decremented to 0. This is consistent with 'count'
9209 * == 0 meaning the remainder of the union is whatever is left in the
9210 * non-exhausted list. */
9215 IV copy_count = len_a - i_a;
9216 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9217 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9219 else { /* The non-exhausted input is b */
9220 copy_count = len_b - i_b;
9221 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9223 len_u = i_u + copy_count;
9226 /* Set the result to the final length, which can change the pointer to
9227 * array_u, so re-find it. (Note that it is unlikely that this will
9228 * change, as we are shrinking the space, not enlarging it) */
9229 if (len_u != _invlist_len(u)) {
9230 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9232 array_u = invlist_array(u);
9235 if (*output == NULL) { /* Simply return the new inversion list */
9239 /* Otherwise, overwrite the inversion list that was in '*output'. We
9240 * could instead free '*output', and then set it to 'u', but experience
9241 * has shown [perl #127392] that if the input is a mortal, we can get a
9242 * huge build-up of these during regex compilation before they get
9244 invlist_replace_list_destroys_src(*output, u);
9252 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9253 const bool complement_b, SV** i)
9255 /* Take the intersection of two inversion lists and point '*i' to it. On
9256 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9257 * even 'a' or 'b'). If to an inversion list, the contents of the original
9258 * list will be replaced by the intersection. The first list, 'a', may be
9259 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9260 * TRUE, the result will be the intersection of 'a' and the complement (or
9261 * inversion) of 'b' instead of 'b' directly.
9263 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9264 * Richard Gillam, published by Addison-Wesley, and explained at some
9265 * length there. The preface says to incorporate its examples into your
9266 * code at your own risk. In fact, it had bugs
9268 * The algorithm is like a merge sort, and is essentially the same as the
9272 const UV* array_a; /* a's array */
9274 UV len_a; /* length of a's array */
9277 SV* r; /* the resulting intersection */
9281 UV i_a = 0; /* current index into a's array */
9285 /* running count of how many of the two inputs are postitioned at ranges
9286 * that are in their sets. As explained in the algorithm source book,
9287 * items are stopped accumulating and are output when the count changes
9288 * to/from 2. The count is incremented when we start a range that's in an
9289 * input's set, and decremented when we start a range that's not in a set.
9290 * Only when it is 2 are we in the intersection. */
9293 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9295 assert(*i == NULL || SvTYPE(*i) == SVt_INVLIST);
9297 /* Special case if either one is empty */
9298 len_a = (a == NULL) ? 0 : _invlist_len(a);
9299 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9300 if (len_a != 0 && complement_b) {
9302 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9303 * must be empty. Here, also we are using 'b's complement, which
9304 * hence must be every possible code point. Thus the intersection
9307 if (*i == a) { /* No-op */
9312 *i = invlist_clone(a);
9316 r = invlist_clone(a);
9317 invlist_replace_list_destroys_src(*i, r);
9322 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9323 * intersection must be empty */
9325 *i = _new_invlist(0);
9333 /* Here both lists exist and are non-empty */
9334 array_a = invlist_array(a);
9335 array_b = invlist_array(b);
9337 /* If are to take the intersection of 'a' with the complement of b, set it
9338 * up so are looking at b's complement. */
9341 /* To complement, we invert: if the first element is 0, remove it. To
9342 * do this, we just pretend the array starts one later */
9343 if (array_b[0] == 0) {
9349 /* But if the first element is not zero, we pretend the list starts
9350 * at the 0 that is always stored immediately before the array. */
9356 /* Size the intersection for the worst case: that the intersection ends up
9357 * fragmenting everything to be completely disjoint */
9358 r= _new_invlist(len_a + len_b);
9360 /* Will contain U+0000 iff both components do */
9361 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9362 && len_b > 0 && array_b[0] == 0);
9364 /* Go through each list item by item, stopping when have exhausted one of
9366 while (i_a < len_a && i_b < len_b) {
9367 UV cp; /* The element to potentially add to the intersection's
9369 bool cp_in_set; /* Is it in the input list's set or not */
9371 /* We need to take one or the other of the two inputs for the
9372 * intersection. Since we are merging two sorted lists, we take the
9373 * smaller of the next items. In case of a tie, we take first the one
9374 * that is not in its set (a difference from the union algorithm). If
9375 * we first took the one in its set, it would increment the count,
9376 * possibly to 2 which would cause it to be output as starting a range
9377 * in the intersection, and the next time through we would take that
9378 * same number, and output it again as ending the set. By doing the
9379 * opposite of this, there is no possibility that the count will be
9380 * momentarily incremented to 2. (In a tie and both are in the set or
9381 * both not in the set, it doesn't matter which we take first.) */
9382 if ( array_a[i_a] < array_b[i_b]
9383 || ( array_a[i_a] == array_b[i_b]
9384 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9386 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9387 cp = array_a[i_a++];
9390 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9394 /* Here, have chosen which of the two inputs to look at. Only output
9395 * if the running count changes to/from 2, which marks the
9396 * beginning/end of a range that's in the intersection */
9400 array_r[i_r++] = cp;
9405 array_r[i_r++] = cp;
9412 /* The loop above increments the index into exactly one of the input lists
9413 * each iteration, and ends when either index gets to its list end. That
9414 * means the other index is lower than its end, and so something is
9415 * remaining in that one. We increment 'count', as explained below, if the
9416 * exhausted list was in its set. (i_a and i_b each currently index the
9417 * element beyond the one we care about.) */
9418 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9419 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9424 /* Above we incremented 'count' if the exhausted list was in its set. This
9425 * has made it so that 'count' being below 2 means there is nothing left to
9426 * output; otheriwse what's left to add to the intersection is precisely
9427 * that which is left in the non-exhausted input list.
9429 * To see why, note first that the exhausted input obviously has nothing
9430 * left to affect the intersection. If it was in its set at its end, that
9431 * means the set extends from here to the platform's infinity, and hence
9432 * anything in the non-exhausted's list will be in the intersection, and
9433 * anything not in it won't be. Hence, the rest of the intersection is
9434 * precisely what's in the non-exhausted list The exhausted set also
9435 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9436 * it means 'count' is now at least 2. This is consistent with the
9437 * incremented 'count' being >= 2 means to add the non-exhausted list to
9440 * But if the exhausted input wasn't in its set, it contributed 0 to
9441 * 'count', and the intersection can't include anything further; the
9442 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9443 * incremented. This is consistent with 'count' being < 2 meaning nothing
9444 * further to add to the intersection. */
9445 if (count < 2) { /* Nothing left to put in the intersection. */
9448 else { /* copy the non-exhausted list, unchanged. */
9449 IV copy_count = len_a - i_a;
9450 if (copy_count > 0) { /* a is the one with stuff left */
9451 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9453 else { /* b is the one with stuff left */
9454 copy_count = len_b - i_b;
9455 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9457 len_r = i_r + copy_count;
9460 /* Set the result to the final length, which can change the pointer to
9461 * array_r, so re-find it. (Note that it is unlikely that this will
9462 * change, as we are shrinking the space, not enlarging it) */
9463 if (len_r != _invlist_len(r)) {
9464 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9466 array_r = invlist_array(r);
9469 if (*i == NULL) { /* Simply return the calculated intersection */
9472 else { /* Otherwise, replace the existing inversion list in '*i'. We could
9473 instead free '*i', and then set it to 'r', but experience has
9474 shown [perl #127392] that if the input is a mortal, we can get a
9475 huge build-up of these during regex compilation before they get
9478 invlist_replace_list_destroys_src(*i, r);
9490 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9492 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9493 * set. A pointer to the inversion list is returned. This may actually be
9494 * a new list, in which case the passed in one has been destroyed. The
9495 * passed-in inversion list can be NULL, in which case a new one is created
9496 * with just the one range in it. The new list is not necessarily
9497 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9498 * result of this function. The gain would not be large, and in many
9499 * cases, this is called multiple times on a single inversion list, so
9500 * anything freed may almost immediately be needed again.
9502 * This used to mostly call the 'union' routine, but that is much more
9503 * heavyweight than really needed for a single range addition */
9505 UV* array; /* The array implementing the inversion list */
9506 UV len; /* How many elements in 'array' */
9507 SSize_t i_s; /* index into the invlist array where 'start'
9509 SSize_t i_e = 0; /* And the index where 'end' should go */
9510 UV cur_highest; /* The highest code point in the inversion list
9511 upon entry to this function */
9513 /* This range becomes the whole inversion list if none already existed */
9514 if (invlist == NULL) {
9515 invlist = _new_invlist(2);
9516 _append_range_to_invlist(invlist, start, end);
9520 /* Likewise, if the inversion list is currently empty */
9521 len = _invlist_len(invlist);
9523 _append_range_to_invlist(invlist, start, end);
9527 /* Starting here, we have to know the internals of the list */
9528 array = invlist_array(invlist);
9530 /* If the new range ends higher than the current highest ... */
9531 cur_highest = invlist_highest(invlist);
9532 if (end > cur_highest) {
9534 /* If the whole range is higher, we can just append it */
9535 if (start > cur_highest) {
9536 _append_range_to_invlist(invlist, start, end);
9540 /* Otherwise, add the portion that is higher ... */
9541 _append_range_to_invlist(invlist, cur_highest + 1, end);
9543 /* ... and continue on below to handle the rest. As a result of the
9544 * above append, we know that the index of the end of the range is the
9545 * final even numbered one of the array. Recall that the final element
9546 * always starts a range that extends to infinity. If that range is in
9547 * the set (meaning the set goes from here to infinity), it will be an
9548 * even index, but if it isn't in the set, it's odd, and the final
9549 * range in the set is one less, which is even. */
9550 if (end == UV_MAX) {
9558 /* We have dealt with appending, now see about prepending. If the new
9559 * range starts lower than the current lowest ... */
9560 if (start < array[0]) {
9562 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
9563 * Let the union code handle it, rather than having to know the
9564 * trickiness in two code places. */
9565 if (UNLIKELY(start == 0)) {
9568 range_invlist = _new_invlist(2);
9569 _append_range_to_invlist(range_invlist, start, end);
9571 _invlist_union(invlist, range_invlist, &invlist);
9573 SvREFCNT_dec_NN(range_invlist);
9578 /* If the whole new range comes before the first entry, and doesn't
9579 * extend it, we have to insert it as an additional range */
9580 if (end < array[0] - 1) {
9582 goto splice_in_new_range;
9585 /* Here the new range adjoins the existing first range, extending it
9589 /* And continue on below to handle the rest. We know that the index of
9590 * the beginning of the range is the first one of the array */
9593 else { /* Not prepending any part of the new range to the existing list.
9594 * Find where in the list it should go. This finds i_s, such that:
9595 * invlist[i_s] <= start < array[i_s+1]
9597 i_s = _invlist_search(invlist, start);
9600 /* At this point, any extending before the beginning of the inversion list
9601 * and/or after the end has been done. This has made it so that, in the
9602 * code below, each endpoint of the new range is either in a range that is
9603 * in the set, or is in a gap between two ranges that are. This means we
9604 * don't have to worry about exceeding the array bounds.
9606 * Find where in the list the new range ends (but we can skip this if we
9607 * have already determined what it is, or if it will be the same as i_s,
9608 * which we already have computed) */
9610 i_e = (start == end)
9612 : _invlist_search(invlist, end);
9615 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
9616 * is a range that goes to infinity there is no element at invlist[i_e+1],
9617 * so only the first relation holds. */
9619 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9621 /* Here, the ranges on either side of the beginning of the new range
9622 * are in the set, and this range starts in the gap between them.
9624 * The new range extends the range above it downwards if the new range
9625 * ends at or above that range's start */
9626 const bool extends_the_range_above = ( end == UV_MAX
9627 || end + 1 >= array[i_s+1]);
9629 /* The new range extends the range below it upwards if it begins just
9630 * after where that range ends */
9631 if (start == array[i_s]) {
9633 /* If the new range fills the entire gap between the other ranges,
9634 * they will get merged together. Other ranges may also get
9635 * merged, depending on how many of them the new range spans. In
9636 * the general case, we do the merge later, just once, after we
9637 * figure out how many to merge. But in the case where the new
9638 * range exactly spans just this one gap (possibly extending into
9639 * the one above), we do the merge here, and an early exit. This
9640 * is done here to avoid having to special case later. */
9641 if (i_e - i_s <= 1) {
9643 /* If i_e - i_s == 1, it means that the new range terminates
9644 * within the range above, and hence 'extends_the_range_above'
9645 * must be true. (If the range above it extends to infinity,
9646 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
9647 * will be 0, so no harm done.) */
9648 if (extends_the_range_above) {
9649 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
9650 invlist_set_len(invlist,
9652 *(get_invlist_offset_addr(invlist)));
9656 /* Here, i_e must == i_s. We keep them in sync, as they apply
9657 * to the same range, and below we are about to decrement i_s
9662 /* Here, the new range is adjacent to the one below. (It may also
9663 * span beyond the range above, but that will get resolved later.)
9664 * Extend the range below to include this one. */
9665 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
9669 else if (extends_the_range_above) {
9671 /* Here the new range only extends the range above it, but not the
9672 * one below. It merges with the one above. Again, we keep i_e
9673 * and i_s in sync if they point to the same range */
9682 /* Here, we've dealt with the new range start extending any adjoining
9685 * If the new range extends to infinity, it is now the final one,
9686 * regardless of what was there before */
9687 if (UNLIKELY(end == UV_MAX)) {
9688 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
9692 /* If i_e started as == i_s, it has also been dealt with,
9693 * and been updated to the new i_s, which will fail the following if */
9694 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
9696 /* Here, the ranges on either side of the end of the new range are in
9697 * the set, and this range ends in the gap between them.
9699 * If this range is adjacent to (hence extends) the range above it, it
9700 * becomes part of that range; likewise if it extends the range below,
9701 * it becomes part of that range */
9702 if (end + 1 == array[i_e+1]) {
9706 else if (start <= array[i_e]) {
9707 array[i_e] = end + 1;
9714 /* If the range fits entirely in an existing range (as possibly already
9715 * extended above), it doesn't add anything new */
9716 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9720 /* Here, no part of the range is in the list. Must add it. It will
9721 * occupy 2 more slots */
9722 splice_in_new_range:
9724 invlist_extend(invlist, len + 2);
9725 array = invlist_array(invlist);
9726 /* Move the rest of the array down two slots. Don't include any
9728 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
9730 /* Do the actual splice */
9731 array[i_e+1] = start;
9732 array[i_e+2] = end + 1;
9733 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
9737 /* Here the new range crossed the boundaries of a pre-existing range. The
9738 * code above has adjusted things so that both ends are in ranges that are
9739 * in the set. This means everything in between must also be in the set.
9740 * Just squash things together */
9741 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
9742 invlist_set_len(invlist,
9744 *(get_invlist_offset_addr(invlist)));
9750 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9751 UV** other_elements_ptr)
9753 /* Create and return an inversion list whose contents are to be populated
9754 * by the caller. The caller gives the number of elements (in 'size') and
9755 * the very first element ('element0'). This function will set
9756 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9759 * Obviously there is some trust involved that the caller will properly
9760 * fill in the other elements of the array.
9762 * (The first element needs to be passed in, as the underlying code does
9763 * things differently depending on whether it is zero or non-zero) */
9765 SV* invlist = _new_invlist(size);
9768 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9770 invlist = add_cp_to_invlist(invlist, element0);
9771 offset = *get_invlist_offset_addr(invlist);
9773 invlist_set_len(invlist, size, offset);
9774 *other_elements_ptr = invlist_array(invlist) + 1;
9780 PERL_STATIC_INLINE SV*
9781 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9782 return _add_range_to_invlist(invlist, cp, cp);
9785 #ifndef PERL_IN_XSUB_RE
9787 Perl__invlist_invert(pTHX_ SV* const invlist)
9789 /* Complement the input inversion list. This adds a 0 if the list didn't
9790 * have a zero; removes it otherwise. As described above, the data
9791 * structure is set up so that this is very efficient */
9793 PERL_ARGS_ASSERT__INVLIST_INVERT;
9795 assert(! invlist_is_iterating(invlist));
9797 /* The inverse of matching nothing is matching everything */
9798 if (_invlist_len(invlist) == 0) {
9799 _append_range_to_invlist(invlist, 0, UV_MAX);
9803 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9808 PERL_STATIC_INLINE SV*
9809 S_invlist_clone(pTHX_ SV* const invlist)
9812 /* Return a new inversion list that is a copy of the input one, which is
9813 * unchanged. The new list will not be mortal even if the old one was. */
9815 /* Need to allocate extra space to accommodate Perl's addition of a
9816 * trailing NUL to SvPV's, since it thinks they are always strings */
9817 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9818 STRLEN physical_length = SvCUR(invlist);
9819 bool offset = *(get_invlist_offset_addr(invlist));
9821 PERL_ARGS_ASSERT_INVLIST_CLONE;
9823 *(get_invlist_offset_addr(new_invlist)) = offset;
9824 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9825 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9830 PERL_STATIC_INLINE STRLEN*
9831 S_get_invlist_iter_addr(SV* invlist)
9833 /* Return the address of the UV that contains the current iteration
9836 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9838 assert(SvTYPE(invlist) == SVt_INVLIST);
9840 return &(((XINVLIST*) SvANY(invlist))->iterator);
9843 PERL_STATIC_INLINE void
9844 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9846 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9848 *get_invlist_iter_addr(invlist) = 0;
9851 PERL_STATIC_INLINE void
9852 S_invlist_iterfinish(SV* invlist)
9854 /* Terminate iterator for invlist. This is to catch development errors.
9855 * Any iteration that is interrupted before completed should call this
9856 * function. Functions that add code points anywhere else but to the end
9857 * of an inversion list assert that they are not in the middle of an
9858 * iteration. If they were, the addition would make the iteration
9859 * problematical: if the iteration hadn't reached the place where things
9860 * were being added, it would be ok */
9862 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9864 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9868 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9870 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9871 * This call sets in <*start> and <*end>, the next range in <invlist>.
9872 * Returns <TRUE> if successful and the next call will return the next
9873 * range; <FALSE> if was already at the end of the list. If the latter,
9874 * <*start> and <*end> are unchanged, and the next call to this function
9875 * will start over at the beginning of the list */
9877 STRLEN* pos = get_invlist_iter_addr(invlist);
9878 UV len = _invlist_len(invlist);
9881 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9884 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9888 array = invlist_array(invlist);
9890 *start = array[(*pos)++];
9896 *end = array[(*pos)++] - 1;
9902 PERL_STATIC_INLINE UV
9903 S_invlist_highest(SV* const invlist)
9905 /* Returns the highest code point that matches an inversion list. This API
9906 * has an ambiguity, as it returns 0 under either the highest is actually
9907 * 0, or if the list is empty. If this distinction matters to you, check
9908 * for emptiness before calling this function */
9910 UV len = _invlist_len(invlist);
9913 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9919 array = invlist_array(invlist);
9921 /* The last element in the array in the inversion list always starts a
9922 * range that goes to infinity. That range may be for code points that are
9923 * matched in the inversion list, or it may be for ones that aren't
9924 * matched. In the latter case, the highest code point in the set is one
9925 * less than the beginning of this range; otherwise it is the final element
9926 * of this range: infinity */
9927 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
9929 : array[len - 1] - 1;
9933 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
9935 /* Get the contents of an inversion list into a string SV so that they can
9936 * be printed out. If 'traditional_style' is TRUE, it uses the format
9937 * traditionally done for debug tracing; otherwise it uses a format
9938 * suitable for just copying to the output, with blanks between ranges and
9939 * a dash between range components */
9943 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
9944 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
9946 if (traditional_style) {
9947 output = newSVpvs("\n");
9950 output = newSVpvs("");
9953 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
9955 assert(! invlist_is_iterating(invlist));
9957 invlist_iterinit(invlist);
9958 while (invlist_iternext(invlist, &start, &end)) {
9959 if (end == UV_MAX) {
9960 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFINITY%c",
9961 start, intra_range_delimiter,
9962 inter_range_delimiter);
9964 else if (end != start) {
9965 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
9967 intra_range_delimiter,
9968 end, inter_range_delimiter);
9971 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
9972 start, inter_range_delimiter);
9976 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
9977 SvCUR_set(output, SvCUR(output) - 1);
9983 #ifndef PERL_IN_XSUB_RE
9985 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
9986 const char * const indent, SV* const invlist)
9988 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
9989 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
9990 * the string 'indent'. The output looks like this:
9991 [0] 0x000A .. 0x000D
9993 [4] 0x2028 .. 0x2029
9994 [6] 0x3104 .. INFINITY
9995 * This means that the first range of code points matched by the list are
9996 * 0xA through 0xD; the second range contains only the single code point
9997 * 0x85, etc. An inversion list is an array of UVs. Two array elements
9998 * are used to define each range (except if the final range extends to
9999 * infinity, only a single element is needed). The array index of the
10000 * first element for the corresponding range is given in brackets. */
10005 PERL_ARGS_ASSERT__INVLIST_DUMP;
10007 if (invlist_is_iterating(invlist)) {
10008 Perl_dump_indent(aTHX_ level, file,
10009 "%sCan't dump inversion list because is in middle of iterating\n",
10014 invlist_iterinit(invlist);
10015 while (invlist_iternext(invlist, &start, &end)) {
10016 if (end == UV_MAX) {
10017 Perl_dump_indent(aTHX_ level, file,
10018 "%s[%" UVuf "] 0x%04" UVXf " .. INFINITY\n",
10019 indent, (UV)count, start);
10021 else if (end != start) {
10022 Perl_dump_indent(aTHX_ level, file,
10023 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10024 indent, (UV)count, start, end);
10027 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10028 indent, (UV)count, start);
10035 Perl__load_PL_utf8_foldclosures (pTHX)
10037 assert(! PL_utf8_foldclosures);
10039 /* If the folds haven't been read in, call a fold function
10041 if (! PL_utf8_tofold) {
10042 U8 dummy[UTF8_MAXBYTES_CASE+1];
10043 const U8 hyphen[] = HYPHEN_UTF8;
10045 /* This string is just a short named one above \xff */
10046 toFOLD_utf8_safe(hyphen, hyphen + sizeof(hyphen) - 1, dummy, NULL);
10047 assert(PL_utf8_tofold); /* Verify that worked */
10049 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
10053 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10055 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10057 /* Return a boolean as to if the two passed in inversion lists are
10058 * identical. The final argument, if TRUE, says to take the complement of
10059 * the second inversion list before doing the comparison */
10061 const UV* array_a = invlist_array(a);
10062 const UV* array_b = invlist_array(b);
10063 UV len_a = _invlist_len(a);
10064 UV len_b = _invlist_len(b);
10066 PERL_ARGS_ASSERT__INVLISTEQ;
10068 /* If are to compare 'a' with the complement of b, set it
10069 * up so are looking at b's complement. */
10070 if (complement_b) {
10072 /* The complement of nothing is everything, so <a> would have to have
10073 * just one element, starting at zero (ending at infinity) */
10075 return (len_a == 1 && array_a[0] == 0);
10077 else if (array_b[0] == 0) {
10079 /* Otherwise, to complement, we invert. Here, the first element is
10080 * 0, just remove it. To do this, we just pretend the array starts
10088 /* But if the first element is not zero, we pretend the list starts
10089 * at the 0 that is always stored immediately before the array. */
10095 return len_a == len_b
10096 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10102 * As best we can, determine the characters that can match the start of
10103 * the given EXACTF-ish node.
10105 * Returns the invlist as a new SV*; it is the caller's responsibility to
10106 * call SvREFCNT_dec() when done with it.
10109 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10111 const U8 * s = (U8*)STRING(node);
10112 SSize_t bytelen = STR_LEN(node);
10114 /* Start out big enough for 2 separate code points */
10115 SV* invlist = _new_invlist(4);
10117 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
10122 /* We punt and assume can match anything if the node begins
10123 * with a multi-character fold. Things are complicated. For
10124 * example, /ffi/i could match any of:
10125 * "\N{LATIN SMALL LIGATURE FFI}"
10126 * "\N{LATIN SMALL LIGATURE FF}I"
10127 * "F\N{LATIN SMALL LIGATURE FI}"
10128 * plus several other things; and making sure we have all the
10129 * possibilities is hard. */
10130 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10131 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10134 /* Any Latin1 range character can potentially match any
10135 * other depending on the locale */
10136 if (OP(node) == EXACTFL) {
10137 _invlist_union(invlist, PL_Latin1, &invlist);
10140 /* But otherwise, it matches at least itself. We can
10141 * quickly tell if it has a distinct fold, and if so,
10142 * it matches that as well */
10143 invlist = add_cp_to_invlist(invlist, uc);
10144 if (IS_IN_SOME_FOLD_L1(uc))
10145 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10148 /* Some characters match above-Latin1 ones under /i. This
10149 * is true of EXACTFL ones when the locale is UTF-8 */
10150 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10151 && (! isASCII(uc) || (OP(node) != EXACTFA
10152 && OP(node) != EXACTFA_NO_TRIE)))
10154 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10158 else { /* Pattern is UTF-8 */
10159 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10160 STRLEN foldlen = UTF8SKIP(s);
10161 const U8* e = s + bytelen;
10164 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10166 /* The only code points that aren't folded in a UTF EXACTFish
10167 * node are are the problematic ones in EXACTFL nodes */
10168 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10169 /* We need to check for the possibility that this EXACTFL
10170 * node begins with a multi-char fold. Therefore we fold
10171 * the first few characters of it so that we can make that
10176 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10178 *(d++) = (U8) toFOLD(*s);
10183 toFOLD_utf8_safe(s, e, d, &len);
10189 /* And set up so the code below that looks in this folded
10190 * buffer instead of the node's string */
10192 foldlen = UTF8SKIP(folded);
10196 /* When we reach here 's' points to the fold of the first
10197 * character(s) of the node; and 'e' points to far enough along
10198 * the folded string to be just past any possible multi-char
10199 * fold. 'foldlen' is the length in bytes of the first
10202 * Unlike the non-UTF-8 case, the macro for determining if a
10203 * string is a multi-char fold requires all the characters to
10204 * already be folded. This is because of all the complications
10205 * if not. Note that they are folded anyway, except in EXACTFL
10206 * nodes. Like the non-UTF case above, we punt if the node
10207 * begins with a multi-char fold */
10209 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10210 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10212 else { /* Single char fold */
10214 /* It matches all the things that fold to it, which are
10215 * found in PL_utf8_foldclosures (including itself) */
10216 invlist = add_cp_to_invlist(invlist, uc);
10217 if (! PL_utf8_foldclosures)
10218 _load_PL_utf8_foldclosures();
10219 if ((listp = hv_fetch(PL_utf8_foldclosures,
10220 (char *) s, foldlen, FALSE)))
10222 AV* list = (AV*) *listp;
10224 for (k = 0; k <= av_tindex_nomg(list); k++) {
10225 SV** c_p = av_fetch(list, k, FALSE);
10231 /* /aa doesn't allow folds between ASCII and non- */
10232 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
10233 && isASCII(c) != isASCII(uc))
10238 invlist = add_cp_to_invlist(invlist, c);
10247 #undef HEADER_LENGTH
10248 #undef TO_INTERNAL_SIZE
10249 #undef FROM_INTERNAL_SIZE
10250 #undef INVLIST_VERSION_ID
10252 /* End of inversion list object */
10255 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10257 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10258 * constructs, and updates RExC_flags with them. On input, RExC_parse
10259 * should point to the first flag; it is updated on output to point to the
10260 * final ')' or ':'. There needs to be at least one flag, or this will
10263 /* for (?g), (?gc), and (?o) warnings; warning
10264 about (?c) will warn about (?g) -- japhy */
10266 #define WASTED_O 0x01
10267 #define WASTED_G 0x02
10268 #define WASTED_C 0x04
10269 #define WASTED_GC (WASTED_G|WASTED_C)
10270 I32 wastedflags = 0x00;
10271 U32 posflags = 0, negflags = 0;
10272 U32 *flagsp = &posflags;
10273 char has_charset_modifier = '\0';
10275 bool has_use_defaults = FALSE;
10276 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10277 int x_mod_count = 0;
10279 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10281 /* '^' as an initial flag sets certain defaults */
10282 if (UCHARAT(RExC_parse) == '^') {
10284 has_use_defaults = TRUE;
10285 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10286 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
10287 ? REGEX_UNICODE_CHARSET
10288 : REGEX_DEPENDS_CHARSET);
10291 cs = get_regex_charset(RExC_flags);
10292 if (cs == REGEX_DEPENDS_CHARSET
10293 && (RExC_utf8 || RExC_uni_semantics))
10295 cs = REGEX_UNICODE_CHARSET;
10298 while (RExC_parse < RExC_end) {
10299 /* && strchr("iogcmsx", *RExC_parse) */
10300 /* (?g), (?gc) and (?o) are useless here
10301 and must be globally applied -- japhy */
10302 switch (*RExC_parse) {
10304 /* Code for the imsxn flags */
10305 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10307 case LOCALE_PAT_MOD:
10308 if (has_charset_modifier) {
10309 goto excess_modifier;
10311 else if (flagsp == &negflags) {
10314 cs = REGEX_LOCALE_CHARSET;
10315 has_charset_modifier = LOCALE_PAT_MOD;
10317 case UNICODE_PAT_MOD:
10318 if (has_charset_modifier) {
10319 goto excess_modifier;
10321 else if (flagsp == &negflags) {
10324 cs = REGEX_UNICODE_CHARSET;
10325 has_charset_modifier = UNICODE_PAT_MOD;
10327 case ASCII_RESTRICT_PAT_MOD:
10328 if (flagsp == &negflags) {
10331 if (has_charset_modifier) {
10332 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10333 goto excess_modifier;
10335 /* Doubled modifier implies more restricted */
10336 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10339 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10341 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10343 case DEPENDS_PAT_MOD:
10344 if (has_use_defaults) {
10345 goto fail_modifiers;
10347 else if (flagsp == &negflags) {
10350 else if (has_charset_modifier) {
10351 goto excess_modifier;
10354 /* The dual charset means unicode semantics if the
10355 * pattern (or target, not known until runtime) are
10356 * utf8, or something in the pattern indicates unicode
10358 cs = (RExC_utf8 || RExC_uni_semantics)
10359 ? REGEX_UNICODE_CHARSET
10360 : REGEX_DEPENDS_CHARSET;
10361 has_charset_modifier = DEPENDS_PAT_MOD;
10365 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10366 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10368 else if (has_charset_modifier == *(RExC_parse - 1)) {
10369 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10370 *(RExC_parse - 1));
10373 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10375 NOT_REACHED; /*NOTREACHED*/
10378 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10379 *(RExC_parse - 1));
10380 NOT_REACHED; /*NOTREACHED*/
10381 case ONCE_PAT_MOD: /* 'o' */
10382 case GLOBAL_PAT_MOD: /* 'g' */
10383 if (PASS2 && ckWARN(WARN_REGEXP)) {
10384 const I32 wflagbit = *RExC_parse == 'o'
10387 if (! (wastedflags & wflagbit) ) {
10388 wastedflags |= wflagbit;
10389 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10392 "Useless (%s%c) - %suse /%c modifier",
10393 flagsp == &negflags ? "?-" : "?",
10395 flagsp == &negflags ? "don't " : "",
10402 case CONTINUE_PAT_MOD: /* 'c' */
10403 if (PASS2 && ckWARN(WARN_REGEXP)) {
10404 if (! (wastedflags & WASTED_C) ) {
10405 wastedflags |= WASTED_GC;
10406 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10409 "Useless (%sc) - %suse /gc modifier",
10410 flagsp == &negflags ? "?-" : "?",
10411 flagsp == &negflags ? "don't " : ""
10416 case KEEPCOPY_PAT_MOD: /* 'p' */
10417 if (flagsp == &negflags) {
10419 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10421 *flagsp |= RXf_PMf_KEEPCOPY;
10425 /* A flag is a default iff it is following a minus, so
10426 * if there is a minus, it means will be trying to
10427 * re-specify a default which is an error */
10428 if (has_use_defaults || flagsp == &negflags) {
10429 goto fail_modifiers;
10431 flagsp = &negflags;
10432 wastedflags = 0; /* reset so (?g-c) warns twice */
10438 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
10439 negflags |= RXf_PMf_EXTENDED_MORE;
10441 RExC_flags |= posflags;
10443 if (negflags & RXf_PMf_EXTENDED) {
10444 negflags |= RXf_PMf_EXTENDED_MORE;
10446 RExC_flags &= ~negflags;
10447 set_regex_charset(&RExC_flags, cs);
10452 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10453 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10454 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
10455 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10456 NOT_REACHED; /*NOTREACHED*/
10459 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10462 vFAIL("Sequence (?... not terminated");
10466 - reg - regular expression, i.e. main body or parenthesized thing
10468 * Caller must absorb opening parenthesis.
10470 * Combining parenthesis handling with the base level of regular expression
10471 * is a trifle forced, but the need to tie the tails of the branches to what
10472 * follows makes it hard to avoid.
10474 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10476 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10478 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10481 PERL_STATIC_INLINE regnode *
10482 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10484 char * parse_start,
10489 char* name_start = RExC_parse;
10491 SV *sv_dat = reg_scan_name(pRExC_state, SIZE_ONLY
10492 ? REG_RSN_RETURN_NULL
10493 : REG_RSN_RETURN_DATA);
10494 GET_RE_DEBUG_FLAGS_DECL;
10496 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10498 if (RExC_parse == name_start || *RExC_parse != ch) {
10499 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10500 vFAIL2("Sequence %.3s... not terminated",parse_start);
10504 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10505 RExC_rxi->data->data[num]=(void*)sv_dat;
10506 SvREFCNT_inc_simple_void(sv_dat);
10509 ret = reganode(pRExC_state,
10512 : (ASCII_FOLD_RESTRICTED)
10514 : (AT_LEAST_UNI_SEMANTICS)
10520 *flagp |= HASWIDTH;
10522 Set_Node_Offset(ret, parse_start+1);
10523 Set_Node_Cur_Length(ret, parse_start);
10525 nextchar(pRExC_state);
10529 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
10530 flags. Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan
10531 needs to be restarted, or'd with NEED_UTF8 if the pattern needs to be
10532 upgraded to UTF-8. Otherwise would only return NULL if regbranch() returns
10533 NULL, which cannot happen. */
10535 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
10536 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10537 * 2 is like 1, but indicates that nextchar() has been called to advance
10538 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10539 * this flag alerts us to the need to check for that */
10541 regnode *ret; /* Will be the head of the group. */
10544 regnode *ender = NULL;
10547 U32 oregflags = RExC_flags;
10548 bool have_branch = 0;
10550 I32 freeze_paren = 0;
10551 I32 after_freeze = 0;
10552 I32 num; /* numeric backreferences */
10554 char * parse_start = RExC_parse; /* MJD */
10555 char * const oregcomp_parse = RExC_parse;
10557 GET_RE_DEBUG_FLAGS_DECL;
10559 PERL_ARGS_ASSERT_REG;
10560 DEBUG_PARSE("reg ");
10562 *flagp = 0; /* Tentatively. */
10564 /* Having this true makes it feasible to have a lot fewer tests for the
10565 * parse pointer being in scope. For example, we can write
10566 * while(isFOO(*RExC_parse)) RExC_parse++;
10568 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
10570 assert(*RExC_end == '\0');
10572 /* Make an OPEN node, if parenthesized. */
10575 /* Under /x, space and comments can be gobbled up between the '(' and
10576 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
10577 * intervening space, as the sequence is a token, and a token should be
10579 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
10581 if (RExC_parse >= RExC_end) {
10582 vFAIL("Unmatched (");
10585 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
10586 char *start_verb = RExC_parse + 1;
10588 char *start_arg = NULL;
10589 unsigned char op = 0;
10590 int arg_required = 0;
10591 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
10593 if (has_intervening_patws) {
10594 RExC_parse++; /* past the '*' */
10595 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
10597 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
10598 if ( *RExC_parse == ':' ) {
10599 start_arg = RExC_parse + 1;
10602 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10604 verb_len = RExC_parse - start_verb;
10606 if (RExC_parse >= RExC_end) {
10607 goto unterminated_verb_pattern;
10609 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10610 while ( RExC_parse < RExC_end && *RExC_parse != ')' )
10611 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10612 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10613 unterminated_verb_pattern:
10614 vFAIL("Unterminated verb pattern argument");
10615 if ( RExC_parse == start_arg )
10618 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10619 vFAIL("Unterminated verb pattern");
10622 /* Here, we know that RExC_parse < RExC_end */
10624 switch ( *start_verb ) {
10625 case 'A': /* (*ACCEPT) */
10626 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
10628 internal_argval = RExC_nestroot;
10631 case 'C': /* (*COMMIT) */
10632 if ( memEQs(start_verb,verb_len,"COMMIT") )
10635 case 'F': /* (*FAIL) */
10636 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
10640 case ':': /* (*:NAME) */
10641 case 'M': /* (*MARK:NAME) */
10642 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
10647 case 'P': /* (*PRUNE) */
10648 if ( memEQs(start_verb,verb_len,"PRUNE") )
10651 case 'S': /* (*SKIP) */
10652 if ( memEQs(start_verb,verb_len,"SKIP") )
10655 case 'T': /* (*THEN) */
10656 /* [19:06] <TimToady> :: is then */
10657 if ( memEQs(start_verb,verb_len,"THEN") ) {
10659 RExC_seen |= REG_CUTGROUP_SEEN;
10664 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10666 "Unknown verb pattern '%" UTF8f "'",
10667 UTF8fARG(UTF, verb_len, start_verb));
10669 if ( arg_required && !start_arg ) {
10670 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
10671 verb_len, start_verb);
10673 if (internal_argval == -1) {
10674 ret = reganode(pRExC_state, op, 0);
10676 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
10678 RExC_seen |= REG_VERBARG_SEEN;
10679 if ( ! SIZE_ONLY ) {
10681 SV *sv = newSVpvn( start_arg,
10682 RExC_parse - start_arg);
10683 ARG(ret) = add_data( pRExC_state,
10684 STR_WITH_LEN("S"));
10685 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
10690 if ( internal_argval != -1 )
10691 ARG2L_SET(ret, internal_argval);
10693 nextchar(pRExC_state);
10696 else if (*RExC_parse == '?') { /* (?...) */
10697 bool is_logical = 0;
10698 const char * const seqstart = RExC_parse;
10699 const char * endptr;
10700 if (has_intervening_patws) {
10702 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
10705 RExC_parse++; /* past the '?' */
10706 paren = *RExC_parse; /* might be a trailing NUL, if not
10708 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10709 if (RExC_parse > RExC_end) {
10712 ret = NULL; /* For look-ahead/behind. */
10715 case 'P': /* (?P...) variants for those used to PCRE/Python */
10716 paren = *RExC_parse;
10717 if ( paren == '<') { /* (?P<...>) named capture */
10719 if (RExC_parse >= RExC_end) {
10720 vFAIL("Sequence (?P<... not terminated");
10722 goto named_capture;
10724 else if (paren == '>') { /* (?P>name) named recursion */
10726 if (RExC_parse >= RExC_end) {
10727 vFAIL("Sequence (?P>... not terminated");
10729 goto named_recursion;
10731 else if (paren == '=') { /* (?P=...) named backref */
10733 return handle_named_backref(pRExC_state, flagp,
10736 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10737 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10738 vFAIL3("Sequence (%.*s...) not recognized",
10739 RExC_parse-seqstart, seqstart);
10740 NOT_REACHED; /*NOTREACHED*/
10741 case '<': /* (?<...) */
10742 if (*RExC_parse == '!')
10744 else if (*RExC_parse != '=')
10751 case '\'': /* (?'...') */
10752 name_start = RExC_parse;
10753 svname = reg_scan_name(pRExC_state,
10754 SIZE_ONLY /* reverse test from the others */
10755 ? REG_RSN_RETURN_NAME
10756 : REG_RSN_RETURN_NULL);
10757 if ( RExC_parse == name_start
10758 || RExC_parse >= RExC_end
10759 || *RExC_parse != paren)
10761 vFAIL2("Sequence (?%c... not terminated",
10762 paren=='>' ? '<' : paren);
10767 if (!svname) /* shouldn't happen */
10769 "panic: reg_scan_name returned NULL");
10770 if (!RExC_paren_names) {
10771 RExC_paren_names= newHV();
10772 sv_2mortal(MUTABLE_SV(RExC_paren_names));
10774 RExC_paren_name_list= newAV();
10775 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10778 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10780 sv_dat = HeVAL(he_str);
10782 /* croak baby croak */
10784 "panic: paren_name hash element allocation failed");
10785 } else if ( SvPOK(sv_dat) ) {
10786 /* (?|...) can mean we have dupes so scan to check
10787 its already been stored. Maybe a flag indicating
10788 we are inside such a construct would be useful,
10789 but the arrays are likely to be quite small, so
10790 for now we punt -- dmq */
10791 IV count = SvIV(sv_dat);
10792 I32 *pv = (I32*)SvPVX(sv_dat);
10794 for ( i = 0 ; i < count ; i++ ) {
10795 if ( pv[i] == RExC_npar ) {
10801 pv = (I32*)SvGROW(sv_dat,
10802 SvCUR(sv_dat) + sizeof(I32)+1);
10803 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10804 pv[count] = RExC_npar;
10805 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10808 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10809 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10812 SvIV_set(sv_dat, 1);
10815 /* Yes this does cause a memory leak in debugging Perls
10817 if (!av_store(RExC_paren_name_list,
10818 RExC_npar, SvREFCNT_inc(svname)))
10819 SvREFCNT_dec_NN(svname);
10822 /*sv_dump(sv_dat);*/
10824 nextchar(pRExC_state);
10826 goto capturing_parens;
10828 RExC_seen |= REG_LOOKBEHIND_SEEN;
10829 RExC_in_lookbehind++;
10831 if (RExC_parse >= RExC_end) {
10832 vFAIL("Sequence (?... not terminated");
10836 case '=': /* (?=...) */
10837 RExC_seen_zerolen++;
10839 case '!': /* (?!...) */
10840 RExC_seen_zerolen++;
10841 /* check if we're really just a "FAIL" assertion */
10842 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
10843 FALSE /* Don't force to /x */ );
10844 if (*RExC_parse == ')') {
10845 ret=reganode(pRExC_state, OPFAIL, 0);
10846 nextchar(pRExC_state);
10850 case '|': /* (?|...) */
10851 /* branch reset, behave like a (?:...) except that
10852 buffers in alternations share the same numbers */
10854 after_freeze = freeze_paren = RExC_npar;
10856 case ':': /* (?:...) */
10857 case '>': /* (?>...) */
10859 case '$': /* (?$...) */
10860 case '@': /* (?@...) */
10861 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10863 case '0' : /* (?0) */
10864 case 'R' : /* (?R) */
10865 if (RExC_parse == RExC_end || *RExC_parse != ')')
10866 FAIL("Sequence (?R) not terminated");
10868 RExC_seen |= REG_RECURSE_SEEN;
10869 *flagp |= POSTPONED;
10870 goto gen_recurse_regop;
10872 /* named and numeric backreferences */
10873 case '&': /* (?&NAME) */
10874 parse_start = RExC_parse - 1;
10877 SV *sv_dat = reg_scan_name(pRExC_state,
10878 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10879 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10881 if (RExC_parse >= RExC_end || *RExC_parse != ')')
10882 vFAIL("Sequence (?&... not terminated");
10883 goto gen_recurse_regop;
10886 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10888 vFAIL("Illegal pattern");
10890 goto parse_recursion;
10892 case '-': /* (?-1) */
10893 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10894 RExC_parse--; /* rewind to let it be handled later */
10898 case '1': case '2': case '3': case '4': /* (?1) */
10899 case '5': case '6': case '7': case '8': case '9':
10900 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
10903 bool is_neg = FALSE;
10905 parse_start = RExC_parse - 1; /* MJD */
10906 if (*RExC_parse == '-') {
10910 if (grok_atoUV(RExC_parse, &unum, &endptr)
10914 RExC_parse = (char*)endptr;
10918 /* Some limit for num? */
10922 if (*RExC_parse!=')')
10923 vFAIL("Expecting close bracket");
10926 if ( paren == '-' ) {
10928 Diagram of capture buffer numbering.
10929 Top line is the normal capture buffer numbers
10930 Bottom line is the negative indexing as from
10934 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
10938 num = RExC_npar + num;
10941 vFAIL("Reference to nonexistent group");
10943 } else if ( paren == '+' ) {
10944 num = RExC_npar + num - 1;
10946 /* We keep track how many GOSUB items we have produced.
10947 To start off the ARG2L() of the GOSUB holds its "id",
10948 which is used later in conjunction with RExC_recurse
10949 to calculate the offset we need to jump for the GOSUB,
10950 which it will store in the final representation.
10951 We have to defer the actual calculation until much later
10952 as the regop may move.
10955 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
10957 if (num > (I32)RExC_rx->nparens) {
10959 vFAIL("Reference to nonexistent group");
10961 RExC_recurse_count++;
10962 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
10963 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
10964 22, "| |", (int)(depth * 2 + 1), "",
10965 (UV)ARG(ret), (IV)ARG2L(ret)));
10967 RExC_seen |= REG_RECURSE_SEEN;
10969 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
10970 Set_Node_Offset(ret, parse_start); /* MJD */
10972 *flagp |= POSTPONED;
10973 assert(*RExC_parse == ')');
10974 nextchar(pRExC_state);
10979 case '?': /* (??...) */
10981 if (*RExC_parse != '{') {
10982 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10983 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10985 "Sequence (%" UTF8f "...) not recognized",
10986 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10987 NOT_REACHED; /*NOTREACHED*/
10989 *flagp |= POSTPONED;
10993 case '{': /* (?{...}) */
10996 struct reg_code_block *cb;
10998 RExC_seen_zerolen++;
11000 if ( !pRExC_state->num_code_blocks
11001 || pRExC_state->code_index >= pRExC_state->num_code_blocks
11002 || pRExC_state->code_blocks[pRExC_state->code_index].start
11003 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11006 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11007 FAIL("panic: Sequence (?{...}): no code block found\n");
11008 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11010 /* this is a pre-compiled code block (?{...}) */
11011 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
11012 RExC_parse = RExC_start + cb->end;
11015 if (cb->src_regex) {
11016 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11017 RExC_rxi->data->data[n] =
11018 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11019 RExC_rxi->data->data[n+1] = (void*)o;
11022 n = add_data(pRExC_state,
11023 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11024 RExC_rxi->data->data[n] = (void*)o;
11027 pRExC_state->code_index++;
11028 nextchar(pRExC_state);
11032 ret = reg_node(pRExC_state, LOGICAL);
11034 eval = reg2Lanode(pRExC_state, EVAL,
11037 /* for later propagation into (??{})
11039 RExC_flags & RXf_PMf_COMPILETIME
11044 REGTAIL(pRExC_state, ret, eval);
11045 /* deal with the length of this later - MJD */
11048 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11049 Set_Node_Length(ret, RExC_parse - parse_start + 1);
11050 Set_Node_Offset(ret, parse_start);
11053 case '(': /* (?(?{...})...) and (?(?=...)...) */
11056 const int DEFINE_len = sizeof("DEFINE") - 1;
11057 if (RExC_parse[0] == '?') { /* (?(?...)) */
11058 if ( RExC_parse < RExC_end - 1
11059 && ( RExC_parse[1] == '='
11060 || RExC_parse[1] == '!'
11061 || RExC_parse[1] == '<'
11062 || RExC_parse[1] == '{')
11063 ) { /* Lookahead or eval. */
11067 ret = reg_node(pRExC_state, LOGICAL);
11071 tail = reg(pRExC_state, 1, &flag, depth+1);
11072 if (flag & (RESTART_PASS1|NEED_UTF8)) {
11073 *flagp = flag & (RESTART_PASS1|NEED_UTF8);
11076 REGTAIL(pRExC_state, ret, tail);
11079 /* Fall through to ‘Unknown switch condition’ at the
11080 end of the if/else chain. */
11082 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11083 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11085 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11086 char *name_start= RExC_parse++;
11088 SV *sv_dat=reg_scan_name(pRExC_state,
11089 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11090 if ( RExC_parse == name_start
11091 || RExC_parse >= RExC_end
11092 || *RExC_parse != ch)
11094 vFAIL2("Sequence (?(%c... not terminated",
11095 (ch == '>' ? '<' : ch));
11099 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11100 RExC_rxi->data->data[num]=(void*)sv_dat;
11101 SvREFCNT_inc_simple_void(sv_dat);
11103 ret = reganode(pRExC_state,NGROUPP,num);
11104 goto insert_if_check_paren;
11106 else if (RExC_end - RExC_parse >= DEFINE_len
11107 && strnEQ(RExC_parse, "DEFINE", DEFINE_len))
11109 ret = reganode(pRExC_state,DEFINEP,0);
11110 RExC_parse += DEFINE_len;
11112 goto insert_if_check_paren;
11114 else if (RExC_parse[0] == 'R') {
11116 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11117 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11118 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11121 if (RExC_parse[0] == '0') {
11125 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11127 if (grok_atoUV(RExC_parse, &uv, &endptr)
11130 parno = (I32)uv + 1;
11131 RExC_parse = (char*)endptr;
11133 /* else "Switch condition not recognized" below */
11134 } else if (RExC_parse[0] == '&') {
11137 sv_dat = reg_scan_name(pRExC_state,
11139 ? REG_RSN_RETURN_NULL
11140 : REG_RSN_RETURN_DATA);
11142 /* we should only have a false sv_dat when
11143 * SIZE_ONLY is true, and we always have false
11144 * sv_dat when SIZE_ONLY is true.
11145 * reg_scan_name() will VFAIL() if the name is
11146 * unknown when SIZE_ONLY is false, and otherwise
11147 * will return something, and when SIZE_ONLY is
11148 * true, reg_scan_name() just parses the string,
11149 * and doesnt return anything. (in theory) */
11150 assert(SIZE_ONLY ? !sv_dat : !!sv_dat);
11153 parno = 1 + *((I32 *)SvPVX(sv_dat));
11155 ret = reganode(pRExC_state,INSUBP,parno);
11156 goto insert_if_check_paren;
11158 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11162 if (grok_atoUV(RExC_parse, &uv, &endptr)
11166 RExC_parse = (char*)endptr;
11169 vFAIL("panic: grok_atoUV returned FALSE");
11171 ret = reganode(pRExC_state, GROUPP, parno);
11173 insert_if_check_paren:
11174 if (UCHARAT(RExC_parse) != ')') {
11175 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11176 vFAIL("Switch condition not recognized");
11178 nextchar(pRExC_state);
11180 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
11181 br = regbranch(pRExC_state, &flags, 1,depth+1);
11183 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11184 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11187 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11190 REGTAIL(pRExC_state, br, reganode(pRExC_state,
11192 c = UCHARAT(RExC_parse);
11193 nextchar(pRExC_state);
11194 if (flags&HASWIDTH)
11195 *flagp |= HASWIDTH;
11198 vFAIL("(?(DEFINE)....) does not allow branches");
11200 /* Fake one for optimizer. */
11201 lastbr = reganode(pRExC_state, IFTHEN, 0);
11203 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
11204 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11205 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11208 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11211 REGTAIL(pRExC_state, ret, lastbr);
11212 if (flags&HASWIDTH)
11213 *flagp |= HASWIDTH;
11214 c = UCHARAT(RExC_parse);
11215 nextchar(pRExC_state);
11220 if (RExC_parse >= RExC_end)
11221 vFAIL("Switch (?(condition)... not terminated");
11223 vFAIL("Switch (?(condition)... contains too many branches");
11225 ender = reg_node(pRExC_state, TAIL);
11226 REGTAIL(pRExC_state, br, ender);
11228 REGTAIL(pRExC_state, lastbr, ender);
11229 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11232 REGTAIL(pRExC_state, ret, ender);
11233 RExC_size++; /* XXX WHY do we need this?!!
11234 For large programs it seems to be required
11235 but I can't figure out why. -- dmq*/
11238 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11239 vFAIL("Unknown switch condition (?(...))");
11241 case '[': /* (?[ ... ]) */
11242 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
11244 case 0: /* A NUL */
11245 RExC_parse--; /* for vFAIL to print correctly */
11246 vFAIL("Sequence (? incomplete");
11248 default: /* e.g., (?i) */
11249 RExC_parse = (char *) seqstart + 1;
11251 parse_lparen_question_flags(pRExC_state);
11252 if (UCHARAT(RExC_parse) != ':') {
11253 if (RExC_parse < RExC_end)
11254 nextchar(pRExC_state);
11259 nextchar(pRExC_state);
11264 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11269 ret = reganode(pRExC_state, OPEN, parno);
11271 if (!RExC_nestroot)
11272 RExC_nestroot = parno;
11273 if (RExC_open_parens && !RExC_open_parens[parno])
11275 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11276 "%*s%*s Setting open paren #%" IVdf " to %d\n",
11277 22, "| |", (int)(depth * 2 + 1), "",
11278 (IV)parno, REG_NODE_NUM(ret)));
11279 RExC_open_parens[parno]= ret;
11282 Set_Node_Length(ret, 1); /* MJD */
11283 Set_Node_Offset(ret, RExC_parse); /* MJD */
11286 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
11295 /* Pick up the branches, linking them together. */
11296 parse_start = RExC_parse; /* MJD */
11297 br = regbranch(pRExC_state, &flags, 1,depth+1);
11299 /* branch_len = (paren != 0); */
11302 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11303 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11306 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11308 if (*RExC_parse == '|') {
11309 if (!SIZE_ONLY && RExC_extralen) {
11310 reginsert(pRExC_state, BRANCHJ, br, depth+1);
11313 reginsert(pRExC_state, BRANCH, br, depth+1);
11314 Set_Node_Length(br, paren != 0);
11315 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
11319 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
11321 else if (paren == ':') {
11322 *flagp |= flags&SIMPLE;
11324 if (is_open) { /* Starts with OPEN. */
11325 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
11327 else if (paren != '?') /* Not Conditional */
11329 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11331 while (*RExC_parse == '|') {
11332 if (!SIZE_ONLY && RExC_extralen) {
11333 ender = reganode(pRExC_state, LONGJMP,0);
11335 /* Append to the previous. */
11336 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11339 RExC_extralen += 2; /* Account for LONGJMP. */
11340 nextchar(pRExC_state);
11341 if (freeze_paren) {
11342 if (RExC_npar > after_freeze)
11343 after_freeze = RExC_npar;
11344 RExC_npar = freeze_paren;
11346 br = regbranch(pRExC_state, &flags, 0, depth+1);
11349 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11350 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11353 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11355 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
11357 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11360 if (have_branch || paren != ':') {
11361 /* Make a closing node, and hook it on the end. */
11364 ender = reg_node(pRExC_state, TAIL);
11367 ender = reganode(pRExC_state, CLOSE, parno);
11368 if ( RExC_close_parens ) {
11369 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11370 "%*s%*s Setting close paren #%" IVdf " to %d\n",
11371 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
11372 RExC_close_parens[parno]= ender;
11373 if (RExC_nestroot == parno)
11376 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
11377 Set_Node_Length(ender,1); /* MJD */
11383 *flagp &= ~HASWIDTH;
11386 ender = reg_node(pRExC_state, SUCCEED);
11389 ender = reg_node(pRExC_state, END);
11391 assert(!RExC_end_op); /* there can only be one! */
11392 RExC_end_op = ender;
11393 if (RExC_close_parens) {
11394 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11395 "%*s%*s Setting close paren #0 (END) to %d\n",
11396 22, "| |", (int)(depth * 2 + 1), "", REG_NODE_NUM(ender)));
11398 RExC_close_parens[0]= ender;
11403 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11404 DEBUG_PARSE_MSG("lsbr");
11405 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
11406 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11407 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11408 SvPV_nolen_const(RExC_mysv1),
11409 (IV)REG_NODE_NUM(lastbr),
11410 SvPV_nolen_const(RExC_mysv2),
11411 (IV)REG_NODE_NUM(ender),
11412 (IV)(ender - lastbr)
11415 REGTAIL(pRExC_state, lastbr, ender);
11417 if (have_branch && !SIZE_ONLY) {
11418 char is_nothing= 1;
11420 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
11422 /* Hook the tails of the branches to the closing node. */
11423 for (br = ret; br; br = regnext(br)) {
11424 const U8 op = PL_regkind[OP(br)];
11425 if (op == BRANCH) {
11426 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
11427 if ( OP(NEXTOPER(br)) != NOTHING
11428 || regnext(NEXTOPER(br)) != ender)
11431 else if (op == BRANCHJ) {
11432 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
11433 /* for now we always disable this optimisation * /
11434 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
11435 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
11441 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
11442 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11443 DEBUG_PARSE_MSG("NADA");
11444 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
11445 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11446 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11447 SvPV_nolen_const(RExC_mysv1),
11448 (IV)REG_NODE_NUM(ret),
11449 SvPV_nolen_const(RExC_mysv2),
11450 (IV)REG_NODE_NUM(ender),
11455 if (OP(ender) == TAIL) {
11460 for ( opt= br + 1; opt < ender ; opt++ )
11461 OP(opt)= OPTIMIZED;
11462 NEXT_OFF(br)= ender - br;
11470 static const char parens[] = "=!<,>";
11472 if (paren && (p = strchr(parens, paren))) {
11473 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
11474 int flag = (p - parens) > 1;
11477 node = SUSPEND, flag = 0;
11478 reginsert(pRExC_state, node,ret, depth+1);
11479 Set_Node_Cur_Length(ret, parse_start);
11480 Set_Node_Offset(ret, parse_start + 1);
11482 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
11486 /* Check for proper termination. */
11488 /* restore original flags, but keep (?p) and, if we've changed from /d
11489 * rules to /u, keep the /u */
11490 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
11491 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
11492 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
11494 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
11495 RExC_parse = oregcomp_parse;
11496 vFAIL("Unmatched (");
11498 nextchar(pRExC_state);
11500 else if (!paren && RExC_parse < RExC_end) {
11501 if (*RExC_parse == ')') {
11503 vFAIL("Unmatched )");
11506 FAIL("Junk on end of regexp"); /* "Can't happen". */
11507 NOT_REACHED; /* NOTREACHED */
11510 if (RExC_in_lookbehind) {
11511 RExC_in_lookbehind--;
11513 if (after_freeze > RExC_npar)
11514 RExC_npar = after_freeze;
11519 - regbranch - one alternative of an | operator
11521 * Implements the concatenation operator.
11523 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11524 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11527 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
11530 regnode *chain = NULL;
11532 I32 flags = 0, c = 0;
11533 GET_RE_DEBUG_FLAGS_DECL;
11535 PERL_ARGS_ASSERT_REGBRANCH;
11537 DEBUG_PARSE("brnc");
11542 if (!SIZE_ONLY && RExC_extralen)
11543 ret = reganode(pRExC_state, BRANCHJ,0);
11545 ret = reg_node(pRExC_state, BRANCH);
11546 Set_Node_Length(ret, 1);
11550 if (!first && SIZE_ONLY)
11551 RExC_extralen += 1; /* BRANCHJ */
11553 *flagp = WORST; /* Tentatively. */
11555 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11556 FALSE /* Don't force to /x */ );
11557 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
11558 flags &= ~TRYAGAIN;
11559 latest = regpiece(pRExC_state, &flags,depth+1);
11560 if (latest == NULL) {
11561 if (flags & TRYAGAIN)
11563 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11564 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11567 FAIL2("panic: regpiece returned NULL, flags=%#" UVxf, (UV) flags);
11569 else if (ret == NULL)
11571 *flagp |= flags&(HASWIDTH|POSTPONED);
11572 if (chain == NULL) /* First piece. */
11573 *flagp |= flags&SPSTART;
11575 /* FIXME adding one for every branch after the first is probably
11576 * excessive now we have TRIE support. (hv) */
11578 REGTAIL(pRExC_state, chain, latest);
11583 if (chain == NULL) { /* Loop ran zero times. */
11584 chain = reg_node(pRExC_state, NOTHING);
11589 *flagp |= flags&SIMPLE;
11596 - regpiece - something followed by possible quantifier * + ? {n,m}
11598 * Note that the branching code sequences used for ? and the general cases
11599 * of * and + are somewhat optimized: they use the same NOTHING node as
11600 * both the endmarker for their branch list and the body of the last branch.
11601 * It might seem that this node could be dispensed with entirely, but the
11602 * endmarker role is not redundant.
11604 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
11606 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11607 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11610 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11616 const char * const origparse = RExC_parse;
11618 I32 max = REG_INFTY;
11619 #ifdef RE_TRACK_PATTERN_OFFSETS
11622 const char *maxpos = NULL;
11625 /* Save the original in case we change the emitted regop to a FAIL. */
11626 regnode * const orig_emit = RExC_emit;
11628 GET_RE_DEBUG_FLAGS_DECL;
11630 PERL_ARGS_ASSERT_REGPIECE;
11632 DEBUG_PARSE("piec");
11634 ret = regatom(pRExC_state, &flags,depth+1);
11636 if (flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8))
11637 *flagp |= flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8);
11639 FAIL2("panic: regatom returned NULL, flags=%#" UVxf, (UV) flags);
11645 if (op == '{' && regcurly(RExC_parse)) {
11647 #ifdef RE_TRACK_PATTERN_OFFSETS
11648 parse_start = RExC_parse; /* MJD */
11650 next = RExC_parse + 1;
11651 while (isDIGIT(*next) || *next == ',') {
11652 if (*next == ',') {
11660 if (*next == '}') { /* got one */
11661 const char* endptr;
11665 if (isDIGIT(*RExC_parse)) {
11666 if (!grok_atoUV(RExC_parse, &uv, &endptr))
11667 vFAIL("Invalid quantifier in {,}");
11668 if (uv >= REG_INFTY)
11669 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11674 if (*maxpos == ',')
11677 maxpos = RExC_parse;
11678 if (isDIGIT(*maxpos)) {
11679 if (!grok_atoUV(maxpos, &uv, &endptr))
11680 vFAIL("Invalid quantifier in {,}");
11681 if (uv >= REG_INFTY)
11682 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11685 max = REG_INFTY; /* meaning "infinity" */
11688 nextchar(pRExC_state);
11689 if (max < min) { /* If can't match, warn and optimize to fail
11693 /* We can't back off the size because we have to reserve
11694 * enough space for all the things we are about to throw
11695 * away, but we can shrink it by the amount we are about
11696 * to re-use here */
11697 RExC_size += PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
11700 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
11701 RExC_emit = orig_emit;
11703 ret = reganode(pRExC_state, OPFAIL, 0);
11706 else if (min == max && *RExC_parse == '?')
11709 ckWARN2reg(RExC_parse + 1,
11710 "Useless use of greediness modifier '%c'",
11716 if ((flags&SIMPLE)) {
11717 if (min == 0 && max == REG_INFTY) {
11718 reginsert(pRExC_state, STAR, ret, depth+1);
11721 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11724 if (min == 1 && max == REG_INFTY) {
11725 reginsert(pRExC_state, PLUS, ret, depth+1);
11728 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11731 MARK_NAUGHTY_EXP(2, 2);
11732 reginsert(pRExC_state, CURLY, ret, depth+1);
11733 Set_Node_Offset(ret, parse_start+1); /* MJD */
11734 Set_Node_Cur_Length(ret, parse_start);
11737 regnode * const w = reg_node(pRExC_state, WHILEM);
11740 REGTAIL(pRExC_state, ret, w);
11741 if (!SIZE_ONLY && RExC_extralen) {
11742 reginsert(pRExC_state, LONGJMP,ret, depth+1);
11743 reginsert(pRExC_state, NOTHING,ret, depth+1);
11744 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
11746 reginsert(pRExC_state, CURLYX,ret, depth+1);
11748 Set_Node_Offset(ret, parse_start+1);
11749 Set_Node_Length(ret,
11750 op == '{' ? (RExC_parse - parse_start) : 1);
11752 if (!SIZE_ONLY && RExC_extralen)
11753 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
11754 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
11756 RExC_whilem_seen++, RExC_extralen += 3;
11757 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
11764 *flagp |= HASWIDTH;
11766 ARG1_SET(ret, (U16)min);
11767 ARG2_SET(ret, (U16)max);
11769 if (max == REG_INFTY)
11770 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11776 if (!ISMULT1(op)) {
11781 #if 0 /* Now runtime fix should be reliable. */
11783 /* if this is reinstated, don't forget to put this back into perldiag:
11785 =item Regexp *+ operand could be empty at {#} in regex m/%s/
11787 (F) The part of the regexp subject to either the * or + quantifier
11788 could match an empty string. The {#} shows in the regular
11789 expression about where the problem was discovered.
11793 if (!(flags&HASWIDTH) && op != '?')
11794 vFAIL("Regexp *+ operand could be empty");
11797 #ifdef RE_TRACK_PATTERN_OFFSETS
11798 parse_start = RExC_parse;
11800 nextchar(pRExC_state);
11802 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
11808 else if (op == '+') {
11812 else if (op == '?') {
11817 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
11818 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
11819 ckWARN2reg(RExC_parse,
11820 "%" UTF8f " matches null string many times",
11821 UTF8fARG(UTF, (RExC_parse >= origparse
11822 ? RExC_parse - origparse
11825 (void)ReREFCNT_inc(RExC_rx_sv);
11828 if (*RExC_parse == '?') {
11829 nextchar(pRExC_state);
11830 reginsert(pRExC_state, MINMOD, ret, depth+1);
11831 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
11833 else if (*RExC_parse == '+') {
11835 nextchar(pRExC_state);
11836 ender = reg_node(pRExC_state, SUCCEED);
11837 REGTAIL(pRExC_state, ret, ender);
11838 reginsert(pRExC_state, SUSPEND, ret, depth+1);
11840 ender = reg_node(pRExC_state, TAIL);
11841 REGTAIL(pRExC_state, ret, ender);
11844 if (ISMULT2(RExC_parse)) {
11846 vFAIL("Nested quantifiers");
11853 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
11862 /* This routine teases apart the various meanings of \N and returns
11863 * accordingly. The input parameters constrain which meaning(s) is/are valid
11864 * in the current context.
11866 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
11868 * If <code_point_p> is not NULL, the context is expecting the result to be a
11869 * single code point. If this \N instance turns out to a single code point,
11870 * the function returns TRUE and sets *code_point_p to that code point.
11872 * If <node_p> is not NULL, the context is expecting the result to be one of
11873 * the things representable by a regnode. If this \N instance turns out to be
11874 * one such, the function generates the regnode, returns TRUE and sets *node_p
11875 * to point to that regnode.
11877 * If this instance of \N isn't legal in any context, this function will
11878 * generate a fatal error and not return.
11880 * On input, RExC_parse should point to the first char following the \N at the
11881 * time of the call. On successful return, RExC_parse will have been updated
11882 * to point to just after the sequence identified by this routine. Also
11883 * *flagp has been updated as needed.
11885 * When there is some problem with the current context and this \N instance,
11886 * the function returns FALSE, without advancing RExC_parse, nor setting
11887 * *node_p, nor *code_point_p, nor *flagp.
11889 * If <cp_count> is not NULL, the caller wants to know the length (in code
11890 * points) that this \N sequence matches. This is set even if the function
11891 * returns FALSE, as detailed below.
11893 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
11895 * Probably the most common case is for the \N to specify a single code point.
11896 * *cp_count will be set to 1, and *code_point_p will be set to that code
11899 * Another possibility is for the input to be an empty \N{}, which for
11900 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
11901 * will be set to a generated NOTHING node.
11903 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
11904 * set to 0. *node_p will be set to a generated REG_ANY node.
11906 * The fourth possibility is that \N resolves to a sequence of more than one
11907 * code points. *cp_count will be set to the number of code points in the
11908 * sequence. *node_p * will be set to a generated node returned by this
11909 * function calling S_reg().
11911 * The final possibility is that it is premature to be calling this function;
11912 * that pass1 needs to be restarted. This can happen when this changes from
11913 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
11914 * latter occurs only when the fourth possibility would otherwise be in
11915 * effect, and is because one of those code points requires the pattern to be
11916 * recompiled as UTF-8. The function returns FALSE, and sets the
11917 * RESTART_PASS1 and NEED_UTF8 flags in *flagp, as appropriate. When this
11918 * happens, the caller needs to desist from continuing parsing, and return
11919 * this information to its caller. This is not set for when there is only one
11920 * code point, as this can be called as part of an ANYOF node, and they can
11921 * store above-Latin1 code points without the pattern having to be in UTF-8.
11923 * For non-single-quoted regexes, the tokenizer has resolved character and
11924 * sequence names inside \N{...} into their Unicode values, normalizing the
11925 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
11926 * hex-represented code points in the sequence. This is done there because
11927 * the names can vary based on what charnames pragma is in scope at the time,
11928 * so we need a way to take a snapshot of what they resolve to at the time of
11929 * the original parse. [perl #56444].
11931 * That parsing is skipped for single-quoted regexes, so we may here get
11932 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
11933 * parser. But if the single-quoted regex is something like '\N{U+41}', that
11934 * is legal and handled here. The code point is Unicode, and has to be
11935 * translated into the native character set for non-ASCII platforms.
11938 char * endbrace; /* points to '}' following the name */
11939 char *endchar; /* Points to '.' or '}' ending cur char in the input
11941 char* p = RExC_parse; /* Temporary */
11943 GET_RE_DEBUG_FLAGS_DECL;
11945 PERL_ARGS_ASSERT_GROK_BSLASH_N;
11947 GET_RE_DEBUG_FLAGS;
11949 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
11950 assert(! (node_p && cp_count)); /* At most 1 should be set */
11952 if (cp_count) { /* Initialize return for the most common case */
11956 /* The [^\n] meaning of \N ignores spaces and comments under the /x
11957 * modifier. The other meanings do not, so use a temporary until we find
11958 * out which we are being called with */
11959 skip_to_be_ignored_text(pRExC_state, &p,
11960 FALSE /* Don't force to /x */ );
11962 /* Disambiguate between \N meaning a named character versus \N meaning
11963 * [^\n]. The latter is assumed when the {...} following the \N is a legal
11964 * quantifier, or there is no '{' at all */
11965 if (*p != '{' || regcurly(p)) {
11975 *node_p = reg_node(pRExC_state, REG_ANY);
11976 *flagp |= HASWIDTH|SIMPLE;
11978 Set_Node_Length(*node_p, 1); /* MJD */
11982 /* Here, we have decided it should be a named character or sequence */
11984 /* The test above made sure that the next real character is a '{', but
11985 * under the /x modifier, it could be separated by space (or a comment and
11986 * \n) and this is not allowed (for consistency with \x{...} and the
11987 * tokenizer handling of \N{NAME}). */
11988 if (*RExC_parse != '{') {
11989 vFAIL("Missing braces on \\N{}");
11992 RExC_parse++; /* Skip past the '{' */
11994 if (! (endbrace = strchr(RExC_parse, '}'))) { /* no trailing brace */
11995 vFAIL2("Missing right brace on \\%c{}", 'N');
11997 else if(!(endbrace == RExC_parse /* nothing between the {} */
11998 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked... */
11999 && strnEQ(RExC_parse, "U+", 2)))) /* ... below for a better
12002 RExC_parse = endbrace; /* position msg's '<--HERE' */
12003 vFAIL("\\N{NAME} must be resolved by the lexer");
12006 REQUIRE_UNI_RULES(flagp, FALSE); /* Unicode named chars imply Unicode
12009 if (endbrace == RExC_parse) { /* empty: \N{} */
12011 RExC_parse++; /* Position after the "}" */
12012 vFAIL("Zero length \\N{}");
12017 nextchar(pRExC_state);
12022 *node_p = reg_node(pRExC_state,NOTHING);
12026 RExC_parse += 2; /* Skip past the 'U+' */
12028 /* Because toke.c has generated a special construct for us guaranteed not
12029 * to have NULs, we can use a str function */
12030 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12032 /* Code points are separated by dots. If none, there is only one code
12033 * point, and is terminated by the brace */
12035 if (endchar >= endbrace) {
12036 STRLEN length_of_hex;
12037 I32 grok_hex_flags;
12039 /* Here, exactly one code point. If that isn't what is wanted, fail */
12040 if (! code_point_p) {
12045 /* Convert code point from hex */
12046 length_of_hex = (STRLEN)(endchar - RExC_parse);
12047 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
12048 | PERL_SCAN_DISALLOW_PREFIX
12050 /* No errors in the first pass (See [perl
12051 * #122671].) We let the code below find the
12052 * errors when there are multiple chars. */
12054 ? PERL_SCAN_SILENT_ILLDIGIT
12057 /* This routine is the one place where both single- and double-quotish
12058 * \N{U+xxxx} are evaluated. The value is a Unicode code point which
12059 * must be converted to native. */
12060 *code_point_p = UNI_TO_NATIVE(grok_hex(RExC_parse,
12065 /* The tokenizer should have guaranteed validity, but it's possible to
12066 * bypass it by using single quoting, so check. Don't do the check
12067 * here when there are multiple chars; we do it below anyway. */
12068 if (length_of_hex == 0
12069 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
12071 RExC_parse += length_of_hex; /* Includes all the valid */
12072 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12073 ? UTF8SKIP(RExC_parse)
12075 /* Guard against malformed utf8 */
12076 if (RExC_parse >= endchar) {
12077 RExC_parse = endchar;
12079 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12082 RExC_parse = endbrace + 1;
12085 else { /* Is a multiple character sequence */
12086 SV * substitute_parse;
12088 char *orig_end = RExC_end;
12089 char *save_start = RExC_start;
12092 /* Count the code points, if desired, in the sequence */
12095 while (RExC_parse < endbrace) {
12096 /* Point to the beginning of the next character in the sequence. */
12097 RExC_parse = endchar + 1;
12098 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12103 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12104 * But don't backup up the pointer if the caller want to know how many
12105 * code points there are (they can then handle things) */
12113 /* What is done here is to convert this to a sub-pattern of the form
12114 * \x{char1}\x{char2}... and then call reg recursively to parse it
12115 * (enclosing in "(?: ... )" ). That way, it retains its atomicness,
12116 * while not having to worry about special handling that some code
12117 * points may have. */
12119 substitute_parse = newSVpvs("?:");
12121 while (RExC_parse < endbrace) {
12123 /* Convert to notation the rest of the code understands */
12124 sv_catpv(substitute_parse, "\\x{");
12125 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
12126 sv_catpv(substitute_parse, "}");
12128 /* Point to the beginning of the next character in the sequence. */
12129 RExC_parse = endchar + 1;
12130 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12133 sv_catpv(substitute_parse, ")");
12135 RExC_parse = RExC_start = RExC_adjusted_start = SvPV(substitute_parse,
12138 /* Don't allow empty number */
12139 if (len < (STRLEN) 8) {
12140 RExC_parse = endbrace;
12141 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12143 RExC_end = RExC_parse + len;
12145 /* The values are Unicode, and therefore not subject to recoding, but
12146 * have to be converted to native on a non-Unicode (meaning non-ASCII)
12149 RExC_recode_x_to_native = 1;
12153 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
12154 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12155 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12158 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#" UVxf,
12161 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12164 /* Restore the saved values */
12165 RExC_start = RExC_adjusted_start = save_start;
12166 RExC_parse = endbrace;
12167 RExC_end = orig_end;
12169 RExC_recode_x_to_native = 0;
12172 SvREFCNT_dec_NN(substitute_parse);
12173 nextchar(pRExC_state);
12180 PERL_STATIC_INLINE U8
12181 S_compute_EXACTish(RExC_state_t *pRExC_state)
12185 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
12193 op = get_regex_charset(RExC_flags);
12194 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
12195 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
12196 been, so there is no hole */
12199 return op + EXACTF;
12202 PERL_STATIC_INLINE void
12203 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
12204 regnode *node, I32* flagp, STRLEN len, UV code_point,
12207 /* This knows the details about sizing an EXACTish node, setting flags for
12208 * it (by setting <*flagp>, and potentially populating it with a single
12211 * If <len> (the length in bytes) is non-zero, this function assumes that
12212 * the node has already been populated, and just does the sizing. In this
12213 * case <code_point> should be the final code point that has already been
12214 * placed into the node. This value will be ignored except that under some
12215 * circumstances <*flagp> is set based on it.
12217 * If <len> is zero, the function assumes that the node is to contain only
12218 * the single character given by <code_point> and calculates what <len>
12219 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
12220 * additionally will populate the node's STRING with <code_point> or its
12223 * In both cases <*flagp> is appropriately set
12225 * It knows that under FOLD, the Latin Sharp S and UTF characters above
12226 * 255, must be folded (the former only when the rules indicate it can
12229 * When it does the populating, it looks at the flag 'downgradable'. If
12230 * true with a node that folds, it checks if the single code point
12231 * participates in a fold, and if not downgrades the node to an EXACT.
12232 * This helps the optimizer */
12234 bool len_passed_in = cBOOL(len != 0);
12235 U8 character[UTF8_MAXBYTES_CASE+1];
12237 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
12239 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
12240 * sizing difference, and is extra work that is thrown away */
12241 if (downgradable && ! PASS2) {
12242 downgradable = FALSE;
12245 if (! len_passed_in) {
12247 if (UVCHR_IS_INVARIANT(code_point)) {
12248 if (LOC || ! FOLD) { /* /l defers folding until runtime */
12249 *character = (U8) code_point;
12251 else { /* Here is /i and not /l. (toFOLD() is defined on just
12252 ASCII, which isn't the same thing as INVARIANT on
12253 EBCDIC, but it works there, as the extra invariants
12254 fold to themselves) */
12255 *character = toFOLD((U8) code_point);
12257 /* We can downgrade to an EXACT node if this character
12258 * isn't a folding one. Note that this assumes that
12259 * nothing above Latin1 folds to some other invariant than
12260 * one of these alphabetics; otherwise we would also have
12262 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12263 * || ASCII_FOLD_RESTRICTED))
12265 if (downgradable && PL_fold[code_point] == code_point) {
12271 else if (FOLD && (! LOC
12272 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
12273 { /* Folding, and ok to do so now */
12274 UV folded = _to_uni_fold_flags(
12278 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12279 ? FOLD_FLAGS_NOMIX_ASCII
12282 && folded == code_point /* This quickly rules out many
12283 cases, avoiding the
12284 _invlist_contains_cp() overhead
12286 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
12293 else if (code_point <= MAX_UTF8_TWO_BYTE) {
12295 /* Not folding this cp, and can output it directly */
12296 *character = UTF8_TWO_BYTE_HI(code_point);
12297 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
12301 uvchr_to_utf8( character, code_point);
12302 len = UTF8SKIP(character);
12304 } /* Else pattern isn't UTF8. */
12306 *character = (U8) code_point;
12308 } /* Else is folded non-UTF8 */
12309 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12310 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12311 || UNICODE_DOT_DOT_VERSION > 0)
12312 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
12316 /* We don't fold any non-UTF8 except possibly the Sharp s (see
12317 * comments at join_exact()); */
12318 *character = (U8) code_point;
12321 /* Can turn into an EXACT node if we know the fold at compile time,
12322 * and it folds to itself and doesn't particpate in other folds */
12325 && PL_fold_latin1[code_point] == code_point
12326 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12327 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
12331 } /* else is Sharp s. May need to fold it */
12332 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
12334 *(character + 1) = 's';
12338 *character = LATIN_SMALL_LETTER_SHARP_S;
12344 RExC_size += STR_SZ(len);
12347 RExC_emit += STR_SZ(len);
12348 STR_LEN(node) = len;
12349 if (! len_passed_in) {
12350 Copy((char *) character, STRING(node), len, char);
12354 *flagp |= HASWIDTH;
12356 /* A single character node is SIMPLE, except for the special-cased SHARP S
12358 if ((len == 1 || (UTF && len == UVCHR_SKIP(code_point)))
12359 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12360 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12361 || UNICODE_DOT_DOT_VERSION > 0)
12362 && ( code_point != LATIN_SMALL_LETTER_SHARP_S
12363 || ! FOLD || ! DEPENDS_SEMANTICS)
12369 /* The OP may not be well defined in PASS1 */
12370 if (PASS2 && OP(node) == EXACTFL) {
12371 RExC_contains_locale = 1;
12376 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
12377 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
12380 S_backref_value(char *p)
12382 const char* endptr;
12384 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
12391 - regatom - the lowest level
12393 Try to identify anything special at the start of the current parse position.
12394 If there is, then handle it as required. This may involve generating a
12395 single regop, such as for an assertion; or it may involve recursing, such as
12396 to handle a () structure.
12398 If the string doesn't start with something special then we gobble up
12399 as much literal text as we can. If we encounter a quantifier, we have to
12400 back off the final literal character, as that quantifier applies to just it
12401 and not to the whole string of literals.
12403 Once we have been able to handle whatever type of thing started the
12404 sequence, we return.
12406 Note: we have to be careful with escapes, as they can be both literal
12407 and special, and in the case of \10 and friends, context determines which.
12409 A summary of the code structure is:
12411 switch (first_byte) {
12412 cases for each special:
12413 handle this special;
12416 switch (2nd byte) {
12417 cases for each unambiguous special:
12418 handle this special;
12420 cases for each ambigous special/literal:
12422 if (special) handle here
12424 default: // unambiguously literal:
12427 default: // is a literal char
12430 create EXACTish node for literal;
12431 while (more input and node isn't full) {
12432 switch (input_byte) {
12433 cases for each special;
12434 make sure parse pointer is set so that the next call to
12435 regatom will see this special first
12436 goto loopdone; // EXACTish node terminated by prev. char
12438 append char to EXACTISH node;
12440 get next input byte;
12444 return the generated node;
12446 Specifically there are two separate switches for handling
12447 escape sequences, with the one for handling literal escapes requiring
12448 a dummy entry for all of the special escapes that are actually handled
12451 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
12453 Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
12454 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
12455 Otherwise does not return NULL.
12459 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12461 regnode *ret = NULL;
12468 GET_RE_DEBUG_FLAGS_DECL;
12470 *flagp = WORST; /* Tentatively. */
12472 DEBUG_PARSE("atom");
12474 PERL_ARGS_ASSERT_REGATOM;
12477 parse_start = RExC_parse;
12478 assert(RExC_parse < RExC_end);
12479 switch ((U8)*RExC_parse) {
12481 RExC_seen_zerolen++;
12482 nextchar(pRExC_state);
12483 if (RExC_flags & RXf_PMf_MULTILINE)
12484 ret = reg_node(pRExC_state, MBOL);
12486 ret = reg_node(pRExC_state, SBOL);
12487 Set_Node_Length(ret, 1); /* MJD */
12490 nextchar(pRExC_state);
12492 RExC_seen_zerolen++;
12493 if (RExC_flags & RXf_PMf_MULTILINE)
12494 ret = reg_node(pRExC_state, MEOL);
12496 ret = reg_node(pRExC_state, SEOL);
12497 Set_Node_Length(ret, 1); /* MJD */
12500 nextchar(pRExC_state);
12501 if (RExC_flags & RXf_PMf_SINGLELINE)
12502 ret = reg_node(pRExC_state, SANY);
12504 ret = reg_node(pRExC_state, REG_ANY);
12505 *flagp |= HASWIDTH|SIMPLE;
12507 Set_Node_Length(ret, 1); /* MJD */
12511 char * const oregcomp_parse = ++RExC_parse;
12512 ret = regclass(pRExC_state, flagp,depth+1,
12513 FALSE, /* means parse the whole char class */
12514 TRUE, /* allow multi-char folds */
12515 FALSE, /* don't silence non-portable warnings. */
12516 (bool) RExC_strict,
12517 TRUE, /* Allow an optimized regnode result */
12521 if (*flagp & (RESTART_PASS1|NEED_UTF8))
12523 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12526 if (*RExC_parse != ']') {
12527 RExC_parse = oregcomp_parse;
12528 vFAIL("Unmatched [");
12530 nextchar(pRExC_state);
12531 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
12535 nextchar(pRExC_state);
12536 ret = reg(pRExC_state, 2, &flags,depth+1);
12538 if (flags & TRYAGAIN) {
12539 if (RExC_parse >= RExC_end) {
12540 /* Make parent create an empty node if needed. */
12541 *flagp |= TRYAGAIN;
12546 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12547 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12550 FAIL2("panic: reg returned NULL to regatom, flags=%#" UVxf,
12553 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12557 if (flags & TRYAGAIN) {
12558 *flagp |= TRYAGAIN;
12561 vFAIL("Internal urp");
12562 /* Supposed to be caught earlier. */
12568 vFAIL("Quantifier follows nothing");
12573 This switch handles escape sequences that resolve to some kind
12574 of special regop and not to literal text. Escape sequnces that
12575 resolve to literal text are handled below in the switch marked
12578 Every entry in this switch *must* have a corresponding entry
12579 in the literal escape switch. However, the opposite is not
12580 required, as the default for this switch is to jump to the
12581 literal text handling code.
12584 switch ((U8)*RExC_parse) {
12585 /* Special Escapes */
12587 RExC_seen_zerolen++;
12588 ret = reg_node(pRExC_state, SBOL);
12589 /* SBOL is shared with /^/ so we set the flags so we can tell
12590 * /\A/ from /^/ in split. We check ret because first pass we
12591 * have no regop struct to set the flags on. */
12595 goto finish_meta_pat;
12597 ret = reg_node(pRExC_state, GPOS);
12598 RExC_seen |= REG_GPOS_SEEN;
12600 goto finish_meta_pat;
12602 RExC_seen_zerolen++;
12603 ret = reg_node(pRExC_state, KEEPS);
12605 /* XXX:dmq : disabling in-place substitution seems to
12606 * be necessary here to avoid cases of memory corruption, as
12607 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
12609 RExC_seen |= REG_LOOKBEHIND_SEEN;
12610 goto finish_meta_pat;
12612 ret = reg_node(pRExC_state, SEOL);
12614 RExC_seen_zerolen++; /* Do not optimize RE away */
12615 goto finish_meta_pat;
12617 ret = reg_node(pRExC_state, EOS);
12619 RExC_seen_zerolen++; /* Do not optimize RE away */
12620 goto finish_meta_pat;
12622 vFAIL("\\C no longer supported");
12624 ret = reg_node(pRExC_state, CLUMP);
12625 *flagp |= HASWIDTH;
12626 goto finish_meta_pat;
12632 arg = ANYOF_WORDCHAR;
12640 regex_charset charset = get_regex_charset(RExC_flags);
12642 RExC_seen_zerolen++;
12643 RExC_seen |= REG_LOOKBEHIND_SEEN;
12644 op = BOUND + charset;
12646 if (op == BOUNDL) {
12647 RExC_contains_locale = 1;
12650 ret = reg_node(pRExC_state, op);
12652 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
12653 FLAGS(ret) = TRADITIONAL_BOUND;
12654 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
12660 char name = *RExC_parse;
12663 endbrace = strchr(RExC_parse, '}');
12666 vFAIL2("Missing right brace on \\%c{}", name);
12668 /* XXX Need to decide whether to take spaces or not. Should be
12669 * consistent with \p{}, but that currently is SPACE, which
12670 * means vertical too, which seems wrong
12671 * while (isBLANK(*RExC_parse)) {
12674 if (endbrace == RExC_parse) {
12675 RExC_parse++; /* After the '}' */
12676 vFAIL2("Empty \\%c{}", name);
12678 length = endbrace - RExC_parse;
12679 /*while (isBLANK(*(RExC_parse + length - 1))) {
12682 switch (*RExC_parse) {
12685 && (length != 3 || strnNE(RExC_parse + 1, "cb", 2)))
12687 goto bad_bound_type;
12689 FLAGS(ret) = GCB_BOUND;
12692 if (length != 2 || *(RExC_parse + 1) != 'b') {
12693 goto bad_bound_type;
12695 FLAGS(ret) = LB_BOUND;
12698 if (length != 2 || *(RExC_parse + 1) != 'b') {
12699 goto bad_bound_type;
12701 FLAGS(ret) = SB_BOUND;
12704 if (length != 2 || *(RExC_parse + 1) != 'b') {
12705 goto bad_bound_type;
12707 FLAGS(ret) = WB_BOUND;
12711 RExC_parse = endbrace;
12713 "'%" UTF8f "' is an unknown bound type",
12714 UTF8fARG(UTF, length, endbrace - length));
12715 NOT_REACHED; /*NOTREACHED*/
12717 RExC_parse = endbrace;
12718 REQUIRE_UNI_RULES(flagp, NULL);
12720 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
12724 /* Don't have to worry about UTF-8, in this message because
12725 * to get here the contents of the \b must be ASCII */
12726 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
12727 "Using /u for '%.*s' instead of /%s",
12729 endbrace - length + 1,
12730 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
12731 ? ASCII_RESTRICT_PAT_MODS
12732 : ASCII_MORE_RESTRICT_PAT_MODS);
12736 if (PASS2 && invert) {
12737 OP(ret) += NBOUND - BOUND;
12739 goto finish_meta_pat;
12747 if (! DEPENDS_SEMANTICS) {
12751 /* \d doesn't have any matches in the upper Latin1 range, hence /d
12752 * is equivalent to /u. Changing to /u saves some branches at
12755 goto join_posix_op_known;
12758 ret = reg_node(pRExC_state, LNBREAK);
12759 *flagp |= HASWIDTH|SIMPLE;
12760 goto finish_meta_pat;
12768 goto join_posix_op_known;
12774 arg = ANYOF_VERTWS;
12776 goto join_posix_op_known;
12786 op = POSIXD + get_regex_charset(RExC_flags);
12787 if (op > POSIXA) { /* /aa is same as /a */
12790 else if (op == POSIXL) {
12791 RExC_contains_locale = 1;
12794 join_posix_op_known:
12797 op += NPOSIXD - POSIXD;
12800 ret = reg_node(pRExC_state, op);
12802 FLAGS(ret) = namedclass_to_classnum(arg);
12805 *flagp |= HASWIDTH|SIMPLE;
12809 nextchar(pRExC_state);
12810 Set_Node_Length(ret, 2); /* MJD */
12816 ret = regclass(pRExC_state, flagp,depth+1,
12817 TRUE, /* means just parse this element */
12818 FALSE, /* don't allow multi-char folds */
12819 FALSE, /* don't silence non-portable warnings. It
12820 would be a bug if these returned
12822 (bool) RExC_strict,
12823 TRUE, /* Allow an optimized regnode result */
12826 if (*flagp & RESTART_PASS1)
12828 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
12829 * multi-char folds are allowed. */
12831 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12836 Set_Node_Offset(ret, parse_start);
12837 Set_Node_Cur_Length(ret, parse_start - 2);
12838 nextchar(pRExC_state);
12841 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
12842 * \N{...} evaluates to a sequence of more than one code points).
12843 * The function call below returns a regnode, which is our result.
12844 * The parameters cause it to fail if the \N{} evaluates to a
12845 * single code point; we handle those like any other literal. The
12846 * reason that the multicharacter case is handled here and not as
12847 * part of the EXACtish code is because of quantifiers. In
12848 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
12849 * this way makes that Just Happen. dmq.
12850 * join_exact() will join this up with adjacent EXACTish nodes
12851 * later on, if appropriate. */
12853 if (grok_bslash_N(pRExC_state,
12854 &ret, /* Want a regnode returned */
12855 NULL, /* Fail if evaluates to a single code
12857 NULL, /* Don't need a count of how many code
12866 if (*flagp & RESTART_PASS1)
12869 /* Here, evaluates to a single code point. Go get that */
12870 RExC_parse = parse_start;
12873 case 'k': /* Handle \k<NAME> and \k'NAME' */
12877 if ( RExC_parse >= RExC_end - 1
12878 || (( ch = RExC_parse[1]) != '<'
12883 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12884 vFAIL2("Sequence %.2s... not terminated",parse_start);
12887 ret = handle_named_backref(pRExC_state,
12899 case '1': case '2': case '3': case '4':
12900 case '5': case '6': case '7': case '8': case '9':
12905 if (*RExC_parse == 'g') {
12909 if (*RExC_parse == '{') {
12913 if (*RExC_parse == '-') {
12917 if (hasbrace && !isDIGIT(*RExC_parse)) {
12918 if (isrel) RExC_parse--;
12920 goto parse_named_seq;
12923 if (RExC_parse >= RExC_end) {
12924 goto unterminated_g;
12926 num = S_backref_value(RExC_parse);
12928 vFAIL("Reference to invalid group 0");
12929 else if (num == I32_MAX) {
12930 if (isDIGIT(*RExC_parse))
12931 vFAIL("Reference to nonexistent group");
12934 vFAIL("Unterminated \\g... pattern");
12938 num = RExC_npar - num;
12940 vFAIL("Reference to nonexistent or unclosed group");
12944 num = S_backref_value(RExC_parse);
12945 /* bare \NNN might be backref or octal - if it is larger
12946 * than or equal RExC_npar then it is assumed to be an
12947 * octal escape. Note RExC_npar is +1 from the actual
12948 * number of parens. */
12949 /* Note we do NOT check if num == I32_MAX here, as that is
12950 * handled by the RExC_npar check */
12953 /* any numeric escape < 10 is always a backref */
12955 /* any numeric escape < RExC_npar is a backref */
12956 && num >= RExC_npar
12957 /* cannot be an octal escape if it starts with 8 */
12958 && *RExC_parse != '8'
12959 /* cannot be an octal escape it it starts with 9 */
12960 && *RExC_parse != '9'
12963 /* Probably not a backref, instead likely to be an
12964 * octal character escape, e.g. \35 or \777.
12965 * The above logic should make it obvious why using
12966 * octal escapes in patterns is problematic. - Yves */
12967 RExC_parse = parse_start;
12972 /* At this point RExC_parse points at a numeric escape like
12973 * \12 or \88 or something similar, which we should NOT treat
12974 * as an octal escape. It may or may not be a valid backref
12975 * escape. For instance \88888888 is unlikely to be a valid
12977 while (isDIGIT(*RExC_parse))
12980 if (*RExC_parse != '}')
12981 vFAIL("Unterminated \\g{...} pattern");
12985 if (num > (I32)RExC_rx->nparens)
12986 vFAIL("Reference to nonexistent group");
12989 ret = reganode(pRExC_state,
12992 : (ASCII_FOLD_RESTRICTED)
12994 : (AT_LEAST_UNI_SEMANTICS)
13000 *flagp |= HASWIDTH;
13002 /* override incorrect value set in reganode MJD */
13003 Set_Node_Offset(ret, parse_start);
13004 Set_Node_Cur_Length(ret, parse_start-1);
13005 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13006 FALSE /* Don't force to /x */ );
13010 if (RExC_parse >= RExC_end)
13011 FAIL("Trailing \\");
13014 /* Do not generate "unrecognized" warnings here, we fall
13015 back into the quick-grab loop below */
13016 RExC_parse = parse_start;
13018 } /* end of switch on a \foo sequence */
13023 /* '#' comments should have been spaced over before this function was
13025 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13027 if (RExC_flags & RXf_PMf_EXTENDED) {
13028 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13029 if (RExC_parse < RExC_end)
13039 /* Here, we have determined that the next thing is probably a
13040 * literal character. RExC_parse points to the first byte of its
13041 * definition. (It still may be an escape sequence that evaluates
13042 * to a single character) */
13048 #define MAX_NODE_STRING_SIZE 127
13049 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
13051 U8 upper_parse = MAX_NODE_STRING_SIZE;
13052 U8 node_type = compute_EXACTish(pRExC_state);
13053 bool next_is_quantifier;
13054 char * oldp = NULL;
13056 /* We can convert EXACTF nodes to EXACTFU if they contain only
13057 * characters that match identically regardless of the target
13058 * string's UTF8ness. The reason to do this is that EXACTF is not
13059 * trie-able, EXACTFU is.
13061 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13062 * contain only above-Latin1 characters (hence must be in UTF8),
13063 * which don't participate in folds with Latin1-range characters,
13064 * as the latter's folds aren't known until runtime. (We don't
13065 * need to figure this out until pass 2) */
13066 bool maybe_exactfu = PASS2
13067 && (node_type == EXACTF || node_type == EXACTFL);
13069 /* If a folding node contains only code points that don't
13070 * participate in folds, it can be changed into an EXACT node,
13071 * which allows the optimizer more things to look for */
13074 ret = reg_node(pRExC_state, node_type);
13076 /* In pass1, folded, we use a temporary buffer instead of the
13077 * actual node, as the node doesn't exist yet */
13078 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
13084 /* We look for the EXACTFish to EXACT node optimizaton only if
13085 * folding. (And we don't need to figure this out until pass 2).
13086 * XXX It might actually make sense to split the node into portions
13087 * that are exact and ones that aren't, so that we could later use
13088 * the exact ones to find the longest fixed and floating strings.
13089 * One would want to join them back into a larger node. One could
13090 * use a pseudo regnode like 'EXACT_ORIG_FOLD' */
13091 maybe_exact = FOLD && PASS2;
13093 /* XXX The node can hold up to 255 bytes, yet this only goes to
13094 * 127. I (khw) do not know why. Keeping it somewhat less than
13095 * 255 allows us to not have to worry about overflow due to
13096 * converting to utf8 and fold expansion, but that value is
13097 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
13098 * split up by this limit into a single one using the real max of
13099 * 255. Even at 127, this breaks under rare circumstances. If
13100 * folding, we do not want to split a node at a character that is a
13101 * non-final in a multi-char fold, as an input string could just
13102 * happen to want to match across the node boundary. The join
13103 * would solve that problem if the join actually happens. But a
13104 * series of more than two nodes in a row each of 127 would cause
13105 * the first join to succeed to get to 254, but then there wouldn't
13106 * be room for the next one, which could at be one of those split
13107 * multi-char folds. I don't know of any fool-proof solution. One
13108 * could back off to end with only a code point that isn't such a
13109 * non-final, but it is possible for there not to be any in the
13112 assert( ! UTF /* Is at the beginning of a character */
13113 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13114 || UTF8_IS_START(UCHARAT(RExC_parse)));
13116 /* Here, we have a literal character. Find the maximal string of
13117 * them in the input that we can fit into a single EXACTish node.
13118 * We quit at the first non-literal or when the node gets full */
13119 for (p = RExC_parse;
13120 len < upper_parse && p < RExC_end;
13125 /* White space has already been ignored */
13126 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13127 || ! is_PATWS_safe((p), RExC_end, UTF));
13139 /* Literal Escapes Switch
13141 This switch is meant to handle escape sequences that
13142 resolve to a literal character.
13144 Every escape sequence that represents something
13145 else, like an assertion or a char class, is handled
13146 in the switch marked 'Special Escapes' above in this
13147 routine, but also has an entry here as anything that
13148 isn't explicitly mentioned here will be treated as
13149 an unescaped equivalent literal.
13152 switch ((U8)*++p) {
13153 /* These are all the special escapes. */
13154 case 'A': /* Start assertion */
13155 case 'b': case 'B': /* Word-boundary assertion*/
13156 case 'C': /* Single char !DANGEROUS! */
13157 case 'd': case 'D': /* digit class */
13158 case 'g': case 'G': /* generic-backref, pos assertion */
13159 case 'h': case 'H': /* HORIZWS */
13160 case 'k': case 'K': /* named backref, keep marker */
13161 case 'p': case 'P': /* Unicode property */
13162 case 'R': /* LNBREAK */
13163 case 's': case 'S': /* space class */
13164 case 'v': case 'V': /* VERTWS */
13165 case 'w': case 'W': /* word class */
13166 case 'X': /* eXtended Unicode "combining
13167 character sequence" */
13168 case 'z': case 'Z': /* End of line/string assertion */
13172 /* Anything after here is an escape that resolves to a
13173 literal. (Except digits, which may or may not)
13179 case 'N': /* Handle a single-code point named character. */
13180 RExC_parse = p + 1;
13181 if (! grok_bslash_N(pRExC_state,
13182 NULL, /* Fail if evaluates to
13183 anything other than a
13184 single code point */
13185 &ender, /* The returned single code
13187 NULL, /* Don't need a count of
13188 how many code points */
13193 if (*flagp & NEED_UTF8)
13194 FAIL("panic: grok_bslash_N set NEED_UTF8");
13195 if (*flagp & RESTART_PASS1)
13198 /* Here, it wasn't a single code point. Go close
13199 * up this EXACTish node. The switch() prior to
13200 * this switch handles the other cases */
13201 RExC_parse = p = oldp;
13205 if (ender > 0xff) {
13206 REQUIRE_UTF8(flagp);
13222 ender = ESC_NATIVE;
13232 const char* error_msg;
13234 bool valid = grok_bslash_o(&p,
13237 PASS2, /* out warnings */
13238 (bool) RExC_strict,
13239 TRUE, /* Output warnings
13244 RExC_parse = p; /* going to die anyway; point
13245 to exact spot of failure */
13249 if (ender > 0xff) {
13250 REQUIRE_UTF8(flagp);
13256 UV result = UV_MAX; /* initialize to erroneous
13258 const char* error_msg;
13260 bool valid = grok_bslash_x(&p,
13263 PASS2, /* out warnings */
13264 (bool) RExC_strict,
13265 TRUE, /* Silence warnings
13270 RExC_parse = p; /* going to die anyway; point
13271 to exact spot of failure */
13276 if (ender < 0x100) {
13278 if (RExC_recode_x_to_native) {
13279 ender = LATIN1_TO_NATIVE(ender);
13284 REQUIRE_UTF8(flagp);
13290 ender = grok_bslash_c(*p++, PASS2);
13292 case '8': case '9': /* must be a backreference */
13294 /* we have an escape like \8 which cannot be an octal escape
13295 * so we exit the loop, and let the outer loop handle this
13296 * escape which may or may not be a legitimate backref. */
13298 case '1': case '2': case '3':case '4':
13299 case '5': case '6': case '7':
13300 /* When we parse backslash escapes there is ambiguity
13301 * between backreferences and octal escapes. Any escape
13302 * from \1 - \9 is a backreference, any multi-digit
13303 * escape which does not start with 0 and which when
13304 * evaluated as decimal could refer to an already
13305 * parsed capture buffer is a back reference. Anything
13308 * Note this implies that \118 could be interpreted as
13309 * 118 OR as "\11" . "8" depending on whether there
13310 * were 118 capture buffers defined already in the
13313 /* NOTE, RExC_npar is 1 more than the actual number of
13314 * parens we have seen so far, hence the < RExC_npar below. */
13316 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
13317 { /* Not to be treated as an octal constant, go
13325 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13327 ender = grok_oct(p, &numlen, &flags, NULL);
13328 if (ender > 0xff) {
13329 REQUIRE_UTF8(flagp);
13332 if (PASS2 /* like \08, \178 */
13334 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
13336 reg_warn_non_literal_string(
13338 form_short_octal_warning(p, numlen));
13344 FAIL("Trailing \\");
13347 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
13348 /* Include any left brace following the alpha to emphasize
13349 * that it could be part of an escape at some point
13351 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
13352 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
13354 goto normal_default;
13355 } /* End of switch on '\' */
13358 /* Currently we don't care if the lbrace is at the start
13359 * of a construct. This catches it in the middle of a
13360 * literal string, or when it's the first thing after
13361 * something like "\b" */
13362 if (len || (p > RExC_start && isALPHA_A(*(p -1)))) {
13363 RExC_parse = p + 1;
13364 vFAIL("Unescaped left brace in regex is illegal here");
13366 goto normal_default;
13369 if (PASS2 && p > RExC_parse && RExC_strict) {
13370 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
13373 default: /* A literal character */
13375 if (! UTF8_IS_INVARIANT(*p) && UTF) {
13377 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
13378 &numlen, UTF8_ALLOW_DEFAULT);
13384 } /* End of switch on the literal */
13386 /* Here, have looked at the literal character and <ender>
13387 * contains its ordinal, <p> points to the character after it.
13388 * We need to check if the next non-ignored thing is a
13389 * quantifier. Move <p> to after anything that should be
13390 * ignored, which, as a side effect, positions <p> for the next
13391 * loop iteration */
13392 skip_to_be_ignored_text(pRExC_state, &p,
13393 FALSE /* Don't force to /x */ );
13395 /* If the next thing is a quantifier, it applies to this
13396 * character only, which means that this character has to be in
13397 * its own node and can't just be appended to the string in an
13398 * existing node, so if there are already other characters in
13399 * the node, close the node with just them, and set up to do
13400 * this character again next time through, when it will be the
13401 * only thing in its new node */
13403 if ((next_is_quantifier = ( LIKELY(p < RExC_end)
13404 && UNLIKELY(ISMULT2(p))))
13411 /* Ready to add 'ender' to the node */
13413 if (! FOLD) { /* The simple case, just append the literal */
13415 /* In the sizing pass, we need only the size of the
13416 * character we are appending, hence we can delay getting
13417 * its representation until PASS2. */
13420 const STRLEN unilen = UVCHR_SKIP(ender);
13423 /* We have to subtract 1 just below (and again in
13424 * the corresponding PASS2 code) because the loop
13425 * increments <len> each time, as all but this path
13426 * (and one other) through it add a single byte to
13427 * the EXACTish node. But these paths would change
13428 * len to be the correct final value, so cancel out
13429 * the increment that follows */
13435 } else { /* PASS2 */
13438 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
13439 len += (char *) new_s - s - 1;
13440 s = (char *) new_s;
13443 *(s++) = (char) ender;
13447 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
13449 /* Here are folding under /l, and the code point is
13450 * problematic. First, we know we can't simplify things */
13451 maybe_exact = FALSE;
13452 maybe_exactfu = FALSE;
13454 /* A problematic code point in this context means that its
13455 * fold isn't known until runtime, so we can't fold it now.
13456 * (The non-problematic code points are the above-Latin1
13457 * ones that fold to also all above-Latin1. Their folds
13458 * don't vary no matter what the locale is.) But here we
13459 * have characters whose fold depends on the locale.
13460 * Unlike the non-folding case above, we have to keep track
13461 * of these in the sizing pass, so that we can make sure we
13462 * don't split too-long nodes in the middle of a potential
13463 * multi-char fold. And unlike the regular fold case
13464 * handled in the else clauses below, we don't actually
13465 * fold and don't have special cases to consider. What we
13466 * do for both passes is the PASS2 code for non-folding */
13467 goto not_fold_common;
13469 else /* A regular FOLD code point */
13471 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13472 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13473 || UNICODE_DOT_DOT_VERSION > 0)
13474 /* See comments for join_exact() as to why we fold
13475 * this non-UTF at compile time */
13476 || ( node_type == EXACTFU
13477 && ender == LATIN_SMALL_LETTER_SHARP_S)
13480 /* Here, are folding and are not UTF-8 encoded; therefore
13481 * the character must be in the range 0-255, and is not /l
13482 * (Not /l because we already handled these under /l in
13483 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
13484 if (IS_IN_SOME_FOLD_L1(ender)) {
13485 maybe_exact = FALSE;
13487 /* See if the character's fold differs between /d and
13488 * /u. This includes the multi-char fold SHARP S to
13490 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
13491 RExC_seen_unfolded_sharp_s = 1;
13492 maybe_exactfu = FALSE;
13494 else if (maybe_exactfu
13495 && (PL_fold[ender] != PL_fold_latin1[ender]
13496 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13497 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13498 || UNICODE_DOT_DOT_VERSION > 0)
13500 && isALPHA_FOLD_EQ(ender, 's')
13501 && isALPHA_FOLD_EQ(*(s-1), 's'))
13504 maybe_exactfu = FALSE;
13508 /* Even when folding, we store just the input character, as
13509 * we have an array that finds its fold quickly */
13510 *(s++) = (char) ender;
13512 else { /* FOLD, and UTF (or sharp s) */
13513 /* Unlike the non-fold case, we do actually have to
13514 * calculate the results here in pass 1. This is for two
13515 * reasons, the folded length may be longer than the
13516 * unfolded, and we have to calculate how many EXACTish
13517 * nodes it will take; and we may run out of room in a node
13518 * in the middle of a potential multi-char fold, and have
13519 * to back off accordingly. */
13522 if (isASCII_uni(ender)) {
13523 folded = toFOLD(ender);
13524 *(s)++ = (U8) folded;
13529 folded = _to_uni_fold_flags(
13533 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
13534 ? FOLD_FLAGS_NOMIX_ASCII
13538 /* The loop increments <len> each time, as all but this
13539 * path (and one other) through it add a single byte to
13540 * the EXACTish node. But this one has changed len to
13541 * be the correct final value, so subtract one to
13542 * cancel out the increment that follows */
13543 len += foldlen - 1;
13545 /* If this node only contains non-folding code points so
13546 * far, see if this new one is also non-folding */
13548 if (folded != ender) {
13549 maybe_exact = FALSE;
13552 /* Here the fold is the original; we have to check
13553 * further to see if anything folds to it */
13554 if (_invlist_contains_cp(PL_utf8_foldable,
13557 maybe_exact = FALSE;
13564 if (next_is_quantifier) {
13566 /* Here, the next input is a quantifier, and to get here,
13567 * the current character is the only one in the node.
13568 * Also, here <len> doesn't include the final byte for this
13574 } /* End of loop through literal characters */
13576 /* Here we have either exhausted the input or ran out of room in
13577 * the node. (If we encountered a character that can't be in the
13578 * node, transfer is made directly to <loopdone>, and so we
13579 * wouldn't have fallen off the end of the loop.) In the latter
13580 * case, we artificially have to split the node into two, because
13581 * we just don't have enough space to hold everything. This
13582 * creates a problem if the final character participates in a
13583 * multi-character fold in the non-final position, as a match that
13584 * should have occurred won't, due to the way nodes are matched,
13585 * and our artificial boundary. So back off until we find a non-
13586 * problematic character -- one that isn't at the beginning or
13587 * middle of such a fold. (Either it doesn't participate in any
13588 * folds, or appears only in the final position of all the folds it
13589 * does participate in.) A better solution with far fewer false
13590 * positives, and that would fill the nodes more completely, would
13591 * be to actually have available all the multi-character folds to
13592 * test against, and to back-off only far enough to be sure that
13593 * this node isn't ending with a partial one. <upper_parse> is set
13594 * further below (if we need to reparse the node) to include just
13595 * up through that final non-problematic character that this code
13596 * identifies, so when it is set to less than the full node, we can
13597 * skip the rest of this */
13598 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
13600 const STRLEN full_len = len;
13602 assert(len >= MAX_NODE_STRING_SIZE);
13604 /* Here, <s> points to the final byte of the final character.
13605 * Look backwards through the string until find a non-
13606 * problematic character */
13610 /* This has no multi-char folds to non-UTF characters */
13611 if (ASCII_FOLD_RESTRICTED) {
13615 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
13619 if (! PL_NonL1NonFinalFold) {
13620 PL_NonL1NonFinalFold = _new_invlist_C_array(
13621 NonL1_Perl_Non_Final_Folds_invlist);
13624 /* Point to the first byte of the final character */
13625 s = (char *) utf8_hop((U8 *) s, -1);
13627 while (s >= s0) { /* Search backwards until find
13628 non-problematic char */
13629 if (UTF8_IS_INVARIANT(*s)) {
13631 /* There are no ascii characters that participate
13632 * in multi-char folds under /aa. In EBCDIC, the
13633 * non-ascii invariants are all control characters,
13634 * so don't ever participate in any folds. */
13635 if (ASCII_FOLD_RESTRICTED
13636 || ! IS_NON_FINAL_FOLD(*s))
13641 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
13642 if (! IS_NON_FINAL_FOLD(EIGHT_BIT_UTF8_TO_NATIVE(
13648 else if (! _invlist_contains_cp(
13649 PL_NonL1NonFinalFold,
13650 valid_utf8_to_uvchr((U8 *) s, NULL)))
13655 /* Here, the current character is problematic in that
13656 * it does occur in the non-final position of some
13657 * fold, so try the character before it, but have to
13658 * special case the very first byte in the string, so
13659 * we don't read outside the string */
13660 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
13661 } /* End of loop backwards through the string */
13663 /* If there were only problematic characters in the string,
13664 * <s> will point to before s0, in which case the length
13665 * should be 0, otherwise include the length of the
13666 * non-problematic character just found */
13667 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
13670 /* Here, have found the final character, if any, that is
13671 * non-problematic as far as ending the node without splitting
13672 * it across a potential multi-char fold. <len> contains the
13673 * number of bytes in the node up-to and including that
13674 * character, or is 0 if there is no such character, meaning
13675 * the whole node contains only problematic characters. In
13676 * this case, give up and just take the node as-is. We can't
13681 /* If the node ends in an 's' we make sure it stays EXACTF,
13682 * as if it turns into an EXACTFU, it could later get
13683 * joined with another 's' that would then wrongly match
13685 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
13687 maybe_exactfu = FALSE;
13691 /* Here, the node does contain some characters that aren't
13692 * problematic. If one such is the final character in the
13693 * node, we are done */
13694 if (len == full_len) {
13697 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
13699 /* If the final character is problematic, but the
13700 * penultimate is not, back-off that last character to
13701 * later start a new node with it */
13706 /* Here, the final non-problematic character is earlier
13707 * in the input than the penultimate character. What we do
13708 * is reparse from the beginning, going up only as far as
13709 * this final ok one, thus guaranteeing that the node ends
13710 * in an acceptable character. The reason we reparse is
13711 * that we know how far in the character is, but we don't
13712 * know how to correlate its position with the input parse.
13713 * An alternate implementation would be to build that
13714 * correlation as we go along during the original parse,
13715 * but that would entail extra work for every node, whereas
13716 * this code gets executed only when the string is too
13717 * large for the node, and the final two characters are
13718 * problematic, an infrequent occurrence. Yet another
13719 * possible strategy would be to save the tail of the
13720 * string, and the next time regatom is called, initialize
13721 * with that. The problem with this is that unless you
13722 * back off one more character, you won't be guaranteed
13723 * regatom will get called again, unless regbranch,
13724 * regpiece ... are also changed. If you do back off that
13725 * extra character, so that there is input guaranteed to
13726 * force calling regatom, you can't handle the case where
13727 * just the first character in the node is acceptable. I
13728 * (khw) decided to try this method which doesn't have that
13729 * pitfall; if performance issues are found, we can do a
13730 * combination of the current approach plus that one */
13736 } /* End of verifying node ends with an appropriate char */
13738 loopdone: /* Jumped to when encounters something that shouldn't be
13741 /* I (khw) don't know if you can get here with zero length, but the
13742 * old code handled this situation by creating a zero-length EXACT
13743 * node. Might as well be NOTHING instead */
13749 /* If 'maybe_exact' is still set here, means there are no
13750 * code points in the node that participate in folds;
13751 * similarly for 'maybe_exactfu' and code points that match
13752 * differently depending on UTF8ness of the target string
13753 * (for /u), or depending on locale for /l */
13759 else if (maybe_exactfu) {
13765 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
13766 FALSE /* Don't look to see if could
13767 be turned into an EXACT
13768 node, as we have already
13773 RExC_parse = p - 1;
13774 Set_Node_Cur_Length(ret, parse_start);
13777 /* len is STRLEN which is unsigned, need to copy to signed */
13780 vFAIL("Internal disaster");
13783 } /* End of label 'defchar:' */
13785 } /* End of giant switch on input character */
13787 /* Position parse to next real character */
13788 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13789 FALSE /* Don't force to /x */ );
13790 if (PASS2 && *RExC_parse == '{' && OP(ret) != SBOL && ! regcurly(RExC_parse)) {
13791 ckWARNregdep(RExC_parse + 1, "Unescaped left brace in regex is deprecated here, passed through");
13799 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
13801 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
13802 * sets up the bitmap and any flags, removing those code points from the
13803 * inversion list, setting it to NULL should it become completely empty */
13805 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
13806 assert(PL_regkind[OP(node)] == ANYOF);
13808 ANYOF_BITMAP_ZERO(node);
13809 if (*invlist_ptr) {
13811 /* This gets set if we actually need to modify things */
13812 bool change_invlist = FALSE;
13816 /* Start looking through *invlist_ptr */
13817 invlist_iterinit(*invlist_ptr);
13818 while (invlist_iternext(*invlist_ptr, &start, &end)) {
13822 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
13823 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
13826 /* Quit if are above what we should change */
13827 if (start >= NUM_ANYOF_CODE_POINTS) {
13831 change_invlist = TRUE;
13833 /* Set all the bits in the range, up to the max that we are doing */
13834 high = (end < NUM_ANYOF_CODE_POINTS - 1)
13836 : NUM_ANYOF_CODE_POINTS - 1;
13837 for (i = start; i <= (int) high; i++) {
13838 if (! ANYOF_BITMAP_TEST(node, i)) {
13839 ANYOF_BITMAP_SET(node, i);
13843 invlist_iterfinish(*invlist_ptr);
13845 /* Done with loop; remove any code points that are in the bitmap from
13846 * *invlist_ptr; similarly for code points above the bitmap if we have
13847 * a flag to match all of them anyways */
13848 if (change_invlist) {
13849 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
13851 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
13852 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
13855 /* If have completely emptied it, remove it completely */
13856 if (_invlist_len(*invlist_ptr) == 0) {
13857 SvREFCNT_dec_NN(*invlist_ptr);
13858 *invlist_ptr = NULL;
13863 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
13864 Character classes ([:foo:]) can also be negated ([:^foo:]).
13865 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
13866 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
13867 but trigger failures because they are currently unimplemented. */
13869 #define POSIXCC_DONE(c) ((c) == ':')
13870 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
13871 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
13872 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
13874 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
13875 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
13876 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
13878 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
13880 /* 'posix_warnings' and 'warn_text' are names of variables in the following
13882 #define ADD_POSIX_WARNING(p, text) STMT_START { \
13883 if (posix_warnings) { \
13884 if (! RExC_warn_text ) RExC_warn_text = (AV *) sv_2mortal((SV *) newAV()); \
13885 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
13889 REPORT_LOCATION_ARGS(p))); \
13894 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
13896 const char * const s, /* Where the putative posix class begins.
13897 Normally, this is one past the '['. This
13898 parameter exists so it can be somewhere
13899 besides RExC_parse. */
13900 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
13902 AV ** posix_warnings, /* Where to place any generated warnings, or
13904 const bool check_only /* Don't die if error */
13907 /* This parses what the caller thinks may be one of the three POSIX
13909 * 1) a character class, like [:blank:]
13910 * 2) a collating symbol, like [. .]
13911 * 3) an equivalence class, like [= =]
13912 * In the latter two cases, it croaks if it finds a syntactically legal
13913 * one, as these are not handled by Perl.
13915 * The main purpose is to look for a POSIX character class. It returns:
13916 * a) the class number
13917 * if it is a completely syntactically and semantically legal class.
13918 * 'updated_parse_ptr', if not NULL, is set to point to just after the
13919 * closing ']' of the class
13920 * b) OOB_NAMEDCLASS
13921 * if it appears that one of the three POSIX constructs was meant, but
13922 * its specification was somehow defective. 'updated_parse_ptr', if
13923 * not NULL, is set to point to the character just after the end
13924 * character of the class. See below for handling of warnings.
13925 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
13926 * if it doesn't appear that a POSIX construct was intended.
13927 * 'updated_parse_ptr' is not changed. No warnings nor errors are
13930 * In b) there may be errors or warnings generated. If 'check_only' is
13931 * TRUE, then any errors are discarded. Warnings are returned to the
13932 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
13933 * instead it is NULL, warnings are suppressed. This is done in all
13934 * passes. The reason for this is that the rest of the parsing is heavily
13935 * dependent on whether this routine found a valid posix class or not. If
13936 * it did, the closing ']' is absorbed as part of the class. If no class,
13937 * or an invalid one is found, any ']' will be considered the terminator of
13938 * the outer bracketed character class, leading to very different results.
13939 * In particular, a '(?[ ])' construct will likely have a syntax error if
13940 * the class is parsed other than intended, and this will happen in pass1,
13941 * before the warnings would normally be output. This mechanism allows the
13942 * caller to output those warnings in pass1 just before dieing, giving a
13943 * much better clue as to what is wrong.
13945 * The reason for this function, and its complexity is that a bracketed
13946 * character class can contain just about anything. But it's easy to
13947 * mistype the very specific posix class syntax but yielding a valid
13948 * regular bracketed class, so it silently gets compiled into something
13949 * quite unintended.
13951 * The solution adopted here maintains backward compatibility except that
13952 * it adds a warning if it looks like a posix class was intended but
13953 * improperly specified. The warning is not raised unless what is input
13954 * very closely resembles one of the 14 legal posix classes. To do this,
13955 * it uses fuzzy parsing. It calculates how many single-character edits it
13956 * would take to transform what was input into a legal posix class. Only
13957 * if that number is quite small does it think that the intention was a
13958 * posix class. Obviously these are heuristics, and there will be cases
13959 * where it errs on one side or another, and they can be tweaked as
13960 * experience informs.
13962 * The syntax for a legal posix class is:
13964 * qr/(?xa: \[ : \^? [:lower:]{4,6} : \] )/
13966 * What this routine considers syntactically to be an intended posix class
13967 * is this (the comments indicate some restrictions that the pattern
13970 * qr/(?x: \[? # The left bracket, possibly
13972 * \h* # possibly followed by blanks
13973 * (?: \^ \h* )? # possibly a misplaced caret
13974 * [:;]? # The opening class character,
13975 * # possibly omitted. A typo
13976 * # semi-colon can also be used.
13978 * \^? # possibly a correctly placed
13979 * # caret, but not if there was also
13980 * # a misplaced one
13982 * .{3,15} # The class name. If there are
13983 * # deviations from the legal syntax,
13984 * # its edit distance must be close
13985 * # to a real class name in order
13986 * # for it to be considered to be
13987 * # an intended posix class.
13989 * [:punct:]? # The closing class character,
13990 * # possibly omitted. If not a colon
13991 * # nor semi colon, the class name
13992 * # must be even closer to a valid
13995 * \]? # The right bracket, possibly
13999 * In the above, \h must be ASCII-only.
14001 * These are heuristics, and can be tweaked as field experience dictates.
14002 * There will be cases when someone didn't intend to specify a posix class
14003 * that this warns as being so. The goal is to minimize these, while
14004 * maximizing the catching of things intended to be a posix class that
14005 * aren't parsed as such.
14009 const char * const e = RExC_end;
14010 unsigned complement = 0; /* If to complement the class */
14011 bool found_problem = FALSE; /* Assume OK until proven otherwise */
14012 bool has_opening_bracket = FALSE;
14013 bool has_opening_colon = FALSE;
14014 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
14016 const char * possible_end = NULL; /* used for a 2nd parse pass */
14017 const char* name_start; /* ptr to class name first char */
14019 /* If the number of single-character typos the input name is away from a
14020 * legal name is no more than this number, it is considered to have meant
14021 * the legal name */
14022 int max_distance = 2;
14024 /* to store the name. The size determines the maximum length before we
14025 * decide that no posix class was intended. Should be at least
14026 * sizeof("alphanumeric") */
14029 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
14031 if (posix_warnings && RExC_warn_text)
14032 av_clear(RExC_warn_text);
14035 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14038 if (*(p - 1) != '[') {
14039 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
14040 found_problem = TRUE;
14043 has_opening_bracket = TRUE;
14046 /* They could be confused and think you can put spaces between the
14049 found_problem = TRUE;
14053 } while (p < e && isBLANK(*p));
14055 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14058 /* For [. .] and [= =]. These are quite different internally from [: :],
14059 * so they are handled separately. */
14060 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
14061 and 1 for at least one char in it
14064 const char open_char = *p;
14065 const char * temp_ptr = p + 1;
14067 /* These two constructs are not handled by perl, and if we find a
14068 * syntactically valid one, we croak. khw, who wrote this code, finds
14069 * this explanation of them very unclear:
14070 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
14071 * And searching the rest of the internet wasn't very helpful either.
14072 * It looks like just about any byte can be in these constructs,
14073 * depending on the locale. But unless the pattern is being compiled
14074 * under /l, which is very rare, Perl runs under the C or POSIX locale.
14075 * In that case, it looks like [= =] isn't allowed at all, and that
14076 * [. .] could be any single code point, but for longer strings the
14077 * constituent characters would have to be the ASCII alphabetics plus
14078 * the minus-hyphen. Any sensible locale definition would limit itself
14079 * to these. And any portable one definitely should. Trying to parse
14080 * the general case is a nightmare (see [perl #127604]). So, this code
14081 * looks only for interiors of these constructs that match:
14083 * Using \w relaxes the apparent rules a little, without adding much
14084 * danger of mistaking something else for one of these constructs.
14086 * [. .] in some implementations described on the internet is usable to
14087 * escape a character that otherwise is special in bracketed character
14088 * classes. For example [.].] means a literal right bracket instead of
14089 * the ending of the class
14091 * [= =] can legitimately contain a [. .] construct, but we don't
14092 * handle this case, as that [. .] construct will later get parsed
14093 * itself and croak then. And [= =] is checked for even when not under
14094 * /l, as Perl has long done so.
14096 * The code below relies on there being a trailing NUL, so it doesn't
14097 * have to keep checking if the parse ptr < e.
14099 if (temp_ptr[1] == open_char) {
14102 else while ( temp_ptr < e
14103 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
14108 if (*temp_ptr == open_char) {
14110 if (*temp_ptr == ']') {
14112 if (! found_problem && ! check_only) {
14113 RExC_parse = (char *) temp_ptr;
14114 vFAIL3("POSIX syntax [%c %c] is reserved for future "
14115 "extensions", open_char, open_char);
14118 /* Here, the syntax wasn't completely valid, or else the call
14119 * is to check-only */
14120 if (updated_parse_ptr) {
14121 *updated_parse_ptr = (char *) temp_ptr;
14124 return OOB_NAMEDCLASS;
14128 /* If we find something that started out to look like one of these
14129 * constructs, but isn't, we continue below so that it can be checked
14130 * for being a class name with a typo of '.' or '=' instead of a colon.
14134 /* Here, we think there is a possibility that a [: :] class was meant, and
14135 * we have the first real character. It could be they think the '^' comes
14138 found_problem = TRUE;
14139 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
14144 found_problem = TRUE;
14148 } while (p < e && isBLANK(*p));
14150 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14154 /* But the first character should be a colon, which they could have easily
14155 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
14156 * distinguish from a colon, so treat that as a colon). */
14159 has_opening_colon = TRUE;
14161 else if (*p == ';') {
14162 found_problem = TRUE;
14164 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14165 has_opening_colon = TRUE;
14168 found_problem = TRUE;
14169 ADD_POSIX_WARNING(p, "there must be a starting ':'");
14171 /* Consider an initial punctuation (not one of the recognized ones) to
14172 * be a left terminator */
14173 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
14178 /* They may think that you can put spaces between the components */
14180 found_problem = TRUE;
14184 } while (p < e && isBLANK(*p));
14186 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14191 /* We consider something like [^:^alnum:]] to not have been intended to
14192 * be a posix class, but XXX maybe we should */
14194 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14201 /* Again, they may think that you can put spaces between the components */
14203 found_problem = TRUE;
14207 } while (p < e && isBLANK(*p));
14209 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14214 /* XXX This ']' may be a typo, and something else was meant. But
14215 * treating it as such creates enough complications, that that
14216 * possibility isn't currently considered here. So we assume that the
14217 * ']' is what is intended, and if we've already found an initial '[',
14218 * this leaves this construct looking like [:] or [:^], which almost
14219 * certainly weren't intended to be posix classes */
14220 if (has_opening_bracket) {
14221 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14224 /* But this function can be called when we parse the colon for
14225 * something like qr/[alpha:]]/, so we back up to look for the
14230 found_problem = TRUE;
14231 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14233 else if (*p != ':') {
14235 /* XXX We are currently very restrictive here, so this code doesn't
14236 * consider the possibility that, say, /[alpha.]]/ was intended to
14237 * be a posix class. */
14238 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14241 /* Here we have something like 'foo:]'. There was no initial colon,
14242 * and we back up over 'foo. XXX Unlike the going forward case, we
14243 * don't handle typos of non-word chars in the middle */
14244 has_opening_colon = FALSE;
14247 while (p > RExC_start && isWORDCHAR(*p)) {
14252 /* Here, we have positioned ourselves to where we think the first
14253 * character in the potential class is */
14256 /* Now the interior really starts. There are certain key characters that
14257 * can end the interior, or these could just be typos. To catch both
14258 * cases, we may have to do two passes. In the first pass, we keep on
14259 * going unless we come to a sequence that matches
14260 * qr/ [[:punct:]] [[:blank:]]* \] /xa
14261 * This means it takes a sequence to end the pass, so two typos in a row if
14262 * that wasn't what was intended. If the class is perfectly formed, just
14263 * this one pass is needed. We also stop if there are too many characters
14264 * being accumulated, but this number is deliberately set higher than any
14265 * real class. It is set high enough so that someone who thinks that
14266 * 'alphanumeric' is a correct name would get warned that it wasn't.
14267 * While doing the pass, we keep track of where the key characters were in
14268 * it. If we don't find an end to the class, and one of the key characters
14269 * was found, we redo the pass, but stop when we get to that character.
14270 * Thus the key character was considered a typo in the first pass, but a
14271 * terminator in the second. If two key characters are found, we stop at
14272 * the second one in the first pass. Again this can miss two typos, but
14273 * catches a single one
14275 * In the first pass, 'possible_end' starts as NULL, and then gets set to
14276 * point to the first key character. For the second pass, it starts as -1.
14282 bool has_blank = FALSE;
14283 bool has_upper = FALSE;
14284 bool has_terminating_colon = FALSE;
14285 bool has_terminating_bracket = FALSE;
14286 bool has_semi_colon = FALSE;
14287 unsigned int name_len = 0;
14288 int punct_count = 0;
14292 /* Squeeze out blanks when looking up the class name below */
14293 if (isBLANK(*p) ) {
14295 found_problem = TRUE;
14300 /* The name will end with a punctuation */
14302 const char * peek = p + 1;
14304 /* Treat any non-']' punctuation followed by a ']' (possibly
14305 * with intervening blanks) as trying to terminate the class.
14306 * ']]' is very likely to mean a class was intended (but
14307 * missing the colon), but the warning message that gets
14308 * generated shows the error position better if we exit the
14309 * loop at the bottom (eventually), so skip it here. */
14311 if (peek < e && isBLANK(*peek)) {
14313 found_problem = TRUE;
14316 } while (peek < e && isBLANK(*peek));
14319 if (peek < e && *peek == ']') {
14320 has_terminating_bracket = TRUE;
14322 has_terminating_colon = TRUE;
14324 else if (*p == ';') {
14325 has_semi_colon = TRUE;
14326 has_terminating_colon = TRUE;
14329 found_problem = TRUE;
14336 /* Here we have punctuation we thought didn't end the class.
14337 * Keep track of the position of the key characters that are
14338 * more likely to have been class-enders */
14339 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
14341 /* Allow just one such possible class-ender not actually
14342 * ending the class. */
14343 if (possible_end) {
14349 /* If we have too many punctuation characters, no use in
14351 if (++punct_count > max_distance) {
14355 /* Treat the punctuation as a typo. */
14356 input_text[name_len++] = *p;
14359 else if (isUPPER(*p)) { /* Use lowercase for lookup */
14360 input_text[name_len++] = toLOWER(*p);
14362 found_problem = TRUE;
14364 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
14365 input_text[name_len++] = *p;
14369 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
14373 /* The declaration of 'input_text' is how long we allow a potential
14374 * class name to be, before saying they didn't mean a class name at
14376 if (name_len >= C_ARRAY_LENGTH(input_text)) {
14381 /* We get to here when the possible class name hasn't been properly
14382 * terminated before:
14383 * 1) we ran off the end of the pattern; or
14384 * 2) found two characters, each of which might have been intended to
14385 * be the name's terminator
14386 * 3) found so many punctuation characters in the purported name,
14387 * that the edit distance to a valid one is exceeded
14388 * 4) we decided it was more characters than anyone could have
14389 * intended to be one. */
14391 found_problem = TRUE;
14393 /* In the final two cases, we know that looking up what we've
14394 * accumulated won't lead to a match, even a fuzzy one. */
14395 if ( name_len >= C_ARRAY_LENGTH(input_text)
14396 || punct_count > max_distance)
14398 /* If there was an intermediate key character that could have been
14399 * an intended end, redo the parse, but stop there */
14400 if (possible_end && possible_end != (char *) -1) {
14401 possible_end = (char *) -1; /* Special signal value to say
14402 we've done a first pass */
14407 /* Otherwise, it can't have meant to have been a class */
14408 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14411 /* If we ran off the end, and the final character was a punctuation
14412 * one, back up one, to look at that final one just below. Later, we
14413 * will restore the parse pointer if appropriate */
14414 if (name_len && p == e && isPUNCT(*(p-1))) {
14419 if (p < e && isPUNCT(*p)) {
14421 has_terminating_bracket = TRUE;
14423 /* If this is a 2nd ']', and the first one is just below this
14424 * one, consider that to be the real terminator. This gives a
14425 * uniform and better positioning for the warning message */
14427 && possible_end != (char *) -1
14428 && *possible_end == ']'
14429 && name_len && input_text[name_len - 1] == ']')
14434 /* And this is actually equivalent to having done the 2nd
14435 * pass now, so set it to not try again */
14436 possible_end = (char *) -1;
14441 has_terminating_colon = TRUE;
14443 else if (*p == ';') {
14444 has_semi_colon = TRUE;
14445 has_terminating_colon = TRUE;
14453 /* Here, we have a class name to look up. We can short circuit the
14454 * stuff below for short names that can't possibly be meant to be a
14455 * class name. (We can do this on the first pass, as any second pass
14456 * will yield an even shorter name) */
14457 if (name_len < 3) {
14458 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14461 /* Find which class it is. Initially switch on the length of the name.
14463 switch (name_len) {
14465 if (memEQ(name_start, "word", 4)) {
14466 /* this is not POSIX, this is the Perl \w */
14467 class_number = ANYOF_WORDCHAR;
14471 /* Names all of length 5: alnum alpha ascii blank cntrl digit
14472 * graph lower print punct space upper
14473 * Offset 4 gives the best switch position. */
14474 switch (name_start[4]) {
14476 if (memEQ(name_start, "alph", 4)) /* alpha */
14477 class_number = ANYOF_ALPHA;
14480 if (memEQ(name_start, "spac", 4)) /* space */
14481 class_number = ANYOF_SPACE;
14484 if (memEQ(name_start, "grap", 4)) /* graph */
14485 class_number = ANYOF_GRAPH;
14488 if (memEQ(name_start, "asci", 4)) /* ascii */
14489 class_number = ANYOF_ASCII;
14492 if (memEQ(name_start, "blan", 4)) /* blank */
14493 class_number = ANYOF_BLANK;
14496 if (memEQ(name_start, "cntr", 4)) /* cntrl */
14497 class_number = ANYOF_CNTRL;
14500 if (memEQ(name_start, "alnu", 4)) /* alnum */
14501 class_number = ANYOF_ALPHANUMERIC;
14504 if (memEQ(name_start, "lowe", 4)) /* lower */
14505 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
14506 else if (memEQ(name_start, "uppe", 4)) /* upper */
14507 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
14510 if (memEQ(name_start, "digi", 4)) /* digit */
14511 class_number = ANYOF_DIGIT;
14512 else if (memEQ(name_start, "prin", 4)) /* print */
14513 class_number = ANYOF_PRINT;
14514 else if (memEQ(name_start, "punc", 4)) /* punct */
14515 class_number = ANYOF_PUNCT;
14520 if (memEQ(name_start, "xdigit", 6))
14521 class_number = ANYOF_XDIGIT;
14525 /* If the name exactly matches a posix class name the class number will
14526 * here be set to it, and the input almost certainly was meant to be a
14527 * posix class, so we can skip further checking. If instead the syntax
14528 * is exactly correct, but the name isn't one of the legal ones, we
14529 * will return that as an error below. But if neither of these apply,
14530 * it could be that no posix class was intended at all, or that one
14531 * was, but there was a typo. We tease these apart by doing fuzzy
14532 * matching on the name */
14533 if (class_number == OOB_NAMEDCLASS && found_problem) {
14534 const UV posix_names[][6] = {
14535 { 'a', 'l', 'n', 'u', 'm' },
14536 { 'a', 'l', 'p', 'h', 'a' },
14537 { 'a', 's', 'c', 'i', 'i' },
14538 { 'b', 'l', 'a', 'n', 'k' },
14539 { 'c', 'n', 't', 'r', 'l' },
14540 { 'd', 'i', 'g', 'i', 't' },
14541 { 'g', 'r', 'a', 'p', 'h' },
14542 { 'l', 'o', 'w', 'e', 'r' },
14543 { 'p', 'r', 'i', 'n', 't' },
14544 { 'p', 'u', 'n', 'c', 't' },
14545 { 's', 'p', 'a', 'c', 'e' },
14546 { 'u', 'p', 'p', 'e', 'r' },
14547 { 'w', 'o', 'r', 'd' },
14548 { 'x', 'd', 'i', 'g', 'i', 't' }
14550 /* The names of the above all have added NULs to make them the same
14551 * size, so we need to also have the real lengths */
14552 const UV posix_name_lengths[] = {
14553 sizeof("alnum") - 1,
14554 sizeof("alpha") - 1,
14555 sizeof("ascii") - 1,
14556 sizeof("blank") - 1,
14557 sizeof("cntrl") - 1,
14558 sizeof("digit") - 1,
14559 sizeof("graph") - 1,
14560 sizeof("lower") - 1,
14561 sizeof("print") - 1,
14562 sizeof("punct") - 1,
14563 sizeof("space") - 1,
14564 sizeof("upper") - 1,
14565 sizeof("word") - 1,
14566 sizeof("xdigit")- 1
14569 int temp_max = max_distance; /* Use a temporary, so if we
14570 reparse, we haven't changed the
14573 /* Use a smaller max edit distance if we are missing one of the
14575 if ( has_opening_bracket + has_opening_colon < 2
14576 || has_terminating_bracket + has_terminating_colon < 2)
14581 /* See if the input name is close to a legal one */
14582 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
14584 /* Short circuit call if the lengths are too far apart to be
14586 if (abs( (int) (name_len - posix_name_lengths[i]))
14592 if (edit_distance(input_text,
14595 posix_name_lengths[i],
14599 { /* If it is close, it probably was intended to be a class */
14600 goto probably_meant_to_be;
14604 /* Here the input name is not close enough to a valid class name
14605 * for us to consider it to be intended to be a posix class. If
14606 * we haven't already done so, and the parse found a character that
14607 * could have been terminators for the name, but which we absorbed
14608 * as typos during the first pass, repeat the parse, signalling it
14609 * to stop at that character */
14610 if (possible_end && possible_end != (char *) -1) {
14611 possible_end = (char *) -1;
14616 /* Here neither pass found a close-enough class name */
14617 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14620 probably_meant_to_be:
14622 /* Here we think that a posix specification was intended. Update any
14624 if (updated_parse_ptr) {
14625 *updated_parse_ptr = (char *) p;
14628 /* If a posix class name was intended but incorrectly specified, we
14629 * output or return the warnings */
14630 if (found_problem) {
14632 /* We set flags for these issues in the parse loop above instead of
14633 * adding them to the list of warnings, because we can parse it
14634 * twice, and we only want one warning instance */
14636 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
14639 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14641 if (has_semi_colon) {
14642 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14644 else if (! has_terminating_colon) {
14645 ADD_POSIX_WARNING(p, "there is no terminating ':'");
14647 if (! has_terminating_bracket) {
14648 ADD_POSIX_WARNING(p, "there is no terminating ']'");
14651 if (posix_warnings && RExC_warn_text && av_top_index(RExC_warn_text) > -1) {
14652 *posix_warnings = RExC_warn_text;
14655 else if (class_number != OOB_NAMEDCLASS) {
14656 /* If it is a known class, return the class. The class number
14657 * #defines are structured so each complement is +1 to the normal
14659 return class_number + complement;
14661 else if (! check_only) {
14663 /* Here, it is an unrecognized class. This is an error (unless the
14664 * call is to check only, which we've already handled above) */
14665 const char * const complement_string = (complement)
14668 RExC_parse = (char *) p;
14669 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
14671 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
14675 return OOB_NAMEDCLASS;
14677 #undef ADD_POSIX_WARNING
14679 STATIC unsigned int
14680 S_regex_set_precedence(const U8 my_operator) {
14682 /* Returns the precedence in the (?[...]) construct of the input operator,
14683 * specified by its character representation. The precedence follows
14684 * general Perl rules, but it extends this so that ')' and ']' have (low)
14685 * precedence even though they aren't really operators */
14687 switch (my_operator) {
14703 NOT_REACHED; /* NOTREACHED */
14704 return 0; /* Silence compiler warning */
14708 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
14709 I32 *flagp, U32 depth,
14710 char * const oregcomp_parse)
14712 /* Handle the (?[...]) construct to do set operations */
14714 U8 curchar; /* Current character being parsed */
14715 UV start, end; /* End points of code point ranges */
14716 SV* final = NULL; /* The end result inversion list */
14717 SV* result_string; /* 'final' stringified */
14718 AV* stack; /* stack of operators and operands not yet
14720 AV* fence_stack = NULL; /* A stack containing the positions in
14721 'stack' of where the undealt-with left
14722 parens would be if they were actually
14724 /* The 'VOL' (expanding to 'volatile') is a workaround for an optimiser bug
14725 * in Solaris Studio 12.3. See RT #127455 */
14726 VOL IV fence = 0; /* Position of where most recent undealt-
14727 with left paren in stack is; -1 if none.
14729 STRLEN len; /* Temporary */
14730 regnode* node; /* Temporary, and final regnode returned by
14732 const bool save_fold = FOLD; /* Temporary */
14733 char *save_end, *save_parse; /* Temporaries */
14734 const bool in_locale = LOC; /* we turn off /l during processing */
14735 AV* posix_warnings = NULL;
14737 GET_RE_DEBUG_FLAGS_DECL;
14739 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
14742 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
14745 REQUIRE_UNI_RULES(flagp, NULL); /* The use of this operator implies /u.
14746 This is required so that the compile
14747 time values are valid in all runtime
14750 /* This will return only an ANYOF regnode, or (unlikely) something smaller
14751 * (such as EXACT). Thus we can skip most everything if just sizing. We
14752 * call regclass to handle '[]' so as to not have to reinvent its parsing
14753 * rules here (throwing away the size it computes each time). And, we exit
14754 * upon an unescaped ']' that isn't one ending a regclass. To do both
14755 * these things, we need to realize that something preceded by a backslash
14756 * is escaped, so we have to keep track of backslashes */
14758 UV depth = 0; /* how many nested (?[...]) constructs */
14760 while (RExC_parse < RExC_end) {
14761 SV* current = NULL;
14763 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14764 TRUE /* Force /x */ );
14766 switch (*RExC_parse) {
14768 if (RExC_parse[1] == '[') depth++, RExC_parse++;
14773 /* Skip past this, so the next character gets skipped, after
14776 if (*RExC_parse == 'c') {
14777 /* Skip the \cX notation for control characters */
14778 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14784 /* See if this is a [:posix:] class. */
14785 bool is_posix_class = (OOB_NAMEDCLASS
14786 < handle_possible_posix(pRExC_state,
14790 TRUE /* checking only */));
14791 /* If it is a posix class, leave the parse pointer at the
14792 * '[' to fool regclass() into thinking it is part of a
14793 * '[[:posix:]]'. */
14794 if (! is_posix_class) {
14798 /* regclass() can only return RESTART_PASS1 and NEED_UTF8
14799 * if multi-char folds are allowed. */
14800 if (!regclass(pRExC_state, flagp,depth+1,
14801 is_posix_class, /* parse the whole char
14802 class only if not a
14804 FALSE, /* don't allow multi-char folds */
14805 TRUE, /* silence non-portable warnings. */
14807 FALSE, /* Require return to be an ANYOF */
14811 FAIL2("panic: regclass returned NULL to handle_sets, "
14812 "flags=%#" UVxf, (UV) *flagp);
14814 /* function call leaves parse pointing to the ']', except
14815 * if we faked it */
14816 if (is_posix_class) {
14820 SvREFCNT_dec(current); /* In case it returned something */
14825 if (depth--) break;
14827 if (*RExC_parse == ')') {
14828 node = reganode(pRExC_state, ANYOF, 0);
14829 RExC_size += ANYOF_SKIP;
14830 nextchar(pRExC_state);
14831 Set_Node_Length(node,
14832 RExC_parse - oregcomp_parse + 1); /* MJD */
14834 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
14842 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14846 /* We output the messages even if warnings are off, because we'll fail
14847 * the very next thing, and these give a likely diagnosis for that */
14848 if (posix_warnings && av_tindex_nomg(posix_warnings) >= 0) {
14849 output_or_return_posix_warnings(pRExC_state, posix_warnings, NULL);
14852 FAIL("Syntax error in (?[...])");
14855 /* Pass 2 only after this. */
14856 Perl_ck_warner_d(aTHX_
14857 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
14858 "The regex_sets feature is experimental" REPORT_LOCATION,
14859 REPORT_LOCATION_ARGS(RExC_parse));
14861 /* Everything in this construct is a metacharacter. Operands begin with
14862 * either a '\' (for an escape sequence), or a '[' for a bracketed
14863 * character class. Any other character should be an operator, or
14864 * parenthesis for grouping. Both types of operands are handled by calling
14865 * regclass() to parse them. It is called with a parameter to indicate to
14866 * return the computed inversion list. The parsing here is implemented via
14867 * a stack. Each entry on the stack is a single character representing one
14868 * of the operators; or else a pointer to an operand inversion list. */
14870 #define IS_OPERATOR(a) SvIOK(a)
14871 #define IS_OPERAND(a) (! IS_OPERATOR(a))
14873 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
14874 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
14875 * with pronouncing it called it Reverse Polish instead, but now that YOU
14876 * know how to pronounce it you can use the correct term, thus giving due
14877 * credit to the person who invented it, and impressing your geek friends.
14878 * Wikipedia says that the pronounciation of "Ł" has been changing so that
14879 * it is now more like an English initial W (as in wonk) than an L.)
14881 * This means that, for example, 'a | b & c' is stored on the stack as
14889 * where the numbers in brackets give the stack [array] element number.
14890 * In this implementation, parentheses are not stored on the stack.
14891 * Instead a '(' creates a "fence" so that the part of the stack below the
14892 * fence is invisible except to the corresponding ')' (this allows us to
14893 * replace testing for parens, by using instead subtraction of the fence
14894 * position). As new operands are processed they are pushed onto the stack
14895 * (except as noted in the next paragraph). New operators of higher
14896 * precedence than the current final one are inserted on the stack before
14897 * the lhs operand (so that when the rhs is pushed next, everything will be
14898 * in the correct positions shown above. When an operator of equal or
14899 * lower precedence is encountered in parsing, all the stacked operations
14900 * of equal or higher precedence are evaluated, leaving the result as the
14901 * top entry on the stack. This makes higher precedence operations
14902 * evaluate before lower precedence ones, and causes operations of equal
14903 * precedence to left associate.
14905 * The only unary operator '!' is immediately pushed onto the stack when
14906 * encountered. When an operand is encountered, if the top of the stack is
14907 * a '!", the complement is immediately performed, and the '!' popped. The
14908 * resulting value is treated as a new operand, and the logic in the
14909 * previous paragraph is executed. Thus in the expression
14911 * the stack looks like
14917 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
14924 * A ')' is treated as an operator with lower precedence than all the
14925 * aforementioned ones, which causes all operations on the stack above the
14926 * corresponding '(' to be evaluated down to a single resultant operand.
14927 * Then the fence for the '(' is removed, and the operand goes through the
14928 * algorithm above, without the fence.
14930 * A separate stack is kept of the fence positions, so that the position of
14931 * the latest so-far unbalanced '(' is at the top of it.
14933 * The ']' ending the construct is treated as the lowest operator of all,
14934 * so that everything gets evaluated down to a single operand, which is the
14937 sv_2mortal((SV *)(stack = newAV()));
14938 sv_2mortal((SV *)(fence_stack = newAV()));
14940 while (RExC_parse < RExC_end) {
14941 I32 top_index; /* Index of top-most element in 'stack' */
14942 SV** top_ptr; /* Pointer to top 'stack' element */
14943 SV* current = NULL; /* To contain the current inversion list
14945 SV* only_to_avoid_leaks;
14947 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14948 TRUE /* Force /x */ );
14949 if (RExC_parse >= RExC_end) {
14950 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
14953 curchar = UCHARAT(RExC_parse);
14957 #ifdef ENABLE_REGEX_SETS_DEBUGGING
14958 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
14959 DEBUG_U(dump_regex_sets_structures(pRExC_state,
14960 stack, fence, fence_stack));
14963 top_index = av_tindex_nomg(stack);
14966 SV** stacked_ptr; /* Ptr to something already on 'stack' */
14967 char stacked_operator; /* The topmost operator on the 'stack'. */
14968 SV* lhs; /* Operand to the left of the operator */
14969 SV* rhs; /* Operand to the right of the operator */
14970 SV* fence_ptr; /* Pointer to top element of the fence
14975 if ( RExC_parse < RExC_end - 1
14976 && (UCHARAT(RExC_parse + 1) == '?'))
14978 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
14979 * This happens when we have some thing like
14981 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
14983 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
14985 * Here we would be handling the interpolated
14986 * '$thai_or_lao'. We handle this by a recursive call to
14987 * ourselves which returns the inversion list the
14988 * interpolated expression evaluates to. We use the flags
14989 * from the interpolated pattern. */
14990 U32 save_flags = RExC_flags;
14991 const char * save_parse;
14993 RExC_parse += 2; /* Skip past the '(?' */
14994 save_parse = RExC_parse;
14996 /* Parse any flags for the '(?' */
14997 parse_lparen_question_flags(pRExC_state);
14999 if (RExC_parse == save_parse /* Makes sure there was at
15000 least one flag (or else
15001 this embedding wasn't
15003 || RExC_parse >= RExC_end - 4
15004 || UCHARAT(RExC_parse) != ':'
15005 || UCHARAT(++RExC_parse) != '('
15006 || UCHARAT(++RExC_parse) != '?'
15007 || UCHARAT(++RExC_parse) != '[')
15010 /* In combination with the above, this moves the
15011 * pointer to the point just after the first erroneous
15012 * character (or if there are no flags, to where they
15013 * should have been) */
15014 if (RExC_parse >= RExC_end - 4) {
15015 RExC_parse = RExC_end;
15017 else if (RExC_parse != save_parse) {
15018 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15020 vFAIL("Expecting '(?flags:(?[...'");
15023 /* Recurse, with the meat of the embedded expression */
15025 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
15026 depth+1, oregcomp_parse);
15028 /* Here, 'current' contains the embedded expression's
15029 * inversion list, and RExC_parse points to the trailing
15030 * ']'; the next character should be the ')' */
15032 assert(UCHARAT(RExC_parse) == ')');
15034 /* Then the ')' matching the original '(' handled by this
15035 * case: statement */
15037 assert(UCHARAT(RExC_parse) == ')');
15040 RExC_flags = save_flags;
15041 goto handle_operand;
15044 /* A regular '('. Look behind for illegal syntax */
15045 if (top_index - fence >= 0) {
15046 /* If the top entry on the stack is an operator, it had
15047 * better be a '!', otherwise the entry below the top
15048 * operand should be an operator */
15049 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
15050 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
15051 || ( IS_OPERAND(*top_ptr)
15052 && ( top_index - fence < 1
15053 || ! (stacked_ptr = av_fetch(stack,
15056 || ! IS_OPERATOR(*stacked_ptr))))
15059 vFAIL("Unexpected '(' with no preceding operator");
15063 /* Stack the position of this undealt-with left paren */
15064 av_push(fence_stack, newSViv(fence));
15065 fence = top_index + 1;
15069 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15070 * multi-char folds are allowed. */
15071 if (!regclass(pRExC_state, flagp,depth+1,
15072 TRUE, /* means parse just the next thing */
15073 FALSE, /* don't allow multi-char folds */
15074 FALSE, /* don't silence non-portable warnings. */
15076 FALSE, /* Require return to be an ANYOF */
15080 FAIL2("panic: regclass returned NULL to handle_sets, "
15081 "flags=%#" UVxf, (UV) *flagp);
15084 /* regclass() will return with parsing just the \ sequence,
15085 * leaving the parse pointer at the next thing to parse */
15087 goto handle_operand;
15089 case '[': /* Is a bracketed character class */
15091 /* See if this is a [:posix:] class. */
15092 bool is_posix_class = (OOB_NAMEDCLASS
15093 < handle_possible_posix(pRExC_state,
15097 TRUE /* checking only */));
15098 /* If it is a posix class, leave the parse pointer at the '['
15099 * to fool regclass() into thinking it is part of a
15100 * '[[:posix:]]'. */
15101 if (! is_posix_class) {
15105 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15106 * multi-char folds are allowed. */
15107 if (!regclass(pRExC_state, flagp,depth+1,
15108 is_posix_class, /* parse the whole char
15109 class only if not a
15111 FALSE, /* don't allow multi-char folds */
15112 TRUE, /* silence non-portable warnings. */
15114 FALSE, /* Require return to be an ANYOF */
15119 FAIL2("panic: regclass returned NULL to handle_sets, "
15120 "flags=%#" UVxf, (UV) *flagp);
15123 /* function call leaves parse pointing to the ']', except if we
15125 if (is_posix_class) {
15129 goto handle_operand;
15133 if (top_index >= 1) {
15134 goto join_operators;
15137 /* Only a single operand on the stack: are done */
15141 if (av_tindex_nomg(fence_stack) < 0) {
15143 vFAIL("Unexpected ')'");
15146 /* If nothing after the fence, is missing an operand */
15147 if (top_index - fence < 0) {
15151 /* If at least two things on the stack, treat this as an
15153 if (top_index - fence >= 1) {
15154 goto join_operators;
15157 /* Here only a single thing on the fenced stack, and there is a
15158 * fence. Get rid of it */
15159 fence_ptr = av_pop(fence_stack);
15161 fence = SvIV(fence_ptr) - 1;
15162 SvREFCNT_dec_NN(fence_ptr);
15169 /* Having gotten rid of the fence, we pop the operand at the
15170 * stack top and process it as a newly encountered operand */
15171 current = av_pop(stack);
15172 if (IS_OPERAND(current)) {
15173 goto handle_operand;
15185 /* These binary operators should have a left operand already
15187 if ( top_index - fence < 0
15188 || top_index - fence == 1
15189 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
15190 || ! IS_OPERAND(*top_ptr))
15192 goto unexpected_binary;
15195 /* If only the one operand is on the part of the stack visible
15196 * to us, we just place this operator in the proper position */
15197 if (top_index - fence < 2) {
15199 /* Place the operator before the operand */
15201 SV* lhs = av_pop(stack);
15202 av_push(stack, newSVuv(curchar));
15203 av_push(stack, lhs);
15207 /* But if there is something else on the stack, we need to
15208 * process it before this new operator if and only if the
15209 * stacked operation has equal or higher precedence than the
15214 /* The operator on the stack is supposed to be below both its
15216 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
15217 || IS_OPERAND(*stacked_ptr))
15219 /* But if not, it's legal and indicates we are completely
15220 * done if and only if we're currently processing a ']',
15221 * which should be the final thing in the expression */
15222 if (curchar == ']') {
15228 vFAIL2("Unexpected binary operator '%c' with no "
15229 "preceding operand", curchar);
15231 stacked_operator = (char) SvUV(*stacked_ptr);
15233 if (regex_set_precedence(curchar)
15234 > regex_set_precedence(stacked_operator))
15236 /* Here, the new operator has higher precedence than the
15237 * stacked one. This means we need to add the new one to
15238 * the stack to await its rhs operand (and maybe more
15239 * stuff). We put it before the lhs operand, leaving
15240 * untouched the stacked operator and everything below it
15242 lhs = av_pop(stack);
15243 assert(IS_OPERAND(lhs));
15245 av_push(stack, newSVuv(curchar));
15246 av_push(stack, lhs);
15250 /* Here, the new operator has equal or lower precedence than
15251 * what's already there. This means the operation already
15252 * there should be performed now, before the new one. */
15254 rhs = av_pop(stack);
15255 if (! IS_OPERAND(rhs)) {
15257 /* This can happen when a ! is not followed by an operand,
15258 * like in /(?[\t &!])/ */
15262 lhs = av_pop(stack);
15264 if (! IS_OPERAND(lhs)) {
15266 /* This can happen when there is an empty (), like in
15267 * /(?[[0]+()+])/ */
15271 switch (stacked_operator) {
15273 _invlist_intersection(lhs, rhs, &rhs);
15278 _invlist_union(lhs, rhs, &rhs);
15282 _invlist_subtract(lhs, rhs, &rhs);
15285 case '^': /* The union minus the intersection */
15290 _invlist_union(lhs, rhs, &u);
15291 _invlist_intersection(lhs, rhs, &i);
15292 _invlist_subtract(u, i, &rhs);
15293 SvREFCNT_dec_NN(i);
15294 SvREFCNT_dec_NN(u);
15300 /* Here, the higher precedence operation has been done, and the
15301 * result is in 'rhs'. We overwrite the stacked operator with
15302 * the result. Then we redo this code to either push the new
15303 * operator onto the stack or perform any higher precedence
15304 * stacked operation */
15305 only_to_avoid_leaks = av_pop(stack);
15306 SvREFCNT_dec(only_to_avoid_leaks);
15307 av_push(stack, rhs);
15310 case '!': /* Highest priority, right associative */
15312 /* If what's already at the top of the stack is another '!",
15313 * they just cancel each other out */
15314 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
15315 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
15317 only_to_avoid_leaks = av_pop(stack);
15318 SvREFCNT_dec(only_to_avoid_leaks);
15320 else { /* Otherwise, since it's right associative, just push
15322 av_push(stack, newSVuv(curchar));
15327 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15328 vFAIL("Unexpected character");
15332 /* Here 'current' is the operand. If something is already on the
15333 * stack, we have to check if it is a !. But first, the code above
15334 * may have altered the stack in the time since we earlier set
15337 top_index = av_tindex_nomg(stack);
15338 if (top_index - fence >= 0) {
15339 /* If the top entry on the stack is an operator, it had better
15340 * be a '!', otherwise the entry below the top operand should
15341 * be an operator */
15342 top_ptr = av_fetch(stack, top_index, FALSE);
15344 if (IS_OPERATOR(*top_ptr)) {
15346 /* The only permissible operator at the top of the stack is
15347 * '!', which is applied immediately to this operand. */
15348 curchar = (char) SvUV(*top_ptr);
15349 if (curchar != '!') {
15350 SvREFCNT_dec(current);
15351 vFAIL2("Unexpected binary operator '%c' with no "
15352 "preceding operand", curchar);
15355 _invlist_invert(current);
15357 only_to_avoid_leaks = av_pop(stack);
15358 SvREFCNT_dec(only_to_avoid_leaks);
15360 /* And we redo with the inverted operand. This allows
15361 * handling multiple ! in a row */
15362 goto handle_operand;
15364 /* Single operand is ok only for the non-binary ')'
15366 else if ((top_index - fence == 0 && curchar != ')')
15367 || (top_index - fence > 0
15368 && (! (stacked_ptr = av_fetch(stack,
15371 || IS_OPERAND(*stacked_ptr))))
15373 SvREFCNT_dec(current);
15374 vFAIL("Operand with no preceding operator");
15378 /* Here there was nothing on the stack or the top element was
15379 * another operand. Just add this new one */
15380 av_push(stack, current);
15382 } /* End of switch on next parse token */
15384 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15385 } /* End of loop parsing through the construct */
15388 if (av_tindex_nomg(fence_stack) >= 0) {
15389 vFAIL("Unmatched (");
15392 if (av_tindex_nomg(stack) < 0 /* Was empty */
15393 || ((final = av_pop(stack)) == NULL)
15394 || ! IS_OPERAND(final)
15395 || SvTYPE(final) != SVt_INVLIST
15396 || av_tindex_nomg(stack) >= 0) /* More left on stack */
15399 SvREFCNT_dec(final);
15400 vFAIL("Incomplete expression within '(?[ ])'");
15403 /* Here, 'final' is the resultant inversion list from evaluating the
15404 * expression. Return it if so requested */
15405 if (return_invlist) {
15406 *return_invlist = final;
15410 /* Otherwise generate a resultant node, based on 'final'. regclass() is
15411 * expecting a string of ranges and individual code points */
15412 invlist_iterinit(final);
15413 result_string = newSVpvs("");
15414 while (invlist_iternext(final, &start, &end)) {
15415 if (start == end) {
15416 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
15419 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%" UVXf "}",
15424 /* About to generate an ANYOF (or similar) node from the inversion list we
15425 * have calculated */
15426 save_parse = RExC_parse;
15427 RExC_parse = SvPV(result_string, len);
15428 save_end = RExC_end;
15429 RExC_end = RExC_parse + len;
15431 /* We turn off folding around the call, as the class we have constructed
15432 * already has all folding taken into consideration, and we don't want
15433 * regclass() to add to that */
15434 RExC_flags &= ~RXf_PMf_FOLD;
15435 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if multi-char
15436 * folds are allowed. */
15437 node = regclass(pRExC_state, flagp,depth+1,
15438 FALSE, /* means parse the whole char class */
15439 FALSE, /* don't allow multi-char folds */
15440 TRUE, /* silence non-portable warnings. The above may very
15441 well have generated non-portable code points, but
15442 they're valid on this machine */
15443 FALSE, /* similarly, no need for strict */
15444 FALSE, /* Require return to be an ANYOF */
15449 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#" UVxf,
15452 /* Fix up the node type if we are in locale. (We have pretended we are
15453 * under /u for the purposes of regclass(), as this construct will only
15454 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
15455 * as to cause any warnings about bad locales to be output in regexec.c),
15456 * and add the flag that indicates to check if not in a UTF-8 locale. The
15457 * reason we above forbid optimization into something other than an ANYOF
15458 * node is simply to minimize the number of code changes in regexec.c.
15459 * Otherwise we would have to create new EXACTish node types and deal with
15460 * them. This decision could be revisited should this construct become
15463 * (One might think we could look at the resulting ANYOF node and suppress
15464 * the flag if everything is above 255, as those would be UTF-8 only,
15465 * but this isn't true, as the components that led to that result could
15466 * have been locale-affected, and just happen to cancel each other out
15467 * under UTF-8 locales.) */
15469 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
15471 assert(OP(node) == ANYOF);
15475 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
15479 RExC_flags |= RXf_PMf_FOLD;
15482 RExC_parse = save_parse + 1;
15483 RExC_end = save_end;
15484 SvREFCNT_dec_NN(final);
15485 SvREFCNT_dec_NN(result_string);
15487 nextchar(pRExC_state);
15488 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
15492 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15495 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
15496 AV * stack, const IV fence, AV * fence_stack)
15497 { /* Dumps the stacks in handle_regex_sets() */
15499 const SSize_t stack_top = av_tindex_nomg(stack);
15500 const SSize_t fence_stack_top = av_tindex_nomg(fence_stack);
15503 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
15505 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
15507 if (stack_top < 0) {
15508 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
15511 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
15512 for (i = stack_top; i >= 0; i--) {
15513 SV ** element_ptr = av_fetch(stack, i, FALSE);
15514 if (! element_ptr) {
15517 if (IS_OPERATOR(*element_ptr)) {
15518 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
15519 (int) i, (int) SvIV(*element_ptr));
15522 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
15523 sv_dump(*element_ptr);
15528 if (fence_stack_top < 0) {
15529 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
15532 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
15533 for (i = fence_stack_top; i >= 0; i--) {
15534 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
15535 if (! element_ptr) {
15538 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
15539 (int) i, (int) SvIV(*element_ptr));
15550 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
15552 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
15553 * innocent-looking character class, like /[ks]/i won't have to go out to
15554 * disk to find the possible matches.
15556 * This should be called only for a Latin1-range code points, cp, which is
15557 * known to be involved in a simple fold with other code points above
15558 * Latin1. It would give false results if /aa has been specified.
15559 * Multi-char folds are outside the scope of this, and must be handled
15562 * XXX It would be better to generate these via regen, in case a new
15563 * version of the Unicode standard adds new mappings, though that is not
15564 * really likely, and may be caught by the default: case of the switch
15567 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
15569 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
15575 add_cp_to_invlist(*invlist, KELVIN_SIGN);
15579 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
15582 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
15583 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
15585 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
15586 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
15587 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
15589 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
15590 *invlist = add_cp_to_invlist(*invlist,
15591 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
15594 #ifdef LATIN_CAPITAL_LETTER_SHARP_S /* not defined in early Unicode releases */
15596 case LATIN_SMALL_LETTER_SHARP_S:
15597 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
15602 #if UNICODE_MAJOR_VERSION < 3 \
15603 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0)
15605 /* In 3.0 and earlier, U+0130 folded simply to 'i'; and in 3.0.1 so did
15610 add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
15611 # if UNICODE_DOT_DOT_VERSION == 1
15612 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_DOTLESS_I);
15618 /* Use deprecated warning to increase the chances of this being
15621 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
15628 S_output_or_return_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings, AV** return_posix_warnings)
15630 /* If the final parameter is NULL, output the elements of the array given
15631 * by '*posix_warnings' as REGEXP warnings. Otherwise, the elements are
15632 * pushed onto it, (creating if necessary) */
15635 const bool first_is_fatal = ! return_posix_warnings
15636 && ckDEAD(packWARN(WARN_REGEXP));
15638 PERL_ARGS_ASSERT_OUTPUT_OR_RETURN_POSIX_WARNINGS;
15640 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
15641 if (return_posix_warnings) {
15642 if (! *return_posix_warnings) { /* mortalize to not leak if
15643 warnings are fatal */
15644 *return_posix_warnings = (AV *) sv_2mortal((SV *) newAV());
15646 av_push(*return_posix_warnings, msg);
15649 if (first_is_fatal) { /* Avoid leaking this */
15650 av_undef(posix_warnings); /* This isn't necessary if the
15651 array is mortal, but is a
15653 (void) sv_2mortal(msg);
15655 SAVEFREESV(RExC_rx_sv);
15658 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
15659 SvREFCNT_dec_NN(msg);
15665 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
15667 /* This adds the string scalar <multi_string> to the array
15668 * <multi_char_matches>. <multi_string> is known to have exactly
15669 * <cp_count> code points in it. This is used when constructing a
15670 * bracketed character class and we find something that needs to match more
15671 * than a single character.
15673 * <multi_char_matches> is actually an array of arrays. Each top-level
15674 * element is an array that contains all the strings known so far that are
15675 * the same length. And that length (in number of code points) is the same
15676 * as the index of the top-level array. Hence, the [2] element is an
15677 * array, each element thereof is a string containing TWO code points;
15678 * while element [3] is for strings of THREE characters, and so on. Since
15679 * this is for multi-char strings there can never be a [0] nor [1] element.
15681 * When we rewrite the character class below, we will do so such that the
15682 * longest strings are written first, so that it prefers the longest
15683 * matching strings first. This is done even if it turns out that any
15684 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
15685 * Christiansen has agreed that this is ok. This makes the test for the
15686 * ligature 'ffi' come before the test for 'ff', for example */
15689 AV** this_array_ptr;
15691 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
15693 if (! multi_char_matches) {
15694 multi_char_matches = newAV();
15697 if (av_exists(multi_char_matches, cp_count)) {
15698 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
15699 this_array = *this_array_ptr;
15702 this_array = newAV();
15703 av_store(multi_char_matches, cp_count,
15706 av_push(this_array, multi_string);
15708 return multi_char_matches;
15711 /* The names of properties whose definitions are not known at compile time are
15712 * stored in this SV, after a constant heading. So if the length has been
15713 * changed since initialization, then there is a run-time definition. */
15714 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
15715 (SvCUR(listsv) != initial_listsv_len)
15717 /* There is a restricted set of white space characters that are legal when
15718 * ignoring white space in a bracketed character class. This generates the
15719 * code to skip them.
15721 * There is a line below that uses the same white space criteria but is outside
15722 * this macro. Both here and there must use the same definition */
15723 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
15726 while (isBLANK_A(UCHARAT(p))) \
15734 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
15735 const bool stop_at_1, /* Just parse the next thing, don't
15736 look for a full character class */
15737 bool allow_multi_folds,
15738 const bool silence_non_portable, /* Don't output warnings
15742 bool optimizable, /* ? Allow a non-ANYOF return
15744 SV** ret_invlist, /* Return an inversion list, not a node */
15745 AV** return_posix_warnings
15748 /* parse a bracketed class specification. Most of these will produce an
15749 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
15750 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
15751 * under /i with multi-character folds: it will be rewritten following the
15752 * paradigm of this example, where the <multi-fold>s are characters which
15753 * fold to multiple character sequences:
15754 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
15755 * gets effectively rewritten as:
15756 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
15757 * reg() gets called (recursively) on the rewritten version, and this
15758 * function will return what it constructs. (Actually the <multi-fold>s
15759 * aren't physically removed from the [abcdefghi], it's just that they are
15760 * ignored in the recursion by means of a flag:
15761 * <RExC_in_multi_char_class>.)
15763 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
15764 * characters, with the corresponding bit set if that character is in the
15765 * list. For characters above this, a range list or swash is used. There
15766 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
15767 * determinable at compile time
15769 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs
15770 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded
15771 * to UTF-8. This can only happen if ret_invlist is non-NULL.
15774 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
15776 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
15779 int namedclass = OOB_NAMEDCLASS;
15780 char *rangebegin = NULL;
15781 bool need_class = 0;
15783 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
15784 than just initialized. */
15785 SV* properties = NULL; /* Code points that match \p{} \P{} */
15786 SV* posixes = NULL; /* Code points that match classes like [:word:],
15787 extended beyond the Latin1 range. These have to
15788 be kept separate from other code points for much
15789 of this function because their handling is
15790 different under /i, and for most classes under
15792 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
15793 separate for a while from the non-complemented
15794 versions because of complications with /d
15796 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
15797 treated more simply than the general case,
15798 leading to less compilation and execution
15800 UV element_count = 0; /* Number of distinct elements in the class.
15801 Optimizations may be possible if this is tiny */
15802 AV * multi_char_matches = NULL; /* Code points that fold to more than one
15803 character; used under /i */
15805 char * stop_ptr = RExC_end; /* where to stop parsing */
15807 /* ignore unescaped whitespace? */
15808 const bool skip_white = cBOOL( ret_invlist
15809 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
15811 /* Unicode properties are stored in a swash; this holds the current one
15812 * being parsed. If this swash is the only above-latin1 component of the
15813 * character class, an optimization is to pass it directly on to the
15814 * execution engine. Otherwise, it is set to NULL to indicate that there
15815 * are other things in the class that have to be dealt with at execution
15817 SV* swash = NULL; /* Code points that match \p{} \P{} */
15819 /* Set if a component of this character class is user-defined; just passed
15820 * on to the engine */
15821 bool has_user_defined_property = FALSE;
15823 /* inversion list of code points this node matches only when the target
15824 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
15826 SV* has_upper_latin1_only_utf8_matches = NULL;
15828 /* Inversion list of code points this node matches regardless of things
15829 * like locale, folding, utf8ness of the target string */
15830 SV* cp_list = NULL;
15832 /* Like cp_list, but code points on this list need to be checked for things
15833 * that fold to/from them under /i */
15834 SV* cp_foldable_list = NULL;
15836 /* Like cp_list, but code points on this list are valid only when the
15837 * runtime locale is UTF-8 */
15838 SV* only_utf8_locale_list = NULL;
15840 /* In a range, if one of the endpoints is non-character-set portable,
15841 * meaning that it hard-codes a code point that may mean a different
15842 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
15843 * mnemonic '\t' which each mean the same character no matter which
15844 * character set the platform is on. */
15845 unsigned int non_portable_endpoint = 0;
15847 /* Is the range unicode? which means on a platform that isn't 1-1 native
15848 * to Unicode (i.e. non-ASCII), each code point in it should be considered
15849 * to be a Unicode value. */
15850 bool unicode_range = FALSE;
15851 bool invert = FALSE; /* Is this class to be complemented */
15853 bool warn_super = ALWAYS_WARN_SUPER;
15855 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
15856 case we need to change the emitted regop to an EXACT. */
15857 const char * orig_parse = RExC_parse;
15858 const SSize_t orig_size = RExC_size;
15859 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
15861 /* This variable is used to mark where the end in the input is of something
15862 * that looks like a POSIX construct but isn't. During the parse, when
15863 * something looks like it could be such a construct is encountered, it is
15864 * checked for being one, but not if we've already checked this area of the
15865 * input. Only after this position is reached do we check again */
15866 char *not_posix_region_end = RExC_parse - 1;
15868 AV* posix_warnings = NULL;
15869 const bool do_posix_warnings = return_posix_warnings
15870 || (PASS2 && ckWARN(WARN_REGEXP));
15872 GET_RE_DEBUG_FLAGS_DECL;
15874 PERL_ARGS_ASSERT_REGCLASS;
15876 PERL_UNUSED_ARG(depth);
15879 DEBUG_PARSE("clas");
15881 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
15882 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
15883 && UNICODE_DOT_DOT_VERSION == 0)
15884 allow_multi_folds = FALSE;
15887 /* Assume we are going to generate an ANYOF node. */
15888 ret = reganode(pRExC_state,
15895 RExC_size += ANYOF_SKIP;
15896 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
15899 ANYOF_FLAGS(ret) = 0;
15901 RExC_emit += ANYOF_SKIP;
15902 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
15903 initial_listsv_len = SvCUR(listsv);
15904 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
15907 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15909 assert(RExC_parse <= RExC_end);
15911 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
15914 allow_multi_folds = FALSE;
15916 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15919 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
15920 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
15921 int maybe_class = handle_possible_posix(pRExC_state,
15923 ¬_posix_region_end,
15925 TRUE /* checking only */);
15926 if (PASS2 && maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
15927 SAVEFREESV(RExC_rx_sv);
15928 ckWARN4reg(not_posix_region_end,
15929 "POSIX syntax [%c %c] belongs inside character classes%s",
15930 *RExC_parse, *RExC_parse,
15931 (maybe_class == OOB_NAMEDCLASS)
15932 ? ((POSIXCC_NOTYET(*RExC_parse))
15933 ? " (but this one isn't implemented)"
15934 : " (but this one isn't fully valid)")
15937 (void)ReREFCNT_inc(RExC_rx_sv);
15941 /* If the caller wants us to just parse a single element, accomplish this
15942 * by faking the loop ending condition */
15943 if (stop_at_1 && RExC_end > RExC_parse) {
15944 stop_ptr = RExC_parse + 1;
15947 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
15948 if (UCHARAT(RExC_parse) == ']')
15949 goto charclassloop;
15953 if ( posix_warnings
15954 && av_tindex_nomg(posix_warnings) >= 0
15955 && RExC_parse > not_posix_region_end)
15957 /* Warnings about posix class issues are considered tentative until
15958 * we are far enough along in the parse that we can no longer
15959 * change our mind, at which point we either output them or add
15960 * them, if it has so specified, to what gets returned to the
15961 * caller. This is done each time through the loop so that a later
15962 * class won't zap them before they have been dealt with. */
15963 output_or_return_posix_warnings(pRExC_state, posix_warnings,
15964 return_posix_warnings);
15967 if (RExC_parse >= stop_ptr) {
15971 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15973 if (UCHARAT(RExC_parse) == ']') {
15979 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
15980 save_value = value;
15981 save_prevvalue = prevvalue;
15984 rangebegin = RExC_parse;
15986 non_portable_endpoint = 0;
15988 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
15989 value = utf8n_to_uvchr((U8*)RExC_parse,
15990 RExC_end - RExC_parse,
15991 &numlen, UTF8_ALLOW_DEFAULT);
15992 RExC_parse += numlen;
15995 value = UCHARAT(RExC_parse++);
15997 if (value == '[') {
15998 char * posix_class_end;
15999 namedclass = handle_possible_posix(pRExC_state,
16002 do_posix_warnings ? &posix_warnings : NULL,
16003 FALSE /* die if error */);
16004 if (namedclass > OOB_NAMEDCLASS) {
16006 /* If there was an earlier attempt to parse this particular
16007 * posix class, and it failed, it was a false alarm, as this
16008 * successful one proves */
16009 if ( posix_warnings
16010 && av_tindex_nomg(posix_warnings) >= 0
16011 && not_posix_region_end >= RExC_parse
16012 && not_posix_region_end <= posix_class_end)
16014 av_undef(posix_warnings);
16017 RExC_parse = posix_class_end;
16019 else if (namedclass == OOB_NAMEDCLASS) {
16020 not_posix_region_end = posix_class_end;
16023 namedclass = OOB_NAMEDCLASS;
16026 else if ( RExC_parse - 1 > not_posix_region_end
16027 && MAYBE_POSIXCC(value))
16029 (void) handle_possible_posix(
16031 RExC_parse - 1, /* -1 because parse has already been
16033 ¬_posix_region_end,
16034 do_posix_warnings ? &posix_warnings : NULL,
16035 TRUE /* checking only */);
16037 else if (value == '\\') {
16038 /* Is a backslash; get the code point of the char after it */
16040 if (RExC_parse >= RExC_end) {
16041 vFAIL("Unmatched [");
16044 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
16045 value = utf8n_to_uvchr((U8*)RExC_parse,
16046 RExC_end - RExC_parse,
16047 &numlen, UTF8_ALLOW_DEFAULT);
16048 RExC_parse += numlen;
16051 value = UCHARAT(RExC_parse++);
16053 /* Some compilers cannot handle switching on 64-bit integer
16054 * values, therefore value cannot be an UV. Yes, this will
16055 * be a problem later if we want switch on Unicode.
16056 * A similar issue a little bit later when switching on
16057 * namedclass. --jhi */
16059 /* If the \ is escaping white space when white space is being
16060 * skipped, it means that that white space is wanted literally, and
16061 * is already in 'value'. Otherwise, need to translate the escape
16062 * into what it signifies. */
16063 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
16065 case 'w': namedclass = ANYOF_WORDCHAR; break;
16066 case 'W': namedclass = ANYOF_NWORDCHAR; break;
16067 case 's': namedclass = ANYOF_SPACE; break;
16068 case 'S': namedclass = ANYOF_NSPACE; break;
16069 case 'd': namedclass = ANYOF_DIGIT; break;
16070 case 'D': namedclass = ANYOF_NDIGIT; break;
16071 case 'v': namedclass = ANYOF_VERTWS; break;
16072 case 'V': namedclass = ANYOF_NVERTWS; break;
16073 case 'h': namedclass = ANYOF_HORIZWS; break;
16074 case 'H': namedclass = ANYOF_NHORIZWS; break;
16075 case 'N': /* Handle \N{NAME} in class */
16077 const char * const backslash_N_beg = RExC_parse - 2;
16080 if (! grok_bslash_N(pRExC_state,
16081 NULL, /* No regnode */
16082 &value, /* Yes single value */
16083 &cp_count, /* Multiple code pt count */
16089 if (*flagp & NEED_UTF8)
16090 FAIL("panic: grok_bslash_N set NEED_UTF8");
16091 if (*flagp & RESTART_PASS1)
16094 if (cp_count < 0) {
16095 vFAIL("\\N in a character class must be a named character: \\N{...}");
16097 else if (cp_count == 0) {
16099 ckWARNreg(RExC_parse,
16100 "Ignoring zero length \\N{} in character class");
16103 else { /* cp_count > 1 */
16104 if (! RExC_in_multi_char_class) {
16105 if (invert || range || *RExC_parse == '-') {
16108 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
16111 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
16113 break; /* <value> contains the first code
16114 point. Drop out of the switch to
16118 SV * multi_char_N = newSVpvn(backslash_N_beg,
16119 RExC_parse - backslash_N_beg);
16121 = add_multi_match(multi_char_matches,
16126 } /* End of cp_count != 1 */
16128 /* This element should not be processed further in this
16131 value = save_value;
16132 prevvalue = save_prevvalue;
16133 continue; /* Back to top of loop to get next char */
16136 /* Here, is a single code point, and <value> contains it */
16137 unicode_range = TRUE; /* \N{} are Unicode */
16145 /* We will handle any undefined properties ourselves */
16146 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
16147 /* And we actually would prefer to get
16148 * the straight inversion list of the
16149 * swash, since we will be accessing it
16150 * anyway, to save a little time */
16151 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
16153 if (RExC_parse >= RExC_end)
16154 vFAIL2("Empty \\%c", (U8)value);
16155 if (*RExC_parse == '{') {
16156 const U8 c = (U8)value;
16157 e = strchr(RExC_parse, '}');
16160 vFAIL2("Missing right brace on \\%c{}", c);
16164 while (isSPACE(*RExC_parse)) {
16168 if (UCHARAT(RExC_parse) == '^') {
16170 /* toggle. (The rhs xor gets the single bit that
16171 * differs between P and p; the other xor inverts just
16173 value ^= 'P' ^ 'p';
16176 while (isSPACE(*RExC_parse)) {
16181 if (e == RExC_parse)
16182 vFAIL2("Empty \\%c{}", c);
16184 n = e - RExC_parse;
16185 while (isSPACE(*(RExC_parse + n - 1)))
16187 } /* The \p isn't immediately followed by a '{' */
16188 else if (! isALPHA(*RExC_parse)) {
16189 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16190 vFAIL2("Character following \\%c must be '{' or a "
16191 "single-character Unicode property name",
16201 char* base_name; /* name after any packages are stripped */
16202 char* lookup_name = NULL;
16203 const char * const colon_colon = "::";
16205 /* Try to get the definition of the property into
16206 * <invlist>. If /i is in effect, the effective property
16207 * will have its name be <__NAME_i>. The design is
16208 * discussed in commit
16209 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
16210 name = savepv(Perl_form(aTHX_ "%.*s", (int)n, RExC_parse));
16213 lookup_name = savepv(Perl_form(aTHX_ "__%s_i", name));
16215 /* The function call just below that uses this can fail
16216 * to return, leaking memory if we don't do this */
16217 SAVEFREEPV(lookup_name);
16220 /* Look up the property name, and get its swash and
16221 * inversion list, if the property is found */
16222 SvREFCNT_dec(swash); /* Free any left-overs */
16223 swash = _core_swash_init("utf8",
16230 NULL, /* No inversion list */
16233 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
16234 HV* curpkg = (IN_PERL_COMPILETIME)
16236 : CopSTASH(PL_curcop);
16240 if (swash) { /* Got a swash but no inversion list.
16241 Something is likely wrong that will
16242 be sorted-out later */
16243 SvREFCNT_dec_NN(swash);
16247 /* Here didn't find it. It could be a an error (like a
16248 * typo) in specifying a Unicode property, or it could
16249 * be a user-defined property that will be available at
16250 * run-time. The names of these must begin with 'In'
16251 * or 'Is' (after any packages are stripped off). So
16252 * if not one of those, or if we accept only
16253 * compile-time properties, is an error; otherwise add
16254 * it to the list for run-time look up. */
16255 if ((base_name = rninstr(name, name + n,
16256 colon_colon, colon_colon + 2)))
16257 { /* Has ::. We know this must be a user-defined
16260 final_n -= base_name - name;
16269 || base_name[0] != 'I'
16270 || (base_name[1] != 's' && base_name[1] != 'n')
16273 const char * const msg
16275 ? "Illegal user-defined property name"
16276 : "Can't find Unicode property definition";
16277 RExC_parse = e + 1;
16279 /* diag_listed_as: Can't find Unicode property definition "%s" */
16280 vFAIL3utf8f("%s \"%" UTF8f "\"",
16281 msg, UTF8fARG(UTF, n, name));
16284 /* If the property name doesn't already have a package
16285 * name, add the current one to it so that it can be
16286 * referred to outside it. [perl #121777] */
16287 if (! has_pkg && curpkg) {
16288 char* pkgname = HvNAME(curpkg);
16289 if (strNE(pkgname, "main")) {
16290 char* full_name = Perl_form(aTHX_
16294 n = strlen(full_name);
16295 name = savepvn(full_name, n);
16299 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s%" UTF8f "%s\n",
16300 (value == 'p' ? '+' : '!'),
16301 (FOLD) ? "__" : "",
16302 UTF8fARG(UTF, n, name),
16303 (FOLD) ? "_i" : "");
16304 has_user_defined_property = TRUE;
16305 optimizable = FALSE; /* Will have to leave this an
16308 /* We don't know yet what this matches, so have to flag
16310 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
16314 /* Here, did get the swash and its inversion list. If
16315 * the swash is from a user-defined property, then this
16316 * whole character class should be regarded as such */
16317 if (swash_init_flags
16318 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
16320 has_user_defined_property = TRUE;
16323 /* We warn on matching an above-Unicode code point
16324 * if the match would return true, except don't
16325 * warn for \p{All}, which has exactly one element
16327 (_invlist_contains_cp(invlist, 0x110000)
16328 && (! (_invlist_len(invlist) == 1
16329 && *invlist_array(invlist) == 0)))
16335 /* Invert if asking for the complement */
16336 if (value == 'P') {
16337 _invlist_union_complement_2nd(properties,
16341 /* The swash can't be used as-is, because we've
16342 * inverted things; delay removing it to here after
16343 * have copied its invlist above */
16344 SvREFCNT_dec_NN(swash);
16348 _invlist_union(properties, invlist, &properties);
16352 RExC_parse = e + 1;
16353 namedclass = ANYOF_UNIPROP; /* no official name, but it's
16356 /* \p means they want Unicode semantics */
16357 REQUIRE_UNI_RULES(flagp, NULL);
16360 case 'n': value = '\n'; break;
16361 case 'r': value = '\r'; break;
16362 case 't': value = '\t'; break;
16363 case 'f': value = '\f'; break;
16364 case 'b': value = '\b'; break;
16365 case 'e': value = ESC_NATIVE; break;
16366 case 'a': value = '\a'; break;
16368 RExC_parse--; /* function expects to be pointed at the 'o' */
16370 const char* error_msg;
16371 bool valid = grok_bslash_o(&RExC_parse,
16374 PASS2, /* warnings only in
16377 silence_non_portable,
16383 non_portable_endpoint++;
16386 RExC_parse--; /* function expects to be pointed at the 'x' */
16388 const char* error_msg;
16389 bool valid = grok_bslash_x(&RExC_parse,
16392 PASS2, /* Output warnings */
16394 silence_non_portable,
16400 non_portable_endpoint++;
16403 value = grok_bslash_c(*RExC_parse++, PASS2);
16404 non_portable_endpoint++;
16406 case '0': case '1': case '2': case '3': case '4':
16407 case '5': case '6': case '7':
16409 /* Take 1-3 octal digits */
16410 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
16411 numlen = (strict) ? 4 : 3;
16412 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
16413 RExC_parse += numlen;
16416 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16417 vFAIL("Need exactly 3 octal digits");
16419 else if (! SIZE_ONLY /* like \08, \178 */
16421 && RExC_parse < RExC_end
16422 && isDIGIT(*RExC_parse)
16423 && ckWARN(WARN_REGEXP))
16425 SAVEFREESV(RExC_rx_sv);
16426 reg_warn_non_literal_string(
16428 form_short_octal_warning(RExC_parse, numlen));
16429 (void)ReREFCNT_inc(RExC_rx_sv);
16432 non_portable_endpoint++;
16436 /* Allow \_ to not give an error */
16437 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
16439 vFAIL2("Unrecognized escape \\%c in character class",
16443 SAVEFREESV(RExC_rx_sv);
16444 ckWARN2reg(RExC_parse,
16445 "Unrecognized escape \\%c in character class passed through",
16447 (void)ReREFCNT_inc(RExC_rx_sv);
16451 } /* End of switch on char following backslash */
16452 } /* end of handling backslash escape sequences */
16454 /* Here, we have the current token in 'value' */
16456 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
16459 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
16460 * literal, as is the character that began the false range, i.e.
16461 * the 'a' in the examples */
16464 const int w = (RExC_parse >= rangebegin)
16465 ? RExC_parse - rangebegin
16469 "False [] range \"%" UTF8f "\"",
16470 UTF8fARG(UTF, w, rangebegin));
16473 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
16474 ckWARN2reg(RExC_parse,
16475 "False [] range \"%" UTF8f "\"",
16476 UTF8fARG(UTF, w, rangebegin));
16477 (void)ReREFCNT_inc(RExC_rx_sv);
16478 cp_list = add_cp_to_invlist(cp_list, '-');
16479 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
16484 range = 0; /* this was not a true range */
16485 element_count += 2; /* So counts for three values */
16488 classnum = namedclass_to_classnum(namedclass);
16490 if (LOC && namedclass < ANYOF_POSIXL_MAX
16491 #ifndef HAS_ISASCII
16492 && classnum != _CC_ASCII
16495 /* What the Posix classes (like \w, [:space:]) match in locale
16496 * isn't knowable under locale until actual match time. Room
16497 * must be reserved (one time per outer bracketed class) to
16498 * store such classes. The space will contain a bit for each
16499 * named class that is to be matched against. This isn't
16500 * needed for \p{} and pseudo-classes, as they are not affected
16501 * by locale, and hence are dealt with separately */
16502 if (! need_class) {
16505 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16508 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16510 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
16511 ANYOF_POSIXL_ZERO(ret);
16513 /* We can't change this into some other type of node
16514 * (unless this is the only element, in which case there
16515 * are nodes that mean exactly this) as has runtime
16517 optimizable = FALSE;
16520 /* Coverity thinks it is possible for this to be negative; both
16521 * jhi and khw think it's not, but be safer */
16522 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16523 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
16525 /* See if it already matches the complement of this POSIX
16527 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16528 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
16532 posixl_matches_all = TRUE;
16533 break; /* No need to continue. Since it matches both
16534 e.g., \w and \W, it matches everything, and the
16535 bracketed class can be optimized into qr/./s */
16538 /* Add this class to those that should be checked at runtime */
16539 ANYOF_POSIXL_SET(ret, namedclass);
16541 /* The above-Latin1 characters are not subject to locale rules.
16542 * Just add them, in the second pass, to the
16543 * unconditionally-matched list */
16545 SV* scratch_list = NULL;
16547 /* Get the list of the above-Latin1 code points this
16549 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
16550 PL_XPosix_ptrs[classnum],
16552 /* Odd numbers are complements, like
16553 * NDIGIT, NASCII, ... */
16554 namedclass % 2 != 0,
16556 /* Checking if 'cp_list' is NULL first saves an extra
16557 * clone. Its reference count will be decremented at the
16558 * next union, etc, or if this is the only instance, at the
16559 * end of the routine */
16561 cp_list = scratch_list;
16564 _invlist_union(cp_list, scratch_list, &cp_list);
16565 SvREFCNT_dec_NN(scratch_list);
16567 continue; /* Go get next character */
16570 else if (! SIZE_ONLY) {
16572 /* Here, not in pass1 (in that pass we skip calculating the
16573 * contents of this class), and is not /l, or is a POSIX class
16574 * for which /l doesn't matter (or is a Unicode property, which
16575 * is skipped here). */
16576 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
16577 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
16579 /* Here, should be \h, \H, \v, or \V. None of /d, /i
16580 * nor /l make a difference in what these match,
16581 * therefore we just add what they match to cp_list. */
16582 if (classnum != _CC_VERTSPACE) {
16583 assert( namedclass == ANYOF_HORIZWS
16584 || namedclass == ANYOF_NHORIZWS);
16586 /* It turns out that \h is just a synonym for
16588 classnum = _CC_BLANK;
16591 _invlist_union_maybe_complement_2nd(
16593 PL_XPosix_ptrs[classnum],
16594 namedclass % 2 != 0, /* Complement if odd
16595 (NHORIZWS, NVERTWS)
16600 else if ( UNI_SEMANTICS
16601 || classnum == _CC_ASCII
16602 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
16603 || classnum == _CC_XDIGIT)))
16605 /* We usually have to worry about /d and /a affecting what
16606 * POSIX classes match, with special code needed for /d
16607 * because we won't know until runtime what all matches.
16608 * But there is no extra work needed under /u, and
16609 * [:ascii:] is unaffected by /a and /d; and :digit: and
16610 * :xdigit: don't have runtime differences under /d. So we
16611 * can special case these, and avoid some extra work below,
16612 * and at runtime. */
16613 _invlist_union_maybe_complement_2nd(
16615 PL_XPosix_ptrs[classnum],
16616 namedclass % 2 != 0,
16619 else { /* Garden variety class. If is NUPPER, NALPHA, ...
16620 complement and use nposixes */
16621 SV** posixes_ptr = namedclass % 2 == 0
16624 _invlist_union_maybe_complement_2nd(
16626 PL_XPosix_ptrs[classnum],
16627 namedclass % 2 != 0,
16631 } /* end of namedclass \blah */
16633 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16635 /* If 'range' is set, 'value' is the ending of a range--check its
16636 * validity. (If value isn't a single code point in the case of a
16637 * range, we should have figured that out above in the code that
16638 * catches false ranges). Later, we will handle each individual code
16639 * point in the range. If 'range' isn't set, this could be the
16640 * beginning of a range, so check for that by looking ahead to see if
16641 * the next real character to be processed is the range indicator--the
16646 /* For unicode ranges, we have to test that the Unicode as opposed
16647 * to the native values are not decreasing. (Above 255, there is
16648 * no difference between native and Unicode) */
16649 if (unicode_range && prevvalue < 255 && value < 255) {
16650 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
16651 goto backwards_range;
16656 if (prevvalue > value) /* b-a */ {
16661 w = RExC_parse - rangebegin;
16663 "Invalid [] range \"%" UTF8f "\"",
16664 UTF8fARG(UTF, w, rangebegin));
16665 NOT_REACHED; /* NOTREACHED */
16669 prevvalue = value; /* save the beginning of the potential range */
16670 if (! stop_at_1 /* Can't be a range if parsing just one thing */
16671 && *RExC_parse == '-')
16673 char* next_char_ptr = RExC_parse + 1;
16675 /* Get the next real char after the '-' */
16676 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
16678 /* If the '-' is at the end of the class (just before the ']',
16679 * it is a literal minus; otherwise it is a range */
16680 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
16681 RExC_parse = next_char_ptr;
16683 /* a bad range like \w-, [:word:]- ? */
16684 if (namedclass > OOB_NAMEDCLASS) {
16685 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
16686 const int w = RExC_parse >= rangebegin
16687 ? RExC_parse - rangebegin
16690 vFAIL4("False [] range \"%*.*s\"",
16695 "False [] range \"%*.*s\"",
16700 cp_list = add_cp_to_invlist(cp_list, '-');
16704 range = 1; /* yeah, it's a range! */
16705 continue; /* but do it the next time */
16710 if (namedclass > OOB_NAMEDCLASS) {
16714 /* Here, we have a single value this time through the loop, and
16715 * <prevvalue> is the beginning of the range, if any; or <value> if
16718 /* non-Latin1 code point implies unicode semantics. Must be set in
16719 * pass1 so is there for the whole of pass 2 */
16721 REQUIRE_UNI_RULES(flagp, NULL);
16724 /* Ready to process either the single value, or the completed range.
16725 * For single-valued non-inverted ranges, we consider the possibility
16726 * of multi-char folds. (We made a conscious decision to not do this
16727 * for the other cases because it can often lead to non-intuitive
16728 * results. For example, you have the peculiar case that:
16729 * "s s" =~ /^[^\xDF]+$/i => Y
16730 * "ss" =~ /^[^\xDF]+$/i => N
16732 * See [perl #89750] */
16733 if (FOLD && allow_multi_folds && value == prevvalue) {
16734 if (value == LATIN_SMALL_LETTER_SHARP_S
16735 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
16738 /* Here <value> is indeed a multi-char fold. Get what it is */
16740 U8 foldbuf[UTF8_MAXBYTES_CASE];
16743 UV folded = _to_uni_fold_flags(
16747 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
16748 ? FOLD_FLAGS_NOMIX_ASCII
16752 /* Here, <folded> should be the first character of the
16753 * multi-char fold of <value>, with <foldbuf> containing the
16754 * whole thing. But, if this fold is not allowed (because of
16755 * the flags), <fold> will be the same as <value>, and should
16756 * be processed like any other character, so skip the special
16758 if (folded != value) {
16760 /* Skip if we are recursed, currently parsing the class
16761 * again. Otherwise add this character to the list of
16762 * multi-char folds. */
16763 if (! RExC_in_multi_char_class) {
16764 STRLEN cp_count = utf8_length(foldbuf,
16765 foldbuf + foldlen);
16766 SV* multi_fold = sv_2mortal(newSVpvs(""));
16768 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
16771 = add_multi_match(multi_char_matches,
16777 /* This element should not be processed further in this
16780 value = save_value;
16781 prevvalue = save_prevvalue;
16787 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
16790 /* If the range starts above 255, everything is portable and
16791 * likely to be so for any forseeable character set, so don't
16793 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
16794 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
16796 else if (prevvalue != value) {
16798 /* Under strict, ranges that stop and/or end in an ASCII
16799 * printable should have each end point be a portable value
16800 * for it (preferably like 'A', but we don't warn if it is
16801 * a (portable) Unicode name or code point), and the range
16802 * must be be all digits or all letters of the same case.
16803 * Otherwise, the range is non-portable and unclear as to
16804 * what it contains */
16805 if ((isPRINT_A(prevvalue) || isPRINT_A(value))
16806 && (non_portable_endpoint
16807 || ! ((isDIGIT_A(prevvalue) && isDIGIT_A(value))
16808 || (isLOWER_A(prevvalue) && isLOWER_A(value))
16809 || (isUPPER_A(prevvalue) && isUPPER_A(value)))))
16811 vWARN(RExC_parse, "Ranges of ASCII printables should be some subset of \"0-9\", \"A-Z\", or \"a-z\"");
16813 else if (prevvalue >= 0x660) { /* ARABIC_INDIC_DIGIT_ZERO */
16815 /* But the nature of Unicode and languages mean we
16816 * can't do the same checks for above-ASCII ranges,
16817 * except in the case of digit ones. These should
16818 * contain only digits from the same group of 10. The
16819 * ASCII case is handled just above. 0x660 is the
16820 * first digit character beyond ASCII. Hence here, the
16821 * range could be a range of digits. Find out. */
16822 IV index_start = _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
16824 IV index_final = _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
16827 /* If the range start and final points are in the same
16828 * inversion list element, it means that either both
16829 * are not digits, or both are digits in a consecutive
16830 * sequence of digits. (So far, Unicode has kept all
16831 * such sequences as distinct groups of 10, but assert
16832 * to make sure). If the end points are not in the
16833 * same element, neither should be a digit. */
16834 if (index_start == index_final) {
16835 assert(! ELEMENT_RANGE_MATCHES_INVLIST(index_start)
16836 || (invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start+1]
16837 - invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
16839 /* But actually Unicode did have one group of 11
16840 * 'digits' in 5.2, so in case we are operating
16841 * on that version, let that pass */
16842 || (invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start+1]
16843 - invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
16845 && invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
16849 else if ((index_start >= 0
16850 && ELEMENT_RANGE_MATCHES_INVLIST(index_start))
16851 || (index_final >= 0
16852 && ELEMENT_RANGE_MATCHES_INVLIST(index_final)))
16854 vWARN(RExC_parse, "Ranges of digits should be from the same group of 10");
16859 if ((! range || prevvalue == value) && non_portable_endpoint) {
16860 if (isPRINT_A(value)) {
16863 if (isBACKSLASHED_PUNCT(value)) {
16864 literal[d++] = '\\';
16866 literal[d++] = (char) value;
16867 literal[d++] = '\0';
16870 "\"%.*s\" is more clearly written simply as \"%s\"",
16871 (int) (RExC_parse - rangebegin),
16876 else if isMNEMONIC_CNTRL(value) {
16878 "\"%.*s\" is more clearly written simply as \"%s\"",
16879 (int) (RExC_parse - rangebegin),
16881 cntrl_to_mnemonic((U8) value)
16887 /* Deal with this element of the class */
16891 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16894 /* On non-ASCII platforms, for ranges that span all of 0..255, and
16895 * ones that don't require special handling, we can just add the
16896 * range like we do for ASCII platforms */
16897 if ((UNLIKELY(prevvalue == 0) && value >= 255)
16898 || ! (prevvalue < 256
16900 || (! non_portable_endpoint
16901 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
16902 || (isUPPER_A(prevvalue)
16903 && isUPPER_A(value)))))))
16905 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16909 /* Here, requires special handling. This can be because it is
16910 * a range whose code points are considered to be Unicode, and
16911 * so must be individually translated into native, or because
16912 * its a subrange of 'A-Z' or 'a-z' which each aren't
16913 * contiguous in EBCDIC, but we have defined them to include
16914 * only the "expected" upper or lower case ASCII alphabetics.
16915 * Subranges above 255 are the same in native and Unicode, so
16916 * can be added as a range */
16917 U8 start = NATIVE_TO_LATIN1(prevvalue);
16919 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
16920 for (j = start; j <= end; j++) {
16921 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
16924 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16931 range = 0; /* this range (if it was one) is done now */
16932 } /* End of loop through all the text within the brackets */
16935 if ( posix_warnings && av_tindex_nomg(posix_warnings) >= 0) {
16936 output_or_return_posix_warnings(pRExC_state, posix_warnings,
16937 return_posix_warnings);
16940 /* If anything in the class expands to more than one character, we have to
16941 * deal with them by building up a substitute parse string, and recursively
16942 * calling reg() on it, instead of proceeding */
16943 if (multi_char_matches) {
16944 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
16947 char *save_end = RExC_end;
16948 char *save_parse = RExC_parse;
16949 char *save_start = RExC_start;
16950 STRLEN prefix_end = 0; /* We copy the character class after a
16951 prefix supplied here. This is the size
16952 + 1 of that prefix */
16953 bool first_time = TRUE; /* First multi-char occurrence doesn't get
16958 assert(RExC_precomp_adj == 0); /* Only one level of recursion allowed */
16960 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
16961 because too confusing */
16963 sv_catpv(substitute_parse, "(?:");
16967 /* Look at the longest folds first */
16968 for (cp_count = av_tindex_nomg(multi_char_matches);
16973 if (av_exists(multi_char_matches, cp_count)) {
16974 AV** this_array_ptr;
16977 this_array_ptr = (AV**) av_fetch(multi_char_matches,
16979 while ((this_sequence = av_pop(*this_array_ptr)) !=
16982 if (! first_time) {
16983 sv_catpv(substitute_parse, "|");
16985 first_time = FALSE;
16987 sv_catpv(substitute_parse, SvPVX(this_sequence));
16992 /* If the character class contains anything else besides these
16993 * multi-character folds, have to include it in recursive parsing */
16994 if (element_count) {
16995 sv_catpv(substitute_parse, "|[");
16996 prefix_end = SvCUR(substitute_parse);
16997 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
16999 /* Put in a closing ']' only if not going off the end, as otherwise
17000 * we are adding something that really isn't there */
17001 if (RExC_parse < RExC_end) {
17002 sv_catpv(substitute_parse, "]");
17006 sv_catpv(substitute_parse, ")");
17009 /* This is a way to get the parse to skip forward a whole named
17010 * sequence instead of matching the 2nd character when it fails the
17012 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
17016 /* Set up the data structure so that any errors will be properly
17017 * reported. See the comments at the definition of
17018 * REPORT_LOCATION_ARGS for details */
17019 RExC_precomp_adj = orig_parse - RExC_precomp;
17020 RExC_start = RExC_parse = SvPV(substitute_parse, len);
17021 RExC_adjusted_start = RExC_start + prefix_end;
17022 RExC_end = RExC_parse + len;
17023 RExC_in_multi_char_class = 1;
17024 RExC_emit = (regnode *)orig_emit;
17026 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17028 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PASS1|NEED_UTF8);
17030 /* And restore so can parse the rest of the pattern */
17031 RExC_parse = save_parse;
17032 RExC_start = RExC_adjusted_start = save_start;
17033 RExC_precomp_adj = 0;
17034 RExC_end = save_end;
17035 RExC_in_multi_char_class = 0;
17036 SvREFCNT_dec_NN(multi_char_matches);
17040 /* Here, we've gone through the entire class and dealt with multi-char
17041 * folds. We are now in a position that we can do some checks to see if we
17042 * can optimize this ANYOF node into a simpler one, even in Pass 1.
17043 * Currently we only do two checks:
17044 * 1) is in the unlikely event that the user has specified both, eg. \w and
17045 * \W under /l, then the class matches everything. (This optimization
17046 * is done only to make the optimizer code run later work.)
17047 * 2) if the character class contains only a single element (including a
17048 * single range), we see if there is an equivalent node for it.
17049 * Other checks are possible */
17051 && ! ret_invlist /* Can't optimize if returning the constructed
17053 && (UNLIKELY(posixl_matches_all) || element_count == 1))
17058 if (UNLIKELY(posixl_matches_all)) {
17061 else if (namedclass > OOB_NAMEDCLASS) { /* this is a single named
17062 class, like \w or [:digit:]
17065 /* All named classes are mapped into POSIXish nodes, with its FLAG
17066 * argument giving which class it is */
17067 switch ((I32)namedclass) {
17068 case ANYOF_UNIPROP:
17071 /* These don't depend on the charset modifiers. They always
17072 * match under /u rules */
17073 case ANYOF_NHORIZWS:
17074 case ANYOF_HORIZWS:
17075 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
17078 case ANYOF_NVERTWS:
17083 /* The actual POSIXish node for all the rest depends on the
17084 * charset modifier. The ones in the first set depend only on
17085 * ASCII or, if available on this platform, also locale */
17089 op = (LOC) ? POSIXL : POSIXA;
17095 /* The following don't have any matches in the upper Latin1
17096 * range, hence /d is equivalent to /u for them. Making it /u
17097 * saves some branches at runtime */
17101 case ANYOF_NXDIGIT:
17102 if (! DEPENDS_SEMANTICS) {
17103 goto treat_as_default;
17109 /* The following change to CASED under /i */
17115 namedclass = ANYOF_CASED + (namedclass % 2);
17119 /* The rest have more possibilities depending on the charset.
17120 * We take advantage of the enum ordering of the charset
17121 * modifiers to get the exact node type, */
17124 op = POSIXD + get_regex_charset(RExC_flags);
17125 if (op > POSIXA) { /* /aa is same as /a */
17130 /* The odd numbered ones are the complements of the
17131 * next-lower even number one */
17132 if (namedclass % 2 == 1) {
17136 arg = namedclass_to_classnum(namedclass);
17140 else if (value == prevvalue) {
17142 /* Here, the class consists of just a single code point */
17145 if (! LOC && value == '\n') {
17146 op = REG_ANY; /* Optimize [^\n] */
17147 *flagp |= HASWIDTH|SIMPLE;
17151 else if (value < 256 || UTF) {
17153 /* Optimize a single value into an EXACTish node, but not if it
17154 * would require converting the pattern to UTF-8. */
17155 op = compute_EXACTish(pRExC_state);
17157 } /* Otherwise is a range */
17158 else if (! LOC) { /* locale could vary these */
17159 if (prevvalue == '0') {
17160 if (value == '9') {
17165 else if (! FOLD || ASCII_FOLD_RESTRICTED) {
17166 /* We can optimize A-Z or a-z, but not if they could match
17167 * something like the KELVIN SIGN under /i. */
17168 if (prevvalue == 'A') {
17171 && ! non_portable_endpoint
17174 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
17178 else if (prevvalue == 'a') {
17181 && ! non_portable_endpoint
17184 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
17191 /* Here, we have changed <op> away from its initial value iff we found
17192 * an optimization */
17195 /* Throw away this ANYOF regnode, and emit the calculated one,
17196 * which should correspond to the beginning, not current, state of
17198 const char * cur_parse = RExC_parse;
17199 RExC_parse = (char *)orig_parse;
17203 /* To get locale nodes to not use the full ANYOF size would
17204 * require moving the code above that writes the portions
17205 * of it that aren't in other nodes to after this point.
17206 * e.g. ANYOF_POSIXL_SET */
17207 RExC_size = orig_size;
17211 RExC_emit = (regnode *)orig_emit;
17212 if (PL_regkind[op] == POSIXD) {
17213 if (op == POSIXL) {
17214 RExC_contains_locale = 1;
17217 op += NPOSIXD - POSIXD;
17222 ret = reg_node(pRExC_state, op);
17224 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17228 *flagp |= HASWIDTH|SIMPLE;
17230 else if (PL_regkind[op] == EXACT) {
17231 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17232 TRUE /* downgradable to EXACT */
17236 RExC_parse = (char *) cur_parse;
17238 SvREFCNT_dec(posixes);
17239 SvREFCNT_dec(nposixes);
17240 SvREFCNT_dec(simple_posixes);
17241 SvREFCNT_dec(cp_list);
17242 SvREFCNT_dec(cp_foldable_list);
17249 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
17251 /* If folding, we calculate all characters that could fold to or from the
17252 * ones already on the list */
17253 if (cp_foldable_list) {
17255 UV start, end; /* End points of code point ranges */
17257 SV* fold_intersection = NULL;
17260 /* Our calculated list will be for Unicode rules. For locale
17261 * matching, we have to keep a separate list that is consulted at
17262 * runtime only when the locale indicates Unicode rules. For
17263 * non-locale, we just use the general list */
17265 use_list = &only_utf8_locale_list;
17268 use_list = &cp_list;
17271 /* Only the characters in this class that participate in folds need
17272 * be checked. Get the intersection of this class and all the
17273 * possible characters that are foldable. This can quickly narrow
17274 * down a large class */
17275 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
17276 &fold_intersection);
17278 /* The folds for all the Latin1 characters are hard-coded into this
17279 * program, but we have to go out to disk to get the others. */
17280 if (invlist_highest(cp_foldable_list) >= 256) {
17282 /* This is a hash that for a particular fold gives all
17283 * characters that are involved in it */
17284 if (! PL_utf8_foldclosures) {
17285 _load_PL_utf8_foldclosures();
17289 /* Now look at the foldable characters in this class individually */
17290 invlist_iterinit(fold_intersection);
17291 while (invlist_iternext(fold_intersection, &start, &end)) {
17294 /* Look at every character in the range */
17295 for (j = start; j <= end; j++) {
17296 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17302 if (IS_IN_SOME_FOLD_L1(j)) {
17304 /* ASCII is always matched; non-ASCII is matched
17305 * only under Unicode rules (which could happen
17306 * under /l if the locale is a UTF-8 one */
17307 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
17308 *use_list = add_cp_to_invlist(*use_list,
17309 PL_fold_latin1[j]);
17312 has_upper_latin1_only_utf8_matches
17313 = add_cp_to_invlist(
17314 has_upper_latin1_only_utf8_matches,
17315 PL_fold_latin1[j]);
17319 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
17320 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
17322 add_above_Latin1_folds(pRExC_state,
17329 /* Here is an above Latin1 character. We don't have the
17330 * rules hard-coded for it. First, get its fold. This is
17331 * the simple fold, as the multi-character folds have been
17332 * handled earlier and separated out */
17333 _to_uni_fold_flags(j, foldbuf, &foldlen,
17334 (ASCII_FOLD_RESTRICTED)
17335 ? FOLD_FLAGS_NOMIX_ASCII
17338 /* Single character fold of above Latin1. Add everything in
17339 * its fold closure to the list that this node should match.
17340 * The fold closures data structure is a hash with the keys
17341 * being the UTF-8 of every character that is folded to, like
17342 * 'k', and the values each an array of all code points that
17343 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
17344 * Multi-character folds are not included */
17345 if ((listp = hv_fetch(PL_utf8_foldclosures,
17346 (char *) foldbuf, foldlen, FALSE)))
17348 AV* list = (AV*) *listp;
17350 for (k = 0; k <= av_tindex_nomg(list); k++) {
17351 SV** c_p = av_fetch(list, k, FALSE);
17357 /* /aa doesn't allow folds between ASCII and non- */
17358 if ((ASCII_FOLD_RESTRICTED
17359 && (isASCII(c) != isASCII(j))))
17364 /* Folds under /l which cross the 255/256 boundary
17365 * are added to a separate list. (These are valid
17366 * only when the locale is UTF-8.) */
17367 if (c < 256 && LOC) {
17368 *use_list = add_cp_to_invlist(*use_list, c);
17372 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
17374 cp_list = add_cp_to_invlist(cp_list, c);
17377 /* Similarly folds involving non-ascii Latin1
17378 * characters under /d are added to their list */
17379 has_upper_latin1_only_utf8_matches
17380 = add_cp_to_invlist(
17381 has_upper_latin1_only_utf8_matches,
17388 SvREFCNT_dec_NN(fold_intersection);
17391 /* Now that we have finished adding all the folds, there is no reason
17392 * to keep the foldable list separate */
17393 _invlist_union(cp_list, cp_foldable_list, &cp_list);
17394 SvREFCNT_dec_NN(cp_foldable_list);
17397 /* And combine the result (if any) with any inversion lists from posix
17398 * classes. The lists are kept separate up to now because we don't want to
17399 * fold the classes (folding of those is automatically handled by the swash
17400 * fetching code) */
17401 if (simple_posixes) { /* These are the classes known to be unaffected by
17404 _invlist_union(cp_list, simple_posixes, &cp_list);
17405 SvREFCNT_dec_NN(simple_posixes);
17408 cp_list = simple_posixes;
17411 if (posixes || nposixes) {
17413 /* We have to adjust /a and /aa */
17414 if (AT_LEAST_ASCII_RESTRICTED) {
17416 /* Under /a and /aa, nothing above ASCII matches these */
17418 _invlist_intersection(posixes,
17419 PL_XPosix_ptrs[_CC_ASCII],
17423 /* Under /a and /aa, everything above ASCII matches these
17426 _invlist_union_complement_2nd(nposixes,
17427 PL_XPosix_ptrs[_CC_ASCII],
17432 if (! DEPENDS_SEMANTICS) {
17434 /* For everything but /d, we can just add the current 'posixes' and
17435 * 'nposixes' to the main list */
17438 _invlist_union(cp_list, posixes, &cp_list);
17439 SvREFCNT_dec_NN(posixes);
17447 _invlist_union(cp_list, nposixes, &cp_list);
17448 SvREFCNT_dec_NN(nposixes);
17451 cp_list = nposixes;
17456 /* Under /d, things like \w match upper Latin1 characters only if
17457 * the target string is in UTF-8. But things like \W match all the
17458 * upper Latin1 characters if the target string is not in UTF-8.
17460 * Handle the case where there something like \W separately */
17462 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1);
17464 /* A complemented posix class matches all upper Latin1
17465 * characters if not in UTF-8. And it matches just certain
17466 * ones when in UTF-8. That means those certain ones are
17467 * matched regardless, so can just be added to the
17468 * unconditional list */
17470 _invlist_union(cp_list, nposixes, &cp_list);
17471 SvREFCNT_dec_NN(nposixes);
17475 cp_list = nposixes;
17478 /* Likewise for 'posixes' */
17479 _invlist_union(posixes, cp_list, &cp_list);
17481 /* Likewise for anything else in the range that matched only
17483 if (has_upper_latin1_only_utf8_matches) {
17484 _invlist_union(cp_list,
17485 has_upper_latin1_only_utf8_matches,
17487 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17488 has_upper_latin1_only_utf8_matches = NULL;
17491 /* If we don't match all the upper Latin1 characters regardless
17492 * of UTF-8ness, we have to set a flag to match the rest when
17494 _invlist_subtract(only_non_utf8_list, cp_list,
17495 &only_non_utf8_list);
17496 if (_invlist_len(only_non_utf8_list) != 0) {
17497 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17501 /* Here there were no complemented posix classes. That means
17502 * the upper Latin1 characters in 'posixes' match only when the
17503 * target string is in UTF-8. So we have to add them to the
17504 * list of those types of code points, while adding the
17505 * remainder to the unconditional list.
17507 * First calculate what they are */
17508 SV* nonascii_but_latin1_properties = NULL;
17509 _invlist_intersection(posixes, PL_UpperLatin1,
17510 &nonascii_but_latin1_properties);
17512 /* And add them to the final list of such characters. */
17513 _invlist_union(has_upper_latin1_only_utf8_matches,
17514 nonascii_but_latin1_properties,
17515 &has_upper_latin1_only_utf8_matches);
17517 /* Remove them from what now becomes the unconditional list */
17518 _invlist_subtract(posixes, nonascii_but_latin1_properties,
17521 /* And add those unconditional ones to the final list */
17523 _invlist_union(cp_list, posixes, &cp_list);
17524 SvREFCNT_dec_NN(posixes);
17531 SvREFCNT_dec(nonascii_but_latin1_properties);
17533 /* Get rid of any characters that we now know are matched
17534 * unconditionally from the conditional list, which may make
17535 * that list empty */
17536 _invlist_subtract(has_upper_latin1_only_utf8_matches,
17538 &has_upper_latin1_only_utf8_matches);
17539 if (_invlist_len(has_upper_latin1_only_utf8_matches) == 0) {
17540 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17541 has_upper_latin1_only_utf8_matches = NULL;
17547 /* And combine the result (if any) with any inversion list from properties.
17548 * The lists are kept separate up to now so that we can distinguish the two
17549 * in regards to matching above-Unicode. A run-time warning is generated
17550 * if a Unicode property is matched against a non-Unicode code point. But,
17551 * we allow user-defined properties to match anything, without any warning,
17552 * and we also suppress the warning if there is a portion of the character
17553 * class that isn't a Unicode property, and which matches above Unicode, \W
17554 * or [\x{110000}] for example.
17555 * (Note that in this case, unlike the Posix one above, there is no
17556 * <has_upper_latin1_only_utf8_matches>, because having a Unicode property
17557 * forces Unicode semantics */
17561 /* If it matters to the final outcome, see if a non-property
17562 * component of the class matches above Unicode. If so, the
17563 * warning gets suppressed. This is true even if just a single
17564 * such code point is specified, as, though not strictly correct if
17565 * another such code point is matched against, the fact that they
17566 * are using above-Unicode code points indicates they should know
17567 * the issues involved */
17569 warn_super = ! (invert
17570 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
17573 _invlist_union(properties, cp_list, &cp_list);
17574 SvREFCNT_dec_NN(properties);
17577 cp_list = properties;
17582 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17584 /* Because an ANYOF node is the only one that warns, this node
17585 * can't be optimized into something else */
17586 optimizable = FALSE;
17590 /* Here, we have calculated what code points should be in the character
17593 * Now we can see about various optimizations. Fold calculation (which we
17594 * did above) needs to take place before inversion. Otherwise /[^k]/i
17595 * would invert to include K, which under /i would match k, which it
17596 * shouldn't. Therefore we can't invert folded locale now, as it won't be
17597 * folded until runtime */
17599 /* If we didn't do folding, it's because some information isn't available
17600 * until runtime; set the run-time fold flag for these. (We don't have to
17601 * worry about properties folding, as that is taken care of by the swash
17602 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
17603 * locales, or the class matches at least one 0-255 range code point */
17606 /* Some things on the list might be unconditionally included because of
17607 * other components. Remove them, and clean up the list if it goes to
17609 if (only_utf8_locale_list && cp_list) {
17610 _invlist_subtract(only_utf8_locale_list, cp_list,
17611 &only_utf8_locale_list);
17613 if (_invlist_len(only_utf8_locale_list) == 0) {
17614 SvREFCNT_dec_NN(only_utf8_locale_list);
17615 only_utf8_locale_list = NULL;
17618 if (only_utf8_locale_list) {
17621 |ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17623 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
17625 invlist_iterinit(cp_list);
17626 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
17627 ANYOF_FLAGS(ret) |= ANYOFL_FOLD;
17629 invlist_iterfinish(cp_list);
17632 else if ( DEPENDS_SEMANTICS
17633 && ( has_upper_latin1_only_utf8_matches
17634 || (ANYOF_FLAGS(ret) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
17637 optimizable = FALSE;
17641 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
17642 * at compile time. Besides not inverting folded locale now, we can't
17643 * invert if there are things such as \w, which aren't known until runtime
17647 && OP(ret) != ANYOFD
17648 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
17649 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17651 _invlist_invert(cp_list);
17653 /* Any swash can't be used as-is, because we've inverted things */
17655 SvREFCNT_dec_NN(swash);
17659 /* Clear the invert flag since have just done it here */
17666 *ret_invlist = cp_list;
17667 SvREFCNT_dec(swash);
17669 /* Discard the generated node */
17671 RExC_size = orig_size;
17674 RExC_emit = orig_emit;
17679 /* Some character classes are equivalent to other nodes. Such nodes take
17680 * up less room and generally fewer operations to execute than ANYOF nodes.
17681 * Above, we checked for and optimized into some such equivalents for
17682 * certain common classes that are easy to test. Getting to this point in
17683 * the code means that the class didn't get optimized there. Since this
17684 * code is only executed in Pass 2, it is too late to save space--it has
17685 * been allocated in Pass 1, and currently isn't given back. But turning
17686 * things into an EXACTish node can allow the optimizer to join it to any
17687 * adjacent such nodes. And if the class is equivalent to things like /./,
17688 * expensive run-time swashes can be avoided. Now that we have more
17689 * complete information, we can find things necessarily missed by the
17690 * earlier code. Another possible "optimization" that isn't done is that
17691 * something like [Ee] could be changed into an EXACTFU. khw tried this
17692 * and found that the ANYOF is faster, including for code points not in the
17693 * bitmap. This still might make sense to do, provided it got joined with
17694 * an adjacent node(s) to create a longer EXACTFU one. This could be
17695 * accomplished by creating a pseudo ANYOF_EXACTFU node type that the join
17696 * routine would know is joinable. If that didn't happen, the node type
17697 * could then be made a straight ANYOF */
17699 if (optimizable && cp_list && ! invert) {
17701 U8 op = END; /* The optimzation node-type */
17702 int posix_class = -1; /* Illegal value */
17703 const char * cur_parse= RExC_parse;
17705 invlist_iterinit(cp_list);
17706 if (! invlist_iternext(cp_list, &start, &end)) {
17708 /* Here, the list is empty. This happens, for example, when a
17709 * Unicode property that doesn't match anything is the only element
17710 * in the character class (perluniprops.pod notes such properties).
17713 *flagp |= HASWIDTH|SIMPLE;
17715 else if (start == end) { /* The range is a single code point */
17716 if (! invlist_iternext(cp_list, &start, &end)
17718 /* Don't do this optimization if it would require changing
17719 * the pattern to UTF-8 */
17720 && (start < 256 || UTF))
17722 /* Here, the list contains a single code point. Can optimize
17723 * into an EXACTish node */
17734 /* A locale node under folding with one code point can be
17735 * an EXACTFL, as its fold won't be calculated until
17741 /* Here, we are generally folding, but there is only one
17742 * code point to match. If we have to, we use an EXACT
17743 * node, but it would be better for joining with adjacent
17744 * nodes in the optimization pass if we used the same
17745 * EXACTFish node that any such are likely to be. We can
17746 * do this iff the code point doesn't participate in any
17747 * folds. For example, an EXACTF of a colon is the same as
17748 * an EXACT one, since nothing folds to or from a colon. */
17750 if (IS_IN_SOME_FOLD_L1(value)) {
17755 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
17760 /* If we haven't found the node type, above, it means we
17761 * can use the prevailing one */
17763 op = compute_EXACTish(pRExC_state);
17767 } /* End of first range contains just a single code point */
17768 else if (start == 0) {
17769 if (end == UV_MAX) {
17771 *flagp |= HASWIDTH|SIMPLE;
17774 else if (end == '\n' - 1
17775 && invlist_iternext(cp_list, &start, &end)
17776 && start == '\n' + 1 && end == UV_MAX)
17779 *flagp |= HASWIDTH|SIMPLE;
17783 invlist_iterfinish(cp_list);
17786 const UV cp_list_len = _invlist_len(cp_list);
17787 const UV* cp_list_array = invlist_array(cp_list);
17789 /* Here, didn't find an optimization. See if this matches any of
17790 * the POSIX classes. These run slightly faster for above-Unicode
17791 * code points, so don't bother with POSIXA ones nor the 2 that
17792 * have no above-Unicode matches. We can avoid these checks unless
17793 * the ANYOF matches at least as high as the lowest POSIX one
17794 * (which was manually found to be \v. The actual code point may
17795 * increase in later Unicode releases, if a higher code point is
17796 * assigned to be \v, but this code will never break. It would
17797 * just mean we could execute the checks for posix optimizations
17798 * unnecessarily) */
17800 if (cp_list_array[cp_list_len-1] > 0x2029) {
17801 for (posix_class = 0;
17802 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
17806 if (posix_class == _CC_ASCII || posix_class == _CC_CNTRL) {
17809 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
17811 /* Check if matches normal or inverted */
17812 if (_invlistEQ(cp_list,
17813 PL_XPosix_ptrs[posix_class],
17816 op = (try_inverted)
17819 *flagp |= HASWIDTH|SIMPLE;
17829 RExC_parse = (char *)orig_parse;
17830 RExC_emit = (regnode *)orig_emit;
17832 if (regarglen[op]) {
17833 ret = reganode(pRExC_state, op, 0);
17835 ret = reg_node(pRExC_state, op);
17838 RExC_parse = (char *)cur_parse;
17840 if (PL_regkind[op] == EXACT) {
17841 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17842 TRUE /* downgradable to EXACT */
17845 else if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17846 FLAGS(ret) = posix_class;
17849 SvREFCNT_dec_NN(cp_list);
17854 /* Here, <cp_list> contains all the code points we can determine at
17855 * compile time that match under all conditions. Go through it, and
17856 * for things that belong in the bitmap, put them there, and delete from
17857 * <cp_list>. While we are at it, see if everything above 255 is in the
17858 * list, and if so, set a flag to speed up execution */
17860 populate_ANYOF_from_invlist(ret, &cp_list);
17863 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
17866 /* Here, the bitmap has been populated with all the Latin1 code points that
17867 * always match. Can now add to the overall list those that match only
17868 * when the target string is UTF-8 (<has_upper_latin1_only_utf8_matches>).
17870 if (has_upper_latin1_only_utf8_matches) {
17872 _invlist_union(cp_list,
17873 has_upper_latin1_only_utf8_matches,
17875 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17878 cp_list = has_upper_latin1_only_utf8_matches;
17880 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
17883 /* If there is a swash and more than one element, we can't use the swash in
17884 * the optimization below. */
17885 if (swash && element_count > 1) {
17886 SvREFCNT_dec_NN(swash);
17890 /* Note that the optimization of using 'swash' if it is the only thing in
17891 * the class doesn't have us change swash at all, so it can include things
17892 * that are also in the bitmap; otherwise we have purposely deleted that
17893 * duplicate information */
17894 set_ANYOF_arg(pRExC_state, ret, cp_list,
17895 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17897 only_utf8_locale_list,
17898 swash, has_user_defined_property);
17900 *flagp |= HASWIDTH|SIMPLE;
17902 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
17903 RExC_contains_locale = 1;
17909 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
17912 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
17913 regnode* const node,
17915 SV* const runtime_defns,
17916 SV* const only_utf8_locale_list,
17918 const bool has_user_defined_property)
17920 /* Sets the arg field of an ANYOF-type node 'node', using information about
17921 * the node passed-in. If there is nothing outside the node's bitmap, the
17922 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
17923 * the count returned by add_data(), having allocated and stored an array,
17924 * av, that that count references, as follows:
17925 * av[0] stores the character class description in its textual form.
17926 * This is used later (regexec.c:Perl_regclass_swash()) to
17927 * initialize the appropriate swash, and is also useful for dumping
17928 * the regnode. This is set to &PL_sv_undef if the textual
17929 * description is not needed at run-time (as happens if the other
17930 * elements completely define the class)
17931 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
17932 * computed from av[0]. But if no further computation need be done,
17933 * the swash is stored here now (and av[0] is &PL_sv_undef).
17934 * av[2] stores the inversion list of code points that match only if the
17935 * current locale is UTF-8
17936 * av[3] stores the cp_list inversion list for use in addition or instead
17937 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
17938 * (Otherwise everything needed is already in av[0] and av[1])
17939 * av[4] is set if any component of the class is from a user-defined
17940 * property; used only if av[3] exists */
17944 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
17946 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
17947 assert(! (ANYOF_FLAGS(node)
17948 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
17949 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
17952 AV * const av = newAV();
17955 av_store(av, 0, (runtime_defns)
17956 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
17959 av_store(av, 1, swash);
17960 SvREFCNT_dec_NN(cp_list);
17963 av_store(av, 1, &PL_sv_undef);
17965 av_store(av, 3, cp_list);
17966 av_store(av, 4, newSVuv(has_user_defined_property));
17970 if (only_utf8_locale_list) {
17971 av_store(av, 2, only_utf8_locale_list);
17974 av_store(av, 2, &PL_sv_undef);
17977 rv = newRV_noinc(MUTABLE_SV(av));
17978 n = add_data(pRExC_state, STR_WITH_LEN("s"));
17979 RExC_rxi->data->data[n] = (void*)rv;
17984 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
17986 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
17987 const regnode* node,
17990 SV** only_utf8_locale_ptr,
17991 SV** output_invlist)
17994 /* For internal core use only.
17995 * Returns the swash for the input 'node' in the regex 'prog'.
17996 * If <doinit> is 'true', will attempt to create the swash if not already
17998 * If <listsvp> is non-null, will return the printable contents of the
17999 * swash. This can be used to get debugging information even before the
18000 * swash exists, by calling this function with 'doinit' set to false, in
18001 * which case the components that will be used to eventually create the
18002 * swash are returned (in a printable form).
18003 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
18004 * store an inversion list of code points that should match only if the
18005 * execution-time locale is a UTF-8 one.
18006 * If <output_invlist> is not NULL, it is where this routine is to store an
18007 * inversion list of the code points that would be instead returned in
18008 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
18009 * when this parameter is used, is just the non-code point data that
18010 * will go into creating the swash. This currently should be just
18011 * user-defined properties whose definitions were not known at compile
18012 * time. Using this parameter allows for easier manipulation of the
18013 * swash's data by the caller. It is illegal to call this function with
18014 * this parameter set, but not <listsvp>
18016 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
18017 * that, in spite of this function's name, the swash it returns may include
18018 * the bitmap data as well */
18021 SV *si = NULL; /* Input swash initialization string */
18022 SV* invlist = NULL;
18024 RXi_GET_DECL(prog,progi);
18025 const struct reg_data * const data = prog ? progi->data : NULL;
18027 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
18028 assert(! output_invlist || listsvp);
18030 if (data && data->count) {
18031 const U32 n = ARG(node);
18033 if (data->what[n] == 's') {
18034 SV * const rv = MUTABLE_SV(data->data[n]);
18035 AV * const av = MUTABLE_AV(SvRV(rv));
18036 SV **const ary = AvARRAY(av);
18037 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
18039 si = *ary; /* ary[0] = the string to initialize the swash with */
18041 if (av_tindex_nomg(av) >= 2) {
18042 if (only_utf8_locale_ptr
18044 && ary[2] != &PL_sv_undef)
18046 *only_utf8_locale_ptr = ary[2];
18049 assert(only_utf8_locale_ptr);
18050 *only_utf8_locale_ptr = NULL;
18053 /* Elements 3 and 4 are either both present or both absent. [3]
18054 * is any inversion list generated at compile time; [4]
18055 * indicates if that inversion list has any user-defined
18056 * properties in it. */
18057 if (av_tindex_nomg(av) >= 3) {
18059 if (SvUV(ary[4])) {
18060 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
18068 /* Element [1] is reserved for the set-up swash. If already there,
18069 * return it; if not, create it and store it there */
18070 if (ary[1] && SvROK(ary[1])) {
18073 else if (doinit && ((si && si != &PL_sv_undef)
18074 || (invlist && invlist != &PL_sv_undef))) {
18076 sw = _core_swash_init("utf8", /* the utf8 package */
18080 0, /* not from tr/// */
18082 &swash_init_flags);
18083 (void)av_store(av, 1, sw);
18088 /* If requested, return a printable version of what this swash matches */
18090 SV* matches_string = NULL;
18092 /* The swash should be used, if possible, to get the data, as it
18093 * contains the resolved data. But this function can be called at
18094 * compile-time, before everything gets resolved, in which case we
18095 * return the currently best available information, which is the string
18096 * that will eventually be used to do that resolving, 'si' */
18097 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
18098 && (si && si != &PL_sv_undef))
18100 /* Here, we only have 'si' (and possibly some passed-in data in
18101 * 'invlist', which is handled below) If the caller only wants
18102 * 'si', use that. */
18103 if (! output_invlist) {
18104 matches_string = newSVsv(si);
18107 /* But if the caller wants an inversion list of the node, we
18108 * need to parse 'si' and place as much as possible in the
18109 * desired output inversion list, making 'matches_string' only
18110 * contain the currently unresolvable things */
18111 const char *si_string = SvPVX(si);
18112 STRLEN remaining = SvCUR(si);
18116 /* Ignore everything before the first new-line */
18117 while (*si_string != '\n' && remaining > 0) {
18121 assert(remaining > 0);
18126 while (remaining > 0) {
18128 /* The data consists of just strings defining user-defined
18129 * property names, but in prior incarnations, and perhaps
18130 * somehow from pluggable regex engines, it could still
18131 * hold hex code point definitions. Each component of a
18132 * range would be separated by a tab, and each range by a
18133 * new-line. If these are found, instead add them to the
18134 * inversion list */
18135 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
18136 |PERL_SCAN_SILENT_NON_PORTABLE;
18137 STRLEN len = remaining;
18138 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
18140 /* If the hex decode routine found something, it should go
18141 * up to the next \n */
18142 if ( *(si_string + len) == '\n') {
18143 if (count) { /* 2nd code point on line */
18144 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
18147 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
18150 goto prepare_for_next_iteration;
18153 /* If the hex decode was instead for the lower range limit,
18154 * save it, and go parse the upper range limit */
18155 if (*(si_string + len) == '\t') {
18156 assert(count == 0);
18160 prepare_for_next_iteration:
18161 si_string += len + 1;
18162 remaining -= len + 1;
18166 /* Here, didn't find a legal hex number. Just add it from
18167 * here to the next \n */
18170 while (*(si_string + len) != '\n' && remaining > 0) {
18174 if (*(si_string + len) == '\n') {
18178 if (matches_string) {
18179 sv_catpvn(matches_string, si_string, len - 1);
18182 matches_string = newSVpvn(si_string, len - 1);
18185 sv_catpvs(matches_string, " ");
18186 } /* end of loop through the text */
18188 assert(matches_string);
18189 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
18190 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
18192 } /* end of has an 'si' but no swash */
18195 /* If we have a swash in place, its equivalent inversion list was above
18196 * placed into 'invlist'. If not, this variable may contain a stored
18197 * inversion list which is information beyond what is in 'si' */
18200 /* Again, if the caller doesn't want the output inversion list, put
18201 * everything in 'matches-string' */
18202 if (! output_invlist) {
18203 if ( ! matches_string) {
18204 matches_string = newSVpvs("\n");
18206 sv_catsv(matches_string, invlist_contents(invlist,
18207 TRUE /* traditional style */
18210 else if (! *output_invlist) {
18211 *output_invlist = invlist_clone(invlist);
18214 _invlist_union(*output_invlist, invlist, output_invlist);
18218 *listsvp = matches_string;
18223 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
18225 /* reg_skipcomment()
18227 Absorbs an /x style # comment from the input stream,
18228 returning a pointer to the first character beyond the comment, or if the
18229 comment terminates the pattern without anything following it, this returns
18230 one past the final character of the pattern (in other words, RExC_end) and
18231 sets the REG_RUN_ON_COMMENT_SEEN flag.
18233 Note it's the callers responsibility to ensure that we are
18234 actually in /x mode
18238 PERL_STATIC_INLINE char*
18239 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
18241 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
18245 while (p < RExC_end) {
18246 if (*(++p) == '\n') {
18251 /* we ran off the end of the pattern without ending the comment, so we have
18252 * to add an \n when wrapping */
18253 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
18258 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
18260 const bool force_to_xmod
18263 /* If the text at the current parse position '*p' is a '(?#...)' comment,
18264 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
18265 * is /x whitespace, advance '*p' so that on exit it points to the first
18266 * byte past all such white space and comments */
18268 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
18270 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
18272 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
18275 if (RExC_end - (*p) >= 3
18277 && *(*p + 1) == '?'
18278 && *(*p + 2) == '#')
18280 while (*(*p) != ')') {
18281 if ((*p) == RExC_end)
18282 FAIL("Sequence (?#... not terminated");
18290 const char * save_p = *p;
18291 while ((*p) < RExC_end) {
18293 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
18296 else if (*(*p) == '#') {
18297 (*p) = reg_skipcomment(pRExC_state, (*p));
18303 if (*p != save_p) {
18316 Advances the parse position by one byte, unless that byte is the beginning
18317 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
18318 those two cases, the parse position is advanced beyond all such comments and
18321 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
18325 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
18327 PERL_ARGS_ASSERT_NEXTCHAR;
18329 if (RExC_parse < RExC_end) {
18331 || UTF8_IS_INVARIANT(*RExC_parse)
18332 || UTF8_IS_START(*RExC_parse));
18334 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
18336 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
18337 FALSE /* Don't force /x */ );
18342 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
18344 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
18345 * space. In pass1, it aligns and increments RExC_size; in pass2,
18348 regnode * const ret = RExC_emit;
18349 GET_RE_DEBUG_FLAGS_DECL;
18351 PERL_ARGS_ASSERT_REGNODE_GUTS;
18353 assert(extra_size >= regarglen[op]);
18356 SIZE_ALIGN(RExC_size);
18357 RExC_size += 1 + extra_size;
18360 if (RExC_emit >= RExC_emit_bound)
18361 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
18362 op, (void*)RExC_emit, (void*)RExC_emit_bound);
18364 NODE_ALIGN_FILL(ret);
18365 #ifndef RE_TRACK_PATTERN_OFFSETS
18366 PERL_UNUSED_ARG(name);
18368 if (RExC_offsets) { /* MJD */
18370 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
18373 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
18374 ? "Overwriting end of array!\n" : "OK",
18375 (UV)(RExC_emit - RExC_emit_start),
18376 (UV)(RExC_parse - RExC_start),
18377 (UV)RExC_offsets[0]));
18378 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
18385 - reg_node - emit a node
18387 STATIC regnode * /* Location. */
18388 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
18390 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
18392 PERL_ARGS_ASSERT_REG_NODE;
18394 assert(regarglen[op] == 0);
18397 regnode *ptr = ret;
18398 FILL_ADVANCE_NODE(ptr, op);
18405 - reganode - emit a node with an argument
18407 STATIC regnode * /* Location. */
18408 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
18410 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
18412 PERL_ARGS_ASSERT_REGANODE;
18414 assert(regarglen[op] == 1);
18417 regnode *ptr = ret;
18418 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
18425 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
18427 /* emit a node with U32 and I32 arguments */
18429 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
18431 PERL_ARGS_ASSERT_REG2LANODE;
18433 assert(regarglen[op] == 2);
18436 regnode *ptr = ret;
18437 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
18444 - reginsert - insert an operator in front of already-emitted operand
18446 * Means relocating the operand.
18449 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
18454 const int offset = regarglen[(U8)op];
18455 const int size = NODE_STEP_REGNODE + offset;
18456 GET_RE_DEBUG_FLAGS_DECL;
18458 PERL_ARGS_ASSERT_REGINSERT;
18459 PERL_UNUSED_CONTEXT;
18460 PERL_UNUSED_ARG(depth);
18461 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
18462 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
18467 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
18468 studying. If this is wrong then we need to adjust RExC_recurse
18469 below like we do with RExC_open_parens/RExC_close_parens. */
18473 if (RExC_open_parens) {
18475 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
18476 /* remember that RExC_npar is rex->nparens + 1,
18477 * iow it is 1 more than the number of parens seen in
18478 * the pattern so far. */
18479 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
18480 /* note, RExC_open_parens[0] is the start of the
18481 * regex, it can't move. RExC_close_parens[0] is the end
18482 * of the regex, it *can* move. */
18483 if ( paren && RExC_open_parens[paren] >= opnd ) {
18484 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
18485 RExC_open_parens[paren] += size;
18487 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
18489 if ( RExC_close_parens[paren] >= opnd ) {
18490 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
18491 RExC_close_parens[paren] += size;
18493 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
18498 RExC_end_op += size;
18500 while (src > opnd) {
18501 StructCopy(--src, --dst, regnode);
18502 #ifdef RE_TRACK_PATTERN_OFFSETS
18503 if (RExC_offsets) { /* MJD 20010112 */
18505 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
18509 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
18510 ? "Overwriting end of array!\n" : "OK",
18511 (UV)(src - RExC_emit_start),
18512 (UV)(dst - RExC_emit_start),
18513 (UV)RExC_offsets[0]));
18514 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
18515 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
18521 place = opnd; /* Op node, where operand used to be. */
18522 #ifdef RE_TRACK_PATTERN_OFFSETS
18523 if (RExC_offsets) { /* MJD */
18525 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
18529 (UV)(place - RExC_emit_start) > RExC_offsets[0]
18530 ? "Overwriting end of array!\n" : "OK",
18531 (UV)(place - RExC_emit_start),
18532 (UV)(RExC_parse - RExC_start),
18533 (UV)RExC_offsets[0]));
18534 Set_Node_Offset(place, RExC_parse);
18535 Set_Node_Length(place, 1);
18538 src = NEXTOPER(place);
18539 FILL_ADVANCE_NODE(place, op);
18540 Zero(src, offset, regnode);
18544 - regtail - set the next-pointer at the end of a node chain of p to val.
18545 - SEE ALSO: regtail_study
18548 S_regtail(pTHX_ RExC_state_t * pRExC_state,
18549 const regnode * const p,
18550 const regnode * const val,
18554 GET_RE_DEBUG_FLAGS_DECL;
18556 PERL_ARGS_ASSERT_REGTAIL;
18558 PERL_UNUSED_ARG(depth);
18564 /* Find last node. */
18565 scan = (regnode *) p;
18567 regnode * const temp = regnext(scan);
18569 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
18570 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18571 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
18572 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
18573 (temp == NULL ? "->" : ""),
18574 (temp == NULL ? PL_reg_name[OP(val)] : "")
18582 if (reg_off_by_arg[OP(scan)]) {
18583 ARG_SET(scan, val - scan);
18586 NEXT_OFF(scan) = val - scan;
18592 - regtail_study - set the next-pointer at the end of a node chain of p to val.
18593 - Look for optimizable sequences at the same time.
18594 - currently only looks for EXACT chains.
18596 This is experimental code. The idea is to use this routine to perform
18597 in place optimizations on branches and groups as they are constructed,
18598 with the long term intention of removing optimization from study_chunk so
18599 that it is purely analytical.
18601 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
18602 to control which is which.
18605 /* TODO: All four parms should be const */
18608 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
18609 const regnode *val,U32 depth)
18613 #ifdef EXPERIMENTAL_INPLACESCAN
18616 GET_RE_DEBUG_FLAGS_DECL;
18618 PERL_ARGS_ASSERT_REGTAIL_STUDY;
18624 /* Find last node. */
18628 regnode * const temp = regnext(scan);
18629 #ifdef EXPERIMENTAL_INPLACESCAN
18630 if (PL_regkind[OP(scan)] == EXACT) {
18631 bool unfolded_multi_char; /* Unexamined in this routine */
18632 if (join_exact(pRExC_state, scan, &min,
18633 &unfolded_multi_char, 1, val, depth+1))
18638 switch (OP(scan)) {
18642 case EXACTFA_NO_TRIE:
18648 if( exact == PSEUDO )
18650 else if ( exact != OP(scan) )
18659 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
18660 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18661 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
18662 SvPV_nolen_const(RExC_mysv),
18663 REG_NODE_NUM(scan),
18664 PL_reg_name[exact]);
18671 DEBUG_PARSE_MSG("");
18672 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
18673 Perl_re_printf( aTHX_
18674 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
18675 SvPV_nolen_const(RExC_mysv),
18676 (IV)REG_NODE_NUM(val),
18680 if (reg_off_by_arg[OP(scan)]) {
18681 ARG_SET(scan, val - scan);
18684 NEXT_OFF(scan) = val - scan;
18692 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
18697 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
18702 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18704 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
18705 if (flags & (1<<bit)) {
18706 if (!set++ && lead)
18707 Perl_re_printf( aTHX_ "%s",lead);
18708 Perl_re_printf( aTHX_ "%s ",PL_reg_intflags_name[bit]);
18713 Perl_re_printf( aTHX_ "\n");
18715 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18720 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
18726 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18728 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
18729 if (flags & (1<<bit)) {
18730 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
18733 if (!set++ && lead)
18734 Perl_re_printf( aTHX_ "%s",lead);
18735 Perl_re_printf( aTHX_ "%s ",PL_reg_extflags_name[bit]);
18738 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
18739 if (!set++ && lead) {
18740 Perl_re_printf( aTHX_ "%s",lead);
18743 case REGEX_UNICODE_CHARSET:
18744 Perl_re_printf( aTHX_ "UNICODE");
18746 case REGEX_LOCALE_CHARSET:
18747 Perl_re_printf( aTHX_ "LOCALE");
18749 case REGEX_ASCII_RESTRICTED_CHARSET:
18750 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
18752 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
18753 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
18756 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
18762 Perl_re_printf( aTHX_ "\n");
18764 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18770 Perl_regdump(pTHX_ const regexp *r)
18773 SV * const sv = sv_newmortal();
18774 SV *dsv= sv_newmortal();
18775 RXi_GET_DECL(r,ri);
18776 GET_RE_DEBUG_FLAGS_DECL;
18778 PERL_ARGS_ASSERT_REGDUMP;
18780 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
18782 /* Header fields of interest. */
18783 if (r->anchored_substr) {
18784 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
18785 RE_SV_DUMPLEN(r->anchored_substr), 30);
18786 Perl_re_printf( aTHX_
18787 "anchored %s%s at %" IVdf " ",
18788 s, RE_SV_TAIL(r->anchored_substr),
18789 (IV)r->anchored_offset);
18790 } else if (r->anchored_utf8) {
18791 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
18792 RE_SV_DUMPLEN(r->anchored_utf8), 30);
18793 Perl_re_printf( aTHX_
18794 "anchored utf8 %s%s at %" IVdf " ",
18795 s, RE_SV_TAIL(r->anchored_utf8),
18796 (IV)r->anchored_offset);
18798 if (r->float_substr) {
18799 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
18800 RE_SV_DUMPLEN(r->float_substr), 30);
18801 Perl_re_printf( aTHX_
18802 "floating %s%s at %" IVdf "..%" UVuf " ",
18803 s, RE_SV_TAIL(r->float_substr),
18804 (IV)r->float_min_offset, (UV)r->float_max_offset);
18805 } else if (r->float_utf8) {
18806 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
18807 RE_SV_DUMPLEN(r->float_utf8), 30);
18808 Perl_re_printf( aTHX_
18809 "floating utf8 %s%s at %" IVdf "..%" UVuf " ",
18810 s, RE_SV_TAIL(r->float_utf8),
18811 (IV)r->float_min_offset, (UV)r->float_max_offset);
18813 if (r->check_substr || r->check_utf8)
18814 Perl_re_printf( aTHX_
18816 (r->check_substr == r->float_substr
18817 && r->check_utf8 == r->float_utf8
18818 ? "(checking floating" : "(checking anchored"));
18819 if (r->intflags & PREGf_NOSCAN)
18820 Perl_re_printf( aTHX_ " noscan");
18821 if (r->extflags & RXf_CHECK_ALL)
18822 Perl_re_printf( aTHX_ " isall");
18823 if (r->check_substr || r->check_utf8)
18824 Perl_re_printf( aTHX_ ") ");
18826 if (ri->regstclass) {
18827 regprop(r, sv, ri->regstclass, NULL, NULL);
18828 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
18830 if (r->intflags & PREGf_ANCH) {
18831 Perl_re_printf( aTHX_ "anchored");
18832 if (r->intflags & PREGf_ANCH_MBOL)
18833 Perl_re_printf( aTHX_ "(MBOL)");
18834 if (r->intflags & PREGf_ANCH_SBOL)
18835 Perl_re_printf( aTHX_ "(SBOL)");
18836 if (r->intflags & PREGf_ANCH_GPOS)
18837 Perl_re_printf( aTHX_ "(GPOS)");
18838 Perl_re_printf( aTHX_ " ");
18840 if (r->intflags & PREGf_GPOS_SEEN)
18841 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
18842 if (r->intflags & PREGf_SKIP)
18843 Perl_re_printf( aTHX_ "plus ");
18844 if (r->intflags & PREGf_IMPLICIT)
18845 Perl_re_printf( aTHX_ "implicit ");
18846 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
18847 if (r->extflags & RXf_EVAL_SEEN)
18848 Perl_re_printf( aTHX_ "with eval ");
18849 Perl_re_printf( aTHX_ "\n");
18851 regdump_extflags("r->extflags: ",r->extflags);
18852 regdump_intflags("r->intflags: ",r->intflags);
18855 PERL_ARGS_ASSERT_REGDUMP;
18856 PERL_UNUSED_CONTEXT;
18857 PERL_UNUSED_ARG(r);
18858 #endif /* DEBUGGING */
18861 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
18864 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
18865 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
18866 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
18867 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
18868 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
18869 || _CC_VERTSPACE != 15
18870 # error Need to adjust order of anyofs[]
18872 static const char * const anyofs[] = {
18909 - regprop - printable representation of opcode, with run time support
18913 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
18917 RXi_GET_DECL(prog,progi);
18918 GET_RE_DEBUG_FLAGS_DECL;
18920 PERL_ARGS_ASSERT_REGPROP;
18924 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
18925 /* It would be nice to FAIL() here, but this may be called from
18926 regexec.c, and it would be hard to supply pRExC_state. */
18927 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
18928 (int)OP(o), (int)REGNODE_MAX);
18929 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
18931 k = PL_regkind[OP(o)];
18934 sv_catpvs(sv, " ");
18935 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
18936 * is a crude hack but it may be the best for now since
18937 * we have no flag "this EXACTish node was UTF-8"
18939 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
18940 PERL_PV_ESCAPE_UNI_DETECT |
18941 PERL_PV_ESCAPE_NONASCII |
18942 PERL_PV_PRETTY_ELLIPSES |
18943 PERL_PV_PRETTY_LTGT |
18944 PERL_PV_PRETTY_NOCLEAR
18946 } else if (k == TRIE) {
18947 /* print the details of the trie in dumpuntil instead, as
18948 * progi->data isn't available here */
18949 const char op = OP(o);
18950 const U32 n = ARG(o);
18951 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
18952 (reg_ac_data *)progi->data->data[n] :
18954 const reg_trie_data * const trie
18955 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
18957 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
18958 DEBUG_TRIE_COMPILE_r({
18960 sv_catpvs(sv, "(JUMP)");
18961 Perl_sv_catpvf(aTHX_ sv,
18962 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
18963 (UV)trie->startstate,
18964 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
18965 (UV)trie->wordcount,
18968 (UV)TRIE_CHARCOUNT(trie),
18969 (UV)trie->uniquecharcount
18972 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
18973 sv_catpvs(sv, "[");
18974 (void) put_charclass_bitmap_innards(sv,
18975 ((IS_ANYOF_TRIE(op))
18977 : TRIE_BITMAP(trie)),
18983 sv_catpvs(sv, "]");
18985 } else if (k == CURLY) {
18986 U32 lo = ARG1(o), hi = ARG2(o);
18987 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
18988 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
18989 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
18990 if (hi == REG_INFTY)
18991 sv_catpvs(sv, "INFTY");
18993 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
18994 sv_catpvs(sv, "}");
18996 else if (k == WHILEM && o->flags) /* Ordinal/of */
18997 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
18998 else if (k == REF || k == OPEN || k == CLOSE
18999 || k == GROUPP || OP(o)==ACCEPT)
19001 AV *name_list= NULL;
19002 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
19003 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
19004 if ( RXp_PAREN_NAMES(prog) ) {
19005 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19006 } else if ( pRExC_state ) {
19007 name_list= RExC_paren_name_list;
19010 if ( k != REF || (OP(o) < NREF)) {
19011 SV **name= av_fetch(name_list, parno, 0 );
19013 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19016 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
19017 I32 *nums=(I32*)SvPVX(sv_dat);
19018 SV **name= av_fetch(name_list, nums[0], 0 );
19021 for ( n=0; n<SvIVX(sv_dat); n++ ) {
19022 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
19023 (n ? "," : ""), (IV)nums[n]);
19025 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19029 if ( k == REF && reginfo) {
19030 U32 n = ARG(o); /* which paren pair */
19031 I32 ln = prog->offs[n].start;
19032 if (prog->lastparen < n || ln == -1)
19033 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
19034 else if (ln == prog->offs[n].end)
19035 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
19037 const char *s = reginfo->strbeg + ln;
19038 Perl_sv_catpvf(aTHX_ sv, ": ");
19039 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
19040 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
19043 } else if (k == GOSUB) {
19044 AV *name_list= NULL;
19045 if ( RXp_PAREN_NAMES(prog) ) {
19046 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19047 } else if ( pRExC_state ) {
19048 name_list= RExC_paren_name_list;
19051 /* Paren and offset */
19052 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
19053 (int)((o + (int)ARG2L(o)) - progi->program) );
19055 SV **name= av_fetch(name_list, ARG(o), 0 );
19057 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19060 else if (k == LOGICAL)
19061 /* 2: embedded, otherwise 1 */
19062 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
19063 else if (k == ANYOF) {
19064 const U8 flags = ANYOF_FLAGS(o);
19065 bool do_sep = FALSE; /* Do we need to separate various components of
19067 /* Set if there is still an unresolved user-defined property */
19068 SV *unresolved = NULL;
19070 /* Things that are ignored except when the runtime locale is UTF-8 */
19071 SV *only_utf8_locale_invlist = NULL;
19073 /* Code points that don't fit in the bitmap */
19074 SV *nonbitmap_invlist = NULL;
19076 /* And things that aren't in the bitmap, but are small enough to be */
19077 SV* bitmap_range_not_in_bitmap = NULL;
19079 const bool inverted = flags & ANYOF_INVERT;
19081 if (OP(o) == ANYOFL) {
19082 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
19083 sv_catpvs(sv, "{utf8-locale-reqd}");
19085 if (flags & ANYOFL_FOLD) {
19086 sv_catpvs(sv, "{i}");
19090 /* If there is stuff outside the bitmap, get it */
19091 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
19092 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
19094 &only_utf8_locale_invlist,
19095 &nonbitmap_invlist);
19096 /* The non-bitmap data may contain stuff that could fit in the
19097 * bitmap. This could come from a user-defined property being
19098 * finally resolved when this call was done; or much more likely
19099 * because there are matches that require UTF-8 to be valid, and so
19100 * aren't in the bitmap. This is teased apart later */
19101 _invlist_intersection(nonbitmap_invlist,
19103 &bitmap_range_not_in_bitmap);
19104 /* Leave just the things that don't fit into the bitmap */
19105 _invlist_subtract(nonbitmap_invlist,
19107 &nonbitmap_invlist);
19110 /* Obey this flag to add all above-the-bitmap code points */
19111 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
19112 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
19113 NUM_ANYOF_CODE_POINTS,
19117 /* Ready to start outputting. First, the initial left bracket */
19118 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
19120 /* Then all the things that could fit in the bitmap */
19121 do_sep = put_charclass_bitmap_innards(sv,
19123 bitmap_range_not_in_bitmap,
19124 only_utf8_locale_invlist,
19127 /* Can't try inverting for a
19128 * better display if there are
19129 * things that haven't been
19131 unresolved != NULL);
19132 SvREFCNT_dec(bitmap_range_not_in_bitmap);
19134 /* If there are user-defined properties which haven't been defined yet,
19135 * output them. If the result is not to be inverted, it is clearest to
19136 * output them in a separate [] from the bitmap range stuff. If the
19137 * result is to be complemented, we have to show everything in one [],
19138 * as the inversion applies to the whole thing. Use {braces} to
19139 * separate them from anything in the bitmap and anything above the
19143 if (! do_sep) { /* If didn't output anything in the bitmap */
19144 sv_catpvs(sv, "^");
19146 sv_catpvs(sv, "{");
19149 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19151 sv_catsv(sv, unresolved);
19153 sv_catpvs(sv, "}");
19155 do_sep = ! inverted;
19158 /* And, finally, add the above-the-bitmap stuff */
19159 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
19162 /* See if truncation size is overridden */
19163 const STRLEN dump_len = (PL_dump_re_max_len)
19164 ? PL_dump_re_max_len
19167 /* This is output in a separate [] */
19169 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19172 /* And, for easy of understanding, it is shown in the
19173 * uncomplemented form if possible. The one exception being if
19174 * there are unresolved items, where the inversion has to be
19175 * delayed until runtime */
19176 if (inverted && ! unresolved) {
19177 _invlist_invert(nonbitmap_invlist);
19178 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
19181 contents = invlist_contents(nonbitmap_invlist,
19182 FALSE /* output suitable for catsv */
19185 /* If the output is shorter than the permissible maximum, just do it. */
19186 if (SvCUR(contents) <= dump_len) {
19187 sv_catsv(sv, contents);
19190 const char * contents_string = SvPVX(contents);
19191 STRLEN i = dump_len;
19193 /* Otherwise, start at the permissible max and work back to the
19194 * first break possibility */
19195 while (i > 0 && contents_string[i] != ' ') {
19198 if (i == 0) { /* Fail-safe. Use the max if we couldn't
19199 find a legal break */
19203 sv_catpvn(sv, contents_string, i);
19204 sv_catpvs(sv, "...");
19207 SvREFCNT_dec_NN(contents);
19208 SvREFCNT_dec_NN(nonbitmap_invlist);
19211 /* And finally the matching, closing ']' */
19212 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
19214 SvREFCNT_dec(unresolved);
19216 else if (k == POSIXD || k == NPOSIXD) {
19217 U8 index = FLAGS(o) * 2;
19218 if (index < C_ARRAY_LENGTH(anyofs)) {
19219 if (*anyofs[index] != '[') {
19222 sv_catpv(sv, anyofs[index]);
19223 if (*anyofs[index] != '[') {
19228 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
19231 else if (k == BOUND || k == NBOUND) {
19232 /* Must be synced with order of 'bound_type' in regcomp.h */
19233 const char * const bounds[] = {
19234 "", /* Traditional */
19240 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
19241 sv_catpv(sv, bounds[FLAGS(o)]);
19243 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
19244 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
19245 else if (OP(o) == SBOL)
19246 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
19248 /* add on the verb argument if there is one */
19249 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
19250 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
19251 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
19254 PERL_UNUSED_CONTEXT;
19255 PERL_UNUSED_ARG(sv);
19256 PERL_UNUSED_ARG(o);
19257 PERL_UNUSED_ARG(prog);
19258 PERL_UNUSED_ARG(reginfo);
19259 PERL_UNUSED_ARG(pRExC_state);
19260 #endif /* DEBUGGING */
19266 Perl_re_intuit_string(pTHX_ REGEXP * const r)
19267 { /* Assume that RE_INTUIT is set */
19268 struct regexp *const prog = ReANY(r);
19269 GET_RE_DEBUG_FLAGS_DECL;
19271 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
19272 PERL_UNUSED_CONTEXT;
19276 const char * const s = SvPV_nolen_const(RX_UTF8(r)
19277 ? prog->check_utf8 : prog->check_substr);
19279 if (!PL_colorset) reginitcolors();
19280 Perl_re_printf( aTHX_
19281 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
19283 RX_UTF8(r) ? "utf8 " : "",
19284 PL_colors[5],PL_colors[0],
19287 (strlen(s) > 60 ? "..." : ""));
19290 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
19291 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
19297 handles refcounting and freeing the perl core regexp structure. When
19298 it is necessary to actually free the structure the first thing it
19299 does is call the 'free' method of the regexp_engine associated to
19300 the regexp, allowing the handling of the void *pprivate; member
19301 first. (This routine is not overridable by extensions, which is why
19302 the extensions free is called first.)
19304 See regdupe and regdupe_internal if you change anything here.
19306 #ifndef PERL_IN_XSUB_RE
19308 Perl_pregfree(pTHX_ REGEXP *r)
19314 Perl_pregfree2(pTHX_ REGEXP *rx)
19316 struct regexp *const r = ReANY(rx);
19317 GET_RE_DEBUG_FLAGS_DECL;
19319 PERL_ARGS_ASSERT_PREGFREE2;
19321 if (r->mother_re) {
19322 ReREFCNT_dec(r->mother_re);
19324 CALLREGFREE_PVT(rx); /* free the private data */
19325 SvREFCNT_dec(RXp_PAREN_NAMES(r));
19326 Safefree(r->xpv_len_u.xpvlenu_pv);
19329 SvREFCNT_dec(r->anchored_substr);
19330 SvREFCNT_dec(r->anchored_utf8);
19331 SvREFCNT_dec(r->float_substr);
19332 SvREFCNT_dec(r->float_utf8);
19333 Safefree(r->substrs);
19335 RX_MATCH_COPY_FREE(rx);
19336 #ifdef PERL_ANY_COW
19337 SvREFCNT_dec(r->saved_copy);
19340 SvREFCNT_dec(r->qr_anoncv);
19341 if (r->recurse_locinput)
19342 Safefree(r->recurse_locinput);
19343 rx->sv_u.svu_rx = 0;
19348 This is a hacky workaround to the structural issue of match results
19349 being stored in the regexp structure which is in turn stored in
19350 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
19351 could be PL_curpm in multiple contexts, and could require multiple
19352 result sets being associated with the pattern simultaneously, such
19353 as when doing a recursive match with (??{$qr})
19355 The solution is to make a lightweight copy of the regexp structure
19356 when a qr// is returned from the code executed by (??{$qr}) this
19357 lightweight copy doesn't actually own any of its data except for
19358 the starp/end and the actual regexp structure itself.
19364 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
19366 struct regexp *ret;
19367 struct regexp *const r = ReANY(rx);
19368 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
19370 PERL_ARGS_ASSERT_REG_TEMP_COPY;
19373 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
19375 SvOK_off((SV *)ret_x);
19377 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
19378 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
19379 made both spots point to the same regexp body.) */
19380 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
19381 assert(!SvPVX(ret_x));
19382 ret_x->sv_u.svu_rx = temp->sv_any;
19383 temp->sv_any = NULL;
19384 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
19385 SvREFCNT_dec_NN(temp);
19386 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
19387 ing below will not set it. */
19388 SvCUR_set(ret_x, SvCUR(rx));
19391 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
19392 sv_force_normal(sv) is called. */
19394 ret = ReANY(ret_x);
19396 SvFLAGS(ret_x) |= SvUTF8(rx);
19397 /* We share the same string buffer as the original regexp, on which we
19398 hold a reference count, incremented when mother_re is set below.
19399 The string pointer is copied here, being part of the regexp struct.
19401 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
19402 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
19404 const I32 npar = r->nparens+1;
19405 Newx(ret->offs, npar, regexp_paren_pair);
19406 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19409 Newx(ret->substrs, 1, struct reg_substr_data);
19410 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19412 SvREFCNT_inc_void(ret->anchored_substr);
19413 SvREFCNT_inc_void(ret->anchored_utf8);
19414 SvREFCNT_inc_void(ret->float_substr);
19415 SvREFCNT_inc_void(ret->float_utf8);
19417 /* check_substr and check_utf8, if non-NULL, point to either their
19418 anchored or float namesakes, and don't hold a second reference. */
19420 RX_MATCH_COPIED_off(ret_x);
19421 #ifdef PERL_ANY_COW
19422 ret->saved_copy = NULL;
19424 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
19425 SvREFCNT_inc_void(ret->qr_anoncv);
19426 if (r->recurse_locinput)
19427 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19433 /* regfree_internal()
19435 Free the private data in a regexp. This is overloadable by
19436 extensions. Perl takes care of the regexp structure in pregfree(),
19437 this covers the *pprivate pointer which technically perl doesn't
19438 know about, however of course we have to handle the
19439 regexp_internal structure when no extension is in use.
19441 Note this is called before freeing anything in the regexp
19446 Perl_regfree_internal(pTHX_ REGEXP * const rx)
19448 struct regexp *const r = ReANY(rx);
19449 RXi_GET_DECL(r,ri);
19450 GET_RE_DEBUG_FLAGS_DECL;
19452 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
19458 SV *dsv= sv_newmortal();
19459 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
19460 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
19461 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
19462 PL_colors[4],PL_colors[5],s);
19465 #ifdef RE_TRACK_PATTERN_OFFSETS
19467 Safefree(ri->u.offsets); /* 20010421 MJD */
19469 if (ri->code_blocks) {
19471 for (n = 0; n < ri->num_code_blocks; n++)
19472 SvREFCNT_dec(ri->code_blocks[n].src_regex);
19473 Safefree(ri->code_blocks);
19477 int n = ri->data->count;
19480 /* If you add a ->what type here, update the comment in regcomp.h */
19481 switch (ri->data->what[n]) {
19487 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
19490 Safefree(ri->data->data[n]);
19496 { /* Aho Corasick add-on structure for a trie node.
19497 Used in stclass optimization only */
19499 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
19500 #ifdef USE_ITHREADS
19504 refcount = --aho->refcount;
19507 PerlMemShared_free(aho->states);
19508 PerlMemShared_free(aho->fail);
19509 /* do this last!!!! */
19510 PerlMemShared_free(ri->data->data[n]);
19511 /* we should only ever get called once, so
19512 * assert as much, and also guard the free
19513 * which /might/ happen twice. At the least
19514 * it will make code anlyzers happy and it
19515 * doesn't cost much. - Yves */
19516 assert(ri->regstclass);
19517 if (ri->regstclass) {
19518 PerlMemShared_free(ri->regstclass);
19519 ri->regstclass = 0;
19526 /* trie structure. */
19528 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
19529 #ifdef USE_ITHREADS
19533 refcount = --trie->refcount;
19536 PerlMemShared_free(trie->charmap);
19537 PerlMemShared_free(trie->states);
19538 PerlMemShared_free(trie->trans);
19540 PerlMemShared_free(trie->bitmap);
19542 PerlMemShared_free(trie->jump);
19543 PerlMemShared_free(trie->wordinfo);
19544 /* do this last!!!! */
19545 PerlMemShared_free(ri->data->data[n]);
19550 Perl_croak(aTHX_ "panic: regfree data code '%c'",
19551 ri->data->what[n]);
19554 Safefree(ri->data->what);
19555 Safefree(ri->data);
19561 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
19562 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
19563 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
19566 re_dup_guts - duplicate a regexp.
19568 This routine is expected to clone a given regexp structure. It is only
19569 compiled under USE_ITHREADS.
19571 After all of the core data stored in struct regexp is duplicated
19572 the regexp_engine.dupe method is used to copy any private data
19573 stored in the *pprivate pointer. This allows extensions to handle
19574 any duplication it needs to do.
19576 See pregfree() and regfree_internal() if you change anything here.
19578 #if defined(USE_ITHREADS)
19579 #ifndef PERL_IN_XSUB_RE
19581 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
19585 const struct regexp *r = ReANY(sstr);
19586 struct regexp *ret = ReANY(dstr);
19588 PERL_ARGS_ASSERT_RE_DUP_GUTS;
19590 npar = r->nparens+1;
19591 Newx(ret->offs, npar, regexp_paren_pair);
19592 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19594 if (ret->substrs) {
19595 /* Do it this way to avoid reading from *r after the StructCopy().
19596 That way, if any of the sv_dup_inc()s dislodge *r from the L1
19597 cache, it doesn't matter. */
19598 const bool anchored = r->check_substr
19599 ? r->check_substr == r->anchored_substr
19600 : r->check_utf8 == r->anchored_utf8;
19601 Newx(ret->substrs, 1, struct reg_substr_data);
19602 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19604 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
19605 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
19606 ret->float_substr = sv_dup_inc(ret->float_substr, param);
19607 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
19609 /* check_substr and check_utf8, if non-NULL, point to either their
19610 anchored or float namesakes, and don't hold a second reference. */
19612 if (ret->check_substr) {
19614 assert(r->check_utf8 == r->anchored_utf8);
19615 ret->check_substr = ret->anchored_substr;
19616 ret->check_utf8 = ret->anchored_utf8;
19618 assert(r->check_substr == r->float_substr);
19619 assert(r->check_utf8 == r->float_utf8);
19620 ret->check_substr = ret->float_substr;
19621 ret->check_utf8 = ret->float_utf8;
19623 } else if (ret->check_utf8) {
19625 ret->check_utf8 = ret->anchored_utf8;
19627 ret->check_utf8 = ret->float_utf8;
19632 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
19633 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
19634 if (r->recurse_locinput)
19635 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19638 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
19640 if (RX_MATCH_COPIED(dstr))
19641 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
19643 ret->subbeg = NULL;
19644 #ifdef PERL_ANY_COW
19645 ret->saved_copy = NULL;
19648 /* Whether mother_re be set or no, we need to copy the string. We
19649 cannot refrain from copying it when the storage points directly to
19650 our mother regexp, because that's
19651 1: a buffer in a different thread
19652 2: something we no longer hold a reference on
19653 so we need to copy it locally. */
19654 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
19655 ret->mother_re = NULL;
19657 #endif /* PERL_IN_XSUB_RE */
19662 This is the internal complement to regdupe() which is used to copy
19663 the structure pointed to by the *pprivate pointer in the regexp.
19664 This is the core version of the extension overridable cloning hook.
19665 The regexp structure being duplicated will be copied by perl prior
19666 to this and will be provided as the regexp *r argument, however
19667 with the /old/ structures pprivate pointer value. Thus this routine
19668 may override any copying normally done by perl.
19670 It returns a pointer to the new regexp_internal structure.
19674 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
19677 struct regexp *const r = ReANY(rx);
19678 regexp_internal *reti;
19680 RXi_GET_DECL(r,ri);
19682 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
19686 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
19687 char, regexp_internal);
19688 Copy(ri->program, reti->program, len+1, regnode);
19691 reti->num_code_blocks = ri->num_code_blocks;
19692 if (ri->code_blocks) {
19694 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
19695 struct reg_code_block);
19696 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
19697 struct reg_code_block);
19698 for (n = 0; n < ri->num_code_blocks; n++)
19699 reti->code_blocks[n].src_regex = (REGEXP*)
19700 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
19703 reti->code_blocks = NULL;
19705 reti->regstclass = NULL;
19708 struct reg_data *d;
19709 const int count = ri->data->count;
19712 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
19713 char, struct reg_data);
19714 Newx(d->what, count, U8);
19717 for (i = 0; i < count; i++) {
19718 d->what[i] = ri->data->what[i];
19719 switch (d->what[i]) {
19720 /* see also regcomp.h and regfree_internal() */
19721 case 'a': /* actually an AV, but the dup function is identical. */
19725 case 'u': /* actually an HV, but the dup function is identical. */
19726 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
19729 /* This is cheating. */
19730 Newx(d->data[i], 1, regnode_ssc);
19731 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
19732 reti->regstclass = (regnode*)d->data[i];
19735 /* Trie stclasses are readonly and can thus be shared
19736 * without duplication. We free the stclass in pregfree
19737 * when the corresponding reg_ac_data struct is freed.
19739 reti->regstclass= ri->regstclass;
19743 ((reg_trie_data*)ri->data->data[i])->refcount++;
19748 d->data[i] = ri->data->data[i];
19751 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
19752 ri->data->what[i]);
19761 reti->name_list_idx = ri->name_list_idx;
19763 #ifdef RE_TRACK_PATTERN_OFFSETS
19764 if (ri->u.offsets) {
19765 Newx(reti->u.offsets, 2*len+1, U32);
19766 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
19769 SetProgLen(reti,len);
19772 return (void*)reti;
19775 #endif /* USE_ITHREADS */
19777 #ifndef PERL_IN_XSUB_RE
19780 - regnext - dig the "next" pointer out of a node
19783 Perl_regnext(pTHX_ regnode *p)
19790 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
19791 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
19792 (int)OP(p), (int)REGNODE_MAX);
19795 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
19804 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
19807 STRLEN l1 = strlen(pat1);
19808 STRLEN l2 = strlen(pat2);
19811 const char *message;
19813 PERL_ARGS_ASSERT_RE_CROAK2;
19819 Copy(pat1, buf, l1 , char);
19820 Copy(pat2, buf + l1, l2 , char);
19821 buf[l1 + l2] = '\n';
19822 buf[l1 + l2 + 1] = '\0';
19823 va_start(args, pat2);
19824 msv = vmess(buf, &args);
19826 message = SvPV_const(msv,l1);
19829 Copy(message, buf, l1 , char);
19830 /* l1-1 to avoid \n */
19831 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, l1-1, buf));
19834 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
19836 #ifndef PERL_IN_XSUB_RE
19838 Perl_save_re_context(pTHX)
19843 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
19846 const REGEXP * const rx = PM_GETRE(PL_curpm);
19848 nparens = RX_NPARENS(rx);
19851 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
19852 * that PL_curpm will be null, but that utf8.pm and the modules it
19853 * loads will only use $1..$3.
19854 * The t/porting/re_context.t test file checks this assumption.
19859 for (i = 1; i <= nparens; i++) {
19860 char digits[TYPE_CHARS(long)];
19861 const STRLEN len = my_snprintf(digits, sizeof(digits),
19863 GV *const *const gvp
19864 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
19867 GV * const gv = *gvp;
19868 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
19878 S_put_code_point(pTHX_ SV *sv, UV c)
19880 PERL_ARGS_ASSERT_PUT_CODE_POINT;
19883 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
19885 else if (isPRINT(c)) {
19886 const char string = (char) c;
19888 /* We use {phrase} as metanotation in the class, so also escape literal
19890 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
19891 sv_catpvs(sv, "\\");
19892 sv_catpvn(sv, &string, 1);
19894 else if (isMNEMONIC_CNTRL(c)) {
19895 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
19898 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
19902 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
19905 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
19907 /* Appends to 'sv' a displayable version of the range of code points from
19908 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
19909 * that have them, when they occur at the beginning or end of the range.
19910 * It uses hex to output the remaining code points, unless 'allow_literals'
19911 * is true, in which case the printable ASCII ones are output as-is (though
19912 * some of these will be escaped by put_code_point()).
19914 * NOTE: This is designed only for printing ranges of code points that fit
19915 * inside an ANYOF bitmap. Higher code points are simply suppressed
19918 const unsigned int min_range_count = 3;
19920 assert(start <= end);
19922 PERL_ARGS_ASSERT_PUT_RANGE;
19924 while (start <= end) {
19926 const char * format;
19928 if (end - start < min_range_count) {
19930 /* Output chars individually when they occur in short ranges */
19931 for (; start <= end; start++) {
19932 put_code_point(sv, start);
19937 /* If permitted by the input options, and there is a possibility that
19938 * this range contains a printable literal, look to see if there is
19940 if (allow_literals && start <= MAX_PRINT_A) {
19942 /* If the character at the beginning of the range isn't an ASCII
19943 * printable, effectively split the range into two parts:
19944 * 1) the portion before the first such printable,
19946 * and output them separately. */
19947 if (! isPRINT_A(start)) {
19948 UV temp_end = start + 1;
19950 /* There is no point looking beyond the final possible
19951 * printable, in MAX_PRINT_A */
19952 UV max = MIN(end, MAX_PRINT_A);
19954 while (temp_end <= max && ! isPRINT_A(temp_end)) {
19958 /* Here, temp_end points to one beyond the first printable if
19959 * found, or to one beyond 'max' if not. If none found, make
19960 * sure that we use the entire range */
19961 if (temp_end > MAX_PRINT_A) {
19962 temp_end = end + 1;
19965 /* Output the first part of the split range: the part that
19966 * doesn't have printables, with the parameter set to not look
19967 * for literals (otherwise we would infinitely recurse) */
19968 put_range(sv, start, temp_end - 1, FALSE);
19970 /* The 2nd part of the range (if any) starts here. */
19973 /* We do a continue, instead of dropping down, because even if
19974 * the 2nd part is non-empty, it could be so short that we want
19975 * to output it as individual characters, as tested for at the
19976 * top of this loop. */
19980 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
19981 * output a sub-range of just the digits or letters, then process
19982 * the remaining portion as usual. */
19983 if (isALPHANUMERIC_A(start)) {
19984 UV mask = (isDIGIT_A(start))
19989 UV temp_end = start + 1;
19991 /* Find the end of the sub-range that includes just the
19992 * characters in the same class as the first character in it */
19993 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
19998 /* For short ranges, don't duplicate the code above to output
19999 * them; just call recursively */
20000 if (temp_end - start < min_range_count) {
20001 put_range(sv, start, temp_end, FALSE);
20003 else { /* Output as a range */
20004 put_code_point(sv, start);
20005 sv_catpvs(sv, "-");
20006 put_code_point(sv, temp_end);
20008 start = temp_end + 1;
20012 /* We output any other printables as individual characters */
20013 if (isPUNCT_A(start) || isSPACE_A(start)) {
20014 while (start <= end && (isPUNCT_A(start)
20015 || isSPACE_A(start)))
20017 put_code_point(sv, start);
20022 } /* End of looking for literals */
20024 /* Here is not to output as a literal. Some control characters have
20025 * mnemonic names. Split off any of those at the beginning and end of
20026 * the range to print mnemonically. It isn't possible for many of
20027 * these to be in a row, so this won't overwhelm with output */
20029 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
20031 while (isMNEMONIC_CNTRL(start) && start <= end) {
20032 put_code_point(sv, start);
20036 /* If this didn't take care of the whole range ... */
20037 if (start <= end) {
20039 /* Look backwards from the end to find the final non-mnemonic
20042 while (isMNEMONIC_CNTRL(temp_end)) {
20046 /* And separately output the interior range that doesn't start
20047 * or end with mnemonics */
20048 put_range(sv, start, temp_end, FALSE);
20050 /* Then output the mnemonic trailing controls */
20051 start = temp_end + 1;
20052 while (start <= end) {
20053 put_code_point(sv, start);
20060 /* As a final resort, output the range or subrange as hex. */
20062 this_end = (end < NUM_ANYOF_CODE_POINTS)
20064 : NUM_ANYOF_CODE_POINTS - 1;
20065 #if NUM_ANYOF_CODE_POINTS > 256
20066 format = (this_end < 256)
20067 ? "\\x%02" UVXf "-\\x%02" UVXf
20068 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
20070 format = "\\x%02" UVXf "-\\x%02" UVXf;
20072 GCC_DIAG_IGNORE(-Wformat-nonliteral);
20073 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
20080 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
20082 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
20086 bool allow_literals = TRUE;
20088 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
20090 /* Generally, it is more readable if printable characters are output as
20091 * literals, but if a range (nearly) spans all of them, it's best to output
20092 * it as a single range. This code will use a single range if all but 2
20093 * ASCII printables are in it */
20094 invlist_iterinit(invlist);
20095 while (invlist_iternext(invlist, &start, &end)) {
20097 /* If the range starts beyond the final printable, it doesn't have any
20099 if (start > MAX_PRINT_A) {
20103 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
20104 * all but two, the range must start and end no later than 2 from
20106 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
20107 if (end > MAX_PRINT_A) {
20113 if (end - start >= MAX_PRINT_A - ' ' - 2) {
20114 allow_literals = FALSE;
20119 invlist_iterfinish(invlist);
20121 /* Here we have figured things out. Output each range */
20122 invlist_iterinit(invlist);
20123 while (invlist_iternext(invlist, &start, &end)) {
20124 if (start >= NUM_ANYOF_CODE_POINTS) {
20127 put_range(sv, start, end, allow_literals);
20129 invlist_iterfinish(invlist);
20135 S_put_charclass_bitmap_innards_common(pTHX_
20136 SV* invlist, /* The bitmap */
20137 SV* posixes, /* Under /l, things like [:word:], \S */
20138 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
20139 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
20140 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
20141 const bool invert /* Is the result to be inverted? */
20144 /* Create and return an SV containing a displayable version of the bitmap
20145 * and associated information determined by the input parameters. If the
20146 * output would have been only the inversion indicator '^', NULL is instead
20151 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
20154 output = newSVpvs("^");
20157 output = newSVpvs("");
20160 /* First, the code points in the bitmap that are unconditionally there */
20161 put_charclass_bitmap_innards_invlist(output, invlist);
20163 /* Traditionally, these have been placed after the main code points */
20165 sv_catsv(output, posixes);
20168 if (only_utf8 && _invlist_len(only_utf8)) {
20169 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
20170 put_charclass_bitmap_innards_invlist(output, only_utf8);
20173 if (not_utf8 && _invlist_len(not_utf8)) {
20174 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
20175 put_charclass_bitmap_innards_invlist(output, not_utf8);
20178 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
20179 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
20180 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
20182 /* This is the only list in this routine that can legally contain code
20183 * points outside the bitmap range. The call just above to
20184 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
20185 * output them here. There's about a half-dozen possible, and none in
20186 * contiguous ranges longer than 2 */
20187 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20189 SV* above_bitmap = NULL;
20191 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
20193 invlist_iterinit(above_bitmap);
20194 while (invlist_iternext(above_bitmap, &start, &end)) {
20197 for (i = start; i <= end; i++) {
20198 put_code_point(output, i);
20201 invlist_iterfinish(above_bitmap);
20202 SvREFCNT_dec_NN(above_bitmap);
20206 if (invert && SvCUR(output) == 1) {
20214 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
20216 SV *nonbitmap_invlist,
20217 SV *only_utf8_locale_invlist,
20218 const regnode * const node,
20219 const bool force_as_is_display)
20221 /* Appends to 'sv' a displayable version of the innards of the bracketed
20222 * character class defined by the other arguments:
20223 * 'bitmap' points to the bitmap.
20224 * 'nonbitmap_invlist' is an inversion list of the code points that are in
20225 * the bitmap range, but for some reason aren't in the bitmap; NULL if
20226 * none. The reasons for this could be that they require some
20227 * condition such as the target string being or not being in UTF-8
20228 * (under /d), or because they came from a user-defined property that
20229 * was not resolved at the time of the regex compilation (under /u)
20230 * 'only_utf8_locale_invlist' is an inversion list of the code points that
20231 * are valid only if the runtime locale is a UTF-8 one; NULL if none
20232 * 'node' is the regex pattern node. It is needed only when the above two
20233 * parameters are not null, and is passed so that this routine can
20234 * tease apart the various reasons for them.
20235 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
20236 * to invert things to see if that leads to a cleaner display. If
20237 * FALSE, this routine is free to use its judgment about doing this.
20239 * It returns TRUE if there was actually something output. (It may be that
20240 * the bitmap, etc is empty.)
20242 * When called for outputting the bitmap of a non-ANYOF node, just pass the
20243 * bitmap, with the succeeding parameters set to NULL, and the final one to
20247 /* In general, it tries to display the 'cleanest' representation of the
20248 * innards, choosing whether to display them inverted or not, regardless of
20249 * whether the class itself is to be inverted. However, there are some
20250 * cases where it can't try inverting, as what actually matches isn't known
20251 * until runtime, and hence the inversion isn't either. */
20252 bool inverting_allowed = ! force_as_is_display;
20255 STRLEN orig_sv_cur = SvCUR(sv);
20257 SV* invlist; /* Inversion list we accumulate of code points that
20258 are unconditionally matched */
20259 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
20261 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
20263 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
20264 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
20267 SV* as_is_display; /* The output string when we take the inputs
20269 SV* inverted_display; /* The output string when we invert the inputs */
20271 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
20273 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
20275 /* We are biased in favor of displaying things without them being inverted,
20276 * as that is generally easier to understand */
20277 const int bias = 5;
20279 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
20281 /* Start off with whatever code points are passed in. (We clone, so we
20282 * don't change the caller's list) */
20283 if (nonbitmap_invlist) {
20284 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
20285 invlist = invlist_clone(nonbitmap_invlist);
20287 else { /* Worst case size is every other code point is matched */
20288 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
20292 if (OP(node) == ANYOFD) {
20294 /* This flag indicates that the code points below 0x100 in the
20295 * nonbitmap list are precisely the ones that match only when the
20296 * target is UTF-8 (they should all be non-ASCII). */
20297 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
20299 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
20300 _invlist_subtract(invlist, only_utf8, &invlist);
20303 /* And this flag for matching all non-ASCII 0xFF and below */
20304 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
20306 not_utf8 = invlist_clone(PL_UpperLatin1);
20309 else if (OP(node) == ANYOFL) {
20311 /* If either of these flags are set, what matches isn't
20312 * determinable except during execution, so don't know enough here
20314 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
20315 inverting_allowed = FALSE;
20318 /* What the posix classes match also varies at runtime, so these
20319 * will be output symbolically. */
20320 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
20323 posixes = newSVpvs("");
20324 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
20325 if (ANYOF_POSIXL_TEST(node,i)) {
20326 sv_catpv(posixes, anyofs[i]);
20333 /* Accumulate the bit map into the unconditional match list */
20334 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
20335 if (BITMAP_TEST(bitmap, i)) {
20337 for (; i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i); i++) {
20340 invlist = _add_range_to_invlist(invlist, start, i-1);
20344 /* Make sure that the conditional match lists don't have anything in them
20345 * that match unconditionally; otherwise the output is quite confusing.
20346 * This could happen if the code that populates these misses some
20349 _invlist_subtract(only_utf8, invlist, &only_utf8);
20352 _invlist_subtract(not_utf8, invlist, ¬_utf8);
20355 if (only_utf8_locale_invlist) {
20357 /* Since this list is passed in, we have to make a copy before
20359 only_utf8_locale = invlist_clone(only_utf8_locale_invlist);
20361 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
20363 /* And, it can get really weird for us to try outputting an inverted
20364 * form of this list when it has things above the bitmap, so don't even
20366 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20367 inverting_allowed = FALSE;
20371 /* Calculate what the output would be if we take the input as-is */
20372 as_is_display = put_charclass_bitmap_innards_common(invlist,
20379 /* If have to take the output as-is, just do that */
20380 if (! inverting_allowed) {
20381 if (as_is_display) {
20382 sv_catsv(sv, as_is_display);
20383 SvREFCNT_dec_NN(as_is_display);
20386 else { /* But otherwise, create the output again on the inverted input, and
20387 use whichever version is shorter */
20389 int inverted_bias, as_is_bias;
20391 /* We will apply our bias to whichever of the the results doesn't have
20401 inverted_bias = bias;
20404 /* Now invert each of the lists that contribute to the output,
20405 * excluding from the result things outside the possible range */
20407 /* For the unconditional inversion list, we have to add in all the
20408 * conditional code points, so that when inverted, they will be gone
20410 _invlist_union(only_utf8, invlist, &invlist);
20411 _invlist_union(not_utf8, invlist, &invlist);
20412 _invlist_union(only_utf8_locale, invlist, &invlist);
20413 _invlist_invert(invlist);
20414 _invlist_intersection(invlist, PL_InBitmap, &invlist);
20417 _invlist_invert(only_utf8);
20418 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
20420 else if (not_utf8) {
20422 /* If a code point matches iff the target string is not in UTF-8,
20423 * then complementing the result has it not match iff not in UTF-8,
20424 * which is the same thing as matching iff it is UTF-8. */
20425 only_utf8 = not_utf8;
20429 if (only_utf8_locale) {
20430 _invlist_invert(only_utf8_locale);
20431 _invlist_intersection(only_utf8_locale,
20433 &only_utf8_locale);
20436 inverted_display = put_charclass_bitmap_innards_common(
20441 only_utf8_locale, invert);
20443 /* Use the shortest representation, taking into account our bias
20444 * against showing it inverted */
20445 if ( inverted_display
20446 && ( ! as_is_display
20447 || ( SvCUR(inverted_display) + inverted_bias
20448 < SvCUR(as_is_display) + as_is_bias)))
20450 sv_catsv(sv, inverted_display);
20452 else if (as_is_display) {
20453 sv_catsv(sv, as_is_display);
20456 SvREFCNT_dec(as_is_display);
20457 SvREFCNT_dec(inverted_display);
20460 SvREFCNT_dec_NN(invlist);
20461 SvREFCNT_dec(only_utf8);
20462 SvREFCNT_dec(not_utf8);
20463 SvREFCNT_dec(posixes);
20464 SvREFCNT_dec(only_utf8_locale);
20466 return SvCUR(sv) > orig_sv_cur;
20469 #define CLEAR_OPTSTART \
20470 if (optstart) STMT_START { \
20471 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
20472 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
20476 #define DUMPUNTIL(b,e) \
20478 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
20480 STATIC const regnode *
20481 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
20482 const regnode *last, const regnode *plast,
20483 SV* sv, I32 indent, U32 depth)
20485 U8 op = PSEUDO; /* Arbitrary non-END op. */
20486 const regnode *next;
20487 const regnode *optstart= NULL;
20489 RXi_GET_DECL(r,ri);
20490 GET_RE_DEBUG_FLAGS_DECL;
20492 PERL_ARGS_ASSERT_DUMPUNTIL;
20494 #ifdef DEBUG_DUMPUNTIL
20495 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n",indent,node-start,
20496 last ? last-start : 0,plast ? plast-start : 0);
20499 if (plast && plast < last)
20502 while (PL_regkind[op] != END && (!last || node < last)) {
20504 /* While that wasn't END last time... */
20507 if (op == CLOSE || op == WHILEM)
20509 next = regnext((regnode *)node);
20512 if (OP(node) == OPTIMIZED) {
20513 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
20520 regprop(r, sv, node, NULL, NULL);
20521 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
20522 (int)(2*indent + 1), "", SvPVX_const(sv));
20524 if (OP(node) != OPTIMIZED) {
20525 if (next == NULL) /* Next ptr. */
20526 Perl_re_printf( aTHX_ " (0)");
20527 else if (PL_regkind[(U8)op] == BRANCH
20528 && PL_regkind[OP(next)] != BRANCH )
20529 Perl_re_printf( aTHX_ " (FAIL)");
20531 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
20532 Perl_re_printf( aTHX_ "\n");
20536 if (PL_regkind[(U8)op] == BRANCHJ) {
20539 const regnode *nnode = (OP(next) == LONGJMP
20540 ? regnext((regnode *)next)
20542 if (last && nnode > last)
20544 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
20547 else if (PL_regkind[(U8)op] == BRANCH) {
20549 DUMPUNTIL(NEXTOPER(node), next);
20551 else if ( PL_regkind[(U8)op] == TRIE ) {
20552 const regnode *this_trie = node;
20553 const char op = OP(node);
20554 const U32 n = ARG(node);
20555 const reg_ac_data * const ac = op>=AHOCORASICK ?
20556 (reg_ac_data *)ri->data->data[n] :
20558 const reg_trie_data * const trie =
20559 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
20561 AV *const trie_words
20562 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
20564 const regnode *nextbranch= NULL;
20567 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
20568 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
20570 Perl_re_indentf( aTHX_ "%s ",
20573 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
20574 SvCUR(*elem_ptr), 60,
20575 PL_colors[0], PL_colors[1],
20577 ? PERL_PV_ESCAPE_UNI
20579 | PERL_PV_PRETTY_ELLIPSES
20580 | PERL_PV_PRETTY_LTGT
20585 U16 dist= trie->jump[word_idx+1];
20586 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
20587 (UV)((dist ? this_trie + dist : next) - start));
20590 nextbranch= this_trie + trie->jump[0];
20591 DUMPUNTIL(this_trie + dist, nextbranch);
20593 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
20594 nextbranch= regnext((regnode *)nextbranch);
20596 Perl_re_printf( aTHX_ "\n");
20599 if (last && next > last)
20604 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
20605 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
20606 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
20608 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
20610 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
20612 else if ( op == PLUS || op == STAR) {
20613 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
20615 else if (PL_regkind[(U8)op] == ANYOF) {
20616 /* arglen 1 + class block */
20617 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
20618 ? ANYOF_POSIXL_SKIP
20620 node = NEXTOPER(node);
20622 else if (PL_regkind[(U8)op] == EXACT) {
20623 /* Literal string, where present. */
20624 node += NODE_SZ_STR(node) - 1;
20625 node = NEXTOPER(node);
20628 node = NEXTOPER(node);
20629 node += regarglen[(U8)op];
20631 if (op == CURLYX || op == OPEN)
20635 #ifdef DEBUG_DUMPUNTIL
20636 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
20641 #endif /* DEBUGGING */
20644 * ex: set ts=8 sts=4 sw=4 et: