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 */
181 I32 override_recoding;
183 I32 recode_x_to_native;
185 I32 in_multi_char_class;
186 struct reg_code_block *code_blocks; /* positions of literal (?{})
188 int num_code_blocks; /* size of code_blocks[] */
189 int code_index; /* next code_blocks[] slot */
190 SSize_t maxlen; /* mininum possible number of chars in string to match */
191 scan_frame *frame_head;
192 scan_frame *frame_last;
195 #ifdef ADD_TO_REGEXEC
196 char *starttry; /* -Dr: where regtry was called. */
197 #define RExC_starttry (pRExC_state->starttry)
199 SV *runtime_code_qr; /* qr with the runtime code blocks */
201 const char *lastparse;
203 AV *paren_name_list; /* idx -> name */
204 U32 study_chunk_recursed_count;
207 #define RExC_lastparse (pRExC_state->lastparse)
208 #define RExC_lastnum (pRExC_state->lastnum)
209 #define RExC_paren_name_list (pRExC_state->paren_name_list)
210 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
211 #define RExC_mysv (pRExC_state->mysv1)
212 #define RExC_mysv1 (pRExC_state->mysv1)
213 #define RExC_mysv2 (pRExC_state->mysv2)
216 bool seen_unfolded_sharp_s;
221 #define RExC_flags (pRExC_state->flags)
222 #define RExC_pm_flags (pRExC_state->pm_flags)
223 #define RExC_precomp (pRExC_state->precomp)
224 #define RExC_precomp_adj (pRExC_state->precomp_adj)
225 #define RExC_adjusted_start (pRExC_state->adjusted_start)
226 #define RExC_precomp_end (pRExC_state->precomp_end)
227 #define RExC_rx_sv (pRExC_state->rx_sv)
228 #define RExC_rx (pRExC_state->rx)
229 #define RExC_rxi (pRExC_state->rxi)
230 #define RExC_start (pRExC_state->start)
231 #define RExC_end (pRExC_state->end)
232 #define RExC_parse (pRExC_state->parse)
233 #define RExC_whilem_seen (pRExC_state->whilem_seen)
235 /* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
236 * EXACTF node, hence was parsed under /di rules. If later in the parse,
237 * something forces the pattern into using /ui rules, the sharp s should be
238 * folded into the sequence 'ss', which takes up more space than previously
239 * calculated. This means that the sizing pass needs to be restarted. (The
240 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
241 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
242 * so there is no need to resize [perl #125990]. */
243 #define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
245 #ifdef RE_TRACK_PATTERN_OFFSETS
246 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
249 #define RExC_emit (pRExC_state->emit)
250 #define RExC_emit_dummy (pRExC_state->emit_dummy)
251 #define RExC_emit_start (pRExC_state->emit_start)
252 #define RExC_emit_bound (pRExC_state->emit_bound)
253 #define RExC_sawback (pRExC_state->sawback)
254 #define RExC_seen (pRExC_state->seen)
255 #define RExC_size (pRExC_state->size)
256 #define RExC_maxlen (pRExC_state->maxlen)
257 #define RExC_npar (pRExC_state->npar)
258 #define RExC_nestroot (pRExC_state->nestroot)
259 #define RExC_extralen (pRExC_state->extralen)
260 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
261 #define RExC_utf8 (pRExC_state->utf8)
262 #define RExC_uni_semantics (pRExC_state->uni_semantics)
263 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
264 #define RExC_open_parens (pRExC_state->open_parens)
265 #define RExC_close_parens (pRExC_state->close_parens)
266 #define RExC_end_op (pRExC_state->end_op)
267 #define RExC_paren_names (pRExC_state->paren_names)
268 #define RExC_recurse (pRExC_state->recurse)
269 #define RExC_recurse_count (pRExC_state->recurse_count)
270 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
271 #define RExC_study_chunk_recursed_bytes \
272 (pRExC_state->study_chunk_recursed_bytes)
273 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
274 #define RExC_contains_locale (pRExC_state->contains_locale)
275 #define RExC_contains_i (pRExC_state->contains_i)
276 #define RExC_override_recoding (pRExC_state->override_recoding)
278 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
280 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
281 #define RExC_frame_head (pRExC_state->frame_head)
282 #define RExC_frame_last (pRExC_state->frame_last)
283 #define RExC_frame_count (pRExC_state->frame_count)
284 #define RExC_strict (pRExC_state->strict)
285 #define RExC_study_started (pRExC_state->study_started)
286 #define RExC_warn_text (pRExC_state->warn_text)
288 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
289 * a flag to disable back-off on the fixed/floating substrings - if it's
290 * a high complexity pattern we assume the benefit of avoiding a full match
291 * is worth the cost of checking for the substrings even if they rarely help.
293 #define RExC_naughty (pRExC_state->naughty)
294 #define TOO_NAUGHTY (10)
295 #define MARK_NAUGHTY(add) \
296 if (RExC_naughty < TOO_NAUGHTY) \
297 RExC_naughty += (add)
298 #define MARK_NAUGHTY_EXP(exp, add) \
299 if (RExC_naughty < TOO_NAUGHTY) \
300 RExC_naughty += RExC_naughty / (exp) + (add)
302 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
303 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
304 ((*s) == '{' && regcurly(s)))
307 * Flags to be passed up and down.
309 #define WORST 0 /* Worst case. */
310 #define HASWIDTH 0x01 /* Known to match non-null strings. */
312 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
313 * character. (There needs to be a case: in the switch statement in regexec.c
314 * for any node marked SIMPLE.) Note that this is not the same thing as
317 #define SPSTART 0x04 /* Starts with * or + */
318 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
319 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
320 #define RESTART_PASS1 0x20 /* Need to restart sizing pass */
321 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
322 calcuate sizes as UTF-8 */
324 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
326 /* whether trie related optimizations are enabled */
327 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
328 #define TRIE_STUDY_OPT
329 #define FULL_TRIE_STUDY
335 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
336 #define PBITVAL(paren) (1 << ((paren) & 7))
337 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
338 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
339 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
341 #define REQUIRE_UTF8(flagp) STMT_START { \
344 *flagp = RESTART_PASS1|NEED_UTF8; \
349 /* Change from /d into /u rules, and restart the parse if we've already seen
350 * something whose size would increase as a result, by setting *flagp and
351 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
352 * we've change to /u during the parse. */
353 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
355 if (DEPENDS_SEMANTICS) { \
357 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
358 RExC_uni_semantics = 1; \
359 if (RExC_seen_unfolded_sharp_s) { \
360 *flagp |= RESTART_PASS1; \
361 return restart_retval; \
366 /* This converts the named class defined in regcomp.h to its equivalent class
367 * number defined in handy.h. */
368 #define namedclass_to_classnum(class) ((int) ((class) / 2))
369 #define classnum_to_namedclass(classnum) ((classnum) * 2)
371 #define _invlist_union_complement_2nd(a, b, output) \
372 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
373 #define _invlist_intersection_complement_2nd(a, b, output) \
374 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
376 /* About scan_data_t.
378 During optimisation we recurse through the regexp program performing
379 various inplace (keyhole style) optimisations. In addition study_chunk
380 and scan_commit populate this data structure with information about
381 what strings MUST appear in the pattern. We look for the longest
382 string that must appear at a fixed location, and we look for the
383 longest string that may appear at a floating location. So for instance
388 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
389 strings (because they follow a .* construct). study_chunk will identify
390 both FOO and BAR as being the longest fixed and floating strings respectively.
392 The strings can be composites, for instance
396 will result in a composite fixed substring 'foo'.
398 For each string some basic information is maintained:
400 - offset or min_offset
401 This is the position the string must appear at, or not before.
402 It also implicitly (when combined with minlenp) tells us how many
403 characters must match before the string we are searching for.
404 Likewise when combined with minlenp and the length of the string it
405 tells us how many characters must appear after the string we have
409 Only used for floating strings. This is the rightmost point that
410 the string can appear at. If set to SSize_t_MAX it indicates that the
411 string can occur infinitely far to the right.
414 A pointer to the minimum number of characters of the pattern that the
415 string was found inside. This is important as in the case of positive
416 lookahead or positive lookbehind we can have multiple patterns
421 The minimum length of the pattern overall is 3, the minimum length
422 of the lookahead part is 3, but the minimum length of the part that
423 will actually match is 1. So 'FOO's minimum length is 3, but the
424 minimum length for the F is 1. This is important as the minimum length
425 is used to determine offsets in front of and behind the string being
426 looked for. Since strings can be composites this is the length of the
427 pattern at the time it was committed with a scan_commit. Note that
428 the length is calculated by study_chunk, so that the minimum lengths
429 are not known until the full pattern has been compiled, thus the
430 pointer to the value.
434 In the case of lookbehind the string being searched for can be
435 offset past the start point of the final matching string.
436 If this value was just blithely removed from the min_offset it would
437 invalidate some of the calculations for how many chars must match
438 before or after (as they are derived from min_offset and minlen and
439 the length of the string being searched for).
440 When the final pattern is compiled and the data is moved from the
441 scan_data_t structure into the regexp structure the information
442 about lookbehind is factored in, with the information that would
443 have been lost precalculated in the end_shift field for the
446 The fields pos_min and pos_delta are used to store the minimum offset
447 and the delta to the maximum offset at the current point in the pattern.
451 typedef struct scan_data_t {
452 /*I32 len_min; unused */
453 /*I32 len_delta; unused */
457 SSize_t last_end; /* min value, <0 unless valid. */
458 SSize_t last_start_min;
459 SSize_t last_start_max;
460 SV **longest; /* Either &l_fixed, or &l_float. */
461 SV *longest_fixed; /* longest fixed string found in pattern */
462 SSize_t offset_fixed; /* offset where it starts */
463 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
464 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
465 SV *longest_float; /* longest floating string found in pattern */
466 SSize_t offset_float_min; /* earliest point in string it can appear */
467 SSize_t offset_float_max; /* latest point in string it can appear */
468 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
469 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
472 SSize_t *last_closep;
473 regnode_ssc *start_class;
477 * Forward declarations for pregcomp()'s friends.
480 static const scan_data_t zero_scan_data =
481 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
483 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
484 #define SF_BEFORE_SEOL 0x0001
485 #define SF_BEFORE_MEOL 0x0002
486 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
487 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
489 #define SF_FIX_SHIFT_EOL (+2)
490 #define SF_FL_SHIFT_EOL (+4)
492 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
493 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
495 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
496 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
497 #define SF_IS_INF 0x0040
498 #define SF_HAS_PAR 0x0080
499 #define SF_IN_PAR 0x0100
500 #define SF_HAS_EVAL 0x0200
503 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
504 * longest substring in the pattern. When it is not set the optimiser keeps
505 * track of position, but does not keep track of the actual strings seen,
507 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
510 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
511 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
512 * turned off because of the alternation (BRANCH). */
513 #define SCF_DO_SUBSTR 0x0400
515 #define SCF_DO_STCLASS_AND 0x0800
516 #define SCF_DO_STCLASS_OR 0x1000
517 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
518 #define SCF_WHILEM_VISITED_POS 0x2000
520 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
521 #define SCF_SEEN_ACCEPT 0x8000
522 #define SCF_TRIE_DOING_RESTUDY 0x10000
523 #define SCF_IN_DEFINE 0x20000
528 #define UTF cBOOL(RExC_utf8)
530 /* The enums for all these are ordered so things work out correctly */
531 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
532 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
533 == REGEX_DEPENDS_CHARSET)
534 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
535 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
536 >= REGEX_UNICODE_CHARSET)
537 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
538 == REGEX_ASCII_RESTRICTED_CHARSET)
539 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
540 >= REGEX_ASCII_RESTRICTED_CHARSET)
541 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
542 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
544 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
546 /* For programs that want to be strictly Unicode compatible by dying if any
547 * attempt is made to match a non-Unicode code point against a Unicode
549 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
551 #define OOB_NAMEDCLASS -1
553 /* There is no code point that is out-of-bounds, so this is problematic. But
554 * its only current use is to initialize a variable that is always set before
556 #define OOB_UNICODE 0xDEADBEEF
558 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
559 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
562 /* length of regex to show in messages that don't mark a position within */
563 #define RegexLengthToShowInErrorMessages 127
566 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
567 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
568 * op/pragma/warn/regcomp.
570 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
571 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
573 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
574 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
576 /* The code in this file in places uses one level of recursion with parsing
577 * rebased to an alternate string constructed by us in memory. This can take
578 * the form of something that is completely different from the input, or
579 * something that uses the input as part of the alternate. In the first case,
580 * there should be no possibility of an error, as we are in complete control of
581 * the alternate string. But in the second case we don't control the input
582 * portion, so there may be errors in that. Here's an example:
584 * is handled specially because \x{df} folds to a sequence of more than one
585 * character, 'ss'. What is done is to create and parse an alternate string,
586 * which looks like this:
587 * /(?:\x{DF}|[abc\x{DF}def])/ui
588 * where it uses the input unchanged in the middle of something it constructs,
589 * which is a branch for the DF outside the character class, and clustering
590 * parens around the whole thing. (It knows enough to skip the DF inside the
591 * class while in this substitute parse.) 'abc' and 'def' may have errors that
592 * need to be reported. The general situation looks like this:
595 * Input: ----------------------------------------------------
596 * Constructed: ---------------------------------------------------
599 * The input string sI..eI is the input pattern. The string sC..EC is the
600 * constructed substitute parse string. The portions sC..tC and eC..EC are
601 * constructed by us. The portion tC..eC is an exact duplicate of the input
602 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
603 * while parsing, we find an error at xC. We want to display a message showing
604 * the real input string. Thus we need to find the point xI in it which
605 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
606 * been constructed by us, and so shouldn't have errors. We get:
608 * xI = sI + (tI - sI) + (xC - tC)
610 * and, the offset into sI is:
612 * (xI - sI) = (tI - sI) + (xC - tC)
614 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
615 * and we save tC as RExC_adjusted_start.
617 * During normal processing of the input pattern, everything points to that,
618 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
621 #define tI_sI RExC_precomp_adj
622 #define tC RExC_adjusted_start
623 #define sC RExC_precomp
624 #define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
625 #define xI(xC) (sC + xI_offset(xC))
626 #define eC RExC_precomp_end
628 #define REPORT_LOCATION_ARGS(xC) \
630 (xI(xC) > eC) /* Don't run off end */ \
631 ? eC - sC /* Length before the <--HERE */ \
633 sC), /* The input pattern printed up to the <--HERE */ \
635 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
636 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
638 /* Used to point after bad bytes for an error message, but avoid skipping
639 * past a nul byte. */
640 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
643 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
644 * arg. Show regex, up to a maximum length. If it's too long, chop and add
647 #define _FAIL(code) STMT_START { \
648 const char *ellipses = ""; \
649 IV len = RExC_precomp_end - RExC_precomp; \
652 SAVEFREESV(RExC_rx_sv); \
653 if (len > RegexLengthToShowInErrorMessages) { \
654 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
655 len = RegexLengthToShowInErrorMessages - 10; \
661 #define FAIL(msg) _FAIL( \
662 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
663 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
665 #define FAIL2(msg,arg) _FAIL( \
666 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
667 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
670 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
672 #define Simple_vFAIL(m) STMT_START { \
673 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
674 m, REPORT_LOCATION_ARGS(RExC_parse)); \
678 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
680 #define vFAIL(m) STMT_START { \
682 SAVEFREESV(RExC_rx_sv); \
687 * Like Simple_vFAIL(), but accepts two arguments.
689 #define Simple_vFAIL2(m,a1) STMT_START { \
690 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
691 REPORT_LOCATION_ARGS(RExC_parse)); \
695 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
697 #define vFAIL2(m,a1) STMT_START { \
699 SAVEFREESV(RExC_rx_sv); \
700 Simple_vFAIL2(m, a1); \
705 * Like Simple_vFAIL(), but accepts three arguments.
707 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
708 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
709 REPORT_LOCATION_ARGS(RExC_parse)); \
713 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
715 #define vFAIL3(m,a1,a2) STMT_START { \
717 SAVEFREESV(RExC_rx_sv); \
718 Simple_vFAIL3(m, a1, a2); \
722 * Like Simple_vFAIL(), but accepts four arguments.
724 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
725 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
726 REPORT_LOCATION_ARGS(RExC_parse)); \
729 #define vFAIL4(m,a1,a2,a3) STMT_START { \
731 SAVEFREESV(RExC_rx_sv); \
732 Simple_vFAIL4(m, a1, a2, a3); \
735 /* A specialized version of vFAIL2 that works with UTF8f */
736 #define vFAIL2utf8f(m, a1) STMT_START { \
738 SAVEFREESV(RExC_rx_sv); \
739 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
740 REPORT_LOCATION_ARGS(RExC_parse)); \
743 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
745 SAVEFREESV(RExC_rx_sv); \
746 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
747 REPORT_LOCATION_ARGS(RExC_parse)); \
750 /* These have asserts in them because of [perl #122671] Many warnings in
751 * regcomp.c can occur twice. If they get output in pass1 and later in that
752 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
753 * would get output again. So they should be output in pass2, and these
754 * asserts make sure new warnings follow that paradigm. */
756 /* m is not necessarily a "literal string", in this macro */
757 #define reg_warn_non_literal_string(loc, m) STMT_START { \
758 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
759 "%s" REPORT_LOCATION, \
760 m, REPORT_LOCATION_ARGS(loc)); \
763 #define ckWARNreg(loc,m) STMT_START { \
764 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
766 REPORT_LOCATION_ARGS(loc)); \
769 #define vWARN(loc, m) STMT_START { \
770 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
772 REPORT_LOCATION_ARGS(loc)); \
775 #define vWARN_dep(loc, m) STMT_START { \
776 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
778 REPORT_LOCATION_ARGS(loc)); \
781 #define ckWARNdep(loc,m) STMT_START { \
782 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
784 REPORT_LOCATION_ARGS(loc)); \
787 #define ckWARNregdep(loc,m) STMT_START { \
788 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
791 REPORT_LOCATION_ARGS(loc)); \
794 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
795 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
797 a1, REPORT_LOCATION_ARGS(loc)); \
800 #define ckWARN2reg(loc, m, a1) STMT_START { \
801 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
803 a1, REPORT_LOCATION_ARGS(loc)); \
806 #define vWARN3(loc, m, a1, a2) STMT_START { \
807 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
809 a1, a2, REPORT_LOCATION_ARGS(loc)); \
812 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
813 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
816 REPORT_LOCATION_ARGS(loc)); \
819 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
820 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
823 REPORT_LOCATION_ARGS(loc)); \
826 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
827 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
830 REPORT_LOCATION_ARGS(loc)); \
833 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
834 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
837 REPORT_LOCATION_ARGS(loc)); \
840 /* Macros for recording node offsets. 20001227 mjd@plover.com
841 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
842 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
843 * Element 0 holds the number n.
844 * Position is 1 indexed.
846 #ifndef RE_TRACK_PATTERN_OFFSETS
847 #define Set_Node_Offset_To_R(node,byte)
848 #define Set_Node_Offset(node,byte)
849 #define Set_Cur_Node_Offset
850 #define Set_Node_Length_To_R(node,len)
851 #define Set_Node_Length(node,len)
852 #define Set_Node_Cur_Length(node,start)
853 #define Node_Offset(n)
854 #define Node_Length(n)
855 #define Set_Node_Offset_Length(node,offset,len)
856 #define ProgLen(ri) ri->u.proglen
857 #define SetProgLen(ri,x) ri->u.proglen = x
859 #define ProgLen(ri) ri->u.offsets[0]
860 #define SetProgLen(ri,x) ri->u.offsets[0] = x
861 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
863 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
864 __LINE__, (int)(node), (int)(byte))); \
866 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
869 RExC_offsets[2*(node)-1] = (byte); \
874 #define Set_Node_Offset(node,byte) \
875 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
876 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
878 #define Set_Node_Length_To_R(node,len) STMT_START { \
880 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
881 __LINE__, (int)(node), (int)(len))); \
883 Perl_croak(aTHX_ "value of node is %d in Length macro", \
886 RExC_offsets[2*(node)] = (len); \
891 #define Set_Node_Length(node,len) \
892 Set_Node_Length_To_R((node)-RExC_emit_start, len)
893 #define Set_Node_Cur_Length(node, start) \
894 Set_Node_Length(node, RExC_parse - start)
896 /* Get offsets and lengths */
897 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
898 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
900 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
901 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
902 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
906 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
907 #define EXPERIMENTAL_INPLACESCAN
908 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
912 Perl_re_printf(pTHX_ const char *fmt, ...)
916 PerlIO *f= Perl_debug_log;
917 PERL_ARGS_ASSERT_RE_PRINTF;
919 result = PerlIO_vprintf(f, fmt, ap);
925 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
929 PerlIO *f= Perl_debug_log;
930 PERL_ARGS_ASSERT_RE_INDENTF;
932 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
933 result = PerlIO_vprintf(f, fmt, ap);
937 #endif /* DEBUGGING */
939 #define DEBUG_RExC_seen() \
940 DEBUG_OPTIMISE_MORE_r({ \
941 Perl_re_printf( aTHX_ "RExC_seen: "); \
943 if (RExC_seen & REG_ZERO_LEN_SEEN) \
944 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
946 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
947 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
949 if (RExC_seen & REG_GPOS_SEEN) \
950 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
952 if (RExC_seen & REG_RECURSE_SEEN) \
953 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
955 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
956 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
958 if (RExC_seen & REG_VERBARG_SEEN) \
959 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
961 if (RExC_seen & REG_CUTGROUP_SEEN) \
962 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
964 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
965 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
967 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
968 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
970 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
971 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
973 Perl_re_printf( aTHX_ "\n"); \
976 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
977 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
979 #define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
981 Perl_re_printf( aTHX_ "%s", open_str); \
982 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
983 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
984 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
985 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
986 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
987 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
988 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
989 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
990 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
991 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
992 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
993 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
994 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
995 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
996 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
997 Perl_re_printf( aTHX_ "%s", close_str); \
1001 #define DEBUG_STUDYDATA(str,data,depth) \
1002 DEBUG_OPTIMISE_MORE_r(if(data){ \
1003 Perl_re_indentf( aTHX_ "" str "Pos:%"IVdf"/%"IVdf \
1004 " Flags: 0x%"UVXf, \
1006 (IV)((data)->pos_min), \
1007 (IV)((data)->pos_delta), \
1008 (UV)((data)->flags) \
1010 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
1011 Perl_re_printf( aTHX_ \
1012 " Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1013 (IV)((data)->whilem_c), \
1014 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
1015 is_inf ? "INF " : "" \
1017 if ((data)->last_found) \
1018 Perl_re_printf( aTHX_ \
1019 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
1020 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
1021 SvPVX_const((data)->last_found), \
1022 (IV)((data)->last_end), \
1023 (IV)((data)->last_start_min), \
1024 (IV)((data)->last_start_max), \
1025 ((data)->longest && \
1026 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
1027 SvPVX_const((data)->longest_fixed), \
1028 (IV)((data)->offset_fixed), \
1029 ((data)->longest && \
1030 (data)->longest==&((data)->longest_float)) ? "*" : "", \
1031 SvPVX_const((data)->longest_float), \
1032 (IV)((data)->offset_float_min), \
1033 (IV)((data)->offset_float_max) \
1035 Perl_re_printf( aTHX_ "\n"); \
1039 /* =========================================================
1040 * BEGIN edit_distance stuff.
1042 * This calculates how many single character changes of any type are needed to
1043 * transform a string into another one. It is taken from version 3.1 of
1045 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1048 /* Our unsorted dictionary linked list. */
1049 /* Note we use UVs, not chars. */
1054 struct dictionary* next;
1056 typedef struct dictionary item;
1059 PERL_STATIC_INLINE item*
1060 push(UV key,item* curr)
1063 Newxz(head, 1, item);
1071 PERL_STATIC_INLINE item*
1072 find(item* head, UV key)
1074 item* iterator = head;
1076 if (iterator->key == key){
1079 iterator = iterator->next;
1085 PERL_STATIC_INLINE item*
1086 uniquePush(item* head,UV key)
1088 item* iterator = head;
1091 if (iterator->key == key) {
1094 iterator = iterator->next;
1097 return push(key,head);
1100 PERL_STATIC_INLINE void
1101 dict_free(item* head)
1103 item* iterator = head;
1106 item* temp = iterator;
1107 iterator = iterator->next;
1114 /* End of Dictionary Stuff */
1116 /* All calculations/work are done here */
1118 S_edit_distance(const UV* src,
1120 const STRLEN x, /* length of src[] */
1121 const STRLEN y, /* length of tgt[] */
1122 const SSize_t maxDistance
1126 UV swapCount,swapScore,targetCharCount,i,j;
1128 UV score_ceil = x + y;
1130 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1132 /* intialize matrix start values */
1133 Newxz(scores, ( (x + 2) * (y + 2)), UV);
1134 scores[0] = score_ceil;
1135 scores[1 * (y + 2) + 0] = score_ceil;
1136 scores[0 * (y + 2) + 1] = score_ceil;
1137 scores[1 * (y + 2) + 1] = 0;
1138 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1143 for (i=1;i<=x;i++) {
1145 head = uniquePush(head,src[i]);
1146 scores[(i+1) * (y + 2) + 1] = i;
1147 scores[(i+1) * (y + 2) + 0] = score_ceil;
1150 for (j=1;j<=y;j++) {
1153 head = uniquePush(head,tgt[j]);
1154 scores[1 * (y + 2) + (j + 1)] = j;
1155 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1158 targetCharCount = find(head,tgt[j-1])->value;
1159 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1161 if (src[i-1] != tgt[j-1]){
1162 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));
1166 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1170 find(head,src[i-1])->value = i;
1174 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1177 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1181 /* END of edit_distance() stuff
1182 * ========================================================= */
1184 /* is c a control character for which we have a mnemonic? */
1185 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1188 S_cntrl_to_mnemonic(const U8 c)
1190 /* Returns the mnemonic string that represents character 'c', if one
1191 * exists; NULL otherwise. The only ones that exist for the purposes of
1192 * this routine are a few control characters */
1195 case '\a': return "\\a";
1196 case '\b': return "\\b";
1197 case ESC_NATIVE: return "\\e";
1198 case '\f': return "\\f";
1199 case '\n': return "\\n";
1200 case '\r': return "\\r";
1201 case '\t': return "\\t";
1207 /* Mark that we cannot extend a found fixed substring at this point.
1208 Update the longest found anchored substring and the longest found
1209 floating substrings if needed. */
1212 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1213 SSize_t *minlenp, int is_inf)
1215 const STRLEN l = CHR_SVLEN(data->last_found);
1216 const STRLEN old_l = CHR_SVLEN(*data->longest);
1217 GET_RE_DEBUG_FLAGS_DECL;
1219 PERL_ARGS_ASSERT_SCAN_COMMIT;
1221 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1222 SvSetMagicSV(*data->longest, data->last_found);
1223 if (*data->longest == data->longest_fixed) {
1224 data->offset_fixed = l ? data->last_start_min : data->pos_min;
1225 if (data->flags & SF_BEFORE_EOL)
1227 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
1229 data->flags &= ~SF_FIX_BEFORE_EOL;
1230 data->minlen_fixed=minlenp;
1231 data->lookbehind_fixed=0;
1233 else { /* *data->longest == data->longest_float */
1234 data->offset_float_min = l ? data->last_start_min : data->pos_min;
1235 data->offset_float_max = (l
1236 ? data->last_start_max
1237 : (data->pos_delta > SSize_t_MAX - data->pos_min
1239 : data->pos_min + data->pos_delta));
1241 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
1242 data->offset_float_max = SSize_t_MAX;
1243 if (data->flags & SF_BEFORE_EOL)
1245 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
1247 data->flags &= ~SF_FL_BEFORE_EOL;
1248 data->minlen_float=minlenp;
1249 data->lookbehind_float=0;
1252 SvCUR_set(data->last_found, 0);
1254 SV * const sv = data->last_found;
1255 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1256 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1261 data->last_end = -1;
1262 data->flags &= ~SF_BEFORE_EOL;
1263 DEBUG_STUDYDATA("commit: ",data,0);
1266 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1267 * list that describes which code points it matches */
1270 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1272 /* Set the SSC 'ssc' to match an empty string or any code point */
1274 PERL_ARGS_ASSERT_SSC_ANYTHING;
1276 assert(is_ANYOF_SYNTHETIC(ssc));
1278 /* mortalize so won't leak */
1279 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1280 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1284 S_ssc_is_anything(const regnode_ssc *ssc)
1286 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1287 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1288 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1289 * in any way, so there's no point in using it */
1294 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1296 assert(is_ANYOF_SYNTHETIC(ssc));
1298 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1302 /* See if the list consists solely of the range 0 - Infinity */
1303 invlist_iterinit(ssc->invlist);
1304 ret = invlist_iternext(ssc->invlist, &start, &end)
1308 invlist_iterfinish(ssc->invlist);
1314 /* If e.g., both \w and \W are set, matches everything */
1315 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1317 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1318 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1328 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1330 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1331 * string, any code point, or any posix class under locale */
1333 PERL_ARGS_ASSERT_SSC_INIT;
1335 Zero(ssc, 1, regnode_ssc);
1336 set_ANYOF_SYNTHETIC(ssc);
1337 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1340 /* If any portion of the regex is to operate under locale rules that aren't
1341 * fully known at compile time, initialization includes it. The reason
1342 * this isn't done for all regexes is that the optimizer was written under
1343 * the assumption that locale was all-or-nothing. Given the complexity and
1344 * lack of documentation in the optimizer, and that there are inadequate
1345 * test cases for locale, many parts of it may not work properly, it is
1346 * safest to avoid locale unless necessary. */
1347 if (RExC_contains_locale) {
1348 ANYOF_POSIXL_SETALL(ssc);
1351 ANYOF_POSIXL_ZERO(ssc);
1356 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1357 const regnode_ssc *ssc)
1359 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1360 * to the list of code points matched, and locale posix classes; hence does
1361 * not check its flags) */
1366 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1368 assert(is_ANYOF_SYNTHETIC(ssc));
1370 invlist_iterinit(ssc->invlist);
1371 ret = invlist_iternext(ssc->invlist, &start, &end)
1375 invlist_iterfinish(ssc->invlist);
1381 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1389 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1390 const regnode_charclass* const node)
1392 /* Returns a mortal inversion list defining which code points are matched
1393 * by 'node', which is of type ANYOF. Handles complementing the result if
1394 * appropriate. If some code points aren't knowable at this time, the
1395 * returned list must, and will, contain every code point that is a
1399 SV* only_utf8_locale_invlist = NULL;
1401 const U32 n = ARG(node);
1402 bool new_node_has_latin1 = FALSE;
1404 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1406 /* Look at the data structure created by S_set_ANYOF_arg() */
1407 if (n != ANYOF_ONLY_HAS_BITMAP) {
1408 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1409 AV * const av = MUTABLE_AV(SvRV(rv));
1410 SV **const ary = AvARRAY(av);
1411 assert(RExC_rxi->data->what[n] == 's');
1413 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1414 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1416 else if (ary[0] && ary[0] != &PL_sv_undef) {
1418 /* Here, no compile-time swash, and there are things that won't be
1419 * known until runtime -- we have to assume it could be anything */
1420 invlist = sv_2mortal(_new_invlist(1));
1421 return _add_range_to_invlist(invlist, 0, UV_MAX);
1423 else if (ary[3] && ary[3] != &PL_sv_undef) {
1425 /* Here no compile-time swash, and no run-time only data. Use the
1426 * node's inversion list */
1427 invlist = sv_2mortal(invlist_clone(ary[3]));
1430 /* Get the code points valid only under UTF-8 locales */
1431 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1432 && ary[2] && ary[2] != &PL_sv_undef)
1434 only_utf8_locale_invlist = ary[2];
1439 invlist = sv_2mortal(_new_invlist(0));
1442 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1443 * code points, and an inversion list for the others, but if there are code
1444 * points that should match only conditionally on the target string being
1445 * UTF-8, those are placed in the inversion list, and not the bitmap.
1446 * Since there are circumstances under which they could match, they are
1447 * included in the SSC. But if the ANYOF node is to be inverted, we have
1448 * to exclude them here, so that when we invert below, the end result
1449 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1450 * have to do this here before we add the unconditionally matched code
1452 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1453 _invlist_intersection_complement_2nd(invlist,
1458 /* Add in the points from the bit map */
1459 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1460 if (ANYOF_BITMAP_TEST(node, i)) {
1461 unsigned int start = i++;
1463 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1466 invlist = _add_range_to_invlist(invlist, start, i-1);
1467 new_node_has_latin1 = TRUE;
1471 /* If this can match all upper Latin1 code points, have to add them
1472 * as well. But don't add them if inverting, as when that gets done below,
1473 * it would exclude all these characters, including the ones it shouldn't
1474 * that were added just above */
1475 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1476 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1478 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1481 /* Similarly for these */
1482 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1483 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1486 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1487 _invlist_invert(invlist);
1489 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1491 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1492 * locale. We can skip this if there are no 0-255 at all. */
1493 _invlist_union(invlist, PL_Latin1, &invlist);
1496 /* Similarly add the UTF-8 locale possible matches. These have to be
1497 * deferred until after the non-UTF-8 locale ones are taken care of just
1498 * above, or it leads to wrong results under ANYOF_INVERT */
1499 if (only_utf8_locale_invlist) {
1500 _invlist_union_maybe_complement_2nd(invlist,
1501 only_utf8_locale_invlist,
1502 ANYOF_FLAGS(node) & ANYOF_INVERT,
1509 /* These two functions currently do the exact same thing */
1510 #define ssc_init_zero ssc_init
1512 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1513 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1515 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1516 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1517 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1520 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1521 const regnode_charclass *and_with)
1523 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1524 * another SSC or a regular ANYOF class. Can create false positives. */
1529 PERL_ARGS_ASSERT_SSC_AND;
1531 assert(is_ANYOF_SYNTHETIC(ssc));
1533 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1534 * the code point inversion list and just the relevant flags */
1535 if (is_ANYOF_SYNTHETIC(and_with)) {
1536 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1537 anded_flags = ANYOF_FLAGS(and_with);
1539 /* XXX This is a kludge around what appears to be deficiencies in the
1540 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1541 * there are paths through the optimizer where it doesn't get weeded
1542 * out when it should. And if we don't make some extra provision for
1543 * it like the code just below, it doesn't get added when it should.
1544 * This solution is to add it only when AND'ing, which is here, and
1545 * only when what is being AND'ed is the pristine, original node
1546 * matching anything. Thus it is like adding it to ssc_anything() but
1547 * only when the result is to be AND'ed. Probably the same solution
1548 * could be adopted for the same problem we have with /l matching,
1549 * which is solved differently in S_ssc_init(), and that would lead to
1550 * fewer false positives than that solution has. But if this solution
1551 * creates bugs, the consequences are only that a warning isn't raised
1552 * that should be; while the consequences for having /l bugs is
1553 * incorrect matches */
1554 if (ssc_is_anything((regnode_ssc *)and_with)) {
1555 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1559 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1560 if (OP(and_with) == ANYOFD) {
1561 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1564 anded_flags = ANYOF_FLAGS(and_with)
1565 &( ANYOF_COMMON_FLAGS
1566 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1567 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1568 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1570 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1575 ANYOF_FLAGS(ssc) &= anded_flags;
1577 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1578 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1579 * 'and_with' may be inverted. When not inverted, we have the situation of
1581 * (C1 | P1) & (C2 | P2)
1582 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1583 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1584 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1585 * <= ((C1 & C2) | P1 | P2)
1586 * Alternatively, the last few steps could be:
1587 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1588 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1589 * <= (C1 | C2 | (P1 & P2))
1590 * We favor the second approach if either P1 or P2 is non-empty. This is
1591 * because these components are a barrier to doing optimizations, as what
1592 * they match cannot be known until the moment of matching as they are
1593 * dependent on the current locale, 'AND"ing them likely will reduce or
1595 * But we can do better if we know that C1,P1 are in their initial state (a
1596 * frequent occurrence), each matching everything:
1597 * (<everything>) & (C2 | P2) = C2 | P2
1598 * Similarly, if C2,P2 are in their initial state (again a frequent
1599 * occurrence), the result is a no-op
1600 * (C1 | P1) & (<everything>) = C1 | P1
1603 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1604 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1605 * <= (C1 & ~C2) | (P1 & ~P2)
1608 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1609 && ! is_ANYOF_SYNTHETIC(and_with))
1613 ssc_intersection(ssc,
1615 FALSE /* Has already been inverted */
1618 /* If either P1 or P2 is empty, the intersection will be also; can skip
1620 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1621 ANYOF_POSIXL_ZERO(ssc);
1623 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1625 /* Note that the Posix class component P from 'and_with' actually
1627 * P = Pa | Pb | ... | Pn
1628 * where each component is one posix class, such as in [\w\s].
1630 * ~P = ~(Pa | Pb | ... | Pn)
1631 * = ~Pa & ~Pb & ... & ~Pn
1632 * <= ~Pa | ~Pb | ... | ~Pn
1633 * The last is something we can easily calculate, but unfortunately
1634 * is likely to have many false positives. We could do better
1635 * in some (but certainly not all) instances if two classes in
1636 * P have known relationships. For example
1637 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1639 * :lower: & :print: = :lower:
1640 * And similarly for classes that must be disjoint. For example,
1641 * since \s and \w can have no elements in common based on rules in
1642 * the POSIX standard,
1643 * \w & ^\S = nothing
1644 * Unfortunately, some vendor locales do not meet the Posix
1645 * standard, in particular almost everything by Microsoft.
1646 * The loop below just changes e.g., \w into \W and vice versa */
1648 regnode_charclass_posixl temp;
1649 int add = 1; /* To calculate the index of the complement */
1651 ANYOF_POSIXL_ZERO(&temp);
1652 for (i = 0; i < ANYOF_MAX; i++) {
1654 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1655 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1657 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1658 ANYOF_POSIXL_SET(&temp, i + add);
1660 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1662 ANYOF_POSIXL_AND(&temp, ssc);
1664 } /* else ssc already has no posixes */
1665 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1666 in its initial state */
1667 else if (! is_ANYOF_SYNTHETIC(and_with)
1668 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1670 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1671 * copy it over 'ssc' */
1672 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1673 if (is_ANYOF_SYNTHETIC(and_with)) {
1674 StructCopy(and_with, ssc, regnode_ssc);
1677 ssc->invlist = anded_cp_list;
1678 ANYOF_POSIXL_ZERO(ssc);
1679 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1680 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1684 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1685 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1687 /* One or the other of P1, P2 is non-empty. */
1688 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1689 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1691 ssc_union(ssc, anded_cp_list, FALSE);
1693 else { /* P1 = P2 = empty */
1694 ssc_intersection(ssc, anded_cp_list, FALSE);
1700 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1701 const regnode_charclass *or_with)
1703 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1704 * another SSC or a regular ANYOF class. Can create false positives if
1705 * 'or_with' is to be inverted. */
1710 PERL_ARGS_ASSERT_SSC_OR;
1712 assert(is_ANYOF_SYNTHETIC(ssc));
1714 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1715 * the code point inversion list and just the relevant flags */
1716 if (is_ANYOF_SYNTHETIC(or_with)) {
1717 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1718 ored_flags = ANYOF_FLAGS(or_with);
1721 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1722 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1723 if (OP(or_with) != ANYOFD) {
1725 |= ANYOF_FLAGS(or_with)
1726 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1727 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1728 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1730 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1735 ANYOF_FLAGS(ssc) |= ored_flags;
1737 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1738 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1739 * 'or_with' may be inverted. When not inverted, we have the simple
1740 * situation of computing:
1741 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1742 * If P1|P2 yields a situation with both a class and its complement are
1743 * set, like having both \w and \W, this matches all code points, and we
1744 * can delete these from the P component of the ssc going forward. XXX We
1745 * might be able to delete all the P components, but I (khw) am not certain
1746 * about this, and it is better to be safe.
1749 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1750 * <= (C1 | P1) | ~C2
1751 * <= (C1 | ~C2) | P1
1752 * (which results in actually simpler code than the non-inverted case)
1755 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1756 && ! is_ANYOF_SYNTHETIC(or_with))
1758 /* We ignore P2, leaving P1 going forward */
1759 } /* else Not inverted */
1760 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1761 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1762 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1764 for (i = 0; i < ANYOF_MAX; i += 2) {
1765 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1767 ssc_match_all_cp(ssc);
1768 ANYOF_POSIXL_CLEAR(ssc, i);
1769 ANYOF_POSIXL_CLEAR(ssc, i+1);
1777 FALSE /* Already has been inverted */
1781 PERL_STATIC_INLINE void
1782 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1784 PERL_ARGS_ASSERT_SSC_UNION;
1786 assert(is_ANYOF_SYNTHETIC(ssc));
1788 _invlist_union_maybe_complement_2nd(ssc->invlist,
1794 PERL_STATIC_INLINE void
1795 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1797 const bool invert2nd)
1799 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1801 assert(is_ANYOF_SYNTHETIC(ssc));
1803 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1809 PERL_STATIC_INLINE void
1810 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1812 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1814 assert(is_ANYOF_SYNTHETIC(ssc));
1816 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1819 PERL_STATIC_INLINE void
1820 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1822 /* AND just the single code point 'cp' into the SSC 'ssc' */
1824 SV* cp_list = _new_invlist(2);
1826 PERL_ARGS_ASSERT_SSC_CP_AND;
1828 assert(is_ANYOF_SYNTHETIC(ssc));
1830 cp_list = add_cp_to_invlist(cp_list, cp);
1831 ssc_intersection(ssc, cp_list,
1832 FALSE /* Not inverted */
1834 SvREFCNT_dec_NN(cp_list);
1837 PERL_STATIC_INLINE void
1838 S_ssc_clear_locale(regnode_ssc *ssc)
1840 /* Set the SSC 'ssc' to not match any locale things */
1841 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1843 assert(is_ANYOF_SYNTHETIC(ssc));
1845 ANYOF_POSIXL_ZERO(ssc);
1846 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1849 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1852 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1854 /* The synthetic start class is used to hopefully quickly winnow down
1855 * places where a pattern could start a match in the target string. If it
1856 * doesn't really narrow things down that much, there isn't much point to
1857 * having the overhead of using it. This function uses some very crude
1858 * heuristics to decide if to use the ssc or not.
1860 * It returns TRUE if 'ssc' rules out more than half what it considers to
1861 * be the "likely" possible matches, but of course it doesn't know what the
1862 * actual things being matched are going to be; these are only guesses
1864 * For /l matches, it assumes that the only likely matches are going to be
1865 * in the 0-255 range, uniformly distributed, so half of that is 127
1866 * For /a and /d matches, it assumes that the likely matches will be just
1867 * the ASCII range, so half of that is 63
1868 * For /u and there isn't anything matching above the Latin1 range, it
1869 * assumes that that is the only range likely to be matched, and uses
1870 * half that as the cut-off: 127. If anything matches above Latin1,
1871 * it assumes that all of Unicode could match (uniformly), except for
1872 * non-Unicode code points and things in the General Category "Other"
1873 * (unassigned, private use, surrogates, controls and formats). This
1874 * is a much large number. */
1876 U32 count = 0; /* Running total of number of code points matched by
1878 UV start, end; /* Start and end points of current range in inversion
1880 const U32 max_code_points = (LOC)
1882 : (( ! UNI_SEMANTICS
1883 || invlist_highest(ssc->invlist) < 256)
1886 const U32 max_match = max_code_points / 2;
1888 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1890 invlist_iterinit(ssc->invlist);
1891 while (invlist_iternext(ssc->invlist, &start, &end)) {
1892 if (start >= max_code_points) {
1895 end = MIN(end, max_code_points - 1);
1896 count += end - start + 1;
1897 if (count >= max_match) {
1898 invlist_iterfinish(ssc->invlist);
1908 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1910 /* The inversion list in the SSC is marked mortal; now we need a more
1911 * permanent copy, which is stored the same way that is done in a regular
1912 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1915 SV* invlist = invlist_clone(ssc->invlist);
1917 PERL_ARGS_ASSERT_SSC_FINALIZE;
1919 assert(is_ANYOF_SYNTHETIC(ssc));
1921 /* The code in this file assumes that all but these flags aren't relevant
1922 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1923 * by the time we reach here */
1924 assert(! (ANYOF_FLAGS(ssc)
1925 & ~( ANYOF_COMMON_FLAGS
1926 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1927 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
1929 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1931 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1932 NULL, NULL, NULL, FALSE);
1934 /* Make sure is clone-safe */
1935 ssc->invlist = NULL;
1937 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1938 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1941 if (RExC_contains_locale) {
1945 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1948 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1949 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1950 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1951 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1952 ? (TRIE_LIST_CUR( idx ) - 1) \
1958 dump_trie(trie,widecharmap,revcharmap)
1959 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1960 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1962 These routines dump out a trie in a somewhat readable format.
1963 The _interim_ variants are used for debugging the interim
1964 tables that are used to generate the final compressed
1965 representation which is what dump_trie expects.
1967 Part of the reason for their existence is to provide a form
1968 of documentation as to how the different representations function.
1973 Dumps the final compressed table form of the trie to Perl_debug_log.
1974 Used for debugging make_trie().
1978 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1979 AV *revcharmap, U32 depth)
1982 SV *sv=sv_newmortal();
1983 int colwidth= widecharmap ? 6 : 4;
1985 GET_RE_DEBUG_FLAGS_DECL;
1987 PERL_ARGS_ASSERT_DUMP_TRIE;
1989 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
1990 depth+1, "Match","Base","Ofs" );
1992 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1993 SV ** const tmp = av_fetch( revcharmap, state, 0);
1995 Perl_re_printf( aTHX_ "%*s",
1997 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1998 PL_colors[0], PL_colors[1],
1999 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2000 PERL_PV_ESCAPE_FIRSTCHAR
2005 Perl_re_printf( aTHX_ "\n");
2006 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2008 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2009 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2010 Perl_re_printf( aTHX_ "\n");
2012 for( state = 1 ; state < trie->statecount ; state++ ) {
2013 const U32 base = trie->states[ state ].trans.base;
2015 Perl_re_indentf( aTHX_ "#%4"UVXf"|", depth+1, (UV)state);
2017 if ( trie->states[ state ].wordnum ) {
2018 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2020 Perl_re_printf( aTHX_ "%6s", "" );
2023 Perl_re_printf( aTHX_ " @%4"UVXf" ", (UV)base );
2028 while( ( base + ofs < trie->uniquecharcount ) ||
2029 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2030 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2034 Perl_re_printf( aTHX_ "+%2"UVXf"[ ", (UV)ofs);
2036 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2037 if ( ( base + ofs >= trie->uniquecharcount )
2038 && ( base + ofs - trie->uniquecharcount
2040 && trie->trans[ base + ofs
2041 - trie->uniquecharcount ].check == state )
2043 Perl_re_printf( aTHX_ "%*"UVXf, colwidth,
2044 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2047 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
2051 Perl_re_printf( aTHX_ "]");
2054 Perl_re_printf( aTHX_ "\n" );
2056 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2058 for (word=1; word <= trie->wordcount; word++) {
2059 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2060 (int)word, (int)(trie->wordinfo[word].prev),
2061 (int)(trie->wordinfo[word].len));
2063 Perl_re_printf( aTHX_ "\n" );
2066 Dumps a fully constructed but uncompressed trie in list form.
2067 List tries normally only are used for construction when the number of
2068 possible chars (trie->uniquecharcount) is very high.
2069 Used for debugging make_trie().
2072 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2073 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2077 SV *sv=sv_newmortal();
2078 int colwidth= widecharmap ? 6 : 4;
2079 GET_RE_DEBUG_FLAGS_DECL;
2081 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2083 /* print out the table precompression. */
2084 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2086 Perl_re_indentf( aTHX_ "%s",
2087 depth+1, "------:-----+-----------------\n" );
2089 for( state=1 ; state < next_alloc ; state ++ ) {
2092 Perl_re_indentf( aTHX_ " %4"UVXf" :",
2093 depth+1, (UV)state );
2094 if ( ! trie->states[ state ].wordnum ) {
2095 Perl_re_printf( aTHX_ "%5s| ","");
2097 Perl_re_printf( aTHX_ "W%4x| ",
2098 trie->states[ state ].wordnum
2101 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2102 SV ** const tmp = av_fetch( revcharmap,
2103 TRIE_LIST_ITEM(state,charid).forid, 0);
2105 Perl_re_printf( aTHX_ "%*s:%3X=%4"UVXf" | ",
2107 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2109 PL_colors[0], PL_colors[1],
2110 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2111 | PERL_PV_ESCAPE_FIRSTCHAR
2113 TRIE_LIST_ITEM(state,charid).forid,
2114 (UV)TRIE_LIST_ITEM(state,charid).newstate
2117 Perl_re_printf( aTHX_ "\n%*s| ",
2118 (int)((depth * 2) + 14), "");
2121 Perl_re_printf( aTHX_ "\n");
2126 Dumps a fully constructed but uncompressed trie in table form.
2127 This is the normal DFA style state transition table, with a few
2128 twists to facilitate compression later.
2129 Used for debugging make_trie().
2132 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2133 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2138 SV *sv=sv_newmortal();
2139 int colwidth= widecharmap ? 6 : 4;
2140 GET_RE_DEBUG_FLAGS_DECL;
2142 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2145 print out the table precompression so that we can do a visual check
2146 that they are identical.
2149 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2151 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2152 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2154 Perl_re_printf( aTHX_ "%*s",
2156 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2157 PL_colors[0], PL_colors[1],
2158 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2159 PERL_PV_ESCAPE_FIRSTCHAR
2165 Perl_re_printf( aTHX_ "\n");
2166 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2168 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2169 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2172 Perl_re_printf( aTHX_ "\n" );
2174 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2176 Perl_re_indentf( aTHX_ "%4"UVXf" : ",
2178 (UV)TRIE_NODENUM( state ) );
2180 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2181 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2183 Perl_re_printf( aTHX_ "%*"UVXf, colwidth, v );
2185 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2187 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2188 Perl_re_printf( aTHX_ " (%4"UVXf")\n",
2189 (UV)trie->trans[ state ].check );
2191 Perl_re_printf( aTHX_ " (%4"UVXf") W%4X\n",
2192 (UV)trie->trans[ state ].check,
2193 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2201 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2202 startbranch: the first branch in the whole branch sequence
2203 first : start branch of sequence of branch-exact nodes.
2204 May be the same as startbranch
2205 last : Thing following the last branch.
2206 May be the same as tail.
2207 tail : item following the branch sequence
2208 count : words in the sequence
2209 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2210 depth : indent depth
2212 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2214 A trie is an N'ary tree where the branches are determined by digital
2215 decomposition of the key. IE, at the root node you look up the 1st character and
2216 follow that branch repeat until you find the end of the branches. Nodes can be
2217 marked as "accepting" meaning they represent a complete word. Eg:
2221 would convert into the following structure. Numbers represent states, letters
2222 following numbers represent valid transitions on the letter from that state, if
2223 the number is in square brackets it represents an accepting state, otherwise it
2224 will be in parenthesis.
2226 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2230 (1) +-i->(6)-+-s->[7]
2232 +-s->(3)-+-h->(4)-+-e->[5]
2234 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2236 This shows that when matching against the string 'hers' we will begin at state 1
2237 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2238 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2239 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2240 single traverse. We store a mapping from accepting to state to which word was
2241 matched, and then when we have multiple possibilities we try to complete the
2242 rest of the regex in the order in which they occurred in the alternation.
2244 The only prior NFA like behaviour that would be changed by the TRIE support is
2245 the silent ignoring of duplicate alternations which are of the form:
2247 / (DUPE|DUPE) X? (?{ ... }) Y /x
2249 Thus EVAL blocks following a trie may be called a different number of times with
2250 and without the optimisation. With the optimisations dupes will be silently
2251 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2252 the following demonstrates:
2254 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2256 which prints out 'word' three times, but
2258 'words'=~/(word|word|word)(?{ print $1 })S/
2260 which doesnt print it out at all. This is due to other optimisations kicking in.
2262 Example of what happens on a structural level:
2264 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2266 1: CURLYM[1] {1,32767}(18)
2277 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2278 and should turn into:
2280 1: CURLYM[1] {1,32767}(18)
2282 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2290 Cases where tail != last would be like /(?foo|bar)baz/:
2300 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2301 and would end up looking like:
2304 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2311 d = uvchr_to_utf8_flags(d, uv, 0);
2313 is the recommended Unicode-aware way of saying
2318 #define TRIE_STORE_REVCHAR(val) \
2321 SV *zlopp = newSV(UTF8_MAXBYTES); \
2322 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2323 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2324 SvCUR_set(zlopp, kapow - flrbbbbb); \
2327 av_push(revcharmap, zlopp); \
2329 char ooooff = (char)val; \
2330 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2334 /* This gets the next character from the input, folding it if not already
2336 #define TRIE_READ_CHAR STMT_START { \
2339 /* if it is UTF then it is either already folded, or does not need \
2341 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2343 else if (folder == PL_fold_latin1) { \
2344 /* This folder implies Unicode rules, which in the range expressible \
2345 * by not UTF is the lower case, with the two exceptions, one of \
2346 * which should have been taken care of before calling this */ \
2347 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2348 uvc = toLOWER_L1(*uc); \
2349 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2352 /* raw data, will be folded later if needed */ \
2360 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2361 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2362 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2363 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2365 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2366 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2367 TRIE_LIST_CUR( state )++; \
2370 #define TRIE_LIST_NEW(state) STMT_START { \
2371 Newxz( trie->states[ state ].trans.list, \
2372 4, reg_trie_trans_le ); \
2373 TRIE_LIST_CUR( state ) = 1; \
2374 TRIE_LIST_LEN( state ) = 4; \
2377 #define TRIE_HANDLE_WORD(state) STMT_START { \
2378 U16 dupe= trie->states[ state ].wordnum; \
2379 regnode * const noper_next = regnext( noper ); \
2382 /* store the word for dumping */ \
2384 if (OP(noper) != NOTHING) \
2385 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2387 tmp = newSVpvn_utf8( "", 0, UTF ); \
2388 av_push( trie_words, tmp ); \
2392 trie->wordinfo[curword].prev = 0; \
2393 trie->wordinfo[curword].len = wordlen; \
2394 trie->wordinfo[curword].accept = state; \
2396 if ( noper_next < tail ) { \
2398 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2400 trie->jump[curword] = (U16)(noper_next - convert); \
2402 jumper = noper_next; \
2404 nextbranch= regnext(cur); \
2408 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2409 /* chain, so that when the bits of chain are later */\
2410 /* linked together, the dups appear in the chain */\
2411 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2412 trie->wordinfo[dupe].prev = curword; \
2414 /* we haven't inserted this word yet. */ \
2415 trie->states[ state ].wordnum = curword; \
2420 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2421 ( ( base + charid >= ucharcount \
2422 && base + charid < ubound \
2423 && state == trie->trans[ base - ucharcount + charid ].check \
2424 && trie->trans[ base - ucharcount + charid ].next ) \
2425 ? trie->trans[ base - ucharcount + charid ].next \
2426 : ( state==1 ? special : 0 ) \
2430 #define MADE_JUMP_TRIE 2
2431 #define MADE_EXACT_TRIE 4
2434 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2435 regnode *first, regnode *last, regnode *tail,
2436 U32 word_count, U32 flags, U32 depth)
2438 /* first pass, loop through and scan words */
2439 reg_trie_data *trie;
2440 HV *widecharmap = NULL;
2441 AV *revcharmap = newAV();
2447 regnode *jumper = NULL;
2448 regnode *nextbranch = NULL;
2449 regnode *convert = NULL;
2450 U32 *prev_states; /* temp array mapping each state to previous one */
2451 /* we just use folder as a flag in utf8 */
2452 const U8 * folder = NULL;
2455 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2456 AV *trie_words = NULL;
2457 /* along with revcharmap, this only used during construction but both are
2458 * useful during debugging so we store them in the struct when debugging.
2461 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2462 STRLEN trie_charcount=0;
2464 SV *re_trie_maxbuff;
2465 GET_RE_DEBUG_FLAGS_DECL;
2467 PERL_ARGS_ASSERT_MAKE_TRIE;
2469 PERL_UNUSED_ARG(depth);
2473 case EXACT: case EXACTL: break;
2477 case EXACTFLU8: folder = PL_fold_latin1; break;
2478 case EXACTF: folder = PL_fold; break;
2479 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2482 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2484 trie->startstate = 1;
2485 trie->wordcount = word_count;
2486 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2487 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2488 if (flags == EXACT || flags == EXACTL)
2489 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2490 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2491 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2494 trie_words = newAV();
2497 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2498 assert(re_trie_maxbuff);
2499 if (!SvIOK(re_trie_maxbuff)) {
2500 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2502 DEBUG_TRIE_COMPILE_r({
2503 Perl_re_indentf( aTHX_
2504 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2506 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2507 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2510 /* Find the node we are going to overwrite */
2511 if ( first == startbranch && OP( last ) != BRANCH ) {
2512 /* whole branch chain */
2515 /* branch sub-chain */
2516 convert = NEXTOPER( first );
2519 /* -- First loop and Setup --
2521 We first traverse the branches and scan each word to determine if it
2522 contains widechars, and how many unique chars there are, this is
2523 important as we have to build a table with at least as many columns as we
2526 We use an array of integers to represent the character codes 0..255
2527 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2528 the native representation of the character value as the key and IV's for
2531 *TODO* If we keep track of how many times each character is used we can
2532 remap the columns so that the table compression later on is more
2533 efficient in terms of memory by ensuring the most common value is in the
2534 middle and the least common are on the outside. IMO this would be better
2535 than a most to least common mapping as theres a decent chance the most
2536 common letter will share a node with the least common, meaning the node
2537 will not be compressible. With a middle is most common approach the worst
2538 case is when we have the least common nodes twice.
2542 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2543 regnode *noper = NEXTOPER( cur );
2547 U32 wordlen = 0; /* required init */
2548 STRLEN minchars = 0;
2549 STRLEN maxchars = 0;
2550 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2553 if (OP(noper) == NOTHING) {
2554 regnode *noper_next= regnext(noper);
2555 if (noper_next < tail)
2559 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2560 uc= (U8*)STRING(noper);
2561 e= uc + STR_LEN(noper);
2568 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2569 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2570 regardless of encoding */
2571 if (OP( noper ) == EXACTFU_SS) {
2572 /* false positives are ok, so just set this */
2573 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2576 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2578 TRIE_CHARCOUNT(trie)++;
2581 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2582 * is in effect. Under /i, this character can match itself, or
2583 * anything that folds to it. If not under /i, it can match just
2584 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2585 * all fold to k, and all are single characters. But some folds
2586 * expand to more than one character, so for example LATIN SMALL
2587 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2588 * the string beginning at 'uc' is 'ffi', it could be matched by
2589 * three characters, or just by the one ligature character. (It
2590 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2591 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2592 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2593 * match.) The trie needs to know the minimum and maximum number
2594 * of characters that could match so that it can use size alone to
2595 * quickly reject many match attempts. The max is simple: it is
2596 * the number of folded characters in this branch (since a fold is
2597 * never shorter than what folds to it. */
2601 /* And the min is equal to the max if not under /i (indicated by
2602 * 'folder' being NULL), or there are no multi-character folds. If
2603 * there is a multi-character fold, the min is incremented just
2604 * once, for the character that folds to the sequence. Each
2605 * character in the sequence needs to be added to the list below of
2606 * characters in the trie, but we count only the first towards the
2607 * min number of characters needed. This is done through the
2608 * variable 'foldlen', which is returned by the macros that look
2609 * for these sequences as the number of bytes the sequence
2610 * occupies. Each time through the loop, we decrement 'foldlen' by
2611 * how many bytes the current char occupies. Only when it reaches
2612 * 0 do we increment 'minchars' or look for another multi-character
2614 if (folder == NULL) {
2617 else if (foldlen > 0) {
2618 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2623 /* See if *uc is the beginning of a multi-character fold. If
2624 * so, we decrement the length remaining to look at, to account
2625 * for the current character this iteration. (We can use 'uc'
2626 * instead of the fold returned by TRIE_READ_CHAR because for
2627 * non-UTF, the latin1_safe macro is smart enough to account
2628 * for all the unfolded characters, and because for UTF, the
2629 * string will already have been folded earlier in the
2630 * compilation process */
2632 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2633 foldlen -= UTF8SKIP(uc);
2636 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2641 /* The current character (and any potential folds) should be added
2642 * to the possible matching characters for this position in this
2646 U8 folded= folder[ (U8) uvc ];
2647 if ( !trie->charmap[ folded ] ) {
2648 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2649 TRIE_STORE_REVCHAR( folded );
2652 if ( !trie->charmap[ uvc ] ) {
2653 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2654 TRIE_STORE_REVCHAR( uvc );
2657 /* store the codepoint in the bitmap, and its folded
2659 TRIE_BITMAP_SET(trie, uvc);
2661 /* store the folded codepoint */
2662 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2665 /* store first byte of utf8 representation of
2666 variant codepoints */
2667 if (! UVCHR_IS_INVARIANT(uvc)) {
2668 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2671 set_bit = 0; /* We've done our bit :-) */
2675 /* XXX We could come up with the list of code points that fold
2676 * to this using PL_utf8_foldclosures, except not for
2677 * multi-char folds, as there may be multiple combinations
2678 * there that could work, which needs to wait until runtime to
2679 * resolve (The comment about LIGATURE FFI above is such an
2684 widecharmap = newHV();
2686 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2689 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2691 if ( !SvTRUE( *svpp ) ) {
2692 sv_setiv( *svpp, ++trie->uniquecharcount );
2693 TRIE_STORE_REVCHAR(uvc);
2696 } /* end loop through characters in this branch of the trie */
2698 /* We take the min and max for this branch and combine to find the min
2699 * and max for all branches processed so far */
2700 if( cur == first ) {
2701 trie->minlen = minchars;
2702 trie->maxlen = maxchars;
2703 } else if (minchars < trie->minlen) {
2704 trie->minlen = minchars;
2705 } else if (maxchars > trie->maxlen) {
2706 trie->maxlen = maxchars;
2708 } /* end first pass */
2709 DEBUG_TRIE_COMPILE_r(
2710 Perl_re_indentf( aTHX_
2711 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2713 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2714 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2715 (int)trie->minlen, (int)trie->maxlen )
2719 We now know what we are dealing with in terms of unique chars and
2720 string sizes so we can calculate how much memory a naive
2721 representation using a flat table will take. If it's over a reasonable
2722 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2723 conservative but potentially much slower representation using an array
2726 At the end we convert both representations into the same compressed
2727 form that will be used in regexec.c for matching with. The latter
2728 is a form that cannot be used to construct with but has memory
2729 properties similar to the list form and access properties similar
2730 to the table form making it both suitable for fast searches and
2731 small enough that its feasable to store for the duration of a program.
2733 See the comment in the code where the compressed table is produced
2734 inplace from the flat tabe representation for an explanation of how
2735 the compression works.
2740 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2743 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2744 > SvIV(re_trie_maxbuff) )
2747 Second Pass -- Array Of Lists Representation
2749 Each state will be represented by a list of charid:state records
2750 (reg_trie_trans_le) the first such element holds the CUR and LEN
2751 points of the allocated array. (See defines above).
2753 We build the initial structure using the lists, and then convert
2754 it into the compressed table form which allows faster lookups
2755 (but cant be modified once converted).
2758 STRLEN transcount = 1;
2760 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2763 trie->states = (reg_trie_state *)
2764 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2765 sizeof(reg_trie_state) );
2769 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2771 regnode *noper = NEXTOPER( cur );
2772 U32 state = 1; /* required init */
2773 U16 charid = 0; /* sanity init */
2774 U32 wordlen = 0; /* required init */
2776 if (OP(noper) == NOTHING) {
2777 regnode *noper_next= regnext(noper);
2778 if (noper_next < tail)
2782 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2783 const U8 *uc= (U8*)STRING(noper);
2784 const U8 *e= uc + STR_LEN(noper);
2786 for ( ; uc < e ; uc += len ) {
2791 charid = trie->charmap[ uvc ];
2793 SV** const svpp = hv_fetch( widecharmap,
2800 charid=(U16)SvIV( *svpp );
2803 /* charid is now 0 if we dont know the char read, or
2804 * nonzero if we do */
2811 if ( !trie->states[ state ].trans.list ) {
2812 TRIE_LIST_NEW( state );
2815 check <= TRIE_LIST_USED( state );
2818 if ( TRIE_LIST_ITEM( state, check ).forid
2821 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2826 newstate = next_alloc++;
2827 prev_states[newstate] = state;
2828 TRIE_LIST_PUSH( state, charid, newstate );
2833 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2837 TRIE_HANDLE_WORD(state);
2839 } /* end second pass */
2841 /* next alloc is the NEXT state to be allocated */
2842 trie->statecount = next_alloc;
2843 trie->states = (reg_trie_state *)
2844 PerlMemShared_realloc( trie->states,
2846 * sizeof(reg_trie_state) );
2848 /* and now dump it out before we compress it */
2849 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2850 revcharmap, next_alloc,
2854 trie->trans = (reg_trie_trans *)
2855 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2862 for( state=1 ; state < next_alloc ; state ++ ) {
2866 DEBUG_TRIE_COMPILE_MORE_r(
2867 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
2871 if (trie->states[state].trans.list) {
2872 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2876 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2877 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2878 if ( forid < minid ) {
2880 } else if ( forid > maxid ) {
2884 if ( transcount < tp + maxid - minid + 1) {
2886 trie->trans = (reg_trie_trans *)
2887 PerlMemShared_realloc( trie->trans,
2889 * sizeof(reg_trie_trans) );
2890 Zero( trie->trans + (transcount / 2),
2894 base = trie->uniquecharcount + tp - minid;
2895 if ( maxid == minid ) {
2897 for ( ; zp < tp ; zp++ ) {
2898 if ( ! trie->trans[ zp ].next ) {
2899 base = trie->uniquecharcount + zp - minid;
2900 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2902 trie->trans[ zp ].check = state;
2908 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2910 trie->trans[ tp ].check = state;
2915 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2916 const U32 tid = base
2917 - trie->uniquecharcount
2918 + TRIE_LIST_ITEM( state, idx ).forid;
2919 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2921 trie->trans[ tid ].check = state;
2923 tp += ( maxid - minid + 1 );
2925 Safefree(trie->states[ state ].trans.list);
2928 DEBUG_TRIE_COMPILE_MORE_r(
2929 Perl_re_printf( aTHX_ " base: %d\n",base);
2932 trie->states[ state ].trans.base=base;
2934 trie->lasttrans = tp + 1;
2938 Second Pass -- Flat Table Representation.
2940 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2941 each. We know that we will need Charcount+1 trans at most to store
2942 the data (one row per char at worst case) So we preallocate both
2943 structures assuming worst case.
2945 We then construct the trie using only the .next slots of the entry
2948 We use the .check field of the first entry of the node temporarily
2949 to make compression both faster and easier by keeping track of how
2950 many non zero fields are in the node.
2952 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2955 There are two terms at use here: state as a TRIE_NODEIDX() which is
2956 a number representing the first entry of the node, and state as a
2957 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2958 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2959 if there are 2 entrys per node. eg:
2967 The table is internally in the right hand, idx form. However as we
2968 also have to deal with the states array which is indexed by nodenum
2969 we have to use TRIE_NODENUM() to convert.
2972 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
2975 trie->trans = (reg_trie_trans *)
2976 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2977 * trie->uniquecharcount + 1,
2978 sizeof(reg_trie_trans) );
2979 trie->states = (reg_trie_state *)
2980 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2981 sizeof(reg_trie_state) );
2982 next_alloc = trie->uniquecharcount + 1;
2985 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2987 regnode *noper = NEXTOPER( cur );
2989 U32 state = 1; /* required init */
2991 U16 charid = 0; /* sanity init */
2992 U32 accept_state = 0; /* sanity init */
2994 U32 wordlen = 0; /* required init */
2996 if (OP(noper) == NOTHING) {
2997 regnode *noper_next= regnext(noper);
2998 if (noper_next < tail)
3002 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3003 const U8 *uc= (U8*)STRING(noper);
3004 const U8 *e= uc + STR_LEN(noper);
3006 for ( ; uc < e ; uc += len ) {
3011 charid = trie->charmap[ uvc ];
3013 SV* const * const svpp = hv_fetch( widecharmap,
3017 charid = svpp ? (U16)SvIV(*svpp) : 0;
3021 if ( !trie->trans[ state + charid ].next ) {
3022 trie->trans[ state + charid ].next = next_alloc;
3023 trie->trans[ state ].check++;
3024 prev_states[TRIE_NODENUM(next_alloc)]
3025 = TRIE_NODENUM(state);
3026 next_alloc += trie->uniquecharcount;
3028 state = trie->trans[ state + charid ].next;
3030 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
3032 /* charid is now 0 if we dont know the char read, or
3033 * nonzero if we do */
3036 accept_state = TRIE_NODENUM( state );
3037 TRIE_HANDLE_WORD(accept_state);
3039 } /* end second pass */
3041 /* and now dump it out before we compress it */
3042 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3044 next_alloc, depth+1));
3048 * Inplace compress the table.*
3050 For sparse data sets the table constructed by the trie algorithm will
3051 be mostly 0/FAIL transitions or to put it another way mostly empty.
3052 (Note that leaf nodes will not contain any transitions.)
3054 This algorithm compresses the tables by eliminating most such
3055 transitions, at the cost of a modest bit of extra work during lookup:
3057 - Each states[] entry contains a .base field which indicates the
3058 index in the state[] array wheres its transition data is stored.
3060 - If .base is 0 there are no valid transitions from that node.
3062 - If .base is nonzero then charid is added to it to find an entry in
3065 -If trans[states[state].base+charid].check!=state then the
3066 transition is taken to be a 0/Fail transition. Thus if there are fail
3067 transitions at the front of the node then the .base offset will point
3068 somewhere inside the previous nodes data (or maybe even into a node
3069 even earlier), but the .check field determines if the transition is
3073 The following process inplace converts the table to the compressed
3074 table: We first do not compress the root node 1,and mark all its
3075 .check pointers as 1 and set its .base pointer as 1 as well. This
3076 allows us to do a DFA construction from the compressed table later,
3077 and ensures that any .base pointers we calculate later are greater
3080 - We set 'pos' to indicate the first entry of the second node.
3082 - We then iterate over the columns of the node, finding the first and
3083 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3084 and set the .check pointers accordingly, and advance pos
3085 appropriately and repreat for the next node. Note that when we copy
3086 the next pointers we have to convert them from the original
3087 NODEIDX form to NODENUM form as the former is not valid post
3090 - If a node has no transitions used we mark its base as 0 and do not
3091 advance the pos pointer.
3093 - If a node only has one transition we use a second pointer into the
3094 structure to fill in allocated fail transitions from other states.
3095 This pointer is independent of the main pointer and scans forward
3096 looking for null transitions that are allocated to a state. When it
3097 finds one it writes the single transition into the "hole". If the
3098 pointer doesnt find one the single transition is appended as normal.
3100 - Once compressed we can Renew/realloc the structures to release the
3103 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3104 specifically Fig 3.47 and the associated pseudocode.
3108 const U32 laststate = TRIE_NODENUM( next_alloc );
3111 trie->statecount = laststate;
3113 for ( state = 1 ; state < laststate ; state++ ) {
3115 const U32 stateidx = TRIE_NODEIDX( state );
3116 const U32 o_used = trie->trans[ stateidx ].check;
3117 U32 used = trie->trans[ stateidx ].check;
3118 trie->trans[ stateidx ].check = 0;
3121 used && charid < trie->uniquecharcount;
3124 if ( flag || trie->trans[ stateidx + charid ].next ) {
3125 if ( trie->trans[ stateidx + charid ].next ) {
3127 for ( ; zp < pos ; zp++ ) {
3128 if ( ! trie->trans[ zp ].next ) {
3132 trie->states[ state ].trans.base
3134 + trie->uniquecharcount
3136 trie->trans[ zp ].next
3137 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3139 trie->trans[ zp ].check = state;
3140 if ( ++zp > pos ) pos = zp;
3147 trie->states[ state ].trans.base
3148 = pos + trie->uniquecharcount - charid ;
3150 trie->trans[ pos ].next
3151 = SAFE_TRIE_NODENUM(
3152 trie->trans[ stateidx + charid ].next );
3153 trie->trans[ pos ].check = state;
3158 trie->lasttrans = pos + 1;
3159 trie->states = (reg_trie_state *)
3160 PerlMemShared_realloc( trie->states, laststate
3161 * sizeof(reg_trie_state) );
3162 DEBUG_TRIE_COMPILE_MORE_r(
3163 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
3165 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3169 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3172 } /* end table compress */
3174 DEBUG_TRIE_COMPILE_MORE_r(
3175 Perl_re_indentf( aTHX_ "Statecount:%"UVxf" Lasttrans:%"UVxf"\n",
3177 (UV)trie->statecount,
3178 (UV)trie->lasttrans)
3180 /* resize the trans array to remove unused space */
3181 trie->trans = (reg_trie_trans *)
3182 PerlMemShared_realloc( trie->trans, trie->lasttrans
3183 * sizeof(reg_trie_trans) );
3185 { /* Modify the program and insert the new TRIE node */
3186 U8 nodetype =(U8)(flags & 0xFF);
3190 regnode *optimize = NULL;
3191 #ifdef RE_TRACK_PATTERN_OFFSETS
3194 U32 mjd_nodelen = 0;
3195 #endif /* RE_TRACK_PATTERN_OFFSETS */
3196 #endif /* DEBUGGING */
3198 This means we convert either the first branch or the first Exact,
3199 depending on whether the thing following (in 'last') is a branch
3200 or not and whther first is the startbranch (ie is it a sub part of
3201 the alternation or is it the whole thing.)
3202 Assuming its a sub part we convert the EXACT otherwise we convert
3203 the whole branch sequence, including the first.
3205 /* Find the node we are going to overwrite */
3206 if ( first != startbranch || OP( last ) == BRANCH ) {
3207 /* branch sub-chain */
3208 NEXT_OFF( first ) = (U16)(last - first);
3209 #ifdef RE_TRACK_PATTERN_OFFSETS
3211 mjd_offset= Node_Offset((convert));
3212 mjd_nodelen= Node_Length((convert));
3215 /* whole branch chain */
3217 #ifdef RE_TRACK_PATTERN_OFFSETS
3220 const regnode *nop = NEXTOPER( convert );
3221 mjd_offset= Node_Offset((nop));
3222 mjd_nodelen= Node_Length((nop));
3226 Perl_re_indentf( aTHX_ "MJD offset:%"UVuf" MJD length:%"UVuf"\n",
3228 (UV)mjd_offset, (UV)mjd_nodelen)
3231 /* But first we check to see if there is a common prefix we can
3232 split out as an EXACT and put in front of the TRIE node. */
3233 trie->startstate= 1;
3234 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3236 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3240 const U32 base = trie->states[ state ].trans.base;
3242 if ( trie->states[state].wordnum )
3245 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3246 if ( ( base + ofs >= trie->uniquecharcount ) &&
3247 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3248 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3250 if ( ++count > 1 ) {
3251 SV **tmp = av_fetch( revcharmap, ofs, 0);
3252 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
3253 if ( state == 1 ) break;
3255 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3257 Perl_re_indentf( aTHX_ "New Start State=%"UVuf" Class: [",
3261 SV ** const tmp = av_fetch( revcharmap, idx, 0);
3262 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3264 TRIE_BITMAP_SET(trie,*ch);
3266 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3268 Perl_re_printf( aTHX_ "%s", (char*)ch)
3272 TRIE_BITMAP_SET(trie,*ch);
3274 TRIE_BITMAP_SET(trie,folder[ *ch ]);
3275 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3281 SV **tmp = av_fetch( revcharmap, idx, 0);
3283 char *ch = SvPV( *tmp, len );
3285 SV *sv=sv_newmortal();
3286 Perl_re_indentf( aTHX_ "Prefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
3289 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3290 PL_colors[0], PL_colors[1],
3291 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3292 PERL_PV_ESCAPE_FIRSTCHAR
3297 OP( convert ) = nodetype;
3298 str=STRING(convert);
3301 STR_LEN(convert) += len;
3307 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3312 trie->prefixlen = (state-1);
3314 regnode *n = convert+NODE_SZ_STR(convert);
3315 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3316 trie->startstate = state;
3317 trie->minlen -= (state - 1);
3318 trie->maxlen -= (state - 1);
3320 /* At least the UNICOS C compiler choked on this
3321 * being argument to DEBUG_r(), so let's just have
3324 #ifdef PERL_EXT_RE_BUILD
3330 regnode *fix = convert;
3331 U32 word = trie->wordcount;
3333 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3334 while( ++fix < n ) {
3335 Set_Node_Offset_Length(fix, 0, 0);
3338 SV ** const tmp = av_fetch( trie_words, word, 0 );
3340 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3341 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3343 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3351 NEXT_OFF(convert) = (U16)(tail - convert);
3352 DEBUG_r(optimize= n);
3358 if ( trie->maxlen ) {
3359 NEXT_OFF( convert ) = (U16)(tail - convert);
3360 ARG_SET( convert, data_slot );
3361 /* Store the offset to the first unabsorbed branch in
3362 jump[0], which is otherwise unused by the jump logic.
3363 We use this when dumping a trie and during optimisation. */
3365 trie->jump[0] = (U16)(nextbranch - convert);
3367 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3368 * and there is a bitmap
3369 * and the first "jump target" node we found leaves enough room
3370 * then convert the TRIE node into a TRIEC node, with the bitmap
3371 * embedded inline in the opcode - this is hypothetically faster.
3373 if ( !trie->states[trie->startstate].wordnum
3375 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3377 OP( convert ) = TRIEC;
3378 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3379 PerlMemShared_free(trie->bitmap);
3382 OP( convert ) = TRIE;
3384 /* store the type in the flags */
3385 convert->flags = nodetype;
3389 + regarglen[ OP( convert ) ];
3391 /* XXX We really should free up the resource in trie now,
3392 as we won't use them - (which resources?) dmq */
3394 /* needed for dumping*/
3395 DEBUG_r(if (optimize) {
3396 regnode *opt = convert;
3398 while ( ++opt < optimize) {
3399 Set_Node_Offset_Length(opt,0,0);
3402 Try to clean up some of the debris left after the
3405 while( optimize < jumper ) {
3406 mjd_nodelen += Node_Length((optimize));
3407 OP( optimize ) = OPTIMIZED;
3408 Set_Node_Offset_Length(optimize,0,0);
3411 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3413 } /* end node insert */
3415 /* Finish populating the prev field of the wordinfo array. Walk back
3416 * from each accept state until we find another accept state, and if
3417 * so, point the first word's .prev field at the second word. If the
3418 * second already has a .prev field set, stop now. This will be the
3419 * case either if we've already processed that word's accept state,
3420 * or that state had multiple words, and the overspill words were
3421 * already linked up earlier.
3428 for (word=1; word <= trie->wordcount; word++) {
3430 if (trie->wordinfo[word].prev)
3432 state = trie->wordinfo[word].accept;
3434 state = prev_states[state];
3437 prev = trie->states[state].wordnum;
3441 trie->wordinfo[word].prev = prev;
3443 Safefree(prev_states);
3447 /* and now dump out the compressed format */
3448 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3450 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3452 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3453 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3455 SvREFCNT_dec_NN(revcharmap);
3459 : trie->startstate>1
3465 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3467 /* The Trie is constructed and compressed now so we can build a fail array if
3470 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3472 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3476 We find the fail state for each state in the trie, this state is the longest
3477 proper suffix of the current state's 'word' that is also a proper prefix of
3478 another word in our trie. State 1 represents the word '' and is thus the
3479 default fail state. This allows the DFA not to have to restart after its
3480 tried and failed a word at a given point, it simply continues as though it
3481 had been matching the other word in the first place.
3483 'abcdgu'=~/abcdefg|cdgu/
3484 When we get to 'd' we are still matching the first word, we would encounter
3485 'g' which would fail, which would bring us to the state representing 'd' in
3486 the second word where we would try 'g' and succeed, proceeding to match
3489 /* add a fail transition */
3490 const U32 trie_offset = ARG(source);
3491 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3493 const U32 ucharcount = trie->uniquecharcount;
3494 const U32 numstates = trie->statecount;
3495 const U32 ubound = trie->lasttrans + ucharcount;
3499 U32 base = trie->states[ 1 ].trans.base;
3502 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3504 GET_RE_DEBUG_FLAGS_DECL;
3506 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3507 PERL_UNUSED_CONTEXT;
3509 PERL_UNUSED_ARG(depth);
3512 if ( OP(source) == TRIE ) {
3513 struct regnode_1 *op = (struct regnode_1 *)
3514 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3515 StructCopy(source,op,struct regnode_1);
3516 stclass = (regnode *)op;
3518 struct regnode_charclass *op = (struct regnode_charclass *)
3519 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3520 StructCopy(source,op,struct regnode_charclass);
3521 stclass = (regnode *)op;
3523 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3525 ARG_SET( stclass, data_slot );
3526 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3527 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3528 aho->trie=trie_offset;
3529 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3530 Copy( trie->states, aho->states, numstates, reg_trie_state );
3531 Newxz( q, numstates, U32);
3532 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3535 /* initialize fail[0..1] to be 1 so that we always have
3536 a valid final fail state */
3537 fail[ 0 ] = fail[ 1 ] = 1;
3539 for ( charid = 0; charid < ucharcount ; charid++ ) {
3540 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3542 q[ q_write ] = newstate;
3543 /* set to point at the root */
3544 fail[ q[ q_write++ ] ]=1;
3547 while ( q_read < q_write) {
3548 const U32 cur = q[ q_read++ % numstates ];
3549 base = trie->states[ cur ].trans.base;
3551 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3552 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3554 U32 fail_state = cur;
3557 fail_state = fail[ fail_state ];
3558 fail_base = aho->states[ fail_state ].trans.base;
3559 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3561 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3562 fail[ ch_state ] = fail_state;
3563 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3565 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3567 q[ q_write++ % numstates] = ch_state;
3571 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3572 when we fail in state 1, this allows us to use the
3573 charclass scan to find a valid start char. This is based on the principle
3574 that theres a good chance the string being searched contains lots of stuff
3575 that cant be a start char.
3577 fail[ 0 ] = fail[ 1 ] = 0;
3578 DEBUG_TRIE_COMPILE_r({
3579 Perl_re_indentf( aTHX_ "Stclass Failtable (%"UVuf" states): 0",
3580 depth, (UV)numstates
3582 for( q_read=1; q_read<numstates; q_read++ ) {
3583 Perl_re_printf( aTHX_ ", %"UVuf, (UV)fail[q_read]);
3585 Perl_re_printf( aTHX_ "\n");
3588 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3593 #define DEBUG_PEEP(str,scan,depth) \
3594 DEBUG_OPTIMISE_r({if (scan){ \
3595 regnode *Next = regnext(scan); \
3596 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);\
3597 Perl_re_indentf( aTHX_ "" str ">%3d: %s (%d)", \
3598 depth, REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3599 Next ? (REG_NODE_NUM(Next)) : 0 );\
3600 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3601 Perl_re_printf( aTHX_ "\n"); \
3604 /* The below joins as many adjacent EXACTish nodes as possible into a single
3605 * one. The regop may be changed if the node(s) contain certain sequences that
3606 * require special handling. The joining is only done if:
3607 * 1) there is room in the current conglomerated node to entirely contain the
3609 * 2) they are the exact same node type
3611 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3612 * these get optimized out
3614 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3615 * as possible, even if that means splitting an existing node so that its first
3616 * part is moved to the preceeding node. This would maximise the efficiency of
3617 * memEQ during matching. Elsewhere in this file, khw proposes splitting
3618 * EXACTFish nodes into portions that don't change under folding vs those that
3619 * do. Those portions that don't change may be the only things in the pattern that
3620 * could be used to find fixed and floating strings.
3622 * If a node is to match under /i (folded), the number of characters it matches
3623 * can be different than its character length if it contains a multi-character
3624 * fold. *min_subtract is set to the total delta number of characters of the
3627 * And *unfolded_multi_char is set to indicate whether or not the node contains
3628 * an unfolded multi-char fold. This happens when whether the fold is valid or
3629 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3630 * SMALL LETTER SHARP S, as only if the target string being matched against
3631 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3632 * folding rules depend on the locale in force at runtime. (Multi-char folds
3633 * whose components are all above the Latin1 range are not run-time locale
3634 * dependent, and have already been folded by the time this function is
3637 * This is as good a place as any to discuss the design of handling these
3638 * multi-character fold sequences. It's been wrong in Perl for a very long
3639 * time. There are three code points in Unicode whose multi-character folds
3640 * were long ago discovered to mess things up. The previous designs for
3641 * dealing with these involved assigning a special node for them. This
3642 * approach doesn't always work, as evidenced by this example:
3643 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3644 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3645 * would match just the \xDF, it won't be able to handle the case where a
3646 * successful match would have to cross the node's boundary. The new approach
3647 * that hopefully generally solves the problem generates an EXACTFU_SS node
3648 * that is "sss" in this case.
3650 * It turns out that there are problems with all multi-character folds, and not
3651 * just these three. Now the code is general, for all such cases. The
3652 * approach taken is:
3653 * 1) This routine examines each EXACTFish node that could contain multi-
3654 * character folded sequences. Since a single character can fold into
3655 * such a sequence, the minimum match length for this node is less than
3656 * the number of characters in the node. This routine returns in
3657 * *min_subtract how many characters to subtract from the the actual
3658 * length of the string to get a real minimum match length; it is 0 if
3659 * there are no multi-char foldeds. This delta is used by the caller to
3660 * adjust the min length of the match, and the delta between min and max,
3661 * so that the optimizer doesn't reject these possibilities based on size
3663 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3664 * is used for an EXACTFU node that contains at least one "ss" sequence in
3665 * it. For non-UTF-8 patterns and strings, this is the only case where
3666 * there is a possible fold length change. That means that a regular
3667 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3668 * with length changes, and so can be processed faster. regexec.c takes
3669 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3670 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3671 * known until runtime). This saves effort in regex matching. However,
3672 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3673 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3674 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3675 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3676 * possibilities for the non-UTF8 patterns are quite simple, except for
3677 * the sharp s. All the ones that don't involve a UTF-8 target string are
3678 * members of a fold-pair, and arrays are set up for all of them so that
3679 * the other member of the pair can be found quickly. Code elsewhere in
3680 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3681 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3682 * described in the next item.
3683 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3684 * validity of the fold won't be known until runtime, and so must remain
3685 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3686 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3687 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3688 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3689 * The reason this is a problem is that the optimizer part of regexec.c
3690 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3691 * that a character in the pattern corresponds to at most a single
3692 * character in the target string. (And I do mean character, and not byte
3693 * here, unlike other parts of the documentation that have never been
3694 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3695 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3696 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3697 * nodes, violate the assumption, and they are the only instances where it
3698 * is violated. I'm reluctant to try to change the assumption, as the
3699 * code involved is impenetrable to me (khw), so instead the code here
3700 * punts. This routine examines EXACTFL nodes, and (when the pattern
3701 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3702 * boolean indicating whether or not the node contains such a fold. When
3703 * it is true, the caller sets a flag that later causes the optimizer in
3704 * this file to not set values for the floating and fixed string lengths,
3705 * and thus avoids the optimizer code in regexec.c that makes the invalid
3706 * assumption. Thus, there is no optimization based on string lengths for
3707 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3708 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3709 * assumption is wrong only in these cases is that all other non-UTF-8
3710 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3711 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3712 * EXACTF nodes because we don't know at compile time if it actually
3713 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3714 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3715 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3716 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3717 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3718 * string would require the pattern to be forced into UTF-8, the overhead
3719 * of which we want to avoid. Similarly the unfolded multi-char folds in
3720 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3723 * Similarly, the code that generates tries doesn't currently handle
3724 * not-already-folded multi-char folds, and it looks like a pain to change
3725 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3726 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3727 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3728 * using /iaa matching will be doing so almost entirely with ASCII
3729 * strings, so this should rarely be encountered in practice */
3731 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3732 if (PL_regkind[OP(scan)] == EXACT) \
3733 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3736 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3737 UV *min_subtract, bool *unfolded_multi_char,
3738 U32 flags,regnode *val, U32 depth)
3740 /* Merge several consecutive EXACTish nodes into one. */
3741 regnode *n = regnext(scan);
3743 regnode *next = scan + NODE_SZ_STR(scan);
3747 regnode *stop = scan;
3748 GET_RE_DEBUG_FLAGS_DECL;
3750 PERL_UNUSED_ARG(depth);
3753 PERL_ARGS_ASSERT_JOIN_EXACT;
3754 #ifndef EXPERIMENTAL_INPLACESCAN
3755 PERL_UNUSED_ARG(flags);
3756 PERL_UNUSED_ARG(val);
3758 DEBUG_PEEP("join",scan,depth);
3760 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3761 * EXACT ones that are mergeable to the current one. */
3763 && (PL_regkind[OP(n)] == NOTHING
3764 || (stringok && OP(n) == OP(scan)))
3766 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3769 if (OP(n) == TAIL || n > next)
3771 if (PL_regkind[OP(n)] == NOTHING) {
3772 DEBUG_PEEP("skip:",n,depth);
3773 NEXT_OFF(scan) += NEXT_OFF(n);
3774 next = n + NODE_STEP_REGNODE;
3781 else if (stringok) {
3782 const unsigned int oldl = STR_LEN(scan);
3783 regnode * const nnext = regnext(n);
3785 /* XXX I (khw) kind of doubt that this works on platforms (should
3786 * Perl ever run on one) where U8_MAX is above 255 because of lots
3787 * of other assumptions */
3788 /* Don't join if the sum can't fit into a single node */
3789 if (oldl + STR_LEN(n) > U8_MAX)
3792 DEBUG_PEEP("merg",n,depth);
3795 NEXT_OFF(scan) += NEXT_OFF(n);
3796 STR_LEN(scan) += STR_LEN(n);
3797 next = n + NODE_SZ_STR(n);
3798 /* Now we can overwrite *n : */
3799 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3807 #ifdef EXPERIMENTAL_INPLACESCAN
3808 if (flags && !NEXT_OFF(n)) {
3809 DEBUG_PEEP("atch", val, depth);
3810 if (reg_off_by_arg[OP(n)]) {
3811 ARG_SET(n, val - n);
3814 NEXT_OFF(n) = val - n;
3822 *unfolded_multi_char = FALSE;
3824 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3825 * can now analyze for sequences of problematic code points. (Prior to
3826 * this final joining, sequences could have been split over boundaries, and
3827 * hence missed). The sequences only happen in folding, hence for any
3828 * non-EXACT EXACTish node */
3829 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3830 U8* s0 = (U8*) STRING(scan);
3832 U8* s_end = s0 + STR_LEN(scan);
3834 int total_count_delta = 0; /* Total delta number of characters that
3835 multi-char folds expand to */
3837 /* One pass is made over the node's string looking for all the
3838 * possibilities. To avoid some tests in the loop, there are two main
3839 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3844 if (OP(scan) == EXACTFL) {
3847 /* An EXACTFL node would already have been changed to another
3848 * node type unless there is at least one character in it that
3849 * is problematic; likely a character whose fold definition
3850 * won't be known until runtime, and so has yet to be folded.
3851 * For all but the UTF-8 locale, folds are 1-1 in length, but
3852 * to handle the UTF-8 case, we need to create a temporary
3853 * folded copy using UTF-8 locale rules in order to analyze it.
3854 * This is because our macros that look to see if a sequence is
3855 * a multi-char fold assume everything is folded (otherwise the
3856 * tests in those macros would be too complicated and slow).
3857 * Note that here, the non-problematic folds will have already
3858 * been done, so we can just copy such characters. We actually
3859 * don't completely fold the EXACTFL string. We skip the
3860 * unfolded multi-char folds, as that would just create work
3861 * below to figure out the size they already are */
3863 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3866 STRLEN s_len = UTF8SKIP(s);
3867 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3868 Copy(s, d, s_len, U8);
3871 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3872 *unfolded_multi_char = TRUE;
3873 Copy(s, d, s_len, U8);
3876 else if (isASCII(*s)) {
3877 *(d++) = toFOLD(*s);
3881 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3887 /* Point the remainder of the routine to look at our temporary
3891 } /* End of creating folded copy of EXACTFL string */
3893 /* Examine the string for a multi-character fold sequence. UTF-8
3894 * patterns have all characters pre-folded by the time this code is
3896 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3897 length sequence we are looking for is 2 */
3899 int count = 0; /* How many characters in a multi-char fold */
3900 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3901 if (! len) { /* Not a multi-char fold: get next char */
3906 /* Nodes with 'ss' require special handling, except for
3907 * EXACTFA-ish for which there is no multi-char fold to this */
3908 if (len == 2 && *s == 's' && *(s+1) == 's'
3909 && OP(scan) != EXACTFA
3910 && OP(scan) != EXACTFA_NO_TRIE)
3913 if (OP(scan) != EXACTFL) {
3914 OP(scan) = EXACTFU_SS;
3918 else { /* Here is a generic multi-char fold. */
3919 U8* multi_end = s + len;
3921 /* Count how many characters are in it. In the case of
3922 * /aa, no folds which contain ASCII code points are
3923 * allowed, so check for those, and skip if found. */
3924 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3925 count = utf8_length(s, multi_end);
3929 while (s < multi_end) {
3932 goto next_iteration;
3942 /* The delta is how long the sequence is minus 1 (1 is how long
3943 * the character that folds to the sequence is) */
3944 total_count_delta += count - 1;
3948 /* We created a temporary folded copy of the string in EXACTFL
3949 * nodes. Therefore we need to be sure it doesn't go below zero,
3950 * as the real string could be shorter */
3951 if (OP(scan) == EXACTFL) {
3952 int total_chars = utf8_length((U8*) STRING(scan),
3953 (U8*) STRING(scan) + STR_LEN(scan));
3954 if (total_count_delta > total_chars) {
3955 total_count_delta = total_chars;
3959 *min_subtract += total_count_delta;
3962 else if (OP(scan) == EXACTFA) {
3964 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3965 * fold to the ASCII range (and there are no existing ones in the
3966 * upper latin1 range). But, as outlined in the comments preceding
3967 * this function, we need to flag any occurrences of the sharp s.
3968 * This character forbids trie formation (because of added
3970 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
3971 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
3972 || UNICODE_DOT_DOT_VERSION > 0)
3974 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3975 OP(scan) = EXACTFA_NO_TRIE;
3976 *unfolded_multi_char = TRUE;
3984 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3985 * folds that are all Latin1. As explained in the comments
3986 * preceding this function, we look also for the sharp s in EXACTF
3987 * and EXACTFL nodes; it can be in the final position. Otherwise
3988 * we can stop looking 1 byte earlier because have to find at least
3989 * two characters for a multi-fold */
3990 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3995 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3996 if (! len) { /* Not a multi-char fold. */
3997 if (*s == LATIN_SMALL_LETTER_SHARP_S
3998 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4000 *unfolded_multi_char = TRUE;
4007 && isALPHA_FOLD_EQ(*s, 's')
4008 && isALPHA_FOLD_EQ(*(s+1), 's'))
4011 /* EXACTF nodes need to know that the minimum length
4012 * changed so that a sharp s in the string can match this
4013 * ss in the pattern, but they remain EXACTF nodes, as they
4014 * won't match this unless the target string is is UTF-8,
4015 * which we don't know until runtime. EXACTFL nodes can't
4016 * transform into EXACTFU nodes */
4017 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4018 OP(scan) = EXACTFU_SS;
4022 *min_subtract += len - 1;
4030 /* Allow dumping but overwriting the collection of skipped
4031 * ops and/or strings with fake optimized ops */
4032 n = scan + NODE_SZ_STR(scan);
4040 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
4044 /* REx optimizer. Converts nodes into quicker variants "in place".
4045 Finds fixed substrings. */
4047 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4048 to the position after last scanned or to NULL. */
4050 #define INIT_AND_WITHP \
4051 assert(!and_withp); \
4052 Newx(and_withp,1, regnode_ssc); \
4053 SAVEFREEPV(and_withp)
4057 S_unwind_scan_frames(pTHX_ const void *p)
4059 scan_frame *f= (scan_frame *)p;
4061 scan_frame *n= f->next_frame;
4069 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4070 SSize_t *minlenp, SSize_t *deltap,
4075 regnode_ssc *and_withp,
4076 U32 flags, U32 depth)
4077 /* scanp: Start here (read-write). */
4078 /* deltap: Write maxlen-minlen here. */
4079 /* last: Stop before this one. */
4080 /* data: string data about the pattern */
4081 /* stopparen: treat close N as END */
4082 /* recursed: which subroutines have we recursed into */
4083 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4085 /* There must be at least this number of characters to match */
4088 regnode *scan = *scanp, *next;
4090 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4091 int is_inf_internal = 0; /* The studied chunk is infinite */
4092 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4093 scan_data_t data_fake;
4094 SV *re_trie_maxbuff = NULL;
4095 regnode *first_non_open = scan;
4096 SSize_t stopmin = SSize_t_MAX;
4097 scan_frame *frame = NULL;
4098 GET_RE_DEBUG_FLAGS_DECL;
4100 PERL_ARGS_ASSERT_STUDY_CHUNK;
4101 RExC_study_started= 1;
4105 while (first_non_open && OP(first_non_open) == OPEN)
4106 first_non_open=regnext(first_non_open);
4112 RExC_study_chunk_recursed_count++;
4114 DEBUG_OPTIMISE_MORE_r(
4116 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4117 depth, (long)stopparen,
4118 (unsigned long)RExC_study_chunk_recursed_count,
4119 (unsigned long)depth, (unsigned long)recursed_depth,
4122 if (recursed_depth) {
4125 for ( j = 0 ; j < recursed_depth ; j++ ) {
4126 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
4128 PAREN_TEST(RExC_study_chunk_recursed +
4129 ( j * RExC_study_chunk_recursed_bytes), i )
4132 !PAREN_TEST(RExC_study_chunk_recursed +
4133 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
4136 Perl_re_printf( aTHX_ " %d",(int)i);
4140 if ( j + 1 < recursed_depth ) {
4141 Perl_re_printf( aTHX_ ",");
4145 Perl_re_printf( aTHX_ "\n");
4148 while ( scan && OP(scan) != END && scan < last ){
4149 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4150 node length to get a real minimum (because
4151 the folded version may be shorter) */
4152 bool unfolded_multi_char = FALSE;
4153 /* Peephole optimizer: */
4154 DEBUG_STUDYDATA("Peep:", data, depth);
4155 DEBUG_PEEP("Peep", scan, depth);
4158 /* The reason we do this here is that we need to deal with things like
4159 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4160 * parsing code, as each (?:..) is handled by a different invocation of
4163 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
4165 /* Follow the next-chain of the current node and optimize
4166 away all the NOTHINGs from it. */
4167 if (OP(scan) != CURLYX) {
4168 const int max = (reg_off_by_arg[OP(scan)]
4170 /* I32 may be smaller than U16 on CRAYs! */
4171 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4172 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
4176 /* Skip NOTHING and LONGJMP. */
4177 while ((n = regnext(n))
4178 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4179 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
4180 && off + noff < max)
4182 if (reg_off_by_arg[OP(scan)])
4185 NEXT_OFF(scan) = off;
4188 /* The principal pseudo-switch. Cannot be a switch, since we
4189 look into several different things. */
4190 if ( OP(scan) == DEFINEP ) {
4192 SSize_t deltanext = 0;
4193 SSize_t fake_last_close = 0;
4194 I32 f = SCF_IN_DEFINE;
4196 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4197 scan = regnext(scan);
4198 assert( OP(scan) == IFTHEN );
4199 DEBUG_PEEP("expect IFTHEN", scan, depth);
4201 data_fake.last_closep= &fake_last_close;
4203 next = regnext(scan);
4204 scan = NEXTOPER(NEXTOPER(scan));
4205 DEBUG_PEEP("scan", scan, depth);
4206 DEBUG_PEEP("next", next, depth);
4208 /* we suppose the run is continuous, last=next...
4209 * NOTE we dont use the return here! */
4210 (void)study_chunk(pRExC_state, &scan, &minlen,
4211 &deltanext, next, &data_fake, stopparen,
4212 recursed_depth, NULL, f, depth+1);
4217 OP(scan) == BRANCH ||
4218 OP(scan) == BRANCHJ ||
4221 next = regnext(scan);
4224 /* The op(next)==code check below is to see if we
4225 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4226 * IFTHEN is special as it might not appear in pairs.
4227 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4228 * we dont handle it cleanly. */
4229 if (OP(next) == code || code == IFTHEN) {
4230 /* NOTE - There is similar code to this block below for
4231 * handling TRIE nodes on a re-study. If you change stuff here
4232 * check there too. */
4233 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
4235 regnode * const startbranch=scan;
4237 if (flags & SCF_DO_SUBSTR) {
4238 /* Cannot merge strings after this. */
4239 scan_commit(pRExC_state, data, minlenp, is_inf);
4242 if (flags & SCF_DO_STCLASS)
4243 ssc_init_zero(pRExC_state, &accum);
4245 while (OP(scan) == code) {
4246 SSize_t deltanext, minnext, fake;
4248 regnode_ssc this_class;
4250 DEBUG_PEEP("Branch", scan, depth);
4253 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4255 data_fake.whilem_c = data->whilem_c;
4256 data_fake.last_closep = data->last_closep;
4259 data_fake.last_closep = &fake;
4261 data_fake.pos_delta = delta;
4262 next = regnext(scan);
4264 scan = NEXTOPER(scan); /* everything */
4265 if (code != BRANCH) /* everything but BRANCH */
4266 scan = NEXTOPER(scan);
4268 if (flags & SCF_DO_STCLASS) {
4269 ssc_init(pRExC_state, &this_class);
4270 data_fake.start_class = &this_class;
4271 f = SCF_DO_STCLASS_AND;
4273 if (flags & SCF_WHILEM_VISITED_POS)
4274 f |= SCF_WHILEM_VISITED_POS;
4276 /* we suppose the run is continuous, last=next...*/
4277 minnext = study_chunk(pRExC_state, &scan, minlenp,
4278 &deltanext, next, &data_fake, stopparen,
4279 recursed_depth, NULL, f,depth+1);
4283 if (deltanext == SSize_t_MAX) {
4284 is_inf = is_inf_internal = 1;
4286 } else if (max1 < minnext + deltanext)
4287 max1 = minnext + deltanext;
4289 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4291 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4292 if ( stopmin > minnext)
4293 stopmin = min + min1;
4294 flags &= ~SCF_DO_SUBSTR;
4296 data->flags |= SCF_SEEN_ACCEPT;
4299 if (data_fake.flags & SF_HAS_EVAL)
4300 data->flags |= SF_HAS_EVAL;
4301 data->whilem_c = data_fake.whilem_c;
4303 if (flags & SCF_DO_STCLASS)
4304 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4306 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4308 if (flags & SCF_DO_SUBSTR) {
4309 data->pos_min += min1;
4310 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
4311 data->pos_delta = SSize_t_MAX;
4313 data->pos_delta += max1 - min1;
4314 if (max1 != min1 || is_inf)
4315 data->longest = &(data->longest_float);
4318 if (delta == SSize_t_MAX
4319 || SSize_t_MAX - delta - (max1 - min1) < 0)
4320 delta = SSize_t_MAX;
4322 delta += max1 - min1;
4323 if (flags & SCF_DO_STCLASS_OR) {
4324 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4326 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4327 flags &= ~SCF_DO_STCLASS;
4330 else if (flags & SCF_DO_STCLASS_AND) {
4332 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4333 flags &= ~SCF_DO_STCLASS;
4336 /* Switch to OR mode: cache the old value of
4337 * data->start_class */
4339 StructCopy(data->start_class, and_withp, regnode_ssc);
4340 flags &= ~SCF_DO_STCLASS_AND;
4341 StructCopy(&accum, data->start_class, regnode_ssc);
4342 flags |= SCF_DO_STCLASS_OR;
4346 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4347 OP( startbranch ) == BRANCH )
4351 Assuming this was/is a branch we are dealing with: 'scan'
4352 now points at the item that follows the branch sequence,
4353 whatever it is. We now start at the beginning of the
4354 sequence and look for subsequences of
4360 which would be constructed from a pattern like
4363 If we can find such a subsequence we need to turn the first
4364 element into a trie and then add the subsequent branch exact
4365 strings to the trie.
4369 1. patterns where the whole set of branches can be
4372 2. patterns where only a subset can be converted.
4374 In case 1 we can replace the whole set with a single regop
4375 for the trie. In case 2 we need to keep the start and end
4378 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4379 becomes BRANCH TRIE; BRANCH X;
4381 There is an additional case, that being where there is a
4382 common prefix, which gets split out into an EXACT like node
4383 preceding the TRIE node.
4385 If x(1..n)==tail then we can do a simple trie, if not we make
4386 a "jump" trie, such that when we match the appropriate word
4387 we "jump" to the appropriate tail node. Essentially we turn
4388 a nested if into a case structure of sorts.
4393 if (!re_trie_maxbuff) {
4394 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4395 if (!SvIOK(re_trie_maxbuff))
4396 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4398 if ( SvIV(re_trie_maxbuff)>=0 ) {
4400 regnode *first = (regnode *)NULL;
4401 regnode *last = (regnode *)NULL;
4402 regnode *tail = scan;
4406 /* var tail is used because there may be a TAIL
4407 regop in the way. Ie, the exacts will point to the
4408 thing following the TAIL, but the last branch will
4409 point at the TAIL. So we advance tail. If we
4410 have nested (?:) we may have to move through several
4414 while ( OP( tail ) == TAIL ) {
4415 /* this is the TAIL generated by (?:) */
4416 tail = regnext( tail );
4420 DEBUG_TRIE_COMPILE_r({
4421 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4422 Perl_re_indentf( aTHX_ "%s %"UVuf":%s\n",
4424 "Looking for TRIE'able sequences. Tail node is ",
4425 (UV)(tail - RExC_emit_start),
4426 SvPV_nolen_const( RExC_mysv )
4432 Step through the branches
4433 cur represents each branch,
4434 noper is the first thing to be matched as part
4436 noper_next is the regnext() of that node.
4438 We normally handle a case like this
4439 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4440 support building with NOJUMPTRIE, which restricts
4441 the trie logic to structures like /FOO|BAR/.
4443 If noper is a trieable nodetype then the branch is
4444 a possible optimization target. If we are building
4445 under NOJUMPTRIE then we require that noper_next is
4446 the same as scan (our current position in the regex
4449 Once we have two or more consecutive such branches
4450 we can create a trie of the EXACT's contents and
4451 stitch it in place into the program.
4453 If the sequence represents all of the branches in
4454 the alternation we replace the entire thing with a
4457 Otherwise when it is a subsequence we need to
4458 stitch it in place and replace only the relevant
4459 branches. This means the first branch has to remain
4460 as it is used by the alternation logic, and its
4461 next pointer, and needs to be repointed at the item
4462 on the branch chain following the last branch we
4463 have optimized away.
4465 This could be either a BRANCH, in which case the
4466 subsequence is internal, or it could be the item
4467 following the branch sequence in which case the
4468 subsequence is at the end (which does not
4469 necessarily mean the first node is the start of the
4472 TRIE_TYPE(X) is a define which maps the optype to a
4476 ----------------+-----------
4480 EXACTFU_SS | EXACTFU
4483 EXACTFLU8 | EXACTFLU8
4487 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4489 : ( EXACT == (X) ) \
4491 : ( EXACTFU == (X) || EXACTFU_SS == (X) ) \
4493 : ( EXACTFA == (X) ) \
4495 : ( EXACTL == (X) ) \
4497 : ( EXACTFLU8 == (X) ) \
4501 /* dont use tail as the end marker for this traverse */
4502 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4503 regnode * const noper = NEXTOPER( cur );
4504 U8 noper_type = OP( noper );
4505 U8 noper_trietype = TRIE_TYPE( noper_type );
4506 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4507 regnode * const noper_next = regnext( noper );
4508 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4509 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4512 DEBUG_TRIE_COMPILE_r({
4513 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4514 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4516 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4518 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4519 Perl_re_printf( aTHX_ " -> %d:%s",
4520 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
4523 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4524 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
4525 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
4527 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
4528 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4529 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4533 /* Is noper a trieable nodetype that can be merged
4534 * with the current trie (if there is one)? */
4538 ( noper_trietype == NOTHING )
4539 || ( trietype == NOTHING )
4540 || ( trietype == noper_trietype )
4543 && noper_next >= tail
4547 /* Handle mergable triable node Either we are
4548 * the first node in a new trieable sequence,
4549 * in which case we do some bookkeeping,
4550 * otherwise we update the end pointer. */
4553 if ( noper_trietype == NOTHING ) {
4554 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4555 regnode * const noper_next = regnext( noper );
4556 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4557 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4560 if ( noper_next_trietype ) {
4561 trietype = noper_next_trietype;
4562 } else if (noper_next_type) {
4563 /* a NOTHING regop is 1 regop wide.
4564 * We need at least two for a trie
4565 * so we can't merge this in */
4569 trietype = noper_trietype;
4572 if ( trietype == NOTHING )
4573 trietype = noper_trietype;
4578 } /* end handle mergable triable node */
4580 /* handle unmergable node -
4581 * noper may either be a triable node which can
4582 * not be tried together with the current trie,
4583 * or a non triable node */
4585 /* If last is set and trietype is not
4586 * NOTHING then we have found at least two
4587 * triable branch sequences in a row of a
4588 * similar trietype so we can turn them
4589 * into a trie. If/when we allow NOTHING to
4590 * start a trie sequence this condition
4591 * will be required, and it isn't expensive
4592 * so we leave it in for now. */
4593 if ( trietype && trietype != NOTHING )
4594 make_trie( pRExC_state,
4595 startbranch, first, cur, tail,
4596 count, trietype, depth+1 );
4597 last = NULL; /* note: we clear/update
4598 first, trietype etc below,
4599 so we dont do it here */
4603 && noper_next >= tail
4606 /* noper is triable, so we can start a new
4610 trietype = noper_trietype;
4612 /* if we already saw a first but the
4613 * current node is not triable then we have
4614 * to reset the first information. */
4619 } /* end handle unmergable node */
4620 } /* loop over branches */
4621 DEBUG_TRIE_COMPILE_r({
4622 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4623 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
4624 depth+1, SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4625 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
4626 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4627 PL_reg_name[trietype]
4631 if ( last && trietype ) {
4632 if ( trietype != NOTHING ) {
4633 /* the last branch of the sequence was part of
4634 * a trie, so we have to construct it here
4635 * outside of the loop */
4636 made= make_trie( pRExC_state, startbranch,
4637 first, scan, tail, count,
4638 trietype, depth+1 );
4639 #ifdef TRIE_STUDY_OPT
4640 if ( ((made == MADE_EXACT_TRIE &&
4641 startbranch == first)
4642 || ( first_non_open == first )) &&
4644 flags |= SCF_TRIE_RESTUDY;
4645 if ( startbranch == first
4648 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4653 /* at this point we know whatever we have is a
4654 * NOTHING sequence/branch AND if 'startbranch'
4655 * is 'first' then we can turn the whole thing
4658 if ( startbranch == first ) {
4660 /* the entire thing is a NOTHING sequence,
4661 * something like this: (?:|) So we can
4662 * turn it into a plain NOTHING op. */
4663 DEBUG_TRIE_COMPILE_r({
4664 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4665 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
4667 SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4670 OP(startbranch)= NOTHING;
4671 NEXT_OFF(startbranch)= tail - startbranch;
4672 for ( opt= startbranch + 1; opt < tail ; opt++ )
4676 } /* end if ( last) */
4677 } /* TRIE_MAXBUF is non zero */
4682 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4683 scan = NEXTOPER(NEXTOPER(scan));
4684 } else /* single branch is optimized. */
4685 scan = NEXTOPER(scan);
4687 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
4689 regnode *start = NULL;
4690 regnode *end = NULL;
4691 U32 my_recursed_depth= recursed_depth;
4693 if (OP(scan) != SUSPEND) { /* GOSUB */
4694 /* Do setup, note this code has side effects beyond
4695 * the rest of this block. Specifically setting
4696 * RExC_recurse[] must happen at least once during
4699 RExC_recurse[ARG2L(scan)] = scan;
4700 start = RExC_open_parens[paren];
4701 end = RExC_close_parens[paren];
4703 /* NOTE we MUST always execute the above code, even
4704 * if we do nothing with a GOSUB */
4706 ( flags & SCF_IN_DEFINE )
4709 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
4711 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
4714 /* no need to do anything here if we are in a define. */
4715 /* or we are after some kind of infinite construct
4716 * so we can skip recursing into this item.
4717 * Since it is infinite we will not change the maxlen
4718 * or delta, and if we miss something that might raise
4719 * the minlen it will merely pessimise a little.
4721 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
4722 * might result in a minlen of 1 and not of 4,
4723 * but this doesn't make us mismatch, just try a bit
4724 * harder than we should.
4726 scan= regnext(scan);
4733 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4735 /* it is quite possible that there are more efficient ways
4736 * to do this. We maintain a bitmap per level of recursion
4737 * of which patterns we have entered so we can detect if a
4738 * pattern creates a possible infinite loop. When we
4739 * recurse down a level we copy the previous levels bitmap
4740 * down. When we are at recursion level 0 we zero the top
4741 * level bitmap. It would be nice to implement a different
4742 * more efficient way of doing this. In particular the top
4743 * level bitmap may be unnecessary.
4745 if (!recursed_depth) {
4746 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4748 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4749 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4750 RExC_study_chunk_recursed_bytes, U8);
4752 /* we havent recursed into this paren yet, so recurse into it */
4753 DEBUG_STUDYDATA("gosub-set:", data,depth);
4754 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4755 my_recursed_depth= recursed_depth + 1;
4757 DEBUG_STUDYDATA("gosub-inf:", data,depth);
4758 /* some form of infinite recursion, assume infinite length
4760 if (flags & SCF_DO_SUBSTR) {
4761 scan_commit(pRExC_state, data, minlenp, is_inf);
4762 data->longest = &(data->longest_float);
4764 is_inf = is_inf_internal = 1;
4765 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4766 ssc_anything(data->start_class);
4767 flags &= ~SCF_DO_STCLASS;
4769 start= NULL; /* reset start so we dont recurse later on. */
4774 end = regnext(scan);
4777 scan_frame *newframe;
4779 if (!RExC_frame_last) {
4780 Newxz(newframe, 1, scan_frame);
4781 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
4782 RExC_frame_head= newframe;
4784 } else if (!RExC_frame_last->next_frame) {
4785 Newxz(newframe,1,scan_frame);
4786 RExC_frame_last->next_frame= newframe;
4787 newframe->prev_frame= RExC_frame_last;
4790 newframe= RExC_frame_last->next_frame;
4792 RExC_frame_last= newframe;
4794 newframe->next_regnode = regnext(scan);
4795 newframe->last_regnode = last;
4796 newframe->stopparen = stopparen;
4797 newframe->prev_recursed_depth = recursed_depth;
4798 newframe->this_prev_frame= frame;
4800 DEBUG_STUDYDATA("frame-new:",data,depth);
4801 DEBUG_PEEP("fnew", scan, depth);
4808 recursed_depth= my_recursed_depth;
4813 else if (OP(scan) == EXACT || OP(scan) == EXACTL) {
4814 SSize_t l = STR_LEN(scan);
4817 const U8 * const s = (U8*)STRING(scan);
4818 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4819 l = utf8_length(s, s + l);
4821 uc = *((U8*)STRING(scan));
4824 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4825 /* The code below prefers earlier match for fixed
4826 offset, later match for variable offset. */
4827 if (data->last_end == -1) { /* Update the start info. */
4828 data->last_start_min = data->pos_min;
4829 data->last_start_max = is_inf
4830 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4832 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4834 SvUTF8_on(data->last_found);
4836 SV * const sv = data->last_found;
4837 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4838 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4839 if (mg && mg->mg_len >= 0)
4840 mg->mg_len += utf8_length((U8*)STRING(scan),
4841 (U8*)STRING(scan)+STR_LEN(scan));
4843 data->last_end = data->pos_min + l;
4844 data->pos_min += l; /* As in the first entry. */
4845 data->flags &= ~SF_BEFORE_EOL;
4848 /* ANDing the code point leaves at most it, and not in locale, and
4849 * can't match null string */
4850 if (flags & SCF_DO_STCLASS_AND) {
4851 ssc_cp_and(data->start_class, uc);
4852 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4853 ssc_clear_locale(data->start_class);
4855 else if (flags & SCF_DO_STCLASS_OR) {
4856 ssc_add_cp(data->start_class, uc);
4857 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4859 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4860 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4862 flags &= ~SCF_DO_STCLASS;
4864 else if (PL_regkind[OP(scan)] == EXACT) {
4865 /* But OP != EXACT!, so is EXACTFish */
4866 SSize_t l = STR_LEN(scan);
4867 const U8 * s = (U8*)STRING(scan);
4869 /* Search for fixed substrings supports EXACT only. */
4870 if (flags & SCF_DO_SUBSTR) {
4872 scan_commit(pRExC_state, data, minlenp, is_inf);
4875 l = utf8_length(s, s + l);
4877 if (unfolded_multi_char) {
4878 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4880 min += l - min_subtract;
4882 delta += min_subtract;
4883 if (flags & SCF_DO_SUBSTR) {
4884 data->pos_min += l - min_subtract;
4885 if (data->pos_min < 0) {
4888 data->pos_delta += min_subtract;
4890 data->longest = &(data->longest_float);
4894 if (flags & SCF_DO_STCLASS) {
4895 SV* EXACTF_invlist = _make_exactf_invlist(pRExC_state, scan);
4897 assert(EXACTF_invlist);
4898 if (flags & SCF_DO_STCLASS_AND) {
4899 if (OP(scan) != EXACTFL)
4900 ssc_clear_locale(data->start_class);
4901 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4902 ANYOF_POSIXL_ZERO(data->start_class);
4903 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4905 else { /* SCF_DO_STCLASS_OR */
4906 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4907 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4909 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4910 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4912 flags &= ~SCF_DO_STCLASS;
4913 SvREFCNT_dec(EXACTF_invlist);
4916 else if (REGNODE_VARIES(OP(scan))) {
4917 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4918 I32 fl = 0, f = flags;
4919 regnode * const oscan = scan;
4920 regnode_ssc this_class;
4921 regnode_ssc *oclass = NULL;
4922 I32 next_is_eval = 0;
4924 switch (PL_regkind[OP(scan)]) {
4925 case WHILEM: /* End of (?:...)* . */
4926 scan = NEXTOPER(scan);
4929 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4930 next = NEXTOPER(scan);
4931 if (OP(next) == EXACT
4932 || OP(next) == EXACTL
4933 || (flags & SCF_DO_STCLASS))
4936 maxcount = REG_INFTY;
4937 next = regnext(scan);
4938 scan = NEXTOPER(scan);
4942 if (flags & SCF_DO_SUBSTR)
4947 if (flags & SCF_DO_STCLASS) {
4949 maxcount = REG_INFTY;
4950 next = regnext(scan);
4951 scan = NEXTOPER(scan);
4954 if (flags & SCF_DO_SUBSTR) {
4955 scan_commit(pRExC_state, data, minlenp, is_inf);
4956 /* Cannot extend fixed substrings */
4957 data->longest = &(data->longest_float);
4959 is_inf = is_inf_internal = 1;
4960 scan = regnext(scan);
4961 goto optimize_curly_tail;
4963 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
4964 && (scan->flags == stopparen))
4969 mincount = ARG1(scan);
4970 maxcount = ARG2(scan);
4972 next = regnext(scan);
4973 if (OP(scan) == CURLYX) {
4974 I32 lp = (data ? *(data->last_closep) : 0);
4975 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
4977 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
4978 next_is_eval = (OP(scan) == EVAL);
4980 if (flags & SCF_DO_SUBSTR) {
4982 scan_commit(pRExC_state, data, minlenp, is_inf);
4983 /* Cannot extend fixed substrings */
4984 pos_before = data->pos_min;
4988 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
4990 data->flags |= SF_IS_INF;
4992 if (flags & SCF_DO_STCLASS) {
4993 ssc_init(pRExC_state, &this_class);
4994 oclass = data->start_class;
4995 data->start_class = &this_class;
4996 f |= SCF_DO_STCLASS_AND;
4997 f &= ~SCF_DO_STCLASS_OR;
4999 /* Exclude from super-linear cache processing any {n,m}
5000 regops for which the combination of input pos and regex
5001 pos is not enough information to determine if a match
5004 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5005 regex pos at the \s*, the prospects for a match depend not
5006 only on the input position but also on how many (bar\s*)
5007 repeats into the {4,8} we are. */
5008 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5009 f &= ~SCF_WHILEM_VISITED_POS;
5011 /* This will finish on WHILEM, setting scan, or on NULL: */
5012 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5013 last, data, stopparen, recursed_depth, NULL,
5015 ? (f & ~SCF_DO_SUBSTR)
5019 if (flags & SCF_DO_STCLASS)
5020 data->start_class = oclass;
5021 if (mincount == 0 || minnext == 0) {
5022 if (flags & SCF_DO_STCLASS_OR) {
5023 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5025 else if (flags & SCF_DO_STCLASS_AND) {
5026 /* Switch to OR mode: cache the old value of
5027 * data->start_class */
5029 StructCopy(data->start_class, and_withp, regnode_ssc);
5030 flags &= ~SCF_DO_STCLASS_AND;
5031 StructCopy(&this_class, data->start_class, regnode_ssc);
5032 flags |= SCF_DO_STCLASS_OR;
5033 ANYOF_FLAGS(data->start_class)
5034 |= SSC_MATCHES_EMPTY_STRING;
5036 } else { /* Non-zero len */
5037 if (flags & SCF_DO_STCLASS_OR) {
5038 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5039 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5041 else if (flags & SCF_DO_STCLASS_AND)
5042 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5043 flags &= ~SCF_DO_STCLASS;
5045 if (!scan) /* It was not CURLYX, but CURLY. */
5047 if (!(flags & SCF_TRIE_DOING_RESTUDY)
5048 /* ? quantifier ok, except for (?{ ... }) */
5049 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5050 && (minnext == 0) && (deltanext == 0)
5051 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5052 && maxcount <= REG_INFTY/3) /* Complement check for big
5055 /* Fatal warnings may leak the regexp without this: */
5056 SAVEFREESV(RExC_rx_sv);
5057 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5058 "Quantifier unexpected on zero-length expression "
5059 "in regex m/%"UTF8f"/",
5060 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5062 (void)ReREFCNT_inc(RExC_rx_sv);
5065 min += minnext * mincount;
5066 is_inf_internal |= deltanext == SSize_t_MAX
5067 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5068 is_inf |= is_inf_internal;
5070 delta = SSize_t_MAX;
5072 delta += (minnext + deltanext) * maxcount
5073 - minnext * mincount;
5075 /* Try powerful optimization CURLYX => CURLYN. */
5076 if ( OP(oscan) == CURLYX && data
5077 && data->flags & SF_IN_PAR
5078 && !(data->flags & SF_HAS_EVAL)
5079 && !deltanext && minnext == 1 ) {
5080 /* Try to optimize to CURLYN. */
5081 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5082 regnode * const nxt1 = nxt;
5089 if (!REGNODE_SIMPLE(OP(nxt))
5090 && !(PL_regkind[OP(nxt)] == EXACT
5091 && STR_LEN(nxt) == 1))
5097 if (OP(nxt) != CLOSE)
5099 if (RExC_open_parens) {
5100 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5101 RExC_close_parens[ARG(nxt1)]=nxt+2; /*close->while*/
5103 /* Now we know that nxt2 is the only contents: */
5104 oscan->flags = (U8)ARG(nxt);
5106 OP(nxt1) = NOTHING; /* was OPEN. */
5109 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5110 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5111 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5112 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5113 OP(nxt + 1) = OPTIMIZED; /* was count. */
5114 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5119 /* Try optimization CURLYX => CURLYM. */
5120 if ( OP(oscan) == CURLYX && data
5121 && !(data->flags & SF_HAS_PAR)
5122 && !(data->flags & SF_HAS_EVAL)
5123 && !deltanext /* atom is fixed width */
5124 && minnext != 0 /* CURLYM can't handle zero width */
5126 /* Nor characters whose fold at run-time may be
5127 * multi-character */
5128 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5130 /* XXXX How to optimize if data == 0? */
5131 /* Optimize to a simpler form. */
5132 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5136 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5137 && (OP(nxt2) != WHILEM))
5139 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5140 /* Need to optimize away parenths. */
5141 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5142 /* Set the parenth number. */
5143 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5145 oscan->flags = (U8)ARG(nxt);
5146 if (RExC_open_parens) {
5147 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5148 RExC_close_parens[ARG(nxt1)]=nxt2+1; /*close->NOTHING*/
5150 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5151 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5154 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5155 OP(nxt + 1) = OPTIMIZED; /* was count. */
5156 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5157 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5160 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5161 regnode *nnxt = regnext(nxt1);
5163 if (reg_off_by_arg[OP(nxt1)])
5164 ARG_SET(nxt1, nxt2 - nxt1);
5165 else if (nxt2 - nxt1 < U16_MAX)
5166 NEXT_OFF(nxt1) = nxt2 - nxt1;
5168 OP(nxt) = NOTHING; /* Cannot beautify */
5173 /* Optimize again: */
5174 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5175 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
5180 else if ((OP(oscan) == CURLYX)
5181 && (flags & SCF_WHILEM_VISITED_POS)
5182 /* See the comment on a similar expression above.
5183 However, this time it's not a subexpression
5184 we care about, but the expression itself. */
5185 && (maxcount == REG_INFTY)
5186 && data && ++data->whilem_c < 16) {
5187 /* This stays as CURLYX, we can put the count/of pair. */
5188 /* Find WHILEM (as in regexec.c) */
5189 regnode *nxt = oscan + NEXT_OFF(oscan);
5191 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5193 PREVOPER(nxt)->flags = (U8)(data->whilem_c
5194 | (RExC_whilem_seen << 4)); /* On WHILEM */
5196 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5198 if (flags & SCF_DO_SUBSTR) {
5199 SV *last_str = NULL;
5200 STRLEN last_chrs = 0;
5201 int counted = mincount != 0;
5203 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5205 SSize_t b = pos_before >= data->last_start_min
5206 ? pos_before : data->last_start_min;
5208 const char * const s = SvPV_const(data->last_found, l);
5209 SSize_t old = b - data->last_start_min;
5212 old = utf8_hop((U8*)s, old) - (U8*)s;
5214 /* Get the added string: */
5215 last_str = newSVpvn_utf8(s + old, l, UTF);
5216 last_chrs = UTF ? utf8_length((U8*)(s + old),
5217 (U8*)(s + old + l)) : l;
5218 if (deltanext == 0 && pos_before == b) {
5219 /* What was added is a constant string */
5222 SvGROW(last_str, (mincount * l) + 1);
5223 repeatcpy(SvPVX(last_str) + l,
5224 SvPVX_const(last_str), l,
5226 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5227 /* Add additional parts. */
5228 SvCUR_set(data->last_found,
5229 SvCUR(data->last_found) - l);
5230 sv_catsv(data->last_found, last_str);
5232 SV * sv = data->last_found;
5234 SvUTF8(sv) && SvMAGICAL(sv) ?
5235 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5236 if (mg && mg->mg_len >= 0)
5237 mg->mg_len += last_chrs * (mincount-1);
5239 last_chrs *= mincount;
5240 data->last_end += l * (mincount - 1);
5243 /* start offset must point into the last copy */
5244 data->last_start_min += minnext * (mincount - 1);
5245 data->last_start_max =
5248 : data->last_start_max +
5249 (maxcount - 1) * (minnext + data->pos_delta);
5252 /* It is counted once already... */
5253 data->pos_min += minnext * (mincount - counted);
5255 Perl_re_printf( aTHX_ "counted=%"UVuf" deltanext=%"UVuf
5256 " SSize_t_MAX=%"UVuf" minnext=%"UVuf
5257 " maxcount=%"UVuf" mincount=%"UVuf"\n",
5258 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
5260 if (deltanext != SSize_t_MAX)
5261 Perl_re_printf( aTHX_ "LHS=%"UVuf" RHS=%"UVuf"\n",
5262 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5263 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
5265 if (deltanext == SSize_t_MAX
5266 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
5267 data->pos_delta = SSize_t_MAX;
5269 data->pos_delta += - counted * deltanext +
5270 (minnext + deltanext) * maxcount - minnext * mincount;
5271 if (mincount != maxcount) {
5272 /* Cannot extend fixed substrings found inside
5274 scan_commit(pRExC_state, data, minlenp, is_inf);
5275 if (mincount && last_str) {
5276 SV * const sv = data->last_found;
5277 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5278 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5282 sv_setsv(sv, last_str);
5283 data->last_end = data->pos_min;
5284 data->last_start_min = data->pos_min - last_chrs;
5285 data->last_start_max = is_inf
5287 : data->pos_min + data->pos_delta - last_chrs;
5289 data->longest = &(data->longest_float);
5291 SvREFCNT_dec(last_str);
5293 if (data && (fl & SF_HAS_EVAL))
5294 data->flags |= SF_HAS_EVAL;
5295 optimize_curly_tail:
5296 if (OP(oscan) != CURLYX) {
5297 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
5299 NEXT_OFF(oscan) += NEXT_OFF(next);
5305 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5310 if (flags & SCF_DO_SUBSTR) {
5311 /* Cannot expect anything... */
5312 scan_commit(pRExC_state, data, minlenp, is_inf);
5313 data->longest = &(data->longest_float);
5315 is_inf = is_inf_internal = 1;
5316 if (flags & SCF_DO_STCLASS_OR) {
5317 if (OP(scan) == CLUMP) {
5318 /* Actually is any start char, but very few code points
5319 * aren't start characters */
5320 ssc_match_all_cp(data->start_class);
5323 ssc_anything(data->start_class);
5326 flags &= ~SCF_DO_STCLASS;
5330 else if (OP(scan) == LNBREAK) {
5331 if (flags & SCF_DO_STCLASS) {
5332 if (flags & SCF_DO_STCLASS_AND) {
5333 ssc_intersection(data->start_class,
5334 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5335 ssc_clear_locale(data->start_class);
5336 ANYOF_FLAGS(data->start_class)
5337 &= ~SSC_MATCHES_EMPTY_STRING;
5339 else if (flags & SCF_DO_STCLASS_OR) {
5340 ssc_union(data->start_class,
5341 PL_XPosix_ptrs[_CC_VERTSPACE],
5343 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5345 /* See commit msg for
5346 * 749e076fceedeb708a624933726e7989f2302f6a */
5347 ANYOF_FLAGS(data->start_class)
5348 &= ~SSC_MATCHES_EMPTY_STRING;
5350 flags &= ~SCF_DO_STCLASS;
5353 if (delta != SSize_t_MAX)
5354 delta++; /* Because of the 2 char string cr-lf */
5355 if (flags & SCF_DO_SUBSTR) {
5356 /* Cannot expect anything... */
5357 scan_commit(pRExC_state, data, minlenp, is_inf);
5359 data->pos_delta += 1;
5360 data->longest = &(data->longest_float);
5363 else if (REGNODE_SIMPLE(OP(scan))) {
5365 if (flags & SCF_DO_SUBSTR) {
5366 scan_commit(pRExC_state, data, minlenp, is_inf);
5370 if (flags & SCF_DO_STCLASS) {
5372 SV* my_invlist = NULL;
5375 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5376 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5378 /* Some of the logic below assumes that switching
5379 locale on will only add false positives. */
5384 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5388 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5389 ssc_match_all_cp(data->start_class);
5394 SV* REG_ANY_invlist = _new_invlist(2);
5395 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5397 if (flags & SCF_DO_STCLASS_OR) {
5398 ssc_union(data->start_class,
5400 TRUE /* TRUE => invert, hence all but \n
5404 else if (flags & SCF_DO_STCLASS_AND) {
5405 ssc_intersection(data->start_class,
5407 TRUE /* TRUE => invert */
5409 ssc_clear_locale(data->start_class);
5411 SvREFCNT_dec_NN(REG_ANY_invlist);
5418 if (flags & SCF_DO_STCLASS_AND)
5419 ssc_and(pRExC_state, data->start_class,
5420 (regnode_charclass *) scan);
5422 ssc_or(pRExC_state, data->start_class,
5423 (regnode_charclass *) scan);
5431 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5432 if (flags & SCF_DO_STCLASS_AND) {
5433 bool was_there = cBOOL(
5434 ANYOF_POSIXL_TEST(data->start_class,
5436 ANYOF_POSIXL_ZERO(data->start_class);
5437 if (was_there) { /* Do an AND */
5438 ANYOF_POSIXL_SET(data->start_class, namedclass);
5440 /* No individual code points can now match */
5441 data->start_class->invlist
5442 = sv_2mortal(_new_invlist(0));
5445 int complement = namedclass + ((invert) ? -1 : 1);
5447 assert(flags & SCF_DO_STCLASS_OR);
5449 /* If the complement of this class was already there,
5450 * the result is that they match all code points,
5451 * (\d + \D == everything). Remove the classes from
5452 * future consideration. Locale is not relevant in
5454 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5455 ssc_match_all_cp(data->start_class);
5456 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5457 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5459 else { /* The usual case; just add this class to the
5461 ANYOF_POSIXL_SET(data->start_class, namedclass);
5466 case NPOSIXA: /* For these, we always know the exact set of
5471 if (FLAGS(scan) == _CC_ASCII) {
5472 my_invlist = invlist_clone(PL_XPosix_ptrs[_CC_ASCII]);
5475 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
5476 PL_XPosix_ptrs[_CC_ASCII],
5487 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5489 /* NPOSIXD matches all upper Latin1 code points unless the
5490 * target string being matched is UTF-8, which is
5491 * unknowable until match time. Since we are going to
5492 * invert, we want to get rid of all of them so that the
5493 * inversion will match all */
5494 if (OP(scan) == NPOSIXD) {
5495 _invlist_subtract(my_invlist, PL_UpperLatin1,
5501 if (flags & SCF_DO_STCLASS_AND) {
5502 ssc_intersection(data->start_class, my_invlist, invert);
5503 ssc_clear_locale(data->start_class);
5506 assert(flags & SCF_DO_STCLASS_OR);
5507 ssc_union(data->start_class, my_invlist, invert);
5509 SvREFCNT_dec(my_invlist);
5511 if (flags & SCF_DO_STCLASS_OR)
5512 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5513 flags &= ~SCF_DO_STCLASS;
5516 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5517 data->flags |= (OP(scan) == MEOL
5520 scan_commit(pRExC_state, data, minlenp, is_inf);
5523 else if ( PL_regkind[OP(scan)] == BRANCHJ
5524 /* Lookbehind, or need to calculate parens/evals/stclass: */
5525 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5526 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5528 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5529 || OP(scan) == UNLESSM )
5531 /* Negative Lookahead/lookbehind
5532 In this case we can't do fixed string optimisation.
5535 SSize_t deltanext, minnext, fake = 0;
5540 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5542 data_fake.whilem_c = data->whilem_c;
5543 data_fake.last_closep = data->last_closep;
5546 data_fake.last_closep = &fake;
5547 data_fake.pos_delta = delta;
5548 if ( flags & SCF_DO_STCLASS && !scan->flags
5549 && OP(scan) == IFMATCH ) { /* Lookahead */
5550 ssc_init(pRExC_state, &intrnl);
5551 data_fake.start_class = &intrnl;
5552 f |= SCF_DO_STCLASS_AND;
5554 if (flags & SCF_WHILEM_VISITED_POS)
5555 f |= SCF_WHILEM_VISITED_POS;
5556 next = regnext(scan);
5557 nscan = NEXTOPER(NEXTOPER(scan));
5558 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5559 last, &data_fake, stopparen,
5560 recursed_depth, NULL, f, depth+1);
5563 FAIL("Variable length lookbehind not implemented");
5565 else if (minnext > (I32)U8_MAX) {
5566 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5569 scan->flags = (U8)minnext;
5572 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5574 if (data_fake.flags & SF_HAS_EVAL)
5575 data->flags |= SF_HAS_EVAL;
5576 data->whilem_c = data_fake.whilem_c;
5578 if (f & SCF_DO_STCLASS_AND) {
5579 if (flags & SCF_DO_STCLASS_OR) {
5580 /* OR before, AND after: ideally we would recurse with
5581 * data_fake to get the AND applied by study of the
5582 * remainder of the pattern, and then derecurse;
5583 * *** HACK *** for now just treat as "no information".
5584 * See [perl #56690].
5586 ssc_init(pRExC_state, data->start_class);
5588 /* AND before and after: combine and continue. These
5589 * assertions are zero-length, so can match an EMPTY
5591 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5592 ANYOF_FLAGS(data->start_class)
5593 |= SSC_MATCHES_EMPTY_STRING;
5597 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5599 /* Positive Lookahead/lookbehind
5600 In this case we can do fixed string optimisation,
5601 but we must be careful about it. Note in the case of
5602 lookbehind the positions will be offset by the minimum
5603 length of the pattern, something we won't know about
5604 until after the recurse.
5606 SSize_t deltanext, fake = 0;
5610 /* We use SAVEFREEPV so that when the full compile
5611 is finished perl will clean up the allocated
5612 minlens when it's all done. This way we don't
5613 have to worry about freeing them when we know
5614 they wont be used, which would be a pain.
5617 Newx( minnextp, 1, SSize_t );
5618 SAVEFREEPV(minnextp);
5621 StructCopy(data, &data_fake, scan_data_t);
5622 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5625 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5626 data_fake.last_found=newSVsv(data->last_found);
5630 data_fake.last_closep = &fake;
5631 data_fake.flags = 0;
5632 data_fake.pos_delta = delta;
5634 data_fake.flags |= SF_IS_INF;
5635 if ( flags & SCF_DO_STCLASS && !scan->flags
5636 && OP(scan) == IFMATCH ) { /* Lookahead */
5637 ssc_init(pRExC_state, &intrnl);
5638 data_fake.start_class = &intrnl;
5639 f |= SCF_DO_STCLASS_AND;
5641 if (flags & SCF_WHILEM_VISITED_POS)
5642 f |= SCF_WHILEM_VISITED_POS;
5643 next = regnext(scan);
5644 nscan = NEXTOPER(NEXTOPER(scan));
5646 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5647 &deltanext, last, &data_fake,
5648 stopparen, recursed_depth, NULL,
5652 FAIL("Variable length lookbehind not implemented");
5654 else if (*minnextp > (I32)U8_MAX) {
5655 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5658 scan->flags = (U8)*minnextp;
5663 if (f & SCF_DO_STCLASS_AND) {
5664 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5665 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
5668 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5670 if (data_fake.flags & SF_HAS_EVAL)
5671 data->flags |= SF_HAS_EVAL;
5672 data->whilem_c = data_fake.whilem_c;
5673 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5674 if (RExC_rx->minlen<*minnextp)
5675 RExC_rx->minlen=*minnextp;
5676 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5677 SvREFCNT_dec_NN(data_fake.last_found);
5679 if ( data_fake.minlen_fixed != minlenp )
5681 data->offset_fixed= data_fake.offset_fixed;
5682 data->minlen_fixed= data_fake.minlen_fixed;
5683 data->lookbehind_fixed+= scan->flags;
5685 if ( data_fake.minlen_float != minlenp )
5687 data->minlen_float= data_fake.minlen_float;
5688 data->offset_float_min=data_fake.offset_float_min;
5689 data->offset_float_max=data_fake.offset_float_max;
5690 data->lookbehind_float+= scan->flags;
5697 else if (OP(scan) == OPEN) {
5698 if (stopparen != (I32)ARG(scan))
5701 else if (OP(scan) == CLOSE) {
5702 if (stopparen == (I32)ARG(scan)) {
5705 if ((I32)ARG(scan) == is_par) {
5706 next = regnext(scan);
5708 if ( next && (OP(next) != WHILEM) && next < last)
5709 is_par = 0; /* Disable optimization */
5712 *(data->last_closep) = ARG(scan);
5714 else if (OP(scan) == EVAL) {
5716 data->flags |= SF_HAS_EVAL;
5718 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5719 if (flags & SCF_DO_SUBSTR) {
5720 scan_commit(pRExC_state, data, minlenp, is_inf);
5721 flags &= ~SCF_DO_SUBSTR;
5723 if (data && OP(scan)==ACCEPT) {
5724 data->flags |= SCF_SEEN_ACCEPT;
5729 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5731 if (flags & SCF_DO_SUBSTR) {
5732 scan_commit(pRExC_state, data, minlenp, is_inf);
5733 data->longest = &(data->longest_float);
5735 is_inf = is_inf_internal = 1;
5736 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5737 ssc_anything(data->start_class);
5738 flags &= ~SCF_DO_STCLASS;
5740 else if (OP(scan) == GPOS) {
5741 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5742 !(delta || is_inf || (data && data->pos_delta)))
5744 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5745 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5746 if (RExC_rx->gofs < (STRLEN)min)
5747 RExC_rx->gofs = min;
5749 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5753 #ifdef TRIE_STUDY_OPT
5754 #ifdef FULL_TRIE_STUDY
5755 else if (PL_regkind[OP(scan)] == TRIE) {
5756 /* NOTE - There is similar code to this block above for handling
5757 BRANCH nodes on the initial study. If you change stuff here
5759 regnode *trie_node= scan;
5760 regnode *tail= regnext(scan);
5761 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5762 SSize_t max1 = 0, min1 = SSize_t_MAX;
5765 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5766 /* Cannot merge strings after this. */
5767 scan_commit(pRExC_state, data, minlenp, is_inf);
5769 if (flags & SCF_DO_STCLASS)
5770 ssc_init_zero(pRExC_state, &accum);
5776 const regnode *nextbranch= NULL;
5779 for ( word=1 ; word <= trie->wordcount ; word++)
5781 SSize_t deltanext=0, minnext=0, f = 0, fake;
5782 regnode_ssc this_class;
5784 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5786 data_fake.whilem_c = data->whilem_c;
5787 data_fake.last_closep = data->last_closep;
5790 data_fake.last_closep = &fake;
5791 data_fake.pos_delta = delta;
5792 if (flags & SCF_DO_STCLASS) {
5793 ssc_init(pRExC_state, &this_class);
5794 data_fake.start_class = &this_class;
5795 f = SCF_DO_STCLASS_AND;
5797 if (flags & SCF_WHILEM_VISITED_POS)
5798 f |= SCF_WHILEM_VISITED_POS;
5800 if (trie->jump[word]) {
5802 nextbranch = trie_node + trie->jump[0];
5803 scan= trie_node + trie->jump[word];
5804 /* We go from the jump point to the branch that follows
5805 it. Note this means we need the vestigal unused
5806 branches even though they arent otherwise used. */
5807 minnext = study_chunk(pRExC_state, &scan, minlenp,
5808 &deltanext, (regnode *)nextbranch, &data_fake,
5809 stopparen, recursed_depth, NULL, f,depth+1);
5811 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5812 nextbranch= regnext((regnode*)nextbranch);
5814 if (min1 > (SSize_t)(minnext + trie->minlen))
5815 min1 = minnext + trie->minlen;
5816 if (deltanext == SSize_t_MAX) {
5817 is_inf = is_inf_internal = 1;
5819 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5820 max1 = minnext + deltanext + trie->maxlen;
5822 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5824 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5825 if ( stopmin > min + min1)
5826 stopmin = min + min1;
5827 flags &= ~SCF_DO_SUBSTR;
5829 data->flags |= SCF_SEEN_ACCEPT;
5832 if (data_fake.flags & SF_HAS_EVAL)
5833 data->flags |= SF_HAS_EVAL;
5834 data->whilem_c = data_fake.whilem_c;
5836 if (flags & SCF_DO_STCLASS)
5837 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5840 if (flags & SCF_DO_SUBSTR) {
5841 data->pos_min += min1;
5842 data->pos_delta += max1 - min1;
5843 if (max1 != min1 || is_inf)
5844 data->longest = &(data->longest_float);
5847 if (delta != SSize_t_MAX)
5848 delta += max1 - min1;
5849 if (flags & SCF_DO_STCLASS_OR) {
5850 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5852 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5853 flags &= ~SCF_DO_STCLASS;
5856 else if (flags & SCF_DO_STCLASS_AND) {
5858 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5859 flags &= ~SCF_DO_STCLASS;
5862 /* Switch to OR mode: cache the old value of
5863 * data->start_class */
5865 StructCopy(data->start_class, and_withp, regnode_ssc);
5866 flags &= ~SCF_DO_STCLASS_AND;
5867 StructCopy(&accum, data->start_class, regnode_ssc);
5868 flags |= SCF_DO_STCLASS_OR;
5875 else if (PL_regkind[OP(scan)] == TRIE) {
5876 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5879 min += trie->minlen;
5880 delta += (trie->maxlen - trie->minlen);
5881 flags &= ~SCF_DO_STCLASS; /* xxx */
5882 if (flags & SCF_DO_SUBSTR) {
5883 /* Cannot expect anything... */
5884 scan_commit(pRExC_state, data, minlenp, is_inf);
5885 data->pos_min += trie->minlen;
5886 data->pos_delta += (trie->maxlen - trie->minlen);
5887 if (trie->maxlen != trie->minlen)
5888 data->longest = &(data->longest_float);
5890 if (trie->jump) /* no more substrings -- for now /grr*/
5891 flags &= ~SCF_DO_SUBSTR;
5893 #endif /* old or new */
5894 #endif /* TRIE_STUDY_OPT */
5896 /* Else: zero-length, ignore. */
5897 scan = regnext(scan);
5902 /* we need to unwind recursion. */
5905 DEBUG_STUDYDATA("frame-end:",data,depth);
5906 DEBUG_PEEP("fend", scan, depth);
5908 /* restore previous context */
5909 last = frame->last_regnode;
5910 scan = frame->next_regnode;
5911 stopparen = frame->stopparen;
5912 recursed_depth = frame->prev_recursed_depth;
5914 RExC_frame_last = frame->prev_frame;
5915 frame = frame->this_prev_frame;
5916 goto fake_study_recurse;
5920 DEBUG_STUDYDATA("pre-fin:",data,depth);
5923 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5925 if (flags & SCF_DO_SUBSTR && is_inf)
5926 data->pos_delta = SSize_t_MAX - data->pos_min;
5927 if (is_par > (I32)U8_MAX)
5929 if (is_par && pars==1 && data) {
5930 data->flags |= SF_IN_PAR;
5931 data->flags &= ~SF_HAS_PAR;
5933 else if (pars && data) {
5934 data->flags |= SF_HAS_PAR;
5935 data->flags &= ~SF_IN_PAR;
5937 if (flags & SCF_DO_STCLASS_OR)
5938 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5939 if (flags & SCF_TRIE_RESTUDY)
5940 data->flags |= SCF_TRIE_RESTUDY;
5942 DEBUG_STUDYDATA("post-fin:",data,depth);
5945 SSize_t final_minlen= min < stopmin ? min : stopmin;
5947 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
5948 if (final_minlen > SSize_t_MAX - delta)
5949 RExC_maxlen = SSize_t_MAX;
5950 else if (RExC_maxlen < final_minlen + delta)
5951 RExC_maxlen = final_minlen + delta;
5953 return final_minlen;
5955 NOT_REACHED; /* NOTREACHED */
5959 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5961 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5963 PERL_ARGS_ASSERT_ADD_DATA;
5965 Renewc(RExC_rxi->data,
5966 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
5967 char, struct reg_data);
5969 Renew(RExC_rxi->data->what, count + n, U8);
5971 Newx(RExC_rxi->data->what, n, U8);
5972 RExC_rxi->data->count = count + n;
5973 Copy(s, RExC_rxi->data->what + count, n, U8);
5977 /*XXX: todo make this not included in a non debugging perl, but appears to be
5978 * used anyway there, in 'use re' */
5979 #ifndef PERL_IN_XSUB_RE
5981 Perl_reginitcolors(pTHX)
5983 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
5985 char *t = savepv(s);
5989 t = strchr(t, '\t');
5995 PL_colors[i] = t = (char *)"";
6000 PL_colors[i++] = (char *)"";
6007 #ifdef TRIE_STUDY_OPT
6008 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6011 (data.flags & SCF_TRIE_RESTUDY) \
6019 #define CHECK_RESTUDY_GOTO_butfirst
6023 * pregcomp - compile a regular expression into internal code
6025 * Decides which engine's compiler to call based on the hint currently in
6029 #ifndef PERL_IN_XSUB_RE
6031 /* return the currently in-scope regex engine (or the default if none) */
6033 regexp_engine const *
6034 Perl_current_re_engine(pTHX)
6036 if (IN_PERL_COMPILETIME) {
6037 HV * const table = GvHV(PL_hintgv);
6040 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6041 return &PL_core_reg_engine;
6042 ptr = hv_fetchs(table, "regcomp", FALSE);
6043 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6044 return &PL_core_reg_engine;
6045 return INT2PTR(regexp_engine*,SvIV(*ptr));
6049 if (!PL_curcop->cop_hints_hash)
6050 return &PL_core_reg_engine;
6051 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6052 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6053 return &PL_core_reg_engine;
6054 return INT2PTR(regexp_engine*,SvIV(ptr));
6060 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6062 regexp_engine const *eng = current_re_engine();
6063 GET_RE_DEBUG_FLAGS_DECL;
6065 PERL_ARGS_ASSERT_PREGCOMP;
6067 /* Dispatch a request to compile a regexp to correct regexp engine. */
6069 Perl_re_printf( aTHX_ "Using engine %"UVxf"\n",
6072 return CALLREGCOMP_ENG(eng, pattern, flags);
6076 /* public(ish) entry point for the perl core's own regex compiling code.
6077 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6078 * pattern rather than a list of OPs, and uses the internal engine rather
6079 * than the current one */
6082 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6084 SV *pat = pattern; /* defeat constness! */
6085 PERL_ARGS_ASSERT_RE_COMPILE;
6086 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6087 #ifdef PERL_IN_XSUB_RE
6090 &PL_core_reg_engine,
6092 NULL, NULL, rx_flags, 0);
6096 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6097 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6098 * point to the realloced string and length.
6100 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6104 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6105 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6107 U8 *const src = (U8*)*pat_p;
6112 GET_RE_DEBUG_FLAGS_DECL;
6114 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6115 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6117 Newx(dst, *plen_p * 2 + 1, U8);
6120 while (s < *plen_p) {
6121 append_utf8_from_native_byte(src[s], &d);
6122 if (n < num_code_blocks) {
6123 if (!do_end && pRExC_state->code_blocks[n].start == s) {
6124 pRExC_state->code_blocks[n].start = d - dst - 1;
6125 assert(*(d - 1) == '(');
6128 else if (do_end && pRExC_state->code_blocks[n].end == s) {
6129 pRExC_state->code_blocks[n].end = d - dst - 1;
6130 assert(*(d - 1) == ')');
6139 *pat_p = (char*) dst;
6141 RExC_orig_utf8 = RExC_utf8 = 1;
6146 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6147 * while recording any code block indices, and handling overloading,
6148 * nested qr// objects etc. If pat is null, it will allocate a new
6149 * string, or just return the first arg, if there's only one.
6151 * Returns the malloced/updated pat.
6152 * patternp and pat_count is the array of SVs to be concatted;
6153 * oplist is the optional list of ops that generated the SVs;
6154 * recompile_p is a pointer to a boolean that will be set if
6155 * the regex will need to be recompiled.
6156 * delim, if non-null is an SV that will be inserted between each element
6160 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6161 SV *pat, SV ** const patternp, int pat_count,
6162 OP *oplist, bool *recompile_p, SV *delim)
6166 bool use_delim = FALSE;
6167 bool alloced = FALSE;
6169 /* if we know we have at least two args, create an empty string,
6170 * then concatenate args to that. For no args, return an empty string */
6171 if (!pat && pat_count != 1) {
6177 for (svp = patternp; svp < patternp + pat_count; svp++) {
6180 STRLEN orig_patlen = 0;
6182 SV *msv = use_delim ? delim : *svp;
6183 if (!msv) msv = &PL_sv_undef;
6185 /* if we've got a delimiter, we go round the loop twice for each
6186 * svp slot (except the last), using the delimiter the second
6195 if (SvTYPE(msv) == SVt_PVAV) {
6196 /* we've encountered an interpolated array within
6197 * the pattern, e.g. /...@a..../. Expand the list of elements,
6198 * then recursively append elements.
6199 * The code in this block is based on S_pushav() */
6201 AV *const av = (AV*)msv;
6202 const SSize_t maxarg = AvFILL(av) + 1;
6206 assert(oplist->op_type == OP_PADAV
6207 || oplist->op_type == OP_RV2AV);
6208 oplist = OpSIBLING(oplist);
6211 if (SvRMAGICAL(av)) {
6214 Newx(array, maxarg, SV*);
6216 for (i=0; i < maxarg; i++) {
6217 SV ** const svp = av_fetch(av, i, FALSE);
6218 array[i] = svp ? *svp : &PL_sv_undef;
6222 array = AvARRAY(av);
6224 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6225 array, maxarg, NULL, recompile_p,
6227 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6233 /* we make the assumption here that each op in the list of
6234 * op_siblings maps to one SV pushed onto the stack,
6235 * except for code blocks, with have both an OP_NULL and
6237 * This allows us to match up the list of SVs against the
6238 * list of OPs to find the next code block.
6240 * Note that PUSHMARK PADSV PADSV ..
6242 * PADRANGE PADSV PADSV ..
6243 * so the alignment still works. */
6246 if (oplist->op_type == OP_NULL
6247 && (oplist->op_flags & OPf_SPECIAL))
6249 assert(n < pRExC_state->num_code_blocks);
6250 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
6251 pRExC_state->code_blocks[n].block = oplist;
6252 pRExC_state->code_blocks[n].src_regex = NULL;
6255 oplist = OpSIBLING(oplist); /* skip CONST */
6258 oplist = OpSIBLING(oplist);;
6261 /* apply magic and QR overloading to arg */
6264 if (SvROK(msv) && SvAMAGIC(msv)) {
6265 SV *sv = AMG_CALLunary(msv, regexp_amg);
6269 if (SvTYPE(sv) != SVt_REGEXP)
6270 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6275 /* try concatenation overload ... */
6276 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6277 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6280 /* overloading involved: all bets are off over literal
6281 * code. Pretend we haven't seen it */
6282 pRExC_state->num_code_blocks -= n;
6286 /* ... or failing that, try "" overload */
6287 while (SvAMAGIC(msv)
6288 && (sv = AMG_CALLunary(msv, string_amg))
6292 && SvRV(msv) == SvRV(sv))
6297 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6301 /* this is a partially unrolled
6302 * sv_catsv_nomg(pat, msv);
6303 * that allows us to adjust code block indices if
6306 char *dst = SvPV_force_nomg(pat, dlen);
6308 if (SvUTF8(msv) && !SvUTF8(pat)) {
6309 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6310 sv_setpvn(pat, dst, dlen);
6313 sv_catsv_nomg(pat, msv);
6317 /* We have only one SV to process, but we need to verify
6318 * it is properly null terminated or we will fail asserts
6319 * later. In theory we probably shouldn't get such SV's,
6320 * but if we do we should handle it gracefully. */
6321 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) ) {
6322 /* not a string, or a string with a trailing null */
6325 /* a string with no trailing null, we need to copy it
6326 * so it we have a trailing null */
6332 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
6335 /* extract any code blocks within any embedded qr//'s */
6336 if (rx && SvTYPE(rx) == SVt_REGEXP
6337 && RX_ENGINE((REGEXP*)rx)->op_comp)
6340 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6341 if (ri->num_code_blocks) {
6343 /* the presence of an embedded qr// with code means
6344 * we should always recompile: the text of the
6345 * qr// may not have changed, but it may be a
6346 * different closure than last time */
6348 Renew(pRExC_state->code_blocks,
6349 pRExC_state->num_code_blocks + ri->num_code_blocks,
6350 struct reg_code_block);
6351 pRExC_state->num_code_blocks += ri->num_code_blocks;
6353 for (i=0; i < ri->num_code_blocks; i++) {
6354 struct reg_code_block *src, *dst;
6355 STRLEN offset = orig_patlen
6356 + ReANY((REGEXP *)rx)->pre_prefix;
6357 assert(n < pRExC_state->num_code_blocks);
6358 src = &ri->code_blocks[i];
6359 dst = &pRExC_state->code_blocks[n];
6360 dst->start = src->start + offset;
6361 dst->end = src->end + offset;
6362 dst->block = src->block;
6363 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6372 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6381 /* see if there are any run-time code blocks in the pattern.
6382 * False positives are allowed */
6385 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6386 char *pat, STRLEN plen)
6391 PERL_UNUSED_CONTEXT;
6393 for (s = 0; s < plen; s++) {
6394 if (n < pRExC_state->num_code_blocks
6395 && s == pRExC_state->code_blocks[n].start)
6397 s = pRExC_state->code_blocks[n].end;
6401 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6403 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6405 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6412 /* Handle run-time code blocks. We will already have compiled any direct
6413 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6414 * copy of it, but with any literal code blocks blanked out and
6415 * appropriate chars escaped; then feed it into
6417 * eval "qr'modified_pattern'"
6421 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6425 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6427 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6428 * and merge them with any code blocks of the original regexp.
6430 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6431 * instead, just save the qr and return FALSE; this tells our caller that
6432 * the original pattern needs upgrading to utf8.
6436 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6437 char *pat, STRLEN plen)
6441 GET_RE_DEBUG_FLAGS_DECL;
6443 if (pRExC_state->runtime_code_qr) {
6444 /* this is the second time we've been called; this should
6445 * only happen if the main pattern got upgraded to utf8
6446 * during compilation; re-use the qr we compiled first time
6447 * round (which should be utf8 too)
6449 qr = pRExC_state->runtime_code_qr;
6450 pRExC_state->runtime_code_qr = NULL;
6451 assert(RExC_utf8 && SvUTF8(qr));
6457 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
6461 /* determine how many extra chars we need for ' and \ escaping */
6462 for (s = 0; s < plen; s++) {
6463 if (pat[s] == '\'' || pat[s] == '\\')
6467 Newx(newpat, newlen, char);
6469 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6471 for (s = 0; s < plen; s++) {
6472 if (n < pRExC_state->num_code_blocks
6473 && s == pRExC_state->code_blocks[n].start)
6475 /* blank out literal code block */
6476 assert(pat[s] == '(');
6477 while (s <= pRExC_state->code_blocks[n].end) {
6485 if (pat[s] == '\'' || pat[s] == '\\')
6490 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
6494 Perl_re_printf( aTHX_
6495 "%sre-parsing pattern for runtime code:%s %s\n",
6496 PL_colors[4],PL_colors[5],newpat);
6499 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6505 PUSHSTACKi(PERLSI_REQUIRE);
6506 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6507 * parsing qr''; normally only q'' does this. It also alters
6509 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6510 SvREFCNT_dec_NN(sv);
6515 SV * const errsv = ERRSV;
6516 if (SvTRUE_NN(errsv))
6518 Safefree(pRExC_state->code_blocks);
6519 /* use croak_sv ? */
6520 Perl_croak_nocontext("%"SVf, SVfARG(errsv));
6523 assert(SvROK(qr_ref));
6525 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6526 /* the leaving below frees the tmp qr_ref.
6527 * Give qr a life of its own */
6535 if (!RExC_utf8 && SvUTF8(qr)) {
6536 /* first time through; the pattern got upgraded; save the
6537 * qr for the next time through */
6538 assert(!pRExC_state->runtime_code_qr);
6539 pRExC_state->runtime_code_qr = qr;
6544 /* extract any code blocks within the returned qr// */
6547 /* merge the main (r1) and run-time (r2) code blocks into one */
6549 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6550 struct reg_code_block *new_block, *dst;
6551 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6554 if (!r2->num_code_blocks) /* we guessed wrong */
6556 SvREFCNT_dec_NN(qr);
6561 r1->num_code_blocks + r2->num_code_blocks,
6562 struct reg_code_block);
6565 while ( i1 < r1->num_code_blocks
6566 || i2 < r2->num_code_blocks)
6568 struct reg_code_block *src;
6571 if (i1 == r1->num_code_blocks) {
6572 src = &r2->code_blocks[i2++];
6575 else if (i2 == r2->num_code_blocks)
6576 src = &r1->code_blocks[i1++];
6577 else if ( r1->code_blocks[i1].start
6578 < r2->code_blocks[i2].start)
6580 src = &r1->code_blocks[i1++];
6581 assert(src->end < r2->code_blocks[i2].start);
6584 assert( r1->code_blocks[i1].start
6585 > r2->code_blocks[i2].start);
6586 src = &r2->code_blocks[i2++];
6588 assert(src->end < r1->code_blocks[i1].start);
6591 assert(pat[src->start] == '(');
6592 assert(pat[src->end] == ')');
6593 dst->start = src->start;
6594 dst->end = src->end;
6595 dst->block = src->block;
6596 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6600 r1->num_code_blocks += r2->num_code_blocks;
6601 Safefree(r1->code_blocks);
6602 r1->code_blocks = new_block;
6605 SvREFCNT_dec_NN(qr);
6611 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6612 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6613 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6614 STRLEN longest_length, bool eol, bool meol)
6616 /* This is the common code for setting up the floating and fixed length
6617 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6618 * as to whether succeeded or not */
6623 if (! (longest_length
6624 || (eol /* Can't have SEOL and MULTI */
6625 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6627 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6628 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6633 /* copy the information about the longest from the reg_scan_data
6634 over to the program. */
6635 if (SvUTF8(sv_longest)) {
6636 *rx_utf8 = sv_longest;
6639 *rx_substr = sv_longest;
6642 /* end_shift is how many chars that must be matched that
6643 follow this item. We calculate it ahead of time as once the
6644 lookbehind offset is added in we lose the ability to correctly
6646 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6647 *rx_end_shift = ml - offset
6648 - longest_length + (SvTAIL(sv_longest) != 0)
6651 t = (eol/* Can't have SEOL and MULTI */
6652 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6653 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6659 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6660 * regular expression into internal code.
6661 * The pattern may be passed either as:
6662 * a list of SVs (patternp plus pat_count)
6663 * a list of OPs (expr)
6664 * If both are passed, the SV list is used, but the OP list indicates
6665 * which SVs are actually pre-compiled code blocks
6667 * The SVs in the list have magic and qr overloading applied to them (and
6668 * the list may be modified in-place with replacement SVs in the latter
6671 * If the pattern hasn't changed from old_re, then old_re will be
6674 * eng is the current engine. If that engine has an op_comp method, then
6675 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6676 * do the initial concatenation of arguments and pass on to the external
6679 * If is_bare_re is not null, set it to a boolean indicating whether the
6680 * arg list reduced (after overloading) to a single bare regex which has
6681 * been returned (i.e. /$qr/).
6683 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6685 * pm_flags contains the PMf_* flags, typically based on those from the
6686 * pm_flags field of the related PMOP. Currently we're only interested in
6687 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6689 * We can't allocate space until we know how big the compiled form will be,
6690 * but we can't compile it (and thus know how big it is) until we've got a
6691 * place to put the code. So we cheat: we compile it twice, once with code
6692 * generation turned off and size counting turned on, and once "for real".
6693 * This also means that we don't allocate space until we are sure that the
6694 * thing really will compile successfully, and we never have to move the
6695 * code and thus invalidate pointers into it. (Note that it has to be in
6696 * one piece because free() must be able to free it all.) [NB: not true in perl]
6698 * Beware that the optimization-preparation code in here knows about some
6699 * of the structure of the compiled regexp. [I'll say.]
6703 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6704 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6705 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6709 regexp_internal *ri;
6717 SV *code_blocksv = NULL;
6718 SV** new_patternp = patternp;
6720 /* these are all flags - maybe they should be turned
6721 * into a single int with different bit masks */
6722 I32 sawlookahead = 0;
6727 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6729 bool runtime_code = 0;
6731 RExC_state_t RExC_state;
6732 RExC_state_t * const pRExC_state = &RExC_state;
6733 #ifdef TRIE_STUDY_OPT
6735 RExC_state_t copyRExC_state;
6737 GET_RE_DEBUG_FLAGS_DECL;
6739 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6741 DEBUG_r(if (!PL_colorset) reginitcolors());
6743 /* Initialize these here instead of as-needed, as is quick and avoids
6744 * having to test them each time otherwise */
6745 if (! PL_AboveLatin1) {
6747 char * dump_len_string;
6750 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6751 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6752 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6753 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6754 PL_HasMultiCharFold =
6755 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6757 /* This is calculated here, because the Perl program that generates the
6758 * static global ones doesn't currently have access to
6759 * NUM_ANYOF_CODE_POINTS */
6760 PL_InBitmap = _new_invlist(2);
6761 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
6762 NUM_ANYOF_CODE_POINTS - 1);
6764 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
6765 if ( ! dump_len_string
6766 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
6768 PL_dump_re_max_len = 0;
6773 pRExC_state->warn_text = NULL;
6774 pRExC_state->code_blocks = NULL;
6775 pRExC_state->num_code_blocks = 0;
6778 *is_bare_re = FALSE;
6780 if (expr && (expr->op_type == OP_LIST ||
6781 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6782 /* allocate code_blocks if needed */
6786 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
6787 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6788 ncode++; /* count of DO blocks */
6790 pRExC_state->num_code_blocks = ncode;
6791 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
6796 /* compile-time pattern with just OP_CONSTs and DO blocks */
6801 /* find how many CONSTs there are */
6804 if (expr->op_type == OP_CONST)
6807 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6808 if (o->op_type == OP_CONST)
6812 /* fake up an SV array */
6814 assert(!new_patternp);
6815 Newx(new_patternp, n, SV*);
6816 SAVEFREEPV(new_patternp);
6820 if (expr->op_type == OP_CONST)
6821 new_patternp[n] = cSVOPx_sv(expr);
6823 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6824 if (o->op_type == OP_CONST)
6825 new_patternp[n++] = cSVOPo_sv;
6830 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6831 "Assembling pattern from %d elements%s\n", pat_count,
6832 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6834 /* set expr to the first arg op */
6836 if (pRExC_state->num_code_blocks
6837 && expr->op_type != OP_CONST)
6839 expr = cLISTOPx(expr)->op_first;
6840 assert( expr->op_type == OP_PUSHMARK
6841 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6842 || expr->op_type == OP_PADRANGE);
6843 expr = OpSIBLING(expr);
6846 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6847 expr, &recompile, NULL);
6849 /* handle bare (possibly after overloading) regex: foo =~ $re */
6854 if (SvTYPE(re) == SVt_REGEXP) {
6858 Safefree(pRExC_state->code_blocks);
6859 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6860 "Precompiled pattern%s\n",
6861 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6867 exp = SvPV_nomg(pat, plen);
6869 if (!eng->op_comp) {
6870 if ((SvUTF8(pat) && IN_BYTES)
6871 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6873 /* make a temporary copy; either to convert to bytes,
6874 * or to avoid repeating get-magic / overloaded stringify */
6875 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6876 (IN_BYTES ? 0 : SvUTF8(pat)));
6878 Safefree(pRExC_state->code_blocks);
6879 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6882 /* ignore the utf8ness if the pattern is 0 length */
6883 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6885 RExC_uni_semantics = 0;
6886 RExC_seen_unfolded_sharp_s = 0;
6887 RExC_contains_locale = 0;
6888 RExC_contains_i = 0;
6889 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
6890 RExC_study_started = 0;
6891 pRExC_state->runtime_code_qr = NULL;
6892 RExC_frame_head= NULL;
6893 RExC_frame_last= NULL;
6894 RExC_frame_count= 0;
6897 RExC_mysv1= sv_newmortal();
6898 RExC_mysv2= sv_newmortal();
6901 SV *dsv= sv_newmortal();
6902 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6903 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
6904 PL_colors[4],PL_colors[5],s);
6908 /* we jump here if we have to recompile, e.g., from upgrading the pattern
6911 if ((pm_flags & PMf_USE_RE_EVAL)
6912 /* this second condition covers the non-regex literal case,
6913 * i.e. $foo =~ '(?{})'. */
6914 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6916 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6918 /* return old regex if pattern hasn't changed */
6919 /* XXX: note in the below we have to check the flags as well as the
6922 * Things get a touch tricky as we have to compare the utf8 flag
6923 * independently from the compile flags. */
6927 && !!RX_UTF8(old_re) == !!RExC_utf8
6928 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
6929 && RX_PRECOMP(old_re)
6930 && RX_PRELEN(old_re) == plen
6931 && memEQ(RX_PRECOMP(old_re), exp, plen)
6932 && !runtime_code /* with runtime code, always recompile */ )
6934 Safefree(pRExC_state->code_blocks);
6938 rx_flags = orig_rx_flags;
6940 if (rx_flags & PMf_FOLD) {
6941 RExC_contains_i = 1;
6943 if ( initial_charset == REGEX_DEPENDS_CHARSET
6944 && (RExC_utf8 ||RExC_uni_semantics))
6947 /* Set to use unicode semantics if the pattern is in utf8 and has the
6948 * 'depends' charset specified, as it means unicode when utf8 */
6949 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6953 RExC_precomp_adj = 0;
6954 RExC_flags = rx_flags;
6955 RExC_pm_flags = pm_flags;
6958 assert(TAINTING_get || !TAINT_get);
6960 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
6962 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
6963 /* whoops, we have a non-utf8 pattern, whilst run-time code
6964 * got compiled as utf8. Try again with a utf8 pattern */
6965 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6966 pRExC_state->num_code_blocks);
6967 goto redo_first_pass;
6970 assert(!pRExC_state->runtime_code_qr);
6976 RExC_in_lookbehind = 0;
6977 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
6979 RExC_override_recoding = 0;
6981 RExC_recode_x_to_native = 0;
6983 RExC_in_multi_char_class = 0;
6985 /* First pass: determine size, legality. */
6987 RExC_start = RExC_adjusted_start = exp;
6988 RExC_end = exp + plen;
6989 RExC_precomp_end = RExC_end;
6994 RExC_emit = (regnode *) &RExC_emit_dummy;
6995 RExC_whilem_seen = 0;
6996 RExC_open_parens = NULL;
6997 RExC_close_parens = NULL;
6999 RExC_paren_names = NULL;
7001 RExC_paren_name_list = NULL;
7003 RExC_recurse = NULL;
7004 RExC_study_chunk_recursed = NULL;
7005 RExC_study_chunk_recursed_bytes= 0;
7006 RExC_recurse_count = 0;
7007 pRExC_state->code_index = 0;
7009 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7010 * code makes sure the final byte is an uncounted NUL. But should this
7011 * ever not be the case, lots of things could read beyond the end of the
7012 * buffer: loops like
7013 * while(isFOO(*RExC_parse)) RExC_parse++;
7014 * strchr(RExC_parse, "foo");
7015 * etc. So it is worth noting. */
7016 assert(*RExC_end == '\0');
7019 Perl_re_printf( aTHX_ "Starting first pass (sizing)\n");
7021 RExC_lastparse=NULL;
7023 /* reg may croak on us, not giving us a chance to free
7024 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
7025 need it to survive as long as the regexp (qr/(?{})/).
7026 We must check that code_blocksv is not already set, because we may
7027 have jumped back to restart the sizing pass. */
7028 if (pRExC_state->code_blocks && !code_blocksv) {
7029 code_blocksv = newSV_type(SVt_PV);
7030 SAVEFREESV(code_blocksv);
7031 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
7032 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
7034 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7035 /* It's possible to write a regexp in ascii that represents Unicode
7036 codepoints outside of the byte range, such as via \x{100}. If we
7037 detect such a sequence we have to convert the entire pattern to utf8
7038 and then recompile, as our sizing calculation will have been based
7039 on 1 byte == 1 character, but we will need to use utf8 to encode
7040 at least some part of the pattern, and therefore must convert the whole
7043 if (flags & RESTART_PASS1) {
7044 if (flags & NEED_UTF8) {
7045 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7046 pRExC_state->num_code_blocks);
7049 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7050 "Need to redo pass 1\n"));
7053 goto redo_first_pass;
7055 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
7058 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
7061 Perl_re_printf( aTHX_
7062 "Required size %"IVdf" nodes\n"
7063 "Starting second pass (creation)\n",
7066 RExC_lastparse=NULL;
7069 /* The first pass could have found things that force Unicode semantics */
7070 if ((RExC_utf8 || RExC_uni_semantics)
7071 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
7073 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7076 /* Small enough for pointer-storage convention?
7077 If extralen==0, this means that we will not need long jumps. */
7078 if (RExC_size >= 0x10000L && RExC_extralen)
7079 RExC_size += RExC_extralen;
7082 if (RExC_whilem_seen > 15)
7083 RExC_whilem_seen = 15;
7085 /* Allocate space and zero-initialize. Note, the two step process
7086 of zeroing when in debug mode, thus anything assigned has to
7087 happen after that */
7088 rx = (REGEXP*) newSV_type(SVt_REGEXP);
7090 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7091 char, regexp_internal);
7092 if ( r == NULL || ri == NULL )
7093 FAIL("Regexp out of space");
7095 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
7096 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7099 /* bulk initialize base fields with 0. */
7100 Zero(ri, sizeof(regexp_internal), char);
7103 /* non-zero initialization begins here */
7106 r->extflags = rx_flags;
7107 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7109 if (pm_flags & PMf_IS_QR) {
7110 ri->code_blocks = pRExC_state->code_blocks;
7111 ri->num_code_blocks = pRExC_state->num_code_blocks;
7116 for (n = 0; n < pRExC_state->num_code_blocks; n++)
7117 if (pRExC_state->code_blocks[n].src_regex)
7118 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
7119 if(pRExC_state->code_blocks)
7120 SAVEFREEPV(pRExC_state->code_blocks); /* often null */
7124 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7125 bool has_charset = (get_regex_charset(r->extflags)
7126 != REGEX_DEPENDS_CHARSET);
7128 /* The caret is output if there are any defaults: if not all the STD
7129 * flags are set, or if no character set specifier is needed */
7131 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7133 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7134 == REG_RUN_ON_COMMENT_SEEN);
7135 U8 reganch = (U8)((r->extflags & RXf_PMf_STD_PMMOD)
7136 >> RXf_PMf_STD_PMMOD_SHIFT);
7137 const char *fptr = STD_PAT_MODS; /*"msixn"*/
7140 /* We output all the necessary flags; we never output a minus, as all
7141 * those are defaults, so are
7142 * covered by the caret */
7143 const STRLEN wraplen = plen + has_p + has_runon
7144 + has_default /* If needs a caret */
7145 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7147 /* If needs a character set specifier */
7148 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7149 + (sizeof("(?:)") - 1);
7151 /* make sure PL_bitcount bounds not exceeded */
7152 assert(sizeof(STD_PAT_MODS) <= 8);
7154 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
7155 r->xpv_len_u.xpvlenu_pv = p;
7157 SvFLAGS(rx) |= SVf_UTF8;
7160 /* If a default, cover it using the caret */
7162 *p++= DEFAULT_PAT_MOD;
7166 const char* const name = get_regex_charset_name(r->extflags, &len);
7167 Copy(name, p, len, char);
7171 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7174 while((ch = *fptr++)) {
7182 Copy(RExC_precomp, p, plen, char);
7183 assert ((RX_WRAPPED(rx) - p) < 16);
7184 r->pre_prefix = p - RX_WRAPPED(rx);
7190 SvCUR_set(rx, p - RX_WRAPPED(rx));
7194 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
7196 /* Useful during FAIL. */
7197 #ifdef RE_TRACK_PATTERN_OFFSETS
7198 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
7199 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7200 "%s %"UVuf" bytes for offset annotations.\n",
7201 ri->u.offsets ? "Got" : "Couldn't get",
7202 (UV)((2*RExC_size+1) * sizeof(U32))));
7204 SetProgLen(ri,RExC_size);
7209 /* Second pass: emit code. */
7210 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7211 RExC_pm_flags = pm_flags;
7213 RExC_end = exp + plen;
7215 RExC_emit_start = ri->program;
7216 RExC_emit = ri->program;
7217 RExC_emit_bound = ri->program + RExC_size + 1;
7218 pRExC_state->code_index = 0;
7220 *((char*) RExC_emit++) = (char) REG_MAGIC;
7221 /* setup various meta data about recursion, this all requires
7222 * RExC_npar to be correctly set, and a bit later on we clear it */
7223 if (RExC_seen & REG_RECURSE_SEEN) {
7224 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
7225 "%*s%*s Setting up open/close parens\n",
7226 22, "| |", (int)(0 * 2 + 1), ""));
7228 /* setup RExC_open_parens, which holds the address of each
7229 * OPEN tag, and to make things simpler for the 0 index
7230 * the start of the program - this is used later for offsets */
7231 Newxz(RExC_open_parens, RExC_npar,regnode *);
7232 SAVEFREEPV(RExC_open_parens);
7233 RExC_open_parens[0] = RExC_emit;
7235 /* setup RExC_close_parens, which holds the address of each
7236 * CLOSE tag, and to make things simpler for the 0 index
7237 * the end of the program - this is used later for offsets */
7238 Newxz(RExC_close_parens, RExC_npar,regnode *);
7239 SAVEFREEPV(RExC_close_parens);
7240 /* we dont know where end op starts yet, so we dont
7241 * need to set RExC_close_parens[0] like we do RExC_open_parens[0] above */
7243 /* Note, RExC_npar is 1 + the number of parens in a pattern.
7244 * So its 1 if there are no parens. */
7245 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
7246 ((RExC_npar & 0x07) != 0);
7247 Newx(RExC_study_chunk_recursed,
7248 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7249 SAVEFREEPV(RExC_study_chunk_recursed);
7252 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7254 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#"UVxf"", (UV) flags);
7257 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7260 /* XXXX To minimize changes to RE engine we always allocate
7261 3-units-long substrs field. */
7262 Newx(r->substrs, 1, struct reg_substr_data);
7263 if (RExC_recurse_count) {
7264 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
7265 SAVEFREEPV(RExC_recurse);
7269 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7271 RExC_study_chunk_recursed_count= 0;
7273 Zero(r->substrs, 1, struct reg_substr_data);
7274 if (RExC_study_chunk_recursed) {
7275 Zero(RExC_study_chunk_recursed,
7276 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7280 #ifdef TRIE_STUDY_OPT
7282 StructCopy(&zero_scan_data, &data, scan_data_t);
7283 copyRExC_state = RExC_state;
7286 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
7288 RExC_state = copyRExC_state;
7289 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
7290 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
7292 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
7293 StructCopy(&zero_scan_data, &data, scan_data_t);
7296 StructCopy(&zero_scan_data, &data, scan_data_t);
7299 /* Dig out information for optimizations. */
7300 r->extflags = RExC_flags; /* was pm_op */
7301 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
7304 SvUTF8_on(rx); /* Unicode in it? */
7305 ri->regstclass = NULL;
7306 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
7307 r->intflags |= PREGf_NAUGHTY;
7308 scan = ri->program + 1; /* First BRANCH. */
7310 /* testing for BRANCH here tells us whether there is "must appear"
7311 data in the pattern. If there is then we can use it for optimisations */
7312 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7315 STRLEN longest_float_length, longest_fixed_length;
7316 regnode_ssc ch_class; /* pointed to by data */
7318 SSize_t last_close = 0; /* pointed to by data */
7319 regnode *first= scan;
7320 regnode *first_next= regnext(first);
7322 * Skip introductions and multiplicators >= 1
7323 * so that we can extract the 'meat' of the pattern that must
7324 * match in the large if() sequence following.
7325 * NOTE that EXACT is NOT covered here, as it is normally
7326 * picked up by the optimiser separately.
7328 * This is unfortunate as the optimiser isnt handling lookahead
7329 * properly currently.
7332 while ((OP(first) == OPEN && (sawopen = 1)) ||
7333 /* An OR of *one* alternative - should not happen now. */
7334 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7335 /* for now we can't handle lookbehind IFMATCH*/
7336 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7337 (OP(first) == PLUS) ||
7338 (OP(first) == MINMOD) ||
7339 /* An {n,m} with n>0 */
7340 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7341 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7344 * the only op that could be a regnode is PLUS, all the rest
7345 * will be regnode_1 or regnode_2.
7347 * (yves doesn't think this is true)
7349 if (OP(first) == PLUS)
7352 if (OP(first) == MINMOD)
7354 first += regarglen[OP(first)];
7356 first = NEXTOPER(first);
7357 first_next= regnext(first);
7360 /* Starting-point info. */
7362 DEBUG_PEEP("first:",first,0);
7363 /* Ignore EXACT as we deal with it later. */
7364 if (PL_regkind[OP(first)] == EXACT) {
7365 if (OP(first) == EXACT || OP(first) == EXACTL)
7366 NOOP; /* Empty, get anchored substr later. */
7368 ri->regstclass = first;
7371 else if (PL_regkind[OP(first)] == TRIE &&
7372 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
7374 /* this can happen only on restudy */
7375 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
7378 else if (REGNODE_SIMPLE(OP(first)))
7379 ri->regstclass = first;
7380 else if (PL_regkind[OP(first)] == BOUND ||
7381 PL_regkind[OP(first)] == NBOUND)
7382 ri->regstclass = first;
7383 else if (PL_regkind[OP(first)] == BOL) {
7384 r->intflags |= (OP(first) == MBOL
7387 first = NEXTOPER(first);
7390 else if (OP(first) == GPOS) {
7391 r->intflags |= PREGf_ANCH_GPOS;
7392 first = NEXTOPER(first);
7395 else if ((!sawopen || !RExC_sawback) &&
7397 (OP(first) == STAR &&
7398 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7399 !(r->intflags & PREGf_ANCH) && !pRExC_state->num_code_blocks)
7401 /* turn .* into ^.* with an implied $*=1 */
7403 (OP(NEXTOPER(first)) == REG_ANY)
7406 r->intflags |= (type | PREGf_IMPLICIT);
7407 first = NEXTOPER(first);
7410 if (sawplus && !sawminmod && !sawlookahead
7411 && (!sawopen || !RExC_sawback)
7412 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
7413 /* x+ must match at the 1st pos of run of x's */
7414 r->intflags |= PREGf_SKIP;
7416 /* Scan is after the zeroth branch, first is atomic matcher. */
7417 #ifdef TRIE_STUDY_OPT
7420 Perl_re_printf( aTHX_ "first at %"IVdf"\n",
7421 (IV)(first - scan + 1))
7425 Perl_re_printf( aTHX_ "first at %"IVdf"\n",
7426 (IV)(first - scan + 1))
7432 * If there's something expensive in the r.e., find the
7433 * longest literal string that must appear and make it the
7434 * regmust. Resolve ties in favor of later strings, since
7435 * the regstart check works with the beginning of the r.e.
7436 * and avoiding duplication strengthens checking. Not a
7437 * strong reason, but sufficient in the absence of others.
7438 * [Now we resolve ties in favor of the earlier string if
7439 * it happens that c_offset_min has been invalidated, since the
7440 * earlier string may buy us something the later one won't.]
7443 data.longest_fixed = newSVpvs("");
7444 data.longest_float = newSVpvs("");
7445 data.last_found = newSVpvs("");
7446 data.longest = &(data.longest_fixed);
7447 ENTER_with_name("study_chunk");
7448 SAVEFREESV(data.longest_fixed);
7449 SAVEFREESV(data.longest_float);
7450 SAVEFREESV(data.last_found);
7452 if (!ri->regstclass) {
7453 ssc_init(pRExC_state, &ch_class);
7454 data.start_class = &ch_class;
7455 stclass_flag = SCF_DO_STCLASS_AND;
7456 } else /* XXXX Check for BOUND? */
7458 data.last_closep = &last_close;
7461 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
7462 scan + RExC_size, /* Up to end */
7464 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
7465 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
7469 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
7472 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
7473 && data.last_start_min == 0 && data.last_end > 0
7474 && !RExC_seen_zerolen
7475 && !(RExC_seen & REG_VERBARG_SEEN)
7476 && !(RExC_seen & REG_GPOS_SEEN)
7478 r->extflags |= RXf_CHECK_ALL;
7480 scan_commit(pRExC_state, &data,&minlen,0);
7482 longest_float_length = CHR_SVLEN(data.longest_float);
7484 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
7485 && data.offset_fixed == data.offset_float_min
7486 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
7487 && S_setup_longest (aTHX_ pRExC_state,
7491 &(r->float_end_shift),
7492 data.lookbehind_float,
7493 data.offset_float_min,
7495 longest_float_length,
7496 cBOOL(data.flags & SF_FL_BEFORE_EOL),
7497 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
7499 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
7500 r->float_max_offset = data.offset_float_max;
7501 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
7502 r->float_max_offset -= data.lookbehind_float;
7503 SvREFCNT_inc_simple_void_NN(data.longest_float);
7506 r->float_substr = r->float_utf8 = NULL;
7507 longest_float_length = 0;
7510 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
7512 if (S_setup_longest (aTHX_ pRExC_state,
7514 &(r->anchored_utf8),
7515 &(r->anchored_substr),
7516 &(r->anchored_end_shift),
7517 data.lookbehind_fixed,
7520 longest_fixed_length,
7521 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
7522 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
7524 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
7525 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
7528 r->anchored_substr = r->anchored_utf8 = NULL;
7529 longest_fixed_length = 0;
7531 LEAVE_with_name("study_chunk");
7534 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
7535 ri->regstclass = NULL;
7537 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
7539 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7540 && is_ssc_worth_it(pRExC_state, data.start_class))
7542 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7544 ssc_finalize(pRExC_state, data.start_class);
7546 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7547 StructCopy(data.start_class,
7548 (regnode_ssc*)RExC_rxi->data->data[n],
7550 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7551 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7552 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7553 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7554 Perl_re_printf( aTHX_
7555 "synthetic stclass \"%s\".\n",
7556 SvPVX_const(sv));});
7557 data.start_class = NULL;
7560 /* A temporary algorithm prefers floated substr to fixed one to dig
7562 if (longest_fixed_length > longest_float_length) {
7563 r->substrs->check_ix = 0;
7564 r->check_end_shift = r->anchored_end_shift;
7565 r->check_substr = r->anchored_substr;
7566 r->check_utf8 = r->anchored_utf8;
7567 r->check_offset_min = r->check_offset_max = r->anchored_offset;
7568 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
7569 r->intflags |= PREGf_NOSCAN;
7572 r->substrs->check_ix = 1;
7573 r->check_end_shift = r->float_end_shift;
7574 r->check_substr = r->float_substr;
7575 r->check_utf8 = r->float_utf8;
7576 r->check_offset_min = r->float_min_offset;
7577 r->check_offset_max = r->float_max_offset;
7579 if ((r->check_substr || r->check_utf8) ) {
7580 r->extflags |= RXf_USE_INTUIT;
7581 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7582 r->extflags |= RXf_INTUIT_TAIL;
7584 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
7586 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7587 if ( (STRLEN)minlen < longest_float_length )
7588 minlen= longest_float_length;
7589 if ( (STRLEN)minlen < longest_fixed_length )
7590 minlen= longest_fixed_length;
7594 /* Several toplevels. Best we can is to set minlen. */
7596 regnode_ssc ch_class;
7597 SSize_t last_close = 0;
7599 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
7601 scan = ri->program + 1;
7602 ssc_init(pRExC_state, &ch_class);
7603 data.start_class = &ch_class;
7604 data.last_closep = &last_close;
7607 minlen = study_chunk(pRExC_state,
7608 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7609 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7610 ? SCF_TRIE_DOING_RESTUDY
7614 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7616 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7617 = r->float_substr = r->float_utf8 = NULL;
7619 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7620 && is_ssc_worth_it(pRExC_state, data.start_class))
7622 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7624 ssc_finalize(pRExC_state, data.start_class);
7626 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7627 StructCopy(data.start_class,
7628 (regnode_ssc*)RExC_rxi->data->data[n],
7630 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7631 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7632 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7633 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7634 Perl_re_printf( aTHX_
7635 "synthetic stclass \"%s\".\n",
7636 SvPVX_const(sv));});
7637 data.start_class = NULL;
7641 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7642 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7643 r->maxlen = REG_INFTY;
7646 r->maxlen = RExC_maxlen;
7649 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7650 the "real" pattern. */
7652 Perl_re_printf( aTHX_ "minlen: %"IVdf" r->minlen:%"IVdf" maxlen:%"IVdf"\n",
7653 (IV)minlen, (IV)r->minlen, (IV)RExC_maxlen);
7655 r->minlenret = minlen;
7656 if (r->minlen < minlen)
7659 if (RExC_seen & REG_RECURSE_SEEN ) {
7660 r->intflags |= PREGf_RECURSE_SEEN;
7661 Newxz(r->recurse_locinput, r->nparens + 1, char *);
7663 if (RExC_seen & REG_GPOS_SEEN)
7664 r->intflags |= PREGf_GPOS_SEEN;
7665 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7666 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7668 if (pRExC_state->num_code_blocks)
7669 r->extflags |= RXf_EVAL_SEEN;
7670 if (RExC_seen & REG_VERBARG_SEEN)
7672 r->intflags |= PREGf_VERBARG_SEEN;
7673 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7675 if (RExC_seen & REG_CUTGROUP_SEEN)
7676 r->intflags |= PREGf_CUTGROUP_SEEN;
7677 if (pm_flags & PMf_USE_RE_EVAL)
7678 r->intflags |= PREGf_USE_RE_EVAL;
7679 if (RExC_paren_names)
7680 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7682 RXp_PAREN_NAMES(r) = NULL;
7684 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7685 * so it can be used in pp.c */
7686 if (r->intflags & PREGf_ANCH)
7687 r->extflags |= RXf_IS_ANCHORED;
7691 /* this is used to identify "special" patterns that might result
7692 * in Perl NOT calling the regex engine and instead doing the match "itself",
7693 * particularly special cases in split//. By having the regex compiler
7694 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7695 * we avoid weird issues with equivalent patterns resulting in different behavior,
7696 * AND we allow non Perl engines to get the same optimizations by the setting the
7697 * flags appropriately - Yves */
7698 regnode *first = ri->program + 1;
7700 regnode *next = regnext(first);
7703 if (PL_regkind[fop] == NOTHING && nop == END)
7704 r->extflags |= RXf_NULL;
7705 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
7706 /* when fop is SBOL first->flags will be true only when it was
7707 * produced by parsing /\A/, and not when parsing /^/. This is
7708 * very important for the split code as there we want to
7709 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
7710 * See rt #122761 for more details. -- Yves */
7711 r->extflags |= RXf_START_ONLY;
7712 else if (fop == PLUS
7713 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7715 r->extflags |= RXf_WHITE;
7716 else if ( r->extflags & RXf_SPLIT
7717 && (fop == EXACT || fop == EXACTL)
7718 && STR_LEN(first) == 1
7719 && *(STRING(first)) == ' '
7721 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7725 if (RExC_contains_locale) {
7726 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7730 if (RExC_paren_names) {
7731 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7732 ri->data->data[ri->name_list_idx]
7733 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7736 ri->name_list_idx = 0;
7738 while ( RExC_recurse_count > 0 ) {
7739 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
7740 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - scan );
7743 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7744 /* assume we don't need to swap parens around before we match */
7746 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
7747 (unsigned long)RExC_study_chunk_recursed_count);
7751 Perl_re_printf( aTHX_ "Final program:\n");
7754 #ifdef RE_TRACK_PATTERN_OFFSETS
7755 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7756 const STRLEN len = ri->u.offsets[0];
7758 GET_RE_DEBUG_FLAGS_DECL;
7759 Perl_re_printf( aTHX_
7760 "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
7761 for (i = 1; i <= len; i++) {
7762 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7763 Perl_re_printf( aTHX_ "%"UVuf":%"UVuf"[%"UVuf"] ",
7764 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7766 Perl_re_printf( aTHX_ "\n");
7771 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7772 * by setting the regexp SV to readonly-only instead. If the
7773 * pattern's been recompiled, the USEDness should remain. */
7774 if (old_re && SvREADONLY(old_re))
7782 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7785 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7787 PERL_UNUSED_ARG(value);
7789 if (flags & RXapif_FETCH) {
7790 return reg_named_buff_fetch(rx, key, flags);
7791 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7792 Perl_croak_no_modify();
7794 } else if (flags & RXapif_EXISTS) {
7795 return reg_named_buff_exists(rx, key, flags)
7798 } else if (flags & RXapif_REGNAMES) {
7799 return reg_named_buff_all(rx, flags);
7800 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7801 return reg_named_buff_scalar(rx, flags);
7803 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7809 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7812 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7813 PERL_UNUSED_ARG(lastkey);
7815 if (flags & RXapif_FIRSTKEY)
7816 return reg_named_buff_firstkey(rx, flags);
7817 else if (flags & RXapif_NEXTKEY)
7818 return reg_named_buff_nextkey(rx, flags);
7820 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7827 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7830 AV *retarray = NULL;
7832 struct regexp *const rx = ReANY(r);
7834 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7836 if (flags & RXapif_ALL)
7839 if (rx && RXp_PAREN_NAMES(rx)) {
7840 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7843 SV* sv_dat=HeVAL(he_str);
7844 I32 *nums=(I32*)SvPVX(sv_dat);
7845 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7846 if ((I32)(rx->nparens) >= nums[i]
7847 && rx->offs[nums[i]].start != -1
7848 && rx->offs[nums[i]].end != -1)
7851 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7856 ret = newSVsv(&PL_sv_undef);
7859 av_push(retarray, ret);
7862 return newRV_noinc(MUTABLE_SV(retarray));
7869 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7872 struct regexp *const rx = ReANY(r);
7874 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7876 if (rx && RXp_PAREN_NAMES(rx)) {
7877 if (flags & RXapif_ALL) {
7878 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7880 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7882 SvREFCNT_dec_NN(sv);
7894 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7896 struct regexp *const rx = ReANY(r);
7898 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7900 if ( rx && RXp_PAREN_NAMES(rx) ) {
7901 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7903 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7910 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7912 struct regexp *const rx = ReANY(r);
7913 GET_RE_DEBUG_FLAGS_DECL;
7915 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7917 if (rx && RXp_PAREN_NAMES(rx)) {
7918 HV *hv = RXp_PAREN_NAMES(rx);
7920 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7923 SV* sv_dat = HeVAL(temphe);
7924 I32 *nums = (I32*)SvPVX(sv_dat);
7925 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7926 if ((I32)(rx->lastparen) >= nums[i] &&
7927 rx->offs[nums[i]].start != -1 &&
7928 rx->offs[nums[i]].end != -1)
7934 if (parno || flags & RXapif_ALL) {
7935 return newSVhek(HeKEY_hek(temphe));
7943 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
7948 struct regexp *const rx = ReANY(r);
7950 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
7952 if (rx && RXp_PAREN_NAMES(rx)) {
7953 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
7954 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
7955 } else if (flags & RXapif_ONE) {
7956 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
7957 av = MUTABLE_AV(SvRV(ret));
7958 length = av_tindex(av);
7959 SvREFCNT_dec_NN(ret);
7960 return newSViv(length + 1);
7962 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
7967 return &PL_sv_undef;
7971 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
7973 struct regexp *const rx = ReANY(r);
7976 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
7978 if (rx && RXp_PAREN_NAMES(rx)) {
7979 HV *hv= RXp_PAREN_NAMES(rx);
7981 (void)hv_iterinit(hv);
7982 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7985 SV* sv_dat = HeVAL(temphe);
7986 I32 *nums = (I32*)SvPVX(sv_dat);
7987 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7988 if ((I32)(rx->lastparen) >= nums[i] &&
7989 rx->offs[nums[i]].start != -1 &&
7990 rx->offs[nums[i]].end != -1)
7996 if (parno || flags & RXapif_ALL) {
7997 av_push(av, newSVhek(HeKEY_hek(temphe)));
8002 return newRV_noinc(MUTABLE_SV(av));
8006 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8009 struct regexp *const rx = ReANY(r);
8015 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8017 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8018 || n == RX_BUFF_IDX_CARET_FULLMATCH
8019 || n == RX_BUFF_IDX_CARET_POSTMATCH
8022 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8024 /* on something like
8027 * the KEEPCOPY is set on the PMOP rather than the regex */
8028 if (PL_curpm && r == PM_GETRE(PL_curpm))
8029 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8038 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8039 /* no need to distinguish between them any more */
8040 n = RX_BUFF_IDX_FULLMATCH;
8042 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8043 && rx->offs[0].start != -1)
8045 /* $`, ${^PREMATCH} */
8046 i = rx->offs[0].start;
8050 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8051 && rx->offs[0].end != -1)
8053 /* $', ${^POSTMATCH} */
8054 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8055 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8058 if ( 0 <= n && n <= (I32)rx->nparens &&
8059 (s1 = rx->offs[n].start) != -1 &&
8060 (t1 = rx->offs[n].end) != -1)
8062 /* $&, ${^MATCH}, $1 ... */
8064 s = rx->subbeg + s1 - rx->suboffset;
8069 assert(s >= rx->subbeg);
8070 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8072 #ifdef NO_TAINT_SUPPORT
8073 sv_setpvn(sv, s, i);
8075 const int oldtainted = TAINT_get;
8077 sv_setpvn(sv, s, i);
8078 TAINT_set(oldtainted);
8080 if (RXp_MATCH_UTF8(rx))
8085 if (RXp_MATCH_TAINTED(rx)) {
8086 if (SvTYPE(sv) >= SVt_PVMG) {
8087 MAGIC* const mg = SvMAGIC(sv);
8090 SvMAGIC_set(sv, mg->mg_moremagic);
8092 if ((mgt = SvMAGIC(sv))) {
8093 mg->mg_moremagic = mgt;
8094 SvMAGIC_set(sv, mg);
8105 sv_setsv(sv,&PL_sv_undef);
8111 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8112 SV const * const value)
8114 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8116 PERL_UNUSED_ARG(rx);
8117 PERL_UNUSED_ARG(paren);
8118 PERL_UNUSED_ARG(value);
8121 Perl_croak_no_modify();
8125 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8128 struct regexp *const rx = ReANY(r);
8132 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8134 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8135 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8136 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8139 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8141 /* on something like
8144 * the KEEPCOPY is set on the PMOP rather than the regex */
8145 if (PL_curpm && r == PM_GETRE(PL_curpm))
8146 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8152 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8154 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8155 case RX_BUFF_IDX_PREMATCH: /* $` */
8156 if (rx->offs[0].start != -1) {
8157 i = rx->offs[0].start;
8166 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8167 case RX_BUFF_IDX_POSTMATCH: /* $' */
8168 if (rx->offs[0].end != -1) {
8169 i = rx->sublen - rx->offs[0].end;
8171 s1 = rx->offs[0].end;
8178 default: /* $& / ${^MATCH}, $1, $2, ... */
8179 if (paren <= (I32)rx->nparens &&
8180 (s1 = rx->offs[paren].start) != -1 &&
8181 (t1 = rx->offs[paren].end) != -1)
8187 if (ckWARN(WARN_UNINITIALIZED))
8188 report_uninit((const SV *)sv);
8193 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8194 const char * const s = rx->subbeg - rx->suboffset + s1;
8199 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8206 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8208 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8209 PERL_UNUSED_ARG(rx);
8213 return newSVpvs("Regexp");
8216 /* Scans the name of a named buffer from the pattern.
8217 * If flags is REG_RSN_RETURN_NULL returns null.
8218 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8219 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8220 * to the parsed name as looked up in the RExC_paren_names hash.
8221 * If there is an error throws a vFAIL().. type exception.
8224 #define REG_RSN_RETURN_NULL 0
8225 #define REG_RSN_RETURN_NAME 1
8226 #define REG_RSN_RETURN_DATA 2
8229 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8231 char *name_start = RExC_parse;
8233 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8235 assert (RExC_parse <= RExC_end);
8236 if (RExC_parse == RExC_end) NOOP;
8237 else if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
8238 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8239 * using do...while */
8242 RExC_parse += UTF8SKIP(RExC_parse);
8243 } while (isWORDCHAR_utf8((U8*)RExC_parse));
8247 } while (isWORDCHAR(*RExC_parse));
8249 RExC_parse++; /* so the <- from the vFAIL is after the offending
8251 vFAIL("Group name must start with a non-digit word character");
8255 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8256 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8257 if ( flags == REG_RSN_RETURN_NAME)
8259 else if (flags==REG_RSN_RETURN_DATA) {
8262 if ( ! sv_name ) /* should not happen*/
8263 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8264 if (RExC_paren_names)
8265 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8267 sv_dat = HeVAL(he_str);
8269 vFAIL("Reference to nonexistent named group");
8273 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8274 (unsigned long) flags);
8276 NOT_REACHED; /* NOTREACHED */
8281 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8283 if (RExC_lastparse!=RExC_parse) { \
8284 Perl_re_printf( aTHX_ "%s", \
8285 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8286 RExC_end - RExC_parse, 16, \
8288 PERL_PV_ESCAPE_UNI_DETECT | \
8289 PERL_PV_PRETTY_ELLIPSES | \
8290 PERL_PV_PRETTY_LTGT | \
8291 PERL_PV_ESCAPE_RE | \
8292 PERL_PV_PRETTY_EXACTSIZE \
8296 Perl_re_printf( aTHX_ "%16s",""); \
8299 num = RExC_size + 1; \
8301 num=REG_NODE_NUM(RExC_emit); \
8302 if (RExC_lastnum!=num) \
8303 Perl_re_printf( aTHX_ "|%4d",num); \
8305 Perl_re_printf( aTHX_ "|%4s",""); \
8306 Perl_re_printf( aTHX_ "|%*s%-4s", \
8307 (int)((depth*2)), "", \
8311 RExC_lastparse=RExC_parse; \
8316 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8317 DEBUG_PARSE_MSG((funcname)); \
8318 Perl_re_printf( aTHX_ "%4s","\n"); \
8320 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8321 DEBUG_PARSE_MSG((funcname)); \
8322 Perl_re_printf( aTHX_ fmt "\n",args); \
8325 /* This section of code defines the inversion list object and its methods. The
8326 * interfaces are highly subject to change, so as much as possible is static to
8327 * this file. An inversion list is here implemented as a malloc'd C UV array
8328 * as an SVt_INVLIST scalar.
8330 * An inversion list for Unicode is an array of code points, sorted by ordinal
8331 * number. Each element gives the code point that begins a range that extends
8332 * up-to but not including the code point given by the next element. The final
8333 * element gives the first code point of a range that extends to the platform's
8334 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8335 * ...) give ranges whose code points are all in the inversion list. We say
8336 * that those ranges are in the set. The odd-numbered elements give ranges
8337 * whose code points are not in the inversion list, and hence not in the set.
8338 * Thus, element [0] is the first code point in the list. Element [1]
8339 * is the first code point beyond that not in the list; and element [2] is the
8340 * first code point beyond that that is in the list. In other words, the first
8341 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8342 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8343 * all code points in that range are not in the inversion list. The third
8344 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8345 * list, and so forth. Thus every element whose index is divisible by two
8346 * gives the beginning of a range that is in the list, and every element whose
8347 * index is not divisible by two gives the beginning of a range not in the
8348 * list. If the final element's index is divisible by two, the inversion list
8349 * extends to the platform's infinity; otherwise the highest code point in the
8350 * inversion list is the contents of that element minus 1.
8352 * A range that contains just a single code point N will look like
8354 * invlist[i+1] == N+1
8356 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8357 * impossible to represent, so element [i+1] is omitted. The single element
8359 * invlist[0] == UV_MAX
8360 * contains just UV_MAX, but is interpreted as matching to infinity.
8362 * Taking the complement (inverting) an inversion list is quite simple, if the
8363 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8364 * This implementation reserves an element at the beginning of each inversion
8365 * list to always contain 0; there is an additional flag in the header which
8366 * indicates if the list begins at the 0, or is offset to begin at the next
8367 * element. This means that the inversion list can be inverted without any
8368 * copying; just flip the flag.
8370 * More about inversion lists can be found in "Unicode Demystified"
8371 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8373 * The inversion list data structure is currently implemented as an SV pointing
8374 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8375 * array of UV whose memory management is automatically handled by the existing
8376 * facilities for SV's.
8378 * Some of the methods should always be private to the implementation, and some
8379 * should eventually be made public */
8381 /* The header definitions are in F<invlist_inline.h> */
8383 #ifndef PERL_IN_XSUB_RE
8385 PERL_STATIC_INLINE UV*
8386 S__invlist_array_init(SV* const invlist, const bool will_have_0)
8388 /* Returns a pointer to the first element in the inversion list's array.
8389 * This is called upon initialization of an inversion list. Where the
8390 * array begins depends on whether the list has the code point U+0000 in it
8391 * or not. The other parameter tells it whether the code that follows this
8392 * call is about to put a 0 in the inversion list or not. The first
8393 * element is either the element reserved for 0, if TRUE, or the element
8394 * after it, if FALSE */
8396 bool* offset = get_invlist_offset_addr(invlist);
8397 UV* zero_addr = (UV *) SvPVX(invlist);
8399 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8402 assert(! _invlist_len(invlist));
8406 /* 1^1 = 0; 1^0 = 1 */
8407 *offset = 1 ^ will_have_0;
8408 return zero_addr + *offset;
8413 PERL_STATIC_INLINE void
8414 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8416 /* Sets the current number of elements stored in the inversion list.
8417 * Updates SvCUR correspondingly */
8418 PERL_UNUSED_CONTEXT;
8419 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8421 assert(SvTYPE(invlist) == SVt_INVLIST);
8426 : TO_INTERNAL_SIZE(len + offset));
8427 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8430 #ifndef PERL_IN_XSUB_RE
8433 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
8435 /* Replaces the inversion list in 'dest' with the one from 'src'. It
8436 * steals the list from 'src', so 'src' is made to have a NULL list. This
8437 * is similar to what SvSetMagicSV() would do, if it were implemented on
8438 * inversion lists, though this routine avoids a copy */
8440 const UV src_len = _invlist_len(src);
8441 const bool src_offset = *get_invlist_offset_addr(src);
8442 const STRLEN src_byte_len = SvLEN(src);
8443 char * array = SvPVX(src);
8445 const int oldtainted = TAINT_get;
8447 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
8449 assert(SvTYPE(src) == SVt_INVLIST);
8450 assert(SvTYPE(dest) == SVt_INVLIST);
8451 assert(! invlist_is_iterating(src));
8452 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
8454 /* Make sure it ends in the right place with a NUL, as our inversion list
8455 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
8457 array[src_byte_len - 1] = '\0';
8459 TAINT_NOT; /* Otherwise it breaks */
8460 sv_usepvn_flags(dest,
8464 /* This flag is documented to cause a copy to be avoided */
8465 SV_HAS_TRAILING_NUL);
8466 TAINT_set(oldtainted);
8471 /* Finish up copying over the other fields in an inversion list */
8472 *get_invlist_offset_addr(dest) = src_offset;
8473 invlist_set_len(dest, src_len, src_offset);
8474 *get_invlist_previous_index_addr(dest) = 0;
8475 invlist_iterfinish(dest);
8478 PERL_STATIC_INLINE IV*
8479 S_get_invlist_previous_index_addr(SV* invlist)
8481 /* Return the address of the IV that is reserved to hold the cached index
8483 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8485 assert(SvTYPE(invlist) == SVt_INVLIST);
8487 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8490 PERL_STATIC_INLINE IV
8491 S_invlist_previous_index(SV* const invlist)
8493 /* Returns cached index of previous search */
8495 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8497 return *get_invlist_previous_index_addr(invlist);
8500 PERL_STATIC_INLINE void
8501 S_invlist_set_previous_index(SV* const invlist, const IV index)
8503 /* Caches <index> for later retrieval */
8505 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8507 assert(index == 0 || index < (int) _invlist_len(invlist));
8509 *get_invlist_previous_index_addr(invlist) = index;
8512 PERL_STATIC_INLINE void
8513 S_invlist_trim(SV* invlist)
8515 /* Free the not currently-being-used space in an inversion list */
8517 /* But don't free up the space needed for the 0 UV that is always at the
8518 * beginning of the list, nor the trailing NUL */
8519 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
8521 PERL_ARGS_ASSERT_INVLIST_TRIM;
8523 assert(SvTYPE(invlist) == SVt_INVLIST);
8525 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
8528 PERL_STATIC_INLINE void
8529 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
8531 PERL_ARGS_ASSERT_INVLIST_CLEAR;
8533 assert(SvTYPE(invlist) == SVt_INVLIST);
8535 invlist_set_len(invlist, 0, 0);
8536 invlist_trim(invlist);
8539 #endif /* ifndef PERL_IN_XSUB_RE */
8541 PERL_STATIC_INLINE bool
8542 S_invlist_is_iterating(SV* const invlist)
8544 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8546 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8549 #ifndef PERL_IN_XSUB_RE
8551 PERL_STATIC_INLINE UV
8552 S_invlist_max(SV* const invlist)
8554 /* Returns the maximum number of elements storable in the inversion list's
8555 * array, without having to realloc() */
8557 PERL_ARGS_ASSERT_INVLIST_MAX;
8559 assert(SvTYPE(invlist) == SVt_INVLIST);
8561 /* Assumes worst case, in which the 0 element is not counted in the
8562 * inversion list, so subtracts 1 for that */
8563 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8564 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8565 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8568 Perl__new_invlist(pTHX_ IV initial_size)
8571 /* Return a pointer to a newly constructed inversion list, with enough
8572 * space to store 'initial_size' elements. If that number is negative, a
8573 * system default is used instead */
8577 if (initial_size < 0) {
8581 /* Allocate the initial space */
8582 new_list = newSV_type(SVt_INVLIST);
8584 /* First 1 is in case the zero element isn't in the list; second 1 is for
8586 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8587 invlist_set_len(new_list, 0, 0);
8589 /* Force iterinit() to be used to get iteration to work */
8590 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8592 *get_invlist_previous_index_addr(new_list) = 0;
8598 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8600 /* Return a pointer to a newly constructed inversion list, initialized to
8601 * point to <list>, which has to be in the exact correct inversion list
8602 * form, including internal fields. Thus this is a dangerous routine that
8603 * should not be used in the wrong hands. The passed in 'list' contains
8604 * several header fields at the beginning that are not part of the
8605 * inversion list body proper */
8607 const STRLEN length = (STRLEN) list[0];
8608 const UV version_id = list[1];
8609 const bool offset = cBOOL(list[2]);
8610 #define HEADER_LENGTH 3
8611 /* If any of the above changes in any way, you must change HEADER_LENGTH
8612 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8613 * perl -E 'say int(rand 2**31-1)'
8615 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8616 data structure type, so that one being
8617 passed in can be validated to be an
8618 inversion list of the correct vintage.
8621 SV* invlist = newSV_type(SVt_INVLIST);
8623 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8625 if (version_id != INVLIST_VERSION_ID) {
8626 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8629 /* The generated array passed in includes header elements that aren't part
8630 * of the list proper, so start it just after them */
8631 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8633 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8634 shouldn't touch it */
8636 *(get_invlist_offset_addr(invlist)) = offset;
8638 /* The 'length' passed to us is the physical number of elements in the
8639 * inversion list. But if there is an offset the logical number is one
8641 invlist_set_len(invlist, length - offset, offset);
8643 invlist_set_previous_index(invlist, 0);
8645 /* Initialize the iteration pointer. */
8646 invlist_iterfinish(invlist);
8648 SvREADONLY_on(invlist);
8654 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8656 /* Grow the maximum size of an inversion list */
8658 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8660 assert(SvTYPE(invlist) == SVt_INVLIST);
8662 /* Add one to account for the zero element at the beginning which may not
8663 * be counted by the calling parameters */
8664 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8668 S__append_range_to_invlist(pTHX_ SV* const invlist,
8669 const UV start, const UV end)
8671 /* Subject to change or removal. Append the range from 'start' to 'end' at
8672 * the end of the inversion list. The range must be above any existing
8676 UV max = invlist_max(invlist);
8677 UV len = _invlist_len(invlist);
8680 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8682 if (len == 0) { /* Empty lists must be initialized */
8683 offset = start != 0;
8684 array = _invlist_array_init(invlist, ! offset);
8687 /* Here, the existing list is non-empty. The current max entry in the
8688 * list is generally the first value not in the set, except when the
8689 * set extends to the end of permissible values, in which case it is
8690 * the first entry in that final set, and so this call is an attempt to
8691 * append out-of-order */
8693 UV final_element = len - 1;
8694 array = invlist_array(invlist);
8695 if ( array[final_element] > start
8696 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8698 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",
8699 array[final_element], start,
8700 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8703 /* Here, it is a legal append. If the new range begins 1 above the end
8704 * of the range below it, it is extending the range below it, so the
8705 * new first value not in the set is one greater than the newly
8706 * extended range. */
8707 offset = *get_invlist_offset_addr(invlist);
8708 if (array[final_element] == start) {
8709 if (end != UV_MAX) {
8710 array[final_element] = end + 1;
8713 /* But if the end is the maximum representable on the machine,
8714 * assume that infinity was actually what was meant. Just let
8715 * the range that this would extend to have no end */
8716 invlist_set_len(invlist, len - 1, offset);
8722 /* Here the new range doesn't extend any existing set. Add it */
8724 len += 2; /* Includes an element each for the start and end of range */
8726 /* If wll overflow the existing space, extend, which may cause the array to
8729 invlist_extend(invlist, len);
8731 /* Have to set len here to avoid assert failure in invlist_array() */
8732 invlist_set_len(invlist, len, offset);
8734 array = invlist_array(invlist);
8737 invlist_set_len(invlist, len, offset);
8740 /* The next item on the list starts the range, the one after that is
8741 * one past the new range. */
8742 array[len - 2] = start;
8743 if (end != UV_MAX) {
8744 array[len - 1] = end + 1;
8747 /* But if the end is the maximum representable on the machine, just let
8748 * the range have no end */
8749 invlist_set_len(invlist, len - 1, offset);
8754 Perl__invlist_search(SV* const invlist, const UV cp)
8756 /* Searches the inversion list for the entry that contains the input code
8757 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8758 * return value is the index into the list's array of the range that
8759 * contains <cp>, that is, 'i' such that
8760 * array[i] <= cp < array[i+1]
8765 IV high = _invlist_len(invlist);
8766 const IV highest_element = high - 1;
8769 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8771 /* If list is empty, return failure. */
8776 /* (We can't get the array unless we know the list is non-empty) */
8777 array = invlist_array(invlist);
8779 mid = invlist_previous_index(invlist);
8781 if (mid > highest_element) {
8782 mid = highest_element;
8785 /* <mid> contains the cache of the result of the previous call to this
8786 * function (0 the first time). See if this call is for the same result,
8787 * or if it is for mid-1. This is under the theory that calls to this
8788 * function will often be for related code points that are near each other.
8789 * And benchmarks show that caching gives better results. We also test
8790 * here if the code point is within the bounds of the list. These tests
8791 * replace others that would have had to be made anyway to make sure that
8792 * the array bounds were not exceeded, and these give us extra information
8793 * at the same time */
8794 if (cp >= array[mid]) {
8795 if (cp >= array[highest_element]) {
8796 return highest_element;
8799 /* Here, array[mid] <= cp < array[highest_element]. This means that
8800 * the final element is not the answer, so can exclude it; it also
8801 * means that <mid> is not the final element, so can refer to 'mid + 1'
8803 if (cp < array[mid + 1]) {
8809 else { /* cp < aray[mid] */
8810 if (cp < array[0]) { /* Fail if outside the array */
8814 if (cp >= array[mid - 1]) {
8819 /* Binary search. What we are looking for is <i> such that
8820 * array[i] <= cp < array[i+1]
8821 * The loop below converges on the i+1. Note that there may not be an
8822 * (i+1)th element in the array, and things work nonetheless */
8823 while (low < high) {
8824 mid = (low + high) / 2;
8825 assert(mid <= highest_element);
8826 if (array[mid] <= cp) { /* cp >= array[mid] */
8829 /* We could do this extra test to exit the loop early.
8830 if (cp < array[low]) {
8835 else { /* cp < array[mid] */
8842 invlist_set_previous_index(invlist, high);
8847 Perl__invlist_populate_swatch(SV* const invlist,
8848 const UV start, const UV end, U8* swatch)
8850 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8851 * but is used when the swash has an inversion list. This makes this much
8852 * faster, as it uses a binary search instead of a linear one. This is
8853 * intimately tied to that function, and perhaps should be in utf8.c,
8854 * except it is intimately tied to inversion lists as well. It assumes
8855 * that <swatch> is all 0's on input */
8858 const IV len = _invlist_len(invlist);
8862 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8864 if (len == 0) { /* Empty inversion list */
8868 array = invlist_array(invlist);
8870 /* Find which element it is */
8871 i = _invlist_search(invlist, start);
8873 /* We populate from <start> to <end> */
8874 while (current < end) {
8877 /* The inversion list gives the results for every possible code point
8878 * after the first one in the list. Only those ranges whose index is
8879 * even are ones that the inversion list matches. For the odd ones,
8880 * and if the initial code point is not in the list, we have to skip
8881 * forward to the next element */
8882 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8884 if (i >= len) { /* Finished if beyond the end of the array */
8888 if (current >= end) { /* Finished if beyond the end of what we
8890 if (LIKELY(end < UV_MAX)) {
8894 /* We get here when the upper bound is the maximum
8895 * representable on the machine, and we are looking for just
8896 * that code point. Have to special case it */
8898 goto join_end_of_list;
8901 assert(current >= start);
8903 /* The current range ends one below the next one, except don't go past
8906 upper = (i < len && array[i] < end) ? array[i] : end;
8908 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8909 * for each code point in it */
8910 for (; current < upper; current++) {
8911 const STRLEN offset = (STRLEN)(current - start);
8912 swatch[offset >> 3] |= 1 << (offset & 7);
8917 /* Quit if at the end of the list */
8920 /* But first, have to deal with the highest possible code point on
8921 * the platform. The previous code assumes that <end> is one
8922 * beyond where we want to populate, but that is impossible at the
8923 * platform's infinity, so have to handle it specially */
8924 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8926 const STRLEN offset = (STRLEN)(end - start);
8927 swatch[offset >> 3] |= 1 << (offset & 7);
8932 /* Advance to the next range, which will be for code points not in the
8941 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8942 const bool complement_b, SV** output)
8944 /* Take the union of two inversion lists and point 'output. to it. *output
8945 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8946 * the reference count to that list will be decremented if not already a
8947 * temporary (mortal); otherwise just its contents will be modified to be
8948 * the union. The first list, 'a., may be NULL, in which case a copy of
8949 * the second list is returned. If 'complement_b. is TRUE, the union is
8950 * taken of the complement (inversion) of 'b. instead of b itself.
8952 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8953 * Richard Gillam, published by Addison-Wesley, and explained at some
8954 * length there. The preface says to incorporate its examples into your
8955 * code at your own risk.
8957 * The algorithm is like a merge sort. */
8959 const UV* array_a; /* a's array */
8961 UV len_a; /* length of a's array */
8964 SV* u; /* the resulting union */
8968 UV i_a = 0; /* current index into a's array */
8972 /* running count, as explained in the algorithm source book; items are
8973 * stopped accumulating and are output when the count changes to/from 0.
8974 * The count is incremented when we start a range that's in an input's set,
8975 * and decremented when we start a range that's not in a set. So this
8976 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
8977 * and hence nothing goes into the union; 1, just one of the inputs is in
8978 * its set (and its current range gets added to the union); and 2 when both
8979 * inputs are in their sets. */
8982 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
8985 len_b = _invlist_len(b);
8988 /* Here, 'b' is empty, hence it's complement is all possible code
8989 * points. So if the union includes the complement of 'b', it includes
8990 * everything, and we need not even look at 'a'. It's easiest to
8991 * create a new inversion list that matches everything. */
8993 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
8995 if (*output == NULL) { /* If the output didn't exist, just point it
8997 *output = everything;
9001 /* Otherwise, replace its contents with the new list */
9002 invlist_replace_list_destroys_src(*output, everything);
9003 SvREFCNT_dec_NN(everything);
9007 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9008 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9009 * output will be empty */
9012 *output = _new_invlist(0);
9016 if (_invlist_len(a) == 0) {
9017 invlist_clear(*output);
9021 /* Here, 'a' is not empty, and entirely determines the union. If the
9022 * output is not to overwrite 'b', we can just return 'a'. */
9025 /* If the output is to overwrite 'a', we have a no-op, as it's
9031 /* But otherwise we have to copy 'a' to the output */
9032 *output = invlist_clone(a);
9036 /* Here, 'b' is to be overwritten by the output, which will be 'a' */
9037 u = invlist_clone(a);
9038 invlist_replace_list_destroys_src(*output, u);
9044 /* Here 'b' is not empty. See about 'a' */
9046 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9048 /* Here, 'a' is empty (and b is not). That means the union will come
9049 * entirely from 'b'. If the output is not to overwrite 'a', we can
9050 * just return what's in 'b'. */
9053 /* If the output is to overwrite 'b', it's already in 'b', but
9054 * otherwise we have to copy 'b' to the output */
9056 *output = invlist_clone(b);
9059 /* And if the output is to be the inversion of 'b', do that */
9061 _invlist_invert(*output);
9067 /* Here, 'a', which is empty or even NULL, is to be overwritten by the
9068 * output, which will either be 'b' or the complement of 'b' */
9071 *output = invlist_clone(b);
9074 u = invlist_clone(b);
9075 invlist_replace_list_destroys_src(*output, u);
9080 _invlist_invert(*output);
9086 /* Here both lists exist and are non-empty */
9087 array_a = invlist_array(a);
9088 array_b = invlist_array(b);
9090 /* If are to take the union of 'a' with the complement of b, set it
9091 * up so are looking at b's complement. */
9094 /* To complement, we invert: if the first element is 0, remove it. To
9095 * do this, we just pretend the array starts one later */
9096 if (array_b[0] == 0) {
9102 /* But if the first element is not zero, we pretend the list starts
9103 * at the 0 that is always stored immediately before the array. */
9109 /* Size the union for the worst case: that the sets are completely
9111 u = _new_invlist(len_a + len_b);
9113 /* Will contain U+0000 if either component does */
9114 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9115 || (len_b > 0 && array_b[0] == 0));
9117 /* Go through each input list item by item, stopping when have exhausted
9119 while (i_a < len_a && i_b < len_b) {
9120 UV cp; /* The element to potentially add to the union's array */
9121 bool cp_in_set; /* is it in the the input list's set or not */
9123 /* We need to take one or the other of the two inputs for the union.
9124 * Since we are merging two sorted lists, we take the smaller of the
9125 * next items. In case of a tie, we take first the one that is in its
9126 * set. If we first took the one not in its set, it would decrement
9127 * the count, possibly to 0 which would cause it to be output as ending
9128 * the range, and the next time through we would take the same number,
9129 * and output it again as beginning the next range. By doing it the
9130 * opposite way, there is no possibility that the count will be
9131 * momentarily decremented to 0, and thus the two adjoining ranges will
9132 * be seamlessly merged. (In a tie and both are in the set or both not
9133 * in the set, it doesn't matter which we take first.) */
9134 if ( array_a[i_a] < array_b[i_b]
9135 || ( array_a[i_a] == array_b[i_b]
9136 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9138 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9139 cp = array_a[i_a++];
9142 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9143 cp = array_b[i_b++];
9146 /* Here, have chosen which of the two inputs to look at. Only output
9147 * if the running count changes to/from 0, which marks the
9148 * beginning/end of a range that's in the set */
9151 array_u[i_u++] = cp;
9158 array_u[i_u++] = cp;
9164 /* The loop above increments the index into exactly one of the input lists
9165 * each iteration, and ends when either index gets to its list end. That
9166 * means the other index is lower than its end, and so something is
9167 * remaining in that one. We decrement 'count', as explained below, if
9168 * that list is in its set. (i_a and i_b each currently index the element
9169 * beyond the one we care about.) */
9170 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9171 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9176 /* Above we decremented 'count' if the list that had unexamined elements in
9177 * it was in its set. This has made it so that 'count' being non-zero
9178 * means there isn't anything left to output; and 'count' equal to 0 means
9179 * that what is left to output is precisely that which is left in the
9180 * non-exhausted input list.
9182 * To see why, note first that the exhausted input obviously has nothing
9183 * left to add to the union. If it was in its set at its end, that means
9184 * the set extends from here to the platform's infinity, and hence so does
9185 * the union and the non-exhausted set is irrelevant. The exhausted set
9186 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9187 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9188 * 'count' remains at 1. This is consistent with the decremented 'count'
9189 * != 0 meaning there's nothing left to add to the union.
9191 * But if the exhausted input wasn't in its set, it contributed 0 to
9192 * 'count', and the rest of the union will be whatever the other input is.
9193 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9194 * otherwise it gets decremented to 0. This is consistent with 'count'
9195 * == 0 meaning the remainder of the union is whatever is left in the
9196 * non-exhausted list. */
9201 IV copy_count = len_a - i_a;
9202 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9203 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9205 else { /* The non-exhausted input is b */
9206 copy_count = len_b - i_b;
9207 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9209 len_u = i_u + copy_count;
9212 /* Set the result to the final length, which can change the pointer to
9213 * array_u, so re-find it. (Note that it is unlikely that this will
9214 * change, as we are shrinking the space, not enlarging it) */
9215 if (len_u != _invlist_len(u)) {
9216 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9218 array_u = invlist_array(u);
9221 /* If the output is not to overwrite either of the inputs, just return the
9222 * calculated union */
9223 if (a != *output && b != *output) {
9227 /* Here, the output is to be the same as one of the input scalars,
9228 * hence replacing it. The simple thing to do is to free the input
9229 * scalar, making it instead be the output one. But experience has
9230 * shown [perl #127392] that if the input is a mortal, we can get a
9231 * huge build-up of these during regex compilation before they get
9232 * freed. So for that case, replace just the input's interior with
9233 * the union's, and then free the union */
9235 assert(! invlist_is_iterating(*output));
9237 if (! SvTEMP(*output)) {
9238 SvREFCNT_dec_NN(*output);
9242 invlist_replace_list_destroys_src(*output, u);
9251 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9252 const bool complement_b, SV** i)
9254 /* Take the intersection of two inversion lists and point 'i' to it. *i
9255 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
9256 * the reference count to that list will be decremented if not already a
9257 * temporary (mortal); otherwise just its contents will be modified to be
9258 * the intersection. The first list, 'a', may be NULL, in which case an
9259 * empty list is returned. If 'complement_b' is TRUE, the result will be
9260 * the intersection of 'a' and the complement (or inversion) of 'b' instead
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;
9296 /* Special case if either one is empty */
9297 len_a = (a == NULL) ? 0 : _invlist_len(a);
9298 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9299 if (len_a != 0 && complement_b) {
9301 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9302 * must be empty. Here, also we are using 'b's complement, which
9303 * hence must be every possible code point. Thus the intersection
9306 if (*i == a) { /* No-op */
9310 /* If not overwriting either input, just make a copy of 'a' */
9312 *i = invlist_clone(a);
9316 /* Here we are overwriting 'b' with 'a's contents */
9317 r = invlist_clone(a);
9318 invlist_replace_list_destroys_src(*i, r);
9323 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9324 * intersection must be empty */
9326 *i = _new_invlist(0);
9334 /* Here both lists exist and are non-empty */
9335 array_a = invlist_array(a);
9336 array_b = invlist_array(b);
9338 /* If are to take the intersection of 'a' with the complement of b, set it
9339 * up so are looking at b's complement. */
9342 /* To complement, we invert: if the first element is 0, remove it. To
9343 * do this, we just pretend the array starts one later */
9344 if (array_b[0] == 0) {
9350 /* But if the first element is not zero, we pretend the list starts
9351 * at the 0 that is always stored immediately before the array. */
9357 /* Size the intersection for the worst case: that the intersection ends up
9358 * fragmenting everything to be completely disjoint */
9359 r= _new_invlist(len_a + len_b);
9361 /* Will contain U+0000 iff both components do */
9362 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9363 && len_b > 0 && array_b[0] == 0);
9365 /* Go through each list item by item, stopping when have exhausted one of
9367 while (i_a < len_a && i_b < len_b) {
9368 UV cp; /* The element to potentially add to the intersection's
9370 bool cp_in_set; /* Is it in the input list's set or not */
9372 /* We need to take one or the other of the two inputs for the
9373 * intersection. Since we are merging two sorted lists, we take the
9374 * smaller of the next items. In case of a tie, we take first the one
9375 * that is not in its set (a difference from the union algorithm). If
9376 * we first took the one in its set, it would increment the count,
9377 * possibly to 2 which would cause it to be output as starting a range
9378 * in the intersection, and the next time through we would take that
9379 * same number, and output it again as ending the set. By doing the
9380 * opposite of this, there is no possibility that the count will be
9381 * momentarily incremented to 2. (In a tie and both are in the set or
9382 * both not in the set, it doesn't matter which we take first.) */
9383 if ( array_a[i_a] < array_b[i_b]
9384 || ( array_a[i_a] == array_b[i_b]
9385 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9387 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9388 cp = array_a[i_a++];
9391 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9395 /* Here, have chosen which of the two inputs to look at. Only output
9396 * if the running count changes to/from 2, which marks the
9397 * beginning/end of a range that's in the intersection */
9401 array_r[i_r++] = cp;
9406 array_r[i_r++] = cp;
9413 /* The loop above increments the index into exactly one of the input lists
9414 * each iteration, and ends when either index gets to its list end. That
9415 * means the other index is lower than its end, and so something is
9416 * remaining in that one. We increment 'count', as explained below, if the
9417 * exhausted list was in its set. (i_a and i_b each currently index the
9418 * element beyond the one we care about.) */
9419 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9420 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9425 /* Above we incremented 'count' if the exhausted list was in its set. This
9426 * has made it so that 'count' being below 2 means there is nothing left to
9427 * output; otheriwse what's left to add to the intersection is precisely
9428 * that which is left in the non-exhausted input list.
9430 * To see why, note first that the exhausted input obviously has nothing
9431 * left to affect the intersection. If it was in its set at its end, that
9432 * means the set extends from here to the platform's infinity, and hence
9433 * anything in the non-exhausted's list will be in the intersection, and
9434 * anything not in it won't be. Hence, the rest of the intersection is
9435 * precisely what's in the non-exhausted list The exhausted set also
9436 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9437 * it means 'count' is now at least 2. This is consistent with the
9438 * incremented 'count' being >= 2 means to add the non-exhausted list to
9441 * But if the exhausted input wasn't in its set, it contributed 0 to
9442 * 'count', and the intersection can't include anything further; the
9443 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9444 * incremented. This is consistent with 'count' being < 2 meaning nothing
9445 * further to add to the intersection. */
9446 if (count < 2) { /* Nothing left to put in the intersection. */
9449 else { /* copy the non-exhausted list, unchanged. */
9450 IV copy_count = len_a - i_a;
9451 if (copy_count > 0) { /* a is the one with stuff left */
9452 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9454 else { /* b is the one with stuff left */
9455 copy_count = len_b - i_b;
9456 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9458 len_r = i_r + copy_count;
9461 /* Set the result to the final length, which can change the pointer to
9462 * array_r, so re-find it. (Note that it is unlikely that this will
9463 * change, as we are shrinking the space, not enlarging it) */
9464 if (len_r != _invlist_len(r)) {
9465 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9467 array_r = invlist_array(r);
9470 /* Finish outputting any remaining */
9471 if (count >= 2) { /* At most one will have a non-zero copy count */
9473 if ((copy_count = len_a - i_a) > 0) {
9474 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9476 else if ((copy_count = len_b - i_b) > 0) {
9477 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9481 /* If the output is not to overwrite either of the inputs, just return the
9482 * calculated intersection */
9483 if (a != *i && b != *i) {
9487 /* Here, the output is to be the same as one of the input scalars,
9488 * hence replacing it. The simple thing to do is to free the input
9489 * scalar, making it instead be the output one. But experience has
9490 * shown [perl #127392] that if the input is a mortal, we can get a
9491 * huge build-up of these during regex compilation before they get
9492 * freed. So for that case, replace just the input's interior with
9493 * the output's, and then free the output. A short-cut in this case
9494 * is if the output is empty, we can just set the input to be empty */
9496 assert(! invlist_is_iterating(*i));
9499 SvREFCNT_dec_NN(*i);
9504 invlist_replace_list_destroys_src(*i, r);
9517 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9519 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9520 * set. A pointer to the inversion list is returned. This may actually be
9521 * a new list, in which case the passed in one has been destroyed. The
9522 * passed-in inversion list can be NULL, in which case a new one is created
9523 * with just the one range in it. The new list is not necessarily
9524 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9525 * result of this function. The gain would not be large, and in many
9526 * cases, this is called multiple times on a single inversion list, so
9527 * anything freed may almost immediately be needed again.
9529 * This used to mostly call the 'union' routine, but that is much more
9530 * heavyweight than really needed for a single range addition */
9532 UV* array; /* The array implementing the inversion list */
9533 UV len; /* How many elements in 'array' */
9534 SSize_t i_s; /* index into the invlist array where 'start'
9536 SSize_t i_e = 0; /* And the index where 'end' should go */
9537 UV cur_highest; /* The highest code point in the inversion list
9538 upon entry to this function */
9540 /* This range becomes the whole inversion list if none already existed */
9541 if (invlist == NULL) {
9542 invlist = _new_invlist(2);
9543 _append_range_to_invlist(invlist, start, end);
9547 /* Likewise, if the inversion list is currently empty */
9548 len = _invlist_len(invlist);
9550 _append_range_to_invlist(invlist, start, end);
9554 /* Starting here, we have to know the internals of the list */
9555 array = invlist_array(invlist);
9557 /* If the new range ends higher than the current highest ... */
9558 cur_highest = invlist_highest(invlist);
9559 if (end > cur_highest) {
9561 /* If the whole range is higher, we can just append it */
9562 if (start > cur_highest) {
9563 _append_range_to_invlist(invlist, start, end);
9567 /* Otherwise, add the portion that is higher ... */
9568 _append_range_to_invlist(invlist, cur_highest + 1, end);
9570 /* ... and continue on below to handle the rest. As a result of the
9571 * above append, we know that the index of the end of the range is the
9572 * final even numbered one of the array. Recall that the final element
9573 * always starts a range that extends to infinity. If that range is in
9574 * the set (meaning the set goes from here to infinity), it will be an
9575 * even index, but if it isn't in the set, it's odd, and the final
9576 * range in the set is one less, which is even. */
9577 if (end == UV_MAX) {
9585 /* We have dealt with appending, now see about prepending. If the new
9586 * range starts lower than the current lowest ... */
9587 if (start < array[0]) {
9589 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
9590 * Let the union code handle it, rather than having to know the
9591 * trickiness in two code places. */
9592 if (UNLIKELY(start == 0)) {
9595 range_invlist = _new_invlist(2);
9596 _append_range_to_invlist(range_invlist, start, end);
9598 _invlist_union(invlist, range_invlist, &invlist);
9600 SvREFCNT_dec_NN(range_invlist);
9605 /* If the whole new range comes before the first entry, and doesn't
9606 * extend it, we have to insert it as an additional range */
9607 if (end < array[0] - 1) {
9609 goto splice_in_new_range;
9612 /* Here the new range adjoins the existing first range, extending it
9616 /* And continue on below to handle the rest. We know that the index of
9617 * the beginning of the range is the first one of the array */
9620 else { /* Not prepending any part of the new range to the existing list.
9621 * Find where in the list it should go. This finds i_s, such that:
9622 * invlist[i_s] <= start < array[i_s+1]
9624 i_s = _invlist_search(invlist, start);
9627 /* At this point, any extending before the beginning of the inversion list
9628 * and/or after the end has been done. This has made it so that, in the
9629 * code below, each endpoint of the new range is either in a range that is
9630 * in the set, or is in a gap between two ranges that are. This means we
9631 * don't have to worry about exceeding the array bounds.
9633 * Find where in the list the new range ends (but we can skip this if we
9634 * have already determined what it is, or if it will be the same as i_s,
9635 * which we already have computed) */
9637 i_e = (start == end)
9639 : _invlist_search(invlist, end);
9642 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
9643 * is a range that goes to infinity there is no element at invlist[i_e+1],
9644 * so only the first relation holds. */
9646 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9648 /* Here, the ranges on either side of the beginning of the new range
9649 * are in the set, and this range starts in the gap between them.
9651 * The new range extends the range above it downwards if the new range
9652 * ends at or above that range's start */
9653 const bool extends_the_range_above = ( end == UV_MAX
9654 || end + 1 >= array[i_s+1]);
9656 /* The new range extends the range below it upwards if it begins just
9657 * after where that range ends */
9658 if (start == array[i_s]) {
9660 /* If the new range fills the entire gap between the other ranges,
9661 * they will get merged together. Other ranges may also get
9662 * merged, depending on how many of them the new range spans. In
9663 * the general case, we do the merge later, just once, after we
9664 * figure out how many to merge. But in the case where the new
9665 * range exactly spans just this one gap (possibly extending into
9666 * the one above), we do the merge here, and an early exit. This
9667 * is done here to avoid having to special case later. */
9668 if (i_e - i_s <= 1) {
9670 /* If i_e - i_s == 1, it means that the new range terminates
9671 * within the range above, and hence 'extends_the_range_above'
9672 * must be true. (If the range above it extends to infinity,
9673 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
9674 * will be 0, so no harm done.) */
9675 if (extends_the_range_above) {
9676 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
9677 invlist_set_len(invlist,
9679 *(get_invlist_offset_addr(invlist)));
9683 /* Here, i_e must == i_s. We keep them in sync, as they apply
9684 * to the same range, and below we are about to decrement i_s
9689 /* Here, the new range is adjacent to the one below. (It may also
9690 * span beyond the range above, but that will get resolved later.)
9691 * Extend the range below to include this one. */
9692 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
9696 else if (extends_the_range_above) {
9698 /* Here the new range only extends the range above it, but not the
9699 * one below. It merges with the one above. Again, we keep i_e
9700 * and i_s in sync if they point to the same range */
9709 /* Here, we've dealt with the new range start extending any adjoining
9712 * If the new range extends to infinity, it is now the final one,
9713 * regardless of what was there before */
9714 if (UNLIKELY(end == UV_MAX)) {
9715 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
9719 /* If i_e started as == i_s, it has also been dealt with,
9720 * and been updated to the new i_s, which will fail the following if */
9721 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
9723 /* Here, the ranges on either side of the end of the new range are in
9724 * the set, and this range ends in the gap between them.
9726 * If this range is adjacent to (hence extends) the range above it, it
9727 * becomes part of that range; likewise if it extends the range below,
9728 * it becomes part of that range */
9729 if (end + 1 == array[i_e+1]) {
9733 else if (start <= array[i_e]) {
9734 array[i_e] = end + 1;
9741 /* If the range fits entirely in an existing range (as possibly already
9742 * extended above), it doesn't add anything new */
9743 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9747 /* Here, no part of the range is in the list. Must add it. It will
9748 * occupy 2 more slots */
9749 splice_in_new_range:
9751 invlist_extend(invlist, len + 2);
9752 array = invlist_array(invlist);
9753 /* Move the rest of the array down two slots. Don't include any
9755 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
9757 /* Do the actual splice */
9758 array[i_e+1] = start;
9759 array[i_e+2] = end + 1;
9760 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
9764 /* Here the new range crossed the boundaries of a pre-existing range. The
9765 * code above has adjusted things so that both ends are in ranges that are
9766 * in the set. This means everything in between must also be in the set.
9767 * Just squash things together */
9768 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
9769 invlist_set_len(invlist,
9771 *(get_invlist_offset_addr(invlist)));
9777 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9778 UV** other_elements_ptr)
9780 /* Create and return an inversion list whose contents are to be populated
9781 * by the caller. The caller gives the number of elements (in 'size') and
9782 * the very first element ('element0'). This function will set
9783 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9786 * Obviously there is some trust involved that the caller will properly
9787 * fill in the other elements of the array.
9789 * (The first element needs to be passed in, as the underlying code does
9790 * things differently depending on whether it is zero or non-zero) */
9792 SV* invlist = _new_invlist(size);
9795 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9797 invlist = add_cp_to_invlist(invlist, element0);
9798 offset = *get_invlist_offset_addr(invlist);
9800 invlist_set_len(invlist, size, offset);
9801 *other_elements_ptr = invlist_array(invlist) + 1;
9807 PERL_STATIC_INLINE SV*
9808 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9809 return _add_range_to_invlist(invlist, cp, cp);
9812 #ifndef PERL_IN_XSUB_RE
9814 Perl__invlist_invert(pTHX_ SV* const invlist)
9816 /* Complement the input inversion list. This adds a 0 if the list didn't
9817 * have a zero; removes it otherwise. As described above, the data
9818 * structure is set up so that this is very efficient */
9820 PERL_ARGS_ASSERT__INVLIST_INVERT;
9822 assert(! invlist_is_iterating(invlist));
9824 /* The inverse of matching nothing is matching everything */
9825 if (_invlist_len(invlist) == 0) {
9826 _append_range_to_invlist(invlist, 0, UV_MAX);
9830 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9835 PERL_STATIC_INLINE SV*
9836 S_invlist_clone(pTHX_ SV* const invlist)
9839 /* Return a new inversion list that is a copy of the input one, which is
9840 * unchanged. The new list will not be mortal even if the old one was. */
9842 /* Need to allocate extra space to accommodate Perl's addition of a
9843 * trailing NUL to SvPV's, since it thinks they are always strings */
9844 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9845 STRLEN physical_length = SvCUR(invlist);
9846 bool offset = *(get_invlist_offset_addr(invlist));
9848 PERL_ARGS_ASSERT_INVLIST_CLONE;
9850 *(get_invlist_offset_addr(new_invlist)) = offset;
9851 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9852 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9857 PERL_STATIC_INLINE STRLEN*
9858 S_get_invlist_iter_addr(SV* invlist)
9860 /* Return the address of the UV that contains the current iteration
9863 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9865 assert(SvTYPE(invlist) == SVt_INVLIST);
9867 return &(((XINVLIST*) SvANY(invlist))->iterator);
9870 PERL_STATIC_INLINE void
9871 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9873 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9875 *get_invlist_iter_addr(invlist) = 0;
9878 PERL_STATIC_INLINE void
9879 S_invlist_iterfinish(SV* invlist)
9881 /* Terminate iterator for invlist. This is to catch development errors.
9882 * Any iteration that is interrupted before completed should call this
9883 * function. Functions that add code points anywhere else but to the end
9884 * of an inversion list assert that they are not in the middle of an
9885 * iteration. If they were, the addition would make the iteration
9886 * problematical: if the iteration hadn't reached the place where things
9887 * were being added, it would be ok */
9889 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9891 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9895 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9897 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9898 * This call sets in <*start> and <*end>, the next range in <invlist>.
9899 * Returns <TRUE> if successful and the next call will return the next
9900 * range; <FALSE> if was already at the end of the list. If the latter,
9901 * <*start> and <*end> are unchanged, and the next call to this function
9902 * will start over at the beginning of the list */
9904 STRLEN* pos = get_invlist_iter_addr(invlist);
9905 UV len = _invlist_len(invlist);
9908 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9911 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9915 array = invlist_array(invlist);
9917 *start = array[(*pos)++];
9923 *end = array[(*pos)++] - 1;
9929 PERL_STATIC_INLINE UV
9930 S_invlist_highest(SV* const invlist)
9932 /* Returns the highest code point that matches an inversion list. This API
9933 * has an ambiguity, as it returns 0 under either the highest is actually
9934 * 0, or if the list is empty. If this distinction matters to you, check
9935 * for emptiness before calling this function */
9937 UV len = _invlist_len(invlist);
9940 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9946 array = invlist_array(invlist);
9948 /* The last element in the array in the inversion list always starts a
9949 * range that goes to infinity. That range may be for code points that are
9950 * matched in the inversion list, or it may be for ones that aren't
9951 * matched. In the latter case, the highest code point in the set is one
9952 * less than the beginning of this range; otherwise it is the final element
9953 * of this range: infinity */
9954 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
9956 : array[len - 1] - 1;
9960 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
9962 /* Get the contents of an inversion list into a string SV so that they can
9963 * be printed out. If 'traditional_style' is TRUE, it uses the format
9964 * traditionally done for debug tracing; otherwise it uses a format
9965 * suitable for just copying to the output, with blanks between ranges and
9966 * a dash between range components */
9970 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
9971 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
9973 if (traditional_style) {
9974 output = newSVpvs("\n");
9977 output = newSVpvs("");
9980 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
9982 assert(! invlist_is_iterating(invlist));
9984 invlist_iterinit(invlist);
9985 while (invlist_iternext(invlist, &start, &end)) {
9986 if (end == UV_MAX) {
9987 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"%cINFINITY%c",
9988 start, intra_range_delimiter,
9989 inter_range_delimiter);
9991 else if (end != start) {
9992 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"%c%04"UVXf"%c",
9994 intra_range_delimiter,
9995 end, inter_range_delimiter);
9998 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"%c",
9999 start, inter_range_delimiter);
10003 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10004 SvCUR_set(output, SvCUR(output) - 1);
10010 #ifndef PERL_IN_XSUB_RE
10012 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10013 const char * const indent, SV* const invlist)
10015 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10016 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10017 * the string 'indent'. The output looks like this:
10018 [0] 0x000A .. 0x000D
10020 [4] 0x2028 .. 0x2029
10021 [6] 0x3104 .. INFINITY
10022 * This means that the first range of code points matched by the list are
10023 * 0xA through 0xD; the second range contains only the single code point
10024 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10025 * are used to define each range (except if the final range extends to
10026 * infinity, only a single element is needed). The array index of the
10027 * first element for the corresponding range is given in brackets. */
10032 PERL_ARGS_ASSERT__INVLIST_DUMP;
10034 if (invlist_is_iterating(invlist)) {
10035 Perl_dump_indent(aTHX_ level, file,
10036 "%sCan't dump inversion list because is in middle of iterating\n",
10041 invlist_iterinit(invlist);
10042 while (invlist_iternext(invlist, &start, &end)) {
10043 if (end == UV_MAX) {
10044 Perl_dump_indent(aTHX_ level, file,
10045 "%s[%"UVuf"] 0x%04"UVXf" .. INFINITY\n",
10046 indent, (UV)count, start);
10048 else if (end != start) {
10049 Perl_dump_indent(aTHX_ level, file,
10050 "%s[%"UVuf"] 0x%04"UVXf" .. 0x%04"UVXf"\n",
10051 indent, (UV)count, start, end);
10054 Perl_dump_indent(aTHX_ level, file, "%s[%"UVuf"] 0x%04"UVXf"\n",
10055 indent, (UV)count, start);
10062 Perl__load_PL_utf8_foldclosures (pTHX)
10064 assert(! PL_utf8_foldclosures);
10066 /* If the folds haven't been read in, call a fold function
10068 if (! PL_utf8_tofold) {
10069 U8 dummy[UTF8_MAXBYTES_CASE+1];
10071 /* This string is just a short named one above \xff */
10072 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
10073 assert(PL_utf8_tofold); /* Verify that worked */
10075 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
10079 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10081 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10083 /* Return a boolean as to if the two passed in inversion lists are
10084 * identical. The final argument, if TRUE, says to take the complement of
10085 * the second inversion list before doing the comparison */
10087 const UV* array_a = invlist_array(a);
10088 const UV* array_b = invlist_array(b);
10089 UV len_a = _invlist_len(a);
10090 UV len_b = _invlist_len(b);
10092 UV i = 0; /* current index into the arrays */
10093 bool retval = TRUE; /* Assume are identical until proven otherwise */
10095 PERL_ARGS_ASSERT__INVLISTEQ;
10097 /* If are to compare 'a' with the complement of b, set it
10098 * up so are looking at b's complement. */
10099 if (complement_b) {
10101 /* The complement of nothing is everything, so <a> would have to have
10102 * just one element, starting at zero (ending at infinity) */
10104 return (len_a == 1 && array_a[0] == 0);
10106 else if (array_b[0] == 0) {
10108 /* Otherwise, to complement, we invert. Here, the first element is
10109 * 0, just remove it. To do this, we just pretend the array starts
10117 /* But if the first element is not zero, we pretend the list starts
10118 * at the 0 that is always stored immediately before the array. */
10124 /* Make sure that the lengths are the same, as well as the final element
10125 * before looping through the remainder. (Thus we test the length, final,
10126 * and first elements right off the bat) */
10127 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
10130 else for (i = 0; i < len_a - 1; i++) {
10131 if (array_a[i] != array_b[i]) {
10142 * As best we can, determine the characters that can match the start of
10143 * the given EXACTF-ish node.
10145 * Returns the invlist as a new SV*; it is the caller's responsibility to
10146 * call SvREFCNT_dec() when done with it.
10149 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10151 const U8 * s = (U8*)STRING(node);
10152 SSize_t bytelen = STR_LEN(node);
10154 /* Start out big enough for 2 separate code points */
10155 SV* invlist = _new_invlist(4);
10157 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
10162 /* We punt and assume can match anything if the node begins
10163 * with a multi-character fold. Things are complicated. For
10164 * example, /ffi/i could match any of:
10165 * "\N{LATIN SMALL LIGATURE FFI}"
10166 * "\N{LATIN SMALL LIGATURE FF}I"
10167 * "F\N{LATIN SMALL LIGATURE FI}"
10168 * plus several other things; and making sure we have all the
10169 * possibilities is hard. */
10170 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10171 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10174 /* Any Latin1 range character can potentially match any
10175 * other depending on the locale */
10176 if (OP(node) == EXACTFL) {
10177 _invlist_union(invlist, PL_Latin1, &invlist);
10180 /* But otherwise, it matches at least itself. We can
10181 * quickly tell if it has a distinct fold, and if so,
10182 * it matches that as well */
10183 invlist = add_cp_to_invlist(invlist, uc);
10184 if (IS_IN_SOME_FOLD_L1(uc))
10185 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10188 /* Some characters match above-Latin1 ones under /i. This
10189 * is true of EXACTFL ones when the locale is UTF-8 */
10190 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10191 && (! isASCII(uc) || (OP(node) != EXACTFA
10192 && OP(node) != EXACTFA_NO_TRIE)))
10194 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10198 else { /* Pattern is UTF-8 */
10199 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10200 STRLEN foldlen = UTF8SKIP(s);
10201 const U8* e = s + bytelen;
10204 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10206 /* The only code points that aren't folded in a UTF EXACTFish
10207 * node are are the problematic ones in EXACTFL nodes */
10208 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10209 /* We need to check for the possibility that this EXACTFL
10210 * node begins with a multi-char fold. Therefore we fold
10211 * the first few characters of it so that we can make that
10216 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10218 *(d++) = (U8) toFOLD(*s);
10223 to_utf8_fold(s, d, &len);
10229 /* And set up so the code below that looks in this folded
10230 * buffer instead of the node's string */
10232 foldlen = UTF8SKIP(folded);
10236 /* When we reach here 's' points to the fold of the first
10237 * character(s) of the node; and 'e' points to far enough along
10238 * the folded string to be just past any possible multi-char
10239 * fold. 'foldlen' is the length in bytes of the first
10242 * Unlike the non-UTF-8 case, the macro for determining if a
10243 * string is a multi-char fold requires all the characters to
10244 * already be folded. This is because of all the complications
10245 * if not. Note that they are folded anyway, except in EXACTFL
10246 * nodes. Like the non-UTF case above, we punt if the node
10247 * begins with a multi-char fold */
10249 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10250 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10252 else { /* Single char fold */
10254 /* It matches all the things that fold to it, which are
10255 * found in PL_utf8_foldclosures (including itself) */
10256 invlist = add_cp_to_invlist(invlist, uc);
10257 if (! PL_utf8_foldclosures)
10258 _load_PL_utf8_foldclosures();
10259 if ((listp = hv_fetch(PL_utf8_foldclosures,
10260 (char *) s, foldlen, FALSE)))
10262 AV* list = (AV*) *listp;
10264 for (k = 0; k <= av_tindex_nomg(list); k++) {
10265 SV** c_p = av_fetch(list, k, FALSE);
10271 /* /aa doesn't allow folds between ASCII and non- */
10272 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
10273 && isASCII(c) != isASCII(uc))
10278 invlist = add_cp_to_invlist(invlist, c);
10287 #undef HEADER_LENGTH
10288 #undef TO_INTERNAL_SIZE
10289 #undef FROM_INTERNAL_SIZE
10290 #undef INVLIST_VERSION_ID
10292 /* End of inversion list object */
10295 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10297 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10298 * constructs, and updates RExC_flags with them. On input, RExC_parse
10299 * should point to the first flag; it is updated on output to point to the
10300 * final ')' or ':'. There needs to be at least one flag, or this will
10303 /* for (?g), (?gc), and (?o) warnings; warning
10304 about (?c) will warn about (?g) -- japhy */
10306 #define WASTED_O 0x01
10307 #define WASTED_G 0x02
10308 #define WASTED_C 0x04
10309 #define WASTED_GC (WASTED_G|WASTED_C)
10310 I32 wastedflags = 0x00;
10311 U32 posflags = 0, negflags = 0;
10312 U32 *flagsp = &posflags;
10313 char has_charset_modifier = '\0';
10315 bool has_use_defaults = FALSE;
10316 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10317 int x_mod_count = 0;
10319 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10321 /* '^' as an initial flag sets certain defaults */
10322 if (UCHARAT(RExC_parse) == '^') {
10324 has_use_defaults = TRUE;
10325 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10326 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
10327 ? REGEX_UNICODE_CHARSET
10328 : REGEX_DEPENDS_CHARSET);
10331 cs = get_regex_charset(RExC_flags);
10332 if (cs == REGEX_DEPENDS_CHARSET
10333 && (RExC_utf8 || RExC_uni_semantics))
10335 cs = REGEX_UNICODE_CHARSET;
10338 while (RExC_parse < RExC_end) {
10339 /* && strchr("iogcmsx", *RExC_parse) */
10340 /* (?g), (?gc) and (?o) are useless here
10341 and must be globally applied -- japhy */
10342 switch (*RExC_parse) {
10344 /* Code for the imsxn flags */
10345 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10347 case LOCALE_PAT_MOD:
10348 if (has_charset_modifier) {
10349 goto excess_modifier;
10351 else if (flagsp == &negflags) {
10354 cs = REGEX_LOCALE_CHARSET;
10355 has_charset_modifier = LOCALE_PAT_MOD;
10357 case UNICODE_PAT_MOD:
10358 if (has_charset_modifier) {
10359 goto excess_modifier;
10361 else if (flagsp == &negflags) {
10364 cs = REGEX_UNICODE_CHARSET;
10365 has_charset_modifier = UNICODE_PAT_MOD;
10367 case ASCII_RESTRICT_PAT_MOD:
10368 if (flagsp == &negflags) {
10371 if (has_charset_modifier) {
10372 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10373 goto excess_modifier;
10375 /* Doubled modifier implies more restricted */
10376 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10379 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10381 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10383 case DEPENDS_PAT_MOD:
10384 if (has_use_defaults) {
10385 goto fail_modifiers;
10387 else if (flagsp == &negflags) {
10390 else if (has_charset_modifier) {
10391 goto excess_modifier;
10394 /* The dual charset means unicode semantics if the
10395 * pattern (or target, not known until runtime) are
10396 * utf8, or something in the pattern indicates unicode
10398 cs = (RExC_utf8 || RExC_uni_semantics)
10399 ? REGEX_UNICODE_CHARSET
10400 : REGEX_DEPENDS_CHARSET;
10401 has_charset_modifier = DEPENDS_PAT_MOD;
10405 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10406 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10408 else if (has_charset_modifier == *(RExC_parse - 1)) {
10409 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10410 *(RExC_parse - 1));
10413 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10415 NOT_REACHED; /*NOTREACHED*/
10418 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10419 *(RExC_parse - 1));
10420 NOT_REACHED; /*NOTREACHED*/
10421 case ONCE_PAT_MOD: /* 'o' */
10422 case GLOBAL_PAT_MOD: /* 'g' */
10423 if (PASS2 && ckWARN(WARN_REGEXP)) {
10424 const I32 wflagbit = *RExC_parse == 'o'
10427 if (! (wastedflags & wflagbit) ) {
10428 wastedflags |= wflagbit;
10429 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10432 "Useless (%s%c) - %suse /%c modifier",
10433 flagsp == &negflags ? "?-" : "?",
10435 flagsp == &negflags ? "don't " : "",
10442 case CONTINUE_PAT_MOD: /* 'c' */
10443 if (PASS2 && ckWARN(WARN_REGEXP)) {
10444 if (! (wastedflags & WASTED_C) ) {
10445 wastedflags |= WASTED_GC;
10446 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10449 "Useless (%sc) - %suse /gc modifier",
10450 flagsp == &negflags ? "?-" : "?",
10451 flagsp == &negflags ? "don't " : ""
10456 case KEEPCOPY_PAT_MOD: /* 'p' */
10457 if (flagsp == &negflags) {
10459 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10461 *flagsp |= RXf_PMf_KEEPCOPY;
10465 /* A flag is a default iff it is following a minus, so
10466 * if there is a minus, it means will be trying to
10467 * re-specify a default which is an error */
10468 if (has_use_defaults || flagsp == &negflags) {
10469 goto fail_modifiers;
10471 flagsp = &negflags;
10472 wastedflags = 0; /* reset so (?g-c) warns twice */
10476 RExC_flags |= posflags;
10477 RExC_flags &= ~negflags;
10478 set_regex_charset(&RExC_flags, cs);
10479 if (RExC_flags & RXf_PMf_FOLD) {
10480 RExC_contains_i = 1;
10483 if (UNLIKELY((x_mod_count) > 1)) {
10484 vFAIL("Only one /x regex modifier is allowed");
10490 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10491 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10492 vFAIL2utf8f("Sequence (%"UTF8f"...) not recognized",
10493 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10494 NOT_REACHED; /*NOTREACHED*/
10497 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10500 vFAIL("Sequence (?... not terminated");
10504 - reg - regular expression, i.e. main body or parenthesized thing
10506 * Caller must absorb opening parenthesis.
10508 * Combining parenthesis handling with the base level of regular expression
10509 * is a trifle forced, but the need to tie the tails of the branches to what
10510 * follows makes it hard to avoid.
10512 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10514 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10516 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10519 PERL_STATIC_INLINE regnode *
10520 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10522 char * parse_start,
10527 char* name_start = RExC_parse;
10529 SV *sv_dat = reg_scan_name(pRExC_state, SIZE_ONLY
10530 ? REG_RSN_RETURN_NULL
10531 : REG_RSN_RETURN_DATA);
10532 GET_RE_DEBUG_FLAGS_DECL;
10534 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10536 if (RExC_parse == name_start || *RExC_parse != ch) {
10537 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10538 vFAIL2("Sequence %.3s... not terminated",parse_start);
10542 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10543 RExC_rxi->data->data[num]=(void*)sv_dat;
10544 SvREFCNT_inc_simple_void(sv_dat);
10547 ret = reganode(pRExC_state,
10550 : (ASCII_FOLD_RESTRICTED)
10552 : (AT_LEAST_UNI_SEMANTICS)
10558 *flagp |= HASWIDTH;
10560 Set_Node_Offset(ret, parse_start+1);
10561 Set_Node_Cur_Length(ret, parse_start);
10563 nextchar(pRExC_state);
10567 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
10568 flags. Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan
10569 needs to be restarted, or'd with NEED_UTF8 if the pattern needs to be
10570 upgraded to UTF-8. Otherwise would only return NULL if regbranch() returns
10571 NULL, which cannot happen. */
10573 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
10574 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10575 * 2 is like 1, but indicates that nextchar() has been called to advance
10576 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10577 * this flag alerts us to the need to check for that */
10579 regnode *ret; /* Will be the head of the group. */
10582 regnode *ender = NULL;
10585 U32 oregflags = RExC_flags;
10586 bool have_branch = 0;
10588 I32 freeze_paren = 0;
10589 I32 after_freeze = 0;
10590 I32 num; /* numeric backreferences */
10592 char * parse_start = RExC_parse; /* MJD */
10593 char * const oregcomp_parse = RExC_parse;
10595 GET_RE_DEBUG_FLAGS_DECL;
10597 PERL_ARGS_ASSERT_REG;
10598 DEBUG_PARSE("reg ");
10600 *flagp = 0; /* Tentatively. */
10602 /* Having this true makes it feasible to have a lot fewer tests for the
10603 * parse pointer being in scope. For example, we can write
10604 * while(isFOO(*RExC_parse)) RExC_parse++;
10606 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
10608 assert(*RExC_end == '\0');
10610 /* Make an OPEN node, if parenthesized. */
10613 /* Under /x, space and comments can be gobbled up between the '(' and
10614 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
10615 * intervening space, as the sequence is a token, and a token should be
10617 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
10619 if (RExC_parse >= RExC_end) {
10620 vFAIL("Unmatched (");
10623 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
10624 char *start_verb = RExC_parse + 1;
10626 char *start_arg = NULL;
10627 unsigned char op = 0;
10628 int arg_required = 0;
10629 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
10631 if (has_intervening_patws) {
10632 RExC_parse++; /* past the '*' */
10633 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
10635 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
10636 if ( *RExC_parse == ':' ) {
10637 start_arg = RExC_parse + 1;
10640 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10642 verb_len = RExC_parse - start_verb;
10644 if (RExC_parse >= RExC_end) {
10645 goto unterminated_verb_pattern;
10647 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10648 while ( RExC_parse < RExC_end && *RExC_parse != ')' )
10649 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10650 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10651 unterminated_verb_pattern:
10652 vFAIL("Unterminated verb pattern argument");
10653 if ( RExC_parse == start_arg )
10656 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10657 vFAIL("Unterminated verb pattern");
10660 /* Here, we know that RExC_parse < RExC_end */
10662 switch ( *start_verb ) {
10663 case 'A': /* (*ACCEPT) */
10664 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
10666 internal_argval = RExC_nestroot;
10669 case 'C': /* (*COMMIT) */
10670 if ( memEQs(start_verb,verb_len,"COMMIT") )
10673 case 'F': /* (*FAIL) */
10674 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
10678 case ':': /* (*:NAME) */
10679 case 'M': /* (*MARK:NAME) */
10680 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
10685 case 'P': /* (*PRUNE) */
10686 if ( memEQs(start_verb,verb_len,"PRUNE") )
10689 case 'S': /* (*SKIP) */
10690 if ( memEQs(start_verb,verb_len,"SKIP") )
10693 case 'T': /* (*THEN) */
10694 /* [19:06] <TimToady> :: is then */
10695 if ( memEQs(start_verb,verb_len,"THEN") ) {
10697 RExC_seen |= REG_CUTGROUP_SEEN;
10702 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10704 "Unknown verb pattern '%"UTF8f"'",
10705 UTF8fARG(UTF, verb_len, start_verb));
10707 if ( arg_required && !start_arg ) {
10708 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
10709 verb_len, start_verb);
10711 if (internal_argval == -1) {
10712 ret = reganode(pRExC_state, op, 0);
10714 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
10716 RExC_seen |= REG_VERBARG_SEEN;
10717 if ( ! SIZE_ONLY ) {
10719 SV *sv = newSVpvn( start_arg,
10720 RExC_parse - start_arg);
10721 ARG(ret) = add_data( pRExC_state,
10722 STR_WITH_LEN("S"));
10723 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
10728 if ( internal_argval != -1 )
10729 ARG2L_SET(ret, internal_argval);
10731 nextchar(pRExC_state);
10734 else if (*RExC_parse == '?') { /* (?...) */
10735 bool is_logical = 0;
10736 const char * const seqstart = RExC_parse;
10737 const char * endptr;
10738 if (has_intervening_patws) {
10740 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
10743 RExC_parse++; /* past the '?' */
10744 paren = *RExC_parse; /* might be a trailing NUL, if not
10746 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10747 if (RExC_parse > RExC_end) {
10750 ret = NULL; /* For look-ahead/behind. */
10753 case 'P': /* (?P...) variants for those used to PCRE/Python */
10754 paren = *RExC_parse;
10755 if ( paren == '<') { /* (?P<...>) named capture */
10757 if (RExC_parse >= RExC_end) {
10758 vFAIL("Sequence (?P<... not terminated");
10760 goto named_capture;
10762 else if (paren == '>') { /* (?P>name) named recursion */
10764 if (RExC_parse >= RExC_end) {
10765 vFAIL("Sequence (?P>... not terminated");
10767 goto named_recursion;
10769 else if (paren == '=') { /* (?P=...) named backref */
10771 return handle_named_backref(pRExC_state, flagp,
10774 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10775 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10776 vFAIL3("Sequence (%.*s...) not recognized",
10777 RExC_parse-seqstart, seqstart);
10778 NOT_REACHED; /*NOTREACHED*/
10779 case '<': /* (?<...) */
10780 if (*RExC_parse == '!')
10782 else if (*RExC_parse != '=')
10789 case '\'': /* (?'...') */
10790 name_start = RExC_parse;
10791 svname = reg_scan_name(pRExC_state,
10792 SIZE_ONLY /* reverse test from the others */
10793 ? REG_RSN_RETURN_NAME
10794 : REG_RSN_RETURN_NULL);
10795 if ( RExC_parse == name_start
10796 || RExC_parse >= RExC_end
10797 || *RExC_parse != paren)
10799 vFAIL2("Sequence (?%c... not terminated",
10800 paren=='>' ? '<' : paren);
10805 if (!svname) /* shouldn't happen */
10807 "panic: reg_scan_name returned NULL");
10808 if (!RExC_paren_names) {
10809 RExC_paren_names= newHV();
10810 sv_2mortal(MUTABLE_SV(RExC_paren_names));
10812 RExC_paren_name_list= newAV();
10813 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10816 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10818 sv_dat = HeVAL(he_str);
10820 /* croak baby croak */
10822 "panic: paren_name hash element allocation failed");
10823 } else if ( SvPOK(sv_dat) ) {
10824 /* (?|...) can mean we have dupes so scan to check
10825 its already been stored. Maybe a flag indicating
10826 we are inside such a construct would be useful,
10827 but the arrays are likely to be quite small, so
10828 for now we punt -- dmq */
10829 IV count = SvIV(sv_dat);
10830 I32 *pv = (I32*)SvPVX(sv_dat);
10832 for ( i = 0 ; i < count ; i++ ) {
10833 if ( pv[i] == RExC_npar ) {
10839 pv = (I32*)SvGROW(sv_dat,
10840 SvCUR(sv_dat) + sizeof(I32)+1);
10841 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10842 pv[count] = RExC_npar;
10843 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10846 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10847 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10850 SvIV_set(sv_dat, 1);
10853 /* Yes this does cause a memory leak in debugging Perls
10855 if (!av_store(RExC_paren_name_list,
10856 RExC_npar, SvREFCNT_inc(svname)))
10857 SvREFCNT_dec_NN(svname);
10860 /*sv_dump(sv_dat);*/
10862 nextchar(pRExC_state);
10864 goto capturing_parens;
10866 RExC_seen |= REG_LOOKBEHIND_SEEN;
10867 RExC_in_lookbehind++;
10869 if (RExC_parse >= RExC_end) {
10870 vFAIL("Sequence (?... not terminated");
10874 case '=': /* (?=...) */
10875 RExC_seen_zerolen++;
10877 case '!': /* (?!...) */
10878 RExC_seen_zerolen++;
10879 /* check if we're really just a "FAIL" assertion */
10880 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
10881 FALSE /* Don't force to /x */ );
10882 if (*RExC_parse == ')') {
10883 ret=reganode(pRExC_state, OPFAIL, 0);
10884 nextchar(pRExC_state);
10888 case '|': /* (?|...) */
10889 /* branch reset, behave like a (?:...) except that
10890 buffers in alternations share the same numbers */
10892 after_freeze = freeze_paren = RExC_npar;
10894 case ':': /* (?:...) */
10895 case '>': /* (?>...) */
10897 case '$': /* (?$...) */
10898 case '@': /* (?@...) */
10899 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10901 case '0' : /* (?0) */
10902 case 'R' : /* (?R) */
10903 if (RExC_parse == RExC_end || *RExC_parse != ')')
10904 FAIL("Sequence (?R) not terminated");
10906 RExC_seen |= REG_RECURSE_SEEN;
10907 *flagp |= POSTPONED;
10908 goto gen_recurse_regop;
10910 /* named and numeric backreferences */
10911 case '&': /* (?&NAME) */
10912 parse_start = RExC_parse - 1;
10915 SV *sv_dat = reg_scan_name(pRExC_state,
10916 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10917 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10919 if (RExC_parse >= RExC_end || *RExC_parse != ')')
10920 vFAIL("Sequence (?&... not terminated");
10921 goto gen_recurse_regop;
10924 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10926 vFAIL("Illegal pattern");
10928 goto parse_recursion;
10930 case '-': /* (?-1) */
10931 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10932 RExC_parse--; /* rewind to let it be handled later */
10936 case '1': case '2': case '3': case '4': /* (?1) */
10937 case '5': case '6': case '7': case '8': case '9':
10938 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
10941 bool is_neg = FALSE;
10943 parse_start = RExC_parse - 1; /* MJD */
10944 if (*RExC_parse == '-') {
10948 if (grok_atoUV(RExC_parse, &unum, &endptr)
10952 RExC_parse = (char*)endptr;
10956 /* Some limit for num? */
10960 if (*RExC_parse!=')')
10961 vFAIL("Expecting close bracket");
10964 if ( paren == '-' ) {
10966 Diagram of capture buffer numbering.
10967 Top line is the normal capture buffer numbers
10968 Bottom line is the negative indexing as from
10972 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
10976 num = RExC_npar + num;
10979 vFAIL("Reference to nonexistent group");
10981 } else if ( paren == '+' ) {
10982 num = RExC_npar + num - 1;
10984 /* We keep track how many GOSUB items we have produced.
10985 To start off the ARG2L() of the GOSUB holds its "id",
10986 which is used later in conjunction with RExC_recurse
10987 to calculate the offset we need to jump for the GOSUB,
10988 which it will store in the final representation.
10989 We have to defer the actual calculation until much later
10990 as the regop may move.
10993 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
10995 if (num > (I32)RExC_rx->nparens) {
10997 vFAIL("Reference to nonexistent group");
10999 RExC_recurse_count++;
11000 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11001 "%*s%*s Recurse #%"UVuf" to %"IVdf"\n",
11002 22, "| |", (int)(depth * 2 + 1), "",
11003 (UV)ARG(ret), (IV)ARG2L(ret)));
11005 RExC_seen |= REG_RECURSE_SEEN;
11007 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
11008 Set_Node_Offset(ret, parse_start); /* MJD */
11010 *flagp |= POSTPONED;
11011 assert(*RExC_parse == ')');
11012 nextchar(pRExC_state);
11017 case '?': /* (??...) */
11019 if (*RExC_parse != '{') {
11020 RExC_parse += SKIP_IF_CHAR(RExC_parse);
11021 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11023 "Sequence (%"UTF8f"...) not recognized",
11024 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11025 NOT_REACHED; /*NOTREACHED*/
11027 *flagp |= POSTPONED;
11031 case '{': /* (?{...}) */
11034 struct reg_code_block *cb;
11036 RExC_seen_zerolen++;
11038 if ( !pRExC_state->num_code_blocks
11039 || pRExC_state->code_index >= pRExC_state->num_code_blocks
11040 || pRExC_state->code_blocks[pRExC_state->code_index].start
11041 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11044 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11045 FAIL("panic: Sequence (?{...}): no code block found\n");
11046 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11048 /* this is a pre-compiled code block (?{...}) */
11049 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
11050 RExC_parse = RExC_start + cb->end;
11053 if (cb->src_regex) {
11054 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11055 RExC_rxi->data->data[n] =
11056 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11057 RExC_rxi->data->data[n+1] = (void*)o;
11060 n = add_data(pRExC_state,
11061 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11062 RExC_rxi->data->data[n] = (void*)o;
11065 pRExC_state->code_index++;
11066 nextchar(pRExC_state);
11070 ret = reg_node(pRExC_state, LOGICAL);
11072 eval = reg2Lanode(pRExC_state, EVAL,
11075 /* for later propagation into (??{})
11077 RExC_flags & RXf_PMf_COMPILETIME
11082 REGTAIL(pRExC_state, ret, eval);
11083 /* deal with the length of this later - MJD */
11086 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11087 Set_Node_Length(ret, RExC_parse - parse_start + 1);
11088 Set_Node_Offset(ret, parse_start);
11091 case '(': /* (?(?{...})...) and (?(?=...)...) */
11094 const int DEFINE_len = sizeof("DEFINE") - 1;
11095 if (RExC_parse[0] == '?') { /* (?(?...)) */
11096 if ( RExC_parse < RExC_end - 1
11097 && ( RExC_parse[1] == '='
11098 || RExC_parse[1] == '!'
11099 || RExC_parse[1] == '<'
11100 || RExC_parse[1] == '{')
11101 ) { /* Lookahead or eval. */
11105 ret = reg_node(pRExC_state, LOGICAL);
11109 tail = reg(pRExC_state, 1, &flag, depth+1);
11110 if (flag & (RESTART_PASS1|NEED_UTF8)) {
11111 *flagp = flag & (RESTART_PASS1|NEED_UTF8);
11114 REGTAIL(pRExC_state, ret, tail);
11117 /* Fall through to ‘Unknown switch condition’ at the
11118 end of the if/else chain. */
11120 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11121 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11123 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11124 char *name_start= RExC_parse++;
11126 SV *sv_dat=reg_scan_name(pRExC_state,
11127 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11128 if ( RExC_parse == name_start
11129 || RExC_parse >= RExC_end
11130 || *RExC_parse != ch)
11132 vFAIL2("Sequence (?(%c... not terminated",
11133 (ch == '>' ? '<' : ch));
11137 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11138 RExC_rxi->data->data[num]=(void*)sv_dat;
11139 SvREFCNT_inc_simple_void(sv_dat);
11141 ret = reganode(pRExC_state,NGROUPP,num);
11142 goto insert_if_check_paren;
11144 else if (RExC_end - RExC_parse >= DEFINE_len
11145 && strnEQ(RExC_parse, "DEFINE", DEFINE_len))
11147 ret = reganode(pRExC_state,DEFINEP,0);
11148 RExC_parse += DEFINE_len;
11150 goto insert_if_check_paren;
11152 else if (RExC_parse[0] == 'R') {
11154 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11155 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11156 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11159 if (RExC_parse[0] == '0') {
11163 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11165 if (grok_atoUV(RExC_parse, &uv, &endptr)
11168 parno = (I32)uv + 1;
11169 RExC_parse = (char*)endptr;
11171 /* else "Switch condition not recognized" below */
11172 } else if (RExC_parse[0] == '&') {
11175 sv_dat = reg_scan_name(pRExC_state,
11177 ? REG_RSN_RETURN_NULL
11178 : REG_RSN_RETURN_DATA);
11180 /* we should only have a false sv_dat when
11181 * SIZE_ONLY is true, and we always have false
11182 * sv_dat when SIZE_ONLY is true.
11183 * reg_scan_name() will VFAIL() if the name is
11184 * unknown when SIZE_ONLY is false, and otherwise
11185 * will return something, and when SIZE_ONLY is
11186 * true, reg_scan_name() just parses the string,
11187 * and doesnt return anything. (in theory) */
11188 assert(SIZE_ONLY ? !sv_dat : !!sv_dat);
11191 parno = 1 + *((I32 *)SvPVX(sv_dat));
11193 ret = reganode(pRExC_state,INSUBP,parno);
11194 goto insert_if_check_paren;
11196 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11200 if (grok_atoUV(RExC_parse, &uv, &endptr)
11204 RExC_parse = (char*)endptr;
11207 vFAIL("panic: grok_atoUV returned FALSE");
11209 ret = reganode(pRExC_state, GROUPP, parno);
11211 insert_if_check_paren:
11212 if (UCHARAT(RExC_parse) != ')') {
11213 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11214 vFAIL("Switch condition not recognized");
11216 nextchar(pRExC_state);
11218 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
11219 br = regbranch(pRExC_state, &flags, 1,depth+1);
11221 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11222 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11225 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
11228 REGTAIL(pRExC_state, br, reganode(pRExC_state,
11230 c = UCHARAT(RExC_parse);
11231 nextchar(pRExC_state);
11232 if (flags&HASWIDTH)
11233 *flagp |= HASWIDTH;
11236 vFAIL("(?(DEFINE)....) does not allow branches");
11238 /* Fake one for optimizer. */
11239 lastbr = reganode(pRExC_state, IFTHEN, 0);
11241 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
11242 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11243 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11246 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
11249 REGTAIL(pRExC_state, ret, lastbr);
11250 if (flags&HASWIDTH)
11251 *flagp |= HASWIDTH;
11252 c = UCHARAT(RExC_parse);
11253 nextchar(pRExC_state);
11258 if (RExC_parse >= RExC_end)
11259 vFAIL("Switch (?(condition)... not terminated");
11261 vFAIL("Switch (?(condition)... contains too many branches");
11263 ender = reg_node(pRExC_state, TAIL);
11264 REGTAIL(pRExC_state, br, ender);
11266 REGTAIL(pRExC_state, lastbr, ender);
11267 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11270 REGTAIL(pRExC_state, ret, ender);
11271 RExC_size++; /* XXX WHY do we need this?!!
11272 For large programs it seems to be required
11273 but I can't figure out why. -- dmq*/
11276 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11277 vFAIL("Unknown switch condition (?(...))");
11279 case '[': /* (?[ ... ]) */
11280 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
11282 case 0: /* A NUL */
11283 RExC_parse--; /* for vFAIL to print correctly */
11284 vFAIL("Sequence (? incomplete");
11286 default: /* e.g., (?i) */
11287 RExC_parse = (char *) seqstart + 1;
11289 parse_lparen_question_flags(pRExC_state);
11290 if (UCHARAT(RExC_parse) != ':') {
11291 if (RExC_parse < RExC_end)
11292 nextchar(pRExC_state);
11297 nextchar(pRExC_state);
11302 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11307 ret = reganode(pRExC_state, OPEN, parno);
11309 if (!RExC_nestroot)
11310 RExC_nestroot = parno;
11311 if (RExC_open_parens && !RExC_open_parens[parno])
11313 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11314 "%*s%*s Setting open paren #%"IVdf" to %d\n",
11315 22, "| |", (int)(depth * 2 + 1), "",
11316 (IV)parno, REG_NODE_NUM(ret)));
11317 RExC_open_parens[parno]= ret;
11320 Set_Node_Length(ret, 1); /* MJD */
11321 Set_Node_Offset(ret, RExC_parse); /* MJD */
11324 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
11333 /* Pick up the branches, linking them together. */
11334 parse_start = RExC_parse; /* MJD */
11335 br = regbranch(pRExC_state, &flags, 1,depth+1);
11337 /* branch_len = (paren != 0); */
11340 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11341 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11344 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
11346 if (*RExC_parse == '|') {
11347 if (!SIZE_ONLY && RExC_extralen) {
11348 reginsert(pRExC_state, BRANCHJ, br, depth+1);
11351 reginsert(pRExC_state, BRANCH, br, depth+1);
11352 Set_Node_Length(br, paren != 0);
11353 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
11357 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
11359 else if (paren == ':') {
11360 *flagp |= flags&SIMPLE;
11362 if (is_open) { /* Starts with OPEN. */
11363 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
11365 else if (paren != '?') /* Not Conditional */
11367 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11369 while (*RExC_parse == '|') {
11370 if (!SIZE_ONLY && RExC_extralen) {
11371 ender = reganode(pRExC_state, LONGJMP,0);
11373 /* Append to the previous. */
11374 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11377 RExC_extralen += 2; /* Account for LONGJMP. */
11378 nextchar(pRExC_state);
11379 if (freeze_paren) {
11380 if (RExC_npar > after_freeze)
11381 after_freeze = RExC_npar;
11382 RExC_npar = freeze_paren;
11384 br = regbranch(pRExC_state, &flags, 0, depth+1);
11387 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11388 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11391 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
11393 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
11395 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11398 if (have_branch || paren != ':') {
11399 /* Make a closing node, and hook it on the end. */
11402 ender = reg_node(pRExC_state, TAIL);
11405 ender = reganode(pRExC_state, CLOSE, parno);
11406 if ( RExC_close_parens ) {
11407 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11408 "%*s%*s Setting close paren #%"IVdf" to %d\n",
11409 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
11410 RExC_close_parens[parno]= ender;
11411 if (RExC_nestroot == parno)
11414 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
11415 Set_Node_Length(ender,1); /* MJD */
11421 *flagp &= ~HASWIDTH;
11424 ender = reg_node(pRExC_state, SUCCEED);
11427 ender = reg_node(pRExC_state, END);
11429 assert(!RExC_end_op); /* there can only be one! */
11430 RExC_end_op = ender;
11431 if (RExC_close_parens) {
11432 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11433 "%*s%*s Setting close paren #0 (END) to %d\n",
11434 22, "| |", (int)(depth * 2 + 1), "", REG_NODE_NUM(ender)));
11436 RExC_close_parens[0]= ender;
11441 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11442 DEBUG_PARSE_MSG("lsbr");
11443 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
11444 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11445 Perl_re_printf( aTHX_ "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
11446 SvPV_nolen_const(RExC_mysv1),
11447 (IV)REG_NODE_NUM(lastbr),
11448 SvPV_nolen_const(RExC_mysv2),
11449 (IV)REG_NODE_NUM(ender),
11450 (IV)(ender - lastbr)
11453 REGTAIL(pRExC_state, lastbr, ender);
11455 if (have_branch && !SIZE_ONLY) {
11456 char is_nothing= 1;
11458 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
11460 /* Hook the tails of the branches to the closing node. */
11461 for (br = ret; br; br = regnext(br)) {
11462 const U8 op = PL_regkind[OP(br)];
11463 if (op == BRANCH) {
11464 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
11465 if ( OP(NEXTOPER(br)) != NOTHING
11466 || regnext(NEXTOPER(br)) != ender)
11469 else if (op == BRANCHJ) {
11470 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
11471 /* for now we always disable this optimisation * /
11472 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
11473 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
11479 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
11480 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11481 DEBUG_PARSE_MSG("NADA");
11482 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
11483 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11484 Perl_re_printf( aTHX_ "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
11485 SvPV_nolen_const(RExC_mysv1),
11486 (IV)REG_NODE_NUM(ret),
11487 SvPV_nolen_const(RExC_mysv2),
11488 (IV)REG_NODE_NUM(ender),
11493 if (OP(ender) == TAIL) {
11498 for ( opt= br + 1; opt < ender ; opt++ )
11499 OP(opt)= OPTIMIZED;
11500 NEXT_OFF(br)= ender - br;
11508 static const char parens[] = "=!<,>";
11510 if (paren && (p = strchr(parens, paren))) {
11511 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
11512 int flag = (p - parens) > 1;
11515 node = SUSPEND, flag = 0;
11516 reginsert(pRExC_state, node,ret, depth+1);
11517 Set_Node_Cur_Length(ret, parse_start);
11518 Set_Node_Offset(ret, parse_start + 1);
11520 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
11524 /* Check for proper termination. */
11526 /* restore original flags, but keep (?p) and, if we've changed from /d
11527 * rules to /u, keep the /u */
11528 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
11529 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
11530 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
11532 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
11533 RExC_parse = oregcomp_parse;
11534 vFAIL("Unmatched (");
11536 nextchar(pRExC_state);
11538 else if (!paren && RExC_parse < RExC_end) {
11539 if (*RExC_parse == ')') {
11541 vFAIL("Unmatched )");
11544 FAIL("Junk on end of regexp"); /* "Can't happen". */
11545 NOT_REACHED; /* NOTREACHED */
11548 if (RExC_in_lookbehind) {
11549 RExC_in_lookbehind--;
11551 if (after_freeze > RExC_npar)
11552 RExC_npar = after_freeze;
11557 - regbranch - one alternative of an | operator
11559 * Implements the concatenation operator.
11561 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11562 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11565 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
11568 regnode *chain = NULL;
11570 I32 flags = 0, c = 0;
11571 GET_RE_DEBUG_FLAGS_DECL;
11573 PERL_ARGS_ASSERT_REGBRANCH;
11575 DEBUG_PARSE("brnc");
11580 if (!SIZE_ONLY && RExC_extralen)
11581 ret = reganode(pRExC_state, BRANCHJ,0);
11583 ret = reg_node(pRExC_state, BRANCH);
11584 Set_Node_Length(ret, 1);
11588 if (!first && SIZE_ONLY)
11589 RExC_extralen += 1; /* BRANCHJ */
11591 *flagp = WORST; /* Tentatively. */
11593 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11594 FALSE /* Don't force to /x */ );
11595 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
11596 flags &= ~TRYAGAIN;
11597 latest = regpiece(pRExC_state, &flags,depth+1);
11598 if (latest == NULL) {
11599 if (flags & TRYAGAIN)
11601 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11602 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11605 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
11607 else if (ret == NULL)
11609 *flagp |= flags&(HASWIDTH|POSTPONED);
11610 if (chain == NULL) /* First piece. */
11611 *flagp |= flags&SPSTART;
11613 /* FIXME adding one for every branch after the first is probably
11614 * excessive now we have TRIE support. (hv) */
11616 REGTAIL(pRExC_state, chain, latest);
11621 if (chain == NULL) { /* Loop ran zero times. */
11622 chain = reg_node(pRExC_state, NOTHING);
11627 *flagp |= flags&SIMPLE;
11634 - regpiece - something followed by possible [*+?]
11636 * Note that the branching code sequences used for ? and the general cases
11637 * of * and + are somewhat optimized: they use the same NOTHING node as
11638 * both the endmarker for their branch list and the body of the last branch.
11639 * It might seem that this node could be dispensed with entirely, but the
11640 * endmarker role is not redundant.
11642 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
11644 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11645 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11648 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11654 const char * const origparse = RExC_parse;
11656 I32 max = REG_INFTY;
11657 #ifdef RE_TRACK_PATTERN_OFFSETS
11660 const char *maxpos = NULL;
11663 /* Save the original in case we change the emitted regop to a FAIL. */
11664 regnode * const orig_emit = RExC_emit;
11666 GET_RE_DEBUG_FLAGS_DECL;
11668 PERL_ARGS_ASSERT_REGPIECE;
11670 DEBUG_PARSE("piec");
11672 ret = regatom(pRExC_state, &flags,depth+1);
11674 if (flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8))
11675 *flagp |= flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8);
11677 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
11683 if (op == '{' && regcurly(RExC_parse)) {
11685 #ifdef RE_TRACK_PATTERN_OFFSETS
11686 parse_start = RExC_parse; /* MJD */
11688 next = RExC_parse + 1;
11689 while (isDIGIT(*next) || *next == ',') {
11690 if (*next == ',') {
11698 if (*next == '}') { /* got one */
11699 const char* endptr;
11703 if (isDIGIT(*RExC_parse)) {
11704 if (!grok_atoUV(RExC_parse, &uv, &endptr))
11705 vFAIL("Invalid quantifier in {,}");
11706 if (uv >= REG_INFTY)
11707 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11712 if (*maxpos == ',')
11715 maxpos = RExC_parse;
11716 if (isDIGIT(*maxpos)) {
11717 if (!grok_atoUV(maxpos, &uv, &endptr))
11718 vFAIL("Invalid quantifier in {,}");
11719 if (uv >= REG_INFTY)
11720 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11723 max = REG_INFTY; /* meaning "infinity" */
11726 nextchar(pRExC_state);
11727 if (max < min) { /* If can't match, warn and optimize to fail
11731 /* We can't back off the size because we have to reserve
11732 * enough space for all the things we are about to throw
11733 * away, but we can shrink it by the amount we are about
11734 * to re-use here */
11735 RExC_size += PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
11738 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
11739 RExC_emit = orig_emit;
11741 ret = reganode(pRExC_state, OPFAIL, 0);
11744 else if (min == max && *RExC_parse == '?')
11747 ckWARN2reg(RExC_parse + 1,
11748 "Useless use of greediness modifier '%c'",
11754 if ((flags&SIMPLE)) {
11755 if (min == 0 && max == REG_INFTY) {
11756 reginsert(pRExC_state, STAR, ret, depth+1);
11759 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11762 if (min == 1 && max == REG_INFTY) {
11763 reginsert(pRExC_state, PLUS, ret, depth+1);
11766 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11769 MARK_NAUGHTY_EXP(2, 2);
11770 reginsert(pRExC_state, CURLY, ret, depth+1);
11771 Set_Node_Offset(ret, parse_start+1); /* MJD */
11772 Set_Node_Cur_Length(ret, parse_start);
11775 regnode * const w = reg_node(pRExC_state, WHILEM);
11778 REGTAIL(pRExC_state, ret, w);
11779 if (!SIZE_ONLY && RExC_extralen) {
11780 reginsert(pRExC_state, LONGJMP,ret, depth+1);
11781 reginsert(pRExC_state, NOTHING,ret, depth+1);
11782 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
11784 reginsert(pRExC_state, CURLYX,ret, depth+1);
11786 Set_Node_Offset(ret, parse_start+1);
11787 Set_Node_Length(ret,
11788 op == '{' ? (RExC_parse - parse_start) : 1);
11790 if (!SIZE_ONLY && RExC_extralen)
11791 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
11792 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
11794 RExC_whilem_seen++, RExC_extralen += 3;
11795 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
11802 *flagp |= HASWIDTH;
11804 ARG1_SET(ret, (U16)min);
11805 ARG2_SET(ret, (U16)max);
11807 if (max == REG_INFTY)
11808 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11814 if (!ISMULT1(op)) {
11819 #if 0 /* Now runtime fix should be reliable. */
11821 /* if this is reinstated, don't forget to put this back into perldiag:
11823 =item Regexp *+ operand could be empty at {#} in regex m/%s/
11825 (F) The part of the regexp subject to either the * or + quantifier
11826 could match an empty string. The {#} shows in the regular
11827 expression about where the problem was discovered.
11831 if (!(flags&HASWIDTH) && op != '?')
11832 vFAIL("Regexp *+ operand could be empty");
11835 #ifdef RE_TRACK_PATTERN_OFFSETS
11836 parse_start = RExC_parse;
11838 nextchar(pRExC_state);
11840 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
11846 else if (op == '+') {
11850 else if (op == '?') {
11855 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
11856 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
11857 ckWARN2reg(RExC_parse,
11858 "%"UTF8f" matches null string many times",
11859 UTF8fARG(UTF, (RExC_parse >= origparse
11860 ? RExC_parse - origparse
11863 (void)ReREFCNT_inc(RExC_rx_sv);
11866 if (*RExC_parse == '?') {
11867 nextchar(pRExC_state);
11868 reginsert(pRExC_state, MINMOD, ret, depth+1);
11869 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
11871 else if (*RExC_parse == '+') {
11873 nextchar(pRExC_state);
11874 ender = reg_node(pRExC_state, SUCCEED);
11875 REGTAIL(pRExC_state, ret, ender);
11876 reginsert(pRExC_state, SUSPEND, ret, depth+1);
11878 ender = reg_node(pRExC_state, TAIL);
11879 REGTAIL(pRExC_state, ret, ender);
11882 if (ISMULT2(RExC_parse)) {
11884 vFAIL("Nested quantifiers");
11891 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
11900 /* This routine teases apart the various meanings of \N and returns
11901 * accordingly. The input parameters constrain which meaning(s) is/are valid
11902 * in the current context.
11904 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
11906 * If <code_point_p> is not NULL, the context is expecting the result to be a
11907 * single code point. If this \N instance turns out to a single code point,
11908 * the function returns TRUE and sets *code_point_p to that code point.
11910 * If <node_p> is not NULL, the context is expecting the result to be one of
11911 * the things representable by a regnode. If this \N instance turns out to be
11912 * one such, the function generates the regnode, returns TRUE and sets *node_p
11913 * to point to that regnode.
11915 * If this instance of \N isn't legal in any context, this function will
11916 * generate a fatal error and not return.
11918 * On input, RExC_parse should point to the first char following the \N at the
11919 * time of the call. On successful return, RExC_parse will have been updated
11920 * to point to just after the sequence identified by this routine. Also
11921 * *flagp has been updated as needed.
11923 * When there is some problem with the current context and this \N instance,
11924 * the function returns FALSE, without advancing RExC_parse, nor setting
11925 * *node_p, nor *code_point_p, nor *flagp.
11927 * If <cp_count> is not NULL, the caller wants to know the length (in code
11928 * points) that this \N sequence matches. This is set even if the function
11929 * returns FALSE, as detailed below.
11931 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
11933 * Probably the most common case is for the \N to specify a single code point.
11934 * *cp_count will be set to 1, and *code_point_p will be set to that code
11937 * Another possibility is for the input to be an empty \N{}, which for
11938 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
11939 * will be set to a generated NOTHING node.
11941 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
11942 * set to 0. *node_p will be set to a generated REG_ANY node.
11944 * The fourth possibility is that \N resolves to a sequence of more than one
11945 * code points. *cp_count will be set to the number of code points in the
11946 * sequence. *node_p * will be set to a generated node returned by this
11947 * function calling S_reg().
11949 * The final possibility is that it is premature to be calling this function;
11950 * that pass1 needs to be restarted. This can happen when this changes from
11951 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
11952 * latter occurs only when the fourth possibility would otherwise be in
11953 * effect, and is because one of those code points requires the pattern to be
11954 * recompiled as UTF-8. The function returns FALSE, and sets the
11955 * RESTART_PASS1 and NEED_UTF8 flags in *flagp, as appropriate. When this
11956 * happens, the caller needs to desist from continuing parsing, and return
11957 * this information to its caller. This is not set for when there is only one
11958 * code point, as this can be called as part of an ANYOF node, and they can
11959 * store above-Latin1 code points without the pattern having to be in UTF-8.
11961 * For non-single-quoted regexes, the tokenizer has resolved character and
11962 * sequence names inside \N{...} into their Unicode values, normalizing the
11963 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
11964 * hex-represented code points in the sequence. This is done there because
11965 * the names can vary based on what charnames pragma is in scope at the time,
11966 * so we need a way to take a snapshot of what they resolve to at the time of
11967 * the original parse. [perl #56444].
11969 * That parsing is skipped for single-quoted regexes, so we may here get
11970 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
11971 * parser. But if the single-quoted regex is something like '\N{U+41}', that
11972 * is legal and handled here. The code point is Unicode, and has to be
11973 * translated into the native character set for non-ASCII platforms.
11976 char * endbrace; /* points to '}' following the name */
11977 char *endchar; /* Points to '.' or '}' ending cur char in the input
11979 char* p = RExC_parse; /* Temporary */
11981 GET_RE_DEBUG_FLAGS_DECL;
11983 PERL_ARGS_ASSERT_GROK_BSLASH_N;
11985 GET_RE_DEBUG_FLAGS;
11987 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
11988 assert(! (node_p && cp_count)); /* At most 1 should be set */
11990 if (cp_count) { /* Initialize return for the most common case */
11994 /* The [^\n] meaning of \N ignores spaces and comments under the /x
11995 * modifier. The other meanings do not, so use a temporary until we find
11996 * out which we are being called with */
11997 skip_to_be_ignored_text(pRExC_state, &p,
11998 FALSE /* Don't force to /x */ );
12000 /* Disambiguate between \N meaning a named character versus \N meaning
12001 * [^\n]. The latter is assumed when the {...} following the \N is a legal
12002 * quantifier, or there is no '{' at all */
12003 if (*p != '{' || regcurly(p)) {
12013 *node_p = reg_node(pRExC_state, REG_ANY);
12014 *flagp |= HASWIDTH|SIMPLE;
12016 Set_Node_Length(*node_p, 1); /* MJD */
12020 /* Here, we have decided it should be a named character or sequence */
12022 /* The test above made sure that the next real character is a '{', but
12023 * under the /x modifier, it could be separated by space (or a comment and
12024 * \n) and this is not allowed (for consistency with \x{...} and the
12025 * tokenizer handling of \N{NAME}). */
12026 if (*RExC_parse != '{') {
12027 vFAIL("Missing braces on \\N{}");
12030 RExC_parse++; /* Skip past the '{' */
12032 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
12033 || ! (endbrace == RExC_parse /* nothing between the {} */
12034 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked... */
12035 && strnEQ(RExC_parse, "U+", 2)))) /* ... below for a better
12038 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
12039 vFAIL("\\N{NAME} must be resolved by the lexer");
12042 REQUIRE_UNI_RULES(flagp, FALSE); /* Unicode named chars imply Unicode
12045 if (endbrace == RExC_parse) { /* empty: \N{} */
12047 RExC_parse++; /* Position after the "}" */
12048 vFAIL("Zero length \\N{}");
12053 nextchar(pRExC_state);
12058 *node_p = reg_node(pRExC_state,NOTHING);
12062 RExC_parse += 2; /* Skip past the 'U+' */
12064 /* Because toke.c has generated a special construct for us guaranteed not
12065 * to have NULs, we can use a str function */
12066 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12068 /* Code points are separated by dots. If none, there is only one code
12069 * point, and is terminated by the brace */
12071 if (endchar >= endbrace) {
12072 STRLEN length_of_hex;
12073 I32 grok_hex_flags;
12075 /* Here, exactly one code point. If that isn't what is wanted, fail */
12076 if (! code_point_p) {
12081 /* Convert code point from hex */
12082 length_of_hex = (STRLEN)(endchar - RExC_parse);
12083 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
12084 | PERL_SCAN_DISALLOW_PREFIX
12086 /* No errors in the first pass (See [perl
12087 * #122671].) We let the code below find the
12088 * errors when there are multiple chars. */
12090 ? PERL_SCAN_SILENT_ILLDIGIT
12093 /* This routine is the one place where both single- and double-quotish
12094 * \N{U+xxxx} are evaluated. The value is a Unicode code point which
12095 * must be converted to native. */
12096 *code_point_p = UNI_TO_NATIVE(grok_hex(RExC_parse,
12101 /* The tokenizer should have guaranteed validity, but it's possible to
12102 * bypass it by using single quoting, so check. Don't do the check
12103 * here when there are multiple chars; we do it below anyway. */
12104 if (length_of_hex == 0
12105 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
12107 RExC_parse += length_of_hex; /* Includes all the valid */
12108 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12109 ? UTF8SKIP(RExC_parse)
12111 /* Guard against malformed utf8 */
12112 if (RExC_parse >= endchar) {
12113 RExC_parse = endchar;
12115 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12118 RExC_parse = endbrace + 1;
12121 else { /* Is a multiple character sequence */
12122 SV * substitute_parse;
12124 char *orig_end = RExC_end;
12125 char *save_start = RExC_start;
12128 /* Count the code points, if desired, in the sequence */
12131 while (RExC_parse < endbrace) {
12132 /* Point to the beginning of the next character in the sequence. */
12133 RExC_parse = endchar + 1;
12134 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12139 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12140 * But don't backup up the pointer if the caller want to know how many
12141 * code points there are (they can then handle things) */
12149 /* What is done here is to convert this to a sub-pattern of the form
12150 * \x{char1}\x{char2}... and then call reg recursively to parse it
12151 * (enclosing in "(?: ... )" ). That way, it retains its atomicness,
12152 * while not having to worry about special handling that some code
12153 * points may have. */
12155 substitute_parse = newSVpvs("?:");
12157 while (RExC_parse < endbrace) {
12159 /* Convert to notation the rest of the code understands */
12160 sv_catpv(substitute_parse, "\\x{");
12161 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
12162 sv_catpv(substitute_parse, "}");
12164 /* Point to the beginning of the next character in the sequence. */
12165 RExC_parse = endchar + 1;
12166 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12169 sv_catpv(substitute_parse, ")");
12171 RExC_parse = RExC_start = RExC_adjusted_start = SvPV(substitute_parse,
12174 /* Don't allow empty number */
12175 if (len < (STRLEN) 8) {
12176 RExC_parse = endbrace;
12177 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12179 RExC_end = RExC_parse + len;
12181 /* The values are Unicode, and therefore not subject to recoding, but
12182 * have to be converted to native on a non-Unicode (meaning non-ASCII)
12184 RExC_override_recoding = 1;
12186 RExC_recode_x_to_native = 1;
12190 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
12191 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12192 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12195 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
12198 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12201 /* Restore the saved values */
12202 RExC_start = RExC_adjusted_start = save_start;
12203 RExC_parse = endbrace;
12204 RExC_end = orig_end;
12205 RExC_override_recoding = 0;
12207 RExC_recode_x_to_native = 0;
12210 SvREFCNT_dec_NN(substitute_parse);
12211 nextchar(pRExC_state);
12218 PERL_STATIC_INLINE U8
12219 S_compute_EXACTish(RExC_state_t *pRExC_state)
12223 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
12231 op = get_regex_charset(RExC_flags);
12232 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
12233 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
12234 been, so there is no hole */
12237 return op + EXACTF;
12240 PERL_STATIC_INLINE void
12241 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
12242 regnode *node, I32* flagp, STRLEN len, UV code_point,
12245 /* This knows the details about sizing an EXACTish node, setting flags for
12246 * it (by setting <*flagp>, and potentially populating it with a single
12249 * If <len> (the length in bytes) is non-zero, this function assumes that
12250 * the node has already been populated, and just does the sizing. In this
12251 * case <code_point> should be the final code point that has already been
12252 * placed into the node. This value will be ignored except that under some
12253 * circumstances <*flagp> is set based on it.
12255 * If <len> is zero, the function assumes that the node is to contain only
12256 * the single character given by <code_point> and calculates what <len>
12257 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
12258 * additionally will populate the node's STRING with <code_point> or its
12261 * In both cases <*flagp> is appropriately set
12263 * It knows that under FOLD, the Latin Sharp S and UTF characters above
12264 * 255, must be folded (the former only when the rules indicate it can
12267 * When it does the populating, it looks at the flag 'downgradable'. If
12268 * true with a node that folds, it checks if the single code point
12269 * participates in a fold, and if not downgrades the node to an EXACT.
12270 * This helps the optimizer */
12272 bool len_passed_in = cBOOL(len != 0);
12273 U8 character[UTF8_MAXBYTES_CASE+1];
12275 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
12277 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
12278 * sizing difference, and is extra work that is thrown away */
12279 if (downgradable && ! PASS2) {
12280 downgradable = FALSE;
12283 if (! len_passed_in) {
12285 if (UVCHR_IS_INVARIANT(code_point)) {
12286 if (LOC || ! FOLD) { /* /l defers folding until runtime */
12287 *character = (U8) code_point;
12289 else { /* Here is /i and not /l. (toFOLD() is defined on just
12290 ASCII, which isn't the same thing as INVARIANT on
12291 EBCDIC, but it works there, as the extra invariants
12292 fold to themselves) */
12293 *character = toFOLD((U8) code_point);
12295 /* We can downgrade to an EXACT node if this character
12296 * isn't a folding one. Note that this assumes that
12297 * nothing above Latin1 folds to some other invariant than
12298 * one of these alphabetics; otherwise we would also have
12300 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12301 * || ASCII_FOLD_RESTRICTED))
12303 if (downgradable && PL_fold[code_point] == code_point) {
12309 else if (FOLD && (! LOC
12310 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
12311 { /* Folding, and ok to do so now */
12312 UV folded = _to_uni_fold_flags(
12316 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12317 ? FOLD_FLAGS_NOMIX_ASCII
12320 && folded == code_point /* This quickly rules out many
12321 cases, avoiding the
12322 _invlist_contains_cp() overhead
12324 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
12331 else if (code_point <= MAX_UTF8_TWO_BYTE) {
12333 /* Not folding this cp, and can output it directly */
12334 *character = UTF8_TWO_BYTE_HI(code_point);
12335 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
12339 uvchr_to_utf8( character, code_point);
12340 len = UTF8SKIP(character);
12342 } /* Else pattern isn't UTF8. */
12344 *character = (U8) code_point;
12346 } /* Else is folded non-UTF8 */
12347 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12348 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12349 || UNICODE_DOT_DOT_VERSION > 0)
12350 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
12354 /* We don't fold any non-UTF8 except possibly the Sharp s (see
12355 * comments at join_exact()); */
12356 *character = (U8) code_point;
12359 /* Can turn into an EXACT node if we know the fold at compile time,
12360 * and it folds to itself and doesn't particpate in other folds */
12363 && PL_fold_latin1[code_point] == code_point
12364 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12365 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
12369 } /* else is Sharp s. May need to fold it */
12370 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
12372 *(character + 1) = 's';
12376 *character = LATIN_SMALL_LETTER_SHARP_S;
12382 RExC_size += STR_SZ(len);
12385 RExC_emit += STR_SZ(len);
12386 STR_LEN(node) = len;
12387 if (! len_passed_in) {
12388 Copy((char *) character, STRING(node), len, char);
12392 *flagp |= HASWIDTH;
12394 /* A single character node is SIMPLE, except for the special-cased SHARP S
12396 if ((len == 1 || (UTF && len == UVCHR_SKIP(code_point)))
12397 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12398 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12399 || UNICODE_DOT_DOT_VERSION > 0)
12400 && ( code_point != LATIN_SMALL_LETTER_SHARP_S
12401 || ! FOLD || ! DEPENDS_SEMANTICS)
12407 /* The OP may not be well defined in PASS1 */
12408 if (PASS2 && OP(node) == EXACTFL) {
12409 RExC_contains_locale = 1;
12414 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
12415 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
12418 S_backref_value(char *p)
12420 const char* endptr;
12422 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
12429 - regatom - the lowest level
12431 Try to identify anything special at the start of the current parse position.
12432 If there is, then handle it as required. This may involve generating a
12433 single regop, such as for an assertion; or it may involve recursing, such as
12434 to handle a () structure.
12436 If the string doesn't start with something special then we gobble up
12437 as much literal text as we can. If we encounter a quantifier, we have to
12438 back off the final literal character, as that quantifier applies to just it
12439 and not to the whole string of literals.
12441 Once we have been able to handle whatever type of thing started the
12442 sequence, we return.
12444 Note: we have to be careful with escapes, as they can be both literal
12445 and special, and in the case of \10 and friends, context determines which.
12447 A summary of the code structure is:
12449 switch (first_byte) {
12450 cases for each special:
12451 handle this special;
12454 switch (2nd byte) {
12455 cases for each unambiguous special:
12456 handle this special;
12458 cases for each ambigous special/literal:
12460 if (special) handle here
12462 default: // unambiguously literal:
12465 default: // is a literal char
12468 create EXACTish node for literal;
12469 while (more input and node isn't full) {
12470 switch (input_byte) {
12471 cases for each special;
12472 make sure parse pointer is set so that the next call to
12473 regatom will see this special first
12474 goto loopdone; // EXACTish node terminated by prev. char
12476 append char to EXACTISH node;
12478 get next input byte;
12482 return the generated node;
12484 Specifically there are two separate switches for handling
12485 escape sequences, with the one for handling literal escapes requiring
12486 a dummy entry for all of the special escapes that are actually handled
12489 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
12491 Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
12492 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
12493 Otherwise does not return NULL.
12497 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12499 regnode *ret = NULL;
12506 GET_RE_DEBUG_FLAGS_DECL;
12508 *flagp = WORST; /* Tentatively. */
12510 DEBUG_PARSE("atom");
12512 PERL_ARGS_ASSERT_REGATOM;
12515 parse_start = RExC_parse;
12516 assert(RExC_parse < RExC_end);
12517 switch ((U8)*RExC_parse) {
12519 RExC_seen_zerolen++;
12520 nextchar(pRExC_state);
12521 if (RExC_flags & RXf_PMf_MULTILINE)
12522 ret = reg_node(pRExC_state, MBOL);
12524 ret = reg_node(pRExC_state, SBOL);
12525 Set_Node_Length(ret, 1); /* MJD */
12528 nextchar(pRExC_state);
12530 RExC_seen_zerolen++;
12531 if (RExC_flags & RXf_PMf_MULTILINE)
12532 ret = reg_node(pRExC_state, MEOL);
12534 ret = reg_node(pRExC_state, SEOL);
12535 Set_Node_Length(ret, 1); /* MJD */
12538 nextchar(pRExC_state);
12539 if (RExC_flags & RXf_PMf_SINGLELINE)
12540 ret = reg_node(pRExC_state, SANY);
12542 ret = reg_node(pRExC_state, REG_ANY);
12543 *flagp |= HASWIDTH|SIMPLE;
12545 Set_Node_Length(ret, 1); /* MJD */
12549 char * const oregcomp_parse = ++RExC_parse;
12550 ret = regclass(pRExC_state, flagp,depth+1,
12551 FALSE, /* means parse the whole char class */
12552 TRUE, /* allow multi-char folds */
12553 FALSE, /* don't silence non-portable warnings. */
12554 (bool) RExC_strict,
12555 TRUE, /* Allow an optimized regnode result */
12559 if (*flagp & (RESTART_PASS1|NEED_UTF8))
12561 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
12564 if (*RExC_parse != ']') {
12565 RExC_parse = oregcomp_parse;
12566 vFAIL("Unmatched [");
12568 nextchar(pRExC_state);
12569 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
12573 nextchar(pRExC_state);
12574 ret = reg(pRExC_state, 2, &flags,depth+1);
12576 if (flags & TRYAGAIN) {
12577 if (RExC_parse >= RExC_end) {
12578 /* Make parent create an empty node if needed. */
12579 *flagp |= TRYAGAIN;
12584 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12585 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12588 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"",
12591 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12595 if (flags & TRYAGAIN) {
12596 *flagp |= TRYAGAIN;
12599 vFAIL("Internal urp");
12600 /* Supposed to be caught earlier. */
12606 vFAIL("Quantifier follows nothing");
12611 This switch handles escape sequences that resolve to some kind
12612 of special regop and not to literal text. Escape sequnces that
12613 resolve to literal text are handled below in the switch marked
12616 Every entry in this switch *must* have a corresponding entry
12617 in the literal escape switch. However, the opposite is not
12618 required, as the default for this switch is to jump to the
12619 literal text handling code.
12622 switch ((U8)*RExC_parse) {
12623 /* Special Escapes */
12625 RExC_seen_zerolen++;
12626 ret = reg_node(pRExC_state, SBOL);
12627 /* SBOL is shared with /^/ so we set the flags so we can tell
12628 * /\A/ from /^/ in split. We check ret because first pass we
12629 * have no regop struct to set the flags on. */
12633 goto finish_meta_pat;
12635 ret = reg_node(pRExC_state, GPOS);
12636 RExC_seen |= REG_GPOS_SEEN;
12638 goto finish_meta_pat;
12640 RExC_seen_zerolen++;
12641 ret = reg_node(pRExC_state, KEEPS);
12643 /* XXX:dmq : disabling in-place substitution seems to
12644 * be necessary here to avoid cases of memory corruption, as
12645 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
12647 RExC_seen |= REG_LOOKBEHIND_SEEN;
12648 goto finish_meta_pat;
12650 ret = reg_node(pRExC_state, SEOL);
12652 RExC_seen_zerolen++; /* Do not optimize RE away */
12653 goto finish_meta_pat;
12655 ret = reg_node(pRExC_state, EOS);
12657 RExC_seen_zerolen++; /* Do not optimize RE away */
12658 goto finish_meta_pat;
12660 vFAIL("\\C no longer supported");
12662 ret = reg_node(pRExC_state, CLUMP);
12663 *flagp |= HASWIDTH;
12664 goto finish_meta_pat;
12670 arg = ANYOF_WORDCHAR;
12678 regex_charset charset = get_regex_charset(RExC_flags);
12680 RExC_seen_zerolen++;
12681 RExC_seen |= REG_LOOKBEHIND_SEEN;
12682 op = BOUND + charset;
12684 if (op == BOUNDL) {
12685 RExC_contains_locale = 1;
12688 ret = reg_node(pRExC_state, op);
12690 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
12691 FLAGS(ret) = TRADITIONAL_BOUND;
12692 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
12698 char name = *RExC_parse;
12701 endbrace = strchr(RExC_parse, '}');
12704 vFAIL2("Missing right brace on \\%c{}", name);
12706 /* XXX Need to decide whether to take spaces or not. Should be
12707 * consistent with \p{}, but that currently is SPACE, which
12708 * means vertical too, which seems wrong
12709 * while (isBLANK(*RExC_parse)) {
12712 if (endbrace == RExC_parse) {
12713 RExC_parse++; /* After the '}' */
12714 vFAIL2("Empty \\%c{}", name);
12716 length = endbrace - RExC_parse;
12717 /*while (isBLANK(*(RExC_parse + length - 1))) {
12720 switch (*RExC_parse) {
12723 && (length != 3 || strnNE(RExC_parse + 1, "cb", 2)))
12725 goto bad_bound_type;
12727 FLAGS(ret) = GCB_BOUND;
12730 if (length != 2 || *(RExC_parse + 1) != 'b') {
12731 goto bad_bound_type;
12733 FLAGS(ret) = LB_BOUND;
12736 if (length != 2 || *(RExC_parse + 1) != 'b') {
12737 goto bad_bound_type;
12739 FLAGS(ret) = SB_BOUND;
12742 if (length != 2 || *(RExC_parse + 1) != 'b') {
12743 goto bad_bound_type;
12745 FLAGS(ret) = WB_BOUND;
12749 RExC_parse = endbrace;
12751 "'%"UTF8f"' is an unknown bound type",
12752 UTF8fARG(UTF, length, endbrace - length));
12753 NOT_REACHED; /*NOTREACHED*/
12755 RExC_parse = endbrace;
12756 REQUIRE_UNI_RULES(flagp, NULL);
12758 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
12762 /* Don't have to worry about UTF-8, in this message because
12763 * to get here the contents of the \b must be ASCII */
12764 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
12765 "Using /u for '%.*s' instead of /%s",
12767 endbrace - length + 1,
12768 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
12769 ? ASCII_RESTRICT_PAT_MODS
12770 : ASCII_MORE_RESTRICT_PAT_MODS);
12774 if (PASS2 && invert) {
12775 OP(ret) += NBOUND - BOUND;
12777 goto finish_meta_pat;
12785 if (! DEPENDS_SEMANTICS) {
12789 /* \d doesn't have any matches in the upper Latin1 range, hence /d
12790 * is equivalent to /u. Changing to /u saves some branches at
12793 goto join_posix_op_known;
12796 ret = reg_node(pRExC_state, LNBREAK);
12797 *flagp |= HASWIDTH|SIMPLE;
12798 goto finish_meta_pat;
12806 goto join_posix_op_known;
12812 arg = ANYOF_VERTWS;
12814 goto join_posix_op_known;
12824 op = POSIXD + get_regex_charset(RExC_flags);
12825 if (op > POSIXA) { /* /aa is same as /a */
12828 else if (op == POSIXL) {
12829 RExC_contains_locale = 1;
12832 join_posix_op_known:
12835 op += NPOSIXD - POSIXD;
12838 ret = reg_node(pRExC_state, op);
12840 FLAGS(ret) = namedclass_to_classnum(arg);
12843 *flagp |= HASWIDTH|SIMPLE;
12847 nextchar(pRExC_state);
12848 Set_Node_Length(ret, 2); /* MJD */
12854 ret = regclass(pRExC_state, flagp,depth+1,
12855 TRUE, /* means just parse this element */
12856 FALSE, /* don't allow multi-char folds */
12857 FALSE, /* don't silence non-portable warnings. It
12858 would be a bug if these returned
12860 (bool) RExC_strict,
12861 TRUE, /* Allow an optimized regnode result */
12864 if (*flagp & RESTART_PASS1)
12866 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
12867 * multi-char folds are allowed. */
12869 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
12874 Set_Node_Offset(ret, parse_start);
12875 Set_Node_Cur_Length(ret, parse_start - 2);
12876 nextchar(pRExC_state);
12879 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
12880 * \N{...} evaluates to a sequence of more than one code points).
12881 * The function call below returns a regnode, which is our result.
12882 * The parameters cause it to fail if the \N{} evaluates to a
12883 * single code point; we handle those like any other literal. The
12884 * reason that the multicharacter case is handled here and not as
12885 * part of the EXACtish code is because of quantifiers. In
12886 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
12887 * this way makes that Just Happen. dmq.
12888 * join_exact() will join this up with adjacent EXACTish nodes
12889 * later on, if appropriate. */
12891 if (grok_bslash_N(pRExC_state,
12892 &ret, /* Want a regnode returned */
12893 NULL, /* Fail if evaluates to a single code
12895 NULL, /* Don't need a count of how many code
12904 if (*flagp & RESTART_PASS1)
12907 /* Here, evaluates to a single code point. Go get that */
12908 RExC_parse = parse_start;
12911 case 'k': /* Handle \k<NAME> and \k'NAME' */
12915 if ( RExC_parse >= RExC_end - 1
12916 || (( ch = RExC_parse[1]) != '<'
12921 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12922 vFAIL2("Sequence %.2s... not terminated",parse_start);
12925 ret = handle_named_backref(pRExC_state,
12937 case '1': case '2': case '3': case '4':
12938 case '5': case '6': case '7': case '8': case '9':
12943 if (*RExC_parse == 'g') {
12947 if (*RExC_parse == '{') {
12951 if (*RExC_parse == '-') {
12955 if (hasbrace && !isDIGIT(*RExC_parse)) {
12956 if (isrel) RExC_parse--;
12958 goto parse_named_seq;
12961 if (RExC_parse >= RExC_end) {
12962 goto unterminated_g;
12964 num = S_backref_value(RExC_parse);
12966 vFAIL("Reference to invalid group 0");
12967 else if (num == I32_MAX) {
12968 if (isDIGIT(*RExC_parse))
12969 vFAIL("Reference to nonexistent group");
12972 vFAIL("Unterminated \\g... pattern");
12976 num = RExC_npar - num;
12978 vFAIL("Reference to nonexistent or unclosed group");
12982 num = S_backref_value(RExC_parse);
12983 /* bare \NNN might be backref or octal - if it is larger
12984 * than or equal RExC_npar then it is assumed to be an
12985 * octal escape. Note RExC_npar is +1 from the actual
12986 * number of parens. */
12987 /* Note we do NOT check if num == I32_MAX here, as that is
12988 * handled by the RExC_npar check */
12991 /* any numeric escape < 10 is always a backref */
12993 /* any numeric escape < RExC_npar is a backref */
12994 && num >= RExC_npar
12995 /* cannot be an octal escape if it starts with 8 */
12996 && *RExC_parse != '8'
12997 /* cannot be an octal escape it it starts with 9 */
12998 && *RExC_parse != '9'
13001 /* Probably not a backref, instead likely to be an
13002 * octal character escape, e.g. \35 or \777.
13003 * The above logic should make it obvious why using
13004 * octal escapes in patterns is problematic. - Yves */
13005 RExC_parse = parse_start;
13010 /* At this point RExC_parse points at a numeric escape like
13011 * \12 or \88 or something similar, which we should NOT treat
13012 * as an octal escape. It may or may not be a valid backref
13013 * escape. For instance \88888888 is unlikely to be a valid
13015 while (isDIGIT(*RExC_parse))
13018 if (*RExC_parse != '}')
13019 vFAIL("Unterminated \\g{...} pattern");
13023 if (num > (I32)RExC_rx->nparens)
13024 vFAIL("Reference to nonexistent group");
13027 ret = reganode(pRExC_state,
13030 : (ASCII_FOLD_RESTRICTED)
13032 : (AT_LEAST_UNI_SEMANTICS)
13038 *flagp |= HASWIDTH;
13040 /* override incorrect value set in reganode MJD */
13041 Set_Node_Offset(ret, parse_start);
13042 Set_Node_Cur_Length(ret, parse_start-1);
13043 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13044 FALSE /* Don't force to /x */ );
13048 if (RExC_parse >= RExC_end)
13049 FAIL("Trailing \\");
13052 /* Do not generate "unrecognized" warnings here, we fall
13053 back into the quick-grab loop below */
13054 RExC_parse = parse_start;
13056 } /* end of switch on a \foo sequence */
13061 /* '#' comments should have been spaced over before this function was
13063 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13065 if (RExC_flags & RXf_PMf_EXTENDED) {
13066 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13067 if (RExC_parse < RExC_end)
13077 /* Here, we have determined that the next thing is probably a
13078 * literal character. RExC_parse points to the first byte of its
13079 * definition. (It still may be an escape sequence that evaluates
13080 * to a single character) */
13086 #define MAX_NODE_STRING_SIZE 127
13087 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
13089 U8 upper_parse = MAX_NODE_STRING_SIZE;
13090 U8 node_type = compute_EXACTish(pRExC_state);
13091 bool next_is_quantifier;
13092 char * oldp = NULL;
13094 /* We can convert EXACTF nodes to EXACTFU if they contain only
13095 * characters that match identically regardless of the target
13096 * string's UTF8ness. The reason to do this is that EXACTF is not
13097 * trie-able, EXACTFU is.
13099 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13100 * contain only above-Latin1 characters (hence must be in UTF8),
13101 * which don't participate in folds with Latin1-range characters,
13102 * as the latter's folds aren't known until runtime. (We don't
13103 * need to figure this out until pass 2) */
13104 bool maybe_exactfu = PASS2
13105 && (node_type == EXACTF || node_type == EXACTFL);
13107 /* If a folding node contains only code points that don't
13108 * participate in folds, it can be changed into an EXACT node,
13109 * which allows the optimizer more things to look for */
13112 ret = reg_node(pRExC_state, node_type);
13114 /* In pass1, folded, we use a temporary buffer instead of the
13115 * actual node, as the node doesn't exist yet */
13116 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
13122 /* We look for the EXACTFish to EXACT node optimizaton only if
13123 * folding. (And we don't need to figure this out until pass 2).
13124 * XXX It might actually make sense to split the node into portions
13125 * that are exact and ones that aren't, so that we could later use
13126 * the exact ones to find the longest fixed and floating strings.
13127 * One would want to join them back into a larger node. One could
13128 * use a pseudo regnode like 'EXACT_ORIG_FOLD' */
13129 maybe_exact = FOLD && PASS2;
13131 /* XXX The node can hold up to 255 bytes, yet this only goes to
13132 * 127. I (khw) do not know why. Keeping it somewhat less than
13133 * 255 allows us to not have to worry about overflow due to
13134 * converting to utf8 and fold expansion, but that value is
13135 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
13136 * split up by this limit into a single one using the real max of
13137 * 255. Even at 127, this breaks under rare circumstances. If
13138 * folding, we do not want to split a node at a character that is a
13139 * non-final in a multi-char fold, as an input string could just
13140 * happen to want to match across the node boundary. The join
13141 * would solve that problem if the join actually happens. But a
13142 * series of more than two nodes in a row each of 127 would cause
13143 * the first join to succeed to get to 254, but then there wouldn't
13144 * be room for the next one, which could at be one of those split
13145 * multi-char folds. I don't know of any fool-proof solution. One
13146 * could back off to end with only a code point that isn't such a
13147 * non-final, but it is possible for there not to be any in the
13150 assert( ! UTF /* Is at the beginning of a character */
13151 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13152 || UTF8_IS_START(UCHARAT(RExC_parse)));
13154 /* Here, we have a literal character. Find the maximal string of
13155 * them in the input that we can fit into a single EXACTish node.
13156 * We quit at the first non-literal or when the node gets full */
13157 for (p = RExC_parse;
13158 len < upper_parse && p < RExC_end;
13163 /* White space has already been ignored */
13164 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13165 || ! is_PATWS_safe((p), RExC_end, UTF));
13177 /* Literal Escapes Switch
13179 This switch is meant to handle escape sequences that
13180 resolve to a literal character.
13182 Every escape sequence that represents something
13183 else, like an assertion or a char class, is handled
13184 in the switch marked 'Special Escapes' above in this
13185 routine, but also has an entry here as anything that
13186 isn't explicitly mentioned here will be treated as
13187 an unescaped equivalent literal.
13190 switch ((U8)*++p) {
13191 /* These are all the special escapes. */
13192 case 'A': /* Start assertion */
13193 case 'b': case 'B': /* Word-boundary assertion*/
13194 case 'C': /* Single char !DANGEROUS! */
13195 case 'd': case 'D': /* digit class */
13196 case 'g': case 'G': /* generic-backref, pos assertion */
13197 case 'h': case 'H': /* HORIZWS */
13198 case 'k': case 'K': /* named backref, keep marker */
13199 case 'p': case 'P': /* Unicode property */
13200 case 'R': /* LNBREAK */
13201 case 's': case 'S': /* space class */
13202 case 'v': case 'V': /* VERTWS */
13203 case 'w': case 'W': /* word class */
13204 case 'X': /* eXtended Unicode "combining
13205 character sequence" */
13206 case 'z': case 'Z': /* End of line/string assertion */
13210 /* Anything after here is an escape that resolves to a
13211 literal. (Except digits, which may or may not)
13217 case 'N': /* Handle a single-code point named character. */
13218 RExC_parse = p + 1;
13219 if (! grok_bslash_N(pRExC_state,
13220 NULL, /* Fail if evaluates to
13221 anything other than a
13222 single code point */
13223 &ender, /* The returned single code
13225 NULL, /* Don't need a count of
13226 how many code points */
13231 if (*flagp & NEED_UTF8)
13232 FAIL("panic: grok_bslash_N set NEED_UTF8");
13233 if (*flagp & RESTART_PASS1)
13236 /* Here, it wasn't a single code point. Go close
13237 * up this EXACTish node. The switch() prior to
13238 * this switch handles the other cases */
13239 RExC_parse = p = oldp;
13243 if (ender > 0xff) {
13244 REQUIRE_UTF8(flagp);
13260 ender = ESC_NATIVE;
13270 const char* error_msg;
13272 bool valid = grok_bslash_o(&p,
13275 PASS2, /* out warnings */
13276 (bool) RExC_strict,
13277 TRUE, /* Output warnings
13282 RExC_parse = p; /* going to die anyway; point
13283 to exact spot of failure */
13287 if (ender > 0xff) {
13288 REQUIRE_UTF8(flagp);
13294 UV result = UV_MAX; /* initialize to erroneous
13296 const char* error_msg;
13298 bool valid = grok_bslash_x(&p,
13301 PASS2, /* out warnings */
13302 (bool) RExC_strict,
13303 TRUE, /* Silence warnings
13308 RExC_parse = p; /* going to die anyway; point
13309 to exact spot of failure */
13314 if (ender < 0x100) {
13316 if (RExC_recode_x_to_native) {
13317 ender = LATIN1_TO_NATIVE(ender);
13322 REQUIRE_UTF8(flagp);
13328 ender = grok_bslash_c(*p++, PASS2);
13330 case '8': case '9': /* must be a backreference */
13332 /* we have an escape like \8 which cannot be an octal escape
13333 * so we exit the loop, and let the outer loop handle this
13334 * escape which may or may not be a legitimate backref. */
13336 case '1': case '2': case '3':case '4':
13337 case '5': case '6': case '7':
13338 /* When we parse backslash escapes there is ambiguity
13339 * between backreferences and octal escapes. Any escape
13340 * from \1 - \9 is a backreference, any multi-digit
13341 * escape which does not start with 0 and which when
13342 * evaluated as decimal could refer to an already
13343 * parsed capture buffer is a back reference. Anything
13346 * Note this implies that \118 could be interpreted as
13347 * 118 OR as "\11" . "8" depending on whether there
13348 * were 118 capture buffers defined already in the
13351 /* NOTE, RExC_npar is 1 more than the actual number of
13352 * parens we have seen so far, hence the < RExC_npar below. */
13354 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
13355 { /* Not to be treated as an octal constant, go
13363 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13365 ender = grok_oct(p, &numlen, &flags, NULL);
13366 if (ender > 0xff) {
13367 REQUIRE_UTF8(flagp);
13370 if (PASS2 /* like \08, \178 */
13372 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
13374 reg_warn_non_literal_string(
13376 form_short_octal_warning(p, numlen));
13382 FAIL("Trailing \\");
13385 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
13386 /* Include any left brace following the alpha to emphasize
13387 * that it could be part of an escape at some point
13389 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
13390 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
13392 goto normal_default;
13393 } /* End of switch on '\' */
13396 /* Currently we don't care if the lbrace is at the start
13397 * of a construct. This catches it in the middle of a
13398 * literal string, or when it's the first thing after
13399 * something like "\b" */
13400 if (len || (p > RExC_start && isALPHA_A(*(p -1)))) {
13401 RExC_parse = p + 1;
13402 vFAIL("Unescaped left brace in regex is illegal here");
13405 default: /* A literal character */
13407 if (! UTF8_IS_INVARIANT(*p) && UTF) {
13409 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
13410 &numlen, UTF8_ALLOW_DEFAULT);
13416 } /* End of switch on the literal */
13418 /* Here, have looked at the literal character and <ender>
13419 * contains its ordinal, <p> points to the character after it.
13420 * We need to check if the next non-ignored thing is a
13421 * quantifier. Move <p> to after anything that should be
13422 * ignored, which, as a side effect, positions <p> for the next
13423 * loop iteration */
13424 skip_to_be_ignored_text(pRExC_state, &p,
13425 FALSE /* Don't force to /x */ );
13427 /* If the next thing is a quantifier, it applies to this
13428 * character only, which means that this character has to be in
13429 * its own node and can't just be appended to the string in an
13430 * existing node, so if there are already other characters in
13431 * the node, close the node with just them, and set up to do
13432 * this character again next time through, when it will be the
13433 * only thing in its new node */
13435 if ((next_is_quantifier = ( LIKELY(p < RExC_end)
13436 && UNLIKELY(ISMULT2(p))))
13443 /* Ready to add 'ender' to the node */
13445 if (! FOLD) { /* The simple case, just append the literal */
13447 /* In the sizing pass, we need only the size of the
13448 * character we are appending, hence we can delay getting
13449 * its representation until PASS2. */
13452 const STRLEN unilen = UVCHR_SKIP(ender);
13455 /* We have to subtract 1 just below (and again in
13456 * the corresponding PASS2 code) because the loop
13457 * increments <len> each time, as all but this path
13458 * (and one other) through it add a single byte to
13459 * the EXACTish node. But these paths would change
13460 * len to be the correct final value, so cancel out
13461 * the increment that follows */
13467 } else { /* PASS2 */
13470 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
13471 len += (char *) new_s - s - 1;
13472 s = (char *) new_s;
13475 *(s++) = (char) ender;
13479 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
13481 /* Here are folding under /l, and the code point is
13482 * problematic. First, we know we can't simplify things */
13483 maybe_exact = FALSE;
13484 maybe_exactfu = FALSE;
13486 /* A problematic code point in this context means that its
13487 * fold isn't known until runtime, so we can't fold it now.
13488 * (The non-problematic code points are the above-Latin1
13489 * ones that fold to also all above-Latin1. Their folds
13490 * don't vary no matter what the locale is.) But here we
13491 * have characters whose fold depends on the locale.
13492 * Unlike the non-folding case above, we have to keep track
13493 * of these in the sizing pass, so that we can make sure we
13494 * don't split too-long nodes in the middle of a potential
13495 * multi-char fold. And unlike the regular fold case
13496 * handled in the else clauses below, we don't actually
13497 * fold and don't have special cases to consider. What we
13498 * do for both passes is the PASS2 code for non-folding */
13499 goto not_fold_common;
13501 else /* A regular FOLD code point */
13503 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13504 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13505 || UNICODE_DOT_DOT_VERSION > 0)
13506 /* See comments for join_exact() as to why we fold
13507 * this non-UTF at compile time */
13508 || ( node_type == EXACTFU
13509 && ender == LATIN_SMALL_LETTER_SHARP_S)
13512 /* Here, are folding and are not UTF-8 encoded; therefore
13513 * the character must be in the range 0-255, and is not /l
13514 * (Not /l because we already handled these under /l in
13515 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
13516 if (IS_IN_SOME_FOLD_L1(ender)) {
13517 maybe_exact = FALSE;
13519 /* See if the character's fold differs between /d and
13520 * /u. This includes the multi-char fold SHARP S to
13522 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
13523 RExC_seen_unfolded_sharp_s = 1;
13524 maybe_exactfu = FALSE;
13526 else if (maybe_exactfu
13527 && (PL_fold[ender] != PL_fold_latin1[ender]
13528 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13529 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13530 || UNICODE_DOT_DOT_VERSION > 0)
13532 && isALPHA_FOLD_EQ(ender, 's')
13533 && isALPHA_FOLD_EQ(*(s-1), 's'))
13536 maybe_exactfu = FALSE;
13540 /* Even when folding, we store just the input character, as
13541 * we have an array that finds its fold quickly */
13542 *(s++) = (char) ender;
13544 else { /* FOLD, and UTF (or sharp s) */
13545 /* Unlike the non-fold case, we do actually have to
13546 * calculate the results here in pass 1. This is for two
13547 * reasons, the folded length may be longer than the
13548 * unfolded, and we have to calculate how many EXACTish
13549 * nodes it will take; and we may run out of room in a node
13550 * in the middle of a potential multi-char fold, and have
13551 * to back off accordingly. */
13554 if (isASCII_uni(ender)) {
13555 folded = toFOLD(ender);
13556 *(s)++ = (U8) folded;
13561 folded = _to_uni_fold_flags(
13565 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
13566 ? FOLD_FLAGS_NOMIX_ASCII
13570 /* The loop increments <len> each time, as all but this
13571 * path (and one other) through it add a single byte to
13572 * the EXACTish node. But this one has changed len to
13573 * be the correct final value, so subtract one to
13574 * cancel out the increment that follows */
13575 len += foldlen - 1;
13577 /* If this node only contains non-folding code points so
13578 * far, see if this new one is also non-folding */
13580 if (folded != ender) {
13581 maybe_exact = FALSE;
13584 /* Here the fold is the original; we have to check
13585 * further to see if anything folds to it */
13586 if (_invlist_contains_cp(PL_utf8_foldable,
13589 maybe_exact = FALSE;
13596 if (next_is_quantifier) {
13598 /* Here, the next input is a quantifier, and to get here,
13599 * the current character is the only one in the node.
13600 * Also, here <len> doesn't include the final byte for this
13606 } /* End of loop through literal characters */
13608 /* Here we have either exhausted the input or ran out of room in
13609 * the node. (If we encountered a character that can't be in the
13610 * node, transfer is made directly to <loopdone>, and so we
13611 * wouldn't have fallen off the end of the loop.) In the latter
13612 * case, we artificially have to split the node into two, because
13613 * we just don't have enough space to hold everything. This
13614 * creates a problem if the final character participates in a
13615 * multi-character fold in the non-final position, as a match that
13616 * should have occurred won't, due to the way nodes are matched,
13617 * and our artificial boundary. So back off until we find a non-
13618 * problematic character -- one that isn't at the beginning or
13619 * middle of such a fold. (Either it doesn't participate in any
13620 * folds, or appears only in the final position of all the folds it
13621 * does participate in.) A better solution with far fewer false
13622 * positives, and that would fill the nodes more completely, would
13623 * be to actually have available all the multi-character folds to
13624 * test against, and to back-off only far enough to be sure that
13625 * this node isn't ending with a partial one. <upper_parse> is set
13626 * further below (if we need to reparse the node) to include just
13627 * up through that final non-problematic character that this code
13628 * identifies, so when it is set to less than the full node, we can
13629 * skip the rest of this */
13630 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
13632 const STRLEN full_len = len;
13634 assert(len >= MAX_NODE_STRING_SIZE);
13636 /* Here, <s> points to the final byte of the final character.
13637 * Look backwards through the string until find a non-
13638 * problematic character */
13642 /* This has no multi-char folds to non-UTF characters */
13643 if (ASCII_FOLD_RESTRICTED) {
13647 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
13651 if (! PL_NonL1NonFinalFold) {
13652 PL_NonL1NonFinalFold = _new_invlist_C_array(
13653 NonL1_Perl_Non_Final_Folds_invlist);
13656 /* Point to the first byte of the final character */
13657 s = (char *) utf8_hop((U8 *) s, -1);
13659 while (s >= s0) { /* Search backwards until find
13660 non-problematic char */
13661 if (UTF8_IS_INVARIANT(*s)) {
13663 /* There are no ascii characters that participate
13664 * in multi-char folds under /aa. In EBCDIC, the
13665 * non-ascii invariants are all control characters,
13666 * so don't ever participate in any folds. */
13667 if (ASCII_FOLD_RESTRICTED
13668 || ! IS_NON_FINAL_FOLD(*s))
13673 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
13674 if (! IS_NON_FINAL_FOLD(EIGHT_BIT_UTF8_TO_NATIVE(
13680 else if (! _invlist_contains_cp(
13681 PL_NonL1NonFinalFold,
13682 valid_utf8_to_uvchr((U8 *) s, NULL)))
13687 /* Here, the current character is problematic in that
13688 * it does occur in the non-final position of some
13689 * fold, so try the character before it, but have to
13690 * special case the very first byte in the string, so
13691 * we don't read outside the string */
13692 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
13693 } /* End of loop backwards through the string */
13695 /* If there were only problematic characters in the string,
13696 * <s> will point to before s0, in which case the length
13697 * should be 0, otherwise include the length of the
13698 * non-problematic character just found */
13699 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
13702 /* Here, have found the final character, if any, that is
13703 * non-problematic as far as ending the node without splitting
13704 * it across a potential multi-char fold. <len> contains the
13705 * number of bytes in the node up-to and including that
13706 * character, or is 0 if there is no such character, meaning
13707 * the whole node contains only problematic characters. In
13708 * this case, give up and just take the node as-is. We can't
13713 /* If the node ends in an 's' we make sure it stays EXACTF,
13714 * as if it turns into an EXACTFU, it could later get
13715 * joined with another 's' that would then wrongly match
13717 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
13719 maybe_exactfu = FALSE;
13723 /* Here, the node does contain some characters that aren't
13724 * problematic. If one such is the final character in the
13725 * node, we are done */
13726 if (len == full_len) {
13729 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
13731 /* If the final character is problematic, but the
13732 * penultimate is not, back-off that last character to
13733 * later start a new node with it */
13738 /* Here, the final non-problematic character is earlier
13739 * in the input than the penultimate character. What we do
13740 * is reparse from the beginning, going up only as far as
13741 * this final ok one, thus guaranteeing that the node ends
13742 * in an acceptable character. The reason we reparse is
13743 * that we know how far in the character is, but we don't
13744 * know how to correlate its position with the input parse.
13745 * An alternate implementation would be to build that
13746 * correlation as we go along during the original parse,
13747 * but that would entail extra work for every node, whereas
13748 * this code gets executed only when the string is too
13749 * large for the node, and the final two characters are
13750 * problematic, an infrequent occurrence. Yet another
13751 * possible strategy would be to save the tail of the
13752 * string, and the next time regatom is called, initialize
13753 * with that. The problem with this is that unless you
13754 * back off one more character, you won't be guaranteed
13755 * regatom will get called again, unless regbranch,
13756 * regpiece ... are also changed. If you do back off that
13757 * extra character, so that there is input guaranteed to
13758 * force calling regatom, you can't handle the case where
13759 * just the first character in the node is acceptable. I
13760 * (khw) decided to try this method which doesn't have that
13761 * pitfall; if performance issues are found, we can do a
13762 * combination of the current approach plus that one */
13768 } /* End of verifying node ends with an appropriate char */
13770 loopdone: /* Jumped to when encounters something that shouldn't be
13773 /* I (khw) don't know if you can get here with zero length, but the
13774 * old code handled this situation by creating a zero-length EXACT
13775 * node. Might as well be NOTHING instead */
13781 /* If 'maybe_exact' is still set here, means there are no
13782 * code points in the node that participate in folds;
13783 * similarly for 'maybe_exactfu' and code points that match
13784 * differently depending on UTF8ness of the target string
13785 * (for /u), or depending on locale for /l */
13791 else if (maybe_exactfu) {
13797 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
13798 FALSE /* Don't look to see if could
13799 be turned into an EXACT
13800 node, as we have already
13805 RExC_parse = p - 1;
13806 Set_Node_Cur_Length(ret, parse_start);
13809 /* len is STRLEN which is unsigned, need to copy to signed */
13812 vFAIL("Internal disaster");
13815 } /* End of label 'defchar:' */
13817 } /* End of giant switch on input character */
13819 /* Position parse to next real character */
13820 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13821 FALSE /* Don't force to /x */ );
13822 if (PASS2 && *RExC_parse == '{' && OP(ret) != SBOL && ! regcurly(RExC_parse)) {
13823 ckWARNregdep(RExC_parse + 1, "Unescaped left brace in regex is deprecated here, passed through");
13831 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
13833 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
13834 * sets up the bitmap and any flags, removing those code points from the
13835 * inversion list, setting it to NULL should it become completely empty */
13837 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
13838 assert(PL_regkind[OP(node)] == ANYOF);
13840 ANYOF_BITMAP_ZERO(node);
13841 if (*invlist_ptr) {
13843 /* This gets set if we actually need to modify things */
13844 bool change_invlist = FALSE;
13848 /* Start looking through *invlist_ptr */
13849 invlist_iterinit(*invlist_ptr);
13850 while (invlist_iternext(*invlist_ptr, &start, &end)) {
13854 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
13855 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
13858 /* Quit if are above what we should change */
13859 if (start >= NUM_ANYOF_CODE_POINTS) {
13863 change_invlist = TRUE;
13865 /* Set all the bits in the range, up to the max that we are doing */
13866 high = (end < NUM_ANYOF_CODE_POINTS - 1)
13868 : NUM_ANYOF_CODE_POINTS - 1;
13869 for (i = start; i <= (int) high; i++) {
13870 if (! ANYOF_BITMAP_TEST(node, i)) {
13871 ANYOF_BITMAP_SET(node, i);
13875 invlist_iterfinish(*invlist_ptr);
13877 /* Done with loop; remove any code points that are in the bitmap from
13878 * *invlist_ptr; similarly for code points above the bitmap if we have
13879 * a flag to match all of them anyways */
13880 if (change_invlist) {
13881 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
13883 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
13884 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
13887 /* If have completely emptied it, remove it completely */
13888 if (_invlist_len(*invlist_ptr) == 0) {
13889 SvREFCNT_dec_NN(*invlist_ptr);
13890 *invlist_ptr = NULL;
13895 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
13896 Character classes ([:foo:]) can also be negated ([:^foo:]).
13897 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
13898 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
13899 but trigger failures because they are currently unimplemented. */
13901 #define POSIXCC_DONE(c) ((c) == ':')
13902 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
13903 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
13904 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
13906 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
13907 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
13908 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
13910 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
13912 /* 'posix_warnings' and 'warn_text' are names of variables in the following
13914 #define ADD_POSIX_WARNING(p, text) STMT_START { \
13915 if (posix_warnings) { \
13916 if (! RExC_warn_text ) RExC_warn_text = (AV *) sv_2mortal((SV *) newAV()); \
13917 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
13921 REPORT_LOCATION_ARGS(p))); \
13926 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
13928 const char * const s, /* Where the putative posix class begins.
13929 Normally, this is one past the '['. This
13930 parameter exists so it can be somewhere
13931 besides RExC_parse. */
13932 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
13934 AV ** posix_warnings, /* Where to place any generated warnings, or
13936 const bool check_only /* Don't die if error */
13939 /* This parses what the caller thinks may be one of the three POSIX
13941 * 1) a character class, like [:blank:]
13942 * 2) a collating symbol, like [. .]
13943 * 3) an equivalence class, like [= =]
13944 * In the latter two cases, it croaks if it finds a syntactically legal
13945 * one, as these are not handled by Perl.
13947 * The main purpose is to look for a POSIX character class. It returns:
13948 * a) the class number
13949 * if it is a completely syntactically and semantically legal class.
13950 * 'updated_parse_ptr', if not NULL, is set to point to just after the
13951 * closing ']' of the class
13952 * b) OOB_NAMEDCLASS
13953 * if it appears that one of the three POSIX constructs was meant, but
13954 * its specification was somehow defective. 'updated_parse_ptr', if
13955 * not NULL, is set to point to the character just after the end
13956 * character of the class. See below for handling of warnings.
13957 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
13958 * if it doesn't appear that a POSIX construct was intended.
13959 * 'updated_parse_ptr' is not changed. No warnings nor errors are
13962 * In b) there may be errors or warnings generated. If 'check_only' is
13963 * TRUE, then any errors are discarded. Warnings are returned to the
13964 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
13965 * instead it is NULL, warnings are suppressed. This is done in all
13966 * passes. The reason for this is that the rest of the parsing is heavily
13967 * dependent on whether this routine found a valid posix class or not. If
13968 * it did, the closing ']' is absorbed as part of the class. If no class,
13969 * or an invalid one is found, any ']' will be considered the terminator of
13970 * the outer bracketed character class, leading to very different results.
13971 * In particular, a '(?[ ])' construct will likely have a syntax error if
13972 * the class is parsed other than intended, and this will happen in pass1,
13973 * before the warnings would normally be output. This mechanism allows the
13974 * caller to output those warnings in pass1 just before dieing, giving a
13975 * much better clue as to what is wrong.
13977 * The reason for this function, and its complexity is that a bracketed
13978 * character class can contain just about anything. But it's easy to
13979 * mistype the very specific posix class syntax but yielding a valid
13980 * regular bracketed class, so it silently gets compiled into something
13981 * quite unintended.
13983 * The solution adopted here maintains backward compatibility except that
13984 * it adds a warning if it looks like a posix class was intended but
13985 * improperly specified. The warning is not raised unless what is input
13986 * very closely resembles one of the 14 legal posix classes. To do this,
13987 * it uses fuzzy parsing. It calculates how many single-character edits it
13988 * would take to transform what was input into a legal posix class. Only
13989 * if that number is quite small does it think that the intention was a
13990 * posix class. Obviously these are heuristics, and there will be cases
13991 * where it errs on one side or another, and they can be tweaked as
13992 * experience informs.
13994 * The syntax for a legal posix class is:
13996 * qr/(?xa: \[ : \^? [:lower:]{4,6} : \] )/
13998 * What this routine considers syntactically to be an intended posix class
13999 * is this (the comments indicate some restrictions that the pattern
14002 * qr/(?x: \[? # The left bracket, possibly
14004 * \h* # possibly followed by blanks
14005 * (?: \^ \h* )? # possibly a misplaced caret
14006 * [:;]? # The opening class character,
14007 * # possibly omitted. A typo
14008 * # semi-colon can also be used.
14010 * \^? # possibly a correctly placed
14011 * # caret, but not if there was also
14012 * # a misplaced one
14014 * .{3,15} # The class name. If there are
14015 * # deviations from the legal syntax,
14016 * # its edit distance must be close
14017 * # to a real class name in order
14018 * # for it to be considered to be
14019 * # an intended posix class.
14021 * [:punct:]? # The closing class character,
14022 * # possibly omitted. If not a colon
14023 * # nor semi colon, the class name
14024 * # must be even closer to a valid
14027 * \]? # The right bracket, possibly
14031 * In the above, \h must be ASCII-only.
14033 * These are heuristics, and can be tweaked as field experience dictates.
14034 * There will be cases when someone didn't intend to specify a posix class
14035 * that this warns as being so. The goal is to minimize these, while
14036 * maximizing the catching of things intended to be a posix class that
14037 * aren't parsed as such.
14041 const char * const e = RExC_end;
14042 unsigned complement = 0; /* If to complement the class */
14043 bool found_problem = FALSE; /* Assume OK until proven otherwise */
14044 bool has_opening_bracket = FALSE;
14045 bool has_opening_colon = FALSE;
14046 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
14048 const char * possible_end = NULL; /* used for a 2nd parse pass */
14049 const char* name_start; /* ptr to class name first char */
14051 /* If the number of single-character typos the input name is away from a
14052 * legal name is no more than this number, it is considered to have meant
14053 * the legal name */
14054 int max_distance = 2;
14056 /* to store the name. The size determines the maximum length before we
14057 * decide that no posix class was intended. Should be at least
14058 * sizeof("alphanumeric") */
14061 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
14063 if (posix_warnings && RExC_warn_text)
14064 av_clear(RExC_warn_text);
14067 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14070 if (*(p - 1) != '[') {
14071 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
14072 found_problem = TRUE;
14075 has_opening_bracket = TRUE;
14078 /* They could be confused and think you can put spaces between the
14081 found_problem = TRUE;
14085 } while (p < e && isBLANK(*p));
14087 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14090 /* For [. .] and [= =]. These are quite different internally from [: :],
14091 * so they are handled separately. */
14092 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
14093 and 1 for at least one char in it
14096 const char open_char = *p;
14097 const char * temp_ptr = p + 1;
14099 /* These two constructs are not handled by perl, and if we find a
14100 * syntactically valid one, we croak. khw, who wrote this code, finds
14101 * this explanation of them very unclear:
14102 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
14103 * And searching the rest of the internet wasn't very helpful either.
14104 * It looks like just about any byte can be in these constructs,
14105 * depending on the locale. But unless the pattern is being compiled
14106 * under /l, which is very rare, Perl runs under the C or POSIX locale.
14107 * In that case, it looks like [= =] isn't allowed at all, and that
14108 * [. .] could be any single code point, but for longer strings the
14109 * constituent characters would have to be the ASCII alphabetics plus
14110 * the minus-hyphen. Any sensible locale definition would limit itself
14111 * to these. And any portable one definitely should. Trying to parse
14112 * the general case is a nightmare (see [perl #127604]). So, this code
14113 * looks only for interiors of these constructs that match:
14115 * Using \w relaxes the apparent rules a little, without adding much
14116 * danger of mistaking something else for one of these constructs.
14118 * [. .] in some implementations described on the internet is usable to
14119 * escape a character that otherwise is special in bracketed character
14120 * classes. For example [.].] means a literal right bracket instead of
14121 * the ending of the class
14123 * [= =] can legitimately contain a [. .] construct, but we don't
14124 * handle this case, as that [. .] construct will later get parsed
14125 * itself and croak then. And [= =] is checked for even when not under
14126 * /l, as Perl has long done so.
14128 * The code below relies on there being a trailing NUL, so it doesn't
14129 * have to keep checking if the parse ptr < e.
14131 if (temp_ptr[1] == open_char) {
14134 else while ( temp_ptr < e
14135 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
14140 if (*temp_ptr == open_char) {
14142 if (*temp_ptr == ']') {
14144 if (! found_problem && ! check_only) {
14145 RExC_parse = (char *) temp_ptr;
14146 vFAIL3("POSIX syntax [%c %c] is reserved for future "
14147 "extensions", open_char, open_char);
14150 /* Here, the syntax wasn't completely valid, or else the call
14151 * is to check-only */
14152 if (updated_parse_ptr) {
14153 *updated_parse_ptr = (char *) temp_ptr;
14156 return OOB_NAMEDCLASS;
14160 /* If we find something that started out to look like one of these
14161 * constructs, but isn't, we continue below so that it can be checked
14162 * for being a class name with a typo of '.' or '=' instead of a colon.
14166 /* Here, we think there is a possibility that a [: :] class was meant, and
14167 * we have the first real character. It could be they think the '^' comes
14170 found_problem = TRUE;
14171 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
14176 found_problem = TRUE;
14180 } while (p < e && isBLANK(*p));
14182 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14186 /* But the first character should be a colon, which they could have easily
14187 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
14188 * distinguish from a colon, so treat that as a colon). */
14191 has_opening_colon = TRUE;
14193 else if (*p == ';') {
14194 found_problem = TRUE;
14196 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14197 has_opening_colon = TRUE;
14200 found_problem = TRUE;
14201 ADD_POSIX_WARNING(p, "there must be a starting ':'");
14203 /* Consider an initial punctuation (not one of the recognized ones) to
14204 * be a left terminator */
14205 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
14210 /* They may think that you can put spaces between the components */
14212 found_problem = TRUE;
14216 } while (p < e && isBLANK(*p));
14218 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14223 /* We consider something like [^:^alnum:]] to not have been intended to
14224 * be a posix class, but XXX maybe we should */
14226 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14233 /* Again, they may think that you can put spaces between the components */
14235 found_problem = TRUE;
14239 } while (p < e && isBLANK(*p));
14241 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14246 /* XXX This ']' may be a typo, and something else was meant. But
14247 * treating it as such creates enough complications, that that
14248 * possibility isn't currently considered here. So we assume that the
14249 * ']' is what is intended, and if we've already found an initial '[',
14250 * this leaves this construct looking like [:] or [:^], which almost
14251 * certainly weren't intended to be posix classes */
14252 if (has_opening_bracket) {
14253 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14256 /* But this function can be called when we parse the colon for
14257 * something like qr/[alpha:]]/, so we back up to look for the
14262 found_problem = TRUE;
14263 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14265 else if (*p != ':') {
14267 /* XXX We are currently very restrictive here, so this code doesn't
14268 * consider the possibility that, say, /[alpha.]]/ was intended to
14269 * be a posix class. */
14270 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14273 /* Here we have something like 'foo:]'. There was no initial colon,
14274 * and we back up over 'foo. XXX Unlike the going forward case, we
14275 * don't handle typos of non-word chars in the middle */
14276 has_opening_colon = FALSE;
14279 while (p > RExC_start && isWORDCHAR(*p)) {
14284 /* Here, we have positioned ourselves to where we think the first
14285 * character in the potential class is */
14288 /* Now the interior really starts. There are certain key characters that
14289 * can end the interior, or these could just be typos. To catch both
14290 * cases, we may have to do two passes. In the first pass, we keep on
14291 * going unless we come to a sequence that matches
14292 * qr/ [[:punct:]] [[:blank:]]* \] /xa
14293 * This means it takes a sequence to end the pass, so two typos in a row if
14294 * that wasn't what was intended. If the class is perfectly formed, just
14295 * this one pass is needed. We also stop if there are too many characters
14296 * being accumulated, but this number is deliberately set higher than any
14297 * real class. It is set high enough so that someone who thinks that
14298 * 'alphanumeric' is a correct name would get warned that it wasn't.
14299 * While doing the pass, we keep track of where the key characters were in
14300 * it. If we don't find an end to the class, and one of the key characters
14301 * was found, we redo the pass, but stop when we get to that character.
14302 * Thus the key character was considered a typo in the first pass, but a
14303 * terminator in the second. If two key characters are found, we stop at
14304 * the second one in the first pass. Again this can miss two typos, but
14305 * catches a single one
14307 * In the first pass, 'possible_end' starts as NULL, and then gets set to
14308 * point to the first key character. For the second pass, it starts as -1.
14314 bool has_blank = FALSE;
14315 bool has_upper = FALSE;
14316 bool has_terminating_colon = FALSE;
14317 bool has_terminating_bracket = FALSE;
14318 bool has_semi_colon = FALSE;
14319 unsigned int name_len = 0;
14320 int punct_count = 0;
14324 /* Squeeze out blanks when looking up the class name below */
14325 if (isBLANK(*p) ) {
14327 found_problem = TRUE;
14332 /* The name will end with a punctuation */
14334 const char * peek = p + 1;
14336 /* Treat any non-']' punctuation followed by a ']' (possibly
14337 * with intervening blanks) as trying to terminate the class.
14338 * ']]' is very likely to mean a class was intended (but
14339 * missing the colon), but the warning message that gets
14340 * generated shows the error position better if we exit the
14341 * loop at the bottom (eventually), so skip it here. */
14343 if (peek < e && isBLANK(*peek)) {
14345 found_problem = TRUE;
14348 } while (peek < e && isBLANK(*peek));
14351 if (peek < e && *peek == ']') {
14352 has_terminating_bracket = TRUE;
14354 has_terminating_colon = TRUE;
14356 else if (*p == ';') {
14357 has_semi_colon = TRUE;
14358 has_terminating_colon = TRUE;
14361 found_problem = TRUE;
14368 /* Here we have punctuation we thought didn't end the class.
14369 * Keep track of the position of the key characters that are
14370 * more likely to have been class-enders */
14371 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
14373 /* Allow just one such possible class-ender not actually
14374 * ending the class. */
14375 if (possible_end) {
14381 /* If we have too many punctuation characters, no use in
14383 if (++punct_count > max_distance) {
14387 /* Treat the punctuation as a typo. */
14388 input_text[name_len++] = *p;
14391 else if (isUPPER(*p)) { /* Use lowercase for lookup */
14392 input_text[name_len++] = toLOWER(*p);
14394 found_problem = TRUE;
14396 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
14397 input_text[name_len++] = *p;
14401 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
14405 /* The declaration of 'input_text' is how long we allow a potential
14406 * class name to be, before saying they didn't mean a class name at
14408 if (name_len >= C_ARRAY_LENGTH(input_text)) {
14413 /* We get to here when the possible class name hasn't been properly
14414 * terminated before:
14415 * 1) we ran off the end of the pattern; or
14416 * 2) found two characters, each of which might have been intended to
14417 * be the name's terminator
14418 * 3) found so many punctuation characters in the purported name,
14419 * that the edit distance to a valid one is exceeded
14420 * 4) we decided it was more characters than anyone could have
14421 * intended to be one. */
14423 found_problem = TRUE;
14425 /* In the final two cases, we know that looking up what we've
14426 * accumulated won't lead to a match, even a fuzzy one. */
14427 if ( name_len >= C_ARRAY_LENGTH(input_text)
14428 || punct_count > max_distance)
14430 /* If there was an intermediate key character that could have been
14431 * an intended end, redo the parse, but stop there */
14432 if (possible_end && possible_end != (char *) -1) {
14433 possible_end = (char *) -1; /* Special signal value to say
14434 we've done a first pass */
14439 /* Otherwise, it can't have meant to have been a class */
14440 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14443 /* If we ran off the end, and the final character was a punctuation
14444 * one, back up one, to look at that final one just below. Later, we
14445 * will restore the parse pointer if appropriate */
14446 if (name_len && p == e && isPUNCT(*(p-1))) {
14451 if (p < e && isPUNCT(*p)) {
14453 has_terminating_bracket = TRUE;
14455 /* If this is a 2nd ']', and the first one is just below this
14456 * one, consider that to be the real terminator. This gives a
14457 * uniform and better positioning for the warning message */
14459 && possible_end != (char *) -1
14460 && *possible_end == ']'
14461 && name_len && input_text[name_len - 1] == ']')
14466 /* And this is actually equivalent to having done the 2nd
14467 * pass now, so set it to not try again */
14468 possible_end = (char *) -1;
14473 has_terminating_colon = TRUE;
14475 else if (*p == ';') {
14476 has_semi_colon = TRUE;
14477 has_terminating_colon = TRUE;
14485 /* Here, we have a class name to look up. We can short circuit the
14486 * stuff below for short names that can't possibly be meant to be a
14487 * class name. (We can do this on the first pass, as any second pass
14488 * will yield an even shorter name) */
14489 if (name_len < 3) {
14490 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14493 /* Find which class it is. Initially switch on the length of the name.
14495 switch (name_len) {
14497 if (memEQ(name_start, "word", 4)) {
14498 /* this is not POSIX, this is the Perl \w */
14499 class_number = ANYOF_WORDCHAR;
14503 /* Names all of length 5: alnum alpha ascii blank cntrl digit
14504 * graph lower print punct space upper
14505 * Offset 4 gives the best switch position. */
14506 switch (name_start[4]) {
14508 if (memEQ(name_start, "alph", 4)) /* alpha */
14509 class_number = ANYOF_ALPHA;
14512 if (memEQ(name_start, "spac", 4)) /* space */
14513 class_number = ANYOF_SPACE;
14516 if (memEQ(name_start, "grap", 4)) /* graph */
14517 class_number = ANYOF_GRAPH;
14520 if (memEQ(name_start, "asci", 4)) /* ascii */
14521 class_number = ANYOF_ASCII;
14524 if (memEQ(name_start, "blan", 4)) /* blank */
14525 class_number = ANYOF_BLANK;
14528 if (memEQ(name_start, "cntr", 4)) /* cntrl */
14529 class_number = ANYOF_CNTRL;
14532 if (memEQ(name_start, "alnu", 4)) /* alnum */
14533 class_number = ANYOF_ALPHANUMERIC;
14536 if (memEQ(name_start, "lowe", 4)) /* lower */
14537 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
14538 else if (memEQ(name_start, "uppe", 4)) /* upper */
14539 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
14542 if (memEQ(name_start, "digi", 4)) /* digit */
14543 class_number = ANYOF_DIGIT;
14544 else if (memEQ(name_start, "prin", 4)) /* print */
14545 class_number = ANYOF_PRINT;
14546 else if (memEQ(name_start, "punc", 4)) /* punct */
14547 class_number = ANYOF_PUNCT;
14552 if (memEQ(name_start, "xdigit", 6))
14553 class_number = ANYOF_XDIGIT;
14557 /* If the name exactly matches a posix class name the class number will
14558 * here be set to it, and the input almost certainly was meant to be a
14559 * posix class, so we can skip further checking. If instead the syntax
14560 * is exactly correct, but the name isn't one of the legal ones, we
14561 * will return that as an error below. But if neither of these apply,
14562 * it could be that no posix class was intended at all, or that one
14563 * was, but there was a typo. We tease these apart by doing fuzzy
14564 * matching on the name */
14565 if (class_number == OOB_NAMEDCLASS && found_problem) {
14566 const UV posix_names[][6] = {
14567 { 'a', 'l', 'n', 'u', 'm' },
14568 { 'a', 'l', 'p', 'h', 'a' },
14569 { 'a', 's', 'c', 'i', 'i' },
14570 { 'b', 'l', 'a', 'n', 'k' },
14571 { 'c', 'n', 't', 'r', 'l' },
14572 { 'd', 'i', 'g', 'i', 't' },
14573 { 'g', 'r', 'a', 'p', 'h' },
14574 { 'l', 'o', 'w', 'e', 'r' },
14575 { 'p', 'r', 'i', 'n', 't' },
14576 { 'p', 'u', 'n', 'c', 't' },
14577 { 's', 'p', 'a', 'c', 'e' },
14578 { 'u', 'p', 'p', 'e', 'r' },
14579 { 'w', 'o', 'r', 'd' },
14580 { 'x', 'd', 'i', 'g', 'i', 't' }
14582 /* The names of the above all have added NULs to make them the same
14583 * size, so we need to also have the real lengths */
14584 const UV posix_name_lengths[] = {
14585 sizeof("alnum") - 1,
14586 sizeof("alpha") - 1,
14587 sizeof("ascii") - 1,
14588 sizeof("blank") - 1,
14589 sizeof("cntrl") - 1,
14590 sizeof("digit") - 1,
14591 sizeof("graph") - 1,
14592 sizeof("lower") - 1,
14593 sizeof("print") - 1,
14594 sizeof("punct") - 1,
14595 sizeof("space") - 1,
14596 sizeof("upper") - 1,
14597 sizeof("word") - 1,
14598 sizeof("xdigit")- 1
14601 int temp_max = max_distance; /* Use a temporary, so if we
14602 reparse, we haven't changed the
14605 /* Use a smaller max edit distance if we are missing one of the
14607 if ( has_opening_bracket + has_opening_colon < 2
14608 || has_terminating_bracket + has_terminating_colon < 2)
14613 /* See if the input name is close to a legal one */
14614 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
14616 /* Short circuit call if the lengths are too far apart to be
14618 if (abs( (int) (name_len - posix_name_lengths[i]))
14624 if (edit_distance(input_text,
14627 posix_name_lengths[i],
14631 { /* If it is close, it probably was intended to be a class */
14632 goto probably_meant_to_be;
14636 /* Here the input name is not close enough to a valid class name
14637 * for us to consider it to be intended to be a posix class. If
14638 * we haven't already done so, and the parse found a character that
14639 * could have been terminators for the name, but which we absorbed
14640 * as typos during the first pass, repeat the parse, signalling it
14641 * to stop at that character */
14642 if (possible_end && possible_end != (char *) -1) {
14643 possible_end = (char *) -1;
14648 /* Here neither pass found a close-enough class name */
14649 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14652 probably_meant_to_be:
14654 /* Here we think that a posix specification was intended. Update any
14656 if (updated_parse_ptr) {
14657 *updated_parse_ptr = (char *) p;
14660 /* If a posix class name was intended but incorrectly specified, we
14661 * output or return the warnings */
14662 if (found_problem) {
14664 /* We set flags for these issues in the parse loop above instead of
14665 * adding them to the list of warnings, because we can parse it
14666 * twice, and we only want one warning instance */
14668 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
14671 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14673 if (has_semi_colon) {
14674 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14676 else if (! has_terminating_colon) {
14677 ADD_POSIX_WARNING(p, "there is no terminating ':'");
14679 if (! has_terminating_bracket) {
14680 ADD_POSIX_WARNING(p, "there is no terminating ']'");
14683 if (posix_warnings && RExC_warn_text && av_top_index(RExC_warn_text) > -1) {
14684 *posix_warnings = RExC_warn_text;
14687 else if (class_number != OOB_NAMEDCLASS) {
14688 /* If it is a known class, return the class. The class number
14689 * #defines are structured so each complement is +1 to the normal
14691 return class_number + complement;
14693 else if (! check_only) {
14695 /* Here, it is an unrecognized class. This is an error (unless the
14696 * call is to check only, which we've already handled above) */
14697 const char * const complement_string = (complement)
14700 RExC_parse = (char *) p;
14701 vFAIL3utf8f("POSIX class [:%s%"UTF8f":] unknown",
14703 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
14707 return OOB_NAMEDCLASS;
14709 #undef ADD_POSIX_WARNING
14711 STATIC unsigned int
14712 S_regex_set_precedence(const U8 my_operator) {
14714 /* Returns the precedence in the (?[...]) construct of the input operator,
14715 * specified by its character representation. The precedence follows
14716 * general Perl rules, but it extends this so that ')' and ']' have (low)
14717 * precedence even though they aren't really operators */
14719 switch (my_operator) {
14735 NOT_REACHED; /* NOTREACHED */
14736 return 0; /* Silence compiler warning */
14740 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
14741 I32 *flagp, U32 depth,
14742 char * const oregcomp_parse)
14744 /* Handle the (?[...]) construct to do set operations */
14746 U8 curchar; /* Current character being parsed */
14747 UV start, end; /* End points of code point ranges */
14748 SV* final = NULL; /* The end result inversion list */
14749 SV* result_string; /* 'final' stringified */
14750 AV* stack; /* stack of operators and operands not yet
14752 AV* fence_stack = NULL; /* A stack containing the positions in
14753 'stack' of where the undealt-with left
14754 parens would be if they were actually
14756 /* The 'VOL' (expanding to 'volatile') is a workaround for an optimiser bug
14757 * in Solaris Studio 12.3. See RT #127455 */
14758 VOL IV fence = 0; /* Position of where most recent undealt-
14759 with left paren in stack is; -1 if none.
14761 STRLEN len; /* Temporary */
14762 regnode* node; /* Temporary, and final regnode returned by
14764 const bool save_fold = FOLD; /* Temporary */
14765 char *save_end, *save_parse; /* Temporaries */
14766 const bool in_locale = LOC; /* we turn off /l during processing */
14767 AV* posix_warnings = NULL;
14769 GET_RE_DEBUG_FLAGS_DECL;
14771 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
14774 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
14777 REQUIRE_UNI_RULES(flagp, NULL); /* The use of this operator implies /u.
14778 This is required so that the compile
14779 time values are valid in all runtime
14782 /* This will return only an ANYOF regnode, or (unlikely) something smaller
14783 * (such as EXACT). Thus we can skip most everything if just sizing. We
14784 * call regclass to handle '[]' so as to not have to reinvent its parsing
14785 * rules here (throwing away the size it computes each time). And, we exit
14786 * upon an unescaped ']' that isn't one ending a regclass. To do both
14787 * these things, we need to realize that something preceded by a backslash
14788 * is escaped, so we have to keep track of backslashes */
14790 UV depth = 0; /* how many nested (?[...]) constructs */
14792 while (RExC_parse < RExC_end) {
14793 SV* current = NULL;
14795 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14796 TRUE /* Force /x */ );
14798 switch (*RExC_parse) {
14800 if (RExC_parse[1] == '[') depth++, RExC_parse++;
14805 /* Skip past this, so the next character gets skipped, after
14808 if (*RExC_parse == 'c') {
14809 /* Skip the \cX notation for control characters */
14810 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14816 /* See if this is a [:posix:] class. */
14817 bool is_posix_class = (OOB_NAMEDCLASS
14818 < handle_possible_posix(pRExC_state,
14822 TRUE /* checking only */));
14823 /* If it is a posix class, leave the parse pointer at the
14824 * '[' to fool regclass() into thinking it is part of a
14825 * '[[:posix:]]'. */
14826 if (! is_posix_class) {
14830 /* regclass() can only return RESTART_PASS1 and NEED_UTF8
14831 * if multi-char folds are allowed. */
14832 if (!regclass(pRExC_state, flagp,depth+1,
14833 is_posix_class, /* parse the whole char
14834 class only if not a
14836 FALSE, /* don't allow multi-char folds */
14837 TRUE, /* silence non-portable warnings. */
14839 FALSE, /* Require return to be an ANYOF */
14843 FAIL2("panic: regclass returned NULL to handle_sets, "
14844 "flags=%#"UVxf"", (UV) *flagp);
14846 /* function call leaves parse pointing to the ']', except
14847 * if we faked it */
14848 if (is_posix_class) {
14852 SvREFCNT_dec(current); /* In case it returned something */
14857 if (depth--) break;
14859 if (*RExC_parse == ')') {
14860 node = reganode(pRExC_state, ANYOF, 0);
14861 RExC_size += ANYOF_SKIP;
14862 nextchar(pRExC_state);
14863 Set_Node_Length(node,
14864 RExC_parse - oregcomp_parse + 1); /* MJD */
14866 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
14874 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14878 /* We output the messages even if warnings are off, because we'll fail
14879 * the very next thing, and these give a likely diagnosis for that */
14880 if (posix_warnings && av_tindex_nomg(posix_warnings) >= 0) {
14881 output_or_return_posix_warnings(pRExC_state, posix_warnings, NULL);
14884 FAIL("Syntax error in (?[...])");
14887 /* Pass 2 only after this. */
14888 Perl_ck_warner_d(aTHX_
14889 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
14890 "The regex_sets feature is experimental" REPORT_LOCATION,
14891 REPORT_LOCATION_ARGS(RExC_parse));
14893 /* Everything in this construct is a metacharacter. Operands begin with
14894 * either a '\' (for an escape sequence), or a '[' for a bracketed
14895 * character class. Any other character should be an operator, or
14896 * parenthesis for grouping. Both types of operands are handled by calling
14897 * regclass() to parse them. It is called with a parameter to indicate to
14898 * return the computed inversion list. The parsing here is implemented via
14899 * a stack. Each entry on the stack is a single character representing one
14900 * of the operators; or else a pointer to an operand inversion list. */
14902 #define IS_OPERATOR(a) SvIOK(a)
14903 #define IS_OPERAND(a) (! IS_OPERATOR(a))
14905 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
14906 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
14907 * with pronouncing it called it Reverse Polish instead, but now that YOU
14908 * know how to pronounce it you can use the correct term, thus giving due
14909 * credit to the person who invented it, and impressing your geek friends.
14910 * Wikipedia says that the pronounciation of "Ł" has been changing so that
14911 * it is now more like an English initial W (as in wonk) than an L.)
14913 * This means that, for example, 'a | b & c' is stored on the stack as
14921 * where the numbers in brackets give the stack [array] element number.
14922 * In this implementation, parentheses are not stored on the stack.
14923 * Instead a '(' creates a "fence" so that the part of the stack below the
14924 * fence is invisible except to the corresponding ')' (this allows us to
14925 * replace testing for parens, by using instead subtraction of the fence
14926 * position). As new operands are processed they are pushed onto the stack
14927 * (except as noted in the next paragraph). New operators of higher
14928 * precedence than the current final one are inserted on the stack before
14929 * the lhs operand (so that when the rhs is pushed next, everything will be
14930 * in the correct positions shown above. When an operator of equal or
14931 * lower precedence is encountered in parsing, all the stacked operations
14932 * of equal or higher precedence are evaluated, leaving the result as the
14933 * top entry on the stack. This makes higher precedence operations
14934 * evaluate before lower precedence ones, and causes operations of equal
14935 * precedence to left associate.
14937 * The only unary operator '!' is immediately pushed onto the stack when
14938 * encountered. When an operand is encountered, if the top of the stack is
14939 * a '!", the complement is immediately performed, and the '!' popped. The
14940 * resulting value is treated as a new operand, and the logic in the
14941 * previous paragraph is executed. Thus in the expression
14943 * the stack looks like
14949 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
14956 * A ')' is treated as an operator with lower precedence than all the
14957 * aforementioned ones, which causes all operations on the stack above the
14958 * corresponding '(' to be evaluated down to a single resultant operand.
14959 * Then the fence for the '(' is removed, and the operand goes through the
14960 * algorithm above, without the fence.
14962 * A separate stack is kept of the fence positions, so that the position of
14963 * the latest so-far unbalanced '(' is at the top of it.
14965 * The ']' ending the construct is treated as the lowest operator of all,
14966 * so that everything gets evaluated down to a single operand, which is the
14969 sv_2mortal((SV *)(stack = newAV()));
14970 sv_2mortal((SV *)(fence_stack = newAV()));
14972 while (RExC_parse < RExC_end) {
14973 I32 top_index; /* Index of top-most element in 'stack' */
14974 SV** top_ptr; /* Pointer to top 'stack' element */
14975 SV* current = NULL; /* To contain the current inversion list
14977 SV* only_to_avoid_leaks;
14979 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14980 TRUE /* Force /x */ );
14981 if (RExC_parse >= RExC_end) {
14982 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
14985 curchar = UCHARAT(RExC_parse);
14989 #ifdef ENABLE_REGEX_SETS_DEBUGGING
14990 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
14991 DEBUG_U(dump_regex_sets_structures(pRExC_state,
14992 stack, fence, fence_stack));
14995 top_index = av_tindex_nomg(stack);
14998 SV** stacked_ptr; /* Ptr to something already on 'stack' */
14999 char stacked_operator; /* The topmost operator on the 'stack'. */
15000 SV* lhs; /* Operand to the left of the operator */
15001 SV* rhs; /* Operand to the right of the operator */
15002 SV* fence_ptr; /* Pointer to top element of the fence
15007 if ( RExC_parse < RExC_end - 1
15008 && (UCHARAT(RExC_parse + 1) == '?'))
15010 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
15011 * This happens when we have some thing like
15013 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
15015 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
15017 * Here we would be handling the interpolated
15018 * '$thai_or_lao'. We handle this by a recursive call to
15019 * ourselves which returns the inversion list the
15020 * interpolated expression evaluates to. We use the flags
15021 * from the interpolated pattern. */
15022 U32 save_flags = RExC_flags;
15023 const char * save_parse;
15025 RExC_parse += 2; /* Skip past the '(?' */
15026 save_parse = RExC_parse;
15028 /* Parse any flags for the '(?' */
15029 parse_lparen_question_flags(pRExC_state);
15031 if (RExC_parse == save_parse /* Makes sure there was at
15032 least one flag (or else
15033 this embedding wasn't
15035 || RExC_parse >= RExC_end - 4
15036 || UCHARAT(RExC_parse) != ':'
15037 || UCHARAT(++RExC_parse) != '('
15038 || UCHARAT(++RExC_parse) != '?'
15039 || UCHARAT(++RExC_parse) != '[')
15042 /* In combination with the above, this moves the
15043 * pointer to the point just after the first erroneous
15044 * character (or if there are no flags, to where they
15045 * should have been) */
15046 if (RExC_parse >= RExC_end - 4) {
15047 RExC_parse = RExC_end;
15049 else if (RExC_parse != save_parse) {
15050 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15052 vFAIL("Expecting '(?flags:(?[...'");
15055 /* Recurse, with the meat of the embedded expression */
15057 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
15058 depth+1, oregcomp_parse);
15060 /* Here, 'current' contains the embedded expression's
15061 * inversion list, and RExC_parse points to the trailing
15062 * ']'; the next character should be the ')' */
15064 assert(UCHARAT(RExC_parse) == ')');
15066 /* Then the ')' matching the original '(' handled by this
15067 * case: statement */
15069 assert(UCHARAT(RExC_parse) == ')');
15072 RExC_flags = save_flags;
15073 goto handle_operand;
15076 /* A regular '('. Look behind for illegal syntax */
15077 if (top_index - fence >= 0) {
15078 /* If the top entry on the stack is an operator, it had
15079 * better be a '!', otherwise the entry below the top
15080 * operand should be an operator */
15081 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
15082 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
15083 || ( IS_OPERAND(*top_ptr)
15084 && ( top_index - fence < 1
15085 || ! (stacked_ptr = av_fetch(stack,
15088 || ! IS_OPERATOR(*stacked_ptr))))
15091 vFAIL("Unexpected '(' with no preceding operator");
15095 /* Stack the position of this undealt-with left paren */
15096 av_push(fence_stack, newSViv(fence));
15097 fence = top_index + 1;
15101 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15102 * multi-char folds are allowed. */
15103 if (!regclass(pRExC_state, flagp,depth+1,
15104 TRUE, /* means parse just the next thing */
15105 FALSE, /* don't allow multi-char folds */
15106 FALSE, /* don't silence non-portable warnings. */
15108 FALSE, /* Require return to be an ANYOF */
15112 FAIL2("panic: regclass returned NULL to handle_sets, "
15113 "flags=%#"UVxf"", (UV) *flagp);
15116 /* regclass() will return with parsing just the \ sequence,
15117 * leaving the parse pointer at the next thing to parse */
15119 goto handle_operand;
15121 case '[': /* Is a bracketed character class */
15123 /* See if this is a [:posix:] class. */
15124 bool is_posix_class = (OOB_NAMEDCLASS
15125 < handle_possible_posix(pRExC_state,
15129 TRUE /* checking only */));
15130 /* If it is a posix class, leave the parse pointer at the '['
15131 * to fool regclass() into thinking it is part of a
15132 * '[[:posix:]]'. */
15133 if (! is_posix_class) {
15137 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15138 * multi-char folds are allowed. */
15139 if (!regclass(pRExC_state, flagp,depth+1,
15140 is_posix_class, /* parse the whole char
15141 class only if not a
15143 FALSE, /* don't allow multi-char folds */
15144 TRUE, /* silence non-portable warnings. */
15146 FALSE, /* Require return to be an ANYOF */
15151 FAIL2("panic: regclass returned NULL to handle_sets, "
15152 "flags=%#"UVxf"", (UV) *flagp);
15155 /* function call leaves parse pointing to the ']', except if we
15157 if (is_posix_class) {
15161 goto handle_operand;
15165 if (top_index >= 1) {
15166 goto join_operators;
15169 /* Only a single operand on the stack: are done */
15173 if (av_tindex_nomg(fence_stack) < 0) {
15175 vFAIL("Unexpected ')'");
15178 /* If nothing after the fence, is missing an operand */
15179 if (top_index - fence < 0) {
15183 /* If at least two things on the stack, treat this as an
15185 if (top_index - fence >= 1) {
15186 goto join_operators;
15189 /* Here only a single thing on the fenced stack, and there is a
15190 * fence. Get rid of it */
15191 fence_ptr = av_pop(fence_stack);
15193 fence = SvIV(fence_ptr) - 1;
15194 SvREFCNT_dec_NN(fence_ptr);
15201 /* Having gotten rid of the fence, we pop the operand at the
15202 * stack top and process it as a newly encountered operand */
15203 current = av_pop(stack);
15204 if (IS_OPERAND(current)) {
15205 goto handle_operand;
15217 /* These binary operators should have a left operand already
15219 if ( top_index - fence < 0
15220 || top_index - fence == 1
15221 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
15222 || ! IS_OPERAND(*top_ptr))
15224 goto unexpected_binary;
15227 /* If only the one operand is on the part of the stack visible
15228 * to us, we just place this operator in the proper position */
15229 if (top_index - fence < 2) {
15231 /* Place the operator before the operand */
15233 SV* lhs = av_pop(stack);
15234 av_push(stack, newSVuv(curchar));
15235 av_push(stack, lhs);
15239 /* But if there is something else on the stack, we need to
15240 * process it before this new operator if and only if the
15241 * stacked operation has equal or higher precedence than the
15246 /* The operator on the stack is supposed to be below both its
15248 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
15249 || IS_OPERAND(*stacked_ptr))
15251 /* But if not, it's legal and indicates we are completely
15252 * done if and only if we're currently processing a ']',
15253 * which should be the final thing in the expression */
15254 if (curchar == ']') {
15260 vFAIL2("Unexpected binary operator '%c' with no "
15261 "preceding operand", curchar);
15263 stacked_operator = (char) SvUV(*stacked_ptr);
15265 if (regex_set_precedence(curchar)
15266 > regex_set_precedence(stacked_operator))
15268 /* Here, the new operator has higher precedence than the
15269 * stacked one. This means we need to add the new one to
15270 * the stack to await its rhs operand (and maybe more
15271 * stuff). We put it before the lhs operand, leaving
15272 * untouched the stacked operator and everything below it
15274 lhs = av_pop(stack);
15275 assert(IS_OPERAND(lhs));
15277 av_push(stack, newSVuv(curchar));
15278 av_push(stack, lhs);
15282 /* Here, the new operator has equal or lower precedence than
15283 * what's already there. This means the operation already
15284 * there should be performed now, before the new one. */
15286 rhs = av_pop(stack);
15287 if (! IS_OPERAND(rhs)) {
15289 /* This can happen when a ! is not followed by an operand,
15290 * like in /(?[\t &!])/ */
15294 lhs = av_pop(stack);
15296 if (! IS_OPERAND(lhs)) {
15298 /* This can happen when there is an empty (), like in
15299 * /(?[[0]+()+])/ */
15303 switch (stacked_operator) {
15305 _invlist_intersection(lhs, rhs, &rhs);
15310 _invlist_union(lhs, rhs, &rhs);
15314 _invlist_subtract(lhs, rhs, &rhs);
15317 case '^': /* The union minus the intersection */
15323 _invlist_union(lhs, rhs, &u);
15324 _invlist_intersection(lhs, rhs, &i);
15325 /* _invlist_subtract will overwrite rhs
15326 without freeing what it already contains */
15328 _invlist_subtract(u, i, &rhs);
15329 SvREFCNT_dec_NN(i);
15330 SvREFCNT_dec_NN(u);
15331 SvREFCNT_dec_NN(element);
15337 /* Here, the higher precedence operation has been done, and the
15338 * result is in 'rhs'. We overwrite the stacked operator with
15339 * the result. Then we redo this code to either push the new
15340 * operator onto the stack or perform any higher precedence
15341 * stacked operation */
15342 only_to_avoid_leaks = av_pop(stack);
15343 SvREFCNT_dec(only_to_avoid_leaks);
15344 av_push(stack, rhs);
15347 case '!': /* Highest priority, right associative */
15349 /* If what's already at the top of the stack is another '!",
15350 * they just cancel each other out */
15351 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
15352 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
15354 only_to_avoid_leaks = av_pop(stack);
15355 SvREFCNT_dec(only_to_avoid_leaks);
15357 else { /* Otherwise, since it's right associative, just push
15359 av_push(stack, newSVuv(curchar));
15364 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15365 vFAIL("Unexpected character");
15369 /* Here 'current' is the operand. If something is already on the
15370 * stack, we have to check if it is a !. But first, the code above
15371 * may have altered the stack in the time since we earlier set
15374 top_index = av_tindex_nomg(stack);
15375 if (top_index - fence >= 0) {
15376 /* If the top entry on the stack is an operator, it had better
15377 * be a '!', otherwise the entry below the top operand should
15378 * be an operator */
15379 top_ptr = av_fetch(stack, top_index, FALSE);
15381 if (IS_OPERATOR(*top_ptr)) {
15383 /* The only permissible operator at the top of the stack is
15384 * '!', which is applied immediately to this operand. */
15385 curchar = (char) SvUV(*top_ptr);
15386 if (curchar != '!') {
15387 SvREFCNT_dec(current);
15388 vFAIL2("Unexpected binary operator '%c' with no "
15389 "preceding operand", curchar);
15392 _invlist_invert(current);
15394 only_to_avoid_leaks = av_pop(stack);
15395 SvREFCNT_dec(only_to_avoid_leaks);
15397 /* And we redo with the inverted operand. This allows
15398 * handling multiple ! in a row */
15399 goto handle_operand;
15401 /* Single operand is ok only for the non-binary ')'
15403 else if ((top_index - fence == 0 && curchar != ')')
15404 || (top_index - fence > 0
15405 && (! (stacked_ptr = av_fetch(stack,
15408 || IS_OPERAND(*stacked_ptr))))
15410 SvREFCNT_dec(current);
15411 vFAIL("Operand with no preceding operator");
15415 /* Here there was nothing on the stack or the top element was
15416 * another operand. Just add this new one */
15417 av_push(stack, current);
15419 } /* End of switch on next parse token */
15421 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15422 } /* End of loop parsing through the construct */
15425 if (av_tindex_nomg(fence_stack) >= 0) {
15426 vFAIL("Unmatched (");
15429 if (av_tindex_nomg(stack) < 0 /* Was empty */
15430 || ((final = av_pop(stack)) == NULL)
15431 || ! IS_OPERAND(final)
15432 || SvTYPE(final) != SVt_INVLIST
15433 || av_tindex_nomg(stack) >= 0) /* More left on stack */
15436 SvREFCNT_dec(final);
15437 vFAIL("Incomplete expression within '(?[ ])'");
15440 /* Here, 'final' is the resultant inversion list from evaluating the
15441 * expression. Return it if so requested */
15442 if (return_invlist) {
15443 *return_invlist = final;
15447 /* Otherwise generate a resultant node, based on 'final'. regclass() is
15448 * expecting a string of ranges and individual code points */
15449 invlist_iterinit(final);
15450 result_string = newSVpvs("");
15451 while (invlist_iternext(final, &start, &end)) {
15452 if (start == end) {
15453 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
15456 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
15461 /* About to generate an ANYOF (or similar) node from the inversion list we
15462 * have calculated */
15463 save_parse = RExC_parse;
15464 RExC_parse = SvPV(result_string, len);
15465 save_end = RExC_end;
15466 RExC_end = RExC_parse + len;
15468 /* We turn off folding around the call, as the class we have constructed
15469 * already has all folding taken into consideration, and we don't want
15470 * regclass() to add to that */
15471 RExC_flags &= ~RXf_PMf_FOLD;
15472 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if multi-char
15473 * folds are allowed. */
15474 node = regclass(pRExC_state, flagp,depth+1,
15475 FALSE, /* means parse the whole char class */
15476 FALSE, /* don't allow multi-char folds */
15477 TRUE, /* silence non-portable warnings. The above may very
15478 well have generated non-portable code points, but
15479 they're valid on this machine */
15480 FALSE, /* similarly, no need for strict */
15481 FALSE, /* Require return to be an ANYOF */
15486 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
15489 /* Fix up the node type if we are in locale. (We have pretended we are
15490 * under /u for the purposes of regclass(), as this construct will only
15491 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
15492 * as to cause any warnings about bad locales to be output in regexec.c),
15493 * and add the flag that indicates to check if not in a UTF-8 locale. The
15494 * reason we above forbid optimization into something other than an ANYOF
15495 * node is simply to minimize the number of code changes in regexec.c.
15496 * Otherwise we would have to create new EXACTish node types and deal with
15497 * them. This decision could be revisited should this construct become
15500 * (One might think we could look at the resulting ANYOF node and suppress
15501 * the flag if everything is above 255, as those would be UTF-8 only,
15502 * but this isn't true, as the components that led to that result could
15503 * have been locale-affected, and just happen to cancel each other out
15504 * under UTF-8 locales.) */
15506 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
15508 assert(OP(node) == ANYOF);
15512 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
15516 RExC_flags |= RXf_PMf_FOLD;
15519 RExC_parse = save_parse + 1;
15520 RExC_end = save_end;
15521 SvREFCNT_dec_NN(final);
15522 SvREFCNT_dec_NN(result_string);
15524 nextchar(pRExC_state);
15525 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
15529 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15532 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
15533 AV * stack, const IV fence, AV * fence_stack)
15534 { /* Dumps the stacks in handle_regex_sets() */
15536 const SSize_t stack_top = av_tindex_nomg(stack);
15537 const SSize_t fence_stack_top = av_tindex_nomg(fence_stack);
15540 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
15542 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
15544 if (stack_top < 0) {
15545 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
15548 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
15549 for (i = stack_top; i >= 0; i--) {
15550 SV ** element_ptr = av_fetch(stack, i, FALSE);
15551 if (! element_ptr) {
15554 if (IS_OPERATOR(*element_ptr)) {
15555 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
15556 (int) i, (int) SvIV(*element_ptr));
15559 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
15560 sv_dump(*element_ptr);
15565 if (fence_stack_top < 0) {
15566 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
15569 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
15570 for (i = fence_stack_top; i >= 0; i--) {
15571 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
15572 if (! element_ptr) {
15575 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
15576 (int) i, (int) SvIV(*element_ptr));
15587 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
15589 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
15590 * innocent-looking character class, like /[ks]/i won't have to go out to
15591 * disk to find the possible matches.
15593 * This should be called only for a Latin1-range code points, cp, which is
15594 * known to be involved in a simple fold with other code points above
15595 * Latin1. It would give false results if /aa has been specified.
15596 * Multi-char folds are outside the scope of this, and must be handled
15599 * XXX It would be better to generate these via regen, in case a new
15600 * version of the Unicode standard adds new mappings, though that is not
15601 * really likely, and may be caught by the default: case of the switch
15604 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
15606 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
15612 add_cp_to_invlist(*invlist, KELVIN_SIGN);
15616 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
15619 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
15620 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
15622 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
15623 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
15624 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
15626 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
15627 *invlist = add_cp_to_invlist(*invlist,
15628 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
15631 #ifdef LATIN_CAPITAL_LETTER_SHARP_S /* not defined in early Unicode releases */
15633 case LATIN_SMALL_LETTER_SHARP_S:
15634 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
15639 #if UNICODE_MAJOR_VERSION < 3 \
15640 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0)
15642 /* In 3.0 and earlier, U+0130 folded simply to 'i'; and in 3.0.1 so did
15647 add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
15648 # if UNICODE_DOT_DOT_VERSION == 1
15649 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_DOTLESS_I);
15655 /* Use deprecated warning to increase the chances of this being
15658 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
15665 S_output_or_return_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings, AV** return_posix_warnings)
15667 /* If the final parameter is NULL, output the elements of the array given
15668 * by '*posix_warnings' as REGEXP warnings. Otherwise, the elements are
15669 * pushed onto it, (creating if necessary) */
15672 const bool first_is_fatal = ! return_posix_warnings
15673 && ckDEAD(packWARN(WARN_REGEXP));
15675 PERL_ARGS_ASSERT_OUTPUT_OR_RETURN_POSIX_WARNINGS;
15677 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
15678 if (return_posix_warnings) {
15679 if (! *return_posix_warnings) { /* mortalize to not leak if
15680 warnings are fatal */
15681 *return_posix_warnings = (AV *) sv_2mortal((SV *) newAV());
15683 av_push(*return_posix_warnings, msg);
15686 if (first_is_fatal) { /* Avoid leaking this */
15687 av_undef(posix_warnings); /* This isn't necessary if the
15688 array is mortal, but is a
15690 (void) sv_2mortal(msg);
15692 SAVEFREESV(RExC_rx_sv);
15695 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
15696 SvREFCNT_dec_NN(msg);
15702 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
15704 /* This adds the string scalar <multi_string> to the array
15705 * <multi_char_matches>. <multi_string> is known to have exactly
15706 * <cp_count> code points in it. This is used when constructing a
15707 * bracketed character class and we find something that needs to match more
15708 * than a single character.
15710 * <multi_char_matches> is actually an array of arrays. Each top-level
15711 * element is an array that contains all the strings known so far that are
15712 * the same length. And that length (in number of code points) is the same
15713 * as the index of the top-level array. Hence, the [2] element is an
15714 * array, each element thereof is a string containing TWO code points;
15715 * while element [3] is for strings of THREE characters, and so on. Since
15716 * this is for multi-char strings there can never be a [0] nor [1] element.
15718 * When we rewrite the character class below, we will do so such that the
15719 * longest strings are written first, so that it prefers the longest
15720 * matching strings first. This is done even if it turns out that any
15721 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
15722 * Christiansen has agreed that this is ok. This makes the test for the
15723 * ligature 'ffi' come before the test for 'ff', for example */
15726 AV** this_array_ptr;
15728 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
15730 if (! multi_char_matches) {
15731 multi_char_matches = newAV();
15734 if (av_exists(multi_char_matches, cp_count)) {
15735 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
15736 this_array = *this_array_ptr;
15739 this_array = newAV();
15740 av_store(multi_char_matches, cp_count,
15743 av_push(this_array, multi_string);
15745 return multi_char_matches;
15748 /* The names of properties whose definitions are not known at compile time are
15749 * stored in this SV, after a constant heading. So if the length has been
15750 * changed since initialization, then there is a run-time definition. */
15751 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
15752 (SvCUR(listsv) != initial_listsv_len)
15754 /* There is a restricted set of white space characters that are legal when
15755 * ignoring white space in a bracketed character class. This generates the
15756 * code to skip them.
15758 * There is a line below that uses the same white space criteria but is outside
15759 * this macro. Both here and there must use the same definition */
15760 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
15763 while (isBLANK_A(UCHARAT(p))) \
15771 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
15772 const bool stop_at_1, /* Just parse the next thing, don't
15773 look for a full character class */
15774 bool allow_multi_folds,
15775 const bool silence_non_portable, /* Don't output warnings
15779 bool optimizable, /* ? Allow a non-ANYOF return
15781 SV** ret_invlist, /* Return an inversion list, not a node */
15782 AV** return_posix_warnings
15785 /* parse a bracketed class specification. Most of these will produce an
15786 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
15787 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
15788 * under /i with multi-character folds: it will be rewritten following the
15789 * paradigm of this example, where the <multi-fold>s are characters which
15790 * fold to multiple character sequences:
15791 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
15792 * gets effectively rewritten as:
15793 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
15794 * reg() gets called (recursively) on the rewritten version, and this
15795 * function will return what it constructs. (Actually the <multi-fold>s
15796 * aren't physically removed from the [abcdefghi], it's just that they are
15797 * ignored in the recursion by means of a flag:
15798 * <RExC_in_multi_char_class>.)
15800 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
15801 * characters, with the corresponding bit set if that character is in the
15802 * list. For characters above this, a range list or swash is used. There
15803 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
15804 * determinable at compile time
15806 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs
15807 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded
15808 * to UTF-8. This can only happen if ret_invlist is non-NULL.
15811 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
15813 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
15816 int namedclass = OOB_NAMEDCLASS;
15817 char *rangebegin = NULL;
15818 bool need_class = 0;
15820 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
15821 than just initialized. */
15822 SV* properties = NULL; /* Code points that match \p{} \P{} */
15823 SV* posixes = NULL; /* Code points that match classes like [:word:],
15824 extended beyond the Latin1 range. These have to
15825 be kept separate from other code points for much
15826 of this function because their handling is
15827 different under /i, and for most classes under
15829 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
15830 separate for a while from the non-complemented
15831 versions because of complications with /d
15833 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
15834 treated more simply than the general case,
15835 leading to less compilation and execution
15837 UV element_count = 0; /* Number of distinct elements in the class.
15838 Optimizations may be possible if this is tiny */
15839 AV * multi_char_matches = NULL; /* Code points that fold to more than one
15840 character; used under /i */
15842 char * stop_ptr = RExC_end; /* where to stop parsing */
15843 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
15846 /* Unicode properties are stored in a swash; this holds the current one
15847 * being parsed. If this swash is the only above-latin1 component of the
15848 * character class, an optimization is to pass it directly on to the
15849 * execution engine. Otherwise, it is set to NULL to indicate that there
15850 * are other things in the class that have to be dealt with at execution
15852 SV* swash = NULL; /* Code points that match \p{} \P{} */
15854 /* Set if a component of this character class is user-defined; just passed
15855 * on to the engine */
15856 bool has_user_defined_property = FALSE;
15858 /* inversion list of code points this node matches only when the target
15859 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
15861 SV* has_upper_latin1_only_utf8_matches = NULL;
15863 /* Inversion list of code points this node matches regardless of things
15864 * like locale, folding, utf8ness of the target string */
15865 SV* cp_list = NULL;
15867 /* Like cp_list, but code points on this list need to be checked for things
15868 * that fold to/from them under /i */
15869 SV* cp_foldable_list = NULL;
15871 /* Like cp_list, but code points on this list are valid only when the
15872 * runtime locale is UTF-8 */
15873 SV* only_utf8_locale_list = NULL;
15875 /* In a range, if one of the endpoints is non-character-set portable,
15876 * meaning that it hard-codes a code point that may mean a different
15877 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
15878 * mnemonic '\t' which each mean the same character no matter which
15879 * character set the platform is on. */
15880 unsigned int non_portable_endpoint = 0;
15882 /* Is the range unicode? which means on a platform that isn't 1-1 native
15883 * to Unicode (i.e. non-ASCII), each code point in it should be considered
15884 * to be a Unicode value. */
15885 bool unicode_range = FALSE;
15886 bool invert = FALSE; /* Is this class to be complemented */
15888 bool warn_super = ALWAYS_WARN_SUPER;
15890 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
15891 case we need to change the emitted regop to an EXACT. */
15892 const char * orig_parse = RExC_parse;
15893 const SSize_t orig_size = RExC_size;
15894 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
15896 /* This variable is used to mark where the end in the input is of something
15897 * that looks like a POSIX construct but isn't. During the parse, when
15898 * something looks like it could be such a construct is encountered, it is
15899 * checked for being one, but not if we've already checked this area of the
15900 * input. Only after this position is reached do we check again */
15901 char *not_posix_region_end = RExC_parse - 1;
15903 AV* posix_warnings = NULL;
15904 const bool do_posix_warnings = return_posix_warnings
15905 || (PASS2 && ckWARN(WARN_REGEXP));
15907 GET_RE_DEBUG_FLAGS_DECL;
15909 PERL_ARGS_ASSERT_REGCLASS;
15911 PERL_UNUSED_ARG(depth);
15914 DEBUG_PARSE("clas");
15916 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
15917 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
15918 && UNICODE_DOT_DOT_VERSION == 0)
15919 allow_multi_folds = FALSE;
15922 /* Assume we are going to generate an ANYOF node. */
15923 ret = reganode(pRExC_state,
15930 RExC_size += ANYOF_SKIP;
15931 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
15934 ANYOF_FLAGS(ret) = 0;
15936 RExC_emit += ANYOF_SKIP;
15937 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
15938 initial_listsv_len = SvCUR(listsv);
15939 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
15942 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15944 assert(RExC_parse <= RExC_end);
15946 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
15949 allow_multi_folds = FALSE;
15951 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15954 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
15955 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
15956 int maybe_class = handle_possible_posix(pRExC_state,
15958 ¬_posix_region_end,
15960 TRUE /* checking only */);
15961 if (PASS2 && maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
15962 SAVEFREESV(RExC_rx_sv);
15963 ckWARN4reg(not_posix_region_end,
15964 "POSIX syntax [%c %c] belongs inside character classes%s",
15965 *RExC_parse, *RExC_parse,
15966 (maybe_class == OOB_NAMEDCLASS)
15967 ? ((POSIXCC_NOTYET(*RExC_parse))
15968 ? " (but this one isn't implemented)"
15969 : " (but this one isn't fully valid)")
15972 (void)ReREFCNT_inc(RExC_rx_sv);
15976 /* If the caller wants us to just parse a single element, accomplish this
15977 * by faking the loop ending condition */
15978 if (stop_at_1 && RExC_end > RExC_parse) {
15979 stop_ptr = RExC_parse + 1;
15982 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
15983 if (UCHARAT(RExC_parse) == ']')
15984 goto charclassloop;
15988 if ( posix_warnings
15989 && av_tindex_nomg(posix_warnings) >= 0
15990 && RExC_parse > not_posix_region_end)
15992 /* Warnings about posix class issues are considered tentative until
15993 * we are far enough along in the parse that we can no longer
15994 * change our mind, at which point we either output them or add
15995 * them, if it has so specified, to what gets returned to the
15996 * caller. This is done each time through the loop so that a later
15997 * class won't zap them before they have been dealt with. */
15998 output_or_return_posix_warnings(pRExC_state, posix_warnings,
15999 return_posix_warnings);
16002 if (RExC_parse >= stop_ptr) {
16006 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16008 if (UCHARAT(RExC_parse) == ']') {
16014 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
16015 save_value = value;
16016 save_prevvalue = prevvalue;
16019 rangebegin = RExC_parse;
16021 non_portable_endpoint = 0;
16023 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
16024 value = utf8n_to_uvchr((U8*)RExC_parse,
16025 RExC_end - RExC_parse,
16026 &numlen, UTF8_ALLOW_DEFAULT);
16027 RExC_parse += numlen;
16030 value = UCHARAT(RExC_parse++);
16032 if (value == '[') {
16033 char * posix_class_end;
16034 namedclass = handle_possible_posix(pRExC_state,
16037 do_posix_warnings ? &posix_warnings : NULL,
16038 FALSE /* die if error */);
16039 if (namedclass > OOB_NAMEDCLASS) {
16041 /* If there was an earlier attempt to parse this particular
16042 * posix class, and it failed, it was a false alarm, as this
16043 * successful one proves */
16044 if ( posix_warnings
16045 && av_tindex_nomg(posix_warnings) >= 0
16046 && not_posix_region_end >= RExC_parse
16047 && not_posix_region_end <= posix_class_end)
16049 av_undef(posix_warnings);
16052 RExC_parse = posix_class_end;
16054 else if (namedclass == OOB_NAMEDCLASS) {
16055 not_posix_region_end = posix_class_end;
16058 namedclass = OOB_NAMEDCLASS;
16061 else if ( RExC_parse - 1 > not_posix_region_end
16062 && MAYBE_POSIXCC(value))
16064 (void) handle_possible_posix(
16066 RExC_parse - 1, /* -1 because parse has already been
16068 ¬_posix_region_end,
16069 do_posix_warnings ? &posix_warnings : NULL,
16070 TRUE /* checking only */);
16072 else if (value == '\\') {
16073 /* Is a backslash; get the code point of the char after it */
16075 if (RExC_parse >= RExC_end) {
16076 vFAIL("Unmatched [");
16079 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
16080 value = utf8n_to_uvchr((U8*)RExC_parse,
16081 RExC_end - RExC_parse,
16082 &numlen, UTF8_ALLOW_DEFAULT);
16083 RExC_parse += numlen;
16086 value = UCHARAT(RExC_parse++);
16088 /* Some compilers cannot handle switching on 64-bit integer
16089 * values, therefore value cannot be an UV. Yes, this will
16090 * be a problem later if we want switch on Unicode.
16091 * A similar issue a little bit later when switching on
16092 * namedclass. --jhi */
16094 /* If the \ is escaping white space when white space is being
16095 * skipped, it means that that white space is wanted literally, and
16096 * is already in 'value'. Otherwise, need to translate the escape
16097 * into what it signifies. */
16098 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
16100 case 'w': namedclass = ANYOF_WORDCHAR; break;
16101 case 'W': namedclass = ANYOF_NWORDCHAR; break;
16102 case 's': namedclass = ANYOF_SPACE; break;
16103 case 'S': namedclass = ANYOF_NSPACE; break;
16104 case 'd': namedclass = ANYOF_DIGIT; break;
16105 case 'D': namedclass = ANYOF_NDIGIT; break;
16106 case 'v': namedclass = ANYOF_VERTWS; break;
16107 case 'V': namedclass = ANYOF_NVERTWS; break;
16108 case 'h': namedclass = ANYOF_HORIZWS; break;
16109 case 'H': namedclass = ANYOF_NHORIZWS; break;
16110 case 'N': /* Handle \N{NAME} in class */
16112 const char * const backslash_N_beg = RExC_parse - 2;
16115 if (! grok_bslash_N(pRExC_state,
16116 NULL, /* No regnode */
16117 &value, /* Yes single value */
16118 &cp_count, /* Multiple code pt count */
16124 if (*flagp & NEED_UTF8)
16125 FAIL("panic: grok_bslash_N set NEED_UTF8");
16126 if (*flagp & RESTART_PASS1)
16129 if (cp_count < 0) {
16130 vFAIL("\\N in a character class must be a named character: \\N{...}");
16132 else if (cp_count == 0) {
16134 ckWARNreg(RExC_parse,
16135 "Ignoring zero length \\N{} in character class");
16138 else { /* cp_count > 1 */
16139 if (! RExC_in_multi_char_class) {
16140 if (invert || range || *RExC_parse == '-') {
16143 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
16146 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
16148 break; /* <value> contains the first code
16149 point. Drop out of the switch to
16153 SV * multi_char_N = newSVpvn(backslash_N_beg,
16154 RExC_parse - backslash_N_beg);
16156 = add_multi_match(multi_char_matches,
16161 } /* End of cp_count != 1 */
16163 /* This element should not be processed further in this
16166 value = save_value;
16167 prevvalue = save_prevvalue;
16168 continue; /* Back to top of loop to get next char */
16171 /* Here, is a single code point, and <value> contains it */
16172 unicode_range = TRUE; /* \N{} are Unicode */
16180 /* We will handle any undefined properties ourselves */
16181 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
16182 /* And we actually would prefer to get
16183 * the straight inversion list of the
16184 * swash, since we will be accessing it
16185 * anyway, to save a little time */
16186 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
16188 if (RExC_parse >= RExC_end)
16189 vFAIL2("Empty \\%c", (U8)value);
16190 if (*RExC_parse == '{') {
16191 const U8 c = (U8)value;
16192 e = strchr(RExC_parse, '}');
16195 vFAIL2("Missing right brace on \\%c{}", c);
16199 while (isSPACE(*RExC_parse)) {
16203 if (UCHARAT(RExC_parse) == '^') {
16205 /* toggle. (The rhs xor gets the single bit that
16206 * differs between P and p; the other xor inverts just
16208 value ^= 'P' ^ 'p';
16211 while (isSPACE(*RExC_parse)) {
16216 if (e == RExC_parse)
16217 vFAIL2("Empty \\%c{}", c);
16219 n = e - RExC_parse;
16220 while (isSPACE(*(RExC_parse + n - 1)))
16222 } /* The \p isn't immediately followed by a '{' */
16223 else if (! isALPHA(*RExC_parse)) {
16224 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16225 vFAIL2("Character following \\%c must be '{' or a "
16226 "single-character Unicode property name",
16236 char* base_name; /* name after any packages are stripped */
16237 char* lookup_name = NULL;
16238 const char * const colon_colon = "::";
16240 /* Try to get the definition of the property into
16241 * <invlist>. If /i is in effect, the effective property
16242 * will have its name be <__NAME_i>. The design is
16243 * discussed in commit
16244 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
16245 name = savepv(Perl_form(aTHX_ "%.*s", (int)n, RExC_parse));
16248 lookup_name = savepv(Perl_form(aTHX_ "__%s_i", name));
16250 /* The function call just below that uses this can fail
16251 * to return, leaking memory if we don't do this */
16252 SAVEFREEPV(lookup_name);
16255 /* Look up the property name, and get its swash and
16256 * inversion list, if the property is found */
16257 SvREFCNT_dec(swash); /* Free any left-overs */
16258 swash = _core_swash_init("utf8",
16265 NULL, /* No inversion list */
16268 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
16269 HV* curpkg = (IN_PERL_COMPILETIME)
16271 : CopSTASH(PL_curcop);
16275 if (swash) { /* Got a swash but no inversion list.
16276 Something is likely wrong that will
16277 be sorted-out later */
16278 SvREFCNT_dec_NN(swash);
16282 /* Here didn't find it. It could be a an error (like a
16283 * typo) in specifying a Unicode property, or it could
16284 * be a user-defined property that will be available at
16285 * run-time. The names of these must begin with 'In'
16286 * or 'Is' (after any packages are stripped off). So
16287 * if not one of those, or if we accept only
16288 * compile-time properties, is an error; otherwise add
16289 * it to the list for run-time look up. */
16290 if ((base_name = rninstr(name, name + n,
16291 colon_colon, colon_colon + 2)))
16292 { /* Has ::. We know this must be a user-defined
16295 final_n -= base_name - name;
16304 || base_name[0] != 'I'
16305 || (base_name[1] != 's' && base_name[1] != 'n')
16308 const char * const msg
16310 ? "Illegal user-defined property name"
16311 : "Can't find Unicode property definition";
16312 RExC_parse = e + 1;
16314 /* diag_listed_as: Can't find Unicode property definition "%s" */
16315 vFAIL3utf8f("%s \"%"UTF8f"\"",
16316 msg, UTF8fARG(UTF, n, name));
16319 /* If the property name doesn't already have a package
16320 * name, add the current one to it so that it can be
16321 * referred to outside it. [perl #121777] */
16322 if (! has_pkg && curpkg) {
16323 char* pkgname = HvNAME(curpkg);
16324 if (strNE(pkgname, "main")) {
16325 char* full_name = Perl_form(aTHX_
16329 n = strlen(full_name);
16330 name = savepvn(full_name, n);
16334 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s%"UTF8f"%s\n",
16335 (value == 'p' ? '+' : '!'),
16336 (FOLD) ? "__" : "",
16337 UTF8fARG(UTF, n, name),
16338 (FOLD) ? "_i" : "");
16339 has_user_defined_property = TRUE;
16340 optimizable = FALSE; /* Will have to leave this an
16343 /* We don't know yet what this matches, so have to flag
16345 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
16349 /* Here, did get the swash and its inversion list. If
16350 * the swash is from a user-defined property, then this
16351 * whole character class should be regarded as such */
16352 if (swash_init_flags
16353 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
16355 has_user_defined_property = TRUE;
16358 /* We warn on matching an above-Unicode code point
16359 * if the match would return true, except don't
16360 * warn for \p{All}, which has exactly one element
16362 (_invlist_contains_cp(invlist, 0x110000)
16363 && (! (_invlist_len(invlist) == 1
16364 && *invlist_array(invlist) == 0)))
16370 /* Invert if asking for the complement */
16371 if (value == 'P') {
16372 _invlist_union_complement_2nd(properties,
16376 /* The swash can't be used as-is, because we've
16377 * inverted things; delay removing it to here after
16378 * have copied its invlist above */
16379 SvREFCNT_dec_NN(swash);
16383 _invlist_union(properties, invlist, &properties);
16387 RExC_parse = e + 1;
16388 namedclass = ANYOF_UNIPROP; /* no official name, but it's
16391 /* \p means they want Unicode semantics */
16392 REQUIRE_UNI_RULES(flagp, NULL);
16395 case 'n': value = '\n'; break;
16396 case 'r': value = '\r'; break;
16397 case 't': value = '\t'; break;
16398 case 'f': value = '\f'; break;
16399 case 'b': value = '\b'; break;
16400 case 'e': value = ESC_NATIVE; break;
16401 case 'a': value = '\a'; break;
16403 RExC_parse--; /* function expects to be pointed at the 'o' */
16405 const char* error_msg;
16406 bool valid = grok_bslash_o(&RExC_parse,
16409 PASS2, /* warnings only in
16412 silence_non_portable,
16418 non_portable_endpoint++;
16421 RExC_parse--; /* function expects to be pointed at the 'x' */
16423 const char* error_msg;
16424 bool valid = grok_bslash_x(&RExC_parse,
16427 PASS2, /* Output warnings */
16429 silence_non_portable,
16435 non_portable_endpoint++;
16438 value = grok_bslash_c(*RExC_parse++, PASS2);
16439 non_portable_endpoint++;
16441 case '0': case '1': case '2': case '3': case '4':
16442 case '5': case '6': case '7':
16444 /* Take 1-3 octal digits */
16445 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
16446 numlen = (strict) ? 4 : 3;
16447 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
16448 RExC_parse += numlen;
16451 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16452 vFAIL("Need exactly 3 octal digits");
16454 else if (! SIZE_ONLY /* like \08, \178 */
16456 && RExC_parse < RExC_end
16457 && isDIGIT(*RExC_parse)
16458 && ckWARN(WARN_REGEXP))
16460 SAVEFREESV(RExC_rx_sv);
16461 reg_warn_non_literal_string(
16463 form_short_octal_warning(RExC_parse, numlen));
16464 (void)ReREFCNT_inc(RExC_rx_sv);
16467 non_portable_endpoint++;
16471 /* Allow \_ to not give an error */
16472 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
16474 vFAIL2("Unrecognized escape \\%c in character class",
16478 SAVEFREESV(RExC_rx_sv);
16479 ckWARN2reg(RExC_parse,
16480 "Unrecognized escape \\%c in character class passed through",
16482 (void)ReREFCNT_inc(RExC_rx_sv);
16486 } /* End of switch on char following backslash */
16487 } /* end of handling backslash escape sequences */
16489 /* Here, we have the current token in 'value' */
16491 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
16494 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
16495 * literal, as is the character that began the false range, i.e.
16496 * the 'a' in the examples */
16499 const int w = (RExC_parse >= rangebegin)
16500 ? RExC_parse - rangebegin
16504 "False [] range \"%"UTF8f"\"",
16505 UTF8fARG(UTF, w, rangebegin));
16508 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
16509 ckWARN2reg(RExC_parse,
16510 "False [] range \"%"UTF8f"\"",
16511 UTF8fARG(UTF, w, rangebegin));
16512 (void)ReREFCNT_inc(RExC_rx_sv);
16513 cp_list = add_cp_to_invlist(cp_list, '-');
16514 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
16519 range = 0; /* this was not a true range */
16520 element_count += 2; /* So counts for three values */
16523 classnum = namedclass_to_classnum(namedclass);
16525 if (LOC && namedclass < ANYOF_POSIXL_MAX
16526 #ifndef HAS_ISASCII
16527 && classnum != _CC_ASCII
16530 /* What the Posix classes (like \w, [:space:]) match in locale
16531 * isn't knowable under locale until actual match time. Room
16532 * must be reserved (one time per outer bracketed class) to
16533 * store such classes. The space will contain a bit for each
16534 * named class that is to be matched against. This isn't
16535 * needed for \p{} and pseudo-classes, as they are not affected
16536 * by locale, and hence are dealt with separately */
16537 if (! need_class) {
16540 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16543 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16545 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
16546 ANYOF_POSIXL_ZERO(ret);
16548 /* We can't change this into some other type of node
16549 * (unless this is the only element, in which case there
16550 * are nodes that mean exactly this) as has runtime
16552 optimizable = FALSE;
16555 /* Coverity thinks it is possible for this to be negative; both
16556 * jhi and khw think it's not, but be safer */
16557 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16558 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
16560 /* See if it already matches the complement of this POSIX
16562 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16563 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
16567 posixl_matches_all = TRUE;
16568 break; /* No need to continue. Since it matches both
16569 e.g., \w and \W, it matches everything, and the
16570 bracketed class can be optimized into qr/./s */
16573 /* Add this class to those that should be checked at runtime */
16574 ANYOF_POSIXL_SET(ret, namedclass);
16576 /* The above-Latin1 characters are not subject to locale rules.
16577 * Just add them, in the second pass, to the
16578 * unconditionally-matched list */
16580 SV* scratch_list = NULL;
16582 /* Get the list of the above-Latin1 code points this
16584 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
16585 PL_XPosix_ptrs[classnum],
16587 /* Odd numbers are complements, like
16588 * NDIGIT, NASCII, ... */
16589 namedclass % 2 != 0,
16591 /* Checking if 'cp_list' is NULL first saves an extra
16592 * clone. Its reference count will be decremented at the
16593 * next union, etc, or if this is the only instance, at the
16594 * end of the routine */
16596 cp_list = scratch_list;
16599 _invlist_union(cp_list, scratch_list, &cp_list);
16600 SvREFCNT_dec_NN(scratch_list);
16602 continue; /* Go get next character */
16605 else if (! SIZE_ONLY) {
16607 /* Here, not in pass1 (in that pass we skip calculating the
16608 * contents of this class), and is not /l, or is a POSIX class
16609 * for which /l doesn't matter (or is a Unicode property, which
16610 * is skipped here). */
16611 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
16612 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
16614 /* Here, should be \h, \H, \v, or \V. None of /d, /i
16615 * nor /l make a difference in what these match,
16616 * therefore we just add what they match to cp_list. */
16617 if (classnum != _CC_VERTSPACE) {
16618 assert( namedclass == ANYOF_HORIZWS
16619 || namedclass == ANYOF_NHORIZWS);
16621 /* It turns out that \h is just a synonym for
16623 classnum = _CC_BLANK;
16626 _invlist_union_maybe_complement_2nd(
16628 PL_XPosix_ptrs[classnum],
16629 namedclass % 2 != 0, /* Complement if odd
16630 (NHORIZWS, NVERTWS)
16635 else if ( UNI_SEMANTICS
16636 || classnum == _CC_ASCII
16637 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
16638 || classnum == _CC_XDIGIT)))
16640 /* We usually have to worry about /d and /a affecting what
16641 * POSIX classes match, with special code needed for /d
16642 * because we won't know until runtime what all matches.
16643 * But there is no extra work needed under /u, and
16644 * [:ascii:] is unaffected by /a and /d; and :digit: and
16645 * :xdigit: don't have runtime differences under /d. So we
16646 * can special case these, and avoid some extra work below,
16647 * and at runtime. */
16648 _invlist_union_maybe_complement_2nd(
16650 PL_XPosix_ptrs[classnum],
16651 namedclass % 2 != 0,
16654 else { /* Garden variety class. If is NUPPER, NALPHA, ...
16655 complement and use nposixes */
16656 SV** posixes_ptr = namedclass % 2 == 0
16659 _invlist_union_maybe_complement_2nd(
16661 PL_XPosix_ptrs[classnum],
16662 namedclass % 2 != 0,
16666 } /* end of namedclass \blah */
16668 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16670 /* If 'range' is set, 'value' is the ending of a range--check its
16671 * validity. (If value isn't a single code point in the case of a
16672 * range, we should have figured that out above in the code that
16673 * catches false ranges). Later, we will handle each individual code
16674 * point in the range. If 'range' isn't set, this could be the
16675 * beginning of a range, so check for that by looking ahead to see if
16676 * the next real character to be processed is the range indicator--the
16681 /* For unicode ranges, we have to test that the Unicode as opposed
16682 * to the native values are not decreasing. (Above 255, there is
16683 * no difference between native and Unicode) */
16684 if (unicode_range && prevvalue < 255 && value < 255) {
16685 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
16686 goto backwards_range;
16691 if (prevvalue > value) /* b-a */ {
16696 w = RExC_parse - rangebegin;
16698 "Invalid [] range \"%"UTF8f"\"",
16699 UTF8fARG(UTF, w, rangebegin));
16700 NOT_REACHED; /* NOTREACHED */
16704 prevvalue = value; /* save the beginning of the potential range */
16705 if (! stop_at_1 /* Can't be a range if parsing just one thing */
16706 && *RExC_parse == '-')
16708 char* next_char_ptr = RExC_parse + 1;
16710 /* Get the next real char after the '-' */
16711 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
16713 /* If the '-' is at the end of the class (just before the ']',
16714 * it is a literal minus; otherwise it is a range */
16715 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
16716 RExC_parse = next_char_ptr;
16718 /* a bad range like \w-, [:word:]- ? */
16719 if (namedclass > OOB_NAMEDCLASS) {
16720 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
16721 const int w = RExC_parse >= rangebegin
16722 ? RExC_parse - rangebegin
16725 vFAIL4("False [] range \"%*.*s\"",
16730 "False [] range \"%*.*s\"",
16735 cp_list = add_cp_to_invlist(cp_list, '-');
16739 range = 1; /* yeah, it's a range! */
16740 continue; /* but do it the next time */
16745 if (namedclass > OOB_NAMEDCLASS) {
16749 /* Here, we have a single value this time through the loop, and
16750 * <prevvalue> is the beginning of the range, if any; or <value> if
16753 /* non-Latin1 code point implies unicode semantics. Must be set in
16754 * pass1 so is there for the whole of pass 2 */
16756 REQUIRE_UNI_RULES(flagp, NULL);
16759 /* Ready to process either the single value, or the completed range.
16760 * For single-valued non-inverted ranges, we consider the possibility
16761 * of multi-char folds. (We made a conscious decision to not do this
16762 * for the other cases because it can often lead to non-intuitive
16763 * results. For example, you have the peculiar case that:
16764 * "s s" =~ /^[^\xDF]+$/i => Y
16765 * "ss" =~ /^[^\xDF]+$/i => N
16767 * See [perl #89750] */
16768 if (FOLD && allow_multi_folds && value == prevvalue) {
16769 if (value == LATIN_SMALL_LETTER_SHARP_S
16770 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
16773 /* Here <value> is indeed a multi-char fold. Get what it is */
16775 U8 foldbuf[UTF8_MAXBYTES_CASE];
16778 UV folded = _to_uni_fold_flags(
16782 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
16783 ? FOLD_FLAGS_NOMIX_ASCII
16787 /* Here, <folded> should be the first character of the
16788 * multi-char fold of <value>, with <foldbuf> containing the
16789 * whole thing. But, if this fold is not allowed (because of
16790 * the flags), <fold> will be the same as <value>, and should
16791 * be processed like any other character, so skip the special
16793 if (folded != value) {
16795 /* Skip if we are recursed, currently parsing the class
16796 * again. Otherwise add this character to the list of
16797 * multi-char folds. */
16798 if (! RExC_in_multi_char_class) {
16799 STRLEN cp_count = utf8_length(foldbuf,
16800 foldbuf + foldlen);
16801 SV* multi_fold = sv_2mortal(newSVpvs(""));
16803 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
16806 = add_multi_match(multi_char_matches,
16812 /* This element should not be processed further in this
16815 value = save_value;
16816 prevvalue = save_prevvalue;
16822 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
16825 /* If the range starts above 255, everything is portable and
16826 * likely to be so for any forseeable character set, so don't
16828 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
16829 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
16831 else if (prevvalue != value) {
16833 /* Under strict, ranges that stop and/or end in an ASCII
16834 * printable should have each end point be a portable value
16835 * for it (preferably like 'A', but we don't warn if it is
16836 * a (portable) Unicode name or code point), and the range
16837 * must be be all digits or all letters of the same case.
16838 * Otherwise, the range is non-portable and unclear as to
16839 * what it contains */
16840 if ((isPRINT_A(prevvalue) || isPRINT_A(value))
16841 && (non_portable_endpoint
16842 || ! ((isDIGIT_A(prevvalue) && isDIGIT_A(value))
16843 || (isLOWER_A(prevvalue) && isLOWER_A(value))
16844 || (isUPPER_A(prevvalue) && isUPPER_A(value)))))
16846 vWARN(RExC_parse, "Ranges of ASCII printables should be some subset of \"0-9\", \"A-Z\", or \"a-z\"");
16848 else if (prevvalue >= 0x660) { /* ARABIC_INDIC_DIGIT_ZERO */
16850 /* But the nature of Unicode and languages mean we
16851 * can't do the same checks for above-ASCII ranges,
16852 * except in the case of digit ones. These should
16853 * contain only digits from the same group of 10. The
16854 * ASCII case is handled just above. 0x660 is the
16855 * first digit character beyond ASCII. Hence here, the
16856 * range could be a range of digits. Find out. */
16857 IV index_start = _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
16859 IV index_final = _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
16862 /* If the range start and final points are in the same
16863 * inversion list element, it means that either both
16864 * are not digits, or both are digits in a consecutive
16865 * sequence of digits. (So far, Unicode has kept all
16866 * such sequences as distinct groups of 10, but assert
16867 * to make sure). If the end points are not in the
16868 * same element, neither should be a digit. */
16869 if (index_start == index_final) {
16870 assert(! ELEMENT_RANGE_MATCHES_INVLIST(index_start)
16871 || (invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start+1]
16872 - invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
16874 /* But actually Unicode did have one group of 11
16875 * 'digits' in 5.2, so in case we are operating
16876 * on that version, let that pass */
16877 || (invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start+1]
16878 - invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
16880 && invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
16884 else if ((index_start >= 0
16885 && ELEMENT_RANGE_MATCHES_INVLIST(index_start))
16886 || (index_final >= 0
16887 && ELEMENT_RANGE_MATCHES_INVLIST(index_final)))
16889 vWARN(RExC_parse, "Ranges of digits should be from the same group of 10");
16894 if ((! range || prevvalue == value) && non_portable_endpoint) {
16895 if (isPRINT_A(value)) {
16898 if (isBACKSLASHED_PUNCT(value)) {
16899 literal[d++] = '\\';
16901 literal[d++] = (char) value;
16902 literal[d++] = '\0';
16905 "\"%.*s\" is more clearly written simply as \"%s\"",
16906 (int) (RExC_parse - rangebegin),
16911 else if isMNEMONIC_CNTRL(value) {
16913 "\"%.*s\" is more clearly written simply as \"%s\"",
16914 (int) (RExC_parse - rangebegin),
16916 cntrl_to_mnemonic((U8) value)
16922 /* Deal with this element of the class */
16926 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16929 /* On non-ASCII platforms, for ranges that span all of 0..255, and
16930 * ones that don't require special handling, we can just add the
16931 * range like we do for ASCII platforms */
16932 if ((UNLIKELY(prevvalue == 0) && value >= 255)
16933 || ! (prevvalue < 256
16935 || (! non_portable_endpoint
16936 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
16937 || (isUPPER_A(prevvalue)
16938 && isUPPER_A(value)))))))
16940 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16944 /* Here, requires special handling. This can be because it is
16945 * a range whose code points are considered to be Unicode, and
16946 * so must be individually translated into native, or because
16947 * its a subrange of 'A-Z' or 'a-z' which each aren't
16948 * contiguous in EBCDIC, but we have defined them to include
16949 * only the "expected" upper or lower case ASCII alphabetics.
16950 * Subranges above 255 are the same in native and Unicode, so
16951 * can be added as a range */
16952 U8 start = NATIVE_TO_LATIN1(prevvalue);
16954 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
16955 for (j = start; j <= end; j++) {
16956 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
16959 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16966 range = 0; /* this range (if it was one) is done now */
16967 } /* End of loop through all the text within the brackets */
16970 if ( posix_warnings && av_tindex_nomg(posix_warnings) >= 0) {
16971 output_or_return_posix_warnings(pRExC_state, posix_warnings,
16972 return_posix_warnings);
16975 /* If anything in the class expands to more than one character, we have to
16976 * deal with them by building up a substitute parse string, and recursively
16977 * calling reg() on it, instead of proceeding */
16978 if (multi_char_matches) {
16979 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
16982 char *save_end = RExC_end;
16983 char *save_parse = RExC_parse;
16984 char *save_start = RExC_start;
16985 STRLEN prefix_end = 0; /* We copy the character class after a
16986 prefix supplied here. This is the size
16987 + 1 of that prefix */
16988 bool first_time = TRUE; /* First multi-char occurrence doesn't get
16993 assert(RExC_precomp_adj == 0); /* Only one level of recursion allowed */
16995 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
16996 because too confusing */
16998 sv_catpv(substitute_parse, "(?:");
17002 /* Look at the longest folds first */
17003 for (cp_count = av_tindex_nomg(multi_char_matches);
17008 if (av_exists(multi_char_matches, cp_count)) {
17009 AV** this_array_ptr;
17012 this_array_ptr = (AV**) av_fetch(multi_char_matches,
17014 while ((this_sequence = av_pop(*this_array_ptr)) !=
17017 if (! first_time) {
17018 sv_catpv(substitute_parse, "|");
17020 first_time = FALSE;
17022 sv_catpv(substitute_parse, SvPVX(this_sequence));
17027 /* If the character class contains anything else besides these
17028 * multi-character folds, have to include it in recursive parsing */
17029 if (element_count) {
17030 sv_catpv(substitute_parse, "|[");
17031 prefix_end = SvCUR(substitute_parse);
17032 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
17034 /* Put in a closing ']' only if not going off the end, as otherwise
17035 * we are adding something that really isn't there */
17036 if (RExC_parse < RExC_end) {
17037 sv_catpv(substitute_parse, "]");
17041 sv_catpv(substitute_parse, ")");
17044 /* This is a way to get the parse to skip forward a whole named
17045 * sequence instead of matching the 2nd character when it fails the
17047 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
17051 /* Set up the data structure so that any errors will be properly
17052 * reported. See the comments at the definition of
17053 * REPORT_LOCATION_ARGS for details */
17054 RExC_precomp_adj = orig_parse - RExC_precomp;
17055 RExC_start = RExC_parse = SvPV(substitute_parse, len);
17056 RExC_adjusted_start = RExC_start + prefix_end;
17057 RExC_end = RExC_parse + len;
17058 RExC_in_multi_char_class = 1;
17059 RExC_override_recoding = 1;
17060 RExC_emit = (regnode *)orig_emit;
17062 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17064 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PASS1|NEED_UTF8);
17066 /* And restore so can parse the rest of the pattern */
17067 RExC_parse = save_parse;
17068 RExC_start = RExC_adjusted_start = save_start;
17069 RExC_precomp_adj = 0;
17070 RExC_end = save_end;
17071 RExC_in_multi_char_class = 0;
17072 RExC_override_recoding = 0;
17073 SvREFCNT_dec_NN(multi_char_matches);
17077 /* Here, we've gone through the entire class and dealt with multi-char
17078 * folds. We are now in a position that we can do some checks to see if we
17079 * can optimize this ANYOF node into a simpler one, even in Pass 1.
17080 * Currently we only do two checks:
17081 * 1) is in the unlikely event that the user has specified both, eg. \w and
17082 * \W under /l, then the class matches everything. (This optimization
17083 * is done only to make the optimizer code run later work.)
17084 * 2) if the character class contains only a single element (including a
17085 * single range), we see if there is an equivalent node for it.
17086 * Other checks are possible */
17088 && ! ret_invlist /* Can't optimize if returning the constructed
17090 && (UNLIKELY(posixl_matches_all) || element_count == 1))
17095 if (UNLIKELY(posixl_matches_all)) {
17098 else if (namedclass > OOB_NAMEDCLASS) { /* this is a single named
17099 class, like \w or [:digit:]
17102 /* All named classes are mapped into POSIXish nodes, with its FLAG
17103 * argument giving which class it is */
17104 switch ((I32)namedclass) {
17105 case ANYOF_UNIPROP:
17108 /* These don't depend on the charset modifiers. They always
17109 * match under /u rules */
17110 case ANYOF_NHORIZWS:
17111 case ANYOF_HORIZWS:
17112 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
17115 case ANYOF_NVERTWS:
17120 /* The actual POSIXish node for all the rest depends on the
17121 * charset modifier. The ones in the first set depend only on
17122 * ASCII or, if available on this platform, also locale */
17126 op = (LOC) ? POSIXL : POSIXA;
17132 /* The following don't have any matches in the upper Latin1
17133 * range, hence /d is equivalent to /u for them. Making it /u
17134 * saves some branches at runtime */
17138 case ANYOF_NXDIGIT:
17139 if (! DEPENDS_SEMANTICS) {
17140 goto treat_as_default;
17146 /* The following change to CASED under /i */
17152 namedclass = ANYOF_CASED + (namedclass % 2);
17156 /* The rest have more possibilities depending on the charset.
17157 * We take advantage of the enum ordering of the charset
17158 * modifiers to get the exact node type, */
17161 op = POSIXD + get_regex_charset(RExC_flags);
17162 if (op > POSIXA) { /* /aa is same as /a */
17167 /* The odd numbered ones are the complements of the
17168 * next-lower even number one */
17169 if (namedclass % 2 == 1) {
17173 arg = namedclass_to_classnum(namedclass);
17177 else if (value == prevvalue) {
17179 /* Here, the class consists of just a single code point */
17182 if (! LOC && value == '\n') {
17183 op = REG_ANY; /* Optimize [^\n] */
17184 *flagp |= HASWIDTH|SIMPLE;
17188 else if (value < 256 || UTF) {
17190 /* Optimize a single value into an EXACTish node, but not if it
17191 * would require converting the pattern to UTF-8. */
17192 op = compute_EXACTish(pRExC_state);
17194 } /* Otherwise is a range */
17195 else if (! LOC) { /* locale could vary these */
17196 if (prevvalue == '0') {
17197 if (value == '9') {
17202 else if (! FOLD || ASCII_FOLD_RESTRICTED) {
17203 /* We can optimize A-Z or a-z, but not if they could match
17204 * something like the KELVIN SIGN under /i. */
17205 if (prevvalue == 'A') {
17208 && ! non_portable_endpoint
17211 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
17215 else if (prevvalue == 'a') {
17218 && ! non_portable_endpoint
17221 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
17228 /* Here, we have changed <op> away from its initial value iff we found
17229 * an optimization */
17232 /* Throw away this ANYOF regnode, and emit the calculated one,
17233 * which should correspond to the beginning, not current, state of
17235 const char * cur_parse = RExC_parse;
17236 RExC_parse = (char *)orig_parse;
17240 /* To get locale nodes to not use the full ANYOF size would
17241 * require moving the code above that writes the portions
17242 * of it that aren't in other nodes to after this point.
17243 * e.g. ANYOF_POSIXL_SET */
17244 RExC_size = orig_size;
17248 RExC_emit = (regnode *)orig_emit;
17249 if (PL_regkind[op] == POSIXD) {
17250 if (op == POSIXL) {
17251 RExC_contains_locale = 1;
17254 op += NPOSIXD - POSIXD;
17259 ret = reg_node(pRExC_state, op);
17261 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17265 *flagp |= HASWIDTH|SIMPLE;
17267 else if (PL_regkind[op] == EXACT) {
17268 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17269 TRUE /* downgradable to EXACT */
17273 RExC_parse = (char *) cur_parse;
17275 SvREFCNT_dec(posixes);
17276 SvREFCNT_dec(nposixes);
17277 SvREFCNT_dec(simple_posixes);
17278 SvREFCNT_dec(cp_list);
17279 SvREFCNT_dec(cp_foldable_list);
17286 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
17288 /* If folding, we calculate all characters that could fold to or from the
17289 * ones already on the list */
17290 if (cp_foldable_list) {
17292 UV start, end; /* End points of code point ranges */
17294 SV* fold_intersection = NULL;
17297 /* Our calculated list will be for Unicode rules. For locale
17298 * matching, we have to keep a separate list that is consulted at
17299 * runtime only when the locale indicates Unicode rules. For
17300 * non-locale, we just use the general list */
17302 use_list = &only_utf8_locale_list;
17305 use_list = &cp_list;
17308 /* Only the characters in this class that participate in folds need
17309 * be checked. Get the intersection of this class and all the
17310 * possible characters that are foldable. This can quickly narrow
17311 * down a large class */
17312 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
17313 &fold_intersection);
17315 /* The folds for all the Latin1 characters are hard-coded into this
17316 * program, but we have to go out to disk to get the others. */
17317 if (invlist_highest(cp_foldable_list) >= 256) {
17319 /* This is a hash that for a particular fold gives all
17320 * characters that are involved in it */
17321 if (! PL_utf8_foldclosures) {
17322 _load_PL_utf8_foldclosures();
17326 /* Now look at the foldable characters in this class individually */
17327 invlist_iterinit(fold_intersection);
17328 while (invlist_iternext(fold_intersection, &start, &end)) {
17331 /* Look at every character in the range */
17332 for (j = start; j <= end; j++) {
17333 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17339 if (IS_IN_SOME_FOLD_L1(j)) {
17341 /* ASCII is always matched; non-ASCII is matched
17342 * only under Unicode rules (which could happen
17343 * under /l if the locale is a UTF-8 one */
17344 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
17345 *use_list = add_cp_to_invlist(*use_list,
17346 PL_fold_latin1[j]);
17349 has_upper_latin1_only_utf8_matches
17350 = add_cp_to_invlist(
17351 has_upper_latin1_only_utf8_matches,
17352 PL_fold_latin1[j]);
17356 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
17357 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
17359 add_above_Latin1_folds(pRExC_state,
17366 /* Here is an above Latin1 character. We don't have the
17367 * rules hard-coded for it. First, get its fold. This is
17368 * the simple fold, as the multi-character folds have been
17369 * handled earlier and separated out */
17370 _to_uni_fold_flags(j, foldbuf, &foldlen,
17371 (ASCII_FOLD_RESTRICTED)
17372 ? FOLD_FLAGS_NOMIX_ASCII
17375 /* Single character fold of above Latin1. Add everything in
17376 * its fold closure to the list that this node should match.
17377 * The fold closures data structure is a hash with the keys
17378 * being the UTF-8 of every character that is folded to, like
17379 * 'k', and the values each an array of all code points that
17380 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
17381 * Multi-character folds are not included */
17382 if ((listp = hv_fetch(PL_utf8_foldclosures,
17383 (char *) foldbuf, foldlen, FALSE)))
17385 AV* list = (AV*) *listp;
17387 for (k = 0; k <= av_tindex_nomg(list); k++) {
17388 SV** c_p = av_fetch(list, k, FALSE);
17394 /* /aa doesn't allow folds between ASCII and non- */
17395 if ((ASCII_FOLD_RESTRICTED
17396 && (isASCII(c) != isASCII(j))))
17401 /* Folds under /l which cross the 255/256 boundary
17402 * are added to a separate list. (These are valid
17403 * only when the locale is UTF-8.) */
17404 if (c < 256 && LOC) {
17405 *use_list = add_cp_to_invlist(*use_list, c);
17409 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
17411 cp_list = add_cp_to_invlist(cp_list, c);
17414 /* Similarly folds involving non-ascii Latin1
17415 * characters under /d are added to their list */
17416 has_upper_latin1_only_utf8_matches
17417 = add_cp_to_invlist(
17418 has_upper_latin1_only_utf8_matches,
17425 SvREFCNT_dec_NN(fold_intersection);
17428 /* Now that we have finished adding all the folds, there is no reason
17429 * to keep the foldable list separate */
17430 _invlist_union(cp_list, cp_foldable_list, &cp_list);
17431 SvREFCNT_dec_NN(cp_foldable_list);
17434 /* And combine the result (if any) with any inversion lists from posix
17435 * classes. The lists are kept separate up to now because we don't want to
17436 * fold the classes (folding of those is automatically handled by the swash
17437 * fetching code) */
17438 if (simple_posixes) { /* These are the classes known to be unaffected by
17441 _invlist_union(cp_list, simple_posixes, &cp_list);
17442 SvREFCNT_dec_NN(simple_posixes);
17445 cp_list = simple_posixes;
17448 if (posixes || nposixes) {
17450 /* We have to adjust /a and /aa */
17451 if (AT_LEAST_ASCII_RESTRICTED) {
17453 /* Under /a and /aa, nothing above ASCII matches these */
17455 _invlist_intersection(posixes,
17456 PL_XPosix_ptrs[_CC_ASCII],
17460 /* Under /a and /aa, everything above ASCII matches these
17463 _invlist_union_complement_2nd(nposixes,
17464 PL_XPosix_ptrs[_CC_ASCII],
17469 if (! DEPENDS_SEMANTICS) {
17471 /* For everything but /d, we can just add the current 'posixes' and
17472 * 'nposixes' to the main list */
17475 _invlist_union(cp_list, posixes, &cp_list);
17476 SvREFCNT_dec_NN(posixes);
17484 _invlist_union(cp_list, nposixes, &cp_list);
17485 SvREFCNT_dec_NN(nposixes);
17488 cp_list = nposixes;
17493 /* Under /d, things like \w match upper Latin1 characters only if
17494 * the target string is in UTF-8. But things like \W match all the
17495 * upper Latin1 characters if the target string is not in UTF-8.
17497 * Handle the case where there something like \W separately */
17499 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1);
17501 /* A complemented posix class matches all upper Latin1
17502 * characters if not in UTF-8. And it matches just certain
17503 * ones when in UTF-8. That means those certain ones are
17504 * matched regardless, so can just be added to the
17505 * unconditional list */
17507 _invlist_union(cp_list, nposixes, &cp_list);
17508 SvREFCNT_dec_NN(nposixes);
17512 cp_list = nposixes;
17515 /* Likewise for 'posixes' */
17516 _invlist_union(posixes, cp_list, &cp_list);
17518 /* Likewise for anything else in the range that matched only
17520 if (has_upper_latin1_only_utf8_matches) {
17521 _invlist_union(cp_list,
17522 has_upper_latin1_only_utf8_matches,
17524 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17525 has_upper_latin1_only_utf8_matches = NULL;
17528 /* If we don't match all the upper Latin1 characters regardless
17529 * of UTF-8ness, we have to set a flag to match the rest when
17531 _invlist_subtract(only_non_utf8_list, cp_list,
17532 &only_non_utf8_list);
17533 if (_invlist_len(only_non_utf8_list) != 0) {
17534 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17538 /* Here there were no complemented posix classes. That means
17539 * the upper Latin1 characters in 'posixes' match only when the
17540 * target string is in UTF-8. So we have to add them to the
17541 * list of those types of code points, while adding the
17542 * remainder to the unconditional list.
17544 * First calculate what they are */
17545 SV* nonascii_but_latin1_properties = NULL;
17546 _invlist_intersection(posixes, PL_UpperLatin1,
17547 &nonascii_but_latin1_properties);
17549 /* And add them to the final list of such characters. */
17550 _invlist_union(has_upper_latin1_only_utf8_matches,
17551 nonascii_but_latin1_properties,
17552 &has_upper_latin1_only_utf8_matches);
17554 /* Remove them from what now becomes the unconditional list */
17555 _invlist_subtract(posixes, nonascii_but_latin1_properties,
17558 /* And add those unconditional ones to the final list */
17560 _invlist_union(cp_list, posixes, &cp_list);
17561 SvREFCNT_dec_NN(posixes);
17568 SvREFCNT_dec(nonascii_but_latin1_properties);
17570 /* Get rid of any characters that we now know are matched
17571 * unconditionally from the conditional list, which may make
17572 * that list empty */
17573 _invlist_subtract(has_upper_latin1_only_utf8_matches,
17575 &has_upper_latin1_only_utf8_matches);
17576 if (_invlist_len(has_upper_latin1_only_utf8_matches) == 0) {
17577 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17578 has_upper_latin1_only_utf8_matches = NULL;
17584 /* And combine the result (if any) with any inversion list from properties.
17585 * The lists are kept separate up to now so that we can distinguish the two
17586 * in regards to matching above-Unicode. A run-time warning is generated
17587 * if a Unicode property is matched against a non-Unicode code point. But,
17588 * we allow user-defined properties to match anything, without any warning,
17589 * and we also suppress the warning if there is a portion of the character
17590 * class that isn't a Unicode property, and which matches above Unicode, \W
17591 * or [\x{110000}] for example.
17592 * (Note that in this case, unlike the Posix one above, there is no
17593 * <has_upper_latin1_only_utf8_matches>, because having a Unicode property
17594 * forces Unicode semantics */
17598 /* If it matters to the final outcome, see if a non-property
17599 * component of the class matches above Unicode. If so, the
17600 * warning gets suppressed. This is true even if just a single
17601 * such code point is specified, as, though not strictly correct if
17602 * another such code point is matched against, the fact that they
17603 * are using above-Unicode code points indicates they should know
17604 * the issues involved */
17606 warn_super = ! (invert
17607 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
17610 _invlist_union(properties, cp_list, &cp_list);
17611 SvREFCNT_dec_NN(properties);
17614 cp_list = properties;
17619 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17621 /* Because an ANYOF node is the only one that warns, this node
17622 * can't be optimized into something else */
17623 optimizable = FALSE;
17627 /* Here, we have calculated what code points should be in the character
17630 * Now we can see about various optimizations. Fold calculation (which we
17631 * did above) needs to take place before inversion. Otherwise /[^k]/i
17632 * would invert to include K, which under /i would match k, which it
17633 * shouldn't. Therefore we can't invert folded locale now, as it won't be
17634 * folded until runtime */
17636 /* If we didn't do folding, it's because some information isn't available
17637 * until runtime; set the run-time fold flag for these. (We don't have to
17638 * worry about properties folding, as that is taken care of by the swash
17639 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
17640 * locales, or the class matches at least one 0-255 range code point */
17643 /* Some things on the list might be unconditionally included because of
17644 * other components. Remove them, and clean up the list if it goes to
17646 if (only_utf8_locale_list && cp_list) {
17647 _invlist_subtract(only_utf8_locale_list, cp_list,
17648 &only_utf8_locale_list);
17650 if (_invlist_len(only_utf8_locale_list) == 0) {
17651 SvREFCNT_dec_NN(only_utf8_locale_list);
17652 only_utf8_locale_list = NULL;
17655 if (only_utf8_locale_list) {
17658 |ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17660 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
17662 invlist_iterinit(cp_list);
17663 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
17664 ANYOF_FLAGS(ret) |= ANYOFL_FOLD;
17666 invlist_iterfinish(cp_list);
17669 else if ( DEPENDS_SEMANTICS
17670 && ( has_upper_latin1_only_utf8_matches
17671 || (ANYOF_FLAGS(ret) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
17674 optimizable = FALSE;
17678 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
17679 * at compile time. Besides not inverting folded locale now, we can't
17680 * invert if there are things such as \w, which aren't known until runtime
17684 && OP(ret) != ANYOFD
17685 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
17686 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17688 _invlist_invert(cp_list);
17690 /* Any swash can't be used as-is, because we've inverted things */
17692 SvREFCNT_dec_NN(swash);
17696 /* Clear the invert flag since have just done it here */
17703 *ret_invlist = cp_list;
17704 SvREFCNT_dec(swash);
17706 /* Discard the generated node */
17708 RExC_size = orig_size;
17711 RExC_emit = orig_emit;
17716 /* Some character classes are equivalent to other nodes. Such nodes take
17717 * up less room and generally fewer operations to execute than ANYOF nodes.
17718 * Above, we checked for and optimized into some such equivalents for
17719 * certain common classes that are easy to test. Getting to this point in
17720 * the code means that the class didn't get optimized there. Since this
17721 * code is only executed in Pass 2, it is too late to save space--it has
17722 * been allocated in Pass 1, and currently isn't given back. But turning
17723 * things into an EXACTish node can allow the optimizer to join it to any
17724 * adjacent such nodes. And if the class is equivalent to things like /./,
17725 * expensive run-time swashes can be avoided. Now that we have more
17726 * complete information, we can find things necessarily missed by the
17727 * earlier code. Another possible "optimization" that isn't done is that
17728 * something like [Ee] could be changed into an EXACTFU. khw tried this
17729 * and found that the ANYOF is faster, including for code points not in the
17730 * bitmap. This still might make sense to do, provided it got joined with
17731 * an adjacent node(s) to create a longer EXACTFU one. This could be
17732 * accomplished by creating a pseudo ANYOF_EXACTFU node type that the join
17733 * routine would know is joinable. If that didn't happen, the node type
17734 * could then be made a straight ANYOF */
17736 if (optimizable && cp_list && ! invert) {
17738 U8 op = END; /* The optimzation node-type */
17739 int posix_class = -1; /* Illegal value */
17740 const char * cur_parse= RExC_parse;
17742 invlist_iterinit(cp_list);
17743 if (! invlist_iternext(cp_list, &start, &end)) {
17745 /* Here, the list is empty. This happens, for example, when a
17746 * Unicode property that doesn't match anything is the only element
17747 * in the character class (perluniprops.pod notes such properties).
17750 *flagp |= HASWIDTH|SIMPLE;
17752 else if (start == end) { /* The range is a single code point */
17753 if (! invlist_iternext(cp_list, &start, &end)
17755 /* Don't do this optimization if it would require changing
17756 * the pattern to UTF-8 */
17757 && (start < 256 || UTF))
17759 /* Here, the list contains a single code point. Can optimize
17760 * into an EXACTish node */
17771 /* A locale node under folding with one code point can be
17772 * an EXACTFL, as its fold won't be calculated until
17778 /* Here, we are generally folding, but there is only one
17779 * code point to match. If we have to, we use an EXACT
17780 * node, but it would be better for joining with adjacent
17781 * nodes in the optimization pass if we used the same
17782 * EXACTFish node that any such are likely to be. We can
17783 * do this iff the code point doesn't participate in any
17784 * folds. For example, an EXACTF of a colon is the same as
17785 * an EXACT one, since nothing folds to or from a colon. */
17787 if (IS_IN_SOME_FOLD_L1(value)) {
17792 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
17797 /* If we haven't found the node type, above, it means we
17798 * can use the prevailing one */
17800 op = compute_EXACTish(pRExC_state);
17804 } /* End of first range contains just a single code point */
17805 else if (start == 0) {
17806 if (end == UV_MAX) {
17808 *flagp |= HASWIDTH|SIMPLE;
17811 else if (end == '\n' - 1
17812 && invlist_iternext(cp_list, &start, &end)
17813 && start == '\n' + 1 && end == UV_MAX)
17816 *flagp |= HASWIDTH|SIMPLE;
17820 invlist_iterfinish(cp_list);
17823 const UV cp_list_len = _invlist_len(cp_list);
17824 const UV* cp_list_array = invlist_array(cp_list);
17826 /* Here, didn't find an optimization. See if this matches any of
17827 * the POSIX classes. These run slightly faster for above-Unicode
17828 * code points, so don't bother with POSIXA ones nor the 2 that
17829 * have no above-Unicode matches. We can avoid these checks unless
17830 * the ANYOF matches at least as high as the lowest POSIX one
17831 * (which was manually found to be \v. The actual code point may
17832 * increase in later Unicode releases, if a higher code point is
17833 * assigned to be \v, but this code will never break. It would
17834 * just mean we could execute the checks for posix optimizations
17835 * unnecessarily) */
17837 if (cp_list_array[cp_list_len-1] > 0x2029) {
17838 for (posix_class = 0;
17839 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
17843 if (posix_class == _CC_ASCII || posix_class == _CC_CNTRL) {
17846 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
17848 /* Check if matches normal or inverted */
17849 if (_invlistEQ(cp_list,
17850 PL_XPosix_ptrs[posix_class],
17853 op = (try_inverted)
17856 *flagp |= HASWIDTH|SIMPLE;
17866 RExC_parse = (char *)orig_parse;
17867 RExC_emit = (regnode *)orig_emit;
17869 if (regarglen[op]) {
17870 ret = reganode(pRExC_state, op, 0);
17872 ret = reg_node(pRExC_state, op);
17875 RExC_parse = (char *)cur_parse;
17877 if (PL_regkind[op] == EXACT) {
17878 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17879 TRUE /* downgradable to EXACT */
17882 else if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17883 FLAGS(ret) = posix_class;
17886 SvREFCNT_dec_NN(cp_list);
17891 /* Here, <cp_list> contains all the code points we can determine at
17892 * compile time that match under all conditions. Go through it, and
17893 * for things that belong in the bitmap, put them there, and delete from
17894 * <cp_list>. While we are at it, see if everything above 255 is in the
17895 * list, and if so, set a flag to speed up execution */
17897 populate_ANYOF_from_invlist(ret, &cp_list);
17900 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
17903 /* Here, the bitmap has been populated with all the Latin1 code points that
17904 * always match. Can now add to the overall list those that match only
17905 * when the target string is UTF-8 (<has_upper_latin1_only_utf8_matches>).
17907 if (has_upper_latin1_only_utf8_matches) {
17909 _invlist_union(cp_list,
17910 has_upper_latin1_only_utf8_matches,
17912 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17915 cp_list = has_upper_latin1_only_utf8_matches;
17917 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
17920 /* If there is a swash and more than one element, we can't use the swash in
17921 * the optimization below. */
17922 if (swash && element_count > 1) {
17923 SvREFCNT_dec_NN(swash);
17927 /* Note that the optimization of using 'swash' if it is the only thing in
17928 * the class doesn't have us change swash at all, so it can include things
17929 * that are also in the bitmap; otherwise we have purposely deleted that
17930 * duplicate information */
17931 set_ANYOF_arg(pRExC_state, ret, cp_list,
17932 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17934 only_utf8_locale_list,
17935 swash, has_user_defined_property);
17937 *flagp |= HASWIDTH|SIMPLE;
17939 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
17940 RExC_contains_locale = 1;
17946 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
17949 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
17950 regnode* const node,
17952 SV* const runtime_defns,
17953 SV* const only_utf8_locale_list,
17955 const bool has_user_defined_property)
17957 /* Sets the arg field of an ANYOF-type node 'node', using information about
17958 * the node passed-in. If there is nothing outside the node's bitmap, the
17959 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
17960 * the count returned by add_data(), having allocated and stored an array,
17961 * av, that that count references, as follows:
17962 * av[0] stores the character class description in its textual form.
17963 * This is used later (regexec.c:Perl_regclass_swash()) to
17964 * initialize the appropriate swash, and is also useful for dumping
17965 * the regnode. This is set to &PL_sv_undef if the textual
17966 * description is not needed at run-time (as happens if the other
17967 * elements completely define the class)
17968 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
17969 * computed from av[0]. But if no further computation need be done,
17970 * the swash is stored here now (and av[0] is &PL_sv_undef).
17971 * av[2] stores the inversion list of code points that match only if the
17972 * current locale is UTF-8
17973 * av[3] stores the cp_list inversion list for use in addition or instead
17974 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
17975 * (Otherwise everything needed is already in av[0] and av[1])
17976 * av[4] is set if any component of the class is from a user-defined
17977 * property; used only if av[3] exists */
17981 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
17983 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
17984 assert(! (ANYOF_FLAGS(node)
17985 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
17986 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
17989 AV * const av = newAV();
17992 av_store(av, 0, (runtime_defns)
17993 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
17996 av_store(av, 1, swash);
17997 SvREFCNT_dec_NN(cp_list);
18000 av_store(av, 1, &PL_sv_undef);
18002 av_store(av, 3, cp_list);
18003 av_store(av, 4, newSVuv(has_user_defined_property));
18007 if (only_utf8_locale_list) {
18008 av_store(av, 2, only_utf8_locale_list);
18011 av_store(av, 2, &PL_sv_undef);
18014 rv = newRV_noinc(MUTABLE_SV(av));
18015 n = add_data(pRExC_state, STR_WITH_LEN("s"));
18016 RExC_rxi->data->data[n] = (void*)rv;
18021 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
18023 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
18024 const regnode* node,
18027 SV** only_utf8_locale_ptr,
18028 SV** output_invlist)
18031 /* For internal core use only.
18032 * Returns the swash for the input 'node' in the regex 'prog'.
18033 * If <doinit> is 'true', will attempt to create the swash if not already
18035 * If <listsvp> is non-null, will return the printable contents of the
18036 * swash. This can be used to get debugging information even before the
18037 * swash exists, by calling this function with 'doinit' set to false, in
18038 * which case the components that will be used to eventually create the
18039 * swash are returned (in a printable form).
18040 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
18041 * store an inversion list of code points that should match only if the
18042 * execution-time locale is a UTF-8 one.
18043 * If <output_invlist> is not NULL, it is where this routine is to store an
18044 * inversion list of the code points that would be instead returned in
18045 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
18046 * when this parameter is used, is just the non-code point data that
18047 * will go into creating the swash. This currently should be just
18048 * user-defined properties whose definitions were not known at compile
18049 * time. Using this parameter allows for easier manipulation of the
18050 * swash's data by the caller. It is illegal to call this function with
18051 * this parameter set, but not <listsvp>
18053 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
18054 * that, in spite of this function's name, the swash it returns may include
18055 * the bitmap data as well */
18058 SV *si = NULL; /* Input swash initialization string */
18059 SV* invlist = NULL;
18061 RXi_GET_DECL(prog,progi);
18062 const struct reg_data * const data = prog ? progi->data : NULL;
18064 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
18065 assert(! output_invlist || listsvp);
18067 if (data && data->count) {
18068 const U32 n = ARG(node);
18070 if (data->what[n] == 's') {
18071 SV * const rv = MUTABLE_SV(data->data[n]);
18072 AV * const av = MUTABLE_AV(SvRV(rv));
18073 SV **const ary = AvARRAY(av);
18074 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
18076 si = *ary; /* ary[0] = the string to initialize the swash with */
18078 if (av_tindex_nomg(av) >= 2) {
18079 if (only_utf8_locale_ptr
18081 && ary[2] != &PL_sv_undef)
18083 *only_utf8_locale_ptr = ary[2];
18086 assert(only_utf8_locale_ptr);
18087 *only_utf8_locale_ptr = NULL;
18090 /* Elements 3 and 4 are either both present or both absent. [3]
18091 * is any inversion list generated at compile time; [4]
18092 * indicates if that inversion list has any user-defined
18093 * properties in it. */
18094 if (av_tindex_nomg(av) >= 3) {
18096 if (SvUV(ary[4])) {
18097 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
18105 /* Element [1] is reserved for the set-up swash. If already there,
18106 * return it; if not, create it and store it there */
18107 if (ary[1] && SvROK(ary[1])) {
18110 else if (doinit && ((si && si != &PL_sv_undef)
18111 || (invlist && invlist != &PL_sv_undef))) {
18113 sw = _core_swash_init("utf8", /* the utf8 package */
18117 0, /* not from tr/// */
18119 &swash_init_flags);
18120 (void)av_store(av, 1, sw);
18125 /* If requested, return a printable version of what this swash matches */
18127 SV* matches_string = NULL;
18129 /* The swash should be used, if possible, to get the data, as it
18130 * contains the resolved data. But this function can be called at
18131 * compile-time, before everything gets resolved, in which case we
18132 * return the currently best available information, which is the string
18133 * that will eventually be used to do that resolving, 'si' */
18134 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
18135 && (si && si != &PL_sv_undef))
18137 /* Here, we only have 'si' (and possibly some passed-in data in
18138 * 'invlist', which is handled below) If the caller only wants
18139 * 'si', use that. */
18140 if (! output_invlist) {
18141 matches_string = newSVsv(si);
18144 /* But if the caller wants an inversion list of the node, we
18145 * need to parse 'si' and place as much as possible in the
18146 * desired output inversion list, making 'matches_string' only
18147 * contain the currently unresolvable things */
18148 const char *si_string = SvPVX(si);
18149 STRLEN remaining = SvCUR(si);
18153 /* Ignore everything before the first new-line */
18154 while (*si_string != '\n' && remaining > 0) {
18158 assert(remaining > 0);
18163 while (remaining > 0) {
18165 /* The data consists of just strings defining user-defined
18166 * property names, but in prior incarnations, and perhaps
18167 * somehow from pluggable regex engines, it could still
18168 * hold hex code point definitions. Each component of a
18169 * range would be separated by a tab, and each range by a
18170 * new-line. If these are found, instead add them to the
18171 * inversion list */
18172 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
18173 |PERL_SCAN_SILENT_NON_PORTABLE;
18174 STRLEN len = remaining;
18175 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
18177 /* If the hex decode routine found something, it should go
18178 * up to the next \n */
18179 if ( *(si_string + len) == '\n') {
18180 if (count) { /* 2nd code point on line */
18181 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
18184 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
18187 goto prepare_for_next_iteration;
18190 /* If the hex decode was instead for the lower range limit,
18191 * save it, and go parse the upper range limit */
18192 if (*(si_string + len) == '\t') {
18193 assert(count == 0);
18197 prepare_for_next_iteration:
18198 si_string += len + 1;
18199 remaining -= len + 1;
18203 /* Here, didn't find a legal hex number. Just add it from
18204 * here to the next \n */
18207 while (*(si_string + len) != '\n' && remaining > 0) {
18211 if (*(si_string + len) == '\n') {
18215 if (matches_string) {
18216 sv_catpvn(matches_string, si_string, len - 1);
18219 matches_string = newSVpvn(si_string, len - 1);
18222 sv_catpvs(matches_string, " ");
18223 } /* end of loop through the text */
18225 assert(matches_string);
18226 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
18227 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
18229 } /* end of has an 'si' but no swash */
18232 /* If we have a swash in place, its equivalent inversion list was above
18233 * placed into 'invlist'. If not, this variable may contain a stored
18234 * inversion list which is information beyond what is in 'si' */
18237 /* Again, if the caller doesn't want the output inversion list, put
18238 * everything in 'matches-string' */
18239 if (! output_invlist) {
18240 if ( ! matches_string) {
18241 matches_string = newSVpvs("\n");
18243 sv_catsv(matches_string, invlist_contents(invlist,
18244 TRUE /* traditional style */
18247 else if (! *output_invlist) {
18248 *output_invlist = invlist_clone(invlist);
18251 _invlist_union(*output_invlist, invlist, output_invlist);
18255 *listsvp = matches_string;
18260 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
18262 /* reg_skipcomment()
18264 Absorbs an /x style # comment from the input stream,
18265 returning a pointer to the first character beyond the comment, or if the
18266 comment terminates the pattern without anything following it, this returns
18267 one past the final character of the pattern (in other words, RExC_end) and
18268 sets the REG_RUN_ON_COMMENT_SEEN flag.
18270 Note it's the callers responsibility to ensure that we are
18271 actually in /x mode
18275 PERL_STATIC_INLINE char*
18276 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
18278 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
18282 while (p < RExC_end) {
18283 if (*(++p) == '\n') {
18288 /* we ran off the end of the pattern without ending the comment, so we have
18289 * to add an \n when wrapping */
18290 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
18295 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
18297 const bool force_to_xmod
18300 /* If the text at the current parse position '*p' is a '(?#...)' comment,
18301 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
18302 * is /x whitespace, advance '*p' so that on exit it points to the first
18303 * byte past all such white space and comments */
18305 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
18307 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
18309 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
18312 if (RExC_end - (*p) >= 3
18314 && *(*p + 1) == '?'
18315 && *(*p + 2) == '#')
18317 while (*(*p) != ')') {
18318 if ((*p) == RExC_end)
18319 FAIL("Sequence (?#... not terminated");
18327 const char * save_p = *p;
18328 while ((*p) < RExC_end) {
18330 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
18333 else if (*(*p) == '#') {
18334 (*p) = reg_skipcomment(pRExC_state, (*p));
18340 if (*p != save_p) {
18353 Advances the parse position by one byte, unless that byte is the beginning
18354 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
18355 those two cases, the parse position is advanced beyond all such comments and
18358 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
18362 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
18364 PERL_ARGS_ASSERT_NEXTCHAR;
18366 if (RExC_parse < RExC_end) {
18368 || UTF8_IS_INVARIANT(*RExC_parse)
18369 || UTF8_IS_START(*RExC_parse));
18371 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
18373 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
18374 FALSE /* Don't force /x */ );
18379 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
18381 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
18382 * space. In pass1, it aligns and increments RExC_size; in pass2,
18385 regnode * const ret = RExC_emit;
18386 GET_RE_DEBUG_FLAGS_DECL;
18388 PERL_ARGS_ASSERT_REGNODE_GUTS;
18390 assert(extra_size >= regarglen[op]);
18393 SIZE_ALIGN(RExC_size);
18394 RExC_size += 1 + extra_size;
18397 if (RExC_emit >= RExC_emit_bound)
18398 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
18399 op, (void*)RExC_emit, (void*)RExC_emit_bound);
18401 NODE_ALIGN_FILL(ret);
18402 #ifndef RE_TRACK_PATTERN_OFFSETS
18403 PERL_UNUSED_ARG(name);
18405 if (RExC_offsets) { /* MJD */
18407 ("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
18410 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
18411 ? "Overwriting end of array!\n" : "OK",
18412 (UV)(RExC_emit - RExC_emit_start),
18413 (UV)(RExC_parse - RExC_start),
18414 (UV)RExC_offsets[0]));
18415 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
18422 - reg_node - emit a node
18424 STATIC regnode * /* Location. */
18425 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
18427 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
18429 PERL_ARGS_ASSERT_REG_NODE;
18431 assert(regarglen[op] == 0);
18434 regnode *ptr = ret;
18435 FILL_ADVANCE_NODE(ptr, op);
18442 - reganode - emit a node with an argument
18444 STATIC regnode * /* Location. */
18445 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
18447 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
18449 PERL_ARGS_ASSERT_REGANODE;
18451 assert(regarglen[op] == 1);
18454 regnode *ptr = ret;
18455 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
18462 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
18464 /* emit a node with U32 and I32 arguments */
18466 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
18468 PERL_ARGS_ASSERT_REG2LANODE;
18470 assert(regarglen[op] == 2);
18473 regnode *ptr = ret;
18474 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
18481 - reginsert - insert an operator in front of already-emitted operand
18483 * Means relocating the operand.
18486 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
18491 const int offset = regarglen[(U8)op];
18492 const int size = NODE_STEP_REGNODE + offset;
18493 GET_RE_DEBUG_FLAGS_DECL;
18495 PERL_ARGS_ASSERT_REGINSERT;
18496 PERL_UNUSED_CONTEXT;
18497 PERL_UNUSED_ARG(depth);
18498 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
18499 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
18504 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
18505 studying. If this is wrong then we need to adjust RExC_recurse
18506 below like we do with RExC_open_parens/RExC_close_parens. */
18510 if (RExC_open_parens) {
18512 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
18513 /* remember that RExC_npar is rex->nparens + 1,
18514 * iow it is 1 more than the number of parens seen in
18515 * the pattern so far. */
18516 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
18517 /* note, RExC_open_parens[0] is the start of the
18518 * regex, it can't move. RExC_close_parens[0] is the end
18519 * of the regex, it *can* move. */
18520 if ( paren && RExC_open_parens[paren] >= opnd ) {
18521 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
18522 RExC_open_parens[paren] += size;
18524 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
18526 if ( RExC_close_parens[paren] >= opnd ) {
18527 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
18528 RExC_close_parens[paren] += size;
18530 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
18535 RExC_end_op += size;
18537 while (src > opnd) {
18538 StructCopy(--src, --dst, regnode);
18539 #ifdef RE_TRACK_PATTERN_OFFSETS
18540 if (RExC_offsets) { /* MJD 20010112 */
18542 ("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
18546 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
18547 ? "Overwriting end of array!\n" : "OK",
18548 (UV)(src - RExC_emit_start),
18549 (UV)(dst - RExC_emit_start),
18550 (UV)RExC_offsets[0]));
18551 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
18552 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
18558 place = opnd; /* Op node, where operand used to be. */
18559 #ifdef RE_TRACK_PATTERN_OFFSETS
18560 if (RExC_offsets) { /* MJD */
18562 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
18566 (UV)(place - RExC_emit_start) > RExC_offsets[0]
18567 ? "Overwriting end of array!\n" : "OK",
18568 (UV)(place - RExC_emit_start),
18569 (UV)(RExC_parse - RExC_start),
18570 (UV)RExC_offsets[0]));
18571 Set_Node_Offset(place, RExC_parse);
18572 Set_Node_Length(place, 1);
18575 src = NEXTOPER(place);
18576 FILL_ADVANCE_NODE(place, op);
18577 Zero(src, offset, regnode);
18581 - regtail - set the next-pointer at the end of a node chain of p to val.
18582 - SEE ALSO: regtail_study
18585 S_regtail(pTHX_ RExC_state_t * pRExC_state,
18586 const regnode * const p,
18587 const regnode * const val,
18591 GET_RE_DEBUG_FLAGS_DECL;
18593 PERL_ARGS_ASSERT_REGTAIL;
18595 PERL_UNUSED_ARG(depth);
18601 /* Find last node. */
18602 scan = (regnode *) p;
18604 regnode * const temp = regnext(scan);
18606 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
18607 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18608 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
18609 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
18610 (temp == NULL ? "->" : ""),
18611 (temp == NULL ? PL_reg_name[OP(val)] : "")
18619 if (reg_off_by_arg[OP(scan)]) {
18620 ARG_SET(scan, val - scan);
18623 NEXT_OFF(scan) = val - scan;
18629 - regtail_study - set the next-pointer at the end of a node chain of p to val.
18630 - Look for optimizable sequences at the same time.
18631 - currently only looks for EXACT chains.
18633 This is experimental code. The idea is to use this routine to perform
18634 in place optimizations on branches and groups as they are constructed,
18635 with the long term intention of removing optimization from study_chunk so
18636 that it is purely analytical.
18638 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
18639 to control which is which.
18642 /* TODO: All four parms should be const */
18645 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
18646 const regnode *val,U32 depth)
18650 #ifdef EXPERIMENTAL_INPLACESCAN
18653 GET_RE_DEBUG_FLAGS_DECL;
18655 PERL_ARGS_ASSERT_REGTAIL_STUDY;
18661 /* Find last node. */
18665 regnode * const temp = regnext(scan);
18666 #ifdef EXPERIMENTAL_INPLACESCAN
18667 if (PL_regkind[OP(scan)] == EXACT) {
18668 bool unfolded_multi_char; /* Unexamined in this routine */
18669 if (join_exact(pRExC_state, scan, &min,
18670 &unfolded_multi_char, 1, val, depth+1))
18675 switch (OP(scan)) {
18679 case EXACTFA_NO_TRIE:
18685 if( exact == PSEUDO )
18687 else if ( exact != OP(scan) )
18696 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
18697 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18698 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
18699 SvPV_nolen_const(RExC_mysv),
18700 REG_NODE_NUM(scan),
18701 PL_reg_name[exact]);
18708 DEBUG_PARSE_MSG("");
18709 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
18710 Perl_re_printf( aTHX_
18711 "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
18712 SvPV_nolen_const(RExC_mysv),
18713 (IV)REG_NODE_NUM(val),
18717 if (reg_off_by_arg[OP(scan)]) {
18718 ARG_SET(scan, val - scan);
18721 NEXT_OFF(scan) = val - scan;
18729 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
18734 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
18739 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18741 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
18742 if (flags & (1<<bit)) {
18743 if (!set++ && lead)
18744 Perl_re_printf( aTHX_ "%s",lead);
18745 Perl_re_printf( aTHX_ "%s ",PL_reg_intflags_name[bit]);
18750 Perl_re_printf( aTHX_ "\n");
18752 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18757 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
18763 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18765 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
18766 if (flags & (1<<bit)) {
18767 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
18770 if (!set++ && lead)
18771 Perl_re_printf( aTHX_ "%s",lead);
18772 Perl_re_printf( aTHX_ "%s ",PL_reg_extflags_name[bit]);
18775 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
18776 if (!set++ && lead) {
18777 Perl_re_printf( aTHX_ "%s",lead);
18780 case REGEX_UNICODE_CHARSET:
18781 Perl_re_printf( aTHX_ "UNICODE");
18783 case REGEX_LOCALE_CHARSET:
18784 Perl_re_printf( aTHX_ "LOCALE");
18786 case REGEX_ASCII_RESTRICTED_CHARSET:
18787 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
18789 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
18790 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
18793 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
18799 Perl_re_printf( aTHX_ "\n");
18801 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18807 Perl_regdump(pTHX_ const regexp *r)
18810 SV * const sv = sv_newmortal();
18811 SV *dsv= sv_newmortal();
18812 RXi_GET_DECL(r,ri);
18813 GET_RE_DEBUG_FLAGS_DECL;
18815 PERL_ARGS_ASSERT_REGDUMP;
18817 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
18819 /* Header fields of interest. */
18820 if (r->anchored_substr) {
18821 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
18822 RE_SV_DUMPLEN(r->anchored_substr), 30);
18823 Perl_re_printf( aTHX_
18824 "anchored %s%s at %"IVdf" ",
18825 s, RE_SV_TAIL(r->anchored_substr),
18826 (IV)r->anchored_offset);
18827 } else if (r->anchored_utf8) {
18828 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
18829 RE_SV_DUMPLEN(r->anchored_utf8), 30);
18830 Perl_re_printf( aTHX_
18831 "anchored utf8 %s%s at %"IVdf" ",
18832 s, RE_SV_TAIL(r->anchored_utf8),
18833 (IV)r->anchored_offset);
18835 if (r->float_substr) {
18836 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
18837 RE_SV_DUMPLEN(r->float_substr), 30);
18838 Perl_re_printf( aTHX_
18839 "floating %s%s at %"IVdf"..%"UVuf" ",
18840 s, RE_SV_TAIL(r->float_substr),
18841 (IV)r->float_min_offset, (UV)r->float_max_offset);
18842 } else if (r->float_utf8) {
18843 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
18844 RE_SV_DUMPLEN(r->float_utf8), 30);
18845 Perl_re_printf( aTHX_
18846 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
18847 s, RE_SV_TAIL(r->float_utf8),
18848 (IV)r->float_min_offset, (UV)r->float_max_offset);
18850 if (r->check_substr || r->check_utf8)
18851 Perl_re_printf( aTHX_
18853 (r->check_substr == r->float_substr
18854 && r->check_utf8 == r->float_utf8
18855 ? "(checking floating" : "(checking anchored"));
18856 if (r->intflags & PREGf_NOSCAN)
18857 Perl_re_printf( aTHX_ " noscan");
18858 if (r->extflags & RXf_CHECK_ALL)
18859 Perl_re_printf( aTHX_ " isall");
18860 if (r->check_substr || r->check_utf8)
18861 Perl_re_printf( aTHX_ ") ");
18863 if (ri->regstclass) {
18864 regprop(r, sv, ri->regstclass, NULL, NULL);
18865 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
18867 if (r->intflags & PREGf_ANCH) {
18868 Perl_re_printf( aTHX_ "anchored");
18869 if (r->intflags & PREGf_ANCH_MBOL)
18870 Perl_re_printf( aTHX_ "(MBOL)");
18871 if (r->intflags & PREGf_ANCH_SBOL)
18872 Perl_re_printf( aTHX_ "(SBOL)");
18873 if (r->intflags & PREGf_ANCH_GPOS)
18874 Perl_re_printf( aTHX_ "(GPOS)");
18875 Perl_re_printf( aTHX_ " ");
18877 if (r->intflags & PREGf_GPOS_SEEN)
18878 Perl_re_printf( aTHX_ "GPOS:%"UVuf" ", (UV)r->gofs);
18879 if (r->intflags & PREGf_SKIP)
18880 Perl_re_printf( aTHX_ "plus ");
18881 if (r->intflags & PREGf_IMPLICIT)
18882 Perl_re_printf( aTHX_ "implicit ");
18883 Perl_re_printf( aTHX_ "minlen %"IVdf" ", (IV)r->minlen);
18884 if (r->extflags & RXf_EVAL_SEEN)
18885 Perl_re_printf( aTHX_ "with eval ");
18886 Perl_re_printf( aTHX_ "\n");
18888 regdump_extflags("r->extflags: ",r->extflags);
18889 regdump_intflags("r->intflags: ",r->intflags);
18892 PERL_ARGS_ASSERT_REGDUMP;
18893 PERL_UNUSED_CONTEXT;
18894 PERL_UNUSED_ARG(r);
18895 #endif /* DEBUGGING */
18898 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
18901 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
18902 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
18903 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
18904 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
18905 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
18906 || _CC_VERTSPACE != 15
18907 # error Need to adjust order of anyofs[]
18909 static const char * const anyofs[] = {
18946 - regprop - printable representation of opcode, with run time support
18950 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
18954 RXi_GET_DECL(prog,progi);
18955 GET_RE_DEBUG_FLAGS_DECL;
18957 PERL_ARGS_ASSERT_REGPROP;
18961 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
18962 /* It would be nice to FAIL() here, but this may be called from
18963 regexec.c, and it would be hard to supply pRExC_state. */
18964 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
18965 (int)OP(o), (int)REGNODE_MAX);
18966 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
18968 k = PL_regkind[OP(o)];
18971 sv_catpvs(sv, " ");
18972 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
18973 * is a crude hack but it may be the best for now since
18974 * we have no flag "this EXACTish node was UTF-8"
18976 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
18977 PERL_PV_ESCAPE_UNI_DETECT |
18978 PERL_PV_ESCAPE_NONASCII |
18979 PERL_PV_PRETTY_ELLIPSES |
18980 PERL_PV_PRETTY_LTGT |
18981 PERL_PV_PRETTY_NOCLEAR
18983 } else if (k == TRIE) {
18984 /* print the details of the trie in dumpuntil instead, as
18985 * progi->data isn't available here */
18986 const char op = OP(o);
18987 const U32 n = ARG(o);
18988 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
18989 (reg_ac_data *)progi->data->data[n] :
18991 const reg_trie_data * const trie
18992 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
18994 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
18995 DEBUG_TRIE_COMPILE_r({
18997 sv_catpvs(sv, "(JUMP)");
18998 Perl_sv_catpvf(aTHX_ sv,
18999 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
19000 (UV)trie->startstate,
19001 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
19002 (UV)trie->wordcount,
19005 (UV)TRIE_CHARCOUNT(trie),
19006 (UV)trie->uniquecharcount
19009 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
19010 sv_catpvs(sv, "[");
19011 (void) put_charclass_bitmap_innards(sv,
19012 ((IS_ANYOF_TRIE(op))
19014 : TRIE_BITMAP(trie)),
19020 sv_catpvs(sv, "]");
19022 } else if (k == CURLY) {
19023 U32 lo = ARG1(o), hi = ARG2(o);
19024 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
19025 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
19026 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
19027 if (hi == REG_INFTY)
19028 sv_catpvs(sv, "INFTY");
19030 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
19031 sv_catpvs(sv, "}");
19033 else if (k == WHILEM && o->flags) /* Ordinal/of */
19034 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
19035 else if (k == REF || k == OPEN || k == CLOSE
19036 || k == GROUPP || OP(o)==ACCEPT)
19038 AV *name_list= NULL;
19039 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
19040 Perl_sv_catpvf(aTHX_ sv, "%"UVuf, (UV)parno); /* Parenth number */
19041 if ( RXp_PAREN_NAMES(prog) ) {
19042 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19043 } else if ( pRExC_state ) {
19044 name_list= RExC_paren_name_list;
19047 if ( k != REF || (OP(o) < NREF)) {
19048 SV **name= av_fetch(name_list, parno, 0 );
19050 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
19053 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
19054 I32 *nums=(I32*)SvPVX(sv_dat);
19055 SV **name= av_fetch(name_list, nums[0], 0 );
19058 for ( n=0; n<SvIVX(sv_dat); n++ ) {
19059 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
19060 (n ? "," : ""), (IV)nums[n]);
19062 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
19066 if ( k == REF && reginfo) {
19067 U32 n = ARG(o); /* which paren pair */
19068 I32 ln = prog->offs[n].start;
19069 if (prog->lastparen < n || ln == -1)
19070 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
19071 else if (ln == prog->offs[n].end)
19072 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
19074 const char *s = reginfo->strbeg + ln;
19075 Perl_sv_catpvf(aTHX_ sv, ": ");
19076 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
19077 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
19080 } else if (k == GOSUB) {
19081 AV *name_list= NULL;
19082 if ( RXp_PAREN_NAMES(prog) ) {
19083 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19084 } else if ( pRExC_state ) {
19085 name_list= RExC_paren_name_list;
19088 /* Paren and offset */
19089 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
19090 (int)((o + (int)ARG2L(o)) - progi->program) );
19092 SV **name= av_fetch(name_list, ARG(o), 0 );
19094 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
19097 else if (k == LOGICAL)
19098 /* 2: embedded, otherwise 1 */
19099 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
19100 else if (k == ANYOF) {
19101 const U8 flags = ANYOF_FLAGS(o);
19102 bool do_sep = FALSE; /* Do we need to separate various components of
19104 /* Set if there is still an unresolved user-defined property */
19105 SV *unresolved = NULL;
19107 /* Things that are ignored except when the runtime locale is UTF-8 */
19108 SV *only_utf8_locale_invlist = NULL;
19110 /* Code points that don't fit in the bitmap */
19111 SV *nonbitmap_invlist = NULL;
19113 /* And things that aren't in the bitmap, but are small enough to be */
19114 SV* bitmap_range_not_in_bitmap = NULL;
19116 const bool inverted = flags & ANYOF_INVERT;
19118 if (OP(o) == ANYOFL) {
19119 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
19120 sv_catpvs(sv, "{utf8-locale-reqd}");
19122 if (flags & ANYOFL_FOLD) {
19123 sv_catpvs(sv, "{i}");
19127 /* If there is stuff outside the bitmap, get it */
19128 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
19129 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
19131 &only_utf8_locale_invlist,
19132 &nonbitmap_invlist);
19133 /* The non-bitmap data may contain stuff that could fit in the
19134 * bitmap. This could come from a user-defined property being
19135 * finally resolved when this call was done; or much more likely
19136 * because there are matches that require UTF-8 to be valid, and so
19137 * aren't in the bitmap. This is teased apart later */
19138 _invlist_intersection(nonbitmap_invlist,
19140 &bitmap_range_not_in_bitmap);
19141 /* Leave just the things that don't fit into the bitmap */
19142 _invlist_subtract(nonbitmap_invlist,
19144 &nonbitmap_invlist);
19147 /* Obey this flag to add all above-the-bitmap code points */
19148 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
19149 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
19150 NUM_ANYOF_CODE_POINTS,
19154 /* Ready to start outputting. First, the initial left bracket */
19155 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
19157 /* Then all the things that could fit in the bitmap */
19158 do_sep = put_charclass_bitmap_innards(sv,
19160 bitmap_range_not_in_bitmap,
19161 only_utf8_locale_invlist,
19164 /* Can't try inverting for a
19165 * better display if there are
19166 * things that haven't been
19168 unresolved != NULL);
19169 SvREFCNT_dec(bitmap_range_not_in_bitmap);
19171 /* If there are user-defined properties which haven't been defined yet,
19172 * output them. If the result is not to be inverted, it is clearest to
19173 * output them in a separate [] from the bitmap range stuff. If the
19174 * result is to be complemented, we have to show everything in one [],
19175 * as the inversion applies to the whole thing. Use {braces} to
19176 * separate them from anything in the bitmap and anything above the
19180 if (! do_sep) { /* If didn't output anything in the bitmap */
19181 sv_catpvs(sv, "^");
19183 sv_catpvs(sv, "{");
19186 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19188 sv_catsv(sv, unresolved);
19190 sv_catpvs(sv, "}");
19192 do_sep = ! inverted;
19195 /* And, finally, add the above-the-bitmap stuff */
19196 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
19199 /* See if truncation size is overridden */
19200 const STRLEN dump_len = (PL_dump_re_max_len)
19201 ? PL_dump_re_max_len
19204 /* This is output in a separate [] */
19206 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19209 /* And, for easy of understanding, it is shown in the
19210 * uncomplemented form if possible. The one exception being if
19211 * there are unresolved items, where the inversion has to be
19212 * delayed until runtime */
19213 if (inverted && ! unresolved) {
19214 _invlist_invert(nonbitmap_invlist);
19215 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
19218 contents = invlist_contents(nonbitmap_invlist,
19219 FALSE /* output suitable for catsv */
19222 /* If the output is shorter than the permissible maximum, just do it. */
19223 if (SvCUR(contents) <= dump_len) {
19224 sv_catsv(sv, contents);
19227 const char * contents_string = SvPVX(contents);
19228 STRLEN i = dump_len;
19230 /* Otherwise, start at the permissible max and work back to the
19231 * first break possibility */
19232 while (i > 0 && contents_string[i] != ' ') {
19235 if (i == 0) { /* Fail-safe. Use the max if we couldn't
19236 find a legal break */
19240 sv_catpvn(sv, contents_string, i);
19241 sv_catpvs(sv, "...");
19244 SvREFCNT_dec_NN(contents);
19245 SvREFCNT_dec_NN(nonbitmap_invlist);
19248 /* And finally the matching, closing ']' */
19249 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
19251 SvREFCNT_dec(unresolved);
19253 else if (k == POSIXD || k == NPOSIXD) {
19254 U8 index = FLAGS(o) * 2;
19255 if (index < C_ARRAY_LENGTH(anyofs)) {
19256 if (*anyofs[index] != '[') {
19259 sv_catpv(sv, anyofs[index]);
19260 if (*anyofs[index] != '[') {
19265 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
19268 else if (k == BOUND || k == NBOUND) {
19269 /* Must be synced with order of 'bound_type' in regcomp.h */
19270 const char * const bounds[] = {
19271 "", /* Traditional */
19277 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
19278 sv_catpv(sv, bounds[FLAGS(o)]);
19280 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
19281 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
19282 else if (OP(o) == SBOL)
19283 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
19285 /* add on the verb argument if there is one */
19286 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
19287 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
19288 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
19291 PERL_UNUSED_CONTEXT;
19292 PERL_UNUSED_ARG(sv);
19293 PERL_UNUSED_ARG(o);
19294 PERL_UNUSED_ARG(prog);
19295 PERL_UNUSED_ARG(reginfo);
19296 PERL_UNUSED_ARG(pRExC_state);
19297 #endif /* DEBUGGING */
19303 Perl_re_intuit_string(pTHX_ REGEXP * const r)
19304 { /* Assume that RE_INTUIT is set */
19305 struct regexp *const prog = ReANY(r);
19306 GET_RE_DEBUG_FLAGS_DECL;
19308 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
19309 PERL_UNUSED_CONTEXT;
19313 const char * const s = SvPV_nolen_const(RX_UTF8(r)
19314 ? prog->check_utf8 : prog->check_substr);
19316 if (!PL_colorset) reginitcolors();
19317 Perl_re_printf( aTHX_
19318 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
19320 RX_UTF8(r) ? "utf8 " : "",
19321 PL_colors[5],PL_colors[0],
19324 (strlen(s) > 60 ? "..." : ""));
19327 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
19328 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
19334 handles refcounting and freeing the perl core regexp structure. When
19335 it is necessary to actually free the structure the first thing it
19336 does is call the 'free' method of the regexp_engine associated to
19337 the regexp, allowing the handling of the void *pprivate; member
19338 first. (This routine is not overridable by extensions, which is why
19339 the extensions free is called first.)
19341 See regdupe and regdupe_internal if you change anything here.
19343 #ifndef PERL_IN_XSUB_RE
19345 Perl_pregfree(pTHX_ REGEXP *r)
19351 Perl_pregfree2(pTHX_ REGEXP *rx)
19353 struct regexp *const r = ReANY(rx);
19354 GET_RE_DEBUG_FLAGS_DECL;
19356 PERL_ARGS_ASSERT_PREGFREE2;
19358 if (r->mother_re) {
19359 ReREFCNT_dec(r->mother_re);
19361 CALLREGFREE_PVT(rx); /* free the private data */
19362 SvREFCNT_dec(RXp_PAREN_NAMES(r));
19363 Safefree(r->xpv_len_u.xpvlenu_pv);
19366 SvREFCNT_dec(r->anchored_substr);
19367 SvREFCNT_dec(r->anchored_utf8);
19368 SvREFCNT_dec(r->float_substr);
19369 SvREFCNT_dec(r->float_utf8);
19370 Safefree(r->substrs);
19372 RX_MATCH_COPY_FREE(rx);
19373 #ifdef PERL_ANY_COW
19374 SvREFCNT_dec(r->saved_copy);
19377 SvREFCNT_dec(r->qr_anoncv);
19378 if (r->recurse_locinput)
19379 Safefree(r->recurse_locinput);
19380 rx->sv_u.svu_rx = 0;
19385 This is a hacky workaround to the structural issue of match results
19386 being stored in the regexp structure which is in turn stored in
19387 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
19388 could be PL_curpm in multiple contexts, and could require multiple
19389 result sets being associated with the pattern simultaneously, such
19390 as when doing a recursive match with (??{$qr})
19392 The solution is to make a lightweight copy of the regexp structure
19393 when a qr// is returned from the code executed by (??{$qr}) this
19394 lightweight copy doesn't actually own any of its data except for
19395 the starp/end and the actual regexp structure itself.
19401 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
19403 struct regexp *ret;
19404 struct regexp *const r = ReANY(rx);
19405 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
19407 PERL_ARGS_ASSERT_REG_TEMP_COPY;
19410 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
19412 SvOK_off((SV *)ret_x);
19414 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
19415 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
19416 made both spots point to the same regexp body.) */
19417 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
19418 assert(!SvPVX(ret_x));
19419 ret_x->sv_u.svu_rx = temp->sv_any;
19420 temp->sv_any = NULL;
19421 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
19422 SvREFCNT_dec_NN(temp);
19423 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
19424 ing below will not set it. */
19425 SvCUR_set(ret_x, SvCUR(rx));
19428 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
19429 sv_force_normal(sv) is called. */
19431 ret = ReANY(ret_x);
19433 SvFLAGS(ret_x) |= SvUTF8(rx);
19434 /* We share the same string buffer as the original regexp, on which we
19435 hold a reference count, incremented when mother_re is set below.
19436 The string pointer is copied here, being part of the regexp struct.
19438 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
19439 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
19441 const I32 npar = r->nparens+1;
19442 Newx(ret->offs, npar, regexp_paren_pair);
19443 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19446 Newx(ret->substrs, 1, struct reg_substr_data);
19447 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19449 SvREFCNT_inc_void(ret->anchored_substr);
19450 SvREFCNT_inc_void(ret->anchored_utf8);
19451 SvREFCNT_inc_void(ret->float_substr);
19452 SvREFCNT_inc_void(ret->float_utf8);
19454 /* check_substr and check_utf8, if non-NULL, point to either their
19455 anchored or float namesakes, and don't hold a second reference. */
19457 RX_MATCH_COPIED_off(ret_x);
19458 #ifdef PERL_ANY_COW
19459 ret->saved_copy = NULL;
19461 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
19462 SvREFCNT_inc_void(ret->qr_anoncv);
19463 if (r->recurse_locinput)
19464 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19470 /* regfree_internal()
19472 Free the private data in a regexp. This is overloadable by
19473 extensions. Perl takes care of the regexp structure in pregfree(),
19474 this covers the *pprivate pointer which technically perl doesn't
19475 know about, however of course we have to handle the
19476 regexp_internal structure when no extension is in use.
19478 Note this is called before freeing anything in the regexp
19483 Perl_regfree_internal(pTHX_ REGEXP * const rx)
19485 struct regexp *const r = ReANY(rx);
19486 RXi_GET_DECL(r,ri);
19487 GET_RE_DEBUG_FLAGS_DECL;
19489 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
19495 SV *dsv= sv_newmortal();
19496 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
19497 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
19498 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
19499 PL_colors[4],PL_colors[5],s);
19502 #ifdef RE_TRACK_PATTERN_OFFSETS
19504 Safefree(ri->u.offsets); /* 20010421 MJD */
19506 if (ri->code_blocks) {
19508 for (n = 0; n < ri->num_code_blocks; n++)
19509 SvREFCNT_dec(ri->code_blocks[n].src_regex);
19510 Safefree(ri->code_blocks);
19514 int n = ri->data->count;
19517 /* If you add a ->what type here, update the comment in regcomp.h */
19518 switch (ri->data->what[n]) {
19524 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
19527 Safefree(ri->data->data[n]);
19533 { /* Aho Corasick add-on structure for a trie node.
19534 Used in stclass optimization only */
19536 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
19537 #ifdef USE_ITHREADS
19541 refcount = --aho->refcount;
19544 PerlMemShared_free(aho->states);
19545 PerlMemShared_free(aho->fail);
19546 /* do this last!!!! */
19547 PerlMemShared_free(ri->data->data[n]);
19548 /* we should only ever get called once, so
19549 * assert as much, and also guard the free
19550 * which /might/ happen twice. At the least
19551 * it will make code anlyzers happy and it
19552 * doesn't cost much. - Yves */
19553 assert(ri->regstclass);
19554 if (ri->regstclass) {
19555 PerlMemShared_free(ri->regstclass);
19556 ri->regstclass = 0;
19563 /* trie structure. */
19565 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
19566 #ifdef USE_ITHREADS
19570 refcount = --trie->refcount;
19573 PerlMemShared_free(trie->charmap);
19574 PerlMemShared_free(trie->states);
19575 PerlMemShared_free(trie->trans);
19577 PerlMemShared_free(trie->bitmap);
19579 PerlMemShared_free(trie->jump);
19580 PerlMemShared_free(trie->wordinfo);
19581 /* do this last!!!! */
19582 PerlMemShared_free(ri->data->data[n]);
19587 Perl_croak(aTHX_ "panic: regfree data code '%c'",
19588 ri->data->what[n]);
19591 Safefree(ri->data->what);
19592 Safefree(ri->data);
19598 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
19599 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
19600 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
19603 re_dup_guts - duplicate a regexp.
19605 This routine is expected to clone a given regexp structure. It is only
19606 compiled under USE_ITHREADS.
19608 After all of the core data stored in struct regexp is duplicated
19609 the regexp_engine.dupe method is used to copy any private data
19610 stored in the *pprivate pointer. This allows extensions to handle
19611 any duplication it needs to do.
19613 See pregfree() and regfree_internal() if you change anything here.
19615 #if defined(USE_ITHREADS)
19616 #ifndef PERL_IN_XSUB_RE
19618 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
19622 const struct regexp *r = ReANY(sstr);
19623 struct regexp *ret = ReANY(dstr);
19625 PERL_ARGS_ASSERT_RE_DUP_GUTS;
19627 npar = r->nparens+1;
19628 Newx(ret->offs, npar, regexp_paren_pair);
19629 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19631 if (ret->substrs) {
19632 /* Do it this way to avoid reading from *r after the StructCopy().
19633 That way, if any of the sv_dup_inc()s dislodge *r from the L1
19634 cache, it doesn't matter. */
19635 const bool anchored = r->check_substr
19636 ? r->check_substr == r->anchored_substr
19637 : r->check_utf8 == r->anchored_utf8;
19638 Newx(ret->substrs, 1, struct reg_substr_data);
19639 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19641 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
19642 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
19643 ret->float_substr = sv_dup_inc(ret->float_substr, param);
19644 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
19646 /* check_substr and check_utf8, if non-NULL, point to either their
19647 anchored or float namesakes, and don't hold a second reference. */
19649 if (ret->check_substr) {
19651 assert(r->check_utf8 == r->anchored_utf8);
19652 ret->check_substr = ret->anchored_substr;
19653 ret->check_utf8 = ret->anchored_utf8;
19655 assert(r->check_substr == r->float_substr);
19656 assert(r->check_utf8 == r->float_utf8);
19657 ret->check_substr = ret->float_substr;
19658 ret->check_utf8 = ret->float_utf8;
19660 } else if (ret->check_utf8) {
19662 ret->check_utf8 = ret->anchored_utf8;
19664 ret->check_utf8 = ret->float_utf8;
19669 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
19670 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
19671 if (r->recurse_locinput)
19672 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19675 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
19677 if (RX_MATCH_COPIED(dstr))
19678 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
19680 ret->subbeg = NULL;
19681 #ifdef PERL_ANY_COW
19682 ret->saved_copy = NULL;
19685 /* Whether mother_re be set or no, we need to copy the string. We
19686 cannot refrain from copying it when the storage points directly to
19687 our mother regexp, because that's
19688 1: a buffer in a different thread
19689 2: something we no longer hold a reference on
19690 so we need to copy it locally. */
19691 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
19692 ret->mother_re = NULL;
19694 #endif /* PERL_IN_XSUB_RE */
19699 This is the internal complement to regdupe() which is used to copy
19700 the structure pointed to by the *pprivate pointer in the regexp.
19701 This is the core version of the extension overridable cloning hook.
19702 The regexp structure being duplicated will be copied by perl prior
19703 to this and will be provided as the regexp *r argument, however
19704 with the /old/ structures pprivate pointer value. Thus this routine
19705 may override any copying normally done by perl.
19707 It returns a pointer to the new regexp_internal structure.
19711 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
19714 struct regexp *const r = ReANY(rx);
19715 regexp_internal *reti;
19717 RXi_GET_DECL(r,ri);
19719 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
19723 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
19724 char, regexp_internal);
19725 Copy(ri->program, reti->program, len+1, regnode);
19728 reti->num_code_blocks = ri->num_code_blocks;
19729 if (ri->code_blocks) {
19731 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
19732 struct reg_code_block);
19733 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
19734 struct reg_code_block);
19735 for (n = 0; n < ri->num_code_blocks; n++)
19736 reti->code_blocks[n].src_regex = (REGEXP*)
19737 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
19740 reti->code_blocks = NULL;
19742 reti->regstclass = NULL;
19745 struct reg_data *d;
19746 const int count = ri->data->count;
19749 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
19750 char, struct reg_data);
19751 Newx(d->what, count, U8);
19754 for (i = 0; i < count; i++) {
19755 d->what[i] = ri->data->what[i];
19756 switch (d->what[i]) {
19757 /* see also regcomp.h and regfree_internal() */
19758 case 'a': /* actually an AV, but the dup function is identical. */
19762 case 'u': /* actually an HV, but the dup function is identical. */
19763 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
19766 /* This is cheating. */
19767 Newx(d->data[i], 1, regnode_ssc);
19768 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
19769 reti->regstclass = (regnode*)d->data[i];
19772 /* Trie stclasses are readonly and can thus be shared
19773 * without duplication. We free the stclass in pregfree
19774 * when the corresponding reg_ac_data struct is freed.
19776 reti->regstclass= ri->regstclass;
19780 ((reg_trie_data*)ri->data->data[i])->refcount++;
19785 d->data[i] = ri->data->data[i];
19788 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
19789 ri->data->what[i]);
19798 reti->name_list_idx = ri->name_list_idx;
19800 #ifdef RE_TRACK_PATTERN_OFFSETS
19801 if (ri->u.offsets) {
19802 Newx(reti->u.offsets, 2*len+1, U32);
19803 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
19806 SetProgLen(reti,len);
19809 return (void*)reti;
19812 #endif /* USE_ITHREADS */
19814 #ifndef PERL_IN_XSUB_RE
19817 - regnext - dig the "next" pointer out of a node
19820 Perl_regnext(pTHX_ regnode *p)
19827 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
19828 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
19829 (int)OP(p), (int)REGNODE_MAX);
19832 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
19841 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
19844 STRLEN l1 = strlen(pat1);
19845 STRLEN l2 = strlen(pat2);
19848 const char *message;
19850 PERL_ARGS_ASSERT_RE_CROAK2;
19856 Copy(pat1, buf, l1 , char);
19857 Copy(pat2, buf + l1, l2 , char);
19858 buf[l1 + l2] = '\n';
19859 buf[l1 + l2 + 1] = '\0';
19860 va_start(args, pat2);
19861 msv = vmess(buf, &args);
19863 message = SvPV_const(msv,l1);
19866 Copy(message, buf, l1 , char);
19867 /* l1-1 to avoid \n */
19868 Perl_croak(aTHX_ "%"UTF8f, UTF8fARG(utf8, l1-1, buf));
19871 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
19873 #ifndef PERL_IN_XSUB_RE
19875 Perl_save_re_context(pTHX)
19880 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
19883 const REGEXP * const rx = PM_GETRE(PL_curpm);
19885 nparens = RX_NPARENS(rx);
19888 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
19889 * that PL_curpm will be null, but that utf8.pm and the modules it
19890 * loads will only use $1..$3.
19891 * The t/porting/re_context.t test file checks this assumption.
19896 for (i = 1; i <= nparens; i++) {
19897 char digits[TYPE_CHARS(long)];
19898 const STRLEN len = my_snprintf(digits, sizeof(digits),
19900 GV *const *const gvp
19901 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
19904 GV * const gv = *gvp;
19905 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
19915 S_put_code_point(pTHX_ SV *sv, UV c)
19917 PERL_ARGS_ASSERT_PUT_CODE_POINT;
19920 Perl_sv_catpvf(aTHX_ sv, "\\x{%04"UVXf"}", c);
19922 else if (isPRINT(c)) {
19923 const char string = (char) c;
19925 /* We use {phrase} as metanotation in the class, so also escape literal
19927 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
19928 sv_catpvs(sv, "\\");
19929 sv_catpvn(sv, &string, 1);
19931 else if (isMNEMONIC_CNTRL(c)) {
19932 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
19935 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
19939 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
19942 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
19944 /* Appends to 'sv' a displayable version of the range of code points from
19945 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
19946 * that have them, when they occur at the beginning or end of the range.
19947 * It uses hex to output the remaining code points, unless 'allow_literals'
19948 * is true, in which case the printable ASCII ones are output as-is (though
19949 * some of these will be escaped by put_code_point()).
19951 * NOTE: This is designed only for printing ranges of code points that fit
19952 * inside an ANYOF bitmap. Higher code points are simply suppressed
19955 const unsigned int min_range_count = 3;
19957 assert(start <= end);
19959 PERL_ARGS_ASSERT_PUT_RANGE;
19961 while (start <= end) {
19963 const char * format;
19965 if (end - start < min_range_count) {
19967 /* Output chars individually when they occur in short ranges */
19968 for (; start <= end; start++) {
19969 put_code_point(sv, start);
19974 /* If permitted by the input options, and there is a possibility that
19975 * this range contains a printable literal, look to see if there is
19977 if (allow_literals && start <= MAX_PRINT_A) {
19979 /* If the character at the beginning of the range isn't an ASCII
19980 * printable, effectively split the range into two parts:
19981 * 1) the portion before the first such printable,
19983 * and output them separately. */
19984 if (! isPRINT_A(start)) {
19985 UV temp_end = start + 1;
19987 /* There is no point looking beyond the final possible
19988 * printable, in MAX_PRINT_A */
19989 UV max = MIN(end, MAX_PRINT_A);
19991 while (temp_end <= max && ! isPRINT_A(temp_end)) {
19995 /* Here, temp_end points to one beyond the first printable if
19996 * found, or to one beyond 'max' if not. If none found, make
19997 * sure that we use the entire range */
19998 if (temp_end > MAX_PRINT_A) {
19999 temp_end = end + 1;
20002 /* Output the first part of the split range: the part that
20003 * doesn't have printables, with the parameter set to not look
20004 * for literals (otherwise we would infinitely recurse) */
20005 put_range(sv, start, temp_end - 1, FALSE);
20007 /* The 2nd part of the range (if any) starts here. */
20010 /* We do a continue, instead of dropping down, because even if
20011 * the 2nd part is non-empty, it could be so short that we want
20012 * to output it as individual characters, as tested for at the
20013 * top of this loop. */
20017 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
20018 * output a sub-range of just the digits or letters, then process
20019 * the remaining portion as usual. */
20020 if (isALPHANUMERIC_A(start)) {
20021 UV mask = (isDIGIT_A(start))
20026 UV temp_end = start + 1;
20028 /* Find the end of the sub-range that includes just the
20029 * characters in the same class as the first character in it */
20030 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
20035 /* For short ranges, don't duplicate the code above to output
20036 * them; just call recursively */
20037 if (temp_end - start < min_range_count) {
20038 put_range(sv, start, temp_end, FALSE);
20040 else { /* Output as a range */
20041 put_code_point(sv, start);
20042 sv_catpvs(sv, "-");
20043 put_code_point(sv, temp_end);
20045 start = temp_end + 1;
20049 /* We output any other printables as individual characters */
20050 if (isPUNCT_A(start) || isSPACE_A(start)) {
20051 while (start <= end && (isPUNCT_A(start)
20052 || isSPACE_A(start)))
20054 put_code_point(sv, start);
20059 } /* End of looking for literals */
20061 /* Here is not to output as a literal. Some control characters have
20062 * mnemonic names. Split off any of those at the beginning and end of
20063 * the range to print mnemonically. It isn't possible for many of
20064 * these to be in a row, so this won't overwhelm with output */
20066 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
20068 while (isMNEMONIC_CNTRL(start) && start <= end) {
20069 put_code_point(sv, start);
20073 /* If this didn't take care of the whole range ... */
20074 if (start <= end) {
20076 /* Look backwards from the end to find the final non-mnemonic
20079 while (isMNEMONIC_CNTRL(temp_end)) {
20083 /* And separately output the interior range that doesn't start
20084 * or end with mnemonics */
20085 put_range(sv, start, temp_end, FALSE);
20087 /* Then output the mnemonic trailing controls */
20088 start = temp_end + 1;
20089 while (start <= end) {
20090 put_code_point(sv, start);
20097 /* As a final resort, output the range or subrange as hex. */
20099 this_end = (end < NUM_ANYOF_CODE_POINTS)
20101 : NUM_ANYOF_CODE_POINTS - 1;
20102 #if NUM_ANYOF_CODE_POINTS > 256
20103 format = (this_end < 256)
20104 ? "\\x%02"UVXf"-\\x%02"UVXf""
20105 : "\\x{%04"UVXf"}-\\x{%04"UVXf"}";
20107 format = "\\x%02"UVXf"-\\x%02"UVXf"";
20109 GCC_DIAG_IGNORE(-Wformat-nonliteral);
20110 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
20117 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
20119 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
20123 bool allow_literals = TRUE;
20125 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
20127 /* Generally, it is more readable if printable characters are output as
20128 * literals, but if a range (nearly) spans all of them, it's best to output
20129 * it as a single range. This code will use a single range if all but 2
20130 * ASCII printables are in it */
20131 invlist_iterinit(invlist);
20132 while (invlist_iternext(invlist, &start, &end)) {
20134 /* If the range starts beyond the final printable, it doesn't have any
20136 if (start > MAX_PRINT_A) {
20140 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
20141 * all but two, the range must start and end no later than 2 from
20143 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
20144 if (end > MAX_PRINT_A) {
20150 if (end - start >= MAX_PRINT_A - ' ' - 2) {
20151 allow_literals = FALSE;
20156 invlist_iterfinish(invlist);
20158 /* Here we have figured things out. Output each range */
20159 invlist_iterinit(invlist);
20160 while (invlist_iternext(invlist, &start, &end)) {
20161 if (start >= NUM_ANYOF_CODE_POINTS) {
20164 put_range(sv, start, end, allow_literals);
20166 invlist_iterfinish(invlist);
20172 S_put_charclass_bitmap_innards_common(pTHX_
20173 SV* invlist, /* The bitmap */
20174 SV* posixes, /* Under /l, things like [:word:], \S */
20175 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
20176 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
20177 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
20178 const bool invert /* Is the result to be inverted? */
20181 /* Create and return an SV containing a displayable version of the bitmap
20182 * and associated information determined by the input parameters. If the
20183 * output would have been only the inversion indicator '^', NULL is instead
20188 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
20191 output = newSVpvs("^");
20194 output = newSVpvs("");
20197 /* First, the code points in the bitmap that are unconditionally there */
20198 put_charclass_bitmap_innards_invlist(output, invlist);
20200 /* Traditionally, these have been placed after the main code points */
20202 sv_catsv(output, posixes);
20205 if (only_utf8 && _invlist_len(only_utf8)) {
20206 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
20207 put_charclass_bitmap_innards_invlist(output, only_utf8);
20210 if (not_utf8 && _invlist_len(not_utf8)) {
20211 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
20212 put_charclass_bitmap_innards_invlist(output, not_utf8);
20215 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
20216 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
20217 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
20219 /* This is the only list in this routine that can legally contain code
20220 * points outside the bitmap range. The call just above to
20221 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
20222 * output them here. There's about a half-dozen possible, and none in
20223 * contiguous ranges longer than 2 */
20224 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20226 SV* above_bitmap = NULL;
20228 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
20230 invlist_iterinit(above_bitmap);
20231 while (invlist_iternext(above_bitmap, &start, &end)) {
20234 for (i = start; i <= end; i++) {
20235 put_code_point(output, i);
20238 invlist_iterfinish(above_bitmap);
20239 SvREFCNT_dec_NN(above_bitmap);
20243 if (invert && SvCUR(output) == 1) {
20251 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
20253 SV *nonbitmap_invlist,
20254 SV *only_utf8_locale_invlist,
20255 const regnode * const node,
20256 const bool force_as_is_display)
20258 /* Appends to 'sv' a displayable version of the innards of the bracketed
20259 * character class defined by the other arguments:
20260 * 'bitmap' points to the bitmap.
20261 * 'nonbitmap_invlist' is an inversion list of the code points that are in
20262 * the bitmap range, but for some reason aren't in the bitmap; NULL if
20263 * none. The reasons for this could be that they require some
20264 * condition such as the target string being or not being in UTF-8
20265 * (under /d), or because they came from a user-defined property that
20266 * was not resolved at the time of the regex compilation (under /u)
20267 * 'only_utf8_locale_invlist' is an inversion list of the code points that
20268 * are valid only if the runtime locale is a UTF-8 one; NULL if none
20269 * 'node' is the regex pattern node. It is needed only when the above two
20270 * parameters are not null, and is passed so that this routine can
20271 * tease apart the various reasons for them.
20272 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
20273 * to invert things to see if that leads to a cleaner display. If
20274 * FALSE, this routine is free to use its judgment about doing this.
20276 * It returns TRUE if there was actually something output. (It may be that
20277 * the bitmap, etc is empty.)
20279 * When called for outputting the bitmap of a non-ANYOF node, just pass the
20280 * bitmap, with the succeeding parameters set to NULL, and the final one to
20284 /* In general, it tries to display the 'cleanest' representation of the
20285 * innards, choosing whether to display them inverted or not, regardless of
20286 * whether the class itself is to be inverted. However, there are some
20287 * cases where it can't try inverting, as what actually matches isn't known
20288 * until runtime, and hence the inversion isn't either. */
20289 bool inverting_allowed = ! force_as_is_display;
20292 STRLEN orig_sv_cur = SvCUR(sv);
20294 SV* invlist; /* Inversion list we accumulate of code points that
20295 are unconditionally matched */
20296 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
20298 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
20300 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
20301 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
20304 SV* as_is_display; /* The output string when we take the inputs
20306 SV* inverted_display; /* The output string when we invert the inputs */
20308 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
20310 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
20312 /* We are biased in favor of displaying things without them being inverted,
20313 * as that is generally easier to understand */
20314 const int bias = 5;
20316 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
20318 /* Start off with whatever code points are passed in. (We clone, so we
20319 * don't change the caller's list) */
20320 if (nonbitmap_invlist) {
20321 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
20322 invlist = invlist_clone(nonbitmap_invlist);
20324 else { /* Worst case size is every other code point is matched */
20325 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
20329 if (OP(node) == ANYOFD) {
20331 /* This flag indicates that the code points below 0x100 in the
20332 * nonbitmap list are precisely the ones that match only when the
20333 * target is UTF-8 (they should all be non-ASCII). */
20334 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
20336 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
20337 _invlist_subtract(invlist, only_utf8, &invlist);
20340 /* And this flag for matching all non-ASCII 0xFF and below */
20341 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
20343 not_utf8 = invlist_clone(PL_UpperLatin1);
20346 else if (OP(node) == ANYOFL) {
20348 /* If either of these flags are set, what matches isn't
20349 * determinable except during execution, so don't know enough here
20351 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
20352 inverting_allowed = FALSE;
20355 /* What the posix classes match also varies at runtime, so these
20356 * will be output symbolically. */
20357 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
20360 posixes = newSVpvs("");
20361 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
20362 if (ANYOF_POSIXL_TEST(node,i)) {
20363 sv_catpv(posixes, anyofs[i]);
20370 /* Accumulate the bit map into the unconditional match list */
20371 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
20372 if (BITMAP_TEST(bitmap, i)) {
20374 for (; i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i); i++) {
20377 invlist = _add_range_to_invlist(invlist, start, i-1);
20381 /* Make sure that the conditional match lists don't have anything in them
20382 * that match unconditionally; otherwise the output is quite confusing.
20383 * This could happen if the code that populates these misses some
20386 _invlist_subtract(only_utf8, invlist, &only_utf8);
20389 _invlist_subtract(not_utf8, invlist, ¬_utf8);
20392 if (only_utf8_locale_invlist) {
20394 /* Since this list is passed in, we have to make a copy before
20396 only_utf8_locale = invlist_clone(only_utf8_locale_invlist);
20398 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
20400 /* And, it can get really weird for us to try outputting an inverted
20401 * form of this list when it has things above the bitmap, so don't even
20403 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20404 inverting_allowed = FALSE;
20408 /* Calculate what the output would be if we take the input as-is */
20409 as_is_display = put_charclass_bitmap_innards_common(invlist,
20416 /* If have to take the output as-is, just do that */
20417 if (! inverting_allowed) {
20418 if (as_is_display) {
20419 sv_catsv(sv, as_is_display);
20420 SvREFCNT_dec_NN(as_is_display);
20423 else { /* But otherwise, create the output again on the inverted input, and
20424 use whichever version is shorter */
20426 int inverted_bias, as_is_bias;
20428 /* We will apply our bias to whichever of the the results doesn't have
20438 inverted_bias = bias;
20441 /* Now invert each of the lists that contribute to the output,
20442 * excluding from the result things outside the possible range */
20444 /* For the unconditional inversion list, we have to add in all the
20445 * conditional code points, so that when inverted, they will be gone
20447 _invlist_union(only_utf8, invlist, &invlist);
20448 _invlist_union(not_utf8, invlist, &invlist);
20449 _invlist_union(only_utf8_locale, invlist, &invlist);
20450 _invlist_invert(invlist);
20451 _invlist_intersection(invlist, PL_InBitmap, &invlist);
20454 _invlist_invert(only_utf8);
20455 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
20457 else if (not_utf8) {
20459 /* If a code point matches iff the target string is not in UTF-8,
20460 * then complementing the result has it not match iff not in UTF-8,
20461 * which is the same thing as matching iff it is UTF-8. */
20462 only_utf8 = not_utf8;
20466 if (only_utf8_locale) {
20467 _invlist_invert(only_utf8_locale);
20468 _invlist_intersection(only_utf8_locale,
20470 &only_utf8_locale);
20473 inverted_display = put_charclass_bitmap_innards_common(
20478 only_utf8_locale, invert);
20480 /* Use the shortest representation, taking into account our bias
20481 * against showing it inverted */
20482 if ( inverted_display
20483 && ( ! as_is_display
20484 || ( SvCUR(inverted_display) + inverted_bias
20485 < SvCUR(as_is_display) + as_is_bias)))
20487 sv_catsv(sv, inverted_display);
20489 else if (as_is_display) {
20490 sv_catsv(sv, as_is_display);
20493 SvREFCNT_dec(as_is_display);
20494 SvREFCNT_dec(inverted_display);
20497 SvREFCNT_dec_NN(invlist);
20498 SvREFCNT_dec(only_utf8);
20499 SvREFCNT_dec(not_utf8);
20500 SvREFCNT_dec(posixes);
20501 SvREFCNT_dec(only_utf8_locale);
20503 return SvCUR(sv) > orig_sv_cur;
20506 #define CLEAR_OPTSTART \
20507 if (optstart) STMT_START { \
20508 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
20509 " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
20513 #define DUMPUNTIL(b,e) \
20515 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
20517 STATIC const regnode *
20518 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
20519 const regnode *last, const regnode *plast,
20520 SV* sv, I32 indent, U32 depth)
20522 U8 op = PSEUDO; /* Arbitrary non-END op. */
20523 const regnode *next;
20524 const regnode *optstart= NULL;
20526 RXi_GET_DECL(r,ri);
20527 GET_RE_DEBUG_FLAGS_DECL;
20529 PERL_ARGS_ASSERT_DUMPUNTIL;
20531 #ifdef DEBUG_DUMPUNTIL
20532 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n",indent,node-start,
20533 last ? last-start : 0,plast ? plast-start : 0);
20536 if (plast && plast < last)
20539 while (PL_regkind[op] != END && (!last || node < last)) {
20541 /* While that wasn't END last time... */
20544 if (op == CLOSE || op == WHILEM)
20546 next = regnext((regnode *)node);
20549 if (OP(node) == OPTIMIZED) {
20550 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
20557 regprop(r, sv, node, NULL, NULL);
20558 Perl_re_printf( aTHX_ "%4"IVdf":%*s%s", (IV)(node - start),
20559 (int)(2*indent + 1), "", SvPVX_const(sv));
20561 if (OP(node) != OPTIMIZED) {
20562 if (next == NULL) /* Next ptr. */
20563 Perl_re_printf( aTHX_ " (0)");
20564 else if (PL_regkind[(U8)op] == BRANCH
20565 && PL_regkind[OP(next)] != BRANCH )
20566 Perl_re_printf( aTHX_ " (FAIL)");
20568 Perl_re_printf( aTHX_ " (%"IVdf")", (IV)(next - start));
20569 Perl_re_printf( aTHX_ "\n");
20573 if (PL_regkind[(U8)op] == BRANCHJ) {
20576 const regnode *nnode = (OP(next) == LONGJMP
20577 ? regnext((regnode *)next)
20579 if (last && nnode > last)
20581 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
20584 else if (PL_regkind[(U8)op] == BRANCH) {
20586 DUMPUNTIL(NEXTOPER(node), next);
20588 else if ( PL_regkind[(U8)op] == TRIE ) {
20589 const regnode *this_trie = node;
20590 const char op = OP(node);
20591 const U32 n = ARG(node);
20592 const reg_ac_data * const ac = op>=AHOCORASICK ?
20593 (reg_ac_data *)ri->data->data[n] :
20595 const reg_trie_data * const trie =
20596 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
20598 AV *const trie_words
20599 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
20601 const regnode *nextbranch= NULL;
20604 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
20605 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
20607 Perl_re_indentf( aTHX_ "%s ",
20610 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
20611 SvCUR(*elem_ptr), 60,
20612 PL_colors[0], PL_colors[1],
20614 ? PERL_PV_ESCAPE_UNI
20616 | PERL_PV_PRETTY_ELLIPSES
20617 | PERL_PV_PRETTY_LTGT
20622 U16 dist= trie->jump[word_idx+1];
20623 Perl_re_printf( aTHX_ "(%"UVuf")\n",
20624 (UV)((dist ? this_trie + dist : next) - start));
20627 nextbranch= this_trie + trie->jump[0];
20628 DUMPUNTIL(this_trie + dist, nextbranch);
20630 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
20631 nextbranch= regnext((regnode *)nextbranch);
20633 Perl_re_printf( aTHX_ "\n");
20636 if (last && next > last)
20641 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
20642 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
20643 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
20645 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
20647 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
20649 else if ( op == PLUS || op == STAR) {
20650 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
20652 else if (PL_regkind[(U8)op] == ANYOF) {
20653 /* arglen 1 + class block */
20654 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
20655 ? ANYOF_POSIXL_SKIP
20657 node = NEXTOPER(node);
20659 else if (PL_regkind[(U8)op] == EXACT) {
20660 /* Literal string, where present. */
20661 node += NODE_SZ_STR(node) - 1;
20662 node = NEXTOPER(node);
20665 node = NEXTOPER(node);
20666 node += regarglen[(U8)op];
20668 if (op == CURLYX || op == OPEN)
20672 #ifdef DEBUG_DUMPUNTIL
20673 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
20678 #endif /* DEBUGGING */
20681 * ex: set ts=8 sts=4 sw=4 et: