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 'src' with the one in 'dest'. It steals
8436 * the list from 'src', so 'src' is made to have a NULL list. This is
8437 * 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. If the output is the complement of 'b', the
8989 * union is all possible code points, and we need not even look at 'a'.
8990 * It's easiest to create a new inversion list that matches everything.
8993 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
8995 /* If the output didn't exist, just point it at the new list */
8996 if (*output == NULL) {
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 it 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 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9046 /* Here, 'a' is empty (and b is not). That means the union will come
9047 * entirely from 'b'. If the output is not to overwrite 'a', we can
9048 * just return what's in 'b'. */
9051 /* If the output is to overwrite 'b', it's already in 'b', but
9052 * otherwise we have to copy 'b' to the output */
9054 *output = invlist_clone(b);
9057 /* And if the output is to be the inversion of 'b', do that */
9059 _invlist_invert(*output);
9065 /* Here, 'a', which is empty or even NULL, is to be overwritten by the
9066 * output, which will either be 'b' or the complement of 'b' */
9069 *output = invlist_clone(b);
9072 u = invlist_clone(b);
9073 invlist_replace_list_destroys_src(*output, u);
9078 _invlist_invert(*output);
9084 /* Here both lists exist and are non-empty */
9085 array_a = invlist_array(a);
9086 array_b = invlist_array(b);
9088 /* If are to take the union of 'a' with the complement of b, set it
9089 * up so are looking at b's complement. */
9092 /* To complement, we invert: if the first element is 0, remove it. To
9093 * do this, we just pretend the array starts one later */
9094 if (array_b[0] == 0) {
9100 /* But if the first element is not zero, we pretend the list starts
9101 * at the 0 that is always stored immediately before the array. */
9107 /* Size the union for the worst case: that the sets are completely
9109 u = _new_invlist(len_a + len_b);
9111 /* Will contain U+0000 if either component does */
9112 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9113 || (len_b > 0 && array_b[0] == 0));
9115 /* Go through each input list item by item, stopping when exhausted one of
9117 while (i_a < len_a && i_b < len_b) {
9118 UV cp; /* The element to potentially add to the union's array */
9119 bool cp_in_set; /* is it in the the input list's set or not */
9121 /* We need to take one or the other of the two inputs for the union.
9122 * Since we are merging two sorted lists, we take the smaller of the
9123 * next items. In case of a tie, we take first the one that is in its
9124 * set. If we first took the one not in its set, it would decrement
9125 * the count, possibly to 0 which would cause it to be output as ending
9126 * the range, and the next time through we would take the same number,
9127 * and output it again as beginning the next range. By doing it the
9128 * opposite way, there is no possibility that the count will be
9129 * momentarily decremented to 0, and thus the two adjoining ranges will
9130 * be seamlessly merged. (In a tie and both are in the set or both not
9131 * in the set, it doesn't matter which we take first.) */
9132 if ( array_a[i_a] < array_b[i_b]
9133 || ( array_a[i_a] == array_b[i_b]
9134 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9136 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9137 cp = array_a[i_a++];
9140 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9141 cp = array_b[i_b++];
9144 /* Here, have chosen which of the two inputs to look at. Only output
9145 * if the running count changes to/from 0, which marks the
9146 * beginning/end of a range that's in the set */
9149 array_u[i_u++] = cp;
9156 array_u[i_u++] = cp;
9162 /* The loop above increments the index into exactly one of the input lists
9163 * each iteration, and ends when either index gets to its list end. That
9164 * means the other index is lower than its end, and so something is
9165 * remaining in that one. We decrement 'count', as explained below, if
9166 * that list is in its set. (i_a and i_b each currently index the element
9167 * beyond the one we care about.) */
9168 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9169 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9174 /* Above we decremented 'count' if the list that had unexamined elements in
9175 * it was in its set. This has made it so that 'count' being non-zero
9176 * means there isn't anything left to output; and 'count' equal to 0 means
9177 * that what is left to output is precisely that which is left in the
9178 * non-exhausted input list.
9180 * To see why, note first that the exhausted input obviously has nothing
9181 * left to add to the union. If it was in its set at its end, that means
9182 * the set extends from here to the platform's infinity, and hence so does
9183 * the union and the non-exhausted set is irrelevant. The exhausted set
9184 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9185 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9186 * 'count' remains at 1. This is consistent with the decremented 'count'
9187 * != 0 meaning there's nothing left to add to the union.
9189 * But if the exhausted input wasn't in its set, it contributed 0 to
9190 * 'count', and the rest of the union will be whatever the other input is.
9191 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9192 * otherwise it gets decremented to 0. This is consistent with 'count'
9193 * == 0 meaning the remainder of the union is whatever is left in the
9194 * non-exhausted list. */
9199 IV copy_count = len_a - i_a;
9200 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9201 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9203 else { /* The non-exhausted input is b */
9204 copy_count = len_b - i_b;
9205 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9207 len_u = i_u + copy_count;
9210 /* Set the result to the final length, which can change the pointer to
9211 * array_u, so re-find it. (Note that it is unlikely that this will
9212 * change, as we are shrinking the space, not enlarging it) */
9213 if (len_u != _invlist_len(u)) {
9214 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9216 array_u = invlist_array(u);
9219 /* If the output is not to overwrite either of the inputs, just return the
9220 * calculated union */
9221 if (a != *output && b != *output) {
9225 /* Here, the output is to be the same as one of the input scalars,
9226 * hence replacing it. The simple thing to do is to free the input
9227 * scalar, making it instead be the output one. But experience has
9228 * shown [perl #127392] that if the input is a mortal, we can get a
9229 * huge build-up of these during regex compilation before they get
9230 * freed. So for that case, replace just the input's interior with
9231 * the union's, and then free the union */
9233 assert(! invlist_is_iterating(*output));
9235 if (! SvTEMP(*output)) {
9236 SvREFCNT_dec_NN(*output);
9240 invlist_replace_list_destroys_src(*output, u);
9249 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9250 const bool complement_b, SV** i)
9252 /* Take the intersection of two inversion lists and point <i> to it. *i
9253 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
9254 * the reference count to that list will be decremented if not already a
9255 * temporary (mortal); otherwise just its contents will be modified to be
9256 * the intersection. The first list, <a>, may be NULL, in which case an
9257 * empty list is returned. If <complement_b> is TRUE, the result will be
9258 * the intersection of <a> and the complement (or inversion) of <b> instead
9261 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9262 * Richard Gillam, published by Addison-Wesley, and explained at some
9263 * length there. The preface says to incorporate its examples into your
9264 * code at your own risk. In fact, it had bugs
9266 * The algorithm is like a merge sort, and is essentially the same as the
9270 const UV* array_a; /* a's array */
9272 UV len_a; /* length of a's array */
9275 SV* r; /* the resulting intersection */
9279 UV i_a = 0; /* current index into a's array */
9283 /* running count of how many of the two inputs are postitioned at ranges
9284 * that are in their sets. As explained in the algorithm source book,
9285 * items are stopped accumulating and are output when the count changes
9286 * to/from 2. The count is incremented when we start a range that's in an
9287 * input's set, and decremented when we start a range that's not in a set.
9288 * Only when it is 2 are we in the intersection. */
9291 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9294 /* Special case if either one is empty */
9295 len_a = (a == NULL) ? 0 : _invlist_len(a);
9296 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9297 if (len_a != 0 && complement_b) {
9299 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9300 * must be empty. Here, also we are using 'b's complement, which
9301 * hence must be every possible code point. Thus the intersection
9304 if (*i == a) { /* No-op */
9308 /* If not overwriting either input, just make a copy of 'a' */
9310 *i = invlist_clone(a);
9314 /* Here we are overwriting 'b' with 'a's contents */
9315 r = invlist_clone(a);
9316 invlist_replace_list_destroys_src(*i, r);
9321 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9322 * intersection must be empty */
9324 *i = _new_invlist(0);
9332 /* Here both lists exist and are non-empty */
9333 array_a = invlist_array(a);
9334 array_b = invlist_array(b);
9336 /* If are to take the intersection of 'a' with the complement of b, set it
9337 * up so are looking at b's complement. */
9340 /* To complement, we invert: if the first element is 0, remove it. To
9341 * do this, we just pretend the array starts one later */
9342 if (array_b[0] == 0) {
9348 /* But if the first element is not zero, we pretend the list starts
9349 * at the 0 that is always stored immediately before the array. */
9355 /* Size the intersection for the worst case: that the intersection ends up
9356 * fragmenting everything to be completely disjoint */
9357 r= _new_invlist(len_a + len_b);
9359 /* Will contain U+0000 iff both components do */
9360 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9361 && len_b > 0 && array_b[0] == 0);
9363 /* Go through each list item by item, stopping when exhausted one of
9365 while (i_a < len_a && i_b < len_b) {
9366 UV cp; /* The element to potentially add to the intersection's
9368 bool cp_in_set; /* Is it in the input list's set or not */
9370 /* We need to take one or the other of the two inputs for the
9371 * intersection. Since we are merging two sorted lists, we take the
9372 * smaller of the next items. In case of a tie, we take first the one
9373 * that is not in its set (a difference from the union algorithm). If
9374 * we first took the one in its set, it would increment the count,
9375 * possibly to 2 which would cause it to be output as starting a range
9376 * in the intersection, and the next time through we would take that
9377 * same number, and output it again as ending the set. By doing the
9378 * opposite of this, there is no possibility that the count will be
9379 * momentarily incremented to 2. (In a tie and both are in the set or
9380 * both not in the set, it doesn't matter which we take first.) */
9381 if ( array_a[i_a] < array_b[i_b]
9382 || ( array_a[i_a] == array_b[i_b]
9383 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9385 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9386 cp = array_a[i_a++];
9389 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9393 /* Here, have chosen which of the two inputs to look at. Only output
9394 * if the running count changes to/from 2, which marks the
9395 * beginning/end of a range that's in the intersection */
9399 array_r[i_r++] = cp;
9404 array_r[i_r++] = cp;
9411 /* The loop above increments the index into exactly one of the input lists
9412 * each iteration, and ends when either index gets to its list end. That
9413 * means the other index is lower than its end, and so something is
9414 * remaining in that one. We increment 'count', as explained below, if the
9415 * exhausted list was in its set. (i_a and i_b each currently index the
9416 * element beyond the one we care about.) */
9417 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9418 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9423 /* Above we incremented 'count' if the exhausted list was in its set. This
9424 * has made it so that 'count' being below 2 means there is nothing left to
9425 * output; otheriwse what's left to add to the intersection is precisely
9426 * that which is left in the non-exhausted input list.
9428 * To see why, note first that the exhausted input obviously has nothing
9429 * left to affect the intersection. If it was in its set at its end, that
9430 * means the set extends from here to the platform's infinity, and hence
9431 * anything in the non-exhausted's list will be in the intersection, and
9432 * anything not in it won't be. Hence, the rest of the intersection is
9433 * precisely what's in the non-exhausted list The exhausted set also
9434 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9435 * it means 'count' is now at least 2. This is consistent with the
9436 * incremented 'count' being >= 2 means to add the non-exhausted list to
9439 * But if the exhausted input wasn't in its set, it contributed 0 to
9440 * 'count', and the intersection can't include anything further; the
9441 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9442 * incremented. This is consistent with 'count' being < 2 meaning nothing
9443 * further to add to the intersection. */
9444 if (count < 2) { /* Nothing left to put in the intersection. */
9447 else { /* copy the non-exhausted list, unchanged. */
9448 IV copy_count = len_a - i_a;
9449 if (copy_count > 0) { /* a is the one with stuff left */
9450 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9452 else { /* b is the one with stuff left */
9453 copy_count = len_b - i_b;
9454 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9456 len_r = i_r + copy_count;
9459 /* Set the result to the final length, which can change the pointer to
9460 * array_r, so re-find it. (Note that it is unlikely that this will
9461 * change, as we are shrinking the space, not enlarging it) */
9462 if (len_r != _invlist_len(r)) {
9463 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9465 array_r = invlist_array(r);
9468 /* Finish outputting any remaining */
9469 if (count >= 2) { /* At most one will have a non-zero copy count */
9471 if ((copy_count = len_a - i_a) > 0) {
9472 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9474 else if ((copy_count = len_b - i_b) > 0) {
9475 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9479 /* If the output is not to overwrite either of the inputs, just return the
9480 * calculated intersection */
9481 if (a != *i && b != *i) {
9485 /* Here, the output is to be the same as one of the input scalars,
9486 * hence replacing it. The simple thing to do is to free the input
9487 * scalar, making it instead be the output one. But experience has
9488 * shown [perl #127392] that if the input is a mortal, we can get a
9489 * huge build-up of these during regex compilation before they get
9490 * freed. So for that case, replace just the input's interior with
9491 * the output's, and then free the output. A short-cut in this case
9492 * is if the output is empty, we can just set the input to be empty */
9494 assert(! invlist_is_iterating(*i));
9497 SvREFCNT_dec_NN(*i);
9502 invlist_replace_list_destroys_src(*i, r);
9515 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9517 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9518 * set. A pointer to the inversion list is returned. This may actually be
9519 * a new list, in which case the passed in one has been destroyed. The
9520 * passed-in inversion list can be NULL, in which case a new one is created
9521 * with just the one range in it. The new list is not necessarily
9522 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9523 * result of this function. The gain would not be large, and in many
9524 * cases, this is called multiple times on a single inversion list, so
9525 * anything freed may almost immediately be needed again.
9527 * This used to mostly call the 'union' routine, but that is much more
9528 * heavyweight than really needed for a single range addition */
9530 UV* array; /* The array implementing the inversion list */
9531 UV len; /* How many elements in 'array' */
9532 SSize_t i_s; /* index into the invlist array where 'start'
9534 SSize_t i_e = 0; /* And the index where 'end' should go */
9535 UV cur_highest; /* The highest code point in the inversion list
9536 upon entry to this function */
9538 /* This range becomes the whole inversion list if none already existed */
9539 if (invlist == NULL) {
9540 invlist = _new_invlist(2);
9541 _append_range_to_invlist(invlist, start, end);
9545 /* Likewise, if the inversion list is currently empty */
9546 len = _invlist_len(invlist);
9548 _append_range_to_invlist(invlist, start, end);
9552 /* Starting here, we have to know the internals of the list */
9553 array = invlist_array(invlist);
9555 /* If the new range ends higher than the current highest ... */
9556 cur_highest = invlist_highest(invlist);
9557 if (end > cur_highest) {
9559 /* If the whole range is higher, we can just append it */
9560 if (start > cur_highest) {
9561 _append_range_to_invlist(invlist, start, end);
9565 /* Otherwise, add the portion that is higher ... */
9566 _append_range_to_invlist(invlist, cur_highest + 1, end);
9568 /* ... and continue on below to handle the rest. As a result of the
9569 * above append, we know that the index of the end of the range is the
9570 * final even numbered one of the array. Recall that the final element
9571 * always starts a range that extends to infinity. If that range is in
9572 * the set (meaning the set goes from here to infinity), it will be an
9573 * even index, but if it isn't in the set, it's odd, and the final
9574 * range in the set is one less, which is even. */
9575 if (end == UV_MAX) {
9583 /* We have dealt with appending, now see about prepending. If the new
9584 * range starts lower than the current lowest ... */
9585 if (start < array[0]) {
9587 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
9588 * Let the union code handle it, rather than having to know the
9589 * trickiness in two code places. */
9590 if (UNLIKELY(start == 0)) {
9593 range_invlist = _new_invlist(2);
9594 _append_range_to_invlist(range_invlist, start, end);
9596 _invlist_union(invlist, range_invlist, &invlist);
9598 SvREFCNT_dec_NN(range_invlist);
9603 /* If the whole new range comes before the first entry, and doesn't
9604 * extend it, we have to insert it as an additional range */
9605 if (end < array[0] - 1) {
9607 goto splice_in_new_range;
9610 /* Here the new range adjoins the existing first range, extending it
9614 /* And continue on below to handle the rest. We know that the index of
9615 * the beginning of the range is the first one of the array */
9618 else { /* Not prepending any part of the new range to the existing list.
9619 * Find where in the list it should go. This finds i_s, such that:
9620 * invlist[i_s] <= start < array[i_s+1]
9622 i_s = _invlist_search(invlist, start);
9625 /* At this point, any extending before the beginning of the inversion list
9626 * and/or after the end has been done. This has made it so that, in the
9627 * code below, each endpoint of the new range is either in a range that is
9628 * in the set, or is in a gap between two ranges that are. This means we
9629 * don't have to worry about exceeding the array bounds.
9631 * Find where in the list the new range ends (but we can skip this if we
9632 * have already determined what it is, or if it will be the same as i_s,
9633 * which we already have computed) */
9635 i_e = (start == end)
9637 : _invlist_search(invlist, end);
9640 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
9641 * is a range that goes to infinity there is no element at invlist[i_e+1],
9642 * so only the first relation holds. */
9644 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9646 /* Here, the ranges on either side of the beginning of the new range
9647 * are in the set, and this range starts in the gap between them.
9649 * The new range extends the range above it downwards if the new range
9650 * ends at or above that range's start */
9651 const bool extends_the_range_above = ( end == UV_MAX
9652 || end + 1 >= array[i_s+1]);
9654 /* The new range extends the range below it upwards if it begins just
9655 * after where that range ends */
9656 if (start == array[i_s]) {
9658 /* If the new range fills the entire gap between the other ranges,
9659 * they will get merged together. Other ranges may also get
9660 * merged, depending on how many of them the new range spans. In
9661 * the general case, we do the merge later, just once, after we
9662 * figure out how many to merge. But in the case where the new
9663 * range exactly spans just this one gap (possibly extending into
9664 * the one above), we do the merge here, and an early exit. This
9665 * is done here to avoid having to special case later. */
9666 if (i_e - i_s <= 1) {
9668 /* If i_e - i_s == 1, it means that the new range terminates
9669 * within the range above, and hence 'extends_the_range_above'
9670 * must be true. (If the range above it extends to infinity,
9671 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
9672 * will be 0, so no harm done.) */
9673 if (extends_the_range_above) {
9674 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
9675 invlist_set_len(invlist,
9677 *(get_invlist_offset_addr(invlist)));
9681 /* Here, i_e must == i_s. We keep them in sync, as they apply
9682 * to the same range, and below we are about to decrement i_s
9687 /* Here, the new range is adjacent to the one below. (It may also
9688 * span beyond the range above, but that will get resolved later.)
9689 * Extend the range below to include this one. */
9690 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
9694 else if (extends_the_range_above) {
9696 /* Here the new range only extends the range above it, but not the
9697 * one below. It merges with the one above. Again, we keep i_e
9698 * and i_s in sync if they point to the same range */
9707 /* Here, we've dealt with the new range start extending any adjoining
9710 * If the new range extends to infinity, it is now the final one,
9711 * regardless of what was there before */
9712 if (UNLIKELY(end == UV_MAX)) {
9713 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
9717 /* If i_e started as == i_s, it has also been dealt with,
9718 * and been updated to the new i_s, which will fail the following if */
9719 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
9721 /* Here, the ranges on either side of the end of the new range are in
9722 * the set, and this range ends in the gap between them.
9724 * If this range is adjacent to (hence extends) the range above it, it
9725 * becomes part of that range; likewise if it extends the range below,
9726 * it becomes part of that range */
9727 if (end + 1 == array[i_e+1]) {
9731 else if (start <= array[i_e]) {
9732 array[i_e] = end + 1;
9739 /* If the range fits entirely in an existing range (as possibly already
9740 * extended above), it doesn't add anything new */
9741 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9745 /* Here, no part of the range is in the list. Must add it. It will
9746 * occupy 2 more slots */
9747 splice_in_new_range:
9749 invlist_extend(invlist, len + 2);
9750 array = invlist_array(invlist);
9751 /* Move the rest of the array down two slots. Don't include any
9753 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
9755 /* Do the actual splice */
9756 array[i_e+1] = start;
9757 array[i_e+2] = end + 1;
9758 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
9762 /* Here the new range crossed the boundaries of a pre-existing range. The
9763 * code above has adjusted things so that both ends are in ranges that are
9764 * in the set. This means everything in between must also be in the set.
9765 * Just squash things together */
9766 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
9767 invlist_set_len(invlist,
9769 *(get_invlist_offset_addr(invlist)));
9775 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9776 UV** other_elements_ptr)
9778 /* Create and return an inversion list whose contents are to be populated
9779 * by the caller. The caller gives the number of elements (in 'size') and
9780 * the very first element ('element0'). This function will set
9781 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9784 * Obviously there is some trust involved that the caller will properly
9785 * fill in the other elements of the array.
9787 * (The first element needs to be passed in, as the underlying code does
9788 * things differently depending on whether it is zero or non-zero) */
9790 SV* invlist = _new_invlist(size);
9793 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9795 invlist = add_cp_to_invlist(invlist, element0);
9796 offset = *get_invlist_offset_addr(invlist);
9798 invlist_set_len(invlist, size, offset);
9799 *other_elements_ptr = invlist_array(invlist) + 1;
9805 PERL_STATIC_INLINE SV*
9806 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9807 return _add_range_to_invlist(invlist, cp, cp);
9810 #ifndef PERL_IN_XSUB_RE
9812 Perl__invlist_invert(pTHX_ SV* const invlist)
9814 /* Complement the input inversion list. This adds a 0 if the list didn't
9815 * have a zero; removes it otherwise. As described above, the data
9816 * structure is set up so that this is very efficient */
9818 PERL_ARGS_ASSERT__INVLIST_INVERT;
9820 assert(! invlist_is_iterating(invlist));
9822 /* The inverse of matching nothing is matching everything */
9823 if (_invlist_len(invlist) == 0) {
9824 _append_range_to_invlist(invlist, 0, UV_MAX);
9828 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9833 PERL_STATIC_INLINE SV*
9834 S_invlist_clone(pTHX_ SV* const invlist)
9837 /* Return a new inversion list that is a copy of the input one, which is
9838 * unchanged. The new list will not be mortal even if the old one was. */
9840 /* Need to allocate extra space to accommodate Perl's addition of a
9841 * trailing NUL to SvPV's, since it thinks they are always strings */
9842 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9843 STRLEN physical_length = SvCUR(invlist);
9844 bool offset = *(get_invlist_offset_addr(invlist));
9846 PERL_ARGS_ASSERT_INVLIST_CLONE;
9848 *(get_invlist_offset_addr(new_invlist)) = offset;
9849 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9850 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9855 PERL_STATIC_INLINE STRLEN*
9856 S_get_invlist_iter_addr(SV* invlist)
9858 /* Return the address of the UV that contains the current iteration
9861 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9863 assert(SvTYPE(invlist) == SVt_INVLIST);
9865 return &(((XINVLIST*) SvANY(invlist))->iterator);
9868 PERL_STATIC_INLINE void
9869 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9871 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9873 *get_invlist_iter_addr(invlist) = 0;
9876 PERL_STATIC_INLINE void
9877 S_invlist_iterfinish(SV* invlist)
9879 /* Terminate iterator for invlist. This is to catch development errors.
9880 * Any iteration that is interrupted before completed should call this
9881 * function. Functions that add code points anywhere else but to the end
9882 * of an inversion list assert that they are not in the middle of an
9883 * iteration. If they were, the addition would make the iteration
9884 * problematical: if the iteration hadn't reached the place where things
9885 * were being added, it would be ok */
9887 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9889 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9893 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9895 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9896 * This call sets in <*start> and <*end>, the next range in <invlist>.
9897 * Returns <TRUE> if successful and the next call will return the next
9898 * range; <FALSE> if was already at the end of the list. If the latter,
9899 * <*start> and <*end> are unchanged, and the next call to this function
9900 * will start over at the beginning of the list */
9902 STRLEN* pos = get_invlist_iter_addr(invlist);
9903 UV len = _invlist_len(invlist);
9906 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9909 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9913 array = invlist_array(invlist);
9915 *start = array[(*pos)++];
9921 *end = array[(*pos)++] - 1;
9927 PERL_STATIC_INLINE UV
9928 S_invlist_highest(SV* const invlist)
9930 /* Returns the highest code point that matches an inversion list. This API
9931 * has an ambiguity, as it returns 0 under either the highest is actually
9932 * 0, or if the list is empty. If this distinction matters to you, check
9933 * for emptiness before calling this function */
9935 UV len = _invlist_len(invlist);
9938 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9944 array = invlist_array(invlist);
9946 /* The last element in the array in the inversion list always starts a
9947 * range that goes to infinity. That range may be for code points that are
9948 * matched in the inversion list, or it may be for ones that aren't
9949 * matched. In the latter case, the highest code point in the set is one
9950 * less than the beginning of this range; otherwise it is the final element
9951 * of this range: infinity */
9952 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
9954 : array[len - 1] - 1;
9958 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
9960 /* Get the contents of an inversion list into a string SV so that they can
9961 * be printed out. If 'traditional_style' is TRUE, it uses the format
9962 * traditionally done for debug tracing; otherwise it uses a format
9963 * suitable for just copying to the output, with blanks between ranges and
9964 * a dash between range components */
9968 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
9969 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
9971 if (traditional_style) {
9972 output = newSVpvs("\n");
9975 output = newSVpvs("");
9978 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
9980 assert(! invlist_is_iterating(invlist));
9982 invlist_iterinit(invlist);
9983 while (invlist_iternext(invlist, &start, &end)) {
9984 if (end == UV_MAX) {
9985 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"%cINFINITY%c",
9986 start, intra_range_delimiter,
9987 inter_range_delimiter);
9989 else if (end != start) {
9990 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"%c%04"UVXf"%c",
9992 intra_range_delimiter,
9993 end, inter_range_delimiter);
9996 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"%c",
9997 start, inter_range_delimiter);
10001 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10002 SvCUR_set(output, SvCUR(output) - 1);
10008 #ifndef PERL_IN_XSUB_RE
10010 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10011 const char * const indent, SV* const invlist)
10013 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10014 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10015 * the string 'indent'. The output looks like this:
10016 [0] 0x000A .. 0x000D
10018 [4] 0x2028 .. 0x2029
10019 [6] 0x3104 .. INFINITY
10020 * This means that the first range of code points matched by the list are
10021 * 0xA through 0xD; the second range contains only the single code point
10022 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10023 * are used to define each range (except if the final range extends to
10024 * infinity, only a single element is needed). The array index of the
10025 * first element for the corresponding range is given in brackets. */
10030 PERL_ARGS_ASSERT__INVLIST_DUMP;
10032 if (invlist_is_iterating(invlist)) {
10033 Perl_dump_indent(aTHX_ level, file,
10034 "%sCan't dump inversion list because is in middle of iterating\n",
10039 invlist_iterinit(invlist);
10040 while (invlist_iternext(invlist, &start, &end)) {
10041 if (end == UV_MAX) {
10042 Perl_dump_indent(aTHX_ level, file,
10043 "%s[%"UVuf"] 0x%04"UVXf" .. INFINITY\n",
10044 indent, (UV)count, start);
10046 else if (end != start) {
10047 Perl_dump_indent(aTHX_ level, file,
10048 "%s[%"UVuf"] 0x%04"UVXf" .. 0x%04"UVXf"\n",
10049 indent, (UV)count, start, end);
10052 Perl_dump_indent(aTHX_ level, file, "%s[%"UVuf"] 0x%04"UVXf"\n",
10053 indent, (UV)count, start);
10060 Perl__load_PL_utf8_foldclosures (pTHX)
10062 assert(! PL_utf8_foldclosures);
10064 /* If the folds haven't been read in, call a fold function
10066 if (! PL_utf8_tofold) {
10067 U8 dummy[UTF8_MAXBYTES_CASE+1];
10069 /* This string is just a short named one above \xff */
10070 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
10071 assert(PL_utf8_tofold); /* Verify that worked */
10073 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
10077 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10079 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10081 /* Return a boolean as to if the two passed in inversion lists are
10082 * identical. The final argument, if TRUE, says to take the complement of
10083 * the second inversion list before doing the comparison */
10085 const UV* array_a = invlist_array(a);
10086 const UV* array_b = invlist_array(b);
10087 UV len_a = _invlist_len(a);
10088 UV len_b = _invlist_len(b);
10090 UV i = 0; /* current index into the arrays */
10091 bool retval = TRUE; /* Assume are identical until proven otherwise */
10093 PERL_ARGS_ASSERT__INVLISTEQ;
10095 /* If are to compare 'a' with the complement of b, set it
10096 * up so are looking at b's complement. */
10097 if (complement_b) {
10099 /* The complement of nothing is everything, so <a> would have to have
10100 * just one element, starting at zero (ending at infinity) */
10102 return (len_a == 1 && array_a[0] == 0);
10104 else if (array_b[0] == 0) {
10106 /* Otherwise, to complement, we invert. Here, the first element is
10107 * 0, just remove it. To do this, we just pretend the array starts
10115 /* But if the first element is not zero, we pretend the list starts
10116 * at the 0 that is always stored immediately before the array. */
10122 /* Make sure that the lengths are the same, as well as the final element
10123 * before looping through the remainder. (Thus we test the length, final,
10124 * and first elements right off the bat) */
10125 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
10128 else for (i = 0; i < len_a - 1; i++) {
10129 if (array_a[i] != array_b[i]) {
10140 * As best we can, determine the characters that can match the start of
10141 * the given EXACTF-ish node.
10143 * Returns the invlist as a new SV*; it is the caller's responsibility to
10144 * call SvREFCNT_dec() when done with it.
10147 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10149 const U8 * s = (U8*)STRING(node);
10150 SSize_t bytelen = STR_LEN(node);
10152 /* Start out big enough for 2 separate code points */
10153 SV* invlist = _new_invlist(4);
10155 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
10160 /* We punt and assume can match anything if the node begins
10161 * with a multi-character fold. Things are complicated. For
10162 * example, /ffi/i could match any of:
10163 * "\N{LATIN SMALL LIGATURE FFI}"
10164 * "\N{LATIN SMALL LIGATURE FF}I"
10165 * "F\N{LATIN SMALL LIGATURE FI}"
10166 * plus several other things; and making sure we have all the
10167 * possibilities is hard. */
10168 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10169 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10172 /* Any Latin1 range character can potentially match any
10173 * other depending on the locale */
10174 if (OP(node) == EXACTFL) {
10175 _invlist_union(invlist, PL_Latin1, &invlist);
10178 /* But otherwise, it matches at least itself. We can
10179 * quickly tell if it has a distinct fold, and if so,
10180 * it matches that as well */
10181 invlist = add_cp_to_invlist(invlist, uc);
10182 if (IS_IN_SOME_FOLD_L1(uc))
10183 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10186 /* Some characters match above-Latin1 ones under /i. This
10187 * is true of EXACTFL ones when the locale is UTF-8 */
10188 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10189 && (! isASCII(uc) || (OP(node) != EXACTFA
10190 && OP(node) != EXACTFA_NO_TRIE)))
10192 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10196 else { /* Pattern is UTF-8 */
10197 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10198 STRLEN foldlen = UTF8SKIP(s);
10199 const U8* e = s + bytelen;
10202 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10204 /* The only code points that aren't folded in a UTF EXACTFish
10205 * node are are the problematic ones in EXACTFL nodes */
10206 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10207 /* We need to check for the possibility that this EXACTFL
10208 * node begins with a multi-char fold. Therefore we fold
10209 * the first few characters of it so that we can make that
10214 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10216 *(d++) = (U8) toFOLD(*s);
10221 to_utf8_fold(s, d, &len);
10227 /* And set up so the code below that looks in this folded
10228 * buffer instead of the node's string */
10230 foldlen = UTF8SKIP(folded);
10234 /* When we reach here 's' points to the fold of the first
10235 * character(s) of the node; and 'e' points to far enough along
10236 * the folded string to be just past any possible multi-char
10237 * fold. 'foldlen' is the length in bytes of the first
10240 * Unlike the non-UTF-8 case, the macro for determining if a
10241 * string is a multi-char fold requires all the characters to
10242 * already be folded. This is because of all the complications
10243 * if not. Note that they are folded anyway, except in EXACTFL
10244 * nodes. Like the non-UTF case above, we punt if the node
10245 * begins with a multi-char fold */
10247 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10248 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10250 else { /* Single char fold */
10252 /* It matches all the things that fold to it, which are
10253 * found in PL_utf8_foldclosures (including itself) */
10254 invlist = add_cp_to_invlist(invlist, uc);
10255 if (! PL_utf8_foldclosures)
10256 _load_PL_utf8_foldclosures();
10257 if ((listp = hv_fetch(PL_utf8_foldclosures,
10258 (char *) s, foldlen, FALSE)))
10260 AV* list = (AV*) *listp;
10262 for (k = 0; k <= av_tindex_nomg(list); k++) {
10263 SV** c_p = av_fetch(list, k, FALSE);
10269 /* /aa doesn't allow folds between ASCII and non- */
10270 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
10271 && isASCII(c) != isASCII(uc))
10276 invlist = add_cp_to_invlist(invlist, c);
10285 #undef HEADER_LENGTH
10286 #undef TO_INTERNAL_SIZE
10287 #undef FROM_INTERNAL_SIZE
10288 #undef INVLIST_VERSION_ID
10290 /* End of inversion list object */
10293 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10295 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10296 * constructs, and updates RExC_flags with them. On input, RExC_parse
10297 * should point to the first flag; it is updated on output to point to the
10298 * final ')' or ':'. There needs to be at least one flag, or this will
10301 /* for (?g), (?gc), and (?o) warnings; warning
10302 about (?c) will warn about (?g) -- japhy */
10304 #define WASTED_O 0x01
10305 #define WASTED_G 0x02
10306 #define WASTED_C 0x04
10307 #define WASTED_GC (WASTED_G|WASTED_C)
10308 I32 wastedflags = 0x00;
10309 U32 posflags = 0, negflags = 0;
10310 U32 *flagsp = &posflags;
10311 char has_charset_modifier = '\0';
10313 bool has_use_defaults = FALSE;
10314 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10315 int x_mod_count = 0;
10317 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10319 /* '^' as an initial flag sets certain defaults */
10320 if (UCHARAT(RExC_parse) == '^') {
10322 has_use_defaults = TRUE;
10323 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10324 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
10325 ? REGEX_UNICODE_CHARSET
10326 : REGEX_DEPENDS_CHARSET);
10329 cs = get_regex_charset(RExC_flags);
10330 if (cs == REGEX_DEPENDS_CHARSET
10331 && (RExC_utf8 || RExC_uni_semantics))
10333 cs = REGEX_UNICODE_CHARSET;
10336 while (RExC_parse < RExC_end) {
10337 /* && strchr("iogcmsx", *RExC_parse) */
10338 /* (?g), (?gc) and (?o) are useless here
10339 and must be globally applied -- japhy */
10340 switch (*RExC_parse) {
10342 /* Code for the imsxn flags */
10343 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10345 case LOCALE_PAT_MOD:
10346 if (has_charset_modifier) {
10347 goto excess_modifier;
10349 else if (flagsp == &negflags) {
10352 cs = REGEX_LOCALE_CHARSET;
10353 has_charset_modifier = LOCALE_PAT_MOD;
10355 case UNICODE_PAT_MOD:
10356 if (has_charset_modifier) {
10357 goto excess_modifier;
10359 else if (flagsp == &negflags) {
10362 cs = REGEX_UNICODE_CHARSET;
10363 has_charset_modifier = UNICODE_PAT_MOD;
10365 case ASCII_RESTRICT_PAT_MOD:
10366 if (flagsp == &negflags) {
10369 if (has_charset_modifier) {
10370 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10371 goto excess_modifier;
10373 /* Doubled modifier implies more restricted */
10374 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10377 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10379 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10381 case DEPENDS_PAT_MOD:
10382 if (has_use_defaults) {
10383 goto fail_modifiers;
10385 else if (flagsp == &negflags) {
10388 else if (has_charset_modifier) {
10389 goto excess_modifier;
10392 /* The dual charset means unicode semantics if the
10393 * pattern (or target, not known until runtime) are
10394 * utf8, or something in the pattern indicates unicode
10396 cs = (RExC_utf8 || RExC_uni_semantics)
10397 ? REGEX_UNICODE_CHARSET
10398 : REGEX_DEPENDS_CHARSET;
10399 has_charset_modifier = DEPENDS_PAT_MOD;
10403 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10404 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10406 else if (has_charset_modifier == *(RExC_parse - 1)) {
10407 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10408 *(RExC_parse - 1));
10411 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10413 NOT_REACHED; /*NOTREACHED*/
10416 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10417 *(RExC_parse - 1));
10418 NOT_REACHED; /*NOTREACHED*/
10419 case ONCE_PAT_MOD: /* 'o' */
10420 case GLOBAL_PAT_MOD: /* 'g' */
10421 if (PASS2 && ckWARN(WARN_REGEXP)) {
10422 const I32 wflagbit = *RExC_parse == 'o'
10425 if (! (wastedflags & wflagbit) ) {
10426 wastedflags |= wflagbit;
10427 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10430 "Useless (%s%c) - %suse /%c modifier",
10431 flagsp == &negflags ? "?-" : "?",
10433 flagsp == &negflags ? "don't " : "",
10440 case CONTINUE_PAT_MOD: /* 'c' */
10441 if (PASS2 && ckWARN(WARN_REGEXP)) {
10442 if (! (wastedflags & WASTED_C) ) {
10443 wastedflags |= WASTED_GC;
10444 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10447 "Useless (%sc) - %suse /gc modifier",
10448 flagsp == &negflags ? "?-" : "?",
10449 flagsp == &negflags ? "don't " : ""
10454 case KEEPCOPY_PAT_MOD: /* 'p' */
10455 if (flagsp == &negflags) {
10457 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10459 *flagsp |= RXf_PMf_KEEPCOPY;
10463 /* A flag is a default iff it is following a minus, so
10464 * if there is a minus, it means will be trying to
10465 * re-specify a default which is an error */
10466 if (has_use_defaults || flagsp == &negflags) {
10467 goto fail_modifiers;
10469 flagsp = &negflags;
10470 wastedflags = 0; /* reset so (?g-c) warns twice */
10474 RExC_flags |= posflags;
10475 RExC_flags &= ~negflags;
10476 set_regex_charset(&RExC_flags, cs);
10477 if (RExC_flags & RXf_PMf_FOLD) {
10478 RExC_contains_i = 1;
10481 if (UNLIKELY((x_mod_count) > 1)) {
10482 vFAIL("Only one /x regex modifier is allowed");
10488 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10489 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10490 vFAIL2utf8f("Sequence (%"UTF8f"...) not recognized",
10491 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10492 NOT_REACHED; /*NOTREACHED*/
10495 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10498 vFAIL("Sequence (?... not terminated");
10502 - reg - regular expression, i.e. main body or parenthesized thing
10504 * Caller must absorb opening parenthesis.
10506 * Combining parenthesis handling with the base level of regular expression
10507 * is a trifle forced, but the need to tie the tails of the branches to what
10508 * follows makes it hard to avoid.
10510 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10512 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10514 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10517 PERL_STATIC_INLINE regnode *
10518 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10520 char * parse_start,
10525 char* name_start = RExC_parse;
10527 SV *sv_dat = reg_scan_name(pRExC_state, SIZE_ONLY
10528 ? REG_RSN_RETURN_NULL
10529 : REG_RSN_RETURN_DATA);
10530 GET_RE_DEBUG_FLAGS_DECL;
10532 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10534 if (RExC_parse == name_start || *RExC_parse != ch) {
10535 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10536 vFAIL2("Sequence %.3s... not terminated",parse_start);
10540 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10541 RExC_rxi->data->data[num]=(void*)sv_dat;
10542 SvREFCNT_inc_simple_void(sv_dat);
10545 ret = reganode(pRExC_state,
10548 : (ASCII_FOLD_RESTRICTED)
10550 : (AT_LEAST_UNI_SEMANTICS)
10556 *flagp |= HASWIDTH;
10558 Set_Node_Offset(ret, parse_start+1);
10559 Set_Node_Cur_Length(ret, parse_start);
10561 nextchar(pRExC_state);
10565 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
10566 flags. Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan
10567 needs to be restarted, or'd with NEED_UTF8 if the pattern needs to be
10568 upgraded to UTF-8. Otherwise would only return NULL if regbranch() returns
10569 NULL, which cannot happen. */
10571 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
10572 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10573 * 2 is like 1, but indicates that nextchar() has been called to advance
10574 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10575 * this flag alerts us to the need to check for that */
10577 regnode *ret; /* Will be the head of the group. */
10580 regnode *ender = NULL;
10583 U32 oregflags = RExC_flags;
10584 bool have_branch = 0;
10586 I32 freeze_paren = 0;
10587 I32 after_freeze = 0;
10588 I32 num; /* numeric backreferences */
10590 char * parse_start = RExC_parse; /* MJD */
10591 char * const oregcomp_parse = RExC_parse;
10593 GET_RE_DEBUG_FLAGS_DECL;
10595 PERL_ARGS_ASSERT_REG;
10596 DEBUG_PARSE("reg ");
10598 *flagp = 0; /* Tentatively. */
10600 /* Having this true makes it feasible to have a lot fewer tests for the
10601 * parse pointer being in scope. For example, we can write
10602 * while(isFOO(*RExC_parse)) RExC_parse++;
10604 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
10606 assert(*RExC_end == '\0');
10608 /* Make an OPEN node, if parenthesized. */
10611 /* Under /x, space and comments can be gobbled up between the '(' and
10612 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
10613 * intervening space, as the sequence is a token, and a token should be
10615 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
10617 if (RExC_parse >= RExC_end) {
10618 vFAIL("Unmatched (");
10621 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
10622 char *start_verb = RExC_parse + 1;
10624 char *start_arg = NULL;
10625 unsigned char op = 0;
10626 int arg_required = 0;
10627 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
10629 if (has_intervening_patws) {
10630 RExC_parse++; /* past the '*' */
10631 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
10633 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
10634 if ( *RExC_parse == ':' ) {
10635 start_arg = RExC_parse + 1;
10638 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10640 verb_len = RExC_parse - start_verb;
10642 if (RExC_parse >= RExC_end) {
10643 goto unterminated_verb_pattern;
10645 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10646 while ( RExC_parse < RExC_end && *RExC_parse != ')' )
10647 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10648 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10649 unterminated_verb_pattern:
10650 vFAIL("Unterminated verb pattern argument");
10651 if ( RExC_parse == start_arg )
10654 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10655 vFAIL("Unterminated verb pattern");
10658 /* Here, we know that RExC_parse < RExC_end */
10660 switch ( *start_verb ) {
10661 case 'A': /* (*ACCEPT) */
10662 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
10664 internal_argval = RExC_nestroot;
10667 case 'C': /* (*COMMIT) */
10668 if ( memEQs(start_verb,verb_len,"COMMIT") )
10671 case 'F': /* (*FAIL) */
10672 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
10676 case ':': /* (*:NAME) */
10677 case 'M': /* (*MARK:NAME) */
10678 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
10683 case 'P': /* (*PRUNE) */
10684 if ( memEQs(start_verb,verb_len,"PRUNE") )
10687 case 'S': /* (*SKIP) */
10688 if ( memEQs(start_verb,verb_len,"SKIP") )
10691 case 'T': /* (*THEN) */
10692 /* [19:06] <TimToady> :: is then */
10693 if ( memEQs(start_verb,verb_len,"THEN") ) {
10695 RExC_seen |= REG_CUTGROUP_SEEN;
10700 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10702 "Unknown verb pattern '%"UTF8f"'",
10703 UTF8fARG(UTF, verb_len, start_verb));
10705 if ( arg_required && !start_arg ) {
10706 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
10707 verb_len, start_verb);
10709 if (internal_argval == -1) {
10710 ret = reganode(pRExC_state, op, 0);
10712 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
10714 RExC_seen |= REG_VERBARG_SEEN;
10715 if ( ! SIZE_ONLY ) {
10717 SV *sv = newSVpvn( start_arg,
10718 RExC_parse - start_arg);
10719 ARG(ret) = add_data( pRExC_state,
10720 STR_WITH_LEN("S"));
10721 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
10726 if ( internal_argval != -1 )
10727 ARG2L_SET(ret, internal_argval);
10729 nextchar(pRExC_state);
10732 else if (*RExC_parse == '?') { /* (?...) */
10733 bool is_logical = 0;
10734 const char * const seqstart = RExC_parse;
10735 const char * endptr;
10736 if (has_intervening_patws) {
10738 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
10741 RExC_parse++; /* past the '?' */
10742 paren = *RExC_parse; /* might be a trailing NUL, if not
10744 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10745 if (RExC_parse > RExC_end) {
10748 ret = NULL; /* For look-ahead/behind. */
10751 case 'P': /* (?P...) variants for those used to PCRE/Python */
10752 paren = *RExC_parse;
10753 if ( paren == '<') { /* (?P<...>) named capture */
10755 if (RExC_parse >= RExC_end) {
10756 vFAIL("Sequence (?P<... not terminated");
10758 goto named_capture;
10760 else if (paren == '>') { /* (?P>name) named recursion */
10762 if (RExC_parse >= RExC_end) {
10763 vFAIL("Sequence (?P>... not terminated");
10765 goto named_recursion;
10767 else if (paren == '=') { /* (?P=...) named backref */
10769 return handle_named_backref(pRExC_state, flagp,
10772 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10773 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10774 vFAIL3("Sequence (%.*s...) not recognized",
10775 RExC_parse-seqstart, seqstart);
10776 NOT_REACHED; /*NOTREACHED*/
10777 case '<': /* (?<...) */
10778 if (*RExC_parse == '!')
10780 else if (*RExC_parse != '=')
10787 case '\'': /* (?'...') */
10788 name_start = RExC_parse;
10789 svname = reg_scan_name(pRExC_state,
10790 SIZE_ONLY /* reverse test from the others */
10791 ? REG_RSN_RETURN_NAME
10792 : REG_RSN_RETURN_NULL);
10793 if ( RExC_parse == name_start
10794 || RExC_parse >= RExC_end
10795 || *RExC_parse != paren)
10797 vFAIL2("Sequence (?%c... not terminated",
10798 paren=='>' ? '<' : paren);
10803 if (!svname) /* shouldn't happen */
10805 "panic: reg_scan_name returned NULL");
10806 if (!RExC_paren_names) {
10807 RExC_paren_names= newHV();
10808 sv_2mortal(MUTABLE_SV(RExC_paren_names));
10810 RExC_paren_name_list= newAV();
10811 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10814 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10816 sv_dat = HeVAL(he_str);
10818 /* croak baby croak */
10820 "panic: paren_name hash element allocation failed");
10821 } else if ( SvPOK(sv_dat) ) {
10822 /* (?|...) can mean we have dupes so scan to check
10823 its already been stored. Maybe a flag indicating
10824 we are inside such a construct would be useful,
10825 but the arrays are likely to be quite small, so
10826 for now we punt -- dmq */
10827 IV count = SvIV(sv_dat);
10828 I32 *pv = (I32*)SvPVX(sv_dat);
10830 for ( i = 0 ; i < count ; i++ ) {
10831 if ( pv[i] == RExC_npar ) {
10837 pv = (I32*)SvGROW(sv_dat,
10838 SvCUR(sv_dat) + sizeof(I32)+1);
10839 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10840 pv[count] = RExC_npar;
10841 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10844 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10845 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10848 SvIV_set(sv_dat, 1);
10851 /* Yes this does cause a memory leak in debugging Perls
10853 if (!av_store(RExC_paren_name_list,
10854 RExC_npar, SvREFCNT_inc(svname)))
10855 SvREFCNT_dec_NN(svname);
10858 /*sv_dump(sv_dat);*/
10860 nextchar(pRExC_state);
10862 goto capturing_parens;
10864 RExC_seen |= REG_LOOKBEHIND_SEEN;
10865 RExC_in_lookbehind++;
10867 if (RExC_parse >= RExC_end) {
10868 vFAIL("Sequence (?... not terminated");
10872 case '=': /* (?=...) */
10873 RExC_seen_zerolen++;
10875 case '!': /* (?!...) */
10876 RExC_seen_zerolen++;
10877 /* check if we're really just a "FAIL" assertion */
10878 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
10879 FALSE /* Don't force to /x */ );
10880 if (*RExC_parse == ')') {
10881 ret=reganode(pRExC_state, OPFAIL, 0);
10882 nextchar(pRExC_state);
10886 case '|': /* (?|...) */
10887 /* branch reset, behave like a (?:...) except that
10888 buffers in alternations share the same numbers */
10890 after_freeze = freeze_paren = RExC_npar;
10892 case ':': /* (?:...) */
10893 case '>': /* (?>...) */
10895 case '$': /* (?$...) */
10896 case '@': /* (?@...) */
10897 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10899 case '0' : /* (?0) */
10900 case 'R' : /* (?R) */
10901 if (RExC_parse == RExC_end || *RExC_parse != ')')
10902 FAIL("Sequence (?R) not terminated");
10904 RExC_seen |= REG_RECURSE_SEEN;
10905 *flagp |= POSTPONED;
10906 goto gen_recurse_regop;
10908 /* named and numeric backreferences */
10909 case '&': /* (?&NAME) */
10910 parse_start = RExC_parse - 1;
10913 SV *sv_dat = reg_scan_name(pRExC_state,
10914 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10915 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10917 if (RExC_parse >= RExC_end || *RExC_parse != ')')
10918 vFAIL("Sequence (?&... not terminated");
10919 goto gen_recurse_regop;
10922 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10924 vFAIL("Illegal pattern");
10926 goto parse_recursion;
10928 case '-': /* (?-1) */
10929 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10930 RExC_parse--; /* rewind to let it be handled later */
10934 case '1': case '2': case '3': case '4': /* (?1) */
10935 case '5': case '6': case '7': case '8': case '9':
10936 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
10939 bool is_neg = FALSE;
10941 parse_start = RExC_parse - 1; /* MJD */
10942 if (*RExC_parse == '-') {
10946 if (grok_atoUV(RExC_parse, &unum, &endptr)
10950 RExC_parse = (char*)endptr;
10954 /* Some limit for num? */
10958 if (*RExC_parse!=')')
10959 vFAIL("Expecting close bracket");
10962 if ( paren == '-' ) {
10964 Diagram of capture buffer numbering.
10965 Top line is the normal capture buffer numbers
10966 Bottom line is the negative indexing as from
10970 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
10974 num = RExC_npar + num;
10977 vFAIL("Reference to nonexistent group");
10979 } else if ( paren == '+' ) {
10980 num = RExC_npar + num - 1;
10982 /* We keep track how many GOSUB items we have produced.
10983 To start off the ARG2L() of the GOSUB holds its "id",
10984 which is used later in conjunction with RExC_recurse
10985 to calculate the offset we need to jump for the GOSUB,
10986 which it will store in the final representation.
10987 We have to defer the actual calculation until much later
10988 as the regop may move.
10991 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
10993 if (num > (I32)RExC_rx->nparens) {
10995 vFAIL("Reference to nonexistent group");
10997 RExC_recurse_count++;
10998 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
10999 "%*s%*s Recurse #%"UVuf" to %"IVdf"\n",
11000 22, "| |", (int)(depth * 2 + 1), "",
11001 (UV)ARG(ret), (IV)ARG2L(ret)));
11003 RExC_seen |= REG_RECURSE_SEEN;
11005 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
11006 Set_Node_Offset(ret, parse_start); /* MJD */
11008 *flagp |= POSTPONED;
11009 assert(*RExC_parse == ')');
11010 nextchar(pRExC_state);
11015 case '?': /* (??...) */
11017 if (*RExC_parse != '{') {
11018 RExC_parse += SKIP_IF_CHAR(RExC_parse);
11019 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11021 "Sequence (%"UTF8f"...) not recognized",
11022 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11023 NOT_REACHED; /*NOTREACHED*/
11025 *flagp |= POSTPONED;
11029 case '{': /* (?{...}) */
11032 struct reg_code_block *cb;
11034 RExC_seen_zerolen++;
11036 if ( !pRExC_state->num_code_blocks
11037 || pRExC_state->code_index >= pRExC_state->num_code_blocks
11038 || pRExC_state->code_blocks[pRExC_state->code_index].start
11039 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11042 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11043 FAIL("panic: Sequence (?{...}): no code block found\n");
11044 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11046 /* this is a pre-compiled code block (?{...}) */
11047 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
11048 RExC_parse = RExC_start + cb->end;
11051 if (cb->src_regex) {
11052 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11053 RExC_rxi->data->data[n] =
11054 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11055 RExC_rxi->data->data[n+1] = (void*)o;
11058 n = add_data(pRExC_state,
11059 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11060 RExC_rxi->data->data[n] = (void*)o;
11063 pRExC_state->code_index++;
11064 nextchar(pRExC_state);
11068 ret = reg_node(pRExC_state, LOGICAL);
11070 eval = reg2Lanode(pRExC_state, EVAL,
11073 /* for later propagation into (??{})
11075 RExC_flags & RXf_PMf_COMPILETIME
11080 REGTAIL(pRExC_state, ret, eval);
11081 /* deal with the length of this later - MJD */
11084 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11085 Set_Node_Length(ret, RExC_parse - parse_start + 1);
11086 Set_Node_Offset(ret, parse_start);
11089 case '(': /* (?(?{...})...) and (?(?=...)...) */
11092 const int DEFINE_len = sizeof("DEFINE") - 1;
11093 if (RExC_parse[0] == '?') { /* (?(?...)) */
11094 if ( RExC_parse < RExC_end - 1
11095 && ( RExC_parse[1] == '='
11096 || RExC_parse[1] == '!'
11097 || RExC_parse[1] == '<'
11098 || RExC_parse[1] == '{')
11099 ) { /* Lookahead or eval. */
11103 ret = reg_node(pRExC_state, LOGICAL);
11107 tail = reg(pRExC_state, 1, &flag, depth+1);
11108 if (flag & (RESTART_PASS1|NEED_UTF8)) {
11109 *flagp = flag & (RESTART_PASS1|NEED_UTF8);
11112 REGTAIL(pRExC_state, ret, tail);
11115 /* Fall through to ‘Unknown switch condition’ at the
11116 end of the if/else chain. */
11118 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11119 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11121 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11122 char *name_start= RExC_parse++;
11124 SV *sv_dat=reg_scan_name(pRExC_state,
11125 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11126 if ( RExC_parse == name_start
11127 || RExC_parse >= RExC_end
11128 || *RExC_parse != ch)
11130 vFAIL2("Sequence (?(%c... not terminated",
11131 (ch == '>' ? '<' : ch));
11135 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11136 RExC_rxi->data->data[num]=(void*)sv_dat;
11137 SvREFCNT_inc_simple_void(sv_dat);
11139 ret = reganode(pRExC_state,NGROUPP,num);
11140 goto insert_if_check_paren;
11142 else if (RExC_end - RExC_parse >= DEFINE_len
11143 && strnEQ(RExC_parse, "DEFINE", DEFINE_len))
11145 ret = reganode(pRExC_state,DEFINEP,0);
11146 RExC_parse += DEFINE_len;
11148 goto insert_if_check_paren;
11150 else if (RExC_parse[0] == 'R') {
11152 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11153 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11154 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11157 if (RExC_parse[0] == '0') {
11161 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11163 if (grok_atoUV(RExC_parse, &uv, &endptr)
11166 parno = (I32)uv + 1;
11167 RExC_parse = (char*)endptr;
11169 /* else "Switch condition not recognized" below */
11170 } else if (RExC_parse[0] == '&') {
11173 sv_dat = reg_scan_name(pRExC_state,
11175 ? REG_RSN_RETURN_NULL
11176 : REG_RSN_RETURN_DATA);
11178 /* we should only have a false sv_dat when
11179 * SIZE_ONLY is true, and we always have false
11180 * sv_dat when SIZE_ONLY is true.
11181 * reg_scan_name() will VFAIL() if the name is
11182 * unknown when SIZE_ONLY is false, and otherwise
11183 * will return something, and when SIZE_ONLY is
11184 * true, reg_scan_name() just parses the string,
11185 * and doesnt return anything. (in theory) */
11186 assert(SIZE_ONLY ? !sv_dat : !!sv_dat);
11189 parno = 1 + *((I32 *)SvPVX(sv_dat));
11191 ret = reganode(pRExC_state,INSUBP,parno);
11192 goto insert_if_check_paren;
11194 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11198 if (grok_atoUV(RExC_parse, &uv, &endptr)
11202 RExC_parse = (char*)endptr;
11205 vFAIL("panic: grok_atoUV returned FALSE");
11207 ret = reganode(pRExC_state, GROUPP, parno);
11209 insert_if_check_paren:
11210 if (UCHARAT(RExC_parse) != ')') {
11211 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11212 vFAIL("Switch condition not recognized");
11214 nextchar(pRExC_state);
11216 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
11217 br = regbranch(pRExC_state, &flags, 1,depth+1);
11219 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11220 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11223 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
11226 REGTAIL(pRExC_state, br, reganode(pRExC_state,
11228 c = UCHARAT(RExC_parse);
11229 nextchar(pRExC_state);
11230 if (flags&HASWIDTH)
11231 *flagp |= HASWIDTH;
11234 vFAIL("(?(DEFINE)....) does not allow branches");
11236 /* Fake one for optimizer. */
11237 lastbr = reganode(pRExC_state, IFTHEN, 0);
11239 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
11240 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11241 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11244 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
11247 REGTAIL(pRExC_state, ret, lastbr);
11248 if (flags&HASWIDTH)
11249 *flagp |= HASWIDTH;
11250 c = UCHARAT(RExC_parse);
11251 nextchar(pRExC_state);
11256 if (RExC_parse >= RExC_end)
11257 vFAIL("Switch (?(condition)... not terminated");
11259 vFAIL("Switch (?(condition)... contains too many branches");
11261 ender = reg_node(pRExC_state, TAIL);
11262 REGTAIL(pRExC_state, br, ender);
11264 REGTAIL(pRExC_state, lastbr, ender);
11265 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11268 REGTAIL(pRExC_state, ret, ender);
11269 RExC_size++; /* XXX WHY do we need this?!!
11270 For large programs it seems to be required
11271 but I can't figure out why. -- dmq*/
11274 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11275 vFAIL("Unknown switch condition (?(...))");
11277 case '[': /* (?[ ... ]) */
11278 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
11280 case 0: /* A NUL */
11281 RExC_parse--; /* for vFAIL to print correctly */
11282 vFAIL("Sequence (? incomplete");
11284 default: /* e.g., (?i) */
11285 RExC_parse = (char *) seqstart + 1;
11287 parse_lparen_question_flags(pRExC_state);
11288 if (UCHARAT(RExC_parse) != ':') {
11289 if (RExC_parse < RExC_end)
11290 nextchar(pRExC_state);
11295 nextchar(pRExC_state);
11300 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11305 ret = reganode(pRExC_state, OPEN, parno);
11307 if (!RExC_nestroot)
11308 RExC_nestroot = parno;
11309 if (RExC_open_parens && !RExC_open_parens[parno])
11311 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11312 "%*s%*s Setting open paren #%"IVdf" to %d\n",
11313 22, "| |", (int)(depth * 2 + 1), "",
11314 (IV)parno, REG_NODE_NUM(ret)));
11315 RExC_open_parens[parno]= ret;
11318 Set_Node_Length(ret, 1); /* MJD */
11319 Set_Node_Offset(ret, RExC_parse); /* MJD */
11322 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
11331 /* Pick up the branches, linking them together. */
11332 parse_start = RExC_parse; /* MJD */
11333 br = regbranch(pRExC_state, &flags, 1,depth+1);
11335 /* branch_len = (paren != 0); */
11338 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11339 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11342 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
11344 if (*RExC_parse == '|') {
11345 if (!SIZE_ONLY && RExC_extralen) {
11346 reginsert(pRExC_state, BRANCHJ, br, depth+1);
11349 reginsert(pRExC_state, BRANCH, br, depth+1);
11350 Set_Node_Length(br, paren != 0);
11351 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
11355 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
11357 else if (paren == ':') {
11358 *flagp |= flags&SIMPLE;
11360 if (is_open) { /* Starts with OPEN. */
11361 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
11363 else if (paren != '?') /* Not Conditional */
11365 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11367 while (*RExC_parse == '|') {
11368 if (!SIZE_ONLY && RExC_extralen) {
11369 ender = reganode(pRExC_state, LONGJMP,0);
11371 /* Append to the previous. */
11372 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11375 RExC_extralen += 2; /* Account for LONGJMP. */
11376 nextchar(pRExC_state);
11377 if (freeze_paren) {
11378 if (RExC_npar > after_freeze)
11379 after_freeze = RExC_npar;
11380 RExC_npar = freeze_paren;
11382 br = regbranch(pRExC_state, &flags, 0, depth+1);
11385 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11386 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11389 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
11391 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
11393 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11396 if (have_branch || paren != ':') {
11397 /* Make a closing node, and hook it on the end. */
11400 ender = reg_node(pRExC_state, TAIL);
11403 ender = reganode(pRExC_state, CLOSE, parno);
11404 if ( RExC_close_parens ) {
11405 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11406 "%*s%*s Setting close paren #%"IVdf" to %d\n",
11407 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
11408 RExC_close_parens[parno]= ender;
11409 if (RExC_nestroot == parno)
11412 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
11413 Set_Node_Length(ender,1); /* MJD */
11419 *flagp &= ~HASWIDTH;
11422 ender = reg_node(pRExC_state, SUCCEED);
11425 ender = reg_node(pRExC_state, END);
11427 assert(!RExC_end_op); /* there can only be one! */
11428 RExC_end_op = ender;
11429 if (RExC_close_parens) {
11430 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11431 "%*s%*s Setting close paren #0 (END) to %d\n",
11432 22, "| |", (int)(depth * 2 + 1), "", REG_NODE_NUM(ender)));
11434 RExC_close_parens[0]= ender;
11439 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11440 DEBUG_PARSE_MSG("lsbr");
11441 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
11442 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11443 Perl_re_printf( aTHX_ "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
11444 SvPV_nolen_const(RExC_mysv1),
11445 (IV)REG_NODE_NUM(lastbr),
11446 SvPV_nolen_const(RExC_mysv2),
11447 (IV)REG_NODE_NUM(ender),
11448 (IV)(ender - lastbr)
11451 REGTAIL(pRExC_state, lastbr, ender);
11453 if (have_branch && !SIZE_ONLY) {
11454 char is_nothing= 1;
11456 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
11458 /* Hook the tails of the branches to the closing node. */
11459 for (br = ret; br; br = regnext(br)) {
11460 const U8 op = PL_regkind[OP(br)];
11461 if (op == BRANCH) {
11462 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
11463 if ( OP(NEXTOPER(br)) != NOTHING
11464 || regnext(NEXTOPER(br)) != ender)
11467 else if (op == BRANCHJ) {
11468 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
11469 /* for now we always disable this optimisation * /
11470 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
11471 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
11477 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
11478 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11479 DEBUG_PARSE_MSG("NADA");
11480 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
11481 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11482 Perl_re_printf( aTHX_ "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
11483 SvPV_nolen_const(RExC_mysv1),
11484 (IV)REG_NODE_NUM(ret),
11485 SvPV_nolen_const(RExC_mysv2),
11486 (IV)REG_NODE_NUM(ender),
11491 if (OP(ender) == TAIL) {
11496 for ( opt= br + 1; opt < ender ; opt++ )
11497 OP(opt)= OPTIMIZED;
11498 NEXT_OFF(br)= ender - br;
11506 static const char parens[] = "=!<,>";
11508 if (paren && (p = strchr(parens, paren))) {
11509 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
11510 int flag = (p - parens) > 1;
11513 node = SUSPEND, flag = 0;
11514 reginsert(pRExC_state, node,ret, depth+1);
11515 Set_Node_Cur_Length(ret, parse_start);
11516 Set_Node_Offset(ret, parse_start + 1);
11518 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
11522 /* Check for proper termination. */
11524 /* restore original flags, but keep (?p) and, if we've changed from /d
11525 * rules to /u, keep the /u */
11526 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
11527 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
11528 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
11530 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
11531 RExC_parse = oregcomp_parse;
11532 vFAIL("Unmatched (");
11534 nextchar(pRExC_state);
11536 else if (!paren && RExC_parse < RExC_end) {
11537 if (*RExC_parse == ')') {
11539 vFAIL("Unmatched )");
11542 FAIL("Junk on end of regexp"); /* "Can't happen". */
11543 NOT_REACHED; /* NOTREACHED */
11546 if (RExC_in_lookbehind) {
11547 RExC_in_lookbehind--;
11549 if (after_freeze > RExC_npar)
11550 RExC_npar = after_freeze;
11555 - regbranch - one alternative of an | operator
11557 * Implements the concatenation operator.
11559 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11560 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11563 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
11566 regnode *chain = NULL;
11568 I32 flags = 0, c = 0;
11569 GET_RE_DEBUG_FLAGS_DECL;
11571 PERL_ARGS_ASSERT_REGBRANCH;
11573 DEBUG_PARSE("brnc");
11578 if (!SIZE_ONLY && RExC_extralen)
11579 ret = reganode(pRExC_state, BRANCHJ,0);
11581 ret = reg_node(pRExC_state, BRANCH);
11582 Set_Node_Length(ret, 1);
11586 if (!first && SIZE_ONLY)
11587 RExC_extralen += 1; /* BRANCHJ */
11589 *flagp = WORST; /* Tentatively. */
11591 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11592 FALSE /* Don't force to /x */ );
11593 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
11594 flags &= ~TRYAGAIN;
11595 latest = regpiece(pRExC_state, &flags,depth+1);
11596 if (latest == NULL) {
11597 if (flags & TRYAGAIN)
11599 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11600 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11603 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
11605 else if (ret == NULL)
11607 *flagp |= flags&(HASWIDTH|POSTPONED);
11608 if (chain == NULL) /* First piece. */
11609 *flagp |= flags&SPSTART;
11611 /* FIXME adding one for every branch after the first is probably
11612 * excessive now we have TRIE support. (hv) */
11614 REGTAIL(pRExC_state, chain, latest);
11619 if (chain == NULL) { /* Loop ran zero times. */
11620 chain = reg_node(pRExC_state, NOTHING);
11625 *flagp |= flags&SIMPLE;
11632 - regpiece - something followed by possible [*+?]
11634 * Note that the branching code sequences used for ? and the general cases
11635 * of * and + are somewhat optimized: they use the same NOTHING node as
11636 * both the endmarker for their branch list and the body of the last branch.
11637 * It might seem that this node could be dispensed with entirely, but the
11638 * endmarker role is not redundant.
11640 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
11642 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11643 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11646 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11652 const char * const origparse = RExC_parse;
11654 I32 max = REG_INFTY;
11655 #ifdef RE_TRACK_PATTERN_OFFSETS
11658 const char *maxpos = NULL;
11661 /* Save the original in case we change the emitted regop to a FAIL. */
11662 regnode * const orig_emit = RExC_emit;
11664 GET_RE_DEBUG_FLAGS_DECL;
11666 PERL_ARGS_ASSERT_REGPIECE;
11668 DEBUG_PARSE("piec");
11670 ret = regatom(pRExC_state, &flags,depth+1);
11672 if (flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8))
11673 *flagp |= flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8);
11675 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
11681 if (op == '{' && regcurly(RExC_parse)) {
11683 #ifdef RE_TRACK_PATTERN_OFFSETS
11684 parse_start = RExC_parse; /* MJD */
11686 next = RExC_parse + 1;
11687 while (isDIGIT(*next) || *next == ',') {
11688 if (*next == ',') {
11696 if (*next == '}') { /* got one */
11697 const char* endptr;
11701 if (isDIGIT(*RExC_parse)) {
11702 if (!grok_atoUV(RExC_parse, &uv, &endptr))
11703 vFAIL("Invalid quantifier in {,}");
11704 if (uv >= REG_INFTY)
11705 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11710 if (*maxpos == ',')
11713 maxpos = RExC_parse;
11714 if (isDIGIT(*maxpos)) {
11715 if (!grok_atoUV(maxpos, &uv, &endptr))
11716 vFAIL("Invalid quantifier in {,}");
11717 if (uv >= REG_INFTY)
11718 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11721 max = REG_INFTY; /* meaning "infinity" */
11724 nextchar(pRExC_state);
11725 if (max < min) { /* If can't match, warn and optimize to fail
11729 /* We can't back off the size because we have to reserve
11730 * enough space for all the things we are about to throw
11731 * away, but we can shrink it by the amount we are about
11732 * to re-use here */
11733 RExC_size += PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
11736 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
11737 RExC_emit = orig_emit;
11739 ret = reganode(pRExC_state, OPFAIL, 0);
11742 else if (min == max && *RExC_parse == '?')
11745 ckWARN2reg(RExC_parse + 1,
11746 "Useless use of greediness modifier '%c'",
11752 if ((flags&SIMPLE)) {
11753 if (min == 0 && max == REG_INFTY) {
11754 reginsert(pRExC_state, STAR, ret, depth+1);
11757 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11760 if (min == 1 && max == REG_INFTY) {
11761 reginsert(pRExC_state, PLUS, ret, depth+1);
11764 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11767 MARK_NAUGHTY_EXP(2, 2);
11768 reginsert(pRExC_state, CURLY, ret, depth+1);
11769 Set_Node_Offset(ret, parse_start+1); /* MJD */
11770 Set_Node_Cur_Length(ret, parse_start);
11773 regnode * const w = reg_node(pRExC_state, WHILEM);
11776 REGTAIL(pRExC_state, ret, w);
11777 if (!SIZE_ONLY && RExC_extralen) {
11778 reginsert(pRExC_state, LONGJMP,ret, depth+1);
11779 reginsert(pRExC_state, NOTHING,ret, depth+1);
11780 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
11782 reginsert(pRExC_state, CURLYX,ret, depth+1);
11784 Set_Node_Offset(ret, parse_start+1);
11785 Set_Node_Length(ret,
11786 op == '{' ? (RExC_parse - parse_start) : 1);
11788 if (!SIZE_ONLY && RExC_extralen)
11789 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
11790 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
11792 RExC_whilem_seen++, RExC_extralen += 3;
11793 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
11800 *flagp |= HASWIDTH;
11802 ARG1_SET(ret, (U16)min);
11803 ARG2_SET(ret, (U16)max);
11805 if (max == REG_INFTY)
11806 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11812 if (!ISMULT1(op)) {
11817 #if 0 /* Now runtime fix should be reliable. */
11819 /* if this is reinstated, don't forget to put this back into perldiag:
11821 =item Regexp *+ operand could be empty at {#} in regex m/%s/
11823 (F) The part of the regexp subject to either the * or + quantifier
11824 could match an empty string. The {#} shows in the regular
11825 expression about where the problem was discovered.
11829 if (!(flags&HASWIDTH) && op != '?')
11830 vFAIL("Regexp *+ operand could be empty");
11833 #ifdef RE_TRACK_PATTERN_OFFSETS
11834 parse_start = RExC_parse;
11836 nextchar(pRExC_state);
11838 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
11844 else if (op == '+') {
11848 else if (op == '?') {
11853 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
11854 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
11855 ckWARN2reg(RExC_parse,
11856 "%"UTF8f" matches null string many times",
11857 UTF8fARG(UTF, (RExC_parse >= origparse
11858 ? RExC_parse - origparse
11861 (void)ReREFCNT_inc(RExC_rx_sv);
11864 if (*RExC_parse == '?') {
11865 nextchar(pRExC_state);
11866 reginsert(pRExC_state, MINMOD, ret, depth+1);
11867 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
11869 else if (*RExC_parse == '+') {
11871 nextchar(pRExC_state);
11872 ender = reg_node(pRExC_state, SUCCEED);
11873 REGTAIL(pRExC_state, ret, ender);
11874 reginsert(pRExC_state, SUSPEND, ret, depth+1);
11876 ender = reg_node(pRExC_state, TAIL);
11877 REGTAIL(pRExC_state, ret, ender);
11880 if (ISMULT2(RExC_parse)) {
11882 vFAIL("Nested quantifiers");
11889 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
11898 /* This routine teases apart the various meanings of \N and returns
11899 * accordingly. The input parameters constrain which meaning(s) is/are valid
11900 * in the current context.
11902 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
11904 * If <code_point_p> is not NULL, the context is expecting the result to be a
11905 * single code point. If this \N instance turns out to a single code point,
11906 * the function returns TRUE and sets *code_point_p to that code point.
11908 * If <node_p> is not NULL, the context is expecting the result to be one of
11909 * the things representable by a regnode. If this \N instance turns out to be
11910 * one such, the function generates the regnode, returns TRUE and sets *node_p
11911 * to point to that regnode.
11913 * If this instance of \N isn't legal in any context, this function will
11914 * generate a fatal error and not return.
11916 * On input, RExC_parse should point to the first char following the \N at the
11917 * time of the call. On successful return, RExC_parse will have been updated
11918 * to point to just after the sequence identified by this routine. Also
11919 * *flagp has been updated as needed.
11921 * When there is some problem with the current context and this \N instance,
11922 * the function returns FALSE, without advancing RExC_parse, nor setting
11923 * *node_p, nor *code_point_p, nor *flagp.
11925 * If <cp_count> is not NULL, the caller wants to know the length (in code
11926 * points) that this \N sequence matches. This is set even if the function
11927 * returns FALSE, as detailed below.
11929 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
11931 * Probably the most common case is for the \N to specify a single code point.
11932 * *cp_count will be set to 1, and *code_point_p will be set to that code
11935 * Another possibility is for the input to be an empty \N{}, which for
11936 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
11937 * will be set to a generated NOTHING node.
11939 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
11940 * set to 0. *node_p will be set to a generated REG_ANY node.
11942 * The fourth possibility is that \N resolves to a sequence of more than one
11943 * code points. *cp_count will be set to the number of code points in the
11944 * sequence. *node_p * will be set to a generated node returned by this
11945 * function calling S_reg().
11947 * The final possibility is that it is premature to be calling this function;
11948 * that pass1 needs to be restarted. This can happen when this changes from
11949 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
11950 * latter occurs only when the fourth possibility would otherwise be in
11951 * effect, and is because one of those code points requires the pattern to be
11952 * recompiled as UTF-8. The function returns FALSE, and sets the
11953 * RESTART_PASS1 and NEED_UTF8 flags in *flagp, as appropriate. When this
11954 * happens, the caller needs to desist from continuing parsing, and return
11955 * this information to its caller. This is not set for when there is only one
11956 * code point, as this can be called as part of an ANYOF node, and they can
11957 * store above-Latin1 code points without the pattern having to be in UTF-8.
11959 * For non-single-quoted regexes, the tokenizer has resolved character and
11960 * sequence names inside \N{...} into their Unicode values, normalizing the
11961 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
11962 * hex-represented code points in the sequence. This is done there because
11963 * the names can vary based on what charnames pragma is in scope at the time,
11964 * so we need a way to take a snapshot of what they resolve to at the time of
11965 * the original parse. [perl #56444].
11967 * That parsing is skipped for single-quoted regexes, so we may here get
11968 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
11969 * parser. But if the single-quoted regex is something like '\N{U+41}', that
11970 * is legal and handled here. The code point is Unicode, and has to be
11971 * translated into the native character set for non-ASCII platforms.
11974 char * endbrace; /* points to '}' following the name */
11975 char *endchar; /* Points to '.' or '}' ending cur char in the input
11977 char* p = RExC_parse; /* Temporary */
11979 GET_RE_DEBUG_FLAGS_DECL;
11981 PERL_ARGS_ASSERT_GROK_BSLASH_N;
11983 GET_RE_DEBUG_FLAGS;
11985 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
11986 assert(! (node_p && cp_count)); /* At most 1 should be set */
11988 if (cp_count) { /* Initialize return for the most common case */
11992 /* The [^\n] meaning of \N ignores spaces and comments under the /x
11993 * modifier. The other meanings do not, so use a temporary until we find
11994 * out which we are being called with */
11995 skip_to_be_ignored_text(pRExC_state, &p,
11996 FALSE /* Don't force to /x */ );
11998 /* Disambiguate between \N meaning a named character versus \N meaning
11999 * [^\n]. The latter is assumed when the {...} following the \N is a legal
12000 * quantifier, or there is no '{' at all */
12001 if (*p != '{' || regcurly(p)) {
12011 *node_p = reg_node(pRExC_state, REG_ANY);
12012 *flagp |= HASWIDTH|SIMPLE;
12014 Set_Node_Length(*node_p, 1); /* MJD */
12018 /* Here, we have decided it should be a named character or sequence */
12020 /* The test above made sure that the next real character is a '{', but
12021 * under the /x modifier, it could be separated by space (or a comment and
12022 * \n) and this is not allowed (for consistency with \x{...} and the
12023 * tokenizer handling of \N{NAME}). */
12024 if (*RExC_parse != '{') {
12025 vFAIL("Missing braces on \\N{}");
12028 RExC_parse++; /* Skip past the '{' */
12030 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
12031 || ! (endbrace == RExC_parse /* nothing between the {} */
12032 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked... */
12033 && strnEQ(RExC_parse, "U+", 2)))) /* ... below for a better
12036 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
12037 vFAIL("\\N{NAME} must be resolved by the lexer");
12040 REQUIRE_UNI_RULES(flagp, FALSE); /* Unicode named chars imply Unicode
12043 if (endbrace == RExC_parse) { /* empty: \N{} */
12045 RExC_parse++; /* Position after the "}" */
12046 vFAIL("Zero length \\N{}");
12051 nextchar(pRExC_state);
12056 *node_p = reg_node(pRExC_state,NOTHING);
12060 RExC_parse += 2; /* Skip past the 'U+' */
12062 /* Because toke.c has generated a special construct for us guaranteed not
12063 * to have NULs, we can use a str function */
12064 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12066 /* Code points are separated by dots. If none, there is only one code
12067 * point, and is terminated by the brace */
12069 if (endchar >= endbrace) {
12070 STRLEN length_of_hex;
12071 I32 grok_hex_flags;
12073 /* Here, exactly one code point. If that isn't what is wanted, fail */
12074 if (! code_point_p) {
12079 /* Convert code point from hex */
12080 length_of_hex = (STRLEN)(endchar - RExC_parse);
12081 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
12082 | PERL_SCAN_DISALLOW_PREFIX
12084 /* No errors in the first pass (See [perl
12085 * #122671].) We let the code below find the
12086 * errors when there are multiple chars. */
12088 ? PERL_SCAN_SILENT_ILLDIGIT
12091 /* This routine is the one place where both single- and double-quotish
12092 * \N{U+xxxx} are evaluated. The value is a Unicode code point which
12093 * must be converted to native. */
12094 *code_point_p = UNI_TO_NATIVE(grok_hex(RExC_parse,
12099 /* The tokenizer should have guaranteed validity, but it's possible to
12100 * bypass it by using single quoting, so check. Don't do the check
12101 * here when there are multiple chars; we do it below anyway. */
12102 if (length_of_hex == 0
12103 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
12105 RExC_parse += length_of_hex; /* Includes all the valid */
12106 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12107 ? UTF8SKIP(RExC_parse)
12109 /* Guard against malformed utf8 */
12110 if (RExC_parse >= endchar) {
12111 RExC_parse = endchar;
12113 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12116 RExC_parse = endbrace + 1;
12119 else { /* Is a multiple character sequence */
12120 SV * substitute_parse;
12122 char *orig_end = RExC_end;
12123 char *save_start = RExC_start;
12126 /* Count the code points, if desired, in the sequence */
12129 while (RExC_parse < endbrace) {
12130 /* Point to the beginning of the next character in the sequence. */
12131 RExC_parse = endchar + 1;
12132 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12137 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12138 * But don't backup up the pointer if the caller want to know how many
12139 * code points there are (they can then handle things) */
12147 /* What is done here is to convert this to a sub-pattern of the form
12148 * \x{char1}\x{char2}... and then call reg recursively to parse it
12149 * (enclosing in "(?: ... )" ). That way, it retains its atomicness,
12150 * while not having to worry about special handling that some code
12151 * points may have. */
12153 substitute_parse = newSVpvs("?:");
12155 while (RExC_parse < endbrace) {
12157 /* Convert to notation the rest of the code understands */
12158 sv_catpv(substitute_parse, "\\x{");
12159 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
12160 sv_catpv(substitute_parse, "}");
12162 /* Point to the beginning of the next character in the sequence. */
12163 RExC_parse = endchar + 1;
12164 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12167 sv_catpv(substitute_parse, ")");
12169 RExC_parse = RExC_start = RExC_adjusted_start = SvPV(substitute_parse,
12172 /* Don't allow empty number */
12173 if (len < (STRLEN) 8) {
12174 RExC_parse = endbrace;
12175 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12177 RExC_end = RExC_parse + len;
12179 /* The values are Unicode, and therefore not subject to recoding, but
12180 * have to be converted to native on a non-Unicode (meaning non-ASCII)
12182 RExC_override_recoding = 1;
12184 RExC_recode_x_to_native = 1;
12188 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
12189 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12190 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12193 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
12196 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12199 /* Restore the saved values */
12200 RExC_start = RExC_adjusted_start = save_start;
12201 RExC_parse = endbrace;
12202 RExC_end = orig_end;
12203 RExC_override_recoding = 0;
12205 RExC_recode_x_to_native = 0;
12208 SvREFCNT_dec_NN(substitute_parse);
12209 nextchar(pRExC_state);
12216 PERL_STATIC_INLINE U8
12217 S_compute_EXACTish(RExC_state_t *pRExC_state)
12221 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
12229 op = get_regex_charset(RExC_flags);
12230 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
12231 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
12232 been, so there is no hole */
12235 return op + EXACTF;
12238 PERL_STATIC_INLINE void
12239 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
12240 regnode *node, I32* flagp, STRLEN len, UV code_point,
12243 /* This knows the details about sizing an EXACTish node, setting flags for
12244 * it (by setting <*flagp>, and potentially populating it with a single
12247 * If <len> (the length in bytes) is non-zero, this function assumes that
12248 * the node has already been populated, and just does the sizing. In this
12249 * case <code_point> should be the final code point that has already been
12250 * placed into the node. This value will be ignored except that under some
12251 * circumstances <*flagp> is set based on it.
12253 * If <len> is zero, the function assumes that the node is to contain only
12254 * the single character given by <code_point> and calculates what <len>
12255 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
12256 * additionally will populate the node's STRING with <code_point> or its
12259 * In both cases <*flagp> is appropriately set
12261 * It knows that under FOLD, the Latin Sharp S and UTF characters above
12262 * 255, must be folded (the former only when the rules indicate it can
12265 * When it does the populating, it looks at the flag 'downgradable'. If
12266 * true with a node that folds, it checks if the single code point
12267 * participates in a fold, and if not downgrades the node to an EXACT.
12268 * This helps the optimizer */
12270 bool len_passed_in = cBOOL(len != 0);
12271 U8 character[UTF8_MAXBYTES_CASE+1];
12273 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
12275 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
12276 * sizing difference, and is extra work that is thrown away */
12277 if (downgradable && ! PASS2) {
12278 downgradable = FALSE;
12281 if (! len_passed_in) {
12283 if (UVCHR_IS_INVARIANT(code_point)) {
12284 if (LOC || ! FOLD) { /* /l defers folding until runtime */
12285 *character = (U8) code_point;
12287 else { /* Here is /i and not /l. (toFOLD() is defined on just
12288 ASCII, which isn't the same thing as INVARIANT on
12289 EBCDIC, but it works there, as the extra invariants
12290 fold to themselves) */
12291 *character = toFOLD((U8) code_point);
12293 /* We can downgrade to an EXACT node if this character
12294 * isn't a folding one. Note that this assumes that
12295 * nothing above Latin1 folds to some other invariant than
12296 * one of these alphabetics; otherwise we would also have
12298 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12299 * || ASCII_FOLD_RESTRICTED))
12301 if (downgradable && PL_fold[code_point] == code_point) {
12307 else if (FOLD && (! LOC
12308 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
12309 { /* Folding, and ok to do so now */
12310 UV folded = _to_uni_fold_flags(
12314 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12315 ? FOLD_FLAGS_NOMIX_ASCII
12318 && folded == code_point /* This quickly rules out many
12319 cases, avoiding the
12320 _invlist_contains_cp() overhead
12322 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
12329 else if (code_point <= MAX_UTF8_TWO_BYTE) {
12331 /* Not folding this cp, and can output it directly */
12332 *character = UTF8_TWO_BYTE_HI(code_point);
12333 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
12337 uvchr_to_utf8( character, code_point);
12338 len = UTF8SKIP(character);
12340 } /* Else pattern isn't UTF8. */
12342 *character = (U8) code_point;
12344 } /* Else is folded non-UTF8 */
12345 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12346 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12347 || UNICODE_DOT_DOT_VERSION > 0)
12348 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
12352 /* We don't fold any non-UTF8 except possibly the Sharp s (see
12353 * comments at join_exact()); */
12354 *character = (U8) code_point;
12357 /* Can turn into an EXACT node if we know the fold at compile time,
12358 * and it folds to itself and doesn't particpate in other folds */
12361 && PL_fold_latin1[code_point] == code_point
12362 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12363 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
12367 } /* else is Sharp s. May need to fold it */
12368 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
12370 *(character + 1) = 's';
12374 *character = LATIN_SMALL_LETTER_SHARP_S;
12380 RExC_size += STR_SZ(len);
12383 RExC_emit += STR_SZ(len);
12384 STR_LEN(node) = len;
12385 if (! len_passed_in) {
12386 Copy((char *) character, STRING(node), len, char);
12390 *flagp |= HASWIDTH;
12392 /* A single character node is SIMPLE, except for the special-cased SHARP S
12394 if ((len == 1 || (UTF && len == UVCHR_SKIP(code_point)))
12395 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12396 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12397 || UNICODE_DOT_DOT_VERSION > 0)
12398 && ( code_point != LATIN_SMALL_LETTER_SHARP_S
12399 || ! FOLD || ! DEPENDS_SEMANTICS)
12405 /* The OP may not be well defined in PASS1 */
12406 if (PASS2 && OP(node) == EXACTFL) {
12407 RExC_contains_locale = 1;
12412 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
12413 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
12416 S_backref_value(char *p)
12418 const char* endptr;
12420 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
12427 - regatom - the lowest level
12429 Try to identify anything special at the start of the current parse position.
12430 If there is, then handle it as required. This may involve generating a
12431 single regop, such as for an assertion; or it may involve recursing, such as
12432 to handle a () structure.
12434 If the string doesn't start with something special then we gobble up
12435 as much literal text as we can. If we encounter a quantifier, we have to
12436 back off the final literal character, as that quantifier applies to just it
12437 and not to the whole string of literals.
12439 Once we have been able to handle whatever type of thing started the
12440 sequence, we return.
12442 Note: we have to be careful with escapes, as they can be both literal
12443 and special, and in the case of \10 and friends, context determines which.
12445 A summary of the code structure is:
12447 switch (first_byte) {
12448 cases for each special:
12449 handle this special;
12452 switch (2nd byte) {
12453 cases for each unambiguous special:
12454 handle this special;
12456 cases for each ambigous special/literal:
12458 if (special) handle here
12460 default: // unambiguously literal:
12463 default: // is a literal char
12466 create EXACTish node for literal;
12467 while (more input and node isn't full) {
12468 switch (input_byte) {
12469 cases for each special;
12470 make sure parse pointer is set so that the next call to
12471 regatom will see this special first
12472 goto loopdone; // EXACTish node terminated by prev. char
12474 append char to EXACTISH node;
12476 get next input byte;
12480 return the generated node;
12482 Specifically there are two separate switches for handling
12483 escape sequences, with the one for handling literal escapes requiring
12484 a dummy entry for all of the special escapes that are actually handled
12487 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
12489 Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
12490 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
12491 Otherwise does not return NULL.
12495 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12497 regnode *ret = NULL;
12504 GET_RE_DEBUG_FLAGS_DECL;
12506 *flagp = WORST; /* Tentatively. */
12508 DEBUG_PARSE("atom");
12510 PERL_ARGS_ASSERT_REGATOM;
12513 parse_start = RExC_parse;
12514 assert(RExC_parse < RExC_end);
12515 switch ((U8)*RExC_parse) {
12517 RExC_seen_zerolen++;
12518 nextchar(pRExC_state);
12519 if (RExC_flags & RXf_PMf_MULTILINE)
12520 ret = reg_node(pRExC_state, MBOL);
12522 ret = reg_node(pRExC_state, SBOL);
12523 Set_Node_Length(ret, 1); /* MJD */
12526 nextchar(pRExC_state);
12528 RExC_seen_zerolen++;
12529 if (RExC_flags & RXf_PMf_MULTILINE)
12530 ret = reg_node(pRExC_state, MEOL);
12532 ret = reg_node(pRExC_state, SEOL);
12533 Set_Node_Length(ret, 1); /* MJD */
12536 nextchar(pRExC_state);
12537 if (RExC_flags & RXf_PMf_SINGLELINE)
12538 ret = reg_node(pRExC_state, SANY);
12540 ret = reg_node(pRExC_state, REG_ANY);
12541 *flagp |= HASWIDTH|SIMPLE;
12543 Set_Node_Length(ret, 1); /* MJD */
12547 char * const oregcomp_parse = ++RExC_parse;
12548 ret = regclass(pRExC_state, flagp,depth+1,
12549 FALSE, /* means parse the whole char class */
12550 TRUE, /* allow multi-char folds */
12551 FALSE, /* don't silence non-portable warnings. */
12552 (bool) RExC_strict,
12553 TRUE, /* Allow an optimized regnode result */
12557 if (*flagp & (RESTART_PASS1|NEED_UTF8))
12559 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
12562 if (*RExC_parse != ']') {
12563 RExC_parse = oregcomp_parse;
12564 vFAIL("Unmatched [");
12566 nextchar(pRExC_state);
12567 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
12571 nextchar(pRExC_state);
12572 ret = reg(pRExC_state, 2, &flags,depth+1);
12574 if (flags & TRYAGAIN) {
12575 if (RExC_parse >= RExC_end) {
12576 /* Make parent create an empty node if needed. */
12577 *flagp |= TRYAGAIN;
12582 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12583 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12586 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"",
12589 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12593 if (flags & TRYAGAIN) {
12594 *flagp |= TRYAGAIN;
12597 vFAIL("Internal urp");
12598 /* Supposed to be caught earlier. */
12604 vFAIL("Quantifier follows nothing");
12609 This switch handles escape sequences that resolve to some kind
12610 of special regop and not to literal text. Escape sequnces that
12611 resolve to literal text are handled below in the switch marked
12614 Every entry in this switch *must* have a corresponding entry
12615 in the literal escape switch. However, the opposite is not
12616 required, as the default for this switch is to jump to the
12617 literal text handling code.
12620 switch ((U8)*RExC_parse) {
12621 /* Special Escapes */
12623 RExC_seen_zerolen++;
12624 ret = reg_node(pRExC_state, SBOL);
12625 /* SBOL is shared with /^/ so we set the flags so we can tell
12626 * /\A/ from /^/ in split. We check ret because first pass we
12627 * have no regop struct to set the flags on. */
12631 goto finish_meta_pat;
12633 ret = reg_node(pRExC_state, GPOS);
12634 RExC_seen |= REG_GPOS_SEEN;
12636 goto finish_meta_pat;
12638 RExC_seen_zerolen++;
12639 ret = reg_node(pRExC_state, KEEPS);
12641 /* XXX:dmq : disabling in-place substitution seems to
12642 * be necessary here to avoid cases of memory corruption, as
12643 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
12645 RExC_seen |= REG_LOOKBEHIND_SEEN;
12646 goto finish_meta_pat;
12648 ret = reg_node(pRExC_state, SEOL);
12650 RExC_seen_zerolen++; /* Do not optimize RE away */
12651 goto finish_meta_pat;
12653 ret = reg_node(pRExC_state, EOS);
12655 RExC_seen_zerolen++; /* Do not optimize RE away */
12656 goto finish_meta_pat;
12658 vFAIL("\\C no longer supported");
12660 ret = reg_node(pRExC_state, CLUMP);
12661 *flagp |= HASWIDTH;
12662 goto finish_meta_pat;
12668 arg = ANYOF_WORDCHAR;
12676 regex_charset charset = get_regex_charset(RExC_flags);
12678 RExC_seen_zerolen++;
12679 RExC_seen |= REG_LOOKBEHIND_SEEN;
12680 op = BOUND + charset;
12682 if (op == BOUNDL) {
12683 RExC_contains_locale = 1;
12686 ret = reg_node(pRExC_state, op);
12688 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
12689 FLAGS(ret) = TRADITIONAL_BOUND;
12690 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
12696 char name = *RExC_parse;
12699 endbrace = strchr(RExC_parse, '}');
12702 vFAIL2("Missing right brace on \\%c{}", name);
12704 /* XXX Need to decide whether to take spaces or not. Should be
12705 * consistent with \p{}, but that currently is SPACE, which
12706 * means vertical too, which seems wrong
12707 * while (isBLANK(*RExC_parse)) {
12710 if (endbrace == RExC_parse) {
12711 RExC_parse++; /* After the '}' */
12712 vFAIL2("Empty \\%c{}", name);
12714 length = endbrace - RExC_parse;
12715 /*while (isBLANK(*(RExC_parse + length - 1))) {
12718 switch (*RExC_parse) {
12721 && (length != 3 || strnNE(RExC_parse + 1, "cb", 2)))
12723 goto bad_bound_type;
12725 FLAGS(ret) = GCB_BOUND;
12728 if (length != 2 || *(RExC_parse + 1) != 'b') {
12729 goto bad_bound_type;
12731 FLAGS(ret) = LB_BOUND;
12734 if (length != 2 || *(RExC_parse + 1) != 'b') {
12735 goto bad_bound_type;
12737 FLAGS(ret) = SB_BOUND;
12740 if (length != 2 || *(RExC_parse + 1) != 'b') {
12741 goto bad_bound_type;
12743 FLAGS(ret) = WB_BOUND;
12747 RExC_parse = endbrace;
12749 "'%"UTF8f"' is an unknown bound type",
12750 UTF8fARG(UTF, length, endbrace - length));
12751 NOT_REACHED; /*NOTREACHED*/
12753 RExC_parse = endbrace;
12754 REQUIRE_UNI_RULES(flagp, NULL);
12756 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
12760 /* Don't have to worry about UTF-8, in this message because
12761 * to get here the contents of the \b must be ASCII */
12762 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
12763 "Using /u for '%.*s' instead of /%s",
12765 endbrace - length + 1,
12766 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
12767 ? ASCII_RESTRICT_PAT_MODS
12768 : ASCII_MORE_RESTRICT_PAT_MODS);
12772 if (PASS2 && invert) {
12773 OP(ret) += NBOUND - BOUND;
12775 goto finish_meta_pat;
12783 if (! DEPENDS_SEMANTICS) {
12787 /* \d doesn't have any matches in the upper Latin1 range, hence /d
12788 * is equivalent to /u. Changing to /u saves some branches at
12791 goto join_posix_op_known;
12794 ret = reg_node(pRExC_state, LNBREAK);
12795 *flagp |= HASWIDTH|SIMPLE;
12796 goto finish_meta_pat;
12804 goto join_posix_op_known;
12810 arg = ANYOF_VERTWS;
12812 goto join_posix_op_known;
12822 op = POSIXD + get_regex_charset(RExC_flags);
12823 if (op > POSIXA) { /* /aa is same as /a */
12826 else if (op == POSIXL) {
12827 RExC_contains_locale = 1;
12830 join_posix_op_known:
12833 op += NPOSIXD - POSIXD;
12836 ret = reg_node(pRExC_state, op);
12838 FLAGS(ret) = namedclass_to_classnum(arg);
12841 *flagp |= HASWIDTH|SIMPLE;
12845 nextchar(pRExC_state);
12846 Set_Node_Length(ret, 2); /* MJD */
12852 ret = regclass(pRExC_state, flagp,depth+1,
12853 TRUE, /* means just parse this element */
12854 FALSE, /* don't allow multi-char folds */
12855 FALSE, /* don't silence non-portable warnings. It
12856 would be a bug if these returned
12858 (bool) RExC_strict,
12859 TRUE, /* Allow an optimized regnode result */
12862 if (*flagp & RESTART_PASS1)
12864 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
12865 * multi-char folds are allowed. */
12867 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
12872 Set_Node_Offset(ret, parse_start);
12873 Set_Node_Cur_Length(ret, parse_start - 2);
12874 nextchar(pRExC_state);
12877 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
12878 * \N{...} evaluates to a sequence of more than one code points).
12879 * The function call below returns a regnode, which is our result.
12880 * The parameters cause it to fail if the \N{} evaluates to a
12881 * single code point; we handle those like any other literal. The
12882 * reason that the multicharacter case is handled here and not as
12883 * part of the EXACtish code is because of quantifiers. In
12884 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
12885 * this way makes that Just Happen. dmq.
12886 * join_exact() will join this up with adjacent EXACTish nodes
12887 * later on, if appropriate. */
12889 if (grok_bslash_N(pRExC_state,
12890 &ret, /* Want a regnode returned */
12891 NULL, /* Fail if evaluates to a single code
12893 NULL, /* Don't need a count of how many code
12902 if (*flagp & RESTART_PASS1)
12905 /* Here, evaluates to a single code point. Go get that */
12906 RExC_parse = parse_start;
12909 case 'k': /* Handle \k<NAME> and \k'NAME' */
12913 if ( RExC_parse >= RExC_end - 1
12914 || (( ch = RExC_parse[1]) != '<'
12919 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12920 vFAIL2("Sequence %.2s... not terminated",parse_start);
12923 ret = handle_named_backref(pRExC_state,
12935 case '1': case '2': case '3': case '4':
12936 case '5': case '6': case '7': case '8': case '9':
12941 if (*RExC_parse == 'g') {
12945 if (*RExC_parse == '{') {
12949 if (*RExC_parse == '-') {
12953 if (hasbrace && !isDIGIT(*RExC_parse)) {
12954 if (isrel) RExC_parse--;
12956 goto parse_named_seq;
12959 if (RExC_parse >= RExC_end) {
12960 goto unterminated_g;
12962 num = S_backref_value(RExC_parse);
12964 vFAIL("Reference to invalid group 0");
12965 else if (num == I32_MAX) {
12966 if (isDIGIT(*RExC_parse))
12967 vFAIL("Reference to nonexistent group");
12970 vFAIL("Unterminated \\g... pattern");
12974 num = RExC_npar - num;
12976 vFAIL("Reference to nonexistent or unclosed group");
12980 num = S_backref_value(RExC_parse);
12981 /* bare \NNN might be backref or octal - if it is larger
12982 * than or equal RExC_npar then it is assumed to be an
12983 * octal escape. Note RExC_npar is +1 from the actual
12984 * number of parens. */
12985 /* Note we do NOT check if num == I32_MAX here, as that is
12986 * handled by the RExC_npar check */
12989 /* any numeric escape < 10 is always a backref */
12991 /* any numeric escape < RExC_npar is a backref */
12992 && num >= RExC_npar
12993 /* cannot be an octal escape if it starts with 8 */
12994 && *RExC_parse != '8'
12995 /* cannot be an octal escape it it starts with 9 */
12996 && *RExC_parse != '9'
12999 /* Probably not a backref, instead likely to be an
13000 * octal character escape, e.g. \35 or \777.
13001 * The above logic should make it obvious why using
13002 * octal escapes in patterns is problematic. - Yves */
13003 RExC_parse = parse_start;
13008 /* At this point RExC_parse points at a numeric escape like
13009 * \12 or \88 or something similar, which we should NOT treat
13010 * as an octal escape. It may or may not be a valid backref
13011 * escape. For instance \88888888 is unlikely to be a valid
13013 while (isDIGIT(*RExC_parse))
13016 if (*RExC_parse != '}')
13017 vFAIL("Unterminated \\g{...} pattern");
13021 if (num > (I32)RExC_rx->nparens)
13022 vFAIL("Reference to nonexistent group");
13025 ret = reganode(pRExC_state,
13028 : (ASCII_FOLD_RESTRICTED)
13030 : (AT_LEAST_UNI_SEMANTICS)
13036 *flagp |= HASWIDTH;
13038 /* override incorrect value set in reganode MJD */
13039 Set_Node_Offset(ret, parse_start);
13040 Set_Node_Cur_Length(ret, parse_start-1);
13041 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13042 FALSE /* Don't force to /x */ );
13046 if (RExC_parse >= RExC_end)
13047 FAIL("Trailing \\");
13050 /* Do not generate "unrecognized" warnings here, we fall
13051 back into the quick-grab loop below */
13052 RExC_parse = parse_start;
13054 } /* end of switch on a \foo sequence */
13059 /* '#' comments should have been spaced over before this function was
13061 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13063 if (RExC_flags & RXf_PMf_EXTENDED) {
13064 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13065 if (RExC_parse < RExC_end)
13075 /* Here, we have determined that the next thing is probably a
13076 * literal character. RExC_parse points to the first byte of its
13077 * definition. (It still may be an escape sequence that evaluates
13078 * to a single character) */
13084 #define MAX_NODE_STRING_SIZE 127
13085 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
13087 U8 upper_parse = MAX_NODE_STRING_SIZE;
13088 U8 node_type = compute_EXACTish(pRExC_state);
13089 bool next_is_quantifier;
13090 char * oldp = NULL;
13092 /* We can convert EXACTF nodes to EXACTFU if they contain only
13093 * characters that match identically regardless of the target
13094 * string's UTF8ness. The reason to do this is that EXACTF is not
13095 * trie-able, EXACTFU is.
13097 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13098 * contain only above-Latin1 characters (hence must be in UTF8),
13099 * which don't participate in folds with Latin1-range characters,
13100 * as the latter's folds aren't known until runtime. (We don't
13101 * need to figure this out until pass 2) */
13102 bool maybe_exactfu = PASS2
13103 && (node_type == EXACTF || node_type == EXACTFL);
13105 /* If a folding node contains only code points that don't
13106 * participate in folds, it can be changed into an EXACT node,
13107 * which allows the optimizer more things to look for */
13110 ret = reg_node(pRExC_state, node_type);
13112 /* In pass1, folded, we use a temporary buffer instead of the
13113 * actual node, as the node doesn't exist yet */
13114 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
13120 /* We look for the EXACTFish to EXACT node optimizaton only if
13121 * folding. (And we don't need to figure this out until pass 2).
13122 * XXX It might actually make sense to split the node into portions
13123 * that are exact and ones that aren't, so that we could later use
13124 * the exact ones to find the longest fixed and floating strings.
13125 * One would want to join them back into a larger node. One could
13126 * use a pseudo regnode like 'EXACT_ORIG_FOLD' */
13127 maybe_exact = FOLD && PASS2;
13129 /* XXX The node can hold up to 255 bytes, yet this only goes to
13130 * 127. I (khw) do not know why. Keeping it somewhat less than
13131 * 255 allows us to not have to worry about overflow due to
13132 * converting to utf8 and fold expansion, but that value is
13133 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
13134 * split up by this limit into a single one using the real max of
13135 * 255. Even at 127, this breaks under rare circumstances. If
13136 * folding, we do not want to split a node at a character that is a
13137 * non-final in a multi-char fold, as an input string could just
13138 * happen to want to match across the node boundary. The join
13139 * would solve that problem if the join actually happens. But a
13140 * series of more than two nodes in a row each of 127 would cause
13141 * the first join to succeed to get to 254, but then there wouldn't
13142 * be room for the next one, which could at be one of those split
13143 * multi-char folds. I don't know of any fool-proof solution. One
13144 * could back off to end with only a code point that isn't such a
13145 * non-final, but it is possible for there not to be any in the
13148 assert( ! UTF /* Is at the beginning of a character */
13149 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13150 || UTF8_IS_START(UCHARAT(RExC_parse)));
13152 /* Here, we have a literal character. Find the maximal string of
13153 * them in the input that we can fit into a single EXACTish node.
13154 * We quit at the first non-literal or when the node gets full */
13155 for (p = RExC_parse;
13156 len < upper_parse && p < RExC_end;
13161 /* White space has already been ignored */
13162 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13163 || ! is_PATWS_safe((p), RExC_end, UTF));
13175 /* Literal Escapes Switch
13177 This switch is meant to handle escape sequences that
13178 resolve to a literal character.
13180 Every escape sequence that represents something
13181 else, like an assertion or a char class, is handled
13182 in the switch marked 'Special Escapes' above in this
13183 routine, but also has an entry here as anything that
13184 isn't explicitly mentioned here will be treated as
13185 an unescaped equivalent literal.
13188 switch ((U8)*++p) {
13189 /* These are all the special escapes. */
13190 case 'A': /* Start assertion */
13191 case 'b': case 'B': /* Word-boundary assertion*/
13192 case 'C': /* Single char !DANGEROUS! */
13193 case 'd': case 'D': /* digit class */
13194 case 'g': case 'G': /* generic-backref, pos assertion */
13195 case 'h': case 'H': /* HORIZWS */
13196 case 'k': case 'K': /* named backref, keep marker */
13197 case 'p': case 'P': /* Unicode property */
13198 case 'R': /* LNBREAK */
13199 case 's': case 'S': /* space class */
13200 case 'v': case 'V': /* VERTWS */
13201 case 'w': case 'W': /* word class */
13202 case 'X': /* eXtended Unicode "combining
13203 character sequence" */
13204 case 'z': case 'Z': /* End of line/string assertion */
13208 /* Anything after here is an escape that resolves to a
13209 literal. (Except digits, which may or may not)
13215 case 'N': /* Handle a single-code point named character. */
13216 RExC_parse = p + 1;
13217 if (! grok_bslash_N(pRExC_state,
13218 NULL, /* Fail if evaluates to
13219 anything other than a
13220 single code point */
13221 &ender, /* The returned single code
13223 NULL, /* Don't need a count of
13224 how many code points */
13229 if (*flagp & NEED_UTF8)
13230 FAIL("panic: grok_bslash_N set NEED_UTF8");
13231 if (*flagp & RESTART_PASS1)
13234 /* Here, it wasn't a single code point. Go close
13235 * up this EXACTish node. The switch() prior to
13236 * this switch handles the other cases */
13237 RExC_parse = p = oldp;
13241 if (ender > 0xff) {
13242 REQUIRE_UTF8(flagp);
13258 ender = ESC_NATIVE;
13268 const char* error_msg;
13270 bool valid = grok_bslash_o(&p,
13273 PASS2, /* out warnings */
13274 (bool) RExC_strict,
13275 TRUE, /* Output warnings
13280 RExC_parse = p; /* going to die anyway; point
13281 to exact spot of failure */
13285 if (ender > 0xff) {
13286 REQUIRE_UTF8(flagp);
13292 UV result = UV_MAX; /* initialize to erroneous
13294 const char* error_msg;
13296 bool valid = grok_bslash_x(&p,
13299 PASS2, /* out warnings */
13300 (bool) RExC_strict,
13301 TRUE, /* Silence warnings
13306 RExC_parse = p; /* going to die anyway; point
13307 to exact spot of failure */
13312 if (ender < 0x100) {
13314 if (RExC_recode_x_to_native) {
13315 ender = LATIN1_TO_NATIVE(ender);
13320 REQUIRE_UTF8(flagp);
13326 ender = grok_bslash_c(*p++, PASS2);
13328 case '8': case '9': /* must be a backreference */
13330 /* we have an escape like \8 which cannot be an octal escape
13331 * so we exit the loop, and let the outer loop handle this
13332 * escape which may or may not be a legitimate backref. */
13334 case '1': case '2': case '3':case '4':
13335 case '5': case '6': case '7':
13336 /* When we parse backslash escapes there is ambiguity
13337 * between backreferences and octal escapes. Any escape
13338 * from \1 - \9 is a backreference, any multi-digit
13339 * escape which does not start with 0 and which when
13340 * evaluated as decimal could refer to an already
13341 * parsed capture buffer is a back reference. Anything
13344 * Note this implies that \118 could be interpreted as
13345 * 118 OR as "\11" . "8" depending on whether there
13346 * were 118 capture buffers defined already in the
13349 /* NOTE, RExC_npar is 1 more than the actual number of
13350 * parens we have seen so far, hence the < RExC_npar below. */
13352 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
13353 { /* Not to be treated as an octal constant, go
13361 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13363 ender = grok_oct(p, &numlen, &flags, NULL);
13364 if (ender > 0xff) {
13365 REQUIRE_UTF8(flagp);
13368 if (PASS2 /* like \08, \178 */
13370 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
13372 reg_warn_non_literal_string(
13374 form_short_octal_warning(p, numlen));
13380 FAIL("Trailing \\");
13383 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
13384 /* Include any left brace following the alpha to emphasize
13385 * that it could be part of an escape at some point
13387 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
13388 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
13390 goto normal_default;
13391 } /* End of switch on '\' */
13394 /* Currently we don't care if the lbrace is at the start
13395 * of a construct. This catches it in the middle of a
13396 * literal string, or when it's the first thing after
13397 * something like "\b" */
13398 if (len || (p > RExC_start && isALPHA_A(*(p -1)))) {
13399 RExC_parse = p + 1;
13400 vFAIL("Unescaped left brace in regex is illegal here");
13403 default: /* A literal character */
13405 if (! UTF8_IS_INVARIANT(*p) && UTF) {
13407 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
13408 &numlen, UTF8_ALLOW_DEFAULT);
13414 } /* End of switch on the literal */
13416 /* Here, have looked at the literal character and <ender>
13417 * contains its ordinal, <p> points to the character after it.
13418 * We need to check if the next non-ignored thing is a
13419 * quantifier. Move <p> to after anything that should be
13420 * ignored, which, as a side effect, positions <p> for the next
13421 * loop iteration */
13422 skip_to_be_ignored_text(pRExC_state, &p,
13423 FALSE /* Don't force to /x */ );
13425 /* If the next thing is a quantifier, it applies to this
13426 * character only, which means that this character has to be in
13427 * its own node and can't just be appended to the string in an
13428 * existing node, so if there are already other characters in
13429 * the node, close the node with just them, and set up to do
13430 * this character again next time through, when it will be the
13431 * only thing in its new node */
13433 if ((next_is_quantifier = ( LIKELY(p < RExC_end)
13434 && UNLIKELY(ISMULT2(p))))
13441 /* Ready to add 'ender' to the node */
13443 if (! FOLD) { /* The simple case, just append the literal */
13445 /* In the sizing pass, we need only the size of the
13446 * character we are appending, hence we can delay getting
13447 * its representation until PASS2. */
13450 const STRLEN unilen = UVCHR_SKIP(ender);
13453 /* We have to subtract 1 just below (and again in
13454 * the corresponding PASS2 code) because the loop
13455 * increments <len> each time, as all but this path
13456 * (and one other) through it add a single byte to
13457 * the EXACTish node. But these paths would change
13458 * len to be the correct final value, so cancel out
13459 * the increment that follows */
13465 } else { /* PASS2 */
13468 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
13469 len += (char *) new_s - s - 1;
13470 s = (char *) new_s;
13473 *(s++) = (char) ender;
13477 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
13479 /* Here are folding under /l, and the code point is
13480 * problematic. First, we know we can't simplify things */
13481 maybe_exact = FALSE;
13482 maybe_exactfu = FALSE;
13484 /* A problematic code point in this context means that its
13485 * fold isn't known until runtime, so we can't fold it now.
13486 * (The non-problematic code points are the above-Latin1
13487 * ones that fold to also all above-Latin1. Their folds
13488 * don't vary no matter what the locale is.) But here we
13489 * have characters whose fold depends on the locale.
13490 * Unlike the non-folding case above, we have to keep track
13491 * of these in the sizing pass, so that we can make sure we
13492 * don't split too-long nodes in the middle of a potential
13493 * multi-char fold. And unlike the regular fold case
13494 * handled in the else clauses below, we don't actually
13495 * fold and don't have special cases to consider. What we
13496 * do for both passes is the PASS2 code for non-folding */
13497 goto not_fold_common;
13499 else /* A regular FOLD code point */
13501 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13502 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13503 || UNICODE_DOT_DOT_VERSION > 0)
13504 /* See comments for join_exact() as to why we fold
13505 * this non-UTF at compile time */
13506 || ( node_type == EXACTFU
13507 && ender == LATIN_SMALL_LETTER_SHARP_S)
13510 /* Here, are folding and are not UTF-8 encoded; therefore
13511 * the character must be in the range 0-255, and is not /l
13512 * (Not /l because we already handled these under /l in
13513 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
13514 if (IS_IN_SOME_FOLD_L1(ender)) {
13515 maybe_exact = FALSE;
13517 /* See if the character's fold differs between /d and
13518 * /u. This includes the multi-char fold SHARP S to
13520 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
13521 RExC_seen_unfolded_sharp_s = 1;
13522 maybe_exactfu = FALSE;
13524 else if (maybe_exactfu
13525 && (PL_fold[ender] != PL_fold_latin1[ender]
13526 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13527 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13528 || UNICODE_DOT_DOT_VERSION > 0)
13530 && isALPHA_FOLD_EQ(ender, 's')
13531 && isALPHA_FOLD_EQ(*(s-1), 's'))
13534 maybe_exactfu = FALSE;
13538 /* Even when folding, we store just the input character, as
13539 * we have an array that finds its fold quickly */
13540 *(s++) = (char) ender;
13542 else { /* FOLD, and UTF (or sharp s) */
13543 /* Unlike the non-fold case, we do actually have to
13544 * calculate the results here in pass 1. This is for two
13545 * reasons, the folded length may be longer than the
13546 * unfolded, and we have to calculate how many EXACTish
13547 * nodes it will take; and we may run out of room in a node
13548 * in the middle of a potential multi-char fold, and have
13549 * to back off accordingly. */
13552 if (isASCII_uni(ender)) {
13553 folded = toFOLD(ender);
13554 *(s)++ = (U8) folded;
13559 folded = _to_uni_fold_flags(
13563 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
13564 ? FOLD_FLAGS_NOMIX_ASCII
13568 /* The loop increments <len> each time, as all but this
13569 * path (and one other) through it add a single byte to
13570 * the EXACTish node. But this one has changed len to
13571 * be the correct final value, so subtract one to
13572 * cancel out the increment that follows */
13573 len += foldlen - 1;
13575 /* If this node only contains non-folding code points so
13576 * far, see if this new one is also non-folding */
13578 if (folded != ender) {
13579 maybe_exact = FALSE;
13582 /* Here the fold is the original; we have to check
13583 * further to see if anything folds to it */
13584 if (_invlist_contains_cp(PL_utf8_foldable,
13587 maybe_exact = FALSE;
13594 if (next_is_quantifier) {
13596 /* Here, the next input is a quantifier, and to get here,
13597 * the current character is the only one in the node.
13598 * Also, here <len> doesn't include the final byte for this
13604 } /* End of loop through literal characters */
13606 /* Here we have either exhausted the input or ran out of room in
13607 * the node. (If we encountered a character that can't be in the
13608 * node, transfer is made directly to <loopdone>, and so we
13609 * wouldn't have fallen off the end of the loop.) In the latter
13610 * case, we artificially have to split the node into two, because
13611 * we just don't have enough space to hold everything. This
13612 * creates a problem if the final character participates in a
13613 * multi-character fold in the non-final position, as a match that
13614 * should have occurred won't, due to the way nodes are matched,
13615 * and our artificial boundary. So back off until we find a non-
13616 * problematic character -- one that isn't at the beginning or
13617 * middle of such a fold. (Either it doesn't participate in any
13618 * folds, or appears only in the final position of all the folds it
13619 * does participate in.) A better solution with far fewer false
13620 * positives, and that would fill the nodes more completely, would
13621 * be to actually have available all the multi-character folds to
13622 * test against, and to back-off only far enough to be sure that
13623 * this node isn't ending with a partial one. <upper_parse> is set
13624 * further below (if we need to reparse the node) to include just
13625 * up through that final non-problematic character that this code
13626 * identifies, so when it is set to less than the full node, we can
13627 * skip the rest of this */
13628 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
13630 const STRLEN full_len = len;
13632 assert(len >= MAX_NODE_STRING_SIZE);
13634 /* Here, <s> points to the final byte of the final character.
13635 * Look backwards through the string until find a non-
13636 * problematic character */
13640 /* This has no multi-char folds to non-UTF characters */
13641 if (ASCII_FOLD_RESTRICTED) {
13645 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
13649 if (! PL_NonL1NonFinalFold) {
13650 PL_NonL1NonFinalFold = _new_invlist_C_array(
13651 NonL1_Perl_Non_Final_Folds_invlist);
13654 /* Point to the first byte of the final character */
13655 s = (char *) utf8_hop((U8 *) s, -1);
13657 while (s >= s0) { /* Search backwards until find
13658 non-problematic char */
13659 if (UTF8_IS_INVARIANT(*s)) {
13661 /* There are no ascii characters that participate
13662 * in multi-char folds under /aa. In EBCDIC, the
13663 * non-ascii invariants are all control characters,
13664 * so don't ever participate in any folds. */
13665 if (ASCII_FOLD_RESTRICTED
13666 || ! IS_NON_FINAL_FOLD(*s))
13671 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
13672 if (! IS_NON_FINAL_FOLD(EIGHT_BIT_UTF8_TO_NATIVE(
13678 else if (! _invlist_contains_cp(
13679 PL_NonL1NonFinalFold,
13680 valid_utf8_to_uvchr((U8 *) s, NULL)))
13685 /* Here, the current character is problematic in that
13686 * it does occur in the non-final position of some
13687 * fold, so try the character before it, but have to
13688 * special case the very first byte in the string, so
13689 * we don't read outside the string */
13690 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
13691 } /* End of loop backwards through the string */
13693 /* If there were only problematic characters in the string,
13694 * <s> will point to before s0, in which case the length
13695 * should be 0, otherwise include the length of the
13696 * non-problematic character just found */
13697 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
13700 /* Here, have found the final character, if any, that is
13701 * non-problematic as far as ending the node without splitting
13702 * it across a potential multi-char fold. <len> contains the
13703 * number of bytes in the node up-to and including that
13704 * character, or is 0 if there is no such character, meaning
13705 * the whole node contains only problematic characters. In
13706 * this case, give up and just take the node as-is. We can't
13711 /* If the node ends in an 's' we make sure it stays EXACTF,
13712 * as if it turns into an EXACTFU, it could later get
13713 * joined with another 's' that would then wrongly match
13715 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
13717 maybe_exactfu = FALSE;
13721 /* Here, the node does contain some characters that aren't
13722 * problematic. If one such is the final character in the
13723 * node, we are done */
13724 if (len == full_len) {
13727 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
13729 /* If the final character is problematic, but the
13730 * penultimate is not, back-off that last character to
13731 * later start a new node with it */
13736 /* Here, the final non-problematic character is earlier
13737 * in the input than the penultimate character. What we do
13738 * is reparse from the beginning, going up only as far as
13739 * this final ok one, thus guaranteeing that the node ends
13740 * in an acceptable character. The reason we reparse is
13741 * that we know how far in the character is, but we don't
13742 * know how to correlate its position with the input parse.
13743 * An alternate implementation would be to build that
13744 * correlation as we go along during the original parse,
13745 * but that would entail extra work for every node, whereas
13746 * this code gets executed only when the string is too
13747 * large for the node, and the final two characters are
13748 * problematic, an infrequent occurrence. Yet another
13749 * possible strategy would be to save the tail of the
13750 * string, and the next time regatom is called, initialize
13751 * with that. The problem with this is that unless you
13752 * back off one more character, you won't be guaranteed
13753 * regatom will get called again, unless regbranch,
13754 * regpiece ... are also changed. If you do back off that
13755 * extra character, so that there is input guaranteed to
13756 * force calling regatom, you can't handle the case where
13757 * just the first character in the node is acceptable. I
13758 * (khw) decided to try this method which doesn't have that
13759 * pitfall; if performance issues are found, we can do a
13760 * combination of the current approach plus that one */
13766 } /* End of verifying node ends with an appropriate char */
13768 loopdone: /* Jumped to when encounters something that shouldn't be
13771 /* I (khw) don't know if you can get here with zero length, but the
13772 * old code handled this situation by creating a zero-length EXACT
13773 * node. Might as well be NOTHING instead */
13779 /* If 'maybe_exact' is still set here, means there are no
13780 * code points in the node that participate in folds;
13781 * similarly for 'maybe_exactfu' and code points that match
13782 * differently depending on UTF8ness of the target string
13783 * (for /u), or depending on locale for /l */
13789 else if (maybe_exactfu) {
13795 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
13796 FALSE /* Don't look to see if could
13797 be turned into an EXACT
13798 node, as we have already
13803 RExC_parse = p - 1;
13804 Set_Node_Cur_Length(ret, parse_start);
13807 /* len is STRLEN which is unsigned, need to copy to signed */
13810 vFAIL("Internal disaster");
13813 } /* End of label 'defchar:' */
13815 } /* End of giant switch on input character */
13817 /* Position parse to next real character */
13818 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13819 FALSE /* Don't force to /x */ );
13820 if (PASS2 && *RExC_parse == '{' && OP(ret) != SBOL && ! regcurly(RExC_parse)) {
13821 ckWARNregdep(RExC_parse + 1, "Unescaped left brace in regex is deprecated here, passed through");
13829 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
13831 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
13832 * sets up the bitmap and any flags, removing those code points from the
13833 * inversion list, setting it to NULL should it become completely empty */
13835 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
13836 assert(PL_regkind[OP(node)] == ANYOF);
13838 ANYOF_BITMAP_ZERO(node);
13839 if (*invlist_ptr) {
13841 /* This gets set if we actually need to modify things */
13842 bool change_invlist = FALSE;
13846 /* Start looking through *invlist_ptr */
13847 invlist_iterinit(*invlist_ptr);
13848 while (invlist_iternext(*invlist_ptr, &start, &end)) {
13852 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
13853 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
13856 /* Quit if are above what we should change */
13857 if (start >= NUM_ANYOF_CODE_POINTS) {
13861 change_invlist = TRUE;
13863 /* Set all the bits in the range, up to the max that we are doing */
13864 high = (end < NUM_ANYOF_CODE_POINTS - 1)
13866 : NUM_ANYOF_CODE_POINTS - 1;
13867 for (i = start; i <= (int) high; i++) {
13868 if (! ANYOF_BITMAP_TEST(node, i)) {
13869 ANYOF_BITMAP_SET(node, i);
13873 invlist_iterfinish(*invlist_ptr);
13875 /* Done with loop; remove any code points that are in the bitmap from
13876 * *invlist_ptr; similarly for code points above the bitmap if we have
13877 * a flag to match all of them anyways */
13878 if (change_invlist) {
13879 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
13881 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
13882 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
13885 /* If have completely emptied it, remove it completely */
13886 if (_invlist_len(*invlist_ptr) == 0) {
13887 SvREFCNT_dec_NN(*invlist_ptr);
13888 *invlist_ptr = NULL;
13893 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
13894 Character classes ([:foo:]) can also be negated ([:^foo:]).
13895 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
13896 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
13897 but trigger failures because they are currently unimplemented. */
13899 #define POSIXCC_DONE(c) ((c) == ':')
13900 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
13901 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
13902 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
13904 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
13905 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
13906 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
13908 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
13910 /* 'posix_warnings' and 'warn_text' are names of variables in the following
13912 #define ADD_POSIX_WARNING(p, text) STMT_START { \
13913 if (posix_warnings) { \
13914 if (! RExC_warn_text ) RExC_warn_text = (AV *) sv_2mortal((SV *) newAV()); \
13915 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
13919 REPORT_LOCATION_ARGS(p))); \
13924 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
13926 const char * const s, /* Where the putative posix class begins.
13927 Normally, this is one past the '['. This
13928 parameter exists so it can be somewhere
13929 besides RExC_parse. */
13930 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
13932 AV ** posix_warnings, /* Where to place any generated warnings, or
13934 const bool check_only /* Don't die if error */
13937 /* This parses what the caller thinks may be one of the three POSIX
13939 * 1) a character class, like [:blank:]
13940 * 2) a collating symbol, like [. .]
13941 * 3) an equivalence class, like [= =]
13942 * In the latter two cases, it croaks if it finds a syntactically legal
13943 * one, as these are not handled by Perl.
13945 * The main purpose is to look for a POSIX character class. It returns:
13946 * a) the class number
13947 * if it is a completely syntactically and semantically legal class.
13948 * 'updated_parse_ptr', if not NULL, is set to point to just after the
13949 * closing ']' of the class
13950 * b) OOB_NAMEDCLASS
13951 * if it appears that one of the three POSIX constructs was meant, but
13952 * its specification was somehow defective. 'updated_parse_ptr', if
13953 * not NULL, is set to point to the character just after the end
13954 * character of the class. See below for handling of warnings.
13955 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
13956 * if it doesn't appear that a POSIX construct was intended.
13957 * 'updated_parse_ptr' is not changed. No warnings nor errors are
13960 * In b) there may be errors or warnings generated. If 'check_only' is
13961 * TRUE, then any errors are discarded. Warnings are returned to the
13962 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
13963 * instead it is NULL, warnings are suppressed. This is done in all
13964 * passes. The reason for this is that the rest of the parsing is heavily
13965 * dependent on whether this routine found a valid posix class or not. If
13966 * it did, the closing ']' is absorbed as part of the class. If no class,
13967 * or an invalid one is found, any ']' will be considered the terminator of
13968 * the outer bracketed character class, leading to very different results.
13969 * In particular, a '(?[ ])' construct will likely have a syntax error if
13970 * the class is parsed other than intended, and this will happen in pass1,
13971 * before the warnings would normally be output. This mechanism allows the
13972 * caller to output those warnings in pass1 just before dieing, giving a
13973 * much better clue as to what is wrong.
13975 * The reason for this function, and its complexity is that a bracketed
13976 * character class can contain just about anything. But it's easy to
13977 * mistype the very specific posix class syntax but yielding a valid
13978 * regular bracketed class, so it silently gets compiled into something
13979 * quite unintended.
13981 * The solution adopted here maintains backward compatibility except that
13982 * it adds a warning if it looks like a posix class was intended but
13983 * improperly specified. The warning is not raised unless what is input
13984 * very closely resembles one of the 14 legal posix classes. To do this,
13985 * it uses fuzzy parsing. It calculates how many single-character edits it
13986 * would take to transform what was input into a legal posix class. Only
13987 * if that number is quite small does it think that the intention was a
13988 * posix class. Obviously these are heuristics, and there will be cases
13989 * where it errs on one side or another, and they can be tweaked as
13990 * experience informs.
13992 * The syntax for a legal posix class is:
13994 * qr/(?xa: \[ : \^? [:lower:]{4,6} : \] )/
13996 * What this routine considers syntactically to be an intended posix class
13997 * is this (the comments indicate some restrictions that the pattern
14000 * qr/(?x: \[? # The left bracket, possibly
14002 * \h* # possibly followed by blanks
14003 * (?: \^ \h* )? # possibly a misplaced caret
14004 * [:;]? # The opening class character,
14005 * # possibly omitted. A typo
14006 * # semi-colon can also be used.
14008 * \^? # possibly a correctly placed
14009 * # caret, but not if there was also
14010 * # a misplaced one
14012 * .{3,15} # The class name. If there are
14013 * # deviations from the legal syntax,
14014 * # its edit distance must be close
14015 * # to a real class name in order
14016 * # for it to be considered to be
14017 * # an intended posix class.
14019 * [:punct:]? # The closing class character,
14020 * # possibly omitted. If not a colon
14021 * # nor semi colon, the class name
14022 * # must be even closer to a valid
14025 * \]? # The right bracket, possibly
14029 * In the above, \h must be ASCII-only.
14031 * These are heuristics, and can be tweaked as field experience dictates.
14032 * There will be cases when someone didn't intend to specify a posix class
14033 * that this warns as being so. The goal is to minimize these, while
14034 * maximizing the catching of things intended to be a posix class that
14035 * aren't parsed as such.
14039 const char * const e = RExC_end;
14040 unsigned complement = 0; /* If to complement the class */
14041 bool found_problem = FALSE; /* Assume OK until proven otherwise */
14042 bool has_opening_bracket = FALSE;
14043 bool has_opening_colon = FALSE;
14044 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
14046 const char * possible_end = NULL; /* used for a 2nd parse pass */
14047 const char* name_start; /* ptr to class name first char */
14049 /* If the number of single-character typos the input name is away from a
14050 * legal name is no more than this number, it is considered to have meant
14051 * the legal name */
14052 int max_distance = 2;
14054 /* to store the name. The size determines the maximum length before we
14055 * decide that no posix class was intended. Should be at least
14056 * sizeof("alphanumeric") */
14059 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
14061 if (posix_warnings && RExC_warn_text)
14062 av_clear(RExC_warn_text);
14065 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14068 if (*(p - 1) != '[') {
14069 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
14070 found_problem = TRUE;
14073 has_opening_bracket = TRUE;
14076 /* They could be confused and think you can put spaces between the
14079 found_problem = TRUE;
14083 } while (p < e && isBLANK(*p));
14085 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14088 /* For [. .] and [= =]. These are quite different internally from [: :],
14089 * so they are handled separately. */
14090 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
14091 and 1 for at least one char in it
14094 const char open_char = *p;
14095 const char * temp_ptr = p + 1;
14097 /* These two constructs are not handled by perl, and if we find a
14098 * syntactically valid one, we croak. khw, who wrote this code, finds
14099 * this explanation of them very unclear:
14100 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
14101 * And searching the rest of the internet wasn't very helpful either.
14102 * It looks like just about any byte can be in these constructs,
14103 * depending on the locale. But unless the pattern is being compiled
14104 * under /l, which is very rare, Perl runs under the C or POSIX locale.
14105 * In that case, it looks like [= =] isn't allowed at all, and that
14106 * [. .] could be any single code point, but for longer strings the
14107 * constituent characters would have to be the ASCII alphabetics plus
14108 * the minus-hyphen. Any sensible locale definition would limit itself
14109 * to these. And any portable one definitely should. Trying to parse
14110 * the general case is a nightmare (see [perl #127604]). So, this code
14111 * looks only for interiors of these constructs that match:
14113 * Using \w relaxes the apparent rules a little, without adding much
14114 * danger of mistaking something else for one of these constructs.
14116 * [. .] in some implementations described on the internet is usable to
14117 * escape a character that otherwise is special in bracketed character
14118 * classes. For example [.].] means a literal right bracket instead of
14119 * the ending of the class
14121 * [= =] can legitimately contain a [. .] construct, but we don't
14122 * handle this case, as that [. .] construct will later get parsed
14123 * itself and croak then. And [= =] is checked for even when not under
14124 * /l, as Perl has long done so.
14126 * The code below relies on there being a trailing NUL, so it doesn't
14127 * have to keep checking if the parse ptr < e.
14129 if (temp_ptr[1] == open_char) {
14132 else while ( temp_ptr < e
14133 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
14138 if (*temp_ptr == open_char) {
14140 if (*temp_ptr == ']') {
14142 if (! found_problem && ! check_only) {
14143 RExC_parse = (char *) temp_ptr;
14144 vFAIL3("POSIX syntax [%c %c] is reserved for future "
14145 "extensions", open_char, open_char);
14148 /* Here, the syntax wasn't completely valid, or else the call
14149 * is to check-only */
14150 if (updated_parse_ptr) {
14151 *updated_parse_ptr = (char *) temp_ptr;
14154 return OOB_NAMEDCLASS;
14158 /* If we find something that started out to look like one of these
14159 * constructs, but isn't, we continue below so that it can be checked
14160 * for being a class name with a typo of '.' or '=' instead of a colon.
14164 /* Here, we think there is a possibility that a [: :] class was meant, and
14165 * we have the first real character. It could be they think the '^' comes
14168 found_problem = TRUE;
14169 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
14174 found_problem = TRUE;
14178 } while (p < e && isBLANK(*p));
14180 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14184 /* But the first character should be a colon, which they could have easily
14185 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
14186 * distinguish from a colon, so treat that as a colon). */
14189 has_opening_colon = TRUE;
14191 else if (*p == ';') {
14192 found_problem = TRUE;
14194 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14195 has_opening_colon = TRUE;
14198 found_problem = TRUE;
14199 ADD_POSIX_WARNING(p, "there must be a starting ':'");
14201 /* Consider an initial punctuation (not one of the recognized ones) to
14202 * be a left terminator */
14203 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
14208 /* They may think that you can put spaces between the components */
14210 found_problem = TRUE;
14214 } while (p < e && isBLANK(*p));
14216 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14221 /* We consider something like [^:^alnum:]] to not have been intended to
14222 * be a posix class, but XXX maybe we should */
14224 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14231 /* Again, they may think that you can put spaces between the components */
14233 found_problem = TRUE;
14237 } while (p < e && isBLANK(*p));
14239 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14244 /* XXX This ']' may be a typo, and something else was meant. But
14245 * treating it as such creates enough complications, that that
14246 * possibility isn't currently considered here. So we assume that the
14247 * ']' is what is intended, and if we've already found an initial '[',
14248 * this leaves this construct looking like [:] or [:^], which almost
14249 * certainly weren't intended to be posix classes */
14250 if (has_opening_bracket) {
14251 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14254 /* But this function can be called when we parse the colon for
14255 * something like qr/[alpha:]]/, so we back up to look for the
14260 found_problem = TRUE;
14261 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14263 else if (*p != ':') {
14265 /* XXX We are currently very restrictive here, so this code doesn't
14266 * consider the possibility that, say, /[alpha.]]/ was intended to
14267 * be a posix class. */
14268 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14271 /* Here we have something like 'foo:]'. There was no initial colon,
14272 * and we back up over 'foo. XXX Unlike the going forward case, we
14273 * don't handle typos of non-word chars in the middle */
14274 has_opening_colon = FALSE;
14277 while (p > RExC_start && isWORDCHAR(*p)) {
14282 /* Here, we have positioned ourselves to where we think the first
14283 * character in the potential class is */
14286 /* Now the interior really starts. There are certain key characters that
14287 * can end the interior, or these could just be typos. To catch both
14288 * cases, we may have to do two passes. In the first pass, we keep on
14289 * going unless we come to a sequence that matches
14290 * qr/ [[:punct:]] [[:blank:]]* \] /xa
14291 * This means it takes a sequence to end the pass, so two typos in a row if
14292 * that wasn't what was intended. If the class is perfectly formed, just
14293 * this one pass is needed. We also stop if there are too many characters
14294 * being accumulated, but this number is deliberately set higher than any
14295 * real class. It is set high enough so that someone who thinks that
14296 * 'alphanumeric' is a correct name would get warned that it wasn't.
14297 * While doing the pass, we keep track of where the key characters were in
14298 * it. If we don't find an end to the class, and one of the key characters
14299 * was found, we redo the pass, but stop when we get to that character.
14300 * Thus the key character was considered a typo in the first pass, but a
14301 * terminator in the second. If two key characters are found, we stop at
14302 * the second one in the first pass. Again this can miss two typos, but
14303 * catches a single one
14305 * In the first pass, 'possible_end' starts as NULL, and then gets set to
14306 * point to the first key character. For the second pass, it starts as -1.
14312 bool has_blank = FALSE;
14313 bool has_upper = FALSE;
14314 bool has_terminating_colon = FALSE;
14315 bool has_terminating_bracket = FALSE;
14316 bool has_semi_colon = FALSE;
14317 unsigned int name_len = 0;
14318 int punct_count = 0;
14322 /* Squeeze out blanks when looking up the class name below */
14323 if (isBLANK(*p) ) {
14325 found_problem = TRUE;
14330 /* The name will end with a punctuation */
14332 const char * peek = p + 1;
14334 /* Treat any non-']' punctuation followed by a ']' (possibly
14335 * with intervening blanks) as trying to terminate the class.
14336 * ']]' is very likely to mean a class was intended (but
14337 * missing the colon), but the warning message that gets
14338 * generated shows the error position better if we exit the
14339 * loop at the bottom (eventually), so skip it here. */
14341 if (peek < e && isBLANK(*peek)) {
14343 found_problem = TRUE;
14346 } while (peek < e && isBLANK(*peek));
14349 if (peek < e && *peek == ']') {
14350 has_terminating_bracket = TRUE;
14352 has_terminating_colon = TRUE;
14354 else if (*p == ';') {
14355 has_semi_colon = TRUE;
14356 has_terminating_colon = TRUE;
14359 found_problem = TRUE;
14366 /* Here we have punctuation we thought didn't end the class.
14367 * Keep track of the position of the key characters that are
14368 * more likely to have been class-enders */
14369 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
14371 /* Allow just one such possible class-ender not actually
14372 * ending the class. */
14373 if (possible_end) {
14379 /* If we have too many punctuation characters, no use in
14381 if (++punct_count > max_distance) {
14385 /* Treat the punctuation as a typo. */
14386 input_text[name_len++] = *p;
14389 else if (isUPPER(*p)) { /* Use lowercase for lookup */
14390 input_text[name_len++] = toLOWER(*p);
14392 found_problem = TRUE;
14394 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
14395 input_text[name_len++] = *p;
14399 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
14403 /* The declaration of 'input_text' is how long we allow a potential
14404 * class name to be, before saying they didn't mean a class name at
14406 if (name_len >= C_ARRAY_LENGTH(input_text)) {
14411 /* We get to here when the possible class name hasn't been properly
14412 * terminated before:
14413 * 1) we ran off the end of the pattern; or
14414 * 2) found two characters, each of which might have been intended to
14415 * be the name's terminator
14416 * 3) found so many punctuation characters in the purported name,
14417 * that the edit distance to a valid one is exceeded
14418 * 4) we decided it was more characters than anyone could have
14419 * intended to be one. */
14421 found_problem = TRUE;
14423 /* In the final two cases, we know that looking up what we've
14424 * accumulated won't lead to a match, even a fuzzy one. */
14425 if ( name_len >= C_ARRAY_LENGTH(input_text)
14426 || punct_count > max_distance)
14428 /* If there was an intermediate key character that could have been
14429 * an intended end, redo the parse, but stop there */
14430 if (possible_end && possible_end != (char *) -1) {
14431 possible_end = (char *) -1; /* Special signal value to say
14432 we've done a first pass */
14437 /* Otherwise, it can't have meant to have been a class */
14438 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14441 /* If we ran off the end, and the final character was a punctuation
14442 * one, back up one, to look at that final one just below. Later, we
14443 * will restore the parse pointer if appropriate */
14444 if (name_len && p == e && isPUNCT(*(p-1))) {
14449 if (p < e && isPUNCT(*p)) {
14451 has_terminating_bracket = TRUE;
14453 /* If this is a 2nd ']', and the first one is just below this
14454 * one, consider that to be the real terminator. This gives a
14455 * uniform and better positioning for the warning message */
14457 && possible_end != (char *) -1
14458 && *possible_end == ']'
14459 && name_len && input_text[name_len - 1] == ']')
14464 /* And this is actually equivalent to having done the 2nd
14465 * pass now, so set it to not try again */
14466 possible_end = (char *) -1;
14471 has_terminating_colon = TRUE;
14473 else if (*p == ';') {
14474 has_semi_colon = TRUE;
14475 has_terminating_colon = TRUE;
14483 /* Here, we have a class name to look up. We can short circuit the
14484 * stuff below for short names that can't possibly be meant to be a
14485 * class name. (We can do this on the first pass, as any second pass
14486 * will yield an even shorter name) */
14487 if (name_len < 3) {
14488 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14491 /* Find which class it is. Initially switch on the length of the name.
14493 switch (name_len) {
14495 if (memEQ(name_start, "word", 4)) {
14496 /* this is not POSIX, this is the Perl \w */
14497 class_number = ANYOF_WORDCHAR;
14501 /* Names all of length 5: alnum alpha ascii blank cntrl digit
14502 * graph lower print punct space upper
14503 * Offset 4 gives the best switch position. */
14504 switch (name_start[4]) {
14506 if (memEQ(name_start, "alph", 4)) /* alpha */
14507 class_number = ANYOF_ALPHA;
14510 if (memEQ(name_start, "spac", 4)) /* space */
14511 class_number = ANYOF_SPACE;
14514 if (memEQ(name_start, "grap", 4)) /* graph */
14515 class_number = ANYOF_GRAPH;
14518 if (memEQ(name_start, "asci", 4)) /* ascii */
14519 class_number = ANYOF_ASCII;
14522 if (memEQ(name_start, "blan", 4)) /* blank */
14523 class_number = ANYOF_BLANK;
14526 if (memEQ(name_start, "cntr", 4)) /* cntrl */
14527 class_number = ANYOF_CNTRL;
14530 if (memEQ(name_start, "alnu", 4)) /* alnum */
14531 class_number = ANYOF_ALPHANUMERIC;
14534 if (memEQ(name_start, "lowe", 4)) /* lower */
14535 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
14536 else if (memEQ(name_start, "uppe", 4)) /* upper */
14537 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
14540 if (memEQ(name_start, "digi", 4)) /* digit */
14541 class_number = ANYOF_DIGIT;
14542 else if (memEQ(name_start, "prin", 4)) /* print */
14543 class_number = ANYOF_PRINT;
14544 else if (memEQ(name_start, "punc", 4)) /* punct */
14545 class_number = ANYOF_PUNCT;
14550 if (memEQ(name_start, "xdigit", 6))
14551 class_number = ANYOF_XDIGIT;
14555 /* If the name exactly matches a posix class name the class number will
14556 * here be set to it, and the input almost certainly was meant to be a
14557 * posix class, so we can skip further checking. If instead the syntax
14558 * is exactly correct, but the name isn't one of the legal ones, we
14559 * will return that as an error below. But if neither of these apply,
14560 * it could be that no posix class was intended at all, or that one
14561 * was, but there was a typo. We tease these apart by doing fuzzy
14562 * matching on the name */
14563 if (class_number == OOB_NAMEDCLASS && found_problem) {
14564 const UV posix_names[][6] = {
14565 { 'a', 'l', 'n', 'u', 'm' },
14566 { 'a', 'l', 'p', 'h', 'a' },
14567 { 'a', 's', 'c', 'i', 'i' },
14568 { 'b', 'l', 'a', 'n', 'k' },
14569 { 'c', 'n', 't', 'r', 'l' },
14570 { 'd', 'i', 'g', 'i', 't' },
14571 { 'g', 'r', 'a', 'p', 'h' },
14572 { 'l', 'o', 'w', 'e', 'r' },
14573 { 'p', 'r', 'i', 'n', 't' },
14574 { 'p', 'u', 'n', 'c', 't' },
14575 { 's', 'p', 'a', 'c', 'e' },
14576 { 'u', 'p', 'p', 'e', 'r' },
14577 { 'w', 'o', 'r', 'd' },
14578 { 'x', 'd', 'i', 'g', 'i', 't' }
14580 /* The names of the above all have added NULs to make them the same
14581 * size, so we need to also have the real lengths */
14582 const UV posix_name_lengths[] = {
14583 sizeof("alnum") - 1,
14584 sizeof("alpha") - 1,
14585 sizeof("ascii") - 1,
14586 sizeof("blank") - 1,
14587 sizeof("cntrl") - 1,
14588 sizeof("digit") - 1,
14589 sizeof("graph") - 1,
14590 sizeof("lower") - 1,
14591 sizeof("print") - 1,
14592 sizeof("punct") - 1,
14593 sizeof("space") - 1,
14594 sizeof("upper") - 1,
14595 sizeof("word") - 1,
14596 sizeof("xdigit")- 1
14599 int temp_max = max_distance; /* Use a temporary, so if we
14600 reparse, we haven't changed the
14603 /* Use a smaller max edit distance if we are missing one of the
14605 if ( has_opening_bracket + has_opening_colon < 2
14606 || has_terminating_bracket + has_terminating_colon < 2)
14611 /* See if the input name is close to a legal one */
14612 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
14614 /* Short circuit call if the lengths are too far apart to be
14616 if (abs( (int) (name_len - posix_name_lengths[i]))
14622 if (edit_distance(input_text,
14625 posix_name_lengths[i],
14629 { /* If it is close, it probably was intended to be a class */
14630 goto probably_meant_to_be;
14634 /* Here the input name is not close enough to a valid class name
14635 * for us to consider it to be intended to be a posix class. If
14636 * we haven't already done so, and the parse found a character that
14637 * could have been terminators for the name, but which we absorbed
14638 * as typos during the first pass, repeat the parse, signalling it
14639 * to stop at that character */
14640 if (possible_end && possible_end != (char *) -1) {
14641 possible_end = (char *) -1;
14646 /* Here neither pass found a close-enough class name */
14647 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14650 probably_meant_to_be:
14652 /* Here we think that a posix specification was intended. Update any
14654 if (updated_parse_ptr) {
14655 *updated_parse_ptr = (char *) p;
14658 /* If a posix class name was intended but incorrectly specified, we
14659 * output or return the warnings */
14660 if (found_problem) {
14662 /* We set flags for these issues in the parse loop above instead of
14663 * adding them to the list of warnings, because we can parse it
14664 * twice, and we only want one warning instance */
14666 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
14669 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14671 if (has_semi_colon) {
14672 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14674 else if (! has_terminating_colon) {
14675 ADD_POSIX_WARNING(p, "there is no terminating ':'");
14677 if (! has_terminating_bracket) {
14678 ADD_POSIX_WARNING(p, "there is no terminating ']'");
14681 if (posix_warnings && RExC_warn_text && av_top_index(RExC_warn_text) > -1) {
14682 *posix_warnings = RExC_warn_text;
14685 else if (class_number != OOB_NAMEDCLASS) {
14686 /* If it is a known class, return the class. The class number
14687 * #defines are structured so each complement is +1 to the normal
14689 return class_number + complement;
14691 else if (! check_only) {
14693 /* Here, it is an unrecognized class. This is an error (unless the
14694 * call is to check only, which we've already handled above) */
14695 const char * const complement_string = (complement)
14698 RExC_parse = (char *) p;
14699 vFAIL3utf8f("POSIX class [:%s%"UTF8f":] unknown",
14701 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
14705 return OOB_NAMEDCLASS;
14707 #undef ADD_POSIX_WARNING
14709 STATIC unsigned int
14710 S_regex_set_precedence(const U8 my_operator) {
14712 /* Returns the precedence in the (?[...]) construct of the input operator,
14713 * specified by its character representation. The precedence follows
14714 * general Perl rules, but it extends this so that ')' and ']' have (low)
14715 * precedence even though they aren't really operators */
14717 switch (my_operator) {
14733 NOT_REACHED; /* NOTREACHED */
14734 return 0; /* Silence compiler warning */
14738 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
14739 I32 *flagp, U32 depth,
14740 char * const oregcomp_parse)
14742 /* Handle the (?[...]) construct to do set operations */
14744 U8 curchar; /* Current character being parsed */
14745 UV start, end; /* End points of code point ranges */
14746 SV* final = NULL; /* The end result inversion list */
14747 SV* result_string; /* 'final' stringified */
14748 AV* stack; /* stack of operators and operands not yet
14750 AV* fence_stack = NULL; /* A stack containing the positions in
14751 'stack' of where the undealt-with left
14752 parens would be if they were actually
14754 /* The 'VOL' (expanding to 'volatile') is a workaround for an optimiser bug
14755 * in Solaris Studio 12.3. See RT #127455 */
14756 VOL IV fence = 0; /* Position of where most recent undealt-
14757 with left paren in stack is; -1 if none.
14759 STRLEN len; /* Temporary */
14760 regnode* node; /* Temporary, and final regnode returned by
14762 const bool save_fold = FOLD; /* Temporary */
14763 char *save_end, *save_parse; /* Temporaries */
14764 const bool in_locale = LOC; /* we turn off /l during processing */
14765 AV* posix_warnings = NULL;
14767 GET_RE_DEBUG_FLAGS_DECL;
14769 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
14772 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
14775 REQUIRE_UNI_RULES(flagp, NULL); /* The use of this operator implies /u.
14776 This is required so that the compile
14777 time values are valid in all runtime
14780 /* This will return only an ANYOF regnode, or (unlikely) something smaller
14781 * (such as EXACT). Thus we can skip most everything if just sizing. We
14782 * call regclass to handle '[]' so as to not have to reinvent its parsing
14783 * rules here (throwing away the size it computes each time). And, we exit
14784 * upon an unescaped ']' that isn't one ending a regclass. To do both
14785 * these things, we need to realize that something preceded by a backslash
14786 * is escaped, so we have to keep track of backslashes */
14788 UV depth = 0; /* how many nested (?[...]) constructs */
14790 while (RExC_parse < RExC_end) {
14791 SV* current = NULL;
14793 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14794 TRUE /* Force /x */ );
14796 switch (*RExC_parse) {
14798 if (RExC_parse[1] == '[') depth++, RExC_parse++;
14803 /* Skip past this, so the next character gets skipped, after
14806 if (*RExC_parse == 'c') {
14807 /* Skip the \cX notation for control characters */
14808 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14814 /* See if this is a [:posix:] class. */
14815 bool is_posix_class = (OOB_NAMEDCLASS
14816 < handle_possible_posix(pRExC_state,
14820 TRUE /* checking only */));
14821 /* If it is a posix class, leave the parse pointer at the
14822 * '[' to fool regclass() into thinking it is part of a
14823 * '[[:posix:]]'. */
14824 if (! is_posix_class) {
14828 /* regclass() can only return RESTART_PASS1 and NEED_UTF8
14829 * if multi-char folds are allowed. */
14830 if (!regclass(pRExC_state, flagp,depth+1,
14831 is_posix_class, /* parse the whole char
14832 class only if not a
14834 FALSE, /* don't allow multi-char folds */
14835 TRUE, /* silence non-portable warnings. */
14837 FALSE, /* Require return to be an ANYOF */
14841 FAIL2("panic: regclass returned NULL to handle_sets, "
14842 "flags=%#"UVxf"", (UV) *flagp);
14844 /* function call leaves parse pointing to the ']', except
14845 * if we faked it */
14846 if (is_posix_class) {
14850 SvREFCNT_dec(current); /* In case it returned something */
14855 if (depth--) break;
14857 if (*RExC_parse == ')') {
14858 node = reganode(pRExC_state, ANYOF, 0);
14859 RExC_size += ANYOF_SKIP;
14860 nextchar(pRExC_state);
14861 Set_Node_Length(node,
14862 RExC_parse - oregcomp_parse + 1); /* MJD */
14864 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
14872 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14876 /* We output the messages even if warnings are off, because we'll fail
14877 * the very next thing, and these give a likely diagnosis for that */
14878 if (posix_warnings && av_tindex_nomg(posix_warnings) >= 0) {
14879 output_or_return_posix_warnings(pRExC_state, posix_warnings, NULL);
14882 FAIL("Syntax error in (?[...])");
14885 /* Pass 2 only after this. */
14886 Perl_ck_warner_d(aTHX_
14887 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
14888 "The regex_sets feature is experimental" REPORT_LOCATION,
14889 REPORT_LOCATION_ARGS(RExC_parse));
14891 /* Everything in this construct is a metacharacter. Operands begin with
14892 * either a '\' (for an escape sequence), or a '[' for a bracketed
14893 * character class. Any other character should be an operator, or
14894 * parenthesis for grouping. Both types of operands are handled by calling
14895 * regclass() to parse them. It is called with a parameter to indicate to
14896 * return the computed inversion list. The parsing here is implemented via
14897 * a stack. Each entry on the stack is a single character representing one
14898 * of the operators; or else a pointer to an operand inversion list. */
14900 #define IS_OPERATOR(a) SvIOK(a)
14901 #define IS_OPERAND(a) (! IS_OPERATOR(a))
14903 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
14904 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
14905 * with pronouncing it called it Reverse Polish instead, but now that YOU
14906 * know how to pronounce it you can use the correct term, thus giving due
14907 * credit to the person who invented it, and impressing your geek friends.
14908 * Wikipedia says that the pronounciation of "Ł" has been changing so that
14909 * it is now more like an English initial W (as in wonk) than an L.)
14911 * This means that, for example, 'a | b & c' is stored on the stack as
14919 * where the numbers in brackets give the stack [array] element number.
14920 * In this implementation, parentheses are not stored on the stack.
14921 * Instead a '(' creates a "fence" so that the part of the stack below the
14922 * fence is invisible except to the corresponding ')' (this allows us to
14923 * replace testing for parens, by using instead subtraction of the fence
14924 * position). As new operands are processed they are pushed onto the stack
14925 * (except as noted in the next paragraph). New operators of higher
14926 * precedence than the current final one are inserted on the stack before
14927 * the lhs operand (so that when the rhs is pushed next, everything will be
14928 * in the correct positions shown above. When an operator of equal or
14929 * lower precedence is encountered in parsing, all the stacked operations
14930 * of equal or higher precedence are evaluated, leaving the result as the
14931 * top entry on the stack. This makes higher precedence operations
14932 * evaluate before lower precedence ones, and causes operations of equal
14933 * precedence to left associate.
14935 * The only unary operator '!' is immediately pushed onto the stack when
14936 * encountered. When an operand is encountered, if the top of the stack is
14937 * a '!", the complement is immediately performed, and the '!' popped. The
14938 * resulting value is treated as a new operand, and the logic in the
14939 * previous paragraph is executed. Thus in the expression
14941 * the stack looks like
14947 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
14954 * A ')' is treated as an operator with lower precedence than all the
14955 * aforementioned ones, which causes all operations on the stack above the
14956 * corresponding '(' to be evaluated down to a single resultant operand.
14957 * Then the fence for the '(' is removed, and the operand goes through the
14958 * algorithm above, without the fence.
14960 * A separate stack is kept of the fence positions, so that the position of
14961 * the latest so-far unbalanced '(' is at the top of it.
14963 * The ']' ending the construct is treated as the lowest operator of all,
14964 * so that everything gets evaluated down to a single operand, which is the
14967 sv_2mortal((SV *)(stack = newAV()));
14968 sv_2mortal((SV *)(fence_stack = newAV()));
14970 while (RExC_parse < RExC_end) {
14971 I32 top_index; /* Index of top-most element in 'stack' */
14972 SV** top_ptr; /* Pointer to top 'stack' element */
14973 SV* current = NULL; /* To contain the current inversion list
14975 SV* only_to_avoid_leaks;
14977 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14978 TRUE /* Force /x */ );
14979 if (RExC_parse >= RExC_end) {
14980 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
14983 curchar = UCHARAT(RExC_parse);
14987 top_index = av_tindex_nomg(stack);
14990 SV** stacked_ptr; /* Ptr to something already on 'stack' */
14991 char stacked_operator; /* The topmost operator on the 'stack'. */
14992 SV* lhs; /* Operand to the left of the operator */
14993 SV* rhs; /* Operand to the right of the operator */
14994 SV* fence_ptr; /* Pointer to top element of the fence
14999 if ( RExC_parse < RExC_end - 1
15000 && (UCHARAT(RExC_parse + 1) == '?'))
15002 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
15003 * This happens when we have some thing like
15005 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
15007 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
15009 * Here we would be handling the interpolated
15010 * '$thai_or_lao'. We handle this by a recursive call to
15011 * ourselves which returns the inversion list the
15012 * interpolated expression evaluates to. We use the flags
15013 * from the interpolated pattern. */
15014 U32 save_flags = RExC_flags;
15015 const char * save_parse;
15017 RExC_parse += 2; /* Skip past the '(?' */
15018 save_parse = RExC_parse;
15020 /* Parse any flags for the '(?' */
15021 parse_lparen_question_flags(pRExC_state);
15023 if (RExC_parse == save_parse /* Makes sure there was at
15024 least one flag (or else
15025 this embedding wasn't
15027 || RExC_parse >= RExC_end - 4
15028 || UCHARAT(RExC_parse) != ':'
15029 || UCHARAT(++RExC_parse) != '('
15030 || UCHARAT(++RExC_parse) != '?'
15031 || UCHARAT(++RExC_parse) != '[')
15034 /* In combination with the above, this moves the
15035 * pointer to the point just after the first erroneous
15036 * character (or if there are no flags, to where they
15037 * should have been) */
15038 if (RExC_parse >= RExC_end - 4) {
15039 RExC_parse = RExC_end;
15041 else if (RExC_parse != save_parse) {
15042 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15044 vFAIL("Expecting '(?flags:(?[...'");
15047 /* Recurse, with the meat of the embedded expression */
15049 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
15050 depth+1, oregcomp_parse);
15052 /* Here, 'current' contains the embedded expression's
15053 * inversion list, and RExC_parse points to the trailing
15054 * ']'; the next character should be the ')' */
15056 assert(UCHARAT(RExC_parse) == ')');
15058 /* Then the ')' matching the original '(' handled by this
15059 * case: statement */
15061 assert(UCHARAT(RExC_parse) == ')');
15064 RExC_flags = save_flags;
15065 goto handle_operand;
15068 /* A regular '('. Look behind for illegal syntax */
15069 if (top_index - fence >= 0) {
15070 /* If the top entry on the stack is an operator, it had
15071 * better be a '!', otherwise the entry below the top
15072 * operand should be an operator */
15073 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
15074 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
15075 || ( IS_OPERAND(*top_ptr)
15076 && ( top_index - fence < 1
15077 || ! (stacked_ptr = av_fetch(stack,
15080 || ! IS_OPERATOR(*stacked_ptr))))
15083 vFAIL("Unexpected '(' with no preceding operator");
15087 /* Stack the position of this undealt-with left paren */
15088 av_push(fence_stack, newSViv(fence));
15089 fence = top_index + 1;
15093 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15094 * multi-char folds are allowed. */
15095 if (!regclass(pRExC_state, flagp,depth+1,
15096 TRUE, /* means parse just the next thing */
15097 FALSE, /* don't allow multi-char folds */
15098 FALSE, /* don't silence non-portable warnings. */
15100 FALSE, /* Require return to be an ANYOF */
15104 FAIL2("panic: regclass returned NULL to handle_sets, "
15105 "flags=%#"UVxf"", (UV) *flagp);
15108 /* regclass() will return with parsing just the \ sequence,
15109 * leaving the parse pointer at the next thing to parse */
15111 goto handle_operand;
15113 case '[': /* Is a bracketed character class */
15115 /* See if this is a [:posix:] class. */
15116 bool is_posix_class = (OOB_NAMEDCLASS
15117 < handle_possible_posix(pRExC_state,
15121 TRUE /* checking only */));
15122 /* If it is a posix class, leave the parse pointer at the '['
15123 * to fool regclass() into thinking it is part of a
15124 * '[[:posix:]]'. */
15125 if (! is_posix_class) {
15129 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15130 * multi-char folds are allowed. */
15131 if (!regclass(pRExC_state, flagp,depth+1,
15132 is_posix_class, /* parse the whole char
15133 class only if not a
15135 FALSE, /* don't allow multi-char folds */
15136 TRUE, /* silence non-portable warnings. */
15138 FALSE, /* Require return to be an ANYOF */
15143 FAIL2("panic: regclass returned NULL to handle_sets, "
15144 "flags=%#"UVxf"", (UV) *flagp);
15147 /* function call leaves parse pointing to the ']', except if we
15149 if (is_posix_class) {
15153 goto handle_operand;
15157 if (top_index >= 1) {
15158 goto join_operators;
15161 /* Only a single operand on the stack: are done */
15165 if (av_tindex_nomg(fence_stack) < 0) {
15167 vFAIL("Unexpected ')'");
15170 /* If nothing after the fence, is missing an operand */
15171 if (top_index - fence < 0) {
15175 /* If at least two things on the stack, treat this as an
15177 if (top_index - fence >= 1) {
15178 goto join_operators;
15181 /* Here only a single thing on the fenced stack, and there is a
15182 * fence. Get rid of it */
15183 fence_ptr = av_pop(fence_stack);
15185 fence = SvIV(fence_ptr) - 1;
15186 SvREFCNT_dec_NN(fence_ptr);
15193 /* Having gotten rid of the fence, we pop the operand at the
15194 * stack top and process it as a newly encountered operand */
15195 current = av_pop(stack);
15196 if (IS_OPERAND(current)) {
15197 goto handle_operand;
15209 /* These binary operators should have a left operand already
15211 if ( top_index - fence < 0
15212 || top_index - fence == 1
15213 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
15214 || ! IS_OPERAND(*top_ptr))
15216 goto unexpected_binary;
15219 /* If only the one operand is on the part of the stack visible
15220 * to us, we just place this operator in the proper position */
15221 if (top_index - fence < 2) {
15223 /* Place the operator before the operand */
15225 SV* lhs = av_pop(stack);
15226 av_push(stack, newSVuv(curchar));
15227 av_push(stack, lhs);
15231 /* But if there is something else on the stack, we need to
15232 * process it before this new operator if and only if the
15233 * stacked operation has equal or higher precedence than the
15238 /* The operator on the stack is supposed to be below both its
15240 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
15241 || IS_OPERAND(*stacked_ptr))
15243 /* But if not, it's legal and indicates we are completely
15244 * done if and only if we're currently processing a ']',
15245 * which should be the final thing in the expression */
15246 if (curchar == ']') {
15252 vFAIL2("Unexpected binary operator '%c' with no "
15253 "preceding operand", curchar);
15255 stacked_operator = (char) SvUV(*stacked_ptr);
15257 if (regex_set_precedence(curchar)
15258 > regex_set_precedence(stacked_operator))
15260 /* Here, the new operator has higher precedence than the
15261 * stacked one. This means we need to add the new one to
15262 * the stack to await its rhs operand (and maybe more
15263 * stuff). We put it before the lhs operand, leaving
15264 * untouched the stacked operator and everything below it
15266 lhs = av_pop(stack);
15267 assert(IS_OPERAND(lhs));
15269 av_push(stack, newSVuv(curchar));
15270 av_push(stack, lhs);
15274 /* Here, the new operator has equal or lower precedence than
15275 * what's already there. This means the operation already
15276 * there should be performed now, before the new one. */
15278 rhs = av_pop(stack);
15279 if (! IS_OPERAND(rhs)) {
15281 /* This can happen when a ! is not followed by an operand,
15282 * like in /(?[\t &!])/ */
15286 lhs = av_pop(stack);
15288 if (! IS_OPERAND(lhs)) {
15290 /* This can happen when there is an empty (), like in
15291 * /(?[[0]+()+])/ */
15295 switch (stacked_operator) {
15297 _invlist_intersection(lhs, rhs, &rhs);
15302 _invlist_union(lhs, rhs, &rhs);
15306 _invlist_subtract(lhs, rhs, &rhs);
15309 case '^': /* The union minus the intersection */
15315 _invlist_union(lhs, rhs, &u);
15316 _invlist_intersection(lhs, rhs, &i);
15317 /* _invlist_subtract will overwrite rhs
15318 without freeing what it already contains */
15320 _invlist_subtract(u, i, &rhs);
15321 SvREFCNT_dec_NN(i);
15322 SvREFCNT_dec_NN(u);
15323 SvREFCNT_dec_NN(element);
15329 /* Here, the higher precedence operation has been done, and the
15330 * result is in 'rhs'. We overwrite the stacked operator with
15331 * the result. Then we redo this code to either push the new
15332 * operator onto the stack or perform any higher precedence
15333 * stacked operation */
15334 only_to_avoid_leaks = av_pop(stack);
15335 SvREFCNT_dec(only_to_avoid_leaks);
15336 av_push(stack, rhs);
15339 case '!': /* Highest priority, right associative */
15341 /* If what's already at the top of the stack is another '!",
15342 * they just cancel each other out */
15343 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
15344 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
15346 only_to_avoid_leaks = av_pop(stack);
15347 SvREFCNT_dec(only_to_avoid_leaks);
15349 else { /* Otherwise, since it's right associative, just push
15351 av_push(stack, newSVuv(curchar));
15356 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15357 vFAIL("Unexpected character");
15361 /* Here 'current' is the operand. If something is already on the
15362 * stack, we have to check if it is a !. But first, the code above
15363 * may have altered the stack in the time since we earlier set
15366 top_index = av_tindex_nomg(stack);
15367 if (top_index - fence >= 0) {
15368 /* If the top entry on the stack is an operator, it had better
15369 * be a '!', otherwise the entry below the top operand should
15370 * be an operator */
15371 top_ptr = av_fetch(stack, top_index, FALSE);
15373 if (IS_OPERATOR(*top_ptr)) {
15375 /* The only permissible operator at the top of the stack is
15376 * '!', which is applied immediately to this operand. */
15377 curchar = (char) SvUV(*top_ptr);
15378 if (curchar != '!') {
15379 SvREFCNT_dec(current);
15380 vFAIL2("Unexpected binary operator '%c' with no "
15381 "preceding operand", curchar);
15384 _invlist_invert(current);
15386 only_to_avoid_leaks = av_pop(stack);
15387 SvREFCNT_dec(only_to_avoid_leaks);
15389 /* And we redo with the inverted operand. This allows
15390 * handling multiple ! in a row */
15391 goto handle_operand;
15393 /* Single operand is ok only for the non-binary ')'
15395 else if ((top_index - fence == 0 && curchar != ')')
15396 || (top_index - fence > 0
15397 && (! (stacked_ptr = av_fetch(stack,
15400 || IS_OPERAND(*stacked_ptr))))
15402 SvREFCNT_dec(current);
15403 vFAIL("Operand with no preceding operator");
15407 /* Here there was nothing on the stack or the top element was
15408 * another operand. Just add this new one */
15409 av_push(stack, current);
15411 } /* End of switch on next parse token */
15413 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15414 } /* End of loop parsing through the construct */
15417 if (av_tindex_nomg(fence_stack) >= 0) {
15418 vFAIL("Unmatched (");
15421 if (av_tindex_nomg(stack) < 0 /* Was empty */
15422 || ((final = av_pop(stack)) == NULL)
15423 || ! IS_OPERAND(final)
15424 || SvTYPE(final) != SVt_INVLIST
15425 || av_tindex_nomg(stack) >= 0) /* More left on stack */
15428 SvREFCNT_dec(final);
15429 vFAIL("Incomplete expression within '(?[ ])'");
15432 /* Here, 'final' is the resultant inversion list from evaluating the
15433 * expression. Return it if so requested */
15434 if (return_invlist) {
15435 *return_invlist = final;
15439 /* Otherwise generate a resultant node, based on 'final'. regclass() is
15440 * expecting a string of ranges and individual code points */
15441 invlist_iterinit(final);
15442 result_string = newSVpvs("");
15443 while (invlist_iternext(final, &start, &end)) {
15444 if (start == end) {
15445 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
15448 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
15453 /* About to generate an ANYOF (or similar) node from the inversion list we
15454 * have calculated */
15455 save_parse = RExC_parse;
15456 RExC_parse = SvPV(result_string, len);
15457 save_end = RExC_end;
15458 RExC_end = RExC_parse + len;
15460 /* We turn off folding around the call, as the class we have constructed
15461 * already has all folding taken into consideration, and we don't want
15462 * regclass() to add to that */
15463 RExC_flags &= ~RXf_PMf_FOLD;
15464 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if multi-char
15465 * folds are allowed. */
15466 node = regclass(pRExC_state, flagp,depth+1,
15467 FALSE, /* means parse the whole char class */
15468 FALSE, /* don't allow multi-char folds */
15469 TRUE, /* silence non-portable warnings. The above may very
15470 well have generated non-portable code points, but
15471 they're valid on this machine */
15472 FALSE, /* similarly, no need for strict */
15473 FALSE, /* Require return to be an ANYOF */
15478 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
15481 /* Fix up the node type if we are in locale. (We have pretended we are
15482 * under /u for the purposes of regclass(), as this construct will only
15483 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
15484 * as to cause any warnings about bad locales to be output in regexec.c),
15485 * and add the flag that indicates to check if not in a UTF-8 locale. The
15486 * reason we above forbid optimization into something other than an ANYOF
15487 * node is simply to minimize the number of code changes in regexec.c.
15488 * Otherwise we would have to create new EXACTish node types and deal with
15489 * them. This decision could be revisited should this construct become
15492 * (One might think we could look at the resulting ANYOF node and suppress
15493 * the flag if everything is above 255, as those would be UTF-8 only,
15494 * but this isn't true, as the components that led to that result could
15495 * have been locale-affected, and just happen to cancel each other out
15496 * under UTF-8 locales.) */
15498 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
15500 assert(OP(node) == ANYOF);
15504 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
15508 RExC_flags |= RXf_PMf_FOLD;
15511 RExC_parse = save_parse + 1;
15512 RExC_end = save_end;
15513 SvREFCNT_dec_NN(final);
15514 SvREFCNT_dec_NN(result_string);
15516 nextchar(pRExC_state);
15517 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
15524 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
15526 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
15527 * innocent-looking character class, like /[ks]/i won't have to go out to
15528 * disk to find the possible matches.
15530 * This should be called only for a Latin1-range code points, cp, which is
15531 * known to be involved in a simple fold with other code points above
15532 * Latin1. It would give false results if /aa has been specified.
15533 * Multi-char folds are outside the scope of this, and must be handled
15536 * XXX It would be better to generate these via regen, in case a new
15537 * version of the Unicode standard adds new mappings, though that is not
15538 * really likely, and may be caught by the default: case of the switch
15541 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
15543 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
15549 add_cp_to_invlist(*invlist, KELVIN_SIGN);
15553 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
15556 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
15557 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
15559 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
15560 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
15561 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
15563 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
15564 *invlist = add_cp_to_invlist(*invlist,
15565 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
15568 #ifdef LATIN_CAPITAL_LETTER_SHARP_S /* not defined in early Unicode releases */
15570 case LATIN_SMALL_LETTER_SHARP_S:
15571 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
15576 #if UNICODE_MAJOR_VERSION < 3 \
15577 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0)
15579 /* In 3.0 and earlier, U+0130 folded simply to 'i'; and in 3.0.1 so did
15584 add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
15585 # if UNICODE_DOT_DOT_VERSION == 1
15586 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_DOTLESS_I);
15592 /* Use deprecated warning to increase the chances of this being
15595 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
15602 S_output_or_return_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings, AV** return_posix_warnings)
15604 /* If the final parameter is NULL, output the elements of the array given
15605 * by '*posix_warnings' as REGEXP warnings. Otherwise, the elements are
15606 * pushed onto it, (creating if necessary) */
15609 const bool first_is_fatal = ! return_posix_warnings
15610 && ckDEAD(packWARN(WARN_REGEXP));
15612 PERL_ARGS_ASSERT_OUTPUT_OR_RETURN_POSIX_WARNINGS;
15614 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
15615 if (return_posix_warnings) {
15616 if (! *return_posix_warnings) { /* mortalize to not leak if
15617 warnings are fatal */
15618 *return_posix_warnings = (AV *) sv_2mortal((SV *) newAV());
15620 av_push(*return_posix_warnings, msg);
15623 if (first_is_fatal) { /* Avoid leaking this */
15624 av_undef(posix_warnings); /* This isn't necessary if the
15625 array is mortal, but is a
15627 (void) sv_2mortal(msg);
15629 SAVEFREESV(RExC_rx_sv);
15632 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
15633 SvREFCNT_dec_NN(msg);
15639 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
15641 /* This adds the string scalar <multi_string> to the array
15642 * <multi_char_matches>. <multi_string> is known to have exactly
15643 * <cp_count> code points in it. This is used when constructing a
15644 * bracketed character class and we find something that needs to match more
15645 * than a single character.
15647 * <multi_char_matches> is actually an array of arrays. Each top-level
15648 * element is an array that contains all the strings known so far that are
15649 * the same length. And that length (in number of code points) is the same
15650 * as the index of the top-level array. Hence, the [2] element is an
15651 * array, each element thereof is a string containing TWO code points;
15652 * while element [3] is for strings of THREE characters, and so on. Since
15653 * this is for multi-char strings there can never be a [0] nor [1] element.
15655 * When we rewrite the character class below, we will do so such that the
15656 * longest strings are written first, so that it prefers the longest
15657 * matching strings first. This is done even if it turns out that any
15658 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
15659 * Christiansen has agreed that this is ok. This makes the test for the
15660 * ligature 'ffi' come before the test for 'ff', for example */
15663 AV** this_array_ptr;
15665 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
15667 if (! multi_char_matches) {
15668 multi_char_matches = newAV();
15671 if (av_exists(multi_char_matches, cp_count)) {
15672 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
15673 this_array = *this_array_ptr;
15676 this_array = newAV();
15677 av_store(multi_char_matches, cp_count,
15680 av_push(this_array, multi_string);
15682 return multi_char_matches;
15685 /* The names of properties whose definitions are not known at compile time are
15686 * stored in this SV, after a constant heading. So if the length has been
15687 * changed since initialization, then there is a run-time definition. */
15688 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
15689 (SvCUR(listsv) != initial_listsv_len)
15691 /* There is a restricted set of white space characters that are legal when
15692 * ignoring white space in a bracketed character class. This generates the
15693 * code to skip them.
15695 * There is a line below that uses the same white space criteria but is outside
15696 * this macro. Both here and there must use the same definition */
15697 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
15700 while (isBLANK_A(UCHARAT(p))) \
15708 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
15709 const bool stop_at_1, /* Just parse the next thing, don't
15710 look for a full character class */
15711 bool allow_multi_folds,
15712 const bool silence_non_portable, /* Don't output warnings
15716 bool optimizable, /* ? Allow a non-ANYOF return
15718 SV** ret_invlist, /* Return an inversion list, not a node */
15719 AV** return_posix_warnings
15722 /* parse a bracketed class specification. Most of these will produce an
15723 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
15724 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
15725 * under /i with multi-character folds: it will be rewritten following the
15726 * paradigm of this example, where the <multi-fold>s are characters which
15727 * fold to multiple character sequences:
15728 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
15729 * gets effectively rewritten as:
15730 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
15731 * reg() gets called (recursively) on the rewritten version, and this
15732 * function will return what it constructs. (Actually the <multi-fold>s
15733 * aren't physically removed from the [abcdefghi], it's just that they are
15734 * ignored in the recursion by means of a flag:
15735 * <RExC_in_multi_char_class>.)
15737 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
15738 * characters, with the corresponding bit set if that character is in the
15739 * list. For characters above this, a range list or swash is used. There
15740 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
15741 * determinable at compile time
15743 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs
15744 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded
15745 * to UTF-8. This can only happen if ret_invlist is non-NULL.
15748 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
15750 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
15753 int namedclass = OOB_NAMEDCLASS;
15754 char *rangebegin = NULL;
15755 bool need_class = 0;
15757 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
15758 than just initialized. */
15759 SV* properties = NULL; /* Code points that match \p{} \P{} */
15760 SV* posixes = NULL; /* Code points that match classes like [:word:],
15761 extended beyond the Latin1 range. These have to
15762 be kept separate from other code points for much
15763 of this function because their handling is
15764 different under /i, and for most classes under
15766 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
15767 separate for a while from the non-complemented
15768 versions because of complications with /d
15770 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
15771 treated more simply than the general case,
15772 leading to less compilation and execution
15774 UV element_count = 0; /* Number of distinct elements in the class.
15775 Optimizations may be possible if this is tiny */
15776 AV * multi_char_matches = NULL; /* Code points that fold to more than one
15777 character; used under /i */
15779 char * stop_ptr = RExC_end; /* where to stop parsing */
15780 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
15783 /* Unicode properties are stored in a swash; this holds the current one
15784 * being parsed. If this swash is the only above-latin1 component of the
15785 * character class, an optimization is to pass it directly on to the
15786 * execution engine. Otherwise, it is set to NULL to indicate that there
15787 * are other things in the class that have to be dealt with at execution
15789 SV* swash = NULL; /* Code points that match \p{} \P{} */
15791 /* Set if a component of this character class is user-defined; just passed
15792 * on to the engine */
15793 bool has_user_defined_property = FALSE;
15795 /* inversion list of code points this node matches only when the target
15796 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
15798 SV* has_upper_latin1_only_utf8_matches = NULL;
15800 /* Inversion list of code points this node matches regardless of things
15801 * like locale, folding, utf8ness of the target string */
15802 SV* cp_list = NULL;
15804 /* Like cp_list, but code points on this list need to be checked for things
15805 * that fold to/from them under /i */
15806 SV* cp_foldable_list = NULL;
15808 /* Like cp_list, but code points on this list are valid only when the
15809 * runtime locale is UTF-8 */
15810 SV* only_utf8_locale_list = NULL;
15812 /* In a range, if one of the endpoints is non-character-set portable,
15813 * meaning that it hard-codes a code point that may mean a different
15814 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
15815 * mnemonic '\t' which each mean the same character no matter which
15816 * character set the platform is on. */
15817 unsigned int non_portable_endpoint = 0;
15819 /* Is the range unicode? which means on a platform that isn't 1-1 native
15820 * to Unicode (i.e. non-ASCII), each code point in it should be considered
15821 * to be a Unicode value. */
15822 bool unicode_range = FALSE;
15823 bool invert = FALSE; /* Is this class to be complemented */
15825 bool warn_super = ALWAYS_WARN_SUPER;
15827 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
15828 case we need to change the emitted regop to an EXACT. */
15829 const char * orig_parse = RExC_parse;
15830 const SSize_t orig_size = RExC_size;
15831 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
15833 /* This variable is used to mark where the end in the input is of something
15834 * that looks like a POSIX construct but isn't. During the parse, when
15835 * something looks like it could be such a construct is encountered, it is
15836 * checked for being one, but not if we've already checked this area of the
15837 * input. Only after this position is reached do we check again */
15838 char *not_posix_region_end = RExC_parse - 1;
15840 AV* posix_warnings = NULL;
15841 const bool do_posix_warnings = return_posix_warnings
15842 || (PASS2 && ckWARN(WARN_REGEXP));
15844 GET_RE_DEBUG_FLAGS_DECL;
15846 PERL_ARGS_ASSERT_REGCLASS;
15848 PERL_UNUSED_ARG(depth);
15851 DEBUG_PARSE("clas");
15853 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
15854 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
15855 && UNICODE_DOT_DOT_VERSION == 0)
15856 allow_multi_folds = FALSE;
15859 /* Assume we are going to generate an ANYOF node. */
15860 ret = reganode(pRExC_state,
15867 RExC_size += ANYOF_SKIP;
15868 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
15871 ANYOF_FLAGS(ret) = 0;
15873 RExC_emit += ANYOF_SKIP;
15874 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
15875 initial_listsv_len = SvCUR(listsv);
15876 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
15879 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15881 assert(RExC_parse <= RExC_end);
15883 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
15886 allow_multi_folds = FALSE;
15888 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15891 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
15892 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
15893 int maybe_class = handle_possible_posix(pRExC_state,
15895 ¬_posix_region_end,
15897 TRUE /* checking only */);
15898 if (PASS2 && maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
15899 SAVEFREESV(RExC_rx_sv);
15900 ckWARN4reg(not_posix_region_end,
15901 "POSIX syntax [%c %c] belongs inside character classes%s",
15902 *RExC_parse, *RExC_parse,
15903 (maybe_class == OOB_NAMEDCLASS)
15904 ? ((POSIXCC_NOTYET(*RExC_parse))
15905 ? " (but this one isn't implemented)"
15906 : " (but this one isn't fully valid)")
15909 (void)ReREFCNT_inc(RExC_rx_sv);
15913 /* If the caller wants us to just parse a single element, accomplish this
15914 * by faking the loop ending condition */
15915 if (stop_at_1 && RExC_end > RExC_parse) {
15916 stop_ptr = RExC_parse + 1;
15919 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
15920 if (UCHARAT(RExC_parse) == ']')
15921 goto charclassloop;
15925 if ( posix_warnings
15926 && av_tindex_nomg(posix_warnings) >= 0
15927 && RExC_parse > not_posix_region_end)
15929 /* Warnings about posix class issues are considered tentative until
15930 * we are far enough along in the parse that we can no longer
15931 * change our mind, at which point we either output them or add
15932 * them, if it has so specified, to what gets returned to the
15933 * caller. This is done each time through the loop so that a later
15934 * class won't zap them before they have been dealt with. */
15935 output_or_return_posix_warnings(pRExC_state, posix_warnings,
15936 return_posix_warnings);
15939 if (RExC_parse >= stop_ptr) {
15943 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
15945 if (UCHARAT(RExC_parse) == ']') {
15951 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
15952 save_value = value;
15953 save_prevvalue = prevvalue;
15956 rangebegin = RExC_parse;
15958 non_portable_endpoint = 0;
15960 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
15961 value = utf8n_to_uvchr((U8*)RExC_parse,
15962 RExC_end - RExC_parse,
15963 &numlen, UTF8_ALLOW_DEFAULT);
15964 RExC_parse += numlen;
15967 value = UCHARAT(RExC_parse++);
15969 if (value == '[') {
15970 char * posix_class_end;
15971 namedclass = handle_possible_posix(pRExC_state,
15974 do_posix_warnings ? &posix_warnings : NULL,
15975 FALSE /* die if error */);
15976 if (namedclass > OOB_NAMEDCLASS) {
15978 /* If there was an earlier attempt to parse this particular
15979 * posix class, and it failed, it was a false alarm, as this
15980 * successful one proves */
15981 if ( posix_warnings
15982 && av_tindex_nomg(posix_warnings) >= 0
15983 && not_posix_region_end >= RExC_parse
15984 && not_posix_region_end <= posix_class_end)
15986 av_undef(posix_warnings);
15989 RExC_parse = posix_class_end;
15991 else if (namedclass == OOB_NAMEDCLASS) {
15992 not_posix_region_end = posix_class_end;
15995 namedclass = OOB_NAMEDCLASS;
15998 else if ( RExC_parse - 1 > not_posix_region_end
15999 && MAYBE_POSIXCC(value))
16001 (void) handle_possible_posix(
16003 RExC_parse - 1, /* -1 because parse has already been
16005 ¬_posix_region_end,
16006 do_posix_warnings ? &posix_warnings : NULL,
16007 TRUE /* checking only */);
16009 else if (value == '\\') {
16010 /* Is a backslash; get the code point of the char after it */
16012 if (RExC_parse >= RExC_end) {
16013 vFAIL("Unmatched [");
16016 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
16017 value = utf8n_to_uvchr((U8*)RExC_parse,
16018 RExC_end - RExC_parse,
16019 &numlen, UTF8_ALLOW_DEFAULT);
16020 RExC_parse += numlen;
16023 value = UCHARAT(RExC_parse++);
16025 /* Some compilers cannot handle switching on 64-bit integer
16026 * values, therefore value cannot be an UV. Yes, this will
16027 * be a problem later if we want switch on Unicode.
16028 * A similar issue a little bit later when switching on
16029 * namedclass. --jhi */
16031 /* If the \ is escaping white space when white space is being
16032 * skipped, it means that that white space is wanted literally, and
16033 * is already in 'value'. Otherwise, need to translate the escape
16034 * into what it signifies. */
16035 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
16037 case 'w': namedclass = ANYOF_WORDCHAR; break;
16038 case 'W': namedclass = ANYOF_NWORDCHAR; break;
16039 case 's': namedclass = ANYOF_SPACE; break;
16040 case 'S': namedclass = ANYOF_NSPACE; break;
16041 case 'd': namedclass = ANYOF_DIGIT; break;
16042 case 'D': namedclass = ANYOF_NDIGIT; break;
16043 case 'v': namedclass = ANYOF_VERTWS; break;
16044 case 'V': namedclass = ANYOF_NVERTWS; break;
16045 case 'h': namedclass = ANYOF_HORIZWS; break;
16046 case 'H': namedclass = ANYOF_NHORIZWS; break;
16047 case 'N': /* Handle \N{NAME} in class */
16049 const char * const backslash_N_beg = RExC_parse - 2;
16052 if (! grok_bslash_N(pRExC_state,
16053 NULL, /* No regnode */
16054 &value, /* Yes single value */
16055 &cp_count, /* Multiple code pt count */
16061 if (*flagp & NEED_UTF8)
16062 FAIL("panic: grok_bslash_N set NEED_UTF8");
16063 if (*flagp & RESTART_PASS1)
16066 if (cp_count < 0) {
16067 vFAIL("\\N in a character class must be a named character: \\N{...}");
16069 else if (cp_count == 0) {
16071 ckWARNreg(RExC_parse,
16072 "Ignoring zero length \\N{} in character class");
16075 else { /* cp_count > 1 */
16076 if (! RExC_in_multi_char_class) {
16077 if (invert || range || *RExC_parse == '-') {
16080 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
16083 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
16085 break; /* <value> contains the first code
16086 point. Drop out of the switch to
16090 SV * multi_char_N = newSVpvn(backslash_N_beg,
16091 RExC_parse - backslash_N_beg);
16093 = add_multi_match(multi_char_matches,
16098 } /* End of cp_count != 1 */
16100 /* This element should not be processed further in this
16103 value = save_value;
16104 prevvalue = save_prevvalue;
16105 continue; /* Back to top of loop to get next char */
16108 /* Here, is a single code point, and <value> contains it */
16109 unicode_range = TRUE; /* \N{} are Unicode */
16117 /* We will handle any undefined properties ourselves */
16118 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
16119 /* And we actually would prefer to get
16120 * the straight inversion list of the
16121 * swash, since we will be accessing it
16122 * anyway, to save a little time */
16123 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
16125 if (RExC_parse >= RExC_end)
16126 vFAIL2("Empty \\%c", (U8)value);
16127 if (*RExC_parse == '{') {
16128 const U8 c = (U8)value;
16129 e = strchr(RExC_parse, '}');
16132 vFAIL2("Missing right brace on \\%c{}", c);
16136 while (isSPACE(*RExC_parse)) {
16140 if (UCHARAT(RExC_parse) == '^') {
16142 /* toggle. (The rhs xor gets the single bit that
16143 * differs between P and p; the other xor inverts just
16145 value ^= 'P' ^ 'p';
16148 while (isSPACE(*RExC_parse)) {
16153 if (e == RExC_parse)
16154 vFAIL2("Empty \\%c{}", c);
16156 n = e - RExC_parse;
16157 while (isSPACE(*(RExC_parse + n - 1)))
16159 } /* The \p isn't immediately followed by a '{' */
16160 else if (! isALPHA(*RExC_parse)) {
16161 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16162 vFAIL2("Character following \\%c must be '{' or a "
16163 "single-character Unicode property name",
16173 char* base_name; /* name after any packages are stripped */
16174 char* lookup_name = NULL;
16175 const char * const colon_colon = "::";
16177 /* Try to get the definition of the property into
16178 * <invlist>. If /i is in effect, the effective property
16179 * will have its name be <__NAME_i>. The design is
16180 * discussed in commit
16181 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
16182 name = savepv(Perl_form(aTHX_ "%.*s", (int)n, RExC_parse));
16185 lookup_name = savepv(Perl_form(aTHX_ "__%s_i", name));
16187 /* The function call just below that uses this can fail
16188 * to return, leaking memory if we don't do this */
16189 SAVEFREEPV(lookup_name);
16192 /* Look up the property name, and get its swash and
16193 * inversion list, if the property is found */
16194 SvREFCNT_dec(swash); /* Free any left-overs */
16195 swash = _core_swash_init("utf8",
16202 NULL, /* No inversion list */
16205 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
16206 HV* curpkg = (IN_PERL_COMPILETIME)
16208 : CopSTASH(PL_curcop);
16212 if (swash) { /* Got a swash but no inversion list.
16213 Something is likely wrong that will
16214 be sorted-out later */
16215 SvREFCNT_dec_NN(swash);
16219 /* Here didn't find it. It could be a an error (like a
16220 * typo) in specifying a Unicode property, or it could
16221 * be a user-defined property that will be available at
16222 * run-time. The names of these must begin with 'In'
16223 * or 'Is' (after any packages are stripped off). So
16224 * if not one of those, or if we accept only
16225 * compile-time properties, is an error; otherwise add
16226 * it to the list for run-time look up. */
16227 if ((base_name = rninstr(name, name + n,
16228 colon_colon, colon_colon + 2)))
16229 { /* Has ::. We know this must be a user-defined
16232 final_n -= base_name - name;
16241 || base_name[0] != 'I'
16242 || (base_name[1] != 's' && base_name[1] != 'n')
16245 const char * const msg
16247 ? "Illegal user-defined property name"
16248 : "Can't find Unicode property definition";
16249 RExC_parse = e + 1;
16251 /* diag_listed_as: Can't find Unicode property definition "%s" */
16252 vFAIL3utf8f("%s \"%"UTF8f"\"",
16253 msg, UTF8fARG(UTF, n, name));
16256 /* If the property name doesn't already have a package
16257 * name, add the current one to it so that it can be
16258 * referred to outside it. [perl #121777] */
16259 if (! has_pkg && curpkg) {
16260 char* pkgname = HvNAME(curpkg);
16261 if (strNE(pkgname, "main")) {
16262 char* full_name = Perl_form(aTHX_
16266 n = strlen(full_name);
16267 name = savepvn(full_name, n);
16271 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s%"UTF8f"%s\n",
16272 (value == 'p' ? '+' : '!'),
16273 (FOLD) ? "__" : "",
16274 UTF8fARG(UTF, n, name),
16275 (FOLD) ? "_i" : "");
16276 has_user_defined_property = TRUE;
16277 optimizable = FALSE; /* Will have to leave this an
16280 /* We don't know yet what this matches, so have to flag
16282 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
16286 /* Here, did get the swash and its inversion list. If
16287 * the swash is from a user-defined property, then this
16288 * whole character class should be regarded as such */
16289 if (swash_init_flags
16290 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
16292 has_user_defined_property = TRUE;
16295 /* We warn on matching an above-Unicode code point
16296 * if the match would return true, except don't
16297 * warn for \p{All}, which has exactly one element
16299 (_invlist_contains_cp(invlist, 0x110000)
16300 && (! (_invlist_len(invlist) == 1
16301 && *invlist_array(invlist) == 0)))
16307 /* Invert if asking for the complement */
16308 if (value == 'P') {
16309 _invlist_union_complement_2nd(properties,
16313 /* The swash can't be used as-is, because we've
16314 * inverted things; delay removing it to here after
16315 * have copied its invlist above */
16316 SvREFCNT_dec_NN(swash);
16320 _invlist_union(properties, invlist, &properties);
16324 RExC_parse = e + 1;
16325 namedclass = ANYOF_UNIPROP; /* no official name, but it's
16328 /* \p means they want Unicode semantics */
16329 REQUIRE_UNI_RULES(flagp, NULL);
16332 case 'n': value = '\n'; break;
16333 case 'r': value = '\r'; break;
16334 case 't': value = '\t'; break;
16335 case 'f': value = '\f'; break;
16336 case 'b': value = '\b'; break;
16337 case 'e': value = ESC_NATIVE; break;
16338 case 'a': value = '\a'; break;
16340 RExC_parse--; /* function expects to be pointed at the 'o' */
16342 const char* error_msg;
16343 bool valid = grok_bslash_o(&RExC_parse,
16346 PASS2, /* warnings only in
16349 silence_non_portable,
16355 non_portable_endpoint++;
16358 RExC_parse--; /* function expects to be pointed at the 'x' */
16360 const char* error_msg;
16361 bool valid = grok_bslash_x(&RExC_parse,
16364 PASS2, /* Output warnings */
16366 silence_non_portable,
16372 non_portable_endpoint++;
16375 value = grok_bslash_c(*RExC_parse++, PASS2);
16376 non_portable_endpoint++;
16378 case '0': case '1': case '2': case '3': case '4':
16379 case '5': case '6': case '7':
16381 /* Take 1-3 octal digits */
16382 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
16383 numlen = (strict) ? 4 : 3;
16384 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
16385 RExC_parse += numlen;
16388 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16389 vFAIL("Need exactly 3 octal digits");
16391 else if (! SIZE_ONLY /* like \08, \178 */
16393 && RExC_parse < RExC_end
16394 && isDIGIT(*RExC_parse)
16395 && ckWARN(WARN_REGEXP))
16397 SAVEFREESV(RExC_rx_sv);
16398 reg_warn_non_literal_string(
16400 form_short_octal_warning(RExC_parse, numlen));
16401 (void)ReREFCNT_inc(RExC_rx_sv);
16404 non_portable_endpoint++;
16408 /* Allow \_ to not give an error */
16409 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
16411 vFAIL2("Unrecognized escape \\%c in character class",
16415 SAVEFREESV(RExC_rx_sv);
16416 ckWARN2reg(RExC_parse,
16417 "Unrecognized escape \\%c in character class passed through",
16419 (void)ReREFCNT_inc(RExC_rx_sv);
16423 } /* End of switch on char following backslash */
16424 } /* end of handling backslash escape sequences */
16426 /* Here, we have the current token in 'value' */
16428 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
16431 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
16432 * literal, as is the character that began the false range, i.e.
16433 * the 'a' in the examples */
16436 const int w = (RExC_parse >= rangebegin)
16437 ? RExC_parse - rangebegin
16441 "False [] range \"%"UTF8f"\"",
16442 UTF8fARG(UTF, w, rangebegin));
16445 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
16446 ckWARN2reg(RExC_parse,
16447 "False [] range \"%"UTF8f"\"",
16448 UTF8fARG(UTF, w, rangebegin));
16449 (void)ReREFCNT_inc(RExC_rx_sv);
16450 cp_list = add_cp_to_invlist(cp_list, '-');
16451 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
16456 range = 0; /* this was not a true range */
16457 element_count += 2; /* So counts for three values */
16460 classnum = namedclass_to_classnum(namedclass);
16462 if (LOC && namedclass < ANYOF_POSIXL_MAX
16463 #ifndef HAS_ISASCII
16464 && classnum != _CC_ASCII
16467 /* What the Posix classes (like \w, [:space:]) match in locale
16468 * isn't knowable under locale until actual match time. Room
16469 * must be reserved (one time per outer bracketed class) to
16470 * store such classes. The space will contain a bit for each
16471 * named class that is to be matched against. This isn't
16472 * needed for \p{} and pseudo-classes, as they are not affected
16473 * by locale, and hence are dealt with separately */
16474 if (! need_class) {
16477 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16480 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16482 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
16483 ANYOF_POSIXL_ZERO(ret);
16485 /* We can't change this into some other type of node
16486 * (unless this is the only element, in which case there
16487 * are nodes that mean exactly this) as has runtime
16489 optimizable = FALSE;
16492 /* Coverity thinks it is possible for this to be negative; both
16493 * jhi and khw think it's not, but be safer */
16494 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16495 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
16497 /* See if it already matches the complement of this POSIX
16499 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16500 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
16504 posixl_matches_all = TRUE;
16505 break; /* No need to continue. Since it matches both
16506 e.g., \w and \W, it matches everything, and the
16507 bracketed class can be optimized into qr/./s */
16510 /* Add this class to those that should be checked at runtime */
16511 ANYOF_POSIXL_SET(ret, namedclass);
16513 /* The above-Latin1 characters are not subject to locale rules.
16514 * Just add them, in the second pass, to the
16515 * unconditionally-matched list */
16517 SV* scratch_list = NULL;
16519 /* Get the list of the above-Latin1 code points this
16521 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
16522 PL_XPosix_ptrs[classnum],
16524 /* Odd numbers are complements, like
16525 * NDIGIT, NASCII, ... */
16526 namedclass % 2 != 0,
16528 /* Checking if 'cp_list' is NULL first saves an extra
16529 * clone. Its reference count will be decremented at the
16530 * next union, etc, or if this is the only instance, at the
16531 * end of the routine */
16533 cp_list = scratch_list;
16536 _invlist_union(cp_list, scratch_list, &cp_list);
16537 SvREFCNT_dec_NN(scratch_list);
16539 continue; /* Go get next character */
16542 else if (! SIZE_ONLY) {
16544 /* Here, not in pass1 (in that pass we skip calculating the
16545 * contents of this class), and is not /l, or is a POSIX class
16546 * for which /l doesn't matter (or is a Unicode property, which
16547 * is skipped here). */
16548 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
16549 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
16551 /* Here, should be \h, \H, \v, or \V. None of /d, /i
16552 * nor /l make a difference in what these match,
16553 * therefore we just add what they match to cp_list. */
16554 if (classnum != _CC_VERTSPACE) {
16555 assert( namedclass == ANYOF_HORIZWS
16556 || namedclass == ANYOF_NHORIZWS);
16558 /* It turns out that \h is just a synonym for
16560 classnum = _CC_BLANK;
16563 _invlist_union_maybe_complement_2nd(
16565 PL_XPosix_ptrs[classnum],
16566 namedclass % 2 != 0, /* Complement if odd
16567 (NHORIZWS, NVERTWS)
16572 else if ( UNI_SEMANTICS
16573 || classnum == _CC_ASCII
16574 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
16575 || classnum == _CC_XDIGIT)))
16577 /* We usually have to worry about /d and /a affecting what
16578 * POSIX classes match, with special code needed for /d
16579 * because we won't know until runtime what all matches.
16580 * But there is no extra work needed under /u, and
16581 * [:ascii:] is unaffected by /a and /d; and :digit: and
16582 * :xdigit: don't have runtime differences under /d. So we
16583 * can special case these, and avoid some extra work below,
16584 * and at runtime. */
16585 _invlist_union_maybe_complement_2nd(
16587 PL_XPosix_ptrs[classnum],
16588 namedclass % 2 != 0,
16591 else { /* Garden variety class. If is NUPPER, NALPHA, ...
16592 complement and use nposixes */
16593 SV** posixes_ptr = namedclass % 2 == 0
16596 _invlist_union_maybe_complement_2nd(
16598 PL_XPosix_ptrs[classnum],
16599 namedclass % 2 != 0,
16603 } /* end of namedclass \blah */
16605 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16607 /* If 'range' is set, 'value' is the ending of a range--check its
16608 * validity. (If value isn't a single code point in the case of a
16609 * range, we should have figured that out above in the code that
16610 * catches false ranges). Later, we will handle each individual code
16611 * point in the range. If 'range' isn't set, this could be the
16612 * beginning of a range, so check for that by looking ahead to see if
16613 * the next real character to be processed is the range indicator--the
16618 /* For unicode ranges, we have to test that the Unicode as opposed
16619 * to the native values are not decreasing. (Above 255, there is
16620 * no difference between native and Unicode) */
16621 if (unicode_range && prevvalue < 255 && value < 255) {
16622 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
16623 goto backwards_range;
16628 if (prevvalue > value) /* b-a */ {
16633 w = RExC_parse - rangebegin;
16635 "Invalid [] range \"%"UTF8f"\"",
16636 UTF8fARG(UTF, w, rangebegin));
16637 NOT_REACHED; /* NOTREACHED */
16641 prevvalue = value; /* save the beginning of the potential range */
16642 if (! stop_at_1 /* Can't be a range if parsing just one thing */
16643 && *RExC_parse == '-')
16645 char* next_char_ptr = RExC_parse + 1;
16647 /* Get the next real char after the '-' */
16648 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
16650 /* If the '-' is at the end of the class (just before the ']',
16651 * it is a literal minus; otherwise it is a range */
16652 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
16653 RExC_parse = next_char_ptr;
16655 /* a bad range like \w-, [:word:]- ? */
16656 if (namedclass > OOB_NAMEDCLASS) {
16657 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
16658 const int w = RExC_parse >= rangebegin
16659 ? RExC_parse - rangebegin
16662 vFAIL4("False [] range \"%*.*s\"",
16667 "False [] range \"%*.*s\"",
16672 cp_list = add_cp_to_invlist(cp_list, '-');
16676 range = 1; /* yeah, it's a range! */
16677 continue; /* but do it the next time */
16682 if (namedclass > OOB_NAMEDCLASS) {
16686 /* Here, we have a single value this time through the loop, and
16687 * <prevvalue> is the beginning of the range, if any; or <value> if
16690 /* non-Latin1 code point implies unicode semantics. Must be set in
16691 * pass1 so is there for the whole of pass 2 */
16693 REQUIRE_UNI_RULES(flagp, NULL);
16696 /* Ready to process either the single value, or the completed range.
16697 * For single-valued non-inverted ranges, we consider the possibility
16698 * of multi-char folds. (We made a conscious decision to not do this
16699 * for the other cases because it can often lead to non-intuitive
16700 * results. For example, you have the peculiar case that:
16701 * "s s" =~ /^[^\xDF]+$/i => Y
16702 * "ss" =~ /^[^\xDF]+$/i => N
16704 * See [perl #89750] */
16705 if (FOLD && allow_multi_folds && value == prevvalue) {
16706 if (value == LATIN_SMALL_LETTER_SHARP_S
16707 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
16710 /* Here <value> is indeed a multi-char fold. Get what it is */
16712 U8 foldbuf[UTF8_MAXBYTES_CASE];
16715 UV folded = _to_uni_fold_flags(
16719 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
16720 ? FOLD_FLAGS_NOMIX_ASCII
16724 /* Here, <folded> should be the first character of the
16725 * multi-char fold of <value>, with <foldbuf> containing the
16726 * whole thing. But, if this fold is not allowed (because of
16727 * the flags), <fold> will be the same as <value>, and should
16728 * be processed like any other character, so skip the special
16730 if (folded != value) {
16732 /* Skip if we are recursed, currently parsing the class
16733 * again. Otherwise add this character to the list of
16734 * multi-char folds. */
16735 if (! RExC_in_multi_char_class) {
16736 STRLEN cp_count = utf8_length(foldbuf,
16737 foldbuf + foldlen);
16738 SV* multi_fold = sv_2mortal(newSVpvs(""));
16740 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
16743 = add_multi_match(multi_char_matches,
16749 /* This element should not be processed further in this
16752 value = save_value;
16753 prevvalue = save_prevvalue;
16759 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
16762 /* If the range starts above 255, everything is portable and
16763 * likely to be so for any forseeable character set, so don't
16765 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
16766 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
16768 else if (prevvalue != value) {
16770 /* Under strict, ranges that stop and/or end in an ASCII
16771 * printable should have each end point be a portable value
16772 * for it (preferably like 'A', but we don't warn if it is
16773 * a (portable) Unicode name or code point), and the range
16774 * must be be all digits or all letters of the same case.
16775 * Otherwise, the range is non-portable and unclear as to
16776 * what it contains */
16777 if ((isPRINT_A(prevvalue) || isPRINT_A(value))
16778 && (non_portable_endpoint
16779 || ! ((isDIGIT_A(prevvalue) && isDIGIT_A(value))
16780 || (isLOWER_A(prevvalue) && isLOWER_A(value))
16781 || (isUPPER_A(prevvalue) && isUPPER_A(value)))))
16783 vWARN(RExC_parse, "Ranges of ASCII printables should be some subset of \"0-9\", \"A-Z\", or \"a-z\"");
16785 else if (prevvalue >= 0x660) { /* ARABIC_INDIC_DIGIT_ZERO */
16787 /* But the nature of Unicode and languages mean we
16788 * can't do the same checks for above-ASCII ranges,
16789 * except in the case of digit ones. These should
16790 * contain only digits from the same group of 10. The
16791 * ASCII case is handled just above. 0x660 is the
16792 * first digit character beyond ASCII. Hence here, the
16793 * range could be a range of digits. Find out. */
16794 IV index_start = _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
16796 IV index_final = _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
16799 /* If the range start and final points are in the same
16800 * inversion list element, it means that either both
16801 * are not digits, or both are digits in a consecutive
16802 * sequence of digits. (So far, Unicode has kept all
16803 * such sequences as distinct groups of 10, but assert
16804 * to make sure). If the end points are not in the
16805 * same element, neither should be a digit. */
16806 if (index_start == index_final) {
16807 assert(! ELEMENT_RANGE_MATCHES_INVLIST(index_start)
16808 || (invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start+1]
16809 - invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
16811 /* But actually Unicode did have one group of 11
16812 * 'digits' in 5.2, so in case we are operating
16813 * on that version, let that pass */
16814 || (invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start+1]
16815 - invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
16817 && invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
16821 else if ((index_start >= 0
16822 && ELEMENT_RANGE_MATCHES_INVLIST(index_start))
16823 || (index_final >= 0
16824 && ELEMENT_RANGE_MATCHES_INVLIST(index_final)))
16826 vWARN(RExC_parse, "Ranges of digits should be from the same group of 10");
16831 if ((! range || prevvalue == value) && non_portable_endpoint) {
16832 if (isPRINT_A(value)) {
16835 if (isBACKSLASHED_PUNCT(value)) {
16836 literal[d++] = '\\';
16838 literal[d++] = (char) value;
16839 literal[d++] = '\0';
16842 "\"%.*s\" is more clearly written simply as \"%s\"",
16843 (int) (RExC_parse - rangebegin),
16848 else if isMNEMONIC_CNTRL(value) {
16850 "\"%.*s\" is more clearly written simply as \"%s\"",
16851 (int) (RExC_parse - rangebegin),
16853 cntrl_to_mnemonic((U8) value)
16859 /* Deal with this element of the class */
16863 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16866 /* On non-ASCII platforms, for ranges that span all of 0..255, and
16867 * ones that don't require special handling, we can just add the
16868 * range like we do for ASCII platforms */
16869 if ((UNLIKELY(prevvalue == 0) && value >= 255)
16870 || ! (prevvalue < 256
16872 || (! non_portable_endpoint
16873 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
16874 || (isUPPER_A(prevvalue)
16875 && isUPPER_A(value)))))))
16877 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16881 /* Here, requires special handling. This can be because it is
16882 * a range whose code points are considered to be Unicode, and
16883 * so must be individually translated into native, or because
16884 * its a subrange of 'A-Z' or 'a-z' which each aren't
16885 * contiguous in EBCDIC, but we have defined them to include
16886 * only the "expected" upper or lower case ASCII alphabetics.
16887 * Subranges above 255 are the same in native and Unicode, so
16888 * can be added as a range */
16889 U8 start = NATIVE_TO_LATIN1(prevvalue);
16891 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
16892 for (j = start; j <= end; j++) {
16893 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
16896 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
16903 range = 0; /* this range (if it was one) is done now */
16904 } /* End of loop through all the text within the brackets */
16907 if ( posix_warnings && av_tindex_nomg(posix_warnings) >= 0) {
16908 output_or_return_posix_warnings(pRExC_state, posix_warnings,
16909 return_posix_warnings);
16912 /* If anything in the class expands to more than one character, we have to
16913 * deal with them by building up a substitute parse string, and recursively
16914 * calling reg() on it, instead of proceeding */
16915 if (multi_char_matches) {
16916 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
16919 char *save_end = RExC_end;
16920 char *save_parse = RExC_parse;
16921 char *save_start = RExC_start;
16922 STRLEN prefix_end = 0; /* We copy the character class after a
16923 prefix supplied here. This is the size
16924 + 1 of that prefix */
16925 bool first_time = TRUE; /* First multi-char occurrence doesn't get
16930 assert(RExC_precomp_adj == 0); /* Only one level of recursion allowed */
16932 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
16933 because too confusing */
16935 sv_catpv(substitute_parse, "(?:");
16939 /* Look at the longest folds first */
16940 for (cp_count = av_tindex_nomg(multi_char_matches);
16945 if (av_exists(multi_char_matches, cp_count)) {
16946 AV** this_array_ptr;
16949 this_array_ptr = (AV**) av_fetch(multi_char_matches,
16951 while ((this_sequence = av_pop(*this_array_ptr)) !=
16954 if (! first_time) {
16955 sv_catpv(substitute_parse, "|");
16957 first_time = FALSE;
16959 sv_catpv(substitute_parse, SvPVX(this_sequence));
16964 /* If the character class contains anything else besides these
16965 * multi-character folds, have to include it in recursive parsing */
16966 if (element_count) {
16967 sv_catpv(substitute_parse, "|[");
16968 prefix_end = SvCUR(substitute_parse);
16969 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
16971 /* Put in a closing ']' only if not going off the end, as otherwise
16972 * we are adding something that really isn't there */
16973 if (RExC_parse < RExC_end) {
16974 sv_catpv(substitute_parse, "]");
16978 sv_catpv(substitute_parse, ")");
16981 /* This is a way to get the parse to skip forward a whole named
16982 * sequence instead of matching the 2nd character when it fails the
16984 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
16988 /* Set up the data structure so that any errors will be properly
16989 * reported. See the comments at the definition of
16990 * REPORT_LOCATION_ARGS for details */
16991 RExC_precomp_adj = orig_parse - RExC_precomp;
16992 RExC_start = RExC_parse = SvPV(substitute_parse, len);
16993 RExC_adjusted_start = RExC_start + prefix_end;
16994 RExC_end = RExC_parse + len;
16995 RExC_in_multi_char_class = 1;
16996 RExC_override_recoding = 1;
16997 RExC_emit = (regnode *)orig_emit;
16999 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17001 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PASS1|NEED_UTF8);
17003 /* And restore so can parse the rest of the pattern */
17004 RExC_parse = save_parse;
17005 RExC_start = RExC_adjusted_start = save_start;
17006 RExC_precomp_adj = 0;
17007 RExC_end = save_end;
17008 RExC_in_multi_char_class = 0;
17009 RExC_override_recoding = 0;
17010 SvREFCNT_dec_NN(multi_char_matches);
17014 /* Here, we've gone through the entire class and dealt with multi-char
17015 * folds. We are now in a position that we can do some checks to see if we
17016 * can optimize this ANYOF node into a simpler one, even in Pass 1.
17017 * Currently we only do two checks:
17018 * 1) is in the unlikely event that the user has specified both, eg. \w and
17019 * \W under /l, then the class matches everything. (This optimization
17020 * is done only to make the optimizer code run later work.)
17021 * 2) if the character class contains only a single element (including a
17022 * single range), we see if there is an equivalent node for it.
17023 * Other checks are possible */
17025 && ! ret_invlist /* Can't optimize if returning the constructed
17027 && (UNLIKELY(posixl_matches_all) || element_count == 1))
17032 if (UNLIKELY(posixl_matches_all)) {
17035 else if (namedclass > OOB_NAMEDCLASS) { /* this is a single named
17036 class, like \w or [:digit:]
17039 /* All named classes are mapped into POSIXish nodes, with its FLAG
17040 * argument giving which class it is */
17041 switch ((I32)namedclass) {
17042 case ANYOF_UNIPROP:
17045 /* These don't depend on the charset modifiers. They always
17046 * match under /u rules */
17047 case ANYOF_NHORIZWS:
17048 case ANYOF_HORIZWS:
17049 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
17052 case ANYOF_NVERTWS:
17057 /* The actual POSIXish node for all the rest depends on the
17058 * charset modifier. The ones in the first set depend only on
17059 * ASCII or, if available on this platform, also locale */
17063 op = (LOC) ? POSIXL : POSIXA;
17069 /* The following don't have any matches in the upper Latin1
17070 * range, hence /d is equivalent to /u for them. Making it /u
17071 * saves some branches at runtime */
17075 case ANYOF_NXDIGIT:
17076 if (! DEPENDS_SEMANTICS) {
17077 goto treat_as_default;
17083 /* The following change to CASED under /i */
17089 namedclass = ANYOF_CASED + (namedclass % 2);
17093 /* The rest have more possibilities depending on the charset.
17094 * We take advantage of the enum ordering of the charset
17095 * modifiers to get the exact node type, */
17098 op = POSIXD + get_regex_charset(RExC_flags);
17099 if (op > POSIXA) { /* /aa is same as /a */
17104 /* The odd numbered ones are the complements of the
17105 * next-lower even number one */
17106 if (namedclass % 2 == 1) {
17110 arg = namedclass_to_classnum(namedclass);
17114 else if (value == prevvalue) {
17116 /* Here, the class consists of just a single code point */
17119 if (! LOC && value == '\n') {
17120 op = REG_ANY; /* Optimize [^\n] */
17121 *flagp |= HASWIDTH|SIMPLE;
17125 else if (value < 256 || UTF) {
17127 /* Optimize a single value into an EXACTish node, but not if it
17128 * would require converting the pattern to UTF-8. */
17129 op = compute_EXACTish(pRExC_state);
17131 } /* Otherwise is a range */
17132 else if (! LOC) { /* locale could vary these */
17133 if (prevvalue == '0') {
17134 if (value == '9') {
17139 else if (! FOLD || ASCII_FOLD_RESTRICTED) {
17140 /* We can optimize A-Z or a-z, but not if they could match
17141 * something like the KELVIN SIGN under /i. */
17142 if (prevvalue == 'A') {
17145 && ! non_portable_endpoint
17148 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
17152 else if (prevvalue == 'a') {
17155 && ! non_portable_endpoint
17158 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
17165 /* Here, we have changed <op> away from its initial value iff we found
17166 * an optimization */
17169 /* Throw away this ANYOF regnode, and emit the calculated one,
17170 * which should correspond to the beginning, not current, state of
17172 const char * cur_parse = RExC_parse;
17173 RExC_parse = (char *)orig_parse;
17177 /* To get locale nodes to not use the full ANYOF size would
17178 * require moving the code above that writes the portions
17179 * of it that aren't in other nodes to after this point.
17180 * e.g. ANYOF_POSIXL_SET */
17181 RExC_size = orig_size;
17185 RExC_emit = (regnode *)orig_emit;
17186 if (PL_regkind[op] == POSIXD) {
17187 if (op == POSIXL) {
17188 RExC_contains_locale = 1;
17191 op += NPOSIXD - POSIXD;
17196 ret = reg_node(pRExC_state, op);
17198 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17202 *flagp |= HASWIDTH|SIMPLE;
17204 else if (PL_regkind[op] == EXACT) {
17205 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17206 TRUE /* downgradable to EXACT */
17210 RExC_parse = (char *) cur_parse;
17212 SvREFCNT_dec(posixes);
17213 SvREFCNT_dec(nposixes);
17214 SvREFCNT_dec(simple_posixes);
17215 SvREFCNT_dec(cp_list);
17216 SvREFCNT_dec(cp_foldable_list);
17223 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
17225 /* If folding, we calculate all characters that could fold to or from the
17226 * ones already on the list */
17227 if (cp_foldable_list) {
17229 UV start, end; /* End points of code point ranges */
17231 SV* fold_intersection = NULL;
17234 /* Our calculated list will be for Unicode rules. For locale
17235 * matching, we have to keep a separate list that is consulted at
17236 * runtime only when the locale indicates Unicode rules. For
17237 * non-locale, we just use the general list */
17239 use_list = &only_utf8_locale_list;
17242 use_list = &cp_list;
17245 /* Only the characters in this class that participate in folds need
17246 * be checked. Get the intersection of this class and all the
17247 * possible characters that are foldable. This can quickly narrow
17248 * down a large class */
17249 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
17250 &fold_intersection);
17252 /* The folds for all the Latin1 characters are hard-coded into this
17253 * program, but we have to go out to disk to get the others. */
17254 if (invlist_highest(cp_foldable_list) >= 256) {
17256 /* This is a hash that for a particular fold gives all
17257 * characters that are involved in it */
17258 if (! PL_utf8_foldclosures) {
17259 _load_PL_utf8_foldclosures();
17263 /* Now look at the foldable characters in this class individually */
17264 invlist_iterinit(fold_intersection);
17265 while (invlist_iternext(fold_intersection, &start, &end)) {
17268 /* Look at every character in the range */
17269 for (j = start; j <= end; j++) {
17270 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17276 if (IS_IN_SOME_FOLD_L1(j)) {
17278 /* ASCII is always matched; non-ASCII is matched
17279 * only under Unicode rules (which could happen
17280 * under /l if the locale is a UTF-8 one */
17281 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
17282 *use_list = add_cp_to_invlist(*use_list,
17283 PL_fold_latin1[j]);
17286 has_upper_latin1_only_utf8_matches
17287 = add_cp_to_invlist(
17288 has_upper_latin1_only_utf8_matches,
17289 PL_fold_latin1[j]);
17293 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
17294 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
17296 add_above_Latin1_folds(pRExC_state,
17303 /* Here is an above Latin1 character. We don't have the
17304 * rules hard-coded for it. First, get its fold. This is
17305 * the simple fold, as the multi-character folds have been
17306 * handled earlier and separated out */
17307 _to_uni_fold_flags(j, foldbuf, &foldlen,
17308 (ASCII_FOLD_RESTRICTED)
17309 ? FOLD_FLAGS_NOMIX_ASCII
17312 /* Single character fold of above Latin1. Add everything in
17313 * its fold closure to the list that this node should match.
17314 * The fold closures data structure is a hash with the keys
17315 * being the UTF-8 of every character that is folded to, like
17316 * 'k', and the values each an array of all code points that
17317 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
17318 * Multi-character folds are not included */
17319 if ((listp = hv_fetch(PL_utf8_foldclosures,
17320 (char *) foldbuf, foldlen, FALSE)))
17322 AV* list = (AV*) *listp;
17324 for (k = 0; k <= av_tindex_nomg(list); k++) {
17325 SV** c_p = av_fetch(list, k, FALSE);
17331 /* /aa doesn't allow folds between ASCII and non- */
17332 if ((ASCII_FOLD_RESTRICTED
17333 && (isASCII(c) != isASCII(j))))
17338 /* Folds under /l which cross the 255/256 boundary
17339 * are added to a separate list. (These are valid
17340 * only when the locale is UTF-8.) */
17341 if (c < 256 && LOC) {
17342 *use_list = add_cp_to_invlist(*use_list, c);
17346 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
17348 cp_list = add_cp_to_invlist(cp_list, c);
17351 /* Similarly folds involving non-ascii Latin1
17352 * characters under /d are added to their list */
17353 has_upper_latin1_only_utf8_matches
17354 = add_cp_to_invlist(
17355 has_upper_latin1_only_utf8_matches,
17362 SvREFCNT_dec_NN(fold_intersection);
17365 /* Now that we have finished adding all the folds, there is no reason
17366 * to keep the foldable list separate */
17367 _invlist_union(cp_list, cp_foldable_list, &cp_list);
17368 SvREFCNT_dec_NN(cp_foldable_list);
17371 /* And combine the result (if any) with any inversion lists from posix
17372 * classes. The lists are kept separate up to now because we don't want to
17373 * fold the classes (folding of those is automatically handled by the swash
17374 * fetching code) */
17375 if (simple_posixes) { /* These are the classes known to be unaffected by
17378 _invlist_union(cp_list, simple_posixes, &cp_list);
17379 SvREFCNT_dec_NN(simple_posixes);
17382 cp_list = simple_posixes;
17385 if (posixes || nposixes) {
17387 /* We have to adjust /a and /aa */
17388 if (AT_LEAST_ASCII_RESTRICTED) {
17390 /* Under /a and /aa, nothing above ASCII matches these */
17392 _invlist_intersection(posixes,
17393 PL_XPosix_ptrs[_CC_ASCII],
17397 /* Under /a and /aa, everything above ASCII matches these
17400 _invlist_union_complement_2nd(nposixes,
17401 PL_XPosix_ptrs[_CC_ASCII],
17406 if (! DEPENDS_SEMANTICS) {
17408 /* For everything but /d, we can just add the current 'posixes' and
17409 * 'nposixes' to the main list */
17412 _invlist_union(cp_list, posixes, &cp_list);
17413 SvREFCNT_dec_NN(posixes);
17421 _invlist_union(cp_list, nposixes, &cp_list);
17422 SvREFCNT_dec_NN(nposixes);
17425 cp_list = nposixes;
17430 /* Under /d, things like \w match upper Latin1 characters only if
17431 * the target string is in UTF-8. But things like \W match all the
17432 * upper Latin1 characters if the target string is not in UTF-8.
17434 * Handle the case where there something like \W separately */
17436 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1);
17438 /* A complemented posix class matches all upper Latin1
17439 * characters if not in UTF-8. And it matches just certain
17440 * ones when in UTF-8. That means those certain ones are
17441 * matched regardless, so can just be added to the
17442 * unconditional list */
17444 _invlist_union(cp_list, nposixes, &cp_list);
17445 SvREFCNT_dec_NN(nposixes);
17449 cp_list = nposixes;
17452 /* Likewise for 'posixes' */
17453 _invlist_union(posixes, cp_list, &cp_list);
17455 /* Likewise for anything else in the range that matched only
17457 if (has_upper_latin1_only_utf8_matches) {
17458 _invlist_union(cp_list,
17459 has_upper_latin1_only_utf8_matches,
17461 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17462 has_upper_latin1_only_utf8_matches = NULL;
17465 /* If we don't match all the upper Latin1 characters regardless
17466 * of UTF-8ness, we have to set a flag to match the rest when
17468 _invlist_subtract(only_non_utf8_list, cp_list,
17469 &only_non_utf8_list);
17470 if (_invlist_len(only_non_utf8_list) != 0) {
17471 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17475 /* Here there were no complemented posix classes. That means
17476 * the upper Latin1 characters in 'posixes' match only when the
17477 * target string is in UTF-8. So we have to add them to the
17478 * list of those types of code points, while adding the
17479 * remainder to the unconditional list.
17481 * First calculate what they are */
17482 SV* nonascii_but_latin1_properties = NULL;
17483 _invlist_intersection(posixes, PL_UpperLatin1,
17484 &nonascii_but_latin1_properties);
17486 /* And add them to the final list of such characters. */
17487 _invlist_union(has_upper_latin1_only_utf8_matches,
17488 nonascii_but_latin1_properties,
17489 &has_upper_latin1_only_utf8_matches);
17491 /* Remove them from what now becomes the unconditional list */
17492 _invlist_subtract(posixes, nonascii_but_latin1_properties,
17495 /* And add those unconditional ones to the final list */
17497 _invlist_union(cp_list, posixes, &cp_list);
17498 SvREFCNT_dec_NN(posixes);
17505 SvREFCNT_dec(nonascii_but_latin1_properties);
17507 /* Get rid of any characters that we now know are matched
17508 * unconditionally from the conditional list, which may make
17509 * that list empty */
17510 _invlist_subtract(has_upper_latin1_only_utf8_matches,
17512 &has_upper_latin1_only_utf8_matches);
17513 if (_invlist_len(has_upper_latin1_only_utf8_matches) == 0) {
17514 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17515 has_upper_latin1_only_utf8_matches = NULL;
17521 /* And combine the result (if any) with any inversion list from properties.
17522 * The lists are kept separate up to now so that we can distinguish the two
17523 * in regards to matching above-Unicode. A run-time warning is generated
17524 * if a Unicode property is matched against a non-Unicode code point. But,
17525 * we allow user-defined properties to match anything, without any warning,
17526 * and we also suppress the warning if there is a portion of the character
17527 * class that isn't a Unicode property, and which matches above Unicode, \W
17528 * or [\x{110000}] for example.
17529 * (Note that in this case, unlike the Posix one above, there is no
17530 * <has_upper_latin1_only_utf8_matches>, because having a Unicode property
17531 * forces Unicode semantics */
17535 /* If it matters to the final outcome, see if a non-property
17536 * component of the class matches above Unicode. If so, the
17537 * warning gets suppressed. This is true even if just a single
17538 * such code point is specified, as, though not strictly correct if
17539 * another such code point is matched against, the fact that they
17540 * are using above-Unicode code points indicates they should know
17541 * the issues involved */
17543 warn_super = ! (invert
17544 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
17547 _invlist_union(properties, cp_list, &cp_list);
17548 SvREFCNT_dec_NN(properties);
17551 cp_list = properties;
17556 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17558 /* Because an ANYOF node is the only one that warns, this node
17559 * can't be optimized into something else */
17560 optimizable = FALSE;
17564 /* Here, we have calculated what code points should be in the character
17567 * Now we can see about various optimizations. Fold calculation (which we
17568 * did above) needs to take place before inversion. Otherwise /[^k]/i
17569 * would invert to include K, which under /i would match k, which it
17570 * shouldn't. Therefore we can't invert folded locale now, as it won't be
17571 * folded until runtime */
17573 /* If we didn't do folding, it's because some information isn't available
17574 * until runtime; set the run-time fold flag for these. (We don't have to
17575 * worry about properties folding, as that is taken care of by the swash
17576 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
17577 * locales, or the class matches at least one 0-255 range code point */
17580 /* Some things on the list might be unconditionally included because of
17581 * other components. Remove them, and clean up the list if it goes to
17583 if (only_utf8_locale_list && cp_list) {
17584 _invlist_subtract(only_utf8_locale_list, cp_list,
17585 &only_utf8_locale_list);
17587 if (_invlist_len(only_utf8_locale_list) == 0) {
17588 SvREFCNT_dec_NN(only_utf8_locale_list);
17589 only_utf8_locale_list = NULL;
17592 if (only_utf8_locale_list) {
17595 |ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17597 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
17599 invlist_iterinit(cp_list);
17600 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
17601 ANYOF_FLAGS(ret) |= ANYOFL_FOLD;
17603 invlist_iterfinish(cp_list);
17606 else if ( DEPENDS_SEMANTICS
17607 && ( has_upper_latin1_only_utf8_matches
17608 || (ANYOF_FLAGS(ret) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
17611 optimizable = FALSE;
17615 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
17616 * at compile time. Besides not inverting folded locale now, we can't
17617 * invert if there are things such as \w, which aren't known until runtime
17621 && OP(ret) != ANYOFD
17622 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
17623 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17625 _invlist_invert(cp_list);
17627 /* Any swash can't be used as-is, because we've inverted things */
17629 SvREFCNT_dec_NN(swash);
17633 /* Clear the invert flag since have just done it here */
17640 *ret_invlist = cp_list;
17641 SvREFCNT_dec(swash);
17643 /* Discard the generated node */
17645 RExC_size = orig_size;
17648 RExC_emit = orig_emit;
17653 /* Some character classes are equivalent to other nodes. Such nodes take
17654 * up less room and generally fewer operations to execute than ANYOF nodes.
17655 * Above, we checked for and optimized into some such equivalents for
17656 * certain common classes that are easy to test. Getting to this point in
17657 * the code means that the class didn't get optimized there. Since this
17658 * code is only executed in Pass 2, it is too late to save space--it has
17659 * been allocated in Pass 1, and currently isn't given back. But turning
17660 * things into an EXACTish node can allow the optimizer to join it to any
17661 * adjacent such nodes. And if the class is equivalent to things like /./,
17662 * expensive run-time swashes can be avoided. Now that we have more
17663 * complete information, we can find things necessarily missed by the
17664 * earlier code. Another possible "optimization" that isn't done is that
17665 * something like [Ee] could be changed into an EXACTFU. khw tried this
17666 * and found that the ANYOF is faster, including for code points not in the
17667 * bitmap. This still might make sense to do, provided it got joined with
17668 * an adjacent node(s) to create a longer EXACTFU one. This could be
17669 * accomplished by creating a pseudo ANYOF_EXACTFU node type that the join
17670 * routine would know is joinable. If that didn't happen, the node type
17671 * could then be made a straight ANYOF */
17673 if (optimizable && cp_list && ! invert) {
17675 U8 op = END; /* The optimzation node-type */
17676 int posix_class = -1; /* Illegal value */
17677 const char * cur_parse= RExC_parse;
17679 invlist_iterinit(cp_list);
17680 if (! invlist_iternext(cp_list, &start, &end)) {
17682 /* Here, the list is empty. This happens, for example, when a
17683 * Unicode property that doesn't match anything is the only element
17684 * in the character class (perluniprops.pod notes such properties).
17687 *flagp |= HASWIDTH|SIMPLE;
17689 else if (start == end) { /* The range is a single code point */
17690 if (! invlist_iternext(cp_list, &start, &end)
17692 /* Don't do this optimization if it would require changing
17693 * the pattern to UTF-8 */
17694 && (start < 256 || UTF))
17696 /* Here, the list contains a single code point. Can optimize
17697 * into an EXACTish node */
17708 /* A locale node under folding with one code point can be
17709 * an EXACTFL, as its fold won't be calculated until
17715 /* Here, we are generally folding, but there is only one
17716 * code point to match. If we have to, we use an EXACT
17717 * node, but it would be better for joining with adjacent
17718 * nodes in the optimization pass if we used the same
17719 * EXACTFish node that any such are likely to be. We can
17720 * do this iff the code point doesn't participate in any
17721 * folds. For example, an EXACTF of a colon is the same as
17722 * an EXACT one, since nothing folds to or from a colon. */
17724 if (IS_IN_SOME_FOLD_L1(value)) {
17729 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
17734 /* If we haven't found the node type, above, it means we
17735 * can use the prevailing one */
17737 op = compute_EXACTish(pRExC_state);
17741 } /* End of first range contains just a single code point */
17742 else if (start == 0) {
17743 if (end == UV_MAX) {
17745 *flagp |= HASWIDTH|SIMPLE;
17748 else if (end == '\n' - 1
17749 && invlist_iternext(cp_list, &start, &end)
17750 && start == '\n' + 1 && end == UV_MAX)
17753 *flagp |= HASWIDTH|SIMPLE;
17757 invlist_iterfinish(cp_list);
17760 const UV cp_list_len = _invlist_len(cp_list);
17761 const UV* cp_list_array = invlist_array(cp_list);
17763 /* Here, didn't find an optimization. See if this matches any of
17764 * the POSIX classes. These run slightly faster for above-Unicode
17765 * code points, so don't bother with POSIXA ones nor the 2 that
17766 * have no above-Unicode matches. We can avoid these checks unless
17767 * the ANYOF matches at least as high as the lowest POSIX one
17768 * (which was manually found to be \v. The actual code point may
17769 * increase in later Unicode releases, if a higher code point is
17770 * assigned to be \v, but this code will never break. It would
17771 * just mean we could execute the checks for posix optimizations
17772 * unnecessarily) */
17774 if (cp_list_array[cp_list_len-1] > 0x2029) {
17775 for (posix_class = 0;
17776 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
17780 if (posix_class == _CC_ASCII || posix_class == _CC_CNTRL) {
17783 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
17785 /* Check if matches normal or inverted */
17786 if (_invlistEQ(cp_list,
17787 PL_XPosix_ptrs[posix_class],
17790 op = (try_inverted)
17793 *flagp |= HASWIDTH|SIMPLE;
17803 RExC_parse = (char *)orig_parse;
17804 RExC_emit = (regnode *)orig_emit;
17806 if (regarglen[op]) {
17807 ret = reganode(pRExC_state, op, 0);
17809 ret = reg_node(pRExC_state, op);
17812 RExC_parse = (char *)cur_parse;
17814 if (PL_regkind[op] == EXACT) {
17815 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17816 TRUE /* downgradable to EXACT */
17819 else if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17820 FLAGS(ret) = posix_class;
17823 SvREFCNT_dec_NN(cp_list);
17828 /* Here, <cp_list> contains all the code points we can determine at
17829 * compile time that match under all conditions. Go through it, and
17830 * for things that belong in the bitmap, put them there, and delete from
17831 * <cp_list>. While we are at it, see if everything above 255 is in the
17832 * list, and if so, set a flag to speed up execution */
17834 populate_ANYOF_from_invlist(ret, &cp_list);
17837 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
17840 /* Here, the bitmap has been populated with all the Latin1 code points that
17841 * always match. Can now add to the overall list those that match only
17842 * when the target string is UTF-8 (<has_upper_latin1_only_utf8_matches>).
17844 if (has_upper_latin1_only_utf8_matches) {
17846 _invlist_union(cp_list,
17847 has_upper_latin1_only_utf8_matches,
17849 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17852 cp_list = has_upper_latin1_only_utf8_matches;
17854 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
17857 /* If there is a swash and more than one element, we can't use the swash in
17858 * the optimization below. */
17859 if (swash && element_count > 1) {
17860 SvREFCNT_dec_NN(swash);
17864 /* Note that the optimization of using 'swash' if it is the only thing in
17865 * the class doesn't have us change swash at all, so it can include things
17866 * that are also in the bitmap; otherwise we have purposely deleted that
17867 * duplicate information */
17868 set_ANYOF_arg(pRExC_state, ret, cp_list,
17869 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17871 only_utf8_locale_list,
17872 swash, has_user_defined_property);
17874 *flagp |= HASWIDTH|SIMPLE;
17876 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
17877 RExC_contains_locale = 1;
17883 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
17886 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
17887 regnode* const node,
17889 SV* const runtime_defns,
17890 SV* const only_utf8_locale_list,
17892 const bool has_user_defined_property)
17894 /* Sets the arg field of an ANYOF-type node 'node', using information about
17895 * the node passed-in. If there is nothing outside the node's bitmap, the
17896 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
17897 * the count returned by add_data(), having allocated and stored an array,
17898 * av, that that count references, as follows:
17899 * av[0] stores the character class description in its textual form.
17900 * This is used later (regexec.c:Perl_regclass_swash()) to
17901 * initialize the appropriate swash, and is also useful for dumping
17902 * the regnode. This is set to &PL_sv_undef if the textual
17903 * description is not needed at run-time (as happens if the other
17904 * elements completely define the class)
17905 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
17906 * computed from av[0]. But if no further computation need be done,
17907 * the swash is stored here now (and av[0] is &PL_sv_undef).
17908 * av[2] stores the inversion list of code points that match only if the
17909 * current locale is UTF-8
17910 * av[3] stores the cp_list inversion list for use in addition or instead
17911 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
17912 * (Otherwise everything needed is already in av[0] and av[1])
17913 * av[4] is set if any component of the class is from a user-defined
17914 * property; used only if av[3] exists */
17918 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
17920 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
17921 assert(! (ANYOF_FLAGS(node)
17922 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
17923 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
17926 AV * const av = newAV();
17929 av_store(av, 0, (runtime_defns)
17930 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
17933 av_store(av, 1, swash);
17934 SvREFCNT_dec_NN(cp_list);
17937 av_store(av, 1, &PL_sv_undef);
17939 av_store(av, 3, cp_list);
17940 av_store(av, 4, newSVuv(has_user_defined_property));
17944 if (only_utf8_locale_list) {
17945 av_store(av, 2, only_utf8_locale_list);
17948 av_store(av, 2, &PL_sv_undef);
17951 rv = newRV_noinc(MUTABLE_SV(av));
17952 n = add_data(pRExC_state, STR_WITH_LEN("s"));
17953 RExC_rxi->data->data[n] = (void*)rv;
17958 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
17960 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
17961 const regnode* node,
17964 SV** only_utf8_locale_ptr,
17965 SV** output_invlist)
17968 /* For internal core use only.
17969 * Returns the swash for the input 'node' in the regex 'prog'.
17970 * If <doinit> is 'true', will attempt to create the swash if not already
17972 * If <listsvp> is non-null, will return the printable contents of the
17973 * swash. This can be used to get debugging information even before the
17974 * swash exists, by calling this function with 'doinit' set to false, in
17975 * which case the components that will be used to eventually create the
17976 * swash are returned (in a printable form).
17977 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
17978 * store an inversion list of code points that should match only if the
17979 * execution-time locale is a UTF-8 one.
17980 * If <output_invlist> is not NULL, it is where this routine is to store an
17981 * inversion list of the code points that would be instead returned in
17982 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
17983 * when this parameter is used, is just the non-code point data that
17984 * will go into creating the swash. This currently should be just
17985 * user-defined properties whose definitions were not known at compile
17986 * time. Using this parameter allows for easier manipulation of the
17987 * swash's data by the caller. It is illegal to call this function with
17988 * this parameter set, but not <listsvp>
17990 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
17991 * that, in spite of this function's name, the swash it returns may include
17992 * the bitmap data as well */
17995 SV *si = NULL; /* Input swash initialization string */
17996 SV* invlist = NULL;
17998 RXi_GET_DECL(prog,progi);
17999 const struct reg_data * const data = prog ? progi->data : NULL;
18001 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
18002 assert(! output_invlist || listsvp);
18004 if (data && data->count) {
18005 const U32 n = ARG(node);
18007 if (data->what[n] == 's') {
18008 SV * const rv = MUTABLE_SV(data->data[n]);
18009 AV * const av = MUTABLE_AV(SvRV(rv));
18010 SV **const ary = AvARRAY(av);
18011 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
18013 si = *ary; /* ary[0] = the string to initialize the swash with */
18015 if (av_tindex_nomg(av) >= 2) {
18016 if (only_utf8_locale_ptr
18018 && ary[2] != &PL_sv_undef)
18020 *only_utf8_locale_ptr = ary[2];
18023 assert(only_utf8_locale_ptr);
18024 *only_utf8_locale_ptr = NULL;
18027 /* Elements 3 and 4 are either both present or both absent. [3]
18028 * is any inversion list generated at compile time; [4]
18029 * indicates if that inversion list has any user-defined
18030 * properties in it. */
18031 if (av_tindex_nomg(av) >= 3) {
18033 if (SvUV(ary[4])) {
18034 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
18042 /* Element [1] is reserved for the set-up swash. If already there,
18043 * return it; if not, create it and store it there */
18044 if (ary[1] && SvROK(ary[1])) {
18047 else if (doinit && ((si && si != &PL_sv_undef)
18048 || (invlist && invlist != &PL_sv_undef))) {
18050 sw = _core_swash_init("utf8", /* the utf8 package */
18054 0, /* not from tr/// */
18056 &swash_init_flags);
18057 (void)av_store(av, 1, sw);
18062 /* If requested, return a printable version of what this swash matches */
18064 SV* matches_string = NULL;
18066 /* The swash should be used, if possible, to get the data, as it
18067 * contains the resolved data. But this function can be called at
18068 * compile-time, before everything gets resolved, in which case we
18069 * return the currently best available information, which is the string
18070 * that will eventually be used to do that resolving, 'si' */
18071 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
18072 && (si && si != &PL_sv_undef))
18074 /* Here, we only have 'si' (and possibly some passed-in data in
18075 * 'invlist', which is handled below) If the caller only wants
18076 * 'si', use that. */
18077 if (! output_invlist) {
18078 matches_string = newSVsv(si);
18081 /* But if the caller wants an inversion list of the node, we
18082 * need to parse 'si' and place as much as possible in the
18083 * desired output inversion list, making 'matches_string' only
18084 * contain the currently unresolvable things */
18085 const char *si_string = SvPVX(si);
18086 STRLEN remaining = SvCUR(si);
18090 /* Ignore everything before the first new-line */
18091 while (*si_string != '\n' && remaining > 0) {
18095 assert(remaining > 0);
18100 while (remaining > 0) {
18102 /* The data consists of just strings defining user-defined
18103 * property names, but in prior incarnations, and perhaps
18104 * somehow from pluggable regex engines, it could still
18105 * hold hex code point definitions. Each component of a
18106 * range would be separated by a tab, and each range by a
18107 * new-line. If these are found, instead add them to the
18108 * inversion list */
18109 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
18110 |PERL_SCAN_SILENT_NON_PORTABLE;
18111 STRLEN len = remaining;
18112 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
18114 /* If the hex decode routine found something, it should go
18115 * up to the next \n */
18116 if ( *(si_string + len) == '\n') {
18117 if (count) { /* 2nd code point on line */
18118 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
18121 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
18124 goto prepare_for_next_iteration;
18127 /* If the hex decode was instead for the lower range limit,
18128 * save it, and go parse the upper range limit */
18129 if (*(si_string + len) == '\t') {
18130 assert(count == 0);
18134 prepare_for_next_iteration:
18135 si_string += len + 1;
18136 remaining -= len + 1;
18140 /* Here, didn't find a legal hex number. Just add it from
18141 * here to the next \n */
18144 while (*(si_string + len) != '\n' && remaining > 0) {
18148 if (*(si_string + len) == '\n') {
18152 if (matches_string) {
18153 sv_catpvn(matches_string, si_string, len - 1);
18156 matches_string = newSVpvn(si_string, len - 1);
18159 sv_catpvs(matches_string, " ");
18160 } /* end of loop through the text */
18162 assert(matches_string);
18163 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
18164 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
18166 } /* end of has an 'si' but no swash */
18169 /* If we have a swash in place, its equivalent inversion list was above
18170 * placed into 'invlist'. If not, this variable may contain a stored
18171 * inversion list which is information beyond what is in 'si' */
18174 /* Again, if the caller doesn't want the output inversion list, put
18175 * everything in 'matches-string' */
18176 if (! output_invlist) {
18177 if ( ! matches_string) {
18178 matches_string = newSVpvs("\n");
18180 sv_catsv(matches_string, invlist_contents(invlist,
18181 TRUE /* traditional style */
18184 else if (! *output_invlist) {
18185 *output_invlist = invlist_clone(invlist);
18188 _invlist_union(*output_invlist, invlist, output_invlist);
18192 *listsvp = matches_string;
18197 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
18199 /* reg_skipcomment()
18201 Absorbs an /x style # comment from the input stream,
18202 returning a pointer to the first character beyond the comment, or if the
18203 comment terminates the pattern without anything following it, this returns
18204 one past the final character of the pattern (in other words, RExC_end) and
18205 sets the REG_RUN_ON_COMMENT_SEEN flag.
18207 Note it's the callers responsibility to ensure that we are
18208 actually in /x mode
18212 PERL_STATIC_INLINE char*
18213 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
18215 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
18219 while (p < RExC_end) {
18220 if (*(++p) == '\n') {
18225 /* we ran off the end of the pattern without ending the comment, so we have
18226 * to add an \n when wrapping */
18227 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
18232 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
18234 const bool force_to_xmod
18237 /* If the text at the current parse position '*p' is a '(?#...)' comment,
18238 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
18239 * is /x whitespace, advance '*p' so that on exit it points to the first
18240 * byte past all such white space and comments */
18242 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
18244 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
18246 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
18249 if (RExC_end - (*p) >= 3
18251 && *(*p + 1) == '?'
18252 && *(*p + 2) == '#')
18254 while (*(*p) != ')') {
18255 if ((*p) == RExC_end)
18256 FAIL("Sequence (?#... not terminated");
18264 const char * save_p = *p;
18265 while ((*p) < RExC_end) {
18267 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
18270 else if (*(*p) == '#') {
18271 (*p) = reg_skipcomment(pRExC_state, (*p));
18277 if (*p != save_p) {
18290 Advances the parse position by one byte, unless that byte is the beginning
18291 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
18292 those two cases, the parse position is advanced beyond all such comments and
18295 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
18299 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
18301 PERL_ARGS_ASSERT_NEXTCHAR;
18303 if (RExC_parse < RExC_end) {
18305 || UTF8_IS_INVARIANT(*RExC_parse)
18306 || UTF8_IS_START(*RExC_parse));
18308 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
18310 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
18311 FALSE /* Don't force /x */ );
18316 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
18318 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
18319 * space. In pass1, it aligns and increments RExC_size; in pass2,
18322 regnode * const ret = RExC_emit;
18323 GET_RE_DEBUG_FLAGS_DECL;
18325 PERL_ARGS_ASSERT_REGNODE_GUTS;
18327 assert(extra_size >= regarglen[op]);
18330 SIZE_ALIGN(RExC_size);
18331 RExC_size += 1 + extra_size;
18334 if (RExC_emit >= RExC_emit_bound)
18335 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
18336 op, (void*)RExC_emit, (void*)RExC_emit_bound);
18338 NODE_ALIGN_FILL(ret);
18339 #ifndef RE_TRACK_PATTERN_OFFSETS
18340 PERL_UNUSED_ARG(name);
18342 if (RExC_offsets) { /* MJD */
18344 ("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
18347 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
18348 ? "Overwriting end of array!\n" : "OK",
18349 (UV)(RExC_emit - RExC_emit_start),
18350 (UV)(RExC_parse - RExC_start),
18351 (UV)RExC_offsets[0]));
18352 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
18359 - reg_node - emit a node
18361 STATIC regnode * /* Location. */
18362 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
18364 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
18366 PERL_ARGS_ASSERT_REG_NODE;
18368 assert(regarglen[op] == 0);
18371 regnode *ptr = ret;
18372 FILL_ADVANCE_NODE(ptr, op);
18379 - reganode - emit a node with an argument
18381 STATIC regnode * /* Location. */
18382 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
18384 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
18386 PERL_ARGS_ASSERT_REGANODE;
18388 assert(regarglen[op] == 1);
18391 regnode *ptr = ret;
18392 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
18399 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
18401 /* emit a node with U32 and I32 arguments */
18403 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
18405 PERL_ARGS_ASSERT_REG2LANODE;
18407 assert(regarglen[op] == 2);
18410 regnode *ptr = ret;
18411 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
18418 - reginsert - insert an operator in front of already-emitted operand
18420 * Means relocating the operand.
18423 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
18428 const int offset = regarglen[(U8)op];
18429 const int size = NODE_STEP_REGNODE + offset;
18430 GET_RE_DEBUG_FLAGS_DECL;
18432 PERL_ARGS_ASSERT_REGINSERT;
18433 PERL_UNUSED_CONTEXT;
18434 PERL_UNUSED_ARG(depth);
18435 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
18436 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
18441 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
18442 studying. If this is wrong then we need to adjust RExC_recurse
18443 below like we do with RExC_open_parens/RExC_close_parens. */
18447 if (RExC_open_parens) {
18449 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
18450 /* remember that RExC_npar is rex->nparens + 1,
18451 * iow it is 1 more than the number of parens seen in
18452 * the pattern so far. */
18453 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
18454 /* note, RExC_open_parens[0] is the start of the
18455 * regex, it can't move. RExC_close_parens[0] is the end
18456 * of the regex, it *can* move. */
18457 if ( paren && RExC_open_parens[paren] >= opnd ) {
18458 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
18459 RExC_open_parens[paren] += size;
18461 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
18463 if ( RExC_close_parens[paren] >= opnd ) {
18464 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
18465 RExC_close_parens[paren] += size;
18467 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
18472 RExC_end_op += size;
18474 while (src > opnd) {
18475 StructCopy(--src, --dst, regnode);
18476 #ifdef RE_TRACK_PATTERN_OFFSETS
18477 if (RExC_offsets) { /* MJD 20010112 */
18479 ("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
18483 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
18484 ? "Overwriting end of array!\n" : "OK",
18485 (UV)(src - RExC_emit_start),
18486 (UV)(dst - RExC_emit_start),
18487 (UV)RExC_offsets[0]));
18488 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
18489 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
18495 place = opnd; /* Op node, where operand used to be. */
18496 #ifdef RE_TRACK_PATTERN_OFFSETS
18497 if (RExC_offsets) { /* MJD */
18499 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
18503 (UV)(place - RExC_emit_start) > RExC_offsets[0]
18504 ? "Overwriting end of array!\n" : "OK",
18505 (UV)(place - RExC_emit_start),
18506 (UV)(RExC_parse - RExC_start),
18507 (UV)RExC_offsets[0]));
18508 Set_Node_Offset(place, RExC_parse);
18509 Set_Node_Length(place, 1);
18512 src = NEXTOPER(place);
18513 FILL_ADVANCE_NODE(place, op);
18514 Zero(src, offset, regnode);
18518 - regtail - set the next-pointer at the end of a node chain of p to val.
18519 - SEE ALSO: regtail_study
18522 S_regtail(pTHX_ RExC_state_t * pRExC_state,
18523 const regnode * const p,
18524 const regnode * const val,
18528 GET_RE_DEBUG_FLAGS_DECL;
18530 PERL_ARGS_ASSERT_REGTAIL;
18532 PERL_UNUSED_ARG(depth);
18538 /* Find last node. */
18539 scan = (regnode *) p;
18541 regnode * const temp = regnext(scan);
18543 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
18544 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18545 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
18546 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
18547 (temp == NULL ? "->" : ""),
18548 (temp == NULL ? PL_reg_name[OP(val)] : "")
18556 if (reg_off_by_arg[OP(scan)]) {
18557 ARG_SET(scan, val - scan);
18560 NEXT_OFF(scan) = val - scan;
18566 - regtail_study - set the next-pointer at the end of a node chain of p to val.
18567 - Look for optimizable sequences at the same time.
18568 - currently only looks for EXACT chains.
18570 This is experimental code. The idea is to use this routine to perform
18571 in place optimizations on branches and groups as they are constructed,
18572 with the long term intention of removing optimization from study_chunk so
18573 that it is purely analytical.
18575 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
18576 to control which is which.
18579 /* TODO: All four parms should be const */
18582 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
18583 const regnode *val,U32 depth)
18587 #ifdef EXPERIMENTAL_INPLACESCAN
18590 GET_RE_DEBUG_FLAGS_DECL;
18592 PERL_ARGS_ASSERT_REGTAIL_STUDY;
18598 /* Find last node. */
18602 regnode * const temp = regnext(scan);
18603 #ifdef EXPERIMENTAL_INPLACESCAN
18604 if (PL_regkind[OP(scan)] == EXACT) {
18605 bool unfolded_multi_char; /* Unexamined in this routine */
18606 if (join_exact(pRExC_state, scan, &min,
18607 &unfolded_multi_char, 1, val, depth+1))
18612 switch (OP(scan)) {
18616 case EXACTFA_NO_TRIE:
18622 if( exact == PSEUDO )
18624 else if ( exact != OP(scan) )
18633 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
18634 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18635 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
18636 SvPV_nolen_const(RExC_mysv),
18637 REG_NODE_NUM(scan),
18638 PL_reg_name[exact]);
18645 DEBUG_PARSE_MSG("");
18646 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
18647 Perl_re_printf( aTHX_
18648 "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
18649 SvPV_nolen_const(RExC_mysv),
18650 (IV)REG_NODE_NUM(val),
18654 if (reg_off_by_arg[OP(scan)]) {
18655 ARG_SET(scan, val - scan);
18658 NEXT_OFF(scan) = val - scan;
18666 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
18671 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
18676 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18678 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
18679 if (flags & (1<<bit)) {
18680 if (!set++ && lead)
18681 Perl_re_printf( aTHX_ "%s",lead);
18682 Perl_re_printf( aTHX_ "%s ",PL_reg_intflags_name[bit]);
18687 Perl_re_printf( aTHX_ "\n");
18689 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18694 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
18700 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18702 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
18703 if (flags & (1<<bit)) {
18704 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
18707 if (!set++ && lead)
18708 Perl_re_printf( aTHX_ "%s",lead);
18709 Perl_re_printf( aTHX_ "%s ",PL_reg_extflags_name[bit]);
18712 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
18713 if (!set++ && lead) {
18714 Perl_re_printf( aTHX_ "%s",lead);
18717 case REGEX_UNICODE_CHARSET:
18718 Perl_re_printf( aTHX_ "UNICODE");
18720 case REGEX_LOCALE_CHARSET:
18721 Perl_re_printf( aTHX_ "LOCALE");
18723 case REGEX_ASCII_RESTRICTED_CHARSET:
18724 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
18726 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
18727 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
18730 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
18736 Perl_re_printf( aTHX_ "\n");
18738 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18744 Perl_regdump(pTHX_ const regexp *r)
18747 SV * const sv = sv_newmortal();
18748 SV *dsv= sv_newmortal();
18749 RXi_GET_DECL(r,ri);
18750 GET_RE_DEBUG_FLAGS_DECL;
18752 PERL_ARGS_ASSERT_REGDUMP;
18754 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
18756 /* Header fields of interest. */
18757 if (r->anchored_substr) {
18758 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
18759 RE_SV_DUMPLEN(r->anchored_substr), 30);
18760 Perl_re_printf( aTHX_
18761 "anchored %s%s at %"IVdf" ",
18762 s, RE_SV_TAIL(r->anchored_substr),
18763 (IV)r->anchored_offset);
18764 } else if (r->anchored_utf8) {
18765 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
18766 RE_SV_DUMPLEN(r->anchored_utf8), 30);
18767 Perl_re_printf( aTHX_
18768 "anchored utf8 %s%s at %"IVdf" ",
18769 s, RE_SV_TAIL(r->anchored_utf8),
18770 (IV)r->anchored_offset);
18772 if (r->float_substr) {
18773 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
18774 RE_SV_DUMPLEN(r->float_substr), 30);
18775 Perl_re_printf( aTHX_
18776 "floating %s%s at %"IVdf"..%"UVuf" ",
18777 s, RE_SV_TAIL(r->float_substr),
18778 (IV)r->float_min_offset, (UV)r->float_max_offset);
18779 } else if (r->float_utf8) {
18780 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
18781 RE_SV_DUMPLEN(r->float_utf8), 30);
18782 Perl_re_printf( aTHX_
18783 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
18784 s, RE_SV_TAIL(r->float_utf8),
18785 (IV)r->float_min_offset, (UV)r->float_max_offset);
18787 if (r->check_substr || r->check_utf8)
18788 Perl_re_printf( aTHX_
18790 (r->check_substr == r->float_substr
18791 && r->check_utf8 == r->float_utf8
18792 ? "(checking floating" : "(checking anchored"));
18793 if (r->intflags & PREGf_NOSCAN)
18794 Perl_re_printf( aTHX_ " noscan");
18795 if (r->extflags & RXf_CHECK_ALL)
18796 Perl_re_printf( aTHX_ " isall");
18797 if (r->check_substr || r->check_utf8)
18798 Perl_re_printf( aTHX_ ") ");
18800 if (ri->regstclass) {
18801 regprop(r, sv, ri->regstclass, NULL, NULL);
18802 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
18804 if (r->intflags & PREGf_ANCH) {
18805 Perl_re_printf( aTHX_ "anchored");
18806 if (r->intflags & PREGf_ANCH_MBOL)
18807 Perl_re_printf( aTHX_ "(MBOL)");
18808 if (r->intflags & PREGf_ANCH_SBOL)
18809 Perl_re_printf( aTHX_ "(SBOL)");
18810 if (r->intflags & PREGf_ANCH_GPOS)
18811 Perl_re_printf( aTHX_ "(GPOS)");
18812 Perl_re_printf( aTHX_ " ");
18814 if (r->intflags & PREGf_GPOS_SEEN)
18815 Perl_re_printf( aTHX_ "GPOS:%"UVuf" ", (UV)r->gofs);
18816 if (r->intflags & PREGf_SKIP)
18817 Perl_re_printf( aTHX_ "plus ");
18818 if (r->intflags & PREGf_IMPLICIT)
18819 Perl_re_printf( aTHX_ "implicit ");
18820 Perl_re_printf( aTHX_ "minlen %"IVdf" ", (IV)r->minlen);
18821 if (r->extflags & RXf_EVAL_SEEN)
18822 Perl_re_printf( aTHX_ "with eval ");
18823 Perl_re_printf( aTHX_ "\n");
18825 regdump_extflags("r->extflags: ",r->extflags);
18826 regdump_intflags("r->intflags: ",r->intflags);
18829 PERL_ARGS_ASSERT_REGDUMP;
18830 PERL_UNUSED_CONTEXT;
18831 PERL_UNUSED_ARG(r);
18832 #endif /* DEBUGGING */
18835 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
18838 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
18839 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
18840 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
18841 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
18842 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
18843 || _CC_VERTSPACE != 15
18844 # error Need to adjust order of anyofs[]
18846 static const char * const anyofs[] = {
18883 - regprop - printable representation of opcode, with run time support
18887 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
18891 RXi_GET_DECL(prog,progi);
18892 GET_RE_DEBUG_FLAGS_DECL;
18894 PERL_ARGS_ASSERT_REGPROP;
18898 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
18899 /* It would be nice to FAIL() here, but this may be called from
18900 regexec.c, and it would be hard to supply pRExC_state. */
18901 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
18902 (int)OP(o), (int)REGNODE_MAX);
18903 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
18905 k = PL_regkind[OP(o)];
18908 sv_catpvs(sv, " ");
18909 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
18910 * is a crude hack but it may be the best for now since
18911 * we have no flag "this EXACTish node was UTF-8"
18913 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
18914 PERL_PV_ESCAPE_UNI_DETECT |
18915 PERL_PV_ESCAPE_NONASCII |
18916 PERL_PV_PRETTY_ELLIPSES |
18917 PERL_PV_PRETTY_LTGT |
18918 PERL_PV_PRETTY_NOCLEAR
18920 } else if (k == TRIE) {
18921 /* print the details of the trie in dumpuntil instead, as
18922 * progi->data isn't available here */
18923 const char op = OP(o);
18924 const U32 n = ARG(o);
18925 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
18926 (reg_ac_data *)progi->data->data[n] :
18928 const reg_trie_data * const trie
18929 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
18931 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
18932 DEBUG_TRIE_COMPILE_r(
18933 Perl_sv_catpvf(aTHX_ sv,
18934 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
18935 (UV)trie->startstate,
18936 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
18937 (UV)trie->wordcount,
18940 (UV)TRIE_CHARCOUNT(trie),
18941 (UV)trie->uniquecharcount
18944 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
18945 sv_catpvs(sv, "[");
18946 (void) put_charclass_bitmap_innards(sv,
18947 ((IS_ANYOF_TRIE(op))
18949 : TRIE_BITMAP(trie)),
18955 sv_catpvs(sv, "]");
18958 } else if (k == CURLY) {
18959 U32 lo = ARG1(o), hi = ARG2(o);
18960 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
18961 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
18962 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
18963 if (hi == REG_INFTY)
18964 sv_catpvs(sv, "INFTY");
18966 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
18967 sv_catpvs(sv, "}");
18969 else if (k == WHILEM && o->flags) /* Ordinal/of */
18970 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
18971 else if (k == REF || k == OPEN || k == CLOSE
18972 || k == GROUPP || OP(o)==ACCEPT)
18974 AV *name_list= NULL;
18975 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
18976 Perl_sv_catpvf(aTHX_ sv, "%"UVuf, (UV)parno); /* Parenth number */
18977 if ( RXp_PAREN_NAMES(prog) ) {
18978 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
18979 } else if ( pRExC_state ) {
18980 name_list= RExC_paren_name_list;
18983 if ( k != REF || (OP(o) < NREF)) {
18984 SV **name= av_fetch(name_list, parno, 0 );
18986 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
18989 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
18990 I32 *nums=(I32*)SvPVX(sv_dat);
18991 SV **name= av_fetch(name_list, nums[0], 0 );
18994 for ( n=0; n<SvIVX(sv_dat); n++ ) {
18995 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
18996 (n ? "," : ""), (IV)nums[n]);
18998 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
19002 if ( k == REF && reginfo) {
19003 U32 n = ARG(o); /* which paren pair */
19004 I32 ln = prog->offs[n].start;
19005 if (prog->lastparen < n || ln == -1)
19006 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
19007 else if (ln == prog->offs[n].end)
19008 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
19010 const char *s = reginfo->strbeg + ln;
19011 Perl_sv_catpvf(aTHX_ sv, ": ");
19012 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
19013 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
19016 } else if (k == GOSUB) {
19017 AV *name_list= NULL;
19018 if ( RXp_PAREN_NAMES(prog) ) {
19019 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19020 } else if ( pRExC_state ) {
19021 name_list= RExC_paren_name_list;
19024 /* Paren and offset */
19025 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
19026 (int)((o + (int)ARG2L(o)) - progi->program) );
19028 SV **name= av_fetch(name_list, ARG(o), 0 );
19030 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
19033 else if (k == LOGICAL)
19034 /* 2: embedded, otherwise 1 */
19035 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
19036 else if (k == ANYOF) {
19037 const U8 flags = ANYOF_FLAGS(o);
19038 bool do_sep = FALSE; /* Do we need to separate various components of
19040 /* Set if there is still an unresolved user-defined property */
19041 SV *unresolved = NULL;
19043 /* Things that are ignored except when the runtime locale is UTF-8 */
19044 SV *only_utf8_locale_invlist = NULL;
19046 /* Code points that don't fit in the bitmap */
19047 SV *nonbitmap_invlist = NULL;
19049 /* And things that aren't in the bitmap, but are small enough to be */
19050 SV* bitmap_range_not_in_bitmap = NULL;
19052 const bool inverted = flags & ANYOF_INVERT;
19054 if (OP(o) == ANYOFL) {
19055 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
19056 sv_catpvs(sv, "{utf8-locale-reqd}");
19058 if (flags & ANYOFL_FOLD) {
19059 sv_catpvs(sv, "{i}");
19063 /* If there is stuff outside the bitmap, get it */
19064 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
19065 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
19067 &only_utf8_locale_invlist,
19068 &nonbitmap_invlist);
19069 /* The non-bitmap data may contain stuff that could fit in the
19070 * bitmap. This could come from a user-defined property being
19071 * finally resolved when this call was done; or much more likely
19072 * because there are matches that require UTF-8 to be valid, and so
19073 * aren't in the bitmap. This is teased apart later */
19074 _invlist_intersection(nonbitmap_invlist,
19076 &bitmap_range_not_in_bitmap);
19077 /* Leave just the things that don't fit into the bitmap */
19078 _invlist_subtract(nonbitmap_invlist,
19080 &nonbitmap_invlist);
19083 /* Obey this flag to add all above-the-bitmap code points */
19084 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
19085 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
19086 NUM_ANYOF_CODE_POINTS,
19090 /* Ready to start outputting. First, the initial left bracket */
19091 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
19093 /* Then all the things that could fit in the bitmap */
19094 do_sep = put_charclass_bitmap_innards(sv,
19096 bitmap_range_not_in_bitmap,
19097 only_utf8_locale_invlist,
19100 /* Can't try inverting for a
19101 * better display if there are
19102 * things that haven't been
19104 unresolved != NULL);
19105 SvREFCNT_dec(bitmap_range_not_in_bitmap);
19107 /* If there are user-defined properties which haven't been defined yet,
19108 * output them. If the result is not to be inverted, it is clearest to
19109 * output them in a separate [] from the bitmap range stuff. If the
19110 * result is to be complemented, we have to show everything in one [],
19111 * as the inversion applies to the whole thing. Use {braces} to
19112 * separate them from anything in the bitmap and anything above the
19116 if (! do_sep) { /* If didn't output anything in the bitmap */
19117 sv_catpvs(sv, "^");
19119 sv_catpvs(sv, "{");
19122 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19124 sv_catsv(sv, unresolved);
19126 sv_catpvs(sv, "}");
19128 do_sep = ! inverted;
19131 /* And, finally, add the above-the-bitmap stuff */
19132 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
19135 /* See if truncation size is overridden */
19136 const STRLEN dump_len = (PL_dump_re_max_len)
19137 ? PL_dump_re_max_len
19140 /* This is output in a separate [] */
19142 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19145 /* And, for easy of understanding, it is shown in the
19146 * uncomplemented form if possible. The one exception being if
19147 * there are unresolved items, where the inversion has to be
19148 * delayed until runtime */
19149 if (inverted && ! unresolved) {
19150 _invlist_invert(nonbitmap_invlist);
19151 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
19154 contents = invlist_contents(nonbitmap_invlist,
19155 FALSE /* output suitable for catsv */
19158 /* If the output is shorter than the permissible maximum, just do it. */
19159 if (SvCUR(contents) <= dump_len) {
19160 sv_catsv(sv, contents);
19163 const char * contents_string = SvPVX(contents);
19164 STRLEN i = dump_len;
19166 /* Otherwise, start at the permissible max and work back to the
19167 * first break possibility */
19168 while (i > 0 && contents_string[i] != ' ') {
19171 if (i == 0) { /* Fail-safe. Use the max if we couldn't
19172 find a legal break */
19176 sv_catpvn(sv, contents_string, i);
19177 sv_catpvs(sv, "...");
19180 SvREFCNT_dec_NN(contents);
19181 SvREFCNT_dec_NN(nonbitmap_invlist);
19184 /* And finally the matching, closing ']' */
19185 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
19187 SvREFCNT_dec(unresolved);
19189 else if (k == POSIXD || k == NPOSIXD) {
19190 U8 index = FLAGS(o) * 2;
19191 if (index < C_ARRAY_LENGTH(anyofs)) {
19192 if (*anyofs[index] != '[') {
19195 sv_catpv(sv, anyofs[index]);
19196 if (*anyofs[index] != '[') {
19201 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
19204 else if (k == BOUND || k == NBOUND) {
19205 /* Must be synced with order of 'bound_type' in regcomp.h */
19206 const char * const bounds[] = {
19207 "", /* Traditional */
19213 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
19214 sv_catpv(sv, bounds[FLAGS(o)]);
19216 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
19217 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
19218 else if (OP(o) == SBOL)
19219 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
19221 /* add on the verb argument if there is one */
19222 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
19223 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
19224 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
19227 PERL_UNUSED_CONTEXT;
19228 PERL_UNUSED_ARG(sv);
19229 PERL_UNUSED_ARG(o);
19230 PERL_UNUSED_ARG(prog);
19231 PERL_UNUSED_ARG(reginfo);
19232 PERL_UNUSED_ARG(pRExC_state);
19233 #endif /* DEBUGGING */
19239 Perl_re_intuit_string(pTHX_ REGEXP * const r)
19240 { /* Assume that RE_INTUIT is set */
19241 struct regexp *const prog = ReANY(r);
19242 GET_RE_DEBUG_FLAGS_DECL;
19244 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
19245 PERL_UNUSED_CONTEXT;
19249 const char * const s = SvPV_nolen_const(RX_UTF8(r)
19250 ? prog->check_utf8 : prog->check_substr);
19252 if (!PL_colorset) reginitcolors();
19253 Perl_re_printf( aTHX_
19254 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
19256 RX_UTF8(r) ? "utf8 " : "",
19257 PL_colors[5],PL_colors[0],
19260 (strlen(s) > 60 ? "..." : ""));
19263 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
19264 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
19270 handles refcounting and freeing the perl core regexp structure. When
19271 it is necessary to actually free the structure the first thing it
19272 does is call the 'free' method of the regexp_engine associated to
19273 the regexp, allowing the handling of the void *pprivate; member
19274 first. (This routine is not overridable by extensions, which is why
19275 the extensions free is called first.)
19277 See regdupe and regdupe_internal if you change anything here.
19279 #ifndef PERL_IN_XSUB_RE
19281 Perl_pregfree(pTHX_ REGEXP *r)
19287 Perl_pregfree2(pTHX_ REGEXP *rx)
19289 struct regexp *const r = ReANY(rx);
19290 GET_RE_DEBUG_FLAGS_DECL;
19292 PERL_ARGS_ASSERT_PREGFREE2;
19294 if (r->mother_re) {
19295 ReREFCNT_dec(r->mother_re);
19297 CALLREGFREE_PVT(rx); /* free the private data */
19298 SvREFCNT_dec(RXp_PAREN_NAMES(r));
19299 Safefree(r->xpv_len_u.xpvlenu_pv);
19302 SvREFCNT_dec(r->anchored_substr);
19303 SvREFCNT_dec(r->anchored_utf8);
19304 SvREFCNT_dec(r->float_substr);
19305 SvREFCNT_dec(r->float_utf8);
19306 Safefree(r->substrs);
19308 RX_MATCH_COPY_FREE(rx);
19309 #ifdef PERL_ANY_COW
19310 SvREFCNT_dec(r->saved_copy);
19313 SvREFCNT_dec(r->qr_anoncv);
19314 if (r->recurse_locinput)
19315 Safefree(r->recurse_locinput);
19316 rx->sv_u.svu_rx = 0;
19321 This is a hacky workaround to the structural issue of match results
19322 being stored in the regexp structure which is in turn stored in
19323 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
19324 could be PL_curpm in multiple contexts, and could require multiple
19325 result sets being associated with the pattern simultaneously, such
19326 as when doing a recursive match with (??{$qr})
19328 The solution is to make a lightweight copy of the regexp structure
19329 when a qr// is returned from the code executed by (??{$qr}) this
19330 lightweight copy doesn't actually own any of its data except for
19331 the starp/end and the actual regexp structure itself.
19337 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
19339 struct regexp *ret;
19340 struct regexp *const r = ReANY(rx);
19341 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
19343 PERL_ARGS_ASSERT_REG_TEMP_COPY;
19346 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
19348 SvOK_off((SV *)ret_x);
19350 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
19351 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
19352 made both spots point to the same regexp body.) */
19353 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
19354 assert(!SvPVX(ret_x));
19355 ret_x->sv_u.svu_rx = temp->sv_any;
19356 temp->sv_any = NULL;
19357 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
19358 SvREFCNT_dec_NN(temp);
19359 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
19360 ing below will not set it. */
19361 SvCUR_set(ret_x, SvCUR(rx));
19364 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
19365 sv_force_normal(sv) is called. */
19367 ret = ReANY(ret_x);
19369 SvFLAGS(ret_x) |= SvUTF8(rx);
19370 /* We share the same string buffer as the original regexp, on which we
19371 hold a reference count, incremented when mother_re is set below.
19372 The string pointer is copied here, being part of the regexp struct.
19374 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
19375 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
19377 const I32 npar = r->nparens+1;
19378 Newx(ret->offs, npar, regexp_paren_pair);
19379 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19382 Newx(ret->substrs, 1, struct reg_substr_data);
19383 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19385 SvREFCNT_inc_void(ret->anchored_substr);
19386 SvREFCNT_inc_void(ret->anchored_utf8);
19387 SvREFCNT_inc_void(ret->float_substr);
19388 SvREFCNT_inc_void(ret->float_utf8);
19390 /* check_substr and check_utf8, if non-NULL, point to either their
19391 anchored or float namesakes, and don't hold a second reference. */
19393 RX_MATCH_COPIED_off(ret_x);
19394 #ifdef PERL_ANY_COW
19395 ret->saved_copy = NULL;
19397 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
19398 SvREFCNT_inc_void(ret->qr_anoncv);
19399 if (r->recurse_locinput)
19400 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19406 /* regfree_internal()
19408 Free the private data in a regexp. This is overloadable by
19409 extensions. Perl takes care of the regexp structure in pregfree(),
19410 this covers the *pprivate pointer which technically perl doesn't
19411 know about, however of course we have to handle the
19412 regexp_internal structure when no extension is in use.
19414 Note this is called before freeing anything in the regexp
19419 Perl_regfree_internal(pTHX_ REGEXP * const rx)
19421 struct regexp *const r = ReANY(rx);
19422 RXi_GET_DECL(r,ri);
19423 GET_RE_DEBUG_FLAGS_DECL;
19425 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
19431 SV *dsv= sv_newmortal();
19432 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
19433 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
19434 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
19435 PL_colors[4],PL_colors[5],s);
19438 #ifdef RE_TRACK_PATTERN_OFFSETS
19440 Safefree(ri->u.offsets); /* 20010421 MJD */
19442 if (ri->code_blocks) {
19444 for (n = 0; n < ri->num_code_blocks; n++)
19445 SvREFCNT_dec(ri->code_blocks[n].src_regex);
19446 Safefree(ri->code_blocks);
19450 int n = ri->data->count;
19453 /* If you add a ->what type here, update the comment in regcomp.h */
19454 switch (ri->data->what[n]) {
19460 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
19463 Safefree(ri->data->data[n]);
19469 { /* Aho Corasick add-on structure for a trie node.
19470 Used in stclass optimization only */
19472 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
19473 #ifdef USE_ITHREADS
19477 refcount = --aho->refcount;
19480 PerlMemShared_free(aho->states);
19481 PerlMemShared_free(aho->fail);
19482 /* do this last!!!! */
19483 PerlMemShared_free(ri->data->data[n]);
19484 /* we should only ever get called once, so
19485 * assert as much, and also guard the free
19486 * which /might/ happen twice. At the least
19487 * it will make code anlyzers happy and it
19488 * doesn't cost much. - Yves */
19489 assert(ri->regstclass);
19490 if (ri->regstclass) {
19491 PerlMemShared_free(ri->regstclass);
19492 ri->regstclass = 0;
19499 /* trie structure. */
19501 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
19502 #ifdef USE_ITHREADS
19506 refcount = --trie->refcount;
19509 PerlMemShared_free(trie->charmap);
19510 PerlMemShared_free(trie->states);
19511 PerlMemShared_free(trie->trans);
19513 PerlMemShared_free(trie->bitmap);
19515 PerlMemShared_free(trie->jump);
19516 PerlMemShared_free(trie->wordinfo);
19517 /* do this last!!!! */
19518 PerlMemShared_free(ri->data->data[n]);
19523 Perl_croak(aTHX_ "panic: regfree data code '%c'",
19524 ri->data->what[n]);
19527 Safefree(ri->data->what);
19528 Safefree(ri->data);
19534 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
19535 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
19536 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
19539 re_dup_guts - duplicate a regexp.
19541 This routine is expected to clone a given regexp structure. It is only
19542 compiled under USE_ITHREADS.
19544 After all of the core data stored in struct regexp is duplicated
19545 the regexp_engine.dupe method is used to copy any private data
19546 stored in the *pprivate pointer. This allows extensions to handle
19547 any duplication it needs to do.
19549 See pregfree() and regfree_internal() if you change anything here.
19551 #if defined(USE_ITHREADS)
19552 #ifndef PERL_IN_XSUB_RE
19554 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
19558 const struct regexp *r = ReANY(sstr);
19559 struct regexp *ret = ReANY(dstr);
19561 PERL_ARGS_ASSERT_RE_DUP_GUTS;
19563 npar = r->nparens+1;
19564 Newx(ret->offs, npar, regexp_paren_pair);
19565 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19567 if (ret->substrs) {
19568 /* Do it this way to avoid reading from *r after the StructCopy().
19569 That way, if any of the sv_dup_inc()s dislodge *r from the L1
19570 cache, it doesn't matter. */
19571 const bool anchored = r->check_substr
19572 ? r->check_substr == r->anchored_substr
19573 : r->check_utf8 == r->anchored_utf8;
19574 Newx(ret->substrs, 1, struct reg_substr_data);
19575 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19577 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
19578 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
19579 ret->float_substr = sv_dup_inc(ret->float_substr, param);
19580 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
19582 /* check_substr and check_utf8, if non-NULL, point to either their
19583 anchored or float namesakes, and don't hold a second reference. */
19585 if (ret->check_substr) {
19587 assert(r->check_utf8 == r->anchored_utf8);
19588 ret->check_substr = ret->anchored_substr;
19589 ret->check_utf8 = ret->anchored_utf8;
19591 assert(r->check_substr == r->float_substr);
19592 assert(r->check_utf8 == r->float_utf8);
19593 ret->check_substr = ret->float_substr;
19594 ret->check_utf8 = ret->float_utf8;
19596 } else if (ret->check_utf8) {
19598 ret->check_utf8 = ret->anchored_utf8;
19600 ret->check_utf8 = ret->float_utf8;
19605 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
19606 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
19607 if (r->recurse_locinput)
19608 Newxz(ret->recurse_locinput,r->nparens + 1,char *);
19611 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
19613 if (RX_MATCH_COPIED(dstr))
19614 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
19616 ret->subbeg = NULL;
19617 #ifdef PERL_ANY_COW
19618 ret->saved_copy = NULL;
19621 /* Whether mother_re be set or no, we need to copy the string. We
19622 cannot refrain from copying it when the storage points directly to
19623 our mother regexp, because that's
19624 1: a buffer in a different thread
19625 2: something we no longer hold a reference on
19626 so we need to copy it locally. */
19627 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
19628 ret->mother_re = NULL;
19630 #endif /* PERL_IN_XSUB_RE */
19635 This is the internal complement to regdupe() which is used to copy
19636 the structure pointed to by the *pprivate pointer in the regexp.
19637 This is the core version of the extension overridable cloning hook.
19638 The regexp structure being duplicated will be copied by perl prior
19639 to this and will be provided as the regexp *r argument, however
19640 with the /old/ structures pprivate pointer value. Thus this routine
19641 may override any copying normally done by perl.
19643 It returns a pointer to the new regexp_internal structure.
19647 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
19650 struct regexp *const r = ReANY(rx);
19651 regexp_internal *reti;
19653 RXi_GET_DECL(r,ri);
19655 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
19659 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
19660 char, regexp_internal);
19661 Copy(ri->program, reti->program, len+1, regnode);
19664 reti->num_code_blocks = ri->num_code_blocks;
19665 if (ri->code_blocks) {
19667 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
19668 struct reg_code_block);
19669 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
19670 struct reg_code_block);
19671 for (n = 0; n < ri->num_code_blocks; n++)
19672 reti->code_blocks[n].src_regex = (REGEXP*)
19673 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
19676 reti->code_blocks = NULL;
19678 reti->regstclass = NULL;
19681 struct reg_data *d;
19682 const int count = ri->data->count;
19685 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
19686 char, struct reg_data);
19687 Newx(d->what, count, U8);
19690 for (i = 0; i < count; i++) {
19691 d->what[i] = ri->data->what[i];
19692 switch (d->what[i]) {
19693 /* see also regcomp.h and regfree_internal() */
19694 case 'a': /* actually an AV, but the dup function is identical. */
19698 case 'u': /* actually an HV, but the dup function is identical. */
19699 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
19702 /* This is cheating. */
19703 Newx(d->data[i], 1, regnode_ssc);
19704 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
19705 reti->regstclass = (regnode*)d->data[i];
19708 /* Trie stclasses are readonly and can thus be shared
19709 * without duplication. We free the stclass in pregfree
19710 * when the corresponding reg_ac_data struct is freed.
19712 reti->regstclass= ri->regstclass;
19716 ((reg_trie_data*)ri->data->data[i])->refcount++;
19721 d->data[i] = ri->data->data[i];
19724 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
19725 ri->data->what[i]);
19734 reti->name_list_idx = ri->name_list_idx;
19736 #ifdef RE_TRACK_PATTERN_OFFSETS
19737 if (ri->u.offsets) {
19738 Newx(reti->u.offsets, 2*len+1, U32);
19739 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
19742 SetProgLen(reti,len);
19745 return (void*)reti;
19748 #endif /* USE_ITHREADS */
19750 #ifndef PERL_IN_XSUB_RE
19753 - regnext - dig the "next" pointer out of a node
19756 Perl_regnext(pTHX_ regnode *p)
19763 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
19764 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
19765 (int)OP(p), (int)REGNODE_MAX);
19768 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
19777 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
19780 STRLEN l1 = strlen(pat1);
19781 STRLEN l2 = strlen(pat2);
19784 const char *message;
19786 PERL_ARGS_ASSERT_RE_CROAK2;
19792 Copy(pat1, buf, l1 , char);
19793 Copy(pat2, buf + l1, l2 , char);
19794 buf[l1 + l2] = '\n';
19795 buf[l1 + l2 + 1] = '\0';
19796 va_start(args, pat2);
19797 msv = vmess(buf, &args);
19799 message = SvPV_const(msv,l1);
19802 Copy(message, buf, l1 , char);
19803 /* l1-1 to avoid \n */
19804 Perl_croak(aTHX_ "%"UTF8f, UTF8fARG(utf8, l1-1, buf));
19807 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
19809 #ifndef PERL_IN_XSUB_RE
19811 Perl_save_re_context(pTHX)
19816 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
19819 const REGEXP * const rx = PM_GETRE(PL_curpm);
19821 nparens = RX_NPARENS(rx);
19824 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
19825 * that PL_curpm will be null, but that utf8.pm and the modules it
19826 * loads will only use $1..$3.
19827 * The t/porting/re_context.t test file checks this assumption.
19832 for (i = 1; i <= nparens; i++) {
19833 char digits[TYPE_CHARS(long)];
19834 const STRLEN len = my_snprintf(digits, sizeof(digits),
19836 GV *const *const gvp
19837 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
19840 GV * const gv = *gvp;
19841 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
19851 S_put_code_point(pTHX_ SV *sv, UV c)
19853 PERL_ARGS_ASSERT_PUT_CODE_POINT;
19856 Perl_sv_catpvf(aTHX_ sv, "\\x{%04"UVXf"}", c);
19858 else if (isPRINT(c)) {
19859 const char string = (char) c;
19861 /* We use {phrase} as metanotation in the class, so also escape literal
19863 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
19864 sv_catpvs(sv, "\\");
19865 sv_catpvn(sv, &string, 1);
19867 else if (isMNEMONIC_CNTRL(c)) {
19868 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
19871 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
19875 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
19878 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
19880 /* Appends to 'sv' a displayable version of the range of code points from
19881 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
19882 * that have them, when they occur at the beginning or end of the range.
19883 * It uses hex to output the remaining code points, unless 'allow_literals'
19884 * is true, in which case the printable ASCII ones are output as-is (though
19885 * some of these will be escaped by put_code_point()).
19887 * NOTE: This is designed only for printing ranges of code points that fit
19888 * inside an ANYOF bitmap. Higher code points are simply suppressed
19891 const unsigned int min_range_count = 3;
19893 assert(start <= end);
19895 PERL_ARGS_ASSERT_PUT_RANGE;
19897 while (start <= end) {
19899 const char * format;
19901 if (end - start < min_range_count) {
19903 /* Output chars individually when they occur in short ranges */
19904 for (; start <= end; start++) {
19905 put_code_point(sv, start);
19910 /* If permitted by the input options, and there is a possibility that
19911 * this range contains a printable literal, look to see if there is
19913 if (allow_literals && start <= MAX_PRINT_A) {
19915 /* If the character at the beginning of the range isn't an ASCII
19916 * printable, effectively split the range into two parts:
19917 * 1) the portion before the first such printable,
19919 * and output them separately. */
19920 if (! isPRINT_A(start)) {
19921 UV temp_end = start + 1;
19923 /* There is no point looking beyond the final possible
19924 * printable, in MAX_PRINT_A */
19925 UV max = MIN(end, MAX_PRINT_A);
19927 while (temp_end <= max && ! isPRINT_A(temp_end)) {
19931 /* Here, temp_end points to one beyond the first printable if
19932 * found, or to one beyond 'max' if not. If none found, make
19933 * sure that we use the entire range */
19934 if (temp_end > MAX_PRINT_A) {
19935 temp_end = end + 1;
19938 /* Output the first part of the split range: the part that
19939 * doesn't have printables, with the parameter set to not look
19940 * for literals (otherwise we would infinitely recurse) */
19941 put_range(sv, start, temp_end - 1, FALSE);
19943 /* The 2nd part of the range (if any) starts here. */
19946 /* We do a continue, instead of dropping down, because even if
19947 * the 2nd part is non-empty, it could be so short that we want
19948 * to output it as individual characters, as tested for at the
19949 * top of this loop. */
19953 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
19954 * output a sub-range of just the digits or letters, then process
19955 * the remaining portion as usual. */
19956 if (isALPHANUMERIC_A(start)) {
19957 UV mask = (isDIGIT_A(start))
19962 UV temp_end = start + 1;
19964 /* Find the end of the sub-range that includes just the
19965 * characters in the same class as the first character in it */
19966 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
19971 /* For short ranges, don't duplicate the code above to output
19972 * them; just call recursively */
19973 if (temp_end - start < min_range_count) {
19974 put_range(sv, start, temp_end, FALSE);
19976 else { /* Output as a range */
19977 put_code_point(sv, start);
19978 sv_catpvs(sv, "-");
19979 put_code_point(sv, temp_end);
19981 start = temp_end + 1;
19985 /* We output any other printables as individual characters */
19986 if (isPUNCT_A(start) || isSPACE_A(start)) {
19987 while (start <= end && (isPUNCT_A(start)
19988 || isSPACE_A(start)))
19990 put_code_point(sv, start);
19995 } /* End of looking for literals */
19997 /* Here is not to output as a literal. Some control characters have
19998 * mnemonic names. Split off any of those at the beginning and end of
19999 * the range to print mnemonically. It isn't possible for many of
20000 * these to be in a row, so this won't overwhelm with output */
20002 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
20004 while (isMNEMONIC_CNTRL(start) && start <= end) {
20005 put_code_point(sv, start);
20009 /* If this didn't take care of the whole range ... */
20010 if (start <= end) {
20012 /* Look backwards from the end to find the final non-mnemonic
20015 while (isMNEMONIC_CNTRL(temp_end)) {
20019 /* And separately output the interior range that doesn't start
20020 * or end with mnemonics */
20021 put_range(sv, start, temp_end, FALSE);
20023 /* Then output the mnemonic trailing controls */
20024 start = temp_end + 1;
20025 while (start <= end) {
20026 put_code_point(sv, start);
20033 /* As a final resort, output the range or subrange as hex. */
20035 this_end = (end < NUM_ANYOF_CODE_POINTS)
20037 : NUM_ANYOF_CODE_POINTS - 1;
20038 #if NUM_ANYOF_CODE_POINTS > 256
20039 format = (this_end < 256)
20040 ? "\\x%02"UVXf"-\\x%02"UVXf""
20041 : "\\x{%04"UVXf"}-\\x{%04"UVXf"}";
20043 format = "\\x%02"UVXf"-\\x%02"UVXf"";
20045 GCC_DIAG_IGNORE(-Wformat-nonliteral);
20046 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
20053 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
20055 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
20059 bool allow_literals = TRUE;
20061 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
20063 /* Generally, it is more readable if printable characters are output as
20064 * literals, but if a range (nearly) spans all of them, it's best to output
20065 * it as a single range. This code will use a single range if all but 2
20066 * ASCII printables are in it */
20067 invlist_iterinit(invlist);
20068 while (invlist_iternext(invlist, &start, &end)) {
20070 /* If the range starts beyond the final printable, it doesn't have any
20072 if (start > MAX_PRINT_A) {
20076 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
20077 * all but two, the range must start and end no later than 2 from
20079 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
20080 if (end > MAX_PRINT_A) {
20086 if (end - start >= MAX_PRINT_A - ' ' - 2) {
20087 allow_literals = FALSE;
20092 invlist_iterfinish(invlist);
20094 /* Here we have figured things out. Output each range */
20095 invlist_iterinit(invlist);
20096 while (invlist_iternext(invlist, &start, &end)) {
20097 if (start >= NUM_ANYOF_CODE_POINTS) {
20100 put_range(sv, start, end, allow_literals);
20102 invlist_iterfinish(invlist);
20108 S_put_charclass_bitmap_innards_common(pTHX_
20109 SV* invlist, /* The bitmap */
20110 SV* posixes, /* Under /l, things like [:word:], \S */
20111 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
20112 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
20113 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
20114 const bool invert /* Is the result to be inverted? */
20117 /* Create and return an SV containing a displayable version of the bitmap
20118 * and associated information determined by the input parameters. If the
20119 * output would have been only the inversion indicator '^', NULL is instead
20124 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
20127 output = newSVpvs("^");
20130 output = newSVpvs("");
20133 /* First, the code points in the bitmap that are unconditionally there */
20134 put_charclass_bitmap_innards_invlist(output, invlist);
20136 /* Traditionally, these have been placed after the main code points */
20138 sv_catsv(output, posixes);
20141 if (only_utf8 && _invlist_len(only_utf8)) {
20142 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
20143 put_charclass_bitmap_innards_invlist(output, only_utf8);
20146 if (not_utf8 && _invlist_len(not_utf8)) {
20147 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
20148 put_charclass_bitmap_innards_invlist(output, not_utf8);
20151 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
20152 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
20153 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
20155 /* This is the only list in this routine that can legally contain code
20156 * points outside the bitmap range. The call just above to
20157 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
20158 * output them here. There's about a half-dozen possible, and none in
20159 * contiguous ranges longer than 2 */
20160 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20162 SV* above_bitmap = NULL;
20164 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
20166 invlist_iterinit(above_bitmap);
20167 while (invlist_iternext(above_bitmap, &start, &end)) {
20170 for (i = start; i <= end; i++) {
20171 put_code_point(output, i);
20174 invlist_iterfinish(above_bitmap);
20175 SvREFCNT_dec_NN(above_bitmap);
20179 if (invert && SvCUR(output) == 1) {
20187 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
20189 SV *nonbitmap_invlist,
20190 SV *only_utf8_locale_invlist,
20191 const regnode * const node,
20192 const bool force_as_is_display)
20194 /* Appends to 'sv' a displayable version of the innards of the bracketed
20195 * character class defined by the other arguments:
20196 * 'bitmap' points to the bitmap.
20197 * 'nonbitmap_invlist' is an inversion list of the code points that are in
20198 * the bitmap range, but for some reason aren't in the bitmap; NULL if
20199 * none. The reasons for this could be that they require some
20200 * condition such as the target string being or not being in UTF-8
20201 * (under /d), or because they came from a user-defined property that
20202 * was not resolved at the time of the regex compilation (under /u)
20203 * 'only_utf8_locale_invlist' is an inversion list of the code points that
20204 * are valid only if the runtime locale is a UTF-8 one; NULL if none
20205 * 'node' is the regex pattern node. It is needed only when the above two
20206 * parameters are not null, and is passed so that this routine can
20207 * tease apart the various reasons for them.
20208 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
20209 * to invert things to see if that leads to a cleaner display. If
20210 * FALSE, this routine is free to use its judgment about doing this.
20212 * It returns TRUE if there was actually something output. (It may be that
20213 * the bitmap, etc is empty.)
20215 * When called for outputting the bitmap of a non-ANYOF node, just pass the
20216 * bitmap, with the succeeding parameters set to NULL, and the final one to
20220 /* In general, it tries to display the 'cleanest' representation of the
20221 * innards, choosing whether to display them inverted or not, regardless of
20222 * whether the class itself is to be inverted. However, there are some
20223 * cases where it can't try inverting, as what actually matches isn't known
20224 * until runtime, and hence the inversion isn't either. */
20225 bool inverting_allowed = ! force_as_is_display;
20228 STRLEN orig_sv_cur = SvCUR(sv);
20230 SV* invlist; /* Inversion list we accumulate of code points that
20231 are unconditionally matched */
20232 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
20234 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
20236 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
20237 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
20240 SV* as_is_display; /* The output string when we take the inputs
20242 SV* inverted_display; /* The output string when we invert the inputs */
20244 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
20246 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
20248 /* We are biased in favor of displaying things without them being inverted,
20249 * as that is generally easier to understand */
20250 const int bias = 5;
20252 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
20254 /* Start off with whatever code points are passed in. (We clone, so we
20255 * don't change the caller's list) */
20256 if (nonbitmap_invlist) {
20257 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
20258 invlist = invlist_clone(nonbitmap_invlist);
20260 else { /* Worst case size is every other code point is matched */
20261 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
20265 if (OP(node) == ANYOFD) {
20267 /* This flag indicates that the code points below 0x100 in the
20268 * nonbitmap list are precisely the ones that match only when the
20269 * target is UTF-8 (they should all be non-ASCII). */
20270 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
20272 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
20273 _invlist_subtract(invlist, only_utf8, &invlist);
20276 /* And this flag for matching all non-ASCII 0xFF and below */
20277 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
20279 not_utf8 = invlist_clone(PL_UpperLatin1);
20282 else if (OP(node) == ANYOFL) {
20284 /* If either of these flags are set, what matches isn't
20285 * determinable except during execution, so don't know enough here
20287 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
20288 inverting_allowed = FALSE;
20291 /* What the posix classes match also varies at runtime, so these
20292 * will be output symbolically. */
20293 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
20296 posixes = newSVpvs("");
20297 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
20298 if (ANYOF_POSIXL_TEST(node,i)) {
20299 sv_catpv(posixes, anyofs[i]);
20306 /* Accumulate the bit map into the unconditional match list */
20307 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
20308 if (BITMAP_TEST(bitmap, i)) {
20310 for (; i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i); i++) {
20313 invlist = _add_range_to_invlist(invlist, start, i-1);
20317 /* Make sure that the conditional match lists don't have anything in them
20318 * that match unconditionally; otherwise the output is quite confusing.
20319 * This could happen if the code that populates these misses some
20322 _invlist_subtract(only_utf8, invlist, &only_utf8);
20325 _invlist_subtract(not_utf8, invlist, ¬_utf8);
20328 if (only_utf8_locale_invlist) {
20330 /* Since this list is passed in, we have to make a copy before
20332 only_utf8_locale = invlist_clone(only_utf8_locale_invlist);
20334 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
20336 /* And, it can get really weird for us to try outputting an inverted
20337 * form of this list when it has things above the bitmap, so don't even
20339 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20340 inverting_allowed = FALSE;
20344 /* Calculate what the output would be if we take the input as-is */
20345 as_is_display = put_charclass_bitmap_innards_common(invlist,
20352 /* If have to take the output as-is, just do that */
20353 if (! inverting_allowed) {
20354 if (as_is_display) {
20355 sv_catsv(sv, as_is_display);
20356 SvREFCNT_dec_NN(as_is_display);
20359 else { /* But otherwise, create the output again on the inverted input, and
20360 use whichever version is shorter */
20362 int inverted_bias, as_is_bias;
20364 /* We will apply our bias to whichever of the the results doesn't have
20374 inverted_bias = bias;
20377 /* Now invert each of the lists that contribute to the output,
20378 * excluding from the result things outside the possible range */
20380 /* For the unconditional inversion list, we have to add in all the
20381 * conditional code points, so that when inverted, they will be gone
20383 _invlist_union(only_utf8, invlist, &invlist);
20384 _invlist_union(not_utf8, invlist, &invlist);
20385 _invlist_union(only_utf8_locale, invlist, &invlist);
20386 _invlist_invert(invlist);
20387 _invlist_intersection(invlist, PL_InBitmap, &invlist);
20390 _invlist_invert(only_utf8);
20391 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
20393 else if (not_utf8) {
20395 /* If a code point matches iff the target string is not in UTF-8,
20396 * then complementing the result has it not match iff not in UTF-8,
20397 * which is the same thing as matching iff it is UTF-8. */
20398 only_utf8 = not_utf8;
20402 if (only_utf8_locale) {
20403 _invlist_invert(only_utf8_locale);
20404 _invlist_intersection(only_utf8_locale,
20406 &only_utf8_locale);
20409 inverted_display = put_charclass_bitmap_innards_common(
20414 only_utf8_locale, invert);
20416 /* Use the shortest representation, taking into account our bias
20417 * against showing it inverted */
20418 if ( inverted_display
20419 && ( ! as_is_display
20420 || ( SvCUR(inverted_display) + inverted_bias
20421 < SvCUR(as_is_display) + as_is_bias)))
20423 sv_catsv(sv, inverted_display);
20425 else if (as_is_display) {
20426 sv_catsv(sv, as_is_display);
20429 SvREFCNT_dec(as_is_display);
20430 SvREFCNT_dec(inverted_display);
20433 SvREFCNT_dec_NN(invlist);
20434 SvREFCNT_dec(only_utf8);
20435 SvREFCNT_dec(not_utf8);
20436 SvREFCNT_dec(posixes);
20437 SvREFCNT_dec(only_utf8_locale);
20439 return SvCUR(sv) > orig_sv_cur;
20442 #define CLEAR_OPTSTART \
20443 if (optstart) STMT_START { \
20444 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
20445 " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
20449 #define DUMPUNTIL(b,e) \
20451 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
20453 STATIC const regnode *
20454 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
20455 const regnode *last, const regnode *plast,
20456 SV* sv, I32 indent, U32 depth)
20458 U8 op = PSEUDO; /* Arbitrary non-END op. */
20459 const regnode *next;
20460 const regnode *optstart= NULL;
20462 RXi_GET_DECL(r,ri);
20463 GET_RE_DEBUG_FLAGS_DECL;
20465 PERL_ARGS_ASSERT_DUMPUNTIL;
20467 #ifdef DEBUG_DUMPUNTIL
20468 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n",indent,node-start,
20469 last ? last-start : 0,plast ? plast-start : 0);
20472 if (plast && plast < last)
20475 while (PL_regkind[op] != END && (!last || node < last)) {
20477 /* While that wasn't END last time... */
20480 if (op == CLOSE || op == WHILEM)
20482 next = regnext((regnode *)node);
20485 if (OP(node) == OPTIMIZED) {
20486 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
20493 regprop(r, sv, node, NULL, NULL);
20494 Perl_re_printf( aTHX_ "%4"IVdf":%*s%s", (IV)(node - start),
20495 (int)(2*indent + 1), "", SvPVX_const(sv));
20497 if (OP(node) != OPTIMIZED) {
20498 if (next == NULL) /* Next ptr. */
20499 Perl_re_printf( aTHX_ " (0)");
20500 else if (PL_regkind[(U8)op] == BRANCH
20501 && PL_regkind[OP(next)] != BRANCH )
20502 Perl_re_printf( aTHX_ " (FAIL)");
20504 Perl_re_printf( aTHX_ " (%"IVdf")", (IV)(next - start));
20505 Perl_re_printf( aTHX_ "\n");
20509 if (PL_regkind[(U8)op] == BRANCHJ) {
20512 const regnode *nnode = (OP(next) == LONGJMP
20513 ? regnext((regnode *)next)
20515 if (last && nnode > last)
20517 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
20520 else if (PL_regkind[(U8)op] == BRANCH) {
20522 DUMPUNTIL(NEXTOPER(node), next);
20524 else if ( PL_regkind[(U8)op] == TRIE ) {
20525 const regnode *this_trie = node;
20526 const char op = OP(node);
20527 const U32 n = ARG(node);
20528 const reg_ac_data * const ac = op>=AHOCORASICK ?
20529 (reg_ac_data *)ri->data->data[n] :
20531 const reg_trie_data * const trie =
20532 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
20534 AV *const trie_words
20535 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
20537 const regnode *nextbranch= NULL;
20540 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
20541 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
20543 Perl_re_indentf( aTHX_ "%s ",
20546 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
20547 SvCUR(*elem_ptr), 60,
20548 PL_colors[0], PL_colors[1],
20550 ? PERL_PV_ESCAPE_UNI
20552 | PERL_PV_PRETTY_ELLIPSES
20553 | PERL_PV_PRETTY_LTGT
20558 U16 dist= trie->jump[word_idx+1];
20559 Perl_re_printf( aTHX_ "(%"UVuf")\n",
20560 (UV)((dist ? this_trie + dist : next) - start));
20563 nextbranch= this_trie + trie->jump[0];
20564 DUMPUNTIL(this_trie + dist, nextbranch);
20566 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
20567 nextbranch= regnext((regnode *)nextbranch);
20569 Perl_re_printf( aTHX_ "\n");
20572 if (last && next > last)
20577 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
20578 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
20579 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
20581 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
20583 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
20585 else if ( op == PLUS || op == STAR) {
20586 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
20588 else if (PL_regkind[(U8)op] == ANYOF) {
20589 /* arglen 1 + class block */
20590 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
20591 ? ANYOF_POSIXL_SKIP
20593 node = NEXTOPER(node);
20595 else if (PL_regkind[(U8)op] == EXACT) {
20596 /* Literal string, where present. */
20597 node += NODE_SZ_STR(node) - 1;
20598 node = NEXTOPER(node);
20601 node = NEXTOPER(node);
20602 node += regarglen[(U8)op];
20604 if (op == CURLYX || op == OPEN)
20608 #ifdef DEBUG_DUMPUNTIL
20609 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
20614 #endif /* DEBUGGING */
20617 * ex: set ts=8 sts=4 sw=4 et: